var __defProp = Object.defineProperty;
var __require = /* @__PURE__ */ ((x) => typeof require !== "undefined" ? require : typeof Proxy !== "undefined" ? new Proxy(x, {
  get: (a, b) => (typeof require !== "undefined" ? require : a)[b]
}) : x)(function(x) {
  if (typeof require !== "undefined") return require.apply(this, arguments);
  throw Error('Dynamic require of "' + x + '" is not supported');
});
var __export = (target, all) => {
  for (var name in all)
    __defProp(target, name, { get: all[name], enumerable: true });
};

// src/env.js
import fs from "fs";
import path from "path";
import url from "url";
var VERSION = "4.2.0";
var HAS_SELF = typeof self !== "undefined";
var IS_FS_AVAILABLE = !isEmpty(fs);
var IS_PATH_AVAILABLE = !isEmpty(path);
var IS_WEB_CACHE_AVAILABLE = HAS_SELF && "caches" in self;
var IS_DENO_RUNTIME = typeof globalThis.Deno !== "undefined";
var IS_BUN_RUNTIME = typeof globalThis.Bun !== "undefined";
var IS_DENO_WEB_RUNTIME = IS_DENO_RUNTIME && IS_WEB_CACHE_AVAILABLE && !IS_FS_AVAILABLE;
var IS_PROCESS_AVAILABLE = typeof process !== "undefined";
var IS_NODE_ENV = IS_PROCESS_AVAILABLE && process?.release?.name === "node" && !IS_DENO_WEB_RUNTIME;
var IS_BROWSER_ENV = typeof window !== "undefined" && typeof window.document !== "undefined";
var IS_WEBWORKER_ENV = HAS_SELF && ["DedicatedWorkerGlobalScope", "ServiceWorkerGlobalScope", "SharedWorkerGlobalScope"].includes(
  self.constructor?.name
);
var IS_WEB_ENV = IS_BROWSER_ENV || IS_WEBWORKER_ENV || IS_DENO_WEB_RUNTIME;
var IS_WEBGPU_AVAILABLE = IS_NODE_ENV || typeof navigator !== "undefined" && "gpu" in navigator;
var IS_WEBNN_AVAILABLE = typeof navigator !== "undefined" && "ml" in navigator;
var IS_CRYPTO_AVAILABLE = typeof crypto !== "undefined" && typeof crypto.getRandomValues === "function";
var IS_CHROME_AVAILABLE = (
  // @ts-ignore - chrome may not exist in all environments
  typeof chrome !== "undefined" && typeof chrome.runtime !== "undefined" && typeof chrome.runtime.id === "string"
);
var IS_SERVICE_WORKER_ENV = (
  // @ts-ignore - ServiceWorkerGlobalScope may not exist in all environments
  typeof ServiceWorkerGlobalScope !== "undefined" && HAS_SELF && self instanceof ServiceWorkerGlobalScope
);
var isSafari = () => {
  if (typeof navigator === "undefined") {
    return false;
  }
  const userAgent = navigator.userAgent;
  const vendor = navigator.vendor || "";
  const isAppleVendor = vendor.indexOf("Apple") > -1;
  const notOtherBrowser = !userAgent.match(/CriOS|FxiOS|EdgiOS|OPiOS|mercury|brave/i) && !userAgent.includes("Chrome") && !userAgent.includes("Android");
  return isAppleVendor && notOtherBrowser;
};
var IS_SAFARI = isSafari();
var apis = Object.freeze({
  /** Whether we are running in a browser environment (and not a web worker) */
  IS_BROWSER_ENV,
  /** Whether we are running in a web worker environment */
  IS_WEBWORKER_ENV,
  /** Whether we are running in a web-like environment (browser, web worker, or Deno web runtime) */
  IS_WEB_ENV,
  /** Whether we are running in a service worker environment */
  IS_SERVICE_WORKER_ENV,
  /** Whether we are running in Deno's web runtime (CDN imports, Cache API available, no filesystem) */
  IS_DENO_WEB_RUNTIME,
  /** Whether the Cache API is available */
  IS_WEB_CACHE_AVAILABLE,
  /** Whether the WebGPU API is available */
  IS_WEBGPU_AVAILABLE,
  /** Whether the WebNN API is available */
  IS_WEBNN_AVAILABLE,
  /** Whether we are running in a Safari browser */
  IS_SAFARI,
  /** Whether the Node.js process API is available */
  IS_PROCESS_AVAILABLE,
  /** Whether we are running in a Node.js-like environment (node, deno, bun) */
  IS_NODE_ENV,
  /** Whether the filesystem API is available */
  IS_FS_AVAILABLE,
  /** Whether the path API is available */
  IS_PATH_AVAILABLE,
  /** Whether the crypto API is available */
  IS_CRYPTO_AVAILABLE,
  /** Whether the Chrome runtime API is available */
  IS_CHROME_AVAILABLE
});
var RUNNING_LOCALLY = IS_FS_AVAILABLE && IS_PATH_AVAILABLE;
var dirname__ = "./";
if (RUNNING_LOCALLY) {
  const _import_meta_url = Object(import.meta).url;
  if (_import_meta_url) {
    dirname__ = path.dirname(path.dirname(url.fileURLToPath(_import_meta_url)));
  } else if (typeof __dirname !== "undefined") {
    dirname__ = path.dirname(__dirname);
  }
}
var DEFAULT_CACHE_DIR = RUNNING_LOCALLY ? path.join(dirname__, "/.cache/") : null;
var DEFAULT_LOCAL_MODEL_PATH = "/models/";
var localModelPath = RUNNING_LOCALLY ? path.join(dirname__, DEFAULT_LOCAL_MODEL_PATH) : DEFAULT_LOCAL_MODEL_PATH;
var DEFAULT_FETCH = typeof globalThis.fetch === "function" ? globalThis.fetch.bind(globalThis) : void 0;
var LogLevel = Object.freeze({
  /** All messages including debug output (value: 10) */
  DEBUG: 10,
  /** Errors, warnings, and info messages (value: 20) */
  INFO: 20,
  /** Errors and warnings (value: 30) */
  WARNING: 30,
  /** Only error messages (value: 40) */
  ERROR: 40,
  /** No logging output (value: 50) */
  NONE: 50
});
var logLevel = LogLevel.WARNING;
var env = {
  version: VERSION,
  /////////////////// Backends settings ///////////////////
  // NOTE: These will be populated later by the backends themselves.
  backends: {
    // onnxruntime-web/onnxruntime-node
    onnx: {}
  },
  /////////////////// Logging settings ///////////////////
  get logLevel() {
    return logLevel;
  },
  set logLevel(level) {
    logLevel = level;
    env.backends.onnx?.setLogLevel?.(level);
  },
  /////////////////// Model settings ///////////////////
  allowRemoteModels: true,
  remoteHost: "https://huggingface.co/",
  remotePathTemplate: "{model}/resolve/{revision}/",
  allowLocalModels: !(IS_BROWSER_ENV || IS_WEBWORKER_ENV || IS_DENO_WEB_RUNTIME),
  // Default to true for non-web environments, false for web environments
  localModelPath,
  useFS: IS_FS_AVAILABLE,
  /////////////////// Cache settings ///////////////////
  useBrowserCache: IS_WEB_CACHE_AVAILABLE,
  useFSCache: IS_FS_AVAILABLE,
  cacheDir: DEFAULT_CACHE_DIR,
  useCustomCache: false,
  customCache: null,
  useWasmCache: IS_WEB_CACHE_AVAILABLE || IS_FS_AVAILABLE,
  cacheKey: "transformers-cache",
  experimental_useCrossOriginStorage: false,
  /////////////////// Custom fetch /////////////////////
  fetch: DEFAULT_FETCH
  //////////////////////////////////////////////////////
};
function isEmpty(obj) {
  return Object.keys(obj).length === 0;
}

// src/utils/generic.js
var Callable = (
  /** @type {any} */
  class {
    /**
     * Creates a new instance of the Callable class.
     */
    constructor() {
      let closure = function(...args) {
        return closure._call(...args);
      };
      return Object.setPrototypeOf(closure, new.target.prototype);
    }
    /**
     * This method should be implemented in subclasses to provide the
     * functionality of the callable object.
     *
     * @param {any[]} args
     * @throws {Error} If the subclass does not implement the `_call` method.
     */
    _call(...args) {
      throw Error("Must implement _call method in subclass");
    }
  }
);

// src/utils/core.js
function dispatchCallback(progress_callback, data) {
  if (progress_callback) progress_callback(data);
}
var DefaultProgressCallback = class extends Callable {
  /**
   * @param {ProgressCallback} callback The original callback.
   * @param {FilesLoadingMap} files_loading Mutable map storing per-file progress.
   */
  constructor(callback, files_loading) {
    super();
    this.callback = callback;
    this.files_loading = files_loading;
  }
  /**
   * @param {ProgressInfo} info
   */
  _call(info) {
    if (info.status === "progress") {
      this.files_loading[info.file] = {
        loaded: info.loaded,
        total: info.total
      };
      const loaded = Object.values(this.files_loading).reduce((acc, curr) => acc + curr.loaded, 0);
      const total = Object.values(this.files_loading).reduce((acc, curr) => acc + curr.total, 0);
      const progress = total > 0 ? loaded / total * 100 : 0;
      this.callback({
        status: "progress_total",
        name: info.name,
        progress,
        loaded,
        total,
        files: structuredClone(this.files_loading)
      });
    }
    this.callback(info);
  }
};
function isIntegralNumber(x) {
  return Number.isInteger(x) || typeof x === "bigint";
}
function isNullishDimension(x) {
  return x === null || x === void 0 || x === -1;
}
function calculateDimensions(arr) {
  const dimensions = [];
  let current = arr;
  while (Array.isArray(current)) {
    dimensions.push(current.length);
    current = current[0];
  }
  return dimensions;
}
function mergeArrays(...arrs) {
  return Array.prototype.concat.apply([], arrs);
}
function product(...a) {
  return a.reduce((a2, b) => a2.flatMap((d) => b.map((e) => [d, e])));
}
function calculateReflectOffset(i, w) {
  return Math.abs((i + w) % (2 * w) - w);
}
function pick(o, props) {
  return Object.assign(
    {},
    ...props.map((prop) => {
      if (o[prop] !== void 0) {
        return { [prop]: o[prop] };
      }
    })
  );
}
function count(arr, value) {
  let count2 = 0;
  for (const v of arr) {
    if (v === value) ++count2;
  }
  return count2;
}

// src/utils/logger.js
var logger = {
  /**
   * Log an error message. Only suppressed when logLevel is NONE.
   * @param  {...any} args - Arguments to log
   */
  error(...args) {
    if (env.logLevel <= LogLevel.ERROR) {
      console.error(...args);
    }
  },
  /**
   * Log a warning message. Shown when logLevel <= WARNING.
   * @param  {...any} args - Arguments to log
   */
  warn(...args) {
    if (env.logLevel <= LogLevel.WARNING) {
      console.warn(...args);
    }
  },
  /**
   * Log an info message. Shown when logLevel <= INFO.
   * @param  {...any} args - Arguments to log
   */
  info(...args) {
    if (env.logLevel <= LogLevel.INFO) {
      console.log(...args);
    }
  },
  /**
   * Log a debug message. Only shown when logLevel is DEBUG.
   * @param  {...any} args - Arguments to log
   */
  debug(...args) {
    if (env.logLevel <= LogLevel.DEBUG) {
      console.log(...args);
    }
  },
  /**
   * Log a message (alias for info). Shown when logLevel <= INFO.
   * @param  {...any} args - Arguments to log
   */
  log(...args) {
    this.info(...args);
  }
};

// ../../node_modules/.pnpm/@huggingface+tokenizers@0.1.3/node_modules/@huggingface/tokenizers/dist/tokenizers.mjs
var DictionarySplitter = class {
  /**
   * @param dictionary The dictionary of words to use for splitting.
   */
  constructor(dictionary) {
    this.trie = this._build_trie(dictionary);
  }
  /**
   * Builds a trie from the given dictionary.
   * @param dictionary The dictionary of words to build the trie from.
   * @returns The root node of the trie.
   * @private
   */
  _build_trie(dictionary) {
    const trie = /* @__PURE__ */ Object.create(null);
    for (const word of dictionary) {
      let node = trie;
      for (let i = 0; i < word.length; ++i) {
        const char = word[i];
        node = node[char] ??= /* @__PURE__ */ Object.create(null);
      }
      node.end = word;
    }
    return trie;
  }
  /**
   * Splits the input text into tokens based on the dictionary.
   * @param text The input text to split.
   * @returns An array of tokens.
   */
  split(text) {
    const result = [];
    const n = text.length;
    let start = 0;
    let i = 0;
    while (i < n) {
      let node = this.trie;
      let match = null;
      let j = i;
      while (j < n && (node = node[text[j]])) {
        if (node.end) {
          match = node.end;
        }
        ++j;
      }
      if (match) {
        if (i > start) {
          result.push(text.slice(start, i));
        }
        result.push(match);
        i += match.length;
        start = i;
      } else {
        ++i;
      }
    }
    if (start < n) {
      result.push(text.slice(start));
    }
    return result;
  }
};
var DictionarySplitter_default = DictionarySplitter;
var AddedToken = class {
  /**
   * Creates a new instance of AddedToken.
   * @param config Added token configuration object.
   */
  constructor(config) {
    this.content = config.content;
    this.id = config.id;
    this.single_word = config.single_word ?? false;
    this.lstrip = config.lstrip ?? false;
    this.rstrip = config.rstrip ?? false;
    this.special = config.special ?? false;
    this.normalized = config.normalized ?? !this.special;
  }
};
var AddedToken_default = AddedToken;
var BYTES_TO_UNICODE = (() => {
  const bs2 = [
    ...Array.from(
      { length: "~".charCodeAt(0) - "!".charCodeAt(0) + 1 },
      (_, i) => i + "!".charCodeAt(0)
    ),
    ...Array.from(
      { length: "\xAC".charCodeAt(0) - "\xA1".charCodeAt(0) + 1 },
      (_, i) => i + "\xA1".charCodeAt(0)
    ),
    ...Array.from(
      { length: "\xFF".charCodeAt(0) - "\xAE".charCodeAt(0) + 1 },
      (_, i) => i + "\xAE".charCodeAt(0)
    )
  ];
  const cs2 = bs2.slice();
  let n = 0;
  for (let b = 0; b < 256; ++b) {
    if (!bs2.includes(b)) {
      bs2.push(b);
      cs2.push(256 + n);
      n += 1;
    }
  }
  const ccs = cs2.map((n2) => String.fromCharCode(n2));
  return Object.fromEntries(bs2.map((b, i) => [b, ccs[i]]));
})();
var reverse_dictionary = (data) => Object.fromEntries(Object.entries(data).map(([key, value]) => [value, key]));
var UNICODE_TO_BYTES = reverse_dictionary(BYTES_TO_UNICODE);
var BLOOM_SPLIT_CHARS = ".,!?\u2026\u3002\uFF0C\u3001\u0964\u06D4\u060C";
var PROBLEMATIC_REGEX_MAP = /* @__PURE__ */ new Map([
  // These uses the case insensitive group modifier, which is not supported in JavaScript.
  // When parsing the regex, an "Invalid group" error is thrown.
  [
    "(?i:'s|'t|'re|'ve|'m|'ll|'d)",
    "(?:'([sS]|[tT]|[rR][eE]|[vV][eE]|[mM]|[lL][lL]|[dD]))"
  ],
  [
    "(?i:[sdmt]|ll|ve|re)",
    "(?:[sS]|[dD]|[mM]|[tT]|[lL][lL]|[vV][eE]|[rR][eE])"
  ],
  // JS doesn't support possessive quantifiers (these are used in recent OpenAI tokenizers).
  ["[^\\r\\n\\p{L}\\p{N}]?+", "[^\\r\\n\\p{L}\\p{N}]?"],
  ["[^\\s\\p{L}\\p{N}]++", "[^\\s\\p{L}\\p{N}]+"],
  // JS doesn't support atomic groups (these are used in AFMoE tokenizers).
  ["(?>\\p{Nd}{510})", "(?:\\p{Nd}{510})"],
  // JS doesn't support stacking quantifiers.
  // Uncaught SyntaxError: Invalid regular expression: /\p{Nd}{3}+/u: Nothing to repeat
  ["\\p{Nd}{3}+", "(?:\\p{Nd}{3})+"],
  // \G is an invalid escape in JS, and in most cases is just used as an optimization.
  // So, we can safely remove it.
  ["\\G", ""],
  // Used to override the default (invalid) regex of the bloom pretokenizer.
  // For more information, see https://github.com/huggingface/transformers.js/issues/94
  [` ?[^(\\s|[${BLOOM_SPLIT_CHARS}])]+`, ` ?[^\\s${BLOOM_SPLIT_CHARS}]+`]
]);
var PUNCTUATION_REGEX = "\\p{P}\\u0021-\\u002F\\u003A-\\u0040\\u005B-\\u0060\\u007B-\\u007E";
var clean_up_tokenization = (text) => text.replace(/ \./g, ".").replace(/ \?/g, "?").replace(/ \!/g, "!").replace(/ ,/g, ",").replace(/ \' /g, "'").replace(/ n't/g, "n't").replace(/ 'm/g, "'m").replace(/ 's/g, "'s").replace(/ 've/g, "'ve").replace(/ 're/g, "'re");
var create_pattern = (pattern, invert = true) => {
  if (pattern.Regex !== void 0) {
    let regex = pattern.Regex.replace(/\\([#&~])/g, "$1");
    regex = regex.replace(/\\A/g, "^").replace(/\\z/g, "$").replace(/\\Z/g, "(?=\\r?\\n?$)");
    for (const [key, value] of PROBLEMATIC_REGEX_MAP) {
      regex = regex.replaceAll(key, value);
    }
    try {
      return new RegExp(regex, "gu");
    } catch (error) {
      if (!(error instanceof SyntaxError) || !error.message.toLowerCase().includes("invalid property name"))
        throw error;
      let changed = false;
      const fixed = regex.replace(/(\\[pP])\{([^}=]+)\}/g, (_, p, n) => {
        try {
          new RegExp(`\\p{${n}}`, "u");
          return `${p}{${n}}`;
        } catch {
          changed = true;
          return `${p}{Script=${n}}`;
        }
      });
      if (!changed) throw error;
      try {
        return new RegExp(fixed, "gu");
      } catch (e) {
        throw error;
      }
    }
  } else if (pattern.String !== void 0) {
    const escaped = escape_reg_exp(pattern.String);
    return new RegExp(invert ? escaped : `(${escaped})`, "gu");
  } else {
    console.warn("Unknown pattern type:", pattern);
    return null;
  }
};
var escape_reg_exp = (string) => string.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
var fuse_unk = (arr, tokens_to_ids, unk_token_id) => {
  const fused = [];
  let i = 0;
  while (i < arr.length) {
    fused.push(arr[i]);
    const token_id = tokens_to_ids.get(arr[i]) ?? unk_token_id;
    if (token_id !== unk_token_id) {
      ++i;
      continue;
    }
    while (++i < arr.length && (tokens_to_ids.get(arr[i]) ?? unk_token_id) === unk_token_id) {
      if (tokens_to_ids.get(fused.at(-1)) !== unk_token_id) {
        fused[fused.length - 1] += arr[i];
      }
    }
  }
  return fused;
};
var is_chinese_char = (cp) => cp >= 19968 && cp <= 40959 || cp >= 13312 && cp <= 19903 || cp >= 131072 && cp <= 173791 || cp >= 173824 && cp <= 177983 || cp >= 177984 && cp <= 178207 || cp >= 178208 && cp <= 183983 || cp >= 63744 && cp <= 64255 || cp >= 194560 && cp <= 195103;
var is_integral_number = (x) => Number.isInteger(x) || typeof x === "bigint";
var len = (s) => {
  let length = 0;
  for (const c of s) ++length;
  return length;
};
var lowercase_and_remove_accents = (text) => remove_accents(text.toLowerCase());
var merge_arrays = (...arrs) => Array.prototype.concat.apply([], arrs);
var object_to_map = (obj) => new Map(Object.entries(obj));
var regex_split = (text, regex) => {
  const result = [];
  let prev = 0;
  for (const match of text.matchAll(regex)) {
    const full_match = match[0];
    if (prev < match.index) {
      result.push(text.slice(prev, match.index));
    }
    if (full_match.length > 0) {
      result.push(full_match);
    }
    prev = match.index + full_match.length;
  }
  if (prev < text.length) {
    result.push(text.slice(prev));
  }
  return result;
};
var remove_accents = (text) => text.replace(/\p{M}/gu, "");
var validate_object = (obj, name, required_keys = []) => {
  if (!obj || Array.isArray(obj) || typeof obj !== "object") {
    return `${name} must be a valid object`;
  }
  for (const key of required_keys) {
    if (!(key in obj)) {
      return `${name} must contain a "${key}" property`;
    }
  }
  return null;
};
var whitespace_split = (text) => text.match(/\S+/g) || [];
var Callable2 = class {
  /**
   * Creates a new instance of the Callable class.
   */
  constructor() {
    const closure = function(...args) {
      return closure._call(...args);
    };
    return Object.setPrototypeOf(closure, new.target.prototype);
  }
};
var Callable_default = Callable2;
var Normalizer = class extends Callable_default {
  /**
   * @param config The configuration object for the normalizer.
   */
  constructor(config) {
    super();
    this.config = config;
  }
  /**
   * Alias for {@link Normalizer#normalize}.
   * @param text The text to normalize.
   * @returns The normalized text.
   */
  _call(text) {
    return this.normalize(text);
  }
};
var Normalizer_default = Normalizer;
var BertNormalizer = class extends Normalizer_default {
  /**
   * Adds whitespace around any CJK (Chinese, Japanese, or Korean) character in the input text.
   *
   * @param text The input text to tokenize.
   * @returns The tokenized text with whitespace added around CJK characters.
   */
  tokenize_chinese_chars(text) {
    const output = [];
    for (let i = 0; i < text.length; ++i) {
      const char = text[i];
      const cp = char.charCodeAt(0);
      if (is_chinese_char(cp)) {
        output.push(" ");
        output.push(char);
        output.push(" ");
      } else {
        output.push(char);
      }
    }
    return output.join("");
  }
  /**
   * Strips accents from the given text.
   * @param text The text to strip accents from.
   * @returns The text with accents removed.
   */
  strip_accents(text) {
    return text.normalize("NFD").replace(/\p{Mn}/gu, "");
  }
  /**
   * Checks whether `char` is a control character.
   * @param char The character to check.
   * @returns Whether `char` is a control character.
   */
  is_control(char) {
    switch (char) {
      case "	":
      case "\n":
      case "\r":
        return false;
      default:
        return /^\p{Cc}|\p{Cf}|\p{Co}|\p{Cs}$/u.test(char);
    }
  }
  /**
   * Performs invalid character removal and whitespace cleanup on text.
   * @param text The text to clean.
   * @returns The cleaned text.
   */
  clean_text(text) {
    const output = [];
    for (const char of text) {
      const cp = char.charCodeAt(0);
      if (cp === 0 || cp === 65533 || this.is_control(char)) {
        continue;
      }
      if (/^\s$/.test(char)) {
        output.push(" ");
      } else {
        output.push(char);
      }
    }
    return output.join("");
  }
  /**
   * Normalizes the given text based on the configuration.
   * @param text The text to normalize.
   * @returns The normalized text.
   */
  normalize(text) {
    if (this.config.clean_text) {
      text = this.clean_text(text);
    }
    if (this.config.handle_chinese_chars) {
      text = this.tokenize_chinese_chars(text);
    }
    if (this.config.lowercase) {
      text = text.toLowerCase();
      if (this.config.strip_accents !== false) {
        text = this.strip_accents(text);
      }
    } else if (this.config.strip_accents) {
      text = this.strip_accents(text);
    }
    return text;
  }
};
var BertNormalizer_default = BertNormalizer;
var Precompiled = class extends Normalizer_default {
  /**
   * Create a new instance of Precompiled normalizer.
   * @param config The configuration object.
   */
  constructor(config) {
    super(config);
    this.charsmap = config.precompiled_charsmap ?? null;
  }
  /**
   * Normalizes the given text by applying the precompiled charsmap.
   * @param text The text to normalize.
   * @returns The normalized text.
   */
  normalize(text) {
    text = text.replace(
      /[\u0001-\u0008\u000B\u000E-\u001F\u007F\u008F\u009F]/gm,
      ""
    );
    text = text.replace(
      /[\u0009\u000A\u000C\u000D\u00A0\u1680\u2000-\u200F\u2028\u2029\u202F\u205F\u2581\u3000\uFEFF\uFFFD]/gm,
      " "
    );
    if (text.includes("\uFF5E")) {
      const parts = text.split("\uFF5E");
      text = parts.map((part) => part.normalize("NFKC")).join("\uFF5E");
    } else {
      text = text.normalize("NFKC");
    }
    return text;
  }
};
var Precompiled_default = Precompiled;
var Sequence = class extends Normalizer_default {
  /**
   * Create a new instance of NormalizerSequence.
   * @param config The configuration object.
   */
  constructor(config) {
    super(config);
    this.normalizers = (config.normalizers ?? []).map(
      (x) => create_normalizer_default(x)
    );
  }
  /**
   * Apply a sequence of Normalizers to the input text.
   * @param text The text to normalize.
   * @returns The normalized text.
   */
  normalize(text) {
    return this.normalizers.reduce((t, normalizer) => {
      return normalizer ? normalizer.normalize(t) : t;
    }, text);
  }
};
var Sequence_default = Sequence;
var Replace = class extends Normalizer_default {
  /**
   * Normalize the input text by replacing the pattern with the content.
   * @param text The input text to be normalized.
   * @returns The normalized text after replacing the pattern with the content.
   */
  normalize(text) {
    const pattern = create_pattern(this.config.pattern ?? {});
    return pattern === null ? text : text.replaceAll(pattern, this.config.content ?? "");
  }
};
var Replace_default = Replace;
var UnicodeNormalizer = class extends Normalizer_default {
  constructor() {
    super(...arguments);
    this.form = "NFC";
  }
  /**
   * Normalize the input text by applying Unicode normalization.
   * @param text The input text to be normalized.
   * @returns The normalized text.
   */
  normalize(text) {
    text = text.normalize(this.form);
    return text;
  }
};
var UnicodeNormalizer_default = UnicodeNormalizer;
var NFC = class extends UnicodeNormalizer_default {
  constructor() {
    super(...arguments);
    this.form = "NFC";
  }
};
var NFC_default = NFC;
var NFD = class extends UnicodeNormalizer_default {
  constructor() {
    super(...arguments);
    this.form = "NFD";
  }
};
var NFD_default = NFD;
var NFKC = class extends UnicodeNormalizer_default {
  constructor() {
    super(...arguments);
    this.form = "NFKC";
  }
};
var NFKC_default = NFKC;
var NFKD = class extends UnicodeNormalizer_default {
  constructor() {
    super(...arguments);
    this.form = "NFKD";
  }
};
var NFKD_default = NFKD;
var Strip = class extends Normalizer_default {
  /**
   * Strip leading and/or trailing whitespace from the input text.
   * @param text The input text.
   * @returns The normalized text.
   */
  normalize(text) {
    if (this.config.strip_left && this.config.strip_right) {
      text = text.trim();
    } else {
      if (this.config.strip_left) {
        text = text.trimStart();
      }
      if (this.config.strip_right) {
        text = text.trimEnd();
      }
    }
    return text;
  }
};
var Strip_default = Strip;
var StripAccents = class extends Normalizer_default {
  /**
   * Remove all accents from the text.
   * @param text The input text.
   * @returns The normalized text without accents.
   */
  normalize(text) {
    return remove_accents(text);
  }
};
var StripAccents_default = StripAccents;
var Lowercase = class extends Normalizer_default {
  /**
   * Lowercases the input string.
   * @param {string} text The text to normalize.
   * @returns {string} The normalized text.
   */
  normalize(text) {
    return text.toLowerCase();
  }
};
var Lowercase_default = Lowercase;
var Prepend = class extends Normalizer_default {
  /**
   * Prepends the input string.
   * @param text The text to normalize.
   * @returns The normalized text.
   */
  normalize(text) {
    text = this.config.prepend + text;
    return text;
  }
};
var Prepend_default = Prepend;
function create_normalizer(config) {
  if (config === null) return null;
  switch (config.type) {
    case "BertNormalizer":
      return new BertNormalizer_default(config);
    case "Precompiled":
      return new Precompiled_default(config);
    case "Sequence":
      return new Sequence_default(config);
    case "Replace":
      return new Replace_default(config);
    case "NFC":
      return new NFC_default(config);
    case "NFD":
      return new NFD_default(config);
    case "NFKC":
      return new NFKC_default(config);
    case "NFKD":
      return new NFKD_default(config);
    case "Strip":
      return new Strip_default(config);
    case "StripAccents":
      return new StripAccents_default(config);
    case "Lowercase":
      return new Lowercase_default(config);
    case "Prepend":
      return new Prepend_default(config);
    default:
      throw new Error(`Unknown Normalizer type: ${config.type}`);
  }
}
var create_normalizer_default = create_normalizer;
var PreTokenizer = class extends Callable_default {
  /**
   * Tokenizes the given text into pre-tokens.
   * @param text The text or array of texts to pre-tokenize.
   * @param options Additional options for the pre-tokenization logic.
   * @returns An array of pre-tokens.
   */
  pre_tokenize(text, options) {
    return (Array.isArray(text) ? text.map((x) => this.pre_tokenize_text(x, options)) : this.pre_tokenize_text(text, options)).flat();
  }
  /**
   * Alias for {@link PreTokenizer#pre_tokenize}.
   * @param text The text or array of texts to pre-tokenize.
   * @param options Additional options for the pre-tokenization logic.
   * @returns An array of pre-tokens.
   */
  _call(text, options) {
    return this.pre_tokenize(text, options);
  }
};
var PreTokenizer_default = PreTokenizer;
var ByteLevel = class extends PreTokenizer_default {
  /**
   * Creates a new instance of the `ByteLevelPreTokenizer` class.
   * @param config The configuration object.
   */
  constructor(config) {
    super();
    this.config = config;
    this.add_prefix_space = this.config.add_prefix_space ?? false;
    this.trim_offsets = this.config.trim_offsets ?? false;
    this.use_regex = this.config.use_regex ?? true;
    this.pattern = /'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+/gu;
    this.byte_encoder = BYTES_TO_UNICODE;
    this.text_encoder = new TextEncoder();
  }
  /**
   * Tokenizes a single piece of text using byte-level tokenization.
   * @param text The text to tokenize.
   * @param options Additional options for the pre-tokenization logic.
   * @returns An array of tokens.
   */
  pre_tokenize_text(text, options) {
    if (this.add_prefix_space && !text.startsWith(" ")) {
      text = " " + text;
    }
    const tokens = this.use_regex ? text.match(this.pattern) || [] : [text];
    return tokens.map(
      (token) => Array.from(
        this.text_encoder.encode(token),
        (byte) => this.byte_encoder[byte]
      ).join("")
    );
  }
};
var ByteLevel_default = ByteLevel;
var Whitespace = class extends PreTokenizer_default {
  /**
   * Pre-tokenizes the input text by splitting it on word boundaries.
   * @param text The text to be pre-tokenized.
   * @param options Additional options for the pre-tokenization logic.
   * @returns An array of tokens produced by splitting the input text on whitespace.
   */
  pre_tokenize_text(text, options) {
    return text.match(/\w+|[^\w\s]+/g) || [];
  }
};
var Whitespace_default = Whitespace;
var Metaspace = class extends PreTokenizer_default {
  /**
   * @param config The configuration object for the MetaspacePreTokenizer.
   */
  constructor(config) {
    super();
    this.replacement = config.replacement ?? "\u2581";
    this.str_rep = config.str_rep || this.replacement;
    this.prepend_scheme = config.prepend_scheme ?? "always";
  }
  /**
   * This method takes a string, replaces spaces with the replacement character,
   * adds a prefix space if requested, and returns a new list of tokens.
   * @param text The text to pre-tokenize.
   * @param options The options for the pre-tokenization.
   * @returns A new list of pre-tokenized tokens.
   */
  pre_tokenize_text(text, options) {
    const { section_index = void 0 } = options ?? {};
    let normalized = text.replaceAll(" ", this.str_rep);
    if (
      // We add a prefix space if:
      //  (1) The normalized token does not already start with the replacement character.
      !normalized.startsWith(this.replacement) && // and (2) either:
      //  (a) prepend_scheme is 'always'
      //  (b) prepend_scheme is 'first' and this is the first section
      (this.prepend_scheme === "always" || this.prepend_scheme === "first" && section_index === 0)
    ) {
      normalized = this.str_rep + normalized;
    }
    return [normalized];
  }
};
var Metaspace_default = Metaspace;
var Split = class extends PreTokenizer_default {
  /**
   * @param config The configuration options for the pre-tokenizer.
   */
  constructor(config) {
    super();
    this.config = config;
    this.pattern = create_pattern(
      this.config.pattern ?? {},
      this.config.invert ?? true
    );
  }
  /**
   * Tokenizes text by splitting it using the given pattern.
   * @param text The text to tokenize.
   * @returns An array of tokens.
   */
  pre_tokenize_text(text) {
    if (this.pattern === null) {
      return [];
    }
    if (this.config.invert) {
      return text.match(this.pattern) || [];
    } else if (this.config.behavior?.toLowerCase() === "removed") {
      return text.split(this.pattern).filter((x) => x);
    } else {
      return regex_split(text, this.pattern);
    }
  }
};
var Split_default = Split;
var Punctuation = class extends PreTokenizer_default {
  /**
   * @param config The configuration options for the pre-tokenizer.
   */
  constructor(config) {
    super();
    this.config = config;
    this.pattern = new RegExp(
      `[^${PUNCTUATION_REGEX}]+|[${PUNCTUATION_REGEX}]+`,
      "gu"
    );
  }
  /**
   * Tokenizes text by splitting it using the given pattern.
   * @param text The text to tokenize.
   * @returns An array of tokens.
   */
  pre_tokenize_text(text) {
    return text.match(this.pattern) || [];
  }
};
var Punctuation_default = Punctuation;
var Digits = class extends PreTokenizer_default {
  /**
   * @param config The configuration options for the pre-tokenizer.
   */
  constructor(config) {
    super();
    this.config = config;
    const digit_pattern = `[^\\d]+|\\d${this.config.individual_digits ? "" : "+"}`;
    this.pattern = new RegExp(digit_pattern, "gu");
  }
  /**
   * Tokenizes text by splitting it using the given pattern.
   * @param text The text to tokenize.
   * @returns An array of tokens.
   */
  pre_tokenize_text(text) {
    return text.match(this.pattern) || [];
  }
};
var Digits_default = Digits;
var BertPreTokenizer = class extends PreTokenizer_default {
  /**
   * A PreTokenizer that splits text into wordpieces using a basic tokenization scheme
   * similar to that used in the original implementation of BERT.
   */
  constructor() {
    super();
    this.pattern = new RegExp(
      `[^\\s${PUNCTUATION_REGEX}]+|[${PUNCTUATION_REGEX}]`,
      "gu"
    );
  }
  /**
   * Tokenizes a single text using the BERT pre-tokenization scheme.
   *
   * @param text The text to tokenize.
   * @param options Additional options for the pre-tokenization logic.
   * @returns An array of tokens.
   */
  pre_tokenize_text(text, options) {
    return text.trim().match(this.pattern) || [];
  }
};
var BertPreTokenizer_default = BertPreTokenizer;
var Replace2 = class extends PreTokenizer_default {
  /**
   * @param config The configuration options for the pre-tokenizer.
   */
  constructor(config) {
    super();
    this.config = config;
    this.pattern = create_pattern(this.config.pattern ?? {});
    this.content = this.config.content ?? "";
  }
  /**
   * Pre-tokenizes the input text by replacing certain characters.
   * @param text The text to be pre-tokenized.
   * @returns An array of tokens produced by replacing certain characters.
   */
  pre_tokenize_text(text) {
    if (this.pattern === null) {
      return [text];
    }
    return [text.replaceAll(this.pattern, this.config.content ?? "")];
  }
};
var Replace_default2 = Replace2;
var Sequence2 = class extends PreTokenizer_default {
  /**
   * Creates an instance of PreTokenizerSequence.
   * @param config The configuration object for the pre-tokenizer sequence.
   */
  constructor(config) {
    super();
    this.tokenizers = (config.pretokenizers ?? []).map(
      (x) => create_pre_tokenizer_default(x)
    );
  }
  /**
   * Applies each pre-tokenizer in the sequence to the input text in turn.
   * @param text The text to pre-tokenize.
   * @param options Additional options for the pre-tokenization logic.
   * @returns The pre-tokenized text.
   */
  pre_tokenize_text(text, options) {
    return this.tokenizers.reduce(
      (pre_tokenized_text, tokenizer) => {
        return tokenizer ? tokenizer.pre_tokenize(pre_tokenized_text, options) : pre_tokenized_text;
      },
      [text]
    );
  }
};
var Sequence_default2 = Sequence2;
var WhitespaceSplit = class extends PreTokenizer_default {
  /**
   * Pre-tokenizes the input text by splitting it on whitespace characters.
   * @param text The text to be pre-tokenized.
   * @returns An array of tokens produced by splitting the input text on whitespace.
   */
  pre_tokenize_text(text) {
    return whitespace_split(text);
  }
};
var WhitespaceSplit_default = WhitespaceSplit;
var FixedLength = class extends PreTokenizer_default {
  /**
   * @param config The configuration options for the pre-tokenizer.
   */
  constructor(config) {
    super();
    this.config = config;
    this._length = config.length;
  }
  /**
   * Pre-tokenizes the input text by splitting it into fixed-length tokens.
   * @param text The text to be pre-tokenized.
   * @returns An array of tokens produced by splitting the input text into fixed-length tokens.
   */
  pre_tokenize_text(text) {
    const tokens = [];
    for (let i = 0; i < text.length; i += this._length) {
      tokens.push(text.slice(i, i + this._length));
    }
    return tokens;
  }
};
var FixedLength_default = FixedLength;
function create_pre_tokenizer(config) {
  if (config === null) return null;
  switch (config.type) {
    case "BertPreTokenizer":
      return new BertPreTokenizer_default();
    case "Sequence":
      return new Sequence_default2(config);
    case "Whitespace":
      return new Whitespace_default();
    case "WhitespaceSplit":
      return new WhitespaceSplit_default();
    case "Metaspace":
      return new Metaspace_default(config);
    case "ByteLevel":
      return new ByteLevel_default(config);
    case "Split":
      return new Split_default(config);
    case "Punctuation":
      return new Punctuation_default(config);
    case "Digits":
      return new Digits_default(config);
    case "Replace":
      return new Replace_default2(config);
    case "FixedLength":
      return new FixedLength_default(config);
    default:
      throw new Error(`Unknown PreTokenizer type: ${config.type}`);
  }
}
var create_pre_tokenizer_default = create_pre_tokenizer;
var TokenizerModel = class extends Callable_default {
  /**
   * Creates a new instance of TokenizerModel.
   * @param config The configuration object for the TokenizerModel.
   */
  constructor(config) {
    super();
    this.config = config;
    this.vocab = [];
    this.tokens_to_ids = /* @__PURE__ */ new Map();
    this.unk_token_id = void 0;
    this.unk_token = void 0;
    this.end_of_word_suffix = void 0;
    this.fuse_unk = this.config.fuse_unk ?? false;
  }
  /**
   * Internal function to call the TokenizerModel instance.
   * @param tokens The tokens to encode.
   * @returns The encoded tokens.
   */
  _call(tokens) {
    let result = this.encode(tokens);
    if (this.fuse_unk) {
      result = fuse_unk(result, this.tokens_to_ids, this.unk_token_id);
    }
    return result;
  }
};
var TokenizerModel_default = TokenizerModel;
var WordPieceTokenizer = class extends TokenizerModel_default {
  /**
   * @param config The configuration object.
   */
  constructor(config) {
    super(config);
    this.max_input_chars_per_word = 100;
    this.tokens_to_ids = object_to_map(config.vocab);
    this.unk_token_id = this.tokens_to_ids.get(config.unk_token);
    this.unk_token = config.unk_token;
    this.max_input_chars_per_word = config.max_input_chars_per_word ?? 100;
    this.vocab = new Array(this.tokens_to_ids.size);
    for (const [key, value] of this.tokens_to_ids) {
      this.vocab[value] = key;
    }
  }
  /**
   * Encodes an array of tokens using WordPiece encoding.
   * @param tokens The tokens to encode.
   * @returns An array of encoded tokens.
   */
  encode(tokens) {
    const output_tokens = [];
    for (const token of tokens) {
      const chars = [...token];
      if (chars.length > this.max_input_chars_per_word) {
        output_tokens.push(this.unk_token);
        continue;
      }
      let is_unknown = false;
      let start = 0;
      const sub_tokens = [];
      while (start < chars.length) {
        let end = chars.length;
        let current_substring = null;
        while (start < end) {
          let substr = chars.slice(start, end).join("");
          if (start > 0) {
            substr = this.config.continuing_subword_prefix + substr;
          }
          if (this.tokens_to_ids.has(substr)) {
            current_substring = substr;
            break;
          }
          --end;
        }
        if (current_substring === null) {
          is_unknown = true;
          break;
        }
        sub_tokens.push(current_substring);
        start = end;
      }
      if (is_unknown) {
        output_tokens.push(this.unk_token);
      } else {
        output_tokens.push(...sub_tokens);
      }
    }
    return output_tokens;
  }
};
var WordPiece_default = WordPieceTokenizer;
var CharTrieNode = class _CharTrieNode {
  /**
   * Create a new CharTrieNode.
   * @param is_leaf Whether the node is a leaf node or not.
   * @param children A map containing the node's children, where the key is a character and the value is a `CharTrieNode`.
   */
  constructor(is_leaf, children) {
    this.is_leaf = is_leaf;
    this.children = children;
  }
  /**
   * Returns a new `CharTrieNode` instance with default values.
   * @returns A new `CharTrieNode` instance with `is_leaf` set to `false` and an empty `children` map.
   */
  static default() {
    return new _CharTrieNode(false, /* @__PURE__ */ new Map());
  }
};
var CharTrie = class {
  constructor() {
    this.root = CharTrieNode.default();
  }
  /**
   * Adds one or more `texts` to the trie.
   * @param texts The strings to add to the trie.
   */
  extend(texts) {
    for (const text of texts) {
      this.push(text);
    }
  }
  /**
   * Adds text to the trie.
   * @param text The string to add to the trie.
   */
  push(text) {
    let node = this.root;
    for (const ch of text) {
      let child = node.children.get(ch);
      if (child === void 0) {
        child = CharTrieNode.default();
        node.children.set(ch, child);
      }
      node = child;
    }
    node.is_leaf = true;
  }
  /**
   * Searches the trie for all strings with a common prefix of `text`.
   * @param text The common prefix to search for.
   * @yields Each string in the trie that has `text` as a prefix.
   */
  *common_prefix_search(text) {
    let node = this.root;
    if (node === void 0) return;
    let prefix = "";
    for (const ch of text) {
      prefix += ch;
      node = node.children.get(ch);
      if (node === void 0) return;
      if (node.is_leaf) {
        yield prefix;
      }
    }
  }
};
var CharTrie_default = CharTrie;
var TokenLatticeNode = class _TokenLatticeNode {
  /**
   * Represents a node in a token lattice for a given sentence.
   * @param token_id The ID of the token associated with this node.
   * @param node_id The ID of this node.
   * @param pos The starting position of the token in the sentence.
   * @param length The length of the token.
   * @param score The score associated with the token.
   */
  constructor(token_id, node_id, pos, length, score) {
    this.token_id = token_id;
    this.node_id = node_id;
    this.pos = pos;
    this.length = length;
    this.score = score;
    this.prev = null;
    this.backtrace_score = 0;
  }
  /**
   * Returns a clone of this node.
   * @returns A clone of this node.
   */
  clone() {
    const n = new _TokenLatticeNode(
      this.token_id,
      this.node_id,
      this.pos,
      this.length,
      this.score
    );
    n.prev = this.prev;
    n.backtrace_score = this.backtrace_score;
    return n;
  }
};
var TokenLattice = class {
  /**
   * Creates a new TokenLattice instance.
   *
   * @param sentence The input sentence to be tokenized.
   * @param bos_token_id The beginning-of-sequence token ID.
   * @param eos_token_id The end-of-sequence token ID.
   */
  constructor(sentence, bos_token_id, eos_token_id) {
    this.chars = Array.from(sentence);
    this.len = this.chars.length;
    this.bos_token_id = bos_token_id;
    this.eos_token_id = eos_token_id;
    this.nodes = [];
    this.begin_nodes = Array.from(
      { length: this.len + 1 },
      () => []
    );
    this.end_nodes = Array.from({ length: this.len + 1 }, () => []);
    const bos = new TokenLatticeNode(this.bos_token_id ?? 0, 0, 0, 0, 0);
    const eos = new TokenLatticeNode(
      this.eos_token_id ?? 0,
      1,
      this.len,
      0,
      0
    );
    this.nodes.push(bos.clone());
    this.nodes.push(eos.clone());
    this.begin_nodes[this.len].push(eos);
    this.end_nodes[0].push(bos);
  }
  /**
   * Inserts a new token node into the token lattice.
   *
   * @param pos The starting position of the token.
   * @param length The length of the token.
   * @param score The score of the token.
   * @param token_id The token ID of the token.
   */
  insert(pos, length, score, token_id) {
    const node_id = this.nodes.length;
    const node = new TokenLatticeNode(token_id, node_id, pos, length, score);
    this.begin_nodes[pos].push(node);
    this.end_nodes[pos + length].push(node);
    this.nodes.push(node);
  }
  /**
   * Implements the Viterbi algorithm to compute the most likely sequence of tokens.
   *
   * @returns The most likely sequence of tokens.
   */
  viterbi() {
    const len2 = this.len;
    let pos = 0;
    while (pos <= len2) {
      if (this.begin_nodes[pos].length == 0) {
        return [];
      }
      for (let rnode of this.begin_nodes[pos]) {
        rnode.prev = null;
        let best_score = 0;
        let best_node = null;
        for (let lnode of this.end_nodes[pos]) {
          const score = lnode.backtrace_score + rnode.score;
          if (best_node === null || score > best_score) {
            best_node = lnode.clone();
            best_score = score;
          }
        }
        if (best_node !== null) {
          rnode.prev = best_node;
          rnode.backtrace_score = best_score;
        } else {
          return [];
        }
      }
      ++pos;
    }
    const results = [];
    const root = this.begin_nodes[len2][0];
    const prev = root.prev;
    if (prev === null) {
      return [];
    }
    let node = prev.clone();
    while (node.prev !== null) {
      results.push(node.clone());
      const n = node.clone();
      node = n.prev.clone();
    }
    results.reverse();
    return results;
  }
  /**
   * Get the text piece for a given node.
   * @param node The node to get the piece for.
   * @returns The array of nodes representing the most likely sequence of tokens.
   */
  piece(node) {
    return this.chars.slice(node.pos, node.pos + node.length).join("");
  }
  /**
   * @returns The most likely sequence of tokens.
   */
  tokens() {
    const nodes = this.viterbi();
    return nodes.map((x) => this.piece(x));
  }
  /**
   * @returns The most likely sequence of token ids.
   */
  token_ids() {
    const nodes = this.viterbi();
    return nodes.map((x) => x.token_id);
  }
};
var TokenLattice_default = TokenLattice;
function min(arr) {
  if (arr.length === 0) throw new Error("Array must not be empty");
  let min_value = arr[0];
  let index_of_min = 0;
  for (let i = 1; i < arr.length; ++i) {
    if (arr[i] < min_value) {
      min_value = arr[i];
      index_of_min = i;
    }
  }
  return [min_value, index_of_min];
}
var Unigram = class extends TokenizerModel_default {
  /**
   * Create a new Unigram tokenizer model.
   * @param config The configuration object for the Unigram model.
   * @param eos_token
   */
  constructor(config, eos_token) {
    super(config);
    const vocab_size = config.vocab.length;
    this.vocab = new Array(vocab_size);
    this.scores = new Array(vocab_size);
    for (let i = 0; i < vocab_size; ++i) {
      [this.vocab[i], this.scores[i]] = config.vocab[i];
    }
    this.unk_token_id = config.unk_id;
    this.unk_token = this.vocab[config.unk_id];
    this.tokens_to_ids = new Map(this.vocab.map((x, i) => [x, i]));
    this.bos_token = " ";
    this.bos_token_id = this.tokens_to_ids.get(this.bos_token);
    this.eos_token = eos_token;
    this.eos_token_id = this.tokens_to_ids.get(this.eos_token);
    this.unk_token = this.vocab[this.unk_token_id];
    this.min_score = min(this.scores)[0];
    this.unk_score = this.min_score - 10;
    this.scores[this.unk_token_id] = this.unk_score;
    this.trie = new CharTrie_default();
    this.trie.extend(this.vocab);
    this.fuse_unk = true;
  }
  /**
   * Populates lattice nodes.
   * @param lattice The token lattice to populate with nodes.
   */
  populate_nodes(lattice) {
    const chars = lattice.chars;
    const mblen = 1;
    let begin_pos = 0;
    while (begin_pos < chars.length) {
      let has_single_node = false;
      const tokens = [];
      const sliced = chars.slice(begin_pos).join("");
      const prefixed_tokens = this.trie.common_prefix_search(sliced);
      for (const token of prefixed_tokens) {
        tokens.push(token);
        const token_id = this.tokens_to_ids.get(token);
        const token_score = this.scores[token_id];
        const n = len(token);
        lattice.insert(begin_pos, n, token_score, token_id);
        if (!has_single_node && n === mblen) {
          has_single_node = true;
        }
      }
      if (!has_single_node) {
        lattice.insert(begin_pos, mblen, this.unk_score, this.unk_token_id);
      }
      begin_pos += mblen;
    }
  }
  /**
   * Encodes an array of tokens into an array of subtokens using the unigram model.
   *
   * @param normalized The normalized string.
   * @returns An array of subtokens obtained by encoding the input tokens using the unigram model.
   */
  tokenize(normalized) {
    const lattice = new TokenLattice_default(
      normalized,
      this.bos_token_id,
      this.eos_token_id
    );
    this.populate_nodes(lattice);
    return lattice.tokens();
  }
  /**
   * Encodes an array of tokens using Unigram encoding.
   * @param tokens The tokens to encode.
   * @returns An array of encoded tokens.
   */
  encode(tokens) {
    const to_return = [];
    for (const token of tokens) {
      const tokenized = this.tokenize(token);
      to_return.push(...tokenized);
    }
    return to_return;
  }
};
var Unigram_default = Unigram;
var PriorityQueue = class {
  /**
   * Create a new PriorityQueue.
   * @param comparator Comparator function to determine priority. Defaults to a MaxHeap.
   * @param max_size Maximum size of the queue. Defaults to Infinity.
   */
  constructor(comparator = (a, b) => a > b, max_size = Infinity) {
    this._heap = [];
    this._comparator = comparator;
    this._max_size = max_size;
  }
  /**
   * The size of the queue
   */
  get size() {
    return this._heap.length;
  }
  /**
   * Check if the queue is empty.
   * @returns `true` if the queue is empty, `false` otherwise.
   */
  is_empty() {
    return this.size === 0;
  }
  /**
   * Return the element with the highest priority in the queue.
   * @returns The highest priority element in the queue.
   */
  peek() {
    return this._heap[0];
  }
  /**
   * Add one or more elements to the queue.
   * @param values The values to push into the queue.
   * @returns The new size of the queue.
   */
  push(...values) {
    return this.extend(values);
  }
  /**
   * Add multiple elements to the queue.
   * @param values The values to push into the queue.
   * @returns The new size of the queue.
   */
  extend(values) {
    for (const value of values) {
      if (this.size < this._max_size) {
        this._heap.push(value);
        this._sift_up();
      } else {
        const smallest = this._smallest();
        if (this._comparator(value, this._heap[smallest])) {
          this._heap[smallest] = value;
          this._sift_up_from(smallest);
        }
      }
    }
    return this.size;
  }
  /**
   * Remove and return the element with the highest priority in the queue.
   * @returns The element with the highest priority in the queue.
   */
  pop() {
    const popped_value = this.peek();
    const bottom = this.size - 1;
    if (bottom > 0) {
      this._swap(0, bottom);
    }
    this._heap.pop();
    this._sift_down();
    return popped_value;
  }
  /**
   * Replace the element with the highest priority in the queue with a new value.
   * @param value The new value.
   * @returns The replaced value.
   */
  replace(value) {
    const replaced_value = this.peek();
    this._heap[0] = value;
    this._sift_down();
    return replaced_value;
  }
  /**
   * Compute the index for the parent of the node at index `i`.
   * @param i The index of the node to get the parent of.
   * @returns The index of the parent node.
   * @private
   */
  _parent(i) {
    return (i + 1 >>> 1) - 1;
  }
  /**
   * Compute the index for the left child of the node at index `i`.
   * @param i The index of the node to get the left child of.
   * @returns The index of the left child.
   * @private
   */
  _left(i) {
    return (i << 1) + 1;
  }
  /**
   * Compute the index for the right child of the node at index `i`.
   * @param i The index of the node to get the right child of.
   * @returns The index of the right child.
   * @private
   */
  _right(i) {
    return i + 1 << 1;
  }
  /**
   * Check if the element at index `i` is greater than the element at index `j`.
   * @param i The index of the first element to compare.
   * @param j The index of the second element to compare.
   * @returns `true` if the element at index `i` is greater than the element at index `j`, `false` otherwise.
   * @private
   */
  _greater(i, j) {
    return this._comparator(this._heap[i], this._heap[j]);
  }
  /**
   * Swap the elements at indices `i` and `j`.
   * @param i The index of the first element to swap.
   * @param j The index of the second element to swap.
   * @private
   */
  _swap(i, j) {
    const temp = this._heap[i];
    this._heap[i] = this._heap[j];
    this._heap[j] = temp;
  }
  /**
   * Maintain the heap property by updating positions in the heap,
   * starting at the last element and moving up the heap.
   * @private
   */
  _sift_up() {
    this._sift_up_from(this.size - 1);
  }
  /**
   * Helper function to sift up from a given node.
   * @param node The index of the node to start sifting up from.
   */
  _sift_up_from(node) {
    while (node > 0 && this._greater(node, this._parent(node))) {
      this._swap(node, this._parent(node));
      node = this._parent(node);
    }
  }
  /**
   * Maintain the heap property by updating positions in the heap,
   * starting at the first element and moving down the heap.
   * @private
   */
  _sift_down() {
    let node = 0;
    while (this._left(node) < this.size && this._greater(this._left(node), node) || this._right(node) < this.size && this._greater(this._right(node), node)) {
      const max_child = this._right(node) < this.size && this._greater(this._right(node), this._left(node)) ? this._right(node) : this._left(node);
      this._swap(node, max_child);
      node = max_child;
    }
  }
  /**
   * Get the index of the smallest element in the heap. Since we use an array-based heap,
   * the index can be computed without needing to traverse the heap.
   * @private
   */
  _smallest() {
    return 2 ** Math.floor(Math.log2(this.size)) - 1;
  }
};
var PriorityQueue_default = PriorityQueue;
var LRUCache = class {
  /**
   * Creates an LRUCache instance.
   * @param capacity The maximum number of items the cache can hold.
   */
  constructor(capacity) {
    this.capacity = capacity;
    this.cache = /* @__PURE__ */ new Map();
  }
  /**
   * Retrieves the value associated with the given key and marks the key as recently used.
   * @param key The key to retrieve.
   * @returns The value associated with the key, or undefined if the key does not exist.
   */
  get(key) {
    if (!this.cache.has(key)) return void 0;
    const value = this.cache.get(key);
    this.cache.delete(key);
    this.cache.set(key, value);
    return value;
  }
  /**
   * Inserts or updates the key-value pair in the cache.
   * If the key already exists, it is updated and marked as recently used.
   * If the cache exceeds its capacity, the least recently used item is evicted.
   * @param key The key to add or update.
   * @param value The value to associate with the key.
   */
  put(key, value) {
    if (this.cache.has(key)) {
      this.cache.delete(key);
    }
    this.cache.set(key, value);
    if (this.cache.size > this.capacity) {
      this.cache.delete(this.cache.keys().next().value);
    }
  }
  /**
   * Clears the cache.
   */
  clear() {
    this.cache.clear();
  }
};
var LRUCache_default = LRUCache;
var BPE = class extends TokenizerModel_default {
  /**
   * Create a BPE instance.
   * @param config The configuration object for BPE.
   */
  constructor(config) {
    super(config);
    this.tokens_to_ids = object_to_map(config.vocab);
    this.unk_token_id = this.tokens_to_ids.get(config.unk_token);
    this.unk_token = config.unk_token;
    this.vocab = new Array(this.tokens_to_ids.size);
    for (const [key, value] of this.tokens_to_ids) {
      this.vocab[value] = key;
    }
    const use_new_merge_format = Array.isArray(config.merges[0]);
    this.merges = use_new_merge_format ? config.merges : config.merges.map(
      (x) => x.split(" ", 2)
    );
    this.bpe_ranks = new Map(this.merges.map((x, i) => [JSON.stringify(x), i]));
    this.end_of_word_suffix = config.end_of_word_suffix;
    this.continuing_subword_suffix = config.continuing_subword_suffix ?? null;
    this.byte_fallback = this.config.byte_fallback ?? false;
    if (this.byte_fallback) {
      this.text_encoder = new TextEncoder();
    }
    this.ignore_merges = this.config.ignore_merges ?? false;
    this.max_length_to_cache = 256;
    this.cache_capacity = 1e4;
    this.cache = new LRUCache_default(this.cache_capacity);
  }
  /**
   * Clears the cache.
   */
  clear_cache() {
    this.cache.clear();
  }
  /**
   * Apply Byte-Pair-Encoding (BPE) to a given token. Efficient heap-based priority
   * queue implementation adapted from https://github.com/belladoreai/llama-tokenizer-js.
   * @param token The token to encode.
   * @returns The BPE encoded tokens.
   */
  bpe(token) {
    if (token.length === 0) {
      return [];
    }
    const cached = this.cache.get(token);
    if (cached !== void 0) {
      return cached;
    }
    const word = Array.from(token);
    if (this.end_of_word_suffix) {
      word[word.length - 1] += this.end_of_word_suffix;
    }
    let result = [];
    if (word.length > 1) {
      const queue = new PriorityQueue_default((a, b) => a.score < b.score);
      let starting_node = {
        token: word[0],
        bias: 0,
        prev: null,
        next: null
      };
      let previous_node = starting_node;
      for (let i = 1; i < word.length; ++i) {
        const current_node = {
          bias: i / word.length,
          // Add fractional component to break ties
          token: word[i],
          prev: previous_node,
          next: null
        };
        previous_node.next = current_node;
        this.add_node(queue, previous_node);
        previous_node = current_node;
      }
      while (!queue.is_empty()) {
        const node = queue.pop();
        if (node.deleted || !node.next || node.next.deleted) continue;
        node.deleted = true;
        node.next.deleted = true;
        if (node.prev) {
          const new_previous_node = { ...node.prev };
          node.prev.deleted = true;
          node.prev = new_previous_node;
          if (new_previous_node.prev) {
            new_previous_node.prev.next = new_previous_node;
          } else {
            starting_node = new_previous_node;
          }
        }
        const merged = {
          token: node.token + node.next.token,
          bias: node.bias,
          prev: node.prev,
          next: node.next.next
        };
        if (merged.prev) {
          merged.prev.next = merged;
          this.add_node(queue, merged.prev);
        } else {
          starting_node = merged;
        }
        if (merged.next) {
          merged.next.prev = merged;
          this.add_node(queue, merged);
        }
      }
      for (let current_node = starting_node; current_node !== null; current_node = current_node.next) {
        result.push(current_node.token);
      }
    } else {
      result = word;
    }
    if (this.continuing_subword_suffix) {
      for (let i = 0; i < result.length - 1; ++i) {
        result[i] += this.continuing_subword_suffix;
      }
    }
    if (token.length < this.max_length_to_cache) {
      this.cache.put(token, result);
    }
    return result;
  }
  /**
   * Helper function to add a node to the priority queue.
   * @param queue
   * @param node
   */
  add_node(queue, node) {
    const rank = this.bpe_ranks.get(
      JSON.stringify([node.token, node.next.token])
    );
    if (rank !== void 0) {
      node.score = rank + node.bias;
      queue.push(node);
    }
  }
  /**
   * Encodes the input sequence of tokens using the BPE algorithm and returns the resulting subword tokens.
   * @param tokens The input sequence of tokens to encode.
   * @returns The resulting subword tokens after applying the BPE algorithm to the input sequence of tokens.
   */
  encode(tokens) {
    const output_tokens = [];
    for (const token of tokens) {
      if (this.ignore_merges && this.tokens_to_ids.has(token)) {
        output_tokens.push(token);
        continue;
      }
      const bpe_token_list = this.bpe(token);
      for (const t of bpe_token_list) {
        if (this.tokens_to_ids.has(t)) {
          output_tokens.push(t);
        } else if (this.byte_fallback) {
          const byte_tokens = Array.from(this.text_encoder.encode(t)).map(
            (x) => `<0x${x.toString(16).toUpperCase().padStart(2, "0")}>`
          );
          if (byte_tokens.every((x) => this.tokens_to_ids.has(x))) {
            output_tokens.push(...byte_tokens);
          } else if (this.unk_token != null) {
            output_tokens.push(this.unk_token);
          }
        } else if (this.unk_token != null) {
          output_tokens.push(this.unk_token);
        }
      }
    }
    return output_tokens;
  }
};
var BPE_default = BPE;
var Legacy = class extends TokenizerModel_default {
  /**
   * Create a Legacy tokenizer model instance.
   * @param config The configuration object for Legacy tokenizer model.
   * @param more_config Additional configuration object for the Legacy tokenizer model.
   */
  constructor(config, more_config) {
    super(config);
    const vocab = config.vocab;
    this.tokens_to_ids = object_to_map(
      more_config.target_lang ? vocab[more_config.target_lang] : vocab
    );
    this.bos_token = more_config.bos_token;
    this.bos_token_id = this.tokens_to_ids.get(this.bos_token);
    this.eos_token = more_config.eos_token;
    this.eos_token_id = this.tokens_to_ids.get(this.eos_token);
    this.pad_token = more_config.pad_token;
    this.pad_token_id = this.tokens_to_ids.get(this.pad_token);
    this.unk_token = more_config.unk_token;
    this.unk_token_id = this.tokens_to_ids.get(this.unk_token);
    this.vocab = new Array(this.tokens_to_ids.size);
    for (const [key, value] of this.tokens_to_ids) {
      this.vocab[value] = key;
    }
  }
  encode(tokens) {
    return tokens;
  }
};
var Legacy_default = Legacy;
function create_tokenizer_model(model_config, config) {
  switch (model_config.type) {
    case "WordPiece":
      return new WordPiece_default(model_config);
    case "Unigram":
      return new Unigram_default(model_config, config.eos_token);
    case "BPE":
      return new BPE_default(model_config);
    default:
      if (model_config.vocab) {
        if (Array.isArray(model_config.vocab)) {
          return new Unigram_default(model_config, config.eos_token);
        } else if (Object.hasOwn(model_config, "continuing_subword_prefix") && Object.hasOwn(model_config, "unk_token")) {
          if (Object.hasOwn(model_config, "merges")) {
            return new BPE_default(model_config);
          } else {
            return new WordPiece_default(model_config);
          }
        } else {
          return new Legacy_default(model_config, {
            target_lang: config.target_lang,
            bos_token: config.bos_token,
            eos_token: config.eos_token,
            pad_token: config.pad_token,
            unk_token: config.unk_token
          });
        }
      }
      throw new Error(
        `Unknown TokenizerModel type: ${model_config?.type}`
      );
  }
}
var create_tokenizer_model_default = create_tokenizer_model;
var PostProcessor = class extends Callable_default {
  /**
   * @param config The configuration for the post-processor.
   */
  constructor(config) {
    super();
    this.config = config;
  }
  /**
   * Alias for {@link PostProcessor#post_process}.
   * @param tokens The text or array of texts to post-process.
   * @param args Additional arguments required by the post-processing logic.
   * @returns The post-processed tokens.
   */
  _call(tokens, ...args) {
    return this.post_process(tokens, ...args);
  }
};
var PostProcessor_default = PostProcessor;
var TemplateProcessing = class extends PostProcessor_default {
  /**
   * Replaces special tokens in the template with actual tokens.
   * @param tokens The list of tokens for the first sequence.
   * @param tokens_pair The list of tokens for the second sequence (optional).
   * @param add_special_tokens Whether to add the special tokens to the beginning and end of the input.
   * @returns An object containing the list of tokens with the special tokens replaced with actual tokens.
   */
  post_process(tokens, tokens_pair = null, add_special_tokens = true) {
    const type = tokens_pair === null ? this.config.single : this.config.pair;
    let processed_tokens = [];
    let types = [];
    for (const item of type) {
      if ("SpecialToken" in item) {
        if (add_special_tokens) {
          processed_tokens.push(item.SpecialToken.id);
          types.push(item.SpecialToken.type_id);
        }
      } else if ("Sequence" in item) {
        if (item.Sequence.id === "A") {
          processed_tokens = merge_arrays(processed_tokens, tokens);
          types = merge_arrays(
            types,
            new Array(tokens.length).fill(item.Sequence.type_id)
          );
        } else if (item.Sequence.id === "B") {
          processed_tokens = merge_arrays(processed_tokens, tokens_pair);
          types = merge_arrays(
            types,
            new Array(tokens_pair.length).fill(item.Sequence.type_id)
          );
        }
      }
    }
    return { tokens: processed_tokens, token_type_ids: types };
  }
};
var TemplateProcessing_default = TemplateProcessing;
var ByteLevel2 = class extends PostProcessor_default {
  /**
   * Post process the given tokens.
   * @param tokens The list of tokens for the first sequence.
   * @param tokens_pair The list of tokens for the second sequence (optional).
   * @returns An object containing the post-processed tokens.
   */
  post_process(tokens, tokens_pair = null) {
    return { tokens, tokens_pair };
  }
};
var ByteLevel_default2 = ByteLevel2;
var BertProcessing = class extends PostProcessor_default {
  /**
   * @param config The configuration for the post-processor.
   * @param config.cls The special tokens to add to the beginning of the input.
   * @param config.sep The special tokens to add to the end of the input.
   */
  constructor(config) {
    super(config);
    this.sep = config.sep;
    this.cls = config.cls;
  }
  /**
   * Adds the special tokens to the beginning and end of the input.
   * @param tokens The input tokens.
   * @param tokens_pair An optional second set of input tokens.
   * @param add_special_tokens Whether to add the special tokens to the beginning and end of the input.
   * @returns The post-processed tokens with the special tokens added to the beginning and end.
   */
  post_process(tokens, tokens_pair = null, add_special_tokens = true) {
    if (add_special_tokens) {
      tokens = merge_arrays([this.cls[0]], tokens, [this.sep[0]]);
    }
    let token_type_ids = new Array(tokens.length).fill(0);
    if (tokens_pair) {
      const middle = [];
      const after = add_special_tokens ? [this.sep[0]] : [];
      tokens = merge_arrays(tokens, middle, tokens_pair, after);
      token_type_ids = merge_arrays(
        token_type_ids,
        new Array(tokens_pair.length + middle.length + after.length).fill(1)
      );
    }
    return { tokens, token_type_ids };
  }
};
var BertProcessing_default = BertProcessing;
var RobertaProcessing = class extends PostProcessor_default {
  /**
   * @param config The configuration for the post-processor.
   * @param config.cls The special tokens to add to the beginning of the input.
   * @param config.sep The special tokens to add to the end of the input.
   */
  constructor(config) {
    super(config);
    this.sep = config.sep;
    this.cls = config.cls;
  }
  /**
   * Adds the special tokens to the beginning and end of the input.
   * @param tokens The input tokens.
   * @param tokens_pair An optional second set of input tokens.
   * @param add_special_tokens Whether to add the special tokens to the beginning and end of the input.
   * @returns The post-processed tokens with the special tokens added to the beginning and end.
   */
  post_process(tokens, tokens_pair, add_special_tokens = true) {
    if (add_special_tokens) {
      tokens = merge_arrays([this.cls[0]], tokens, [this.sep[0]]);
    }
    let token_type_ids = new Array(tokens.length).fill(0);
    if (tokens_pair) {
      const middle = add_special_tokens ? [this.sep[0]] : [];
      const after = add_special_tokens ? [this.sep[0]] : [];
      tokens = merge_arrays(tokens, middle, tokens_pair, after);
      token_type_ids = merge_arrays(
        token_type_ids,
        new Array(tokens_pair.length + middle.length + after.length).fill(1)
      );
    }
    return { tokens, token_type_ids };
  }
};
var RobertaProcessing_default = RobertaProcessing;
var Sequence3 = class extends PostProcessor_default {
  /**
   * Creates a new instance of Sequence post-processor.
   * @param config The configuration object.
   */
  constructor(config) {
    super(config);
    this.processors = (config.processors ?? []).map((x) => create_post_processor_default(x));
  }
  /**
   * Post process the given tokens.
   * @param tokens The list of tokens for the first sequence.
   * @param tokens_pair The list of tokens for the second sequence (optional).
   * @param add_special_tokens Whether to add the special tokens to the beginning and end of the input.
   * @returns An object containing the post-processed tokens.
   */
  post_process(tokens, tokens_pair = null, add_special_tokens = true) {
    let processed_tokens = { tokens, tokens_pair };
    for (const processor of this.processors) {
      processed_tokens = processor.post_process(
        processed_tokens.tokens,
        processed_tokens.tokens_pair,
        add_special_tokens
      );
    }
    return processed_tokens;
  }
};
var Sequence_default3 = Sequence3;
function create_post_processor(config) {
  if (config === null) return null;
  switch (config.type) {
    case "TemplateProcessing":
      return new TemplateProcessing_default(config);
    case "ByteLevel":
      return new ByteLevel_default2(config);
    case "BertProcessing":
      return new BertProcessing_default(config);
    case "RobertaProcessing":
      return new RobertaProcessing_default(config);
    case "Sequence":
      return new Sequence_default3(config);
    default:
      throw new Error(`Unknown PostProcessor type: ${config.type}`);
  }
}
var create_post_processor_default = create_post_processor;
var Decoder = class extends Callable_default {
  /**
   * Creates an instance of `Decoder`.
   * @param config The configuration object.
   **/
  constructor(config) {
    super();
    this.config = config;
    this.added_tokens = [];
    this.end_of_word_suffix = null;
    this.trim_offsets = "trim_offsets" in config ? config.trim_offsets : false;
  }
  /**
   * Calls the `decode` method.
   *
   * @param tokens The list of tokens.
   * @returns The decoded string.
   */
  _call(tokens) {
    return this.decode(tokens);
  }
  /**
   * Decodes a list of tokens.
   * @param tokens The list of tokens.
   * @returns The decoded string.
   */
  decode(tokens) {
    return this.decode_chain(tokens).join("");
  }
};
var Decoder_default = Decoder;
var ByteLevel3 = class extends Decoder_default {
  /**
   * Create a `ByteLevelDecoder` object.
   */
  constructor(config) {
    super(config);
    this.byte_decoder = UNICODE_TO_BYTES;
    this.text_decoder = new TextDecoder("utf-8", {
      fatal: false,
      // eslint-disable-next-line @typescript-eslint/naming-convention
      ignoreBOM: true
    });
    this.end_of_word_suffix = null;
  }
  /**
   * Convert an array of tokens to string by decoding each byte.
   * @param tokens Array of tokens to be decoded.
   * @returns The decoded string.
   */
  convert_tokens_to_string(tokens) {
    const text = tokens.join("");
    const byte_array = new Uint8Array(
      [...text].map((c) => this.byte_decoder[c])
    );
    return this.text_decoder.decode(byte_array);
  }
  decode_chain(tokens) {
    const sub_texts = [];
    let current_sub_text = [];
    for (const token of tokens) {
      if (this.added_tokens.find((x) => x.content === token) !== void 0) {
        if (current_sub_text.length > 0) {
          sub_texts.push(this.convert_tokens_to_string(current_sub_text));
          current_sub_text = [];
        }
        sub_texts.push(token);
      } else {
        current_sub_text.push(token);
      }
    }
    if (current_sub_text.length > 0) {
      sub_texts.push(this.convert_tokens_to_string(current_sub_text));
    }
    return sub_texts;
  }
};
var ByteLevel_default3 = ByteLevel3;
var WordPiece = class extends Decoder_default {
  /**
   * Creates a new instance of WordPieceDecoder.
   * @param config The configuration object.
   */
  constructor(config) {
    super(config);
    this.cleanup = config.cleanup;
  }
  decode_chain(tokens) {
    return tokens.map((token, i) => {
      if (i !== 0) {
        const prefix = this.config.prefix;
        if (prefix && token.startsWith(prefix)) {
          token = token.replace(prefix, "");
        } else {
          token = " " + token;
        }
      }
      if (this.cleanup) {
        token = clean_up_tokenization(token);
      }
      return token;
    });
  }
};
var WordPiece_default2 = WordPiece;
var Metaspace2 = class extends Decoder_default {
  /**
   * Constructs a new MetaspaceDecoder object.
   * @param config The configuration object for the MetaspaceDecoder.
   */
  constructor(config) {
    super(config);
    this.replacement = config.replacement ?? "\u2581";
  }
  decode_chain(tokens) {
    const result = [];
    for (let i = 0; i < tokens.length; ++i) {
      let normalized = tokens[i].replaceAll(this.replacement, " ");
      if (i == 0 && normalized.startsWith(" ")) {
        normalized = normalized.substring(1);
      }
      result.push(normalized);
    }
    return result;
  }
};
var Metaspace_default2 = Metaspace2;
var BPE2 = class extends Decoder_default {
  constructor(config) {
    super(config);
    this.suffix = config.suffix ?? "";
  }
  decode_chain(tokens) {
    return tokens.map((token, i) => {
      return token.replaceAll(this.suffix, i === tokens.length - 1 ? "" : " ");
    });
  }
};
var BPE_default2 = BPE2;
var CTC = class extends Decoder_default {
  constructor(config) {
    super(config);
    this.pad_token = config.pad_token ?? "";
    this.word_delimiter_token = config.word_delimiter_token ?? "";
    this.cleanup = config.cleanup;
  }
  /**
   * Converts a connectionist-temporal-classification (CTC) output tokens into a single string.
   * @param tokens Array of tokens to be decoded.
   * @returns The decoded string.
   */
  convert_tokens_to_string(tokens) {
    if (tokens.length === 0) return "";
    const grouped_tokens = [tokens[0]];
    for (let i = 1; i < tokens.length; ++i) {
      if (tokens[i] !== grouped_tokens.at(-1)) {
        grouped_tokens.push(tokens[i]);
      }
    }
    const filtered_tokens = grouped_tokens.filter(
      (token) => token !== this.pad_token
    );
    let text = filtered_tokens.join("");
    if (this.cleanup) {
      text = clean_up_tokenization(text).replaceAll(this.word_delimiter_token, " ").trim();
    }
    return text;
  }
  decode_chain(tokens) {
    return [this.convert_tokens_to_string(tokens)];
  }
};
var CTC_default = CTC;
var Sequence4 = class extends Decoder_default {
  /**
   * Creates a new instance of DecoderSequence.
   * @param config The configuration object.
   */
  constructor(config) {
    super(config);
    this.decoders = (config.decoders ?? []).map((x) => create_decoder_default(x));
  }
  decode_chain(tokens) {
    return this.decoders.reduce((toks, decoder) => {
      return decoder.decode_chain(toks);
    }, tokens);
  }
};
var Sequence_default4 = Sequence4;
var Replace3 = class extends Decoder_default {
  decode_chain(tokens) {
    const pattern = create_pattern(this.config.pattern);
    const content = this.config.content ?? "";
    return pattern === null ? tokens : tokens.map((token) => token.replaceAll(pattern, content));
  }
};
var Replace_default3 = Replace3;
var Fuse = class extends Decoder_default {
  decode_chain(tokens) {
    return [tokens.join("")];
  }
};
var Fuse_default = Fuse;
var Strip2 = class extends Decoder_default {
  constructor(config) {
    super(config);
    this.content = config.content ?? "";
    this.start = config.start ?? 0;
    this.stop = config.stop ?? 0;
  }
  decode_chain(tokens) {
    return tokens.map((token) => {
      let start_cut = 0;
      for (let i = 0; i < this.start; ++i) {
        if (token[i] === this.content) {
          start_cut = i + 1;
          continue;
        } else {
          break;
        }
      }
      let stop_cut = token.length;
      for (let i = 0; i < this.stop; ++i) {
        const index = token.length - i - 1;
        if (token[index] === this.content) {
          stop_cut = index;
          continue;
        } else {
          break;
        }
      }
      return token.slice(start_cut, stop_cut);
    });
  }
};
var Strip_default2 = Strip2;
var ByteFallback = class extends Decoder_default {
  constructor(config) {
    super(config);
    this.text_decoder = new TextDecoder();
  }
  decode_chain(tokens) {
    const new_tokens = [];
    let previous_byte_tokens = [];
    for (const token of tokens) {
      let bytes = null;
      if (token.length === 6 && token.startsWith("<0x") && token.endsWith(">")) {
        const byte = parseInt(token.slice(3, 5), 16);
        if (!isNaN(byte)) {
          bytes = byte;
        }
      }
      if (bytes !== null) {
        previous_byte_tokens.push(bytes);
      } else {
        if (previous_byte_tokens.length > 0) {
          const string = this.text_decoder.decode(
            Uint8Array.from(previous_byte_tokens)
          );
          new_tokens.push(string);
          previous_byte_tokens = [];
        }
        new_tokens.push(token);
      }
    }
    if (previous_byte_tokens.length > 0) {
      const string = this.text_decoder.decode(
        Uint8Array.from(previous_byte_tokens)
      );
      new_tokens.push(string);
      previous_byte_tokens = [];
    }
    return new_tokens;
  }
};
var ByteFallback_default = ByteFallback;
function create_decoder(config) {
  if (config === null) return null;
  switch (config.type) {
    case "ByteLevel":
      return new ByteLevel_default3(config);
    case "WordPiece":
      return new WordPiece_default2(config);
    case "Metaspace":
      return new Metaspace_default2(config);
    case "BPEDecoder":
      return new BPE_default2(config);
    case "CTC":
      return new CTC_default(config);
    case "Sequence":
      return new Sequence_default4(config);
    case "Replace":
      return new Replace_default3(config);
    case "Fuse":
      return new Fuse_default(config);
    case "Strip":
      return new Strip_default2(config);
    case "ByteFallback":
      return new ByteFallback_default(config);
    default:
      throw new Error(`Unknown Decoder type: ${config.type}`);
  }
}
var create_decoder_default = create_decoder;
var Tokenizer = class {
  constructor(tokenizer, config) {
    const tokenizer_error = validate_object(tokenizer, "Tokenizer", [
      "model",
      "decoder",
      "post_processor",
      "pre_tokenizer",
      "normalizer"
    ]);
    if (tokenizer_error) {
      throw new Error(tokenizer_error);
    }
    const config_error = validate_object(config, "Config");
    if (config_error) {
      throw new Error(config_error);
    }
    this.tokenizer = tokenizer;
    this.config = config;
    this.normalizer = create_normalizer_default(this.tokenizer.normalizer);
    this.pre_tokenizer = create_pre_tokenizer_default(this.tokenizer.pre_tokenizer);
    this.model = create_tokenizer_model_default(this.tokenizer.model, this.config);
    this.post_processor = create_post_processor_default(this.tokenizer.post_processor);
    this.decoder = create_decoder_default(this.tokenizer.decoder);
    this.special_tokens = [];
    this.all_special_ids = [];
    this.added_tokens = [];
    const unnormalized_contents = [];
    const normalized_contents = [];
    this.added_tokens_map = /* @__PURE__ */ new Map();
    for (const added_token of this.tokenizer.added_tokens) {
      const token = new AddedToken_default(added_token);
      this.added_tokens.push(token);
      this.model.tokens_to_ids.set(token.content, token.id);
      this.model.vocab[token.id] = token.content;
      if (token.special) {
        this.special_tokens.push(token.content);
        this.all_special_ids.push(token.id);
      }
      this.added_tokens_map.set(token.content, token);
      if (token.normalized && this.normalizer !== null) {
        const normalized_content = this.normalizer(token.content);
        normalized_contents.push(normalized_content);
        this.added_tokens_map.set(normalized_content, token);
      } else {
        unnormalized_contents.push(token.content);
      }
    }
    (this.config.additional_special_tokens ?? []).forEach((token) => {
      if (!this.special_tokens.includes(token)) this.special_tokens.push(token);
    });
    if (this.decoder) {
      this.decoder.added_tokens = this.added_tokens;
      this.decoder.end_of_word_suffix = this.model.end_of_word_suffix;
    }
    this.splitter_unnormalized = new DictionarySplitter_default(unnormalized_contents);
    this.splitter_normalized = new DictionarySplitter_default(normalized_contents);
    this.remove_space = this.config.remove_space;
    this.clean_up_tokenization_spaces = this.config.clean_up_tokenization_spaces ?? true;
    this.do_lowercase_and_remove_accent = this.config.do_lowercase_and_remove_accent ?? false;
  }
  // Implementation
  encode(text, {
    text_pair = null,
    add_special_tokens = true,
    return_token_type_ids = null
  } = {}) {
    const { tokens, token_type_ids } = this.tokenize_helper(text, {
      text_pair,
      add_special_tokens
    });
    const input_ids = tokens.map(
      (t) => this.added_tokens_map.get(t)?.id ?? this.model.tokens_to_ids.get(t) ?? this.model.unk_token_id
    );
    const result = {
      ids: input_ids,
      tokens,
      attention_mask: new Array(input_ids.length).fill(1)
    };
    if (return_token_type_ids && token_type_ids) {
      result.token_type_ids = token_type_ids;
    }
    return result;
  }
  decode(token_ids, options = {}) {
    if (!Array.isArray(token_ids) || token_ids.length === 0 || !is_integral_number(token_ids[0])) {
      throw Error("token_ids must be a non-empty array of integers.");
    }
    let tokens = token_ids.map(
      (i) => this.model.vocab[Number(i)] ?? this.model.unk_token
    );
    if (options.skip_special_tokens) {
      tokens = tokens.filter((x) => !this.special_tokens.includes(x));
    }
    let decoded = this.decoder ? this.decoder(tokens) : tokens.join(" ");
    if (this.decoder && this.decoder.end_of_word_suffix) {
      decoded = decoded.replaceAll(this.decoder.end_of_word_suffix, " ");
      if (options.skip_special_tokens) {
        decoded = decoded.trim();
      }
    }
    if (options.clean_up_tokenization_spaces ?? this.clean_up_tokenization_spaces) {
      decoded = clean_up_tokenization(decoded);
    }
    return decoded;
  }
  /**
   * Converts a string into a sequence of tokens.
   * @param text The sequence to be encoded.
   * @param options An optional object containing the following properties:
   * @returns The list of tokens.
   */
  tokenize(text, { text_pair = null, add_special_tokens = false } = {}) {
    return this.tokenize_helper(text, { text_pair, add_special_tokens }).tokens;
  }
  encode_text(text) {
    if (text === null) {
      return null;
    }
    const sections = this.splitter_unnormalized.split(text);
    sections.forEach((section, i) => {
      const added_token = this.added_tokens_map.get(section);
      if (added_token) {
        if (added_token.lstrip && i > 0) {
          sections[i - 1] = sections[i - 1].trimEnd();
        }
        if (added_token.rstrip && i < sections.length - 1) {
          sections[i + 1] = sections[i + 1].trimStart();
        }
      }
    });
    return sections.flatMap((processed_text, section_index) => {
      if (processed_text.length === 0) {
        return [];
      }
      if (this.added_tokens_map.has(processed_text)) {
        return [processed_text];
      }
      if (this.remove_space === true) {
        processed_text = processed_text.trim().split(/\s+/).join(" ");
      }
      if (this.do_lowercase_and_remove_accent) {
        processed_text = lowercase_and_remove_accents(processed_text);
      }
      if (this.normalizer !== null) {
        processed_text = this.normalizer(processed_text);
      }
      if (processed_text.length === 0) {
        return [];
      }
      const subsections = this.splitter_normalized.split(processed_text);
      subsections.forEach((subsection, j) => {
        const added_token = this.added_tokens_map.get(subsection);
        if (added_token) {
          if (added_token.lstrip && j > 0) {
            subsections[j - 1] = subsections[j - 1].trimEnd();
          }
          if (added_token.rstrip && j < subsections.length - 1) {
            subsections[j + 1] = subsections[j + 1].trimStart();
          }
        }
      });
      return subsections.flatMap((subsection) => {
        if (subsection.length === 0) {
          return [];
        }
        if (this.added_tokens_map.has(subsection)) {
          return [subsection];
        }
        const section_tokens = this.pre_tokenizer !== null ? this.pre_tokenizer(subsection, {
          section_index
        }) : [subsection];
        return this.model(section_tokens);
      });
    });
  }
  tokenize_helper(text, { text_pair = null, add_special_tokens = true }) {
    const tokens1 = this.encode_text(text);
    const tokens2 = this.encode_text(text_pair || null);
    return this.post_processor ? this.post_processor(tokens1, tokens2, add_special_tokens) : { tokens: merge_arrays(tokens1 ?? [], tokens2 ?? []) };
  }
  /**
   * Converts a token string to its corresponding token ID.
   * @param token The token string to convert.
   * @returns The token ID, or undefined if the token is not in the vocabulary.
   */
  token_to_id(token) {
    return this.model.tokens_to_ids.get(token);
  }
  /**
   * Converts a token ID to its corresponding token string.
   * @param id The token ID to convert.
   * @returns The token string, or undefined if the ID is not in the vocabulary.
   */
  id_to_token(id) {
    return this.model.vocab[id];
  }
  /**
   * Returns a mapping of token IDs to AddedToken objects for all added tokens.
   * @returns A Map where keys are token IDs and values are AddedToken objects.
   */
  get_added_tokens_decoder() {
    const decoder = /* @__PURE__ */ new Map();
    for (const token of this.added_tokens) {
      decoder.set(token.id, token);
    }
    return decoder;
  }
  /**
   * Get the underlying vocabulary
   * @param with_added_tokens Whether to include the added tokens
   * @returns The vocabulary
   */
  get_vocab(with_added_tokens = true) {
    const vocab = /* @__PURE__ */ new Map();
    for (let i = 0; i < this.model.vocab.length; ++i) {
      const token = this.model.vocab[i];
      if (with_added_tokens || !this.added_tokens_map.has(token)) {
        vocab.set(token, i);
      }
    }
    return vocab;
  }
};
var Tokenizer_default = Tokenizer;

// ../../node_modules/.pnpm/@huggingface+jinja@0.5.6/node_modules/@huggingface/jinja/dist/index.js
var TOKEN_TYPES = Object.freeze({
  Text: "Text",
  // The text between Jinja statements or expressions
  NumericLiteral: "NumericLiteral",
  // e.g., 123, 1.0
  StringLiteral: "StringLiteral",
  // 'string'
  Identifier: "Identifier",
  // Variables, functions, statements, booleans, etc.
  Equals: "Equals",
  // =
  OpenParen: "OpenParen",
  // (
  CloseParen: "CloseParen",
  // )
  OpenStatement: "OpenStatement",
  // {%
  CloseStatement: "CloseStatement",
  // %}
  OpenExpression: "OpenExpression",
  // {{
  CloseExpression: "CloseExpression",
  // }}
  OpenSquareBracket: "OpenSquareBracket",
  // [
  CloseSquareBracket: "CloseSquareBracket",
  // ]
  OpenCurlyBracket: "OpenCurlyBracket",
  // {
  CloseCurlyBracket: "CloseCurlyBracket",
  // }
  Comma: "Comma",
  // ,
  Dot: "Dot",
  // .
  Colon: "Colon",
  // :
  Pipe: "Pipe",
  // |
  CallOperator: "CallOperator",
  // ()
  AdditiveBinaryOperator: "AdditiveBinaryOperator",
  // + - ~
  MultiplicativeBinaryOperator: "MultiplicativeBinaryOperator",
  // * / %
  ComparisonBinaryOperator: "ComparisonBinaryOperator",
  // < > <= >= == !=
  UnaryOperator: "UnaryOperator",
  // ! - +
  Comment: "Comment"
  // {# ... #}
});
var Token = class {
  /**
   * Constructs a new Token.
   * @param {string} value The raw value as seen inside the source code.
   * @param {TokenType} type The type of token.
   */
  constructor(value, type) {
    this.value = value;
    this.type = type;
  }
};
function isWord(char) {
  return /\w/.test(char);
}
function isInteger(char) {
  return /[0-9]/.test(char);
}
function isWhitespace(char) {
  return /\s/.test(char);
}
var ORDERED_MAPPING_TABLE = [
  // Control sequences
  ["{%", TOKEN_TYPES.OpenStatement],
  ["%}", TOKEN_TYPES.CloseStatement],
  ["{{", TOKEN_TYPES.OpenExpression],
  ["}}", TOKEN_TYPES.CloseExpression],
  // Single character tokens
  ["(", TOKEN_TYPES.OpenParen],
  [")", TOKEN_TYPES.CloseParen],
  ["{", TOKEN_TYPES.OpenCurlyBracket],
  ["}", TOKEN_TYPES.CloseCurlyBracket],
  ["[", TOKEN_TYPES.OpenSquareBracket],
  ["]", TOKEN_TYPES.CloseSquareBracket],
  [",", TOKEN_TYPES.Comma],
  [".", TOKEN_TYPES.Dot],
  [":", TOKEN_TYPES.Colon],
  ["|", TOKEN_TYPES.Pipe],
  // Comparison operators
  ["<=", TOKEN_TYPES.ComparisonBinaryOperator],
  [">=", TOKEN_TYPES.ComparisonBinaryOperator],
  ["==", TOKEN_TYPES.ComparisonBinaryOperator],
  ["!=", TOKEN_TYPES.ComparisonBinaryOperator],
  ["<", TOKEN_TYPES.ComparisonBinaryOperator],
  [">", TOKEN_TYPES.ComparisonBinaryOperator],
  // Arithmetic operators
  ["+", TOKEN_TYPES.AdditiveBinaryOperator],
  ["-", TOKEN_TYPES.AdditiveBinaryOperator],
  ["~", TOKEN_TYPES.AdditiveBinaryOperator],
  ["*", TOKEN_TYPES.MultiplicativeBinaryOperator],
  ["/", TOKEN_TYPES.MultiplicativeBinaryOperator],
  ["%", TOKEN_TYPES.MultiplicativeBinaryOperator],
  // Assignment operator
  ["=", TOKEN_TYPES.Equals]
];
var ESCAPE_CHARACTERS = /* @__PURE__ */ new Map([
  ["n", "\n"],
  // New line
  ["t", "	"],
  // Horizontal tab
  ["r", "\r"],
  // Carriage return
  ["b", "\b"],
  // Backspace
  ["f", "\f"],
  // Form feed
  ["v", "\v"],
  // Vertical tab
  ["'", "'"],
  // Single quote
  ['"', '"'],
  // Double quote
  ["\\", "\\"]
  // Backslash
]);
function preprocess(template, options = {}) {
  if (template.endsWith("\n")) {
    template = template.slice(0, -1);
  }
  if (options.lstrip_blocks) {
    template = template.replace(/^[ \t]*({[#%-])/gm, "$1");
  }
  if (options.trim_blocks) {
    template = template.replace(/([#%-]})\n/g, "$1");
  }
  return template.replace(/{%\s*(end)?generation\s*%}/gs, "");
}
function tokenize(source, options = {}) {
  const tokens = [];
  const src = preprocess(source, options);
  let cursorPosition = 0;
  let curlyBracketDepth = 0;
  const consumeWhile = (predicate) => {
    let str = "";
    while (predicate(src[cursorPosition])) {
      if (src[cursorPosition] === "\\") {
        ++cursorPosition;
        if (cursorPosition >= src.length)
          throw new SyntaxError("Unexpected end of input");
        const escaped = src[cursorPosition++];
        const unescaped = ESCAPE_CHARACTERS.get(escaped);
        if (unescaped === void 0) {
          throw new SyntaxError(`Unexpected escaped character: ${escaped}`);
        }
        str += unescaped;
        continue;
      }
      str += src[cursorPosition++];
      if (cursorPosition >= src.length)
        throw new SyntaxError("Unexpected end of input");
    }
    return str;
  };
  const stripTrailingWhitespace = () => {
    const lastToken = tokens.at(-1);
    if (lastToken && lastToken.type === TOKEN_TYPES.Text) {
      lastToken.value = lastToken.value.trimEnd();
      if (lastToken.value === "") {
        tokens.pop();
      }
    }
  };
  const skipLeadingWhitespace = () => {
    while (cursorPosition < src.length && isWhitespace(src[cursorPosition])) {
      ++cursorPosition;
    }
  };
  main:
    while (cursorPosition < src.length) {
      const lastTokenType = tokens.at(-1)?.type;
      if (lastTokenType === void 0 || lastTokenType === TOKEN_TYPES.CloseStatement || lastTokenType === TOKEN_TYPES.CloseExpression || lastTokenType === TOKEN_TYPES.Comment) {
        let text = "";
        while (cursorPosition < src.length && // Keep going until we hit the next Jinja statement or expression
        !(src[cursorPosition] === "{" && (src[cursorPosition + 1] === "%" || src[cursorPosition + 1] === "{" || src[cursorPosition + 1] === "#"))) {
          text += src[cursorPosition++];
        }
        if (text.length > 0) {
          tokens.push(new Token(text, TOKEN_TYPES.Text));
          continue;
        }
      }
      if (src[cursorPosition] === "{" && src[cursorPosition + 1] === "#") {
        cursorPosition += 2;
        const stripBefore = src[cursorPosition] === "-";
        if (stripBefore) {
          ++cursorPosition;
        }
        let comment = "";
        while (src[cursorPosition] !== "#" || src[cursorPosition + 1] !== "}") {
          if (cursorPosition + 2 >= src.length) {
            throw new SyntaxError("Missing end of comment tag");
          }
          comment += src[cursorPosition++];
        }
        const stripAfter = comment.endsWith("-");
        if (stripAfter) {
          comment = comment.slice(0, -1);
        }
        if (stripBefore) {
          stripTrailingWhitespace();
        }
        tokens.push(new Token(comment, TOKEN_TYPES.Comment));
        cursorPosition += 2;
        if (stripAfter) {
          skipLeadingWhitespace();
        }
        continue;
      }
      if (src.slice(cursorPosition, cursorPosition + 3) === "{%-") {
        stripTrailingWhitespace();
        tokens.push(new Token("{%", TOKEN_TYPES.OpenStatement));
        cursorPosition += 3;
        continue;
      }
      if (src.slice(cursorPosition, cursorPosition + 3) === "{{-") {
        stripTrailingWhitespace();
        tokens.push(new Token("{{", TOKEN_TYPES.OpenExpression));
        curlyBracketDepth = 0;
        cursorPosition += 3;
        continue;
      }
      consumeWhile(isWhitespace);
      if (src.slice(cursorPosition, cursorPosition + 3) === "-%}") {
        tokens.push(new Token("%}", TOKEN_TYPES.CloseStatement));
        cursorPosition += 3;
        skipLeadingWhitespace();
        continue;
      }
      if (src.slice(cursorPosition, cursorPosition + 3) === "-}}") {
        tokens.push(new Token("}}", TOKEN_TYPES.CloseExpression));
        cursorPosition += 3;
        skipLeadingWhitespace();
        continue;
      }
      const char = src[cursorPosition];
      if (char === "-" || char === "+") {
        const lastTokenType2 = tokens.at(-1)?.type;
        if (lastTokenType2 === TOKEN_TYPES.Text || lastTokenType2 === void 0) {
          throw new SyntaxError(`Unexpected character: ${char}`);
        }
        switch (lastTokenType2) {
          case TOKEN_TYPES.Identifier:
          case TOKEN_TYPES.NumericLiteral:
          case TOKEN_TYPES.StringLiteral:
          case TOKEN_TYPES.CloseParen:
          case TOKEN_TYPES.CloseSquareBracket:
            break;
          default: {
            ++cursorPosition;
            const num = consumeWhile(isInteger);
            tokens.push(
              new Token(`${char}${num}`, num.length > 0 ? TOKEN_TYPES.NumericLiteral : TOKEN_TYPES.UnaryOperator)
            );
            continue;
          }
        }
      }
      for (const [seq, type] of ORDERED_MAPPING_TABLE) {
        if (seq === "}}" && curlyBracketDepth > 0) {
          continue;
        }
        const slice22 = src.slice(cursorPosition, cursorPosition + seq.length);
        if (slice22 === seq) {
          tokens.push(new Token(seq, type));
          if (type === TOKEN_TYPES.OpenExpression) {
            curlyBracketDepth = 0;
          } else if (type === TOKEN_TYPES.OpenCurlyBracket) {
            ++curlyBracketDepth;
          } else if (type === TOKEN_TYPES.CloseCurlyBracket) {
            --curlyBracketDepth;
          }
          cursorPosition += seq.length;
          continue main;
        }
      }
      if (char === "'" || char === '"') {
        ++cursorPosition;
        const str = consumeWhile((c) => c !== char);
        tokens.push(new Token(str, TOKEN_TYPES.StringLiteral));
        ++cursorPosition;
        continue;
      }
      if (isInteger(char)) {
        let num = consumeWhile(isInteger);
        if (src[cursorPosition] === "." && isInteger(src[cursorPosition + 1])) {
          ++cursorPosition;
          const frac = consumeWhile(isInteger);
          num = `${num}.${frac}`;
        }
        tokens.push(new Token(num, TOKEN_TYPES.NumericLiteral));
        continue;
      }
      if (isWord(char)) {
        const word = consumeWhile(isWord);
        tokens.push(new Token(word, TOKEN_TYPES.Identifier));
        continue;
      }
      throw new SyntaxError(`Unexpected character: ${char}`);
    }
  return tokens;
}
var Statement = class {
  type = "Statement";
};
var Program = class extends Statement {
  constructor(body) {
    super();
    this.body = body;
  }
  type = "Program";
};
var If = class extends Statement {
  constructor(test, body, alternate) {
    super();
    this.test = test;
    this.body = body;
    this.alternate = alternate;
  }
  type = "If";
};
var For = class extends Statement {
  constructor(loopvar, iterable, body, defaultBlock) {
    super();
    this.loopvar = loopvar;
    this.iterable = iterable;
    this.body = body;
    this.defaultBlock = defaultBlock;
  }
  type = "For";
};
var Break = class extends Statement {
  type = "Break";
};
var Continue = class extends Statement {
  type = "Continue";
};
var SetStatement = class extends Statement {
  constructor(assignee, value, body) {
    super();
    this.assignee = assignee;
    this.value = value;
    this.body = body;
  }
  type = "Set";
};
var Macro = class extends Statement {
  constructor(name, args, body) {
    super();
    this.name = name;
    this.args = args;
    this.body = body;
  }
  type = "Macro";
};
var Comment = class extends Statement {
  constructor(value) {
    super();
    this.value = value;
  }
  type = "Comment";
};
var Expression = class extends Statement {
  type = "Expression";
};
var MemberExpression = class extends Expression {
  constructor(object, property, computed) {
    super();
    this.object = object;
    this.property = property;
    this.computed = computed;
  }
  type = "MemberExpression";
};
var CallExpression = class extends Expression {
  constructor(callee, args) {
    super();
    this.callee = callee;
    this.args = args;
  }
  type = "CallExpression";
};
var Identifier = class extends Expression {
  /**
   * @param {string} value The name of the identifier
   */
  constructor(value) {
    super();
    this.value = value;
  }
  type = "Identifier";
};
var Literal = class extends Expression {
  constructor(value) {
    super();
    this.value = value;
  }
  type = "Literal";
};
var IntegerLiteral = class extends Literal {
  type = "IntegerLiteral";
};
var FloatLiteral = class extends Literal {
  type = "FloatLiteral";
};
var StringLiteral = class extends Literal {
  type = "StringLiteral";
};
var ArrayLiteral = class extends Literal {
  type = "ArrayLiteral";
};
var TupleLiteral = class extends Literal {
  type = "TupleLiteral";
};
var ObjectLiteral = class extends Literal {
  type = "ObjectLiteral";
};
var BinaryExpression = class extends Expression {
  constructor(operator, left, right) {
    super();
    this.operator = operator;
    this.left = left;
    this.right = right;
  }
  type = "BinaryExpression";
};
var FilterExpression = class extends Expression {
  constructor(operand, filter) {
    super();
    this.operand = operand;
    this.filter = filter;
  }
  type = "FilterExpression";
};
var FilterStatement = class extends Statement {
  constructor(filter, body) {
    super();
    this.filter = filter;
    this.body = body;
  }
  type = "FilterStatement";
};
var SelectExpression = class extends Expression {
  constructor(lhs, test) {
    super();
    this.lhs = lhs;
    this.test = test;
  }
  type = "SelectExpression";
};
var TestExpression = class extends Expression {
  constructor(operand, negate, test) {
    super();
    this.operand = operand;
    this.negate = negate;
    this.test = test;
  }
  type = "TestExpression";
};
var UnaryExpression = class extends Expression {
  constructor(operator, argument) {
    super();
    this.operator = operator;
    this.argument = argument;
  }
  type = "UnaryExpression";
};
var SliceExpression = class extends Expression {
  constructor(start = void 0, stop = void 0, step = void 0) {
    super();
    this.start = start;
    this.stop = stop;
    this.step = step;
  }
  type = "SliceExpression";
};
var KeywordArgumentExpression = class extends Expression {
  constructor(key, value) {
    super();
    this.key = key;
    this.value = value;
  }
  type = "KeywordArgumentExpression";
};
var SpreadExpression = class extends Expression {
  constructor(argument) {
    super();
    this.argument = argument;
  }
  type = "SpreadExpression";
};
var CallStatement = class extends Statement {
  constructor(call, callerArgs, body) {
    super();
    this.call = call;
    this.callerArgs = callerArgs;
    this.body = body;
  }
  type = "CallStatement";
};
var Ternary = class extends Expression {
  constructor(condition, trueExpr, falseExpr) {
    super();
    this.condition = condition;
    this.trueExpr = trueExpr;
    this.falseExpr = falseExpr;
  }
  type = "Ternary";
};
function parse(tokens) {
  const program = new Program([]);
  let current = 0;
  function expect(type, error) {
    const prev = tokens[current++];
    if (!prev || prev.type !== type) {
      throw new Error(`Parser Error: ${error}. ${prev.type} !== ${type}.`);
    }
    return prev;
  }
  function expectIdentifier(name) {
    if (!isIdentifier(name)) {
      throw new SyntaxError(`Expected ${name}`);
    }
    ++current;
  }
  function parseAny() {
    switch (tokens[current].type) {
      case TOKEN_TYPES.Comment:
        return new Comment(tokens[current++].value);
      case TOKEN_TYPES.Text:
        return parseText();
      case TOKEN_TYPES.OpenStatement:
        return parseJinjaStatement();
      case TOKEN_TYPES.OpenExpression:
        return parseJinjaExpression();
      default:
        throw new SyntaxError(`Unexpected token type: ${tokens[current].type}`);
    }
  }
  function is2(...types) {
    return current + types.length <= tokens.length && types.every((type, i) => type === tokens[current + i].type);
  }
  function isStatement(...names) {
    return tokens[current]?.type === TOKEN_TYPES.OpenStatement && tokens[current + 1]?.type === TOKEN_TYPES.Identifier && names.includes(tokens[current + 1]?.value);
  }
  function isIdentifier(...names) {
    return current + names.length <= tokens.length && names.every((name, i) => tokens[current + i].type === "Identifier" && name === tokens[current + i].value);
  }
  function parseText() {
    return new StringLiteral(expect(TOKEN_TYPES.Text, "Expected text token").value);
  }
  function parseJinjaStatement() {
    expect(TOKEN_TYPES.OpenStatement, "Expected opening statement token");
    if (tokens[current].type !== TOKEN_TYPES.Identifier) {
      throw new SyntaxError(`Unknown statement, got ${tokens[current].type}`);
    }
    const name = tokens[current].value;
    let result;
    switch (name) {
      case "set":
        ++current;
        result = parseSetStatement();
        break;
      case "if":
        ++current;
        result = parseIfStatement();
        expect(TOKEN_TYPES.OpenStatement, "Expected {% token");
        expectIdentifier("endif");
        expect(TOKEN_TYPES.CloseStatement, "Expected %} token");
        break;
      case "macro":
        ++current;
        result = parseMacroStatement();
        expect(TOKEN_TYPES.OpenStatement, "Expected {% token");
        expectIdentifier("endmacro");
        expect(TOKEN_TYPES.CloseStatement, "Expected %} token");
        break;
      case "for":
        ++current;
        result = parseForStatement();
        expect(TOKEN_TYPES.OpenStatement, "Expected {% token");
        expectIdentifier("endfor");
        expect(TOKEN_TYPES.CloseStatement, "Expected %} token");
        break;
      case "call": {
        ++current;
        let callerArgs = null;
        if (is2(TOKEN_TYPES.OpenParen)) {
          callerArgs = parseArgs();
        }
        const callee = parsePrimaryExpression();
        if (callee.type !== "Identifier") {
          throw new SyntaxError(`Expected identifier following call statement`);
        }
        const callArgs = parseArgs();
        expect(TOKEN_TYPES.CloseStatement, "Expected closing statement token");
        const body = [];
        while (!isStatement("endcall")) {
          body.push(parseAny());
        }
        expect(TOKEN_TYPES.OpenStatement, "Expected '{%'");
        expectIdentifier("endcall");
        expect(TOKEN_TYPES.CloseStatement, "Expected closing statement token");
        const callExpr = new CallExpression(callee, callArgs);
        result = new CallStatement(callExpr, callerArgs, body);
        break;
      }
      case "break":
        ++current;
        expect(TOKEN_TYPES.CloseStatement, "Expected closing statement token");
        result = new Break();
        break;
      case "continue":
        ++current;
        expect(TOKEN_TYPES.CloseStatement, "Expected closing statement token");
        result = new Continue();
        break;
      case "filter": {
        ++current;
        let filterNode = parsePrimaryExpression();
        if (filterNode instanceof Identifier && is2(TOKEN_TYPES.OpenParen)) {
          filterNode = parseCallExpression(filterNode);
        }
        expect(TOKEN_TYPES.CloseStatement, "Expected closing statement token");
        const filterBody = [];
        while (!isStatement("endfilter")) {
          filterBody.push(parseAny());
        }
        expect(TOKEN_TYPES.OpenStatement, "Expected '{%'");
        expectIdentifier("endfilter");
        expect(TOKEN_TYPES.CloseStatement, "Expected '%}'");
        result = new FilterStatement(filterNode, filterBody);
        break;
      }
      default:
        throw new SyntaxError(`Unknown statement type: ${name}`);
    }
    return result;
  }
  function parseJinjaExpression() {
    expect(TOKEN_TYPES.OpenExpression, "Expected opening expression token");
    const result = parseExpression();
    expect(TOKEN_TYPES.CloseExpression, "Expected closing expression token");
    return result;
  }
  function parseSetStatement() {
    const left = parseExpressionSequence();
    let value = null;
    const body = [];
    if (is2(TOKEN_TYPES.Equals)) {
      ++current;
      value = parseExpressionSequence();
    } else {
      expect(TOKEN_TYPES.CloseStatement, "Expected %} token");
      while (!isStatement("endset")) {
        body.push(parseAny());
      }
      expect(TOKEN_TYPES.OpenStatement, "Expected {% token");
      expectIdentifier("endset");
    }
    expect(TOKEN_TYPES.CloseStatement, "Expected closing statement token");
    return new SetStatement(left, value, body);
  }
  function parseIfStatement() {
    const test = parseExpression();
    expect(TOKEN_TYPES.CloseStatement, "Expected closing statement token");
    const body = [];
    const alternate = [];
    while (!isStatement("elif", "else", "endif")) {
      body.push(parseAny());
    }
    if (isStatement("elif")) {
      ++current;
      ++current;
      const result = parseIfStatement();
      alternate.push(result);
    } else if (isStatement("else")) {
      ++current;
      ++current;
      expect(TOKEN_TYPES.CloseStatement, "Expected closing statement token");
      while (!isStatement("endif")) {
        alternate.push(parseAny());
      }
    }
    return new If(test, body, alternate);
  }
  function parseMacroStatement() {
    const name = parsePrimaryExpression();
    if (name.type !== "Identifier") {
      throw new SyntaxError(`Expected identifier following macro statement`);
    }
    const args = parseArgs();
    expect(TOKEN_TYPES.CloseStatement, "Expected closing statement token");
    const body = [];
    while (!isStatement("endmacro")) {
      body.push(parseAny());
    }
    return new Macro(name, args, body);
  }
  function parseExpressionSequence(primary = false) {
    const fn2 = primary ? parsePrimaryExpression : parseExpression;
    const expressions = [fn2()];
    const isTuple = is2(TOKEN_TYPES.Comma);
    while (isTuple) {
      ++current;
      expressions.push(fn2());
      if (!is2(TOKEN_TYPES.Comma)) {
        break;
      }
    }
    return isTuple ? new TupleLiteral(expressions) : expressions[0];
  }
  function parseForStatement() {
    const loopVariable = parseExpressionSequence(true);
    if (!(loopVariable instanceof Identifier || loopVariable instanceof TupleLiteral)) {
      throw new SyntaxError(`Expected identifier/tuple for the loop variable, got ${loopVariable.type} instead`);
    }
    if (!isIdentifier("in")) {
      throw new SyntaxError("Expected `in` keyword following loop variable");
    }
    ++current;
    const iterable = parseExpression();
    expect(TOKEN_TYPES.CloseStatement, "Expected closing statement token");
    const body = [];
    while (!isStatement("endfor", "else")) {
      body.push(parseAny());
    }
    const alternative = [];
    if (isStatement("else")) {
      ++current;
      ++current;
      expect(TOKEN_TYPES.CloseStatement, "Expected closing statement token");
      while (!isStatement("endfor")) {
        alternative.push(parseAny());
      }
    }
    return new For(loopVariable, iterable, body, alternative);
  }
  function parseExpression() {
    return parseIfExpression();
  }
  function parseIfExpression() {
    const a = parseLogicalOrExpression();
    if (isIdentifier("if")) {
      ++current;
      const test = parseLogicalOrExpression();
      if (isIdentifier("else")) {
        ++current;
        const falseExpr = parseIfExpression();
        return new Ternary(test, a, falseExpr);
      } else {
        return new SelectExpression(a, test);
      }
    }
    return a;
  }
  function parseLogicalOrExpression() {
    let left = parseLogicalAndExpression();
    while (isIdentifier("or")) {
      const operator = tokens[current];
      ++current;
      const right = parseLogicalAndExpression();
      left = new BinaryExpression(operator, left, right);
    }
    return left;
  }
  function parseLogicalAndExpression() {
    let left = parseLogicalNegationExpression();
    while (isIdentifier("and")) {
      const operator = tokens[current];
      ++current;
      const right = parseLogicalNegationExpression();
      left = new BinaryExpression(operator, left, right);
    }
    return left;
  }
  function parseLogicalNegationExpression() {
    let right;
    while (isIdentifier("not")) {
      const operator = tokens[current];
      ++current;
      const arg = parseLogicalNegationExpression();
      right = new UnaryExpression(operator, arg);
    }
    return right ?? parseComparisonExpression();
  }
  function parseComparisonExpression() {
    let left = parseAdditiveExpression();
    while (true) {
      let operator;
      if (isIdentifier("not", "in")) {
        operator = new Token("not in", TOKEN_TYPES.Identifier);
        current += 2;
      } else if (isIdentifier("in")) {
        operator = tokens[current++];
      } else if (is2(TOKEN_TYPES.ComparisonBinaryOperator)) {
        operator = tokens[current++];
      } else {
        break;
      }
      const right = parseAdditiveExpression();
      left = new BinaryExpression(operator, left, right);
    }
    return left;
  }
  function parseAdditiveExpression() {
    let left = parseMultiplicativeExpression();
    while (is2(TOKEN_TYPES.AdditiveBinaryOperator)) {
      const operator = tokens[current];
      ++current;
      const right = parseMultiplicativeExpression();
      left = new BinaryExpression(operator, left, right);
    }
    return left;
  }
  function parseCallMemberExpression() {
    const member = parseMemberExpression(parsePrimaryExpression());
    if (is2(TOKEN_TYPES.OpenParen)) {
      return parseCallExpression(member);
    }
    return member;
  }
  function parseCallExpression(callee) {
    let expression = new CallExpression(callee, parseArgs());
    expression = parseMemberExpression(expression);
    if (is2(TOKEN_TYPES.OpenParen)) {
      expression = parseCallExpression(expression);
    }
    return expression;
  }
  function parseArgs() {
    expect(TOKEN_TYPES.OpenParen, "Expected opening parenthesis for arguments list");
    const args = parseArgumentsList();
    expect(TOKEN_TYPES.CloseParen, "Expected closing parenthesis for arguments list");
    return args;
  }
  function parseArgumentsList() {
    const args = [];
    while (!is2(TOKEN_TYPES.CloseParen)) {
      let argument;
      if (tokens[current].type === TOKEN_TYPES.MultiplicativeBinaryOperator && tokens[current].value === "*") {
        ++current;
        const expr = parseExpression();
        argument = new SpreadExpression(expr);
      } else {
        argument = parseExpression();
        if (is2(TOKEN_TYPES.Equals)) {
          ++current;
          if (!(argument instanceof Identifier)) {
            throw new SyntaxError(`Expected identifier for keyword argument`);
          }
          const value = parseExpression();
          argument = new KeywordArgumentExpression(argument, value);
        }
      }
      args.push(argument);
      if (is2(TOKEN_TYPES.Comma)) {
        ++current;
      }
    }
    return args;
  }
  function parseMemberExpressionArgumentsList() {
    const slices = [];
    let isSlice = false;
    while (!is2(TOKEN_TYPES.CloseSquareBracket)) {
      if (is2(TOKEN_TYPES.Colon)) {
        slices.push(void 0);
        ++current;
        isSlice = true;
      } else {
        slices.push(parseExpression());
        if (is2(TOKEN_TYPES.Colon)) {
          ++current;
          isSlice = true;
        }
      }
    }
    if (slices.length === 0) {
      throw new SyntaxError(`Expected at least one argument for member/slice expression`);
    }
    if (isSlice) {
      if (slices.length > 3) {
        throw new SyntaxError(`Expected 0-3 arguments for slice expression`);
      }
      return new SliceExpression(...slices);
    }
    return slices[0];
  }
  function parseMemberExpression(object) {
    while (is2(TOKEN_TYPES.Dot) || is2(TOKEN_TYPES.OpenSquareBracket)) {
      const operator = tokens[current];
      ++current;
      let property;
      const computed = operator.type === TOKEN_TYPES.OpenSquareBracket;
      if (computed) {
        property = parseMemberExpressionArgumentsList();
        expect(TOKEN_TYPES.CloseSquareBracket, "Expected closing square bracket");
      } else {
        property = parsePrimaryExpression();
        if (property.type !== "Identifier") {
          throw new SyntaxError(`Expected identifier following dot operator`);
        }
      }
      object = new MemberExpression(object, property, computed);
    }
    return object;
  }
  function parseMultiplicativeExpression() {
    let left = parseTestExpression();
    while (is2(TOKEN_TYPES.MultiplicativeBinaryOperator)) {
      const operator = tokens[current++];
      const right = parseTestExpression();
      left = new BinaryExpression(operator, left, right);
    }
    return left;
  }
  function parseTestExpression() {
    let operand = parseFilterExpression();
    while (isIdentifier("is")) {
      ++current;
      const negate = isIdentifier("not");
      if (negate) {
        ++current;
      }
      const filter = parsePrimaryExpression();
      if (!(filter instanceof Identifier)) {
        throw new SyntaxError(`Expected identifier for the test`);
      }
      operand = new TestExpression(operand, negate, filter);
    }
    return operand;
  }
  function parseFilterExpression() {
    let operand = parseCallMemberExpression();
    while (is2(TOKEN_TYPES.Pipe)) {
      ++current;
      let filter = parsePrimaryExpression();
      if (!(filter instanceof Identifier)) {
        throw new SyntaxError(`Expected identifier for the filter`);
      }
      if (is2(TOKEN_TYPES.OpenParen)) {
        filter = parseCallExpression(filter);
      }
      operand = new FilterExpression(operand, filter);
    }
    return operand;
  }
  function parsePrimaryExpression() {
    const token = tokens[current++];
    switch (token.type) {
      case TOKEN_TYPES.NumericLiteral: {
        const num = token.value;
        return num.includes(".") ? new FloatLiteral(Number(num)) : new IntegerLiteral(Number(num));
      }
      case TOKEN_TYPES.StringLiteral: {
        let value = token.value;
        while (is2(TOKEN_TYPES.StringLiteral)) {
          value += tokens[current++].value;
        }
        return new StringLiteral(value);
      }
      case TOKEN_TYPES.Identifier:
        return new Identifier(token.value);
      case TOKEN_TYPES.OpenParen: {
        const expression = parseExpressionSequence();
        expect(TOKEN_TYPES.CloseParen, "Expected closing parenthesis, got ${tokens[current].type} instead.");
        return expression;
      }
      case TOKEN_TYPES.OpenSquareBracket: {
        const values = [];
        while (!is2(TOKEN_TYPES.CloseSquareBracket)) {
          values.push(parseExpression());
          if (is2(TOKEN_TYPES.Comma)) {
            ++current;
          }
        }
        ++current;
        return new ArrayLiteral(values);
      }
      case TOKEN_TYPES.OpenCurlyBracket: {
        const values = /* @__PURE__ */ new Map();
        while (!is2(TOKEN_TYPES.CloseCurlyBracket)) {
          const key = parseExpression();
          expect(TOKEN_TYPES.Colon, "Expected colon between key and value in object literal");
          const value = parseExpression();
          values.set(key, value);
          if (is2(TOKEN_TYPES.Comma)) {
            ++current;
          }
        }
        ++current;
        return new ObjectLiteral(values);
      }
      default:
        throw new SyntaxError(`Unexpected token: ${token.type}`);
    }
  }
  while (current < tokens.length) {
    program.body.push(parseAny());
  }
  return program;
}
function range(start, stop, step = 1) {
  if (stop === void 0) {
    stop = start;
    start = 0;
  }
  if (step === 0) {
    throw new Error("range() step must not be zero");
  }
  const result = [];
  if (step > 0) {
    for (let i = start; i < stop; i += step) {
      result.push(i);
    }
  } else {
    for (let i = start; i > stop; i += step) {
      result.push(i);
    }
  }
  return result;
}
function slice(array, start, stop, step = 1) {
  const direction = Math.sign(step);
  if (direction >= 0) {
    start = (start ??= 0) < 0 ? Math.max(array.length + start, 0) : Math.min(start, array.length);
    stop = (stop ??= array.length) < 0 ? Math.max(array.length + stop, 0) : Math.min(stop, array.length);
  } else {
    start = (start ??= array.length - 1) < 0 ? Math.max(array.length + start, -1) : Math.min(start, array.length - 1);
    stop = (stop ??= -1) < -1 ? Math.max(array.length + stop, -1) : Math.min(stop, array.length - 1);
  }
  const result = [];
  for (let i = start; direction * i < direction * stop; i += step) {
    result.push(array[i]);
  }
  return result;
}
function titleCase(value) {
  return value.replace(/\b\w/g, (c) => c.toUpperCase());
}
function strftime_now(format2) {
  return strftime(/* @__PURE__ */ new Date(), format2);
}
function strftime(date, format2) {
  const monthFormatterLong = new Intl.DateTimeFormat(void 0, { month: "long" });
  const monthFormatterShort = new Intl.DateTimeFormat(void 0, { month: "short" });
  const pad2 = (n) => n < 10 ? "0" + n : n.toString();
  return format2.replace(/%[YmdbBHM%]/g, (token) => {
    switch (token) {
      case "%Y":
        return date.getFullYear().toString();
      case "%m":
        return pad2(date.getMonth() + 1);
      case "%d":
        return pad2(date.getDate());
      case "%b":
        return monthFormatterShort.format(date);
      case "%B":
        return monthFormatterLong.format(date);
      case "%H":
        return pad2(date.getHours());
      case "%M":
        return pad2(date.getMinutes());
      case "%%":
        return "%";
      default:
        return token;
    }
  });
}
function escapeRegExp(s) {
  return s.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
}
function replace(str, oldvalue, newvalue, count2) {
  if (count2 === 0)
    return str;
  let remaining = count2 == null || count2 < 0 ? Infinity : count2;
  const pattern = oldvalue.length === 0 ? new RegExp("(?=)", "gu") : new RegExp(escapeRegExp(oldvalue), "gu");
  return str.replaceAll(pattern, (match) => {
    if (remaining > 0) {
      --remaining;
      return newvalue;
    }
    return match;
  });
}
var BreakControl = class extends Error {
};
var ContinueControl = class extends Error {
};
var RuntimeValue = class {
  type = "RuntimeValue";
  value;
  /**
   * A collection of built-in functions for this type.
   */
  builtins = /* @__PURE__ */ new Map();
  /**
   * Creates a new RuntimeValue.
   */
  constructor(value = void 0) {
    this.value = value;
  }
  /**
   * Determines truthiness or falsiness of the runtime value.
   * This function should be overridden by subclasses if it has custom truthiness criteria.
   * @returns {BooleanValue} BooleanValue(true) if the value is truthy, BooleanValue(false) otherwise.
   */
  __bool__() {
    return new BooleanValue(!!this.value);
  }
  toString() {
    return String(this.value);
  }
};
var IntegerValue = class extends RuntimeValue {
  type = "IntegerValue";
};
var FloatValue = class extends RuntimeValue {
  type = "FloatValue";
  toString() {
    return this.value % 1 === 0 ? this.value.toFixed(1) : this.value.toString();
  }
};
var StringValue = class extends RuntimeValue {
  type = "StringValue";
  builtins = /* @__PURE__ */ new Map([
    [
      "upper",
      new FunctionValue(() => {
        return new StringValue(this.value.toUpperCase());
      })
    ],
    [
      "lower",
      new FunctionValue(() => {
        return new StringValue(this.value.toLowerCase());
      })
    ],
    [
      "strip",
      new FunctionValue(() => {
        return new StringValue(this.value.trim());
      })
    ],
    [
      "title",
      new FunctionValue(() => {
        return new StringValue(titleCase(this.value));
      })
    ],
    [
      "capitalize",
      new FunctionValue(() => {
        return new StringValue(this.value.charAt(0).toUpperCase() + this.value.slice(1));
      })
    ],
    ["length", new IntegerValue(this.value.length)],
    [
      "rstrip",
      new FunctionValue(() => {
        return new StringValue(this.value.trimEnd());
      })
    ],
    [
      "lstrip",
      new FunctionValue(() => {
        return new StringValue(this.value.trimStart());
      })
    ],
    [
      "startswith",
      new FunctionValue((args) => {
        if (args.length === 0) {
          throw new Error("startswith() requires at least one argument");
        }
        const pattern = args[0];
        if (pattern instanceof StringValue) {
          return new BooleanValue(this.value.startsWith(pattern.value));
        } else if (pattern instanceof ArrayValue) {
          for (const item of pattern.value) {
            if (!(item instanceof StringValue)) {
              throw new Error("startswith() tuple elements must be strings");
            }
            if (this.value.startsWith(item.value)) {
              return new BooleanValue(true);
            }
          }
          return new BooleanValue(false);
        }
        throw new Error("startswith() argument must be a string or tuple of strings");
      })
    ],
    [
      "endswith",
      new FunctionValue((args) => {
        if (args.length === 0) {
          throw new Error("endswith() requires at least one argument");
        }
        const pattern = args[0];
        if (pattern instanceof StringValue) {
          return new BooleanValue(this.value.endsWith(pattern.value));
        } else if (pattern instanceof ArrayValue) {
          for (const item of pattern.value) {
            if (!(item instanceof StringValue)) {
              throw new Error("endswith() tuple elements must be strings");
            }
            if (this.value.endsWith(item.value)) {
              return new BooleanValue(true);
            }
          }
          return new BooleanValue(false);
        }
        throw new Error("endswith() argument must be a string or tuple of strings");
      })
    ],
    [
      "split",
      // follows Python's `str.split(sep=None, maxsplit=-1)` function behavior
      // https://docs.python.org/3.13/library/stdtypes.html#str.split
      new FunctionValue((args) => {
        const sep = args[0] ?? new NullValue();
        if (!(sep instanceof StringValue || sep instanceof NullValue)) {
          throw new Error("sep argument must be a string or null");
        }
        const maxsplit = args[1] ?? new IntegerValue(-1);
        if (!(maxsplit instanceof IntegerValue)) {
          throw new Error("maxsplit argument must be a number");
        }
        let result = [];
        if (sep instanceof NullValue) {
          const text = this.value.trimStart();
          for (const { 0: match, index } of text.matchAll(/\S+/g)) {
            if (maxsplit.value !== -1 && result.length >= maxsplit.value && index !== void 0) {
              result.push(match + text.slice(index + match.length));
              break;
            }
            result.push(match);
          }
        } else {
          if (sep.value === "") {
            throw new Error("empty separator");
          }
          result = this.value.split(sep.value);
          if (maxsplit.value !== -1 && result.length > maxsplit.value) {
            result.push(result.splice(maxsplit.value).join(sep.value));
          }
        }
        return new ArrayValue(result.map((part) => new StringValue(part)));
      })
    ],
    [
      "replace",
      new FunctionValue((args) => {
        if (args.length < 2) {
          throw new Error("replace() requires at least two arguments");
        }
        const oldValue = args[0];
        const newValue = args[1];
        if (!(oldValue instanceof StringValue && newValue instanceof StringValue)) {
          throw new Error("replace() arguments must be strings");
        }
        let count2;
        if (args.length > 2) {
          if (args[2].type === "KeywordArgumentsValue") {
            count2 = args[2].value.get("count") ?? new NullValue();
          } else {
            count2 = args[2];
          }
        } else {
          count2 = new NullValue();
        }
        if (!(count2 instanceof IntegerValue || count2 instanceof NullValue)) {
          throw new Error("replace() count argument must be a number or null");
        }
        return new StringValue(replace(this.value, oldValue.value, newValue.value, count2.value));
      })
    ]
  ]);
};
var BooleanValue = class extends RuntimeValue {
  type = "BooleanValue";
};
var NON_ASCII_CHARS = /[\x7f-\uffff]/g;
function makeAsciiSafe(str) {
  return str.replace(NON_ASCII_CHARS, (char) => "\\u" + char.charCodeAt(0).toString(16).padStart(4, "0"));
}
function toJSON(input, options = {}, depth = 0, convertUndefinedToNull = true) {
  const { indent = null, ensureAscii = false, separators = null, sortKeys = false } = options;
  let itemSeparator;
  let keySeparator;
  if (separators) {
    [itemSeparator, keySeparator] = separators;
  } else if (indent) {
    itemSeparator = ",";
    keySeparator = ": ";
  } else {
    itemSeparator = ", ";
    keySeparator = ": ";
  }
  switch (input.type) {
    case "NullValue":
      return "null";
    case "UndefinedValue":
      return convertUndefinedToNull ? "null" : "undefined";
    case "IntegerValue":
    case "FloatValue":
    case "BooleanValue":
      return JSON.stringify(input.value);
    case "StringValue": {
      let result = JSON.stringify(input.value);
      if (ensureAscii) {
        result = makeAsciiSafe(result);
      }
      return result;
    }
    case "ArrayValue":
    case "ObjectValue": {
      const indentValue = indent ? " ".repeat(indent) : "";
      const basePadding = "\n" + indentValue.repeat(depth);
      const childrenPadding = basePadding + indentValue;
      if (input.type === "ArrayValue") {
        const core = input.value.map((x) => toJSON(x, options, depth + 1, convertUndefinedToNull));
        return indent ? `[${childrenPadding}${core.join(`${itemSeparator}${childrenPadding}`)}${basePadding}]` : `[${core.join(itemSeparator)}]`;
      } else {
        let entries = Array.from(input.value.entries());
        if (sortKeys) {
          entries = entries.sort(([a], [b]) => a.localeCompare(b));
        }
        const core = entries.map(([key, value]) => {
          let keyStr = JSON.stringify(key);
          if (ensureAscii) {
            keyStr = makeAsciiSafe(keyStr);
          }
          const v = `${keyStr}${keySeparator}${toJSON(value, options, depth + 1, convertUndefinedToNull)}`;
          return indent ? `${childrenPadding}${v}` : v;
        });
        return indent ? `{${core.join(itemSeparator)}${basePadding}}` : `{${core.join(itemSeparator)}}`;
      }
    }
    default:
      throw new Error(`Cannot convert to JSON: ${input.type}`);
  }
}
var ObjectValue = class extends RuntimeValue {
  type = "ObjectValue";
  /**
   * NOTE: necessary to override since all JavaScript arrays are considered truthy,
   * while only non-empty Python arrays are consider truthy.
   *
   * e.g.,
   *  - JavaScript:  {} && 5 -> 5
   *  - Python:      {} and 5 -> {}
   */
  __bool__() {
    return new BooleanValue(this.value.size > 0);
  }
  builtins = /* @__PURE__ */ new Map([
    [
      "get",
      new FunctionValue(([key, defaultValue]) => {
        if (!(key instanceof StringValue)) {
          throw new Error(`Object key must be a string: got ${key.type}`);
        }
        return this.value.get(key.value) ?? defaultValue ?? new NullValue();
      })
    ],
    ["items", new FunctionValue(() => this.items())],
    ["keys", new FunctionValue(() => this.keys())],
    ["values", new FunctionValue(() => this.values())],
    [
      "dictsort",
      new FunctionValue((args) => {
        let kwargs = /* @__PURE__ */ new Map();
        const positionalArgs = args.filter((arg) => {
          if (arg instanceof KeywordArgumentsValue) {
            kwargs = arg.value;
            return false;
          }
          return true;
        });
        const caseSensitive = positionalArgs.at(0) ?? kwargs.get("case_sensitive") ?? new BooleanValue(false);
        if (!(caseSensitive instanceof BooleanValue)) {
          throw new Error("case_sensitive must be a boolean");
        }
        const by = positionalArgs.at(1) ?? kwargs.get("by") ?? new StringValue("key");
        if (!(by instanceof StringValue)) {
          throw new Error("by must be a string");
        }
        if (!["key", "value"].includes(by.value)) {
          throw new Error("by must be either 'key' or 'value'");
        }
        const reverse = positionalArgs.at(2) ?? kwargs.get("reverse") ?? new BooleanValue(false);
        if (!(reverse instanceof BooleanValue)) {
          throw new Error("reverse must be a boolean");
        }
        const items = Array.from(this.value.entries()).map(([key, value]) => new ArrayValue([new StringValue(key), value])).sort((a, b) => {
          const index = by.value === "key" ? 0 : 1;
          const aVal = a.value[index];
          const bVal = b.value[index];
          const result = compareRuntimeValues(aVal, bVal, caseSensitive.value);
          return reverse.value ? -result : result;
        });
        return new ArrayValue(items);
      })
    ]
  ]);
  items() {
    return new ArrayValue(
      Array.from(this.value.entries()).map(([key, value]) => new ArrayValue([new StringValue(key), value]))
    );
  }
  keys() {
    return new ArrayValue(Array.from(this.value.keys()).map((key) => new StringValue(key)));
  }
  values() {
    return new ArrayValue(Array.from(this.value.values()));
  }
  toString() {
    return toJSON(this, {}, 0, false);
  }
};
var KeywordArgumentsValue = class extends ObjectValue {
  type = "KeywordArgumentsValue";
};
var ArrayValue = class extends RuntimeValue {
  type = "ArrayValue";
  builtins = /* @__PURE__ */ new Map([["length", new IntegerValue(this.value.length)]]);
  /**
   * NOTE: necessary to override since all JavaScript arrays are considered truthy,
   * while only non-empty Python arrays are consider truthy.
   *
   * e.g.,
   *  - JavaScript:  [] && 5 -> 5
   *  - Python:      [] and 5 -> []
   */
  __bool__() {
    return new BooleanValue(this.value.length > 0);
  }
  toString() {
    return toJSON(this, {}, 0, false);
  }
};
var TupleValue = class extends ArrayValue {
  type = "TupleValue";
};
var FunctionValue = class extends RuntimeValue {
  type = "FunctionValue";
};
var NullValue = class extends RuntimeValue {
  type = "NullValue";
};
var UndefinedValue = class extends RuntimeValue {
  type = "UndefinedValue";
};
var Environment = class {
  constructor(parent) {
    this.parent = parent;
  }
  /**
   * The variables declared in this environment.
   */
  variables = /* @__PURE__ */ new Map([
    [
      "namespace",
      new FunctionValue((args) => {
        if (args.length === 0) {
          return new ObjectValue(/* @__PURE__ */ new Map());
        }
        if (args.length !== 1 || !(args[0] instanceof ObjectValue)) {
          throw new Error("`namespace` expects either zero arguments or a single object argument");
        }
        return args[0];
      })
    ]
  ]);
  /**
   * The tests available in this environment.
   */
  tests = /* @__PURE__ */ new Map([
    ["boolean", (operand) => operand.type === "BooleanValue"],
    ["callable", (operand) => operand instanceof FunctionValue],
    [
      "odd",
      (operand) => {
        if (!(operand instanceof IntegerValue)) {
          throw new Error(`cannot odd on ${operand.type}`);
        }
        return operand.value % 2 !== 0;
      }
    ],
    [
      "even",
      (operand) => {
        if (!(operand instanceof IntegerValue)) {
          throw new Error(`cannot even on ${operand.type}`);
        }
        return operand.value % 2 === 0;
      }
    ],
    ["false", (operand) => operand.type === "BooleanValue" && !operand.value],
    ["true", (operand) => operand.type === "BooleanValue" && operand.value],
    ["none", (operand) => operand.type === "NullValue"],
    ["string", (operand) => operand.type === "StringValue"],
    ["number", (operand) => operand instanceof IntegerValue || operand instanceof FloatValue],
    ["integer", (operand) => operand instanceof IntegerValue],
    ["iterable", (operand) => operand.type === "ArrayValue" || operand.type === "StringValue"],
    ["mapping", (operand) => operand instanceof ObjectValue],
    [
      "sequence",
      (operand) => operand instanceof ArrayValue || operand instanceof ObjectValue || operand instanceof StringValue
    ],
    [
      "lower",
      (operand) => {
        const str = operand.value;
        return operand.type === "StringValue" && str === str.toLowerCase();
      }
    ],
    [
      "upper",
      (operand) => {
        const str = operand.value;
        return operand.type === "StringValue" && str === str.toUpperCase();
      }
    ],
    ["none", (operand) => operand.type === "NullValue"],
    ["defined", (operand) => operand.type !== "UndefinedValue"],
    ["undefined", (operand) => operand.type === "UndefinedValue"],
    ["equalto", (a, b) => a.value === b.value],
    ["eq", (a, b) => a.value === b.value]
  ]);
  /**
   * Set the value of a variable in the current environment.
   */
  set(name, value) {
    return this.declareVariable(name, convertToRuntimeValues(value));
  }
  declareVariable(name, value) {
    if (this.variables.has(name)) {
      throw new SyntaxError(`Variable already declared: ${name}`);
    }
    this.variables.set(name, value);
    return value;
  }
  // private assignVariable(name: string, value: AnyRuntimeValue): AnyRuntimeValue {
  // 	const env = this.resolve(name);
  // 	env.variables.set(name, value);
  // 	return value;
  // }
  /**
   * Set variable in the current scope.
   * See https://jinja.palletsprojects.com/en/3.0.x/templates/#assignments for more information.
   */
  setVariable(name, value) {
    this.variables.set(name, value);
    return value;
  }
  /**
   * Resolve the environment in which the variable is declared.
   * @param {string} name The name of the variable.
   * @returns {Environment} The environment in which the variable is declared.
   */
  resolve(name) {
    if (this.variables.has(name)) {
      return this;
    }
    if (this.parent) {
      return this.parent.resolve(name);
    }
    throw new Error(`Unknown variable: ${name}`);
  }
  lookupVariable(name) {
    try {
      return this.resolve(name).variables.get(name) ?? new UndefinedValue();
    } catch {
      return new UndefinedValue();
    }
  }
};
function setupGlobals(env2) {
  env2.set("false", false);
  env2.set("true", true);
  env2.set("none", null);
  env2.set("raise_exception", (args) => {
    throw new Error(args);
  });
  env2.set("range", range);
  env2.set("strftime_now", strftime_now);
  env2.set("True", true);
  env2.set("False", false);
  env2.set("None", null);
}
function getAttributeValue(item, attributePath) {
  const parts = attributePath.split(".");
  let value = item;
  for (const part of parts) {
    if (value instanceof ObjectValue) {
      value = value.value.get(part) ?? new UndefinedValue();
    } else if (value instanceof ArrayValue) {
      const index = parseInt(part, 10);
      if (!isNaN(index) && index >= 0 && index < value.value.length) {
        value = value.value[index];
      } else {
        return new UndefinedValue();
      }
    } else {
      return new UndefinedValue();
    }
  }
  return value;
}
function compareRuntimeValues(a, b, caseSensitive = false) {
  if (a instanceof NullValue && b instanceof NullValue) {
    return 0;
  }
  if (a instanceof NullValue || b instanceof NullValue) {
    throw new Error(`Cannot compare ${a.type} with ${b.type}`);
  }
  if (a instanceof UndefinedValue && b instanceof UndefinedValue) {
    return 0;
  }
  if (a instanceof UndefinedValue || b instanceof UndefinedValue) {
    throw new Error(`Cannot compare ${a.type} with ${b.type}`);
  }
  const isNumericLike = (v) => v instanceof IntegerValue || v instanceof FloatValue || v instanceof BooleanValue;
  const getNumericValue = (v) => {
    if (v instanceof BooleanValue) {
      return v.value ? 1 : 0;
    }
    return v.value;
  };
  if (isNumericLike(a) && isNumericLike(b)) {
    const aNum = getNumericValue(a);
    const bNum = getNumericValue(b);
    return aNum < bNum ? -1 : aNum > bNum ? 1 : 0;
  }
  if (a.type !== b.type) {
    throw new Error(`Cannot compare different types: ${a.type} and ${b.type}`);
  }
  switch (a.type) {
    case "StringValue": {
      let aStr = a.value;
      let bStr = b.value;
      if (!caseSensitive) {
        aStr = aStr.toLowerCase();
        bStr = bStr.toLowerCase();
      }
      return aStr < bStr ? -1 : aStr > bStr ? 1 : 0;
    }
    default:
      throw new Error(`Cannot compare type: ${a.type}`);
  }
}
var Interpreter = class {
  global;
  constructor(env2) {
    this.global = env2 ?? new Environment();
  }
  /**
   * Run the program.
   */
  run(program) {
    return this.evaluate(program, this.global);
  }
  /**
   * Evaluates expressions following the binary operation type.
   */
  evaluateBinaryExpression(node, environment) {
    const left = this.evaluate(node.left, environment);
    switch (node.operator.value) {
      case "and":
        return left.__bool__().value ? this.evaluate(node.right, environment) : left;
      case "or":
        return left.__bool__().value ? left : this.evaluate(node.right, environment);
    }
    const right = this.evaluate(node.right, environment);
    switch (node.operator.value) {
      case "==":
        return new BooleanValue(left.value == right.value);
      case "!=":
        return new BooleanValue(left.value != right.value);
    }
    if (left instanceof UndefinedValue || right instanceof UndefinedValue) {
      if (right instanceof UndefinedValue && ["in", "not in"].includes(node.operator.value)) {
        return new BooleanValue(node.operator.value === "not in");
      }
      throw new Error(`Cannot perform operation ${node.operator.value} on undefined values`);
    } else if (left instanceof NullValue || right instanceof NullValue) {
      throw new Error("Cannot perform operation on null values");
    } else if (node.operator.value === "~") {
      return new StringValue(left.value.toString() + right.value.toString());
    } else if ((left instanceof IntegerValue || left instanceof FloatValue) && (right instanceof IntegerValue || right instanceof FloatValue)) {
      const a = left.value, b = right.value;
      switch (node.operator.value) {
        case "+":
        case "-":
        case "*": {
          const res = node.operator.value === "+" ? a + b : node.operator.value === "-" ? a - b : a * b;
          const isFloat = left instanceof FloatValue || right instanceof FloatValue;
          return isFloat ? new FloatValue(res) : new IntegerValue(res);
        }
        case "/":
          return new FloatValue(a / b);
        case "%": {
          const rem = a % b;
          const isFloat = left instanceof FloatValue || right instanceof FloatValue;
          return isFloat ? new FloatValue(rem) : new IntegerValue(rem);
        }
        case "<":
          return new BooleanValue(a < b);
        case ">":
          return new BooleanValue(a > b);
        case ">=":
          return new BooleanValue(a >= b);
        case "<=":
          return new BooleanValue(a <= b);
      }
    } else if (left instanceof ArrayValue && right instanceof ArrayValue) {
      switch (node.operator.value) {
        case "+":
          return new ArrayValue(left.value.concat(right.value));
      }
    } else if (right instanceof ArrayValue) {
      const member = right.value.find((x) => x.value === left.value) !== void 0;
      switch (node.operator.value) {
        case "in":
          return new BooleanValue(member);
        case "not in":
          return new BooleanValue(!member);
      }
    }
    if (left instanceof StringValue || right instanceof StringValue) {
      switch (node.operator.value) {
        case "+":
          return new StringValue(left.value.toString() + right.value.toString());
      }
    }
    if (left instanceof StringValue && right instanceof StringValue) {
      switch (node.operator.value) {
        case "in":
          return new BooleanValue(right.value.includes(left.value));
        case "not in":
          return new BooleanValue(!right.value.includes(left.value));
      }
    }
    if (left instanceof StringValue && right instanceof ObjectValue) {
      switch (node.operator.value) {
        case "in":
          return new BooleanValue(right.value.has(left.value));
        case "not in":
          return new BooleanValue(!right.value.has(left.value));
      }
    }
    throw new SyntaxError(`Unknown operator "${node.operator.value}" between ${left.type} and ${right.type}`);
  }
  evaluateArguments(args, environment) {
    const positionalArguments = [];
    const keywordArguments = /* @__PURE__ */ new Map();
    for (const argument of args) {
      if (argument.type === "SpreadExpression") {
        const spreadNode = argument;
        const val = this.evaluate(spreadNode.argument, environment);
        if (!(val instanceof ArrayValue)) {
          throw new Error(`Cannot unpack non-iterable type: ${val.type}`);
        }
        for (const item of val.value) {
          positionalArguments.push(item);
        }
      } else if (argument.type === "KeywordArgumentExpression") {
        const kwarg = argument;
        keywordArguments.set(kwarg.key.value, this.evaluate(kwarg.value, environment));
      } else {
        if (keywordArguments.size > 0) {
          throw new Error("Positional arguments must come before keyword arguments");
        }
        positionalArguments.push(this.evaluate(argument, environment));
      }
    }
    return [positionalArguments, keywordArguments];
  }
  applyFilter(operand, filterNode, environment) {
    if (filterNode.type === "Identifier") {
      const filter = filterNode;
      if (filter.value === "safe") {
        return operand;
      }
      if (filter.value === "tojson") {
        return new StringValue(toJSON(operand, {}));
      }
      if (operand instanceof ArrayValue) {
        switch (filter.value) {
          case "list":
            return operand;
          case "first":
            return operand.value[0];
          case "last":
            return operand.value[operand.value.length - 1];
          case "length":
            return new IntegerValue(operand.value.length);
          case "reverse":
            return new ArrayValue(operand.value.slice().reverse());
          case "sort": {
            return new ArrayValue(operand.value.slice().sort((a, b) => compareRuntimeValues(a, b, false)));
          }
          case "join":
            return new StringValue(operand.value.map((x) => x.value).join(""));
          case "string":
            return new StringValue(toJSON(operand, {}, 0, false));
          case "unique": {
            const seen = /* @__PURE__ */ new Set();
            const output = [];
            for (const item of operand.value) {
              if (!seen.has(item.value)) {
                seen.add(item.value);
                output.push(item);
              }
            }
            return new ArrayValue(output);
          }
          default:
            throw new Error(`Unknown ArrayValue filter: ${filter.value}`);
        }
      } else if (operand instanceof StringValue) {
        switch (filter.value) {
          case "length":
          case "upper":
          case "lower":
          case "title":
          case "capitalize": {
            const builtin = operand.builtins.get(filter.value);
            if (builtin instanceof FunctionValue) {
              return builtin.value(
                /* no arguments */
                [],
                environment
              );
            } else if (builtin instanceof IntegerValue) {
              return builtin;
            } else {
              throw new Error(`Unknown StringValue filter: ${filter.value}`);
            }
          }
          case "trim":
            return new StringValue(operand.value.trim());
          case "indent":
            return new StringValue(
              operand.value.split("\n").map(
                (x, i) => (
                  // By default, don't indent the first line or empty lines
                  i === 0 || x.length === 0 ? x : "    " + x
                )
              ).join("\n")
            );
          case "join":
          case "string":
            return operand;
          case "int": {
            const val = parseInt(operand.value, 10);
            return new IntegerValue(isNaN(val) ? 0 : val);
          }
          case "float": {
            const val = parseFloat(operand.value);
            return new FloatValue(isNaN(val) ? 0 : val);
          }
          default:
            throw new Error(`Unknown StringValue filter: ${filter.value}`);
        }
      } else if (operand instanceof IntegerValue || operand instanceof FloatValue) {
        switch (filter.value) {
          case "abs":
            return operand instanceof IntegerValue ? new IntegerValue(Math.abs(operand.value)) : new FloatValue(Math.abs(operand.value));
          case "int":
            return new IntegerValue(Math.floor(operand.value));
          case "float":
            return new FloatValue(operand.value);
          case "string":
            return new StringValue(operand.toString());
          default:
            throw new Error(`Unknown NumericValue filter: ${filter.value}`);
        }
      } else if (operand instanceof ObjectValue) {
        switch (filter.value) {
          case "items":
            return new ArrayValue(
              Array.from(operand.value.entries()).map(([key, value]) => new ArrayValue([new StringValue(key), value]))
            );
          case "length":
            return new IntegerValue(operand.value.size);
          default: {
            const builtin = operand.builtins.get(filter.value);
            if (builtin) {
              if (builtin instanceof FunctionValue) {
                return builtin.value([], environment);
              }
              return builtin;
            }
            throw new Error(`Unknown ObjectValue filter: ${filter.value}`);
          }
        }
      } else if (operand instanceof BooleanValue) {
        switch (filter.value) {
          case "bool":
            return new BooleanValue(operand.value);
          case "int":
            return new IntegerValue(operand.value ? 1 : 0);
          case "float":
            return new FloatValue(operand.value ? 1 : 0);
          case "string":
            return new StringValue(operand.value ? "true" : "false");
          default:
            throw new Error(`Unknown BooleanValue filter: ${filter.value}`);
        }
      }
      throw new Error(`Cannot apply filter "${filter.value}" to type: ${operand.type}`);
    } else if (filterNode.type === "CallExpression") {
      const filter = filterNode;
      if (filter.callee.type !== "Identifier") {
        throw new Error(`Unknown filter: ${filter.callee.type}`);
      }
      const filterName = filter.callee.value;
      if (filterName === "tojson") {
        const [, kwargs] = this.evaluateArguments(filter.args, environment);
        const indent = kwargs.get("indent") ?? new NullValue();
        if (!(indent instanceof IntegerValue || indent instanceof NullValue)) {
          throw new Error("If set, indent must be a number");
        }
        const ensureAscii = kwargs.get("ensure_ascii") ?? new BooleanValue(false);
        if (!(ensureAscii instanceof BooleanValue)) {
          throw new Error("If set, ensure_ascii must be a boolean");
        }
        const sortKeys = kwargs.get("sort_keys") ?? new BooleanValue(false);
        if (!(sortKeys instanceof BooleanValue)) {
          throw new Error("If set, sort_keys must be a boolean");
        }
        const separatorsArg = kwargs.get("separators") ?? new NullValue();
        let separators = null;
        if (separatorsArg instanceof ArrayValue || separatorsArg instanceof TupleValue) {
          if (separatorsArg.value.length !== 2) {
            throw new Error("separators must be a tuple of two strings");
          }
          const [itemSep, keySep] = separatorsArg.value;
          if (!(itemSep instanceof StringValue) || !(keySep instanceof StringValue)) {
            throw new Error("separators must be a tuple of two strings");
          }
          separators = [itemSep.value, keySep.value];
        } else if (!(separatorsArg instanceof NullValue)) {
          throw new Error("If set, separators must be a tuple of two strings");
        }
        return new StringValue(
          toJSON(operand, {
            indent: indent.value,
            ensureAscii: ensureAscii.value,
            sortKeys: sortKeys.value,
            separators
          })
        );
      } else if (filterName === "join") {
        let value;
        if (operand instanceof StringValue) {
          value = Array.from(operand.value);
        } else if (operand instanceof ArrayValue) {
          value = operand.value.map((x) => x.value);
        } else {
          throw new Error(`Cannot apply filter "${filterName}" to type: ${operand.type}`);
        }
        const [args, kwargs] = this.evaluateArguments(filter.args, environment);
        const separator = args.at(0) ?? kwargs.get("separator") ?? new StringValue("");
        if (!(separator instanceof StringValue)) {
          throw new Error("separator must be a string");
        }
        return new StringValue(value.join(separator.value));
      } else if (filterName === "int" || filterName === "float") {
        const [args, kwargs] = this.evaluateArguments(filter.args, environment);
        const defaultValue = args.at(0) ?? kwargs.get("default") ?? (filterName === "int" ? new IntegerValue(0) : new FloatValue(0));
        if (operand instanceof StringValue) {
          const val = filterName === "int" ? parseInt(operand.value, 10) : parseFloat(operand.value);
          return isNaN(val) ? defaultValue : filterName === "int" ? new IntegerValue(val) : new FloatValue(val);
        } else if (operand instanceof IntegerValue || operand instanceof FloatValue) {
          return operand;
        } else if (operand instanceof BooleanValue) {
          return filterName === "int" ? new IntegerValue(operand.value ? 1 : 0) : new FloatValue(operand.value ? 1 : 0);
        } else {
          throw new Error(`Cannot apply filter "${filterName}" to type: ${operand.type}`);
        }
      } else if (filterName === "default") {
        const [args, kwargs] = this.evaluateArguments(filter.args, environment);
        const defaultValue = args[0] ?? new StringValue("");
        const booleanValue = args[1] ?? kwargs.get("boolean") ?? new BooleanValue(false);
        if (!(booleanValue instanceof BooleanValue)) {
          throw new Error("`default` filter flag must be a boolean");
        }
        if (operand instanceof UndefinedValue || booleanValue.value && !operand.__bool__().value) {
          return defaultValue;
        }
        return operand;
      }
      if (operand instanceof ArrayValue) {
        switch (filterName) {
          case "sort": {
            const [args, kwargs] = this.evaluateArguments(filter.args, environment);
            const reverse = args.at(0) ?? kwargs.get("reverse") ?? new BooleanValue(false);
            if (!(reverse instanceof BooleanValue)) {
              throw new Error("reverse must be a boolean");
            }
            const caseSensitive = args.at(1) ?? kwargs.get("case_sensitive") ?? new BooleanValue(false);
            if (!(caseSensitive instanceof BooleanValue)) {
              throw new Error("case_sensitive must be a boolean");
            }
            const attribute = args.at(2) ?? kwargs.get("attribute") ?? new NullValue();
            if (!(attribute instanceof StringValue || attribute instanceof IntegerValue || attribute instanceof NullValue)) {
              throw new Error("attribute must be a string, integer, or null");
            }
            const getSortValue = (item) => {
              if (attribute instanceof NullValue) {
                return item;
              }
              const attrPath = attribute instanceof IntegerValue ? String(attribute.value) : attribute.value;
              return getAttributeValue(item, attrPath);
            };
            return new ArrayValue(
              operand.value.slice().sort((a, b) => {
                const aVal = getSortValue(a);
                const bVal = getSortValue(b);
                const result = compareRuntimeValues(aVal, bVal, caseSensitive.value);
                return reverse.value ? -result : result;
              })
            );
          }
          case "selectattr":
          case "rejectattr": {
            const select = filterName === "selectattr";
            if (operand.value.some((x) => !(x instanceof ObjectValue))) {
              throw new Error(`\`${filterName}\` can only be applied to array of objects`);
            }
            if (filter.args.some((x) => x.type !== "StringLiteral")) {
              throw new Error(`arguments of \`${filterName}\` must be strings`);
            }
            const [attr, testName, value] = filter.args.map((x) => this.evaluate(x, environment));
            let testFunction;
            if (testName) {
              const test = environment.tests.get(testName.value);
              if (!test) {
                throw new Error(`Unknown test: ${testName.value}`);
              }
              testFunction = test;
            } else {
              testFunction = (...x) => x[0].__bool__().value;
            }
            const filtered = operand.value.filter((item) => {
              const a = item.value.get(attr.value);
              const result = a ? testFunction(a, value) : false;
              return select ? result : !result;
            });
            return new ArrayValue(filtered);
          }
          case "map": {
            const [, kwargs] = this.evaluateArguments(filter.args, environment);
            if (kwargs.has("attribute")) {
              const attr = kwargs.get("attribute");
              if (!(attr instanceof StringValue)) {
                throw new Error("attribute must be a string");
              }
              const defaultValue = kwargs.get("default");
              const mapped = operand.value.map((item) => {
                if (!(item instanceof ObjectValue)) {
                  throw new Error("items in map must be an object");
                }
                const value = getAttributeValue(item, attr.value);
                return value instanceof UndefinedValue ? defaultValue ?? new UndefinedValue() : value;
              });
              return new ArrayValue(mapped);
            } else {
              throw new Error("`map` expressions without `attribute` set are not currently supported.");
            }
          }
        }
        throw new Error(`Unknown ArrayValue filter: ${filterName}`);
      } else if (operand instanceof StringValue) {
        switch (filterName) {
          case "indent": {
            const [args, kwargs] = this.evaluateArguments(filter.args, environment);
            const width = args.at(0) ?? kwargs.get("width") ?? new IntegerValue(4);
            if (!(width instanceof IntegerValue)) {
              throw new Error("width must be a number");
            }
            const first = args.at(1) ?? kwargs.get("first") ?? new BooleanValue(false);
            const blank = args.at(2) ?? kwargs.get("blank") ?? new BooleanValue(false);
            const lines = operand.value.split("\n");
            const indent = " ".repeat(width.value);
            const indented = lines.map(
              (x, i) => !first.value && i === 0 || !blank.value && x.length === 0 ? x : indent + x
            );
            return new StringValue(indented.join("\n"));
          }
          case "replace": {
            const replaceFn = operand.builtins.get("replace");
            if (!(replaceFn instanceof FunctionValue)) {
              throw new Error("replace filter not available");
            }
            const [args, kwargs] = this.evaluateArguments(filter.args, environment);
            return replaceFn.value([...args, new KeywordArgumentsValue(kwargs)], environment);
          }
        }
        throw new Error(`Unknown StringValue filter: ${filterName}`);
      } else if (operand instanceof ObjectValue) {
        const builtin = operand.builtins.get(filterName);
        if (builtin && builtin instanceof FunctionValue) {
          const [args, kwargs] = this.evaluateArguments(filter.args, environment);
          if (kwargs.size > 0) {
            args.push(new KeywordArgumentsValue(kwargs));
          }
          return builtin.value(args, environment);
        }
        throw new Error(`Unknown ObjectValue filter: ${filterName}`);
      } else {
        throw new Error(`Cannot apply filter "${filterName}" to type: ${operand.type}`);
      }
    }
    throw new Error(`Unknown filter: ${filterNode.type}`);
  }
  /**
   * Evaluates expressions following the filter operation type.
   */
  evaluateFilterExpression(node, environment) {
    const operand = this.evaluate(node.operand, environment);
    return this.applyFilter(operand, node.filter, environment);
  }
  /**
   * Evaluates expressions following the test operation type.
   */
  evaluateTestExpression(node, environment) {
    const operand = this.evaluate(node.operand, environment);
    const test = environment.tests.get(node.test.value);
    if (!test) {
      throw new Error(`Unknown test: ${node.test.value}`);
    }
    const result = test(operand);
    return new BooleanValue(node.negate ? !result : result);
  }
  /**
   * Evaluates expressions following the select operation type.
   */
  evaluateSelectExpression(node, environment) {
    const predicate = this.evaluate(node.test, environment);
    if (!predicate.__bool__().value) {
      return new UndefinedValue();
    }
    return this.evaluate(node.lhs, environment);
  }
  /**
   * Evaluates expressions following the unary operation type.
   */
  evaluateUnaryExpression(node, environment) {
    const argument = this.evaluate(node.argument, environment);
    switch (node.operator.value) {
      case "not":
        return new BooleanValue(!argument.value);
      default:
        throw new SyntaxError(`Unknown operator: ${node.operator.value}`);
    }
  }
  evaluateTernaryExpression(node, environment) {
    const cond = this.evaluate(node.condition, environment);
    return cond.__bool__().value ? this.evaluate(node.trueExpr, environment) : this.evaluate(node.falseExpr, environment);
  }
  evalProgram(program, environment) {
    return this.evaluateBlock(program.body, environment);
  }
  evaluateBlock(statements, environment) {
    let result = "";
    for (const statement of statements) {
      const lastEvaluated = this.evaluate(statement, environment);
      if (lastEvaluated.type !== "NullValue" && lastEvaluated.type !== "UndefinedValue") {
        result += lastEvaluated.toString();
      }
    }
    return new StringValue(result);
  }
  evaluateIdentifier(node, environment) {
    return environment.lookupVariable(node.value);
  }
  evaluateCallExpression(expr, environment) {
    const [args, kwargs] = this.evaluateArguments(expr.args, environment);
    if (kwargs.size > 0) {
      args.push(new KeywordArgumentsValue(kwargs));
    }
    const fn2 = this.evaluate(expr.callee, environment);
    if (fn2.type !== "FunctionValue") {
      throw new Error(`Cannot call something that is not a function: got ${fn2.type}`);
    }
    return fn2.value(args, environment);
  }
  evaluateSliceExpression(object, expr, environment) {
    if (!(object instanceof ArrayValue || object instanceof StringValue)) {
      throw new Error("Slice object must be an array or string");
    }
    const start = this.evaluate(expr.start, environment);
    const stop = this.evaluate(expr.stop, environment);
    const step = this.evaluate(expr.step, environment);
    if (!(start instanceof IntegerValue || start instanceof UndefinedValue)) {
      throw new Error("Slice start must be numeric or undefined");
    }
    if (!(stop instanceof IntegerValue || stop instanceof UndefinedValue)) {
      throw new Error("Slice stop must be numeric or undefined");
    }
    if (!(step instanceof IntegerValue || step instanceof UndefinedValue)) {
      throw new Error("Slice step must be numeric or undefined");
    }
    if (object instanceof ArrayValue) {
      return new ArrayValue(slice(object.value, start.value, stop.value, step.value));
    } else {
      return new StringValue(slice(Array.from(object.value), start.value, stop.value, step.value).join(""));
    }
  }
  evaluateMemberExpression(expr, environment) {
    const object = this.evaluate(expr.object, environment);
    let property;
    if (expr.computed) {
      if (expr.property.type === "SliceExpression") {
        return this.evaluateSliceExpression(object, expr.property, environment);
      } else {
        property = this.evaluate(expr.property, environment);
      }
    } else {
      property = new StringValue(expr.property.value);
    }
    let value;
    if (object instanceof ObjectValue) {
      if (!(property instanceof StringValue)) {
        throw new Error(`Cannot access property with non-string: got ${property.type}`);
      }
      value = object.value.get(property.value) ?? object.builtins.get(property.value);
    } else if (object instanceof ArrayValue || object instanceof StringValue) {
      if (property instanceof IntegerValue) {
        value = object.value.at(property.value);
        if (object instanceof StringValue) {
          value = new StringValue(object.value.at(property.value));
        }
      } else if (property instanceof StringValue) {
        value = object.builtins.get(property.value);
      } else {
        throw new Error(`Cannot access property with non-string/non-number: got ${property.type}`);
      }
    } else {
      if (!(property instanceof StringValue)) {
        throw new Error(`Cannot access property with non-string: got ${property.type}`);
      }
      value = object.builtins.get(property.value);
    }
    return value instanceof RuntimeValue ? value : new UndefinedValue();
  }
  evaluateSet(node, environment) {
    const rhs = node.value ? this.evaluate(node.value, environment) : this.evaluateBlock(node.body, environment);
    if (node.assignee.type === "Identifier") {
      const variableName = node.assignee.value;
      environment.setVariable(variableName, rhs);
    } else if (node.assignee.type === "TupleLiteral") {
      const tuple = node.assignee;
      if (!(rhs instanceof ArrayValue)) {
        throw new Error(`Cannot unpack non-iterable type in set: ${rhs.type}`);
      }
      const arr = rhs.value;
      if (arr.length !== tuple.value.length) {
        throw new Error(`Too ${tuple.value.length > arr.length ? "few" : "many"} items to unpack in set`);
      }
      for (let i = 0; i < tuple.value.length; ++i) {
        const elem = tuple.value[i];
        if (elem.type !== "Identifier") {
          throw new Error(`Cannot unpack to non-identifier in set: ${elem.type}`);
        }
        environment.setVariable(elem.value, arr[i]);
      }
    } else if (node.assignee.type === "MemberExpression") {
      const member = node.assignee;
      const object = this.evaluate(member.object, environment);
      if (!(object instanceof ObjectValue)) {
        throw new Error("Cannot assign to member of non-object");
      }
      if (member.property.type !== "Identifier") {
        throw new Error("Cannot assign to member with non-identifier property");
      }
      object.value.set(member.property.value, rhs);
    } else {
      throw new Error(`Invalid LHS inside assignment expression: ${JSON.stringify(node.assignee)}`);
    }
    return new NullValue();
  }
  evaluateIf(node, environment) {
    const test = this.evaluate(node.test, environment);
    return this.evaluateBlock(test.__bool__().value ? node.body : node.alternate, environment);
  }
  evaluateFor(node, environment) {
    const scope = new Environment(environment);
    let test, iterable;
    if (node.iterable.type === "SelectExpression") {
      const select = node.iterable;
      iterable = this.evaluate(select.lhs, scope);
      test = select.test;
    } else {
      iterable = this.evaluate(node.iterable, scope);
    }
    if (!(iterable instanceof ArrayValue || iterable instanceof ObjectValue)) {
      throw new Error(`Expected iterable or object type in for loop: got ${iterable.type}`);
    }
    if (iterable instanceof ObjectValue) {
      iterable = iterable.keys();
    }
    const items = [];
    const scopeUpdateFunctions = [];
    for (let i = 0; i < iterable.value.length; ++i) {
      const loopScope = new Environment(scope);
      const current = iterable.value[i];
      let scopeUpdateFunction;
      if (node.loopvar.type === "Identifier") {
        scopeUpdateFunction = (scope2) => scope2.setVariable(node.loopvar.value, current);
      } else if (node.loopvar.type === "TupleLiteral") {
        const loopvar = node.loopvar;
        if (current.type !== "ArrayValue") {
          throw new Error(`Cannot unpack non-iterable type: ${current.type}`);
        }
        const c = current;
        if (loopvar.value.length !== c.value.length) {
          throw new Error(`Too ${loopvar.value.length > c.value.length ? "few" : "many"} items to unpack`);
        }
        scopeUpdateFunction = (scope2) => {
          for (let j = 0; j < loopvar.value.length; ++j) {
            if (loopvar.value[j].type !== "Identifier") {
              throw new Error(`Cannot unpack non-identifier type: ${loopvar.value[j].type}`);
            }
            scope2.setVariable(loopvar.value[j].value, c.value[j]);
          }
        };
      } else {
        throw new Error(`Invalid loop variable(s): ${node.loopvar.type}`);
      }
      if (test) {
        scopeUpdateFunction(loopScope);
        const testValue = this.evaluate(test, loopScope);
        if (!testValue.__bool__().value) {
          continue;
        }
      }
      items.push(current);
      scopeUpdateFunctions.push(scopeUpdateFunction);
    }
    let result = "";
    let noIteration = true;
    for (let i = 0; i < items.length; ++i) {
      const loop = /* @__PURE__ */ new Map([
        ["index", new IntegerValue(i + 1)],
        ["index0", new IntegerValue(i)],
        ["revindex", new IntegerValue(items.length - i)],
        ["revindex0", new IntegerValue(items.length - i - 1)],
        ["first", new BooleanValue(i === 0)],
        ["last", new BooleanValue(i === items.length - 1)],
        ["length", new IntegerValue(items.length)],
        ["previtem", i > 0 ? items[i - 1] : new UndefinedValue()],
        ["nextitem", i < items.length - 1 ? items[i + 1] : new UndefinedValue()]
      ]);
      scope.setVariable("loop", new ObjectValue(loop));
      scopeUpdateFunctions[i](scope);
      try {
        const evaluated = this.evaluateBlock(node.body, scope);
        result += evaluated.value;
      } catch (err) {
        if (err instanceof ContinueControl) {
          continue;
        }
        if (err instanceof BreakControl) {
          break;
        }
        throw err;
      }
      noIteration = false;
    }
    if (noIteration) {
      const defaultEvaluated = this.evaluateBlock(node.defaultBlock, scope);
      result += defaultEvaluated.value;
    }
    return new StringValue(result);
  }
  /**
   * See https://jinja.palletsprojects.com/en/3.1.x/templates/#macros for more information.
   */
  evaluateMacro(node, environment) {
    environment.setVariable(
      node.name.value,
      new FunctionValue((args, scope) => {
        const macroScope = new Environment(scope);
        args = args.slice();
        let kwargs;
        if (args.at(-1)?.type === "KeywordArgumentsValue") {
          kwargs = args.pop();
        }
        for (let i = 0; i < node.args.length; ++i) {
          const nodeArg = node.args[i];
          const passedArg = args[i];
          if (nodeArg.type === "Identifier") {
            const identifier = nodeArg;
            if (!passedArg) {
              throw new Error(`Missing positional argument: ${identifier.value}`);
            }
            macroScope.setVariable(identifier.value, passedArg);
          } else if (nodeArg.type === "KeywordArgumentExpression") {
            const kwarg = nodeArg;
            const value = passedArg ?? // Try positional arguments first
            kwargs?.value.get(kwarg.key.value) ?? // Look in user-passed kwargs
            this.evaluate(kwarg.value, macroScope);
            macroScope.setVariable(kwarg.key.value, value);
          } else {
            throw new Error(`Unknown argument type: ${nodeArg.type}`);
          }
        }
        return this.evaluateBlock(node.body, macroScope);
      })
    );
    return new NullValue();
  }
  evaluateCallStatement(node, environment) {
    const callerFn = new FunctionValue((callerArgs, callerEnv) => {
      const callBlockEnv = new Environment(callerEnv);
      if (node.callerArgs) {
        for (let i = 0; i < node.callerArgs.length; ++i) {
          const param = node.callerArgs[i];
          if (param.type !== "Identifier") {
            throw new Error(`Caller parameter must be an identifier, got ${param.type}`);
          }
          callBlockEnv.setVariable(param.value, callerArgs[i] ?? new UndefinedValue());
        }
      }
      return this.evaluateBlock(node.body, callBlockEnv);
    });
    const [macroArgs, macroKwargs] = this.evaluateArguments(node.call.args, environment);
    macroArgs.push(new KeywordArgumentsValue(macroKwargs));
    const fn2 = this.evaluate(node.call.callee, environment);
    if (fn2.type !== "FunctionValue") {
      throw new Error(`Cannot call something that is not a function: got ${fn2.type}`);
    }
    const newEnv = new Environment(environment);
    newEnv.setVariable("caller", callerFn);
    return fn2.value(macroArgs, newEnv);
  }
  evaluateFilterStatement(node, environment) {
    const rendered = this.evaluateBlock(node.body, environment);
    return this.applyFilter(rendered, node.filter, environment);
  }
  evaluate(statement, environment) {
    if (!statement)
      return new UndefinedValue();
    switch (statement.type) {
      case "Program":
        return this.evalProgram(statement, environment);
      case "Set":
        return this.evaluateSet(statement, environment);
      case "If":
        return this.evaluateIf(statement, environment);
      case "For":
        return this.evaluateFor(statement, environment);
      case "Macro":
        return this.evaluateMacro(statement, environment);
      case "CallStatement":
        return this.evaluateCallStatement(statement, environment);
      case "Break":
        throw new BreakControl();
      case "Continue":
        throw new ContinueControl();
      case "IntegerLiteral":
        return new IntegerValue(statement.value);
      case "FloatLiteral":
        return new FloatValue(statement.value);
      case "StringLiteral":
        return new StringValue(statement.value);
      case "ArrayLiteral":
        return new ArrayValue(statement.value.map((x) => this.evaluate(x, environment)));
      case "TupleLiteral":
        return new TupleValue(statement.value.map((x) => this.evaluate(x, environment)));
      case "ObjectLiteral": {
        const mapping = /* @__PURE__ */ new Map();
        for (const [key, value] of statement.value) {
          const evaluatedKey = this.evaluate(key, environment);
          if (!(evaluatedKey instanceof StringValue)) {
            throw new Error(`Object keys must be strings: got ${evaluatedKey.type}`);
          }
          mapping.set(evaluatedKey.value, this.evaluate(value, environment));
        }
        return new ObjectValue(mapping);
      }
      case "Identifier":
        return this.evaluateIdentifier(statement, environment);
      case "CallExpression":
        return this.evaluateCallExpression(statement, environment);
      case "MemberExpression":
        return this.evaluateMemberExpression(statement, environment);
      case "UnaryExpression":
        return this.evaluateUnaryExpression(statement, environment);
      case "BinaryExpression":
        return this.evaluateBinaryExpression(statement, environment);
      case "FilterExpression":
        return this.evaluateFilterExpression(statement, environment);
      case "FilterStatement":
        return this.evaluateFilterStatement(statement, environment);
      case "TestExpression":
        return this.evaluateTestExpression(statement, environment);
      case "SelectExpression":
        return this.evaluateSelectExpression(statement, environment);
      case "Ternary":
        return this.evaluateTernaryExpression(statement, environment);
      case "Comment":
        return new NullValue();
      default:
        throw new SyntaxError(`Unknown node type: ${statement.type}`);
    }
  }
};
function convertToRuntimeValues(input) {
  switch (typeof input) {
    case "number":
      return Number.isInteger(input) ? new IntegerValue(input) : new FloatValue(input);
    case "string":
      return new StringValue(input);
    case "boolean":
      return new BooleanValue(input);
    case "undefined":
      return new UndefinedValue();
    case "object":
      if (input === null) {
        return new NullValue();
      } else if (Array.isArray(input)) {
        return new ArrayValue(input.map(convertToRuntimeValues));
      } else {
        return new ObjectValue(
          new Map(Object.entries(input).map(([key, value]) => [key, convertToRuntimeValues(value)]))
        );
      }
    case "function":
      return new FunctionValue((args, _scope) => {
        const result = input(...args.map((x) => x.value)) ?? null;
        return convertToRuntimeValues(result);
      });
    default:
      throw new Error(`Cannot convert to runtime value: ${input}`);
  }
}
var NEWLINE = "\n";
var OPEN_STATEMENT = "{%- ";
var CLOSE_STATEMENT = " -%}";
function getBinaryOperatorPrecedence(expr) {
  switch (expr.operator.type) {
    case "MultiplicativeBinaryOperator":
      return 4;
    case "AdditiveBinaryOperator":
      return 3;
    case "ComparisonBinaryOperator":
      return 2;
    case "Identifier":
      if (expr.operator.value === "and")
        return 1;
      if (expr.operator.value === "in" || expr.operator.value === "not in")
        return 2;
      return 0;
  }
  return 0;
}
function format(program, indent = "	") {
  const indentStr = typeof indent === "number" ? " ".repeat(indent) : indent;
  const body = formatStatements(program.body, 0, indentStr);
  return body.replace(/\n$/, "");
}
function createStatement(...text) {
  return OPEN_STATEMENT + text.join(" ") + CLOSE_STATEMENT;
}
function formatStatements(stmts, depth, indentStr) {
  return stmts.map((stmt) => formatStatement(stmt, depth, indentStr)).join(NEWLINE);
}
function formatStatement(node, depth, indentStr) {
  const pad = indentStr.repeat(depth);
  switch (node.type) {
    case "Program":
      return formatStatements(node.body, depth, indentStr);
    case "If":
      return formatIf(node, depth, indentStr);
    case "For":
      return formatFor(node, depth, indentStr);
    case "Set":
      return formatSet(node, depth, indentStr);
    case "Macro":
      return formatMacro(node, depth, indentStr);
    case "Break":
      return pad + createStatement("break");
    case "Continue":
      return pad + createStatement("continue");
    case "CallStatement":
      return formatCallStatement(node, depth, indentStr);
    case "FilterStatement":
      return formatFilterStatement(node, depth, indentStr);
    case "Comment":
      return pad + "{# " + node.value + " #}";
    default:
      return pad + "{{- " + formatExpression(node) + " -}}";
  }
}
function formatIf(node, depth, indentStr) {
  const pad = indentStr.repeat(depth);
  const clauses = [];
  let current = node;
  while (current) {
    clauses.push({ test: current.test, body: current.body });
    if (current.alternate.length === 1 && current.alternate[0].type === "If") {
      current = current.alternate[0];
    } else {
      break;
    }
  }
  let out = pad + createStatement("if", formatExpression(clauses[0].test)) + NEWLINE + formatStatements(clauses[0].body, depth + 1, indentStr);
  for (let i = 1; i < clauses.length; ++i) {
    out += NEWLINE + pad + createStatement("elif", formatExpression(clauses[i].test)) + NEWLINE + formatStatements(clauses[i].body, depth + 1, indentStr);
  }
  if (current && current.alternate.length > 0) {
    out += NEWLINE + pad + createStatement("else") + NEWLINE + formatStatements(current.alternate, depth + 1, indentStr);
  }
  out += NEWLINE + pad + createStatement("endif");
  return out;
}
function formatFor(node, depth, indentStr) {
  const pad = indentStr.repeat(depth);
  let formattedIterable = "";
  if (node.iterable.type === "SelectExpression") {
    const n = node.iterable;
    formattedIterable = `${formatExpression(n.lhs)} if ${formatExpression(n.test)}`;
  } else {
    formattedIterable = formatExpression(node.iterable);
  }
  let out = pad + createStatement("for", formatExpression(node.loopvar), "in", formattedIterable) + NEWLINE + formatStatements(node.body, depth + 1, indentStr);
  if (node.defaultBlock.length > 0) {
    out += NEWLINE + pad + createStatement("else") + NEWLINE + formatStatements(node.defaultBlock, depth + 1, indentStr);
  }
  out += NEWLINE + pad + createStatement("endfor");
  return out;
}
function formatSet(node, depth, indentStr) {
  const pad = indentStr.repeat(depth);
  const left = formatExpression(node.assignee);
  const right = node.value ? formatExpression(node.value) : "";
  const value = pad + createStatement("set", `${left}${node.value ? " = " + right : ""}`);
  if (node.body.length === 0) {
    return value;
  }
  return value + NEWLINE + formatStatements(node.body, depth + 1, indentStr) + NEWLINE + pad + createStatement("endset");
}
function formatMacro(node, depth, indentStr) {
  const pad = indentStr.repeat(depth);
  const args = node.args.map(formatExpression).join(", ");
  return pad + createStatement("macro", `${node.name.value}(${args})`) + NEWLINE + formatStatements(node.body, depth + 1, indentStr) + NEWLINE + pad + createStatement("endmacro");
}
function formatCallStatement(node, depth, indentStr) {
  const pad = indentStr.repeat(depth);
  const params = node.callerArgs && node.callerArgs.length > 0 ? `(${node.callerArgs.map(formatExpression).join(", ")})` : "";
  const callExpr = formatExpression(node.call);
  let out = pad + createStatement(`call${params}`, callExpr) + NEWLINE;
  out += formatStatements(node.body, depth + 1, indentStr) + NEWLINE;
  out += pad + createStatement("endcall");
  return out;
}
function formatFilterStatement(node, depth, indentStr) {
  const pad = indentStr.repeat(depth);
  const spec = node.filter.type === "Identifier" ? node.filter.value : formatExpression(node.filter);
  let out = pad + createStatement("filter", spec) + NEWLINE;
  out += formatStatements(node.body, depth + 1, indentStr) + NEWLINE;
  out += pad + createStatement("endfilter");
  return out;
}
function formatExpression(node, parentPrec = -1) {
  switch (node.type) {
    case "SpreadExpression": {
      const n = node;
      return `*${formatExpression(n.argument)}`;
    }
    case "Identifier":
      return node.value;
    case "IntegerLiteral":
      return `${node.value}`;
    case "FloatLiteral":
      return `${node.value}`;
    case "StringLiteral":
      return JSON.stringify(node.value);
    case "BinaryExpression": {
      const n = node;
      const thisPrecedence = getBinaryOperatorPrecedence(n);
      const left = formatExpression(n.left, thisPrecedence);
      const right = formatExpression(n.right, thisPrecedence + 1);
      const expr = `${left} ${n.operator.value} ${right}`;
      return thisPrecedence < parentPrec ? `(${expr})` : expr;
    }
    case "UnaryExpression": {
      const n = node;
      const val = n.operator.value + (n.operator.value === "not" ? " " : "") + formatExpression(n.argument, Infinity);
      return val;
    }
    case "CallExpression": {
      const n = node;
      const args = n.args.map(formatExpression).join(", ");
      return `${formatExpression(n.callee)}(${args})`;
    }
    case "MemberExpression": {
      const n = node;
      let obj = formatExpression(n.object);
      if (![
        "Identifier",
        "MemberExpression",
        "CallExpression",
        "StringLiteral",
        "IntegerLiteral",
        "FloatLiteral",
        "ArrayLiteral",
        "TupleLiteral",
        "ObjectLiteral"
      ].includes(n.object.type)) {
        obj = `(${obj})`;
      }
      let prop = formatExpression(n.property);
      if (!n.computed && n.property.type !== "Identifier") {
        prop = `(${prop})`;
      }
      return n.computed ? `${obj}[${prop}]` : `${obj}.${prop}`;
    }
    case "FilterExpression": {
      const n = node;
      const operand = formatExpression(n.operand, Infinity);
      if (n.filter.type === "CallExpression") {
        return `${operand} | ${formatExpression(n.filter)}`;
      }
      return `${operand} | ${n.filter.value}`;
    }
    case "SelectExpression": {
      const n = node;
      return `${formatExpression(n.lhs)} if ${formatExpression(n.test)}`;
    }
    case "TestExpression": {
      const n = node;
      return `${formatExpression(n.operand)} is${n.negate ? " not" : ""} ${n.test.value}`;
    }
    case "ArrayLiteral":
    case "TupleLiteral": {
      const elems = node.value.map(formatExpression);
      const brackets = node.type === "ArrayLiteral" ? "[]" : "()";
      return `${brackets[0]}${elems.join(", ")}${brackets[1]}`;
    }
    case "ObjectLiteral": {
      const entries = Array.from(node.value.entries()).map(
        ([k2, v]) => `${formatExpression(k2)}: ${formatExpression(v)}`
      );
      return `{${entries.join(", ")}}`;
    }
    case "SliceExpression": {
      const n = node;
      const s = n.start ? formatExpression(n.start) : "";
      const t = n.stop ? formatExpression(n.stop) : "";
      const st2 = n.step ? `:${formatExpression(n.step)}` : "";
      return `${s}:${t}${st2}`;
    }
    case "KeywordArgumentExpression": {
      const n = node;
      return `${n.key.value}=${formatExpression(n.value)}`;
    }
    case "Ternary": {
      const n = node;
      const expr = `${formatExpression(n.trueExpr)} if ${formatExpression(n.condition, 0)} else ${formatExpression(
        n.falseExpr
      )}`;
      return parentPrec > -1 ? `(${expr})` : expr;
    }
    default:
      throw new Error(`Unknown expression type: ${node.type}`);
  }
}
var Template = class {
  parsed;
  /**
   * @param {string} template The template string
   */
  constructor(template) {
    const tokens = tokenize(template, {
      lstrip_blocks: true,
      trim_blocks: true
    });
    this.parsed = parse(tokens);
  }
  render(items) {
    const env2 = new Environment();
    setupGlobals(env2);
    if (items) {
      for (const [key, value] of Object.entries(items)) {
        env2.set(key, value);
      }
    }
    const interpreter = new Interpreter(env2);
    const result = interpreter.run(this.parsed);
    return result.value;
  }
  format(options) {
    return format(this.parsed, options?.indent || "	");
  }
};

// src/utils/hub/FileResponse.js
import fs2 from "fs";
var CONTENT_TYPE_MAP = {
  txt: "text/plain",
  html: "text/html",
  css: "text/css",
  js: "text/javascript",
  json: "application/json",
  png: "image/png",
  jpg: "image/jpeg",
  jpeg: "image/jpeg",
  gif: "image/gif"
};
var FileResponse = class _FileResponse {
  /**
   * Creates a new `FileResponse` object.
   * @param {string} filePath
   */
  constructor(filePath) {
    this.filePath = filePath;
    this.headers = new Headers();
    this.exists = fs2.existsSync(filePath);
    if (this.exists) {
      this.status = 200;
      this.statusText = "OK";
      let stats = fs2.statSync(filePath);
      this.headers.set("content-length", stats.size.toString());
      this.updateContentType();
      const stream = fs2.createReadStream(filePath);
      this.body = new ReadableStream({
        start(controller) {
          stream.on("data", (chunk2) => controller.enqueue(chunk2));
          stream.on("end", () => controller.close());
          stream.on("error", (err) => controller.error(err));
        },
        cancel() {
          stream.destroy();
        }
      });
    } else {
      this.status = 404;
      this.statusText = "Not Found";
      this.body = null;
    }
  }
  /**
   * Updates the 'content-type' header property of the response based on the extension of
   * the file specified by the filePath property of the current object.
   * @returns {void}
   */
  updateContentType() {
    const extension = this.filePath.toString().split(".").pop().toLowerCase();
    this.headers.set("content-type", CONTENT_TYPE_MAP[extension] ?? "application/octet-stream");
  }
  /**
   * Clone the current FileResponse object.
   * @returns {FileResponse} A new FileResponse object with the same properties as the current object.
   */
  clone() {
    let response = new _FileResponse(this.filePath);
    response.exists = this.exists;
    response.status = this.status;
    response.statusText = this.statusText;
    response.headers = new Headers(this.headers);
    return response;
  }
  /**
   * Reads the contents of the file specified by the filePath property and returns a Promise that
   * resolves with an ArrayBuffer containing the file's contents.
   * @returns {Promise<ArrayBuffer>} A Promise that resolves with an ArrayBuffer containing the file's contents.
   * @throws {Error} If the file cannot be read.
   */
  async arrayBuffer() {
    const data = await fs2.promises.readFile(this.filePath);
    return (
      /** @type {ArrayBuffer} */
      data.buffer
    );
  }
  /**
   * Reads the contents of the file specified by the filePath property and returns a Promise that
   * resolves with a Blob containing the file's contents.
   * @returns {Promise<Blob>} A Promise that resolves with a Blob containing the file's contents.
   * @throws {Error} If the file cannot be read.
   */
  async blob() {
    const data = await fs2.promises.readFile(this.filePath);
    return new Blob([
      /** @type {any} */
      data
    ], { type: this.headers.get("content-type") });
  }
  /**
   * Reads the contents of the file specified by the filePath property and returns a Promise that
   * resolves with a string containing the file's contents.
   * @returns {Promise<string>} A Promise that resolves with a string containing the file's contents.
   * @throws {Error} If the file cannot be read.
   */
  async text() {
    return await fs2.promises.readFile(this.filePath, "utf8");
  }
  /**
   * Reads the contents of the file specified by the filePath property and returns a Promise that
   * resolves with a parsed JavaScript object containing the file's contents.
   *
   * @returns {Promise<Object>} A Promise that resolves with a parsed JavaScript object containing the file's contents.
   * @throws {Error} If the file cannot be read.
   */
  async json() {
    return JSON.parse(await this.text());
  }
};

// src/utils/cache/FileCache.js
import fs3 from "fs";
import path2 from "path";

// src/utils/random.js
var Random = class {
  constructor(seed) {
    this._mt = new Uint32Array(624);
    this._idx = 625;
    this._gauss_next = null;
    this._random_fn = this.random.bind(this);
    this.seed(seed);
  }
  /**
   * Seeds this instance's PRNG.
   *
   * When called with a number, initializes the state deterministically from that value.
   * When called with no arguments (or `undefined`/`null`), seeds from OS entropy
   * via `crypto.getRandomValues`, matching Python's `random.seed()` behaviour.
   *
   * @param {number} [n] The seed value. Omit to seed from OS entropy.
   */
  seed(n) {
    if (n === void 0 || n === null) {
      if (apis.IS_CRYPTO_AVAILABLE) {
        const buf = new Uint32Array(1);
        crypto.getRandomValues(buf);
        n = buf[0];
      } else {
        n = Date.now() >>> 0;
      }
    }
    const mt2 = this._mt;
    const u = (a, b) => Math.imul(a, b) >>> 0, key = [];
    for (let v = n || 0; v > 0; v = Math.floor(v / 4294967296)) key.push(v & 4294967295);
    if (!key.length) key.push(0);
    mt2[0] = 19650218;
    for (let k2 = 1; k2 < 624; ++k2) mt2[k2] = u(1812433253, mt2[k2 - 1] ^ mt2[k2 - 1] >>> 30) + k2 >>> 0;
    let i = 1, j = 0;
    for (let k2 = Math.max(624, key.length); k2 > 0; --k2, ++i, ++j) {
      if (i >= 624) {
        mt2[0] = mt2[623];
        i = 1;
      }
      if (j >= key.length) j = 0;
      mt2[i] = (mt2[i] ^ u(mt2[i - 1] ^ mt2[i - 1] >>> 30, 1664525)) + key[j] + j >>> 0;
    }
    for (let k2 = 623; k2 > 0; --k2, ++i) {
      if (i >= 624) {
        mt2[0] = mt2[623];
        i = 1;
      }
      mt2[i] = (mt2[i] ^ u(mt2[i - 1] ^ mt2[i - 1] >>> 30, 1566083941)) - i >>> 0;
    }
    mt2[0] = 2147483648;
    this._idx = 624;
    this._gauss_next = null;
  }
  /**
   * Generates a random unsigned 32-bit integer.
   *
   * Performs the "twist" step when the state buffer is exhausted,
   * then applies the standard MT19937 tempering transform.
   *
   * @returns {number} A random integer in the range [0, 2^32 - 1].
   */
  _int32() {
    const mt2 = this._mt;
    if (this._idx >= 624) {
      for (let k2 = 0; k2 < 624; ++k2) {
        const y2 = mt2[k2] & 2147483648 | mt2[(k2 + 1) % 624] & 2147483647;
        mt2[k2] = (mt2[(k2 + 397) % 624] ^ y2 >>> 1 ^ (y2 & 1 ? 2567483615 : 0)) >>> 0;
      }
      this._idx = 0;
    }
    let y = mt2[this._idx++];
    y ^= y >>> 11;
    y ^= y << 7 & 2636928640;
    y ^= y << 15 & 4022730752;
    y ^= y >>> 18;
    return y >>> 0;
  }
  /**
   * Generates a random floating-point number in the half-open interval [0, 1).
   *
   * Combines two 32-bit integers (using 53 bits of precision) to produce
   * a uniformly distributed double, matching Python's `random.random()`.
   *
   * @returns {number} A random float in [0, 1).
   */
  random() {
    return ((this._int32() >>> 5) * 67108864 + (this._int32() >>> 6)) / 9007199254740992;
  }
  /**
   * Generates a random number from a Gaussian (normal) distribution.
   *
   * Uses the Box-Muller transform with a cached spare value,
   * matching Python's `random.gauss()` output for the same seed.
   *
   * @param {number} [mu=0] The mean of the distribution.
   * @param {number} [sigma=1] The standard deviation of the distribution.
   * @returns {number} A normally distributed random value.
   */
  gauss(mu = 0, sigma = 1) {
    let z2 = this._gauss_next;
    this._gauss_next = null;
    if (z2 === null) {
      const x2pi = this.random() * 2 * Math.PI, g2rad = Math.sqrt(-2 * Math.log(1 - this.random()));
      z2 = Math.cos(x2pi) * g2rad;
      this._gauss_next = Math.sin(x2pi) * g2rad;
    }
    return mu + z2 * sigma;
  }
  /**
   * Shuffles an array in-place using the Fisher-Yates algorithm.
   *
   * Uses rejection sampling via `getrandbits`-style bit masking to ensure
   * a uniform distribution, matching Python's `random.shuffle()`.
   *
   * @param {any[]} arr The array to shuffle in-place.
   */
  shuffle(arr) {
    for (let i = arr.length - 1; i > 0; --i) {
      const k2 = 32 - Math.clz32(i + 1);
      let r = this._int32() >>> 32 - k2;
      while (r > i) r = this._int32() >>> 32 - k2;
      const t = arr[i];
      arr[i] = arr[r];
      arr[r] = t;
    }
  }
  /**
   * Selects a single element from a weighted population.
   *
   * Matches Python's `random.choices(population, weights=weights, k=1)[0]`
   *
   * @param {any[]} population The array of items to choose from.
   * @param {number[]} weights An array of non-negative weights, one per population element.
   * @returns {*} A single randomly selected element from the population.
   */
  choices(population, weights) {
    return population[_weightedIndexWith(this._random_fn, weights)];
  }
};
function _weightedIndexWith(randomFn, weights) {
  let sum = 0;
  for (let i = 0; i < weights.length; ++i) sum += weights[i];
  let x = randomFn() * sum;
  for (let i = 0; i < weights.length; ++i) {
    x -= weights[i];
    if (x < 0) return i;
  }
  return weights.length - 1;
}
var _default = new Random();
var random = Object.freeze({
  Random,
  seed: _default.seed.bind(_default),
  random: _default.random.bind(_default),
  gauss: _default.gauss.bind(_default),
  shuffle: _default.shuffle.bind(_default),
  choices: _default.choices.bind(_default)
});
var _weightedIndex = (weights) => _weightedIndexWith(random.random, weights);

// src/utils/cache/FileCache.js
var rng = new Random();
var FileCache = class {
  /**
   * Instantiate a `FileCache` object.
   * @param {string} path
   */
  constructor(path3) {
    this.path = path3;
  }
  /**
   * Checks whether the given request is in the cache.
   * @param {string} request
   * @returns {Promise<FileResponse | undefined>}
   */
  async match(request) {
    let filePath = path2.join(this.path, request);
    let file = new FileResponse(filePath);
    if (file.exists) {
      return file;
    } else {
      return void 0;
    }
  }
  /**
   * Adds the given response to the cache.
   * @param {string} request
   * @param {Response} response
   * @param {(data: {progress: number, loaded: number, total: number}) => void} [progress_callback] Optional.
   * The function to call with progress updates
   * @returns {Promise<void>}
   */
  async put(request, response, progress_callback = void 0) {
    const filePath = path2.join(this.path, request);
    const id = apis.IS_PROCESS_AVAILABLE ? process.pid : Date.now();
    const randomSuffix = rng._int32().toString(36);
    const tmpPath = filePath + `.tmp.${id}.${randomSuffix}`;
    try {
      const contentLength = response.headers.get("Content-Length");
      const total = parseInt(contentLength ?? "0");
      let loaded = 0;
      await fs3.promises.mkdir(path2.dirname(filePath), { recursive: true });
      const fileStream = fs3.createWriteStream(tmpPath);
      const reader = response.body.getReader();
      while (true) {
        const { done, value } = await reader.read();
        if (done) {
          break;
        }
        await new Promise((resolve, reject) => {
          fileStream.write(value, (err) => {
            if (err) {
              reject(err);
              return;
            }
            resolve();
          });
        });
        loaded += value.length;
        const progress = total ? loaded / total * 100 : 0;
        progress_callback?.({ progress, loaded, total });
      }
      await new Promise((resolve, reject) => {
        fileStream.close((err) => err ? reject(err) : resolve());
      });
      await fs3.promises.rename(tmpPath, filePath);
    } catch (error) {
      try {
        await fs3.promises.unlink(tmpPath);
      } catch {
      }
      throw error;
    }
  }
  /**
   * Deletes the cache entry for the given request.
   * @param {string} request
   * @returns {Promise<boolean>} A Promise that resolves to `true` if the cache entry was deleted, `false` otherwise.
   */
  async delete(request) {
    let filePath = path2.join(this.path, request);
    try {
      await fs3.promises.unlink(filePath);
      return true;
    } catch (error) {
      return false;
    }
  }
  // TODO add the rest?
  // addAll(requests: RequestInfo[]): Promise<void>;
  // keys(request?: RequestInfo | URL, options?: CacheQueryOptions): Promise<ReadonlyArray<Request>>;
  // match(request: RequestInfo | URL, options?: CacheQueryOptions): Promise<Response | undefined>;
  // matchAll(request?: RequestInfo | URL, options?: CacheQueryOptions): Promise<ReadonlyArray<Response>>;
};

// src/utils/hub/constants.js
var ERROR_MAPPING = {
  // 4xx errors (https://developer.mozilla.org/en-US/docs/Web/HTTP/Status#client_error_responses)
  400: "Bad request error occurred while trying to load file",
  401: "Unauthorized access to file",
  403: "Forbidden access to file",
  404: "Could not locate file",
  408: "Request timeout error occurred while trying to load file",
  // 5xx errors (https://developer.mozilla.org/en-US/docs/Web/HTTP/Status#server_error_responses)
  500: "Internal server error error occurred while trying to load file",
  502: "Bad gateway error occurred while trying to load file",
  503: "Service unavailable error occurred while trying to load file",
  504: "Gateway timeout error occurred while trying to load file"
};
var MAX_EXTERNAL_DATA_CHUNKS = 100;
var REPO_ID_REGEX = /^(\b[\w\-.]+\b\/)?\b[\w\-.]{1,96}\b$/;

// src/utils/hub/utils.js
function pathJoin(...parts) {
  parts = parts.map((part, index) => {
    if (index) {
      part = part.replace(new RegExp("^/"), "");
    }
    if (index !== parts.length - 1) {
      part = part.replace(new RegExp("/$"), "");
    }
    return part;
  });
  return parts.join("/");
}
function isValidUrl(string, protocols = null, validHosts = null) {
  let url2;
  try {
    url2 = new URL(string);
  } catch (_) {
    return false;
  }
  if (protocols && !protocols.includes(url2.protocol)) {
    return false;
  }
  if (validHosts && !validHosts.includes(url2.hostname)) {
    return false;
  }
  return true;
}
function isValidHfModelId(string) {
  if (!REPO_ID_REGEX.test(string)) return false;
  if (string.includes("..") || string.includes("--")) return false;
  if (string.endsWith(".git") || string.endsWith(".ipynb")) return false;
  return true;
}
function handleError(status, remoteURL, fatal) {
  if (!fatal) {
    return null;
  }
  const message = ERROR_MAPPING[status] ?? `Error (${status}) occurred while trying to load file`;
  throw Error(`${message}: "${remoteURL}".`);
}
async function readResponse(response, progress_callback, expectedSize) {
  const contentLength = response.headers.get("Content-Length");
  let total = contentLength ? parseInt(contentLength, 10) : expectedSize ?? 0;
  if (contentLength === null && !expectedSize) {
    logger.warn("Unable to determine content-length from response headers. Will expand buffer when needed.");
  }
  let buffer = new Uint8Array(total);
  let loaded = 0;
  const reader = response.body.getReader();
  async function read() {
    const { done, value } = await reader.read();
    if (done) return;
    const newLoaded = loaded + value.length;
    if (newLoaded > total) {
      total = newLoaded;
      const newBuffer = new Uint8Array(total);
      newBuffer.set(buffer);
      buffer = newBuffer;
    }
    buffer.set(value, loaded);
    loaded = newLoaded;
    const progress = loaded / total * 100;
    progress_callback({ progress, loaded, total });
    return read();
  }
  await read();
  return buffer;
}
function isBlobURL(url2) {
  return isValidUrl(url2, ["blob:"]);
}
function toAbsoluteURL(url2) {
  let baseURL;
  if (typeof location !== "undefined" && location.href) {
    baseURL = location.href;
  } else if (typeof import.meta !== "undefined" && import.meta.url) {
    baseURL = import.meta.url;
  } else {
    return url2;
  }
  return new URL(url2, baseURL).href;
}

// src/utils/cache/CrossOriginStorageCache.js
var HASH_ALGORITHM = "SHA-256";
var HASH_CACHE_NAME = "experimental_transformers-hash-cache";
var makeHashDescriptor = (value) => ({ algorithm: HASH_ALGORITHM, value });
var CrossOriginStorage = class {
  /** @type {Promise<Cache> | null} */
  #hashCache = null;
  /**
   * Returns (and lazily opens) the hash cache, reusing the same promise across concurrent callers.
   * @returns {Promise<Cache>}
   */
  _getHashCache = () => {
    this.#hashCache ??= caches.open(HASH_CACHE_NAME);
    return this.#hashCache;
  };
  /**
   * Returns whether the `navigator.crossOriginStorage` API is available in the current environment.
   * @returns {boolean}
   */
  static isAvailable = () => typeof navigator !== "undefined" && "crossOriginStorage" in navigator;
  /**
   * Looks up a cached response for the given URL by resolving its SHA-256 hash and requesting
   * the corresponding file handle from cross-origin storage.
   *
   * Implements `CacheInterface.match`.
   *
   * @param {string} request The URL of the resource to look up.
   * @returns {Promise<Response|undefined>} The cached `Response`, or `undefined` if not found.
   */
  match = async (request) => {
    const hashValue = await this._getFileHash(request);
    if (!hashValue) {
      return void 0;
    }
    try {
      const [handle] = await navigator.crossOriginStorage.requestFileHandles([makeHashDescriptor(hashValue)]);
      const blob = await handle.getFile();
      return new Response(blob, {
        headers: {
          "Content-Length": String(blob.size)
        }
      });
    } catch {
      return void 0;
    }
  };
  /**
   * Stores a response in cross-origin storage, keyed by its SHA-256 hash.
   *
   * For LFS-backed URLs the hash is resolved cheaply via `_getFileHash` (which checks
   * `HASH_CACHE_NAME` first, then falls back to fetching the Git LFS pointer file)
   * without reading the response body a second time.
   *
   * For non-LFS resources the hash is unknown upfront.  In that case the body is consumed
   * in the background: the stream is read to compute the content hash, the file is written
   * into cross-origin storage, and the computed hash is persisted to `HASH_CACHE_NAME`
   * so that future `match` calls can resolve the file without a network round-trip.
   *
   * Implements `CacheInterface.put`.
   *
   * @param {string} request The URL of the resource (used as the hash-cache key).
   * @param {Response} response The response whose body will be written to the cache.
   * @returns {Promise<void>}
   */
  put = async (request, response) => {
    const hashValue = await this._getFileHash(request);
    if (hashValue) {
      const blob = await response.blob();
      await this._storeBlobInCOS(blob, hashValue);
    } else {
      this._processAndStore(request, response.body);
    }
  };
  /**
   * Writes a blob into cross-origin storage using the given pre-computed hex hash string.
   *
   * @param {Blob} blob
   * @param {string} hashHex Hex-encoded SHA-256 hash of `blob`.
   * @returns {Promise<void>}
   */
  _storeBlobInCOS = async (blob, hashHex) => {
    const [handle] = await navigator.crossOriginStorage.requestFileHandles([makeHashDescriptor(hashHex)], {
      create: true
    });
    const writableStream = await handle.createWritable();
    await writableStream.write(blob);
    await writableStream.close();
  };
  /**
   * Background task for non-LFS resources: consumes `stream`, computes the SHA-256 hash
   * of the resulting blob, stores it in cross-origin storage, and persists the computed
   * hash to `HASH_CACHE_NAME` keyed by `request` so future `match` calls can resolve the
   * file without a network round-trip.
   *
   * Called fire-and-forget from `put` — errors are swallowed so failures never surface to
   * the caller.
   *
   * @param {string} request The original resource URL.
   * @param {ReadableStream} stream The response body stream to consume.
   * @returns {Promise<void>}
   */
  _processAndStore = async (request, stream) => {
    try {
      const chunks = [];
      for await (const chunk2 of stream) {
        chunks.push(chunk2);
      }
      const blob = new Blob(chunks);
      const hashHex = await this._getBlobHash(blob);
      await this._storeBlobInCOS(blob, hashHex);
      try {
        const hashCache = await this._getHashCache();
        await hashCache.put(request, new Response(hashHex));
      } catch {
      }
    } catch {
    }
  };
  /**
   * Deletes the cache entry for the given request.
   *
   * Removes the hash entry from `HASH_CACHE_NAME`. Note: cross-origin storage itself does not
   * expose a delete API, so only the local hash mapping is removed. For non-LFS URLs this
   * permanently prevents `match` from resolving the file. For LFS-backed URLs, `match` will
   * re-fetch the LFS pointer file on the next call and repopulate the hash cache automatically.
   *
   * Implements `CacheInterface.delete`.
   *
   * @param {string} request
   * @returns {Promise<boolean>} Resolves to `true` if the hash entry was deleted, `false` otherwise.
   */
  delete = async (request) => {
    try {
      const hashCache = await this._getHashCache();
      return await hashCache.delete(request);
    } catch {
      return false;
    }
  };
  /**
   * Resolves the SHA-256 hash for a given URL.
   *
   * Returns the cached hash immediately if one has been persisted to `HASH_CACHE_NAME`.
   * Otherwise falls back to `_getLfsFileHash` to retrieve the hash from the Hugging Face
   * LFS pointer file, persisting the result to `HASH_CACHE_NAME` for future lookups.
   *
   * Returns `null` if the hash cannot be determined (e.g. non-LFS URL with no cached entry).
   *
   * @param {string} url The resource URL to resolve a hash for.
   * @returns {Promise<string|null>} The hex-encoded SHA-256 hash, or `null` if unavailable.
   */
  _getFileHash = async (url2) => {
    try {
      const hashCache = await this._getHashCache();
      const cached = await hashCache.match(url2);
      if (cached) {
        return cached.text();
      }
      const hash = await this._getLfsFileHash(url2);
      if (hash) {
        await hashCache.put(url2, new Response(hash));
        return hash;
      }
      return null;
    } catch {
      return null;
    }
  };
  /**
   * Attempts to retrieve the SHA-256 hash for a Hugging Face resource URL from its raw
   * Git LFS pointer file.
   *
   * Only applicable to URLs containing `/resolve/` (i.e. Hugging Face resolved file URLs).
   * The `/resolve/` segment is rewritten to `/raw/` to fetch the LFS pointer directly.
   * Returns `null` for non-LFS URLs or when the network request fails.
   *
   * @see https://huggingface.co/docs/hub/en/storage-backends#xet
   * @param {string} url The resolved Hugging Face URL of the resource.
   * @returns {Promise<string|null>} The hex-encoded SHA-256 hash, or `null` if unavailable.
   */
  _getLfsFileHash = async (url2) => {
    if (!url2.includes("/resolve/")) {
      return null;
    }
    const rawUrl = url2.replace("/resolve/", "/raw/");
    try {
      const text = await fetch(rawUrl).then((r) => r.text());
      const match = text.match(/^oid sha256:([0-9a-f]+)$/m);
      return match ? match[1] : null;
    } catch {
      return null;
    }
  };
  /**
   * Computes the SHA-256 hash of a `Blob`'s contents.
   *
   * @param {Blob} blob The blob to hash.
   * @returns {Promise<string>} The lowercase hex-encoded SHA-256 hash.
   */
  _getBlobHash = async (blob) => {
    const arrayBuffer = await blob.arrayBuffer();
    const hashBuffer = await crypto.subtle.digest(HASH_ALGORITHM, arrayBuffer);
    const hashArray = Array.from(new Uint8Array(hashBuffer));
    return hashArray.map((byte) => byte.toString(16).padStart(2, "0")).join("");
  };
};

// src/utils/cache.js
async function getCache(file_cache_dir = null) {
  let cache2 = null;
  if (env.useCustomCache) {
    if (!env.customCache) {
      throw Error("`env.useCustomCache=true`, but `env.customCache` is not defined.");
    }
    if (!env.customCache.match || !env.customCache.put) {
      throw new Error(
        "`env.customCache` must be an object which implements the `match` and `put` functions of the Web Cache API. For more information, see https://developer.mozilla.org/en-US/docs/Web/API/Cache"
      );
    }
    cache2 = env.customCache;
  }
  if (!cache2 && env.experimental_useCrossOriginStorage && CrossOriginStorage.isAvailable()) {
    cache2 = new CrossOriginStorage();
  }
  if (!cache2 && env.useBrowserCache) {
    if (typeof caches === "undefined") {
      throw Error("Browser cache is not available in this environment.");
    }
    try {
      cache2 = await caches.open(env.cacheKey);
    } catch (e) {
      logger.warn("An error occurred while opening the browser cache:", e);
    }
  }
  if (!cache2 && env.useFSCache) {
    if (!apis.IS_FS_AVAILABLE) {
      throw Error("File System Cache is not available in this environment.");
    }
    cache2 = new FileCache(file_cache_dir ?? env.cacheDir);
  }
  return cache2;
}
async function tryCache(cache2, ...names) {
  for (let name of names) {
    try {
      let result = await cache2.match(name);
      if (result) return result;
    } catch (e) {
      continue;
    }
  }
  return void 0;
}

// src/utils/lru_cache.js
var LRUCache2 = class {
  /** @type {number} */
  #capacity;
  /** @type {Map<any, any>} */
  #cache;
  /**
   * Creates an LRUCache instance.
   * @param {number} capacity The maximum number of items the cache can hold.
   */
  constructor(capacity) {
    this.#capacity = capacity;
    this.#cache = /* @__PURE__ */ new Map();
  }
  /**
   * Retrieves the value associated with the given key and marks the key as recently used.
   * @param {any} key The key to retrieve.
   * @returns {any} The value associated with the key, or undefined if the key does not exist.
   */
  get(key) {
    if (!this.#cache.has(key)) return void 0;
    const value = this.#cache.get(key);
    this.#cache.delete(key);
    this.#cache.set(key, value);
    return value;
  }
  /**
   * Inserts or updates the key-value pair in the cache.
   * If the key already exists, it is updated and marked as recently used.
   * If the cache exceeds its capacity, the least recently used item is evicted.
   * @param {any} key The key to add or update.
   * @param {any} value The value to associate with the key.
   */
  put(key, value) {
    if (this.#cache.has(key)) {
      this.#cache.delete(key);
    }
    this.#cache.set(key, value);
    if (this.#cache.size > this.#capacity) {
      this.#cache.delete(this.#cache.keys().next().value);
    }
  }
  /**
   * Removes the entry for the given key from the cache.
   * @param {any} key The key to delete.
   * @returns {boolean} `true` if the entry existed and was removed, `false` otherwise.
   */
  delete(key) {
    return this.#cache.delete(key);
  }
  /**
   * Clears the cache.
   */
  clear() {
    this.#cache.clear();
  }
};

// src/utils/memoize_promise.js
var MAX_CACHE_SIZE = 100;
var cache = new LRUCache2(MAX_CACHE_SIZE);
function memoizePromise(key, factory) {
  const cached = cache.get(key);
  if (cached !== void 0) {
    return cached;
  }
  const promise = factory().then(
    (value) => value,
    (err) => {
      cache.delete(key);
      return Promise.reject(err);
    }
  );
  cache.put(key, promise);
  return promise;
}

// src/utils/model_registry/get_file_metadata.js
async function fetch_file_head(urlOrPath) {
  if (!isValidUrl(urlOrPath, ["http:", "https:"])) {
    return null;
  }
  const headers = getFetchHeaders(urlOrPath);
  headers.set("Range", "bytes=0-0");
  return env.fetch(urlOrPath, { method: "GET", headers, cache: "no-store" });
}
function get_file_metadata(path_or_repo_id, filename, options = {}) {
  const key = JSON.stringify([
    path_or_repo_id,
    filename,
    options?.revision,
    options?.cache_dir,
    options?.local_files_only
  ]);
  return memoizePromise(key, () => _get_file_metadata(path_or_repo_id, filename, options));
}
async function _get_file_metadata(path_or_repo_id, filename, options) {
  const cache2 = await getCache(options?.cache_dir);
  const { localPath, remoteURL, proposedCacheKey, validModelId } = buildResourcePaths(
    path_or_repo_id,
    filename,
    options,
    cache2
  );
  const cachedResponse = await checkCachedResource(cache2, localPath, proposedCacheKey);
  if (cachedResponse !== void 0 && typeof cachedResponse !== "string") {
    const size = cachedResponse.headers.get("content-length");
    const contentType = cachedResponse.headers.get("content-type");
    return {
      exists: true,
      size: size ? parseInt(size, 10) : void 0,
      contentType: contentType || void 0,
      fromCache: true
    };
  }
  if (env.allowLocalModels) {
    const isURL = isValidUrl(localPath, ["http:", "https:"]);
    if (!isURL) {
      try {
        const response = await getFile(localPath);
        if (typeof response !== "string" && response.status !== 404) {
          const size = response.headers.get("content-length");
          const contentType = response.headers.get("content-type");
          return {
            exists: true,
            size: size ? parseInt(size, 10) : void 0,
            contentType: contentType || void 0,
            fromCache: false
          };
        }
      } catch (e) {
      }
    }
  }
  if (env.allowRemoteModels && !options.local_files_only && validModelId) {
    try {
      const rangeResponse = await fetch_file_head(remoteURL);
      if (rangeResponse && rangeResponse.status >= 200 && rangeResponse.status < 300) {
        let size;
        const contentType = rangeResponse.headers.get("content-type");
        if (rangeResponse.status === 206) {
          const contentRange = rangeResponse.headers.get("content-range");
          if (contentRange) {
            const match = contentRange.match(/bytes \d+-\d+\/(\d+)/);
            if (match) {
              size = parseInt(match[1], 10);
            }
          }
        } else if (rangeResponse.status === 200) {
          try {
            await rangeResponse.body?.cancel();
          } catch (cancelError) {
          }
        }
        if (size === void 0) {
          const contentLength = rangeResponse.headers.get("content-length");
          size = contentLength ? parseInt(contentLength, 10) : void 0;
        }
        return {
          exists: true,
          size,
          contentType: contentType || void 0,
          fromCache: false
        };
      }
    } catch (e) {
      logger.warn(`Unable to fetch file metadata for "${remoteURL}": ${e}`);
    }
  }
  return { exists: false, fromCache: false };
}

// src/utils/hub.js
async function getFile(urlOrPath) {
  if (env.useFS && !isValidUrl(urlOrPath, ["http:", "https:", "blob:"])) {
    return new FileResponse(
      urlOrPath instanceof URL ? urlOrPath.protocol === "file:" ? urlOrPath.pathname : urlOrPath.toString() : urlOrPath
    );
  } else {
    return env.fetch(urlOrPath, {
      headers: getFetchHeaders(urlOrPath)
    });
  }
}
function getFetchHeaders(urlOrPath) {
  const isNode = typeof process !== "undefined" && process?.release?.name === "node";
  const headers = new Headers();
  if (isNode) {
    const IS_CI = !!process.env?.TESTING_REMOTELY;
    const version = env.version;
    headers.set("User-Agent", `transformers.js/${version}; is_ci/${IS_CI};`);
    const isHFURL = isValidUrl(urlOrPath, ["http:", "https:"], ["huggingface.co", "hf.co"]);
    if (isHFURL) {
      const token = process.env?.HF_TOKEN ?? process.env?.HF_ACCESS_TOKEN;
      if (token) {
        headers.set("Authorization", `Bearer ${token}`);
      }
    }
  } else {
  }
  return headers;
}
function buildResourcePaths(path_or_repo_id, filename, options = {}, cache2 = null) {
  const revision = options.revision ?? "main";
  const requestURL = pathJoin(path_or_repo_id, filename);
  const validModelId = isValidHfModelId(path_or_repo_id);
  const localPath = validModelId ? pathJoin(env.localModelPath, requestURL) : requestURL;
  const remoteURL = pathJoin(
    env.remoteHost,
    env.remotePathTemplate.replaceAll("{model}", path_or_repo_id).replaceAll("{revision}", encodeURIComponent(revision)),
    filename
  );
  const proposedCacheKey = cache2 instanceof FileCache ? (
    // Choose cache key for filesystem cache
    // When using the main revision (default), we use the request URL as the cache key.
    // If a specific revision is requested, we account for this in the cache key.
    revision === "main" ? requestURL : pathJoin(path_or_repo_id, revision, filename)
  ) : remoteURL;
  return {
    requestURL,
    localPath,
    remoteURL,
    proposedCacheKey,
    validModelId
  };
}
async function checkCachedResource(cache2, localPath, proposedCacheKey) {
  if (!cache2) {
    return void 0;
  }
  return await tryCache(cache2, localPath, proposedCacheKey);
}
async function storeCachedResource(path_or_repo_id, filename, cache2, cacheKey, response, result, options = {}) {
  if (await cache2.match(cacheKey) !== void 0) {
    return;
  }
  if (!result) {
    const wrapped_progress = options.progress_callback ? (data) => dispatchCallback(options.progress_callback, {
      status: "progress",
      name: path_or_repo_id,
      file: filename,
      ...data
    }) : void 0;
    await cache2.put(
      cacheKey,
      /** @type {Response} */
      response,
      wrapped_progress
    );
  } else if (typeof response !== "string") {
    const headers = new Headers(response.headers);
    headers.set("content-length", result.byteLength.toString());
    await cache2.put(
      cacheKey,
      new Response(
        /** @type {any} */
        result,
        {
          headers
        }
      )
    ).catch((err) => {
      logger.warn(`Unable to add response to browser cache: ${err}.`);
    });
  }
}
async function loadResourceFile(path_or_repo_id, filename, fatal = true, options = {}, return_path = false, cache2 = null) {
  const { requestURL, localPath, remoteURL, proposedCacheKey, validModelId } = buildResourcePaths(
    path_or_repo_id,
    filename,
    options,
    cache2
  );
  let cacheKey;
  let toCacheResponse = false;
  let response;
  response = await checkCachedResource(cache2, localPath, proposedCacheKey);
  const cacheHit = response !== void 0;
  if (cacheHit) {
    cacheKey = proposedCacheKey;
  } else {
    if (env.allowLocalModels) {
      const isURL = isValidUrl(requestURL, ["http:", "https:"]);
      if (!isURL) {
        try {
          response = await getFile(localPath);
          cacheKey = localPath;
        } catch (e) {
          logger.warn(`Unable to load from local path "${localPath}": "${e}"`);
        }
      } else if (options.local_files_only) {
        throw new Error(`\`local_files_only=true\`, but attempted to load a remote file from: ${requestURL}.`);
      } else if (!env.allowRemoteModels) {
        throw new Error(
          `\`env.allowRemoteModels=false\`, but attempted to load a remote file from: ${requestURL}.`
        );
      }
    }
    if (response === void 0 || typeof response !== "string" && response.status === 404) {
      if (options.local_files_only || !env.allowRemoteModels) {
        if (fatal) {
          throw Error(
            `\`local_files_only=true\` or \`env.allowRemoteModels=false\` and file was not found locally at "${localPath}".`
          );
        } else {
          return null;
        }
      }
      if (!validModelId) {
        throw Error(
          `Local file missing at "${localPath}" and download aborted due to invalid model ID "${path_or_repo_id}".`
        );
      }
      response = await getFile(remoteURL);
      if (response.status !== 200) {
        return handleError(response.status, remoteURL, fatal);
      }
      cacheKey = proposedCacheKey;
    }
    toCacheResponse = cache2 && // 1. A caching system is available
    typeof Response !== "undefined" && // 2. `Response` is defined (i.e., we are in a browser-like environment)
    response instanceof Response && // 3. result is a `Response` object (i.e., not a `FileResponse`)
    response.status === 200;
  }
  dispatchCallback(options.progress_callback, {
    status: "download",
    name: path_or_repo_id,
    file: filename
  });
  let result;
  if (apis.IS_NODE_ENV && return_path) {
  } else {
    let buffer;
    if (typeof response !== "string") {
      if (!options.progress_callback) {
        buffer = new Uint8Array(await response.arrayBuffer());
      } else if (cacheHit && // The item is being read from the cache
      typeof navigator !== "undefined" && /firefox/i.test(navigator.userAgent)) {
        buffer = new Uint8Array(await response.arrayBuffer());
        dispatchCallback(options.progress_callback, {
          status: "progress",
          name: path_or_repo_id,
          file: filename,
          progress: 100,
          loaded: buffer.length,
          total: buffer.length
        });
      } else {
        let expectedSize;
        const contentLength = response.headers.get("content-length");
        if (contentLength) {
          expectedSize = parseInt(contentLength, 10);
        } else {
          try {
            const metadata = await get_file_metadata(path_or_repo_id, filename, options);
            if (metadata.size) {
              expectedSize = metadata.size;
            }
          } catch (e) {
          }
        }
        buffer = await readResponse(
          response,
          (data) => {
            dispatchCallback(options.progress_callback, {
              status: "progress",
              name: path_or_repo_id,
              file: filename,
              ...data
            });
          },
          expectedSize
        );
      }
    }
    result = buffer;
  }
  if (
    // Only cache web responses
    // i.e., do not cache FileResponses (prevents duplication)
    toCacheResponse && cacheKey && typeof response !== "string"
  ) {
    await storeCachedResource(path_or_repo_id, filename, cache2, cacheKey, response, result, options);
  }
  if (apis.IS_NODE_ENV && return_path && options.progress_callback && typeof response !== "string") {
    const size = parseInt(response.headers.get("content-length"), 10) || 0;
    dispatchCallback(options.progress_callback, {
      status: "progress",
      name: path_or_repo_id,
      file: filename,
      progress: 100,
      loaded: size,
      total: size
    });
  }
  dispatchCallback(options.progress_callback, {
    status: "done",
    name: path_or_repo_id,
    file: filename
  });
  if (result) {
    if (!apis.IS_NODE_ENV && return_path) {
      throw new Error("Cannot return path in a browser environment.");
    }
    return result;
  }
  if (response instanceof FileResponse) {
    return response.filePath;
  }
  const cachedResponse = await cache2?.match(cacheKey);
  if (cachedResponse instanceof FileResponse) {
    return cachedResponse.filePath;
  } else if (cachedResponse instanceof Response) {
    return new Uint8Array(await cachedResponse.arrayBuffer());
  } else if (typeof cachedResponse === "string") {
    return cachedResponse;
  }
  throw new Error("Unable to get model file path or buffer.");
}
var INFLIGHT_LOADS = /* @__PURE__ */ new Map();
async function getModelFile(path_or_repo_id, filename, fatal = true, options = {}, return_path = false) {
  if (!env.allowLocalModels) {
    if (options.local_files_only) {
      throw Error(
        "Invalid configuration detected: local models are disabled (`env.allowLocalModels=false`) but you have requested to only use local models (`local_files_only=true`)."
      );
    } else if (!env.allowRemoteModels) {
      throw Error(
        "Invalid configuration detected: both local and remote models are disabled. Fix by setting `env.allowLocalModels` or `env.allowRemoteModels` to `true`."
      );
    }
  }
  dispatchCallback(options.progress_callback, {
    status: "initiate",
    name: path_or_repo_id,
    file: filename
  });
  const key = `${path_or_repo_id}::${filename}`;
  let pending = INFLIGHT_LOADS.get(key);
  if (!pending) {
    const cache2 = await getCache(options?.cache_dir);
    pending = loadResourceFile(path_or_repo_id, filename, fatal, options, return_path, cache2).then(
      (result) => {
        INFLIGHT_LOADS.delete(key);
        return result;
      },
      (err) => {
        INFLIGHT_LOADS.delete(key);
        throw err;
      }
    );
    INFLIGHT_LOADS.set(key, pending);
  }
  return await pending;
}
async function getModelText(modelPath, fileName, fatal = true, options = {}) {
  const buffer = await getModelFile(modelPath, fileName, fatal, options, false);
  if (buffer === null) {
    return null;
  }
  const decoder = new TextDecoder("utf-8");
  return decoder.decode(
    /** @type {Uint8Array} */
    buffer
  );
}
async function getModelJSON(modelPath, fileName, fatal = true, options = {}) {
  const text = await getModelText(modelPath, fileName, fatal, options);
  if (text === null) {
    return {};
  }
  return JSON.parse(text);
}

// src/utils/maths.js
function interpolate_data(input, [in_channels, in_height, in_width], [out_height, out_width], mode = "bilinear", align_corners = false) {
  const x_scale = out_width / in_width;
  const y_scale = out_height / in_height;
  const out_img = new input.constructor(out_height * out_width * in_channels);
  const inStride = in_height * in_width;
  const outStride = out_height * out_width;
  for (let i = 0; i < out_height; ++i) {
    for (let j = 0; j < out_width; ++j) {
      const outOffset = i * out_width + j;
      const x = (j + 0.5) / x_scale - 0.5;
      const y = (i + 0.5) / y_scale - 0.5;
      let x1 = Math.floor(x);
      let y1 = Math.floor(y);
      const x2 = Math.min(x1 + 1, in_width - 1);
      const y2 = Math.min(y1 + 1, in_height - 1);
      x1 = Math.max(x1, 0);
      y1 = Math.max(y1, 0);
      const s = x - x1;
      const t = y - y1;
      const w1 = (1 - s) * (1 - t);
      const w2 = s * (1 - t);
      const w3 = (1 - s) * t;
      const w4 = s * t;
      const yStride = y1 * in_width;
      const xStride = y2 * in_width;
      const idx1 = yStride + x1;
      const idx2 = yStride + x2;
      const idx3 = xStride + x1;
      const idx4 = xStride + x2;
      for (let k2 = 0; k2 < in_channels; ++k2) {
        const cOffset = k2 * inStride;
        out_img[k2 * outStride + outOffset] = w1 * input[cOffset + idx1] + w2 * input[cOffset + idx2] + w3 * input[cOffset + idx3] + w4 * input[cOffset + idx4];
      }
    }
  }
  return out_img;
}
function permute_data(array, dims, axes) {
  const shape = new Array(axes.length);
  const stride = new Array(axes.length);
  for (let i = axes.length - 1, s = 1; i >= 0; --i) {
    stride[i] = s;
    shape[i] = dims[axes[i]];
    s *= shape[i];
  }
  const invStride = axes.map((_, i) => stride[axes.indexOf(i)]);
  const permutedData = new array.constructor(array.length);
  for (let i = 0; i < array.length; ++i) {
    let newIndex = 0;
    for (let j = dims.length - 1, k2 = i; j >= 0; --j) {
      newIndex += k2 % dims[j] * invStride[j];
      k2 = Math.floor(k2 / dims[j]);
    }
    permutedData[newIndex] = array[i];
  }
  return [permutedData, shape];
}
function softmax(arr) {
  const maxVal = max(arr)[0];
  const exps = arr.map((x) => Math.exp(x - maxVal));
  const sumExps = exps.reduce((acc, val) => acc + val, 0);
  const softmaxArr = exps.map((x) => x / sumExps);
  return (
    /** @type {T} */
    softmaxArr
  );
}
function log_softmax(arr) {
  const maxVal = max(arr)[0];
  let sumExps = 0;
  for (let i = 0; i < arr.length; ++i) {
    sumExps += Math.exp(arr[i] - maxVal);
  }
  const logSum = Math.log(sumExps);
  const logSoftmaxArr = arr.map((x) => x - maxVal - logSum);
  return (
    /** @type {T} */
    logSoftmaxArr
  );
}
function dot(arr1, arr2) {
  let result = 0;
  for (let i = 0; i < arr1.length; ++i) {
    result += arr1[i] * arr2[i];
  }
  return result;
}
function cos_sim(arr1, arr2) {
  const dotProduct = dot(arr1, arr2);
  const magnitudeA = magnitude(arr1);
  const magnitudeB = magnitude(arr2);
  const cosineSimilarity = dotProduct / (magnitudeA * magnitudeB);
  return cosineSimilarity;
}
function magnitude(arr) {
  return Math.sqrt(arr.reduce((acc, val) => acc + val * val, 0));
}
function min2(arr) {
  if (arr.length === 0) throw Error("Array must not be empty");
  let min3 = arr[0];
  let indexOfMin = 0;
  for (let i = 1; i < arr.length; ++i) {
    if (arr[i] < min3) {
      min3 = arr[i];
      indexOfMin = i;
    }
  }
  return (
    /** @type {T extends bigint[]|BigTypedArray ? [bigint, number] : [number, number]} */
    [min3, indexOfMin]
  );
}
function max(arr) {
  if (arr.length === 0) throw Error("Array must not be empty");
  let max2 = arr[0];
  let indexOfMax = 0;
  for (let i = 1; i < arr.length; ++i) {
    if (arr[i] > max2) {
      max2 = arr[i];
      indexOfMax = i;
    }
  }
  return (
    /** @type {T extends bigint[]|BigTypedArray ? [bigint, number] : [number, number]} */
    [max2, indexOfMax]
  );
}
function isPowerOfTwo(number) {
  return number > 0 && (number & number - 1) === 0;
}
var P2FFT = class {
  /**
   * @param {number} size The size of the input array. Must be a power of two larger than 1.
   * @throws {Error} FFT size must be a power of two larger than 1.
   */
  constructor(size) {
    this.size = size | 0;
    if (this.size <= 1 || !isPowerOfTwo(this.size))
      throw new Error("FFT size must be a power of two larger than 1");
    this._csize = size << 1;
    this.table = new Float64Array(this.size * 2);
    for (let i = 0; i < this.table.length; i += 2) {
      const angle = Math.PI * i / this.size;
      this.table[i] = Math.cos(angle);
      this.table[i + 1] = -Math.sin(angle);
    }
    let power = 0;
    for (let t = 1; this.size > t; t <<= 1) ++power;
    this._width = power % 2 === 0 ? power - 1 : power;
    this._bitrev = new Int32Array(1 << this._width);
    for (let j = 0; j < this._bitrev.length; ++j) {
      this._bitrev[j] = 0;
      for (let shift = 0; shift < this._width; shift += 2) {
        const revShift = this._width - shift - 2;
        this._bitrev[j] |= (j >>> shift & 3) << revShift;
      }
    }
  }
  /**
   * Create a complex number array with size `2 * size`
   *
   * @returns {Float64Array} A complex number array with size `2 * size`
   */
  createComplexArray() {
    return new Float64Array(this._csize);
  }
  /**
   * Converts a complex number representation stored in a Float64Array to an array of real numbers.
   *
   * @param {Float64Array} complex The complex number representation to be converted.
   * @param {number[]} [storage] An optional array to store the result in.
   * @returns {number[]} An array of real numbers representing the input complex number representation.
   */
  fromComplexArray(complex, storage) {
    const res = storage || new Array(complex.length >>> 1);
    for (let i = 0; i < complex.length; i += 2) res[i >>> 1] = complex[i];
    return res;
  }
  /**
   * Convert a real-valued input array to a complex-valued output array.
   * @param {Float64Array} input The real-valued input array.
   * @param {Float64Array} [storage] Optional buffer to store the output array.
   * @returns {Float64Array} The complex-valued output array.
   */
  toComplexArray(input, storage) {
    const res = storage || this.createComplexArray();
    for (let i = 0; i < res.length; i += 2) {
      res[i] = input[i >>> 1];
      res[i + 1] = 0;
    }
    return res;
  }
  /**
   * Performs a Fast Fourier Transform (FFT) on the given input data and stores the result in the output buffer.
   *
   * @param {Float64Array} out The output buffer to store the result.
   * @param {Float64Array} data The input data to transform.
   *
   * @throws {Error} Input and output buffers must be different.
   *
   * @returns {void}
   */
  transform(out, data) {
    if (out === data) throw new Error("Input and output buffers must be different");
    this._transform4(
      out,
      data,
      1
      /* DONE */
    );
  }
  /**
   * Performs a real-valued forward FFT on the given input buffer and stores the result in the given output buffer.
   * The input buffer must contain real values only, while the output buffer will contain complex values. The input and
   * output buffers must be different.
   *
   * @param {Float64Array} out The output buffer.
   * @param {Float64Array} data The input buffer containing real values.
   *
   * @throws {Error} If the input and output buffers are the same.
   */
  realTransform(out, data) {
    if (out === data) throw new Error("Input and output buffers must be different");
    this._realTransform4(
      out,
      data,
      1
      /* DONE */
    );
  }
  /**
   * Performs an inverse FFT transformation on the given `data` array, and stores the result in `out`.
   * The `out` array must be a different buffer than the `data` array. The `out` array will contain the
   * result of the transformation. The `data` array will not be modified.
   *
   * @param {Float64Array} out The output buffer for the transformed data.
   * @param {Float64Array} data The input data to transform.
   * @throws {Error} If `out` and `data` refer to the same buffer.
   * @returns {void}
   */
  inverseTransform(out, data) {
    if (out === data) throw new Error("Input and output buffers must be different");
    this._transform4(
      out,
      data,
      -1
      /* DONE */
    );
    for (let i = 0; i < out.length; ++i) out[i] /= this.size;
  }
  /**
   * Performs a radix-4 implementation of a discrete Fourier transform on a given set of data.
   *
   * @param {Float64Array} out The output buffer for the transformed data.
   * @param {Float64Array} data The input buffer of data to be transformed.
   * @param {number} inv A scaling factor to apply to the transform.
   * @returns {void}
   */
  _transform4(out, data, inv) {
    const size = this._csize;
    const width = this._width;
    let step = 1 << width;
    let len2 = size / step << 1;
    let outOff;
    let t;
    const bitrev = this._bitrev;
    if (len2 === 4) {
      for (outOff = 0, t = 0; outOff < size; outOff += len2, ++t) {
        const off = bitrev[t];
        this._singleTransform2(data, out, outOff, off, step);
      }
    } else {
      for (outOff = 0, t = 0; outOff < size; outOff += len2, ++t) {
        const off = bitrev[t];
        this._singleTransform4(data, out, outOff, off, step, inv);
      }
    }
    const table = this.table;
    for (step >>= 2; step >= 2; step >>= 2) {
      len2 = size / step << 1;
      const quarterLen = len2 >>> 2;
      for (outOff = 0; outOff < size; outOff += len2) {
        const limit = outOff + quarterLen - 1;
        for (let i = outOff, k2 = 0; i < limit; i += 2, k2 += step) {
          const A = i;
          const B = A + quarterLen;
          const C = B + quarterLen;
          const D = C + quarterLen;
          const Ar = out[A];
          const Ai = out[A + 1];
          const Br = out[B];
          const Bi = out[B + 1];
          const Cr = out[C];
          const Ci = out[C + 1];
          const Dr = out[D];
          const Di = out[D + 1];
          const tableBr = table[k2];
          const tableBi = inv * table[k2 + 1];
          const MBr = Br * tableBr - Bi * tableBi;
          const MBi = Br * tableBi + Bi * tableBr;
          const tableCr = table[2 * k2];
          const tableCi = inv * table[2 * k2 + 1];
          const MCr = Cr * tableCr - Ci * tableCi;
          const MCi = Cr * tableCi + Ci * tableCr;
          const tableDr = table[3 * k2];
          const tableDi = inv * table[3 * k2 + 1];
          const MDr = Dr * tableDr - Di * tableDi;
          const MDi = Dr * tableDi + Di * tableDr;
          const T0r = Ar + MCr;
          const T0i = Ai + MCi;
          const T1r = Ar - MCr;
          const T1i = Ai - MCi;
          const T2r = MBr + MDr;
          const T2i = MBi + MDi;
          const T3r = inv * (MBr - MDr);
          const T3i = inv * (MBi - MDi);
          out[A] = T0r + T2r;
          out[A + 1] = T0i + T2i;
          out[B] = T1r + T3i;
          out[B + 1] = T1i - T3r;
          out[C] = T0r - T2r;
          out[C + 1] = T0i - T2i;
          out[D] = T1r - T3i;
          out[D + 1] = T1i + T3r;
        }
      }
    }
  }
  /**
   * Performs a radix-2 implementation of a discrete Fourier transform on a given set of data.
   *
   * @param {Float64Array} data The input buffer of data to be transformed.
   * @param {Float64Array} out The output buffer for the transformed data.
   * @param {number} outOff The offset at which to write the output data.
   * @param {number} off The offset at which to begin reading the input data.
   * @param {number} step The step size for indexing the input data.
   * @returns {void}
   */
  _singleTransform2(data, out, outOff, off, step) {
    const evenR = data[off];
    const evenI = data[off + 1];
    const oddR = data[off + step];
    const oddI = data[off + step + 1];
    out[outOff] = evenR + oddR;
    out[outOff + 1] = evenI + oddI;
    out[outOff + 2] = evenR - oddR;
    out[outOff + 3] = evenI - oddI;
  }
  /**
   * Performs radix-4 transformation on input data of length 8
   *
   * @param {Float64Array} data Input data array of length 8
   * @param {Float64Array} out Output data array of length 8
   * @param {number} outOff Index of output array to start writing from
   * @param {number} off Index of input array to start reading from
   * @param {number} step Step size between elements in input array
   * @param {number} inv Scaling factor for inverse transform
   *
   * @returns {void}
   */
  _singleTransform4(data, out, outOff, off, step, inv) {
    const step2 = step * 2;
    const step3 = step * 3;
    const Ar = data[off];
    const Ai = data[off + 1];
    const Br = data[off + step];
    const Bi = data[off + step + 1];
    const Cr = data[off + step2];
    const Ci = data[off + step2 + 1];
    const Dr = data[off + step3];
    const Di = data[off + step3 + 1];
    const T0r = Ar + Cr;
    const T0i = Ai + Ci;
    const T1r = Ar - Cr;
    const T1i = Ai - Ci;
    const T2r = Br + Dr;
    const T2i = Bi + Di;
    const T3r = inv * (Br - Dr);
    const T3i = inv * (Bi - Di);
    out[outOff] = T0r + T2r;
    out[outOff + 1] = T0i + T2i;
    out[outOff + 2] = T1r + T3i;
    out[outOff + 3] = T1i - T3r;
    out[outOff + 4] = T0r - T2r;
    out[outOff + 5] = T0i - T2i;
    out[outOff + 6] = T1r - T3i;
    out[outOff + 7] = T1i + T3r;
  }
  /**
   * Real input radix-4 implementation
   * @param {Float64Array} out Output array for the transformed data
   * @param {Float64Array} data Input array of real data to be transformed
   * @param {number} inv The scale factor used to normalize the inverse transform
   */
  _realTransform4(out, data, inv) {
    const size = this._csize;
    const width = this._width;
    let step = 1 << width;
    let len2 = size / step << 1;
    let outOff;
    let t;
    const bitrev = this._bitrev;
    if (len2 === 4) {
      for (outOff = 0, t = 0; outOff < size; outOff += len2, ++t) {
        const off = bitrev[t];
        this._singleRealTransform2(data, out, outOff, off >>> 1, step >>> 1);
      }
    } else {
      for (outOff = 0, t = 0; outOff < size; outOff += len2, ++t) {
        const off = bitrev[t];
        this._singleRealTransform4(data, out, outOff, off >>> 1, step >>> 1, inv);
      }
    }
    const table = this.table;
    for (step >>= 2; step >= 2; step >>= 2) {
      len2 = size / step << 1;
      const halfLen = len2 >>> 1;
      const quarterLen = halfLen >>> 1;
      const hquarterLen = quarterLen >>> 1;
      for (outOff = 0; outOff < size; outOff += len2) {
        for (let i = 0, k2 = 0; i <= hquarterLen; i += 2, k2 += step) {
          const A = outOff + i;
          const B = A + quarterLen;
          const C = B + quarterLen;
          const D = C + quarterLen;
          const Ar = out[A];
          const Ai = out[A + 1];
          const Br = out[B];
          const Bi = out[B + 1];
          const Cr = out[C];
          const Ci = out[C + 1];
          const Dr = out[D];
          const Di = out[D + 1];
          const MAr = Ar;
          const MAi = Ai;
          const tableBr = table[k2];
          const tableBi = inv * table[k2 + 1];
          const MBr = Br * tableBr - Bi * tableBi;
          const MBi = Br * tableBi + Bi * tableBr;
          const tableCr = table[2 * k2];
          const tableCi = inv * table[2 * k2 + 1];
          const MCr = Cr * tableCr - Ci * tableCi;
          const MCi = Cr * tableCi + Ci * tableCr;
          const tableDr = table[3 * k2];
          const tableDi = inv * table[3 * k2 + 1];
          const MDr = Dr * tableDr - Di * tableDi;
          const MDi = Dr * tableDi + Di * tableDr;
          const T0r = MAr + MCr;
          const T0i = MAi + MCi;
          const T1r = MAr - MCr;
          const T1i = MAi - MCi;
          const T2r = MBr + MDr;
          const T2i = MBi + MDi;
          const T3r = inv * (MBr - MDr);
          const T3i = inv * (MBi - MDi);
          out[A] = T0r + T2r;
          out[A + 1] = T0i + T2i;
          out[B] = T1r + T3i;
          out[B + 1] = T1i - T3r;
          if (i === 0) {
            out[C] = T0r - T2r;
            out[C + 1] = T0i - T2i;
            continue;
          }
          if (i === hquarterLen) continue;
          const SA = outOff + quarterLen - i;
          const SB = outOff + halfLen - i;
          out[SA] = T1r - inv * T3i;
          out[SA + 1] = -T1i - inv * T3r;
          out[SB] = T0r - inv * T2r;
          out[SB + 1] = -T0i + inv * T2i;
        }
      }
    }
    const half = size >>> 1;
    for (let i = 2; i < half; i += 2) {
      out[size - i] = out[i];
      out[size - i + 1] = -out[i + 1];
    }
  }
  /**
   * Performs a single real input radix-2 transformation on the provided data
   *
   * @param {Float64Array} data The input data array
   * @param {Float64Array} out The output data array
   * @param {number} outOff The output offset
   * @param {number} off The input offset
   * @param {number} step The step
   *
   * @returns {void}
   */
  _singleRealTransform2(data, out, outOff, off, step) {
    const evenR = data[off];
    const oddR = data[off + step];
    out[outOff] = evenR + oddR;
    out[outOff + 1] = 0;
    out[outOff + 2] = evenR - oddR;
    out[outOff + 3] = 0;
  }
  /**
   * Computes a single real-valued transform using radix-4 algorithm.
   * This method is only called for len=8.
   *
   * @param {Float64Array} data The input data array.
   * @param {Float64Array} out The output data array.
   * @param {number} outOff The offset into the output array.
   * @param {number} off The offset into the input array.
   * @param {number} step The step size for the input array.
   * @param {number} inv The value of inverse.
   */
  _singleRealTransform4(data, out, outOff, off, step, inv) {
    const step2 = step * 2;
    const step3 = step * 3;
    const Ar = data[off];
    const Br = data[off + step];
    const Cr = data[off + step2];
    const Dr = data[off + step3];
    const T0r = Ar + Cr;
    const T1r = Ar - Cr;
    const T2r = Br + Dr;
    const T3r = inv * (Br - Dr);
    out[outOff] = T0r + T2r;
    out[outOff + 1] = 0;
    out[outOff + 2] = T1r;
    out[outOff + 3] = -T3r;
    out[outOff + 4] = T0r - T2r;
    out[outOff + 5] = 0;
    out[outOff + 6] = T1r;
    out[outOff + 7] = T3r;
  }
};
var NP2FFT = class {
  /**
   * Constructs a new NP2FFT object.
   * @param {number} fft_length The length of the FFT
   */
  constructor(fft_length) {
    const a = 2 * (fft_length - 1);
    const b = 2 * (2 * fft_length - 1);
    const nextP2 = 2 ** Math.ceil(Math.log2(b));
    this.bufferSize = nextP2;
    this._a = a;
    const chirp = new Float64Array(b);
    const ichirp = new Float64Array(nextP2);
    this._chirpBuffer = new Float64Array(nextP2);
    this._buffer1 = new Float64Array(nextP2);
    this._buffer2 = new Float64Array(nextP2);
    this._outBuffer1 = new Float64Array(nextP2);
    this._outBuffer2 = new Float64Array(nextP2);
    const theta = -2 * Math.PI / fft_length;
    const baseR = Math.cos(theta);
    const baseI = Math.sin(theta);
    for (let i = 0; i < b >> 1; ++i) {
      const e = (i + 1 - fft_length) ** 2 / 2;
      const result_mod = Math.sqrt(baseR ** 2 + baseI ** 2) ** e;
      const result_arg = e * Math.atan2(baseI, baseR);
      const i2 = 2 * i;
      chirp[i2] = result_mod * Math.cos(result_arg);
      chirp[i2 + 1] = result_mod * Math.sin(result_arg);
      ichirp[i2] = chirp[i2];
      ichirp[i2 + 1] = -chirp[i2 + 1];
    }
    this._slicedChirpBuffer = chirp.subarray(a, b);
    this._f = new P2FFT(nextP2 >> 1);
    this._f.transform(this._chirpBuffer, ichirp);
  }
  _transform(output, input, real) {
    const ib1 = this._buffer1;
    const ib2 = this._buffer2;
    const ob2 = this._outBuffer1;
    const ob3 = this._outBuffer2;
    const cb = this._chirpBuffer;
    const sb = this._slicedChirpBuffer;
    const a = this._a;
    if (real) {
      for (let j = 0; j < sb.length; j += 2) {
        const j2 = j + 1;
        const j3 = j >> 1;
        const a_real = input[j3];
        ib1[j] = a_real * sb[j];
        ib1[j2] = a_real * sb[j2];
      }
    } else {
      for (let j = 0; j < sb.length; j += 2) {
        const j2 = j + 1;
        ib1[j] = input[j] * sb[j] - input[j2] * sb[j2];
        ib1[j2] = input[j] * sb[j2] + input[j2] * sb[j];
      }
    }
    this._f.transform(ob2, ib1);
    for (let j = 0; j < cb.length; j += 2) {
      const j2 = j + 1;
      ib2[j] = ob2[j] * cb[j] - ob2[j2] * cb[j2];
      ib2[j2] = ob2[j] * cb[j2] + ob2[j2] * cb[j];
    }
    this._f.inverseTransform(ob3, ib2);
    for (let j = 0; j < ob3.length; j += 2) {
      const a_real = ob3[j + a];
      const a_imag = ob3[j + a + 1];
      const b_real = sb[j];
      const b_imag = sb[j + 1];
      output[j] = a_real * b_real - a_imag * b_imag;
      output[j + 1] = a_real * b_imag + a_imag * b_real;
    }
  }
  transform(output, input) {
    this._transform(output, input, false);
  }
  realTransform(output, input) {
    this._transform(output, input, true);
  }
};
var FFT = class {
  constructor(fft_length) {
    this.fft_length = fft_length;
    this.isPowerOfTwo = isPowerOfTwo(fft_length);
    if (this.isPowerOfTwo) {
      this.fft = new P2FFT(fft_length);
      this.outputBufferSize = 2 * fft_length;
    } else {
      this.fft = new NP2FFT(fft_length);
      this.outputBufferSize = this.fft.bufferSize;
    }
  }
  realTransform(out, input) {
    this.fft.realTransform(out, input);
  }
  transform(out, input) {
    this.fft.transform(out, input);
  }
};
function medianFilter(data, windowSize) {
  if (windowSize % 2 === 0 || windowSize <= 0) {
    throw new Error("Window size must be a positive odd number");
  }
  const outputArray = new data.constructor(data.length);
  const buffer = new data.constructor(windowSize);
  const halfWindowSize = Math.floor(windowSize / 2);
  for (let i = 0; i < data.length; ++i) {
    let valuesIndex = 0;
    for (let j = -halfWindowSize; j <= halfWindowSize; ++j) {
      let index = i + j;
      if (index < 0) {
        index = Math.abs(index);
      } else if (index >= data.length) {
        index = 2 * (data.length - 1) - index;
      }
      buffer[valuesIndex++] = data[index];
    }
    buffer.sort();
    outputArray[i] = buffer[halfWindowSize];
  }
  return outputArray;
}
function round(num, decimals) {
  const pow = Math.pow(10, decimals);
  return Math.round(num * pow) / pow;
}
function bankers_round(x) {
  const r = Math.round(x);
  const br = Math.abs(x) % 1 === 0.5 ? r % 2 === 0 ? r : r - 1 : r;
  return br;
}
function dynamic_time_warping(matrix) {
  const output_length = matrix.length;
  const input_length = matrix[0].length;
  const outputShape = [output_length + 1, input_length + 1];
  const cost = Array.from({ length: outputShape[0] }, () => Array(outputShape[1]).fill(Infinity));
  cost[0][0] = 0;
  const trace = Array.from({ length: outputShape[0] }, () => Array(outputShape[1]).fill(-1));
  for (let j2 = 1; j2 < outputShape[1]; ++j2) {
    for (let i2 = 1; i2 < outputShape[0]; ++i2) {
      const c0 = cost[i2 - 1][j2 - 1];
      const c1 = cost[i2 - 1][j2];
      const c2 = cost[i2][j2 - 1];
      let c, t;
      if (c0 < c1 && c0 < c2) {
        c = c0;
        t = 0;
      } else if (c1 < c0 && c1 < c2) {
        c = c1;
        t = 1;
      } else {
        c = c2;
        t = 2;
      }
      cost[i2][j2] = matrix[i2 - 1][j2 - 1] + c;
      trace[i2][j2] = t;
    }
  }
  for (let i2 = 0; i2 < outputShape[1]; ++i2) {
    trace[0][i2] = 2;
  }
  for (let i2 = 0; i2 < outputShape[0]; ++i2) {
    trace[i2][0] = 1;
  }
  let i = output_length;
  let j = input_length;
  let text_indices = [];
  let time_indices = [];
  while (i > 0 || j > 0) {
    text_indices.push(i - 1);
    time_indices.push(j - 1);
    switch (trace[i][j]) {
      case 0:
        --i;
        --j;
        break;
      case 1:
        --i;
        break;
      case 2:
        --j;
        break;
      default:
        throw new Error(
          `Internal error in dynamic time warping. Unexpected trace[${i}, ${j}]. Please file a bug report.`
        );
    }
  }
  text_indices.reverse();
  time_indices.reverse();
  return [text_indices, time_indices];
}
var uint16_to_float32 = /* @__PURE__ */ (function() {
  let float16LUT = null;
  return function(u16Array) {
    if (!float16LUT) {
      float16LUT = new Float32Array(65536);
      const buffer = new ArrayBuffer(4);
      const u32 = new Uint32Array(buffer);
      const f32 = new Float32Array(buffer);
      for (let i = 0; i < float16LUT.length; ++i) {
        let outBits = 0;
        const sign = (i & 32768) << 16;
        const exp = (i & 31744) >> 10;
        let mantissa = i & 1023;
        if (exp === 31) {
          outBits = sign | 2139095040 | mantissa << 13;
        } else if (exp === 0) {
          if (mantissa === 0) {
            outBits = sign;
          } else {
            let renormExp = 113;
            while ((mantissa & 1024) === 0) {
              mantissa <<= 1;
              --renormExp;
            }
            mantissa &= ~1024;
            outBits = sign | renormExp << 23 | mantissa << 13;
          }
        } else {
          outBits = sign | exp + 112 << 23 | mantissa << 13;
        }
        u32[0] = outBits;
        float16LUT[i] = f32[0];
      }
    }
    const length = u16Array.length;
    const lut = float16LUT;
    const out = new Float32Array(length);
    for (let i = 0; i < length; ++i) {
      out[i] = lut[u16Array[i]];
    }
    return out;
  };
})();

// src/backends/onnx.js
import * as ONNX_NODE from "onnxruntime-node";

// ../../node_modules/.pnpm/onnxruntime-web@1.26.0-dev.20260416-b7804b056c/node_modules/onnxruntime-web/dist/ort.webgpu.bundle.min.mjs
var ort_webgpu_bundle_min_exports = {};
__export(ort_webgpu_bundle_min_exports, {
  InferenceSession: () => qf,
  TRACE: () => Ga,
  TRACE_EVENT_BEGIN: () => $e,
  TRACE_EVENT_END: () => ze,
  TRACE_FUNC_BEGIN: () => tt,
  TRACE_FUNC_END: () => rt,
  Tensor: () => Le,
  default: () => gl,
  env: () => K,
  registerBackend: () => Ke
});
var jr = Object.defineProperty;
var zf = Object.getOwnPropertyDescriptor;
var Vf = Object.getOwnPropertyNames;
var jf = Object.prototype.hasOwnProperty;
var Hr = ((a) => typeof __require < "u" ? __require : typeof Proxy < "u" ? new Proxy(a, { get: (r, s) => (typeof __require < "u" ? __require : r)[s] }) : a)(function(a) {
  if (typeof __require < "u") return __require.apply(this, arguments);
  throw Error('Dynamic require of "' + a + '" is not supported');
});
var k = (a, r) => () => (a && (r = a(a = 0)), r);
var At = (a, r) => {
  for (var s in r) jr(a, s, { get: r[s], enumerable: true });
};
var Hf = (a, r, s, f) => {
  if (r && typeof r == "object" || typeof r == "function") for (let i of Vf(r)) !jf.call(a, i) && i !== s && jr(a, i, { get: () => r[i], enumerable: !(f = zf(r, i)) || f.enumerable });
  return a;
};
var $t = (a) => Hf(jr({}, "__esModule", { value: true }), a);
var zt;
var Ze;
var Ke;
var Yf;
var Ta;
var Yr = k(() => {
  "use strict";
  zt = /* @__PURE__ */ new Map(), Ze = [], Ke = (a, r, s) => {
    if (r && typeof r.init == "function" && typeof r.createInferenceSessionHandler == "function") {
      let f = zt.get(a);
      if (f === void 0) zt.set(a, { backend: r, priority: s });
      else {
        if (f.priority > s) return;
        if (f.priority === s && f.backend !== r) throw new Error(`cannot register backend "${a}" using priority ${s}`);
      }
      if (s >= 0) {
        let i = Ze.indexOf(a);
        i !== -1 && Ze.splice(i, 1);
        for (let d = 0; d < Ze.length; d++) if (zt.get(Ze[d]).priority <= s) {
          Ze.splice(d, 0, a);
          return;
        }
        Ze.push(a);
      }
      return;
    }
    throw new TypeError("not a valid backend");
  }, Yf = async (a) => {
    let r = zt.get(a);
    if (!r) return "backend not found.";
    if (r.initialized) return r.backend;
    if (r.aborted) return r.error;
    {
      let s = !!r.initPromise;
      try {
        return s || (r.initPromise = r.backend.init(a)), await r.initPromise, r.initialized = true, r.backend;
      } catch (f) {
        return s || (r.error = `${f}`, r.aborted = true), r.error;
      } finally {
        delete r.initPromise;
      }
    }
  }, Ta = async (a) => {
    let r = a.executionProviders || [], s = r.map((y) => typeof y == "string" ? y : y.name), f = s.length === 0 ? Ze : s, i, d = [], l = /* @__PURE__ */ new Set();
    for (let y of f) {
      let w = await Yf(y);
      typeof w == "string" ? d.push({ name: y, err: w }) : (i || (i = w), i === w && l.add(y));
    }
    if (!i) throw new Error(`no available backend found. ERR: ${d.map((y) => `[${y.name}] ${y.err}`).join(", ")}`);
    for (let { name: y, err: w } of d) s.includes(y) && console.warn(`removing requested execution provider "${y}" from session options because it is not available: ${w}`);
    let m = r.filter((y) => l.has(typeof y == "string" ? y : y.name));
    return [i, new Proxy(a, { get: (y, w) => w === "executionProviders" ? m : Reflect.get(y, w) })];
  };
});
var va = k(() => {
  "use strict";
  Yr();
});
var Ea;
var Sa = k(() => {
  "use strict";
  Ea = "1.24.0-dev.20251116-b39e144322";
});
var Aa;
var ie;
var qr = k(() => {
  "use strict";
  Sa();
  Aa = "warning", ie = { wasm: {}, webgl: {}, webgpu: {}, versions: { common: Ea }, set logLevel(a) {
    if (a !== void 0) {
      if (typeof a != "string" || ["verbose", "info", "warning", "error", "fatal"].indexOf(a) === -1) throw new Error(`Unsupported logging level: ${a}`);
      Aa = a;
    }
  }, get logLevel() {
    return Aa;
  } };
  Object.defineProperty(ie, "logLevel", { enumerable: true });
});
var K;
var Ia = k(() => {
  "use strict";
  qr();
  K = ie;
});
var xa;
var La;
var Oa = k(() => {
  "use strict";
  xa = (a, r) => {
    let s = typeof document < "u" ? document.createElement("canvas") : new OffscreenCanvas(1, 1);
    s.width = a.dims[3], s.height = a.dims[2];
    let f = s.getContext("2d");
    if (f != null) {
      let i, d;
      r?.tensorLayout !== void 0 && r.tensorLayout === "NHWC" ? (i = a.dims[2], d = a.dims[3]) : (i = a.dims[3], d = a.dims[2]);
      let l = r?.format !== void 0 ? r.format : "RGB", m = r?.norm, y, w;
      m === void 0 || m.mean === void 0 ? y = [255, 255, 255, 255] : typeof m.mean == "number" ? y = [m.mean, m.mean, m.mean, m.mean] : (y = [m.mean[0], m.mean[1], m.mean[2], 0], m.mean[3] !== void 0 && (y[3] = m.mean[3])), m === void 0 || m.bias === void 0 ? w = [0, 0, 0, 0] : typeof m.bias == "number" ? w = [m.bias, m.bias, m.bias, m.bias] : (w = [m.bias[0], m.bias[1], m.bias[2], 0], m.bias[3] !== void 0 && (w[3] = m.bias[3]));
      let T = d * i, g = 0, v = T, S = T * 2, C = -1;
      l === "RGBA" ? (g = 0, v = T, S = T * 2, C = T * 3) : l === "RGB" ? (g = 0, v = T, S = T * 2) : l === "RBG" && (g = 0, S = T, v = T * 2);
      for (let R = 0; R < d; R++) for (let H = 0; H < i; H++) {
        let U = (a.data[g++] - w[0]) * y[0], M = (a.data[v++] - w[1]) * y[1], Y = (a.data[S++] - w[2]) * y[2], O = C === -1 ? 255 : (a.data[C++] - w[3]) * y[3];
        f.fillStyle = "rgba(" + U + "," + M + "," + Y + "," + O + ")", f.fillRect(H, R, 1, 1);
      }
      if ("toDataURL" in s) return s.toDataURL();
      throw new Error("toDataURL is not supported");
    } else throw new Error("Can not access image data");
  }, La = (a, r) => {
    let s = typeof document < "u" ? document.createElement("canvas").getContext("2d") : new OffscreenCanvas(1, 1).getContext("2d"), f;
    if (s != null) {
      let i, d, l;
      r?.tensorLayout !== void 0 && r.tensorLayout === "NHWC" ? (i = a.dims[2], d = a.dims[1], l = a.dims[3]) : (i = a.dims[3], d = a.dims[2], l = a.dims[1]);
      let m = r !== void 0 && r.format !== void 0 ? r.format : "RGB", y = r?.norm, w, T;
      y === void 0 || y.mean === void 0 ? w = [255, 255, 255, 255] : typeof y.mean == "number" ? w = [y.mean, y.mean, y.mean, y.mean] : (w = [y.mean[0], y.mean[1], y.mean[2], 255], y.mean[3] !== void 0 && (w[3] = y.mean[3])), y === void 0 || y.bias === void 0 ? T = [0, 0, 0, 0] : typeof y.bias == "number" ? T = [y.bias, y.bias, y.bias, y.bias] : (T = [y.bias[0], y.bias[1], y.bias[2], 0], y.bias[3] !== void 0 && (T[3] = y.bias[3]));
      let g = d * i;
      if (r !== void 0 && (r.format !== void 0 && l === 4 && r.format !== "RGBA" || l === 3 && r.format !== "RGB" && r.format !== "BGR")) throw new Error("Tensor format doesn't match input tensor dims");
      let v = 4, S = 0, C = 1, R = 2, H = 3, U = 0, M = g, Y = g * 2, O = -1;
      m === "RGBA" ? (U = 0, M = g, Y = g * 2, O = g * 3) : m === "RGB" ? (U = 0, M = g, Y = g * 2) : m === "RBG" && (U = 0, Y = g, M = g * 2), f = s.createImageData(i, d);
      for (let W = 0; W < d * i; S += v, C += v, R += v, H += v, W++) f.data[S] = (a.data[U++] - T[0]) * w[0], f.data[C] = (a.data[M++] - T[1]) * w[1], f.data[R] = (a.data[Y++] - T[2]) * w[2], f.data[H] = O === -1 ? 255 : (a.data[O++] - T[3]) * w[3];
    } else throw new Error("Can not access image data");
    return f;
  };
});
var Jr;
var Ba;
var Ma;
var Ca;
var Ua;
var Da;
var Pa = k(() => {
  "use strict";
  Vt();
  Jr = (a, r) => {
    if (a === void 0) throw new Error("Image buffer must be defined");
    if (r.height === void 0 || r.width === void 0) throw new Error("Image height and width must be defined");
    if (r.tensorLayout === "NHWC") throw new Error("NHWC Tensor layout is not supported yet");
    let { height: s, width: f } = r, i = r.norm ?? { mean: 255, bias: 0 }, d, l;
    typeof i.mean == "number" ? d = [i.mean, i.mean, i.mean, i.mean] : d = [i.mean[0], i.mean[1], i.mean[2], i.mean[3] ?? 255], typeof i.bias == "number" ? l = [i.bias, i.bias, i.bias, i.bias] : l = [i.bias[0], i.bias[1], i.bias[2], i.bias[3] ?? 0];
    let m = r.format !== void 0 ? r.format : "RGBA", y = r.tensorFormat !== void 0 && r.tensorFormat !== void 0 ? r.tensorFormat : "RGB", w = s * f, T = y === "RGBA" ? new Float32Array(w * 4) : new Float32Array(w * 3), g = 4, v = 0, S = 1, C = 2, R = 3, H = 0, U = w, M = w * 2, Y = -1;
    m === "RGB" && (g = 3, v = 0, S = 1, C = 2, R = -1), y === "RGBA" ? Y = w * 3 : y === "RBG" ? (H = 0, M = w, U = w * 2) : y === "BGR" && (M = 0, U = w, H = w * 2);
    for (let W = 0; W < w; W++, v += g, C += g, S += g, R += g) T[H++] = (a[v] + l[0]) / d[0], T[U++] = (a[S] + l[1]) / d[1], T[M++] = (a[C] + l[2]) / d[2], Y !== -1 && R !== -1 && (T[Y++] = (a[R] + l[3]) / d[3]);
    return y === "RGBA" ? new le("float32", T, [1, 4, s, f]) : new le("float32", T, [1, 3, s, f]);
  }, Ba = async (a, r) => {
    let s = typeof HTMLImageElement < "u" && a instanceof HTMLImageElement, f = typeof ImageData < "u" && a instanceof ImageData, i = typeof ImageBitmap < "u" && a instanceof ImageBitmap, d = typeof a == "string", l, m = r ?? {}, y = () => {
      if (typeof document < "u") return document.createElement("canvas");
      if (typeof OffscreenCanvas < "u") return new OffscreenCanvas(1, 1);
      throw new Error("Canvas is not supported");
    }, w = (T) => typeof HTMLCanvasElement < "u" && T instanceof HTMLCanvasElement || T instanceof OffscreenCanvas ? T.getContext("2d") : null;
    if (s) {
      let T = y();
      T.width = a.width, T.height = a.height;
      let g = w(T);
      if (g != null) {
        let v = a.height, S = a.width;
        if (r !== void 0 && r.resizedHeight !== void 0 && r.resizedWidth !== void 0 && (v = r.resizedHeight, S = r.resizedWidth), r !== void 0) {
          if (m = r, r.tensorFormat !== void 0) throw new Error("Image input config format must be RGBA for HTMLImageElement");
          m.tensorFormat = "RGBA", m.height = v, m.width = S;
        } else m.tensorFormat = "RGBA", m.height = v, m.width = S;
        g.drawImage(a, 0, 0), l = g.getImageData(0, 0, S, v).data;
      } else throw new Error("Can not access image data");
    } else if (f) {
      let T, g;
      if (r !== void 0 && r.resizedWidth !== void 0 && r.resizedHeight !== void 0 ? (T = r.resizedHeight, g = r.resizedWidth) : (T = a.height, g = a.width), r !== void 0 && (m = r), m.format = "RGBA", m.height = T, m.width = g, r !== void 0) {
        let v = y();
        v.width = g, v.height = T;
        let S = w(v);
        if (S != null) S.putImageData(a, 0, 0), l = S.getImageData(0, 0, g, T).data;
        else throw new Error("Can not access image data");
      } else l = a.data;
    } else if (i) {
      if (r === void 0) throw new Error("Please provide image config with format for Imagebitmap");
      let T = y();
      T.width = a.width, T.height = a.height;
      let g = w(T);
      if (g != null) {
        let v = a.height, S = a.width;
        return g.drawImage(a, 0, 0, S, v), l = g.getImageData(0, 0, S, v).data, m.height = v, m.width = S, Jr(l, m);
      } else throw new Error("Can not access image data");
    } else {
      if (d) return new Promise((T, g) => {
        let v = y(), S = w(v);
        if (!a || !S) return g();
        let C = new Image();
        C.crossOrigin = "Anonymous", C.src = a, C.onload = () => {
          v.width = C.width, v.height = C.height, S.drawImage(C, 0, 0, v.width, v.height);
          let R = S.getImageData(0, 0, v.width, v.height);
          m.height = v.height, m.width = v.width, T(Jr(R.data, m));
        };
      });
      throw new Error("Input data provided is not supported - aborted tensor creation");
    }
    if (l !== void 0) return Jr(l, m);
    throw new Error("Input data provided is not supported - aborted tensor creation");
  }, Ma = (a, r) => {
    let { width: s, height: f, download: i, dispose: d } = r, l = [1, f, s, 4];
    return new le({ location: "texture", type: "float32", texture: a, dims: l, download: i, dispose: d });
  }, Ca = (a, r) => {
    let { dataType: s, dims: f, download: i, dispose: d } = r;
    return new le({ location: "gpu-buffer", type: s ?? "float32", gpuBuffer: a, dims: f, download: i, dispose: d });
  }, Ua = (a, r) => {
    let { dataType: s, dims: f, download: i, dispose: d } = r;
    return new le({ location: "ml-tensor", type: s ?? "float32", mlTensor: a, dims: f, download: i, dispose: d });
  }, Da = (a, r, s) => new le({ location: "cpu-pinned", type: a, data: r, dims: s ?? [r.length] });
});
var et;
var It;
var _a;
var Ra;
var Na = k(() => {
  "use strict";
  et = /* @__PURE__ */ new Map([["float32", Float32Array], ["uint8", Uint8Array], ["int8", Int8Array], ["uint16", Uint16Array], ["int16", Int16Array], ["int32", Int32Array], ["bool", Uint8Array], ["float64", Float64Array], ["uint32", Uint32Array], ["int4", Uint8Array], ["uint4", Uint8Array]]), It = /* @__PURE__ */ new Map([[Float32Array, "float32"], [Uint8Array, "uint8"], [Int8Array, "int8"], [Uint16Array, "uint16"], [Int16Array, "int16"], [Int32Array, "int32"], [Float64Array, "float64"], [Uint32Array, "uint32"]]), _a = false, Ra = () => {
    if (!_a) {
      _a = true;
      let a = typeof BigInt64Array < "u" && BigInt64Array.from, r = typeof BigUint64Array < "u" && BigUint64Array.from, s = globalThis.Float16Array, f = typeof s < "u" && s.from;
      a && (et.set("int64", BigInt64Array), It.set(BigInt64Array, "int64")), r && (et.set("uint64", BigUint64Array), It.set(BigUint64Array, "uint64")), f ? (et.set("float16", s), It.set(s, "float16")) : et.set("float16", Uint16Array);
    }
  };
});
var ka;
var Wa;
var Fa = k(() => {
  "use strict";
  Vt();
  ka = (a) => {
    let r = 1;
    for (let s = 0; s < a.length; s++) {
      let f = a[s];
      if (typeof f != "number" || !Number.isSafeInteger(f)) throw new TypeError(`dims[${s}] must be an integer, got: ${f}`);
      if (f < 0) throw new RangeError(`dims[${s}] must be a non-negative integer, got: ${f}`);
      r *= f;
    }
    return r;
  }, Wa = (a, r) => {
    switch (a.location) {
      case "cpu":
        return new le(a.type, a.data, r);
      case "cpu-pinned":
        return new le({ location: "cpu-pinned", data: a.data, type: a.type, dims: r });
      case "texture":
        return new le({ location: "texture", texture: a.texture, type: a.type, dims: r });
      case "gpu-buffer":
        return new le({ location: "gpu-buffer", gpuBuffer: a.gpuBuffer, type: a.type, dims: r });
      case "ml-tensor":
        return new le({ location: "ml-tensor", mlTensor: a.mlTensor, type: a.type, dims: r });
      default:
        throw new Error(`tensorReshape: tensor location ${a.location} is not supported`);
    }
  };
});
var le;
var Vt = k(() => {
  "use strict";
  Oa();
  Pa();
  Na();
  Fa();
  le = class {
    constructor(r, s, f) {
      Ra();
      let i, d;
      if (typeof r == "object" && "location" in r) switch (this.dataLocation = r.location, i = r.type, d = r.dims, r.location) {
        case "cpu-pinned": {
          let m = et.get(i);
          if (!m) throw new TypeError(`unsupported type "${i}" to create tensor from pinned buffer`);
          if (!(r.data instanceof m)) throw new TypeError(`buffer should be of type ${m.name}`);
          this.cpuData = r.data;
          break;
        }
        case "texture": {
          if (i !== "float32") throw new TypeError(`unsupported type "${i}" to create tensor from texture`);
          this.gpuTextureData = r.texture, this.downloader = r.download, this.disposer = r.dispose;
          break;
        }
        case "gpu-buffer": {
          if (i !== "float32" && i !== "float16" && i !== "int32" && i !== "int64" && i !== "uint32" && i !== "uint8" && i !== "bool" && i !== "uint4" && i !== "int4") throw new TypeError(`unsupported type "${i}" to create tensor from gpu buffer`);
          this.gpuBufferData = r.gpuBuffer, this.downloader = r.download, this.disposer = r.dispose;
          break;
        }
        case "ml-tensor": {
          if (i !== "float32" && i !== "float16" && i !== "int32" && i !== "int64" && i !== "uint32" && i !== "uint64" && i !== "int8" && i !== "uint8" && i !== "bool" && i !== "uint4" && i !== "int4") throw new TypeError(`unsupported type "${i}" to create tensor from MLTensor`);
          this.mlTensorData = r.mlTensor, this.downloader = r.download, this.disposer = r.dispose;
          break;
        }
        default:
          throw new Error(`Tensor constructor: unsupported location '${this.dataLocation}'`);
      }
      else {
        let m, y;
        if (typeof r == "string") if (i = r, y = f, r === "string") {
          if (!Array.isArray(s)) throw new TypeError("A string tensor's data must be a string array.");
          m = s;
        } else {
          let w = et.get(r);
          if (w === void 0) throw new TypeError(`Unsupported tensor type: ${r}.`);
          if (Array.isArray(s)) {
            if (r === "float16" && w === Uint16Array || r === "uint4" || r === "int4") throw new TypeError(`Creating a ${r} tensor from number array is not supported. Please use ${w.name} as data.`);
            r === "uint64" || r === "int64" ? m = w.from(s, BigInt) : m = w.from(s);
          } else if (s instanceof w) m = s;
          else if (s instanceof Uint8ClampedArray) if (r === "uint8") m = Uint8Array.from(s);
          else throw new TypeError("A Uint8ClampedArray tensor's data must be type of uint8");
          else if (r === "float16" && s instanceof Uint16Array && w !== Uint16Array) m = new globalThis.Float16Array(s.buffer, s.byteOffset, s.length);
          else throw new TypeError(`A ${i} tensor's data must be type of ${w}`);
        }
        else if (y = s, Array.isArray(r)) {
          if (r.length === 0) throw new TypeError("Tensor type cannot be inferred from an empty array.");
          let w = typeof r[0];
          if (w === "string") i = "string", m = r;
          else if (w === "boolean") i = "bool", m = Uint8Array.from(r);
          else throw new TypeError(`Invalid element type of data array: ${w}.`);
        } else if (r instanceof Uint8ClampedArray) i = "uint8", m = Uint8Array.from(r);
        else {
          let w = It.get(r.constructor);
          if (w === void 0) throw new TypeError(`Unsupported type for tensor data: ${r.constructor}.`);
          i = w, m = r;
        }
        if (y === void 0) y = [m.length];
        else if (!Array.isArray(y)) throw new TypeError("A tensor's dims must be a number array");
        d = y, this.cpuData = m, this.dataLocation = "cpu";
      }
      let l = ka(d);
      if (this.cpuData && l !== this.cpuData.length && !((i === "uint4" || i === "int4") && Math.ceil(l / 2) === this.cpuData.length)) throw new Error(`Tensor's size(${l}) does not match data length(${this.cpuData.length}).`);
      this.type = i, this.dims = d, this.size = l;
    }
    static async fromImage(r, s) {
      return Ba(r, s);
    }
    static fromTexture(r, s) {
      return Ma(r, s);
    }
    static fromGpuBuffer(r, s) {
      return Ca(r, s);
    }
    static fromMLTensor(r, s) {
      return Ua(r, s);
    }
    static fromPinnedBuffer(r, s, f) {
      return Da(r, s, f);
    }
    toDataURL(r) {
      return xa(this, r);
    }
    toImageData(r) {
      return La(this, r);
    }
    get data() {
      if (this.ensureValid(), !this.cpuData) throw new Error("The data is not on CPU. Use `getData()` to download GPU data to CPU, or use `texture` or `gpuBuffer` property to access the GPU data directly.");
      return this.cpuData;
    }
    get location() {
      return this.dataLocation;
    }
    get texture() {
      if (this.ensureValid(), !this.gpuTextureData) throw new Error("The data is not stored as a WebGL texture.");
      return this.gpuTextureData;
    }
    get gpuBuffer() {
      if (this.ensureValid(), !this.gpuBufferData) throw new Error("The data is not stored as a WebGPU buffer.");
      return this.gpuBufferData;
    }
    get mlTensor() {
      if (this.ensureValid(), !this.mlTensorData) throw new Error("The data is not stored as a WebNN MLTensor.");
      return this.mlTensorData;
    }
    async getData(r) {
      switch (this.ensureValid(), this.dataLocation) {
        case "cpu":
        case "cpu-pinned":
          return this.data;
        case "texture":
        case "gpu-buffer":
        case "ml-tensor": {
          if (!this.downloader) throw new Error("The current tensor is not created with a specified data downloader.");
          if (this.isDownloading) throw new Error("The current tensor is being downloaded.");
          try {
            this.isDownloading = true;
            let s = await this.downloader();
            return this.downloader = void 0, this.dataLocation = "cpu", this.cpuData = s, r && this.disposer && (this.disposer(), this.disposer = void 0), s;
          } finally {
            this.isDownloading = false;
          }
        }
        default:
          throw new Error(`cannot get data from location: ${this.dataLocation}`);
      }
    }
    dispose() {
      if (this.isDownloading) throw new Error("The current tensor is being downloaded.");
      this.disposer && (this.disposer(), this.disposer = void 0), this.cpuData = void 0, this.gpuTextureData = void 0, this.gpuBufferData = void 0, this.mlTensorData = void 0, this.downloader = void 0, this.isDownloading = void 0, this.dataLocation = "none";
    }
    ensureValid() {
      if (this.dataLocation === "none") throw new Error("The tensor is disposed.");
    }
    reshape(r) {
      if (this.ensureValid(), this.downloader || this.disposer) throw new Error("Cannot reshape a tensor that owns GPU resource.");
      return Wa(this, r);
    }
  };
});
var Le;
var Xr = k(() => {
  "use strict";
  Vt();
  Le = le;
});
var Ga;
var $a;
var tt;
var rt;
var $e;
var ze;
var Qr = k(() => {
  "use strict";
  qr();
  Ga = (a, r) => {
    (typeof ie.trace > "u" ? !ie.wasm.trace : !ie.trace) || console.timeStamp(`${a}::ORT::${r}`);
  }, $a = (a, r) => {
    let s = new Error().stack?.split(/\r\n|\r|\n/g) || [], f = false;
    for (let i = 0; i < s.length; i++) {
      if (f && !s[i].includes("TRACE_FUNC")) {
        let d = `FUNC_${a}::${s[i].trim().split(" ")[1]}`;
        r && (d += `::${r}`), Ga("CPU", d);
        return;
      }
      s[i].includes("TRACE_FUNC") && (f = true);
    }
  }, tt = (a) => {
    (typeof ie.trace > "u" ? !ie.wasm.trace : !ie.trace) || $a("BEGIN", a);
  }, rt = (a) => {
    (typeof ie.trace > "u" ? !ie.wasm.trace : !ie.trace) || $a("END", a);
  }, $e = (a) => {
    (typeof ie.trace > "u" ? !ie.wasm.trace : !ie.trace) || console.time(`ORT::${a}`);
  }, ze = (a) => {
    (typeof ie.trace > "u" ? !ie.wasm.trace : !ie.trace) || console.timeEnd(`ORT::${a}`);
  };
});
var jt;
var za = k(() => {
  "use strict";
  Yr();
  Xr();
  Qr();
  jt = class a {
    constructor(r) {
      this.handler = r;
    }
    async run(r, s, f) {
      tt(), $e("InferenceSession.run");
      let i = {}, d = {};
      if (typeof r != "object" || r === null || r instanceof Le || Array.isArray(r)) throw new TypeError("'feeds' must be an object that use input names as keys and OnnxValue as corresponding values.");
      let l = true;
      if (typeof s == "object") {
        if (s === null) throw new TypeError("Unexpected argument[1]: cannot be null.");
        if (s instanceof Le) throw new TypeError("'fetches' cannot be a Tensor");
        if (Array.isArray(s)) {
          if (s.length === 0) throw new TypeError("'fetches' cannot be an empty array.");
          l = false;
          for (let w of s) {
            if (typeof w != "string") throw new TypeError("'fetches' must be a string array or an object.");
            if (this.outputNames.indexOf(w) === -1) throw new RangeError(`'fetches' contains invalid output name: ${w}.`);
            i[w] = null;
          }
          if (typeof f == "object" && f !== null) d = f;
          else if (typeof f < "u") throw new TypeError("'options' must be an object.");
        } else {
          let w = false, T = Object.getOwnPropertyNames(s);
          for (let g of this.outputNames) if (T.indexOf(g) !== -1) {
            let v = s[g];
            (v === null || v instanceof Le) && (w = true, l = false, i[g] = v);
          }
          if (w) {
            if (typeof f == "object" && f !== null) d = f;
            else if (typeof f < "u") throw new TypeError("'options' must be an object.");
          } else d = s;
        }
      } else if (typeof s < "u") throw new TypeError("Unexpected argument[1]: must be 'fetches' or 'options'.");
      for (let w of this.inputNames) if (typeof r[w] > "u") throw new Error(`input '${w}' is missing in 'feeds'.`);
      if (l) for (let w of this.outputNames) i[w] = null;
      let m = await this.handler.run(r, i, d), y = {};
      for (let w in m) if (Object.hasOwnProperty.call(m, w)) {
        let T = m[w];
        T instanceof Le ? y[w] = T : y[w] = new Le(T.type, T.data, T.dims);
      }
      return ze("InferenceSession.run"), rt(), y;
    }
    async release() {
      return this.handler.dispose();
    }
    static async create(r, s, f, i) {
      tt(), $e("InferenceSession.create");
      let d, l = {};
      if (typeof r == "string") {
        if (d = r, typeof s == "object" && s !== null) l = s;
        else if (typeof s < "u") throw new TypeError("'options' must be an object.");
      } else if (r instanceof Uint8Array) {
        if (d = r, typeof s == "object" && s !== null) l = s;
        else if (typeof s < "u") throw new TypeError("'options' must be an object.");
      } else if (r instanceof ArrayBuffer || typeof SharedArrayBuffer < "u" && r instanceof SharedArrayBuffer) {
        let T = r, g = 0, v = r.byteLength;
        if (typeof s == "object" && s !== null) l = s;
        else if (typeof s == "number") {
          if (g = s, !Number.isSafeInteger(g)) throw new RangeError("'byteOffset' must be an integer.");
          if (g < 0 || g >= T.byteLength) throw new RangeError(`'byteOffset' is out of range [0, ${T.byteLength}).`);
          if (v = r.byteLength - g, typeof f == "number") {
            if (v = f, !Number.isSafeInteger(v)) throw new RangeError("'byteLength' must be an integer.");
            if (v <= 0 || g + v > T.byteLength) throw new RangeError(`'byteLength' is out of range (0, ${T.byteLength - g}].`);
            if (typeof i == "object" && i !== null) l = i;
            else if (typeof i < "u") throw new TypeError("'options' must be an object.");
          } else if (typeof f < "u") throw new TypeError("'byteLength' must be a number.");
        } else if (typeof s < "u") throw new TypeError("'options' must be an object.");
        d = new Uint8Array(T, g, v);
      } else throw new TypeError("Unexpected argument[0]: must be 'path' or 'buffer'.");
      let [m, y] = await Ta(l), w = await m.createInferenceSessionHandler(d, y);
      return ze("InferenceSession.create"), rt(), new a(w);
    }
    startProfiling() {
      this.handler.startProfiling();
    }
    endProfiling() {
      this.handler.endProfiling();
    }
    get inputNames() {
      return this.handler.inputNames;
    }
    get outputNames() {
      return this.handler.outputNames;
    }
    get inputMetadata() {
      return this.handler.inputMetadata;
    }
    get outputMetadata() {
      return this.handler.outputMetadata;
    }
  };
});
var qf;
var Va = k(() => {
  "use strict";
  za();
  qf = jt;
});
var ja = k(() => {
  "use strict";
});
var Ha = k(() => {
  "use strict";
});
var Ya = k(() => {
  "use strict";
});
var qa = k(() => {
  "use strict";
});
var Zr = {};
At(Zr, { InferenceSession: () => qf, TRACE: () => Ga, TRACE_EVENT_BEGIN: () => $e, TRACE_EVENT_END: () => ze, TRACE_FUNC_BEGIN: () => tt, TRACE_FUNC_END: () => rt, Tensor: () => Le, env: () => K, registerBackend: () => Ke });
var Ve = k(() => {
  "use strict";
  va();
  Ia();
  Va();
  Xr();
  ja();
  Ha();
  Qr();
  Ya();
  qa();
});
var Ht = k(() => {
  "use strict";
});
var Za = {};
At(Za, { default: () => Jf });
var Xa;
var Qa;
var Jf;
var Ka = k(() => {
  "use strict";
  Kr();
  je();
  Yt();
  Xa = "ort-wasm-proxy-worker", Qa = globalThis.self?.name === Xa;
  Qa && (self.onmessage = (a) => {
    let { type: r, in: s } = a.data;
    try {
      switch (r) {
        case "init-wasm":
          qt(s.wasm).then(() => {
            Jt(s).then(() => {
              postMessage({ type: r });
            }, (f) => {
              postMessage({ type: r, err: f });
            });
          }, (f) => {
            postMessage({ type: r, err: f });
          });
          break;
        case "init-ep": {
          let { epName: f, env: i } = s;
          Xt(i, f).then(() => {
            postMessage({ type: r });
          }, (d) => {
            postMessage({ type: r, err: d });
          });
          break;
        }
        case "copy-from": {
          let { buffer: f } = s, i = xt(f);
          postMessage({ type: r, out: i });
          break;
        }
        case "create": {
          let { model: f, options: i } = s;
          Qt(f, i).then((d) => {
            postMessage({ type: r, out: d });
          }, (d) => {
            postMessage({ type: r, err: d });
          });
          break;
        }
        case "release":
          Zt(s), postMessage({ type: r });
          break;
        case "run": {
          let { sessionId: f, inputIndices: i, inputs: d, outputIndices: l, options: m } = s;
          Kt(f, i, d, l, new Array(l.length).fill(null), m).then((y) => {
            y.some((w) => w[3] !== "cpu") ? postMessage({ type: r, err: "Proxy does not support non-cpu tensor location." }) : postMessage({ type: r, out: y }, tr([...d, ...y]));
          }, (y) => {
            postMessage({ type: r, err: y });
          });
          break;
        }
        case "end-profiling":
          er(s), postMessage({ type: r });
          break;
        default:
      }
    } catch (f) {
      postMessage({ type: r, err: f });
    }
  });
  Jf = Qa ? null : (a) => new Worker(a ?? ge, { type: "module", name: Xa });
});
var ts = {};
At(ts, { default: () => Xf });
async function es(a = {}) {
  var r = a, s = !!globalThis.window, f = !!globalThis.WorkerGlobalScope, i = f && self.name?.startsWith("em-pthread");
  r.mountExternalData = (e, t) => {
    e.startsWith("./") && (e = e.substring(2)), (r.Uc || (r.Uc = /* @__PURE__ */ new Map())).set(e, t);
  }, r.unmountExternalData = () => {
    delete r.Uc;
  }, globalThis.SharedArrayBuffer ?? new WebAssembly.Memory({ initial: 0, maximum: 0, shared: true }).buffer.constructor;
  let d = () => {
    let e = (t) => (...n) => {
      let o = Me;
      return n = t(...n), Me != o ? new Promise((u, c) => {
        Lr = { resolve: u, reject: c };
      }) : n;
    };
    (() => {
      for (let t of ["_OrtAppendExecutionProvider", "_OrtCreateSession", "_OrtRun", "_OrtRunWithBinding", "_OrtBindInput"]) r[t] = e(r[t]);
    })(), typeof jsepRunAsync < "u" && (r._OrtRun = jsepRunAsync(r._OrtRun), r._OrtRunWithBinding = jsepRunAsync(r._OrtRunWithBinding)), d = void 0;
  };
  r.asyncInit = () => {
    d?.();
  };
  var l, m, y = (e, t) => {
    throw t;
  }, w = import.meta.url, T = "";
  if (s || f) {
    try {
      T = new URL(".", w).href;
    } catch {
    }
    f && (m = (e) => {
      var t = new XMLHttpRequest();
      return t.open("GET", e, false), t.responseType = "arraybuffer", t.send(null), new Uint8Array(t.response);
    }), l = async (e) => {
      if (oe(e)) return new Promise((n, o) => {
        var u = new XMLHttpRequest();
        u.open("GET", e, true), u.responseType = "arraybuffer", u.onload = () => {
          u.status == 200 || u.status == 0 && u.response ? n(u.response) : o(u.status);
        }, u.onerror = o, u.send(null);
      });
      var t = await fetch(e, { credentials: "same-origin" });
      if (t.ok) return t.arrayBuffer();
      throw Error(t.status + " : " + t.url);
    };
  }
  var g, v, S, C, R, H, U = console.log.bind(console), M = console.error.bind(console), Y = U, O = M, W = false, oe = (e) => e.startsWith("file://");
  function p() {
    Fe.buffer != X.buffer && se();
  }
  if (i) {
    let e = function(t) {
      try {
        var n = t.data, o = n.Oc;
        if (o === "load") {
          let u = [];
          self.onmessage = (c) => u.push(c), H = () => {
            postMessage({ Oc: "loaded" });
            for (let c of u) e(c);
            self.onmessage = e;
          };
          for (let c of n.ce) r[c] && !r[c].proxy || (r[c] = (...h) => {
            postMessage({ Oc: "callHandler", be: c, args: h });
          }, c == "print" && (Y = r[c]), c == "printErr" && (O = r[c]));
          Fe = n.ie, se(), v = n.je, bt(), Gt();
        } else if (o === "run") {
          (function(u) {
            var c = (p(), A)[u + 52 >>> 2 >>> 0];
            u = (p(), A)[u + 56 >>> 2 >>> 0], Co(c, c - u), D(c);
          })(n.Nc), Wr(n.Nc, 0, 0, 1, 0, 0), bn(), Ar(n.Nc), ne || (po(), ne = true);
          try {
            $s(n.ge, n.Wc);
          } catch (u) {
            if (u != "unwind") throw u;
          }
        } else n.target !== "setimmediate" && (o === "checkMailbox" ? ne && Dt() : o && (O(`worker: received unknown command ${o}`), O(n)));
      } catch (u) {
        throw xo(), u;
      }
    };
    var vc = e, ne = false;
    self.onunhandledrejection = (t) => {
      throw t.reason || t;
    }, self.onmessage = e;
  }
  var X, J, Ue, Q, x, A, _, ae, me, q, we, re = false;
  function se() {
    var e = Fe.buffer;
    r.HEAP8 = X = new Int8Array(e), Ue = new Int16Array(e), r.HEAPU8 = J = new Uint8Array(e), Q = new Uint16Array(e), r.HEAP32 = x = new Int32Array(e), r.HEAPU32 = A = new Uint32Array(e), _ = new Float32Array(e), ae = new Float64Array(e), me = new BigInt64Array(e), q = new BigUint64Array(e);
  }
  function hr() {
    re = true, i ? H() : ke._b();
  }
  function Te(e) {
    throw O(e = "Aborted(" + e + ")"), W = true, e = new WebAssembly.RuntimeError(e + ". Build with -sASSERTIONS for more info."), R?.(e), e;
  }
  function Ye() {
    return { a: { f: zs, J: Vs, k: js, p: Hs, l: Ys, sa: qs, b: Js, ca: Xs, Ja: Sn, q: Qs, da: Ln, Za: On, Fa: Bn, Ha: Mn, _a: Cn, Xa: Un, Qa: Dn, Wa: Pn, oa: _n, Ga: Rn, Xb: Nn, Ya: kn, Yb: Wn, db: Zs, Da: ei, Sb: ti, Qb: ni, Ca: ai, M: si, I: ii, Rb: ui, ja: hi, Tb: yi, Ta: bi, Vb: gi, Ka: Ti, Ob: vi, ka: Ei, Sa: Ar, ab: Si, U: Li, n: Ui, c: Er, rb: Di, w: Pi, L: _i, z: Ri, j: Ni, o: Yn, sb: ki, G: Wi, T: Fi, h: Gi, u: $i, m: zi, i: Vi, Na: ji, Oa: Hi, Pa: Yi, La: Qn, Ma: Zn, Pb: Kn, eb: Ji, cb: Zi, Y: Ki, qb: eu, la: tu, bb: Xi, fb: ru, $a: nu, Wb: ou, N: qi, gb: au, X: su, Ub: iu, nb: yu, C: bu, ra: wu, qa: gu, pb: Tu, W: vu, v: Eu, mb: Su, lb: Au, kb: Iu, ob: xu, jb: Lu, ib: Ou, hb: Bu, Ua: ao, Va: so, Ia: br, V: io, na: uo, Ra: fo, ma: co, Cb: Ff, xa: Pf, Db: Wf, ya: Df, F: Ef, e: ff, s: sf, x: af, B: gf, Fb: Mf, ba: Bf, D: lf, za: Cf, $: _f, ga: Of, Gb: Lf, Hb: xf, Ba: Sf, Aa: If2, Ib: Af, wa: kf, aa: Uf, d: uf, A: df, r: cf, Bb: Gf, t: mf, y: Tf, H: pf, E: hf, K: vf, R: Rf, ia: wf, _: Nf, Jb: bf, Kb: yf, g: Cu, a: Fe, Nb: qe, Eb: Uu, ha: Du, O: Pu, pa: _u, Lb: Ru, ta: Nu, Q: ku, yb: Wu, zb: Fu, ua: Gu, ea: $u, P: zu, Ea: Vu, va: ju, Z: Hu, wb: Yu, Zb: qu, S: Ju, Ab: Xu, tb: Qu, ub: Ku, vb: ef, fa: tf, xb: rf, Mb: nf } };
  }
  async function bt() {
    function e(o, u) {
      var c = ke = o.exports;
      o = {};
      for (let [h, b] of Object.entries(c)) typeof b == "function" ? (c = Ai(b), o[h] = c) : o[h] = b;
      return ke = o, ke = (function() {
        var h = ke, b = (I) => (F) => I(F) >>> 0, E = (I) => () => I() >>> 0;
        return (h = Object.assign({}, h)).$b = b(h.$b), h.Cc = E(h.Cc), h.Ec = b(h.Ec), h.rd = /* @__PURE__ */ ((I) => (F, j) => I(F, j) >>> 0)(h.rd), h.wd = b(h.wd), h.xd = E(h.xd), h.Bd = b(h.Bd), h;
      })(), hn.push(ke.id), lo = (o = ke).$b, po = o.ac, r._OrtInit = o.bc, r._OrtGetLastError = o.cc, r._OrtCreateSessionOptions = o.dc, r._OrtAppendExecutionProvider = o.ec, r._OrtAddFreeDimensionOverride = o.fc, r._OrtAddSessionConfigEntry = o.gc, r._OrtReleaseSessionOptions = o.hc, r._OrtCreateSession = o.ic, r._OrtReleaseSession = o.jc, r._OrtGetInputOutputCount = o.kc, r._OrtGetInputOutputMetadata = o.lc, r._OrtFree = o.mc, r._OrtCreateTensor = o.nc, r._OrtGetTensorData = o.oc, r._OrtReleaseTensor = o.pc, r._OrtCreateRunOptions = o.qc, r._OrtAddRunConfigEntry = o.rc, r._OrtReleaseRunOptions = o.sc, r._OrtCreateBinding = o.tc, r._OrtBindInput = o.uc, r._OrtBindOutput = o.vc, r._OrtClearBoundOutputs = o.wc, r._OrtReleaseBinding = o.xc, r._OrtRunWithBinding = o.yc, r._OrtRun = o.zc, r._OrtEndProfiling = o.Ac, Dr = r._OrtGetWebGpuDevice = o.Bc, Wt = o.Cc, xe = r._free = o.Dc, pt = r._malloc = o.Ec, mo = r._wgpuBufferRelease = o.Fc, ho = r._wgpuCreateInstance = o.Gc, yo = o.Hc, bo = o.Ic, wo = o.Jc, go = o.Kc, To = o.Lc, vo = o.Pc, Eo = o.Zc, So = o._c, Ao = o.$c, Pr = o.bd, _r = o.cd, Rr = o.dd, Nr = o.ed, Et = o.fd, kr = o.gd, Io = o.hd, Wr = o.kd, xo = o.ld, Lo = o.md, Oo = o.nd, Fr = o.od, Bo = o.pd, Mo = o.qd, Gr = o.rd, N = o.sd, St = o.td, Co = o.ud, D = o.vd, Ft = o.wd, P = o.xd, Uo = o.yd, $r = o.zd, Do = o.Ad, Po = o.Bd, _o = o.Cd, zr = o.Dd, Ro = o.Ed, No = o.Fd, ko = o.Gd, Wo = o.Hd, Fo = o.Id, Go = o.Jd, $o = o.Kd, zo = o.Ld, Vo = o.Md, jo = o.Nd, Ho = o.Od, Yo = o.Pd, qo = o.Qd, Jo = o.Rd, Xo = o.Td, Qo = o.Ud, Zo = o.Vd, Ko = o.Wd, ea = o.Yd, ta = o.Zd, ra = o._d, na = o.$d, oa = o.ae, aa = o.oe, sa = o.pe, ia = o.qe, ua = o.re, fa = o.se, ca = o.te, da = o.ue, la = o.ve, pa = o.we, ma = o.xe, ha = o.ye, ya = o.Ye, ba = o.Ze, wa = o._e, ga = o.$e, v = u, ke;
    }
    var t, n = Ye();
    return r.instantiateWasm ? new Promise((o) => {
      r.instantiateWasm(n, (u, c) => {
        o(e(u, c));
      });
    }) : i ? e(new WebAssembly.Instance(v, Ye()), v) : (we ??= r.locateFile ? r.locateFile ? r.locateFile("ort-wasm-simd-threaded.asyncify.wasm", T) : T + "ort-wasm-simd-threaded.asyncify.wasm" : new URL("ort-wasm-simd-threaded.asyncify.wasm", import.meta.url).href, t = await (async function(o) {
      var u = we;
      if (!g && !oe(u)) try {
        var c = fetch(u, { credentials: "same-origin" });
        return await WebAssembly.instantiateStreaming(c, o);
      } catch (h) {
        O(`wasm streaming compile failed: ${h}`), O("falling back to ArrayBuffer instantiation");
      }
      return (async function(h, b) {
        try {
          var E = await (async function(I) {
            if (!g) try {
              var F = await l(I);
              return new Uint8Array(F);
            } catch {
            }
            if (I == we && g) I = new Uint8Array(g);
            else {
              if (!m) throw "both async and sync fetching of the wasm failed";
              I = m(I);
            }
            return I;
          })(h);
          return await WebAssembly.instantiate(E, b);
        } catch (I) {
          O(`failed to asynchronously prepare wasm: ${I}`), Te(I);
        }
      })(u, o);
    })(n), e(t.instance, t.module));
  }
  class wt {
    name = "ExitStatus";
    constructor(t) {
      this.message = `Program terminated with exit(${t})`, this.status = t;
    }
  }
  var Se = (e) => {
    e.terminate(), e.onmessage = () => {
    };
  }, Ae = [], Oe = 0, ee = null, Z = (e) => {
    We.length == 0 && (gn(), wn(We[0]));
    var t = We.pop();
    if (!t) return 6;
    gt.push(t), Je[e.Nc] = t, t.Nc = e.Nc;
    var n = { Oc: "run", ge: e.fe, Wc: e.Wc, Nc: e.Nc };
    return t.postMessage(n, e.Yc), 0;
  }, G = 0, V = (e, t, ...n) => {
    var o, u = 16 * n.length, c = P(), h = Ft(u), b = h >>> 3;
    for (o of n) typeof o == "bigint" ? ((p(), me)[b++ >>> 0] = 1n, (p(), me)[b++ >>> 0] = o) : ((p(), me)[b++ >>> 0] = 0n, (p(), ae)[b++ >>> 0] = o);
    return e = Lo(e, 0, u, h, t), D(c), e;
  };
  function qe(e) {
    if (i) return V(0, 1, e);
    if (S = e, !(0 < G)) {
      for (var t of gt) Se(t);
      for (t of We) Se(t);
      We = [], gt = [], Je = {}, W = true;
    }
    y(0, new wt(e));
  }
  function yr(e) {
    if (i) return V(1, 0, e);
    br(e);
  }
  var br = (e) => {
    if (S = e, i) throw yr(e), "unwind";
    qe(e);
  }, We = [], gt = [], hn = [], Je = {}, yn = (e) => {
    var t = e.Nc;
    delete Je[t], We.push(e), gt.splice(gt.indexOf(e), 1), e.Nc = 0, Oo(t);
  };
  function bn() {
    hn.forEach((e) => e());
  }
  var wn = (e) => new Promise((t) => {
    e.onmessage = (u) => {
      var c = u.data;
      if (u = c.Oc, c.Vc && c.Vc != Wt()) {
        var h = Je[c.Vc];
        h ? h.postMessage(c, c.Yc) : O(`Internal error! Worker sent a message "${u}" to target pthread ${c.Vc}, but that thread no longer exists!`);
      } else u === "checkMailbox" ? Dt() : u === "spawnThread" ? Z(c) : u === "cleanupThread" ? he(() => {
        yn(Je[c.he]);
      }) : u === "loaded" ? (e.loaded = true, t(e)) : c.target === "setimmediate" ? e.postMessage(c) : u === "uncaughtException" ? e.onerror(c.error) : u === "callHandler" ? r[c.be](...c.args) : u && O(`worker sent an unknown command ${u}`);
    }, e.onerror = (u) => {
      throw O(`worker sent an error! ${u.filename}:${u.lineno}: ${u.message}`), u;
    };
    var n, o = [];
    for (n of []) r.propertyIsEnumerable(n) && o.push(n);
    e.postMessage({ Oc: "load", ce: o, ie: Fe, je: v });
  });
  function gn() {
    var e = new Worker((() => {
      let t = URL;
      return import.meta.url > "file:" && import.meta.url < "file;" ? new t("ort.webgpu.bundle.min.mjs", import.meta.url) : new URL(import.meta.url);
    })(), { type: "module", workerData: "em-pthread", name: "em-pthread" });
    We.push(e);
  }
  var Fe, $s = (e, t) => {
    G = 0, e = zr(e, t), 0 < G ? S = e : Fr(e);
  }, Ct = [], Ut = 0, ce = (e) => -9007199254740992 > e || 9007199254740992 < e ? NaN : Number(e);
  function zs(e) {
    var t = new wr(e >>>= 0);
    return (p(), X)[t.Qc + 12 >>> 0] == 0 && (Tn(t, true), Ut--), vn(t, false), Ct.push(t), Po(e);
  }
  var ft = 0, Vs = () => {
    N(0, 0);
    var e = Ct.pop();
    Uo(e.Xc), ft = 0;
  };
  function Tn(e, t) {
    t = t ? 1 : 0, (p(), X)[e.Qc + 12 >>> 0] = t;
  }
  function vn(e, t) {
    t = t ? 1 : 0, (p(), X)[e.Qc + 13 >>> 0] = t;
  }
  class wr {
    constructor(t) {
      this.Xc = t, this.Qc = t - 24;
    }
  }
  var gr = (e) => {
    var t = ft;
    if (!t) return St(0), 0;
    var n = new wr(t);
    (p(), A)[n.Qc + 16 >>> 2 >>> 0] = t;
    var o = (p(), A)[n.Qc + 4 >>> 2 >>> 0];
    if (!o) return St(0), t;
    for (var u of e) {
      if (u === 0 || u === o) break;
      if (Do(u, o, n.Qc + 16)) return St(u), t;
    }
    return St(o), t;
  };
  function js() {
    return gr([]);
  }
  function Hs(e) {
    return gr([e >>> 0]);
  }
  function Ys(e, t, n, o) {
    return gr([e >>> 0, t >>> 0, n >>> 0, o >>> 0]);
  }
  var qs = () => {
    var e = Ct.pop();
    e || Te("no exception to throw");
    var t = e.Xc;
    throw (p(), X)[e.Qc + 13 >>> 0] == 0 && (Ct.push(e), vn(e, true), Tn(e, false), Ut++), $r(t), ft = t;
  };
  function Js(e, t, n) {
    var o = new wr(e >>>= 0);
    throw t >>>= 0, n >>>= 0, (p(), A)[o.Qc + 16 >>> 2 >>> 0] = 0, (p(), A)[o.Qc + 4 >>> 2 >>> 0] = t, (p(), A)[o.Qc + 8 >>> 2 >>> 0] = n, $r(e), Ut++, ft = e;
  }
  var Xs = () => Ut;
  function En(e, t, n, o) {
    return i ? V(2, 1, e, t, n, o) : Sn(e, t, n, o);
  }
  function Sn(e, t, n, o) {
    if (e >>>= 0, t >>>= 0, n >>>= 0, o >>>= 0, !globalThis.SharedArrayBuffer) return 6;
    var u = [];
    return i && u.length === 0 ? En(e, t, n, o) : (e = { fe: n, Nc: e, Wc: o, Yc: u }, i ? (e.Oc = "spawnThread", postMessage(e, u), 0) : Z(e));
  }
  function Qs(e) {
    throw ft ||= e >>> 0, ft;
  }
  var An = globalThis.TextDecoder && new TextDecoder(), In = (e, t, n, o) => {
    if (n = t + n, o) return n;
    for (; e[t] && !(t >= n); ) ++t;
    return t;
  }, xn = (e, t = 0, n, o) => {
    if (16 < (n = In(e, t >>>= 0, n, o)) - t && e.buffer && An) return An.decode(e.buffer instanceof ArrayBuffer ? e.subarray(t, n) : e.slice(t, n));
    for (o = ""; t < n; ) {
      var u = e[t++];
      if (128 & u) {
        var c = 63 & e[t++];
        if ((224 & u) == 192) o += String.fromCharCode((31 & u) << 6 | c);
        else {
          var h = 63 & e[t++];
          65536 > (u = (240 & u) == 224 ? (15 & u) << 12 | c << 6 | h : (7 & u) << 18 | c << 12 | h << 6 | 63 & e[t++]) ? o += String.fromCharCode(u) : (u -= 65536, o += String.fromCharCode(55296 | u >> 10, 56320 | 1023 & u));
        }
      } else o += String.fromCharCode(u);
    }
    return o;
  }, ct = (e, t, n) => (e >>>= 0) ? xn((p(), J), e, t, n) : "";
  function Ln(e, t, n) {
    return i ? V(3, 1, e, t, n) : 0;
  }
  function On(e, t) {
    if (i) return V(4, 1, e, t);
  }
  function Bn(e, t) {
    if (i) return V(5, 1, e, t);
  }
  function Mn(e, t, n) {
    if (i) return V(6, 1, e, t, n);
  }
  function Cn(e, t, n) {
    return i ? V(7, 1, e, t, n) : 0;
  }
  function Un(e, t) {
    if (i) return V(8, 1, e, t);
  }
  function Dn(e, t, n) {
    if (i) return V(9, 1, e, t, n);
  }
  function Pn(e, t, n, o) {
    if (i) return V(10, 1, e, t, n, o);
  }
  function _n(e, t, n, o) {
    if (i) return V(11, 1, e, t, n, o);
  }
  function Rn(e, t, n, o) {
    if (i) return V(12, 1, e, t, n, o);
  }
  function Nn(e) {
    if (i) return V(13, 1, e);
  }
  function kn(e, t) {
    if (i) return V(14, 1, e, t);
  }
  function Wn(e, t, n) {
    if (i) return V(15, 1, e, t, n);
  }
  var Zs = () => Te(""), Be = (e) => {
    e >>>= 0;
    for (var t = ""; ; ) {
      var n = (p(), J)[e++ >>> 0];
      if (!n) return t;
      t += String.fromCharCode(n);
    }
  }, Tr = {}, vr = {}, Ks = {}, dt = class extends Error {
    constructor(e) {
      super(e), this.name = "BindingError";
    }
  };
  function De(e, t, n = {}) {
    return (function(o, u, c = {}) {
      var h = u.name;
      if (!o) throw new dt(`type "${h}" must have a positive integer typeid pointer`);
      if (vr.hasOwnProperty(o)) {
        if (c.de) return;
        throw new dt(`Cannot register type '${h}' twice`);
      }
      vr[o] = u, delete Ks[o], Tr.hasOwnProperty(o) && (u = Tr[o], delete Tr[o], u.forEach((b) => b()));
    })(e, t, n);
  }
  var Fn = (e, t, n) => {
    switch (t) {
      case 1:
        return n ? (o) => (p(), X)[o >>> 0] : (o) => (p(), J)[o >>> 0];
      case 2:
        return n ? (o) => (p(), Ue)[o >>> 1 >>> 0] : (o) => (p(), Q)[o >>> 1 >>> 0];
      case 4:
        return n ? (o) => (p(), x)[o >>> 2 >>> 0] : (o) => (p(), A)[o >>> 2 >>> 0];
      case 8:
        return n ? (o) => (p(), me)[o >>> 3 >>> 0] : (o) => (p(), q)[o >>> 3 >>> 0];
      default:
        throw new TypeError(`invalid integer width (${t}): ${e}`);
    }
  };
  function ei(e, t, n, o, u) {
    e >>>= 0, n >>>= 0, t = Be(t >>> 0);
    let c = (h) => h;
    if (o = o === 0n) {
      let h = 8 * n;
      c = (b) => BigInt.asUintN(h, b), u = c(u);
    }
    De(e, { name: t, Mc: c, Sc: (h, b) => (typeof b == "number" && (b = BigInt(b)), b), Rc: Fn(t, n, !o), Tc: null });
  }
  function ti(e, t, n, o) {
    De(e >>>= 0, { name: t = Be(t >>> 0), Mc: function(u) {
      return !!u;
    }, Sc: function(u, c) {
      return c ? n : o;
    }, Rc: function(u) {
      return this.Mc((p(), J)[u >>> 0]);
    }, Tc: null });
  }
  var Gn = [], Xe = [0, 1, , 1, null, 1, true, 1, false, 1];
  function Er(e) {
    9 < (e >>>= 0) && --Xe[e + 1] == 0 && (Xe[e] = void 0, Gn.push(e));
  }
  var ve = (e) => {
    if (!e) throw new dt(`Cannot use deleted val. handle = ${e}`);
    return Xe[e];
  }, Ie = (e) => {
    switch (e) {
      case void 0:
        return 2;
      case null:
        return 4;
      case true:
        return 6;
      case false:
        return 8;
      default:
        let t = Gn.pop() || Xe.length;
        return Xe[t] = e, Xe[t + 1] = 1, t;
    }
  };
  function Sr(e) {
    return this.Mc((p(), A)[e >>> 2 >>> 0]);
  }
  var ri = { name: "emscripten::val", Mc: (e) => {
    var t = ve(e);
    return Er(e), t;
  }, Sc: (e, t) => Ie(t), Rc: Sr, Tc: null };
  function ni(e) {
    return De(e >>> 0, ri);
  }
  var oi = (e, t) => {
    switch (t) {
      case 4:
        return function(n) {
          return this.Mc((p(), _)[n >>> 2 >>> 0]);
        };
      case 8:
        return function(n) {
          return this.Mc((p(), ae)[n >>> 3 >>> 0]);
        };
      default:
        throw new TypeError(`invalid float width (${t}): ${e}`);
    }
  };
  function ai(e, t, n) {
    n >>>= 0, De(e >>>= 0, { name: t = Be(t >>> 0), Mc: (o) => o, Sc: (o, u) => u, Rc: oi(t, n), Tc: null });
  }
  function si(e, t, n, o, u) {
    e >>>= 0, n >>>= 0, t = Be(t >>> 0);
    let c = (b) => b;
    if (o === 0) {
      var h = 32 - 8 * n;
      c = (b) => b << h >>> h, u = c(u);
    }
    De(e, { name: t, Mc: c, Sc: (b, E) => E, Rc: Fn(t, n, o !== 0), Tc: null });
  }
  function ii(e, t, n) {
    function o(c) {
      var h = (p(), A)[c >>> 2 >>> 0];
      return c = (p(), A)[c + 4 >>> 2 >>> 0], new u((p(), X).buffer, c, h);
    }
    var u = [Int8Array, Uint8Array, Int16Array, Uint16Array, Int32Array, Uint32Array, Float32Array, Float64Array, BigInt64Array, BigUint64Array][t];
    De(e >>>= 0, { name: n = Be(n >>> 0), Mc: o, Rc: o }, { de: true });
  }
  var Pe = (e, t, n) => {
    var o = (p(), J);
    if (t >>>= 0, 0 < n) {
      var u = t;
      n = t + n - 1;
      for (var c = 0; c < e.length; ++c) {
        var h = e.codePointAt(c);
        if (127 >= h) {
          if (t >= n) break;
          o[t++ >>> 0] = h;
        } else if (2047 >= h) {
          if (t + 1 >= n) break;
          o[t++ >>> 0] = 192 | h >> 6, o[t++ >>> 0] = 128 | 63 & h;
        } else if (65535 >= h) {
          if (t + 2 >= n) break;
          o[t++ >>> 0] = 224 | h >> 12, o[t++ >>> 0] = 128 | h >> 6 & 63, o[t++ >>> 0] = 128 | 63 & h;
        } else {
          if (t + 3 >= n) break;
          o[t++ >>> 0] = 240 | h >> 18, o[t++ >>> 0] = 128 | h >> 12 & 63, o[t++ >>> 0] = 128 | h >> 6 & 63, o[t++ >>> 0] = 128 | 63 & h, c++;
        }
      }
      o[t >>> 0] = 0, e = t - u;
    } else e = 0;
    return e;
  }, _e = (e) => {
    for (var t = 0, n = 0; n < e.length; ++n) {
      var o = e.charCodeAt(n);
      127 >= o ? t++ : 2047 >= o ? t += 2 : 55296 <= o && 57343 >= o ? (t += 4, ++n) : t += 3;
    }
    return t;
  };
  function ui(e, t) {
    De(e >>>= 0, { name: t = Be(t >>> 0), Mc(n) {
      var o = (p(), A)[n >>> 2 >>> 0];
      return o = ct(n + 4, o, true), xe(n), o;
    }, Sc(n, o) {
      o instanceof ArrayBuffer && (o = new Uint8Array(o));
      var u = typeof o == "string";
      if (!(u || ArrayBuffer.isView(o) && o.BYTES_PER_ELEMENT == 1)) throw new dt("Cannot pass non-string to std::string");
      var c = u ? _e(o) : o.length, h = pt(4 + c + 1), b = h + 4;
      return (p(), A)[h >>> 2 >>> 0] = c, u ? Pe(o, b, c + 1) : (p(), J).set(o, b >>> 0), n !== null && n.push(xe, h), h;
    }, Rc: Sr, Tc(n) {
      xe(n);
    } });
  }
  var $n = globalThis.TextDecoder ? new TextDecoder("utf-16le") : void 0, fi = (e, t, n) => {
    if (e >>>= 1, 16 < (t = In((p(), Q), e, t / 2, n)) - e && $n) return $n.decode((p(), Q).slice(e, t));
    for (n = ""; e < t; ++e) {
      var o = (p(), Q)[e >>> 0];
      n += String.fromCharCode(o);
    }
    return n;
  }, ci = (e, t, n) => {
    if (n ??= 2147483647, 2 > n) return 0;
    var o = t;
    n = (n -= 2) < 2 * e.length ? n / 2 : e.length;
    for (var u = 0; u < n; ++u) {
      var c = e.charCodeAt(u);
      (p(), Ue)[t >>> 1 >>> 0] = c, t += 2;
    }
    return (p(), Ue)[t >>> 1 >>> 0] = 0, t - o;
  }, di = (e) => 2 * e.length, li = (e, t, n) => {
    var o = "";
    e >>>= 2;
    for (var u = 0; !(u >= t / 4); u++) {
      var c = (p(), A)[e + u >>> 0];
      if (!c && !n) break;
      o += String.fromCodePoint(c);
    }
    return o;
  }, pi = (e, t, n) => {
    if (t >>>= 0, n ??= 2147483647, 4 > n) return 0;
    var o = t;
    n = o + n - 4;
    for (var u = 0; u < e.length; ++u) {
      var c = e.codePointAt(u);
      if (65535 < c && u++, (p(), x)[t >>> 2 >>> 0] = c, (t += 4) + 4 > n) break;
    }
    return (p(), x)[t >>> 2 >>> 0] = 0, t - o;
  }, mi = (e) => {
    for (var t = 0, n = 0; n < e.length; ++n) 65535 < e.codePointAt(n) && n++, t += 4;
    return t;
  };
  function hi(e, t, n) {
    if (e >>>= 0, t >>>= 0, n = Be(n >>>= 0), t === 2) var o = fi, u = ci, c = di;
    else o = li, u = pi, c = mi;
    De(e, { name: n, Mc: (h) => {
      var b = (p(), A)[h >>> 2 >>> 0];
      return b = o(h + 4, b * t, true), xe(h), b;
    }, Sc: (h, b) => {
      if (typeof b != "string") throw new dt(`Cannot pass non-string to C++ string type ${n}`);
      var E = c(b), I = pt(4 + E + t);
      return (p(), A)[I >>> 2 >>> 0] = E / t, u(b, I + 4, E + t), h !== null && h.push(xe, I), I;
    }, Rc: Sr, Tc(h) {
      xe(h);
    } });
  }
  function yi(e, t) {
    De(e >>>= 0, { ee: true, name: t = Be(t >>> 0), Mc: () => {
    }, Sc: () => {
    } });
  }
  function bi(e) {
    Wr(e >>> 0, !f, 1, !s, 131072, false), bn();
  }
  var he = (e) => {
    if (!W) try {
      if (e(), !(0 < G)) try {
        i ? Wt() && Fr(S) : br(S);
      } catch (t) {
        t instanceof wt || t == "unwind" || y(0, t);
      }
    } catch (t) {
      t instanceof wt || t == "unwind" || y(0, t);
    }
  }, wi = !Atomics.waitAsync || globalThis.navigator?.userAgent && 91 > Number((navigator.userAgent.match(/Chrom(e|ium)\/([0-9]+)\./) || [])[2]);
  function Ar(e) {
    e >>>= 0, wi || (Atomics.waitAsync((p(), x), e >>> 2, e).value.then(Dt), e += 128, Atomics.store((p(), x), e >>> 2, 1));
  }
  var Dt = () => he(() => {
    var e = Wt();
    e && (Ar(e), Mo());
  });
  function gi(e, t) {
    (e >>>= 0) == t >>> 0 ? setTimeout(Dt) : i ? postMessage({ Vc: e, Oc: "checkMailbox" }) : (e = Je[e]) && e.postMessage({ Oc: "checkMailbox" });
  }
  var Ir = [];
  function Ti(e, t, n, o, u) {
    for (t >>>= 0, u >>>= 0, Ir.length = 0, n = u >>> 3, o = u + o >>> 3; n < o; ) {
      var c;
      c = (p(), me)[n++ >>> 0] ? (p(), me)[n++ >>> 0] : (p(), ae)[n++ >>> 0], Ir.push(c);
    }
    return (t ? Vr[t] : of[e])(...Ir);
  }
  var vi = () => {
    G = 0;
  };
  function Ei(e) {
    e >>>= 0, i ? postMessage({ Oc: "cleanupThread", he: e }) : yn(Je[e]);
  }
  function Si(e) {
  }
  var Pt = (e) => {
    try {
      e();
    } catch (t) {
      Te(t);
    }
  };
  function Ai(e) {
    var t = (...n) => {
      _t.push(e);
      try {
        return e(...n);
      } finally {
        W || (_t.pop(), Me && Ge === 1 && _t.length === 0 && (Ge = 0, G += 1, Pt(ba), typeof Fibers < "u" && Fibers.Be()));
      }
    };
    return jn.set(e, t), t;
  }
  var Ge = 0, Me = null, zn = 0, _t = [], xr = /* @__PURE__ */ new Map(), Vn = /* @__PURE__ */ new Map(), jn = /* @__PURE__ */ new Map(), Ii = 0, Lr = null, xi = [], Hn = (e) => (function(t) {
    if (!W) {
      if (Ge === 0) {
        var n = false, o = false;
        t((u = 0) => {
          if (!W && (zn = u, n = true, o)) {
            Ge = 2, Pt(() => wa(Me)), typeof MainLoop < "u" && MainLoop.Xd && MainLoop.resume(), u = false;
            try {
              var c = (function() {
                var E = (p(), x)[Me + 8 >>> 2 >>> 0];
                return E = Vn.get(E), E = jn.get(E), --G, E();
              })();
            } catch (E) {
              c = E, u = true;
            }
            var h = false;
            if (!Me) {
              var b = Lr;
              b && (Lr = null, (u ? b.reject : b.resolve)(c), h = true);
            }
            if (u && !h) throw c;
          }
        }), o = true, n || (Ge = 1, Me = (function() {
          var u = pt(65548), c = u + 12;
          if ((p(), A)[u >>> 2 >>> 0] = c, (p(), A)[u + 4 >>> 2 >>> 0] = c + 65536, c = _t[0], !xr.has(c)) {
            var h = Ii++;
            xr.set(c, h), Vn.set(h, c);
          }
          return c = xr.get(c), (p(), x)[u + 8 >>> 2 >>> 0] = c, u;
        })(), typeof MainLoop < "u" && MainLoop.Xd && MainLoop.pause(), Pt(() => ya(Me)));
      } else Ge === 2 ? (Ge = 0, Pt(ga), xe(Me), Me = null, xi.forEach(he)) : Te(`invalid state: ${Ge}`);
      return zn;
    }
  })((t) => {
    e().then(t);
  });
  function Li(e) {
    return e >>>= 0, Hn(async () => {
      var t = await ve(e);
      return Ie(t);
    });
  }
  var Or = [], Oi = (e) => {
    var t = Or.length;
    return Or.push(e), t;
  }, Bi = (e, t) => {
    for (var n = Array(e), o = 0; o < e; ++o) {
      var u = o, c = (p(), A)[t + 4 * o >>> 2 >>> 0], h = vr[c];
      if (h === void 0) throw e = `parameter ${o}`, c = lo(c), t = Be(c), xe(c), new dt(`${e} has unknown type ${t}`);
      n[u] = h;
    }
    return n;
  }, Mi = (e, t, n) => {
    var o = [];
    return e = e(o, n), o.length && ((p(), A)[t >>> 2 >>> 0] = Ie(o)), e;
  }, Ci = {}, Rt = (e) => {
    var t = Ci[e];
    return t === void 0 ? Be(e) : t;
  };
  function Ui(e, t, n) {
    var [o, ...u] = Bi(e, t >>> 0);
    t = o.Sc.bind(o);
    var c = u.map((E) => E.Rc.bind(E));
    e--;
    var h = { toValue: ve };
    switch (e = c.map((E, I) => {
      var F = `argFromPtr${I}`;
      return h[F] = E, `${F}(args${I ? "+" + 8 * I : ""})`;
    }), n) {
      case 0:
        var b = "toValue(handle)";
        break;
      case 2:
        b = "new (toValue(handle))";
        break;
      case 3:
        b = "";
        break;
      case 1:
        h.getStringOrSymbol = Rt, b = "toValue(handle)[getStringOrSymbol(methodName)]";
    }
    return b += `(${e})`, o.ee || (h.toReturnWire = t, h.emval_returnValue = Mi, b = `return emval_returnValue(toReturnWire, destructorsRef, ${b})`), b = `return function (handle, methodName, destructorsRef, args) {
  ${b}
  }`, n = new Function(Object.keys(h), b)(...Object.values(h)), b = `methodCaller<(${u.map((E) => E.name)}) => ${o.name}>`, Oi(Object.defineProperty(n, "name", { value: b }));
  }
  function Di(e, t) {
    return t >>>= 0, (e = ve(e >>> 0)) == ve(t);
  }
  function Pi(e) {
    return (e >>>= 0) ? (e = Rt(e), Ie(globalThis[e])) : Ie(globalThis);
  }
  function _i(e) {
    return e = Rt(e >>> 0), Ie(r[e]);
  }
  function Ri(e, t) {
    return t >>>= 0, e = ve(e >>> 0), t = ve(t), Ie(e[t]);
  }
  function Ni(e) {
    9 < (e >>>= 0) && (Xe[e + 1] += 1);
  }
  function Yn(e, t, n, o, u) {
    return Or[e >>> 0](t >>> 0, n >>> 0, o >>> 0, u >>> 0);
  }
  function ki(e, t, n, o, u) {
    return Yn(e >>> 0, t >>> 0, n >>> 0, o >>> 0, u >>> 0);
  }
  function Wi() {
    return Ie([]);
  }
  function Fi(e) {
    e = ve(e >>> 0);
    for (var t = Array(e.length), n = 0; n < e.length; n++) t[n] = e[n];
    return Ie(t);
  }
  function Gi(e) {
    return Ie(Rt(e >>> 0));
  }
  function $i() {
    return Ie({});
  }
  function zi(e) {
    for (var t = ve(e >>>= 0); t.length; ) {
      var n = t.pop();
      t.pop()(n);
    }
    Er(e);
  }
  function Vi(e, t, n) {
    t >>>= 0, n >>>= 0, e = ve(e >>> 0), t = ve(t), n = ve(n), e[t] = n;
  }
  function ji(e, t) {
    e = ce(e), t >>>= 0, e = new Date(1e3 * e), (p(), x)[t >>> 2 >>> 0] = e.getUTCSeconds(), (p(), x)[t + 4 >>> 2 >>> 0] = e.getUTCMinutes(), (p(), x)[t + 8 >>> 2 >>> 0] = e.getUTCHours(), (p(), x)[t + 12 >>> 2 >>> 0] = e.getUTCDate(), (p(), x)[t + 16 >>> 2 >>> 0] = e.getUTCMonth(), (p(), x)[t + 20 >>> 2 >>> 0] = e.getUTCFullYear() - 1900, (p(), x)[t + 24 >>> 2 >>> 0] = e.getUTCDay(), e = (e.getTime() - Date.UTC(e.getUTCFullYear(), 0, 1, 0, 0, 0, 0)) / 864e5 | 0, (p(), x)[t + 28 >>> 2 >>> 0] = e;
  }
  var qn = (e) => e % 4 == 0 && (e % 100 != 0 || e % 400 == 0), Jn = [0, 31, 60, 91, 121, 152, 182, 213, 244, 274, 305, 335], Xn = [0, 31, 59, 90, 120, 151, 181, 212, 243, 273, 304, 334];
  function Hi(e, t) {
    e = ce(e), t >>>= 0, e = new Date(1e3 * e), (p(), x)[t >>> 2 >>> 0] = e.getSeconds(), (p(), x)[t + 4 >>> 2 >>> 0] = e.getMinutes(), (p(), x)[t + 8 >>> 2 >>> 0] = e.getHours(), (p(), x)[t + 12 >>> 2 >>> 0] = e.getDate(), (p(), x)[t + 16 >>> 2 >>> 0] = e.getMonth(), (p(), x)[t + 20 >>> 2 >>> 0] = e.getFullYear() - 1900, (p(), x)[t + 24 >>> 2 >>> 0] = e.getDay();
    var n = (qn(e.getFullYear()) ? Jn : Xn)[e.getMonth()] + e.getDate() - 1 | 0;
    (p(), x)[t + 28 >>> 2 >>> 0] = n, (p(), x)[t + 36 >>> 2 >>> 0] = -60 * e.getTimezoneOffset(), n = new Date(e.getFullYear(), 6, 1).getTimezoneOffset();
    var o = new Date(e.getFullYear(), 0, 1).getTimezoneOffset();
    e = 0 | (n != o && e.getTimezoneOffset() == Math.min(o, n)), (p(), x)[t + 32 >>> 2 >>> 0] = e;
  }
  function Yi(e) {
    e >>>= 0;
    var t = new Date((p(), x)[e + 20 >>> 2 >>> 0] + 1900, (p(), x)[e + 16 >>> 2 >>> 0], (p(), x)[e + 12 >>> 2 >>> 0], (p(), x)[e + 8 >>> 2 >>> 0], (p(), x)[e + 4 >>> 2 >>> 0], (p(), x)[e >>> 2 >>> 0], 0), n = (p(), x)[e + 32 >>> 2 >>> 0], o = t.getTimezoneOffset(), u = new Date(t.getFullYear(), 6, 1).getTimezoneOffset(), c = new Date(t.getFullYear(), 0, 1).getTimezoneOffset(), h = Math.min(c, u);
    return 0 > n ? (p(), x)[e + 32 >>> 2 >>> 0] = +(u != c && h == o) : 0 < n != (h == o) && (u = Math.max(c, u), t.setTime(t.getTime() + 6e4 * ((0 < n ? h : u) - o))), (p(), x)[e + 24 >>> 2 >>> 0] = t.getDay(), n = (qn(t.getFullYear()) ? Jn : Xn)[t.getMonth()] + t.getDate() - 1 | 0, (p(), x)[e + 28 >>> 2 >>> 0] = n, (p(), x)[e >>> 2 >>> 0] = t.getSeconds(), (p(), x)[e + 4 >>> 2 >>> 0] = t.getMinutes(), (p(), x)[e + 8 >>> 2 >>> 0] = t.getHours(), (p(), x)[e + 12 >>> 2 >>> 0] = t.getDate(), (p(), x)[e + 16 >>> 2 >>> 0] = t.getMonth(), (p(), x)[e + 20 >>> 2 >>> 0] = t.getYear(), e = t.getTime(), BigInt(isNaN(e) ? -1 : e / 1e3);
  }
  function Qn(e, t, n, o, u, c, h) {
    return i ? V(16, 1, e, t, n, o, u, c, h) : -52;
  }
  function Zn(e, t, n, o, u, c) {
    if (i) return V(17, 1, e, t, n, o, u, c);
  }
  var Tt = {}, qi = () => performance.timeOrigin + performance.now();
  function Kn(e, t) {
    if (i) return V(18, 1, e, t);
    if (Tt[e] && (clearTimeout(Tt[e].id), delete Tt[e]), !t) return 0;
    var n = setTimeout(() => {
      delete Tt[e], he(() => Bo(e, performance.timeOrigin + performance.now()));
    }, t);
    return Tt[e] = { id: n, Ae: t }, 0;
  }
  function Ji(e, t, n, o) {
    e >>>= 0, t >>>= 0, n >>>= 0, o >>>= 0;
    var u = (/* @__PURE__ */ new Date()).getFullYear(), c = new Date(u, 0, 1).getTimezoneOffset();
    u = new Date(u, 6, 1).getTimezoneOffset();
    var h = Math.max(c, u);
    (p(), A)[e >>> 2 >>> 0] = 60 * h, (p(), x)[t >>> 2 >>> 0] = +(c != u), e = (t = (b) => {
      var E = Math.abs(b);
      return `UTC${0 <= b ? "-" : "+"}${String(Math.floor(E / 60)).padStart(2, "0")}${String(E % 60).padStart(2, "0")}`;
    })(c), t = t(u), u < c ? (Pe(e, n, 17), Pe(t, o, 17)) : (Pe(e, o, 17), Pe(t, n, 17));
  }
  var Xi = () => Date.now(), Qi = 1;
  function Zi(e, t, n) {
    if (n >>>= 0, !(0 <= e && 3 >= e)) return 28;
    if (e === 0) e = Date.now();
    else {
      if (!Qi) return 52;
      e = performance.timeOrigin + performance.now();
    }
    return e = Math.round(1e6 * e), (p(), me)[n >>> 3 >>> 0] = BigInt(e), 0;
  }
  var Br = [], eo = (e, t) => {
    Br.length = 0;
    for (var n; n = (p(), J)[e++ >>> 0]; ) {
      var o = n != 105;
      t += (o &= n != 112) && t % 8 ? 4 : 0, Br.push(n == 112 ? (p(), A)[t >>> 2 >>> 0] : n == 106 ? (p(), me)[t >>> 3 >>> 0] : n == 105 ? (p(), x)[t >>> 2 >>> 0] : (p(), ae)[t >>> 3 >>> 0]), t += o ? 8 : 4;
    }
    return Br;
  };
  function Ki(e, t, n) {
    return e >>>= 0, t = eo(t >>> 0, n >>> 0), Vr[e](...t);
  }
  function eu(e, t, n) {
    return e >>>= 0, t = eo(t >>> 0, n >>> 0), Vr[e](...t);
  }
  var tu = () => {
  };
  function ru(e, t) {
    return O(ct(e >>> 0, t >>> 0));
  }
  var nu = () => {
    throw G += 1, "unwind";
  };
  function ou() {
    return 4294901760;
  }
  var au = () => 1, su = () => navigator.hardwareConcurrency;
  function iu(e) {
    e >>>= 0;
    var t = (p(), J).length;
    if (e <= t || 4294901760 < e) return false;
    for (var n = 1; 4 >= n; n *= 2) {
      var o = t * (1 + 0.2 / n);
      o = Math.min(o, e + 100663296);
      e: {
        o = (Math.min(4294901760, 65536 * Math.ceil(Math.max(e, o) / 65536)) - Fe.buffer.byteLength + 65535) / 65536 | 0;
        try {
          Fe.grow(o), se();
          var u = 1;
          break e;
        } catch {
        }
        u = void 0;
      }
      if (u) return true;
    }
    return false;
  }
  var Ce = (e) => {
    var t = _e(e) + 1, n = Ft(t);
    return Pe(e, n, t), n;
  }, Mr = (e, t) => {
    (p(), A)[e >>> 2 >>> 0] = t;
    var n = (p(), A)[e >>> 2 >>> 0];
    (p(), A)[e + 4 >>> 2 >>> 0] = (t - n) / 4294967296;
  }, vt = (e) => (p(), A)[e >>> 2 >>> 0] + 4294967296 * (p(), x)[e + 4 >>> 2 >>> 0], de = [], uu = (e, t) => {
    de[e >>> 0] = t;
  }, Re = [], Nt = [], lt = (e, t) => {
    Nt[e] = new Promise((n) => t.finally(() => n(e)));
  }, L = (e) => {
    if (e) return de[e >>> 0];
  }, fu = (e, t) => {
    for (e = (p(), A)[e >>> 2 >>> 0]; e; e = (p(), A)[e >>> 2 >>> 0]) t[(p(), x)[e + 4 >>> 2 >>> 0]](e);
  }, kt = (e, t, n) => {
    (p(), A)[e >>> 2 >>> 0] = t, (p(), A)[e + 4 >>> 2 >>> 0] = n;
  }, to = (e) => {
    var t = (p(), A)[e >>> 2 >>> 0];
    return e = (p(), A)[e + 4 >>> 2 >>> 0], ct(t, e);
  }, Ne = (e) => {
    var t = (p(), A)[e >>> 2 >>> 0];
    return e = (p(), A)[e + 4 >>> 2 >>> 0], t ? ct(t, e) : e === 0 ? "" : void 0;
  }, cu = (e) => {
    var t = Ne(e + 4), n = (n = (p(), A)[e + 12 >>> 2 >>> 0]) ? L(n) : "auto";
    if (e += 16) {
      var o = L((p(), A)[e + 4 >>> 2 >>> 0]), u = (p(), A)[e + 16 >>> 2 >>> 0], c = (p(), A)[e + 20 >>> 2 >>> 0];
      if (u) {
        for (var h = {}, b = 0; b < u; ++b) {
          var E = c + 24 * b;
          h[to(E + 4)] = (p(), ae)[E + 16 >>> 3 >>> 0];
        }
        u = h;
      } else u = void 0;
      e = { module: o, constants: u, entryPoint: Ne(e + 8) };
    } else e = void 0;
    return { label: t, layout: n, compute: e };
  }, ro = (e, t) => {
    function n(o, u) {
      o = e[o], (p(), A)[t + u >>> 2 >>> 0] = o;
    }
    n("maxTextureDimension1D", 4), n("maxTextureDimension2D", 8), n("maxTextureDimension3D", 12), n("maxTextureArrayLayers", 16), n("maxBindGroups", 20), n("maxBindGroupsPlusVertexBuffers", 24), n("maxBindingsPerBindGroup", 28), n("maxDynamicUniformBuffersPerPipelineLayout", 32), n("maxDynamicStorageBuffersPerPipelineLayout", 36), n("maxSampledTexturesPerShaderStage", 40), n("maxSamplersPerShaderStage", 44), n("maxStorageBuffersPerShaderStage", 48), n("maxStorageTexturesPerShaderStage", 52), n("maxUniformBuffersPerShaderStage", 56), n("minUniformBufferOffsetAlignment", 80), n("minStorageBufferOffsetAlignment", 84), Mr(t + 64, e.maxUniformBufferBindingSize), Mr(t + 72, e.maxStorageBufferBindingSize), n("maxVertexBuffers", 88), Mr(t + 96, e.maxBufferSize), n("maxVertexAttributes", 104), n("maxVertexBufferArrayStride", 108), n("maxInterStageShaderVariables", 112), n("maxColorAttachments", 116), n("maxColorAttachmentBytesPerSample", 120), n("maxComputeWorkgroupStorageSize", 124), n("maxComputeInvocationsPerWorkgroup", 128), n("maxComputeWorkgroupSizeX", 132), n("maxComputeWorkgroupSizeY", 136), n("maxComputeWorkgroupSizeZ", 140), n("maxComputeWorkgroupsPerDimension", 144), e.ze !== void 0 && n("maxImmediateSize", 148);
  }, du = [, "validation", "out-of-memory", "internal"], lu = [, "compatibility", "core"], no = { 1: "core-features-and-limits", 2: "depth-clip-control", 3: "depth32float-stencil8", 4: "texture-compression-bc", 5: "texture-compression-bc-sliced-3d", 6: "texture-compression-etc2", 7: "texture-compression-astc", 8: "texture-compression-astc-sliced-3d", 9: "timestamp-query", 10: "indirect-first-instance", 11: "shader-f16", 12: "rg11b10ufloat-renderable", 13: "bgra8unorm-storage", 14: "float32-filterable", 15: "float32-blendable", 16: "clip-distances", 17: "dual-source-blending", 18: "subgroups", 19: "texture-formats-tier1", 20: "texture-formats-tier2", 21: "primitive-index", 22: "texture-component-swizzle", 327692: "chromium-experimental-unorm16-texture-formats", 327729: "chromium-experimental-multi-draw-indirect" }, pu = [, "low-power", "high-performance"], mu = [, "occlusion", "timestamp"], hu = { undefined: 1, unknown: 1, destroyed: 2 };
  function yu(e, t, n, o, u, c) {
    t = ce(t), n = ce(n), o >>>= 0, u >>>= 0, c >>>= 0;
    var h = L(e >>> 0);
    if (e = {}, c) {
      var b = (p(), A)[c + 12 >>> 2 >>> 0];
      if (b) {
        var E = (p(), A)[c + 16 >>> 2 >>> 0];
        e.requiredFeatures = Array.from((p(), A).subarray(E >>> 2 >>> 0, E + 4 * b >>> 2 >>> 0), (B) => no[B]);
      }
      var I = (p(), A)[c + 20 >>> 2 >>> 0];
      if (I) {
        let B = function(ye, fe, Qe = false) {
          fe = I + fe, (fe = (p(), A)[fe >>> 2 >>> 0]) == 4294967295 || Qe && fe == 0 || (F[ye] = fe);
        }, ue = function(ye, fe) {
          fe = I + fe;
          var Qe = (p(), A)[fe >>> 2 >>> 0], $f = (p(), A)[fe + 4 >>> 2 >>> 0];
          Qe == 4294967295 && $f == 4294967295 || (F[ye] = vt(fe));
        };
        var j = B, te = ue, F = {};
        B("maxTextureDimension1D", 4), B("maxTextureDimension2D", 8), B("maxTextureDimension3D", 12), B("maxTextureArrayLayers", 16), B("maxBindGroups", 20), B("maxBindGroupsPlusVertexBuffers", 24), B("maxDynamicUniformBuffersPerPipelineLayout", 32), B("maxDynamicStorageBuffersPerPipelineLayout", 36), B("maxSampledTexturesPerShaderStage", 40), B("maxSamplersPerShaderStage", 44), B("maxStorageBuffersPerShaderStage", 48), B("maxStorageTexturesPerShaderStage", 52), B("maxUniformBuffersPerShaderStage", 56), B("minUniformBufferOffsetAlignment", 80), B("minStorageBufferOffsetAlignment", 84), ue("maxUniformBufferBindingSize", 64), ue("maxStorageBufferBindingSize", 72), B("maxVertexBuffers", 88), ue("maxBufferSize", 96), B("maxVertexAttributes", 104), B("maxVertexBufferArrayStride", 108), B("maxInterStageShaderVariables", 112), B("maxColorAttachments", 116), B("maxColorAttachmentBytesPerSample", 120), B("maxComputeWorkgroupStorageSize", 124), B("maxComputeInvocationsPerWorkgroup", 128), B("maxComputeWorkgroupSizeX", 132), B("maxComputeWorkgroupSizeY", 136), B("maxComputeWorkgroupSizeZ", 140), B("maxComputeWorkgroupsPerDimension", 144), B("maxImmediateSize", 148, true), e.requiredLimits = F;
      }
      (b = (p(), A)[c + 24 >>> 2 >>> 0]) && (b = { label: Ne(b + 4) }, e.defaultQueue = b), e.label = Ne(c + 4);
    }
    G += 1, lt(t, h.requestDevice(e).then((B) => {
      --G, he(() => {
        de[u >>> 0] = B.queue, de[o >>> 0] = B, G += 1, lt(n, B.lost.then((ue) => {
          he(() => {
            B.onuncapturederror = () => {
            };
            var ye = P(), fe = Ce(ue.message);
            _r(n, hu[ue.reason], fe), D(ye);
          }), --G;
        })), B.onuncapturederror = (ue) => {
          var ye = 5;
          ue.error instanceof GPUValidationError ? ye = 2 : ue.error instanceof GPUOutOfMemoryError ? ye = 3 : ue.error instanceof GPUInternalError && (ye = 4);
          var fe = P();
          ue = Ce(ue.error.message), Io(o, ye, ue), D(fe);
        }, "adapterInfo" in B || (B.adapterInfo = h.info), kr(t, 1, o, 0);
      });
    }, (B) => {
      --G, he(() => {
        var ue = P(), ye = Ce(B.message);
        kr(t, 3, o, ye), n && _r(n, 4, ye), D(ue);
      });
    }));
  }
  function bu(e) {
    var t = L(e >>>= 0), n = Re[e];
    if (n) {
      for (var o = 0; o < n.length; ++o) n[o]();
      delete Re[e];
    }
    t.destroy();
  }
  function wu(e, t, n) {
    n >>>= 0;
    var o = L(e >>>= 0);
    n == 4294967295 && (n = void 0);
    try {
      var u = o.getMappedRange(t >>> 0, n);
    } catch {
      return 0;
    }
    var c = Gr(16, u.byteLength);
    return (p(), J).set(new Uint8Array(u), c >>> 0), Re[e].push(() => xe(c)), c;
  }
  function gu(e, t, n) {
    n >>>= 0;
    var o = L(e >>>= 0);
    n == 4294967295 && (n = void 0);
    try {
      var u = o.getMappedRange(t >>> 0, n);
    } catch {
      return 0;
    }
    var c = Gr(16, u.byteLength);
    return (p(), J).fill(0, c, u.byteLength), Re[e].push(() => {
      new Uint8Array(u).set((p(), J).subarray(c >>> 0, c + u.byteLength >>> 0)), xe(c);
    }), c;
  }
  function Tu(e, t, n, o, u) {
    e >>>= 0, t = ce(t), n = ce(n), u >>>= 0;
    var c = L(e);
    Re[e] = [], u == 4294967295 && (u = void 0), G += 1, lt(t, c.mapAsync(n, o >>> 0, u).then(() => {
      --G, he(() => {
        Rr(t, 1, 0);
      });
    }, (h) => {
      --G, he(() => {
        P();
        var b = Ce(h.message);
        Rr(t, h.name === "AbortError" ? 4 : h.name === "OperationError" ? 3 : 0, b), delete Re[e];
      });
    }));
  }
  function vu(e) {
    var t = L(e >>>= 0), n = Re[e];
    if (n) {
      for (var o = 0; o < n.length; ++o) n[o]();
      delete Re[e], t.unmap();
    }
  }
  function Eu(e) {
    delete de[e >>> 0];
  }
  function Su(e, t, n) {
    e >>>= 0, t >>>= 0, n >>>= 0;
    var o = !!(p(), A)[t + 32 >>> 2 >>> 0];
    t = { label: Ne(t + 4), usage: (p(), A)[t + 16 >>> 2 >>> 0], size: vt(t + 24), mappedAtCreation: o }, e = L(e);
    try {
      var u = e.createBuffer(t);
    } catch {
      return false;
    }
    return de[n >>> 0] = u, o && (Re[n] = []), true;
  }
  function Au(e, t, n, o) {
    e >>>= 0, t = ce(t), o >>>= 0, n = cu(n >>> 0), e = L(e), G += 1, lt(t, e.createComputePipelineAsync(n).then((u) => {
      --G, he(() => {
        de[o >>> 0] = u, Pr(t, 1, o, 0);
      });
    }, (u) => {
      --G, he(() => {
        var c = P(), h = Ce(u.message);
        Pr(t, u.reason === "validation" ? 3 : u.reason === "internal" ? 4 : 0, o, h), D(c);
      });
    }));
  }
  function Iu(e, t, n) {
    e >>>= 0, t >>>= 0, n >>>= 0;
    var o = (p(), A)[t >>> 2 >>> 0], u = (p(), x)[o + 4 >>> 2 >>> 0];
    t = { label: Ne(t + 4), code: "" }, u === 2 && (t.code = to(o + 8)), e = L(e).createShaderModule(t), de[n >>> 0] = e;
  }
  var xu = (e) => {
    (e = L(e)).onuncapturederror = null, e.destroy();
  };
  function Lu(e, t) {
    t = ce(t), e = L(e >>> 0), G += 1, lt(t, e.popErrorScope().then((n) => {
      --G, he(() => {
        var o = 5;
        n ? n instanceof GPUValidationError ? o = 2 : n instanceof GPUOutOfMemoryError ? o = 3 : n instanceof GPUInternalError && (o = 4) : o = 1;
        var u = P(), c = n ? Ce(n.message) : 0;
        Nr(t, 1, o, c), D(u);
      });
    }, (n) => {
      --G, he(() => {
        var o = P(), u = Ce(n.message);
        Nr(t, 1, 5, u), D(o);
      });
    }));
  }
  function Ou(e, t, n, o) {
    if (t = ce(t), o >>>= 0, n >>>= 0) {
      var u = { featureLevel: lu[(p(), x)[n + 4 >>> 2 >>> 0]], powerPreference: pu[(p(), x)[n + 8 >>> 2 >>> 0]], forceFallbackAdapter: !!(p(), A)[n + 12 >>> 2 >>> 0] };
      (e = (p(), A)[n >>> 2 >>> 0]) !== 0 && (p(), u.De = !!(p(), A)[e + 8 >>> 2 >>> 0]);
    }
    "gpu" in navigator ? (G += 1, lt(t, navigator.gpu.requestAdapter(u).then((c) => {
      --G, he(() => {
        if (c) de[o >>> 0] = c, Et(t, 1, o, 0);
        else {
          var h = P(), b = Ce("WebGPU not available on this browser (requestAdapter returned null)");
          Et(t, 3, o, b), D(h);
        }
      });
    }, (c) => {
      --G, he(() => {
        var h = P(), b = Ce(c.message);
        Et(t, 4, o, b), D(h);
      });
    }))) : (u = P(), e = Ce("WebGPU not available on this browser (navigator.gpu is not available)"), Et(t, 3, o, e), D(u));
  }
  function Bu(e, t, n) {
    return e >>>= 0, t >>>= 0, n >>>= 0, Hn(async () => {
      var o = [];
      if (n) {
        var u = (p(), x)[n >>> 2 >>> 0];
        o.length = t + 1, o[t] = new Promise((b) => setTimeout(b, u, 0));
      } else o.length = t;
      for (var c = 0; c < t; ++c) {
        var h = vt(e + 8 * c);
        if (!(h in Nt)) return h;
        o[c] = Nt[h];
      }
      return o = await Promise.race(o), delete Nt[o], o;
    });
  }
  var Cr, Ur = {}, oo = () => {
    if (!Cr) {
      var e, t = { USER: "web_user", LOGNAME: "web_user", PATH: "/", PWD: "/", HOME: "/home/web_user", LANG: (globalThis.navigator?.language ?? "C").replace("-", "_") + ".UTF-8", _: "./this.program" };
      for (e in Ur) Ur[e] === void 0 ? delete t[e] : t[e] = Ur[e];
      var n = [];
      for (e in t) n.push(`${e}=${t[e]}`);
      Cr = n;
    }
    return Cr;
  };
  function ao(e, t) {
    if (i) return V(19, 1, e, t);
    e >>>= 0, t >>>= 0;
    var n, o = 0, u = 0;
    for (n of oo()) {
      var c = t + o;
      (p(), A)[e + u >>> 2 >>> 0] = c, o += Pe(n, c, 1 / 0) + 1, u += 4;
    }
    return 0;
  }
  function so(e, t) {
    if (i) return V(20, 1, e, t);
    e >>>= 0, t >>>= 0;
    var n = oo();
    for (var o of ((p(), A)[e >>> 2 >>> 0] = n.length, e = 0, n)) e += _e(o) + 1;
    return (p(), A)[t >>> 2 >>> 0] = e, 0;
  }
  function io(e) {
    return i ? V(21, 1, e) : 52;
  }
  function uo(e, t, n, o) {
    return i ? V(22, 1, e, t, n, o) : 52;
  }
  function fo(e, t, n, o) {
    return i ? V(23, 1, e, t, n, o) : 70;
  }
  var Mu = [null, [], []];
  function co(e, t, n, o) {
    if (i) return V(24, 1, e, t, n, o);
    t >>>= 0, n >>>= 0, o >>>= 0;
    for (var u = 0, c = 0; c < n; c++) {
      var h = (p(), A)[t >>> 2 >>> 0], b = (p(), A)[t + 4 >>> 2 >>> 0];
      t += 8;
      for (var E = 0; E < b; E++) {
        var I = e, F = (p(), J)[h + E >>> 0], j = Mu[I];
        F === 0 || F === 10 ? ((I === 1 ? Y : O)(xn(j)), j.length = 0) : j.push(F);
      }
      u += b;
    }
    return (p(), A)[o >>> 2 >>> 0] = u, 0;
  }
  function Cu(e) {
    return e >>> 0;
  }
  function Uu(e, t) {
    return ro(L(e >>> 0).limits, t >>> 0), 1;
  }
  function Du(e, t) {
    return L(e >>> 0).features.has(no[t]);
  }
  function Pu(e) {
    return BigInt(L(e >>> 0).size);
  }
  function _u(e) {
    return BigInt(L(e >>> 0).usage);
  }
  function Ru(e, t) {
    if (e >>>= 0, t >>>= 0) {
      var n = Ne(t + 4);
      n = { label: n, timestampWrites: t = (t = (p(), A)[t + 12 >>> 2 >>> 0]) !== 0 ? { querySet: L((p(), A)[t + 4 >>> 2 >>> 0]), beginningOfPassWriteIndex: (p(), A)[t + 8 >>> 2 >>> 0], endOfPassWriteIndex: (p(), A)[t + 12 >>> 2 >>> 0] } : void 0 };
    }
    return t = L(e), e = To(0), n = t.beginComputePass(n), de[e >>> 0] = n, e;
  }
  function Nu(e, t, n, o) {
    n = ce(n), (o = ce(o)) == -1 && (o = void 0), (e = L(e >>> 0)).clearBuffer(L(t >>> 0), n, o);
  }
  function ku(e, t, n, o, u, c) {
    n = ce(n), u = ce(u), c = ce(c), L(e >>> 0).copyBufferToBuffer(L(t >>> 0), n, L(o >>> 0), u, c);
  }
  function Wu(e) {
    var t = L(e >>> 0);
    return e = wo(0), t = t.finish(), de[e >>> 0] = t, e;
  }
  function Fu(e, t, n, o, u, c) {
    c = ce(c), L(e >>> 0).resolveQuerySet(L(t >>> 0), n, o, L(u >>> 0), c);
  }
  function Gu(e, t, n, o) {
    L(e >>> 0).dispatchWorkgroups(t, n, o);
  }
  function $u(e, t, n) {
    n = ce(n), L(e >>> 0).dispatchWorkgroupsIndirect(L(t >>> 0), n);
  }
  function zu(e) {
    L(e >>> 0).end();
  }
  function Vu(e, t, n, o, u) {
    o >>>= 0, u >>>= 0, e = L(e >>> 0), n = L(n >>> 0), o == 0 ? e.setBindGroup(t, n) : e.setBindGroup(t, n, (p(), A), u >>> 2, o);
  }
  function ju(e, t) {
    L(e >>> 0).setPipeline(L(t >>> 0));
  }
  function Hu(e, t, n) {
    L(e >>> 0).Ce(L(t >>> 0), n);
  }
  function Yu(e, t) {
    var n = L(e >>> 0);
    return e = bo(0), t = n.getBindGroupLayout(t), de[e >>> 0] = t, e;
  }
  function qu(e, t) {
    function n(u) {
      var c = (p(), A)[u + 8 >>> 2 >>> 0], h = (p(), A)[u + 32 >>> 2 >>> 0], b = (p(), A)[u + 36 >>> 2 >>> 0], E = 0;
      return fu(u, { 327681: (I) => {
        E = (p(), A)[I + 8 >>> 2 >>> 0];
      } }), c ? ((h = vt(u + 24)) == -1 && (h = void 0), c = { buffer: L(c), offset: vt(u + 16), size: h }) : c = L(h || b || E), { binding: (p(), A)[u + 4 >>> 2 >>> 0], resource: c };
    }
    e >>>= 0, t = { label: Ne(4 + (t >>>= 0)), layout: L((p(), A)[t + 12 >>> 2 >>> 0]), entries: (function(u, c) {
      for (var h = [], b = 0; b < u; ++b) h.push(n(c + 40 * b));
      return h;
    })((p(), A)[t + 16 >>> 2 >>> 0], (p(), A)[t + 20 >>> 2 >>> 0]) }, e = L(e);
    var o = yo(0);
    return uu(o, e.createBindGroup(t)), o;
  }
  function Ju(e, t) {
    var n;
    return e >>>= 0, (t >>>= 0) && (n = { label: Ne(t + 4) }), t = L(e), e = go(0), n = t.createCommandEncoder(n), de[e >>> 0] = n, e;
  }
  function Xu(e, t) {
    e >>>= 0, t >>>= 0, t = { type: mu[(p(), x)[t + 12 >>> 2 >>> 0]], count: (p(), A)[t + 16 >>> 2 >>> 0] };
    var n = L(e);
    return e = vo(0), t = n.createQuerySet(t), de[e >>> 0] = t, e;
  }
  function Qu(e, t) {
    e = L(e >>> 0).adapterInfo, t >>>= 0, (p(), A)[t + 52 >>> 2 >>> 0] = e.subgroupMinSize, (p(), A)[t + 56 >>> 2 >>> 0] = e.subgroupMaxSize;
    var n = e.vendor + e.architecture + e.device + e.description, o = _e(n) + 1, u = pt(o);
    return u && Pe(n, u, o), n = u, o = _e(e.vendor), kt(t + 4, n, o), n += o, o = _e(e.architecture), kt(t + 12, n, o), n += o, o = _e(e.device), kt(t + 20, n, o), kt(t + 28, n + o, _e(e.description)), (p(), x)[t + 36 >>> 2 >>> 0] = 2, e = e.isFallbackAdapter ? 3 : 4, (p(), x)[t + 40 >>> 2 >>> 0] = e, (p(), A)[t + 44 >>> 2 >>> 0] = 0, (p(), A)[t + 48 >>> 2 >>> 0] = 0, 1;
  }
  var Zu = { "core-features-and-limits": 1, "depth-clip-control": 2, "depth32float-stencil8": 3, "texture-compression-bc": 4, "texture-compression-bc-sliced-3d": 5, "texture-compression-etc2": 6, "texture-compression-astc": 7, "texture-compression-astc-sliced-3d": 8, "timestamp-query": 9, "indirect-first-instance": 10, "shader-f16": 11, "rg11b10ufloat-renderable": 12, "bgra8unorm-storage": 13, "float32-filterable": 14, "float32-blendable": 15, "clip-distances": 16, "dual-source-blending": 17, subgroups: 18, "texture-formats-tier1": 19, "texture-formats-tier2": 20, "primitive-index": 21, "texture-component-swizzle": 22, "chromium-experimental-unorm16-texture-formats": 327692, "chromium-experimental-multi-draw-indirect": 327729 };
  function Ku(e, t) {
    t >>>= 0;
    var n = L(e >>> 0);
    e = pt(4 * n.features.size);
    var o = 0, u = 0;
    for (let c of n.features) 0 <= (n = Zu[c]) && ((p(), x)[e + o >>> 2 >>> 0] = n, o += 4, u++);
    (p(), A)[t + 4 >>> 2 >>> 0] = e, (p(), A)[t >>> 2 >>> 0] = u;
  }
  function ef(e, t) {
    return ro(L(e >>> 0).limits, t >>> 0), 1;
  }
  function tf(e, t) {
    L(e >>> 0).pushErrorScope(du[t]);
  }
  function rf(e, t, n) {
    t >>>= 0, n >>>= 0, e = L(e >>> 0), t = Array.from((p(), x).subarray(n >>> 2 >>> 0, n + 4 * t >>> 2 >>> 0), (o) => L(o)), e.submit(t);
  }
  function nf(e, t, n, o, u) {
    n = ce(n), o >>>= 0, u >>>= 0, e = L(e >>> 0), t = L(t >>> 0), o = (p(), J).subarray(o >>> 0, o + u >>> 0), e.writeBuffer(t, n, o, 0, u);
  }
  i || (function() {
    for (var e = r.numThreads - 1; e--; ) gn();
    Ae.push(async () => {
      var t = (async function() {
        if (!i) return Promise.all(We.map(wn));
      })();
      Oe++, await t, --Oe == 0 && ee && (t = ee, ee = null, t());
    });
  })(), i || (Fe = new WebAssembly.Memory({ initial: 256, maximum: 65536, shared: true }), se()), r.wasmBinary && (g = r.wasmBinary), r.stackSave = () => P(), r.stackRestore = (e) => D(e), r.stackAlloc = (e) => Ft(e), r.setValue = function(e, t, n = "i8") {
    switch (n.endsWith("*") && (n = "*"), n) {
      case "i1":
      case "i8":
        (p(), X)[e >>> 0] = t;
        break;
      case "i16":
        (p(), Ue)[e >>> 1 >>> 0] = t;
        break;
      case "i32":
        (p(), x)[e >>> 2 >>> 0] = t;
        break;
      case "i64":
        (p(), me)[e >>> 3 >>> 0] = BigInt(t);
        break;
      case "float":
        (p(), _)[e >>> 2 >>> 0] = t;
        break;
      case "double":
        (p(), ae)[e >>> 3 >>> 0] = t;
        break;
      case "*":
        (p(), A)[e >>> 2 >>> 0] = t;
        break;
      default:
        Te(`invalid type for setValue: ${n}`);
    }
  }, r.getValue = function(e, t = "i8") {
    switch (t.endsWith("*") && (t = "*"), t) {
      case "i1":
      case "i8":
        return (p(), X)[e >>> 0];
      case "i16":
        return (p(), Ue)[e >>> 1 >>> 0];
      case "i32":
        return (p(), x)[e >>> 2 >>> 0];
      case "i64":
        return (p(), me)[e >>> 3 >>> 0];
      case "float":
        return (p(), _)[e >>> 2 >>> 0];
      case "double":
        return (p(), ae)[e >>> 3 >>> 0];
      case "*":
        return (p(), A)[e >>> 2 >>> 0];
      default:
        Te(`invalid type for getValue: ${t}`);
    }
  }, r.UTF8ToString = ct, r.stringToUTF8 = Pe, r.lengthBytesUTF8 = _e;
  var lo, po, Dr, Wt, xe, pt, mo, ho, yo, bo, wo, go, To, vo, Eo, So, Ao, Pr, _r, Rr, Nr, Et, kr, Io, Wr, xo, Lo, Oo, Fr, Bo, Mo, Gr, N, St, Co, D, Ft, P, Uo, $r, Do, Po, _o, zr, Ro, No, ko, Wo, Fo, Go, $o, zo, Vo, jo, Ho, Yo, qo, Jo, Xo, Qo, Zo, Ko, ea, ta, ra, na, oa, aa, sa, ia, ua, fa, ca, da, la, pa, ma, ha, ya, ba, wa, ga, ke, of = [qe, yr, En, Ln, On, Bn, Mn, Cn, Un, Dn, Pn, _n, Rn, Nn, kn, Wn, Qn, Zn, Kn, ao, so, io, uo, fo, co], Vr = { 970348: (e, t, n, o, u) => {
    if (r === void 0 || !r.Uc) return 1;
    if ((e = ct(Number(e >>> 0))).startsWith("./") && (e = e.substring(2)), !(e = r.Uc.get(e))) return 2;
    if (t = Number(t >>> 0), n = Number(n >>> 0), o = Number(o >>> 0), t + n > e.byteLength) return 3;
    try {
      let c = e.subarray(t, t + n);
      switch (u) {
        case 0:
          (p(), J).set(c, o >>> 0);
          break;
        case 1:
          r.ad ? r.ad(o, c) : r.ne(o, c);
          break;
        default:
          return 4;
      }
      return 0;
    } catch {
      return 4;
    }
  }, 971172: (e, t, n) => {
    r.Sd(e, (p(), J).subarray(t >>> 0, t + n >>> 0));
  }, 971236: () => r.le(), 971278: (e) => {
    r.jd(e);
  }, 971315: () => typeof wasmOffsetConverter < "u" };
  function af(e, t, n, o) {
    var u = P();
    try {
      return zo(e, t, n, o);
    } catch (c) {
      if (D(u), c !== c + 0) throw c;
      N(1, 0);
    }
  }
  function sf(e, t, n) {
    var o = P();
    try {
      return Fo(e, t, n);
    } catch (u) {
      if (D(o), u !== u + 0) throw u;
      N(1, 0);
    }
  }
  function uf(e) {
    var t = P();
    try {
      Ro(e);
    } catch (n) {
      if (D(t), n !== n + 0) throw n;
      N(1, 0);
    }
  }
  function ff(e, t) {
    var n = P();
    try {
      return zr(e, t);
    } catch (o) {
      if (D(n), o !== o + 0) throw o;
      N(1, 0);
    }
  }
  function cf(e, t, n) {
    var o = P();
    try {
      _o(e, t, n);
    } catch (u) {
      if (D(o), u !== u + 0) throw u;
      N(1, 0);
    }
  }
  function df(e, t) {
    var n = P();
    try {
      Vo(e, t);
    } catch (o) {
      if (D(n), o !== o + 0) throw o;
      N(1, 0);
    }
  }
  function lf(e, t, n, o, u, c, h) {
    var b = P();
    try {
      return Wo(e, t, n, o, u, c, h);
    } catch (E) {
      if (D(b), E !== E + 0) throw E;
      N(1, 0);
    }
  }
  function pf(e, t, n, o, u, c) {
    var h = P();
    try {
      No(e, t, n, o, u, c);
    } catch (b) {
      if (D(h), b !== b + 0) throw b;
      N(1, 0);
    }
  }
  function mf(e, t, n, o) {
    var u = P();
    try {
      $o(e, t, n, o);
    } catch (c) {
      if (D(u), c !== c + 0) throw c;
      N(1, 0);
    }
  }
  function hf(e, t, n, o, u, c, h) {
    var b = P();
    try {
      Ho(e, t, n, o, u, c, h);
    } catch (E) {
      if (D(b), E !== E + 0) throw E;
      N(1, 0);
    }
  }
  function yf(e, t, n, o, u, c, h) {
    var b = P();
    try {
      Yo(e, t, n, o, u, c, h);
    } catch (E) {
      if (D(b), E !== E + 0) throw E;
      N(1, 0);
    }
  }
  function bf(e, t, n, o, u, c, h, b) {
    var E = P();
    try {
      ra(e, t, n, o, u, c, h, b);
    } catch (I) {
      if (D(E), I !== I + 0) throw I;
      N(1, 0);
    }
  }
  function wf(e, t, n, o, u, c, h, b, E, I, F, j) {
    var te = P();
    try {
      qo(e, t, n, o, u, c, h, b, E, I, F, j);
    } catch (B) {
      if (D(te), B !== B + 0) throw B;
      N(1, 0);
    }
  }
  function gf(e, t, n, o, u) {
    var c = P();
    try {
      return jo(e, t, n, o, u);
    } catch (h) {
      if (D(c), h !== h + 0) throw h;
      N(1, 0);
    }
  }
  function Tf(e, t, n, o, u) {
    var c = P();
    try {
      ko(e, t, n, o, u);
    } catch (h) {
      if (D(c), h !== h + 0) throw h;
      N(1, 0);
    }
  }
  function vf(e, t, n, o, u, c, h, b) {
    var E = P();
    try {
      Go(e, t, n, o, u, c, h, b);
    } catch (I) {
      if (D(E), I !== I + 0) throw I;
      N(1, 0);
    }
  }
  function Ef(e) {
    var t = P();
    try {
      return na(e);
    } catch (n) {
      if (D(t), n !== n + 0) throw n;
      N(1, 0);
    }
  }
  function Sf(e, t, n) {
    var o = P();
    try {
      return oa(e, t, n);
    } catch (u) {
      if (D(o), u !== u + 0) throw u;
      N(1, 0);
    }
  }
  function Af(e, t) {
    var n = P();
    try {
      return ha(e, t);
    } catch (o) {
      if (D(n), o !== o + 0) throw o;
      return N(1, 0), 0n;
    }
  }
  function If2(e) {
    var t = P();
    try {
      return Jo(e);
    } catch (n) {
      if (D(t), n !== n + 0) throw n;
      return N(1, 0), 0n;
    }
  }
  function xf(e, t, n, o) {
    var u = P();
    try {
      return aa(e, t, n, o);
    } catch (c) {
      if (D(u), c !== c + 0) throw c;
      N(1, 0);
    }
  }
  function Lf(e, t, n, o, u) {
    var c = P();
    try {
      return sa(e, t, n, o, u);
    } catch (h) {
      if (D(c), h !== h + 0) throw h;
      N(1, 0);
    }
  }
  function Of(e, t, n, o, u, c) {
    var h = P();
    try {
      return ia(e, t, n, o, u, c);
    } catch (b) {
      if (D(h), b !== b + 0) throw b;
      N(1, 0);
    }
  }
  function Bf(e, t, n, o, u, c) {
    var h = P();
    try {
      return ea(e, t, n, o, u, c);
    } catch (b) {
      if (D(h), b !== b + 0) throw b;
      N(1, 0);
    }
  }
  function Mf(e, t, n, o, u, c) {
    var h = P();
    try {
      return ua(e, t, n, o, u, c);
    } catch (b) {
      if (D(h), b !== b + 0) throw b;
      N(1, 0);
    }
  }
  function Cf(e, t, n, o, u, c, h, b) {
    var E = P();
    try {
      return ta(e, t, n, o, u, c, h, b);
    } catch (I) {
      if (D(E), I !== I + 0) throw I;
      N(1, 0);
    }
  }
  function Uf(e, t, n, o, u) {
    var c = P();
    try {
      return fa(e, t, n, o, u);
    } catch (h) {
      if (D(c), h !== h + 0) throw h;
      return N(1, 0), 0n;
    }
  }
  function Df(e, t, n, o) {
    var u = P();
    try {
      return ca(e, t, n, o);
    } catch (c) {
      if (D(u), c !== c + 0) throw c;
      N(1, 0);
    }
  }
  function Pf(e, t, n, o) {
    var u = P();
    try {
      return da(e, t, n, o);
    } catch (c) {
      if (D(u), c !== c + 0) throw c;
      N(1, 0);
    }
  }
  function _f(e, t, n, o, u, c, h, b, E, I, F, j) {
    var te = P();
    try {
      return la(e, t, n, o, u, c, h, b, E, I, F, j);
    } catch (B) {
      if (D(te), B !== B + 0) throw B;
      N(1, 0);
    }
  }
  function Rf(e, t, n, o, u, c, h, b, E, I, F) {
    var j = P();
    try {
      pa(e, t, n, o, u, c, h, b, E, I, F);
    } catch (te) {
      if (D(j), te !== te + 0) throw te;
      N(1, 0);
    }
  }
  function Nf(e, t, n, o, u, c, h, b, E, I, F, j, te, B, ue, ye) {
    var fe = P();
    try {
      ma(e, t, n, o, u, c, h, b, E, I, F, j, te, B, ue, ye);
    } catch (Qe) {
      if (D(fe), Qe !== Qe + 0) throw Qe;
      N(1, 0);
    }
  }
  function kf(e, t, n) {
    var o = P();
    try {
      return Qo(e, t, n);
    } catch (u) {
      if (D(o), u !== u + 0) throw u;
      return N(1, 0), 0n;
    }
  }
  function Wf(e, t, n) {
    var o = P();
    try {
      return Xo(e, t, n);
    } catch (u) {
      if (D(o), u !== u + 0) throw u;
      N(1, 0);
    }
  }
  function Ff(e, t, n) {
    var o = P();
    try {
      return Zo(e, t, n);
    } catch (u) {
      if (D(o), u !== u + 0) throw u;
      N(1, 0);
    }
  }
  function Gf(e, t, n, o) {
    var u = P();
    try {
      Ko(e, t, n, o);
    } catch (c) {
      if (D(u), c !== c + 0) throw c;
      N(1, 0);
    }
  }
  function Gt() {
    if (0 < Oe) ee = Gt;
    else if (i) C?.(r), hr();
    else {
      for (var e = Ae; 0 < e.length; ) e.shift()(r);
      0 < Oe ? ee = Gt : (r.calledRun = true, W || (hr(), C?.(r)));
    }
  }
  return i || (ke = await bt(), Gt()), r.PTR_SIZE = 4, r.webgpuInit = (e) => {
    let t = /* @__PURE__ */ new WeakMap(), n, o, u = 1;
    r.webgpuRegisterDevice = (b) => {
      if (o !== void 0) throw Error("another WebGPU EP inference session is being created.");
      if (b) {
        var E = t.get(b);
        if (!E) {
          let I = ((F, j = 0) => {
            var te = Ao(j);
            return j = So(j, te), de[te >>> 0] = F.queue, de[j >>> 0] = F, j;
          })(b, E = ho(0));
          E = [u++, E, I], t.set(b, E);
        }
        return n = b, o = E[0], E;
      }
      n = void 0, o = 0;
    };
    let c = /* @__PURE__ */ new Map();
    r.webgpuOnCreateSession = (b) => {
      if (o !== void 0) {
        var E = o;
        if (o = void 0, b) {
          let I = Dr(E);
          c.set(b, I), E === 0 && e(n ?? L(I));
        }
        n = void 0;
      }
    }, r.webgpuOnReleaseSession = (b) => {
      c.delete(b);
    };
    let h = /* @__PURE__ */ Symbol("gpuBufferMetadata");
    r.webgpuRegisterBuffer = (b, E, I) => {
      if (I) return b[h] = [I, NaN], I;
      if (I = b[h]) return I[1]++, I[0];
      if ((E = c.get(E)) === void 0) throw Error("Invalid session handle passed to webgpuRegisterBuffer");
      return E = ((F, j = 0) => (F.mapState === "unmapped" || Te(), j = Eo(j), de[j >>> 0] = F, j))(b, E), b[h] = [E, 1], E;
    }, r.webgpuUnregisterBuffer = (b) => {
      let E = b[h];
      if (!E) throw Error("Buffer is not registered");
      E[1]--, E[1] === 0 && (mo(E[0]), delete b[h]);
    }, r.webgpuGetBuffer = (b) => L(b), r.webgpuCreateDownloader = (b, E, I) => {
      if ((I = c.get(I)) === void 0) throw Error("Invalid session handle passed to webgpuRegisterBuffer");
      let F = L(I), j = 16 * Math.ceil(Number(E) / 16);
      return async () => {
        let te = F.createBuffer({ size: j, usage: 9 });
        try {
          let B = F.createCommandEncoder();
          return B.copyBufferToBuffer(b, 0, te, 0, j), F.queue.submit([B.finish()]), await te.mapAsync(GPUMapMode.READ), te.getMappedRange().slice(0, E);
        } finally {
          te.destroy();
        }
      };
    }, r.ad = (b, E) => {
      var I = E.buffer;
      let F = E.byteOffset, j = E.byteLength;
      if (E = 16 * Math.ceil(Number(j) / 16), b = L(b), !n) {
        var te = Dr(o);
        n = L(te);
      }
      let B = (te = n.createBuffer({ mappedAtCreation: true, size: E, usage: 6 })).getMappedRange();
      new Uint8Array(B).set(new Uint8Array(I, F, j)), te.unmap(), (I = n.createCommandEncoder()).copyBufferToBuffer(te, 0, b, 0, E), n.queue.submit([I.finish()]), te.destroy();
    };
  }, r.webnnInit = (e) => {
    let t = e[0];
    [r.le, r.jd, r.webnnEnsureTensor, r.Sd, r.webnnDownloadTensor, r.ke, r.webnnEnableTraceEvent] = e.slice(1), r.webnnReleaseTensorId = r.jd, r.webnnUploadTensor = r.Sd, r.webnnRegisterMLContext = r.ke, r.webnnOnRunStart = (n) => t.onRunStart(n), r.webnnOnRunEnd = t.onRunEnd.bind(t), r.webnnOnReleaseSession = (n) => {
      t.onReleaseSession(n);
    }, r.webnnCreateMLTensorDownloader = (n, o) => t.createMLTensorDownloader(n, o), r.webnnRegisterMLTensor = (n, o, u, c) => t.registerMLTensor(n, o, u, c), r.webnnCreateMLContext = (n) => t.createMLContext(n), r.webnnRegisterMLConstant = (n, o, u, c, h, b) => t.registerMLConstant(n, o, u, c, h, r.Uc, b), r.webnnRegisterGraphInput = t.registerGraphInput.bind(t), r.webnnIsGraphInput = t.isGraphInput.bind(t), r.webnnRegisterGraphOutput = t.registerGraphOutput.bind(t), r.webnnIsGraphOutput = t.isGraphOutput.bind(t), r.webnnCreateTemporaryTensor = t.createTemporaryTensor.bind(t), r.webnnIsGraphInputOutputTypeSupported = t.isGraphInputOutputTypeSupported.bind(t);
  }, re ? r : new Promise((e, t) => {
    C = e, R = t;
  });
}
var Xf;
var Qf;
var rs = k(() => {
  "use strict";
  Xf = es, Qf = globalThis.self?.name?.startsWith("em-pthread");
  Qf && es();
});
var as;
var tn;
var Zf;
var ge;
var ss;
var en;
var Kf;
var ec;
var is;
var tc;
var ns;
var us;
var os;
var fs4;
var Yt = k(() => {
  "use strict";
  Ht();
  as = typeof location > "u" ? void 0 : location.origin, tn = import.meta.url > "file:" && import.meta.url < "file;", Zf = () => {
    if (true) {
      if (tn) {
        let a = URL;
        return new URL(new a("ort.webgpu.bundle.min.mjs", import.meta.url).href, as).href;
      }
      return import.meta.url;
    }
  }, ge = Zf(), ss = () => {
    if (ge && !ge.startsWith("blob:")) return ge.substring(0, ge.lastIndexOf("/") + 1);
  }, en = (a, r) => {
    try {
      let s = r ?? ge;
      return (s ? new URL(a, s) : new URL(a)).origin === as;
    } catch {
      return false;
    }
  }, Kf = (a, r) => {
    let s = r ?? ge;
    try {
      return (s ? new URL(a, s) : new URL(a)).href;
    } catch {
      return;
    }
  }, ec = (a, r) => `${r ?? "./"}${a}`, is = async (a) => {
    let s = await (await fetch(a, { credentials: "same-origin" })).blob();
    return URL.createObjectURL(s);
  }, tc = async (a) => (await import(
    /*webpackIgnore:true*/
    /*@vite-ignore*/
    a
  )).default, ns = (Ka(), $t(Za)).default, us = async () => {
    if (!ge) throw new Error("Failed to load proxy worker: cannot determine the script source URL.");
    if (en(ge)) return [void 0, ns()];
    let a = await is(ge);
    return [a, ns(a)];
  }, os = (rs(), $t(ts)).default, fs4 = async (a, r, s, f) => {
    let i = os && !(a || r);
    if (i) if (ge) i = en(ge) || f && !s;
    else if (f && !s) i = true;
    else throw new Error("cannot determine the script source URL.");
    if (i) return [void 0, os];
    {
      let d = "ort-wasm-simd-threaded.asyncify.mjs", l = a ?? Kf(d, r), m = s && l && !en(l, r), y = m ? await is(l) : l ?? ec(d, r);
      return [m ? y : void 0, await tc(y)];
    }
  };
});
var rn;
var nn;
var rr;
var cs;
var rc;
var nc;
var oc;
var qt;
var z;
var je = k(() => {
  "use strict";
  Yt();
  nn = false, rr = false, cs = false, rc = () => {
    if (typeof SharedArrayBuffer > "u") return false;
    try {
      return typeof MessageChannel < "u" && new MessageChannel().port1.postMessage(new SharedArrayBuffer(1)), WebAssembly.validate(new Uint8Array([0, 97, 115, 109, 1, 0, 0, 0, 1, 4, 1, 96, 0, 0, 3, 2, 1, 0, 5, 4, 1, 3, 1, 1, 10, 11, 1, 9, 0, 65, 0, 254, 16, 2, 0, 26, 11]));
    } catch {
      return false;
    }
  }, nc = () => {
    try {
      return WebAssembly.validate(new Uint8Array([0, 97, 115, 109, 1, 0, 0, 0, 1, 4, 1, 96, 0, 0, 3, 2, 1, 0, 10, 30, 1, 28, 0, 65, 0, 253, 15, 253, 12, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 253, 186, 1, 26, 11]));
    } catch {
      return false;
    }
  }, oc = () => {
    try {
      return WebAssembly.validate(new Uint8Array([0, 97, 115, 109, 1, 0, 0, 0, 1, 5, 1, 96, 0, 1, 123, 3, 2, 1, 0, 10, 19, 1, 17, 0, 65, 1, 253, 15, 65, 2, 253, 15, 65, 3, 253, 15, 253, 147, 2, 11]));
    } catch {
      return false;
    }
  }, qt = async (a) => {
    if (nn) return Promise.resolve();
    if (rr) throw new Error("multiple calls to 'initializeWebAssembly()' detected.");
    if (cs) throw new Error("previous call to 'initializeWebAssembly()' failed.");
    rr = true;
    let r = a.initTimeout, s = a.numThreads;
    if (a.simd !== false) {
      if (a.simd === "relaxed") {
        if (!oc()) throw new Error("Relaxed WebAssembly SIMD is not supported in the current environment.");
      } else if (!nc()) throw new Error("WebAssembly SIMD is not supported in the current environment.");
    }
    let f = rc();
    s > 1 && !f && (typeof self < "u" && !self.crossOriginIsolated && console.warn("env.wasm.numThreads is set to " + s + ", but this will not work unless you enable crossOriginIsolated mode. See https://web.dev/cross-origin-isolation-guide/ for more info."), console.warn("WebAssembly multi-threading is not supported in the current environment. Falling back to single-threading."), a.numThreads = s = 1);
    let i = a.wasmPaths, d = typeof i == "string" ? i : void 0, l = i?.mjs, m = l?.href ?? l, y = i?.wasm, w = y?.href ?? y, T = a.wasmBinary, [g, v] = await fs4(m, d, s > 1, !!T || !!w), S = false, C = [];
    if (r > 0 && C.push(new Promise((R) => {
      setTimeout(() => {
        S = true, R();
      }, r);
    })), C.push(new Promise((R, H) => {
      let U = { numThreads: s };
      if (T) U.wasmBinary = T, U.locateFile = (M) => M;
      else if (w || d) U.locateFile = (M) => w ?? d + M;
      else if (m && m.indexOf("blob:") !== 0) U.locateFile = (M) => new URL(M, m).href;
      else if (g) {
        let M = ss();
        M && (U.locateFile = (Y) => M + Y);
      }
      v(U).then((M) => {
        rr = false, nn = true, rn = M, R(), g && URL.revokeObjectURL(g);
      }, (M) => {
        rr = false, cs = true, H(M);
      });
    })), await Promise.race(C), S) throw new Error(`WebAssembly backend initializing failed due to timeout: ${r}ms`);
  }, z = () => {
    if (nn && rn) return rn;
    throw new Error("WebAssembly is not initialized yet.");
  };
});
var be;
var Lt;
var $;
var nr = k(() => {
  "use strict";
  je();
  be = (a, r) => {
    let s = z(), f = s.lengthBytesUTF8(a) + 1, i = s._malloc(f);
    return s.stringToUTF8(a, i, f), r.push(i), i;
  }, Lt = (a, r, s, f) => {
    if (typeof a == "object" && a !== null) {
      if (s.has(a)) throw new Error("Circular reference in options");
      s.add(a);
    }
    Object.entries(a).forEach(([i, d]) => {
      let l = r ? r + i : i;
      if (typeof d == "object") Lt(d, l + ".", s, f);
      else if (typeof d == "string" || typeof d == "number") f(l, d.toString());
      else if (typeof d == "boolean") f(l, d ? "1" : "0");
      else throw new Error(`Can't handle extra config type: ${typeof d}`);
    });
  }, $ = (a) => {
    let r = z(), s = r.stackSave();
    try {
      let f = r.PTR_SIZE, i = r.stackAlloc(2 * f);
      r._OrtGetLastError(i, i + f);
      let d = Number(r.getValue(i, f === 4 ? "i32" : "i64")), l = r.getValue(i + f, "*"), m = l ? r.UTF8ToString(l) : "";
      throw new Error(`${a} ERROR_CODE: ${d}, ERROR_MESSAGE: ${m}`);
    } finally {
      r.stackRestore(s);
    }
  };
});
var ds;
var ls = k(() => {
  "use strict";
  je();
  nr();
  ds = (a) => {
    let r = z(), s = 0, f = [], i = a || {};
    try {
      if (a?.logSeverityLevel === void 0) i.logSeverityLevel = 2;
      else if (typeof a.logSeverityLevel != "number" || !Number.isInteger(a.logSeverityLevel) || a.logSeverityLevel < 0 || a.logSeverityLevel > 4) throw new Error(`log severity level is not valid: ${a.logSeverityLevel}`);
      if (a?.logVerbosityLevel === void 0) i.logVerbosityLevel = 0;
      else if (typeof a.logVerbosityLevel != "number" || !Number.isInteger(a.logVerbosityLevel)) throw new Error(`log verbosity level is not valid: ${a.logVerbosityLevel}`);
      a?.terminate === void 0 && (i.terminate = false);
      let d = 0;
      return a?.tag !== void 0 && (d = be(a.tag, f)), s = r._OrtCreateRunOptions(i.logSeverityLevel, i.logVerbosityLevel, !!i.terminate, d), s === 0 && $("Can't create run options."), a?.extra !== void 0 && Lt(a.extra, "", /* @__PURE__ */ new WeakSet(), (l, m) => {
        let y = be(l, f), w = be(m, f);
        r._OrtAddRunConfigEntry(s, y, w) !== 0 && $(`Can't set a run config entry: ${l} - ${m}.`);
      }), [s, f];
    } catch (d) {
      throw s !== 0 && r._OrtReleaseRunOptions(s), f.forEach((l) => r._free(l)), d;
    }
  };
});
var ac;
var sc;
var ic;
var on;
var ot;
var uc;
var ps;
var ms = k(() => {
  "use strict";
  je();
  nr();
  ac = (a) => {
    switch (a) {
      case "disabled":
        return 0;
      case "basic":
        return 1;
      case "extended":
        return 2;
      case "layout":
        return 3;
      case "all":
        return 99;
      default:
        throw new Error(`unsupported graph optimization level: ${a}`);
    }
  }, sc = (a) => {
    switch (a) {
      case "sequential":
        return 0;
      case "parallel":
        return 1;
      default:
        throw new Error(`unsupported execution mode: ${a}`);
    }
  }, ic = (a) => {
    a.extra || (a.extra = {}), a.extra.session || (a.extra.session = {});
    let r = a.extra.session;
    r.use_ort_model_bytes_directly || (r.use_ort_model_bytes_directly = "1"), a.executionProviders && a.executionProviders.some((s) => (typeof s == "string" ? s : s.name) === "webgpu") && (a.enableMemPattern = false);
  }, on = (a, r, s, f) => {
    let i = be(r, f), d = be(s, f);
    z()._OrtAddSessionConfigEntry(a, i, d) !== 0 && $(`Can't set a session config entry: ${r} - ${s}.`);
  }, ot = (a, r, s, f) => {
    let i = be(r, f), d = be(s, f);
    a.push([i, d]);
  }, uc = async (a, r, s) => {
    let f = r.executionProviders;
    for (let i of f) {
      let d = typeof i == "string" ? i : i.name, l = [];
      switch (d) {
        case "webnn":
          if (d = "WEBNN", typeof i != "string") {
            let v = i?.deviceType;
            v && on(a, "deviceType", v, s);
          }
          break;
        case "webgpu":
          {
            d = "WebGPU";
            let g;
            if (typeof i != "string") {
              let S = i;
              if (S.device) if (typeof GPUDevice < "u" && S.device instanceof GPUDevice) g = S.device;
              else throw new Error("Invalid GPU device set in WebGPU EP options.");
              let { enableGraphCapture: C } = r;
              if (typeof C == "boolean" && C && ot(l, "enableGraphCapture", "1", s), typeof S.preferredLayout == "string" && ot(l, "preferredLayout", S.preferredLayout, s), S.forceCpuNodeNames) {
                let R = Array.isArray(S.forceCpuNodeNames) ? S.forceCpuNodeNames : [S.forceCpuNodeNames];
                ot(l, "forceCpuNodeNames", R.join(`
`), s);
              }
              S.validationMode && ot(l, "validationMode", S.validationMode, s);
            }
            let v = z().webgpuRegisterDevice(g);
            if (v) {
              let [S, C, R] = v;
              ot(l, "deviceId", S.toString(), s), ot(l, "webgpuInstance", C.toString(), s), ot(l, "webgpuDevice", R.toString(), s);
            }
          }
          break;
        case "wasm":
        case "cpu":
          continue;
        default:
          throw new Error(`not supported execution provider: ${d}`);
      }
      let m = be(d, s), y = l.length, w = 0, T = 0;
      if (y > 0) {
        w = z()._malloc(y * z().PTR_SIZE), s.push(w), T = z()._malloc(y * z().PTR_SIZE), s.push(T);
        for (let g = 0; g < y; g++) z().setValue(w + g * z().PTR_SIZE, l[g][0], "*"), z().setValue(T + g * z().PTR_SIZE, l[g][1], "*");
      }
      await z()._OrtAppendExecutionProvider(a, m, w, T, y) !== 0 && $(`Can't append execution provider: ${d}.`);
    }
  }, ps = async (a) => {
    let r = z(), s = 0, f = [], i = a || {};
    ic(i);
    try {
      let d = ac(i.graphOptimizationLevel ?? "all"), l = sc(i.executionMode ?? "sequential"), m = typeof i.logId == "string" ? be(i.logId, f) : 0, y = i.logSeverityLevel ?? 2;
      if (!Number.isInteger(y) || y < 0 || y > 4) throw new Error(`log severity level is not valid: ${y}`);
      let w = i.logVerbosityLevel ?? 0;
      if (!Number.isInteger(w) || w < 0 || w > 4) throw new Error(`log verbosity level is not valid: ${w}`);
      let T = typeof i.optimizedModelFilePath == "string" ? be(i.optimizedModelFilePath, f) : 0;
      if (s = r._OrtCreateSessionOptions(d, !!i.enableCpuMemArena, !!i.enableMemPattern, l, !!i.enableProfiling, 0, m, y, w, T), s === 0 && $("Can't create session options."), i.executionProviders && await uc(s, i, f), i.enableGraphCapture !== void 0) {
        if (typeof i.enableGraphCapture != "boolean") throw new Error(`enableGraphCapture must be a boolean value: ${i.enableGraphCapture}`);
        on(s, "enableGraphCapture", i.enableGraphCapture.toString(), f);
      }
      if (i.freeDimensionOverrides) for (let [g, v] of Object.entries(i.freeDimensionOverrides)) {
        if (typeof g != "string") throw new Error(`free dimension override name must be a string: ${g}`);
        if (typeof v != "number" || !Number.isInteger(v) || v < 0) throw new Error(`free dimension override value must be a non-negative integer: ${v}`);
        let S = be(g, f);
        r._OrtAddFreeDimensionOverride(s, S, v) !== 0 && $(`Can't set a free dimension override: ${g} - ${v}.`);
      }
      return i.extra !== void 0 && Lt(i.extra, "", /* @__PURE__ */ new WeakSet(), (g, v) => {
        on(s, g, v, f);
      }), [s, f];
    } catch (d) {
      throw s !== 0 && r._OrtReleaseSessionOptions(s) !== 0 && $("Can't release session options."), f.forEach((l) => r._free(l)), d;
    }
  };
});
var He;
var or;
var mt;
var at;
var Ot;
var ar;
var sr;
var an;
var st = k(() => {
  "use strict";
  He = (a) => {
    switch (a) {
      case "int8":
        return 3;
      case "uint8":
        return 2;
      case "bool":
        return 9;
      case "int16":
        return 5;
      case "uint16":
        return 4;
      case "int32":
        return 6;
      case "uint32":
        return 12;
      case "float16":
        return 10;
      case "float32":
        return 1;
      case "float64":
        return 11;
      case "string":
        return 8;
      case "int64":
        return 7;
      case "uint64":
        return 13;
      case "int4":
        return 22;
      case "uint4":
        return 21;
      default:
        throw new Error(`unsupported data type: ${a}`);
    }
  }, or = (a) => {
    switch (a) {
      case 3:
        return "int8";
      case 2:
        return "uint8";
      case 9:
        return "bool";
      case 5:
        return "int16";
      case 4:
        return "uint16";
      case 6:
        return "int32";
      case 12:
        return "uint32";
      case 10:
        return "float16";
      case 1:
        return "float32";
      case 11:
        return "float64";
      case 8:
        return "string";
      case 7:
        return "int64";
      case 13:
        return "uint64";
      case 22:
        return "int4";
      case 21:
        return "uint4";
      default:
        throw new Error(`unsupported data type: ${a}`);
    }
  }, mt = (a, r) => {
    let s = [-1, 4, 1, 1, 2, 2, 4, 8, -1, 1, 2, 8, 4, 8, -1, -1, -1, -1, -1, -1, -1, 0.5, 0.5][a], f = typeof r == "number" ? r : r.reduce((i, d) => i * d, 1);
    return s > 0 ? Math.ceil(f * s) : void 0;
  }, at = (a) => {
    switch (a) {
      case "float16":
        return typeof Float16Array < "u" && Float16Array.from ? Float16Array : Uint16Array;
      case "float32":
        return Float32Array;
      case "uint8":
        return Uint8Array;
      case "int8":
        return Int8Array;
      case "uint16":
        return Uint16Array;
      case "int16":
        return Int16Array;
      case "int32":
        return Int32Array;
      case "bool":
        return Uint8Array;
      case "float64":
        return Float64Array;
      case "uint32":
        return Uint32Array;
      case "int64":
        return BigInt64Array;
      case "uint64":
        return BigUint64Array;
      default:
        throw new Error(`unsupported type: ${a}`);
    }
  }, Ot = (a) => {
    switch (a) {
      case "verbose":
        return 0;
      case "info":
        return 1;
      case "warning":
        return 2;
      case "error":
        return 3;
      case "fatal":
        return 4;
      default:
        throw new Error(`unsupported logging level: ${a}`);
    }
  }, ar = (a) => a === "float32" || a === "float16" || a === "int32" || a === "int64" || a === "uint32" || a === "uint8" || a === "bool" || a === "uint4" || a === "int4", sr = (a) => a === "float32" || a === "float16" || a === "int32" || a === "int64" || a === "uint32" || a === "uint64" || a === "int8" || a === "uint8" || a === "bool" || a === "uint4" || a === "int4", an = (a) => {
    switch (a) {
      case "none":
        return 0;
      case "cpu":
        return 1;
      case "cpu-pinned":
        return 2;
      case "texture":
        return 3;
      case "gpu-buffer":
        return 4;
      case "ml-tensor":
        return 5;
      default:
        throw new Error(`unsupported data location: ${a}`);
    }
  };
});
var Bt;
var sn = k(() => {
  "use strict";
  Ht();
  Bt = async (a) => {
    if (typeof a == "string") if (false) try {
      let { readFile: r } = Hr("node:fs/promises");
      return new Uint8Array(await r(a));
    } catch (r) {
      if (r.code === "ERR_FS_FILE_TOO_LARGE") {
        let { createReadStream: s } = Hr("node:fs"), f = s(a), i = [];
        for await (let d of f) i.push(d);
        return new Uint8Array(Buffer.concat(i));
      }
      throw r;
    }
    else {
      let r = await fetch(a);
      if (!r.ok) throw new Error(`failed to load external data file: ${a}`);
      let s = r.headers.get("Content-Length"), f = s ? parseInt(s, 10) : 0;
      if (f < 1073741824) return new Uint8Array(await r.arrayBuffer());
      {
        if (!r.body) throw new Error(`failed to load external data file: ${a}, no response body.`);
        let i = r.body.getReader(), d;
        try {
          d = new ArrayBuffer(f);
        } catch (m) {
          if (m instanceof RangeError) {
            let y = Math.ceil(f / 65536);
            d = new WebAssembly.Memory({ initial: y, maximum: y }).buffer;
          } else throw m;
        }
        let l = 0;
        for (; ; ) {
          let { done: m, value: y } = await i.read();
          if (m) break;
          let w = y.byteLength;
          new Uint8Array(d, l, w).set(y), l += w;
        }
        return new Uint8Array(d, 0, f);
      }
    }
    else return a instanceof Blob ? new Uint8Array(await a.arrayBuffer()) : a instanceof Uint8Array ? a : new Uint8Array(a);
  };
});
var hs;
var ys = k(() => {
  "use strict";
  st();
  hs = (a, r) => new (at(r))(a);
});
var fc;
var cc;
var bs;
var ws;
var gs;
var dc;
var pe;
var un = k(() => {
  "use strict";
  st();
  fc = ["V", "I", "W", "E", "F"], cc = (a, r) => {
    console.log(`[${fc[a]},${(/* @__PURE__ */ new Date()).toISOString()}]${r}`);
  }, gs = (a, r) => {
    bs = a, ws = r;
  }, dc = (a, r) => {
    let s = Ot(a), f = Ot(bs);
    s >= f && cc(s, typeof r == "function" ? r() : r);
  }, pe = (...a) => {
    ws && dc(...a);
  };
});
var vs;
var cn;
var Es;
var lc;
var Ts;
var pc;
var Ss;
var ir;
var ur;
var fn;
var As;
var Is = k(() => {
  "use strict";
  st();
  un();
  vs = /* @__PURE__ */ new Map([["float32", 32], ["float16", 16], ["int32", 32], ["uint32", 32], ["int64", 64], ["uint64", 64], ["int8", 8], ["uint8", 8], ["int4", 4], ["uint4", 4]]), cn = (a, r) => {
    if (r === "int32") return a;
    let s = vs.get(r);
    if (!s) throw new Error(`WebNN backend does not support data type: ${r}`);
    let f = s / 8;
    if (a.byteLength % f !== 0) throw new Error(`Invalid Uint8Array length - must be a multiple of ${f}.`);
    let i = a.byteLength / f, d = new (at(r))(a.buffer, a.byteOffset, i);
    switch (r) {
      case "int64":
      case "uint64": {
        let l = new Int32Array(i);
        for (let m = 0; m < i; m++) {
          let y = d[m];
          if (y > 2147483647n || y < -2147483648n) throw new Error("Can not convert int64 data to int32 - value out of range.");
          l[m] = Number(y);
        }
        return new Uint8Array(l.buffer);
      }
      case "int8":
      case "uint8":
      case "uint32": {
        if (r === "uint32" && d.some((m) => m > 2147483647)) throw new Error("Can not convert uint32 data to int32 - value out of range.");
        let l = Int32Array.from(d, Number);
        return new Uint8Array(l.buffer);
      }
      default:
        throw new Error(`Unsupported data conversion from ${r} to 'int32'`);
    }
  }, Es = (a, r) => {
    if (r === "int32") return a;
    if (a.byteLength % 4 !== 0) throw new Error("Invalid Uint8Array length - must be a multiple of 4 (int32).");
    let s = a.byteLength / 4, f = new Int32Array(a.buffer, a.byteOffset, s);
    switch (r) {
      case "int64": {
        let i = BigInt64Array.from(f, BigInt);
        return new Uint8Array(i.buffer);
      }
      case "uint64": {
        if (f.some((d) => d < 0)) throw new Error("Can not convert int32 data to uin64 - negative value found.");
        let i = BigUint64Array.from(f, BigInt);
        return new Uint8Array(i.buffer);
      }
      case "int8": {
        if (f.some((d) => d < -128 || d > 127)) throw new Error("Can not convert int32 data to int8 - value out of range.");
        let i = Int8Array.from(f, Number);
        return new Uint8Array(i.buffer);
      }
      case "uint8": {
        if (f.some((i) => i < 0 || i > 255)) throw new Error("Can not convert int32 data to uint8 - value out of range.");
        return Uint8Array.from(f, Number);
      }
      case "uint32": {
        if (f.some((d) => d < 0)) throw new Error("Can not convert int32 data to uint32 - negative value found.");
        let i = Uint32Array.from(f, Number);
        return new Uint8Array(i.buffer);
      }
      default:
        throw new Error(`Unsupported data conversion from 'int32' to ${r}`);
    }
  }, lc = 1, Ts = () => lc++, pc = /* @__PURE__ */ new Map([["int8", "int32"], ["uint8", "int32"], ["uint32", "int32"], ["int64", "int32"]]), Ss = (a, r) => {
    let s = vs.get(a);
    if (!s) throw new Error(`WebNN backend does not support data type: ${a}`);
    return r.length > 0 ? Math.ceil(r.reduce((f, i) => f * i) * s / 8) : 0;
  }, ir = class {
    constructor(r) {
      this.isDataConverted = false;
      let { sessionId: s, context: f, tensor: i, dataType: d, shape: l, fallbackDataType: m } = r;
      this.sessionId = s, this.mlContext = f, this.mlTensor = i, this.dataType = d, this.tensorShape = l, this.fallbackDataType = m;
    }
    get tensor() {
      return this.mlTensor;
    }
    get type() {
      return this.dataType;
    }
    get fallbackType() {
      return this.fallbackDataType;
    }
    get shape() {
      return this.tensorShape;
    }
    get byteLength() {
      return Ss(this.dataType, this.tensorShape);
    }
    destroy() {
      pe("verbose", () => "[WebNN] TensorWrapper.destroy"), this.mlTensor.destroy();
    }
    write(r) {
      this.mlContext.writeTensor(this.mlTensor, r);
    }
    async read(r) {
      if (this.fallbackDataType) {
        let s = await this.mlContext.readTensor(this.mlTensor), f = Es(new Uint8Array(s), this.dataType);
        if (r) {
          (r instanceof ArrayBuffer ? new Uint8Array(r) : new Uint8Array(r.buffer, r.byteOffset, r.byteLength)).set(f);
          return;
        } else return f.buffer;
      } else return r ? this.mlContext.readTensor(this.mlTensor, r) : this.mlContext.readTensor(this.mlTensor);
    }
    canReuseTensor(r, s, f) {
      return this.mlContext === r && this.dataType === s && this.tensorShape.length === f.length && this.tensorShape.every((i, d) => i === f[d]);
    }
    setIsDataConverted(r) {
      this.isDataConverted = r;
    }
  }, ur = class {
    constructor(r, s) {
      this.tensorManager = r;
      this.wrapper = s;
    }
    get tensorWrapper() {
      return this.wrapper;
    }
    releaseTensor() {
      this.tensorWrapper && (this.tensorManager.releaseTensor(this.tensorWrapper), this.wrapper = void 0);
    }
    async ensureTensor(r, s, f, i) {
      let d = this.tensorManager.getMLContext(r), l = this.tensorManager.getMLOpSupportLimits(r), m;
      if (!l?.input.dataTypes.includes(s)) {
        if (m = pc.get(s), !m || l?.input.dataTypes.includes(m)) throw new Error(`WebNN backend does not support data type: ${s}`);
        pe("verbose", () => `[WebNN] TensorIdTracker.ensureTensor: fallback dataType from ${s} to ${m}`);
      }
      if (this.wrapper) {
        if (this.wrapper.canReuseTensor(d, s, f)) return this.wrapper.tensor;
        if (i) {
          if (this.wrapper.byteLength !== Ss(s, f)) throw new Error("Unable to copy data to tensor with different size.");
          this.activeUpload = new Uint8Array(await this.wrapper.read());
        }
        this.tensorManager.releaseTensor(this.wrapper);
      }
      let y = typeof MLTensorUsage > "u" ? void 0 : MLTensorUsage.READ | MLTensorUsage.WRITE;
      return this.wrapper = await this.tensorManager.getCachedTensor(r, s, f, y, true, true, m), i && this.activeUpload && (this.wrapper.write(this.activeUpload), this.activeUpload = void 0), this.wrapper.tensor;
    }
    upload(r) {
      let s = r;
      if (this.wrapper) {
        if (this.wrapper.fallbackType) if (this.wrapper.fallbackType === "int32") s = cn(r, this.wrapper.type), this.wrapper.setIsDataConverted(true);
        else throw new Error(`Unsupported fallback data type: ${this.wrapper.fallbackType}`);
        if (r.byteLength === this.wrapper.byteLength) {
          this.wrapper.write(s);
          return;
        } else pe("verbose", () => "Data size does not match tensor size. Releasing tensor."), this.releaseTensor();
      }
      this.activeUpload ? this.activeUpload.set(s) : this.activeUpload = new Uint8Array(s);
    }
    async download(r) {
      if (this.activeUpload) {
        let s = this.wrapper?.isDataConverted ? Es(this.activeUpload, this.wrapper?.type) : this.activeUpload;
        if (r) {
          r instanceof ArrayBuffer ? new Uint8Array(r).set(s) : new Uint8Array(r.buffer, r.byteOffset, r.byteLength).set(s);
          return;
        } else return s.buffer;
      }
      if (!this.wrapper) throw new Error("Tensor has not been created.");
      return r ? this.wrapper.read(r) : this.wrapper.read();
    }
  }, fn = class {
    constructor(r) {
      this.backend = r;
      this.tensorTrackersById = /* @__PURE__ */ new Map();
      this.freeTensors = [];
      this.externalTensors = /* @__PURE__ */ new Set();
    }
    getMLContext(r) {
      let s = this.backend.getMLContext(r);
      if (!s) throw new Error("MLContext not found for session.");
      return s;
    }
    getMLOpSupportLimits(r) {
      return this.backend.getMLOpSupportLimits(r);
    }
    reserveTensorId() {
      let r = Ts();
      return this.tensorTrackersById.set(r, new ur(this)), r;
    }
    releaseTensorId(r) {
      let s = this.tensorTrackersById.get(r);
      s && (this.tensorTrackersById.delete(r), s.tensorWrapper && this.releaseTensor(s.tensorWrapper));
    }
    async ensureTensor(r, s, f, i, d) {
      pe("verbose", () => `[WebNN] TensorManager.ensureTensor {tensorId: ${s}, dataType: ${f}, shape: ${i}, copyOld: ${d}}`);
      let l = this.tensorTrackersById.get(s);
      if (!l) throw new Error("Tensor not found.");
      return l.ensureTensor(r, f, i, d);
    }
    upload(r, s) {
      let f = this.tensorTrackersById.get(r);
      if (!f) throw new Error("Tensor not found.");
      f.upload(s);
    }
    async download(r, s) {
      pe("verbose", () => `[WebNN] TensorManager.download {tensorId: ${r}, dstBuffer: ${s?.byteLength}}`);
      let f = this.tensorTrackersById.get(r);
      if (!f) throw new Error("Tensor not found.");
      return f.download(s);
    }
    releaseTensorsForSession(r) {
      for (let s of this.freeTensors) s.sessionId === r && s.destroy();
      this.freeTensors = this.freeTensors.filter((s) => s.sessionId !== r);
    }
    registerTensor(r, s, f, i) {
      let d = this.getMLContext(r), l = Ts(), m = new ir({ sessionId: r, context: d, tensor: s, dataType: f, shape: i });
      return this.tensorTrackersById.set(l, new ur(this, m)), this.externalTensors.add(m), l;
    }
    async getCachedTensor(r, s, f, i, d, l, m) {
      let y = this.getMLContext(r);
      for (let [T, g] of this.freeTensors.entries()) if (g.canReuseTensor(y, s, f)) {
        pe("verbose", () => `[WebNN] Reusing tensor {dataType: ${s}, ${m ? `fallbackDataType: ${m},` : ""} shape: ${f}`);
        let v = this.freeTensors.splice(T, 1)[0];
        return v.sessionId = r, v;
      }
      pe("verbose", () => `[WebNN] MLContext.createTensor {dataType: ${s}, ${m ? `fallbackDataType: ${m},` : ""} shape: ${f}}`);
      let w = await y.createTensor({ dataType: m ?? s, shape: f, dimensions: f, usage: i, writable: d, readable: l });
      return new ir({ sessionId: r, context: y, tensor: w, dataType: s, shape: f, fallbackDataType: m });
    }
    releaseTensor(r) {
      this.externalTensors.has(r) && this.externalTensors.delete(r), this.freeTensors.push(r);
    }
  }, As = (...a) => new fn(...a);
});
var xs = {};
At(xs, { WebNNBackend: () => dn });
var fr;
var mc;
var dn;
var Ls = k(() => {
  "use strict";
  st();
  je();
  ys();
  Is();
  un();
  fr = /* @__PURE__ */ new Map([[1, "float32"], [10, "float16"], [6, "int32"], [12, "uint32"], [7, "int64"], [13, "uint64"], [22, "int4"], [21, "uint4"], [3, "int8"], [2, "uint8"], [9, "uint8"]]), mc = (a, r) => {
    if (a === r) return true;
    if (a === void 0 || r === void 0) return false;
    let s = Object.keys(a).sort(), f = Object.keys(r).sort();
    return s.length === f.length && s.every((i, d) => i === f[d] && a[i] === r[i]);
  }, dn = class {
    constructor(r) {
      this.tensorManager = As(this);
      this.mlContextBySessionId = /* @__PURE__ */ new Map();
      this.sessionIdsByMLContext = /* @__PURE__ */ new Map();
      this.mlContextCache = [];
      this.sessionGraphInputs = /* @__PURE__ */ new Map();
      this.sessionGraphOutputs = /* @__PURE__ */ new Map();
      this.temporaryGraphInputs = [];
      this.temporaryGraphOutputs = [];
      this.temporarySessionTensorIds = /* @__PURE__ */ new Map();
      this.mlOpSupportLimitsBySessionId = /* @__PURE__ */ new Map();
      gs(r.logLevel, !!r.debug);
    }
    get currentSessionId() {
      if (this.activeSessionId === void 0) throw new Error("No active session");
      return this.activeSessionId;
    }
    onRunStart(r) {
      pe("verbose", () => `[WebNN] onRunStart {sessionId: ${r}}`), this.activeSessionId = r;
    }
    onRunEnd(r) {
      pe("verbose", () => `[WebNN] onRunEnd {sessionId: ${r}}`);
      let s = this.temporarySessionTensorIds.get(r);
      if (s) {
        for (let f of s) pe("verbose", () => `[WebNN] releasing temporary tensor {tensorId: ${f}}`), this.tensorManager.releaseTensorId(f);
        this.temporarySessionTensorIds.delete(r), this.activeSessionId = void 0;
      }
    }
    async createMLContext(r) {
      if (r instanceof GPUDevice) {
        let f = this.mlContextCache.findIndex((i) => i.gpuDevice === r);
        if (f !== -1) return this.mlContextCache[f].mlContext;
        {
          let i = await navigator.ml.createContext(r);
          return this.mlContextCache.push({ gpuDevice: r, mlContext: i }), i;
        }
      } else if (r === void 0) {
        let f = this.mlContextCache.findIndex((i) => i.options === void 0 && i.gpuDevice === void 0);
        if (f !== -1) return this.mlContextCache[f].mlContext;
        {
          let i = await navigator.ml.createContext();
          return this.mlContextCache.push({ mlContext: i }), i;
        }
      }
      let s = this.mlContextCache.findIndex((f) => mc(f.options, r));
      if (s !== -1) return this.mlContextCache[s].mlContext;
      {
        let f = await navigator.ml.createContext(r);
        return this.mlContextCache.push({ options: r, mlContext: f }), f;
      }
    }
    registerMLContext(r, s) {
      this.mlContextBySessionId.set(r, s);
      let f = this.sessionIdsByMLContext.get(s);
      f || (f = /* @__PURE__ */ new Set(), this.sessionIdsByMLContext.set(s, f)), f.add(r), this.mlOpSupportLimitsBySessionId.has(r) || this.mlOpSupportLimitsBySessionId.set(r, s.opSupportLimits()), this.temporaryGraphInputs.length > 0 && (this.sessionGraphInputs.set(r, this.temporaryGraphInputs), this.temporaryGraphInputs = []), this.temporaryGraphOutputs.length > 0 && (this.sessionGraphOutputs.set(r, this.temporaryGraphOutputs), this.temporaryGraphOutputs = []);
    }
    onReleaseSession(r) {
      this.sessionGraphInputs.delete(r), this.sessionGraphOutputs.delete(r);
      let s = this.mlContextBySessionId.get(r);
      if (!s) return;
      this.tensorManager.releaseTensorsForSession(r), this.mlContextBySessionId.delete(r), this.mlOpSupportLimitsBySessionId.delete(r);
      let f = this.sessionIdsByMLContext.get(s);
      if (f.delete(r), f.size === 0) {
        this.sessionIdsByMLContext.delete(s);
        let i = this.mlContextCache.findIndex((d) => d.mlContext === s);
        i !== -1 && this.mlContextCache.splice(i, 1);
      }
    }
    getMLContext(r) {
      return this.mlContextBySessionId.get(r);
    }
    getMLOpSupportLimits(r) {
      return this.mlOpSupportLimitsBySessionId.get(r);
    }
    reserveTensorId() {
      return this.tensorManager.reserveTensorId();
    }
    releaseTensorId(r) {
      pe("verbose", () => `[WebNN] releaseTensorId {tensorId: ${r}}`), this.tensorManager.releaseTensorId(r);
    }
    async ensureTensor(r, s, f, i, d) {
      let l = fr.get(f);
      if (!l) throw new Error(`Unsupported ONNX data type: ${f}`);
      return this.tensorManager.ensureTensor(r ?? this.currentSessionId, s, l, i, d);
    }
    async createTemporaryTensor(r, s, f) {
      pe("verbose", () => `[WebNN] createTemporaryTensor {onnxDataType: ${s}, shape: ${f}}`);
      let i = fr.get(s);
      if (!i) throw new Error(`Unsupported ONNX data type: ${s}`);
      let d = this.tensorManager.reserveTensorId();
      await this.tensorManager.ensureTensor(r, d, i, f, false);
      let l = this.temporarySessionTensorIds.get(r);
      return l ? l.push(d) : this.temporarySessionTensorIds.set(r, [d]), d;
    }
    uploadTensor(r, s) {
      if (!z().shouldTransferToMLTensor) throw new Error("Trying to upload to a MLTensor while shouldTransferToMLTensor is false");
      pe("verbose", () => `[WebNN] uploadTensor {tensorId: ${r}, data: ${s.byteLength}}`), this.tensorManager.upload(r, s);
    }
    async downloadTensor(r, s) {
      return this.tensorManager.download(r, s);
    }
    createMLTensorDownloader(r, s) {
      return async () => {
        let f = await this.tensorManager.download(r);
        return hs(f, s);
      };
    }
    registerMLTensor(r, s, f, i) {
      let d = fr.get(f);
      if (!d) throw new Error(`Unsupported ONNX data type: ${f}`);
      let l = this.tensorManager.registerTensor(r, s, d, i);
      return pe("verbose", () => `[WebNN] registerMLTensor {tensor: ${s}, dataType: ${d}, dimensions: ${i}} -> {tensorId: ${l}}`), l;
    }
    registerMLConstant(r, s, f, i, d, l, m = false) {
      if (!l) throw new Error("External mounted files are not available.");
      let y = r;
      r.startsWith("./") && (y = r.substring(2));
      let w = l.get(y);
      if (!w) throw new Error(`File with name ${y} not found in preloaded files.`);
      if (s + f > w.byteLength) throw new Error("Out of bounds: data offset and length exceed the external file data size.");
      let T = w.slice(s, s + f).buffer, g;
      switch (d.dataType) {
        case "float32":
          g = new Float32Array(T);
          break;
        case "float16":
          g = typeof Float16Array < "u" && Float16Array.from ? new Float16Array(T) : new Uint16Array(T);
          break;
        case "int32":
          g = new Int32Array(T);
          break;
        case "uint32":
          g = new Uint32Array(T);
          break;
        case "int64":
          if (m) {
            let v = cn(new Uint8Array(T), "int64");
            g = new Int32Array(v.buffer), d.dataType = "int32";
          } else g = new BigInt64Array(T);
          break;
        case "uint64":
          g = new BigUint64Array(T);
          break;
        case "int8":
          g = new Int8Array(T);
          break;
        case "int4":
        case "uint4":
        case "uint8":
          g = new Uint8Array(T);
          break;
        default:
          throw new Error(`Unsupported data type: ${d.dataType} in creating WebNN Constant from external data.`);
      }
      return pe("verbose", () => `[WebNN] registerMLConstant {dataType: ${d.dataType}, shape: ${d.shape}}} ${m ? "(Note: it was int64 data type and registered to int32 as workaround)" : ""}`), i.constant(d, g);
    }
    registerGraphInput(r) {
      this.temporaryGraphInputs.push(r);
    }
    registerGraphOutput(r) {
      this.temporaryGraphOutputs.push(r);
    }
    isGraphInput(r, s) {
      let f = this.sessionGraphInputs.get(r);
      return f ? f.includes(s) : false;
    }
    isGraphOutput(r, s) {
      let f = this.sessionGraphOutputs.get(r);
      return f ? f.includes(s) : false;
    }
    isGraphInputOutputTypeSupported(r, s, f = true) {
      let i = fr.get(He(s)), d = this.mlOpSupportLimitsBySessionId.get(r);
      return typeof i > "u" ? false : f ? !!d?.input.dataTypes.includes(i) : !!d?.output.dataTypes.includes(i);
    }
    flush() {
    }
  };
});
var hc;
var Jt;
var Xt;
var it;
var yc;
var Os;
var xt;
var Qt;
var Zt;
var Bs;
var Kt;
var er;
var tr;
var Kr = k(() => {
  "use strict";
  Ve();
  ls();
  ms();
  st();
  je();
  nr();
  sn();
  hc = (a, r) => {
    z()._OrtInit(a, r) !== 0 && $("Can't initialize onnxruntime.");
  }, Jt = async (a) => {
    hc(a.wasm.numThreads, Ot(a.logLevel));
  }, Xt = async (a, r) => {
    z().asyncInit?.();
    let s = a.webgpu.adapter;
    if (r === "webgpu") {
      if (typeof navigator > "u" || !navigator.gpu) throw new Error("WebGPU is not supported in current environment");
      if (s) {
        if (typeof s.limits != "object" || typeof s.features != "object" || typeof s.requestDevice != "function") throw new Error("Invalid GPU adapter set in `env.webgpu.adapter`. It must be a GPUAdapter object.");
      } else {
        let f = a.webgpu.powerPreference;
        if (f !== void 0 && f !== "low-power" && f !== "high-performance") throw new Error(`Invalid powerPreference setting: "${f}"`);
        let i = a.webgpu.forceFallbackAdapter;
        if (i !== void 0 && typeof i != "boolean") throw new Error(`Invalid forceFallbackAdapter setting: "${i}"`);
        if (s = await navigator.gpu.requestAdapter({ powerPreference: f, forceFallbackAdapter: i }), !s) throw new Error('Failed to get GPU adapter. You may need to enable flag "--enable-unsafe-webgpu" if you are using Chrome.');
      }
    }
    if (r === "webnn" && (typeof navigator > "u" || !navigator.ml)) throw new Error("WebNN is not supported in current environment");
    if (r === "webgpu" && z().webgpuInit((f) => {
      a.webgpu.device = f;
    }), r === "webnn") {
      let f = new (Ls(), $t(xs)).WebNNBackend(a);
      z().webnnInit([f, () => f.reserveTensorId(), (i) => f.releaseTensorId(i), async (i, d, l, m, y) => f.ensureTensor(i, d, l, m, y), (i, d) => {
        f.uploadTensor(i, d);
      }, async (i, d) => f.downloadTensor(i, d), (i, d) => f.registerMLContext(i, d), !!a.trace]);
    }
  }, it = /* @__PURE__ */ new Map(), yc = (a) => {
    let r = z(), s = r.stackSave();
    try {
      let f = r.PTR_SIZE, i = r.stackAlloc(2 * f);
      r._OrtGetInputOutputCount(a, i, i + f) !== 0 && $("Can't get session input/output count.");
      let l = f === 4 ? "i32" : "i64";
      return [Number(r.getValue(i, l)), Number(r.getValue(i + f, l))];
    } finally {
      r.stackRestore(s);
    }
  }, Os = (a, r) => {
    let s = z(), f = s.stackSave(), i = 0;
    try {
      let d = s.PTR_SIZE, l = s.stackAlloc(2 * d);
      s._OrtGetInputOutputMetadata(a, r, l, l + d) !== 0 && $("Can't get session input/output metadata.");
      let y = Number(s.getValue(l, "*"));
      i = Number(s.getValue(l + d, "*"));
      let w = s.HEAP32[i / 4];
      if (w === 0) return [y, 0];
      let T = s.HEAPU32[i / 4 + 1], g = [];
      for (let v = 0; v < T; v++) {
        let S = Number(s.getValue(i + 8 + v * d, "*"));
        g.push(S !== 0 ? s.UTF8ToString(S) : Number(s.getValue(i + 8 + (v + T) * d, "*")));
      }
      return [y, w, g];
    } finally {
      s.stackRestore(f), i !== 0 && s._OrtFree(i);
    }
  }, xt = (a) => {
    let r = z(), s = r._malloc(a.byteLength);
    if (s === 0) throw new Error(`Can't create a session. failed to allocate a buffer of size ${a.byteLength}.`);
    return r.HEAPU8.set(a, s), [s, a.byteLength];
  }, Qt = async (a, r) => {
    let s, f, i = z();
    Array.isArray(a) ? [s, f] = a : a.buffer === i.HEAPU8.buffer ? [s, f] = [a.byteOffset, a.byteLength] : [s, f] = xt(a);
    let d = 0, l = 0, m = 0, y = [], w = [], T = [];
    try {
      if ([l, y] = await ps(r), r?.externalData && i.mountExternalData) {
        let O = [];
        for (let W of r.externalData) {
          let oe = typeof W == "string" ? W : W.path;
          O.push(Bt(typeof W == "string" ? W : W.data).then((p) => {
            i.mountExternalData(oe, p);
          }));
        }
        await Promise.all(O);
      }
      for (let O of r?.executionProviders ?? []) if ((typeof O == "string" ? O : O.name) === "webnn") {
        if (i.shouldTransferToMLTensor = false, typeof O != "string") {
          let oe = O, p = oe?.context, ne = oe?.gpuDevice, X = oe?.deviceType, J = oe?.powerPreference;
          p ? i.currentContext = p : ne ? i.currentContext = await i.webnnCreateMLContext(ne) : i.currentContext = await i.webnnCreateMLContext({ deviceType: X, powerPreference: J });
        } else i.currentContext = await i.webnnCreateMLContext();
        break;
      }
      d = await i._OrtCreateSession(s, f, l), i.webgpuOnCreateSession?.(d), d === 0 && $("Can't create a session."), i.jsepOnCreateSession?.(), i.currentContext && (i.webnnRegisterMLContext(d, i.currentContext), i.currentContext = void 0, i.shouldTransferToMLTensor = true);
      let [g, v] = yc(d), S = !!r?.enableGraphCapture, C = [], R = [], H = [], U = [], M = [];
      for (let O = 0; O < g; O++) {
        let [W, oe, p] = Os(d, O);
        W === 0 && $("Can't get an input name."), w.push(W);
        let ne = i.UTF8ToString(W);
        C.push(ne), H.push(oe === 0 ? { name: ne, isTensor: false } : { name: ne, isTensor: true, type: or(oe), shape: p });
      }
      for (let O = 0; O < v; O++) {
        let [W, oe, p] = Os(d, O + g);
        W === 0 && $("Can't get an output name."), T.push(W);
        let ne = i.UTF8ToString(W);
        R.push(ne), U.push(oe === 0 ? { name: ne, isTensor: false } : { name: ne, isTensor: true, type: or(oe), shape: p });
        {
          if (S && r?.preferredOutputLocation === void 0) {
            M.push("gpu-buffer");
            continue;
          }
          let X = typeof r?.preferredOutputLocation == "string" ? r.preferredOutputLocation : r?.preferredOutputLocation?.[ne] ?? "cpu", J = i.webnnIsGraphOutput;
          if (X === "cpu" && J && J(d, ne)) {
            M.push("ml-tensor-cpu-output");
            continue;
          }
          if (X !== "cpu" && X !== "cpu-pinned" && X !== "gpu-buffer" && X !== "ml-tensor") throw new Error(`Not supported preferred output location: ${X}.`);
          if (S && X !== "gpu-buffer") throw new Error(`Not supported preferred output location: ${X}. Only 'gpu-buffer' location is supported when enableGraphCapture is true.`);
          M.push(X);
        }
      }
      let Y = null;
      return M.some((O) => O === "gpu-buffer" || O === "ml-tensor" || O === "ml-tensor-cpu-output") && (m = i._OrtCreateBinding(d), m === 0 && $("Can't create IO binding."), Y = { handle: m, outputPreferredLocations: M, outputPreferredLocationsEncoded: M.map((O) => O === "ml-tensor-cpu-output" ? "ml-tensor" : O).map((O) => an(O)) }), it.set(d, [d, w, T, Y, S, false]), [d, C, R, H, U];
    } catch (g) {
      throw w.forEach((v) => i._OrtFree(v)), T.forEach((v) => i._OrtFree(v)), m !== 0 && i._OrtReleaseBinding(m) !== 0 && $("Can't release IO binding."), d !== 0 && i._OrtReleaseSession(d) !== 0 && $("Can't release session."), g;
    } finally {
      i._free(s), l !== 0 && i._OrtReleaseSessionOptions(l) !== 0 && $("Can't release session options."), y.forEach((g) => i._free(g)), i.unmountExternalData?.();
    }
  }, Zt = (a) => {
    let r = z(), s = it.get(a);
    if (!s) throw new Error(`cannot release session. invalid session id: ${a}`);
    let [f, i, d, l, m] = s;
    l && (m && r._OrtClearBoundOutputs(l.handle) !== 0 && $("Can't clear bound outputs."), r._OrtReleaseBinding(l.handle) !== 0 && $("Can't release IO binding.")), r.jsepOnReleaseSession?.(a), r.webnnOnReleaseSession?.(a), r.webgpuOnReleaseSession?.(a), i.forEach((y) => r._OrtFree(y)), d.forEach((y) => r._OrtFree(y)), r._OrtReleaseSession(f) !== 0 && $("Can't release session."), it.delete(a);
  }, Bs = async (a, r, s, f, i, d, l = false) => {
    if (!a) {
      r.push(0);
      return;
    }
    let m = z(), y = m.PTR_SIZE, w = a[0], T = a[1], g = a[3], v = g, S, C;
    if (w === "string" && (g === "gpu-buffer" || g === "ml-tensor")) throw new Error("String tensor is not supported on GPU.");
    if (l && g !== "gpu-buffer") throw new Error(`External buffer must be provided for input/output index ${d} when enableGraphCapture is true.`);
    if (g === "gpu-buffer") {
      let U = a[2].gpuBuffer;
      C = mt(He(w), T);
      {
        let M = m.webgpuRegisterBuffer;
        if (!M) throw new Error('Tensor location "gpu-buffer" is not supported without using WebGPU.');
        S = M(U, f);
      }
    } else if (g === "ml-tensor") {
      let U = a[2].mlTensor;
      C = mt(He(w), T);
      let M = m.webnnRegisterMLTensor;
      if (!M) throw new Error('Tensor location "ml-tensor" is not supported without using WebNN.');
      S = M(f, U, He(w), T);
    } else {
      let U = a[2];
      if (Array.isArray(U)) {
        C = y * U.length, S = m._malloc(C), s.push(S);
        for (let M = 0; M < U.length; M++) {
          if (typeof U[M] != "string") throw new TypeError(`tensor data at index ${M} is not a string`);
          m.setValue(S + M * y, be(U[M], s), "*");
        }
      } else {
        let M = m.webnnIsGraphInput, Y = m.webnnIsGraphOutput;
        if (w !== "string" && M && Y) {
          let O = m.UTF8ToString(i);
          if (M(f, O) || Y(f, O)) {
            let W = He(w);
            C = mt(W, T), v = "ml-tensor";
            let oe = m.webnnCreateTemporaryTensor, p = m.webnnUploadTensor;
            if (!oe || !p) throw new Error('Tensor location "ml-tensor" is not supported without using WebNN.');
            let ne = await oe(f, W, T);
            p(ne, new Uint8Array(U.buffer, U.byteOffset, U.byteLength)), S = ne;
          } else C = U.byteLength, S = m._malloc(C), s.push(S), m.HEAPU8.set(new Uint8Array(U.buffer, U.byteOffset, C), S);
        } else C = U.byteLength, S = m._malloc(C), s.push(S), m.HEAPU8.set(new Uint8Array(U.buffer, U.byteOffset, C), S);
      }
    }
    let R = m.stackSave(), H = m.stackAlloc(4 * T.length);
    try {
      T.forEach((M, Y) => m.setValue(H + Y * y, M, y === 4 ? "i32" : "i64"));
      let U = m._OrtCreateTensor(He(w), S, C, H, T.length, an(v));
      U === 0 && $(`Can't create tensor for input/output. session=${f}, index=${d}.`), r.push(U);
    } finally {
      m.stackRestore(R);
    }
  }, Kt = async (a, r, s, f, i, d) => {
    let l = z(), m = l.PTR_SIZE, y = it.get(a);
    if (!y) throw new Error(`cannot run inference. invalid session id: ${a}`);
    let w = y[0], T = y[1], g = y[2], v = y[3], S = y[4], C = y[5], R = r.length, H = f.length, U = 0, M = [], Y = [], O = [], W = [], oe = [], p = l.stackSave(), ne = l.stackAlloc(R * m), X = l.stackAlloc(R * m), J = l.stackAlloc(H * m), Ue = l.stackAlloc(H * m);
    try {
      [U, M] = ds(d), $e("wasm prepareInputOutputTensor");
      for (let _ = 0; _ < R; _++) await Bs(s[_], Y, W, a, T[r[_]], r[_], S);
      for (let _ = 0; _ < H; _++) await Bs(i[_], O, W, a, g[f[_]], R + f[_], S);
      ze("wasm prepareInputOutputTensor");
      for (let _ = 0; _ < R; _++) l.setValue(ne + _ * m, Y[_], "*"), l.setValue(X + _ * m, T[r[_]], "*");
      for (let _ = 0; _ < H; _++) l.setValue(J + _ * m, O[_], "*"), l.setValue(Ue + _ * m, g[f[_]], "*");
      if (v && !C) {
        let { handle: _, outputPreferredLocations: ae, outputPreferredLocationsEncoded: me } = v;
        if (T.length !== R) throw new Error(`input count from feeds (${R}) is expected to be always equal to model's input count (${T.length}).`);
        $e("wasm bindInputsOutputs");
        for (let q = 0; q < R; q++) {
          let we = r[q];
          await l._OrtBindInput(_, T[we], Y[q]) !== 0 && $(`Can't bind input[${q}] for session=${a}.`);
        }
        for (let q = 0; q < H; q++) {
          let we = f[q];
          i[q]?.[3] ? (oe.push(O[q]), l._OrtBindOutput(_, g[we], O[q], 0) !== 0 && $(`Can't bind pre-allocated output[${q}] for session=${a}.`)) : l._OrtBindOutput(_, g[we], 0, me[we]) !== 0 && $(`Can't bind output[${q}] to ${ae[q]} for session=${a}.`);
        }
        ze("wasm bindInputsOutputs"), it.set(a, [w, T, g, v, S, true]);
      }
      l.jsepOnRunStart?.(w), l.webnnOnRunStart?.(w);
      let Q;
      v ? Q = await l._OrtRunWithBinding(w, v.handle, H, J, U) : Q = await l._OrtRun(w, X, ne, R, Ue, H, J, U), Q !== 0 && $("failed to call OrtRun().");
      let x = [], A = [];
      $e("wasm ProcessOutputTensor");
      for (let _ = 0; _ < H; _++) {
        let ae = Number(l.getValue(J + _ * m, "*"));
        if (ae === O[_] || oe.includes(O[_])) {
          x.push(i[_]), ae !== O[_] && l._OrtReleaseTensor(ae) !== 0 && $("Can't release tensor.");
          continue;
        }
        let me = l.stackSave(), q = l.stackAlloc(4 * m), we = false, re, se = 0;
        try {
          l._OrtGetTensorData(ae, q, q + m, q + 2 * m, q + 3 * m) !== 0 && $(`Can't access output tensor data on index ${_}.`);
          let Te = m === 4 ? "i32" : "i64", Ye = Number(l.getValue(q, Te));
          se = l.getValue(q + m, "*");
          let bt = l.getValue(q + m * 2, "*"), wt = Number(l.getValue(q + m * 3, Te)), Se = [];
          for (let ee = 0; ee < wt; ee++) Se.push(Number(l.getValue(bt + ee * m, Te)));
          l._OrtFree(bt) !== 0 && $("Can't free memory for tensor dims.");
          let Ae = Se.reduce((ee, Z) => ee * Z, 1);
          re = or(Ye);
          let Oe = v?.outputPreferredLocations[f[_]];
          if (re === "string") {
            if (Oe === "gpu-buffer" || Oe === "ml-tensor") throw new Error("String tensor is not supported on GPU.");
            let ee = [];
            for (let Z = 0; Z < Ae; Z++) {
              let G = l.getValue(se + Z * m, "*"), V = l.getValue(se + (Z + 1) * m, "*"), qe = Z === Ae - 1 ? void 0 : V - G;
              ee.push(l.UTF8ToString(G, qe));
            }
            x.push([re, Se, ee, "cpu"]);
          } else if (Oe === "gpu-buffer" && Ae > 0) {
            let ee = l.webgpuGetBuffer;
            if (!ee) throw new Error('preferredLocation "gpu-buffer" is not supported without using WebGPU.');
            let Z = ee(se), G = mt(Ye, Ae);
            if (G === void 0 || !ar(re)) throw new Error(`Unsupported data type: ${re}`);
            we = true;
            {
              l.webgpuRegisterBuffer(Z, a, se);
              let V = l.webgpuCreateDownloader(Z, G, a);
              x.push([re, Se, { gpuBuffer: Z, download: async () => {
                let qe = await V();
                return new (at(re))(qe);
              }, dispose: () => {
                l._OrtReleaseTensor(ae) !== 0 && $("Can't release tensor.");
              } }, "gpu-buffer"]);
            }
          } else if (Oe === "ml-tensor" && Ae > 0) {
            let ee = l.webnnEnsureTensor, Z = l.webnnIsGraphInputOutputTypeSupported;
            if (!ee || !Z) throw new Error('preferredLocation "ml-tensor" is not supported without using WebNN.');
            if (mt(Ye, Ae) === void 0 || !sr(re)) throw new Error(`Unsupported data type: ${re}`);
            if (!Z(a, re, false)) throw new Error(`preferredLocation "ml-tensor" for ${re} output is not supported by current WebNN Context.`);
            let V = await ee(a, se, Ye, Se, false);
            we = true, x.push([re, Se, { mlTensor: V, download: l.webnnCreateMLTensorDownloader(se, re), dispose: () => {
              l.webnnReleaseTensorId(se), l._OrtReleaseTensor(ae);
            } }, "ml-tensor"]);
          } else if (Oe === "ml-tensor-cpu-output" && Ae > 0) {
            let ee = l.webnnCreateMLTensorDownloader(se, re)(), Z = x.length;
            we = true, A.push((async () => {
              let G = [Z, await ee];
              return l.webnnReleaseTensorId(se), l._OrtReleaseTensor(ae), G;
            })()), x.push([re, Se, [], "cpu"]);
          } else {
            let ee = at(re), Z = new ee(Ae);
            new Uint8Array(Z.buffer, Z.byteOffset, Z.byteLength).set(l.HEAPU8.subarray(se, se + Z.byteLength)), x.push([re, Se, Z, "cpu"]);
          }
        } finally {
          l.stackRestore(me), re === "string" && se && l._free(se), we || l._OrtReleaseTensor(ae);
        }
      }
      v && !S && (l._OrtClearBoundOutputs(v.handle) !== 0 && $("Can't clear bound outputs."), it.set(a, [w, T, g, v, S, false]));
      for (let [_, ae] of await Promise.all(A)) x[_][2] = ae;
      return ze("wasm ProcessOutputTensor"), x;
    } finally {
      l.webnnOnRunEnd?.(w), l.stackRestore(p), s.forEach((Q) => {
        Q && Q[3] === "gpu-buffer" && l.webgpuUnregisterBuffer(Q[2].gpuBuffer);
      }), i.forEach((Q) => {
        Q && Q[3] === "gpu-buffer" && l.webgpuUnregisterBuffer(Q[2].gpuBuffer);
      }), Y.forEach((Q) => l._OrtReleaseTensor(Q)), O.forEach((Q) => l._OrtReleaseTensor(Q)), W.forEach((Q) => l._free(Q)), U !== 0 && l._OrtReleaseRunOptions(U), M.forEach((Q) => l._free(Q));
    }
  }, er = (a) => {
    let r = z(), s = it.get(a);
    if (!s) throw new Error("invalid session id");
    let f = s[0], i = r._OrtEndProfiling(f);
    i === 0 && $("Can't get an profile file name."), r._OrtFree(i);
  }, tr = (a) => {
    let r = [];
    for (let s of a) {
      let f = s[2];
      !Array.isArray(f) && "buffer" in f && r.push(f.buffer);
    }
    return r;
  };
});
var ut;
var Ee;
var Mt;
var dr;
var lr;
var cr;
var ln;
var pn;
var ht;
var yt;
var wc;
var Ms;
var Cs;
var Us;
var Ds;
var Ps;
var _s;
var Rs;
var mn = k(() => {
  "use strict";
  Ve();
  Kr();
  je();
  Yt();
  ut = () => !!K.wasm.proxy && typeof document < "u", Mt = false, dr = false, lr = false, pn = /* @__PURE__ */ new Map(), ht = (a, r) => {
    let s = pn.get(a);
    s ? s.push(r) : pn.set(a, [r]);
  }, yt = () => {
    if (Mt || !dr || lr || !Ee) throw new Error("worker not ready");
  }, wc = (a) => {
    switch (a.data.type) {
      case "init-wasm":
        Mt = false, a.data.err ? (lr = true, ln[1](a.data.err)) : (dr = true, ln[0]()), cr && (URL.revokeObjectURL(cr), cr = void 0);
        break;
      case "init-ep":
      case "copy-from":
      case "create":
      case "release":
      case "run":
      case "end-profiling": {
        let r = pn.get(a.data.type);
        a.data.err ? r.shift()[1](a.data.err) : r.shift()[0](a.data.out);
        break;
      }
      default:
    }
  }, Ms = async () => {
    if (!dr) {
      if (Mt) throw new Error("multiple calls to 'initWasm()' detected.");
      if (lr) throw new Error("previous call to 'initWasm()' failed.");
      if (Mt = true, ut()) return new Promise((a, r) => {
        Ee?.terminate(), us().then(([s, f]) => {
          try {
            Ee = f, Ee.onerror = (d) => r(d), Ee.onmessage = wc, ln = [a, r];
            let i = { type: "init-wasm", in: K };
            !i.in.wasm.wasmPaths && (s || tn) && (i.in.wasm.wasmPaths = { wasm: new URL("ort-wasm-simd-threaded.asyncify.wasm", import.meta.url).href }), Ee.postMessage(i), cr = s;
          } catch (i) {
            r(i);
          }
        }, r);
      });
      try {
        await qt(K.wasm), await Jt(K), dr = true;
      } catch (a) {
        throw lr = true, a;
      } finally {
        Mt = false;
      }
    }
  }, Cs = async (a) => {
    if (ut()) return yt(), new Promise((r, s) => {
      ht("init-ep", [r, s]);
      let f = { type: "init-ep", in: { epName: a, env: K } };
      Ee.postMessage(f);
    });
    await Xt(K, a);
  }, Us = async (a) => ut() ? (yt(), new Promise((r, s) => {
    ht("copy-from", [r, s]);
    let f = { type: "copy-from", in: { buffer: a } };
    Ee.postMessage(f, [a.buffer]);
  })) : xt(a), Ds = async (a, r) => {
    if (ut()) {
      if (r?.preferredOutputLocation) throw new Error('session option "preferredOutputLocation" is not supported for proxy.');
      return yt(), new Promise((s, f) => {
        ht("create", [s, f]);
        let i = { type: "create", in: { model: a, options: { ...r } } }, d = [];
        a instanceof Uint8Array && d.push(a.buffer), Ee.postMessage(i, d);
      });
    } else return Qt(a, r);
  }, Ps = async (a) => {
    if (ut()) return yt(), new Promise((r, s) => {
      ht("release", [r, s]);
      let f = { type: "release", in: a };
      Ee.postMessage(f);
    });
    Zt(a);
  }, _s = async (a, r, s, f, i, d) => {
    if (ut()) {
      if (s.some((l) => l[3] !== "cpu")) throw new Error("input tensor on GPU is not supported for proxy.");
      if (i.some((l) => l)) throw new Error("pre-allocated output tensor is not supported for proxy.");
      return yt(), new Promise((l, m) => {
        ht("run", [l, m]);
        let y = s, w = { type: "run", in: { sessionId: a, inputIndices: r, inputs: y, outputIndices: f, options: d } };
        Ee.postMessage(w, tr(y));
      });
    } else return Kt(a, r, s, f, i, d);
  }, Rs = async (a) => {
    if (ut()) return yt(), new Promise((r, s) => {
      ht("end-profiling", [r, s]);
      let f = { type: "end-profiling", in: a };
      Ee.postMessage(f);
    });
    er(a);
  };
});
var Ns;
var gc;
var pr;
var ks = k(() => {
  "use strict";
  Ve();
  mn();
  st();
  Ht();
  sn();
  Ns = (a, r) => {
    switch (a.location) {
      case "cpu":
        return [a.type, a.dims, a.data, "cpu"];
      case "gpu-buffer":
        return [a.type, a.dims, { gpuBuffer: a.gpuBuffer }, "gpu-buffer"];
      case "ml-tensor":
        return [a.type, a.dims, { mlTensor: a.mlTensor }, "ml-tensor"];
      default:
        throw new Error(`invalid data location: ${a.location} for ${r()}`);
    }
  }, gc = (a) => {
    switch (a[3]) {
      case "cpu":
        return new Le(a[0], a[2], a[1]);
      case "gpu-buffer": {
        let r = a[0];
        if (!ar(r)) throw new Error(`not supported data type: ${r} for deserializing GPU tensor`);
        let { gpuBuffer: s, download: f, dispose: i } = a[2];
        return Le.fromGpuBuffer(s, { dataType: r, dims: a[1], download: f, dispose: i });
      }
      case "ml-tensor": {
        let r = a[0];
        if (!sr(r)) throw new Error(`not supported data type: ${r} for deserializing MLTensor tensor`);
        let { mlTensor: s, download: f, dispose: i } = a[2];
        return Le.fromMLTensor(s, { dataType: r, dims: a[1], download: f, dispose: i });
      }
      default:
        throw new Error(`invalid data location: ${a[3]}`);
    }
  }, pr = class {
    async fetchModelAndCopyToWasmMemory(r) {
      return Us(await Bt(r));
    }
    async loadModel(r, s) {
      tt();
      let f;
      typeof r == "string" ? f = await this.fetchModelAndCopyToWasmMemory(r) : f = r, [this.sessionId, this.inputNames, this.outputNames, this.inputMetadata, this.outputMetadata] = await Ds(f, s), rt();
    }
    async dispose() {
      return Ps(this.sessionId);
    }
    async run(r, s, f) {
      tt();
      let i = [], d = [];
      Object.entries(r).forEach((v) => {
        let S = v[0], C = v[1], R = this.inputNames.indexOf(S);
        if (R === -1) throw new Error(`invalid input '${S}'`);
        i.push(C), d.push(R);
      });
      let l = [], m = [];
      Object.entries(s).forEach((v) => {
        let S = v[0], C = v[1], R = this.outputNames.indexOf(S);
        if (R === -1) throw new Error(`invalid output '${S}'`);
        l.push(C), m.push(R);
      });
      let y = i.map((v, S) => Ns(v, () => `input "${this.inputNames[d[S]]}"`)), w = l.map((v, S) => v ? Ns(v, () => `output "${this.outputNames[m[S]]}"`) : null), T = await _s(this.sessionId, d, y, m, w, f), g = {};
      for (let v = 0; v < T.length; v++) g[this.outputNames[m[v]]] = l[v] ?? gc(T[v]);
      return rt(), g;
    }
    startProfiling() {
    }
    endProfiling() {
      Rs(this.sessionId);
    }
  };
});
var Fs = {};
At(Fs, { OnnxruntimeWebAssemblyBackend: () => mr, initializeFlags: () => Ws, wasmBackend: () => Tc });
var Ws;
var mr;
var Tc;
var Gs = k(() => {
  "use strict";
  Ve();
  mn();
  ks();
  Ws = () => {
    (typeof K.wasm.initTimeout != "number" || K.wasm.initTimeout < 0) && (K.wasm.initTimeout = 0);
    let a = K.wasm.simd;
    if (typeof a != "boolean" && a !== void 0 && a !== "fixed" && a !== "relaxed" && (console.warn(`Property "env.wasm.simd" is set to unknown value "${a}". Reset it to \`false\` and ignore SIMD feature checking.`), K.wasm.simd = false), typeof K.wasm.proxy != "boolean" && (K.wasm.proxy = false), typeof K.wasm.trace != "boolean" && (K.wasm.trace = false), typeof K.wasm.numThreads != "number" || !Number.isInteger(K.wasm.numThreads) || K.wasm.numThreads <= 0) if (typeof self < "u" && !self.crossOriginIsolated) K.wasm.numThreads = 1;
    else {
      let r = typeof navigator > "u" ? Hr("node:os").cpus().length : navigator.hardwareConcurrency;
      K.wasm.numThreads = Math.min(4, Math.ceil((r || 1) / 2));
    }
  }, mr = class {
    async init(r) {
      Ws(), await Ms(), await Cs(r);
    }
    async createInferenceSessionHandler(r, s) {
      let f = new pr();
      return await f.loadModel(r, s), f;
    }
  }, Tc = new mr();
});
Ve();
Ve();
Ve();
var Ja = "1.26.0-dev.20260416-b7804b056c";
var gl = Zr;
{
  let a = (Gs(), $t(Fs)).wasmBackend;
  Ke("webgpu", a, 5), Ke("webnn", a, 5), Ke("cpu", a, 10), Ke("wasm", a, 10);
}
Object.defineProperty(K.versions, "web", { value: Ja, enumerable: true });

// src/backends/utils/cacheWasm.js
async function loadAndCacheFile(url2) {
  const fileName = url2.split("/").pop();
  let cache2;
  try {
    cache2 = await getCache();
    if (cache2) {
      const result = await cache2.match(url2);
      if (result) {
        return result;
      }
    }
  } catch (error) {
    logger.warn(`Failed to load ${fileName} from cache:`, error);
  }
  const response = await env.fetch(url2);
  if (!response.ok) {
    throw new Error(`Failed to fetch ${fileName}: ${response.status} ${response.statusText}`);
  }
  if (cache2) {
    try {
      await cache2.put(url2, response.clone());
    } catch (e) {
      logger.warn(`Failed to cache ${fileName}:`, e);
    }
  }
  return response;
}
async function loadWasmBinary(wasmURL) {
  const response = await loadAndCacheFile(wasmURL);
  if (!response || typeof response === "string") return null;
  try {
    return await response.arrayBuffer();
  } catch (error) {
    logger.warn("Failed to read WASM binary:", error);
    return null;
  }
}
async function loadWasmFactory(libURL) {
  if (apis.IS_SERVICE_WORKER_ENV || apis.IS_CHROME_AVAILABLE) {
    return libURL;
  }
  const response = await loadAndCacheFile(libURL);
  if (!response || typeof response === "string") return null;
  try {
    let code = await response.text();
    code = code.replaceAll("globalThis.process?.versions?.node", "false");
    const blob = new Blob([code], { type: "text/javascript" });
    return URL.createObjectURL(blob);
  } catch (error) {
    logger.warn("Failed to read WASM factory:", error);
    return null;
  }
}

// src/backends/onnx.js
import { Tensor } from "onnxruntime-common";
var DEVICE_TO_EXECUTION_PROVIDER_MAPPING = Object.freeze({
  auto: null,
  // Auto-detect based on device and environment
  gpu: null,
  // Auto-detect GPU
  cpu: "cpu",
  // CPU
  wasm: "wasm",
  // WebAssembly
  webgpu: "webgpu",
  // WebGPU
  cuda: "cuda",
  // CUDA
  dml: "dml",
  // DirectML
  coreml: "coreml",
  // CoreML
  webnn: { name: "webnn", deviceType: "cpu" },
  // WebNN (default)
  "webnn-npu": { name: "webnn", deviceType: "npu" },
  // WebNN NPU
  "webnn-gpu": { name: "webnn", deviceType: "gpu" },
  // WebNN GPU
  "webnn-cpu": { name: "webnn", deviceType: "cpu" }
  // WebNN CPU
});
function getOnnxLogSeverityLevel(logLevel2) {
  if (logLevel2 <= LogLevel.DEBUG) {
    return 0;
  } else if (logLevel2 <= LogLevel.INFO) {
    return 2;
  } else if (logLevel2 <= LogLevel.WARNING) {
    return 3;
  } else if (logLevel2 <= LogLevel.ERROR) {
    return 3;
  } else {
    return 4;
  }
}
var ONNX_LOG_LEVEL_NAMES = {
  0: "verbose",
  1: "info",
  2: "warning",
  3: "error",
  4: "fatal"
};
var supportedDevices = [];
var defaultDevices;
var ONNX;
var ORT_SYMBOL = /* @__PURE__ */ Symbol.for("onnxruntime");
if (ORT_SYMBOL in globalThis) {
  ONNX = globalThis[ORT_SYMBOL];
} else if (apis.IS_NODE_ENV) {
  ONNX = ONNX_NODE;
  switch (process.platform) {
    case "win32":
      supportedDevices.push("dml");
      break;
    case "linux":
      if (process.arch === "x64") {
        supportedDevices.push("cuda");
      }
      break;
    case "darwin":
      supportedDevices.push("coreml");
      break;
  }
  supportedDevices.push("webgpu");
  supportedDevices.push("cpu");
  defaultDevices = ["cpu"];
} else {
  ONNX = ort_webgpu_bundle_min_exports;
  if (apis.IS_WEBNN_AVAILABLE) {
    supportedDevices.push("webnn-npu", "webnn-gpu", "webnn-cpu", "webnn");
  }
  if (apis.IS_WEBGPU_AVAILABLE) {
    supportedDevices.push("webgpu");
  }
  supportedDevices.push("wasm");
  defaultDevices = ["wasm"];
}
var InferenceSession = ONNX.InferenceSession;
function deviceToExecutionProviders(device = null) {
  if (!device) return defaultDevices;
  switch (device) {
    case "auto":
      return supportedDevices;
    case "gpu":
      return supportedDevices.filter((x) => ["webgpu", "cuda", "dml", "webnn-gpu"].includes(x));
  }
  if (supportedDevices.includes(device)) {
    return [DEVICE_TO_EXECUTION_PROVIDER_MAPPING[device] ?? device];
  }
  throw new Error(`Unsupported device: "${device}". Should be one of: ${supportedDevices.join(", ")}.`);
}
var webInitChain = Promise.resolve();
var wasmLoadPromise = null;
async function ensureWasmLoaded() {
  if (wasmLoadPromise) {
    return wasmLoadPromise;
  }
  const shouldUseWasmCache = env.useWasmCache && typeof ONNX_ENV?.wasm?.wasmPaths === "object" && ONNX_ENV?.wasm?.wasmPaths?.wasm && ONNX_ENV?.wasm?.wasmPaths?.mjs;
  if (!shouldUseWasmCache) {
    if (apis.IS_DENO_WEB_RUNTIME) {
      throw new Error(
        "env.useWasmCache=false is not supported in Deno's web runtime. Remove the useWasmCache override."
      );
    }
    wasmLoadPromise = Promise.resolve();
    return wasmLoadPromise;
  }
  wasmLoadPromise = (async () => {
    const urls = (
      /** @type {{ wasm: string, mjs: string }} */
      ONNX_ENV.wasm.wasmPaths
    );
    let wasmBinaryLoaded = false;
    await Promise.all([
      // Load and cache the WASM binary
      urls.wasm && !isBlobURL(urls.wasm) ? (async () => {
        try {
          const wasmBinary = await loadWasmBinary(toAbsoluteURL(urls.wasm));
          if (wasmBinary) {
            ONNX_ENV.wasm.wasmBinary = wasmBinary;
            wasmBinaryLoaded = true;
          }
        } catch (err) {
          logger.warn("Failed to pre-load WASM binary:", err);
        }
      })() : Promise.resolve(),
      // Load and cache the WASM factory as a blob URL
      urls.mjs && !isBlobURL(urls.mjs) ? (async () => {
        try {
          const wasmFactoryBlob = await loadWasmFactory(toAbsoluteURL(urls.mjs));
          if (wasmFactoryBlob) {
            ONNX_ENV.wasm.wasmPaths.mjs = wasmFactoryBlob;
          }
        } catch (err) {
          logger.warn("Failed to pre-load WASM factory:", err);
        }
      })() : Promise.resolve()
    ]);
    if (!wasmBinaryLoaded) {
      ONNX_ENV.wasm.wasmPaths.mjs = urls.mjs;
    }
  })();
  return wasmLoadPromise;
}
async function createInferenceSession(buffer_or_path, session_options, session_config) {
  await ensureWasmLoaded();
  const logSeverityLevel = getOnnxLogSeverityLevel(env.logLevel ?? LogLevel.WARNING);
  const load = () => InferenceSession.create(buffer_or_path, {
    // Set default log severity level, but allow overriding through session options
    logSeverityLevel,
    ...session_options
  });
  const session = await (apis.IS_WEB_ENV ? webInitChain = webInitChain.then(load) : load());
  session.config = session_config;
  return session;
}
var webInferenceChain = Promise.resolve();
async function runInferenceSession(session, ortFeed) {
  const run = () => session.run(ortFeed);
  return apis.IS_WEB_ENV ? webInferenceChain = webInferenceChain.then(run) : run();
}
function isONNXTensor(x) {
  return x instanceof ONNX.Tensor;
}
var ONNX_ENV = ONNX?.env;
function isONNXProxy() {
  return ONNX_ENV?.wasm?.proxy;
}
if (ONNX_ENV) {
  let setLogLevel = function(logLevel2) {
    const severityLevel = getOnnxLogSeverityLevel(logLevel2);
    ONNX_ENV.logLevel = ONNX_LOG_LEVEL_NAMES[severityLevel];
  };
  if (ONNX_ENV.wasm) {
    if (
      // @ts-ignore Cannot find name 'ServiceWorkerGlobalScope'.ts(2304)
      !(typeof ServiceWorkerGlobalScope !== "undefined" && self instanceof ServiceWorkerGlobalScope) && ONNX_ENV.versions?.web && !ONNX_ENV.wasm.wasmPaths
    ) {
      const wasmPathPrefix = `https://cdn.jsdelivr.net/npm/onnxruntime-web@${ONNX_ENV.versions.web}/dist/`;
      ONNX_ENV.wasm.wasmPaths = apis.IS_SAFARI ? {
        mjs: `${wasmPathPrefix}ort-wasm-simd-threaded.mjs`,
        wasm: `${wasmPathPrefix}ort-wasm-simd-threaded.wasm`
      } : {
        mjs: `${wasmPathPrefix}ort-wasm-simd-threaded.asyncify.mjs`,
        wasm: `${wasmPathPrefix}ort-wasm-simd-threaded.asyncify.wasm`
      };
    }
    ONNX_ENV.wasm.proxy = false;
  }
  if (ONNX_ENV.webgpu) {
    ONNX_ENV.webgpu.powerPreference = "high-performance";
  }
  setLogLevel(env.logLevel ?? LogLevel.WARNING);
  env.backends.onnx = {
    ...ONNX_ENV,
    setLogLevel
  };
}

// src/ops/registry.js
var wrap = async (session_bytes, session_options, names) => {
  const session = await createInferenceSession(new Uint8Array(session_bytes), session_options);
  return (
    /** @type {any} */
    (async (inputs) => {
      const proxied = isONNXProxy();
      const ortFeed = Object.fromEntries(
        Object.entries(inputs).map(([k2, v]) => [k2, (proxied ? v.clone() : v).ort_tensor])
      );
      const outputs = await runInferenceSession(session, ortFeed);
      if (Array.isArray(names)) {
        return names.map((n) => new Tensor2(outputs[n]));
      } else {
        return new Tensor2(outputs[
          /** @type {string} */
          names
        ]);
      }
    })
  );
};
var TensorOpRegistry = class {
  static session_options = {
    // TODO: Allow for multiple execution providers
    // executionProviders: ['webgpu'],
  };
  static get nearest_interpolate_4d() {
    if (!this._nearest_interpolate_4d) {
      this._nearest_interpolate_4d = wrap(
        [
          8,
          10,
          18,
          0,
          58,
          129,
          1,
          10,
          41,
          10,
          1,
          120,
          10,
          0,
          10,
          0,
          10,
          1,
          115,
          18,
          1,
          121,
          34,
          6,
          82,
          101,
          115,
          105,
          122,
          101,
          42,
          18,
          10,
          4,
          109,
          111,
          100,
          101,
          34,
          7,
          110,
          101,
          97,
          114,
          101,
          115,
          116,
          160,
          1,
          3,
          18,
          1,
          114,
          90,
          31,
          10,
          1,
          120,
          18,
          26,
          10,
          24,
          8,
          1,
          18,
          20,
          10,
          3,
          18,
          1,
          98,
          10,
          3,
          18,
          1,
          99,
          10,
          3,
          18,
          1,
          104,
          10,
          3,
          18,
          1,
          119,
          90,
          15,
          10,
          1,
          115,
          18,
          10,
          10,
          8,
          8,
          7,
          18,
          4,
          10,
          2,
          8,
          4,
          98,
          31,
          10,
          1,
          121,
          18,
          26,
          10,
          24,
          8,
          1,
          18,
          20,
          10,
          3,
          18,
          1,
          98,
          10,
          3,
          18,
          1,
          99,
          10,
          3,
          18,
          1,
          104,
          10,
          3,
          18,
          1,
          119,
          66,
          2,
          16,
          21
        ],
        this.session_options,
        "y"
      );
    }
    return this._nearest_interpolate_4d;
  }
  static get bilinear_interpolate_4d() {
    if (!this._bilinear_interpolate_4d) {
      this._bilinear_interpolate_4d = wrap(
        [
          8,
          9,
          18,
          0,
          58,
          128,
          1,
          10,
          40,
          10,
          1,
          120,
          10,
          0,
          10,
          0,
          10,
          1,
          115,
          18,
          1,
          121,
          34,
          6,
          82,
          101,
          115,
          105,
          122,
          101,
          42,
          17,
          10,
          4,
          109,
          111,
          100,
          101,
          34,
          6,
          108,
          105,
          110,
          101,
          97,
          114,
          160,
          1,
          3,
          18,
          1,
          114,
          90,
          31,
          10,
          1,
          120,
          18,
          26,
          10,
          24,
          8,
          1,
          18,
          20,
          10,
          3,
          18,
          1,
          98,
          10,
          3,
          18,
          1,
          99,
          10,
          3,
          18,
          1,
          104,
          10,
          3,
          18,
          1,
          119,
          90,
          15,
          10,
          1,
          115,
          18,
          10,
          10,
          8,
          8,
          7,
          18,
          4,
          10,
          2,
          8,
          4,
          98,
          31,
          10,
          1,
          121,
          18,
          26,
          10,
          24,
          8,
          1,
          18,
          20,
          10,
          3,
          18,
          1,
          98,
          10,
          3,
          18,
          1,
          99,
          10,
          3,
          18,
          1,
          104,
          10,
          3,
          18,
          1,
          119,
          66,
          2,
          16,
          20
        ],
        this.session_options,
        "y"
      );
    }
    return this._bilinear_interpolate_4d;
  }
  static get bicubic_interpolate_4d() {
    if (!this._bicubic_interpolate_4d) {
      this._bicubic_interpolate_4d = wrap(
        [
          8,
          9,
          18,
          0,
          58,
          127,
          10,
          39,
          10,
          1,
          120,
          10,
          0,
          10,
          0,
          10,
          1,
          115,
          18,
          1,
          121,
          34,
          6,
          82,
          101,
          115,
          105,
          122,
          101,
          42,
          16,
          10,
          4,
          109,
          111,
          100,
          101,
          34,
          5,
          99,
          117,
          98,
          105,
          99,
          160,
          1,
          3,
          18,
          1,
          114,
          90,
          31,
          10,
          1,
          120,
          18,
          26,
          10,
          24,
          8,
          1,
          18,
          20,
          10,
          3,
          18,
          1,
          98,
          10,
          3,
          18,
          1,
          99,
          10,
          3,
          18,
          1,
          104,
          10,
          3,
          18,
          1,
          119,
          90,
          15,
          10,
          1,
          115,
          18,
          10,
          10,
          8,
          8,
          7,
          18,
          4,
          10,
          2,
          8,
          4,
          98,
          31,
          10,
          1,
          121,
          18,
          26,
          10,
          24,
          8,
          1,
          18,
          20,
          10,
          3,
          18,
          1,
          98,
          10,
          3,
          18,
          1,
          99,
          10,
          3,
          18,
          1,
          104,
          10,
          3,
          18,
          1,
          119,
          66,
          2,
          16,
          20
        ],
        this.session_options,
        "y"
      );
    }
    return this._bicubic_interpolate_4d;
  }
  static get matmul() {
    if (!this._matmul) {
      this._matmul = wrap(
        [
          8,
          9,
          18,
          0,
          58,
          55,
          10,
          17,
          10,
          1,
          97,
          10,
          1,
          98,
          18,
          1,
          99,
          34,
          6,
          77,
          97,
          116,
          77,
          117,
          108,
          18,
          1,
          114,
          90,
          9,
          10,
          1,
          97,
          18,
          4,
          10,
          2,
          8,
          1,
          90,
          9,
          10,
          1,
          98,
          18,
          4,
          10,
          2,
          8,
          1,
          98,
          9,
          10,
          1,
          99,
          18,
          4,
          10,
          2,
          8,
          1,
          66,
          2,
          16,
          20
        ],
        this.session_options,
        "c"
      );
    }
    return this._matmul;
  }
  static get stft() {
    if (!this._stft) {
      this._stft = wrap(
        [
          8,
          7,
          18,
          0,
          58,
          148,
          1,
          10,
          38,
          10,
          1,
          115,
          10,
          1,
          106,
          10,
          1,
          119,
          10,
          1,
          108,
          18,
          1,
          111,
          34,
          4,
          83,
          84,
          70,
          84,
          42,
          15,
          10,
          8,
          111,
          110,
          101,
          115,
          105,
          100,
          101,
          100,
          24,
          1,
          160,
          1,
          2,
          18,
          1,
          115,
          90,
          26,
          10,
          1,
          115,
          18,
          21,
          10,
          19,
          8,
          1,
          18,
          15,
          10,
          3,
          18,
          1,
          98,
          10,
          3,
          18,
          1,
          115,
          10,
          3,
          18,
          1,
          99,
          90,
          11,
          10,
          1,
          106,
          18,
          6,
          10,
          4,
          8,
          7,
          18,
          0,
          90,
          16,
          10,
          1,
          119,
          18,
          11,
          10,
          9,
          8,
          1,
          18,
          5,
          10,
          3,
          18,
          1,
          119,
          90,
          11,
          10,
          1,
          108,
          18,
          6,
          10,
          4,
          8,
          7,
          18,
          0,
          98,
          31,
          10,
          1,
          111,
          18,
          26,
          10,
          24,
          8,
          1,
          18,
          20,
          10,
          3,
          18,
          1,
          98,
          10,
          3,
          18,
          1,
          102,
          10,
          3,
          18,
          1,
          100,
          10,
          3,
          18,
          1,
          99,
          66,
          2,
          16,
          17
        ],
        this.session_options,
        "o"
      );
    }
    return this._stft;
  }
  static get rfft() {
    if (!this._rfft) {
      this._rfft = wrap(
        [
          8,
          9,
          18,
          0,
          58,
          97,
          10,
          33,
          10,
          1,
          120,
          10,
          0,
          10,
          1,
          97,
          18,
          1,
          121,
          34,
          3,
          68,
          70,
          84,
          42,
          15,
          10,
          8,
          111,
          110,
          101,
          115,
          105,
          100,
          101,
          100,
          24,
          1,
          160,
          1,
          2,
          18,
          1,
          100,
          90,
          21,
          10,
          1,
          120,
          18,
          16,
          10,
          14,
          8,
          1,
          18,
          10,
          10,
          3,
          18,
          1,
          115,
          10,
          3,
          18,
          1,
          99,
          90,
          11,
          10,
          1,
          97,
          18,
          6,
          10,
          4,
          8,
          7,
          18,
          0,
          98,
          21,
          10,
          1,
          121,
          18,
          16,
          10,
          14,
          8,
          1,
          18,
          10,
          10,
          3,
          18,
          1,
          115,
          10,
          3,
          18,
          1,
          99,
          66,
          2,
          16,
          20
        ],
        this.session_options,
        "y"
      );
    }
    return this._rfft;
  }
  static get top_k() {
    if (!this._top_k) {
      this._top_k = wrap(
        [
          8,
          10,
          18,
          0,
          58,
          73,
          10,
          18,
          10,
          1,
          120,
          10,
          1,
          107,
          18,
          1,
          118,
          18,
          1,
          105,
          34,
          4,
          84,
          111,
          112,
          75,
          18,
          1,
          116,
          90,
          9,
          10,
          1,
          120,
          18,
          4,
          10,
          2,
          8,
          1,
          90,
          15,
          10,
          1,
          107,
          18,
          10,
          10,
          8,
          8,
          7,
          18,
          4,
          10,
          2,
          8,
          1,
          98,
          9,
          10,
          1,
          118,
          18,
          4,
          10,
          2,
          8,
          1,
          98,
          9,
          10,
          1,
          105,
          18,
          4,
          10,
          2,
          8,
          7,
          66,
          2,
          16,
          21
        ],
        this.session_options,
        [
          /* Values */
          "v",
          /* Indices */
          "i"
        ]
      );
    }
    return this._top_k;
  }
  static get slice() {
    if (!this._slice) {
      this._slice = wrap(
        [
          8,
          7,
          18,
          0,
          58,
          96,
          10,
          25,
          10,
          1,
          120,
          10,
          1,
          115,
          10,
          1,
          101,
          10,
          1,
          97,
          10,
          1,
          116,
          18,
          1,
          121,
          34,
          5,
          83,
          108,
          105,
          99,
          101,
          18,
          1,
          114,
          90,
          9,
          10,
          1,
          120,
          18,
          4,
          10,
          2,
          8,
          1,
          90,
          9,
          10,
          1,
          115,
          18,
          4,
          10,
          2,
          8,
          7,
          90,
          9,
          10,
          1,
          101,
          18,
          4,
          10,
          2,
          8,
          7,
          90,
          9,
          10,
          1,
          97,
          18,
          4,
          10,
          2,
          8,
          7,
          90,
          9,
          10,
          1,
          116,
          18,
          4,
          10,
          2,
          8,
          7,
          98,
          9,
          10,
          1,
          121,
          18,
          4,
          10,
          2,
          8,
          1,
          66,
          2,
          16,
          13
        ],
        this.session_options,
        "y"
      );
    }
    return this._slice;
  }
};

// src/utils/devices.js
var DEVICE_TYPES = Object.freeze({
  auto: "auto",
  // Auto-detect based on device and environment
  gpu: "gpu",
  // Auto-detect GPU
  cpu: "cpu",
  // CPU
  wasm: "wasm",
  // WebAssembly
  webgpu: "webgpu",
  // WebGPU
  cuda: "cuda",
  // CUDA
  dml: "dml",
  // DirectML
  coreml: "coreml",
  // CoreML
  webnn: "webnn",
  // WebNN (default)
  "webnn-npu": "webnn-npu",
  // WebNN NPU
  "webnn-gpu": "webnn-gpu",
  // WebNN GPU
  "webnn-cpu": "webnn-cpu"
  // WebNN CPU
});
var DEFAULT_DEVICE = apis.IS_NODE_ENV ? "cpu" : "wasm";
function selectDevice(deviceConfig, fileName, { warn } = {}) {
  if (!deviceConfig) return DEFAULT_DEVICE;
  if (typeof deviceConfig === "string") return deviceConfig;
  if (deviceConfig.hasOwnProperty(fileName)) return deviceConfig[fileName];
  if (warn) {
    warn(`device not specified for "${fileName}". Using the default device (${DEFAULT_DEVICE}).`);
  }
  return DEFAULT_DEVICE;
}

// src/utils/dtypes.js
var isWebGpuFp16Supported = /* @__PURE__ */ (function() {
  let cachedResult;
  return async function() {
    if (cachedResult === void 0) {
      if (!apis.IS_WEBGPU_AVAILABLE) {
        cachedResult = false;
      } else {
        try {
          const adapter = await navigator.gpu.requestAdapter();
          cachedResult = adapter.features.has("shader-f16");
        } catch (e) {
          cachedResult = false;
        }
      }
    }
    return cachedResult;
  };
})();
var DATA_TYPES = Object.freeze({
  auto: "auto",
  // Auto-detect based on environment
  fp32: "fp32",
  fp16: "fp16",
  q8: "q8",
  int8: "int8",
  uint8: "uint8",
  q4: "q4",
  bnb4: "bnb4",
  q4f16: "q4f16",
  // fp16 model with 4-bit block weight quantization
  q2: "q2",
  q2f16: "q2f16",
  // fp16 model with 2-bit block weight quantization
  q1: "q1",
  q1f16: "q1f16"
  // fp16 model with 1-bit block weight quantization
});
var DEFAULT_DEVICE_DTYPE = DATA_TYPES.fp32;
var DEFAULT_DEVICE_DTYPE_MAPPING = Object.freeze({
  // NOTE: If not specified, will default to fp32
  [DEVICE_TYPES.wasm]: DATA_TYPES.q8
});
var DEFAULT_DTYPE_SUFFIX_MAPPING = Object.freeze({
  [DATA_TYPES.fp32]: "",
  [DATA_TYPES.fp16]: "_fp16",
  [DATA_TYPES.int8]: "_int8",
  [DATA_TYPES.uint8]: "_uint8",
  [DATA_TYPES.q8]: "_quantized",
  [DATA_TYPES.q4]: "_q4",
  [DATA_TYPES.q2]: "_q2",
  [DATA_TYPES.q1]: "_q1",
  [DATA_TYPES.q4f16]: "_q4f16",
  [DATA_TYPES.q2f16]: "_q2f16",
  [DATA_TYPES.q1f16]: "_q1f16",
  [DATA_TYPES.bnb4]: "_bnb4"
});
function selectDtype(dtype, fileName, selectedDevice, { configDtype = null, warn } = {}) {
  let resolved;
  let needsWarn = false;
  if (dtype && typeof dtype !== "string") {
    if (dtype.hasOwnProperty(fileName)) {
      resolved = dtype[fileName];
    } else {
      resolved = null;
      needsWarn = true;
    }
  } else {
    resolved = /** @type {string|null|undefined} */
    dtype;
  }
  let result;
  if (resolved === DATA_TYPES.auto) {
    if (configDtype) {
      const fallback = typeof configDtype === "string" ? configDtype : configDtype?.[fileName];
      if (fallback && fallback !== DATA_TYPES.auto && DATA_TYPES.hasOwnProperty(fallback)) {
        return (
          /** @type {DataType} */
          fallback
        );
      }
    }
    result = DEFAULT_DEVICE_DTYPE_MAPPING[selectedDevice] ?? DEFAULT_DEVICE_DTYPE;
  } else if (resolved && DATA_TYPES.hasOwnProperty(resolved)) {
    result = /** @type {DataType} */
    resolved;
  } else {
    result = DEFAULT_DEVICE_DTYPE_MAPPING[selectedDevice] ?? DEFAULT_DEVICE_DTYPE;
  }
  if (needsWarn && warn) {
    warn(
      `dtype not specified for "${fileName}". Using the default dtype (${result}) for this device (${selectedDevice}).`
    );
  }
  return result;
}
var DataTypeMap = Object.freeze({
  float32: Float32Array,
  // @ts-ignore ts(2552) Limited availability of Float16Array across browsers:
  // https://developer.mozilla.org/en-US/docs/Web/JavaScript/Reference/Global_Objects/Float16Array
  float16: typeof Float16Array !== "undefined" ? Float16Array : Uint16Array,
  float64: Float64Array,
  string: Array,
  // string[]
  int8: Int8Array,
  uint8: Uint8Array,
  int16: Int16Array,
  uint16: Uint16Array,
  int32: Int32Array,
  uint32: Uint32Array,
  int64: BigInt64Array,
  uint64: BigUint64Array,
  bool: Uint8Array,
  uint4: Uint8Array,
  int4: Int8Array
});

// src/utils/tensor.js
var Tensor2 = class _Tensor {
  /**
   * Dimensions of the tensor.
   * @type {number[]}
   */
  get dims() {
    return this.ort_tensor.dims;
  }
  set dims(value) {
    this.ort_tensor.dims = value;
  }
  /**
   * Type of the tensor.
   * @type {DataType}
   */
  get type() {
    return this.ort_tensor.type;
  }
  /**
   * The data stored in the tensor.
   * @type {DataArray}
   */
  get data() {
    return this.ort_tensor.data;
  }
  /**
   * The number of elements in the tensor.
   * @type {number}
   */
  get size() {
    return this.ort_tensor.size;
  }
  /**
   * The location of the tensor data.
   * @type {string}
   */
  get location() {
    return this.ort_tensor.location;
  }
  ort_tensor;
  /**
   * Create a new Tensor or copy an existing Tensor.
   * @param {[DataType, DataArray, number[]]|[ONNXTensor]} args
   */
  constructor(...args) {
    if (isONNXTensor(args[0])) {
      this.ort_tensor = /** @type {ONNXTensor} */
      args[0];
    } else {
      this.ort_tensor = new Tensor(
        /** @type {DataType} */
        args[0],
        // @ts-expect-error ts(2769) Type 'number' is not assignable to type 'bigint'.
        /** @type {Exclude<import('./maths.js').AnyTypedArray, Uint8ClampedArray>} */
        args[1],
        args[2]
      );
    }
    return new Proxy(this, {
      get: (obj, key) => {
        if (typeof key === "string") {
          let index = Number(key);
          if (Number.isInteger(index)) {
            return obj._getitem(index);
          }
        }
        return obj[key];
      },
      set: (obj, key, value) => {
        return obj[key] = value;
      }
    });
  }
  dispose() {
    this.ort_tensor.dispose();
  }
  /**
   * Returns an iterator object for iterating over the tensor data in row-major order.
   * If the tensor has more than one dimension, the iterator will yield subarrays.
   * @returns {Iterator<any>} An iterator object for iterating over the tensor data in row-major order.
   */
  *[Symbol.iterator]() {
    const [iterLength, ...iterDims] = this.dims;
    if (iterDims.length > 0) {
      const iterSize = iterDims.reduce((a, b) => a * b);
      for (let i = 0; i < iterLength; ++i) {
        yield this._subarray(i, iterSize, iterDims);
      }
    } else {
      yield* this.data;
    }
  }
  /**
   * Index into a Tensor object.
   * @param {number} index The index to access.
   * @returns {Tensor} The data at the specified index.
   */
  _getitem(index) {
    const [iterLength, ...iterDims] = this.dims;
    index = safeIndex(index, iterLength);
    if (iterDims.length > 0) {
      const iterSize = iterDims.reduce((a, b) => a * b);
      return this._subarray(index, iterSize, iterDims);
    } else {
      return new _Tensor(this.type, [this.data[index]], iterDims);
    }
  }
  /**
   * @param {number|bigint} item The item to search for in the tensor
   * @returns {number} The index of the first occurrence of item in the tensor data.
   */
  indexOf(item) {
    const this_data = this.data;
    for (let index = 0; index < this_data.length; ++index) {
      if (this_data[index] == item) {
        return index;
      }
    }
    return -1;
  }
  /**
   * @param {number} index
   * @param {number} iterSize
   * @param {any} iterDims
   * @returns {Tensor}
   */
  _subarray(index, iterSize, iterDims) {
    const o1 = index * iterSize;
    const o2 = (index + 1) * iterSize;
    const data = "subarray" in this.data ? this.data.subarray(o1, o2) : this.data.slice(o1, o2);
    return new _Tensor(this.type, data, iterDims);
  }
  /**
   * Returns the value of this tensor as a standard JavaScript Number. This only works
   * for tensors with one element. For other cases, see `Tensor.tolist()`.
   * @returns {number|bigint} The value of this tensor as a standard JavaScript Number.
   * @throws {Error} If the tensor has more than one element.
   */
  item() {
    const this_data = this.data;
    if (this_data.length !== 1) {
      throw new Error(`a Tensor with ${this_data.length} elements cannot be converted to Scalar`);
    }
    return this_data[0];
  }
  /**
   * Convert tensor data to a n-dimensional JS list
   * @returns {any[]}
   */
  tolist() {
    return reshape(this.data, this.dims);
  }
  /**
   * Return a new Tensor with the sigmoid function applied to each element.
   * @returns {Tensor} The tensor with the sigmoid function applied.
   */
  sigmoid() {
    return this.clone().sigmoid_();
  }
  /**
   * Applies the sigmoid function to the tensor in place.
   * @returns {Tensor} Returns `this`.
   */
  sigmoid_() {
    const this_data = this.data;
    for (let i = 0; i < this_data.length; ++i) {
      this_data[i] = 1 / (1 + Math.exp(-this_data[i]));
    }
    return this;
  }
  /**
   * Return a new Tensor with a callback function applied to each element.
   * @param {Function} callback - The function to apply to each element. It should take three arguments:
   *                              the current element, its index, and the tensor's data array.
   * @returns {Tensor} A new Tensor with the callback function applied to each element.
   */
  map(callback) {
    return this.clone().map_(callback);
  }
  /**
   * Apply a callback function to each element of the tensor in place.
   * @param {Function} callback - The function to apply to each element. It should take three arguments:
   *                              the current element, its index, and the tensor's data array.
   * @returns {Tensor} Returns `this`.
   */
  map_(callback) {
    const this_data = this.data;
    for (let i = 0; i < this_data.length; ++i) {
      this_data[i] = callback(this_data[i], i, this_data);
    }
    return this;
  }
  /**
   * Return a new Tensor with every element multiplied by a constant.
   * @param {number} val The value to multiply by.
   * @returns {Tensor} The new tensor.
   */
  mul(val) {
    return this.clone().mul_(val);
  }
  /**
   * Multiply the tensor by a constant in place.
   * @param {number} val The value to multiply by.
   * @returns {Tensor} Returns `this`.
   */
  mul_(val) {
    const this_data = this.data;
    for (let i = 0; i < this_data.length; ++i) {
      this_data[i] *= val;
    }
    return this;
  }
  /**
   * Return a new Tensor with every element divided by a constant.
   * @param {number} val The value to divide by.
   * @returns {Tensor} The new tensor.
   */
  div(val) {
    return this.clone().div_(val);
  }
  /**
   * Divide the tensor by a constant in place.
   * @param {number} val The value to divide by.
   * @returns {Tensor} Returns `this`.
   */
  div_(val) {
    const this_data = this.data;
    for (let i = 0; i < this_data.length; ++i) {
      this_data[i] /= val;
    }
    return this;
  }
  /**
   * Return a new Tensor with every element added by a constant.
   * @param {number} val The value to add by.
   * @returns {Tensor} The new tensor.
   */
  add(val) {
    return this.clone().add_(val);
  }
  /**
   * Add the tensor by a constant in place.
   * @param {number} val The value to add by.
   * @returns {Tensor} Returns `this`.
   */
  add_(val) {
    const this_data = this.data;
    for (let i = 0; i < this_data.length; ++i) {
      this_data[i] += val;
    }
    return this;
  }
  /**
   * Return a new Tensor with every element subtracted by a constant.
   * @param {number} val The value to subtract by.
   * @returns {Tensor} The new tensor.
   */
  sub(val) {
    return this.clone().sub_(val);
  }
  /**
   * Subtract the tensor by a constant in place.
   * @param {number} val The value to subtract by.
   * @returns {Tensor} Returns `this`.
   */
  sub_(val) {
    const this_data = this.data;
    for (let i = 0; i < this_data.length; ++i) {
      this_data[i] -= val;
    }
    return this;
  }
  /**
   * Creates a deep copy of the current Tensor.
   * @returns {Tensor} A new Tensor with the same type, data, and dimensions as the original.
   */
  clone() {
    return new _Tensor(this.type, this.data.slice(), this.dims.slice());
  }
  /**
   * Performs a slice operation on the Tensor along specified dimensions.
   *
   * Consider a Tensor that has a dimension of [4, 7]:
   * ```
   * [ 1,  2,  3,  4,  5,  6,  7]
   * [ 8,  9, 10, 11, 12, 13, 14]
   * [15, 16, 17, 18, 19, 20, 21]
   * [22, 23, 24, 25, 26, 27, 28]
   * ```
   * We can slice against the two dims of row and column, for instance in this
   * case we can start at the second element, and return to the second last,
   * like this:
   * ```
   * tensor.slice([1, -1], [1, -1]);
   * ```
   * which would return:
   * ```
   * [  9, 10, 11, 12, 13 ]
   * [ 16, 17, 18, 19, 20 ]
   * ```
   *
   * @param {...(number|number[]|null)} slices The slice specifications for each dimension.
   * - If a number is given, then a single element is selected.
   * - If an array of two numbers is given, then a range of elements [start, end (exclusive)] is selected.
   * - If null is given, then the entire dimension is selected.
   * @returns {Tensor} A new Tensor containing the selected elements.
   * @throws {Error} If the slice input is invalid.
   */
  slice(...slices) {
    const newTensorDims = [];
    const newOffsets = [];
    for (let sliceIndex = 0; sliceIndex < this.dims.length; ++sliceIndex) {
      let slice3 = slices[sliceIndex];
      if (slice3 === null || slice3 === void 0) {
        newOffsets.push([0, this.dims[sliceIndex]]);
        newTensorDims.push(this.dims[sliceIndex]);
      } else if (typeof slice3 === "number") {
        slice3 = safeIndex(slice3, this.dims[sliceIndex], sliceIndex);
        newOffsets.push([slice3, slice3 + 1]);
      } else if (Array.isArray(slice3) && slice3.length === 2) {
        let [start, end] = slice3;
        start = start === null ? 0 : safeIndex(start, this.dims[sliceIndex], sliceIndex, false);
        end = end === null ? this.dims[sliceIndex] : safeIndex(end, this.dims[sliceIndex], sliceIndex, false);
        if (start > end) {
          throw new Error(`Invalid slice: ${slice3}`);
        }
        const offsets = [Math.max(start, 0), Math.min(end, this.dims[sliceIndex])];
        newOffsets.push(offsets);
        newTensorDims.push(offsets[1] - offsets[0]);
      } else {
        throw new Error(`Invalid slice: ${slice3}`);
      }
    }
    const newDims = newOffsets.map(([start, end]) => end - start);
    const newBufferSize = newDims.reduce((a, b) => a * b);
    const this_data = this.data;
    const data = new this_data.constructor(newBufferSize);
    const stride = this.stride();
    let isContiguous = true;
    for (let i = 1; i < newDims.length; ++i) {
      if (newOffsets[i][0] !== 0 || newOffsets[i][1] !== this.dims[i]) {
        isContiguous = false;
        break;
      }
    }
    if (isContiguous) {
      const start = newOffsets[0][0] * stride[0];
      const end = newOffsets[0][1] * stride[0];
      if (ArrayBuffer.isView(this_data)) {
        data.set(this_data.subarray(start, end));
      } else if (Array.isArray(this_data)) {
        const slicedData = this_data.slice(start, end);
        for (let i = 0; i < slicedData.length; ++i) {
          data[i] = slicedData[i];
        }
      } else {
        throw new Error("Unsupported data type for slicing");
      }
    } else {
      for (let i = 0; i < newBufferSize; ++i) {
        let originalIndex = 0;
        for (let j = newDims.length - 1, num = i; j >= 0; --j) {
          const size = newDims[j];
          originalIndex += (num % size + newOffsets[j][0]) * stride[j];
          num = Math.floor(num / size);
        }
        data[i] = this_data[originalIndex];
      }
    }
    return new _Tensor(this.type, data, newTensorDims);
  }
  /**
   * Return a permuted version of this Tensor, according to the provided dimensions.
   * @param  {...number} dims Dimensions to permute.
   * @returns {Tensor} The permuted tensor.
   */
  permute(...dims) {
    return permute(this, dims);
  }
  // TODO: implement transpose. For now (backwards compatibility), it's just an alias for permute()
  /** @type {Tensor['permute']} */
  transpose(...dims) {
    return this.permute(...dims);
  }
  /**
   * Returns the sum of each row of the input tensor in the given dimension dim.
   *
   * @param {number|null} [dim=null] The dimension or dimensions to reduce. If `null`, all dimensions are reduced.
   * @param {boolean} keepdim Whether the output tensor has `dim` retained or not.
   * @returns The summed tensor
   */
  sum(dim = null, keepdim = false) {
    return this.norm(1, dim, keepdim);
  }
  /**
   * Returns the matrix norm or vector norm of a given tensor.
   * @param {number|string} [p='fro'] The order of norm
   * @param {number|null} [dim=null] Specifies which dimension of the tensor to calculate the norm across.
   * If dim is None, the norm will be calculated across all dimensions of input.
   * @param {boolean} [keepdim=false] Whether the output tensors have dim retained or not.
   * @returns {Tensor} The norm of the tensor.
   */
  norm(p = "fro", dim = null, keepdim = false) {
    if (p === "fro") {
      p = 2;
    } else if (typeof p === "string") {
      throw Error(`Unsupported norm: ${p}`);
    }
    const this_data = this.data;
    const is_bigint = this_data instanceof BigInt64Array || this_data instanceof BigUint64Array;
    if (is_bigint && p !== 1) {
      throw Error(`Expected a floating point tensor as input. Got ${this.type}`);
    }
    let fn2, zero;
    if (is_bigint) {
      fn2 = (a, b) => a + b;
      zero = 0n;
    } else {
      fn2 = (a, b) => a + b ** p;
      zero = 0;
    }
    if (dim === null) {
      let val = this_data.reduce(fn2, zero);
      if (p !== 1) {
        val = val ** (1 / p);
      }
      return new _Tensor(this.type, [val], []);
    }
    const [type, result, resultDims] = reduce_helper(fn2, this, dim, keepdim);
    if (p !== 1) {
      for (let i = 0; i < result.length; ++i) {
        result[i] = result[i] ** (1 / p);
      }
    }
    return new _Tensor(type, result, resultDims);
  }
  /**
   * Performs `L_p` normalization of inputs over specified dimension. Operates in place.
   * @param {number} [p=2] The exponent value in the norm formulation
   * @param {number} [dim=1] The dimension to reduce
   * @returns {Tensor} `this` for operation chaining.
   */
  normalize_(p = 2, dim = 1) {
    dim = safeIndex(dim, this.dims.length);
    const norm = this.norm(p, dim, true);
    const this_data = this.data;
    const norm_data = norm.data;
    for (let i = 0; i < this_data.length; ++i) {
      let resultIndex = 0;
      for (let j = this.dims.length - 1, num = i, resultMultiplier = 1; j >= 0; --j) {
        const size = this.dims[j];
        if (j !== dim) {
          const index = num % size;
          resultIndex += index * resultMultiplier;
          resultMultiplier *= this.dims[j];
        }
        num = Math.floor(num / size);
      }
      this_data[i] /= norm_data[resultIndex];
    }
    return this;
  }
  /**
   * Performs `L_p` normalization of inputs over specified dimension.
   * @param {number} [p=2] The exponent value in the norm formulation
   * @param {number} [dim=1] The dimension to reduce
   * @returns {Tensor} The normalized tensor.
   */
  normalize(p = 2, dim = 1) {
    return this.clone().normalize_(p, dim);
  }
  /**
   * Compute and return the stride of this tensor.
   * Stride is the jump necessary to go from one element to the next one in the specified dimension dim.
   * @returns {number[]} The stride of this tensor.
   */
  stride() {
    return dimsToStride(this.dims);
  }
  /**
   * Returns a tensor with all specified dimensions of input of size 1 removed.
   *
   * NOTE: The returned tensor shares the storage with the input tensor, so changing the contents of one will change the contents of the other.
   * If you would like a copy, use `tensor.clone()` before squeezing.
   *
   * @param {number|number[]|null} [dim=null] If given, the input will be squeezed only in the specified dimensions.
   * @returns {Tensor} The squeezed tensor
   */
  squeeze(dim = null) {
    return new _Tensor(this.type, this.data, calc_squeeze_dims(this.dims, dim));
  }
  /**
   * In-place version of @see {@link Tensor.squeeze}
   */
  squeeze_(dim = null) {
    this.dims = calc_squeeze_dims(this.dims, dim);
    return this;
  }
  /**
   * Returns a new tensor with a dimension of size one inserted at the specified position.
   *
   * NOTE: The returned tensor shares the same underlying data with this tensor.
   *
   * @param {number} dim The index at which to insert the singleton dimension
   * @returns {Tensor} The unsqueezed tensor
   */
  unsqueeze(dim) {
    return new _Tensor(this.type, this.data, calc_unsqueeze_dims(this.dims, dim));
  }
  /**
   * In-place version of @see {@link Tensor.unsqueeze}
   * @type {Tensor['unsqueeze']}
   */
  unsqueeze_(dim) {
    this.dims = calc_unsqueeze_dims(this.dims, dim);
    return this;
  }
  /**
   * In-place version of @see {@link Tensor.flatten}
   */
  flatten_(start_dim = 0, end_dim = -1) {
    end_dim = (end_dim + this.dims.length) % this.dims.length;
    let dimsToKeepBefore = this.dims.slice(0, start_dim);
    let dimsToFlatten = this.dims.slice(start_dim, end_dim + 1);
    let dimsToKeepAfter = this.dims.slice(end_dim + 1);
    this.dims = [...dimsToKeepBefore, dimsToFlatten.reduce((a, b) => a * b, 1), ...dimsToKeepAfter];
    return this;
  }
  /**
   * Flattens input by reshaping it into a one-dimensional tensor.
   * If `start_dim` or `end_dim` are passed, only dimensions starting with `start_dim`
   * and ending with `end_dim` are flattened. The order of elements in input is unchanged.
   * @param {number} start_dim the first dim to flatten
   * @param {number} end_dim the last dim to flatten
   * @returns {Tensor} The flattened tensor.
   */
  flatten(start_dim = 0, end_dim = -1) {
    return this.clone().flatten_(start_dim, end_dim);
  }
  /**
   * Returns a new tensor with the same data as the `self` tensor but of a different `shape`.
   * @param  {...number} dims the desired size
   * @returns {Tensor} The tensor with the same data but different shape
   */
  view(...dims) {
    let inferredIndex = -1;
    for (let i = 0; i < dims.length; ++i) {
      if (dims[i] === -1) {
        if (inferredIndex !== -1) {
          throw new Error("Only one dimension can be inferred");
        }
        inferredIndex = i;
      }
    }
    const this_data = this.data;
    if (inferredIndex !== -1) {
      const productOther = dims.reduce((product2, curr, index) => {
        return index !== inferredIndex ? product2 * curr : product2;
      }, 1);
      dims[inferredIndex] = this_data.length / productOther;
    }
    return new _Tensor(this.type, this_data, dims);
  }
  neg_() {
    const this_data = this.data;
    for (let i = 0; i < this_data.length; ++i) {
      this_data[i] = -this_data[i];
    }
    return this;
  }
  neg() {
    return this.clone().neg_();
  }
  /**
   * Computes input > val element-wise.
   * @param {number} val The value to compare with.
   * @returns {Tensor} A boolean tensor that is `true` where input is greater than other and `false` elsewhere.
   */
  gt(val) {
    const mask = new Uint8Array(this.data.length);
    const this_data = this.data;
    for (let i = 0; i < this_data.length; ++i) {
      mask[i] = this_data[i] > val ? 1 : 0;
    }
    return new _Tensor("bool", mask, this.dims);
  }
  /**
   * Computes input < val element-wise.
   * @param {number} val The value to compare with.
   * @returns {Tensor} A boolean tensor that is `true` where input is less than other and `false` elsewhere.
   */
  lt(val) {
    const mask = new Uint8Array(this.data.length);
    const this_data = this.data;
    for (let i = 0; i < this_data.length; ++i) {
      mask[i] = this_data[i] < val ? 1 : 0;
    }
    return new _Tensor("bool", mask, this.dims);
  }
  /**
   * In-place version of @see {@link Tensor.clamp}
   * @type {Tensor['clamp']}
   */
  clamp_(min3, max2) {
    const this_data = this.data;
    for (let i = 0; i < this_data.length; ++i) {
      this_data[i] = Math.min(Math.max(this_data[i], min3), max2);
    }
    return this;
  }
  /**
   * Clamps all elements in input into the range [ min, max ]
   * @param {number} min lower-bound of the range to be clamped to
   * @param {number} max upper-bound of the range to be clamped to
   * @returns {Tensor} the output tensor.
   */
  clamp(min3, max2) {
    return this.clone().clamp_(min3, max2);
  }
  /**
   * In-place version of @see {@link Tensor.round}
   */
  round_() {
    const this_data = this.data;
    for (let i = 0; i < this_data.length; ++i) {
      this_data[i] = Math.round(this_data[i]);
    }
    return this;
  }
  /**
   * Rounds elements of input to the nearest integer.
   * @returns {Tensor} the output tensor.
   */
  round() {
    return this.clone().round_();
  }
  mean(dim = null, keepdim = false) {
    return mean(this, dim, keepdim);
  }
  min(dim = null, keepdim = false) {
    if (dim === null) {
      const val = min2(this.data)[0];
      return new _Tensor(
        this.type,
        [val],
        [
          /* scalar */
        ]
      );
    }
    const [type, result, resultDims] = reduce_helper((a, b) => Math.min(a, b), this, dim, keepdim, Infinity);
    return new _Tensor(type, result, resultDims);
  }
  max(dim = null, keepdim = false) {
    if (dim === null) {
      const val = max(this.data)[0];
      return new _Tensor(
        this.type,
        [val],
        [
          /* scalar */
        ]
      );
    }
    const [type, result, resultDims] = reduce_helper((a, b) => Math.max(a, b), this, dim, keepdim, -Infinity);
    return new _Tensor(type, result, resultDims);
  }
  argmin(dim = null, keepdim = false) {
    if (dim !== null) {
      throw new Error("`dim !== null` not yet implemented.");
    }
    const index = min2(this.data)[1];
    return new _Tensor("int64", [BigInt(index)], []);
  }
  argmax(dim = null, keepdim = false) {
    if (dim !== null) {
      throw new Error("`dim !== null` not yet implemented.");
    }
    const index = max(this.data)[1];
    return new _Tensor("int64", [BigInt(index)], []);
  }
  /**
   * Repeats this tensor along the specified dimensions.
   * @param {...number} repeats The number of times to repeat this tensor along each dimension.
   * @returns {Tensor} The repeated tensor.
   * @throws {Error} If the number of repeats is less than the number of dimensions.
   */
  repeat(...repeats) {
    if (repeats.length < this.dims.length) {
      throw new Error(
        `Number of dimensions of repeat dims (${repeats.length}) cannot be smaller than number of dimensions of tensor (${this.dims.length})`
      );
    }
    if (repeats.every((r) => r === 1)) {
      if (repeats.length === this.dims.length) {
        return this.clone();
      }
      const numPrependedDims2 = repeats.length - this.dims.length;
      const newDims = Array(numPrependedDims2).fill(1).concat(this.dims);
      return new _Tensor(this.type, this.data.slice(), newDims);
    }
    const numPrependedDims = repeats.length - this.dims.length;
    const expandedDims = Array(numPrependedDims).fill(1).concat(this.dims);
    const outputDims = expandedDims.map((dim, i) => dim * repeats[i]);
    const outputSize = outputDims.reduce((a, b) => a * b, 1);
    const this_data = this.data;
    const data = new this_data.constructor(outputSize);
    const inputStrides = dimsToStride(expandedDims);
    const outputStrides = dimsToStride(outputDims);
    for (let i = 0; i < outputSize; ++i) {
      let remaining = i;
      let inputIndex = 0;
      for (let d = 0; d < outputDims.length; ++d) {
        const outputCoord = Math.floor(remaining / outputStrides[d]);
        remaining = remaining % outputStrides[d];
        const inputCoord = outputCoord % expandedDims[d];
        inputIndex += inputCoord * inputStrides[d];
      }
      data[i] = this_data[inputIndex];
    }
    return new _Tensor(this.type, data, outputDims);
  }
  /**
   * Constructs a tensor by repeating the elements of input. The dims argument specifies the number of repetitions in each dimension.
   * @param  {...number} dims The number of repetitions per dimension.
   * @returns {Tensor} The tiled tensor.
   */
  tile(...dims) {
    if (dims.length < this.dims.length) {
      const numPrependedRepeats = this.dims.length - dims.length;
      dims = Array(numPrependedRepeats).fill(1).concat(dims);
    }
    return this.repeat(...dims);
  }
  /**
   * Performs Tensor dtype conversion.
   * @param {DataType} type The desired data type.
   * @returns {Tensor} The converted tensor.
   */
  to(type) {
    if (this.type === type) return this;
    if (!DataTypeMap.hasOwnProperty(type)) {
      throw new Error(`Unsupported type: ${type}`);
    }
    let map_fn;
    const is_source_bigint = ["int64", "uint64"].includes(this.type);
    const is_dest_bigint = ["int64", "uint64"].includes(type);
    if (is_source_bigint && !is_dest_bigint) {
      map_fn = Number;
    } else if (!is_source_bigint && is_dest_bigint) {
      if (["float16", "float32", "float64"].includes(this.type)) {
        map_fn = (x) => BigInt(Math.floor(x));
      } else {
        map_fn = BigInt;
      }
    } else if (this.type === "float16" && type == "float32" && this.data instanceof Uint16Array) {
      return new _Tensor(type, uint16_to_float32(this.data), this.dims);
    }
    return new _Tensor(
      type,
      // @ts-ignore - TypeScript can't verify .from() across union of TypedArray constructors
      DataTypeMap[type].from(this.data, map_fn),
      this.dims
    );
  }
};
function reshape(data, dimensions) {
  const totalElements = data.length;
  const dimensionSize = dimensions.reduce((a, b) => a * b);
  if (totalElements !== dimensionSize) {
    throw Error(`cannot reshape array of size ${totalElements} into shape (${dimensions})`);
  }
  let reshapedArray = data;
  for (let i = dimensions.length - 1; i >= 0; i--) {
    reshapedArray = reshapedArray.reduce(
      (acc, val) => {
        let lastArray = acc[acc.length - 1];
        if (lastArray.length < dimensions[i]) {
          lastArray.push(val);
        } else {
          acc.push([val]);
        }
        return acc;
      },
      [[]]
    );
  }
  return reshapedArray[0];
}
function permute(tensor, axes) {
  const [permutedData, shape] = permute_data(tensor.data, tensor.dims, axes);
  return new Tensor2(tensor.type, permutedData, shape);
}
function interpolate(input, [out_height, out_width], mode = "bilinear", align_corners = false) {
  const in_channels = input.dims.at(-3) ?? 1;
  const in_height = input.dims.at(-2);
  const in_width = input.dims.at(-1);
  const output = interpolate_data(
    /** @type {import('./maths.js').TypedArray}*/
    input.data,
    [in_channels, in_height, in_width],
    [out_height, out_width],
    mode,
    align_corners
  );
  return new Tensor2(input.type, output, [in_channels, out_height, out_width]);
}
async function interpolate_4d(input, { size = null, mode = "bilinear" } = {}) {
  if (input.dims.length !== 4) {
    throw new Error("`interpolate_4d` currently only supports 4D input.");
  }
  if (!size) {
    throw new Error("`interpolate_4d` requires a `size` argument.");
  }
  let targetDims;
  if (size.length === 2) {
    targetDims = [...input.dims.slice(0, 2), ...size];
  } else if (size.length === 3) {
    targetDims = [input.dims[0], ...size];
  } else if (size.length === 4) {
    targetDims = size;
  } else {
    throw new Error("`size` must be of length 2, 3, or 4.");
  }
  let op;
  if (mode === "nearest") {
    op = await TensorOpRegistry.nearest_interpolate_4d;
  } else if (mode === "bilinear") {
    op = await TensorOpRegistry.bilinear_interpolate_4d;
  } else if (mode === "bicubic") {
    op = await TensorOpRegistry.bicubic_interpolate_4d;
  } else {
    throw new Error(`Unsupported mode: ${mode}`);
  }
  const sizeTensor = new Tensor2("int64", new BigInt64Array(targetDims.map(BigInt)), [targetDims.length]);
  return await op({ x: input, s: sizeTensor });
}
async function matmul(a, b) {
  const op = await TensorOpRegistry.matmul;
  return await op({ a, b });
}
async function rfft(x, a) {
  const op = await TensorOpRegistry.rfft;
  return await op({ x, a });
}
async function topk(x, k2) {
  const op = await TensorOpRegistry.top_k;
  if (k2 == null) {
    k2 = x.dims.at(-1);
  } else {
    k2 = Math.min(k2, x.dims.at(-1));
  }
  return await op({
    x,
    k: new Tensor2("int64", [BigInt(k2)], [1])
  });
}
var arrayToIndexTensor = (array) => new Tensor2("int64", array, [array.length]);
async function slice2(data, starts, ends, axes, steps) {
  const op = await TensorOpRegistry.slice;
  return await op({
    x: data,
    s: arrayToIndexTensor(starts),
    e: arrayToIndexTensor(ends),
    a: arrayToIndexTensor(axes),
    t: arrayToIndexTensor(steps ?? new Array(axes.length).fill(1))
  });
}
function mean_pooling(last_hidden_state, attention_mask) {
  const lastHiddenStateData = last_hidden_state.data;
  const attentionMaskData = attention_mask.data;
  const shape = [last_hidden_state.dims[0], last_hidden_state.dims[2]];
  const returnedData = new lastHiddenStateData.constructor(shape[0] * shape[1]);
  const [batchSize, seqLength, embedDim] = last_hidden_state.dims;
  let outIndex = 0;
  for (let i = 0; i < batchSize; ++i) {
    const offset = i * embedDim * seqLength;
    for (let k2 = 0; k2 < embedDim; ++k2) {
      let sum = 0;
      let count2 = 0;
      const attnMaskOffset = i * seqLength;
      const offset2 = offset + k2;
      for (let j = 0; j < seqLength; ++j) {
        const attn = Number(attentionMaskData[attnMaskOffset + j]);
        count2 += attn;
        sum += lastHiddenStateData[offset2 + j * embedDim] * attn;
      }
      const avg = sum / count2;
      returnedData[outIndex++] = avg;
    }
  }
  return new Tensor2(last_hidden_state.type, returnedData, shape);
}
function layer_norm(input, normalized_shape, { eps = 1e-5 } = {}) {
  if (input.dims.length !== 2) {
    throw new Error("`layer_norm` currently only supports 2D input.");
  }
  const [batchSize, featureDim] = input.dims;
  if (normalized_shape.length !== 1 && normalized_shape[0] !== featureDim) {
    throw new Error("`normalized_shape` must be a 1D array with shape `[input.dims[1]]`.");
  }
  const [std, mean2] = std_mean(input, 1, 0, true);
  const stdData = (
    /** @type {Float32Array} */
    std.data
  );
  const meanData = (
    /** @type {Float32Array} */
    mean2.data
  );
  const inputData = (
    /** @type {Float32Array} */
    input.data
  );
  const returnedData = new inputData.constructor(inputData.length);
  for (let i = 0; i < batchSize; ++i) {
    const offset = i * featureDim;
    for (let j = 0; j < featureDim; ++j) {
      const offset2 = offset + j;
      returnedData[offset2] = (inputData[offset2] - meanData[i]) / (stdData[i] + eps);
    }
  }
  return new Tensor2(input.type, returnedData, input.dims);
}
function calc_squeeze_dims(dims, dim) {
  dims = dims.slice();
  if (dim === null) {
    dims = dims.filter((d) => d !== 1);
  } else if (typeof dim === "number") {
    if (dims[dim] === 1) {
      dims.splice(dim, 1);
    }
  } else if (Array.isArray(dim)) {
    dims = dims.filter((x, i) => {
      return x !== 1 || !dim.includes(i);
    });
  }
  return dims;
}
function calc_unsqueeze_dims(dims, dim) {
  dim = safeIndex(dim, dims.length + 1);
  dims = dims.slice();
  dims.splice(dim, 0, 1);
  return dims;
}
function safeIndex(index, size, dimension = null, boundsCheck = true) {
  if (index < -size || index >= size) {
    if (boundsCheck) {
      throw new Error(
        `IndexError: index ${index} is out of bounds for dimension${dimension === null ? "" : " " + dimension} with size ${size}`
      );
    } else {
      return index < -size ? 0 : size;
    }
  }
  if (index < 0) {
    index = (index % size + size) % size;
  }
  return index;
}
function cat(tensors, dim = 0) {
  dim = safeIndex(dim, tensors[0].dims.length);
  const resultDims = tensors[0].dims.slice();
  resultDims[dim] = tensors.reduce((a, b) => a + b.dims[dim], 0);
  const resultSize = resultDims.reduce((a, b) => a * b, 1);
  const result = new tensors[0].data.constructor(resultSize);
  const resultType = tensors[0].type;
  if (dim === 0) {
    let offset = 0;
    for (const tensor of tensors) {
      const tensorData = tensor.data;
      result.set(tensorData, offset);
      offset += tensorData.length;
    }
  } else {
    let currentDim = 0;
    for (let t = 0; t < tensors.length; ++t) {
      const { data, dims } = tensors[t];
      for (let i = 0; i < data.length; ++i) {
        let resultIndex = 0;
        for (let j = dims.length - 1, num = i, resultMultiplier = 1; j >= 0; --j) {
          const size = dims[j];
          let index = num % size;
          if (j === dim) {
            index += currentDim;
          }
          resultIndex += index * resultMultiplier;
          resultMultiplier *= resultDims[j];
          num = Math.floor(num / size);
        }
        result[resultIndex] = data[i];
      }
      currentDim += dims[dim];
    }
  }
  return new Tensor2(resultType, result, resultDims);
}
function stack(tensors, dim = 0) {
  return cat(
    tensors.map((t) => t.unsqueeze(dim)),
    dim
  );
}
function reduce_helper(callbackfn, input, dim, keepdim = false, initialValue = null) {
  const inputData = input.data;
  const inputDims = input.dims;
  dim = safeIndex(dim, inputDims.length);
  const resultDims = inputDims.slice();
  resultDims[dim] = 1;
  const result = new inputData.constructor(inputData.length / inputDims[dim]);
  if (initialValue !== null) {
    result.fill(initialValue);
  }
  for (let i = 0; i < inputData.length; ++i) {
    let resultIndex = 0;
    for (let j = inputDims.length - 1, num = i, resultMultiplier = 1; j >= 0; --j) {
      const size = inputDims[j];
      if (j !== dim) {
        const index = num % size;
        resultIndex += index * resultMultiplier;
        resultMultiplier *= resultDims[j];
      }
      num = Math.floor(num / size);
    }
    result[resultIndex] = callbackfn(result[resultIndex], inputData[i], i, resultIndex);
  }
  if (!keepdim) resultDims.splice(dim, 1);
  return [input.type, result, resultDims];
}
function std_mean(input, dim = null, correction = 1, keepdim = false) {
  const inputData = (
    /** @type {Float32Array} */
    input.data
  );
  const inputDims = input.dims;
  if (dim === null) {
    const sum = inputData.reduce((a, b) => a + b, 0);
    const mean2 = sum / inputData.length;
    const std = Math.sqrt(inputData.reduce((a, b) => a + (b - mean2) ** 2, 0) / (inputData.length - correction));
    const meanTensor2 = new Tensor2(
      input.type,
      [mean2],
      [
        /* scalar */
      ]
    );
    const stdTensor2 = new Tensor2(
      input.type,
      [std],
      [
        /* scalar */
      ]
    );
    return [stdTensor2, meanTensor2];
  }
  dim = safeIndex(dim, inputDims.length);
  const meanTensor = mean(input, dim, keepdim);
  const meanTensorData = meanTensor.data;
  const [type, result, resultDims] = reduce_helper(
    (a, b, i, j) => a + (b - meanTensorData[j]) ** 2,
    input,
    dim,
    keepdim
  );
  for (let i = 0; i < result.length; ++i) {
    result[i] = Math.sqrt(result[i] / (inputDims[dim] - correction));
  }
  const stdTensor = new Tensor2(type, result, resultDims);
  return [stdTensor, meanTensor];
}
function mean(input, dim = null, keepdim = false) {
  const inputDims = input.dims;
  const inputData = (
    /** @type {Float32Array} */
    input.data
  );
  if (dim === null) {
    const val = inputData.reduce((a, b) => a + b, 0);
    return new Tensor2(
      input.type,
      [val / inputData.length],
      [
        /* scalar */
      ]
    );
  }
  dim = safeIndex(dim, inputDims.length);
  const [type, result, resultDims] = reduce_helper((a, b) => a + b, input, dim, keepdim);
  if (inputDims[dim] !== 1) {
    for (let i = 0; i < result.length; ++i) {
      result[i] /= inputDims[dim];
    }
  }
  return new Tensor2(type, result, resultDims);
}
function dimsToStride(dims) {
  const stride = new Array(dims.length);
  for (let i = dims.length - 1, s2 = 1; i >= 0; --i) {
    stride[i] = s2;
    s2 *= dims[i];
  }
  return stride;
}
function fullHelper(size, fill_value, dtype, cls) {
  const numElements = size.reduce((a, b) => a * b, 1);
  return new Tensor2(dtype, new cls(numElements).fill(fill_value), size);
}
function full(size, fill_value) {
  let dtype;
  let typedArrayCls;
  if (typeof fill_value === "number") {
    dtype = "float32";
    typedArrayCls = Float32Array;
  } else if (typeof fill_value === "bigint") {
    dtype = "int64";
    typedArrayCls = BigInt64Array;
  } else if (typeof fill_value === "boolean") {
    dtype = "bool";
    typedArrayCls = Uint8Array;
  } else {
    throw new Error(`Unsupported data type: ${typeof fill_value}`);
  }
  return fullHelper(size, fill_value, dtype, typedArrayCls);
}
function full_like(tensor, fill_value) {
  return full(tensor.dims, fill_value);
}
function ones(size) {
  return fullHelper(size, 1n, "int64", BigInt64Array);
}
function ones_like(tensor) {
  return ones(tensor.dims);
}
function zeros(size) {
  return fullHelper(size, 0n, "int64", BigInt64Array);
}
function zeros_like(tensor) {
  return zeros(tensor.dims);
}
function rand(size) {
  const length = size.reduce((a, b) => a * b, 1);
  return new Tensor2(
    "float32",
    Float32Array.from({ length }, () => random.random()),
    size
  );
}
function randn(size) {
  const length = size.reduce((a, b) => a * b, 1);
  return new Tensor2(
    "float32",
    Float32Array.from({ length }, () => random.gauss()),
    size
  );
}
function quantize_embeddings(tensor, precision) {
  if (tensor.dims.length !== 2) {
    throw new Error("The tensor must have 2 dimensions");
  }
  if (tensor.dims.at(-1) % 8 !== 0) {
    throw new Error("The last dimension of the tensor must be a multiple of 8");
  }
  if (!["binary", "ubinary"].includes(precision)) {
    throw new Error("The precision must be either 'binary' or 'ubinary'");
  }
  const signed = precision === "binary";
  const dtype = signed ? "int8" : "uint8";
  const cls = signed ? Int8Array : Uint8Array;
  const inputData = tensor.data;
  const outputData = new cls(inputData.length / 8);
  for (let i = 0; i < inputData.length; ++i) {
    const bit = inputData[i] > 0 ? 1 : 0;
    const arrayIndex = Math.floor(i / 8);
    const bitPosition = i % 8;
    outputData[arrayIndex] |= bit << 7 - bitPosition;
    if (signed && bitPosition === 0) {
      outputData[arrayIndex] -= 128;
    }
  }
  return new Tensor2(dtype, outputData, [tensor.dims[0], tensor.dims[1] / 8]);
}

// src/utils/model_registry/get_tokenizer_files.js
async function get_tokenizer_files(modelId) {
  if (!modelId) {
    throw new Error("modelId is required for get_tokenizer_files");
  }
  const metadata = await get_file_metadata(modelId, "tokenizer_config.json", {});
  if (metadata.exists) {
    return ["tokenizer.json", "tokenizer_config.json"];
  }
  return [];
}

// src/tokenization_utils.js
async function loadTokenizer(pretrained_model_name_or_path, options) {
  const tokenizerFiles = await get_tokenizer_files(pretrained_model_name_or_path);
  return await Promise.all(
    tokenizerFiles.map((file) => getModelJSON(pretrained_model_name_or_path, file, true, options))
  );
}
function prepareTensorForDecode(tensor) {
  const dims = tensor.dims;
  switch (dims.length) {
    case 1:
      return tensor.tolist();
    case 2:
      if (dims[0] !== 1) {
        throw new Error(
          "Unable to decode tensor with `batch size !== 1`. Use `tokenizer.batch_decode(...)` for batched inputs."
        );
      }
      return tensor.tolist()[0];
    default:
      throw new Error(`Expected tensor to have 1-2 dimensions, got ${dims.length}.`);
  }
}
var SPECIAL_TOKEN_ATTRIBUTES = [
  "bos_token",
  "eos_token",
  "unk_token",
  "sep_token",
  "pad_token",
  "cls_token",
  "mask_token"
  // additional_special_tokens (TODO)
];
function padHelper(item, length, value_fn, side) {
  for (const key of Object.keys(item)) {
    const diff = length - item[key].length;
    const value = value_fn(key);
    const padData = new Array(diff).fill(value);
    item[key] = side === "right" ? mergeArrays(item[key], padData) : mergeArrays(padData, item[key]);
  }
}
function truncateHelper(item, length) {
  for (const key of Object.keys(item)) {
    item[key].length = length;
  }
}
function getTokenFromConfig(config, ...keys) {
  for (const key of keys) {
    if (!Object.hasOwn(config, key)) continue;
    const item = config[key];
    if (!item) continue;
    if (typeof item === "object") {
      if (item.__type === "AddedToken") {
        return item.content;
      } else {
        throw Error(`Unknown token: ${item}`);
      }
    } else {
      return item;
    }
  }
  return null;
}
function getSpecialTokens(tokenizer) {
  const special = [];
  for (const value of tokenizer.get_added_tokens_decoder().values()) {
    if (value.special) special.push(value);
  }
  return special;
}
var PreTrainedTokenizer = class extends /** @type {new (tokenizerJSON: Object, tokenizerConfig: Object) => PreTrainedTokenizerCallback} */
Callable {
  return_token_type_ids = false;
  padding_side = "right";
  /**
   * Create a new PreTrainedTokenizer instance.
   * @param {Object} tokenizerJSON The JSON of the tokenizer.
   * @param {Object} tokenizerConfig The config of the tokenizer.
   */
  constructor(tokenizerJSON, tokenizerConfig) {
    super();
    this._tokenizerJSON = tokenizerJSON;
    this._tokenizerConfig = tokenizerConfig;
    this._tokenizer = new Tokenizer_default(tokenizerJSON, tokenizerConfig);
    this.config = tokenizerConfig;
    this.padding_side = tokenizerConfig.padding_side ?? this.padding_side;
    this.mask_token = getTokenFromConfig(tokenizerConfig, "mask_token");
    this.mask_token_id = this._tokenizer.token_to_id(this.mask_token);
    this.pad_token = getTokenFromConfig(tokenizerConfig, "pad_token", "eos_token");
    this.pad_token_id = this._tokenizer.token_to_id(this.pad_token);
    this.sep_token = getTokenFromConfig(tokenizerConfig, "sep_token");
    this.sep_token_id = this._tokenizer.token_to_id(this.sep_token);
    this.unk_token = getTokenFromConfig(tokenizerConfig, "unk_token");
    this.unk_token_id = this._tokenizer.token_to_id(this.unk_token);
    this.bos_token = getTokenFromConfig(tokenizerConfig, "bos_token");
    this.bos_token_id = this._tokenizer.token_to_id(this.bos_token);
    this.eos_token = getTokenFromConfig(tokenizerConfig, "eos_token");
    this.eos_token_id = this._tokenizer.token_to_id(this.eos_token);
    this.chat_template = tokenizerConfig.chat_template ?? null;
    if (Array.isArray(this.chat_template)) {
      const chat_template = /* @__PURE__ */ Object.create(null);
      for (const { name, template } of this.chat_template) {
        if (typeof name !== "string" || typeof template !== "string") {
          throw new Error('Chat template must be a list of objects with "name" and "template" properties');
        }
        chat_template[name] = template;
      }
      this.chat_template = chat_template;
    }
    this._compiled_template_cache = /* @__PURE__ */ new Map();
    const special_tokens = getSpecialTokens(this._tokenizer);
    this.all_special_ids = special_tokens.map((t) => t.id);
    this.all_special_tokens = special_tokens.map((t) => t.content);
  }
  /**
   * Loads a pre-trained tokenizer from the given `pretrained_model_name_or_path`.
   *
   * @param {string} pretrained_model_name_or_path The path to the pre-trained tokenizer.
   * @param {PretrainedTokenizerOptions} options Additional options for loading the tokenizer.
   *
   * @throws {Error} Throws an error if the tokenizer.json or tokenizer_config.json files are not found in the `pretrained_model_name_or_path`.
   * @returns {Promise<PreTrainedTokenizer>} A new instance of the `PreTrainedTokenizer` class.
   */
  static async from_pretrained(pretrained_model_name_or_path, { progress_callback = null, config = null, cache_dir = null, local_files_only = false, revision = "main" } = {}) {
    const info = await loadTokenizer(pretrained_model_name_or_path, {
      progress_callback,
      config,
      cache_dir,
      local_files_only,
      revision
    });
    return new this(...info);
  }
  get_vocab() {
    return this._tokenizer.get_vocab();
  }
  get model_max_length() {
    return this._tokenizerConfig.model_max_length ?? Infinity;
  }
  get add_eos_token() {
    return this._tokenizerConfig.add_eos_token;
  }
  get add_bos_token() {
    return this._tokenizerConfig.add_bos_token;
  }
  /**
   * Converts a token string (or a sequence of tokens) into a single integer id (or a sequence of ids), using the vocabulary.
   *
   * @template {string|string[]} T
   * @param {T} tokens One or several token(s) to convert to token id(s).
   * @returns {T extends string ? number : number[]} The token id or list of token ids.
   */
  convert_tokens_to_ids(tokens) {
    if (typeof tokens === "string") {
      return (
        /** @type {any} */
        this._tokenizer.token_to_id(tokens)
      );
    } else {
      return (
        /** @type {any} */
        tokens.map((token) => this._tokenizer.token_to_id(token))
      );
    }
  }
  /**
   * Encode/tokenize the given text(s).
   * @template {string|string[]} TText
   * @template {boolean} [TReturnTensor=true]
   * @param {TText} text The text to tokenize.
   * @param {TokenizerCallOptions<TText, TReturnTensor>} [options] Additional tokenization options.
   * @returns {BatchEncoding<BatchEncodingItem<TText, TReturnTensor>>} Object to be passed to the model.
   */
  _call(text, options = {}) {
    const { text_pair = null, add_special_tokens = true, padding = false, return_token_type_ids = null } = options;
    let { truncation = null, max_length = null } = options;
    const return_tensor = (
      /** @type {TReturnTensor} */
      options.return_tensor ?? true
    );
    const isBatched = Array.isArray(text);
    let encodedTokens;
    if (isBatched) {
      if (text.length === 0) {
        throw Error("text array must be non-empty");
      }
      if (text_pair !== null) {
        if (!Array.isArray(text_pair)) {
          throw Error("text_pair must also be an array");
        } else if (text.length !== text_pair.length) {
          throw Error("text and text_pair must have the same length");
        }
        encodedTokens = text.map(
          (t, i) => this._encode_plus(t, { text_pair: text_pair[i], add_special_tokens, return_token_type_ids })
        );
      } else {
        encodedTokens = text.map((x) => this._encode_plus(x, { add_special_tokens, return_token_type_ids }));
      }
    } else {
      if (text === null || text === void 0) {
        throw Error("text may not be null or undefined");
      }
      if (Array.isArray(text_pair)) {
        throw Error(
          "When specifying `text_pair`, since `text` is a string, `text_pair` must also be a string (i.e., not an array)."
        );
      }
      encodedTokens = [this._encode_plus(text, { text_pair, add_special_tokens, return_token_type_ids })];
    }
    if (max_length === null) {
      max_length = this.model_max_length;
    } else if (truncation === null) {
      if (padding === true) {
        logger.warn(
          "`max_length` is ignored when `padding: true` and there is no truncation strategy. To pad to max length, use `padding: 'max_length'`."
        );
        max_length = this.model_max_length;
      } else if (padding === false) {
        logger.warn(
          "Truncation was not explicitly activated but `max_length` is provided a specific value, please use `truncation: true` to explicitly truncate examples to max length."
        );
        truncation = true;
      }
    }
    if (padding === true) {
      max_length = Math.min(max(encodedTokens.map((x) => x.input_ids.length))[0], max_length ?? Infinity);
    }
    max_length = Math.min(max_length, this.model_max_length ?? Infinity);
    if (padding || truncation) {
      for (let i = 0; i < encodedTokens.length; ++i) {
        if (encodedTokens[i].input_ids.length === max_length) {
          continue;
        } else if (encodedTokens[i].input_ids.length > max_length) {
          if (truncation) {
            truncateHelper(encodedTokens[i], max_length);
          }
        } else {
          if (padding) {
            padHelper(
              encodedTokens[i],
              max_length,
              (key) => key === "input_ids" ? this.pad_token_id : 0,
              this.padding_side
            );
          }
        }
      }
    }
    const result = {};
    if (return_tensor) {
      if (!(padding && truncation)) {
        if (encodedTokens.some((x) => {
          for (const key of Object.keys(x)) {
            if (x[key].length !== encodedTokens[0][key]?.length) {
              return true;
            }
          }
          return false;
        })) {
          throw Error(
            "Unable to create tensor, you should probably activate truncation and/or padding with 'padding=true' and 'truncation=true' to have batched tensors with the same length."
          );
        }
      }
      const dims = [encodedTokens.length, encodedTokens[0].input_ids.length];
      for (const key of Object.keys(encodedTokens[0])) {
        result[key] = new Tensor2(
          "int64",
          BigInt64Array.from(encodedTokens.flatMap((x) => x[key]).map(BigInt)),
          dims
        );
      }
    } else {
      for (const key of Object.keys(encodedTokens[0])) {
        result[key] = encodedTokens.map((x) => x[key]);
      }
      if (!isBatched) {
        for (const key of Object.keys(result)) {
          result[key] = result[key][0];
        }
      }
    }
    return (
      /** @type {BatchEncoding<BatchEncodingItem<TText, TReturnTensor>>} */
      result
    );
  }
  /**
   * Encodes a single text using the preprocessor pipeline of the tokenizer.
   *
   * @param {string|null} text The text to encode.
   * @returns {string[]|null} The encoded tokens.
   */
  _encode_text(text) {
    if (text === null) return null;
    return this._tokenizer.encode(text).tokens;
  }
  /**
   * Encodes a single text or a pair of texts using the model's tokenizer.
   *
   * @param {string} text The text to encode.
   * @param {Object} options An optional object containing the following properties:
   * @param {string|null} [options.text_pair=null] The optional second text to encode.
   * @param {boolean} [options.add_special_tokens=true] Whether or not to add the special tokens associated with the corresponding model.
   * @param {boolean|null} [options.return_token_type_ids=null] Whether to return token_type_ids.
   * @returns {{input_ids: number[], attention_mask: number[], token_type_ids?: number[]}} An object containing the encoded text.
   * @private
   */
  _encode_plus(text, { text_pair = null, add_special_tokens = true, return_token_type_ids = null } = {}) {
    const { ids, attention_mask, token_type_ids } = this._tokenizer.encode(text, {
      text_pair,
      add_special_tokens,
      return_token_type_ids: return_token_type_ids ?? this.return_token_type_ids
    });
    return {
      input_ids: ids,
      attention_mask,
      ...token_type_ids ? { token_type_ids } : {}
    };
  }
  /**
   * Converts a string into a sequence of tokens.
   * @param {string} text The sequence to be encoded.
   * @param {Object} options An optional object containing the following properties:
   * @param {string|null} [options.pair] A second sequence to be encoded with the first.
   * @param {boolean} [options.add_special_tokens=false] Whether or not to add the special tokens associated with the corresponding model.
   * @returns {string[]} The list of tokens.
   */
  tokenize(text, { pair = null, add_special_tokens = false } = {}) {
    return this._tokenizer.tokenize(text, { text_pair: pair, add_special_tokens });
  }
  /**
   * Encodes a single text or a pair of texts using the model's tokenizer.
   *
   * @param {string} text The text to encode.
   * @param {Object} options An optional object containing the following properties:
   * @param {string|null} [options.text_pair=null] The optional second text to encode.
   * @param {boolean} [options.add_special_tokens=true] Whether or not to add the special tokens associated with the corresponding model.
   * @param {boolean|null} [options.return_token_type_ids=null] Whether to return token_type_ids.
   * @returns {number[]} An array of token IDs representing the encoded text(s).
   */
  encode(text, { text_pair = null, add_special_tokens = true, return_token_type_ids = null } = {}) {
    return this._tokenizer.encode(text, {
      text_pair,
      add_special_tokens,
      return_token_type_ids
    }).ids;
  }
  /**
   * Decode a batch of tokenized sequences.
   * @param {number[][]|Tensor} batch List/Tensor of tokenized input sequences.
   * @param {Object} decode_args (Optional) Object with decoding arguments.
   * @returns {string[]} List of decoded sequences.
   */
  batch_decode(batch, decode_args = {}) {
    if (batch instanceof Tensor2) {
      batch = batch.tolist();
    }
    return batch.map((x) => this.decode(x, decode_args));
  }
  /**
   * Decodes a sequence of token IDs back to a string.
   *
   * @param {number[]|bigint[]|Tensor} token_ids List/Tensor of token IDs to decode.
   * @param {Object} [decode_args={}]
   * @param {boolean} [decode_args.skip_special_tokens=false] If true, special tokens are removed from the output string.
   * @param {boolean} [decode_args.clean_up_tokenization_spaces=true] If true, spaces before punctuations and abbreviated forms are removed.
   *
   * @returns {string} The decoded string.
   * @throws {Error} If `token_ids` is not a non-empty array of integers.
   */
  decode(token_ids, decode_args = {}) {
    if (token_ids instanceof Tensor2) {
      token_ids = prepareTensorForDecode(token_ids);
    }
    if (!Array.isArray(token_ids) || token_ids.length === 0 || !isIntegralNumber(token_ids[0])) {
      throw Error("token_ids must be a non-empty array of integers.");
    }
    return this.decode_single(token_ids, decode_args);
  }
  /**
   * Decode a single list of token ids to a string.
   * @param {number[]|bigint[]} token_ids List of token ids to decode
   * @param {Object} decode_args Optional arguments for decoding
   * @param {boolean} [decode_args.skip_special_tokens=false] Whether to skip special tokens during decoding
   * @param {boolean|null} [decode_args.clean_up_tokenization_spaces=null] Whether to clean up tokenization spaces during decoding.
   * If null, the value is set to `this.decoder.cleanup` if it exists, falling back to `this.clean_up_tokenization_spaces` if it exists, falling back to `true`.
   * @returns {string} The decoded string
   */
  decode_single(token_ids, { skip_special_tokens = false, clean_up_tokenization_spaces = null }) {
    return this._tokenizer.decode(token_ids, {
      skip_special_tokens,
      clean_up_tokenization_spaces
    });
  }
  /**
   * Retrieve the chat template string used for tokenizing chat messages. This template is used
   * internally by the `apply_chat_template` method and can also be used externally to retrieve the model's chat
   * template for better generation tracking.
   *
   * @param {Object} options An optional object containing the following properties:
   * @param {string|null} [options.chat_template=null]
   * A Jinja template or the name of a template to use for this conversion.
   * It is usually not necessary to pass anything to this argument,
   * as the model's template will be used by default.
   * @param {Object[]} [options.tools=null]
   * A list of tools (callable functions) that will be accessible to the model. If the template does not
   * support function calling, this argument will have no effect. Each tool should be passed as a JSON Schema,
   * giving the name, description and argument types for the tool. See our
   * [chat templating guide](https://huggingface.co/docs/transformers/main/en/chat_templating#automated-function-conversion-for-tool-use)
   * for more information.
   * @returns {string} The chat template string.
   */
  get_chat_template({ chat_template = null, tools = null } = {}) {
    if (this.chat_template && typeof this.chat_template === "object") {
      const template_dict = this.chat_template;
      if (chat_template !== null && Object.hasOwn(template_dict, chat_template)) {
        chat_template = template_dict[chat_template];
      } else if (chat_template === null) {
        if (tools !== null && "tool_use" in template_dict) {
          chat_template = template_dict["tool_use"];
        } else if ("default" in template_dict) {
          chat_template = template_dict["default"];
        } else {
          throw Error(
            `This model has multiple chat templates with no default specified! Please either pass a chat template or the name of the template you wish to use to the 'chat_template' argument. Available template names are ${Object.keys(template_dict).sort()}.`
          );
        }
      }
    } else if (chat_template === null) {
      if (this.chat_template) {
        chat_template = this.chat_template;
      } else {
        throw Error(
          "Cannot use apply_chat_template() because tokenizer.chat_template is not set and no template argument was passed! For information about writing templates and setting the tokenizer.chat_template attribute, please see the documentation at https://huggingface.co/docs/transformers/main/en/chat_templating"
        );
      }
    }
    return chat_template;
  }
  /**
   * Converts a list of message objects with `"role"` and `"content"` keys to a list of token
   * ids. This method is intended for use with chat models, and will read the tokenizer's chat_template attribute to
   * determine the format and control tokens to use when converting.
   *
   * See [here](https://huggingface.co/docs/transformers/chat_templating) for more information.
   *
   * **Example:** Applying a chat template to a conversation.
   *
   * ```javascript
   * import { AutoTokenizer } from "@huggingface/transformers";
   *
   * const tokenizer = await AutoTokenizer.from_pretrained("Xenova/mistral-tokenizer-v1");
   *
   * const chat = [
   *   { "role": "user", "content": "Hello, how are you?" },
   *   { "role": "assistant", "content": "I'm doing great. How can I help you today?" },
   *   { "role": "user", "content": "I'd like to show off how chat templating works!" },
   * ]
   *
   * const text = tokenizer.apply_chat_template(chat, { tokenize: false });
   * // "<s>[INST] Hello, how are you? [/INST]I'm doing great. How can I help you today?</s> [INST] I'd like to show off how chat templating works! [/INST]"
   *
   * const input_ids = tokenizer.apply_chat_template(chat, { tokenize: true, return_tensor: false });
   * // [1, 733, 16289, 28793, 22557, 28725, 910, 460, 368, 28804, 733, 28748, 16289, 28793, 28737, 28742, 28719, 2548, 1598, 28723, 1602, 541, 315, 1316, 368, 3154, 28804, 2, 28705, 733, 16289, 28793, 315, 28742, 28715, 737, 298, 1347, 805, 910, 10706, 5752, 1077, 3791, 28808, 733, 28748, 16289, 28793]
   * ```
   *
   * @param {Message[]} conversation A list of message objects with `"role"` and `"content"` keys,
   * representing the chat history so far.
   * @template {boolean} [TTokenize=true]
   * @template {boolean} [TReturnTensor=true]
   * @template {boolean} [TReturnDict=true]
   * @param {Object} [options] An optional object containing the following properties:
   * @param {string|null} [options.chat_template=null] A Jinja template to use for this conversion. If
   * this is not passed, the model's chat template will be used instead.
   * @param {Object[]} [options.tools=null]
   * A list of tools (callable functions) that will be accessible to the model. If the template does not
   * support function calling, this argument will have no effect. Each tool should be passed as a JSON Schema,
   * giving the name, description and argument types for the tool. See our
   * [chat templating guide](https://huggingface.co/docs/transformers/main/en/chat_templating#automated-function-conversion-for-tool-use)
   * for more information.
   * @param {Record<string, string>[]} [options.documents=null]
   * A list of dicts representing documents that will be accessible to the model if it is performing RAG
   * (retrieval-augmented generation). If the template does not support RAG, this argument will have no
   * effect. We recommend that each document should be a dict containing "title" and "text" keys. Please
   * see the RAG section of the [chat templating guide](https://huggingface.co/docs/transformers/main/en/chat_templating#arguments-for-RAG)
   * for examples of passing documents with chat templates.
   * @param {boolean} [options.add_generation_prompt=false] Whether to end the prompt with the token(s) that indicate
   * the start of an assistant message. This is useful when you want to generate a response from the model.
   * Note that this argument will be passed to the chat template, and so it must be supported in the
   * template for this argument to have any effect.
   * @param {TTokenize} [options.tokenize=true] Whether to tokenize the output. If false, the output will be a string.
   * @param {boolean} [options.padding=false] Whether to pad sequences to the maximum length. Has no effect if tokenize is false.
   * @param {boolean} [options.truncation=false] Whether to truncate sequences to the maximum length. Has no effect if tokenize is false.
   * @param {number|null} [options.max_length=null] Maximum length (in tokens) to use for padding or truncation. Has no effect if tokenize is false.
   * If not specified, the tokenizer's `max_length` attribute will be used as a default.
   * @param {TReturnTensor} [options.return_tensor=true] Whether to return the output as a Tensor or an Array. Has no effect if tokenize is false.
   * @param {TReturnDict} [options.return_dict=true] Whether to return a dictionary with named outputs. Has no effect if tokenize is false.
   * @param {Object} [options.tokenizer_kwargs={}] Additional options to pass to the tokenizer.
   * @returns {ApplyChatTemplateReturn<TTokenize, TReturnTensor, TReturnDict>} The tokenized output.
   */
  apply_chat_template(conversation, options = (
    /** @type {ApplyChatTemplateOptions<TTokenize, TReturnTensor, TReturnDict>} */
    {}
  )) {
    let {
      tools = null,
      documents = null,
      chat_template = null,
      add_generation_prompt = false,
      tokenize: tokenize2 = (
        /** @type {TTokenize} */
        true
      ),
      padding = false,
      truncation = false,
      max_length = null,
      return_tensor = (
        /** @type {TReturnTensor} */
        true
      ),
      return_dict = (
        /** @type {TReturnDict} */
        true
      ),
      tokenizer_kwargs = {},
      ...kwargs
    } = options;
    chat_template = this.get_chat_template({ chat_template, tools });
    if (typeof chat_template !== "string") {
      throw Error(`chat_template must be a string, but got ${typeof chat_template}`);
    }
    let compiledTemplate = this._compiled_template_cache.get(chat_template);
    if (compiledTemplate === void 0) {
      compiledTemplate = new Template(chat_template);
      this._compiled_template_cache.set(chat_template, compiledTemplate);
    }
    const special_tokens_map = /* @__PURE__ */ Object.create(null);
    for (const key of SPECIAL_TOKEN_ATTRIBUTES) {
      const value = getTokenFromConfig(this.config, key);
      if (value) {
        special_tokens_map[key] = value;
      }
    }
    const rendered = compiledTemplate.render({
      messages: conversation,
      add_generation_prompt,
      tools,
      documents,
      ...special_tokens_map,
      ...kwargs
    });
    if (tokenize2) {
      const out = this._call(rendered, {
        add_special_tokens: false,
        padding,
        truncation,
        max_length,
        return_tensor,
        ...tokenizer_kwargs
      });
      return (
        /** @type {ApplyChatTemplateReturn<TTokenize, TReturnTensor, TReturnDict>} */
        return_dict ? out : out.input_ids
      );
    }
    return (
      /** @type {ApplyChatTemplateReturn<TTokenize, TReturnTensor, TReturnDict>} */
      rendered
    );
  }
};
function _build_translation_inputs(self2, raw_inputs, tokenizer_options, generate_kwargs) {
  if (!("language_codes" in self2) || !Array.isArray(self2.language_codes)) {
    throw new Error(
      "Tokenizer must have `language_codes` attribute set and it should be an array of language ids."
    );
  }
  if (!("languageRegex" in self2) || !(self2.languageRegex instanceof RegExp)) {
    throw new Error("Tokenizer must have `languageRegex` attribute set and it should be a regular expression.");
  }
  if (!("lang_to_token" in self2) || typeof self2.lang_to_token !== "function") {
    throw new Error("Tokenizer must have `lang_to_token` attribute set and it should be a function.");
  }
  const src_lang_token = generate_kwargs.src_lang;
  const tgt_lang_token = generate_kwargs.tgt_lang;
  if (!self2.language_codes.includes(tgt_lang_token)) {
    throw new Error(
      `Target language code "${tgt_lang_token}" is not valid. Must be one of: {${self2.language_codes.join(", ")}}`
    );
  }
  if (src_lang_token !== void 0) {
    if (!self2.language_codes.includes(src_lang_token)) {
      throw new Error(
        `Source language code "${src_lang_token}" is not valid. Must be one of: {${self2.language_codes.join(", ")}}`
      );
    }
    for (const item of self2._tokenizer.post_processor.config.single) {
      if ("SpecialToken" in item && self2.languageRegex.test(item.SpecialToken.id)) {
        item.SpecialToken.id = self2.lang_to_token(src_lang_token);
        break;
      }
    }
  }
  generate_kwargs.forced_bos_token_id = self2._tokenizer.token_to_id(self2.lang_to_token(tgt_lang_token));
  return self2._call(raw_inputs, tokenizer_options);
}

// src/models/tokenizers.js
var tokenizers_exports = {};
__export(tokenizers_exports, {
  AlbertTokenizer: () => AlbertTokenizer,
  AutoTokenizer: () => AutoTokenizer,
  BartTokenizer: () => BartTokenizer,
  BertTokenizer: () => BertTokenizer,
  BlenderbotSmallTokenizer: () => BlenderbotSmallTokenizer,
  BlenderbotTokenizer: () => BlenderbotTokenizer,
  BloomTokenizer: () => BloomTokenizer,
  CLIPTokenizer: () => CLIPTokenizer,
  CamembertTokenizer: () => CamembertTokenizer,
  CodeGenTokenizer: () => CodeGenTokenizer,
  CodeLlamaTokenizer: () => CodeLlamaTokenizer,
  CohereAsrTokenizer: () => CohereAsrTokenizer,
  CohereTokenizer: () => CohereTokenizer,
  ConvBertTokenizer: () => ConvBertTokenizer,
  DebertaTokenizer: () => DebertaTokenizer,
  DebertaV2Tokenizer: () => DebertaV2Tokenizer,
  DistilBertTokenizer: () => DistilBertTokenizer,
  ElectraTokenizer: () => ElectraTokenizer,
  EsmTokenizer: () => EsmTokenizer,
  FalconTokenizer: () => FalconTokenizer,
  GPT2Tokenizer: () => GPT2Tokenizer,
  GPTNeoXTokenizer: () => GPTNeoXTokenizer,
  GemmaTokenizer: () => GemmaTokenizer,
  HerbertTokenizer: () => HerbertTokenizer,
  LlamaTokenizer: () => LlamaTokenizer,
  M2M100Tokenizer: () => M2M100Tokenizer,
  MBart50Tokenizer: () => MBart50Tokenizer,
  MBartTokenizer: () => MBartTokenizer,
  MPNetTokenizer: () => MPNetTokenizer,
  MarianTokenizer: () => MarianTokenizer,
  MgpstrTokenizer: () => MgpstrTokenizer,
  MobileBertTokenizer: () => MobileBertTokenizer,
  NllbTokenizer: () => NllbTokenizer,
  NougatTokenizer: () => NougatTokenizer,
  PreTrainedTokenizer: () => PreTrainedTokenizer,
  Qwen2Tokenizer: () => Qwen2Tokenizer,
  RoFormerTokenizer: () => RoFormerTokenizer,
  RobertaTokenizer: () => RobertaTokenizer,
  SiglipTokenizer: () => SiglipTokenizer,
  SpeechT5Tokenizer: () => SpeechT5Tokenizer,
  SqueezeBertTokenizer: () => SqueezeBertTokenizer,
  T5Tokenizer: () => T5Tokenizer,
  TokenizersBackend: () => PreTrainedTokenizer,
  VitsTokenizer: () => VitsTokenizer,
  Wav2Vec2CTCTokenizer: () => Wav2Vec2CTCTokenizer,
  WhisperTokenizer: () => WhisperTokenizer,
  XLMRobertaTokenizer: () => XLMRobertaTokenizer,
  XLMTokenizer: () => XLMTokenizer
});

// src/models/albert/tokenization_albert.js
var AlbertTokenizer = class extends PreTrainedTokenizer {
  return_token_type_ids = true;
};

// src/models/bart/tokenization_bart.js
var BartTokenizer = class extends PreTrainedTokenizer {
};

// src/models/bert/tokenization_bert.js
var BertTokenizer = class extends PreTrainedTokenizer {
  return_token_type_ids = true;
};

// src/models/blenderbot_small/tokenization_blenderbot_small.js
var BlenderbotSmallTokenizer = class extends PreTrainedTokenizer {
};

// src/models/blenderbot/tokenization_blenderbot.js
var BlenderbotTokenizer = class extends PreTrainedTokenizer {
};

// src/models/bloom/tokenization_bloom.js
var BloomTokenizer = class extends PreTrainedTokenizer {
};

// src/models/camembert/tokenization_camembert.js
var CamembertTokenizer = class extends PreTrainedTokenizer {
};

// src/models/clip/tokenization_clip.js
var CLIPTokenizer = class extends PreTrainedTokenizer {
};

// src/models/code_llama/tokenization_code_llama.js
var CodeLlamaTokenizer = class extends PreTrainedTokenizer {
};

// src/models/codegen/tokenization_codegen.js
var CodeGenTokenizer = class extends PreTrainedTokenizer {
};

// src/models/cohere/tokenization_cohere.js
var CohereTokenizer = class extends PreTrainedTokenizer {
};

// src/models/cohere_asr/tokenization_cohere_asr.js
var CohereAsrTokenizer = class extends PreTrainedTokenizer {
};

// src/models/convbert/tokenization_convbert.js
var ConvBertTokenizer = class extends PreTrainedTokenizer {
  return_token_type_ids = true;
};

// src/models/deberta_v2/tokenization_deberta_v2.js
var DebertaV2Tokenizer = class extends PreTrainedTokenizer {
  return_token_type_ids = true;
};

// src/models/deberta/tokenization_deberta.js
var DebertaTokenizer = class extends PreTrainedTokenizer {
  return_token_type_ids = true;
};

// src/models/distilbert/tokenization_distilbert.js
var DistilBertTokenizer = class extends PreTrainedTokenizer {
};

// src/models/electra/tokenization_electra.js
var ElectraTokenizer = class extends PreTrainedTokenizer {
  return_token_type_ids = true;
};

// src/models/esm/tokenization_esm.js
var EsmTokenizer = class extends PreTrainedTokenizer {
};

// src/models/falcon/tokenization_falcon.js
var FalconTokenizer = class extends PreTrainedTokenizer {
};

// src/models/gemma/tokenization_gemma.js
var GemmaTokenizer = class extends PreTrainedTokenizer {
};

// src/models/gpt_neox/tokenization_gpt_neox.js
var GPTNeoXTokenizer = class extends PreTrainedTokenizer {
};

// src/models/gpt2/tokenization_gpt2.js
var GPT2Tokenizer = class extends PreTrainedTokenizer {
};

// src/models/herbert/tokenization_herbert.js
var HerbertTokenizer = class extends PreTrainedTokenizer {
  return_token_type_ids = true;
};

// src/models/llama/tokenization_llama.js
var LlamaTokenizer = class extends PreTrainedTokenizer {
  padding_side = "left";
};

// src/models/m2m_100/tokenization_m2m_100.js
var M2M100Tokenizer = class extends PreTrainedTokenizer {
  constructor(tokenizerJSON, tokenizerConfig) {
    super(tokenizerJSON, tokenizerConfig);
    this.languageRegex = /^__[a-z]{2,3}__$/;
    this.language_codes = this.all_special_tokens.filter((x) => this.languageRegex.test(x)).map((x) => x.slice(2, -2));
    this.lang_to_token = (x) => `__${x}__`;
  }
  /**
   * Helper function to build translation inputs for an `M2M100Tokenizer`.
   * @param {string|string[]} raw_inputs The text to tokenize.
   * @param {Object} tokenizer_options Options to be sent to the tokenizer
   * @param {Object} generate_kwargs Generation options.
   * @returns {Object} Object to be passed to the model.
   */
  _build_translation_inputs(raw_inputs, tokenizer_options, generate_kwargs) {
    return _build_translation_inputs(this, raw_inputs, tokenizer_options, generate_kwargs);
  }
};

// src/models/marian/tokenization_marian.js
var MarianTokenizer = class extends PreTrainedTokenizer {
  /**
   * Create a new MarianTokenizer instance.
   * @param {Object} tokenizerJSON The JSON of the tokenizer.
   * @param {Object} tokenizerConfig The config of the tokenizer.
   */
  constructor(tokenizerJSON, tokenizerConfig) {
    super(tokenizerJSON, tokenizerConfig);
    this.languageRegex = /^(>>\w+<<)\s*/g;
    this.supported_language_codes = Array.from(this.get_vocab().keys()).filter((x) => this.languageRegex.test(x));
    logger.warn(
      'WARNING: `MarianTokenizer` is not yet supported by Hugging Face\'s "fast" tokenizers library. Therefore, you may experience slightly inaccurate results.'
    );
  }
  /**
   * Encodes a single text. Overriding this method is necessary since the language codes
   * must be removed before encoding with sentencepiece model.
   * @see https://github.com/huggingface/transformers/blob/12d51db243a00726a548a43cc333390ebae731e3/src/transformers/models/marian/tokenization_marian.py#L204-L213
   *
   * @param {string|null} text The text to encode.
   * @returns {string[]|null} The encoded tokens.
   */
  _encode_text(text) {
    if (text === null) return null;
    const [matchInfo, ...remainder] = text.trim().split(this.languageRegex);
    if (remainder.length === 0) {
      return super._encode_text(matchInfo);
    } else if (remainder.length === 2) {
      const [language, text2] = remainder;
      if (!this.supported_language_codes.includes(language)) {
        logger.warn(
          `Unsupported language code "${language}" detected, which may lead to unexpected behavior. Should be one of: ${JSON.stringify(this.supported_language_codes)}`
        );
      }
      return mergeArrays([language], super._encode_text(text2));
    }
  }
};

// src/models/mbart/tokenization_mbart.js
var MBartTokenizer = class extends PreTrainedTokenizer {
  constructor(tokenizerJSON, tokenizerConfig) {
    super(tokenizerJSON, tokenizerConfig);
    this.languageRegex = /^[a-z]{2}_[A-Z]{2}$/;
    this.language_codes = this.all_special_tokens.filter((x) => this.languageRegex.test(x)).map((x) => x);
    this.lang_to_token = (x) => x;
  }
  /**
   * Helper function to build translation inputs for an `MBartTokenizer`.
   * @param {string|string[]} raw_inputs The text to tokenize.
   * @param {Object} tokenizer_options Options to be sent to the tokenizer
   * @param {Object} generate_kwargs Generation options.
   * @returns {Object} Object to be passed to the model.
   */
  _build_translation_inputs(raw_inputs, tokenizer_options, generate_kwargs) {
    return _build_translation_inputs(this, raw_inputs, tokenizer_options, generate_kwargs);
  }
};

// src/models/mbart50/tokenization_mbart50.js
var MBart50Tokenizer = class extends MBartTokenizer {
};

// src/models/mgp_str/tokenization_mgp_str.js
var MgpstrTokenizer = class extends PreTrainedTokenizer {
};

// src/models/mobilebert/tokenization_mobilebert.js
var MobileBertTokenizer = class extends PreTrainedTokenizer {
  return_token_type_ids = true;
};

// src/models/mpnet/tokenization_mpnet.js
var MPNetTokenizer = class extends PreTrainedTokenizer {
};

// src/models/nllb/tokenization_nllb.js
var NllbTokenizer = class extends PreTrainedTokenizer {
  constructor(tokenizerJSON, tokenizerConfig) {
    super(tokenizerJSON, tokenizerConfig);
    this.languageRegex = /^[a-z]{3}_[A-Z][a-z]{3}$/;
    this.language_codes = this.all_special_tokens.filter((x) => this.languageRegex.test(x));
    this.lang_to_token = (x) => x;
  }
  /**
   * Helper function to build translation inputs for an `NllbTokenizer`.
   * @param {string|string[]} raw_inputs The text to tokenize.
   * @param {Object} tokenizer_options Options to be sent to the tokenizer
   * @param {Object} generate_kwargs Generation options.
   * @returns {Object} Object to be passed to the model.
   */
  _build_translation_inputs(raw_inputs, tokenizer_options, generate_kwargs) {
    return _build_translation_inputs(this, raw_inputs, tokenizer_options, generate_kwargs);
  }
};

// src/models/nougat/tokenization_nougat.js
var NougatTokenizer = class extends PreTrainedTokenizer {
};

// src/models/qwen2/tokenization_qwen2.js
var Qwen2Tokenizer = class extends PreTrainedTokenizer {
};

// src/models/roberta/tokenization_roberta.js
var RobertaTokenizer = class extends PreTrainedTokenizer {
};

// src/models/roformer/tokenization_roformer.js
var RoFormerTokenizer = class extends PreTrainedTokenizer {
  return_token_type_ids = true;
};

// src/models/siglip/tokenization_siglip.js
var SiglipTokenizer = class extends PreTrainedTokenizer {
};

// src/models/speecht5/tokenization_speecht5.js
var SpeechT5Tokenizer = class extends PreTrainedTokenizer {
};

// src/models/squeezebert/tokenization_squeezebert.js
var SqueezeBertTokenizer = class extends PreTrainedTokenizer {
  return_token_type_ids = true;
};

// src/models/t5/tokenization_t5.js
var T5Tokenizer = class extends PreTrainedTokenizer {
};

// src/models/vits/tokenization_vits.js
var VitsDecoder = class extends Decoder_default {
  /** @type {Decoder['decode_chain']} */
  decode_chain(tokens) {
    let decoded = "";
    for (let i = 1; i < tokens.length; i += 2) {
      decoded += tokens[i];
    }
    return [decoded];
  }
};
var VitsTokenizer = class extends PreTrainedTokenizer {
  constructor(tokenizerJSON, tokenizerConfig) {
    super(tokenizerJSON, tokenizerConfig);
    this._tokenizer.decoder = new VitsDecoder({ type: "VitsDecoder" });
  }
};

// src/models/wav2vec2/tokenization_wav2vec2.js
var Wav2Vec2CTCTokenizer = class extends PreTrainedTokenizer {
};

// src/models/whisper/common_whisper.js
var WHISPER_LANGUAGES = [
  ["en", "english"],
  ["zh", "chinese"],
  ["de", "german"],
  ["es", "spanish"],
  ["ru", "russian"],
  ["ko", "korean"],
  ["fr", "french"],
  ["ja", "japanese"],
  ["pt", "portuguese"],
  ["tr", "turkish"],
  ["pl", "polish"],
  ["ca", "catalan"],
  ["nl", "dutch"],
  ["ar", "arabic"],
  ["sv", "swedish"],
  ["it", "italian"],
  ["id", "indonesian"],
  ["hi", "hindi"],
  ["fi", "finnish"],
  ["vi", "vietnamese"],
  ["he", "hebrew"],
  ["uk", "ukrainian"],
  ["el", "greek"],
  ["ms", "malay"],
  ["cs", "czech"],
  ["ro", "romanian"],
  ["da", "danish"],
  ["hu", "hungarian"],
  ["ta", "tamil"],
  ["no", "norwegian"],
  ["th", "thai"],
  ["ur", "urdu"],
  ["hr", "croatian"],
  ["bg", "bulgarian"],
  ["lt", "lithuanian"],
  ["la", "latin"],
  ["mi", "maori"],
  ["ml", "malayalam"],
  ["cy", "welsh"],
  ["sk", "slovak"],
  ["te", "telugu"],
  ["fa", "persian"],
  ["lv", "latvian"],
  ["bn", "bengali"],
  ["sr", "serbian"],
  ["az", "azerbaijani"],
  ["sl", "slovenian"],
  ["kn", "kannada"],
  ["et", "estonian"],
  ["mk", "macedonian"],
  ["br", "breton"],
  ["eu", "basque"],
  ["is", "icelandic"],
  ["hy", "armenian"],
  ["ne", "nepali"],
  ["mn", "mongolian"],
  ["bs", "bosnian"],
  ["kk", "kazakh"],
  ["sq", "albanian"],
  ["sw", "swahili"],
  ["gl", "galician"],
  ["mr", "marathi"],
  ["pa", "punjabi"],
  ["si", "sinhala"],
  ["km", "khmer"],
  ["sn", "shona"],
  ["yo", "yoruba"],
  ["so", "somali"],
  ["af", "afrikaans"],
  ["oc", "occitan"],
  ["ka", "georgian"],
  ["be", "belarusian"],
  ["tg", "tajik"],
  ["sd", "sindhi"],
  ["gu", "gujarati"],
  ["am", "amharic"],
  ["yi", "yiddish"],
  ["lo", "lao"],
  ["uz", "uzbek"],
  ["fo", "faroese"],
  ["ht", "haitian creole"],
  ["ps", "pashto"],
  ["tk", "turkmen"],
  ["nn", "nynorsk"],
  ["mt", "maltese"],
  ["sa", "sanskrit"],
  ["lb", "luxembourgish"],
  ["my", "myanmar"],
  ["bo", "tibetan"],
  ["tl", "tagalog"],
  ["mg", "malagasy"],
  ["as", "assamese"],
  ["tt", "tatar"],
  ["haw", "hawaiian"],
  ["ln", "lingala"],
  ["ha", "hausa"],
  ["ba", "bashkir"],
  ["jw", "javanese"],
  ["su", "sundanese"]
];
var WHISPER_LANGUAGE_MAPPING = new Map(WHISPER_LANGUAGES);
var WHISPER_TO_LANGUAGE_CODE_MAPPING = new Map([
  ...WHISPER_LANGUAGES.map(([k2, v]) => [v, k2]),
  ...[
    ["burmese", "my"],
    ["valencian", "ca"],
    ["flemish", "nl"],
    ["haitian", "ht"],
    ["letzeburgesch", "lb"],
    ["pushto", "ps"],
    ["panjabi", "pa"],
    ["moldavian", "ro"],
    ["moldovan", "ro"],
    ["sinhalese", "si"],
    ["castilian", "es"]
  ]
]);
function whisper_language_to_code(language) {
  language = language.toLowerCase();
  let language_code = WHISPER_TO_LANGUAGE_CODE_MAPPING.get(language);
  if (language_code === void 0) {
    const language_special_token = language.match(/^<\|([a-z]{2})\|>$/);
    if (language_special_token) {
      language = language_special_token[1];
    }
    if (WHISPER_LANGUAGE_MAPPING.has(language)) {
      language_code = language;
    } else {
      const is_language_code = language.length === 2;
      const langs = is_language_code ? WHISPER_LANGUAGE_MAPPING.keys() : WHISPER_LANGUAGE_MAPPING.values();
      throw new Error(
        `Language "${language}" is not supported. Must be one of: ${JSON.stringify(Array.from(langs))}`
      );
    }
  }
  return language_code;
}

// src/models/whisper/tokenization_whisper.js
var PUNCTUATION_REGEX2 = "\\p{P}\\u0021-\\u002F\\u003A-\\u0040\\u005B-\\u0060\\u007B-\\u007E";
var PUNCTUATION_ONLY_REGEX = new RegExp(`^[${PUNCTUATION_REGEX2}]+$`, "gu");
var TIMESTAMP_MERGE_TOLERANCE = 0.1;
var WhisperTokenizer = class extends PreTrainedTokenizer {
  get timestamp_begin() {
    return this._tokenizer.token_to_id("<|notimestamps|>") + 1;
  }
  /**
   * Decodes automatic speech recognition (ASR) sequences.
   * @param {Array<{tokens: bigint[], token_timestamps?: number[], stride: number[]}>} sequences The sequences to decode.
   * @param {Object} options The options to use for decoding.
   * @returns {Array<string|{chunks?: undefined|Array<{language: string|null, timestamp: Array<number|null>, text: string}>}>} The decoded sequences.
   */
  _decode_asr(sequences, { return_timestamps = false, return_language = false, time_precision = null, force_full_sequences = true } = {}) {
    if (time_precision === null) {
      throw Error("Must specify time_precision");
    }
    let last_language = null;
    const returnWordTimestamps = return_timestamps === "word";
    function new_chunk() {
      return { language: last_language, timestamp: [null, null], text: "" };
    }
    const chunks = [];
    let chunk2 = new_chunk();
    let time_offset = 0;
    const timestamp_begin = this.timestamp_begin;
    const total_timestamp_tokens = 1500;
    const timestamp_end = timestamp_begin + total_timestamp_tokens;
    let previous_tokens = [];
    let previous_token_timestamps = [];
    let skip = false;
    let right_stride_start = null;
    const all_special_ids = new Set(this.all_special_ids);
    for (const output of sequences) {
      const token_ids = output.tokens;
      const token_timestamps = returnWordTimestamps ? output.token_timestamps : null;
      let last_timestamp = null;
      let first_timestamp = timestamp_begin;
      if ("stride" in output) {
        const [chunk_len, stride_left, stride_right] = output.stride;
        time_offset -= stride_left;
        right_stride_start = chunk_len - stride_right;
        if (stride_left) {
          first_timestamp = stride_left / time_precision + timestamp_begin;
        }
        if (stride_right) {
          for (let i = token_ids.length - 1; i >= 0; --i) {
            const token = Number(token_ids[i]);
            if (token >= timestamp_begin) {
              if (last_timestamp !== null && (token - timestamp_begin) * time_precision < right_stride_start) {
                break;
              }
              last_timestamp = token;
            }
          }
        }
      }
      let current_tokens = [];
      let current_token_timestamps = [];
      for (let i = 0; i < token_ids.length; ++i) {
        const token = Number(token_ids[i]);
        if (all_special_ids.has(token)) {
          const text = this.decode([token]);
          const language = WHISPER_LANGUAGE_MAPPING.get(text.slice(2, -2));
          if (language !== void 0) {
            if (last_language !== null && language !== last_language && !return_timestamps) {
              previous_tokens.push(current_tokens);
              const resolved_tokens = this.findLongestCommonSequence(previous_tokens)[0];
              const resolved_text = this.decode(resolved_tokens);
              chunk2.text = resolved_text;
              chunks.push(chunk2);
              previous_tokens = [];
              current_tokens = [];
              chunk2 = new_chunk();
            }
            last_language = chunk2.language = language;
          } else {
          }
        } else if (token >= timestamp_begin && token <= timestamp_end) {
          const time = (token - timestamp_begin) * time_precision + time_offset;
          const rounded_time = round(time, 2);
          if (last_timestamp !== null && token >= last_timestamp) {
            skip = true;
          } else if (skip || previous_tokens.length > 0 && token < first_timestamp) {
            skip = false;
          } else if (chunk2.timestamp[0] === null) {
            chunk2.timestamp[0] = rounded_time;
          } else {
            if (rounded_time === chunk2.timestamp[0]) {
            } else {
              chunk2.timestamp[1] = rounded_time;
              previous_tokens.push(current_tokens);
              if (returnWordTimestamps) {
                previous_token_timestamps.push(current_token_timestamps);
              }
              const [resolved_tokens, resolved_token_timestamps] = this.findLongestCommonSequence(
                previous_tokens,
                previous_token_timestamps
              );
              const resolved_text = this.decode(resolved_tokens);
              chunk2.text = resolved_text;
              if (returnWordTimestamps) {
                chunk2.words = this.collateWordTimestamps(
                  resolved_tokens,
                  resolved_token_timestamps,
                  last_language
                );
                if (chunk2.words.length > 0 && chunk2.timestamp[1] !== null) {
                  for (const word of chunk2.words) {
                    if (word.timestamp[1] > chunk2.timestamp[1] && chunk2.timestamp[1] >= word.timestamp[0]) {
                      word.timestamp[1] = chunk2.timestamp[1];
                    }
                  }
                }
              }
              chunks.push(chunk2);
              previous_tokens = [];
              current_tokens = [];
              previous_token_timestamps = [];
              current_token_timestamps = [];
              chunk2 = new_chunk();
            }
          }
        } else {
          current_tokens.push(token);
          if (returnWordTimestamps) {
            let start_time = round(token_timestamps[i] + time_offset, 2);
            let end_time;
            if (i + 1 < token_timestamps.length) {
              end_time = round(token_timestamps[i + 1] + time_offset, 2);
              const decoded_text = this.decode([token]);
              if (PUNCTUATION_ONLY_REGEX.test(decoded_text)) {
                end_time = round(Math.min(start_time + time_precision, end_time), 2);
              }
            } else {
              end_time = null;
            }
            current_token_timestamps.push([start_time, end_time]);
          }
        }
      }
      if ("stride" in output) {
        const [chunk_len, stride_left, stride_right] = output.stride;
        time_offset += chunk_len - stride_right;
      }
      if (current_tokens.length > 0) {
        previous_tokens.push(current_tokens);
        if (returnWordTimestamps) {
          previous_token_timestamps.push(current_token_timestamps);
        }
      } else if (previous_tokens.every((p) => p.length === 0)) {
        chunk2 = new_chunk();
        previous_tokens = [];
        current_tokens = [];
        previous_token_timestamps = [];
        current_token_timestamps = [];
      }
    }
    if (previous_tokens.length > 0) {
      if (force_full_sequences && return_timestamps) {
        throw new Error(
          "Whisper did not predict an ending timestamp, which can happen if audio is cut off in the middle of a word. Also make sure WhisperTimeStampLogitsProcessor was used during generation."
        );
      }
      const [resolved_tokens, resolved_token_timestamps] = this.findLongestCommonSequence(
        previous_tokens,
        previous_token_timestamps
      );
      const resolved_text = this.decode(resolved_tokens);
      chunk2.text = resolved_text;
      if (returnWordTimestamps) {
        chunk2.words = this.collateWordTimestamps(resolved_tokens, resolved_token_timestamps, last_language);
      }
      chunks.push(chunk2);
    }
    let optional = /* @__PURE__ */ Object.create(null);
    const full_text = chunks.map((chunk3) => chunk3.text).join("");
    if (return_timestamps || return_language) {
      for (let i = 0; i < chunks.length; ++i) {
        const chunk3 = chunks[i];
        if (!return_timestamps) {
          delete chunk3["timestamp"];
        }
        if (!return_language) {
          delete chunk3["language"];
        }
      }
      if (returnWordTimestamps) {
        const new_chunks = [];
        for (const chunk3 of chunks) {
          for (const word of chunk3.words) {
            new_chunks.push(word);
          }
        }
        optional = { chunks: new_chunks };
      } else {
        optional = { chunks };
      }
    }
    return [full_text, optional];
  }
  /**
   * Finds the longest common sequence among the provided sequences.
   * @param {number[][]} sequences An array of sequences of token ids to compare.
   * @returns {number[][]} The longest common sequence found.
   * @throws {Error} If there is a bug within the function.
   * @private
   */
  findLongestCommonSequence(sequences, token_timestamp_sequences = null) {
    let leftSequence = sequences[0];
    let leftLength = leftSequence.length;
    let totalSequence = [];
    const use_token_timestamp_sequences = Array.isArray(token_timestamp_sequences) && token_timestamp_sequences.length > 0;
    let total_token_timestamp_sequence = use_token_timestamp_sequences ? [] : null;
    let left_token_timestamp_sequence = use_token_timestamp_sequences ? token_timestamp_sequences[0] : null;
    for (let i = 1; i < sequences.length; ++i) {
      const rightSequence = sequences[i];
      let max2 = 0;
      let maxIndices = [leftLength, leftLength, 0, 0];
      const rightLength = rightSequence.length;
      for (let j = 1; j < leftLength + rightLength; ++j) {
        const leftStart2 = Math.max(0, leftLength - j);
        const leftStop2 = Math.min(leftLength, leftLength + rightLength - j);
        const left = leftSequence.slice(leftStart2, leftStop2);
        const rightStart2 = Math.max(0, j - leftLength);
        const rightStop2 = Math.min(rightLength, j);
        const right = rightSequence.slice(rightStart2, rightStop2);
        if (left.length !== right.length) {
          throw new Error(
            "There is a bug within whisper `decode_asr` function, please report it. Dropping to prevent bad inference."
          );
        }
        let matches;
        if (use_token_timestamp_sequences) {
          matches = left.filter(
            (elem, idx) => elem === right[idx] && left_token_timestamp_sequence[leftStart2 + idx][0] - TIMESTAMP_MERGE_TOLERANCE <= token_timestamp_sequences[i][rightStart2 + idx][0]
          ).length;
        } else {
          matches = left.filter((elem, idx) => elem === right[idx]).length;
        }
        const eps = j / 1e4;
        const matching = matches / j + eps;
        if (matches > 1 && matching > max2) {
          max2 = matching;
          maxIndices = [leftStart2, leftStop2, rightStart2, rightStop2];
        }
      }
      const [leftStart, leftStop, rightStart, rightStop] = maxIndices;
      const leftMid = Math.floor((leftStop + leftStart) / 2);
      let rightMid = Math.floor((rightStop + rightStart) / 2);
      if (use_token_timestamp_sequences && max2 === 0 && leftLength > 0) {
        const lastLeftTs = left_token_timestamp_sequence[leftLength - 1][0];
        const idx = token_timestamp_sequences[i].findIndex((ts2) => ts2[0] >= lastLeftTs);
        rightMid = idx === -1 ? rightSequence.length : idx;
      }
      totalSequence.push(...leftSequence.slice(0, leftMid));
      leftSequence = rightSequence.slice(rightMid);
      leftLength = leftSequence.length;
      if (use_token_timestamp_sequences) {
        total_token_timestamp_sequence.push(...left_token_timestamp_sequence.slice(0, leftMid));
        left_token_timestamp_sequence = token_timestamp_sequences[i].slice(rightMid);
      }
    }
    totalSequence.push(...leftSequence);
    if (use_token_timestamp_sequences) {
      total_token_timestamp_sequence.push(...left_token_timestamp_sequence);
      return [totalSequence, total_token_timestamp_sequence];
    } else {
      return [totalSequence, []];
    }
  }
  /** @private */
  collateWordTimestamps(tokens, token_timestamps, language) {
    const [words, _, token_indices] = this.combineTokensIntoWords(tokens, language);
    const timings = [];
    for (let i = 0; i < words.length; ++i) {
      const indices = token_indices[i];
      timings.push({
        text: words[i],
        timestamp: [token_timestamps[indices.at(0)][0], token_timestamps[indices.at(-1)][1]]
      });
    }
    return timings;
  }
  /**
   * Groups tokens by word. Returns a tuple containing a list of strings with the words,
   * and a list of `token_id` sequences with the tokens making up each word.
   * @param {number[]} tokens
   * @param {string} [language]
   * @param {string} prepend_punctionations
   * @param {string} append_punctuations
   *
   * @private
   */
  combineTokensIntoWords(tokens, language, prepend_punctionations = `"'\u201C\xA1\xBF([{-`, append_punctuations = `"'.\u3002,\uFF0C!\uFF01?\uFF1F:\uFF1A\u201D)]}\u3001`) {
    language = language ?? "english";
    let words, word_tokens, token_indices;
    if (["chinese", "japanese", "thai", "lao", "myanmar"].includes(language)) {
      [words, word_tokens, token_indices] = this.splitTokensOnUnicode(tokens);
    } else {
      [words, word_tokens, token_indices] = this.splitTokensOnSpaces(tokens);
    }
    return this.mergePunctuations(words, word_tokens, token_indices, prepend_punctionations, append_punctuations);
  }
  /** @type {PreTrainedTokenizer['decode']} */
  decode(token_ids, decode_args) {
    let text;
    if (decode_args?.decode_with_timestamps) {
      if (token_ids instanceof Tensor2) {
        token_ids = prepareTensorForDecode(token_ids);
      }
      text = this.decodeWithTimestamps(token_ids, decode_args);
    } else {
      text = super.decode(token_ids, decode_args);
    }
    return text;
  }
  /**
   * @param {number[]|bigint[]} token_ids List of token IDs to decode.
   * @param {Object} decode_args Optional arguments for decoding
   * @private
   */
  decodeWithTimestamps(token_ids, decode_args) {
    const time_precision = decode_args?.time_precision ?? 0.02;
    const timestamp_begin = this.all_special_ids.at(-1) + 1;
    let outputs = [[]];
    for (let token of token_ids) {
      token = Number(token);
      if (token >= timestamp_begin) {
        const timestamp = ((token - timestamp_begin) * time_precision).toFixed(2);
        outputs.push(`<|${timestamp}|>`);
        outputs.push([]);
      } else {
        outputs[outputs.length - 1].push(token);
      }
    }
    outputs = outputs.map((s) => typeof s === "string" ? s : super.decode(s, decode_args));
    return outputs.join("");
  }
  /**
   * Combine tokens into words by splitting at any position where the tokens are decoded as valid unicode points.
   * @param {number[]} tokens
   * @returns {*}
   * @private
   */
  splitTokensOnUnicode(tokens) {
    const decoded_full = this.decode(tokens, {
      // @ts-ignore
      decode_with_timestamps: true
    });
    const replacement_char = "\uFFFD";
    const words = [];
    const word_tokens = [];
    const token_indices = [];
    let current_tokens = [];
    let current_indices = [];
    let unicode_offset = 0;
    for (let token_idx = 0; token_idx < tokens.length; ++token_idx) {
      const token = tokens[token_idx];
      current_tokens.push(token);
      current_indices.push(token_idx);
      const decoded = this.decode(current_tokens, {
        // @ts-ignore
        decode_with_timestamps: true
      });
      if (!decoded.includes(replacement_char) || decoded_full[unicode_offset + decoded.indexOf(replacement_char)] === replacement_char) {
        words.push(decoded);
        word_tokens.push(current_tokens);
        token_indices.push(current_indices);
        current_tokens = [];
        current_indices = [];
        unicode_offset += decoded.length;
      }
    }
    return [words, word_tokens, token_indices];
  }
  /**
   * Combine tokens into words by splitting at whitespace and punctuation tokens.
   * @param {number[]} tokens
   * @private
   */
  splitTokensOnSpaces(tokens) {
    const [subwords, subword_tokens_list, subword_indices_list] = this.splitTokensOnUnicode(tokens);
    const words = [];
    const word_tokens = [];
    const token_indices = [];
    for (let i = 0; i < subwords.length; ++i) {
      const subword = subwords[i];
      const subword_tokens = subword_tokens_list[i];
      const subword_indices = subword_indices_list[i];
      const special = subword_tokens[0] >= this._tokenizer.token_to_id("<|endoftext|>");
      const with_space = subword.startsWith(" ");
      const trimmed = subword.trim();
      const punctuation = PUNCTUATION_ONLY_REGEX.test(trimmed);
      if (special || with_space || punctuation || words.length === 0) {
        words.push(subword);
        word_tokens.push(subword_tokens);
        token_indices.push(subword_indices);
      } else {
        const ix = words.length - 1;
        words[ix] += subword;
        word_tokens[ix].push(...subword_tokens);
        token_indices[ix].push(...subword_indices);
      }
    }
    return [words, word_tokens, token_indices];
  }
  /**
   * Merges punctuation tokens with neighboring words.
   * @param {string[]} words
   * @param {number[][]} tokens
   * @param {number[][]} indices
   * @param {string} prepended
   * @param {string} appended
   * @private
   */
  mergePunctuations(words, tokens, indices, prepended, appended) {
    const newWords = structuredClone(words);
    const newTokens = structuredClone(tokens);
    const newIndices = structuredClone(indices);
    let i = newWords.length - 2;
    let j = newWords.length - 1;
    while (i >= 0) {
      if (newWords[i].startsWith(" ") && prepended.includes(newWords[i].trim())) {
        newWords[j] = newWords[i] + newWords[j];
        newTokens[j] = mergeArrays(newTokens[i], newTokens[j]);
        newIndices[j] = mergeArrays(newIndices[i], newIndices[j]);
        newWords[i] = "";
        newTokens[i] = [];
        newIndices[i] = [];
      } else {
        j = i;
      }
      --i;
    }
    i = 0;
    j = 1;
    while (j < newWords.length) {
      if (!newWords[i].endsWith(" ") && appended.includes(newWords[j])) {
        newWords[i] += newWords[j];
        newTokens[i] = mergeArrays(newTokens[i], newTokens[j]);
        newIndices[i] = mergeArrays(newIndices[i], newIndices[j]);
        newWords[j] = "";
        newTokens[j] = [];
        newIndices[j] = [];
      } else {
        i = j;
      }
      ++j;
    }
    return [
      newWords.filter((x) => x),
      newTokens.filter((x) => x.length > 0),
      newIndices.filter((x) => x.length > 0)
    ];
  }
};

// src/models/xlm_roberta/tokenization_xlm_roberta.js
var XLMRobertaTokenizer = class extends PreTrainedTokenizer {
};

// src/models/xlm/tokenization_xlm.js
var XLMTokenizer = class extends PreTrainedTokenizer {
  return_token_type_ids = true;
  constructor(tokenizerJSON, tokenizerConfig) {
    super(tokenizerJSON, tokenizerConfig);
    logger.warn(
      'WARNING: `XLMTokenizer` is not yet supported by Hugging Face\'s "fast" tokenizers library. Therefore, you may experience slightly inaccurate results.'
    );
  }
};

// src/models/auto/tokenization_auto.js
var AutoTokenizer = class {
  /**
   * Instantiate one of the tokenizer classes of the library from a pretrained model.
   *
   * The tokenizer class to instantiate is selected based on the `tokenizer_class` property of the config object
   * (either passed as an argument or loaded from `pretrained_model_name_or_path` if possible)
   *
   * @param {string} pretrained_model_name_or_path The name or path of the pretrained model. Can be either:
   * - A string, the *model id* of a pretrained tokenizer hosted inside a model repo on huggingface.co.
   *   Valid model ids can be located at the root-level, like `bert-base-uncased`, or namespaced under a
   *   user or organization name, like `dbmdz/bert-base-german-cased`.
   * - A path to a *directory* containing tokenizer files, e.g., `./my_model_directory/`.
   * @param {import('../../tokenization_utils.js').PretrainedTokenizerOptions} options Additional options for loading the tokenizer.
   *
   * @returns {Promise<PreTrainedTokenizer>} A new instance of the PreTrainedTokenizer class.
   */
  static async from_pretrained(pretrained_model_name_or_path, { progress_callback = null, config = null, cache_dir = null, local_files_only = false, revision = "main" } = {}) {
    const [tokenizerJSON, tokenizerConfig] = await loadTokenizer(pretrained_model_name_or_path, {
      progress_callback,
      config,
      cache_dir,
      local_files_only,
      revision
    });
    const tokenizerName = tokenizerConfig.tokenizer_class?.replace(/Fast$/, "") ?? "PreTrainedTokenizer";
    let cls = tokenizers_exports[tokenizerName];
    if (!cls) {
      logger.warn(`Unknown tokenizer class "${tokenizerName}", attempting to construct from base class.`);
      cls = PreTrainedTokenizer;
    }
    return new cls(tokenizerJSON, tokenizerConfig);
  }
};

// src/utils/constants.js
var GITHUB_ISSUE_URL = "https://github.com/huggingface/transformers.js/issues/new/choose";
var FEATURE_EXTRACTOR_NAME = "preprocessor_config.json";
var IMAGE_PROCESSOR_NAME = FEATURE_EXTRACTOR_NAME;
var PROCESSOR_NAME = "processor_config.json";
var CHAT_TEMPLATE_NAME = "chat_template.jinja";

// src/processing_utils.js
var Processor = class extends Callable {
  static classes = ["image_processor_class", "tokenizer_class", "feature_extractor_class"];
  static uses_processor_config = false;
  static uses_chat_template_file = false;
  /**
   * Creates a new Processor with the given components
   * @param {Object} config
   * @param {Record<string, Object>} components
   * @param {string} chat_template
   */
  constructor(config, components, chat_template) {
    super();
    this.config = config;
    this.components = components;
    this.chat_template = chat_template;
  }
  /**
   * @returns {import('./image_processors_utils.js').ImageProcessor|undefined} The image processor of the processor, if it exists.
   */
  get image_processor() {
    return this.components.image_processor;
  }
  /**
   * @returns {PreTrainedTokenizer|undefined} The tokenizer of the processor, if it exists.
   */
  get tokenizer() {
    return this.components.tokenizer;
  }
  /**
   * @returns {import('./feature_extraction_utils.js').FeatureExtractor|undefined} The feature extractor of the processor, if it exists.
   */
  get feature_extractor() {
    return this.components.feature_extractor;
  }
  /**
   * @param {Parameters<PreTrainedTokenizer['apply_chat_template']>[0]} messages
   * @param {Parameters<PreTrainedTokenizer['apply_chat_template']>[1]} options
   * @returns {ReturnType<PreTrainedTokenizer['apply_chat_template']>}
   */
  apply_chat_template(messages, options = {}) {
    if (!this.tokenizer) {
      throw new Error("Unable to apply chat template without a tokenizer.");
    }
    return this.tokenizer.apply_chat_template(messages, {
      tokenize: false,
      // default to false
      chat_template: this.chat_template ?? void 0,
      ...options
    });
  }
  /**
   * @param {Parameters<PreTrainedTokenizer['batch_decode']>} args
   * @returns {ReturnType<PreTrainedTokenizer['batch_decode']>}
   */
  batch_decode(...args) {
    if (!this.tokenizer) {
      throw new Error("Unable to decode without a tokenizer.");
    }
    return this.tokenizer.batch_decode(...args);
  }
  /**
   * @param {Parameters<PreTrainedTokenizer['decode']>} args
   * @returns {ReturnType<PreTrainedTokenizer['decode']>}
   */
  decode(...args) {
    if (!this.tokenizer) {
      throw new Error("Unable to decode without a tokenizer.");
    }
    return this.tokenizer.decode(...args);
  }
  /**
   * Calls the feature_extractor function with the given input.
   * @param {any} input The input to extract features from.
   * @param {...any} args Additional arguments.
   * @returns {Promise<any>} A Promise that resolves with the extracted features.
   */
  async _call(input, ...args) {
    for (const item of [this.image_processor, this.feature_extractor, this.tokenizer]) {
      if (item) {
        return item(input, ...args);
      }
    }
    throw new Error("No image processor, feature extractor, or tokenizer found.");
  }
  /**
   * Instantiate one of the processor classes of the library from a pretrained model.
   *
   * The processor class to instantiate is selected based on the `image_processor_type` (or `feature_extractor_type`; legacy)
   * property of the config object (either passed as an argument or loaded from `pretrained_model_name_or_path` if possible)
   *
   * @param {string} pretrained_model_name_or_path The name or path of the pretrained model. Can be either:
   * - A string, the *model id* of a pretrained processor hosted inside a model repo on huggingface.co.
   *   Valid model ids can be located at the root-level, like `bert-base-uncased`, or namespaced under a
   *   user or organization name, like `dbmdz/bert-base-german-cased`.
   * - A path to a *directory* containing processor files, e.g., `./my_model_directory/`.
   * @param {PretrainedProcessorOptions} options Additional options for loading the processor.
   *
   * @returns {Promise<Processor>} A new instance of the Processor class.
   */
  static async from_pretrained(pretrained_model_name_or_path, options = {}) {
    const [config, components, chat_template] = await Promise.all([
      // TODO:
      this.uses_processor_config ? getModelJSON(pretrained_model_name_or_path, PROCESSOR_NAME, true, options) : {},
      Promise.all(
        this.classes.filter((cls) => cls in this).map(async (cls) => {
          const component = await this[cls].from_pretrained(pretrained_model_name_or_path, options);
          return [cls.replace(/_class$/, ""), component];
        })
      ).then(Object.fromEntries),
      this.uses_chat_template_file ? getModelText(pretrained_model_name_or_path, CHAT_TEMPLATE_NAME, true, options) : null
    ]);
    return new this(config, components, chat_template);
  }
};

// src/models/processors.js
var processors_exports = {};
__export(processors_exports, {
  ChatterboxProcessor: () => ChatterboxProcessor,
  CohereAsrProcessor: () => CohereAsrProcessor,
  Florence2Processor: () => Florence2Processor,
  Gemma3Processor: () => Gemma3Processor,
  Gemma3nProcessor: () => Gemma3nProcessor,
  Gemma4Processor: () => Gemma4Processor,
  Glm46VProcessor: () => Glm46VProcessor,
  GraniteSpeechProcessor: () => GraniteSpeechProcessor,
  GroundingDinoProcessor: () => GroundingDinoProcessor,
  Idefics3Processor: () => Idefics3Processor,
  JinaCLIPProcessor: () => JinaCLIPProcessor,
  Lfm2VlProcessor: () => Lfm2VlProcessor,
  LlavaProcessor: () => LlavaProcessor,
  MgpstrProcessor: () => MgpstrProcessor,
  MoonshineProcessor: () => MoonshineProcessor,
  OwlViTProcessor: () => OwlViTProcessor,
  PaliGemmaProcessor: () => PaliGemmaProcessor,
  Phi3VProcessor: () => Phi3VProcessor,
  PixtralProcessor: () => PixtralProcessor,
  Processor: () => Processor,
  PyAnnoteProcessor: () => PyAnnoteProcessor,
  Qwen2VLProcessor: () => Qwen2VLProcessor,
  Qwen2_5_VLProcessor: () => Qwen2_5_VLProcessor,
  Qwen3VLProcessor: () => Qwen3VLProcessor,
  Sam2Processor: () => Sam2Processor,
  Sam2VideoProcessor: () => Sam2VideoProcessor,
  SamProcessor: () => SamProcessor,
  SmolVLMProcessor: () => Idefics3Processor,
  SpeechT5Processor: () => SpeechT5Processor,
  UltravoxProcessor: () => UltravoxProcessor,
  VLChatProcessor: () => VLChatProcessor,
  VoxtralProcessor: () => VoxtralProcessor,
  VoxtralRealtimeProcessor: () => VoxtralRealtimeProcessor,
  Wav2Vec2Processor: () => Wav2Vec2Processor,
  Wav2Vec2ProcessorWithLM: () => Wav2Vec2ProcessorWithLM,
  WhisperProcessor: () => WhisperProcessor
});

// src/feature_extraction_utils.js
var FeatureExtractor = class extends Callable {
  /**
   * Constructs a new FeatureExtractor instance.
   *
   * @param {Object} config The configuration for the feature extractor.
   */
  constructor(config) {
    super();
    this.config = config;
  }
  /**
   * Instantiate one of the feature extractor classes of the library from a pretrained model.
   *
   * The feature extractor class to instantiate is selected based on the `feature_extractor_type` property of
   * the config object (either passed as an argument or loaded from `pretrained_model_name_or_path` if possible)
   *
   * @param {string} pretrained_model_name_or_path The name or path of the pretrained model. Can be either:
   * - A string, the *model id* of a pretrained feature_extractor hosted inside a model repo on huggingface.co.
   *   Valid model ids can be located at the root-level, like `bert-base-uncased`, or namespaced under a
   *   user or organization name, like `dbmdz/bert-base-german-cased`.
   * - A path to a *directory* containing feature_extractor files, e.g., `./my_model_directory/`.
   * @param {import('./utils/hub.js').PretrainedOptions} options Additional options for loading the feature_extractor.
   *
   * @returns {Promise<FeatureExtractor>} A new instance of the Feature Extractor class.
   */
  static async from_pretrained(pretrained_model_name_or_path, options = {}) {
    const config = await getModelJSON(pretrained_model_name_or_path, FEATURE_EXTRACTOR_NAME, true, options);
    return new this(config);
  }
};
function validate_audio_inputs(audio, feature_extractor) {
  if (!(audio instanceof Float32Array || audio instanceof Float64Array)) {
    throw new Error(
      `${feature_extractor} expects input to be a Float32Array or a Float64Array, but got ${audio?.constructor?.name ?? typeof audio} instead. If using the feature extractor directly, remember to use \`read_audio(url, sampling_rate)\` to obtain the raw audio data of the file/url.`
    );
  }
}

// src/models/feature_extractors.js
var feature_extractors_exports = {};
__export(feature_extractors_exports, {
  ASTFeatureExtractor: () => ASTFeatureExtractor,
  ChatterboxFeatureExtractor: () => ChatterboxFeatureExtractor,
  ClapFeatureExtractor: () => ClapFeatureExtractor,
  CohereAsrFeatureExtractor: () => CohereAsrFeatureExtractor,
  DacFeatureExtractor: () => DacFeatureExtractor,
  EncodecFeatureExtractor: () => EncodecFeatureExtractor,
  FeatureExtractor: () => FeatureExtractor,
  Gemma3nAudioFeatureExtractor: () => Gemma3nAudioFeatureExtractor,
  Gemma4AudioFeatureExtractor: () => Gemma4AudioFeatureExtractor,
  GraniteSpeechFeatureExtractor: () => GraniteSpeechFeatureExtractor,
  MoonshineFeatureExtractor: () => MoonshineFeatureExtractor,
  ParakeetFeatureExtractor: () => ParakeetFeatureExtractor,
  PyAnnoteFeatureExtractor: () => PyAnnoteFeatureExtractor,
  SeamlessM4TFeatureExtractor: () => SeamlessM4TFeatureExtractor,
  SnacFeatureExtractor: () => SnacFeatureExtractor,
  SpeechT5FeatureExtractor: () => SpeechT5FeatureExtractor,
  VoxtralRealtimeFeatureExtractor: () => VoxtralRealtimeFeatureExtractor,
  Wav2Vec2FeatureExtractor: () => Wav2Vec2FeatureExtractor,
  WeSpeakerFeatureExtractor: () => WeSpeakerFeatureExtractor,
  WhisperFeatureExtractor: () => WhisperFeatureExtractor
});

// src/utils/io.js
import fs5 from "fs";
import { Readable } from "stream";
import { pipeline as pipe } from "stream/promises";
async function saveBlob(path3, blob) {
  if (apis.IS_BROWSER_ENV) {
    if (apis.IS_WEBWORKER_ENV) {
      throw new Error("Unable to save a file from a Web Worker.");
    }
    const dataURL = URL.createObjectURL(blob);
    const downloadLink = document.createElement("a");
    downloadLink.href = dataURL;
    downloadLink.download = path3;
    downloadLink.click();
    downloadLink.remove();
    URL.revokeObjectURL(dataURL);
  } else if (apis.IS_FS_AVAILABLE) {
    const webStream = blob.stream();
    const nodeStream = Readable.fromWeb(webStream);
    const fileStream = fs5.createWriteStream(path3);
    await pipe(nodeStream, fileStream);
  } else {
    throw new Error("Unable to save because filesystem is disabled in this environment.");
  }
}

// src/utils/audio.js
async function load_audio(url2, sampling_rate) {
  if (typeof AudioContext === "undefined") {
    throw Error(
      "Unable to load audio from path/URL since `AudioContext` is not available in your environment. Instead, audio data should be passed directly to the pipeline/processor. For more information and some example code, see https://huggingface.co/docs/transformers.js/guides/node-audio-processing."
    );
  }
  const response = await (await getFile(url2)).arrayBuffer();
  const audioCTX = new AudioContext({ sampleRate: sampling_rate });
  if (typeof sampling_rate === "undefined") {
    logger.warn(`No sampling rate provided, using default of ${audioCTX.sampleRate}Hz.`);
  }
  const decoded = await audioCTX.decodeAudioData(response);
  let audio;
  if (decoded.numberOfChannels === 2) {
    const SCALING_FACTOR = Math.sqrt(2);
    const left = decoded.getChannelData(0);
    const right = decoded.getChannelData(1);
    audio = new Float32Array(left.length);
    for (let i = 0; i < decoded.length; ++i) {
      audio[i] = SCALING_FACTOR * (left[i] + right[i]) / 2;
    }
  } else {
    audio = decoded.getChannelData(0);
  }
  return audio;
}
var read_audio = load_audio;
function generalized_cosine_window(M, a_0) {
  if (M < 1) {
    return new Float64Array();
  }
  if (M === 1) {
    return new Float64Array([1]);
  }
  const a_1 = 1 - a_0;
  const factor = 2 * Math.PI / (M - 1);
  const cos_vals = new Float64Array(M);
  for (let i = 0; i < M; ++i) {
    cos_vals[i] = a_0 - a_1 * Math.cos(i * factor);
  }
  return cos_vals;
}
function hanning(M) {
  return generalized_cosine_window(M, 0.5);
}
function hamming(M) {
  return generalized_cosine_window(M, 0.54);
}
var HERTZ_TO_MEL_MAPPING = {
  htk: (freq) => 2595 * Math.log10(1 + freq / 700),
  kaldi: (freq) => 1127 * Math.log(1 + freq / 700),
  slaney: (freq, min_log_hertz = 1e3, min_log_mel = 15, logstep = 27 / Math.log(6.4)) => freq >= min_log_hertz ? min_log_mel + Math.log(freq / min_log_hertz) * logstep : 3 * freq / 200
};
function hertz_to_mel(freq, mel_scale = "htk") {
  const fn2 = HERTZ_TO_MEL_MAPPING[mel_scale];
  if (!fn2) {
    throw new Error('mel_scale should be one of "htk", "slaney" or "kaldi".');
  }
  return typeof freq === "number" ? fn2(freq) : freq.map((x) => fn2(x));
}
var MEL_TO_HERTZ_MAPPING = {
  htk: (mels) => 700 * (10 ** (mels / 2595) - 1),
  kaldi: (mels) => 700 * (Math.exp(mels / 1127) - 1),
  slaney: (mels, min_log_hertz = 1e3, min_log_mel = 15, logstep = Math.log(6.4) / 27) => mels >= min_log_mel ? min_log_hertz * Math.exp(logstep * (mels - min_log_mel)) : 200 * mels / 3
};
function mel_to_hertz(mels, mel_scale = "htk") {
  const fn2 = MEL_TO_HERTZ_MAPPING[mel_scale];
  if (!fn2) {
    throw new Error('mel_scale should be one of "htk", "slaney" or "kaldi".');
  }
  return typeof mels === "number" ? fn2(mels) : mels.map((x) => fn2(x));
}
function _create_triangular_filter_bank(fft_freqs, filter_freqs) {
  const filter_diff = Float64Array.from(
    { length: filter_freqs.length - 1 },
    (_, i) => filter_freqs[i + 1] - filter_freqs[i]
  );
  const slopes = Array.from(
    {
      length: fft_freqs.length
    },
    () => new Array(filter_freqs.length)
  );
  for (let j = 0; j < fft_freqs.length; ++j) {
    const slope = slopes[j];
    for (let i = 0; i < filter_freqs.length; ++i) {
      slope[i] = filter_freqs[i] - fft_freqs[j];
    }
  }
  const numFreqs = filter_freqs.length - 2;
  const ret = Array.from({ length: numFreqs }, () => new Array(fft_freqs.length));
  for (let j = 0; j < fft_freqs.length; ++j) {
    const slope = slopes[j];
    for (let i = 0; i < numFreqs; ++i) {
      const down = -slope[i] / filter_diff[i];
      const up = slope[i + 2] / filter_diff[i + 1];
      ret[i][j] = Math.max(0, Math.min(down, up));
    }
  }
  return ret;
}
function linspace(start, end, num) {
  const step = (end - start) / (num - 1);
  return Float64Array.from({ length: num }, (_, i) => start + step * i);
}
function mel_filter_bank(num_frequency_bins, num_mel_filters, min_frequency, max_frequency, sampling_rate, norm = null, mel_scale = "htk", triangularize_in_mel_space = false) {
  if (norm !== null && norm !== "slaney") {
    throw new Error('norm must be one of null or "slaney"');
  }
  if (num_frequency_bins < 2) {
    throw new Error(`Require num_frequency_bins: ${num_frequency_bins} >= 2`);
  }
  if (min_frequency > max_frequency) {
    throw new Error(`Require min_frequency: ${min_frequency} <= max_frequency: ${max_frequency}`);
  }
  const mel_min = hertz_to_mel(min_frequency, mel_scale);
  const mel_max = hertz_to_mel(max_frequency, mel_scale);
  const mel_freqs = linspace(mel_min, mel_max, num_mel_filters + 2);
  let filter_freqs = mel_to_hertz(mel_freqs, mel_scale);
  let fft_freqs;
  if (triangularize_in_mel_space) {
    const fft_bin_width = sampling_rate / ((num_frequency_bins - 1) * 2);
    fft_freqs = hertz_to_mel(
      Float64Array.from({ length: num_frequency_bins }, (_, i) => i * fft_bin_width),
      mel_scale
    );
    filter_freqs = mel_freqs;
  } else {
    fft_freqs = linspace(0, Math.floor(sampling_rate / 2), num_frequency_bins);
  }
  const mel_filters = _create_triangular_filter_bank(fft_freqs, filter_freqs);
  if (norm !== null && norm === "slaney") {
    for (let i = 0; i < num_mel_filters; ++i) {
      const filter = mel_filters[i];
      const enorm = 2 / (filter_freqs[i + 2] - filter_freqs[i]);
      for (let j = 0; j < num_frequency_bins; ++j) {
        filter[j] *= enorm;
      }
    }
  }
  return mel_filters;
}
function padReflect(array, left, right) {
  const padded = new array.constructor(array.length + left + right);
  const w = array.length - 1;
  for (let i = 0; i < array.length; ++i) {
    padded[left + i] = array[i];
  }
  for (let i = 1; i <= left; ++i) {
    padded[left - i] = array[calculateReflectOffset(i, w)];
  }
  for (let i = 1; i <= right; ++i) {
    padded[w + left + i] = array[calculateReflectOffset(w - i, w)];
  }
  return padded;
}
function _db_conversion_helper(spectrogram2, factor, reference, min_value, db_range) {
  if (reference <= 0) {
    throw new Error("reference must be greater than zero");
  }
  if (min_value <= 0) {
    throw new Error("min_value must be greater than zero");
  }
  reference = Math.max(min_value, reference);
  const logReference = Math.log10(reference);
  for (let i = 0; i < spectrogram2.length; ++i) {
    spectrogram2[i] = factor * Math.log10(Math.max(min_value, spectrogram2[i]) - logReference);
  }
  if (db_range !== null) {
    if (db_range <= 0) {
      throw new Error("db_range must be greater than zero");
    }
    const maxValue = max(spectrogram2)[0] - db_range;
    for (let i = 0; i < spectrogram2.length; ++i) {
      spectrogram2[i] = Math.max(spectrogram2[i], maxValue);
    }
  }
  return spectrogram2;
}
function amplitude_to_db(spectrogram2, reference = 1, min_value = 1e-5, db_range = null) {
  return _db_conversion_helper(spectrogram2, 20, reference, min_value, db_range);
}
function power_to_db(spectrogram2, reference = 1, min_value = 1e-10, db_range = null) {
  return _db_conversion_helper(spectrogram2, 10, reference, min_value, db_range);
}
async function spectrogram(waveform, window2, frame_length, hop_length, {
  fft_length = null,
  power = 1,
  center = true,
  pad_mode = "reflect",
  onesided = true,
  preemphasis = null,
  preemphasis_htk_flavor = true,
  mel_filters = null,
  mel_floor = 1e-10,
  log_mel = null,
  max_log_mel = null,
  reference = 1,
  min_value = 1e-10,
  db_range = null,
  remove_dc_offset = null,
  // Custom parameters for efficiency reasons
  min_num_frames = null,
  max_num_frames = null,
  do_pad = true,
  transpose = false,
  mel_offset = 0,
  mel_floor_mode = "clamp"
} = {}) {
  const window_length = window2.length;
  if (fft_length === null) {
    fft_length = frame_length;
  }
  if (frame_length > fft_length) {
    throw Error(`frame_length (${frame_length}) may not be larger than fft_length (${fft_length})`);
  }
  if (window_length !== frame_length) {
    throw new Error(`Length of the window (${window_length}) must equal frame_length (${frame_length})`);
  }
  if (hop_length <= 0) {
    throw new Error("hop_length must be greater than zero");
  }
  if (power === null && mel_filters !== null) {
    throw new Error(
      "You have provided `mel_filters` but `power` is `None`. Mel spectrogram computation is not yet supported for complex-valued spectrogram. Specify `power` to fix this issue."
    );
  }
  if (!preemphasis_htk_flavor) {
    throw new Error("`preemphasis_htk_flavor=false` is not currently supported.");
  }
  if (center) {
    const padding = Math.floor(frame_length / 2);
    switch (pad_mode) {
      case "reflect": {
        waveform = padReflect(waveform, padding, padding);
        break;
      }
      case "constant": {
        const padded = new waveform.constructor(waveform.length + 2 * padding);
        padded.set(waveform, padding);
        waveform = padded;
        break;
      }
      case "semicausal": {
        const padded = new waveform.constructor(waveform.length + padding);
        padded.set(waveform, padding);
        waveform = padded;
        break;
      }
      default:
        throw new Error(`pad_mode="${pad_mode}" not implemented yet.`);
    }
  }
  let num_frames = Math.floor(1 + Math.floor((waveform.length - frame_length) / hop_length));
  if (min_num_frames !== null && num_frames < min_num_frames) {
    num_frames = min_num_frames;
  }
  const num_frequency_bins = onesided ? Math.floor(fft_length / 2) + 1 : fft_length;
  let d1 = num_frames;
  let d1Max = num_frames;
  if (max_num_frames !== null) {
    if (max_num_frames > num_frames) {
      if (do_pad) {
        d1Max = max_num_frames;
      }
    } else {
      d1Max = d1 = max_num_frames;
    }
  }
  const fft = new FFT(fft_length);
  const inputBuffer = new Float64Array(fft_length);
  const outputBuffer = new Float64Array(fft.outputBufferSize);
  const transposedMagnitudeData = new Float32Array(num_frequency_bins * d1Max);
  for (let i = 0; i < d1; ++i) {
    const offset = i * hop_length;
    const buffer_size = Math.min(waveform.length - offset, frame_length);
    if (buffer_size !== frame_length) {
      inputBuffer.fill(0, 0, frame_length);
    }
    for (let j = 0; j < buffer_size; ++j) {
      inputBuffer[j] = waveform[offset + j];
    }
    if (remove_dc_offset) {
      let sum = 0;
      for (let j = 0; j < buffer_size; ++j) {
        sum += inputBuffer[j];
      }
      const mean2 = sum / buffer_size;
      for (let j = 0; j < buffer_size; ++j) {
        inputBuffer[j] -= mean2;
      }
    }
    if (preemphasis !== null) {
      for (let j = buffer_size - 1; j >= 1; --j) {
        inputBuffer[j] -= preemphasis * inputBuffer[j - 1];
      }
      inputBuffer[0] *= 1 - preemphasis;
    }
    for (let j = 0; j < window2.length; ++j) {
      inputBuffer[j] *= window2[j];
    }
    fft.realTransform(outputBuffer, inputBuffer);
    for (let j = 0; j < num_frequency_bins; ++j) {
      const j2 = j << 1;
      transposedMagnitudeData[j * d1Max + i] = outputBuffer[j2] ** 2 + outputBuffer[j2 + 1] ** 2;
    }
  }
  if (power !== null && power !== 2) {
    const pow = power / 2;
    for (let i = 0; i < transposedMagnitudeData.length; ++i) {
      transposedMagnitudeData[i] **= pow;
    }
  }
  const num_mel_filters = mel_filters.length;
  let mel_spec = await matmul(
    // TODO: Make `mel_filters` a Tensor during initialization
    new Tensor2("float32", mel_filters.flat(), [num_mel_filters, num_frequency_bins]),
    new Tensor2("float32", transposedMagnitudeData, [num_frequency_bins, d1Max])
  );
  if (transpose) {
    mel_spec = mel_spec.transpose(1, 0);
  }
  const mel_spec_data = (
    /** @type {Float32Array} */
    mel_spec.data
  );
  if (mel_floor_mode === "add") {
    for (let i = 0; i < mel_spec_data.length; ++i) {
      mel_spec_data[i] = mel_offset + mel_spec_data[i] + mel_floor;
    }
  } else {
    for (let i = 0; i < mel_spec_data.length; ++i) {
      mel_spec_data[i] = mel_offset + Math.max(mel_floor, mel_spec_data[i]);
    }
  }
  if (power !== null && log_mel !== null) {
    const o = Math.min(mel_spec_data.length, d1 * num_mel_filters);
    switch (log_mel) {
      case "log":
        for (let i = 0; i < o; ++i) {
          mel_spec_data[i] = Math.log(mel_spec_data[i]);
        }
        break;
      case "log10":
        for (let i = 0; i < o; ++i) {
          mel_spec_data[i] = Math.log10(mel_spec_data[i]);
        }
        break;
      case "log10_max_norm": {
        for (let i = 0; i < o; ++i) {
          mel_spec_data[i] = Math.log10(mel_spec_data[i]);
        }
        const logMax = max_log_mel ?? max(mel_spec_data)[0];
        const threshold = logMax - 8;
        for (let i = 0; i < o; ++i) {
          mel_spec_data[i] = (Math.max(mel_spec_data[i], threshold) + 4) / 4;
        }
        break;
      }
      case "dB":
        if (power === 1) {
          amplitude_to_db(mel_spec_data, reference, min_value, db_range);
        } else if (power === 2) {
          power_to_db(mel_spec_data, reference, min_value, db_range);
        } else {
          throw new Error(`Cannot use log_mel option '${log_mel}' with power ${power}`);
        }
        break;
      default:
        throw new Error(
          `log_mel must be one of null, 'log', 'log10', 'log10_max_norm', or 'dB'. Got '${log_mel}'`
        );
    }
  }
  return mel_spec;
}
function window_function(window_length, name, { periodic = true, frame_length = null, center = true } = {}) {
  const length = periodic ? window_length + 1 : window_length;
  let window2;
  switch (name) {
    case "boxcar":
      window2 = new Float64Array(length).fill(1);
      break;
    case "hann":
    case "hann_window":
      window2 = hanning(length);
      break;
    case "hamming":
      window2 = hamming(length);
      break;
    case "povey":
      window2 = hanning(length).map((x) => Math.pow(x, 0.85));
      break;
    default:
      throw new Error(`Unknown window type ${name}.`);
  }
  if (periodic) {
    window2 = window2.subarray(0, window_length);
  }
  if (frame_length === null || window_length === frame_length) {
    return window2;
  }
  if (window_length > frame_length) {
    throw new Error(
      `Length of the window (${window_length}) may not be larger than frame_length (${frame_length})`
    );
  }
  const padded = new Float64Array(frame_length);
  const offset = center ? Math.floor((frame_length - window_length) / 2) : 0;
  padded.set(window2, offset);
  return padded;
}
function encodeWAV(chunks, rate) {
  const totalLength = chunks.reduce((acc, chunk2) => acc + chunk2.length, 0);
  const buffer = new ArrayBuffer(44);
  const view = new DataView(buffer);
  writeString(view, 0, "RIFF");
  view.setUint32(4, 36 + totalLength * 4, true);
  writeString(view, 8, "WAVE");
  writeString(view, 12, "fmt ");
  view.setUint32(16, 16, true);
  view.setUint16(20, 3, true);
  view.setUint16(22, 1, true);
  view.setUint32(24, rate, true);
  view.setUint32(28, rate * 4, true);
  view.setUint16(32, 4, true);
  view.setUint16(34, 32, true);
  writeString(view, 36, "data");
  view.setUint32(40, totalLength * 4, true);
  return new Blob([buffer, ...chunks.map((chunk2) => (
    /** @type {ArrayBuffer} */
    chunk2.buffer
  ))], {
    type: "audio/wav"
  });
}
function writeString(view, offset, string) {
  for (let i = 0; i < string.length; ++i) {
    view.setUint8(offset + i, string.charCodeAt(i));
  }
}
var RawAudio = class {
  /**
   * Create a new `RawAudio` object.
   * @param {Float32Array|Float32Array[]} audio Audio data, either as a single `Float32Array` chunk or multiple `Float32Array` chunks.
   * @param {number} sampling_rate Sampling rate of the audio data
   */
  constructor(audio, sampling_rate) {
    this.audio = audio;
    this.sampling_rate = sampling_rate;
  }
  /**
   * Get the audio data, accumulating all chunks if necessary.
   * @returns {Float32Array} The audio data.
   */
  get data() {
    if (Array.isArray(this.audio)) {
      if (this.audio.length === 0) {
        return new Float32Array(0);
      }
      if (this.audio.length === 1) {
        return this.audio[0];
      }
      const totalLength = this.audio.reduce((acc, chunk2) => acc + chunk2.length, 0);
      const result = new Float32Array(totalLength);
      let offset = 0;
      for (const chunk2 of this.audio) {
        result.set(chunk2, offset);
        offset += chunk2.length;
      }
      return result;
    } else {
      return this.audio;
    }
  }
  /**
   * Convert the audio to a blob.
   * @returns {Blob}
   */
  toBlob() {
    let audio = this.audio;
    if (audio instanceof Float32Array) {
      audio = [audio];
    }
    return encodeWAV(audio, this.sampling_rate);
  }
  /**
   * Save the audio to a wav file.
   * @param {string} path
   * @returns {Promise<void>}
   */
  async save(path3) {
    return saveBlob(path3, this.toBlob());
  }
};

// src/models/audio_spectrogram_transformer/feature_extraction_audio_spectrogram_transformer.js
var ASTFeatureExtractor = class extends FeatureExtractor {
  constructor(config) {
    super(config);
    const sampling_rate = this.config.sampling_rate;
    const mel_filters = mel_filter_bank(
      257,
      // num_frequency_bins
      this.config.num_mel_bins,
      // num_mel_filters
      20,
      // min_frequency
      Math.floor(sampling_rate / 2),
      // max_frequency
      sampling_rate,
      // sampling_rate
      null,
      // norm
      "kaldi",
      // mel_scale
      true
      // triangularize_in_mel_space
    );
    this.mel_filters = mel_filters;
    this.window = window_function(400, "hann", {
      periodic: false
    });
    this.mean = this.config.mean;
    this.std = this.config.std;
  }
  /**
   * Computes the log-Mel spectrogram of the provided audio waveform.
   * @param {Float32Array|Float64Array} waveform The audio waveform to process.
   * @param {number} max_length The maximum number of frames to return.
   * @returns {Promise<Tensor>} An object containing the log-Mel spectrogram data as a Float32Array and its dimensions as an array of numbers.
   */
  async _extract_fbank_features(waveform, max_length) {
    return spectrogram(
      waveform,
      this.window,
      // window
      400,
      // frame_length
      160,
      // hop_length
      {
        fft_length: 512,
        power: 2,
        center: false,
        preemphasis: 0.97,
        mel_filters: this.mel_filters,
        log_mel: "log",
        mel_floor: 1192092955078125e-22,
        remove_dc_offset: true,
        // Custom
        max_num_frames: max_length,
        transpose: true
      }
    );
  }
  /**
   * Asynchronously extracts features from a given audio using the provided configuration.
   * @param {Float32Array|Float64Array} audio The audio data as a Float32Array/Float64Array.
   * @returns {Promise<{ input_values: Tensor }>} A Promise resolving to an object containing the extracted input features as a Tensor.
   */
  async _call(audio) {
    validate_audio_inputs(audio, "ASTFeatureExtractor");
    const features = await this._extract_fbank_features(audio, this.config.max_length);
    if (this.config.do_normalize) {
      const denom = this.std * 2;
      const features_data = features.data;
      for (let i = 0; i < features_data.length; ++i) {
        features_data[i] = (features_data[i] - this.mean) / denom;
      }
    }
    return {
      input_values: features.unsqueeze_(0)
    };
  }
};

// src/models/encodec/feature_extraction_encodec.js
var EncodecFeatureExtractor = class extends FeatureExtractor {
  /**
   * Asynchronously extracts input values from a given audio using the provided configuration.
   * @param {Float32Array|Float64Array} audio The audio data as a Float32Array/Float64Array.
   * @returns {Promise<{ input_values: Tensor; }>} The extracted input values.
   */
  async _call(audio) {
    validate_audio_inputs(audio, "EncodecFeatureExtractor");
    if (audio instanceof Float64Array) {
      audio = new Float32Array(audio);
    }
    const num_channels = this.config.feature_size;
    if (audio.length % num_channels !== 0) {
      throw new Error(
        `The length of the audio data must be a multiple of the number of channels (${num_channels}).`
      );
    }
    const shape = [
      1,
      num_channels,
      audio.length / num_channels
    ];
    return {
      input_values: new Tensor2("float32", audio, shape)
    };
  }
};

// src/models/chatterbox/feature_extraction_chatterbox.js
var ChatterboxFeatureExtractor = class extends FeatureExtractor {
  /**
   * Asynchronously extracts input values from a given audio using the provided configuration.
   * @param {Float32Array|Float64Array} audio The audio data as a Float32Array/Float64Array.
   * @returns {Promise<{ input_values: Tensor; }>} The extracted input values.
   */
  async _call(audio) {
    validate_audio_inputs(audio, "ChatterboxFeatureExtractor");
    if (audio instanceof Float64Array) {
      audio = new Float32Array(audio);
    }
    const shape = [
      1,
      audio.length
      /* num_samples */
    ];
    return {
      input_values: new Tensor2("float32", audio, shape)
    };
  }
};

// src/models/clap/feature_extraction_clap.js
var ClapFeatureExtractor = class extends FeatureExtractor {
  constructor(config) {
    super(config);
    this.mel_filters = mel_filter_bank(
      this.config.nb_frequency_bins,
      // num_frequency_bins
      this.config.feature_size,
      // num_mel_filters
      this.config.frequency_min,
      // min_frequency
      this.config.frequency_max,
      // max_frequency
      this.config.sampling_rate,
      // sampling_rate
      null,
      // norm
      "htk"
      // mel_scale
    );
    this.mel_filters_slaney = mel_filter_bank(
      this.config.nb_frequency_bins,
      // num_frequency_bins
      this.config.feature_size,
      // num_mel_filters
      this.config.frequency_min,
      // min_frequency
      this.config.frequency_max,
      // max_frequency
      this.config.sampling_rate,
      // sampling_rate
      "slaney",
      // norm
      "slaney"
      // mel_scale
    );
    this.window = window_function(this.config.fft_window_size, "hann");
  }
  /**
   * Extracts the mel spectrogram and prepares it for the mode based on the `truncation` and `padding` arguments.
   *
   * Four different path are possible:
   *   - `truncation="fusion"` and the length of the waveform is greater than the max length: the mel spectrogram
   *     will be computed on the entire audio. 3 random crops and a dowsampled version of the full mel spectrogram
   *     are then stacked together. They will later be used for `feature_fusion`.
   *   - `truncation="rand_trunc"` and the length of the waveform is smaller than the max length: the audio is
   *     padded based on `padding`.
   *   - `truncation="fusion"` and the length of the waveform is smaller than the max length: the audio is padded
   *     based on `padding`, and is repeated `4` times.
   *   - `truncation="rand_trunc"` and the length of the waveform is greater than the max length: the mel
   *     spectrogram will be computed on a random crop of the waveform.
   *
   * @param {Float32Array|Float64Array} waveform The input waveform.
   * @param {number} max_length The maximum length of the waveform.
   * @param {string} truncation The truncation strategy to use.
   * @param {string} padding The padding strategy to use.
   * @returns {Promise<Tensor>} An object containing the mel spectrogram data as a Float32Array, its dimensions as an array of numbers, and a boolean indicating whether the waveform was longer than the max length.
   * @private
   */
  async _get_input_mel(waveform, max_length, truncation, padding) {
    let input_mel;
    let longer = false;
    const diff = waveform.length - max_length;
    if (diff > 0) {
      if (truncation === "rand_trunc") {
        longer = true;
        const idx = Math.floor(random.random() * (diff + 1));
        waveform = waveform.subarray(idx, idx + max_length);
        input_mel = await this._extract_fbank_features(
          waveform,
          this.mel_filters_slaney,
          this.config.nb_max_samples
        );
      } else {
        throw new Error(`Truncation strategy "${truncation}" not implemented`);
      }
    } else {
      if (diff < 0) {
        let padded = new Float64Array(max_length);
        padded.set(waveform);
        if (padding === "repeat") {
          for (let i = waveform.length; i < max_length; i += waveform.length) {
            padded.set(waveform.subarray(0, Math.min(waveform.length, max_length - i)), i);
          }
        } else if (padding === "repeatpad") {
          for (let i = waveform.length; i < -diff; i += waveform.length) {
            padded.set(waveform, i);
          }
        }
        waveform = padded;
      }
      if (truncation === "fusion") {
        throw new Error(`Truncation strategy "${truncation}" not implemented`);
      }
      input_mel = await this._extract_fbank_features(
        waveform,
        this.mel_filters_slaney,
        this.config.nb_max_samples
      );
    }
    return input_mel.unsqueeze_(0);
  }
  /**
   * Compute the log-mel spectrogram of the provided `waveform` using the Hann window.
   * In CLAP, two different filter banks are used depending on the truncation pattern:
   *  - `self.mel_filters`: they correspond to the default parameters of `torchaudio` which can be obtained from
   *    calling `torchaudio.transforms.MelSpectrogram().mel_scale.fb`. These filters are used when `truncation`
   *    is set to `"fusion"`.
   *  - `self.mel_filteres_slaney` : they correspond to the default parameters of `librosa` which used
   *    `librosa.filters.mel` when computing the mel spectrogram. These filters were only used in the original
   *    implementation when the truncation mode is not `"fusion"`.
   *
   * @param {Float32Array|Float64Array} waveform The audio waveform to process.
   * @param {number[][]} mel_filters The mel filters to use.
   * @param {number} [max_length=null] The maximum number of frames to return.
   * @returns {Promise<Tensor>} An object containing the log-Mel spectrogram data as a Float32Array and its dimensions as an array of numbers.
   */
  async _extract_fbank_features(waveform, mel_filters, max_length = null) {
    return spectrogram(
      waveform,
      this.window,
      // window
      this.config.fft_window_size,
      // frame_length
      this.config.hop_length,
      // hop_length
      {
        power: 2,
        mel_filters,
        log_mel: "dB",
        // Custom
        max_num_frames: max_length,
        do_pad: false,
        transpose: true
      }
    );
  }
  /**
   * Asynchronously extracts features from a given audio using the provided configuration.
   * @param {Float32Array|Float64Array} audio The audio data as a Float32Array/Float64Array.
   * @returns {Promise<{ input_features: Tensor }>} A Promise resolving to an object containing the extracted input features as a Tensor.
   */
  async _call(audio, { max_length = null } = {}) {
    validate_audio_inputs(audio, "ClapFeatureExtractor");
    const padded_inputs = await this._get_input_mel(
      audio,
      max_length ?? this.config.nb_max_samples,
      this.config.truncation,
      this.config.padding
    );
    return {
      input_features: padded_inputs.unsqueeze_(0)
    };
  }
};

// src/models/parakeet/feature_extraction_parakeet.js
var EPSILON = 1e-5;
var ParakeetFeatureExtractor = class extends FeatureExtractor {
  constructor(config) {
    super(config);
    this.config.mel_filters ??= mel_filter_bank(
      Math.floor(1 + this.config.n_fft / 2),
      // num_frequency_bins
      this.config.feature_size,
      // num_mel_filters
      0,
      // min_frequency
      this.config.sampling_rate / 2,
      // max_frequency
      this.config.sampling_rate,
      // sampling_rate
      "slaney",
      // norm
      "slaney"
      // mel_scale
    );
    const window2 = window_function(this.config.win_length, "hann", {
      periodic: false
    });
    this.window = new Float64Array(this.config.n_fft);
    const offset = Math.floor((this.config.n_fft - this.config.win_length) / 2);
    this.window.set(window2, offset);
  }
  /**
   * Computes the log-Mel spectrogram of the provided audio waveform.
   * @param {Float32Array|Float64Array} waveform The audio waveform to process.
   * @returns {Promise<Tensor>} An object containing the log-Mel spectrogram data as a Float32Array and its dimensions as an array of numbers.
   */
  async _extract_fbank_features(waveform) {
    const preemphasis = this.config.preemphasis;
    waveform = new Float64Array(waveform);
    for (let j = waveform.length - 1; j >= 1; --j) {
      waveform[j] -= preemphasis * waveform[j - 1];
    }
    const features = await spectrogram(
      waveform,
      this.window,
      // window
      this.window.length,
      // frame_length
      this.config.hop_length,
      // hop_length
      {
        fft_length: this.config.n_fft,
        power: 2,
        mel_filters: this.config.mel_filters,
        log_mel: "log",
        mel_floor: -Infinity,
        pad_mode: "constant",
        center: true,
        // Custom
        transpose: true,
        mel_offset: 2 ** -24
      }
    );
    return features;
  }
  /**
   * Asynchronously extracts features from a given audio using the provided configuration.
   * @param {Float32Array|Float64Array} audio The audio data as a Float32Array/Float64Array.
   * @returns {Promise<{ input_features: Tensor; attention_mask: Tensor; }>} A Promise resolving to an object containing the extracted input features as a Tensor.
   */
  async _call(audio) {
    validate_audio_inputs(audio, "ParakeetFeatureExtractor");
    const features = await this._extract_fbank_features(audio);
    const features_length = Math.floor(
      (audio.length + Math.floor(this.config.n_fft / 2) * 2 - this.config.n_fft) / this.config.hop_length
    );
    const features_data = (
      /** @type {Float32Array} */
      features.data
    );
    features_data.fill(0, features_length * features.dims[1]);
    const [num_frames, num_features] = features.dims;
    const sum = new Float64Array(num_features);
    const sum_sq = new Float64Array(num_features);
    for (let i = 0; i < features_length; ++i) {
      const offset = i * num_features;
      for (let j = 0; j < num_features; ++j) {
        const val = features_data[offset + j];
        sum[j] += val;
        sum_sq[j] += val * val;
      }
    }
    const divisor = features_length > 1 ? features_length - 1 : 1;
    for (let j = 0; j < num_features; ++j) {
      const mean2 = sum[j] / features_length;
      const variance = (sum_sq[j] - features_length * mean2 * mean2) / divisor;
      const std = Math.sqrt(variance) + EPSILON;
      const inv_std = 1 / std;
      for (let i = 0; i < features_length; ++i) {
        const index = i * num_features + j;
        features_data[index] = (features_data[index] - mean2) * inv_std;
      }
    }
    const mask_data = new BigInt64Array(num_frames);
    mask_data.fill(1n, 0, features_length);
    return {
      input_features: features.unsqueeze_(0),
      attention_mask: new Tensor2("int64", mask_data, [1, num_frames])
    };
  }
};

// src/models/cohere_asr/feature_extraction_cohere_asr.js
var CohereAsrFeatureExtractor = class extends ParakeetFeatureExtractor {
  /**
   * Apply deterministic dithering seeded by the waveform length.
   * @param {Float64Array} waveform
   * @returns {Float64Array} The dithered waveform (mutated in-place).
   */
  _apply_dither(waveform) {
    const dither = this.config.dither ?? 0;
    if (dither <= 0) return waveform;
    const rng2 = new Random(waveform.length);
    for (let i = 0; i < waveform.length; ++i) {
      waveform[i] += dither * rng2.gauss();
    }
    return waveform;
  }
  /**
   * Split audio into chunks at energy-based boundaries for long audio.
   * @param {Float32Array|Float64Array} audio The raw audio waveform.
   * @returns {(Float32Array|Float64Array)[]} Array of audio chunks.
   */
  split_audio(audio) {
    const max_audio_clip_s = this.config.max_audio_clip_s ?? 35;
    const overlap_chunk_second = this.config.overlap_chunk_second ?? 5;
    const min_energy_window_samples = this.config.min_energy_window_samples ?? 1600;
    const sampling_rate = this.config.sampling_rate;
    const chunk_size = Math.max(1, Math.round(max_audio_clip_s * sampling_rate));
    const boundary_context_size = Math.max(1, Math.round(overlap_chunk_second * sampling_rate));
    if (audio.length <= chunk_size) {
      return [audio];
    }
    const chunks = [];
    let idx = 0;
    const total_samples = audio.length;
    while (idx < total_samples) {
      if (idx + chunk_size >= total_samples) {
        chunks.push(audio.slice(idx, total_samples));
        break;
      }
      const search_start = Math.max(idx, idx + chunk_size - boundary_context_size);
      const search_end = Math.min(idx + chunk_size, total_samples);
      let split_point;
      if (search_end <= search_start) {
        split_point = idx + chunk_size;
      } else {
        split_point = this._find_split_point_energy(audio, search_start, search_end, min_energy_window_samples);
      }
      split_point = Math.max(idx + 1, Math.min(split_point, total_samples));
      chunks.push(audio.slice(idx, split_point));
      idx = split_point;
    }
    return chunks;
  }
  /**
   * Find the quietest point (minimum energy) within a segment of audio.
   * @param {Float32Array|Float64Array} waveform
   * @param {number} start_idx
   * @param {number} end_idx
   * @param {number} window_size
   * @returns {number} Index of the quietest point.
   */
  _find_split_point_energy(waveform, start_idx, end_idx, window_size) {
    const segment_len = end_idx - start_idx;
    if (segment_len <= window_size) {
      return Math.floor((start_idx + end_idx) / 2);
    }
    let min_energy = Infinity;
    let quietest_idx = start_idx;
    const upper = segment_len - window_size;
    for (let i = 0; i <= upper; i += window_size) {
      let energy = 0;
      for (let j = 0; j < window_size; ++j) {
        const val = waveform[start_idx + i + j];
        energy += val * val;
      }
      energy = Math.sqrt(energy / window_size);
      if (energy < min_energy) {
        min_energy = energy;
        quietest_idx = start_idx + i;
      }
    }
    return quietest_idx;
  }
  /**
   * Extracts features from a given audio waveform.
   * @param {Float32Array|Float64Array} audio The audio data.
   * @returns {Promise<{ input_features: import('../../utils/tensor.js').Tensor; attention_mask: import('../../utils/tensor.js').Tensor; }>}
   */
  async _call(audio) {
    validate_audio_inputs(audio, "CohereAsrFeatureExtractor");
    const waveform = new Float64Array(audio);
    this._apply_dither(waveform);
    return super._call(waveform);
  }
};

// src/models/dac/feature_extraction_dac.js
var DacFeatureExtractor = class extends EncodecFeatureExtractor {
};

// src/models/gemma3n/feature_extraction_gemma3n.js
var Gemma3nAudioFeatureExtractor = class extends FeatureExtractor {
  constructor(config) {
    super(config);
    const { fft_length, feature_size, min_frequency, max_frequency, sampling_rate, frame_length } = this.config;
    const mel_filters = mel_filter_bank(
      Math.floor(1 + fft_length / 2),
      // num_frequency_bins
      feature_size,
      // num_mel_filters
      min_frequency,
      // min_frequency
      max_frequency,
      // max_frequency
      sampling_rate,
      // sampling_rate
      null,
      // norm
      "htk",
      // mel_scale
      false
      // triangularize_in_mel_space
    );
    this.mel_filters = mel_filters;
    this.window = window_function(frame_length, "hann");
  }
  /**
   * Computes the log-Mel spectrogram of the provided audio waveform.
   * @param {Float32Array|Float64Array} waveform The audio waveform to process.
   * @param {number} max_length The maximum number of frames to return.
   * @returns {Promise<Tensor>} An object containing the log-Mel spectrogram data as a Float32Array and its dimensions as an array of numbers.
   */
  async _extract_fbank_features(waveform, max_length) {
    return spectrogram(
      waveform,
      this.window,
      // window
      this.config.frame_length,
      // frame_length
      this.config.hop_length,
      // hop_length
      {
        fft_length: this.config.fft_length,
        center: false,
        onesided: true,
        preemphasis: this.config.preemphasis,
        preemphasis_htk_flavor: this.config.preemphasis_htk_flavor,
        mel_filters: this.mel_filters,
        log_mel: "log",
        mel_floor: this.config.mel_floor,
        remove_dc_offset: false,
        // Custom
        transpose: true
      }
    );
  }
  /**
   * Asynchronously extracts features from a given audio using the provided configuration.
   * @param {Float32Array|Float64Array} audio The audio data as a Float32Array/Float64Array.
   * @param {Object} options Optional parameters for feature extraction.
   * @param {number} [options.max_length=480_000] If provided, defines the maximum length of the audio to allow.
   * Audio longer than this will be truncated if `truncation=True`.
   * @param {boolean} [options.truncation=true] Whether or not to truncate audio above `max_length`.
   * @param {boolean} [options.padding=true] Whether to pad the sequence to a multiple of `pad_to_multiple_of`.
   * @param {number} [options.pad_to_multiple_of=128] The number to pad the sequence to a multiple of.
   * @returns {Promise<{ input_features: Tensor, input_features_mask: Tensor }>} A Promise resolving to an object containing the extracted input features and attention masks as Tensors.
   */
  async _call(audio, { max_length = 48e4, truncation = true, padding = true, pad_to_multiple_of = 128 } = {}) {
    validate_audio_inputs(audio, "Gemma3nAudioFeatureExtractor");
    if (truncation && audio.length > max_length) {
      audio = audio.slice(0, max_length);
    }
    if (padding && audio.length % pad_to_multiple_of !== 0) {
      const padding_length = pad_to_multiple_of - audio.length % pad_to_multiple_of;
      const padded_audio = new Float64Array(audio.length + padding_length);
      padded_audio.set(audio);
      if (this.config.padding_value !== 0) {
        padded_audio.fill(this.config.padding_value, audio.length);
      }
      audio = padded_audio;
    }
    const features = await this._extract_fbank_features(audio, this.config.max_length);
    const padded_attention_mask = full([1, features.dims[0]], true);
    return {
      input_features: features.unsqueeze_(0),
      input_features_mask: padded_attention_mask
    };
  }
};

// src/models/gemma4/feature_extraction_gemma4.js
var Gemma4AudioFeatureExtractor = class extends Gemma3nAudioFeatureExtractor {
  /**
   * @override
   * Gemma4 uses semicausal padding, unfold(frame_length+1) framing, and
   * additive mel_floor — all controlled via flags on the shared spectrogram().
   */
  async _extract_fbank_features(waveform, max_length) {
    const { frame_length, hop_length, fft_length } = this.config;
    const pad_left = Math.floor(frame_length / 2);
    const num_frames = Math.floor((waveform.length + pad_left - (frame_length + 1)) / hop_length) + 1;
    return spectrogram(waveform, this.window, frame_length, hop_length, {
      fft_length,
      center: true,
      pad_mode: "semicausal",
      onesided: true,
      preemphasis: this.config.preemphasis,
      preemphasis_htk_flavor: this.config.preemphasis_htk_flavor,
      mel_filters: this.mel_filters,
      log_mel: "log",
      mel_floor: this.config.mel_floor,
      mel_floor_mode: "add",
      remove_dc_offset: false,
      transpose: true,
      max_num_frames: num_frames
    });
  }
  /**
   * @override
   * Wraps the base class result with a frame-aware attention mask
   * and zeros out features for invalid (padded) frames.
   */
  async _call(audio, options = {}) {
    validate_audio_inputs(audio, "Gemma4AudioFeatureExtractor");
    const original_length = audio.length;
    const result = await super._call(audio, options);
    const { input_features } = result;
    const [, num_frames, num_features] = input_features.dims;
    const { frame_length, hop_length } = this.config;
    const pad_left = Math.floor(frame_length / 2);
    const frame_size_for_unfold = frame_length + 1;
    const sample_mask = new Uint8Array(original_length + pad_left + (options.pad_to_multiple_of ?? 128));
    sample_mask.fill(1, pad_left, pad_left + original_length);
    const frame_mask = new Uint8Array(num_frames);
    for (let i = 0; i < num_frames; ++i) {
      frame_mask[i] = sample_mask[i * hop_length + frame_size_for_unfold - 1] ? 1 : 0;
    }
    const feat_data = (
      /** @type {Float32Array} */
      input_features.data
    );
    for (let i = 0; i < num_frames; ++i) {
      if (!frame_mask[i]) {
        feat_data.fill(0, i * num_features, (i + 1) * num_features);
      }
    }
    result.input_features_mask = new Tensor2("bool", frame_mask, [1, num_frames]);
    return result;
  }
};

// src/models/granite_speech/feature_extraction_granite_speech.js
var GraniteSpeechFeatureExtractor = class extends FeatureExtractor {
  constructor(config) {
    super(config);
    const { n_fft, win_length, n_mels, sample_rate } = config.melspec_kwargs;
    this.mel_filters = mel_filter_bank(
      Math.floor(1 + n_fft / 2),
      // num_frequency_bins = 257
      n_mels,
      // 80
      0,
      // min_frequency
      sample_rate / 2,
      // max_frequency = 8000
      sample_rate,
      // 16000
      null,
      // norm (torchaudio default: no norm)
      "htk"
      // mel_scale (torchaudio default)
    );
    const raw_window = window_function(win_length, "hann");
    this.window = new Float64Array(n_fft);
    const pad = Math.floor((n_fft - win_length) / 2);
    this.window.set(raw_window, pad);
  }
  /**
   * Extract mel spectrogram features from audio, matching the Python GraniteSpeechFeatureExtractor.
   * @param {Float32Array|Float64Array} audio The audio waveform.
   * @returns {Promise<{input_features: Tensor}>}
   */
  async _call(audio) {
    validate_audio_inputs(audio, "GraniteSpeechFeatureExtractor");
    const { n_fft, hop_length, n_mels } = this.config.melspec_kwargs;
    const num_frames = 1 + Math.floor((audio.length - 1) / hop_length);
    const max_num_frames = num_frames - num_frames % 2;
    const mel = await spectrogram(audio, this.window, n_fft, hop_length, {
      power: 2,
      mel_filters: this.mel_filters,
      log_mel: "log10_max_norm",
      transpose: true,
      // [time, n_mels]
      max_num_frames,
      do_pad: false
    });
    const input_features = mel.view(-1, 2 * n_mels).unsqueeze_(0);
    return { input_features };
  }
};

// src/models/moonshine/feature_extraction_moonshine.js
var MoonshineFeatureExtractor = class extends FeatureExtractor {
  /**
   * Asynchronously extracts input values from a given audio using the provided configuration.
   * @param {Float32Array|Float64Array} audio The audio data as a Float32Array/Float64Array.
   * @returns {Promise<{ input_values: Tensor; }>} The extracted input values.
   */
  async _call(audio) {
    validate_audio_inputs(audio, "MoonshineFeatureExtractor");
    if (audio instanceof Float64Array) {
      audio = new Float32Array(audio);
    }
    const shape = [
      1,
      audio.length
      /* num_samples */
    ];
    return {
      input_values: new Tensor2("float32", audio, shape)
    };
  }
};

// src/models/pyannote/feature_extraction_pyannote.js
var PyAnnoteFeatureExtractor = class extends FeatureExtractor {
  /**
   * Asynchronously extracts features from a given audio using the provided configuration.
   * @param {Float32Array|Float64Array} audio The audio data as a Float32Array/Float64Array.
   * @returns {Promise<{ input_values: Tensor; }>} The extracted input features.
   */
  async _call(audio) {
    validate_audio_inputs(audio, "PyAnnoteFeatureExtractor");
    if (audio instanceof Float64Array) {
      audio = new Float32Array(audio);
    }
    const shape = [
      1,
      1,
      audio.length
      /* num_samples */
    ];
    return {
      input_values: new Tensor2("float32", audio, shape)
    };
  }
  /**
   * NOTE: Can return fractional values. `Math.ceil` will ensure correct value.
   * @param {number} samples The number of frames in the audio.
   * @returns {number} The number of frames in the audio.
   */
  samples_to_frames(samples) {
    return (samples - this.config.offset) / this.config.step;
  }
  /**
   * Post-processes the speaker diarization logits output by the model.
   * @param {import('../../utils/tensor.js').Tensor} logits The speaker diarization logits output by the model.
   * @param {number} num_samples Number of samples in the input audio.
   * @returns {Array<Array<{ id: number, start: number, end: number, confidence: number }>>} The post-processed speaker diarization results.
   */
  post_process_speaker_diarization(logits, num_samples) {
    const ratio = num_samples / this.samples_to_frames(num_samples) / this.config.sampling_rate;
    const results = [];
    for (const scores of logits.tolist()) {
      const accumulated_segments = [];
      let current_speaker = -1;
      for (let i = 0; i < scores.length; ++i) {
        const probabilities = softmax(scores[i]);
        const [score, id] = max(probabilities);
        const [start, end] = [i, i + 1];
        if (id !== current_speaker) {
          current_speaker = id;
          accumulated_segments.push({ id, start, end, score });
        } else {
          accumulated_segments.at(-1).end = end;
          accumulated_segments.at(-1).score += score;
        }
      }
      results.push(
        accumulated_segments.map(
          // Convert frame-space to time-space
          // and compute the confidence
          ({ id, start, end, score }) => ({
            id,
            start: start * ratio,
            end: end * ratio,
            confidence: score / (end - start)
          })
        )
      );
    }
    return results;
  }
};

// src/models/seamless_m4t/feature_extraction_seamless_m4t.js
var SeamlessM4TFeatureExtractor = class extends FeatureExtractor {
  constructor(config) {
    super(config);
    const sampling_rate = this.config.sampling_rate;
    const mel_filters = mel_filter_bank(
      257,
      // num_frequency_bins
      this.config.num_mel_bins,
      // num_mel_filters
      20,
      // min_frequency
      Math.floor(sampling_rate / 2),
      // max_frequency
      sampling_rate,
      // sampling_rate
      null,
      // norm
      "kaldi",
      // mel_scale
      true
      // triangularize_in_mel_space
    );
    this.mel_filters = mel_filters;
    this.window = window_function(400, "povey", {
      periodic: false
    });
  }
  /**
   * Computes the log-Mel spectrogram of the provided audio waveform.
   * @param {Float32Array|Float64Array} waveform The audio waveform to process.
   * @param {number} max_length The maximum number of frames to return.
   * @returns {Promise<Tensor>} An object containing the log-Mel spectrogram data as a Float32Array and its dimensions as an array of numbers.
   */
  async _extract_fbank_features(waveform, max_length) {
    waveform = waveform.map((x) => x * 32768);
    return spectrogram(
      waveform,
      this.window,
      // window
      400,
      // frame_length
      160,
      // hop_length
      {
        fft_length: 512,
        power: 2,
        center: false,
        preemphasis: 0.97,
        mel_filters: this.mel_filters,
        log_mel: "log",
        mel_floor: 1192092955078125e-22,
        remove_dc_offset: true,
        // Custom
        max_num_frames: max_length,
        transpose: true
      }
    );
  }
  /**
   * Asynchronously extracts features from a given audio using the provided configuration.
   * @param {Float32Array|Float64Array} audio The audio data as a Float32Array/Float64Array.
   * @param {Object} options Optional parameters for feature extraction.
   * @param {boolean} [options.padding=true] Whether to pad the sequence to a multiple of `pad_to_multiple_of`.
   * @param {number} [options.pad_to_multiple_of=2] The number to pad the sequence to a multiple of.
   * @param {boolean} [options.do_normalize_per_mel_bins=true] Whether or not to zero-mean unit-variance normalize the input per mel-channel.
   * @param {boolean} [options.return_attention_mask=true] Whether to return the attention mask.
   * @returns {Promise<{ input_features: Tensor, attention_mask?: Tensor }>} A Promise resolving to an object containing the extracted input features and attention masks as Tensors.
   */
  async _call(audio, { padding = true, pad_to_multiple_of = 2, do_normalize_per_mel_bins = true, return_attention_mask = true } = {}) {
    validate_audio_inputs(audio, "SeamlessM4TFeatureExtractor");
    let features = await this._extract_fbank_features(audio, this.config.max_length);
    if (do_normalize_per_mel_bins) {
      const [num_features, feature_size] = features.dims;
      const data = features.data;
      for (let i = 0; i < feature_size; ++i) {
        let sum = 0;
        for (let j = 0; j < num_features; ++j) {
          sum += data[j * feature_size + i];
        }
        const mean2 = sum / num_features;
        let variance = 0;
        for (let j = 0; j < num_features; ++j) {
          variance += (data[j * feature_size + i] - mean2) ** 2;
        }
        variance /= num_features - 1;
        const std = Math.sqrt(variance + 1e-7);
        for (let j = 0; j < num_features; ++j) {
          const index = j * feature_size + i;
          data[index] = (data[index] - mean2) / std;
        }
      }
    }
    let padded_attention_mask;
    if (padding) {
      const [num_frames2, num_channels2] = features.dims;
      const data = (
        /** @type {Float32Array} */
        features.data
      );
      const pad_size = num_frames2 % pad_to_multiple_of;
      if (pad_size > 0) {
        const padded_data = new Float32Array(num_channels2 * (num_frames2 + pad_size));
        padded_data.set(data);
        padded_data.fill(this.config.padding_value, data.length);
        const numPaddedFrames = num_frames2 + pad_size;
        features = new Tensor2(features.type, padded_data, [numPaddedFrames, num_channels2]);
        if (return_attention_mask) {
          padded_attention_mask = new Tensor2("int64", new BigInt64Array(numPaddedFrames), [
            1,
            numPaddedFrames
          ]);
          padded_attention_mask.data.fill(1n, 0, num_frames2);
        }
      }
    }
    const [num_frames, num_channels] = features.dims;
    const stride = this.config.stride;
    const remainder = num_frames % stride;
    if (remainder !== 0) {
      throw new Error(`The number of frames (${num_frames}) must be a multiple of the stride (${stride}).`);
    }
    const input_features = features.view(1, Math.floor(num_frames / stride), num_channels * stride);
    const result = { input_features };
    if (return_attention_mask) {
      const reshapedNumFrames = input_features.dims[1];
      const attention_mask_data = new BigInt64Array(reshapedNumFrames);
      if (padded_attention_mask) {
        const padded_attention_mask_data = padded_attention_mask.data;
        for (let i = 1, j = 0; i < num_frames; i += stride, ++j) {
          attention_mask_data[j] = padded_attention_mask_data[i];
        }
      } else {
        attention_mask_data.fill(1n);
      }
      result.attention_mask = new Tensor2("int64", attention_mask_data, [1, reshapedNumFrames]);
    }
    return result;
  }
};

// src/models/snac/feature_extraction_snac.js
var SnacFeatureExtractor = class extends DacFeatureExtractor {
};

// src/models/speecht5/feature_extraction_speecht5.js
var SpeechT5FeatureExtractor = class extends FeatureExtractor {
};

// src/models/wav2vec2/feature_extraction_wav2vec2.js
var Wav2Vec2FeatureExtractor = class extends FeatureExtractor {
  /**
   * @param {Float32Array} input_values
   * @returns {Float32Array}
   */
  _zero_mean_unit_var_norm(input_values) {
    const sum = input_values.reduce((a, b) => a + b, 0);
    const mean2 = sum / input_values.length;
    const variance = input_values.reduce((a, b) => a + (b - mean2) ** 2, 0) / input_values.length;
    return input_values.map((x) => (x - mean2) / Math.sqrt(variance + 1e-7));
  }
  /**
   * Asynchronously extracts features from a given audio using the provided configuration.
   * @param {Float32Array|Float64Array} audio The audio data as a Float32Array/Float64Array.
   * @returns {Promise<{ input_values: Tensor; attention_mask: Tensor }>} A Promise resolving to an object containing the extracted input features and attention mask as Tensors.
   */
  async _call(audio) {
    validate_audio_inputs(audio, "Wav2Vec2FeatureExtractor");
    if (audio instanceof Float64Array) {
      audio = new Float32Array(audio);
    }
    let input_values = audio;
    if (this.config.do_normalize) {
      input_values = this._zero_mean_unit_var_norm(input_values);
    }
    const shape = [1, input_values.length];
    return {
      input_values: new Tensor2("float32", input_values, shape),
      attention_mask: new Tensor2("int64", new BigInt64Array(input_values.length).fill(1n), shape)
    };
  }
};

// src/models/wespeaker/feature_extraction_wespeaker.js
var WeSpeakerFeatureExtractor = class extends FeatureExtractor {
  constructor(config) {
    super(config);
    const sampling_rate = this.config.sampling_rate;
    const mel_filters = mel_filter_bank(
      257,
      // num_frequency_bins
      this.config.num_mel_bins,
      // num_mel_filters
      20,
      // min_frequency
      Math.floor(sampling_rate / 2),
      // max_frequency
      sampling_rate,
      // sampling_rate
      null,
      // norm
      "kaldi",
      // mel_scale
      true
      // triangularize_in_mel_space
    );
    this.mel_filters = mel_filters;
    this.window = window_function(400, "hamming", {
      periodic: false
    });
    this.min_num_frames = this.config.min_num_frames;
  }
  /**
   * Computes the log-Mel spectrogram of the provided audio waveform.
   * @param {Float32Array|Float64Array} waveform The audio waveform to process.
   * @returns {Promise<Tensor>} An object containing the log-Mel spectrogram data as a Float32Array and its dimensions as an array of numbers.
   */
  async _extract_fbank_features(waveform) {
    waveform = waveform.map((x) => x * 32768);
    return spectrogram(
      waveform,
      this.window,
      // window
      400,
      // frame_length
      160,
      // hop_length
      {
        fft_length: 512,
        power: 2,
        center: false,
        preemphasis: 0.97,
        mel_filters: this.mel_filters,
        log_mel: "log",
        mel_floor: 1192092955078125e-22,
        remove_dc_offset: true,
        // Custom
        transpose: true,
        min_num_frames: this.min_num_frames
      }
    );
  }
  /**
   * Asynchronously extracts features from a given audio using the provided configuration.
   * @param {Float32Array|Float64Array} audio The audio data as a Float32Array/Float64Array.
   * @returns {Promise<{ input_features: Tensor }>} A Promise resolving to an object containing the extracted input features as a Tensor.
   */
  async _call(audio) {
    validate_audio_inputs(audio, "WeSpeakerFeatureExtractor");
    const features = (await this._extract_fbank_features(audio)).unsqueeze_(0);
    if (this.config.fbank_centering_span === null) {
      const meanData = (
        /** @type {Float32Array} */
        features.mean(1).data
      );
      const featuresData = (
        /** @type {Float32Array} */
        features.data
      );
      const [batch_size, num_frames, feature_size] = features.dims;
      for (let i = 0; i < batch_size; ++i) {
        const offset1 = i * num_frames * feature_size;
        const offset2 = i * feature_size;
        for (let j = 0; j < num_frames; ++j) {
          const offset3 = offset1 + j * feature_size;
          for (let k2 = 0; k2 < feature_size; ++k2) {
            featuresData[offset3 + k2] -= meanData[offset2 + k2];
          }
        }
      }
    }
    return {
      input_features: features
    };
  }
};

// src/models/voxtral_realtime/feature_extraction_voxtral_realtime.js
var VoxtralRealtimeFeatureExtractor = class extends FeatureExtractor {
  constructor(config) {
    super(config);
    this.config.mel_filters ??= mel_filter_bank(
      Math.floor(1 + this.config.n_fft / 2),
      // num_frequency_bins
      this.config.feature_size,
      // num_mel_filters
      0,
      // min_frequency
      8e3,
      // max_frequency
      this.config.sampling_rate,
      // sampling_rate
      "slaney",
      // norm
      "slaney"
      // mel_scale
    );
    this.window = window_function(this.config.n_fft, "hann");
  }
  /**
   * Computes the log-Mel spectrogram of the provided audio waveform.
   * @param {Float32Array|Float64Array} waveform The audio waveform to process.
   * @param {Object} [options]
   * @param {boolean} [options.center=true] Whether to center-pad the waveform for STFT.
   * @returns {Promise<import('../../utils/tensor.js').Tensor>} The log-Mel spectrogram tensor of shape [num_mel_bins, num_frames].
   */
  async _extract_fbank_features(waveform, { center = true } = {}) {
    const { n_fft, hop_length, mel_filters, global_log_mel_max } = this.config;
    const max_num_frames = center ? Math.floor(waveform.length / hop_length) : Math.floor((waveform.length - n_fft) / hop_length);
    return await spectrogram(
      waveform,
      this.window,
      n_fft,
      // frame_length
      hop_length,
      {
        power: 2,
        mel_filters,
        log_mel: "log10_max_norm",
        max_log_mel: global_log_mel_max,
        center,
        max_num_frames,
        do_pad: false
      }
    );
  }
  /**
   * Extract mel spectrogram features from audio.
   * @param {Float32Array|Float64Array} audio The audio data.
   * @param {Object} [options]
   * @param {boolean} [options.center=true] Whether to center-pad the waveform.
   * @returns {Promise<{ input_features: import('../../utils/tensor.js').Tensor }>}
   */
  async _call(audio, { center = true } = {}) {
    validate_audio_inputs(audio, "VoxtralRealtimeFeatureExtractor");
    const features = await this._extract_fbank_features(audio, { center });
    return {
      input_features: features.unsqueeze_(0)
    };
  }
};

// src/models/whisper/feature_extraction_whisper.js
var WhisperFeatureExtractor = class extends FeatureExtractor {
  constructor(config) {
    super(config);
    this.config.mel_filters ??= mel_filter_bank(
      Math.floor(1 + this.config.n_fft / 2),
      // num_frequency_bins
      this.config.feature_size,
      // num_mel_filters
      0,
      // min_frequency
      8e3,
      // max_frequency
      this.config.sampling_rate,
      // sampling_rate
      "slaney",
      // norm
      "slaney"
      // mel_scale
    );
    this.window = window_function(this.config.n_fft, "hann");
  }
  /**
   * Computes the log-Mel spectrogram of the provided audio waveform.
   * @param {Float32Array|Float64Array} waveform The audio waveform to process.
   * @returns {Promise<Tensor>} An object containing the log-Mel spectrogram data as a Float32Array and its dimensions as an array of numbers.
   */
  async _extract_fbank_features(waveform) {
    return await spectrogram(
      waveform,
      this.window,
      // window
      this.config.n_fft,
      // frame_length
      this.config.hop_length,
      // hop_length
      {
        power: 2,
        mel_filters: this.config.mel_filters,
        log_mel: "log10_max_norm",
        // Custom
        max_num_frames: Math.min(
          Math.floor(waveform.length / this.config.hop_length),
          this.config.nb_max_frames
          // 3000
        )
      }
    );
  }
  /**
   * Asynchronously extracts features from a given audio using the provided configuration.
   * @param {Float32Array|Float64Array} audio The audio data as a Float32Array/Float64Array.
   * @returns {Promise<{ input_features: Tensor }>} A Promise resolving to an object containing the extracted input features as a Tensor.
   */
  async _call(audio, { max_length = null } = {}) {
    validate_audio_inputs(audio, "WhisperFeatureExtractor");
    let waveform;
    const length = max_length ?? this.config.n_samples;
    if (audio.length > length) {
      if (audio.length > this.config.n_samples) {
        logger.warn(
          "Attempting to extract features for audio longer than 30 seconds. If using a pipeline to extract transcript from a long audio clip, remember to specify `chunk_length_s` and/or `stride_length_s`."
        );
      }
      waveform = audio.slice(0, length);
    } else {
      waveform = new Float32Array(length);
      waveform.set(audio);
    }
    const features = await this._extract_fbank_features(waveform);
    return {
      input_features: features.unsqueeze_(0)
    };
  }
};

// src/models/auto/feature_extraction_auto.js
var AutoFeatureExtractor = class {
  /** @type {typeof FeatureExtractor.from_pretrained} */
  static async from_pretrained(pretrained_model_name_or_path, options = {}) {
    const preprocessorConfig = await getModelJSON(
      pretrained_model_name_or_path,
      FEATURE_EXTRACTOR_NAME,
      true,
      options
    );
    const key = preprocessorConfig.feature_extractor_type;
    const feature_extractor_class = feature_extractors_exports[key];
    if (!feature_extractor_class) {
      throw new Error(`Unknown feature_extractor_type: '${key}'. Please report this at ${GITHUB_ISSUE_URL}.`);
    }
    return new feature_extractor_class(preprocessorConfig);
  }
};

// src/models/chatterbox/processing_chatterbox.js
var ChatterboxProcessor = class extends Processor {
  static tokenizer_class = AutoTokenizer;
  static feature_extractor_class = AutoFeatureExtractor;
  async _call(text, audio = null) {
    const text_features = this.tokenizer(text);
    const audio_features = audio ? await this.feature_extractor(audio) : {};
    return { ...text_features, ...audio_features };
  }
};

// src/models/cohere_asr/processing_cohere_asr.js
var NO_SPACE_LANGUAGES = /* @__PURE__ */ new Set(["ja", "zh"]);
var CohereAsrProcessor = class extends Processor {
  static tokenizer_class = AutoTokenizer;
  static feature_extractor_class = AutoFeatureExtractor;
  static uses_processor_config = true;
  /**
   * Build the 10-token decoder prompt for the given language.
   * @param {string} [language='en'] Language code.
   * @returns {number[]} Token IDs for the decoder prompt.
   */
  get_decoder_prompt_ids(language = "en") {
    const tokens = [
      "\u2581",
      "<|startofcontext|>",
      "<|startoftranscript|>",
      "<|emo:undefined|>",
      `<|${language}|>`,
      `<|${language}|>`,
      "<|pnc|>",
      "<|noitn|>",
      "<|notimestamp|>",
      "<|nodiarize|>"
    ];
    return this.tokenizer.convert_tokens_to_ids(tokens);
  }
  /**
   * Join chunk texts back together, using the appropriate separator for the language.
   * @param {string[]} texts Decoded texts, one per chunk.
   * @param {string} [language='en'] Language code.
   * @returns {string} The joined text.
   */
  static join_chunks(texts, language = "en") {
    const non_empty = texts.filter((t) => t && t.trim());
    if (non_empty.length === 0) return "";
    const separator = NO_SPACE_LANGUAGES.has(language) ? "" : " ";
    const parts = [non_empty[0].trimEnd(), ...non_empty.slice(1).map((t) => t.trim())];
    return parts.join(separator);
  }
  /**
   * Calls the feature_extractor function with the given audio input.
   * @param {any} audio The audio input to extract features from.
   * @returns {Promise<any>}
   */
  async _call(audio) {
    return await this.feature_extractor(audio);
  }
};

// src/utils/image.js
import sharp from "sharp";
var createCanvasFunction;
var ImageDataClass;
var loadImageFunction;
if (apis.IS_WEB_ENV) {
  createCanvasFunction = (width, height) => {
    if (!self.OffscreenCanvas) {
      throw new Error("OffscreenCanvas not supported by this environment.");
    }
    return new self.OffscreenCanvas(width, height);
  };
  loadImageFunction = self.createImageBitmap;
  ImageDataClass = self.ImageData;
} else if (sharp) {
  loadImageFunction = async (img) => {
    const metadata = await img.metadata();
    const rawChannels = metadata.channels;
    const { data, info } = await img.rotate().raw().toBuffer({ resolveWithObject: true });
    const newImage = new RawImage(new Uint8ClampedArray(data), info.width, info.height, info.channels);
    if (rawChannels !== void 0 && rawChannels !== info.channels) {
      newImage.convert(rawChannels);
    }
    return newImage;
  };
} else {
  throw new Error("Unable to load image processing library.");
}
var RESAMPLING_MAPPING = {
  0: "nearest",
  1: "lanczos",
  2: "bilinear",
  3: "bicubic",
  4: "box",
  5: "hamming"
};
var CONTENT_TYPE_MAP2 = /* @__PURE__ */ new Map([
  ["png", "image/png"],
  ["jpg", "image/jpeg"],
  ["jpeg", "image/jpeg"],
  ["gif", "image/gif"]
]);
var RawImage = class _RawImage {
  /**
   * Create a new `RawImage` object.
   * @param {Uint8ClampedArray|Uint8Array} data The pixel data.
   * @param {number} width The width of the image.
   * @param {number} height The height of the image.
   * @param {1|2|3|4} channels The number of channels.
   */
  constructor(data, width, height, channels) {
    this.data = data;
    this.width = width;
    this.height = height;
    this.channels = channels;
  }
  /**
   * Returns the size of the image (width, height).
   * @returns {[number, number]} The size of the image (width, height).
   */
  get size() {
    return [this.width, this.height];
  }
  /**
   * Helper method for reading an image from a variety of input types.
   * @param {RawImage|string|URL|Blob|HTMLCanvasElement|OffscreenCanvas} input
   * @returns {Promise<RawImage>} The image object.
   *
   * **Example:** Read image from a URL.
   * ```javascript
   * let image = await RawImage.read('https://huggingface.co/datasets/Xenova/transformers.js-docs/resolve/main/football-match.jpg');
   * // RawImage {
   * //   "data": Uint8ClampedArray [ 25, 25, 25, 19, 19, 19, ... ],
   * //   "width": 800,
   * //   "height": 533,
   * //   "channels": 3
   * // }
   * ```
   */
  static async read(input) {
    if (input instanceof _RawImage) {
      return input;
    } else if (typeof input === "string" || input instanceof URL) {
      return await this.fromURL(input);
    } else if (input instanceof Blob) {
      return await this.fromBlob(input);
    } else if (typeof HTMLCanvasElement !== "undefined" && input instanceof HTMLCanvasElement || typeof OffscreenCanvas !== "undefined" && input instanceof OffscreenCanvas) {
      return this.fromCanvas(input);
    } else {
      throw new Error(`Unsupported input type: ${typeof input}`);
    }
  }
  /**
   * Read an image from a canvas.
   * @param {HTMLCanvasElement|OffscreenCanvas} canvas The canvas to read the image from.
   * @returns {RawImage} The image object.
   */
  static fromCanvas(canvas) {
    if (!apis.IS_WEB_ENV) {
      throw new Error("fromCanvas() is only supported in browser environments.");
    }
    const ctx = (
      /** @type {CanvasRenderingContext2D | OffscreenCanvasRenderingContext2D} */
      canvas.getContext("2d")
    );
    const data = ctx.getImageData(0, 0, canvas.width, canvas.height).data;
    return new _RawImage(data, canvas.width, canvas.height, 4);
  }
  /**
   * Read an image from a URL or file path.
   * @param {string|URL} url The URL or file path to read the image from.
   * @returns {Promise<RawImage>} The image object.
   */
  static async fromURL(url2) {
    const response = await getFile(url2);
    if (response.status !== 200) {
      throw new Error(`Unable to read image from "${url2}" (${response.status} ${response.statusText})`);
    }
    const blob = await response.blob();
    return this.fromBlob(blob);
  }
  /**
   * Helper method to create a new Image from a blob.
   * @param {Blob} blob The blob to read the image from.
   * @returns {Promise<RawImage>} The image object.
   */
  static async fromBlob(blob) {
    if (apis.IS_WEB_ENV) {
      const img = await loadImageFunction(blob);
      const ctx = createCanvasFunction(img.width, img.height).getContext("2d");
      ctx.drawImage(img, 0, 0);
      return new this(ctx.getImageData(0, 0, img.width, img.height).data, img.width, img.height, 4);
    } else {
      const img = sharp(await blob.arrayBuffer());
      return await loadImageFunction(img);
    }
  }
  /**
   * Helper method to create a new Image from a tensor
   * @param {Tensor} tensor
   */
  static fromTensor(tensor, channel_format = "CHW") {
    if (tensor.dims.length !== 3) {
      throw new Error(`Tensor should have 3 dimensions, but has ${tensor.dims.length} dimensions.`);
    }
    if (channel_format === "CHW") {
      tensor = tensor.transpose(1, 2, 0);
    } else if (channel_format === "HWC") {
    } else {
      throw new Error(`Unsupported channel format: ${channel_format}`);
    }
    if (!(tensor.data instanceof Uint8ClampedArray || tensor.data instanceof Uint8Array)) {
      throw new Error(`Unsupported tensor type: ${tensor.type}`);
    }
    switch (tensor.dims[2]) {
      case 1:
      case 2:
      case 3:
      case 4:
        return new _RawImage(tensor.data, tensor.dims[1], tensor.dims[0], tensor.dims[2]);
      default:
        throw new Error(`Unsupported number of channels: ${tensor.dims[2]}`);
    }
  }
  /**
   * Convert the image to grayscale format.
   * @returns {RawImage} `this` to support chaining.
   */
  grayscale() {
    if (this.channels === 1) {
      return this;
    }
    const newData = new Uint8ClampedArray(this.width * this.height * 1);
    switch (this.channels) {
      case 3:
      // rgb to grayscale
      case 4:
        for (let i = 0, offset = 0; i < this.data.length; i += this.channels) {
          const red = this.data[i];
          const green = this.data[i + 1];
          const blue = this.data[i + 2];
          newData[offset++] = Math.round(0.2989 * red + 0.587 * green + 0.114 * blue);
        }
        break;
      default:
        throw new Error(`Conversion failed due to unsupported number of channels: ${this.channels}`);
    }
    return this._update(newData, this.width, this.height, 1);
  }
  /**
   * Convert the image to RGB format.
   * @returns {RawImage} `this` to support chaining.
   */
  rgb() {
    if (this.channels === 3) {
      return this;
    }
    const newData = new Uint8ClampedArray(this.width * this.height * 3);
    switch (this.channels) {
      case 1:
        for (let i = 0, offset = 0; i < this.data.length; ++i) {
          newData[offset++] = this.data[i];
          newData[offset++] = this.data[i];
          newData[offset++] = this.data[i];
        }
        break;
      case 4:
        for (let i = 0, offset = 0; i < this.data.length; i += 4) {
          newData[offset++] = this.data[i];
          newData[offset++] = this.data[i + 1];
          newData[offset++] = this.data[i + 2];
        }
        break;
      default:
        throw new Error(`Conversion failed due to unsupported number of channels: ${this.channels}`);
    }
    return this._update(newData, this.width, this.height, 3);
  }
  /**
   * Convert the image to RGBA format.
   * @returns {RawImage} `this` to support chaining.
   */
  rgba() {
    if (this.channels === 4) {
      return this;
    }
    const newData = new Uint8ClampedArray(this.width * this.height * 4);
    switch (this.channels) {
      case 1:
        for (let i = 0, offset = 0; i < this.data.length; ++i) {
          newData[offset++] = this.data[i];
          newData[offset++] = this.data[i];
          newData[offset++] = this.data[i];
          newData[offset++] = 255;
        }
        break;
      case 3:
        for (let i = 0, offset = 0; i < this.data.length; i += 3) {
          newData[offset++] = this.data[i];
          newData[offset++] = this.data[i + 1];
          newData[offset++] = this.data[i + 2];
          newData[offset++] = 255;
        }
        break;
      default:
        throw new Error(`Conversion failed due to unsupported number of channels: ${this.channels}`);
    }
    return this._update(newData, this.width, this.height, 4);
  }
  /**
   * Apply an alpha mask to the image. Operates in place.
   * @param {RawImage} mask The mask to apply. It should have a single channel.
   * @returns {RawImage} The masked image.
   * @throws {Error} If the mask is not the same size as the image.
   * @throws {Error} If the image does not have 4 channels.
   * @throws {Error} If the mask is not a single channel.
   */
  putAlpha(mask) {
    if (mask.width !== this.width || mask.height !== this.height) {
      throw new Error(
        `Expected mask size to be ${this.width}x${this.height}, but got ${mask.width}x${mask.height}`
      );
    }
    if (mask.channels !== 1) {
      throw new Error(`Expected mask to have 1 channel, but got ${mask.channels}`);
    }
    const this_data = this.data;
    const mask_data = mask.data;
    const num_pixels = this.width * this.height;
    if (this.channels === 3) {
      const newData = new Uint8ClampedArray(num_pixels * 4);
      for (let i = 0, in_offset = 0, out_offset = 0; i < num_pixels; ++i) {
        newData[out_offset++] = this_data[in_offset++];
        newData[out_offset++] = this_data[in_offset++];
        newData[out_offset++] = this_data[in_offset++];
        newData[out_offset++] = mask_data[i];
      }
      return this._update(newData, this.width, this.height, 4);
    } else if (this.channels === 4) {
      for (let i = 0; i < num_pixels; ++i) {
        this_data[4 * i + 3] = mask_data[i];
      }
      return this;
    }
    throw new Error(`Expected image to have 3 or 4 channels, but got ${this.channels}`);
  }
  /**
   * Resize the image to the given dimensions. This method uses the canvas API to perform the resizing.
   * @param {number} width The width of the new image. `null` or `-1` will preserve the aspect ratio.
   * @param {number} height The height of the new image. `null` or `-1` will preserve the aspect ratio.
   * @param {Object} options Additional options for resizing.
   * @param {0|1|2|3|4|5|string} [options.resample] The resampling method to use.
   * @returns {Promise<RawImage>} `this` to support chaining.
   */
  async resize(width, height, { resample = 2 } = {}) {
    if (this.width === width && this.height === height) {
      return this;
    }
    let resampleMethod = RESAMPLING_MAPPING[resample] ?? resample;
    const nullish_width = isNullishDimension(width);
    const nullish_height = isNullishDimension(height);
    if (nullish_width && nullish_height) {
      return this;
    } else if (nullish_width) {
      width = height / this.height * this.width;
    } else if (nullish_height) {
      height = width / this.width * this.height;
    }
    if (apis.IS_WEB_ENV) {
      const numChannels = this.channels;
      const canvas = this.toCanvas();
      const ctx = createCanvasFunction(width, height).getContext("2d");
      ctx.drawImage(canvas, 0, 0, width, height);
      const resizedImage = new _RawImage(ctx.getImageData(0, 0, width, height).data, width, height, 4);
      return resizedImage.convert(numChannels);
    } else {
      let img = this.toSharp();
      switch (resampleMethod) {
        case "box":
        case "hamming":
          if (resampleMethod === "box" || resampleMethod === "hamming") {
            logger.warn(
              `Resampling method ${resampleMethod} is not yet supported. Using bilinear instead.`
            );
            resampleMethod = "bilinear";
          }
        case "nearest":
        case "bilinear":
        case "bicubic":
          img = img.affine([width / this.width, 0, 0, height / this.height], {
            interpolator: resampleMethod
          });
          break;
        case "lanczos":
          img = img.resize({
            width,
            height,
            fit: "fill",
            kernel: "lanczos3"
            // PIL Lanczos uses a kernel size of 3
          });
          break;
        default:
          throw new Error(`Resampling method ${resampleMethod} is not supported.`);
      }
      return await loadImageFunction(img);
    }
  }
  async pad([left, right, top, bottom]) {
    left = Math.max(left, 0);
    right = Math.max(right, 0);
    top = Math.max(top, 0);
    bottom = Math.max(bottom, 0);
    if (left === 0 && right === 0 && top === 0 && bottom === 0) {
      return this;
    }
    if (apis.IS_WEB_ENV) {
      const numChannels = this.channels;
      const canvas = this.toCanvas();
      const newWidth = this.width + left + right;
      const newHeight = this.height + top + bottom;
      const ctx = createCanvasFunction(newWidth, newHeight).getContext("2d");
      ctx.drawImage(canvas, 0, 0, this.width, this.height, left, top, this.width, this.height);
      const paddedImage = new _RawImage(ctx.getImageData(0, 0, newWidth, newHeight).data, newWidth, newHeight, 4);
      return paddedImage.convert(numChannels);
    } else {
      const img = this.toSharp().extend({ left, right, top, bottom });
      return await loadImageFunction(img);
    }
  }
  async crop([x_min, y_min, x_max, y_max]) {
    x_min = Math.max(x_min, 0);
    y_min = Math.max(y_min, 0);
    x_max = Math.min(x_max, this.width - 1);
    y_max = Math.min(y_max, this.height - 1);
    if (x_min === 0 && y_min === 0 && x_max === this.width - 1 && y_max === this.height - 1) {
      return this;
    }
    const crop_width = x_max - x_min + 1;
    const crop_height = y_max - y_min + 1;
    if (apis.IS_WEB_ENV) {
      const numChannels = this.channels;
      const canvas = this.toCanvas();
      const ctx = createCanvasFunction(crop_width, crop_height).getContext("2d");
      ctx.drawImage(canvas, x_min, y_min, crop_width, crop_height, 0, 0, crop_width, crop_height);
      const resizedImage = new _RawImage(
        ctx.getImageData(0, 0, crop_width, crop_height).data,
        crop_width,
        crop_height,
        4
      );
      return resizedImage.convert(numChannels);
    } else {
      const img = this.toSharp().extract({
        left: x_min,
        top: y_min,
        width: crop_width,
        height: crop_height
      });
      return await loadImageFunction(img);
    }
  }
  async center_crop(crop_width, crop_height) {
    if (this.width === crop_width && this.height === crop_height) {
      return this;
    }
    const width_offset = (this.width - crop_width) / 2;
    const height_offset = (this.height - crop_height) / 2;
    if (apis.IS_WEB_ENV) {
      const numChannels = this.channels;
      const canvas = this.toCanvas();
      const ctx = createCanvasFunction(crop_width, crop_height).getContext("2d");
      let sourceX = 0;
      let sourceY = 0;
      let destX = 0;
      let destY = 0;
      if (width_offset >= 0) {
        sourceX = width_offset;
      } else {
        destX = -width_offset;
      }
      if (height_offset >= 0) {
        sourceY = height_offset;
      } else {
        destY = -height_offset;
      }
      ctx.drawImage(canvas, sourceX, sourceY, crop_width, crop_height, destX, destY, crop_width, crop_height);
      const resizedImage = new _RawImage(
        ctx.getImageData(0, 0, crop_width, crop_height).data,
        crop_width,
        crop_height,
        4
      );
      return resizedImage.convert(numChannels);
    } else {
      let img = this.toSharp();
      if (width_offset >= 0 && height_offset >= 0) {
        img = img.extract({
          left: Math.floor(width_offset),
          top: Math.floor(height_offset),
          width: crop_width,
          height: crop_height
        });
      } else if (width_offset <= 0 && height_offset <= 0) {
        const top = Math.floor(-height_offset);
        const left = Math.floor(-width_offset);
        img = img.extend({
          top,
          left,
          // Ensures the resulting image has the desired dimensions
          right: crop_width - this.width - left,
          bottom: crop_height - this.height - top
        });
      } else {
        let y_padding = [0, 0];
        let y_extract = 0;
        if (height_offset < 0) {
          y_padding[0] = Math.floor(-height_offset);
          y_padding[1] = crop_height - this.height - y_padding[0];
        } else {
          y_extract = Math.floor(height_offset);
        }
        let x_padding = [0, 0];
        let x_extract = 0;
        if (width_offset < 0) {
          x_padding[0] = Math.floor(-width_offset);
          x_padding[1] = crop_width - this.width - x_padding[0];
        } else {
          x_extract = Math.floor(width_offset);
        }
        img = img.extend({
          top: y_padding[0],
          bottom: y_padding[1],
          left: x_padding[0],
          right: x_padding[1]
        }).extract({
          left: x_extract,
          top: y_extract,
          width: crop_width,
          height: crop_height
        });
      }
      return await loadImageFunction(img);
    }
  }
  async toBlob(type = "image/png", quality = 1) {
    if (!apis.IS_WEB_ENV) {
      throw new Error("toBlob() is only supported in browser environments.");
    }
    const canvas = this.toCanvas();
    return await canvas.convertToBlob({ type, quality });
  }
  toTensor(channel_format = "CHW") {
    let tensor = new Tensor2("uint8", new Uint8Array(this.data), [this.height, this.width, this.channels]);
    if (channel_format === "HWC") {
    } else if (channel_format === "CHW") {
      tensor = tensor.permute(2, 0, 1);
    } else {
      throw new Error(`Unsupported channel format: ${channel_format}`);
    }
    return tensor;
  }
  toCanvas() {
    if (!apis.IS_WEB_ENV) {
      throw new Error("toCanvas() is only supported in browser environments.");
    }
    const cloned = this.clone().rgba();
    const clonedCanvas = createCanvasFunction(cloned.width, cloned.height);
    const data = new ImageDataClass(cloned.data, cloned.width, cloned.height);
    clonedCanvas.getContext("2d").putImageData(data, 0, 0);
    return clonedCanvas;
  }
  /**
   * Split this image into individual bands. This method returns an array of individual image bands from an image.
   * For example, splitting an "RGB" image creates three new images each containing a copy of one of the original bands (red, green, blue).
   *
   * Inspired by PIL's `Image.split()` [function](https://pillow.readthedocs.io/en/latest/reference/Image.html#PIL.Image.Image.split).
   * @returns {RawImage[]} An array containing bands.
   */
  split() {
    const { data, width, height, channels } = this;
    const data_type = (
      /** @type {any} */
      data.constructor
    );
    const per_channel_length = data.length / channels;
    const split_data = Array.from({ length: channels }, () => new data_type(per_channel_length));
    for (let i = 0; i < per_channel_length; ++i) {
      const data_offset = channels * i;
      for (let j = 0; j < channels; ++j) {
        split_data[j][i] = data[data_offset + j];
      }
    }
    return split_data.map((data2) => new _RawImage(data2, width, height, 1));
  }
  /**
   * Helper method to update the image data.
   * @param {Uint8ClampedArray} data The new image data.
   * @param {number} width The new width of the image.
   * @param {number} height The new height of the image.
   * @param {1|2|3|4|null} [channels] The new number of channels of the image.
   * @private
   */
  _update(data, width, height, channels = null) {
    this.data = data;
    this.width = width;
    this.height = height;
    if (channels !== null) {
      this.channels = channels;
    }
    return this;
  }
  /**
   * Clone the image
   * @returns {RawImage} The cloned image
   */
  clone() {
    return new _RawImage(this.data.slice(), this.width, this.height, this.channels);
  }
  /**
   * Helper method for converting image to have a certain number of channels
   * @param {number} numChannels The number of channels. Must be 1, 3, or 4.
   * @returns {RawImage} `this` to support chaining.
   */
  convert(numChannels) {
    if (this.channels === numChannels) return this;
    switch (numChannels) {
      case 1:
        this.grayscale();
        break;
      case 3:
        this.rgb();
        break;
      case 4:
        this.rgba();
        break;
      default:
        throw new Error(`Conversion failed due to unsupported number of channels: ${this.channels}`);
    }
    return this;
  }
  /**
   * Save the image to the given path.
   * @param {string} path The path to save the image to.
   * @returns {Promise<void>}
   */
  async save(path3) {
    if (apis.IS_WEB_ENV) {
      if (apis.IS_WEBWORKER_ENV) {
        throw new Error("Unable to save an image from a Web Worker.");
      }
      const extension = path3.split(".").pop().toLowerCase();
      const mime = CONTENT_TYPE_MAP2.get(extension) ?? "image/png";
      const blob = await this.toBlob(mime);
      return saveBlob(path3, blob);
    } else if (apis.IS_FS_AVAILABLE) {
      const img = this.toSharp();
      await img.toFile(path3);
    } else {
      throw new Error("Unable to save the image because filesystem is disabled in this environment.");
    }
  }
  /**
   * Convert the image to a Sharp instance.
   * @returns {import('sharp').Sharp} The Sharp instance.
   */
  toSharp() {
    if (apis.IS_WEB_ENV) {
      throw new Error("toSharp() is only supported in server-side environments.");
    }
    return sharp(this.data, {
      raw: {
        width: this.width,
        height: this.height,
        channels: this.channels
      }
    });
  }
};
var load_image = RawImage.read.bind(RawImage);

// src/image_processors_utils.js
function constraint_to_multiple_of(val, multiple, minVal = 0, maxVal = null) {
  const a = val / multiple;
  let x = bankers_round(a) * multiple;
  if (maxVal !== null && x > maxVal) {
    x = Math.floor(a) * multiple;
  }
  if (x < minVal) {
    x = Math.ceil(a) * multiple;
  }
  return x;
}
function enforce_size_divisibility([width, height], divisor) {
  return [Math.max(Math.floor(width / divisor), 1) * divisor, Math.max(Math.floor(height / divisor), 1) * divisor];
}
function center_to_corners_format([centerX, centerY, width, height]) {
  return [centerX - width / 2, centerY - height / 2, centerX + width / 2, centerY + height / 2];
}
function post_process_object_detection(outputs, threshold = 0.5, target_sizes = null, is_zero_shot = false) {
  const out_logits = outputs.logits;
  const out_bbox = outputs.pred_boxes;
  const [batch_size, num_boxes, num_classes] = out_logits.dims;
  if (target_sizes !== null && target_sizes.length !== batch_size) {
    throw Error("Make sure that you pass in as many target sizes as the batch dimension of the logits");
  }
  let toReturn = [];
  for (let i = 0; i < batch_size; ++i) {
    let target_size = target_sizes !== null ? target_sizes[i] : null;
    let info = {
      boxes: [],
      classes: [],
      scores: []
    };
    let logits = out_logits[i];
    let bbox = out_bbox[i];
    for (let j = 0; j < num_boxes; ++j) {
      let logit = logits[j];
      let indices = [];
      let probs;
      if (is_zero_shot) {
        probs = logit.sigmoid().data;
        for (let k2 = 0; k2 < probs.length; ++k2) {
          if (probs[k2] > threshold) {
            indices.push(k2);
          }
        }
      } else {
        let maxIndex = max(logit.data)[1];
        if (maxIndex === num_classes - 1) {
          continue;
        }
        probs = softmax(logit.data);
        if (probs[maxIndex] < threshold) {
          continue;
        }
        indices.push(maxIndex);
      }
      for (const index of indices) {
        let box = bbox[j].data;
        box = center_to_corners_format(box);
        if (target_size !== null) {
          box = box.map((x, i2) => x * target_size[(i2 + 1) % 2]);
        }
        info.boxes.push(box);
        info.classes.push(index);
        info.scores.push(probs[index]);
      }
    }
    toReturn.push(info);
  }
  return toReturn;
}
function post_process_semantic_segmentation(outputs, target_sizes = null) {
  const logits = outputs.logits;
  const batch_size = logits.dims[0];
  if (target_sizes !== null && target_sizes.length !== batch_size) {
    throw Error("Make sure that you pass in as many target sizes as the batch dimension of the logits");
  }
  const toReturn = [];
  for (let i = 0; i < batch_size; ++i) {
    const target_size = target_sizes !== null ? target_sizes[i] : null;
    let data = logits[i];
    if (target_size !== null) {
      data = interpolate(data, target_size, "bilinear", false);
    }
    const [height, width] = target_size ?? data.dims.slice(-2);
    const segmentation = new Tensor2("int32", new Int32Array(height * width), [height, width]);
    const buffer = data[0].data;
    const segmentation_data = segmentation.data;
    for (let j = 1; j < data.dims[0]; ++j) {
      const row = data[j].data;
      for (let k2 = 0; k2 < row.length; ++k2) {
        if (row[k2] > buffer[k2]) {
          buffer[k2] = row[k2];
          segmentation_data[k2] = j;
        }
      }
    }
    const hasLabel = new Array(data.dims[0]);
    for (let j = 0; j < segmentation_data.length; ++j) {
      const index = segmentation_data[j];
      hasLabel[index] = index;
    }
    const labels = hasLabel.filter((x) => x !== void 0);
    toReturn.push({ segmentation, labels });
  }
  return toReturn;
}
function remove_low_and_no_objects(class_logits, mask_logits, object_mask_threshold, num_labels) {
  const mask_probs_item = [];
  const pred_scores_item = [];
  const pred_labels_item = [];
  for (let j = 0; j < class_logits.dims[0]; ++j) {
    const cls = class_logits[j];
    const mask = mask_logits[j];
    const pred_label = max(cls.data)[1];
    if (pred_label === num_labels) {
      continue;
    }
    const scores = softmax(cls.data);
    const pred_score = scores[pred_label];
    if (pred_score > object_mask_threshold) {
      mask_probs_item.push(mask);
      pred_scores_item.push(pred_score);
      pred_labels_item.push(pred_label);
    }
  }
  return [mask_probs_item, pred_scores_item, pred_labels_item];
}
function check_segment_validity(mask_labels, mask_probs, k2, mask_threshold = 0.5, overlap_mask_area_threshold = 0.8) {
  const mask_k = [];
  let mask_k_area = 0;
  let original_area = 0;
  const mask_probs_k_data = mask_probs[k2].data;
  for (let i = 0; i < mask_labels.length; ++i) {
    if (mask_labels[i] === k2) {
      mask_k.push(i);
      ++mask_k_area;
    }
    if (mask_probs_k_data[i] >= mask_threshold) {
      ++original_area;
    }
  }
  let mask_exists = mask_k_area > 0 && original_area > 0;
  if (mask_exists) {
    let area_ratio = mask_k_area / original_area;
    mask_exists = area_ratio > overlap_mask_area_threshold;
  }
  return [mask_exists, mask_k];
}
function compute_segments(mask_probs, pred_scores, pred_labels, mask_threshold, overlap_mask_area_threshold, label_ids_to_fuse = null, target_size = null) {
  const [height, width] = target_size ?? mask_probs[0].dims;
  const segmentation = new Tensor2("int32", new Int32Array(height * width), [height, width]);
  const segments = [];
  if (target_size !== null) {
    for (let i = 0; i < mask_probs.length; ++i) {
      mask_probs[i] = interpolate(mask_probs[i], target_size, "bilinear", false);
    }
  }
  const mask_labels = new Int32Array(mask_probs[0].data.length);
  const bestScores = new Float32Array(mask_probs[0].data.length);
  for (let i = 0; i < mask_probs.length; ++i) {
    let score = pred_scores[i];
    const mask_probs_i_data = mask_probs[i].data;
    for (let j = 0; j < mask_probs_i_data.length; ++j) {
      mask_probs_i_data[j] *= score;
      if (mask_probs_i_data[j] > bestScores[j]) {
        mask_labels[j] = i;
        bestScores[j] = mask_probs_i_data[j];
      }
    }
  }
  let current_segment_id = 0;
  const segmentation_data = segmentation.data;
  for (let k2 = 0; k2 < pred_labels.length; ++k2) {
    const pred_class = pred_labels[k2];
    const [mask_exists, mask_k] = check_segment_validity(
      mask_labels,
      mask_probs,
      k2,
      mask_threshold,
      overlap_mask_area_threshold
    );
    if (!mask_exists) {
      continue;
    }
    ++current_segment_id;
    for (const index of mask_k) {
      segmentation_data[index] = current_segment_id;
    }
    segments.push({
      id: current_segment_id,
      label_id: pred_class,
      // was_fused: should_fuse, TODO
      score: pred_scores[k2]
    });
  }
  return [segmentation, segments];
}
function smart_resize(height, width, factor = 28, min_pixels = 56 * 56, max_pixels = 14 * 14 * 4 * 1280, temporal_factor = 1) {
  if (height < factor || width < factor) {
    const scale = Math.max(factor / height, factor / width);
    height = Math.round(height * scale);
    width = Math.round(width * scale);
  }
  if (Math.max(height, width) / Math.min(height, width) > 200) {
    throw new Error(
      `absolute aspect ratio must be smaller than 200, got ${Math.max(height, width) / Math.min(height, width)}`
    );
  }
  let h_bar = Math.round(height / factor) * factor;
  let w_bar = Math.round(width / factor) * factor;
  if (temporal_factor * h_bar * w_bar > max_pixels) {
    const beta = Math.sqrt(temporal_factor * height * width / max_pixels);
    h_bar = Math.max(factor, Math.floor(height / beta / factor) * factor);
    w_bar = Math.max(factor, Math.floor(width / beta / factor) * factor);
  } else if (temporal_factor * h_bar * w_bar < min_pixels) {
    const beta = Math.sqrt(min_pixels / (temporal_factor * height * width));
    h_bar = Math.ceil(height * beta / factor) * factor;
    w_bar = Math.ceil(width * beta / factor) * factor;
  }
  return [w_bar, h_bar];
}
function post_process_panoptic_segmentation(outputs, threshold = 0.5, mask_threshold = 0.5, overlap_mask_area_threshold = 0.8, label_ids_to_fuse = null, target_sizes = null) {
  if (label_ids_to_fuse === null) {
    logger.warn("`label_ids_to_fuse` unset. No instance will be fused.");
    label_ids_to_fuse = /* @__PURE__ */ new Set();
  }
  const class_queries_logits = outputs.class_queries_logits ?? outputs.logits;
  const masks_queries_logits = outputs.masks_queries_logits ?? outputs.pred_masks;
  const mask_probs = masks_queries_logits.sigmoid();
  let [batch_size, num_queries, num_labels] = class_queries_logits.dims;
  num_labels -= 1;
  if (target_sizes !== null && target_sizes.length !== batch_size) {
    throw Error("Make sure that you pass in as many target sizes as the batch dimension of the logits");
  }
  let toReturn = [];
  for (let i = 0; i < batch_size; ++i) {
    let target_size = target_sizes !== null ? target_sizes[i] : null;
    let class_logits = class_queries_logits[i];
    let mask_logits = mask_probs[i];
    let [mask_probs_item, pred_scores_item, pred_labels_item] = remove_low_and_no_objects(
      class_logits,
      mask_logits,
      threshold,
      num_labels
    );
    if (pred_labels_item.length === 0) {
      let [height, width] = target_size ?? mask_logits.dims.slice(-2);
      let segmentation2 = new Tensor2("int32", new Int32Array(height * width).fill(-1), [height, width]);
      toReturn.push({
        segmentation: segmentation2,
        segments_info: []
      });
      continue;
    }
    let [segmentation, segments] = compute_segments(
      mask_probs_item,
      pred_scores_item,
      pred_labels_item,
      mask_threshold,
      overlap_mask_area_threshold,
      label_ids_to_fuse,
      target_size
    );
    toReturn.push({
      segmentation,
      segments_info: segments
    });
  }
  return toReturn;
}
function post_process_instance_segmentation(outputs, threshold = 0.5, target_sizes = null) {
  throw new Error("`post_process_instance_segmentation` is not yet implemented.");
}
var ImageProcessor = class extends Callable {
  /**
   * Constructs a new `ImageProcessor`.
   * @param {ImageProcessorConfig} config The configuration object.
   */
  constructor(config) {
    super();
    this.image_mean = config.image_mean ?? config.mean;
    this.image_std = config.image_std ?? config.std;
    this.resample = config.resample ?? 2;
    this.do_rescale = config.do_rescale ?? true;
    this.rescale_factor = config.rescale_factor ?? 1 / 255;
    this.do_normalize = config.do_normalize;
    this.do_thumbnail = config.do_thumbnail;
    this.size = config.size ?? config.image_size;
    this.do_resize = config.do_resize ?? this.size !== void 0;
    this.size_divisibility = config.size_divisibility ?? config.size_divisor;
    this.do_center_crop = config.do_center_crop;
    this.crop_size = config.crop_size;
    this.do_convert_rgb = config.do_convert_rgb ?? true;
    this.do_crop_margin = config.do_crop_margin;
    this.pad_size = config.pad_size;
    this.do_pad = config.do_pad;
    this.min_pixels = config.min_pixels;
    this.max_pixels = config.max_pixels;
    if (this.do_pad && !this.pad_size && !this.size_divisibility && this.size && this.size.width !== void 0 && this.size.height !== void 0) {
      this.pad_size = this.size;
    }
    this.do_flip_channel_order = config.do_flip_channel_order ?? false;
    this.config = config;
  }
  /**
   * Resize the image to make a thumbnail. The image is resized so that no dimension is larger than any
   * corresponding dimension of the specified size.
   * @param {RawImage} image The image to be resized.
   * @param {{height:number, width:number}} size The size `{"height": h, "width": w}` to resize the image to.
   * @param {string | 0 | 1 | 2 | 3 | 4 | 5} [resample=2] The resampling filter to use.
   * @returns {Promise<RawImage>} The resized image.
   */
  async thumbnail(image, size, resample = 2) {
    const input_height = image.height;
    const input_width = image.width;
    const output_height = size.height;
    const output_width = size.width;
    let height = Math.min(input_height, output_height);
    let width = Math.min(input_width, output_width);
    if (height === input_height && width === input_width) {
      return image;
    }
    if (input_height > input_width) {
      width = Math.floor(input_width * height / input_height);
    } else if (input_width > input_height) {
      height = Math.floor(input_height * width / input_width);
    }
    return await image.resize(width, height, { resample });
  }
  /**
   * Crops the margin of the image. Gray pixels are considered margin (i.e., pixels with a value below the threshold).
   * @param {RawImage} image The image to be cropped.
   * @param {number} gray_threshold Value below which pixels are considered to be gray.
   * @returns {Promise<RawImage>} The cropped image.
   */
  async crop_margin(image, gray_threshold = 200) {
    const gray_image = image.clone().grayscale();
    const minValue = min2(gray_image.data)[0];
    const maxValue = max(gray_image.data)[0];
    const diff = maxValue - minValue;
    if (diff === 0) {
      return image;
    }
    const threshold = gray_threshold / 255;
    let x_min = gray_image.width, y_min = gray_image.height, x_max = 0, y_max = 0;
    const gray_image_data = gray_image.data;
    for (let j = 0; j < gray_image.height; ++j) {
      const row = j * gray_image.width;
      for (let i = 0; i < gray_image.width; ++i) {
        if ((gray_image_data[row + i] - minValue) / diff < threshold) {
          x_min = Math.min(x_min, i);
          y_min = Math.min(y_min, j);
          x_max = Math.max(x_max, i);
          y_max = Math.max(y_max, j);
        }
      }
    }
    image = await image.crop([x_min, y_min, x_max, y_max]);
    return image;
  }
  /**
   * Pad the image by a certain amount.
   * @param {Float32Array} pixelData The pixel data to pad.
   * @param {number[]} imgDims The dimensions of the image (height, width, channels).
   * @param {{width:number; height:number}|number|'square'} padSize The dimensions of the padded image.
   * @param {Object} options The options for padding.
   * @param {'constant'|'symmetric'} [options.mode='constant'] The type of padding to add.
   * @param {boolean} [options.center=false] Whether to center the image.
   * @param {number|number[]} [options.constant_values=0] The constant value to use for padding.
   * @returns {[Float32Array, number[]]} The padded pixel data and image dimensions.
   */
  pad_image(pixelData, imgDims, padSize, { mode = "constant", center = false, constant_values = 0 } = {}) {
    const [imageHeight, imageWidth, imageChannels] = imgDims;
    let paddedImageWidth, paddedImageHeight;
    if (typeof padSize === "number") {
      paddedImageWidth = padSize;
      paddedImageHeight = padSize;
    } else if (padSize === "square") {
      paddedImageWidth = paddedImageHeight = Math.max(imageHeight, imageWidth);
    } else {
      paddedImageWidth = padSize.width;
      paddedImageHeight = padSize.height;
    }
    if (paddedImageWidth !== imageWidth || paddedImageHeight !== imageHeight) {
      const paddedPixelData = new Float32Array(paddedImageWidth * paddedImageHeight * imageChannels);
      if (Array.isArray(constant_values)) {
        for (let i = 0; i < paddedPixelData.length; ++i) {
          paddedPixelData[i] = constant_values[i % imageChannels];
        }
      } else if (constant_values !== 0) {
        paddedPixelData.fill(constant_values);
      }
      const [left, top] = center ? [Math.floor((paddedImageWidth - imageWidth) / 2), Math.floor((paddedImageHeight - imageHeight) / 2)] : [0, 0];
      for (let i = 0; i < imageHeight; ++i) {
        const a = (i + top) * paddedImageWidth;
        const b = i * imageWidth;
        for (let j = 0; j < imageWidth; ++j) {
          const c = (a + j + left) * imageChannels;
          const d = (b + j) * imageChannels;
          for (let k2 = 0; k2 < imageChannels; ++k2) {
            paddedPixelData[c + k2] = pixelData[d + k2];
          }
        }
      }
      if (mode === "symmetric") {
        if (center) {
          throw new Error("`center` padding is not supported when `mode` is set to `symmetric`.");
        }
        const h1 = imageHeight - 1;
        const w1 = imageWidth - 1;
        for (let i = 0; i < paddedImageHeight; ++i) {
          const a = i * paddedImageWidth;
          const b = calculateReflectOffset(i, h1) * imageWidth;
          for (let j = 0; j < paddedImageWidth; ++j) {
            if (i < imageHeight && j < imageWidth) continue;
            const c = (a + j) * imageChannels;
            const d = (b + calculateReflectOffset(j, w1)) * imageChannels;
            for (let k2 = 0; k2 < imageChannels; ++k2) {
              paddedPixelData[c + k2] = pixelData[d + k2];
            }
          }
        }
      }
      pixelData = paddedPixelData;
      imgDims = [paddedImageHeight, paddedImageWidth, imageChannels];
    }
    return [pixelData, imgDims];
  }
  /**
   * Rescale the image' pixel values by `this.rescale_factor`.
   * @param {Float32Array} pixelData The pixel data to rescale.
   * @returns {void}
   */
  rescale(pixelData) {
    for (let i = 0; i < pixelData.length; ++i) {
      pixelData[i] = this.rescale_factor * pixelData[i];
    }
  }
  /**
   * Find the target (width, height) dimension of the output image after
   * resizing given the input image and the desired size.
   * @param {RawImage} image The image to resize.
   * @param {any} size The size to use for resizing the image.
   * @returns {[number, number]} The target (width, height) dimension of the output image after resizing.
   */
  get_resize_output_image_size(image, size) {
    const [srcWidth, srcHeight] = image.size;
    let shortest_edge;
    let longest_edge;
    if (this.do_thumbnail) {
      const { height, width } = size;
      shortest_edge = Math.min(height, width);
    } else if (Number.isInteger(size)) {
      shortest_edge = size;
      longest_edge = this.config.max_size ?? shortest_edge;
    } else if (size !== void 0) {
      shortest_edge = size.shortest_edge;
      longest_edge = size.longest_edge;
    }
    if (shortest_edge !== void 0 || longest_edge !== void 0) {
      const shortResizeFactor = shortest_edge === void 0 ? 1 : Math.max(shortest_edge / srcWidth, shortest_edge / srcHeight);
      const newWidth = srcWidth * shortResizeFactor;
      const newHeight = srcHeight * shortResizeFactor;
      const longResizeFactor = longest_edge === void 0 ? 1 : Math.min(longest_edge / newWidth, longest_edge / newHeight);
      let finalWidth = Math.floor(Number((newWidth * longResizeFactor).toFixed(2)));
      let finalHeight = Math.floor(Number((newHeight * longResizeFactor).toFixed(2)));
      if (this.size_divisibility !== void 0) {
        [finalWidth, finalHeight] = enforce_size_divisibility(
          [finalWidth, finalHeight],
          this.size_divisibility
        );
      }
      return [finalWidth, finalHeight];
    } else if (size !== void 0 && size.width !== void 0 && size.height !== void 0) {
      let newWidth = size.width;
      let newHeight = size.height;
      if (this.config.keep_aspect_ratio && this.config.ensure_multiple_of) {
        let scale_height = newHeight / srcHeight;
        let scale_width = newWidth / srcWidth;
        if (Math.abs(1 - scale_width) < Math.abs(1 - scale_height)) {
          scale_height = scale_width;
        } else {
          scale_width = scale_height;
        }
        newHeight = constraint_to_multiple_of(scale_height * srcHeight, this.config.ensure_multiple_of);
        newWidth = constraint_to_multiple_of(scale_width * srcWidth, this.config.ensure_multiple_of);
      }
      return [newWidth, newHeight];
    } else if (this.size_divisibility !== void 0) {
      return enforce_size_divisibility([srcWidth, srcHeight], this.size_divisibility);
    } else {
      throw new Error(
        `Could not resize image due to unsupported \`this.size\` option in config: ${JSON.stringify(size)}`
      );
    }
  }
  /**
   * Resizes the image.
   * @param {RawImage} image The image to resize.
   * @returns {Promise<RawImage>} The resized image.
   */
  async resize(image) {
    const [newWidth, newHeight] = this.get_resize_output_image_size(image, this.size);
    return await image.resize(newWidth, newHeight, {
      // @ts-expect-error TS2322
      resample: this.resample
    });
  }
  /**
   * @typedef {Object} PreprocessedImage
   * @property {HeightWidth} original_size The original size of the image.
   * @property {HeightWidth} reshaped_input_size The reshaped input size of the image.
   * @property {Tensor} pixel_values The pixel values of the preprocessed image.
   */
  /**
   * Preprocesses the given image.
   *
   * @param {RawImage} image The image to preprocess.
   * @param {Object} overrides The overrides for the preprocessing options.
   * @returns {Promise<PreprocessedImage>} The preprocessed image.
   */
  async preprocess(image, {
    do_normalize = null,
    do_pad = null,
    do_convert_rgb = null,
    do_convert_grayscale = null,
    do_flip_channel_order = null
  } = {}) {
    if (this.do_crop_margin) {
      image = await this.crop_margin(image);
    }
    const [srcWidth, srcHeight] = image.size;
    if (do_convert_rgb ?? this.do_convert_rgb) {
      image = image.rgb();
    } else if (do_convert_grayscale) {
      image = image.grayscale();
    }
    if (this.do_resize) {
      image = await this.resize(image);
    }
    if (this.do_thumbnail) {
      image = await this.thumbnail(image, this.size, this.resample);
    }
    if (this.do_center_crop) {
      let crop_width;
      let crop_height;
      if (Number.isInteger(this.crop_size)) {
        crop_width = this.crop_size;
        crop_height = this.crop_size;
      } else {
        crop_width = this.crop_size.width;
        crop_height = this.crop_size.height;
      }
      image = await image.center_crop(crop_width, crop_height);
    }
    const reshaped_input_size = [image.height, image.width];
    let pixelData = Float32Array.from(image.data);
    let imgDims = [image.height, image.width, image.channels];
    if (this.do_rescale) {
      this.rescale(pixelData);
    }
    if (do_normalize ?? this.do_normalize) {
      let image_mean = this.image_mean;
      if (!Array.isArray(this.image_mean)) {
        image_mean = new Array(image.channels).fill(image_mean);
      }
      let image_std = this.image_std;
      if (!Array.isArray(this.image_std)) {
        image_std = new Array(image.channels).fill(image_std);
      }
      if (image_mean.length !== image.channels || image_std.length !== image.channels) {
        throw new Error(
          `When set to arrays, the length of \`image_mean\` (${image_mean.length}) and \`image_std\` (${image_std.length}) must match the number of channels in the image (${image.channels}).`
        );
      }
      for (let i = 0; i < pixelData.length; i += image.channels) {
        for (let j = 0; j < image.channels; ++j) {
          pixelData[i + j] = (pixelData[i + j] - image_mean[j]) / image_std[j];
        }
      }
    }
    if (do_pad ?? this.do_pad) {
      if (this.pad_size) {
        const padded = this.pad_image(pixelData, [image.height, image.width, image.channels], this.pad_size);
        [pixelData, imgDims] = padded;
      } else if (this.size_divisibility) {
        const paddedWidth = Math.ceil(imgDims[1] / this.size_divisibility) * this.size_divisibility;
        const paddedHeight = Math.ceil(imgDims[0] / this.size_divisibility) * this.size_divisibility;
        [pixelData, imgDims] = this.pad_image(pixelData, imgDims, { width: paddedWidth, height: paddedHeight });
      }
    }
    if (do_flip_channel_order ?? this.do_flip_channel_order) {
      if (imgDims[2] !== 3) {
        throw new Error("Flipping channel order is only supported for RGB images.");
      }
      for (let i = 0; i < pixelData.length; i += 3) {
        const temp = pixelData[i];
        pixelData[i] = pixelData[i + 2];
        pixelData[i + 2] = temp;
      }
    }
    const pixel_values = new Tensor2("float32", pixelData, imgDims).permute(2, 0, 1);
    return {
      original_size: [srcHeight, srcWidth],
      reshaped_input_size,
      pixel_values
    };
  }
  /**
   * Calls the feature extraction process on an array of images,
   * preprocesses each image, and concatenates the resulting
   * features into a single Tensor.
   * @param {RawImage[]} images The image(s) to extract features from.
   * @param {...any} args Additional arguments.
   * @returns {Promise<ImageProcessorResult>} An object containing the concatenated pixel values (and other metadata) of the preprocessed images.
   */
  async _call(images, ...args) {
    if (!Array.isArray(images)) {
      images = [images];
    }
    const imageData = await Promise.all(images.map((x) => this.preprocess(x)));
    const pixel_values = stack(
      imageData.map((x) => x.pixel_values),
      0
    );
    return {
      pixel_values,
      // Original sizes of images
      original_sizes: imageData.map((x) => x.original_size),
      // Reshaped sizes of images, before padding or cropping
      reshaped_input_sizes: imageData.map((x) => x.reshaped_input_size)
    };
  }
  /**
   * Instantiate one of the processor classes of the library from a pretrained model.
   *
   * The processor class to instantiate is selected based on the `image_processor_type` (or `feature_extractor_type`; legacy)
   * property of the config object (either passed as an argument or loaded from `pretrained_model_name_or_path` if possible)
   *
   * @param {string} pretrained_model_name_or_path The name or path of the pretrained model. Can be either:
   * - A string, the *model id* of a pretrained processor hosted inside a model repo on huggingface.co.
   *   Valid model ids can be located at the root-level, like `bert-base-uncased`, or namespaced under a
   *   user or organization name, like `dbmdz/bert-base-german-cased`.
   * - A path to a *directory* containing processor files, e.g., `./my_model_directory/`.
   * @param {import('./utils/hub.js').PretrainedOptions} options Additional options for loading the processor.
   *
   * @returns {Promise<ImageProcessor>} A new instance of the Processor class.
   */
  static async from_pretrained(pretrained_model_name_or_path, options = {}) {
    const preprocessorConfig = await getModelJSON(
      pretrained_model_name_or_path,
      IMAGE_PROCESSOR_NAME,
      true,
      options
    );
    return new this(preprocessorConfig);
  }
};

// src/models/image_processors.js
var image_processors_exports = {};
__export(image_processors_exports, {
  BeitFeatureExtractor: () => BeitFeatureExtractor,
  BitImageProcessor: () => BitImageProcessor,
  CHMv2ImageProcessor: () => CHMv2ImageProcessor,
  CLIPFeatureExtractor: () => CLIPFeatureExtractor,
  CLIPImageProcessor: () => CLIPImageProcessor,
  ChineseCLIPFeatureExtractor: () => ChineseCLIPFeatureExtractor,
  ConvNextFeatureExtractor: () => ConvNextFeatureExtractor,
  ConvNextImageProcessor: () => ConvNextImageProcessor,
  DINOv3ViTImageProcessor: () => DINOv3ViTImageProcessor,
  DPTFeatureExtractor: () => DPTFeatureExtractor,
  DPTImageProcessor: () => DPTImageProcessor,
  DeiTFeatureExtractor: () => DeiTFeatureExtractor,
  DeiTImageProcessor: () => DeiTImageProcessor,
  DetrFeatureExtractor: () => DetrFeatureExtractor,
  DetrImageProcessor: () => DetrImageProcessor,
  DonutFeatureExtractor: () => DonutFeatureExtractor,
  DonutImageProcessor: () => DonutImageProcessor,
  EfficientNetImageProcessor: () => EfficientNetImageProcessor,
  GLPNFeatureExtractor: () => GLPNFeatureExtractor,
  Gemma3ImageProcessor: () => Gemma3ImageProcessor,
  Gemma4ImageProcessor: () => Gemma4ImageProcessor,
  Glm46VImageProcessor: () => Glm46VImageProcessor,
  GroundingDinoImageProcessor: () => GroundingDinoImageProcessor,
  Idefics3ImageProcessor: () => Idefics3ImageProcessor,
  ImageFeatureExtractor: () => ImageProcessor,
  ImageProcessor: () => ImageProcessor,
  JinaCLIPImageProcessor: () => JinaCLIPImageProcessor,
  Lfm2VlImageProcessor: () => Lfm2VlImageProcessor,
  LlavaOnevisionImageProcessor: () => LlavaOnevisionImageProcessor,
  Mask2FormerImageProcessor: () => Mask2FormerImageProcessor,
  MaskFormerFeatureExtractor: () => MaskFormerFeatureExtractor,
  MaskFormerImageProcessor: () => MaskFormerImageProcessor,
  MobileNetV1FeatureExtractor: () => MobileNetV1FeatureExtractor,
  MobileNetV1ImageProcessor: () => MobileNetV1ImageProcessor,
  MobileNetV2FeatureExtractor: () => MobileNetV2FeatureExtractor,
  MobileNetV2ImageProcessor: () => MobileNetV2ImageProcessor,
  MobileNetV3FeatureExtractor: () => MobileNetV3FeatureExtractor,
  MobileNetV3ImageProcessor: () => MobileNetV3ImageProcessor,
  MobileNetV4FeatureExtractor: () => MobileNetV4FeatureExtractor,
  MobileNetV4ImageProcessor: () => MobileNetV4ImageProcessor,
  MobileViTFeatureExtractor: () => MobileViTFeatureExtractor,
  MobileViTImageProcessor: () => MobileViTImageProcessor,
  NougatImageProcessor: () => NougatImageProcessor,
  OwlViTFeatureExtractor: () => OwlViTFeatureExtractor,
  OwlViTImageProcessor: () => OwlViTImageProcessor,
  Owlv2ImageProcessor: () => Owlv2ImageProcessor,
  Phi3VImageProcessor: () => Phi3VImageProcessor,
  PixtralImageProcessor: () => PixtralImageProcessor,
  PvtImageProcessor: () => PvtImageProcessor,
  Qwen2VLImageProcessor: () => Qwen2VLImageProcessor,
  RTDetrImageProcessor: () => RTDetrImageProcessor,
  Sam2ImageProcessor: () => SamImageProcessor,
  Sam3ImageProcessor: () => SamImageProcessor,
  SamImageProcessor: () => SamImageProcessor,
  SapiensFeatureExtractor: () => SapiensFeatureExtractor,
  SapiensImageProcessor: () => SapiensImageProcessor,
  SegformerFeatureExtractor: () => SegformerFeatureExtractor,
  SegformerImageProcessor: () => SegformerImageProcessor,
  SiglipImageProcessor: () => SiglipImageProcessor,
  SmolVLMImageProcessor: () => Idefics3ImageProcessor,
  Swin2SRImageProcessor: () => Swin2SRImageProcessor,
  VLMImageProcessor: () => VLMImageProcessor,
  ViTFeatureExtractor: () => ViTFeatureExtractor,
  ViTImageProcessor: () => ViTImageProcessor,
  VitMatteImageProcessor: () => VitMatteImageProcessor,
  VitPoseImageProcessor: () => VitPoseImageProcessor,
  YolosFeatureExtractor: () => YolosFeatureExtractor,
  YolosImageProcessor: () => YolosImageProcessor
});

// src/models/beit/image_processing_beit.js
var BeitFeatureExtractor = class extends ImageProcessor {
};

// src/models/bit/image_processing_bit.js
var BitImageProcessor = class extends ImageProcessor {
};

// src/models/chinese_clip/image_processing_chinese_clip.js
var ChineseCLIPFeatureExtractor = class extends ImageProcessor {
};

// src/models/chmv2/image_processing_chmv2.js
var CHMv2ImageProcessor = class extends ImageProcessor {
};

// src/models/clip/image_processing_clip.js
var CLIPImageProcessor = class extends ImageProcessor {
};
var CLIPFeatureExtractor = class extends CLIPImageProcessor {
};

// src/models/convnext/image_processing_convnext.js
var ConvNextImageProcessor = class extends ImageProcessor {
  constructor(config) {
    super(config);
    this.crop_pct = this.config.crop_pct ?? 224 / 256;
  }
  async resize(image) {
    const shortest_edge = this.size?.shortest_edge;
    if (shortest_edge === void 0) {
      throw new Error(`Size dictionary must contain 'shortest_edge' key.`);
    }
    if (shortest_edge < 384) {
      const resize_shortest_edge = Math.floor(shortest_edge / this.crop_pct);
      const [newWidth, newHeight] = this.get_resize_output_image_size(image, {
        shortest_edge: resize_shortest_edge
      });
      image = await image.resize(newWidth, newHeight, {
        resample: this.resample
      });
      image = await image.center_crop(shortest_edge, shortest_edge);
    } else {
      image = await image.resize(shortest_edge, shortest_edge, {
        resample: this.resample
      });
    }
    return image;
  }
};
var ConvNextFeatureExtractor = class extends ConvNextImageProcessor {
};

// src/models/deit/image_processing_deit.js
var DeiTImageProcessor = class extends ImageProcessor {
};
var DeiTFeatureExtractor = class extends DeiTImageProcessor {
};

// src/models/detr/image_processing_detr.js
var DetrImageProcessor = class extends ImageProcessor {
  /**
   * Calls the feature extraction process on an array of images, preprocesses
   * each image, and concatenates the resulting features into a single Tensor.
   * @param {import('../../utils/image.js').RawImage[]} images The image(s) to extract features from.
   * @returns {Promise<DetrFeatureExtractorResult>} An object containing the concatenated pixel values of the preprocessed images.
   */
  async _call(images) {
    const result = await super._call(images);
    const maskSize = [result.pixel_values.dims[0], 64, 64];
    const pixel_mask = full(maskSize, 1n);
    return { ...result, pixel_mask };
  }
  /** @type {typeof post_process_object_detection} */
  post_process_object_detection(...args) {
    return post_process_object_detection(...args);
  }
  /** @type {typeof post_process_panoptic_segmentation} */
  post_process_panoptic_segmentation(...args) {
    return post_process_panoptic_segmentation(...args);
  }
  /** @type {typeof post_process_instance_segmentation} */
  post_process_instance_segmentation(...args) {
    return post_process_instance_segmentation(...args);
  }
};
var DetrFeatureExtractor = class extends DetrImageProcessor {
};

// src/models/dinov3_vit/image_processing_dinov3_vit.js
var DINOv3ViTImageProcessor = class extends ImageProcessor {
};

// src/models/donut/image_processing_donut.js
var DonutImageProcessor = class extends ImageProcessor {
  pad_image(pixelData, imgDims, padSize, options = {}) {
    const [imageHeight, imageWidth, imageChannels] = imgDims;
    let image_mean = this.image_mean;
    if (!Array.isArray(this.image_mean)) {
      image_mean = new Array(imageChannels).fill(image_mean);
    }
    let image_std = this.image_std;
    if (!Array.isArray(image_std)) {
      image_std = new Array(imageChannels).fill(image_mean);
    }
    const constant_values = image_mean.map((x, i) => -x / image_std[i]);
    return super.pad_image(pixelData, imgDims, padSize, {
      center: true,
      // Since normalization is done after padding, we need to use certain constant values to ensure the same behaviour is observed.
      // For more information, see https://github.com/huggingface/transformers/blob/main/src/transformers/models/donut/image_processing_donut.py#L433-L451
      constant_values,
      ...options
    });
  }
};
var DonutFeatureExtractor = class extends DonutImageProcessor {
};

// src/models/dpt/image_processing_dpt.js
var DPTImageProcessor = class extends ImageProcessor {
};
var DPTFeatureExtractor = class extends DPTImageProcessor {
};

// src/models/efficientnet/image_processing_efficientnet.js
var EfficientNetImageProcessor = class extends ImageProcessor {
  constructor(config) {
    super(config);
    this.include_top = this.config.include_top ?? true;
    if (this.include_top) {
      this.image_std = this.image_std.map((x) => x * x);
    }
  }
};

// src/models/gemma3/image_processing_gemma3.js
var Gemma3ImageProcessor = class extends ImageProcessor {
};

// src/models/gemma4/image_processing_gemma4.js
function get_aspect_ratio_preserving_size(height, width, patch_size, max_patches, pooling_kernel_size) {
  const target_px = max_patches * patch_size ** 2;
  const factor = Math.sqrt(target_px / (height * width));
  const side_mult = pooling_kernel_size * patch_size;
  let target_height = Math.floor(factor * height / side_mult) * side_mult;
  let target_width = Math.floor(factor * width / side_mult) * side_mult;
  if (target_height === 0 && target_width === 0) {
    throw new Error(
      `Attempting to resize to a 0 x 0 image. Resized height should be divisible by \`pooling_kernel_size * patch_size\`=${side_mult}.`
    );
  }
  const max_side_length = Math.floor(max_patches / pooling_kernel_size ** 2) * side_mult;
  if (target_height === 0) {
    target_height = side_mult;
    target_width = Math.min(Math.floor(width / height) * side_mult, max_side_length);
  } else if (target_width === 0) {
    target_width = side_mult;
    target_height = Math.min(Math.floor(height / width) * side_mult, max_side_length);
  }
  return [target_height, target_width];
}
function patchify(hwc_data, H, W, C, patch_size, max_patches, pooling_kernel_size) {
  const num_patches_h = Math.floor(H / patch_size);
  const num_patches_w = Math.floor(W / patch_size);
  const num_patches = num_patches_h * num_patches_w;
  const patch_dim = patch_size * patch_size * C;
  const patch_data = new Float32Array(max_patches * patch_dim);
  let out = 0;
  for (let ph = 0; ph < num_patches_h; ++ph) {
    for (let pw = 0; pw < num_patches_w; ++pw) {
      for (let dy = 0; dy < patch_size; ++dy) {
        const row_offset = (ph * patch_size + dy) * W * C + pw * patch_size * C;
        for (let dx = 0; dx < patch_size; ++dx) {
          const src = row_offset + dx * C;
          for (let c = 0; c < C; ++c) {
            patch_data[out++] = hwc_data[src + c];
          }
        }
      }
    }
  }
  const pos_data = new BigInt64Array(max_patches * 2).fill(-1n);
  let idx = 0;
  for (let row = 0; row < num_patches_h; ++row) {
    for (let col = 0; col < num_patches_w; ++col) {
      pos_data[idx++] = BigInt(col);
      pos_data[idx++] = BigInt(row);
    }
  }
  return {
    patches: new Tensor2("float32", patch_data, [max_patches, patch_dim]),
    positions: new Tensor2("int64", pos_data, [max_patches, 2]),
    num_soft_tokens: Math.floor(num_patches / pooling_kernel_size ** 2)
  };
}
var Gemma4ImageProcessor = class extends Callable {
  /** @param {Record<string, any>} config */
  constructor(config) {
    super();
    this.config = config;
    this.patch_size = config.patch_size ?? 16;
    this.max_soft_tokens = config.max_soft_tokens ?? 280;
    this.pooling_kernel_size = config.pooling_kernel_size ?? 3;
    this.resample = config.resample ?? 3;
    this.rescale_factor = config.rescale_factor ?? 1 / 255;
    this.do_rescale = config.do_rescale ?? true;
    this.do_resize = config.do_resize ?? true;
    this.do_convert_rgb = config.do_convert_rgb ?? true;
  }
  /**
   * @param {RawImage|RawImage[]} images
   * @returns {Promise<{ pixel_values: Tensor, image_position_ids: Tensor, num_soft_tokens_per_image: number[] }>}
   */
  async _call(images) {
    if (!Array.isArray(images)) {
      images = [images];
    }
    const { patch_size, pooling_kernel_size } = this;
    const max_patches = this.max_soft_tokens * pooling_kernel_size ** 2;
    const all_patches = [];
    const all_positions = [];
    const num_soft_tokens_per_image = [];
    for (let image of images) {
      if (this.do_convert_rgb) {
        image = image.rgb();
      }
      if (this.do_resize) {
        const [target_h, target_w] = get_aspect_ratio_preserving_size(
          image.height,
          image.width,
          patch_size,
          max_patches,
          pooling_kernel_size
        );
        if (target_h !== image.height || target_w !== image.width) {
          image = await image.resize(target_w, target_h, { resample: this.resample });
        }
      }
      const pixelData = Float32Array.from(image.data);
      if (this.do_rescale) {
        for (let i = 0; i < pixelData.length; ++i) {
          pixelData[i] *= this.rescale_factor;
        }
      }
      const { patches, positions, num_soft_tokens } = patchify(
        pixelData,
        image.height,
        image.width,
        image.channels,
        patch_size,
        max_patches,
        pooling_kernel_size
      );
      all_patches.push(patches);
      all_positions.push(positions);
      num_soft_tokens_per_image.push(num_soft_tokens);
    }
    return {
      pixel_values: stack(all_patches, 0),
      image_position_ids: stack(all_positions, 0),
      num_soft_tokens_per_image
    };
  }
};

// src/models/qwen2_vl/image_processing_qwen2_vl.js
var Qwen2VLImageProcessor = class extends ImageProcessor {
  constructor(config) {
    super(config);
    this.min_pixels = config.min_pixels ?? config.size?.shortest_edge;
    this.max_pixels = config.max_pixels ?? config.size?.longest_edge;
    this.patch_size = config.patch_size;
    this.merge_size = config.merge_size;
  }
  /** @type {ImageProcessor['get_resize_output_image_size']} */
  get_resize_output_image_size(image, size) {
    const factor = this.patch_size * this.merge_size;
    return smart_resize(image.height, image.width, factor, this.min_pixels, this.max_pixels);
  }
  async _call(images, ...args) {
    const { pixel_values, original_sizes, reshaped_input_sizes } = await super._call(images, ...args);
    let patches = pixel_values;
    const { temporal_patch_size, merge_size, patch_size } = this.config;
    if (patches.dims[0] === 1) {
      patches = cat(
        Array.from({ length: temporal_patch_size }, () => patches),
        0
      );
    }
    const grid_t = patches.dims[0] / temporal_patch_size;
    const channel = patches.dims[1];
    const grid_h = Math.floor(patches.dims[2] / patch_size);
    const grid_w = Math.floor(patches.dims[3] / patch_size);
    const flatten_patches = patches.view(
      grid_t,
      temporal_patch_size,
      channel,
      Math.floor(grid_h / merge_size),
      merge_size,
      patch_size,
      Math.floor(grid_w / merge_size),
      merge_size,
      patch_size
    ).permute(0, 3, 6, 4, 7, 2, 1, 5, 8).view(grid_t * grid_h * grid_w, channel * temporal_patch_size * patch_size * patch_size);
    const image_grid_thw = new Tensor2("int64", [grid_t, grid_h, grid_w], [1, 3]);
    return {
      pixel_values: flatten_patches,
      image_grid_thw,
      original_sizes,
      reshaped_input_sizes
    };
  }
};

// src/models/glm46v/image_processing_glm46v.js
var Glm46VImageProcessor = class extends Qwen2VLImageProcessor {
  /** @type {Qwen2VLImageProcessor['get_resize_output_image_size']} */
  get_resize_output_image_size(image, size) {
    const factor = this.patch_size * this.merge_size;
    const temporal_factor = this.config.temporal_patch_size ?? 2;
    return smart_resize(image.height, image.width, factor, this.min_pixels, this.max_pixels, temporal_factor);
  }
};

// src/models/glpn/image_processing_glpn.js
var GLPNFeatureExtractor = class extends ImageProcessor {
};

// src/models/grounding_dino/image_processing_grounding_dino.js
var GroundingDinoImageProcessor = class extends ImageProcessor {
  /**
   * Calls the feature extraction process on an array of images, preprocesses
   * each image, and concatenates the resulting features into a single Tensor.
   * @param {import('../../utils/image.js').RawImage[]} images The image(s) to extract features from.
   * @returns {Promise<GroundingDinoFeatureExtractorResult>} An object containing the concatenated pixel values of the preprocessed images.
   */
  async _call(images) {
    const result = await super._call(images);
    const dims = result.pixel_values.dims;
    const pixel_mask = ones([dims[0], dims[2], dims[3]]);
    return { ...result, pixel_mask };
  }
};

// src/models/idefics3/image_processing_idefics3.js
var Idefics3ImageProcessor = class extends ImageProcessor {
  constructor(config) {
    super(config);
    this.do_image_splitting = config.do_image_splitting ?? true;
    this.max_image_size = config.max_image_size;
  }
  /**
   * @typedef {import('../../utils/image.js').RawImage} RawImage
   * @typedef {import('../../utils/tensor.js').Tensor} Tensor
   */
  /**
   * Calculate size to resize images to, to be multiples of `vision_encoder_max_size` while preserving the aspect ratio.
   * @param {Tensor} pixel_values Tensor of the image to resize.
   * @param {number} vision_encoder_max_size Maximum size of the output image. If the image is larger than this size,
   * it will be split into patches of this size, and the original image will be concatenated with the patches, resized to max_size.
   */
  get_resize_for_vision_encoder(pixel_values, vision_encoder_max_size) {
    let [height, width] = pixel_values.dims.slice(-2);
    const aspect_ratio = width / height;
    if (width >= height) {
      width = Math.ceil(width / vision_encoder_max_size) * vision_encoder_max_size;
      height = Math.floor(width / aspect_ratio);
      height = Math.ceil(height / vision_encoder_max_size) * vision_encoder_max_size;
    } else {
      height = Math.ceil(height / vision_encoder_max_size) * vision_encoder_max_size;
      width = Math.floor(height * aspect_ratio);
      width = Math.ceil(width / vision_encoder_max_size) * vision_encoder_max_size;
    }
    return { height, width };
  }
  /** @param {RawImage|RawImage[]|RawImage[][]} images */
  async _call(images, { do_image_splitting = null, return_row_col_info = false } = {}) {
    let batched_2d_images;
    if (!Array.isArray(images)) {
      batched_2d_images = [[images]];
    } else {
      if (images.length === 0 || !images[0]) {
        throw new Error("No images provided.");
      }
      if (!Array.isArray(images[0])) {
        batched_2d_images = [
          /** @type {RawImage[]} */
          images
        ];
      } else {
        batched_2d_images = /** @type {RawImage[][]} */
        images;
      }
    }
    let all_pixel_values = [];
    let images_list_rows = [];
    let images_list_cols = [];
    const original_sizes = [];
    const reshaped_input_sizes = [];
    for (const image_batch of batched_2d_images) {
      let images_list = await Promise.all(image_batch.map((x) => this.preprocess(x)));
      original_sizes.push(...images_list.map((x) => x.original_size));
      reshaped_input_sizes.push(...images_list.map((x) => x.reshaped_input_size));
      images_list.forEach((x) => x.pixel_values.unsqueeze_(0));
      const { longest_edge } = this.max_image_size;
      let images_tensor;
      if (do_image_splitting ?? this.do_image_splitting) {
        let image_rows = new Array(images_list.length);
        let image_cols = new Array(images_list.length);
        images_tensor = await Promise.all(
          images_list.map(async (x, i) => {
            const new_size = this.get_resize_for_vision_encoder(x.pixel_values, longest_edge);
            const resized = await interpolate_4d(x.pixel_values, {
              size: [new_size.height, new_size.width]
            });
            const { frames, num_splits_h, num_splits_w } = await this.split_image(
              resized,
              this.max_image_size
            );
            image_rows[i] = num_splits_h;
            image_cols[i] = num_splits_w;
            return cat(frames, 0);
          })
        );
        images_list_rows.push(image_rows);
        images_list_cols.push(image_cols);
      } else {
        const size = [longest_edge, longest_edge];
        images_tensor = await Promise.all(images_list.map((x) => interpolate_4d(x.pixel_values, { size })));
        images_list_rows.push(new Array(images_list.length).fill(0));
        images_list_cols.push(new Array(images_list.length).fill(0));
      }
      all_pixel_values.push(cat(images_tensor, 0));
    }
    const batch_size = all_pixel_values.length;
    const [n, c, h, w] = all_pixel_values[0].dims;
    let pixel_values;
    let pixel_attention_mask;
    if (batch_size === 1) {
      pixel_values = all_pixel_values[0].unsqueeze_(0);
      pixel_attention_mask = full([batch_size, n, h, w], true);
    } else {
      const max_num_patches = Math.max(...all_pixel_values.map((x) => x.dims.at(0)));
      pixel_attention_mask = full([batch_size, max_num_patches, h, w], true);
      const pixel_attention_mask_data = pixel_attention_mask.data;
      const pixel_attention_mask_stride = max_num_patches * h * w;
      for (let i = 0; i < batch_size; ++i) {
        const num_patches = all_pixel_values[i].dims[0];
        if (num_patches < max_num_patches) {
          all_pixel_values[i] = cat(
            [all_pixel_values[i], full([max_num_patches - num_patches, c, h, w], 0)],
            0
          );
          const start_offset = i * pixel_attention_mask_stride + num_patches * h * w;
          const end_offset = (i + 1) * pixel_attention_mask_stride;
          pixel_attention_mask_data.fill(false, start_offset, end_offset);
        }
      }
      pixel_values = stack(all_pixel_values, 0);
    }
    return {
      pixel_values,
      pixel_attention_mask,
      original_sizes,
      reshaped_input_sizes,
      ...return_row_col_info ? { rows: images_list_rows, cols: images_list_cols } : {}
    };
  }
  async split_image(pixel_values, { longest_edge }) {
    const max_height = longest_edge;
    const max_width = longest_edge;
    const frames = [];
    const [height, width] = pixel_values.dims.slice(-2);
    let num_splits_h = 0, num_splits_w = 0;
    if (height > max_height || width > max_width) {
      num_splits_h = Math.ceil(height / max_height);
      num_splits_w = Math.ceil(width / max_width);
      const optimal_height = Math.ceil(height / num_splits_h);
      const optimal_width = Math.ceil(width / num_splits_w);
      for (let r = 0; r < num_splits_h; ++r) {
        for (let c = 0; c < num_splits_w; ++c) {
          let start_x, start_y, end_x, end_y;
          if (r === num_splits_h - 1) {
            start_y = height - optimal_height;
            end_y = height;
          } else {
            start_y = r * optimal_height;
            end_y = (r + 1) * optimal_height;
          }
          if (c === num_splits_w - 1) {
            start_x = width - optimal_width;
            end_x = width;
          } else {
            start_x = c * optimal_width;
            end_x = (c + 1) * optimal_width;
          }
          const starts = [start_y, start_x];
          const ends = [end_y, end_x];
          const patch = await slice2(pixel_values, starts, ends, [2, 3]);
          frames.push(patch);
        }
      }
      const global_image_height = max_height;
      const global_image_width = max_width;
      if (height !== global_image_height || width !== global_image_width) {
        pixel_values = await interpolate_4d(pixel_values, {
          size: [global_image_height, global_image_width]
        });
      }
    }
    frames.push(pixel_values);
    return { frames, num_splits_h, num_splits_w };
  }
};

// src/models/janus/image_processing_janus.js
var VLMImageProcessor = class extends ImageProcessor {
  constructor(config) {
    super({
      do_pad: true,
      pad_size: {
        width: config.image_size,
        height: config.image_size
      },
      ...config
    });
    this.constant_values = this.config.background_color.map((x) => x * this.rescale_factor);
  }
  pad_image(pixelData, imgDims, padSize, options) {
    return super.pad_image(pixelData, imgDims, padSize, {
      constant_values: this.constant_values,
      center: true,
      ...options
    });
  }
};

// src/models/jina_clip/image_processing_jina_clip.js
var JinaCLIPImageProcessor = class extends ImageProcessor {
  constructor(config) {
    const { resize_mode, fill_color, interpolation, size, ...other } = config;
    const new_size = resize_mode === "squash" ? { width: size, height: size } : resize_mode === "shortest" ? { shortest_edge: size } : { longest_edge: size };
    const resample = interpolation === "bicubic" ? 3 : 2;
    super({
      ...other,
      size: new_size,
      resample,
      do_center_crop: true,
      crop_size: size,
      do_normalize: true
    });
  }
};

// src/models/lfm2_vl/image_processing_lfm2_vl.js
function round_by_factor(number, factor) {
  return Math.round(number / factor) * factor;
}
function find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size) {
  let best_ratio_diff = Infinity;
  let best_ratio = [1, 1];
  const area = width * height;
  for (const ratio of target_ratios) {
    const ratio_diff = Math.abs(aspect_ratio - ratio[0] / ratio[1]);
    if (ratio_diff < best_ratio_diff) {
      best_ratio_diff = ratio_diff;
      best_ratio = ratio;
    } else if (ratio_diff === best_ratio_diff && area > 0.5 * image_size * image_size * ratio[0] * ratio[1]) {
      best_ratio = ratio;
    }
  }
  return best_ratio;
}
function get_target_ratios(min_tiles, max_tiles) {
  const ratios = [];
  const seen = /* @__PURE__ */ new Set();
  for (let n = min_tiles; n <= max_tiles; ++n) {
    for (let w = 1; w <= n; ++w) {
      for (let h = 1; h <= n; ++h) {
        const product2 = w * h;
        if (product2 >= min_tiles && product2 <= max_tiles) {
          const key = w << 16 | h;
          if (!seen.has(key)) {
            seen.add(key);
            ratios.push([w, h]);
          }
        }
      }
    }
  }
  return ratios.sort((a, b) => a[0] * a[1] - b[0] * b[1]);
}
function convert_image_to_patches(images, patch_size) {
  const [B, C, H, W] = images.dims;
  const ph = Math.floor(H / patch_size), pw = Math.floor(W / patch_size);
  const patch_dim = patch_size * patch_size * C;
  const data = (
    /** @type {Float32Array} */
    images.data
  );
  const result = new Float32Array(B * ph * pw * patch_dim);
  const ch_stride = H * W;
  for (let b = 0; b < B; ++b) {
    const b_src = b * C * ch_stride;
    const b_dst = b * ph * pw * patch_dim;
    for (let py = 0; py < ph; ++py) {
      for (let px = 0; px < pw; ++px) {
        let off = b_dst + (py * pw + px) * patch_dim;
        for (let dy = 0; dy < patch_size; ++dy) {
          const row = (py * patch_size + dy) * W + px * patch_size;
          for (let dx = 0; dx < patch_size; ++dx) {
            const pixel = row + dx;
            for (let c = 0; c < C; ++c) {
              result[off++] = data[b_src + c * ch_stride + pixel];
            }
          }
        }
      }
    }
  }
  return new Tensor2("float32", result, [B, ph * pw, patch_dim]);
}
function pad_along_first_dim(patches, target_length) {
  const [, len2, dim] = patches.dims;
  const mask_data = new BigInt64Array(target_length);
  mask_data.fill(1n, 0, len2);
  let padded = patches;
  if (len2 < target_length) {
    const padded_data = new Float32Array(target_length * dim);
    padded_data.set(
      /** @type {Float32Array} */
      patches.data
    );
    padded = new Tensor2("float32", padded_data, [1, target_length, dim]);
  }
  return { padded, mask: new Tensor2("int64", mask_data, [target_length]) };
}
var Lfm2VlImageProcessor = class extends ImageProcessor {
  constructor(config) {
    super(config);
    this.downsample_factor = config.downsample_factor ?? 2;
    this.do_image_splitting = config.do_image_splitting ?? true;
    this.min_tiles = config.min_tiles ?? 2;
    this.max_tiles = config.max_tiles ?? 10;
    this.use_thumbnail = config.use_thumbnail ?? true;
    this.min_image_tokens = config.min_image_tokens ?? 64;
    this.max_image_tokens = config.max_image_tokens ?? 256;
    this.encoder_patch_size = config.encoder_patch_size ?? config.patch_size ?? 16;
    this.tile_size = config.tile_size ?? 512;
    this.max_pixels_tolerance = config.max_pixels_tolerance ?? 2;
    this.return_row_col_info = config.return_row_col_info ?? false;
    const max_thumbnail_patches = this.max_image_tokens * this.downsample_factor ** 2;
    const tile_size_patches = this.do_image_splitting ? (this.tile_size / this.encoder_patch_size) ** 2 : 0;
    this.max_num_patches = Math.max(max_thumbnail_patches, tile_size_patches);
  }
  /**
   * Check if the image is too large to be processed as a single tile.
   * @param {number} height
   * @param {number} width
   * @returns {boolean}
   */
  _is_image_too_large(height, width) {
    const total_factor = this.encoder_patch_size * this.downsample_factor;
    const h_bar = Math.max(this.encoder_patch_size, round_by_factor(height, total_factor));
    const w_bar = Math.max(this.encoder_patch_size, round_by_factor(width, total_factor));
    return h_bar * w_bar > this.max_image_tokens * (this.encoder_patch_size * this.downsample_factor) ** 2 * this.max_pixels_tolerance;
  }
  /**
   * Get the grid layout for tiling a large image.
   * @param {number} height
   * @param {number} width
   * @returns {{ grid_width: number, grid_height: number, target_width: number, target_height: number }}
   */
  _get_grid_layout(height, width) {
    const target_ratios = get_target_ratios(this.min_tiles, this.max_tiles);
    const [grid_width, grid_height] = find_closest_aspect_ratio(
      width / height,
      target_ratios,
      width,
      height,
      this.tile_size
    );
    return {
      grid_width,
      grid_height,
      target_width: this.tile_size * grid_width,
      target_height: this.tile_size * grid_height
    };
  }
  /** @param {RawImage|RawImage[]|RawImage[][]} images */
  // @ts-expect-error
  async _call(images, { return_row_col_info = null } = {}) {
    let batched_images;
    if (!Array.isArray(images)) {
      batched_images = [[images]];
    } else if (!Array.isArray(images[0])) {
      batched_images = [
        /** @type {RawImage[]} */
        images
      ];
    } else {
      batched_images = /** @type {RawImage[][]} */
      images;
    }
    const all_pixel_values = [];
    const all_pixel_masks = [];
    const all_spatial_shapes = [];
    const all_rows = [];
    const all_cols = [];
    const all_image_sizes = [];
    for (const image_batch of batched_images) {
      const preprocessed = await Promise.all(image_batch.map((x) => this.preprocess(x, { do_pad: false })));
      for (const { pixel_values } of preprocessed) {
        const [, height, width] = pixel_values.dims;
        const img = pixel_values.unsqueeze_(0);
        const total_factor = this.encoder_patch_size * this.downsample_factor;
        const f2 = total_factor ** 2;
        const [new_width, new_height] = smart_resize(
          Math.max(total_factor, height),
          Math.max(total_factor, width),
          total_factor,
          this.min_image_tokens * f2,
          this.max_image_tokens * f2
        ).map((x) => Math.max(total_factor, x));
        let tiles;
        let num_rows = 1, num_cols = 1;
        const is_large = this._is_image_too_large(height, width);
        const do_splitting = this.do_image_splitting && !(this.min_tiles === 1 && this.max_tiles === 1);
        if (is_large && do_splitting) {
          const { grid_width, grid_height, target_width, target_height } = this._get_grid_layout(
            height,
            width
          );
          num_rows = grid_height;
          num_cols = grid_width;
          const resized = await interpolate_4d(img, {
            size: [target_height, target_width]
          });
          tiles = [];
          for (let r = 0; r < grid_height; ++r) {
            for (let c = 0; c < grid_width; ++c) {
              const y = r * this.tile_size;
              const x = c * this.tile_size;
              tiles.push(resized.slice(null, null, [y, y + this.tile_size], [x, x + this.tile_size]));
            }
          }
          if (this.use_thumbnail && grid_width * grid_height !== 1) {
            tiles.push(await interpolate_4d(img, { size: [new_height, new_width] }));
          }
        } else {
          tiles = [await interpolate_4d(img, { size: [new_height, new_width] })];
        }
        for (const tile of tiles) {
          const [, , th, tw] = tile.dims;
          const patches = convert_image_to_patches(tile, this.encoder_patch_size);
          const { padded, mask } = pad_along_first_dim(patches, this.max_num_patches);
          all_pixel_values.push(padded);
          all_pixel_masks.push(mask);
          all_spatial_shapes.push([
            Math.floor(th / this.encoder_patch_size),
            Math.floor(tw / this.encoder_patch_size)
          ]);
        }
        all_rows.push(num_rows);
        all_cols.push(num_cols);
        all_image_sizes.push([new_height, new_width]);
      }
    }
    const result = {
      pixel_values: cat(all_pixel_values, 0),
      pixel_attention_mask: stack(all_pixel_masks, 0),
      spatial_shapes: new Tensor2("int64", BigInt64Array.from(all_spatial_shapes.flat(), BigInt), [
        all_spatial_shapes.length,
        2
      ])
    };
    if (return_row_col_info ?? this.return_row_col_info) {
      result.image_rows = all_rows;
      result.image_cols = all_cols;
      result.image_sizes = all_image_sizes;
    }
    return result;
  }
};

// src/models/llava_onevision/image_processing_llava_onevision.js
var LlavaOnevisionImageProcessor = class extends ImageProcessor {
};

// src/models/maskformer/image_processing_maskformer.js
var MaskFormerImageProcessor = class extends ImageProcessor {
  /** @type {typeof post_process_panoptic_segmentation} */
  post_process_panoptic_segmentation(...args) {
    return post_process_panoptic_segmentation(...args);
  }
  /** @type {typeof post_process_instance_segmentation} */
  post_process_instance_segmentation(...args) {
    return post_process_instance_segmentation(...args);
  }
};
var MaskFormerFeatureExtractor = class extends MaskFormerImageProcessor {
};

// src/models/mask2former/image_processing_mask2former.js
var Mask2FormerImageProcessor = class extends MaskFormerImageProcessor {
};

// src/models/mobilenet_v1/image_processing_mobilenet_v1.js
var MobileNetV1ImageProcessor = class extends ImageProcessor {
};
var MobileNetV1FeatureExtractor = class extends MobileNetV1ImageProcessor {
};

// src/models/mobilenet_v2/image_processing_mobilenet_v2.js
var MobileNetV2ImageProcessor = class extends ImageProcessor {
};
var MobileNetV2FeatureExtractor = class extends MobileNetV2ImageProcessor {
};

// src/models/mobilenet_v3/image_processing_mobilenet_v3.js
var MobileNetV3ImageProcessor = class extends ImageProcessor {
};
var MobileNetV3FeatureExtractor = class extends MobileNetV3ImageProcessor {
};

// src/models/mobilenet_v4/image_processing_mobilenet_v4.js
var MobileNetV4ImageProcessor = class extends ImageProcessor {
};
var MobileNetV4FeatureExtractor = class extends MobileNetV4ImageProcessor {
};

// src/models/mobilevit/image_processing_mobilevit.js
var MobileViTImageProcessor = class extends ImageProcessor {
};
var MobileViTFeatureExtractor = class extends MobileViTImageProcessor {
};

// src/models/nougat/image_processing_nougat.js
var NougatImageProcessor = class extends DonutImageProcessor {
};

// src/models/owlvit/image_processing_owlvit.js
var OwlViTImageProcessor = class extends ImageProcessor {
  /** @type {typeof post_process_object_detection} */
  post_process_object_detection(...args) {
    return post_process_object_detection(...args);
  }
};
var OwlViTFeatureExtractor = class extends OwlViTImageProcessor {
};

// src/models/owlv2/image_processing_owlv2.js
var Owlv2ImageProcessor = class extends OwlViTImageProcessor {
};

// src/models/phi3_v/image_processing_phi3_v.js
var IMAGE_SIZE = 336;
var SLICE_AXES = [2, 3];
var { ceil, floor, sqrt } = Math;
var Phi3VImageProcessor = class extends ImageProcessor {
  constructor(config) {
    super({
      ...config,
      do_normalize: true,
      do_pad: true,
      pad_size: "custom",
      do_convert_rgb: true,
      do_resize: true
      // Smart resizing "hd_transform"
    });
    this._num_crops = config.num_crops;
  }
  calc_num_image_tokens_from_image_size(width, height) {
    const { num_img_tokens } = this.config;
    return floor(
      (floor(height / IMAGE_SIZE) * floor(width / IMAGE_SIZE) + 1) * num_img_tokens + 1 + (floor(height / IMAGE_SIZE) + 1) * sqrt(num_img_tokens)
    );
  }
  /** @type {ImageProcessor['get_resize_output_image_size']} */
  get_resize_output_image_size(image, size) {
    const hd_num = this._num_crops;
    const [width, height] = image.size;
    let ratio = width / height;
    let scale = 1;
    while (scale * Math.ceil(scale / ratio) <= hd_num) {
      scale += 1;
    }
    scale -= 1;
    const new_w = Math.floor(scale * 336);
    const new_h = Math.floor(new_w / ratio);
    return [new_w, new_h];
  }
  /** @type {ImageProcessor['pad_image']} */
  pad_image(pixelData, imgDims, padSize, options = {}) {
    const [imageHeight, imageWidth] = imgDims;
    const height = IMAGE_SIZE * ceil(imageHeight / IMAGE_SIZE);
    const width = IMAGE_SIZE * ceil(imageWidth / IMAGE_SIZE);
    const constant_values = [1, 1, 1].map((x, i) => (x - this.image_mean[i]) / this.image_std[i]);
    return super.pad_image(
      pixelData,
      imgDims,
      { width, height },
      {
        center: true,
        constant_values,
        ...options
      }
    );
  }
  async _call(images, { num_crops = null } = {}) {
    this._num_crops = num_crops ??= this.config.num_crops;
    if (num_crops < 4 || sqrt(num_crops) % 1 !== 0) {
      throw new Error("num_crops must be a square number >= 4");
    }
    if (!Array.isArray(images)) {
      images = [images];
    }
    const num_images = images.length;
    const imageData = await Promise.all(images.map((x) => this.preprocess(x)));
    const original_sizes = imageData.map((x) => x.original_size);
    const reshaped_input_sizes = imageData.map((x) => x.reshaped_input_size);
    const all_pixel_values = [];
    for (const { pixel_values: pixel_values2 } of imageData) {
      pixel_values2.unsqueeze_(0);
      const [height, width] = pixel_values2.dims.slice(-2);
      const batch_pixel_values = await interpolate_4d(pixel_values2, {
        size: [IMAGE_SIZE, IMAGE_SIZE],
        mode: "bicubic"
      });
      if (num_crops > 0) {
        const patches = [];
        const sqrt_patches = sqrt(num_crops);
        const patch_width = floor(width / sqrt_patches);
        const patch_height = floor(height / sqrt_patches);
        for (let y = 0; y < sqrt_patches; ++y) {
          for (let x = 0; x < sqrt_patches; ++x) {
            let start_x, start_y, end_x, end_y;
            if (y === sqrt_patches - 1) {
              start_y = height - patch_height;
              end_y = height;
            } else {
              start_y = y * patch_height;
              end_y = (y + 1) * patch_height;
            }
            if (x === sqrt_patches - 1) {
              start_x = width - patch_width;
              end_x = width;
            } else {
              start_x = x * patch_width;
              end_x = (x + 1) * patch_width;
            }
            const starts = [start_y, start_x];
            const ends = [end_y, end_x];
            const patch = await slice2(pixel_values2, starts, ends, SLICE_AXES);
            patches.push(patch);
          }
        }
        const resized_tensors = await interpolate_4d(cat(patches, 0), {
          size: [IMAGE_SIZE, IMAGE_SIZE],
          mode: "bicubic"
        });
        all_pixel_values.push(cat([batch_pixel_values, resized_tensors], 0));
      } else {
        all_pixel_values.push(batch_pixel_values);
      }
    }
    const pixel_values = stack(all_pixel_values, 0);
    const sizes = reshaped_input_sizes.map((x) => x.map((y) => IMAGE_SIZE * ceil(y / IMAGE_SIZE)));
    const image_sizes = new Tensor2("int64", sizes.flat(), [num_images, 2]);
    const num_img_tokens = sizes.map(
      ([height, width]) => this.calc_num_image_tokens_from_image_size(width, height)
    );
    return { pixel_values, original_sizes, reshaped_input_sizes, image_sizes, num_img_tokens };
  }
};

// src/models/pixtral/image_processing_pixtral.js
var PixtralImageProcessor = class extends ImageProcessor {
  /** @type {ImageProcessor['get_resize_output_image_size']} */
  get_resize_output_image_size(image, size) {
    const { longest_edge } = size;
    if (longest_edge === void 0) {
      throw new Error("size must contain 'longest_edge'");
    }
    const [srcWidth, srcHeight] = image.size;
    const ratio = Math.max(srcWidth, srcHeight) / longest_edge;
    let newWidth = srcWidth;
    let newHeight = srcHeight;
    if (ratio > 1) {
      newWidth = Math.floor(srcWidth / ratio);
      newHeight = Math.floor(srcHeight / ratio);
    }
    const { patch_size, spatial_merge_size } = this.config;
    if (!spatial_merge_size) {
      throw new Error("config must contain 'spatial_merge_size'");
    }
    const real_patch_size = patch_size * spatial_merge_size;
    const num_width_tokens = Math.floor((newWidth - 1) / real_patch_size) + 1;
    const num_height_tokens = Math.floor((newHeight - 1) / real_patch_size) + 1;
    return [num_width_tokens * real_patch_size, num_height_tokens * real_patch_size];
  }
};

// src/models/pvt/image_processing_pvt.js
var PvtImageProcessor = class extends ImageProcessor {
};

// src/models/rt_detr/image_processing_rt_detr.js
var RTDetrImageProcessor = class extends ImageProcessor {
  /** @type {typeof post_process_object_detection} */
  post_process_object_detection(...args) {
    return post_process_object_detection(...args);
  }
};

// src/models/sam/image_processing_sam.js
var SamImageProcessor = class extends ImageProcessor {
  /**
   *
   * @param {any} input_points
   * @param {import("../../image_processors_utils.js").HeightWidth[]} original_sizes
   * @param {import("../../image_processors_utils.js").HeightWidth[]} reshaped_input_sizes
   * @returns {Tensor}
   */
  reshape_input_points(input_points, original_sizes, reshaped_input_sizes, is_bounding_box = false) {
    input_points = structuredClone(input_points);
    let shape = calculateDimensions(input_points);
    if (shape.length === 3) {
      if (!is_bounding_box) {
        shape = [1, ...shape];
      }
      input_points = [input_points];
    } else if (shape.length !== 4) {
      throw Error(
        "The input_points must be a 4D tensor of shape `batch_size`, `point_batch_size`, `nb_points_per_image`, `2`."
      );
    }
    for (let i = 0; i < input_points.length; ++i) {
      const [originalHeight, originalWidth] = original_sizes[i];
      const [reshapedHeight, reshapedWidth] = reshaped_input_sizes[i];
      const resizeFactors = [reshapedWidth / originalWidth, reshapedHeight / originalHeight];
      for (let j = 0; j < input_points[i].length; ++j) {
        for (let k2 = 0; k2 < input_points[i][j].length; ++k2) {
          for (let w = 0; w < input_points[i][j][k2].length; ++w) {
            input_points[i][j][k2][w] *= resizeFactors[w % 2];
          }
        }
      }
    }
    return new Tensor2("float32", Float32Array.from(input_points.flat(Infinity)), shape);
  }
  /**
   *
   * @param {any} input_labels
   * @param {Tensor} input_points
   * @returns {Tensor}
   */
  add_input_labels(input_labels, input_points) {
    let shape = calculateDimensions(input_labels);
    if (shape.length === 2) {
      shape = [1, ...shape];
      input_labels = [input_labels];
    } else if (shape.length !== 3) {
      throw Error(
        "The input_points must be a 4D tensor of shape `batch_size`, `point_batch_size`, `nb_points_per_image`, `2`."
      );
    }
    if (shape.some((x, i) => x !== input_points.dims[i])) {
      throw Error(`The first ${shape.length} dimensions of 'input_points' and 'input_labels' must be the same.`);
    }
    return new Tensor2("int64", input_labels.flat(Infinity).map(BigInt), shape);
  }
  /**
   * @param {any[]} images The URL(s) of the image(s) to extract features from.
   * @param {Object} [options] Additional options for the processor.
   * @param {any} [options.input_points=null] A 3D or 4D array, representing the input points provided by the user.
   * - 3D: `[point_batch_size, nb_points_per_image, 2]`. In this case, `batch_size` is assumed to be 1.
   * - 4D: `[batch_size, point_batch_size, nb_points_per_image, 2]`.
   * @param {any} [options.input_labels=null] A 2D or 3D array, representing the input labels for the points, used by the prompt encoder to encode the prompt.
   * - 2D: `[point_batch_size, nb_points_per_image]`. In this case, `batch_size` is assumed to be 1.
   * - 3D: `[batch_size, point_batch_size, nb_points_per_image]`.
   * @param {number[][][]} [options.input_boxes=null] A 3D array of shape `(batch_size, num_boxes, 4)`, representing the input boxes provided by the user.
   * This is used by the prompt encoder to encode the prompt. Generally yields to much better generated masks.
   * The processor will generate a tensor, with each dimension corresponding respectively to the image batch size,
   * the number of boxes per image and the coordinates of the top left and botton right point of the box.
   * In the order (`x1`, `y1`, `x2`, `y2`):
   * - `x1`: the x coordinate of the top left point of the input box
   * - `y1`: the y coordinate of the top left point of the input box
   * - `x2`: the x coordinate of the bottom right point of the input box
   * - `y2`: the y coordinate of the bottom right point of the input box
   * @returns {Promise<SamImageProcessorResult>}
   */
  async _call(images, { input_points = null, input_labels = null, input_boxes = null } = {}) {
    const processed = await super._call(images);
    if (input_points) {
      processed.input_points = this.reshape_input_points(
        input_points,
        processed.original_sizes,
        processed.reshaped_input_sizes
      );
    }
    if (input_labels) {
      if (!processed.input_points) {
        throw Error("`input_points` must be provided if `input_labels` are provided.");
      }
      processed.input_labels = this.add_input_labels(input_labels, processed.input_points);
    }
    if (input_boxes) {
      processed.input_boxes = this.reshape_input_points(
        input_boxes,
        processed.original_sizes,
        processed.reshaped_input_sizes,
        true
      );
    }
    return processed;
  }
  /**
   * Remove padding and upscale masks to the original image size.
   * @param {Tensor} masks Batched masks from the mask_decoder in (batch_size, num_channels, height, width) format.
   * @param {[number, number][]} original_sizes The original sizes of each image before it was resized to the model's expected input shape, in (height, width) format.
   * @param {[number, number][]} reshaped_input_sizes The size of each image as it is fed to the model, in (height, width) format. Used to remove padding.
   * @param {Object} options Optional parameters for post-processing.
   * @param {number} [options.mask_threshold] The threshold to use for binarizing the masks.
   * @param {boolean} [options.binarize] Whether to binarize the masks.
   * @param {Object} [options.pad_size] The target size the images were padded to before being passed to the model. If `null`, the target size is assumed to be the processor's `pad_size`.
   * @param {number} [options.pad_size.height] The height the images were padded to.
   * @param {number} [options.pad_size.width] The width the images were padded to.
   * @returns {Promise<Tensor[]>} Batched masks in batch_size, num_channels, height, width) format, where (height, width) is given by original_size.
   */
  async post_process_masks(masks, original_sizes, reshaped_input_sizes, { mask_threshold = 0, binarize = true, pad_size = null } = {}) {
    const output_masks = [];
    pad_size = pad_size ?? this.pad_size ?? this.size;
    const target_image_size = [pad_size.height, pad_size.width];
    for (let i = 0; i < original_sizes.length; ++i) {
      const original_size = original_sizes[i];
      const reshaped_input_size = reshaped_input_sizes[i];
      let interpolated_mask = await interpolate_4d(masks[i], { mode: "bilinear", size: target_image_size });
      interpolated_mask = interpolated_mask.slice(
        null,
        null,
        [0, reshaped_input_size[0]],
        [0, reshaped_input_size[1]]
      );
      interpolated_mask = await interpolate_4d(interpolated_mask, { mode: "bilinear", size: original_size });
      if (binarize) {
        const data = interpolated_mask.data;
        const binarizedMaskData = new Uint8Array(data.length);
        for (let i2 = 0; i2 < data.length; ++i2) {
          if (data[i2] > mask_threshold) {
            binarizedMaskData[i2] = 1;
          }
        }
        interpolated_mask = new Tensor2("bool", binarizedMaskData, interpolated_mask.dims);
      }
      output_masks.push(interpolated_mask);
    }
    return output_masks;
  }
  /**
   * Generates a list of crop boxes of different sizes. Each layer has (2**i)**2 boxes for the ith layer.
   * @param {import("../../utils/image.js").RawImage} image Input original image
   * @param {number} target_size Target size of the resized image
   * @param {Object} options Options for generating crop boxes
   * @param {number} [options.crop_n_layers] If >0, mask prediction will be run again on crops of the image.
   * Sets the number of layers to run, where each layer has 2**i_layer number of image crops.
   * @param {number} [options.overlap_ratio] Sets the degree to which crops overlap. In the first crop layer,
   * crops will overlap by this fraction of the image length. Later layers with more crops scale down this overlap.
   * @param {number} [options.points_per_crop] Number of points to sample from each crop.
   * @param {number} [options.crop_n_points_downscale_factor] The number of points-per-side sampled in layer n is
   * scaled down by crop_n_points_downscale_factor**n.
   * @returns {Object} An object containing the crop boxes, number of points per crop, cropped images, and input labels.
   */
  generate_crop_boxes(image, target_size, {
    crop_n_layers = 0,
    overlap_ratio = 512 / 1500,
    points_per_crop = 32,
    crop_n_points_downscale_factor = 1
  } = {}) {
  }
};

// src/models/sapiens/image_processing_sapiens.js
var SapiensImageProcessor = class extends ImageProcessor {
  /** @type {typeof post_process_semantic_segmentation} */
  post_process_semantic_segmentation(...args) {
    return post_process_semantic_segmentation(...args);
  }
};
var SapiensFeatureExtractor = class extends SapiensImageProcessor {
};

// src/models/segformer/image_processing_segformer.js
var SegformerImageProcessor = class extends ImageProcessor {
  /** @type {typeof post_process_semantic_segmentation} */
  post_process_semantic_segmentation(...args) {
    return post_process_semantic_segmentation(...args);
  }
};
var SegformerFeatureExtractor = class extends SegformerImageProcessor {
};

// src/models/siglip/image_processing_siglip.js
var SiglipImageProcessor = class extends ImageProcessor {
};

// src/models/swin2sr/image_processing_swin2sr.js
var Swin2SRImageProcessor = class extends ImageProcessor {
  pad_image(pixelData, imgDims, padSize, options = {}) {
    const [imageHeight, imageWidth, imageChannels] = imgDims;
    return super.pad_image(
      pixelData,
      imgDims,
      {
        // NOTE: For Swin2SR models, the original python implementation adds padding even when the image's width/height is already
        // a multiple of `pad_size`. However, this is most likely a bug (PR: https://github.com/mv-lab/swin2sr/pull/19).
        // For this reason, we only add padding when the image's width/height is not a multiple of `pad_size`.
        width: imageWidth + (padSize - imageWidth % padSize) % padSize,
        height: imageHeight + (padSize - imageHeight % padSize) % padSize
      },
      {
        mode: "symmetric",
        center: false,
        constant_values: -1,
        ...options
      }
    );
  }
};

// src/models/vit/image_processing_vit.js
var ViTImageProcessor = class extends ImageProcessor {
};
var ViTFeatureExtractor = class extends ViTImageProcessor {
};

// src/models/vitmatte/image_processing_vitmatte.js
var VitMatteImageProcessor = class extends ImageProcessor {
  /**
   * Calls the feature extraction process on an array of images, preprocesses
   * each image, and concatenates the resulting features into a single Tensor.
   * @param {import("../../utils/image.js").RawImage[]} images The image(s) to extract features from.
   * @param {import("../../utils/image.js").RawImage[]} trimaps The trimaps(s) to extract features from.
   * @returns {Promise<import("../../image_processors_utils.js").ImageProcessorResult>} An object containing the concatenated pixel values of the preprocessed images.
   */
  async _call(images, trimaps) {
    if (!Array.isArray(images)) {
      images = [images];
    }
    if (!Array.isArray(trimaps)) {
      trimaps = [trimaps];
    }
    const imageData = await Promise.all(images.map((x) => this.preprocess(x)));
    const trimapData = await Promise.all(
      trimaps.map(
        (x) => this.preprocess(x, {
          do_normalize: false,
          do_convert_rgb: false,
          do_convert_grayscale: true
        })
      )
    );
    const pixel_values = stack(
      imageData.map(
        // Concatenate images and trimaps
        (x, i) => cat([x.pixel_values, trimapData[i].pixel_values], 0)
      ),
      0
    );
    return {
      pixel_values,
      // Original sizes of images
      original_sizes: imageData.map((x) => x.original_size),
      // Reshaped sizes of images, before padding or cropping
      reshaped_input_sizes: imageData.map((x) => x.reshaped_input_size)
    };
  }
};

// src/models/vitpose/image_processing_vitpose.js
var VitPoseImageProcessor = class extends ImageProcessor {
  /**
   * Transform the heatmaps into keypoint predictions and transform them back to the image.
   * NOTE: This is a naive implementation and does not include advanced post-processing techniques,
   * so the results may not be as accurate as the original implementation.
   * @param {import('../../utils/tensor.js').Tensor} outputs The model outputs.
   * @param {[number, number, number, number][][]} boxes List or array of bounding boxes for each image.
   * Each box should be a list of 4 floats representing the bounding box coordinates in COCO format (top_left_x, top_left_y, width, height).
   * @returns {{
   *   bbox: [number, number, number, number],
   *   scores: number[],
   *   labels: number[],
   *   keypoints: [number, number][]
   * }[][]} List of keypoints predictions for each image.
   */
  post_process_pose_estimation(outputs, boxes, {
    threshold = null
    // TODO:
    // kernel_size = 11,
    // target_sizes = null,
  } = {}) {
    const heatmaps = outputs.tolist();
    const [batch_size, num_classes, height, width] = outputs.dims;
    const results = [];
    for (let b = 0; b < batch_size; ++b) {
      const heatmap = heatmaps[b];
      const bboxes = boxes[b];
      const batch_results = [];
      for (let n = 0; n < bboxes.length; ++n) {
        const bbox = bboxes[n];
        const keypoints = [];
        const scores = [];
        const labels = [];
        const xScale = bbox.at(-2) / width;
        const yScale = bbox.at(-1) / height;
        for (let c = 0; c < heatmap.length; ++c) {
          let [xWeightedSum, yWeightedSum] = [0, 0];
          let sum = 0;
          let score = -Infinity;
          const row = heatmap[c];
          for (let y = 0; y < row.length; ++y) {
            const col = row[y];
            for (let x = 0; x < col.length; ++x) {
              const value = col[x];
              sum += value;
              score = Math.max(score, value);
              xWeightedSum += (x + 0.5) * value;
              yWeightedSum += y * value;
            }
          }
          if (threshold != null && score < threshold) continue;
          const keypoint = [xScale * xWeightedSum / sum, yScale * yWeightedSum / sum];
          keypoints.push(keypoint);
          labels.push(c);
          scores.push(score);
        }
        batch_results.push({
          bbox,
          scores,
          labels,
          keypoints
        });
      }
      results.push(batch_results);
    }
    return results;
  }
};

// src/models/yolos/image_processing_yolos.js
var YolosImageProcessor = class extends ImageProcessor {
  /** @type {typeof post_process_object_detection} */
  post_process_object_detection(...args) {
    return post_process_object_detection(...args);
  }
};
var YolosFeatureExtractor = class extends YolosImageProcessor {
};

// src/models/auto/image_processing_auto.js
var AutoImageProcessor = class {
  /** @type {typeof ImageProcessor.from_pretrained} */
  static async from_pretrained(pretrained_model_name_or_path, options = {}) {
    const preprocessorConfig = await getModelJSON(
      pretrained_model_name_or_path,
      IMAGE_PROCESSOR_NAME,
      true,
      options
    );
    const key = preprocessorConfig.image_processor_type ?? preprocessorConfig.feature_extractor_type;
    let image_processor_class = image_processors_exports[key?.replace(/Fast$/, "")];
    if (!image_processor_class) {
      if (key !== void 0) {
        logger.warn(
          `Image processor type '${key}' not found, assuming base ImageProcessor. Please report this at ${GITHUB_ISSUE_URL}.`
        );
      }
      image_processor_class = ImageProcessor;
    }
    return new image_processor_class(preprocessorConfig);
  }
};

// src/models/florence2/processing_florence2.js
var Florence2Processor = class extends Processor {
  static tokenizer_class = AutoTokenizer;
  static image_processor_class = AutoImageProcessor;
  constructor(config, components, chat_template) {
    super(config, components, chat_template);
    const {
      // @ts-expect-error TS2339
      tasks_answer_post_processing_type,
      // @ts-expect-error TS2339
      task_prompts_without_inputs,
      // @ts-expect-error TS2339
      task_prompts_with_input
    } = this.image_processor.config;
    this.tasks_answer_post_processing_type = new Map(Object.entries(tasks_answer_post_processing_type ?? {}));
    this.task_prompts_without_inputs = new Map(Object.entries(task_prompts_without_inputs ?? {}));
    this.task_prompts_with_input = new Map(Object.entries(task_prompts_with_input ?? {}));
    this.regexes = {
      quad_boxes: /(.+?)<loc_(\d+)><loc_(\d+)><loc_(\d+)><loc_(\d+)><loc_(\d+)><loc_(\d+)><loc_(\d+)><loc_(\d+)>/gm,
      bboxes: /([^<]+)?<loc_(\d+)><loc_(\d+)><loc_(\d+)><loc_(\d+)>/gm
    };
    this.size_per_bin = 1e3;
  }
  /**
   * Helper function to construct prompts from input texts
   * @param {string|string[]} text
   * @returns {string[]}
   */
  construct_prompts(text) {
    if (typeof text === "string") {
      text = [text];
    }
    const prompts = [];
    for (const t of text) {
      if (this.task_prompts_without_inputs.has(t)) {
        prompts.push(this.task_prompts_without_inputs.get(t));
      } else {
        for (const [task, prompt] of this.task_prompts_with_input) {
          if (t.includes(task)) {
            prompts.push(prompt.replaceAll("{input}", t).replaceAll(task, ""));
            break;
          }
        }
        if (prompts.length !== text.length) {
          prompts.push(t);
        }
      }
    }
    return prompts;
  }
  /**
   * Post-process the output of the model to each of the task outputs.
   * @param {string} text The text to post-process.
   * @param {string} task The task to post-process the text for.
   * @param {[number, number]} image_size The size of the image. height x width.
   */
  post_process_generation(text, task, image_size) {
    const task_answer_post_processing_type = this.tasks_answer_post_processing_type.get(task) ?? "pure_text";
    text = text.replaceAll("<s>", "").replaceAll("</s>", "");
    let final_answer;
    switch (task_answer_post_processing_type) {
      case "pure_text":
        final_answer = text;
        break;
      case "description_with_bboxes":
      case "bboxes":
      case "phrase_grounding":
      case "ocr":
        const key = task_answer_post_processing_type === "ocr" ? "quad_boxes" : "bboxes";
        const matches = text.matchAll(this.regexes[key]);
        const labels = [];
        const items = [];
        for (const [_, label, ...locations] of matches) {
          labels.push(label ? label.trim() : labels.at(-1) ?? "");
          items.push(
            locations.map(
              (x, i) => (
                // NOTE: Add 0.5 to use the center position of the bin as the coordinate.
                (Number(x) + 0.5) / this.size_per_bin * image_size[i % 2]
              )
            )
          );
        }
        final_answer = { labels, [key]: items };
        break;
      default:
        throw new Error(`Task "${task}" (of type "${task_answer_post_processing_type}") not yet implemented.`);
    }
    return { [task]: final_answer };
  }
  // NOTE: images and text are switched from the python version
  // `images` is required, `text` is optional
  async _call(images, text = null, kwargs = {}) {
    if (!images && !text) {
      throw new Error("Either text or images must be provided");
    }
    const image_inputs = await this.image_processor(images, kwargs);
    const text_inputs = text ? this.tokenizer(this.construct_prompts(text), kwargs) : {};
    return {
      ...image_inputs,
      ...text_inputs
    };
  }
};

// src/models/gemma3/processing_gemma3.js
var Gemma3Processor = class extends Processor {
  static tokenizer_class = AutoTokenizer;
  static image_processor_class = AutoImageProcessor;
  static uses_processor_config = true;
  static uses_chat_template_file = true;
  constructor(config, components, chat_template) {
    super(config, components, chat_template);
    this.image_seq_length = this.config.image_seq_length;
    const { boi_token, image_token, eoi_token } = this.tokenizer.config;
    this.boi_token = boi_token;
    this.image_token = image_token;
    this.eoi_token = eoi_token;
    const image_tokens_expanded = image_token.repeat(this.image_seq_length);
    this.full_image_sequence = `

${boi_token}${image_tokens_expanded}${eoi_token}

`;
  }
  /**
   * @param {string|string[]} text
   * @param {import('../../utils/image.js').RawImage|import('../../utils/image.js').RawImage[]} [images]
   * @param {Object} [options]
   */
  async _call(text, images = null, options = {}) {
    if (typeof text === "string") {
      text = [text];
    }
    let image_inputs;
    if (images) {
      image_inputs = await this.image_processor(images, options);
      text = text.map((prompt) => prompt.replaceAll(this.boi_token, this.full_image_sequence));
    }
    const text_inputs = this.tokenizer(text, options);
    return {
      ...text_inputs,
      ...image_inputs
    };
  }
};

// src/models/gemma3n/processing_gemma3n.js
var Gemma3nProcessor = class extends Processor {
  static image_processor_class = AutoImageProcessor;
  static feature_extractor_class = AutoFeatureExtractor;
  static tokenizer_class = AutoTokenizer;
  static uses_processor_config = true;
  static uses_chat_template_file = true;
  constructor(config, components, chat_template) {
    super(config, components, chat_template);
    this.audio_seq_length = this.config.audio_seq_length;
    this.image_seq_length = this.config.image_seq_length;
    const {
      // Audio tokens
      audio_token_id,
      boa_token,
      audio_token,
      eoa_token,
      // Image tokens
      image_token_id,
      boi_token,
      image_token,
      eoi_token
    } = this.tokenizer.config;
    this.audio_token_id = audio_token_id;
    this.boa_token = boa_token;
    this.audio_token = audio_token;
    const audio_tokens_expanded = audio_token.repeat(this.audio_seq_length);
    this.full_audio_sequence = `

${boa_token}${audio_tokens_expanded}${eoa_token}

`;
    this.image_token_id = image_token_id;
    this.boi_token = boi_token;
    this.image_token = image_token;
    const image_tokens_expanded = image_token.repeat(this.image_seq_length);
    this.full_image_sequence = `

${boi_token}${image_tokens_expanded}${eoi_token}

`;
  }
  /**
   *
   * @param {string|string[]} text
   * @param {RawImage|RawImage[]|RawImage[][]} images
   * @param {RawAudio|RawAudio[]|RawAudio[][]} audio
   * @returns {Promise<any>}
   */
  async _call(text, images = null, audio = null, options = {}) {
    if (typeof text === "string") {
      text = [text];
    }
    let audio_inputs;
    if (audio) {
      audio_inputs = await this.feature_extractor(audio, options);
      text = text.map((prompt) => prompt.replaceAll(this.audio_token, this.full_audio_sequence));
    }
    let image_inputs;
    if (images) {
      image_inputs = await this.image_processor(images, options);
      text = text.map((prompt) => prompt.replaceAll(this.image_token, this.full_image_sequence));
    }
    let text_inputs = this.tokenizer(text, options);
    return {
      ...text_inputs,
      ...image_inputs,
      ...audio_inputs
    };
  }
};

// src/models/gemma4/processing_gemma4.js
var Gemma4Processor = class extends Processor {
  static uses_processor_config = true;
  static uses_chat_template_file = true;
  constructor(config, components, chat_template) {
    super(config, components, chat_template);
    this.audio_ms_per_token = this.config.audio_ms_per_token ?? 40;
    this.audio_seq_length = this.config.audio_seq_length ?? 750;
    this.image_seq_length = this.config.image_seq_length ?? 280;
    const { audio_token, boa_token, eoa_token, image_token, boi_token, eoi_token } = this.tokenizer.config;
    this.audio_token = audio_token;
    this.boa_token = boa_token;
    this.eoa_token = eoa_token;
    this.image_token = image_token;
    this.boi_token = boi_token;
    this.eoi_token = eoi_token;
  }
  static async from_pretrained(pretrained_model_name_or_path, options = {}) {
    const [config, tokenizer, chat_template] = await Promise.all([
      getModelJSON(pretrained_model_name_or_path, PROCESSOR_NAME, true, options),
      AutoTokenizer.from_pretrained(pretrained_model_name_or_path, options),
      getModelText(pretrained_model_name_or_path, CHAT_TEMPLATE_NAME, false, options)
    ]);
    const components = { tokenizer };
    if (config.image_processor) {
      components.image_processor = new Gemma4ImageProcessor(config.image_processor);
    }
    if (config.feature_extractor) {
      components.feature_extractor = new Gemma4AudioFeatureExtractor(config.feature_extractor);
    }
    return new this(config, components, chat_template);
  }
  /**
   * Compute the number of audio soft tokens for a single waveform.
   * Replicates the audio encoder's sequence-length arithmetic:
   * mel framing → two SSCP conv layers (kernel=3, stride=2) → cap.
   * @param {number} num_samples
   * @param {number} sampling_rate
   * @returns {number}
   */
  _compute_audio_num_tokens(num_samples, sampling_rate) {
    const frame_length = Math.round(sampling_rate * 20 / 1e3);
    const hop_length = Math.round(sampling_rate * 10 / 1e3);
    const pad_left = Math.floor(frame_length / 2);
    let t = Math.floor((num_samples + pad_left - frame_length - 1) / hop_length) + 1;
    if (t <= 0) return 0;
    for (let i = 0; i < 2; ++i) t = Math.floor((t - 1) / 2) + 1;
    return Math.min(t, this.audio_seq_length);
  }
  async _call(text, images = null, audio = null, options = {}) {
    if (typeof text === "string") {
      text = [text];
    }
    let image_inputs;
    if (images) {
      image_inputs = await this.image_processor(images, options);
      const counts = image_inputs.num_soft_tokens_per_image;
      let i = 0;
      text = text.map(
        (t) => t.replaceAll(
          this.image_token,
          () => `

${this.boi_token}${this.image_token.repeat(counts[i++])}${this.eoi_token}

`
        )
      );
    }
    let audio_inputs;
    if (audio) {
      const audio_array = Array.isArray(audio) ? audio : [audio];
      audio_inputs = await this.feature_extractor(audio_array[0], options);
      const sampling_rate = this.feature_extractor.config.sampling_rate ?? 16e3;
      let i = 0;
      text = text.map(
        (t) => t.replaceAll(
          this.audio_token,
          () => `

${this.boa_token}${this.audio_token.repeat(this._compute_audio_num_tokens(audio_array[i++].length, sampling_rate))}${this.eoa_token}

`
        )
      );
    }
    return {
      ...this.tokenizer(text, options),
      ...image_inputs,
      ...audio_inputs
    };
  }
};

// src/models/qwen2_vl/processing_qwen2_vl.js
var Qwen2VLProcessor = class extends Processor {
  static image_processor_class = AutoImageProcessor;
  static tokenizer_class = AutoTokenizer;
  static image_token = "<|image_pad|>";
  /**
   *
   * @param {string|string[]} text
   * @param {RawImage|RawImage[]} images
   * @param  {...any} args
   * @returns {Promise<any>}
   */
  async _call(text, images = null, ...args) {
    if (!Array.isArray(text)) {
      text = [text];
    }
    let image_inputs, image_grid_thw;
    if (images) {
      image_inputs = await this.image_processor(images);
      image_grid_thw = image_inputs.image_grid_thw;
    }
    if (image_grid_thw) {
      let merge_length = this.image_processor.config.merge_size ** 2;
      let index = 0;
      const image_token = (
        /** @type {typeof Qwen2VLProcessor} */
        this.constructor.image_token
      );
      const image_grid_thw_list = image_grid_thw.tolist();
      text = text.map((t) => {
        while (t.includes(image_token)) {
          const prod = Number(image_grid_thw_list[index++].reduce((a, b) => a * b, 1n));
          t = t.replace(image_token, "<|placeholder|>".repeat(Math.floor(prod / merge_length)));
        }
        return t.replaceAll("<|placeholder|>", image_token);
      });
    }
    const text_inputs = this.tokenizer(text);
    return {
      ...text_inputs,
      ...image_inputs
    };
  }
};

// src/models/glm46v/processing_glm46v.js
var Glm46VProcessor = class extends Qwen2VLProcessor {
  static image_token = "<|image|>";
};

// src/models/granite_speech/processing_granite_speech.js
var GraniteSpeechProcessor = class extends Processor {
  static tokenizer_class = AutoTokenizer;
  static feature_extractor_class = AutoFeatureExtractor;
  static uses_processor_config = true;
  /**
   * Compute the number of audio tokens for a given raw audio length.
   * @param {number} audioLength Raw audio sample count.
   * @returns {number} Number of projector output tokens.
   */
  _get_num_audio_features(audioLength) {
    const { hop_length } = this.feature_extractor.config.melspec_kwargs;
    const { projector_window_size, projector_downsample_rate } = this.feature_extractor.config;
    const effective_window_size = Math.floor(projector_window_size / projector_downsample_rate);
    const mel_length = Math.floor(audioLength / hop_length) + 1;
    const encoder_length = Math.floor(mel_length / 2);
    const nblocks = Math.ceil(encoder_length / projector_window_size);
    return nblocks * effective_window_size;
  }
  /**
   * @param {string} text The text input to process.
   * @param {Float32Array} audio The audio input to process.
   */
  async _call(text, audio = null, kwargs = {}) {
    if (Array.isArray(text)) {
      throw new Error("Batched inputs are not supported yet.");
    }
    let audio_inputs = {};
    if (audio) {
      const { input_features } = await this.feature_extractor(audio);
      audio_inputs["input_features"] = input_features;
      const audio_embed_size = this._get_num_audio_features(audio.length);
      const mask_data = new Uint8Array(audio_embed_size).fill(1);
      audio_inputs["input_features_mask"] = new Tensor2("bool", mask_data, [1, audio_embed_size]);
      const audio_token = this.config.audio_token ?? "<|audio|>";
      if (!text.includes(audio_token)) {
        throw new Error(`The input text does not contain the audio token ${audio_token}.`);
      }
      text = text.replaceAll(audio_token, audio_token.repeat(audio_embed_size));
    }
    const text_inputs = this.tokenizer(text, {
      add_special_tokens: false,
      ...kwargs
    });
    return {
      ...text_inputs,
      ...audio_inputs
    };
  }
};

// src/models/grounding_dino/processing_grounding_dino.js
function get_phrases_from_posmap(posmaps, input_ids) {
  const left_idx = 0;
  const right_idx = posmaps.dims.at(-1) - 1;
  const posmaps_list = posmaps.tolist();
  posmaps_list.fill(false, 0, left_idx + 1);
  posmaps_list.fill(false, right_idx);
  const input_ids_list = input_ids.tolist();
  return posmaps_list.map((val, idx) => val ? idx : null).filter((idx) => idx !== null).map((i) => input_ids_list[i]);
}
var GroundingDinoProcessor = class extends Processor {
  static tokenizer_class = AutoTokenizer;
  static image_processor_class = AutoImageProcessor;
  /**
   * @typedef {import('../../utils/image.js').RawImage} RawImage
   */
  /**
   *
   * @param {RawImage|RawImage[]|RawImage[][]} images
   * @param {string|string[]} text
   * @returns {Promise<any>}
   */
  async _call(images, text, options = {}) {
    const image_inputs = images ? await this.image_processor(images, options) : {};
    const text_inputs = text ? this.tokenizer(text, options) : {};
    return {
      ...text_inputs,
      ...image_inputs
    };
  }
  post_process_grounded_object_detection(outputs, input_ids, { box_threshold = 0.25, text_threshold = 0.25, target_sizes = null } = {}) {
    const { logits, pred_boxes } = outputs;
    const batch_size = logits.dims[0];
    if (target_sizes !== null && target_sizes.length !== batch_size) {
      throw Error("Make sure that you pass in as many target sizes as the batch dimension of the logits");
    }
    const num_queries = logits.dims.at(1);
    const probs = logits.sigmoid();
    const scores = probs.max(-1).tolist();
    const boxes = pred_boxes.tolist().map((batch) => batch.map((box) => center_to_corners_format(box)));
    const results = [];
    for (let i = 0; i < batch_size; ++i) {
      const target_size = target_sizes !== null ? target_sizes[i] : null;
      if (target_size !== null) {
        boxes[i] = boxes[i].map((box) => box.map((x, j) => x * target_size[(j + 1) % 2]));
      }
      const batch_scores = scores[i];
      const final_scores = [];
      const final_phrases = [];
      const final_boxes = [];
      for (let j = 0; j < num_queries; ++j) {
        const score = batch_scores[j];
        if (score <= box_threshold) {
          continue;
        }
        const box = boxes[i][j];
        const prob = probs[i][j];
        final_scores.push(score);
        final_boxes.push(box);
        const phrases = get_phrases_from_posmap(prob.gt(text_threshold), input_ids[i]);
        final_phrases.push(phrases);
      }
      results.push({ scores: final_scores, boxes: final_boxes, labels: this.batch_decode(final_phrases) });
    }
    return results;
  }
};

// src/models/idefics3/processing_idefics3.js
function _prompt_split_image(image_seq_len, image_rows, image_cols, fake_token_around_image, image_token, global_img_token) {
  let text_split_images = "";
  for (let n_h = 0; n_h < image_rows; ++n_h) {
    for (let n_w = 0; n_w < image_cols; ++n_w) {
      text_split_images += fake_token_around_image + `<row_${n_h + 1}_col_${n_w + 1}>` + image_token.repeat(image_seq_len);
    }
    text_split_images += "\n";
  }
  text_split_images += `
${fake_token_around_image}${global_img_token}` + image_token.repeat(image_seq_len) + `${fake_token_around_image}`;
  return text_split_images;
}
function _prompt_single_image(image_seq_len, fake_token_around_image, image_token, global_img_token) {
  return `${fake_token_around_image}${global_img_token}` + image_token.repeat(image_seq_len) + `${fake_token_around_image}`;
}
function get_image_prompt_string(image_rows, image_cols, image_seq_len, fake_token_around_image, image_token, global_img_token) {
  if (image_rows === 0 && image_cols === 0) {
    return _prompt_single_image(image_seq_len, fake_token_around_image, image_token, global_img_token);
  }
  return _prompt_split_image(
    image_seq_len,
    image_rows,
    image_cols,
    fake_token_around_image,
    image_token,
    global_img_token
  );
}
var Idefics3Processor = class extends Processor {
  static image_processor_class = AutoImageProcessor;
  static tokenizer_class = AutoTokenizer;
  static uses_processor_config = true;
  fake_image_token = "<fake_token_around_image>";
  image_token = "<image>";
  global_img_token = "<global-img>";
  /**
   *
   * @param {string|string[]} text
   * @param {RawImage|RawImage[]|RawImage[][]} images
   * @returns {Promise<any>}
   */
  async _call(text, images = null, options = {}) {
    options.return_row_col_info ??= true;
    let image_inputs;
    if (images) {
      image_inputs = await this.image_processor(images, options);
    }
    if (!Array.isArray(text)) {
      text = [text];
    }
    const image_rows = image_inputs.rows ?? [new Array(text.length).fill(0)];
    const image_cols = image_inputs.cols ?? [new Array(text.length).fill(0)];
    const image_seq_len = this.config.image_seq_len;
    const n_images_in_text = [];
    const prompt_strings = [];
    for (let i = 0; i < text.length; ++i) {
      const sample = text[i];
      const sample_rows = image_rows[i];
      const sample_cols = image_cols[i];
      n_images_in_text.push(count(sample, this.image_token));
      const image_prompt_strings = sample_rows.map(
        (n_rows, j) => get_image_prompt_string(
          n_rows,
          sample_cols[j],
          image_seq_len,
          this.fake_image_token,
          this.image_token,
          this.global_img_token
        )
      );
      const split_sample = sample.split(this.image_token);
      if (split_sample.length === 0) {
        throw new Error("The image token should be present in the text.");
      }
      let new_sample = split_sample[0];
      for (let j = 0; j < image_prompt_strings.length; ++j) {
        new_sample += image_prompt_strings[j] + split_sample[j + 1];
      }
      prompt_strings.push(new_sample);
    }
    const text_inputs = this.tokenizer(prompt_strings);
    return {
      ...text_inputs,
      ...image_inputs
    };
  }
};

// src/models/janus/processing_janus.js
var VLChatProcessor = class extends Processor {
  static image_processor_class = AutoImageProcessor;
  static tokenizer_class = AutoTokenizer;
  static uses_processor_config = true;
  constructor(config, components, chat_template) {
    super(config, components, chat_template);
    this.image_tag = this.config.image_tag;
    this.image_start_tag = this.config.image_start_tag;
    this.image_end_tag = this.config.image_end_tag;
    this.num_image_tokens = this.config.num_image_tokens;
  }
  /**
   * @typedef {Object} MultimodalMessageProperties Additional properties for multimodal messages.
   * @property {(RawImage | string | URL)[]} [images] The images in the message.
   * @typedef {(import('../../tokenization_utils.js').Message & MultimodalMessageProperties)[]} MultimodalConversation The conversation possibly containing multimodal inputs.
   */
  /**
   * @typedef {Object} VLCChatProcessorResult The processed input.
   * @property {Tensor} input_ids The input IDs.
   * @property {Tensor} attention_mask The attention mask.
   * @property {Tensor} images_seq_mask The image sequence mask.
   * @property {Tensor} images_emb_mask The image embedding mask.
   */
  /**
   * @param {MultimodalConversation} conversation The chat messages to process.
   * @param {Object} options Additional options for processing.
   * @param {RawImage|RawImage[]} [options.images] The images to process, if not set in the conversation.
   * @param {string} [options.chat_template="default"] The chat template to use.
   * @returns {Promise<VLCChatProcessorResult | VLCChatProcessorResult & import('../../image_processors_utils.js').ImageProcessorResult>} The processed input.
   */
  async _call(conversation, { images = null, chat_template = "default" } = {}) {
    if (!images) {
      images = await Promise.all(
        conversation.filter((msg) => msg.images).flatMap((msg) => msg.images).map((img) => RawImage.read(img))
      );
    } else if (!Array.isArray(images)) {
      images = [images];
    }
    const tokenizer = this.tokenizer;
    const result = tokenizer.apply_chat_template(conversation, {
      tokenize: false,
      add_generation_prompt: true,
      chat_template
    });
    const encode = (text) => tokenizer.encode(text, { add_special_tokens: false });
    const parts = (
      /** @type {string} */
      result.split(this.image_tag)
    );
    const num_images = parts.length - 1;
    if (images.length !== num_images) {
      throw new Error(
        `Number of images provided (${images.length}) does not match number of "${this.image_tag}" image tags (${num_images})`
      );
    }
    const [image_placeholder_tag_id, image_start_tag_id, image_end_tag_id] = tokenizer.convert_tokens_to_ids([
      this.image_tag,
      this.image_start_tag,
      this.image_end_tag
    ]);
    let input_ids = encode(parts[0]);
    let images_seq_mask = new Array(input_ids.length).fill(false);
    for (let i = 1; i < parts.length; ++i) {
      const placeholder_image_tokens = new Array(this.num_image_tokens).fill(image_placeholder_tag_id);
      const tokens = encode(parts[i]);
      input_ids = mergeArrays(
        input_ids,
        [image_start_tag_id],
        placeholder_image_tokens,
        [image_end_tag_id],
        tokens
      );
      const image_mask = new Array(this.num_image_tokens).fill(true);
      images_seq_mask = mergeArrays(
        images_seq_mask,
        [false],
        image_mask,
        [false],
        new Array(tokens.length).fill(false)
      );
    }
    const dims = [1, input_ids.length];
    const final = {
      input_ids: new Tensor2("int64", input_ids, dims),
      attention_mask: new Tensor2("int64", new Array(input_ids.length).fill(1), dims),
      images_seq_mask: new Tensor2("bool", images_seq_mask, dims),
      images_emb_mask: new Tensor2("bool", new Array(num_images * this.num_image_tokens).fill(true), [
        1,
        num_images,
        this.num_image_tokens
      ])
    };
    if (images && images.length > 0) {
      const image_inputs = await this.image_processor(images);
      image_inputs.pixel_values.unsqueeze_(0);
      return { ...final, ...image_inputs };
    }
    return final;
  }
};

// src/models/jina_clip/processing_jina_clip.js
var JinaCLIPProcessor = class extends Processor {
  static tokenizer_class = AutoTokenizer;
  static image_processor_class = AutoImageProcessor;
  async _call(text = null, images = null, kwargs = {}) {
    if (!text && !images) {
      throw new Error("Either text or images must be provided");
    }
    const text_inputs = text ? this.tokenizer(text, kwargs) : {};
    const image_inputs = images ? await this.image_processor(images, kwargs) : {};
    return {
      ...text_inputs,
      ...image_inputs
    };
  }
};

// src/models/lfm2_vl/processing_lfm2_vl.js
var Lfm2VlProcessor = class extends Processor {
  static tokenizer_class = AutoTokenizer;
  static image_processor_class = AutoImageProcessor;
  /**
   * @param {RawImage|RawImage[]} images
   * @param {string|string[]|null} [text]
   * @param {Record<string, any>} [kwargs]
   */
  async _call(images, text = null, kwargs = {}) {
    const { image_rows, image_cols, image_sizes, ...image_inputs } = await this.image_processor(images, {
      ...kwargs,
      return_row_col_info: true
    });
    if (text) {
      const image_token = this.config.image_token ?? "<image>";
      const {
        tile_size = 512,
        downsample_factor = 2,
        encoder_patch_size = 16,
        use_thumbnail = true
      } = (
        /** @type {Record<string, any>} */
        this.image_processor.config
      );
      const ds2 = (s) => Math.ceil(Math.floor(s / encoder_patch_size) / downsample_factor);
      const tokens_per_tile = ds2(tile_size) ** 2;
      const image_start = this.config.image_start_token ?? "<|image_start|>";
      const image_end = this.config.image_end_token ?? "<|image_end|>";
      const thumbnail_token = this.config.image_thumbnail ?? "<|img_thumbnail|>";
      if (!Array.isArray(text)) text = [text];
      let image_idx = 0;
      text = text.map((sample) => {
        const parts = sample.split(image_token);
        return parts[0] + parts.slice(1).map((part) => {
          const idx = image_idx++;
          const [h, w] = image_sizes[idx];
          const rows = image_rows[idx], cols = image_cols[idx];
          const tokens_for_image = ds2(h) * ds2(w);
          let expanded = image_start;
          if (rows > 1 || cols > 1) {
            const tile_str = image_token.repeat(tokens_per_tile);
            for (let r = 0; r < rows; ++r)
              for (let c = 0; c < cols; ++c)
                expanded += `<|img_row_${r + 1}_col_${c + 1}|>` + tile_str;
            if (use_thumbnail) expanded += thumbnail_token + image_token.repeat(tokens_for_image);
          } else {
            expanded += image_token.repeat(tokens_for_image);
          }
          return expanded + image_end + part;
        }).join("");
      });
    }
    return {
      ...image_inputs,
      ...text ? this.tokenizer(text, kwargs) : {}
    };
  }
};

// src/models/llava/processing_llava.js
var LlavaProcessor = class extends Processor {
  static tokenizer_class = AutoTokenizer;
  static image_processor_class = AutoImageProcessor;
  static uses_processor_config = true;
  /**
   * @typedef {import('../../utils/image.js').RawImage} RawImage
   */
  // `images` is required, `text` is optional
  async _call(images, text = null, kwargs = {}) {
    const image_inputs = await this.image_processor(images, kwargs);
    if (text) {
      const [height, width] = image_inputs.pixel_values.dims.slice(-2);
      const { image_token, patch_size, num_additional_image_tokens } = this.config;
      const num_image_tokens = Math.floor(height / patch_size) * Math.floor(width / patch_size) + num_additional_image_tokens;
      text = structuredClone(text);
      if (!Array.isArray(text)) {
        text = [text];
      }
      for (let i = 0; i < text.length; ++i) {
        text[i] = text[i].replace(image_token, image_token.repeat(num_image_tokens));
      }
    }
    const text_inputs = text ? this.tokenizer(text, kwargs) : {};
    return {
      ...image_inputs,
      ...text_inputs
    };
  }
};

// src/models/mgp_str/processing_mgp_str.js
var DECODE_TYPE_MAPPING = {
  char: ["char_decode", 1],
  bpe: ["bpe_decode", 2],
  wp: ["wp_decode", 102]
};
var MgpstrProcessor = class extends Processor {
  static tokenizer_class = AutoTokenizer;
  static image_processor_class = AutoImageProcessor;
  /**
   * @returns {import('./tokenization_mgp_str.js').MgpstrTokenizer} The character tokenizer.
   */
  get char_tokenizer() {
    return this.components.char_tokenizer;
  }
  /**
   * @returns {import('../gpt2/tokenization_gpt2.js').GPT2Tokenizer} The BPE tokenizer.
   */
  get bpe_tokenizer() {
    return this.components.bpe_tokenizer;
  }
  /**
   * @returns {import('../bert/tokenization_bert.js').BertTokenizer} The WordPiece tokenizer.
   */
  get wp_tokenizer() {
    return this.components.wp_tokenizer;
  }
  /**
   * Helper function to decode the model prediction logits.
   * @param {import('../../utils/tensor.js').Tensor} pred_logits Model prediction logits.
   * @param {string} format Type of model prediction. Must be one of ['char', 'bpe', 'wp'].
   * @returns {[string[], number[]]} The decoded sentences and their confidence scores.
   */
  _decode_helper(pred_logits, format2) {
    if (!DECODE_TYPE_MAPPING.hasOwnProperty(format2)) {
      throw new Error(`Format ${format2} is not supported.`);
    }
    const [decoder_name, eos_token] = DECODE_TYPE_MAPPING[format2];
    const decoder = this[decoder_name].bind(this);
    const [batch_size, batch_max_length] = pred_logits.dims;
    const conf_scores = [];
    const all_ids = [];
    const pred_logits_list = pred_logits.tolist();
    for (let i = 0; i < batch_size; ++i) {
      const logits = pred_logits_list[i];
      const ids = [];
      const scores = [];
      for (let j = 1; j < batch_max_length; ++j) {
        const [max_prob, max_prob_index] = max(softmax(logits[j]));
        scores.push(max_prob);
        if (max_prob_index == eos_token) {
          break;
        }
        ids.push(max_prob_index);
      }
      const confidence_score = scores.length > 0 ? scores.reduce((a, b) => a * b, 1) : 0;
      all_ids.push(ids);
      conf_scores.push(confidence_score);
    }
    const decoded = decoder(all_ids);
    return [decoded, conf_scores];
  }
  /**
   * Convert a list of lists of char token ids into a list of strings by calling char tokenizer.
   * @param {number[][]} sequences List of tokenized input ids.
   * @returns {string[]} The list of char decoded sentences.
   */
  char_decode(sequences) {
    return this.char_tokenizer.batch_decode(sequences).map((str) => str.replaceAll(" ", ""));
  }
  /**
   * Convert a list of lists of BPE token ids into a list of strings by calling BPE tokenizer.
   * @param {number[][]} sequences List of tokenized input ids.
   * @returns {string[]} The list of BPE decoded sentences.
   */
  bpe_decode(sequences) {
    return this.bpe_tokenizer.batch_decode(sequences);
  }
  /**
   * Convert a list of lists of word piece token ids into a list of strings by calling word piece tokenizer.
   * @param {number[][]} sequences List of tokenized input ids.
   * @returns {string[]} The list of wp decoded sentences.
   */
  wp_decode(sequences) {
    return this.wp_tokenizer.batch_decode(sequences).map((str) => str.replaceAll(" ", ""));
  }
  /**
   * Convert a list of lists of token ids into a list of strings by calling decode.
   * @param {[import('../../utils/tensor.js').Tensor, import('../../utils/tensor.js').Tensor, import('../../utils/tensor.js').Tensor]} sequences List of tokenized input ids.
   * @returns {{generated_text: string[], scores: number[], char_preds: string[], bpe_preds: string[], wp_preds: string[]}}
   * Dictionary of all the outputs of the decoded results.
   * - generated_text: The final results after fusion of char, bpe, and wp.
   * - scores: The final scores after fusion of char, bpe, and wp.
   * - char_preds: The list of character decoded sentences.
   * - bpe_preds: The list of BPE decoded sentences.
   * - wp_preds: The list of wp decoded sentences.
   */
  // @ts-expect-error The type of this method is not compatible with the one in the base class.
  batch_decode([char_logits, bpe_logits, wp_logits]) {
    const [char_preds, char_scores] = this._decode_helper(char_logits, "char");
    const [bpe_preds, bpe_scores] = this._decode_helper(bpe_logits, "bpe");
    const [wp_preds, wp_scores] = this._decode_helper(wp_logits, "wp");
    const generated_text = [];
    const scores = [];
    for (let i = 0; i < char_preds.length; ++i) {
      const [max_score, max_score_index] = max([char_scores[i], bpe_scores[i], wp_scores[i]]);
      generated_text.push([char_preds[i], bpe_preds[i], wp_preds[i]][max_score_index]);
      scores.push(max_score);
    }
    return {
      generated_text,
      scores,
      char_preds,
      bpe_preds,
      wp_preds
    };
  }
  /** @type {typeof Processor.from_pretrained} */
  static async from_pretrained(...args) {
    const base = await super.from_pretrained(...args);
    const bpe_tokenizer = await AutoTokenizer.from_pretrained("Xenova/gpt2");
    const wp_tokenizer = await AutoTokenizer.from_pretrained("Xenova/bert-base-uncased");
    base.components = {
      image_processor: base.image_processor,
      char_tokenizer: base.tokenizer,
      bpe_tokenizer,
      wp_tokenizer
    };
    return base;
  }
  async _call(images, text = null) {
    const result = await this.image_processor(images);
    if (text) {
      result.labels = this.tokenizer(text).input_ids;
    }
    return result;
  }
};

// src/models/moonshine/processing_moonshine.js
var MoonshineProcessor = class extends Processor {
  static tokenizer_class = AutoTokenizer;
  static feature_extractor_class = AutoFeatureExtractor;
  /**
   * Calls the feature_extractor function with the given audio input.
   * @param {any} audio The audio input to extract features from.
   * @returns {Promise<any>} A Promise that resolves with the extracted features.
   */
  async _call(audio) {
    return await this.feature_extractor(audio);
  }
};

// src/models/owlvit/processing_owlvit.js
var OwlViTProcessor = class extends Processor {
  static tokenizer_class = AutoTokenizer;
  static image_processor_class = AutoImageProcessor;
};

// src/models/paligemma/processing_paligemma.js
var IMAGE_TOKEN = "<image>";
function build_string_from_input(prompt, bos_token, image_seq_len, image_token, num_images) {
  return `${image_token.repeat(image_seq_len * num_images)}${bos_token}${prompt}
`;
}
var PaliGemmaProcessor = class extends Processor {
  static tokenizer_class = AutoTokenizer;
  static image_processor_class = AutoImageProcessor;
  static uses_processor_config = false;
  /**
   * @typedef {import('../../utils/image.js').RawImage} RawImage
   */
  // `images` is required, `text` is optional
  async _call(images, text = null, kwargs = {}) {
    if (!text) {
      logger.warn(
        "You are using PaliGemma without a text prefix. It will perform as a picture-captioning model."
      );
      text = "";
    }
    if (!Array.isArray(images)) {
      images = [images];
    }
    if (!Array.isArray(text)) {
      text = [text];
    }
    const bos_token = this.tokenizer.bos_token;
    const image_seq_length = this.image_processor.config.image_seq_length;
    let input_strings;
    if (text.some((t) => t.includes(IMAGE_TOKEN))) {
      input_strings = text.map((sample) => {
        const expanded_sample = sample.replaceAll(IMAGE_TOKEN, IMAGE_TOKEN.repeat(image_seq_length));
        const bos_rfind_index = expanded_sample.lastIndexOf(IMAGE_TOKEN);
        const bos_index = bos_rfind_index === -1 ? 0 : bos_rfind_index + IMAGE_TOKEN.length;
        return expanded_sample.slice(0, bos_index) + bos_token + expanded_sample.slice(bos_index) + "\n";
      });
    } else {
      logger.warn(
        "You are passing both `text` and `images` to `PaliGemmaProcessor`. The processor expects special image tokens in the text, as many tokens as there are images per each text. It is recommended to add `<image>` tokens in the very beginning of your text. For this call, we will infer how many images each text has and add special tokens."
      );
      input_strings = text.map(
        (sample) => build_string_from_input(sample, bos_token, image_seq_length, IMAGE_TOKEN, images.length)
      );
    }
    const text_inputs = this.tokenizer(input_strings, kwargs);
    const image_inputs = await this.image_processor(images, kwargs);
    return {
      ...image_inputs,
      ...text_inputs
    };
  }
};

// src/models/phi3_v/processing_phi3_v.js
var IMAGE_TOKEN2 = "<|image|>";
var IMAGE_TOKEN_PATTERN = /<\|image_\d+\|>/g;
var Phi3VProcessor = class extends Processor {
  static image_processor_class = AutoImageProcessor;
  static tokenizer_class = AutoTokenizer;
  /**
   *
   * @param {string|string[]} text
   * @param {RawImage|RawImage[]} images
   * @param  { { padding?: boolean, truncation?: boolean, num_crops?: number } | undefined } options
   * @returns {Promise<any>}
   */
  async _call(text, images = null, { padding = true, truncation = true, num_crops = null } = {}) {
    if (!Array.isArray(text)) {
      text = [text];
    }
    let text_inputs, image_inputs;
    if (images) {
      image_inputs = await this.image_processor(images, { num_crops });
      const { num_img_tokens } = image_inputs;
      const prompt_chunks = text.map(
        (t, i) => t.split(IMAGE_TOKEN_PATTERN).join(IMAGE_TOKEN2.repeat(num_img_tokens[i]))
      );
      text_inputs = this.tokenizer(prompt_chunks, { padding, truncation });
      const image_token_id = this.tokenizer._tokenizer.token_to_id(IMAGE_TOKEN2);
      text_inputs.input_ids.map_((id) => id == image_token_id ? -id : id);
    } else {
      text_inputs = this.tokenizer(text);
    }
    return {
      ...text_inputs,
      ...image_inputs
    };
  }
};

// src/models/pixtral/processing_pixtral.js
var PixtralProcessor = class extends Processor {
  static tokenizer_class = AutoTokenizer;
  static image_processor_class = AutoImageProcessor;
  static uses_processor_config = true;
  /**
   * @typedef {import('../../utils/image.js').RawImage} RawImage
   */
  // `images` is required, `text` is optional
  async _call(images, text = null, kwargs = {}) {
    const image_inputs = await this.image_processor(images, kwargs);
    if (text) {
      const [height, width] = image_inputs.pixel_values.dims.slice(-2);
      const { image_token, image_break_token, image_end_token, patch_size, spatial_merge_size } = this.config;
      const real_patch_size = patch_size * spatial_merge_size;
      const num_height_tokens = Math.floor(height / real_patch_size);
      const num_width_tokens = Math.floor(width / real_patch_size);
      text = structuredClone(text);
      if (!Array.isArray(text)) {
        text = [text];
      }
      for (let i = 0; i < text.length; ++i) {
        const width_tokens = image_token.repeat(num_width_tokens);
        const row = width_tokens + image_break_token;
        const finalRow = width_tokens + image_end_token;
        const full2 = row.repeat(num_height_tokens - 1) + finalRow;
        text[i] = text[i].replace(image_token, full2);
      }
    }
    const text_inputs = text ? this.tokenizer(text, kwargs) : {};
    return {
      ...image_inputs,
      ...text_inputs
    };
  }
};

// src/models/pyannote/processing_pyannote.js
var PyAnnoteProcessor = class extends Processor {
  static feature_extractor_class = PyAnnoteFeatureExtractor;
  /**
   * Calls the feature_extractor function with the given audio input.
   * @param {any} audio The audio input to extract features from.
   * @returns {Promise<any>} A Promise that resolves with the extracted features.
   */
  async _call(audio) {
    return await this.feature_extractor(audio);
  }
  /** @type {PyAnnoteFeatureExtractor['post_process_speaker_diarization']} */
  post_process_speaker_diarization(...args) {
    return (
      /** @type {PyAnnoteFeatureExtractor} */
      this.feature_extractor.post_process_speaker_diarization(
        ...args
      )
    );
  }
  get sampling_rate() {
    return this.feature_extractor.config.sampling_rate;
  }
};

// src/models/qwen2_5_vl/processing_qwen2_5_vl.js
var Qwen2_5_VLProcessor = class extends Qwen2VLProcessor {
};

// src/models/qwen3_vl/processing_qwen3_vl.js
var Qwen3VLProcessor = class extends Qwen2_5_VLProcessor {
};

// src/models/sam/processing_sam.js
var SamProcessor = class extends Processor {
  static image_processor_class = AutoImageProcessor;
  async _call(...args) {
    return await this.image_processor(...args);
  }
  post_process_masks(...args) {
    return this.image_processor.post_process_masks(...args);
  }
  reshape_input_points(...args) {
    return this.image_processor.reshape_input_points(...args);
  }
};

// src/models/sam2/processing_sam2.js
var Sam2Processor = class extends SamProcessor {
};
var Sam2VideoProcessor = class extends Sam2Processor {
};

// src/models/speecht5/processing_speecht5.js
var SpeechT5Processor = class extends Processor {
  static tokenizer_class = AutoTokenizer;
  static feature_extractor_class = AutoFeatureExtractor;
  /**
   * Calls the feature_extractor function with the given input.
   * @param {any} input The input to extract features from.
   * @returns {Promise<any>} A Promise that resolves with the extracted features.
   */
  async _call(input) {
    return await this.feature_extractor(input);
  }
};

// src/models/ultravox/processing_ultravox.js
var UltravoxProcessor = class extends Processor {
  static tokenizer_class = AutoTokenizer;
  static feature_extractor_class = AutoFeatureExtractor;
  static uses_processor_config = true;
  /**
   * @param {string} text The text input to process.
   * @param {Float32Array} audio The audio input to process.
   */
  async _call(text, audio = null, kwargs = {}) {
    if (Array.isArray(text)) {
      throw new Error("Batched inputs are not supported yet.");
    }
    let audio_inputs = {};
    if (audio) {
      const audio_len = audio.length;
      const { input_features } = await this.feature_extractor(audio, {
        ...kwargs,
        max_length: audio_len
      });
      const nb_encoder_frames = Math.round(audio_len / this.config.encoder_ds_factor + 1e-4);
      const audio_embed_frames = 1 + Math.ceil(nb_encoder_frames / this.config.stack_factor);
      audio_inputs["audio_token_len"] = [audio_embed_frames];
      audio_inputs["audio_values"] = input_features;
      const image_token = this.config.audio_placeholder;
      if (!text.includes(image_token)) {
        throw new Error(`The input text does not contain the image token ${image_token}.`);
      }
      text = text.replaceAll(image_token, image_token.repeat(audio_embed_frames));
    }
    const text_inputs = this.tokenizer(text, {
      add_special_tokens: false,
      ...kwargs
    });
    return {
      ...text_inputs,
      ...audio_inputs
    };
  }
};

// src/models/voxtral/processing_voxtral.js
var AUDIO_TOKEN = "[AUDIO]";
var BEGIN_AUDIO_TOKEN = "[BEGIN_AUDIO]";
var NUM_AUDIO_TOKENS = 375;
function chunk(audio, n_samples) {
  const chunks = [];
  for (let i = 0; i < audio.length; i += n_samples) {
    chunks.push(audio.subarray(i, Math.min(i + n_samples, audio.length)));
  }
  return chunks;
}
var VoxtralProcessor = class extends Processor {
  static tokenizer_class = AutoTokenizer;
  static feature_extractor_class = AutoFeatureExtractor;
  static uses_processor_config = false;
  /**
   * @param {string} text The text input to process.
   * @param {Float32Array|Float32Array[]} audio The audio input(s) to process.
   */
  async _call(text, audio = null, kwargs = {}) {
    if (Array.isArray(text)) {
      throw new Error("Batched inputs are not supported yet.");
    }
    const audio_inputs = {};
    if (audio) {
      if (!text.includes(AUDIO_TOKEN)) {
        throw new Error(`The input text does not contain the audio token ${AUDIO_TOKEN}.`);
      }
      if (!Array.isArray(audio)) {
        audio = [audio];
      }
      const text_parts = text.split(AUDIO_TOKEN);
      const num_audio_tokens = text_parts.length - 1;
      if (num_audio_tokens !== audio.length) {
        throw new Error(
          `The number of audio inputs (${audio.length}) does not match the number of audio tokens in the text (${num_audio_tokens}).`
        );
      }
      const n_samples = this.feature_extractor.config.n_samples;
      const audio_chunks = audio.map((a) => chunk(a, n_samples));
      const chunk_counts = audio_chunks.map((chunks) => chunks.length);
      const all_chunks = audio_chunks.flat();
      const features = (await Promise.all(all_chunks.map((audio_input) => this.feature_extractor(audio_input, kwargs)))).map((x) => x.input_features);
      audio_inputs["audio_values"] = features.length > 1 ? cat(features, 0) : features[0];
      let new_text = text_parts[0];
      for (let i = 0; i < chunk_counts.length; ++i) {
        new_text += BEGIN_AUDIO_TOKEN;
        for (let j = 0; j < chunk_counts[i]; ++j) {
          new_text += AUDIO_TOKEN.repeat(NUM_AUDIO_TOKENS);
        }
        new_text += text_parts[i + 1];
      }
      text = new_text;
    }
    const text_inputs = this.tokenizer(text, {
      add_special_tokens: false,
      ...kwargs
    });
    return {
      ...text_inputs,
      ...audio_inputs
    };
  }
};

// src/models/voxtral_realtime/processing_voxtral_realtime.js
var NUM_LEFT_PAD_TOKENS = 32;
var NUM_DELAY_TOKENS = 6;
var AUDIO_LENGTH_PER_TOK = 8;
var OFFLINE_STREAMING_BUFFER_TOKENS = 10;
var STREAMING_PAD_TOKEN_ID = 32;
var VoxtralRealtimeProcessor = class extends Processor {
  static tokenizer_class = AutoTokenizer;
  static feature_extractor_class = AutoFeatureExtractor;
  static uses_processor_config = false;
  /** Number of mel frames in the first audio chunk. */
  get num_mel_frames_first_audio_chunk() {
    return (NUM_DELAY_TOKENS + 1) * AUDIO_LENGTH_PER_TOK;
  }
  /** Number of raw audio samples in the first audio chunk. */
  get num_samples_first_audio_chunk() {
    const { hop_length, n_fft } = this.feature_extractor.config;
    return (this.num_mel_frames_first_audio_chunk - 1) * hop_length + Math.floor(n_fft / 2);
  }
  /** Number of raw audio samples per subsequent audio chunk. */
  get num_samples_per_audio_chunk() {
    const { hop_length, n_fft } = this.feature_extractor.config;
    return AUDIO_LENGTH_PER_TOK * hop_length + n_fft;
  }
  /** Number of right-pad tokens for non-streaming mode. */
  get num_right_pad_tokens() {
    return NUM_DELAY_TOKENS + 1 + OFFLINE_STREAMING_BUFFER_TOKENS;
  }
  /** Number of mel frames per text token. */
  get audio_length_per_tok() {
    return AUDIO_LENGTH_PER_TOK;
  }
  /** Number of raw audio samples per token. */
  get raw_audio_length_per_tok() {
    return AUDIO_LENGTH_PER_TOK * this.feature_extractor.config.hop_length;
  }
  /**
   * Process audio input for VoxtralRealtime.
   *
   * In streaming mode with `is_first_audio_chunk=true`, the audio is left-padded
   * with silence and mel features are extracted with `center=true`.
   * Returns `{ input_ids, input_features }`.
   *
   * In streaming mode with `is_first_audio_chunk=false`, the audio chunk is
   * processed with `center=false` and only `{ input_features }` is returned.
   *
   * In non-streaming mode, the audio is right-padded to ensure the model
   * transcribes the full audio, then processed with `center=true`.
   * Returns `{ input_features }`.
   *
   * @param {Float32Array|Float64Array} audio The audio waveform.
   * @param {Object} [options]
   * @param {boolean} [options.is_streaming=false] Whether processing in streaming mode.
   * @param {boolean} [options.is_first_audio_chunk=true] Whether this is the first audio chunk.
   * @returns {Promise<Object>}
   */
  async _call(audio, { is_streaming = false, is_first_audio_chunk = true } = {}) {
    validate_audio_inputs(audio, "VoxtralRealtimeProcessor");
    if (!is_streaming && !is_first_audio_chunk) {
      throw new Error("In non-streaming mode (`is_streaming=false`), `is_first_audio_chunk` must be `true`.");
    }
    if (is_first_audio_chunk) {
      if (is_streaming) {
        const num_left_pad_samples = NUM_LEFT_PAD_TOKENS * this.raw_audio_length_per_tok;
        const padded_audio = new Float32Array(num_left_pad_samples + audio.length);
        padded_audio.set(audio, num_left_pad_samples);
        const audio_encoding = await this.feature_extractor(padded_audio, { center: true });
        const num_pad_tokens = NUM_LEFT_PAD_TOKENS + NUM_DELAY_TOKENS;
        const num_input_tokens = 1 + num_pad_tokens;
        const input_ids_data = new BigInt64Array(num_input_tokens).fill(BigInt(STREAMING_PAD_TOKEN_ID));
        input_ids_data[0] = 1n;
        const input_ids = new Tensor2("int64", input_ids_data, [1, num_input_tokens]);
        return {
          input_ids,
          ...audio_encoding
        };
      } else {
        const right_pad_samples = this.num_right_pad_tokens * this.raw_audio_length_per_tok;
        const padded_audio = new Float32Array(audio.length + right_pad_samples);
        padded_audio.set(audio);
        return await this.feature_extractor(padded_audio, { center: true });
      }
    } else {
      return await this.feature_extractor(audio, { center: false });
    }
  }
};

// src/models/wav2vec2/processing_wav2vec2.js
var Wav2Vec2Processor = class extends Processor {
  static tokenizer_class = AutoTokenizer;
  static feature_extractor_class = AutoFeatureExtractor;
  /**
   * Calls the feature_extractor function with the given audio input.
   * @param {any} audio The audio input to extract features from.
   * @returns {Promise<any>} A Promise that resolves with the extracted features.
   */
  async _call(audio) {
    return await this.feature_extractor(audio);
  }
};

// src/models/wav2vec2_with_lm/processing_wav2vec2_with_lm.js
var Wav2Vec2ProcessorWithLM = class extends Processor {
  static tokenizer_class = AutoTokenizer;
  static feature_extractor_class = AutoFeatureExtractor;
  /**
   * Calls the feature_extractor function with the given audio input.
   * @param {any} audio The audio input to extract features from.
   * @returns {Promise<any>} A Promise that resolves with the extracted features.
   */
  async _call(audio) {
    return await this.feature_extractor(audio);
  }
};

// src/models/whisper/processing_whisper.js
var WhisperProcessor = class extends Processor {
  static tokenizer_class = AutoTokenizer;
  static feature_extractor_class = AutoFeatureExtractor;
  /**
   * Calls the feature_extractor function with the given audio input.
   * @param {any} audio The audio input to extract features from.
   * @returns {Promise<any>} A Promise that resolves with the extracted features.
   */
  async _call(audio) {
    return await this.feature_extractor(audio);
  }
};

// src/models/auto/processing_auto.js
var AutoProcessor = class {
  /** @type {typeof Processor.from_pretrained} */
  static async from_pretrained(pretrained_model_name_or_path, options = {}) {
    const preprocessorConfig = await getModelJSON(
      pretrained_model_name_or_path,
      IMAGE_PROCESSOR_NAME,
      true,
      options
    );
    const { image_processor_type, feature_extractor_type, processor_class } = preprocessorConfig;
    if (processor_class && processors_exports[processor_class]) {
      return processors_exports[processor_class].from_pretrained(pretrained_model_name_or_path, options);
    }
    if (!image_processor_type && !feature_extractor_type) {
      throw new Error("No `image_processor_type` or `feature_extractor_type` found in the config.");
    }
    const components = {};
    if (image_processor_type) {
      const image_processor_class = image_processors_exports[image_processor_type.replace(/Fast$/, "")];
      if (!image_processor_class) {
        throw new Error(`Unknown image_processor_type: '${image_processor_type}'.`);
      }
      components.image_processor = new image_processor_class(preprocessorConfig);
    }
    if (feature_extractor_type) {
      const image_processor_class = image_processors_exports[feature_extractor_type];
      if (image_processor_class) {
        components.image_processor = new image_processor_class(preprocessorConfig);
      } else {
        const feature_extractor_class = feature_extractors_exports[feature_extractor_type];
        if (!feature_extractor_class) {
          throw new Error(`Unknown feature_extractor_type: '${feature_extractor_type}'.`);
        }
        components.feature_extractor = new feature_extractor_class(preprocessorConfig);
      }
    }
    const config = {};
    return new Processor(config, components, null);
  }
};

// src/configs.js
async function loadConfig(pretrained_model_name_or_path, options) {
  return await getModelJSON(pretrained_model_name_or_path, "config.json", true, options);
}
function getNormalizedConfig(config) {
  const mapping = {};
  let init_normalized_config = {};
  switch (config.model_type) {
    // Sub-configs
    case "llava":
    case "paligemma":
    case "gemma3":
    case "florence2":
    case "llava_onevision":
    case "idefics3":
    case "granite_speech":
    case "ultravox":
    case "voxtral":
    case "voxtral_realtime":
    case "smolvlm":
    case "gemma3n":
    case "gemma4":
    case "lfm2_vl":
    case "chatterbox":
    case "lighton_ocr":
    case "glm_ocr":
    case "mistral3":
    case "qwen2_5_vl":
    case "qwen3_vl":
    case "qwen3_vl_moe":
      init_normalized_config = getNormalizedConfig(config.text_config);
      break;
    case "moondream1":
      init_normalized_config = getNormalizedConfig(config.phi_config);
      break;
    case "musicgen":
      init_normalized_config = getNormalizedConfig(config.decoder);
      break;
    case "multi_modality":
      init_normalized_config = getNormalizedConfig(config.language_config);
      break;
    // Decoder-only models
    case "gpt2":
    case "gptj":
    case "jais":
    case "codegen":
    case "gpt_bigcode":
      mapping["num_heads"] = "n_head";
      mapping["num_layers"] = "n_layer";
      mapping["hidden_size"] = "n_embd";
      break;
    case "gpt_neox":
    case "stablelm":
    case "opt":
    case "falcon":
    case "modernbert-decoder":
      mapping["num_heads"] = "num_attention_heads";
      mapping["num_layers"] = "num_hidden_layers";
      mapping["hidden_size"] = "hidden_size";
      break;
    case "gpt_oss":
    case "llama":
    case "llama4_text":
    case "nanochat":
    case "apertus":
    case "arcee":
    case "afmoe":
    case "lfm2":
    case "lfm2_moe":
    case "smollm3":
    case "olmo":
    case "olmo2":
    case "olmo3":
    case "mobilellm":
    case "granite":
    case "granitemoehybrid":
    case "cohere":
    case "cohere2":
    case "mistral":
    case "voxtral_realtime_text":
    case "voxtral_realtime_encoder":
    case "starcoder2":
    case "qwen2":
    case "qwen2_moe":
    case "qwen2_vl":
    case "qwen2_vl_text":
    case "qwen2_5_vl_text":
    case "qwen3_moe":
    case "qwen3_vl_text":
    case "qwen3_vl_moe_text":
    case "phi":
    case "phi3":
    case "phi3_v":
    case "llava_qwen2":
      mapping["num_heads"] = "num_key_value_heads";
      mapping["num_layers"] = "num_hidden_layers";
      mapping["hidden_size"] = "hidden_size";
      mapping["num_attention_heads"] = "num_attention_heads";
      mapping["dim_kv"] = "head_dim";
      break;
    case "qwen3":
    case "solar_open":
    case "glm_ocr_text":
    case "gemma":
    case "gemma2":
    case "vaultgemma":
    case "gemma3_text":
    case "gemma3n_text":
    case "gemma4_text":
    case "glm":
    case "helium":
    case "ernie4_5":
    case "hunyuan_v1_dense":
    case "falcon_h1":
    case "nemotron_h":
    case "ministral":
    case "ministral3":
      mapping["num_heads"] = "num_key_value_heads";
      mapping["num_layers"] = "num_hidden_layers";
      mapping["dim_kv"] = "head_dim";
      break;
    case "openelm":
      mapping["num_heads"] = "num_kv_heads";
      mapping["num_layers"] = "num_transformer_layers";
      mapping["dim_kv"] = "head_dim";
      break;
    case "gpt_neo":
    case "donut-swin":
      mapping["num_heads"] = "num_heads";
      mapping["num_layers"] = "num_layers";
      mapping["hidden_size"] = "hidden_size";
      break;
    case "bloom":
      mapping["num_heads"] = "n_head";
      mapping["num_layers"] = "n_layer";
      mapping["hidden_size"] = "hidden_size";
      break;
    case "mpt":
      mapping["num_heads"] = "n_heads";
      mapping["num_layers"] = "n_layers";
      mapping["hidden_size"] = "d_model";
      break;
    case "exaone":
      mapping["num_heads"] = "num_key_value_heads";
      mapping["num_layers"] = "num_layers";
      mapping["dim_kv"] = "head_dim";
      mapping["num_attention_heads"] = "num_attention_heads";
      break;
    case "youtu":
    case "deepseek_v3":
    case "glm_moe_dsa":
    case "mistral4":
      mapping["num_heads"] = "num_key_value_heads";
      mapping["num_layers"] = "num_hidden_layers";
      mapping["dim_kv"] = "qk_head_dim";
      mapping["num_attention_heads"] = "num_attention_heads";
      break;
    // Encoder-decoder models
    case "t5":
    case "mt5":
    case "longt5":
      mapping["num_decoder_layers"] = "num_decoder_layers";
      mapping["num_decoder_heads"] = "num_heads";
      mapping["decoder_dim_kv"] = "d_kv";
      mapping["num_encoder_layers"] = "num_layers";
      mapping["num_encoder_heads"] = "num_heads";
      mapping["encoder_dim_kv"] = "d_kv";
      break;
    case "bart":
    case "mbart":
    case "marian":
    case "whisper":
    case "lite-whisper":
    case "m2m_100":
    case "blenderbot":
    case "blenderbot-small":
    case "florence2_language":
      mapping["num_decoder_layers"] = "decoder_layers";
      mapping["num_decoder_heads"] = "decoder_attention_heads";
      mapping["decoder_hidden_size"] = "d_model";
      mapping["num_encoder_layers"] = "encoder_layers";
      mapping["num_encoder_heads"] = "encoder_attention_heads";
      mapping["encoder_hidden_size"] = "d_model";
      break;
    case "speecht5":
      mapping["num_decoder_layers"] = "decoder_layers";
      mapping["num_decoder_heads"] = "decoder_attention_heads";
      mapping["decoder_hidden_size"] = "hidden_size";
      mapping["num_encoder_layers"] = "encoder_layers";
      mapping["num_encoder_heads"] = "encoder_attention_heads";
      mapping["encoder_hidden_size"] = "hidden_size";
      break;
    case "trocr":
      mapping["num_encoder_layers"] = mapping["num_decoder_layers"] = "decoder_layers";
      mapping["num_encoder_heads"] = mapping["num_decoder_heads"] = "decoder_attention_heads";
      mapping["encoder_hidden_size"] = mapping["decoder_hidden_size"] = "d_model";
      break;
    case "musicgen_decoder":
      mapping["num_encoder_layers"] = mapping["num_decoder_layers"] = "num_hidden_layers";
      mapping["num_encoder_heads"] = mapping["num_decoder_heads"] = "num_attention_heads";
      mapping["encoder_hidden_size"] = mapping["decoder_hidden_size"] = "hidden_size";
      break;
    case "moonshine":
      mapping["num_decoder_layers"] = "decoder_num_hidden_layers";
      mapping["num_decoder_heads"] = "decoder_num_key_value_heads";
      mapping["num_encoder_layers"] = "encoder_num_hidden_layers";
      mapping["num_encoder_heads"] = "encoder_num_key_value_heads";
      mapping["encoder_hidden_size"] = mapping["decoder_hidden_size"] = "hidden_size";
      break;
    case "cohere_asr":
      mapping["num_decoder_layers"] = "num_hidden_layers";
      mapping["num_decoder_heads"] = "num_key_value_heads";
      mapping["decoder_hidden_size"] = "hidden_size";
      mapping["decoder_dim_kv"] = "head_dim";
      const {
        num_hidden_layers: num_encoder_layers,
        num_attention_heads: num_encoder_heads,
        hidden_size: encoder_hidden_size
      } = (
        /** @type {any} */
        config.encoder_config
      );
      init_normalized_config = {
        num_encoder_layers,
        num_encoder_heads,
        encoder_hidden_size,
        // @ts-expect-error TS2339
        encoder_dim_kv: config.head_dim
      };
      break;
    case "vision-encoder-decoder":
      const decoderConfig = getNormalizedConfig(config.decoder);
      const add_encoder_pkv = "num_decoder_layers" in decoderConfig;
      const result = pick(config, ["model_type", "is_encoder_decoder"]);
      if (add_encoder_pkv) {
        result.num_decoder_layers = decoderConfig.num_decoder_layers;
        result.num_decoder_heads = decoderConfig.num_decoder_heads;
        result.decoder_hidden_size = decoderConfig.decoder_hidden_size;
        result.num_encoder_layers = decoderConfig.num_encoder_layers;
        result.num_encoder_heads = decoderConfig.num_encoder_heads;
        result.encoder_hidden_size = decoderConfig.encoder_hidden_size;
      } else {
        result.num_layers = decoderConfig.num_layers;
        result.num_heads = decoderConfig.num_heads;
        result.hidden_size = decoderConfig.hidden_size;
      }
      return result;
  }
  const normalized_config = {
    ...init_normalized_config,
    ...pick(config, ["model_type", "multi_query", "is_encoder_decoder"])
  };
  for (const key in mapping) {
    normalized_config[key] = config[mapping[key]];
  }
  return normalized_config;
}
function getCacheNames(config, options) {
  if (!(config instanceof PretrainedConfig)) {
    config = new PretrainedConfig(config);
  }
  const pkv_prefix = options?.prefix ?? "past_key_values";
  const conv_prefix = pkv_prefix === "present" ? "present" : "past";
  const names = /* @__PURE__ */ new Set();
  if (["lfm2", "lfm2_moe"].includes(config.model_type)) {
    const { layer_types } = (
      /** @type {any} */
      config
    );
    for (let i = 0; i < layer_types.length; ++i) {
      if (layer_types[i] === "full_attention") {
        names.add(`${pkv_prefix}.${i}.key`);
        names.add(`${pkv_prefix}.${i}.value`);
      } else if (layer_types[i] === "conv") {
        names.add(`${conv_prefix}_conv.${i}`);
      } else {
        throw new Error(`Unsupported layer type: ${layer_types[i]}`);
      }
    }
    return names;
  } else if (["granitemoehybrid", "falcon_h1", "nemotron_h"].includes(config.model_type)) {
    const c = (
      /** @type {any} */
      config
    );
    const layer_types = c.layer_types ?? c.layers_block_type;
    const num_layers = c.num_hidden_layers ?? layer_types?.length;
    for (let i = 0; i < num_layers; ++i) {
      if (!layer_types || layer_types[i] === "mamba") {
        names.add(`${conv_prefix}_conv.${i}`);
        names.add(`${conv_prefix}_ssm.${i}`);
      }
      if (!layer_types || layer_types[i] === "attention") {
        names.add(`${pkv_prefix}.${i}.key`);
        names.add(`${pkv_prefix}.${i}.value`);
      }
    }
    return names;
  } else if (["qwen3_next", "qwen3_5_text", "qwen3_5_moe_text", "olmo_hybrid"].includes(config.model_type)) {
    const { layer_types } = (
      /** @type {any} */
      config
    );
    for (let i = 0; i < layer_types.length; ++i) {
      if (layer_types[i] === "full_attention") {
        names.add(`${pkv_prefix}.${i}.key`);
        names.add(`${pkv_prefix}.${i}.value`);
      } else if (layer_types[i] === "linear_attention") {
        if (config.model_type === "olmo_hybrid") {
          names.add(`${conv_prefix}_conv.${i}.key`);
          names.add(`${conv_prefix}_conv.${i}.value`);
          names.add(`${conv_prefix}_conv.${i}.query`);
        } else {
          names.add(`${conv_prefix}_conv.${i}`);
        }
        names.add(`${conv_prefix}_recurrent.${i}`);
      } else {
        throw new Error(`Unsupported layer type: ${layer_types[i]}`);
      }
    }
    return names;
  } else if (["gemma4", "gemma4_text"].includes(config.model_type)) {
    const c = (
      /** @type {any} */
      config.model_type === "gemma4" ? (
        /** @type {any} */
        config.text_config
      ) : config
    );
    const num_hidden_layers = c.num_hidden_layers;
    const num_kv_shared_layers = c.num_kv_shared_layers ?? 0;
    const num_kv_layers = num_hidden_layers - num_kv_shared_layers;
    for (let i = 0; i < num_kv_layers; ++i) {
      names.add(`${pkv_prefix}.${i}.key`);
      names.add(`${pkv_prefix}.${i}.value`);
    }
    return names;
  } else if (["lfm2_vl", "qwen3_5", "qwen3_5_moe", "voxtral_realtime"].includes(config.model_type)) {
    let subConfig;
    if (config.model_type === "voxtral_realtime" && options?.session_name === "audio_encoder") {
      subConfig = /** @type {any} */
      config.audio_config;
    } else {
      subConfig = /** @type {any} */
      config.text_config;
    }
    return getCacheNames(subConfig, options);
  }
  return getKeyValueNames(config, { prefix: pkv_prefix });
}
function getKeyValueNames(config, { prefix = "past_key_values" } = {}) {
  const names = /* @__PURE__ */ new Set();
  const normalized_config = config.normalized_config;
  if (normalized_config.is_encoder_decoder && "num_encoder_heads" in normalized_config && "num_decoder_heads" in normalized_config) {
    for (let i = 0; i < normalized_config.num_decoder_layers; ++i) {
      names.add(`${prefix}.${i}.encoder.key`);
      names.add(`${prefix}.${i}.encoder.value`);
      names.add(`${prefix}.${i}.decoder.key`);
      names.add(`${prefix}.${i}.decoder.value`);
    }
  } else if (normalized_config.multi_query) {
    for (let i = 0; i < normalized_config.num_layers; ++i) {
      names.add(`${prefix}.${i}.key_value`);
    }
  } else {
    for (let i = 0; i < normalized_config.num_layers; ++i) {
      names.add(`${prefix}.${i}.key`);
      names.add(`${prefix}.${i}.value`);
    }
  }
  return names;
}
var PretrainedConfig = class _PretrainedConfig {
  // NOTE: Typo in original
  /** @type {string|null} */
  model_type = null;
  /** @type {boolean} */
  is_encoder_decoder = false;
  /** @type {number} */
  max_position_embeddings;
  /** @type {TransformersJSConfig} */
  "transformers.js_config";
  /**
   * Create a new PreTrainedTokenizer instance.
   * @param {Object} configJSON The JSON of the config.
   */
  constructor(configJSON) {
    Object.assign(this, configJSON);
    this.normalized_config = getNormalizedConfig(this);
  }
  /**
   * Loads a pre-trained config from the given `pretrained_model_name_or_path`.
   *
   * @param {string} pretrained_model_name_or_path The path to the pre-trained config.
   * @param {PretrainedOptions} options Additional options for loading the config.
   * @throws {Error} Throws an error if the config.json is not found in the `pretrained_model_name_or_path`.
   *
   * @returns {Promise<PretrainedConfig>} A new instance of the `PretrainedConfig` class.
   */
  static async from_pretrained(pretrained_model_name_or_path, { progress_callback = null, config = null, cache_dir = null, local_files_only = false, revision = "main" } = {}) {
    if (config && !(config instanceof _PretrainedConfig)) {
      config = new _PretrainedConfig(config);
    }
    const data = config ?? await loadConfig(pretrained_model_name_or_path, {
      progress_callback,
      config,
      cache_dir,
      local_files_only,
      revision
    });
    return new this(data);
  }
};
var AutoConfig = class {
  /** @type {typeof PretrainedConfig.from_pretrained} */
  static async from_pretrained(...args) {
    return PretrainedConfig.from_pretrained(...args);
  }
};

// src/utils/model-loader.js
function resolveExternalDataFormat(config, fullName, fileName) {
  if (!config) return 0;
  if (typeof config === "object" && config !== null) {
    if (config.hasOwnProperty(fullName)) return +config[fullName];
    if (config.hasOwnProperty(fileName)) return +config[fileName];
    return 0;
  }
  return +config;
}
function getExternalDataChunkNames(fullName, numChunks) {
  const names = [];
  for (let i = 0; i < numChunks; ++i) {
    names.push(`${fullName}_data${i === 0 ? "" : "_" + i}`);
  }
  return names;
}
async function getCoreModelFile(pretrained_model_name_or_path, fileName, options, suffix) {
  const baseName = `${fileName}${suffix}.onnx`;
  const fullPath = `${options.subfolder ?? ""}/${baseName}`;
  return await getModelFile(pretrained_model_name_or_path, fullPath, true, options, apis.IS_NODE_ENV);
}
async function getModelDataFiles(pretrained_model_name_or_path, fileName, suffix, options, use_external_data_format, session_options = {}) {
  const baseName = `${fileName}${suffix}.onnx`;
  const return_path = apis.IS_NODE_ENV;
  let externalDataPromises = [];
  const num_chunks = resolveExternalDataFormat(use_external_data_format, baseName, fileName);
  if (num_chunks > 0) {
    if (num_chunks > MAX_EXTERNAL_DATA_CHUNKS) {
      throw new Error(
        `The number of external data chunks (${num_chunks}) exceeds the maximum allowed value (${MAX_EXTERNAL_DATA_CHUNKS}).`
      );
    }
    const chunkNames = getExternalDataChunkNames(baseName, num_chunks);
    for (const path3 of chunkNames) {
      const fullPath = `${options.subfolder ?? ""}/${path3}`;
      externalDataPromises.push(
        new Promise(async (resolve, reject) => {
          const data = await getModelFile(
            pretrained_model_name_or_path,
            fullPath,
            true,
            options,
            return_path
          );
          resolve(data instanceof Uint8Array ? { path: path3, data } : path3);
        })
      );
    }
  } else if (session_options.externalData !== void 0) {
    externalDataPromises = session_options.externalData.map(async (ext) => {
      if (typeof ext.data === "string") {
        const ext_buffer = await getModelFile(pretrained_model_name_or_path, ext.data, true, options);
        return { ...ext, data: ext_buffer };
      }
      return ext;
    });
  }
  return Promise.all(externalDataPromises);
}

// src/models/session.js
async function getSession(pretrained_model_name_or_path, fileName, options, cache_config = false, session_name = void 0) {
  let custom_config = options.config?.["transformers.js_config"] ?? {};
  const selectedDevice = (
    /** @type {import("../utils/devices.js").DeviceType} */
    selectDevice(options.device ?? custom_config.device, fileName, {
      warn: (msg) => logger.info(msg)
    })
  );
  const executionProviders = deviceToExecutionProviders(selectedDevice);
  const device_config = custom_config.device_config ?? {};
  if (device_config.hasOwnProperty(selectedDevice)) {
    custom_config = {
      ...custom_config,
      ...device_config[selectedDevice]
    };
  }
  const selectedDtype = (
    /** @type {import("../utils/dtypes.js").DataType} */
    selectDtype(options.dtype ?? custom_config.dtype, fileName, selectedDevice, {
      configDtype: custom_config.dtype,
      warn: (msg) => logger.info(msg)
    })
  );
  if (!DEFAULT_DTYPE_SUFFIX_MAPPING.hasOwnProperty(selectedDtype)) {
    throw new Error(`Invalid dtype: ${selectedDtype}. Should be one of: ${Object.keys(DATA_TYPES).join(", ")}`);
  } else if (selectedDevice === "webgpu" && // NOTE: Currently, we assume that the Native WebGPU EP always supports fp16. In future, we will add a check for this.
  !apis.IS_NODE_ENV && selectedDtype === DATA_TYPES.fp16 && !await isWebGpuFp16Supported()) {
    throw new Error(`The device (${selectedDevice}) does not support fp16.`);
  }
  const suffix = DEFAULT_DTYPE_SUFFIX_MAPPING[selectedDtype];
  const session_options = { ...options.session_options };
  session_options.executionProviders ??= executionProviders;
  const free_dimension_overrides = custom_config.free_dimension_overrides;
  if (free_dimension_overrides) {
    session_options.freeDimensionOverrides ??= free_dimension_overrides;
  } else if (selectedDevice.startsWith("webnn") && !session_options.freeDimensionOverrides) {
    logger.warn(
      `WebNN does not currently support dynamic shapes and requires 'free_dimension_overrides' to be set in config.json, preferably as a field within config["transformers.js_config"]["device_config"]["${selectedDevice}"]. When 'free_dimension_overrides' is not set, you may experience significant performance degradation.`
    );
  }
  const bufferOrPathPromise = getCoreModelFile(pretrained_model_name_or_path, fileName, options, suffix);
  const use_external_data_format = options.use_external_data_format ?? custom_config.use_external_data_format;
  const externalData = await getModelDataFiles(
    pretrained_model_name_or_path,
    fileName,
    suffix,
    options,
    use_external_data_format,
    session_options
  );
  if (externalData.length > 0 && (!apis.IS_NODE_ENV || externalData.some((data) => typeof data !== "string"))) {
    session_options.externalData = externalData;
  }
  if (cache_config && selectedDevice === "webgpu") {
    const names = getCacheNames(options.config, {
      prefix: "present",
      session_name
    });
    if (names.size > 0 && !isONNXProxy()) {
      const preferredOutputLocation = {};
      for (const key of names) {
        preferredOutputLocation[key] = "gpu-buffer";
      }
      session_options.preferredOutputLocation = preferredOutputLocation;
    }
  }
  const buffer_or_path = await bufferOrPathPromise;
  const session_config = {
    dtype: selectedDtype,
    device: selectedDevice
  };
  return { buffer_or_path, session_options, session_config };
}
async function constructSessions(pretrained_model_name_or_path, names, options, cache_sessions = void 0) {
  return Object.fromEntries(
    await Promise.all(
      Object.keys(names).map(async (name) => {
        const cache_config = cache_sessions?.[name] ?? false;
        const { buffer_or_path, session_options, session_config } = await getSession(
          pretrained_model_name_or_path,
          names[name],
          options,
          cache_config,
          name
        );
        const session = await createInferenceSession(buffer_or_path, session_options, session_config);
        return [name, session];
      })
    )
  );
}
function replaceTensors(obj) {
  for (let prop in obj) {
    if (isONNXTensor(obj[prop])) {
      obj[prop] = new Tensor2(obj[prop]);
    } else if (typeof obj[prop] === "object") {
      replaceTensors(obj[prop]);
    }
  }
  return obj;
}
async function sessionRun(session, inputs) {
  const checkedInputs = validateInputs(session, inputs);
  try {
    const ortFeed = Object.fromEntries(
      Object.entries(checkedInputs).map(([k2, v]) => {
        const tensor = (
          /** @type {any} */
          v.ort_tensor
        );
        if (apis.IS_NODE_ENV) {
          if (typeof Float16Array !== "undefined" && tensor.cpuData instanceof Float16Array) {
            tensor.cpuData = new Uint16Array(tensor.cpuData.buffer);
          }
        }
        return [k2, tensor];
      })
    );
    const output = await runInferenceSession(session, ortFeed);
    return replaceTensors(output);
  } catch (e) {
    const formatted = Object.fromEntries(
      Object.entries(checkedInputs).map(([k2, tensor]) => {
        const unpacked = {
          type: tensor.type,
          dims: tensor.dims,
          location: tensor.location
        };
        if (unpacked.location !== "gpu-buffer") {
          unpacked.data = tensor.data;
        }
        return [k2, unpacked];
      })
    );
    logger.error(`An error occurred during model execution: "${e}".`);
    logger.error("Inputs given to model:", formatted);
    throw e;
  }
}
function validateInputs(session, inputs) {
  const checkedInputs = /* @__PURE__ */ Object.create(null);
  const missingInputs = [];
  for (const inputName of session.inputNames) {
    const tensor = inputs[inputName];
    if (!(tensor instanceof Tensor2)) {
      missingInputs.push(inputName);
      continue;
    }
    checkedInputs[inputName] = isONNXProxy() ? tensor.clone() : tensor;
  }
  if (missingInputs.length > 0) {
    throw new Error(
      `An error occurred during model execution: "Missing the following inputs: ${missingInputs.join(", ")}.`
    );
  }
  const numInputsProvided = Object.keys(inputs).length;
  const numInputsNeeded = session.inputNames.length;
  if (numInputsProvided > numInputsNeeded) {
    let ignored = Object.keys(inputs).filter((inputName) => !session.inputNames.includes(inputName));
    logger.warn(
      `WARNING: Too many inputs were provided (${numInputsProvided} > ${numInputsNeeded}). The following inputs will be ignored: "${ignored.join(", ")}".`
    );
  }
  return checkedInputs;
}

// src/models/modeling_outputs.js
var ModelOutput = class {
};
var SequenceClassifierOutput = class extends ModelOutput {
  /**
   * @param {Object} output The output of the model.
   * @param {Tensor} output.logits classification (or regression if config.num_labels==1) scores (before SoftMax).
   * @param {Record<string, Tensor>} [output.attentions] Object of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, sequence_length)`.
   * Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
   */
  constructor({ logits, ...attentions }) {
    super();
    this.logits = logits;
    const attentions_list = Object.values(attentions);
    if (attentions_list.length > 0) {
      this.attentions = attentions_list;
    }
  }
};
var TokenClassifierOutput = class extends ModelOutput {
  /**
   * @param {Object} output The output of the model.
   * @param {Tensor} output.logits Classification scores (before SoftMax).
   */
  constructor({ logits }) {
    super();
    this.logits = logits;
  }
};
var MaskedLMOutput = class extends ModelOutput {
  /**
   * @param {Object} output The output of the model.
   * @param {Tensor} output.logits Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
   */
  constructor({ logits }) {
    super();
    this.logits = logits;
  }
};
var QuestionAnsweringModelOutput = class extends ModelOutput {
  /**
   * @param {Object} output The output of the model.
   * @param {Tensor} output.start_logits Span-start scores (before SoftMax).
   * @param {Tensor} output.end_logits Span-end scores (before SoftMax).
   */
  constructor({ start_logits, end_logits }) {
    super();
    this.start_logits = start_logits;
    this.end_logits = end_logits;
  }
};
var CausalLMOutput = class extends ModelOutput {
  /**
   * @param {Object} output The output of the model.
   * @param {Tensor} output.logits Prediction scores of the language modeling head (scores for each vocabulary token before softmax).
   */
  constructor({ logits }) {
    super();
    this.logits = logits;
  }
};
var ImageMattingOutput = class extends ModelOutput {
  /**
   * @param {Object} output The output of the model.
   * @param {Tensor} output.alphas Estimated alpha values, of shape `(batch_size, num_channels, height, width)`.
   */
  constructor({ alphas }) {
    super();
    this.alphas = alphas;
  }
};

// src/generation/logits_process.js
var LogitsProcessor = class extends Callable {
  /**
   * Apply the processor to the input logits.
   *
   * @abstract
   * @param {bigint[][]} input_ids The input ids.
   * @param {Tensor} logits The logits to process.
   * @throws {Error} Throws an error if `_call` is not implemented in the subclass.
   */
  _call(input_ids, logits) {
    throw Error("`_call` should be implemented in a subclass");
  }
};
var LogitsWarper = class extends Callable {
  /**
   * Apply the processor to the input logits.
   *
   * @abstract
   * @param {bigint[][]} input_ids The input ids.
   * @param {Tensor} logits The logits to process.
   * @throws {Error} Throws an error if `_call` is not implemented in the subclass.
   */
  _call(input_ids, logits) {
    throw Error("`_call` should be implemented in a subclass");
  }
};
var LogitsProcessorList = class extends Callable {
  /**
   * Constructs a new instance of `LogitsProcessorList`.
   */
  constructor() {
    super();
    this.processors = [];
  }
  /**
   * Adds a new logits processor to the list.
   *
   * @param {LogitsProcessor} item The logits processor function to add.
   */
  push(item) {
    this.processors.push(item);
  }
  /**
   * Adds multiple logits processors to the list.
   *
   * @param {LogitsProcessor[]} items The logits processor functions to add.
   */
  extend(items) {
    this.processors.push(...items);
  }
  /**
   * Applies all logits processors in the list to a batch of logits, modifying them in-place.
   *
   * @param {bigint[][]} input_ids The input IDs for the language model.
   * @param {Tensor} logits
   */
  _call(input_ids, logits) {
    let toReturn = logits;
    for (const processor of this.processors) {
      toReturn = processor(input_ids, toReturn);
    }
    return toReturn;
  }
  [Symbol.iterator]() {
    return this.processors.values();
  }
};
var ForcedBOSTokenLogitsProcessor = class extends LogitsProcessor {
  /**
   * Create a ForcedBOSTokenLogitsProcessor.
   * @param {number} bos_token_id The ID of the beginning-of-sequence token to be forced.
   */
  constructor(bos_token_id) {
    super();
    this.bos_token_id = bos_token_id;
  }
  /**
   * Apply the BOS token forcing to the logits.
   * @param {bigint[][]} input_ids The input IDs.
   * @param {Tensor} logits The logits.
   * @returns {Tensor} The logits with BOS token forcing.
   */
  _call(input_ids, logits) {
    for (let i = 0; i < input_ids.length; ++i) {
      if (input_ids[i].length === 1) {
        const batch_logits_data = (
          /** @type {Float32Array} */
          logits[i].data
        );
        batch_logits_data.fill(-Infinity);
        batch_logits_data[this.bos_token_id] = 0;
      }
    }
    return logits;
  }
};
var ForcedEOSTokenLogitsProcessor = class extends LogitsProcessor {
  /**
   * Create a ForcedEOSTokenLogitsProcessor.
   * @param {number} max_length The maximum length of the sequence to be generated.
   * @param {number|number[]} eos_token_id The id(s) of the *end-of-sequence* token.
   */
  constructor(max_length, eos_token_id) {
    super();
    this.max_length = max_length;
    this.eos_token_id = Array.isArray(eos_token_id) ? eos_token_id : [eos_token_id];
  }
  /**
   * Apply the processor to input_ids and logits.
   *
   * @param {bigint[][]} input_ids The input ids.
   * @param {Tensor} logits The logits tensor.
   */
  _call(input_ids, logits) {
    for (let i = 0; i < input_ids.length; ++i) {
      if (input_ids[i].length === this.max_length - 1) {
        const batch_logits_data = (
          /** @type {Float32Array} */
          logits[i].data
        );
        batch_logits_data.fill(-Infinity);
        for (const eos_token of this.eos_token_id) {
          batch_logits_data[eos_token] = 0;
        }
      }
    }
    return logits;
  }
};
var SuppressTokensLogitsProcessor = class extends LogitsProcessor {
  /**
   * Create a SuppressTokensLogitsProcessor.
   * @param {number[]} suppress_tokens The IDs of the tokens to suppress.
   */
  constructor(suppress_tokens) {
    super();
    this.suppress_tokens = suppress_tokens;
  }
  /**
   * Suppress the specified tokens by setting their logits to -Infinity.
   * @param {bigint[][]} input_ids The input IDs.
   * @param {Tensor} logits The logits.
   * @returns {Tensor} The modified logits.
   */
  _call(input_ids, logits) {
    for (let i = 0; i < input_ids.length; ++i) {
      const batch_logits_data = (
        /** @type {Float32Array} */
        logits[i].data
      );
      for (const token_id of this.suppress_tokens) {
        batch_logits_data[token_id] = -Infinity;
      }
    }
    return logits;
  }
};
var SuppressTokensAtBeginLogitsProcessor = class extends LogitsProcessor {
  /**
   * Create a SuppressTokensAtBeginLogitsProcessor.
   * @param {number[]} begin_suppress_tokens The IDs of the tokens to suppress.
   * @param {number} begin_index The number of tokens to generate before suppressing tokens.
   */
  constructor(begin_suppress_tokens, begin_index) {
    super();
    this.begin_suppress_tokens = begin_suppress_tokens;
    this.begin_index = begin_index;
  }
  /**
   * Apply the BOS token forcing to the logits.
   * @param {bigint[][]} input_ids The input IDs.
   * @param {Tensor} logits The logits.
   * @returns {Tensor} The logits with BOS token forcing.
   */
  _call(input_ids, logits) {
    for (let i = 0; i < input_ids.length; ++i) {
      if (input_ids[i].length === this.begin_index) {
        const batch_logits_data = (
          /** @type {Float32Array} */
          logits[i].data
        );
        for (const token_id of this.begin_suppress_tokens) {
          batch_logits_data[token_id] = -Infinity;
        }
      }
    }
    return logits;
  }
};
var WhisperTimeStampLogitsProcessor = class extends LogitsProcessor {
  /**
   * Constructs a new WhisperTimeStampLogitsProcessor.
   * @param {import('../models/whisper/generation_whisper.js').WhisperGenerationConfig} generate_config The config object passed to the `generate()` method of a transformer model.
   * @param {number[]} init_tokens The initial tokens of the input sequence.
   */
  constructor(generate_config, init_tokens) {
    super();
    this.eos_token_id = Array.isArray(generate_config.eos_token_id) ? generate_config.eos_token_id[0] : generate_config.eos_token_id;
    this.no_timestamps_token_id = generate_config.no_timestamps_token_id;
    this.timestamp_begin = this.no_timestamps_token_id + 1;
    this.begin_index = init_tokens.length;
    if (init_tokens.at(-1) === this.no_timestamps_token_id) {
      this.begin_index -= 1;
    }
    this.max_initial_timestamp_index = generate_config.max_initial_timestamp_index;
  }
  /**
   * Modify the logits to handle timestamp tokens.
   * @param {bigint[][]} input_ids The input sequence of tokens.
   * @param {Tensor} logits The logits output by the model.
   * @returns {Tensor} The modified logits.
   */
  _call(input_ids, logits) {
    for (let i = 0; i < input_ids.length; ++i) {
      const batch_logits_data = (
        /** @type {Float32Array} */
        logits[i].data
      );
      batch_logits_data[this.no_timestamps_token_id] = -Infinity;
      if (input_ids[i].length === this.begin_index) {
        batch_logits_data.subarray(0, this.timestamp_begin).fill(-Infinity);
        continue;
      }
      const seq = input_ids[i].slice(this.begin_index);
      const last_was_timestamp = seq.length >= 1 && seq[seq.length - 1] >= this.timestamp_begin;
      const penultimate_was_timestamp = seq.length < 2 || seq[seq.length - 2] >= this.timestamp_begin;
      if (last_was_timestamp) {
        if (penultimate_was_timestamp) {
          batch_logits_data.subarray(this.timestamp_begin).fill(-Infinity);
        } else {
          batch_logits_data.subarray(0, this.eos_token_id).fill(-Infinity);
        }
      }
      if (input_ids[i].length === this.begin_index && this.max_initial_timestamp_index !== null) {
        const last_allowed = this.timestamp_begin + this.max_initial_timestamp_index;
        batch_logits_data.subarray(last_allowed + 1).fill(-Infinity);
      }
      const logprobs = log_softmax(batch_logits_data);
      const timestamp_logprob = Math.log(
        logprobs.subarray(this.timestamp_begin).map(Math.exp).reduce((a, b) => a + b)
      );
      const max_text_token_logprob = max(logprobs.subarray(0, this.timestamp_begin))[0];
      if (timestamp_logprob > max_text_token_logprob) {
        batch_logits_data.subarray(0, this.timestamp_begin).fill(-Infinity);
      }
    }
    return logits;
  }
};
var NoRepeatNGramLogitsProcessor = class extends LogitsProcessor {
  /**
   * Create a NoRepeatNGramLogitsProcessor.
   * @param {number} no_repeat_ngram_size The no-repeat-ngram size. All ngrams of this size can only occur once.
   */
  constructor(no_repeat_ngram_size) {
    super();
    this.no_repeat_ngram_size = no_repeat_ngram_size;
  }
  /**
   * Generate n-grams from a sequence of token ids.
   * @param {bigint[]} prevInputIds List of previous input ids
   * @returns {Map<string, number[]>} Map of generated n-grams
   */
  getNgrams(prevInputIds) {
    const curLen = prevInputIds.length;
    const ngrams = [];
    for (let j = 0; j < curLen + 1 - this.no_repeat_ngram_size; ++j) {
      const ngram = [];
      for (let k2 = 0; k2 < this.no_repeat_ngram_size; ++k2) {
        ngram.push(prevInputIds[j + k2]);
      }
      ngrams.push(ngram.map(Number));
    }
    const generatedNgram = /* @__PURE__ */ new Map();
    for (const ngram of ngrams) {
      const prevNgram = ngram.slice(0, ngram.length - 1);
      const prevNgramKey = JSON.stringify(prevNgram);
      const prevNgramValue = generatedNgram.get(prevNgramKey) ?? [];
      prevNgramValue.push(ngram[ngram.length - 1]);
      generatedNgram.set(prevNgramKey, prevNgramValue);
    }
    return generatedNgram;
  }
  /**
   * Generate n-grams from a sequence of token ids.
   * @param {Map<string, number[]>} bannedNgrams Map of banned n-grams
   * @param {bigint[]} prevInputIds List of previous input ids
   * @returns {number[]} Map of generated n-grams
   */
  getGeneratedNgrams(bannedNgrams, prevInputIds) {
    const ngramIdx = prevInputIds.slice(prevInputIds.length + 1 - this.no_repeat_ngram_size, prevInputIds.length);
    const banned = bannedNgrams.get(JSON.stringify(ngramIdx.map(Number))) ?? [];
    return banned;
  }
  /**
   * Calculate banned n-gram tokens
   * @param {bigint[]} prevInputIds List of previous input ids
   * @returns {number[]} Map of generated n-grams
   */
  calcBannedNgramTokens(prevInputIds) {
    const bannedTokens = [];
    if (prevInputIds.length + 1 < this.no_repeat_ngram_size) {
      return bannedTokens;
    } else {
      const generatedNgrams = this.getNgrams(prevInputIds);
      const bannedTokens2 = this.getGeneratedNgrams(generatedNgrams, prevInputIds);
      return bannedTokens2;
    }
  }
  /**
   * Apply the no-repeat-ngram processor to the logits.
   * @param {bigint[][]} input_ids The input IDs.
   * @param {Tensor} logits The logits.
   * @returns {Tensor} The logits with no-repeat-ngram processing.
   */
  _call(input_ids, logits) {
    for (let i = 0; i < input_ids.length; ++i) {
      const batch_logits_data = (
        /** @type {Float32Array} */
        logits[i].data
      );
      const bannedTokens = this.calcBannedNgramTokens(input_ids[i]);
      for (const token of bannedTokens) {
        batch_logits_data[token] = -Infinity;
      }
    }
    return logits;
  }
};
var RepetitionPenaltyLogitsProcessor = class extends LogitsProcessor {
  /**
   * Create a RepetitionPenaltyLogitsProcessor.
   * @param {number} penalty The parameter for repetition penalty.
   * - 1.0 means no penalty. Above 1.0 penalizes previously generated tokens.
   * - Between 0.0 and 1.0 rewards previously generated tokens.
   */
  constructor(penalty) {
    super();
    this.penalty = penalty;
  }
  /**
   * Apply the repetition penalty to the logits.
   * @param {bigint[][]} input_ids The input IDs.
   * @param {Tensor} logits The logits.
   * @returns {Tensor} The logits with repetition penalty processing.
   */
  _call(input_ids, logits) {
    for (let i = 0; i < input_ids.length; ++i) {
      const batch_logits_data = (
        /** @type {Float32Array} */
        logits[i].data
      );
      for (const input_id of new Set(input_ids[i])) {
        const token = Number(input_id);
        if (batch_logits_data[token] < 0) {
          batch_logits_data[token] *= this.penalty;
        } else {
          batch_logits_data[token] /= this.penalty;
        }
      }
    }
    return logits;
  }
};
var MinLengthLogitsProcessor = class extends LogitsProcessor {
  /**
   * Create a MinLengthLogitsProcessor.
   * @param {number} min_length The minimum length below which the score of `eos_token_id` is set to negative infinity.
   * @param {number|number[]} eos_token_id The ID/IDs of the end-of-sequence token.
   */
  constructor(min_length, eos_token_id) {
    super();
    this.min_length = min_length;
    this.eos_token_id = Array.isArray(eos_token_id) ? eos_token_id : [eos_token_id];
  }
  /**
   * Apply logit processor.
   * @param {bigint[][]} input_ids The input IDs.
   * @param {Tensor} logits The logits.
   * @returns {Tensor} The processed logits.
   */
  _call(input_ids, logits) {
    for (let i = 0; i < input_ids.length; ++i) {
      if (input_ids[i].length < this.min_length) {
        const batch_logits_data = (
          /** @type {Float32Array} */
          logits[i].data
        );
        for (const eos_token of this.eos_token_id) {
          batch_logits_data[eos_token] = -Infinity;
        }
      }
    }
    return logits;
  }
};
var MinNewTokensLengthLogitsProcessor = class extends LogitsProcessor {
  /**
   * Create a MinNewTokensLengthLogitsProcessor.
   * @param {number} prompt_length_to_skip The input tokens length.
   * @param {number} min_new_tokens The minimum *new* tokens length below which the score of `eos_token_id` is set to negative infinity.
   * @param {number|number[]} eos_token_id The ID/IDs of the end-of-sequence token.
   */
  constructor(prompt_length_to_skip, min_new_tokens, eos_token_id) {
    super();
    this.prompt_length_to_skip = prompt_length_to_skip;
    this.min_new_tokens = min_new_tokens;
    this.eos_token_id = Array.isArray(eos_token_id) ? eos_token_id : [eos_token_id];
  }
  /**
   * Apply logit processor.
   * @param {bigint[][]} input_ids The input IDs.
   * @param {Tensor} logits The logits.
   * @returns {Tensor} The processed logits.
   */
  _call(input_ids, logits) {
    for (let i = 0; i < input_ids.length; ++i) {
      const new_tokens_length = input_ids[i].length - this.prompt_length_to_skip;
      if (new_tokens_length < this.min_new_tokens) {
        const batch_logits_data = (
          /** @type {Float32Array} */
          logits[i].data
        );
        for (const eos_token of this.eos_token_id) {
          batch_logits_data[eos_token] = -Infinity;
        }
      }
    }
    return logits;
  }
};
var NoBadWordsLogitsProcessor = class extends LogitsProcessor {
  /**
   * Create a `NoBadWordsLogitsProcessor`.
   * @param {number[][]} bad_words_ids List of list of token ids that are not allowed to be generated.
   * @param {number|number[]} eos_token_id The id of the *end-of-sequence* token. Optionally, use a list to set multiple *end-of-sequence* tokens.
   */
  constructor(bad_words_ids, eos_token_id) {
    super();
    this.bad_words_ids = bad_words_ids;
    this.eos_token_id = Array.isArray(eos_token_id) ? eos_token_id : [eos_token_id];
  }
  /**
   * Apply logit processor.
   * @param {bigint[][]} input_ids The input IDs.
   * @param {Tensor} logits The logits.
   * @returns {Tensor} The processed logits.
   */
  _call(input_ids, logits) {
    for (let i = 0; i < input_ids.length; ++i) {
      const batch_logits_data = (
        /** @type {Float32Array} */
        logits[i].data
      );
      const ids = input_ids[i];
      for (const bad_word_ids of this.bad_words_ids) {
        if (ids.length < bad_word_ids.length - 1) continue;
        let mark = true;
        for (let j = 1; j <= bad_word_ids.length - 1; ++j) {
          if (bad_word_ids.at(-j - 1) != ids.at(-j)) {
            mark = false;
            break;
          }
        }
        if (mark) {
          batch_logits_data[bad_word_ids.at(-1)] = -Infinity;
        }
      }
    }
    return logits;
  }
};
var ClassifierFreeGuidanceLogitsProcessor = class extends LogitsProcessor {
  /**
   * Create a `ClassifierFreeGuidanceLogitsProcessor`.
   * @param {number} guidance_scale The guidance scale for classifier free guidance (CFG). CFG is enabled by setting `guidance_scale > 1`.
   * Higher guidance scale encourages the model to generate samples that are more closely linked to the input
   * prompt, usually at the expense of poorer quality.
   */
  constructor(guidance_scale) {
    super();
    if (guidance_scale <= 1) {
      throw new Error(
        `Require guidance scale >1 to use the classifier free guidance processor, got guidance scale ${guidance_scale}.`
      );
    }
    this.guidance_scale = guidance_scale;
  }
  /**
   * Apply logit processor.
   * @param {bigint[][]} input_ids The input IDs.
   * @param {Tensor} logits The logits.
   * @returns {Tensor} The processed logits.
   */
  _call(input_ids, logits) {
    if (logits.dims[0] !== 2 * input_ids.length) {
      throw new Error(
        `Logits should have twice the batch size of the input ids, the first half of batches corresponding to the conditional inputs, and the second half of batches corresponding to the unconditional inputs. Got batch size ${logits.dims[0]} for the logits and ${input_ids.length} for the input ids.`
      );
    }
    const unguided_bsz = input_ids.length;
    const cond_logits = logits.slice([0, unguided_bsz], null);
    const uncond_logits = logits.slice([unguided_bsz, logits.dims[0]], null);
    for (let i = 0; i < uncond_logits.data.length; ++i) {
      uncond_logits.data[i] += (cond_logits.data[i] - uncond_logits.data[i]) * this.guidance_scale;
    }
    return uncond_logits;
  }
};
var TemperatureLogitsWarper = class extends LogitsWarper {
  /**
   * Create a `TemperatureLogitsWarper`.
   * @param {number} temperature Strictly positive float value used to modulate the logits distribution.
   * A value smaller than `1` decreases randomness (and vice versa), with `0` being equivalent to shifting
   * all probability mass to the most likely token.
   */
  constructor(temperature) {
    super();
    if (typeof temperature !== "number" || temperature <= 0) {
      let errorMessage = `\`temperature\` (=${temperature}) must be a strictly positive float, otherwise your next token scores will be invalid.`;
      if (temperature === 0) {
        errorMessage += " If you're looking for greedy decoding strategies, set `do_sample=false`.";
      }
    }
    this.temperature = temperature;
  }
  /**
   * Apply logit warper.
   * @param {bigint[][]} input_ids The input IDs.
   * @param {Tensor} logits The logits.
   * @returns {Tensor} The processed logits.
   */
  _call(input_ids, logits) {
    const batch_logits_data = (
      /** @type {Float32Array} */
      logits.data
    );
    for (let i = 0; i < batch_logits_data.length; ++i) {
      batch_logits_data[i] /= this.temperature;
    }
    return logits;
  }
};
var TopPLogitsWarper = class extends LogitsWarper {
  /**
   * Create a `TopPLogitsWarper`.
   * @param {number} top_p If set to < 1, only the smallest set of most probable tokens with
   * probabilities that add up to `top_p` or higher are kept for generation.
   * @param {Object} options Additional options for the top-p sampling.
   * @param {number} [options.filter_value=-Infinity] All filtered values will be set to this float value.
   * @param {number} [options.min_tokens_to_keep=1] Minimum number of tokens that cannot be filtered.
   */
  constructor(top_p, { filter_value = -Infinity, min_tokens_to_keep = 1 } = {}) {
    super();
    if (top_p < 0 || top_p > 1) {
      throw new Error(`\`top_p\` must be a float > 0 and < 1, but is ${top_p}`);
    }
    if (!Number.isInteger(min_tokens_to_keep) || min_tokens_to_keep < 1) {
      throw new Error(`\`min_tokens_to_keep\` must be a positive integer, but is ${min_tokens_to_keep}`);
    }
    this.top_p = top_p;
    this.filter_value = filter_value;
    this.min_tokens_to_keep = min_tokens_to_keep;
  }
};
var TopKLogitsWarper = class extends LogitsWarper {
  /**
   * Create a `TopKLogitsWarper`.
   * @param {number} top_k If set to > 0, only the top `top_k` tokens are kept for generation.
   * @param {Object} options Additional options for the top-k sampling.
   * @param {number} [options.filter_value=-Infinity] All filtered values will be set to this float value.
   * @param {number} [options.min_tokens_to_keep=1] Minimum number of tokens that cannot be filtered.
   */
  constructor(top_k, { filter_value = -Infinity, min_tokens_to_keep = 1 } = {}) {
    super();
    if (!Number.isInteger(top_k) || top_k < 0) {
      throw new Error(`\`top_k\` must be a positive integer, but is ${top_k}`);
    }
    this.top_k = Math.max(top_k, min_tokens_to_keep);
    this.filter_value = filter_value;
  }
};

// src/generation/configuration_utils.js
var GenerationConfig = class {
  // Parameters that control the length of the output
  /**
   * The maximum length the generated tokens can have.
   * Corresponds to the length of the input prompt + `max_new_tokens`.
   * Its effect is overridden by `max_new_tokens`, if also set.
   * @type {number}
   * @default 20
   */
  max_length = 20;
  /**
   * The maximum numbers of tokens to generate, ignoring the number of tokens in the prompt.
   * @type {number}
   * @default null
   */
  max_new_tokens = null;
  /**
   * The minimum length of the sequence to be generated.
   * Corresponds to the length of the input prompt + `min_new_tokens`.
   * Its effect is overridden by `min_new_tokens`, if also set.
   * @type {number}
   * @default 0
   */
  min_length = 0;
  /**
   * The minimum numbers of tokens to generate, ignoring the number of tokens in the prompt.
   * @type {number}
   * @default null
   */
  min_new_tokens = null;
  /**
   * Controls the stopping condition for beam-based methods, like beam-search. It accepts the following values:
   * - `true`, where the generation stops as soon as there are `num_beams` complete candidates;
   * - `false`, where an heuristic is applied and the generation stops when is it very unlikely to find better candidates;
   * - `"never"`, where the beam search procedure only stops when there cannot be better candidates (canonical beam search algorithm).
   * @type {boolean|"never"}
   * @default false
   */
  early_stopping = false;
  /**
   * The maximum amount of time you allow the computation to run for in seconds.
   * Generation will still finish the current pass after allocated time has been passed.
   * @type {number}
   * @default null
   */
  max_time = null;
  // Parameters that control the generation strategy used
  /**
   * Whether or not to use sampling; use greedy decoding otherwise.
   * @type {boolean}
   * @default false
   */
  do_sample = false;
  /**
   * Number of beams for beam search. 1 means no beam search.
   * @type {number}
   * @default 1
   */
  num_beams = 1;
  /**
   * Number of groups to divide `num_beams` into in order to ensure diversity among different groups of beams.
   * See [this paper](https://huggingface.co/papers/1610.02424) for more details.
   * @type {number}
   * @default 1
   */
  num_beam_groups = 1;
  /**
   * The values balance the model confidence and the degeneration penalty in contrastive search decoding.
   * @type {number}
   * @default null
   */
  penalty_alpha = null;
  /**
   * Whether or not the model should use the past last key/values attentions (if applicable to the model) to speed up decoding.
   * @type {boolean}
   * @default true
   */
  use_cache = true;
  // Parameters for manipulation of the model output logits
  /**
   * The value used to modulate the next token probabilities.
   * @type {number}
   * @default 1.0
   */
  temperature = 1;
  /**
   * The number of highest probability vocabulary tokens to keep for top-k-filtering.
   * @type {number}
   * @default 50
   */
  top_k = 50;
  /**
   * If set to float < 1, only the smallest set of most probable tokens with probabilities that add up to `top_p` or higher are kept for generation.
   * @type {number}
   * @default 1.0
   */
  top_p = 1;
  /**
   * Local typicality measures how similar the conditional probability of predicting a target token next is to the expected conditional probability of predicting a random token next, given the partial text already generated.
   * If set to float < 1, the smallest set of the most locally typical tokens with probabilities that add up to `typical_p` or higher are kept for generation.
   * See [this paper](https://huggingface.co/papers/2202.00666) for more details.
   * @type {number}
   * @default 1.0
   */
  typical_p = 1;
  /**
   * If set to float strictly between 0 and 1, only tokens with a conditional probability greater than `epsilon_cutoff` will be sampled.
   * In the paper, suggested values range from 3e-4 to 9e-4, depending on the size of the model.
   * See [Truncation Sampling as Language Model Desmoothing](https://huggingface.co/papers/2210.15191) for more details.
   * @type {number}
   * @default 0.0
   */
  epsilon_cutoff = 0;
  /**
   * Eta sampling is a hybrid of locally typical sampling and epsilon sampling.
   * If set to float strictly between 0 and 1, a token is only considered if it is greater than either `eta_cutoff` or `sqrt(eta_cutoff) * exp(-entropy(softmax(next_token_logits)))`.
   * The latter term is intuitively the expected next token probability, scaled by `sqrt(eta_cutoff)`. In the paper, suggested values range from 3e-4 to 2e-3, depending on the size of the model.
   * See [Truncation Sampling as Language Model Desmoothing](https://huggingface.co/papers/2210.15191) for more details.
   * @type {number}
   * @default 0.0
   */
  eta_cutoff = 0;
  /**
   * This value is subtracted from a beam's score if it generates a token same as any beam from other group at a particular time.
   * Note that `diversity_penalty` is only effective if `group beam search` is enabled.
   * @type {number}
   * @default 0.0
   */
  diversity_penalty = 0;
  /**
   * The parameter for repetition penalty. 1.0 means no penalty.
   * See [this paper](https://huggingface.co/papers/1909.05858) for more details.
   * @type {number}
   * @default 1.0
   */
  repetition_penalty = 1;
  /**
   * The paramater for encoder_repetition_penalty.
   * An exponential penalty on sequences that are not in the original input.
   * 1.0 means no penalty.
   * @type {number}
   * @default 1.0
   */
  encoder_repetition_penalty = 1;
  /**
   * Exponential penalty to the length that is used with beam-based generation.
   * It is applied as an exponent to the sequence length, which in turn is used to divide the score of the sequence.
   * Since the score is the log likelihood of the sequence (i.e. negative), `length_penalty` > 0.0 promotes longer sequences, while `length_penalty` < 0.0 encourages shorter sequences.
   * @type {number}
   * @default 1.0
   */
  length_penalty = 1;
  /**
   * If set to int > 0, all ngrams of that size can only occur once.
   * @type {number}
   * @default 0
   */
  no_repeat_ngram_size = 0;
  /**
   * List of token ids that are not allowed to be generated.
   * In order to get the token ids of the words that should not appear in the generated text, use
   * `tokenizer(bad_words, { add_prefix_space: true, add_special_tokens: false }).input_ids`.
   * @type {number[][]}
   * @default null
   */
  bad_words_ids = null;
  /**
   * List of token ids that must be generated.
   * If given a `number[][]`, this is treated as a simple list of words that must be included, the opposite to `bad_words_ids`.
   * If given `number[][][]`, this triggers a [disjunctive constraint](https://github.com/huggingface/transformers/issues/14081), where one can allow different forms of each word.
   * @type {number[][]|number[][][]}
   * @default null
   */
  force_words_ids = null;
  /**
   * Whether to renormalize the logits after applying all the logits processors or warpers (including the custom ones).
   * It's highly recommended to set this flag to `true` as the search algorithms suppose the score logits are normalized but some logit processors or warpers break the normalization.
   * @type {boolean}
   * @default false
   */
  renormalize_logits = false;
  /**
   * Custom constraints that can be added to the generation to ensure that the output will contain the use of certain tokens as defined by `Constraint` objects, in the most sensible way possible.
   * @type {Object[]}
   * @default null
   */
  constraints = null;
  /**
   * The id of the token to force as the first generated token after the `decoder_start_token_id`.
   * Useful for multilingual models like mBART where the first generated token needs to be the target language token.
   * @type {number}
   * @default null
   */
  forced_bos_token_id = null;
  /**
   * The id of the token to force as the last generated token when `max_length` is reached.
   * Optionally, use a list to set multiple *end-of-sequence* tokens.
   * @type {number|number[]}
   * @default null
   */
  forced_eos_token_id = null;
  /**
   * Whether to remove possible *nan* and *inf* outputs of the model to prevent the generation method to crash. Note that using `remove_invalid_values` can slow down generation.
   * @type {boolean}
   */
  remove_invalid_values = false;
  /**
   * This Tuple adds an exponentially increasing length penalty, after a certain amount of tokens have been generated.
   * The tuple shall consist of: `(start_index, decay_factor)` where `start_index` indicates where penalty starts and `decay_factor` represents the factor of exponential decay.
   * @type {[number, number]}
   * @default null
   */
  exponential_decay_length_penalty = null;
  /**
   * A list of tokens that will be suppressed at generation.
   * The `SuppressTokens` logit processor will set their log probs to `-inf` so that they are not sampled.
   * @type {number[]}
   * @default null
   */
  suppress_tokens = null;
  /**
   * A streamer that will be used to stream the generation.
   * @type {import('./streamers.js').TextStreamer}
   * @default null
   */
  streamer = null;
  /**
   * A list of tokens that will be suppressed at the beginning of the generation.
   * The `SuppressBeginTokens` logit processor will set their log probs to `-inf` so that they are not sampled.
   * @type {number[]}
   * @default null
   */
  begin_suppress_tokens = null;
  /**
   * A list of pairs of integers which indicates a mapping from generation indices to token indices that will be forced before sampling.
   * For example, `[[1, 123]]` means the second generated token will always be a token of index 123.
   * @type {[number, number][]}
   * @default null
   */
  forced_decoder_ids = null;
  /**
   * The guidance scale for classifier free guidance (CFG). CFG is enabled by setting `guidance_scale > 1`.
   * Higher guidance scale encourages the model to generate samples that are more closely linked to the input
   * prompt, usually at the expense of poorer quality.
   * @type {number}
   * @default null
   */
  guidance_scale = null;
  // Parameters that define the output variables of `generate`
  /**
   * The number of independently computed returned sequences for each element in the batch.
   * @type {number}
   * @default 1
   */
  num_return_sequences = 1;
  /**
   * Whether or not to return the attentions tensors of all attention layers.
   * See `attentions` under returned tensors for more details.
   * @type {boolean}
   * @default false
   */
  output_attentions = false;
  /**
   * Whether or not to return the hidden states of all layers.
   * See `hidden_states` under returned tensors for more details.
   * @type {boolean}
   * @default false
   */
  output_hidden_states = false;
  /**
   * Whether or not to return the prediction scores.
   * See `scores` under returned tensors for more details.
   * @type {boolean}
   * @default false
   */
  output_scores = false;
  /**
   * Whether or not to return a `ModelOutput` instead of a plain tuple.
   * @type {boolean}
   * @default false
   */
  return_dict_in_generate = false;
  // Special tokens that can be used at generation time
  /**
   * The id of the *padding* token.
   * @type {number}
   * @default null
   */
  pad_token_id = null;
  /**
   * The id of the *beginning-of-sequence* token.
   * @type {number}
   * @default null
   */
  bos_token_id = null;
  /**
   * The id of the *end-of-sequence* token.
   * Optionally, use a list to set multiple *end-of-sequence* tokens.
   * @type {number|number[]}
   * @default null
   */
  eos_token_id = null;
  // Generation parameters exclusive to encoder-decoder models
  /**
   * If set to int > 0, all ngrams of that size that occur in the `encoder_input_ids` cannot occur in the `decoder_input_ids`.
   * @type {number}
   * @default 0
   */
  encoder_no_repeat_ngram_size = 0;
  /**
   * If an encoder-decoder model starts decoding with a different token than *bos*, the id of that token.
   * @type {number}
   * @default null
   */
  decoder_start_token_id = null;
  // Wild card
  /**
   * Additional generation kwargs will be forwarded to the `generate` function of the model.
   * Kwargs that are not present in `generate`'s signature will be used in the model forward pass.
   * @type {Object}
   * @default {}
   */
  generation_kwargs = {};
  /**
   *
   * @param {GenerationConfig|import('../configs.js').PretrainedConfig} config
   */
  constructor(config) {
    Object.assign(this, pick(config, Object.getOwnPropertyNames(this)));
  }
};

// src/generation/stopping_criteria.js
var StoppingCriteria = class extends Callable {
  /**
   *
   * @param {number[][]} input_ids (`number[][]` of shape `(batch_size, sequence_length)`):
   * Indices of input sequence tokens in the vocabulary.
   * @param {number[][]} scores scores (`number[][]` of shape `(batch_size, config.vocab_size)`):
   * Prediction scores of a language modeling head. These can be scores for each vocabulary token before SoftMax
   * or scores for each vocabulary token after SoftMax.
   * @returns {boolean[]} A list of booleans indicating whether each sequence should be stopped.
   */
  _call(input_ids, scores) {
    throw Error("StoppingCriteria needs to be subclassed");
  }
};
var StoppingCriteriaList = class _StoppingCriteriaList extends Callable {
  /**
   * Constructs a new instance of `StoppingCriteriaList`.
   */
  constructor() {
    super();
    this.criteria = [];
  }
  /**
   * Adds a new stopping criterion to the list.
   *
   * @param {StoppingCriteria} item The stopping criterion to add.
   */
  push(item) {
    this.criteria.push(item);
  }
  /**
   * Adds multiple stopping criteria to the list.
   *
   * @param {StoppingCriteria|StoppingCriteriaList|StoppingCriteria[]} items The stopping criteria to add.
   */
  extend(items) {
    if (items instanceof _StoppingCriteriaList) {
      items = items.criteria;
    } else if (items instanceof StoppingCriteria) {
      items = [items];
    }
    this.criteria.push(...items);
  }
  _call(input_ids, scores) {
    const is_done = new Array(input_ids.length).fill(false);
    for (const criterion of this.criteria) {
      const criterion_done = criterion(input_ids, scores);
      for (let i = 0; i < is_done.length; ++i) {
        is_done[i] ||= criterion_done[i];
      }
    }
    return is_done;
  }
  [Symbol.iterator]() {
    return this.criteria.values();
  }
};
var MaxLengthCriteria = class extends StoppingCriteria {
  /**
   *
   * @param {number} max_length The maximum length that the output sequence can have in number of tokens.
   * @param {number} [max_position_embeddings=null] The maximum model length, as defined by the model's `config.max_position_embeddings` attribute.
   */
  constructor(max_length, max_position_embeddings = null) {
    super();
    this.max_length = max_length;
    this.max_position_embeddings = max_position_embeddings;
  }
  _call(input_ids) {
    return input_ids.map((ids) => ids.length >= this.max_length);
  }
};
var EosTokenCriteria = class extends StoppingCriteria {
  /**
   *
   * @param {number|number[]} eos_token_id The id of the *end-of-sequence* token.
   * Optionally, use a list to set multiple *end-of-sequence* tokens.
   */
  constructor(eos_token_id) {
    super();
    if (!Array.isArray(eos_token_id)) {
      eos_token_id = [eos_token_id];
    }
    this.eos_token_id = eos_token_id;
  }
  /**
   *
   * @param {number[][]} input_ids
   * @param {number[][]} scores
   * @returns {boolean[]}
   */
  _call(input_ids, scores) {
    return input_ids.map((ids) => {
      const last = ids.at(-1);
      return this.eos_token_id.some((eos_id) => last == eos_id);
    });
  }
};
var InterruptableStoppingCriteria = class extends StoppingCriteria {
  constructor() {
    super();
    this.interrupted = false;
  }
  interrupt() {
    this.interrupted = true;
  }
  reset() {
    this.interrupted = false;
  }
  _call(input_ids, scores) {
    return new Array(input_ids.length).fill(this.interrupted);
  }
};

// src/generation/logits_sampler.js
var LogitsSampler = class extends Callable {
  /**
   * Creates a new Sampler object with the specified generation config.
   * @param {GenerationConfig} generation_config The generation config.
   */
  constructor(generation_config) {
    super();
    this.generation_config = generation_config;
  }
  /**
   * Executes the sampler, using the specified logits.
   * @param {Tensor} logits
   * @returns {Promise<[bigint, number][]>}
   */
  async _call(logits) {
    return this.sample(logits);
  }
  /**
   * Abstract method for sampling the logits.
   * @param {Tensor} logits
   * @throws {Error} If not implemented in subclass.
   * @returns {Promise<[bigint, number][]>}
   */
  async sample(logits) {
    throw Error("sample should be implemented in subclasses.");
  }
  /**
   * Returns the specified logits as an array, with temperature applied.
   * @param {Tensor} logits
   * @param {number} index
   * @returns {Float32Array}
   */
  getLogits(logits, index) {
    let vocabSize = logits.dims.at(-1);
    let logs = (
      /** @type {Float32Array} */
      logits.data
    );
    if (index === -1) {
      logs = logs.slice(-vocabSize);
    } else {
      let startIndex = index * vocabSize;
      logs = logs.slice(startIndex, startIndex + vocabSize);
    }
    return logs;
  }
  /**
   * Selects an item randomly based on the specified probabilities.
   * @param {Float32Array} probabilities An array of probabilities to use for selection.
   * @returns {number} The index of the selected item.
   */
  randomSelect(probabilities) {
    return _weightedIndex(probabilities);
  }
  /**
   * Returns a Sampler object based on the specified options.
   * @param {GenerationConfig} generation_config An object containing options for the sampler.
   * @returns {LogitsSampler} A Sampler object.
   */
  static getSampler(generation_config) {
    if (generation_config.do_sample) {
      return new MultinomialSampler(generation_config);
    } else if (generation_config.num_beams > 1) {
      return new BeamSearchSampler(generation_config);
    } else {
      if (generation_config.num_return_sequences > 1) {
        throw Error(
          `num_return_sequences has to be 1 when doing greedy search, but is ${generation_config.num_return_sequences}.`
        );
      }
      return new GreedySampler(generation_config);
    }
  }
};
var GreedySampler = class extends LogitsSampler {
  /**
   * Sample the maximum probability of a given logits tensor.
   * @param {Tensor} logits
   * @returns {Promise<[bigint, number][]>} An array with a single tuple, containing the index of the maximum value and a meaningless score (since this is a greedy search).
   */
  async sample(logits) {
    const argmax = max(logits.data)[1];
    return [[BigInt(argmax), 0]];
  }
};
var MultinomialSampler = class extends LogitsSampler {
  /**
   * Sample from the logits.
   * @param {Tensor} logits
   * @returns {Promise<[bigint, number][]>}
   */
  async sample(logits) {
    let k2 = logits.dims.at(-1);
    if (this.generation_config.top_k > 0) {
      k2 = Math.min(this.generation_config.top_k, k2);
    }
    const [v, i] = await topk(logits, k2);
    const probabilities = softmax(
      /** @type {Float32Array} */
      v.data
    );
    return Array.from({ length: this.generation_config.num_beams }, () => {
      const sampledIndex = this.randomSelect(probabilities);
      return [
        i.data[sampledIndex],
        // token id
        Math.log(probabilities[sampledIndex])
        // score
      ];
    });
  }
};
var BeamSearchSampler = class extends LogitsSampler {
  /**
   * Sample from the logits.
   * @param {Tensor} logits
   * @returns {Promise<[bigint, number][]>}
   */
  async sample(logits) {
    let k2 = logits.dims.at(-1);
    if (this.generation_config.top_k > 0) {
      k2 = Math.min(this.generation_config.top_k, k2);
    }
    const [v, i] = await topk(logits, k2);
    const probabilities = softmax(
      /** @type {Float32Array} */
      v.data
    );
    return Array.from({ length: this.generation_config.num_beams }, (_, x) => {
      return [
        i.data[x],
        // token id
        Math.log(probabilities[x])
        // score
      ];
    });
  }
};

// src/cache_utils.js
var _DynamicCache = class {
  /**
   * Create a DynamicCache, optionally pre-populated with entries.
   * @param {Record<string, Tensor>} [entries] Initial name→Tensor mappings.
   */
  constructor(entries) {
    if (!entries) return;
    for (const key in entries) {
      if (key in this) {
        throw new TypeError(`Key "${key}" conflicts with an existing property on DynamicCache`);
      }
      const value = entries[key];
      if (!(value instanceof Tensor2)) {
        throw new TypeError(`Expected a Tensor for key "${key}", got ${typeof value}`);
      }
      this[key] = value;
    }
  }
  /**
   * Get the cached sequence length. This requires at least one attention cache entry to be present.
   * @returns {number} The past sequence length.
   */
  get_seq_length() {
    const self2 = (
      /** @type {any} */
      this
    );
    if (Object.keys(self2).length === 0) {
      return 0;
    }
    for (const name in self2) {
      if (name.startsWith("past_key_values.")) {
        return self2[name].dims.at(-2);
      }
    }
    throw new Error("Unable to determine sequence length from the cache.");
  }
  /**
   * Update the cache in-place with new entries, disposing replaced GPU tensors.
   * @param {Record<string, Tensor>} newEntries The new name → Tensor mappings.
   */
  update(newEntries) {
    for (const key in newEntries) {
      const oldValue = this[key];
      const newValue = newEntries[key];
      if (oldValue && oldValue !== newValue && oldValue.location === "gpu-buffer") {
        oldValue.dispose();
      }
      this[key] = newValue;
    }
  }
  /**
   * Dispose all contained tensors whose data resides on the GPU.
   * Returns a promise that resolves when all disposals are complete.
   * @returns {Promise<void>} Promise that resolves when all GPU tensors are disposed.
   */
  async dispose() {
    const promises = [];
    for (
      const t of
      /** @type {Tensor[]} */
      Object.values(this)
    ) {
      if (t.location === "gpu-buffer") {
        promises.push(t.dispose());
      }
    }
    await Promise.all(promises);
  }
};
var DynamicCache = (
  /** @type {new (entries?: Record<string, Tensor>) => DynamicCache} */
  /** @type {unknown} */
  _DynamicCache
);

// src/models/session_config.js
var MODEL_TYPES = {
  EncoderOnly: 0,
  EncoderDecoder: 1,
  Seq2Seq: 2,
  Vision2Seq: 3,
  DecoderOnly: 4,
  DecoderOnlyWithoutHead: 5,
  MaskGeneration: 6,
  ImageTextToText: 7,
  Musicgen: 8,
  MultiModality: 9,
  Phi3V: 10,
  AudioTextToText: 11,
  AutoEncoder: 12,
  ImageAudioTextToText: 13,
  Supertonic: 14,
  Chatterbox: 15,
  VoxtralRealtime: 16
};
var MODEL_SESSION_CONFIG = {
  [MODEL_TYPES.DecoderOnly]: {
    sessions: (config, options) => ({ model: options.model_file_name ?? "model" }),
    cache_sessions: { model: true },
    optional_configs: { generation_config: "generation_config.json" }
  },
  [MODEL_TYPES.DecoderOnlyWithoutHead]: {
    sessions: (config, options) => ({ model: options.model_file_name ?? "model" })
  },
  [MODEL_TYPES.Seq2Seq]: {
    sessions: () => ({ model: "encoder_model", decoder_model_merged: "decoder_model_merged" }),
    cache_sessions: { decoder_model_merged: true },
    optional_configs: { generation_config: "generation_config.json" }
  },
  [MODEL_TYPES.Vision2Seq]: {
    sessions: () => ({ model: "encoder_model", decoder_model_merged: "decoder_model_merged" }),
    cache_sessions: { decoder_model_merged: true },
    optional_configs: { generation_config: "generation_config.json" }
  },
  [MODEL_TYPES.Musicgen]: {
    sessions: () => ({
      model: "text_encoder",
      decoder_model_merged: "decoder_model_merged",
      encodec_decode: "encodec_decode"
    }),
    cache_sessions: { decoder_model_merged: true },
    optional_configs: { generation_config: "generation_config.json" }
  },
  [MODEL_TYPES.EncoderDecoder]: {
    sessions: () => ({ model: "encoder_model", decoder_model_merged: "decoder_model_merged" }),
    cache_sessions: { decoder_model_merged: true }
  },
  [MODEL_TYPES.MaskGeneration]: {
    sessions: () => ({ model: "vision_encoder", prompt_encoder_mask_decoder: "prompt_encoder_mask_decoder" })
  },
  [MODEL_TYPES.ImageTextToText]: {
    text_only_sessions: { embed_tokens: "embed_tokens", decoder_model_merged: "decoder_model_merged" },
    sessions: (config, options, textOnly) => {
      const s = { ...MODEL_SESSION_CONFIG[MODEL_TYPES.ImageTextToText].text_only_sessions };
      if (!textOnly) s["vision_encoder"] = "vision_encoder";
      if (config.is_encoder_decoder) s["model"] = "encoder_model";
      return s;
    },
    cache_sessions: { decoder_model_merged: true },
    optional_configs: { generation_config: "generation_config.json" }
  },
  [MODEL_TYPES.AudioTextToText]: {
    text_only_sessions: { embed_tokens: "embed_tokens", decoder_model_merged: "decoder_model_merged" },
    sessions: (config, options, textOnly) => {
      const s = { ...MODEL_SESSION_CONFIG[MODEL_TYPES.AudioTextToText].text_only_sessions };
      if (!textOnly) s["audio_encoder"] = "audio_encoder";
      return s;
    },
    cache_sessions: { decoder_model_merged: true },
    optional_configs: { generation_config: "generation_config.json" }
  },
  [MODEL_TYPES.ImageAudioTextToText]: {
    text_only_sessions: { embed_tokens: "embed_tokens", decoder_model_merged: "decoder_model_merged" },
    sessions: (config, options, textOnly) => {
      const s = { ...MODEL_SESSION_CONFIG[MODEL_TYPES.ImageAudioTextToText].text_only_sessions };
      if (!textOnly) {
        s["audio_encoder"] = "audio_encoder";
        s["vision_encoder"] = "vision_encoder";
      }
      return s;
    },
    optional_configs: { generation_config: "generation_config.json" }
  },
  [MODEL_TYPES.Phi3V]: {
    sessions: () => ({
      prepare_inputs_embeds: "prepare_inputs_embeds",
      model: "model",
      vision_encoder: "vision_encoder"
    }),
    cache_sessions: { model: true },
    optional_configs: { generation_config: "generation_config.json" }
  },
  [MODEL_TYPES.MultiModality]: {
    sessions: () => ({
      prepare_inputs_embeds: "prepare_inputs_embeds",
      model: "language_model",
      lm_head: "lm_head",
      gen_head: "gen_head",
      gen_img_embeds: "gen_img_embeds",
      image_decode: "image_decode"
    }),
    cache_sessions: { model: true },
    optional_configs: { generation_config: "generation_config.json" }
  },
  [MODEL_TYPES.AutoEncoder]: {
    sessions: () => ({ encoder_model: "encoder_model", decoder_model: "decoder_model" })
  },
  [MODEL_TYPES.Supertonic]: {
    sessions: () => ({
      text_encoder: "text_encoder",
      latent_denoiser: "latent_denoiser",
      voice_decoder: "voice_decoder"
    })
  },
  [MODEL_TYPES.Chatterbox]: {
    sessions: () => ({
      embed_tokens: "embed_tokens",
      speech_encoder: "speech_encoder",
      model: "language_model",
      conditional_decoder: "conditional_decoder"
    }),
    cache_sessions: { model: true },
    optional_configs: { generation_config: "generation_config.json" }
  },
  [MODEL_TYPES.VoxtralRealtime]: {
    text_only_sessions: { embed_tokens: "embed_tokens", decoder_model_merged: "decoder_model_merged" },
    sessions: (config, options, textOnly) => {
      const s = { ...MODEL_SESSION_CONFIG[MODEL_TYPES.VoxtralRealtime].text_only_sessions };
      if (!textOnly) s["audio_encoder"] = "audio_encoder";
      return s;
    },
    cache_sessions: { decoder_model_merged: true, audio_encoder: true },
    optional_configs: { generation_config: "generation_config.json" }
  },
  default: {
    sessions: (config, options) => ({ model: options.model_file_name ?? "model" })
  }
};
function getTextOnlySessions(modelType) {
  const typeConfig = MODEL_SESSION_CONFIG[modelType];
  return typeConfig?.text_only_sessions ?? null;
}
function getSessionsConfig(modelType, config, options = {}) {
  const typeConfig = MODEL_SESSION_CONFIG[modelType] ?? MODEL_SESSION_CONFIG.default;
  return {
    sessions: typeConfig.sessions(config, options, options.textOnly ?? false),
    cache_sessions: typeConfig.cache_sessions,
    optional_configs: typeConfig.optional_configs
  };
}

// src/utils/model_registry/resolve_model_type.js
function resolve_model_type(config, { warn = true } = {}) {
  const architectures = (
    /** @type {string[]} */
    config.architectures || []
  );
  for (const arch of architectures) {
    const mappedType = MODEL_TYPE_MAPPING.get(arch);
    if (mappedType !== void 0) {
      return mappedType;
    }
  }
  if (config.model_type) {
    const mappedType = MODEL_TYPE_MAPPING.get(config.model_type);
    if (mappedType !== void 0) {
      return mappedType;
    }
    for (const mapping of Object.values(MODEL_MAPPING_NAMES)) {
      if (mapping.has(config.model_type)) {
        const resolved = MODEL_TYPE_MAPPING.get(mapping.get(config.model_type));
        if (resolved !== void 0) {
          return resolved;
        }
      }
    }
  }
  if (warn) {
    const archList = architectures.length > 0 ? architectures.join(", ") : "(none)";
    logger.warn(
      `[resolve_model_type] Architecture(s) not found in MODEL_TYPE_MAPPING: [${archList}] for model type '${config.model_type}'. Falling back to EncoderOnly (single model.onnx file). If you encounter issues, please report at: ${GITHUB_ISSUE_URL}`
    );
  }
  return MODEL_TYPES.EncoderOnly;
}

// src/utils/model_registry/get_model_files.js
function get_config(modelId, { config = null, cache_dir = null, local_files_only = false, revision = "main" } = {}) {
  if (config !== null) {
    return AutoConfig.from_pretrained(modelId, { config, cache_dir, local_files_only, revision });
  }
  const key = JSON.stringify([modelId, cache_dir, local_files_only, revision]);
  return memoizePromise(
    key,
    () => AutoConfig.from_pretrained(modelId, { config, cache_dir, local_files_only, revision })
  );
}
async function get_model_files(modelId, { config = null, dtype: overrideDtype = null, device: overrideDevice = null, model_file_name = null } = {}) {
  config = await get_config(modelId, { config });
  const files = [
    // Add config.json (always loaded)
    "config.json"
  ];
  const custom_config = config["transformers.js_config"] ?? {};
  const use_external_data_format = custom_config.use_external_data_format;
  const subfolder = "onnx";
  const rawDevice = overrideDevice ?? custom_config.device;
  let dtype = overrideDtype ?? custom_config.dtype;
  const modelType = resolve_model_type(config);
  const add_model_file = (fileName, baseName = null) => {
    baseName = baseName ?? fileName;
    const selectedDevice = selectDevice(rawDevice, fileName);
    const selectedDtype = selectDtype(dtype, fileName, selectedDevice);
    const suffix = DEFAULT_DTYPE_SUFFIX_MAPPING[selectedDtype] ?? "";
    const fullName = `${baseName}${suffix}.onnx`;
    const fullPath = subfolder ? `${subfolder}/${fullName}` : fullName;
    files.push(fullPath);
    const num_chunks = resolveExternalDataFormat(use_external_data_format, fullName, fileName);
    for (const dataFileName of getExternalDataChunkNames(fullName, num_chunks)) {
      const dataFilePath = subfolder ? `${subfolder}/${dataFileName}` : dataFileName;
      files.push(dataFilePath);
    }
  };
  const { sessions, optional_configs } = getSessionsConfig(modelType, config, { model_file_name });
  for (const [sessionKey, baseName] of Object.entries(sessions)) {
    add_model_file(sessionKey, baseName);
  }
  if (optional_configs) {
    for (const configFile of Object.values(optional_configs)) {
      files.push(configFile);
    }
  }
  return files;
}

// src/models/modeling_utils.js
var MODEL_MAPPING_NAMES = null;
function registerTaskMappings(mappings) {
  MODEL_MAPPING_NAMES = mappings;
}
function toI64Tensor(items) {
  if (items instanceof Tensor2) {
    return items;
  }
  if (items.length === 0) {
    throw Error("items must be non-empty");
  }
  if (Array.isArray(items[0])) {
    if (items.some((x) => x.length !== items[0].length)) {
      throw Error(
        "Unable to create tensor, you should probably activate truncation and/or padding with 'padding=True' and/or 'truncation=True' to have batched tensors with the same length."
      );
    }
    return new Tensor2("int64", BigInt64Array.from(items.flat().map((x) => BigInt(x))), [
      items.length,
      items[0].length
    ]);
  } else {
    return new Tensor2("int64", BigInt64Array.from(items.map((x) => BigInt(x))), [1, items.length]);
  }
}
function boolTensor(value) {
  return new Tensor2("bool", [value], [1]);
}
var MODEL_RUNTIME_CONFIG = {
  [MODEL_TYPES.DecoderOnly]: {
    can_generate: true,
    forward: decoder_forward,
    prepare_inputs: decoder_prepare_inputs_for_generation
  },
  [MODEL_TYPES.DecoderOnlyWithoutHead]: {
    can_generate: false,
    forward: decoder_forward,
    prepare_inputs: decoder_prepare_inputs_for_generation
  },
  [MODEL_TYPES.Seq2Seq]: {
    can_generate: true,
    forward: seq2seq_forward,
    prepare_inputs: encoder_decoder_prepare_inputs_for_generation
  },
  [MODEL_TYPES.Vision2Seq]: {
    can_generate: true,
    forward: seq2seq_forward,
    prepare_inputs: encoder_decoder_prepare_inputs_for_generation
  },
  [MODEL_TYPES.Musicgen]: {
    can_generate: true,
    forward: seq2seq_forward
  },
  [MODEL_TYPES.EncoderDecoder]: {
    can_generate: false,
    forward: seq2seq_forward
  },
  [MODEL_TYPES.ImageTextToText]: {
    can_generate: true,
    forward: image_text_to_text_forward,
    prepare_inputs: multimodal_text_to_text_prepare_inputs_for_generation
  },
  [MODEL_TYPES.AudioTextToText]: {
    can_generate: true,
    forward: audio_text_to_text_forward,
    prepare_inputs: multimodal_text_to_text_prepare_inputs_for_generation
  },
  [MODEL_TYPES.ImageAudioTextToText]: {
    can_generate: true,
    prepare_inputs: multimodal_text_to_text_prepare_inputs_for_generation
  },
  [MODEL_TYPES.Phi3V]: {
    can_generate: true,
    prepare_inputs: multimodal_text_to_text_prepare_inputs_for_generation
  },
  [MODEL_TYPES.MultiModality]: {
    can_generate: true
  },
  [MODEL_TYPES.AutoEncoder]: {
    can_generate: false,
    forward: auto_encoder_forward
  },
  [MODEL_TYPES.Chatterbox]: {
    can_generate: true,
    forward: encoder_forward
  },
  [MODEL_TYPES.VoxtralRealtime]: {
    can_generate: true,
    prepare_inputs: decoder_prepare_inputs_for_generation
  },
  default: {
    can_generate: false,
    forward: encoder_forward
  }
};
function resolveTypeConfig(modelName, config) {
  let modelType = MODEL_TYPE_MAPPING.get(modelName);
  let textOnly = false;
  const nativeArch = config?.architectures?.[0];
  if (nativeArch && nativeArch !== modelName && modelName?.endsWith("ForCausalLM") && nativeArch.endsWith("ForConditionalGeneration")) {
    const nativeType = MODEL_TYPE_MAPPING.get(nativeArch);
    if (nativeType !== void 0) {
      modelType = nativeType;
      textOnly = true;
    }
  }
  const runtimeConfig = MODEL_RUNTIME_CONFIG[modelType] ?? MODEL_RUNTIME_CONFIG.default;
  const sessionConfig = MODEL_SESSION_CONFIG[modelType] ?? MODEL_SESSION_CONFIG.default;
  return { typeConfig: { ...runtimeConfig, ...sessionConfig }, textOnly, modelType };
}
var MODEL_TYPE_MAPPING = /* @__PURE__ */ new Map();
var MODEL_NAME_TO_CLASS_MAPPING = /* @__PURE__ */ new Map();
var MODEL_CLASS_TO_NAME_MAPPING = /* @__PURE__ */ new Map();
var PreTrainedModel = class extends Callable {
  main_input_name = "input_ids";
  forward_params = ["input_ids", "attention_mask"];
  _return_dict_in_generate_keys = null;
  /**
   * Creates a new instance of the `PreTrainedModel` class.
   * @param {import('../configs.js').PretrainedConfig} config The model configuration.
   * @param {Record<string, any>} sessions The inference sessions for the model.
   * @param {Record<string, Object>} configs Additional configuration files (e.g., generation_config.json).
   */
  constructor(config, sessions, configs) {
    super();
    this.config = config;
    this.sessions = sessions;
    this.configs = configs;
    const modelName = MODEL_CLASS_TO_NAME_MAPPING.get(this.constructor);
    const { typeConfig } = resolveTypeConfig(modelName, config);
    this.can_generate = typeConfig.can_generate;
    this._forward = typeConfig.forward;
    this._prepare_inputs_for_generation = typeConfig.prepare_inputs;
    if (this.can_generate) {
      this.forward_params.push("past_key_values");
    }
    this.custom_config = this.config["transformers.js_config"] ?? {};
  }
  /**
   * Disposes of all the ONNX sessions that were created during inference.
   * @returns {Promise<unknown[]>} An array of promises, one for each ONNX session that is being disposed.
   * @todo Use https://developer.mozilla.org/en-US/docs/Web/JavaScript/Reference/Global_Objects/FinalizationRegistry
   */
  async dispose() {
    const promises = [];
    for (const session of Object.values(this.sessions)) {
      promises.push(session.release?.());
    }
    return await Promise.all(promises);
  }
  /**
   * Instantiate one of the model classes of the library from a pretrained model.
   *
   * The model class to instantiate is selected based on the `model_type` property of the config object
   * (either passed as an argument or loaded from `pretrained_model_name_or_path` if possible)
   *
   * @param {string} pretrained_model_name_or_path The name or path of the pretrained model. Can be either:
   * - A string, the *model id* of a pretrained model hosted inside a model repo on huggingface.co.
   *   Valid model ids can be located at the root-level, like `bert-base-uncased`, or namespaced under a
   *   user or organization name, like `dbmdz/bert-base-german-cased`.
   * - A path to a *directory* containing model weights, e.g., `./my_model_directory/`.
   * @param {import('../utils/hub.js').PretrainedModelOptions} options Additional options for loading the model.
   *
   * @returns {Promise<PreTrainedModel>} A new instance of the `PreTrainedModel` class.
   */
  static async from_pretrained(pretrained_model_name_or_path, {
    progress_callback = null,
    config = null,
    cache_dir = null,
    local_files_only = false,
    revision = "main",
    model_file_name = null,
    subfolder = "onnx",
    device = null,
    dtype = null,
    use_external_data_format = null,
    session_options = {}
  } = {}) {
    const options = {
      progress_callback,
      config,
      cache_dir,
      local_files_only,
      revision,
      model_file_name,
      subfolder,
      device,
      dtype,
      use_external_data_format,
      session_options
    };
    const modelName = MODEL_CLASS_TO_NAME_MAPPING.get(this);
    config = options.config = await AutoConfig.from_pretrained(pretrained_model_name_or_path, options);
    const { typeConfig, textOnly, modelType } = resolveTypeConfig(modelName, config);
    if (modelType === void 0) {
      const type = modelName ?? config?.model_type;
      if (type !== "custom") {
        logger.warn(
          `Model type for '${type}' not found, assuming encoder-only architecture. Please report this at ${GITHUB_ISSUE_URL}.`
        );
      }
    }
    if (progress_callback && !(progress_callback instanceof DefaultProgressCallback)) {
      const files_loading = {};
      try {
        const expected_files = await get_model_files(pretrained_model_name_or_path, {
          config,
          dtype,
          device,
          model_file_name
        });
        const metadata = await Promise.all(
          expected_files.map((file) => get_file_metadata(pretrained_model_name_or_path, file, options))
        );
        metadata.forEach((m, i) => {
          if (m.exists) {
            const isAlreadyLoaded = expected_files[i] === "config.json";
            files_loading[expected_files[i]] = {
              loaded: isAlreadyLoaded ? m.size ?? 0 : 0,
              total: m.size ?? 0
            };
          }
        });
      } catch (e) {
        logger.warn(`Unable to fetch model file metadata for total progress tracking: ${e}`);
      }
      if (Object.keys(files_loading).length > 0) {
        options.progress_callback = new DefaultProgressCallback(progress_callback, files_loading);
      }
    }
    const sessions = typeConfig.sessions(config, options, textOnly);
    const promises = [
      constructSessions(pretrained_model_name_or_path, sessions, options, typeConfig.cache_sessions)
    ];
    if (typeConfig.optional_configs) {
      promises.push(get_optional_configs(pretrained_model_name_or_path, typeConfig.optional_configs, options));
    }
    const info = await Promise.all(promises);
    return new this(config, ...info);
  }
  /**
   * Runs the model with the provided inputs
   * @param {Object} model_inputs Object containing input tensors
   * @returns {Promise<Object>} Object containing output tensors
   */
  async _call(model_inputs) {
    return await this.forward(model_inputs);
  }
  /**
   * Forward method for a pretrained model. If not overridden by a subclass, the correct forward method
   * will be chosen based on the model type.
   * @param {Object} model_inputs The input data to the model in the format specified in the ONNX model.
   * @returns {Promise<Object>} The output data from the model in the format specified in the ONNX model.
   * @throws {Error} This method must be implemented in subclasses.
   */
  async forward(model_inputs) {
    return await this._forward(this, model_inputs);
  }
  /**
   * Get the model's generation config, if it exists.
   * @returns {GenerationConfig|null} The model's generation config if it exists, otherwise `null`.
   */
  get generation_config() {
    return this.configs?.generation_config ?? null;
  }
  /**
   * @param {GenerationConfig} generation_config
   * @param {number} input_ids_seq_length The starting sequence length for the input ids.
   * @returns {LogitsProcessorList}
   * @private
   */
  _get_logits_processor(generation_config, input_ids_seq_length, logits_processor = null) {
    const processors = new LogitsProcessorList();
    if (generation_config.repetition_penalty !== null && generation_config.repetition_penalty !== 1) {
      processors.push(new RepetitionPenaltyLogitsProcessor(generation_config.repetition_penalty));
    }
    if (generation_config.no_repeat_ngram_size !== null && generation_config.no_repeat_ngram_size > 0) {
      processors.push(new NoRepeatNGramLogitsProcessor(generation_config.no_repeat_ngram_size));
    }
    if (generation_config.bad_words_ids !== null) {
      processors.push(
        new NoBadWordsLogitsProcessor(generation_config.bad_words_ids, generation_config.eos_token_id)
      );
    }
    if (generation_config.min_length !== null && generation_config.eos_token_id !== null && generation_config.min_length > 0) {
      processors.push(new MinLengthLogitsProcessor(generation_config.min_length, generation_config.eos_token_id));
    }
    if (generation_config.min_new_tokens !== null && generation_config.eos_token_id !== null && generation_config.min_new_tokens > 0) {
      processors.push(
        new MinNewTokensLengthLogitsProcessor(
          input_ids_seq_length,
          generation_config.min_new_tokens,
          generation_config.eos_token_id
        )
      );
    }
    if (generation_config.forced_bos_token_id !== null) {
      processors.push(new ForcedBOSTokenLogitsProcessor(generation_config.forced_bos_token_id));
    }
    if (generation_config.forced_eos_token_id !== null) {
      processors.push(
        new ForcedEOSTokenLogitsProcessor(generation_config.max_length, generation_config.forced_eos_token_id)
      );
    }
    if (generation_config.suppress_tokens !== null) {
      processors.push(new SuppressTokensLogitsProcessor(generation_config.suppress_tokens));
    }
    if (generation_config.begin_suppress_tokens !== null) {
      const begin_index = input_ids_seq_length > 1 || generation_config.forced_bos_token_id === null ? input_ids_seq_length : input_ids_seq_length + 1;
      processors.push(
        new SuppressTokensAtBeginLogitsProcessor(generation_config.begin_suppress_tokens, begin_index)
      );
    }
    if (generation_config.guidance_scale !== null && generation_config.guidance_scale > 1) {
      processors.push(new ClassifierFreeGuidanceLogitsProcessor(generation_config.guidance_scale));
    }
    if (generation_config.temperature === 0 && generation_config.do_sample) {
      logger.warn(
        "`do_sample` changed to false because `temperature: 0` implies greedy sampling (always selecting the most likely token), which is incompatible with `do_sample: true`."
      );
      generation_config.do_sample = false;
    }
    if (generation_config.do_sample) {
      if (generation_config.temperature !== null && generation_config.temperature !== 1) {
        processors.push(new TemperatureLogitsWarper(generation_config.temperature));
      }
    }
    if (logits_processor !== null) {
      processors.extend(logits_processor);
    }
    return processors;
  }
  /**
   * This function merges multiple generation configs together to form a final generation config to be used by the model for text generation.
   * It first creates an empty `GenerationConfig` object, then it applies the model's own `generation_config` property to it. Finally, if a `generation_config` object was passed in the arguments, it overwrites the corresponding properties in the final config with those of the passed config object.
   * @param {GenerationConfig|null} generation_config A `GenerationConfig` object containing generation parameters.
   * @param {Object} kwargs Additional generation parameters to be used in place of those in the `generation_config` object.
   * @returns {GenerationConfig} The final generation config object to be used by the model for text generation.
   */
  _prepare_generation_config(generation_config, kwargs, cls = GenerationConfig) {
    const config = { ...this.config };
    for (const key of ["decoder", "generator", "text_config"]) {
      if (key in config) {
        Object.assign(config, config[key]);
      }
    }
    const gen_config = new cls(config);
    Object.assign(gen_config, this.generation_config ?? {});
    if (generation_config) {
      Object.assign(gen_config, generation_config);
    }
    if (kwargs) {
      Object.assign(gen_config, pick(kwargs, Object.getOwnPropertyNames(gen_config)));
    }
    return gen_config;
  }
  /**
   *
   * @param {GenerationConfig} generation_config
   * @param {import('../generation/stopping_criteria.js').StoppingCriteria|import('../generation/stopping_criteria.js').StoppingCriteria[]|StoppingCriteriaList} [stopping_criteria=null]
   */
  _get_stopping_criteria(generation_config, stopping_criteria = null) {
    const criteria = new StoppingCriteriaList();
    if (generation_config.max_length !== null) {
      criteria.push(
        new MaxLengthCriteria(generation_config.max_length, this.config.max_position_embeddings ?? null)
      );
    }
    if (generation_config.eos_token_id !== null) {
      criteria.push(new EosTokenCriteria(generation_config.eos_token_id));
    }
    if (stopping_criteria) {
      criteria.extend(stopping_criteria);
    }
    return criteria;
  }
  /**
   * Confirms that the model class is compatible with generation.
   * If not, raises an exception that points to the right class to use.
   */
  _validate_model_class() {
    if (!this.can_generate) {
      const generate_compatible_mappings = [
        MODEL_MAPPING_NAMES.MODEL_FOR_CAUSAL_LM_MAPPING_NAMES,
        // MODEL_MAPPING_NAMES.MODEL_FOR_CAUSAL_IMAGE_MODELING_MAPPING, // TODO
        MODEL_MAPPING_NAMES.MODEL_FOR_VISION_2_SEQ_MAPPING_NAMES,
        MODEL_MAPPING_NAMES.MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES,
        MODEL_MAPPING_NAMES.MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING_NAMES
      ].filter(Boolean);
      const modelName = MODEL_CLASS_TO_NAME_MAPPING.get(this.constructor);
      const generate_compatible_classes = /* @__PURE__ */ new Set();
      const modelType = this.config.model_type;
      for (const model_mapping of generate_compatible_mappings) {
        const supported_models = model_mapping?.get(modelType);
        if (supported_models) {
          generate_compatible_classes.add(supported_models);
        }
      }
      let errorMessage = `The current model class (${modelName}) is not compatible with \`.generate()\`, as it doesn't have a language model head.`;
      if (generate_compatible_classes.size > 0) {
        errorMessage += ` Please use the following class instead: ${[...generate_compatible_classes].join(", ")}`;
      }
      throw Error(errorMessage);
    }
  }
  prepare_inputs_for_generation(...args) {
    if (!this._prepare_inputs_for_generation) {
      throw new Error("prepare_inputs_for_generation is not implemented for this model.");
    }
    return this._prepare_inputs_for_generation(this, ...args);
  }
  /**
   *
   * @param {Object} inputs
   * @param {bigint[][]} inputs.generated_input_ids
   * @param {Object} inputs.outputs
   * @param {Object} inputs.model_inputs
   * @param {boolean} inputs.is_encoder_decoder
   * @returns {Object} The updated model inputs for the next generation iteration.
   */
  _update_model_kwargs_for_generation({ generated_input_ids, outputs, model_inputs, is_encoder_decoder }) {
    model_inputs["past_key_values"] = getPastKeyValues(outputs, model_inputs.past_key_values);
    model_inputs["input_ids"] = new Tensor2("int64", generated_input_ids.flat(), [generated_input_ids.length, 1]);
    if (!is_encoder_decoder) {
      model_inputs.attention_mask = cat(
        [model_inputs.attention_mask, ones([model_inputs.attention_mask.dims[0], 1])],
        1
      );
    } else if ("decoder_attention_mask" in model_inputs) {
      model_inputs.decoder_attention_mask = cat(
        [model_inputs.decoder_attention_mask, ones([model_inputs.decoder_attention_mask.dims[0], 1])],
        1
      );
    }
    model_inputs["position_ids"] = null;
    return model_inputs;
  }
  /**
   * This function extracts the model-specific `inputs` for generation.
   * @param {Object} params
   * @param {Tensor} [params.inputs=null]
   * @param {number} [params.bos_token_id=null]
   * @param {Record<string, Tensor|number[]>} [params.model_kwargs]
   * @returns {{inputs_tensor: Tensor, model_inputs: Record<string, Tensor> & {past_key_values?: DynamicCache}, model_input_name: string}} The model-specific inputs for generation.
   */
  _prepare_model_inputs({ inputs, bos_token_id, model_kwargs }) {
    const model_inputs = pick(model_kwargs, this.forward_params);
    const input_name = this.main_input_name;
    if (input_name in model_inputs) {
      if (inputs) {
        throw new Error(
          "`inputs`: {inputs}` were passed alongside {input_name} which is not allowed. Make sure to either pass {inputs} or {input_name}=..."
        );
      }
    } else {
      model_inputs[input_name] = inputs;
    }
    const inputs_tensor = model_inputs[input_name];
    return { inputs_tensor, model_inputs, model_input_name: input_name };
  }
  async _prepare_encoder_decoder_kwargs_for_generation({
    inputs_tensor,
    model_inputs,
    model_input_name,
    generation_config
  }) {
    if (this.sessions["model"].inputNames.includes("inputs_embeds") && !model_inputs.inputs_embeds && "_prepare_inputs_embeds" in this) {
      const { input_ids, pixel_values, attention_mask, ...kwargs } = model_inputs;
      const prepared_inputs = await this._prepare_inputs_embeds(model_inputs);
      model_inputs = {
        ...kwargs,
        ...pick(prepared_inputs, ["inputs_embeds", "attention_mask"])
      };
    }
    let { last_hidden_state } = await encoder_forward(this, model_inputs);
    if (generation_config.guidance_scale !== null && generation_config.guidance_scale > 1) {
      last_hidden_state = cat([last_hidden_state, full_like(last_hidden_state, 0)], 0);
      if ("attention_mask" in model_inputs) {
        model_inputs["attention_mask"] = cat(
          [model_inputs["attention_mask"], zeros_like(model_inputs["attention_mask"])],
          0
        );
      }
    } else if (model_inputs.decoder_input_ids) {
      const decoder_input_ids_batch_size = toI64Tensor(model_inputs.decoder_input_ids).dims[0];
      if (decoder_input_ids_batch_size !== last_hidden_state.dims[0]) {
        if (last_hidden_state.dims[0] !== 1) {
          throw new Error(
            `The encoder outputs have a different batch size (${last_hidden_state.dims[0]}) than the decoder inputs (${decoder_input_ids_batch_size}).`
          );
        }
        last_hidden_state = cat(
          Array.from({ length: decoder_input_ids_batch_size }, () => last_hidden_state),
          0
        );
      }
    }
    model_inputs["encoder_outputs"] = last_hidden_state;
    return model_inputs;
  }
  /**
   * Prepares `decoder_input_ids` for generation with encoder-decoder models
   * @param {*} param0
   */
  _prepare_decoder_input_ids_for_generation({
    batch_size,
    model_input_name,
    model_kwargs,
    decoder_start_token_id,
    bos_token_id,
    generation_config
  }) {
    let { decoder_input_ids, ...model_inputs } = model_kwargs;
    if (!(decoder_input_ids instanceof Tensor2)) {
      if (!decoder_input_ids) {
        decoder_start_token_id ??= bos_token_id;
        if (this.config.model_type === "musicgen") {
          decoder_input_ids = Array.from(
            {
              // @ts-expect-error TS2339
              length: batch_size * this.config.decoder.num_codebooks
            },
            () => [decoder_start_token_id]
          );
        } else if (Array.isArray(decoder_start_token_id)) {
          if (decoder_start_token_id.length !== batch_size) {
            throw new Error(
              `\`decoder_start_token_id\` expcted to have length ${batch_size} but got ${decoder_start_token_id.length}`
            );
          }
          decoder_input_ids = decoder_start_token_id;
        } else {
          decoder_input_ids = Array.from(
            {
              length: batch_size
            },
            () => [decoder_start_token_id]
          );
        }
      } else if (!Array.isArray(decoder_input_ids[0])) {
        decoder_input_ids = Array.from(
          {
            length: batch_size
          },
          () => decoder_input_ids
        );
      }
      decoder_input_ids = toI64Tensor(decoder_input_ids);
    }
    model_inputs["decoder_attention_mask"] = ones_like(decoder_input_ids);
    return { input_ids: decoder_input_ids, model_inputs };
  }
  /**
   * Generates sequences of token ids for models with a language modeling head.
   * @param {import('../generation/parameters.js').GenerationFunctionParameters} options
   * @returns {Promise<ModelOutput|Tensor>} The output of the model, which can contain the generated token ids, attentions, and scores.
   */
  async generate({
    inputs = null,
    generation_config = null,
    logits_processor = null,
    stopping_criteria = null,
    streamer = null,
    // inputs_attention_mask = null,
    ...kwargs
  }) {
    this._validate_model_class();
    generation_config = this._prepare_generation_config(generation_config, kwargs);
    let { inputs_tensor, model_inputs, model_input_name } = this._prepare_model_inputs({
      inputs,
      model_kwargs: (
        /** @type {Record<string, Tensor|number[]>} */
        kwargs
      )
    });
    const is_encoder_decoder = this.config.is_encoder_decoder;
    if (!is_encoder_decoder) {
    } else if (!("encoder_outputs" in model_inputs)) {
      model_inputs = await this._prepare_encoder_decoder_kwargs_for_generation({
        inputs_tensor,
        model_inputs,
        model_input_name,
        generation_config
      });
    }
    let input_ids;
    if (is_encoder_decoder) {
      ({ input_ids, model_inputs } = this._prepare_decoder_input_ids_for_generation({
        batch_size: model_inputs[model_input_name].dims.at(0),
        model_input_name,
        model_kwargs: model_inputs,
        decoder_start_token_id: generation_config.decoder_start_token_id,
        bos_token_id: generation_config.bos_token_id,
        generation_config
      }));
    } else {
      input_ids = model_inputs[model_input_name];
    }
    let input_ids_length = input_ids.dims.at(-1);
    if (generation_config.max_new_tokens !== null) {
      generation_config.max_length = input_ids_length + generation_config.max_new_tokens;
    }
    const prepared_logits_processor = this._get_logits_processor(
      generation_config,
      input_ids_length,
      logits_processor
    );
    const prepared_stopping_criteria = this._get_stopping_criteria(generation_config, stopping_criteria);
    const numInputs = model_inputs[model_input_name].dims.at(0);
    const sampler = LogitsSampler.getSampler(generation_config);
    const scores = new Array(numInputs).fill(0);
    const all_input_ids = input_ids.tolist();
    if (streamer) {
      streamer.put(all_input_ids);
    }
    let outputs;
    let attentions = {};
    let return_dict_items = {};
    while (true) {
      model_inputs = this.prepare_inputs_for_generation(all_input_ids, model_inputs, generation_config);
      outputs = await this.forward(model_inputs);
      if (generation_config.return_dict_in_generate) {
        if (generation_config.output_attentions) {
          const token_attentions = getAttentions(outputs);
          for (const key in token_attentions) {
            if (!(key in attentions)) {
              attentions[key] = [];
            }
            attentions[key].push(token_attentions[key]);
          }
        } else if (this._return_dict_in_generate_keys) {
          Object.assign(return_dict_items, pick(outputs, this._return_dict_in_generate_keys));
        }
      }
      const logits = outputs.logits.slice(null, -1, null).to("float32");
      const next_tokens_scores = prepared_logits_processor(all_input_ids, logits);
      const generated_input_ids = [];
      for (let batch_idx = 0; batch_idx < next_tokens_scores.dims.at(0); ++batch_idx) {
        const logs = next_tokens_scores[batch_idx];
        const sampledTokens = await sampler(logs);
        for (const [newTokenId, logProb] of sampledTokens) {
          const bigint = BigInt(newTokenId);
          scores[batch_idx] += logProb;
          all_input_ids[batch_idx].push(bigint);
          generated_input_ids.push([bigint]);
          break;
        }
      }
      if (streamer) {
        streamer.put(generated_input_ids);
      }
      const stop = prepared_stopping_criteria(all_input_ids);
      if (stop.every((x) => x)) {
        break;
      }
      model_inputs = this._update_model_kwargs_for_generation({
        generated_input_ids,
        outputs,
        model_inputs,
        is_encoder_decoder
      });
    }
    if (streamer) {
      streamer.end();
    }
    const sequences = new Tensor2("int64", all_input_ids.flat(), [all_input_ids.length, all_input_ids[0].length]);
    const past_key_values = getPastKeyValues(outputs, model_inputs.past_key_values);
    const cachedTensors = new Set(Object.values(past_key_values));
    for (const tensor of Object.values(outputs)) {
      if (tensor.location === "gpu-buffer" && !cachedTensors.has(tensor)) {
        tensor.dispose();
      }
    }
    const keepCacheAlive = "past_key_values" in kwargs || generation_config.return_dict_in_generate;
    if (!keepCacheAlive) {
      await past_key_values.dispose();
    }
    if (generation_config.return_dict_in_generate) {
      return {
        sequences,
        past_key_values,
        ...attentions,
        ...return_dict_items
        // TODO:
        // scores,
        // logits,
      };
    }
    return sequences;
  }
  /**
   * Helper function to select valid inputs and run through the appropriate encoder (vision, text, audio) based on the input type.
   * @param {string} sessionName
   * @param {Record<string, Tensor>} inputs
   * @param {string} outputName
   * @private
   */
  async _encode_input(sessionName, inputs, outputName) {
    if (!Object.hasOwn(this.sessions, sessionName)) {
      throw new Error(`Model does not have a ${sessionName} session.`);
    }
    const session = this.sessions[sessionName];
    const output = await sessionRun(session, pick(inputs, session.inputNames));
    return output[outputName];
  }
  async encode_image(inputs) {
    return this._encode_input("vision_encoder", inputs, "image_features");
  }
  async encode_text(inputs) {
    return this._encode_input("embed_tokens", inputs, "inputs_embeds");
  }
  async encode_audio(inputs) {
    return this._encode_input("audio_encoder", inputs, "audio_features");
  }
};
async function seq2seq_forward(self2, model_inputs) {
  let { encoder_outputs, input_ids, decoder_input_ids, decoder_attention_mask, ...other_decoder_inputs } = model_inputs;
  if (!encoder_outputs) {
    const encoder_inputs = pick(model_inputs, self2.sessions["model"].inputNames);
    encoder_outputs = (await encoder_forward(self2, encoder_inputs)).last_hidden_state;
  }
  other_decoder_inputs.input_ids = decoder_input_ids;
  other_decoder_inputs.encoder_hidden_states = encoder_outputs;
  if (self2.sessions["decoder_model_merged"].inputNames.includes("encoder_attention_mask")) {
    other_decoder_inputs.encoder_attention_mask = model_inputs.attention_mask;
  }
  if (decoder_attention_mask && !other_decoder_inputs.attention_mask) {
    other_decoder_inputs.attention_mask = decoder_attention_mask;
  }
  return await decoder_forward(self2, other_decoder_inputs, true);
}
async function encoder_forward(self2, model_inputs) {
  const session = self2.sessions["model"];
  const encoderFeeds = pick(model_inputs, session.inputNames);
  if (session.inputNames.includes("inputs_embeds") && !encoderFeeds.inputs_embeds) {
    if (!model_inputs.input_ids) {
      throw new Error("Both `input_ids` and `inputs_embeds` are missing in the model inputs.");
    }
    encoderFeeds.inputs_embeds = await self2.encode_text({ input_ids: model_inputs.input_ids });
  }
  if (session.inputNames.includes("token_type_ids") && !encoderFeeds.token_type_ids) {
    if (!encoderFeeds.input_ids) {
      throw new Error("Both `input_ids` and `token_type_ids` are missing in the model inputs.");
    }
    encoderFeeds.token_type_ids = zeros_like(encoderFeeds.input_ids);
  }
  if (session.inputNames.includes("pixel_mask") && !encoderFeeds.pixel_mask) {
    if (!encoderFeeds.pixel_values) {
      throw new Error("Both `pixel_values` and `pixel_mask` are missing in the model inputs.");
    }
    const dims = encoderFeeds.pixel_values.dims;
    encoderFeeds.pixel_mask = ones([dims[0], dims[2], dims[3]]);
  }
  return await sessionRun(session, encoderFeeds);
}
async function auto_encoder_forward(self2, model_inputs) {
  const encoded = await self2.encode(model_inputs);
  const decoded = await self2.decode(encoded);
  return decoded;
}
function getPastKeyValues(decoderResults, pastKeyValues) {
  const pkvs = /* @__PURE__ */ Object.create(null);
  for (const name in decoderResults) {
    if (name.startsWith("present")) {
      const newName = name.replace("present_ssm", "past_ssm").replace("present_conv", "past_conv").replace("present_recurrent", "past_recurrent").replace("present", "past_key_values");
      const is_encoder_pkv = name.includes("encoder");
      if (is_encoder_pkv && pastKeyValues) {
        pkvs[newName] = pastKeyValues[newName];
      } else {
        pkvs[newName] = decoderResults[name];
      }
    }
  }
  if (pastKeyValues) {
    pastKeyValues.update(pkvs);
    return pastKeyValues;
  }
  return new DynamicCache(pkvs);
}
function getAttentions(model_output) {
  const attentions = {};
  for (const attnName of ["cross_attentions", "encoder_attentions", "decoder_attentions"]) {
    for (const name in model_output) {
      if (name.startsWith(attnName)) {
        if (!(attnName in attentions)) {
          attentions[attnName] = [];
        }
        attentions[attnName].push(model_output[name]);
      }
    }
  }
  return attentions;
}
function resolveCacheShape(metadataShape, symbols) {
  return metadataShape.map((d) => {
    if (typeof d === "number") return d;
    return symbols[d] ?? 0;
  });
}
function addPastKeyValues(self2, decoderFeeds, pastKeyValues) {
  if (pastKeyValues && Object.keys(pastKeyValues).length > 0) {
    Object.assign(decoderFeeds, pastKeyValues);
    return pastKeyValues;
  }
  const session = self2.sessions["decoder_model_merged"] ?? self2.sessions["model"];
  const batch_size = (decoderFeeds[self2.main_input_name] ?? decoderFeeds.attention_mask)?.dims?.[0] ?? 1;
  const names = getCacheNames(self2.config);
  const num_heads = self2.config?.normalized_config?.num_heads;
  const symbols = { batch_size };
  if (typeof num_heads === "number") {
    symbols["batch_size x num_heads"] = batch_size * num_heads;
  }
  const entries = /* @__PURE__ */ Object.create(null);
  for (const meta of session.inputMetadata) {
    if (!names.has(meta.name)) continue;
    const shape = resolveCacheShape(meta.shape, symbols);
    const size = shape.reduce((a, b) => a * b, 1);
    const cls = DataTypeMap[meta.type];
    const t = new Tensor2(meta.type, new cls(size), shape);
    decoderFeeds[meta.name] = t;
    entries[meta.name] = t;
  }
  if (pastKeyValues) {
    pastKeyValues.update(entries);
    return pastKeyValues;
  }
  return new DynamicCache(entries);
}
async function decoder_forward(self2, model_inputs, is_encoder_decoder = false) {
  const session = self2.sessions[is_encoder_decoder ? "decoder_model_merged" : "model"];
  const { past_key_values, ...new_model_inputs } = model_inputs;
  if (session.inputNames.includes("use_cache_branch")) {
    new_model_inputs.use_cache_branch = boolTensor(
      past_key_values != null && Object.keys(past_key_values).length > 0
    );
  }
  if (session.inputNames.includes("position_ids") && new_model_inputs.attention_mask && !new_model_inputs.position_ids) {
    const start_index = ["paligemma", "gemma3_text", "gemma3"].includes(self2.config.model_type) ? 1 : 0;
    new_model_inputs.position_ids = create_position_ids(new_model_inputs, past_key_values, start_index);
  }
  if (session.inputNames.includes("num_logits_to_keep") && !new_model_inputs.num_logits_to_keep) {
    new_model_inputs.num_logits_to_keep = new Tensor2("int64", [0n], []);
  }
  addPastKeyValues(self2, new_model_inputs, past_key_values);
  const fixed = pick(new_model_inputs, session.inputNames);
  return await sessionRun(session, fixed);
}
async function generic_text_to_text_forward(self2, {
  // Generic parameters:
  encode_function,
  merge_function,
  modality_input_names,
  modality_output_name,
  // Produced by the tokenizer/processor:
  input_ids = null,
  attention_mask = null,
  // Used during generation:
  position_ids = null,
  inputs_embeds = null,
  past_key_values = null,
  // Generic generation parameters
  generation_config = null,
  logits_processor = null,
  // Additional parameters
  ...kwargs
}) {
  if (!inputs_embeds) {
    inputs_embeds = await self2.encode_text({ input_ids, ...kwargs });
    const modality_values = pick(kwargs, modality_input_names);
    if (Object.keys(modality_values).length > 0) {
      if (input_ids.dims[1] !== 1) {
        const modality_features = await encode_function({
          // Pass the modality values under its expected key.
          // The caller knows whether this is audio or image.
          ...modality_values,
          ...kwargs
        });
        ({ inputs_embeds, attention_mask } = merge_function({
          [modality_output_name]: modality_features,
          inputs_embeds,
          input_ids,
          attention_mask
        }));
      } else if (past_key_values && input_ids.dims[1] === 1) {
        const target_length = input_ids.dims[1];
        const past_length = past_key_values.get_seq_length();
        attention_mask = cat(
          [
            ones([input_ids.dims[0], past_length]),
            attention_mask.slice(null, [attention_mask.dims[1] - target_length, attention_mask.dims[1]])
          ],
          1
        );
      }
    }
  }
  if (!position_ids) {
    if (
      // Handle special case for qwen vl models
      [
        "qwen2_vl",
        "qwen2_vl_text",
        "qwen2_5_vl",
        "qwen2_5_vl_text",
        "qwen3_vl",
        "qwen3_vl_text",
        "qwen3_vl_moe",
        "qwen3_vl_moe_text",
        "qwen3_5",
        "qwen3_5_text",
        "qwen3_5_moe",
        "qwen3_5_moe_text",
        "glm_ocr",
        "glm_ocr_text"
      ].includes(self2.config.model_type)
    ) {
      const { image_grid_thw, video_grid_thw } = kwargs;
      [position_ids] = self2.get_rope_index(input_ids, image_grid_thw, video_grid_thw, attention_mask);
    }
  }
  const outputs = await decoder_forward(
    self2,
    {
      inputs_embeds,
      past_key_values,
      attention_mask,
      position_ids,
      generation_config,
      logits_processor
    },
    true
  );
  return outputs;
}
async function audio_text_to_text_forward(self2, params) {
  return await generic_text_to_text_forward(self2, {
    ...params,
    modality_input_names: ["audio_values", "input_features"],
    modality_output_name: "audio_features",
    encode_function: self2.encode_audio.bind(self2),
    merge_function: self2._merge_input_ids_with_audio_features.bind(self2)
  });
}
async function image_text_to_text_forward(self2, params) {
  return await generic_text_to_text_forward(self2, {
    ...params,
    modality_input_names: ["pixel_values"],
    modality_output_name: "image_features",
    encode_function: self2.encode_image.bind(self2),
    merge_function: self2._merge_input_ids_with_image_features.bind(self2)
  });
}
function cumsum_masked_fill(attention_mask, start_index = 0) {
  const [bz, seq_len] = attention_mask.dims;
  const attn_mask_data = attention_mask.data;
  const data = new BigInt64Array(attn_mask_data.length);
  for (let i = 0; i < bz; ++i) {
    const start = i * seq_len;
    let sum = BigInt(start_index);
    for (let j = 0; j < seq_len; ++j) {
      const index = start + j;
      if (attn_mask_data[index] === 0n) {
        data[index] = BigInt(1);
      } else {
        data[index] = sum;
        sum += attn_mask_data[index];
      }
    }
  }
  return { data, dims: attention_mask.dims };
}
function create_position_ids(model_inputs, past_key_values = null, start_index = 0) {
  const { input_ids, inputs_embeds, attention_mask } = model_inputs;
  const { data, dims } = cumsum_masked_fill(attention_mask, start_index);
  let position_ids = new Tensor2("int64", data, dims);
  if (past_key_values) {
    const offset = -(input_ids ?? inputs_embeds).dims.at(1);
    position_ids = position_ids.slice(null, [offset, null]);
  }
  return position_ids;
}
function decoder_prepare_inputs_for_generation(self2, input_ids, model_inputs, generation_config) {
  const past_length = model_inputs.past_key_values ? model_inputs.past_key_values.get_seq_length() : 0;
  const session = self2.sessions["decoder_model_merged"] ?? self2.sessions["model"];
  if (session?.inputNames.includes("num_logits_to_keep") && !model_inputs.num_logits_to_keep) {
    model_inputs.num_logits_to_keep = new Tensor2("int64", [1n], []);
  }
  if (!model_inputs.attention_mask) {
    let dims;
    for (const key of ["input_ids", "inputs_embeds", "position_ids"]) {
      if (model_inputs[key]) {
        dims = model_inputs[key].dims;
        break;
      }
    }
    if (!dims) {
      throw new Error("attention_mask is not provided, and unable to infer its shape from model inputs.");
    }
    model_inputs.attention_mask = ones([dims[0], past_length + dims[1]]);
  }
  if (model_inputs.past_key_values) {
    const { input_ids: input_ids2, attention_mask } = model_inputs;
    if (attention_mask && attention_mask.dims[1] > input_ids2.dims[1]) {
    } else if (past_length < input_ids2.dims[1]) {
      model_inputs.input_ids = input_ids2.slice(null, [past_length, null]);
    } else {
    }
  }
  return model_inputs;
}
function encoder_decoder_prepare_inputs_for_generation(self2, input_ids, model_inputs, generation_config) {
  if (model_inputs.past_key_values) {
    input_ids = input_ids.map((x) => [x.at(-1)]);
  }
  return {
    ...model_inputs,
    decoder_input_ids: toI64Tensor(input_ids)
  };
}
function multimodal_text_to_text_prepare_inputs_for_generation(self2, ...args) {
  if (self2.config.is_encoder_decoder) {
    return encoder_decoder_prepare_inputs_for_generation(self2, ...args);
  } else {
    return decoder_prepare_inputs_for_generation(self2, ...args);
  }
}
function default_merge_input_ids_with_features({
  modality_token_id,
  inputs_embeds,
  modality_features,
  input_ids,
  attention_mask
}) {
  const token_positions = input_ids.tolist().map(
    (ids) => ids.reduce((acc, x, idx) => {
      if (x == modality_token_id) acc.push(idx);
      return acc;
    }, [])
  );
  const n_tokens = token_positions.reduce((acc, x) => acc + x.length, 0);
  const n_features = modality_features.dims[0];
  if (n_tokens !== n_features) {
    throw new Error(`Number of tokens and features do not match: tokens: ${n_tokens}, features ${n_features}`);
  }
  let img = 0;
  for (let i = 0; i < token_positions.length; ++i) {
    const tokens = token_positions[i];
    const embeds = inputs_embeds[i];
    for (let j = 0; j < tokens.length; ++j) {
      embeds[tokens[j]].data.set(modality_features[img++].data);
    }
  }
  return { inputs_embeds, attention_mask };
}
function default_merge_input_ids_with_image_features({
  image_token_id,
  inputs_embeds,
  image_features,
  input_ids,
  attention_mask
}) {
  return default_merge_input_ids_with_features({
    modality_token_id: image_token_id,
    inputs_embeds,
    modality_features: image_features,
    input_ids,
    attention_mask
  });
}
function default_merge_input_ids_with_audio_features({
  audio_token_id,
  inputs_embeds,
  audio_features,
  input_ids,
  attention_mask
}) {
  return default_merge_input_ids_with_features({
    modality_token_id: audio_token_id,
    inputs_embeds,
    modality_features: audio_features,
    input_ids,
    attention_mask
  });
}
async function get_optional_configs(pretrained_model_name_or_path, names, options) {
  return Object.fromEntries(
    await Promise.all(
      Object.keys(names).map(async (name) => {
        const config = await getModelJSON(pretrained_model_name_or_path, names[name], false, options);
        return [name, config];
      })
    )
  );
}

// src/models/models.js
var models_exports = {};
__export(models_exports, {
  ASTForAudioClassification: () => ASTForAudioClassification,
  ASTModel: () => ASTModel,
  ASTPreTrainedModel: () => ASTPreTrainedModel,
  AfmoeForCausalLM: () => AfmoeForCausalLM,
  AfmoeModel: () => AfmoeModel,
  AfmoePreTrainedModel: () => AfmoePreTrainedModel,
  AlbertForMaskedLM: () => AlbertForMaskedLM,
  AlbertForQuestionAnswering: () => AlbertForQuestionAnswering,
  AlbertForSequenceClassification: () => AlbertForSequenceClassification,
  AlbertModel: () => AlbertModel,
  AlbertPreTrainedModel: () => AlbertPreTrainedModel,
  ApertusForCausalLM: () => ApertusForCausalLM,
  ApertusModel: () => ApertusModel,
  ApertusPreTrainedModel: () => ApertusPreTrainedModel,
  ArceeForCausalLM: () => ArceeForCausalLM,
  ArceeModel: () => ArceeModel,
  ArceePreTrainedModel: () => ArceePreTrainedModel,
  BartForConditionalGeneration: () => BartForConditionalGeneration,
  BartForSequenceClassification: () => BartForSequenceClassification,
  BartModel: () => BartModel,
  BartPretrainedModel: () => BartPretrainedModel,
  BeitForImageClassification: () => BeitForImageClassification,
  BeitModel: () => BeitModel,
  BeitPreTrainedModel: () => BeitPreTrainedModel,
  BertForMaskedLM: () => BertForMaskedLM,
  BertForQuestionAnswering: () => BertForQuestionAnswering,
  BertForSequenceClassification: () => BertForSequenceClassification,
  BertForTokenClassification: () => BertForTokenClassification,
  BertModel: () => BertModel,
  BertPreTrainedModel: () => BertPreTrainedModel,
  BlenderbotForConditionalGeneration: () => BlenderbotForConditionalGeneration,
  BlenderbotModel: () => BlenderbotModel,
  BlenderbotPreTrainedModel: () => BlenderbotPreTrainedModel,
  BlenderbotSmallForConditionalGeneration: () => BlenderbotSmallForConditionalGeneration,
  BlenderbotSmallModel: () => BlenderbotSmallModel,
  BlenderbotSmallPreTrainedModel: () => BlenderbotSmallPreTrainedModel,
  BloomForCausalLM: () => BloomForCausalLM,
  BloomModel: () => BloomModel,
  BloomPreTrainedModel: () => BloomPreTrainedModel,
  CHMv2ForDepthEstimation: () => CHMv2ForDepthEstimation,
  CHMv2PreTrainedModel: () => CHMv2PreTrainedModel,
  CLIPModel: () => CLIPModel,
  CLIPPreTrainedModel: () => CLIPPreTrainedModel,
  CLIPSegForImageSegmentation: () => CLIPSegForImageSegmentation,
  CLIPSegModel: () => CLIPSegModel,
  CLIPSegPreTrainedModel: () => CLIPSegPreTrainedModel,
  CLIPTextModel: () => CLIPTextModel,
  CLIPTextModelWithProjection: () => CLIPTextModelWithProjection,
  CLIPVisionModel: () => CLIPVisionModel,
  CLIPVisionModelWithProjection: () => CLIPVisionModelWithProjection,
  CamembertForMaskedLM: () => CamembertForMaskedLM,
  CamembertForQuestionAnswering: () => CamembertForQuestionAnswering,
  CamembertForSequenceClassification: () => CamembertForSequenceClassification,
  CamembertForTokenClassification: () => CamembertForTokenClassification,
  CamembertModel: () => CamembertModel,
  CamembertPreTrainedModel: () => CamembertPreTrainedModel,
  ChatterboxModel: () => ChatterboxModel,
  ChatterboxPreTrainedModel: () => ChatterboxPreTrainedModel,
  ChineseCLIPModel: () => ChineseCLIPModel,
  ChineseCLIPPreTrainedModel: () => ChineseCLIPPreTrainedModel,
  ClapAudioModelWithProjection: () => ClapAudioModelWithProjection,
  ClapModel: () => ClapModel,
  ClapPreTrainedModel: () => ClapPreTrainedModel,
  ClapTextModelWithProjection: () => ClapTextModelWithProjection,
  CodeGenForCausalLM: () => CodeGenForCausalLM,
  CodeGenModel: () => CodeGenModel,
  CodeGenPreTrainedModel: () => CodeGenPreTrainedModel,
  Cohere2ForCausalLM: () => Cohere2ForCausalLM,
  Cohere2Model: () => Cohere2Model,
  Cohere2PreTrainedModel: () => Cohere2PreTrainedModel,
  CohereAsrForConditionalGeneration: () => CohereAsrForConditionalGeneration,
  CohereAsrModel: () => CohereAsrModel,
  CohereAsrPreTrainedModel: () => CohereAsrPreTrainedModel,
  CohereForCausalLM: () => CohereForCausalLM,
  CohereModel: () => CohereModel,
  CoherePreTrainedModel: () => CoherePreTrainedModel,
  ConvBertForMaskedLM: () => ConvBertForMaskedLM,
  ConvBertForQuestionAnswering: () => ConvBertForQuestionAnswering,
  ConvBertForSequenceClassification: () => ConvBertForSequenceClassification,
  ConvBertForTokenClassification: () => ConvBertForTokenClassification,
  ConvBertModel: () => ConvBertModel,
  ConvBertPreTrainedModel: () => ConvBertPreTrainedModel,
  ConvNextForImageClassification: () => ConvNextForImageClassification,
  ConvNextModel: () => ConvNextModel,
  ConvNextPreTrainedModel: () => ConvNextPreTrainedModel,
  ConvNextV2ForImageClassification: () => ConvNextV2ForImageClassification,
  ConvNextV2Model: () => ConvNextV2Model,
  ConvNextV2PreTrainedModel: () => ConvNextV2PreTrainedModel,
  DFineForObjectDetection: () => DFineForObjectDetection,
  DFineModel: () => DFineModel,
  DFinePreTrainedModel: () => DFinePreTrainedModel,
  DINOv3ConvNextModel: () => DINOv3ConvNextModel,
  DINOv3ConvNextPreTrainedModel: () => DINOv3ConvNextPreTrainedModel,
  DINOv3ViTModel: () => DINOv3ViTModel,
  DINOv3ViTPreTrainedModel: () => DINOv3ViTPreTrainedModel,
  DPTForDepthEstimation: () => DPTForDepthEstimation,
  DPTModel: () => DPTModel,
  DPTPreTrainedModel: () => DPTPreTrainedModel,
  DacDecoderModel: () => DacDecoderModel,
  DacDecoderOutput: () => DacDecoderOutput,
  DacEncoderModel: () => DacEncoderModel,
  DacEncoderOutput: () => DacEncoderOutput,
  DacModel: () => DacModel,
  DacPreTrainedModel: () => DacPreTrainedModel,
  DebertaForMaskedLM: () => DebertaForMaskedLM,
  DebertaForQuestionAnswering: () => DebertaForQuestionAnswering,
  DebertaForSequenceClassification: () => DebertaForSequenceClassification,
  DebertaForTokenClassification: () => DebertaForTokenClassification,
  DebertaModel: () => DebertaModel,
  DebertaPreTrainedModel: () => DebertaPreTrainedModel,
  DebertaV2ForMaskedLM: () => DebertaV2ForMaskedLM,
  DebertaV2ForQuestionAnswering: () => DebertaV2ForQuestionAnswering,
  DebertaV2ForSequenceClassification: () => DebertaV2ForSequenceClassification,
  DebertaV2ForTokenClassification: () => DebertaV2ForTokenClassification,
  DebertaV2Model: () => DebertaV2Model,
  DebertaV2PreTrainedModel: () => DebertaV2PreTrainedModel,
  DecisionTransformerModel: () => DecisionTransformerModel,
  DecisionTransformerPreTrainedModel: () => DecisionTransformerPreTrainedModel,
  DeepseekV3ForCausalLM: () => DeepseekV3ForCausalLM,
  DeepseekV3Model: () => DeepseekV3Model,
  DeepseekV3PreTrainedModel: () => DeepseekV3PreTrainedModel,
  DeiTForImageClassification: () => DeiTForImageClassification,
  DeiTModel: () => DeiTModel,
  DeiTPreTrainedModel: () => DeiTPreTrainedModel,
  DepthAnythingForDepthEstimation: () => DepthAnythingForDepthEstimation,
  DepthAnythingPreTrainedModel: () => DepthAnythingPreTrainedModel,
  DepthProForDepthEstimation: () => DepthProForDepthEstimation,
  DepthProPreTrainedModel: () => DepthProPreTrainedModel,
  DetrForObjectDetection: () => DetrForObjectDetection,
  DetrForSegmentation: () => DetrForSegmentation,
  DetrModel: () => DetrModel,
  DetrObjectDetectionOutput: () => DetrObjectDetectionOutput,
  DetrPreTrainedModel: () => DetrPreTrainedModel,
  DetrSegmentationOutput: () => DetrSegmentationOutput,
  Dinov2ForImageClassification: () => Dinov2ForImageClassification,
  Dinov2Model: () => Dinov2Model,
  Dinov2PreTrainedModel: () => Dinov2PreTrainedModel,
  Dinov2WithRegistersForImageClassification: () => Dinov2WithRegistersForImageClassification,
  Dinov2WithRegistersModel: () => Dinov2WithRegistersModel,
  Dinov2WithRegistersPreTrainedModel: () => Dinov2WithRegistersPreTrainedModel,
  DistilBertForMaskedLM: () => DistilBertForMaskedLM,
  DistilBertForQuestionAnswering: () => DistilBertForQuestionAnswering,
  DistilBertForSequenceClassification: () => DistilBertForSequenceClassification,
  DistilBertForTokenClassification: () => DistilBertForTokenClassification,
  DistilBertModel: () => DistilBertModel,
  DistilBertPreTrainedModel: () => DistilBertPreTrainedModel,
  DonutSwinModel: () => DonutSwinModel,
  DonutSwinPreTrainedModel: () => DonutSwinPreTrainedModel,
  EdgeTamModel: () => EdgeTamModel,
  EfficientNetForImageClassification: () => EfficientNetForImageClassification,
  EfficientNetModel: () => EfficientNetModel,
  EfficientNetPreTrainedModel: () => EfficientNetPreTrainedModel,
  ElectraForMaskedLM: () => ElectraForMaskedLM,
  ElectraForQuestionAnswering: () => ElectraForQuestionAnswering,
  ElectraForSequenceClassification: () => ElectraForSequenceClassification,
  ElectraForTokenClassification: () => ElectraForTokenClassification,
  ElectraModel: () => ElectraModel,
  ElectraPreTrainedModel: () => ElectraPreTrainedModel,
  Ernie4_5ForCausalLM: () => Ernie4_5ForCausalLM,
  Ernie4_5Model: () => Ernie4_5Model,
  Ernie4_5PretrainedModel: () => Ernie4_5PretrainedModel,
  EsmForMaskedLM: () => EsmForMaskedLM,
  EsmForSequenceClassification: () => EsmForSequenceClassification,
  EsmForTokenClassification: () => EsmForTokenClassification,
  EsmModel: () => EsmModel,
  EsmPreTrainedModel: () => EsmPreTrainedModel,
  EuroBertForMaskedLM: () => EuroBertForMaskedLM,
  EuroBertForSequenceClassification: () => EuroBertForSequenceClassification,
  EuroBertForTokenClassification: () => EuroBertForTokenClassification,
  EuroBertModel: () => EuroBertModel,
  EuroBertPreTrainedModel: () => EuroBertPreTrainedModel,
  ExaoneForCausalLM: () => ExaoneForCausalLM,
  ExaoneModel: () => ExaoneModel,
  ExaonePreTrainedModel: () => ExaonePreTrainedModel,
  FalconForCausalLM: () => FalconForCausalLM,
  FalconH1ForCausalLM: () => FalconH1ForCausalLM,
  FalconH1Model: () => FalconH1Model,
  FalconH1PreTrainedModel: () => FalconH1PreTrainedModel,
  FalconModel: () => FalconModel,
  FalconPreTrainedModel: () => FalconPreTrainedModel,
  FastViTForImageClassification: () => FastViTForImageClassification,
  FastViTModel: () => FastViTModel,
  FastViTPreTrainedModel: () => FastViTPreTrainedModel,
  Florence2ForConditionalGeneration: () => Florence2ForConditionalGeneration,
  Florence2PreTrainedModel: () => Florence2PreTrainedModel,
  GLPNForDepthEstimation: () => GLPNForDepthEstimation,
  GLPNModel: () => GLPNModel,
  GLPNPreTrainedModel: () => GLPNPreTrainedModel,
  GPT2LMHeadModel: () => GPT2LMHeadModel,
  GPT2Model: () => GPT2Model,
  GPT2PreTrainedModel: () => GPT2PreTrainedModel,
  GPTBigCodeForCausalLM: () => GPTBigCodeForCausalLM,
  GPTBigCodeModel: () => GPTBigCodeModel,
  GPTBigCodePreTrainedModel: () => GPTBigCodePreTrainedModel,
  GPTJForCausalLM: () => GPTJForCausalLM,
  GPTJModel: () => GPTJModel,
  GPTJPreTrainedModel: () => GPTJPreTrainedModel,
  GPTNeoForCausalLM: () => GPTNeoForCausalLM,
  GPTNeoModel: () => GPTNeoModel,
  GPTNeoPreTrainedModel: () => GPTNeoPreTrainedModel,
  GPTNeoXForCausalLM: () => GPTNeoXForCausalLM,
  GPTNeoXModel: () => GPTNeoXModel,
  GPTNeoXPreTrainedModel: () => GPTNeoXPreTrainedModel,
  Gemma2ForCausalLM: () => Gemma2ForCausalLM,
  Gemma2Model: () => Gemma2Model,
  Gemma2PreTrainedModel: () => Gemma2PreTrainedModel,
  Gemma3ForCausalLM: () => Gemma3ForCausalLM,
  Gemma3ForConditionalGeneration: () => Gemma3ForConditionalGeneration,
  Gemma3Model: () => Gemma3Model,
  Gemma3PreTrainedModel: () => Gemma3PreTrainedModel,
  Gemma3nForCausalLM: () => Gemma3nForCausalLM,
  Gemma3nForConditionalGeneration: () => Gemma3nForConditionalGeneration,
  Gemma3nPreTrainedModel: () => Gemma3nPreTrainedModel,
  Gemma4ForCausalLM: () => Gemma4ForCausalLM,
  Gemma4ForConditionalGeneration: () => Gemma4ForConditionalGeneration,
  GemmaForCausalLM: () => GemmaForCausalLM,
  GemmaModel: () => GemmaModel,
  GemmaPreTrainedModel: () => GemmaPreTrainedModel,
  GlmForCausalLM: () => GlmForCausalLM,
  GlmModel: () => GlmModel,
  GlmMoeDsaForCausalLM: () => GlmMoeDsaForCausalLM,
  GlmMoeDsaModel: () => GlmMoeDsaModel,
  GlmMoeDsaPreTrainedModel: () => GlmMoeDsaPreTrainedModel,
  GlmOcrForConditionalGeneration: () => GlmOcrForConditionalGeneration,
  GlmPreTrainedModel: () => GlmPreTrainedModel,
  GptOssForCausalLM: () => GptOssForCausalLM,
  GptOssModel: () => GptOssModel,
  GptOssPreTrainedModel: () => GptOssPreTrainedModel,
  GraniteForCausalLM: () => GraniteForCausalLM,
  GraniteModel: () => GraniteModel,
  GraniteMoeHybridForCausalLM: () => GraniteMoeHybridForCausalLM,
  GraniteMoeHybridModel: () => GraniteMoeHybridModel,
  GraniteMoeHybridPreTrainedModel: () => GraniteMoeHybridPreTrainedModel,
  GranitePreTrainedModel: () => GranitePreTrainedModel,
  GraniteSpeechForConditionalGeneration: () => GraniteSpeechForConditionalGeneration,
  GroundingDinoForObjectDetection: () => GroundingDinoForObjectDetection,
  GroundingDinoPreTrainedModel: () => GroundingDinoPreTrainedModel,
  GroupViTModel: () => GroupViTModel,
  GroupViTPreTrainedModel: () => GroupViTPreTrainedModel,
  HeliumForCausalLM: () => HeliumForCausalLM,
  HeliumModel: () => HeliumModel,
  HeliumPreTrainedModel: () => HeliumPreTrainedModel,
  HieraForImageClassification: () => HieraForImageClassification,
  HieraModel: () => HieraModel,
  HieraPreTrainedModel: () => HieraPreTrainedModel,
  HubertForCTC: () => HubertForCTC,
  HubertForSequenceClassification: () => HubertForSequenceClassification,
  HubertModel: () => HubertModel,
  HubertPreTrainedModel: () => HubertPreTrainedModel,
  HunYuanDenseV1ForCausalLM: () => HunYuanDenseV1ForCausalLM,
  HunYuanDenseV1Model: () => HunYuanDenseV1Model,
  HunYuanDenseV1PreTrainedModel: () => HunYuanDenseV1PreTrainedModel,
  IJepaForImageClassification: () => IJepaForImageClassification,
  IJepaModel: () => IJepaModel,
  IJepaPreTrainedModel: () => IJepaPreTrainedModel,
  Idefics3ForConditionalGeneration: () => Idefics3ForConditionalGeneration,
  JAISLMHeadModel: () => JAISLMHeadModel,
  JAISModel: () => JAISModel,
  JAISPreTrainedModel: () => JAISPreTrainedModel,
  JinaCLIPModel: () => JinaCLIPModel,
  JinaCLIPPreTrainedModel: () => JinaCLIPPreTrainedModel,
  JinaCLIPTextModel: () => JinaCLIPTextModel,
  JinaCLIPVisionModel: () => JinaCLIPVisionModel,
  Lfm2ForCausalLM: () => Lfm2ForCausalLM,
  Lfm2Model: () => Lfm2Model,
  Lfm2MoeForCausalLM: () => Lfm2MoeForCausalLM,
  Lfm2MoeModel: () => Lfm2MoeModel,
  Lfm2MoePreTrainedModel: () => Lfm2MoePreTrainedModel,
  Lfm2PreTrainedModel: () => Lfm2PreTrainedModel,
  Lfm2VlForConditionalGeneration: () => Lfm2VlForConditionalGeneration,
  LightOnOcrForConditionalGeneration: () => LightOnOcrForConditionalGeneration,
  LiteWhisperForConditionalGeneration: () => LiteWhisperForConditionalGeneration,
  Llama4ForCausalLM: () => Llama4ForCausalLM,
  Llama4PreTrainedModel: () => Llama4PreTrainedModel,
  LlamaForCausalLM: () => LlamaForCausalLM,
  LlamaModel: () => LlamaModel,
  LlamaPreTrainedModel: () => LlamaPreTrainedModel,
  LlavaForConditionalGeneration: () => LlavaForConditionalGeneration,
  LlavaOnevisionForConditionalGeneration: () => LlavaForConditionalGeneration,
  LlavaPreTrainedModel: () => LlavaPreTrainedModel,
  LlavaQwen2ForCausalLM: () => LlavaQwen2ForCausalLM,
  LongT5ForConditionalGeneration: () => LongT5ForConditionalGeneration,
  LongT5Model: () => LongT5Model,
  LongT5PreTrainedModel: () => LongT5PreTrainedModel,
  M2M100ForConditionalGeneration: () => M2M100ForConditionalGeneration,
  M2M100Model: () => M2M100Model,
  M2M100PreTrainedModel: () => M2M100PreTrainedModel,
  MBartForCausalLM: () => MBartForCausalLM,
  MBartForConditionalGeneration: () => MBartForConditionalGeneration,
  MBartForSequenceClassification: () => MBartForSequenceClassification,
  MBartModel: () => MBartModel,
  MBartPreTrainedModel: () => MBartPreTrainedModel,
  MPNetForMaskedLM: () => MPNetForMaskedLM,
  MPNetForQuestionAnswering: () => MPNetForQuestionAnswering,
  MPNetForSequenceClassification: () => MPNetForSequenceClassification,
  MPNetForTokenClassification: () => MPNetForTokenClassification,
  MPNetModel: () => MPNetModel,
  MPNetPreTrainedModel: () => MPNetPreTrainedModel,
  MT5ForConditionalGeneration: () => MT5ForConditionalGeneration,
  MT5Model: () => MT5Model,
  MT5PreTrainedModel: () => MT5PreTrainedModel,
  MarianMTModel: () => MarianMTModel,
  MarianModel: () => MarianModel,
  MarianPreTrainedModel: () => MarianPreTrainedModel,
  MaskFormerForInstanceSegmentation: () => MaskFormerForInstanceSegmentation,
  MaskFormerModel: () => MaskFormerModel,
  MaskFormerPreTrainedModel: () => MaskFormerPreTrainedModel,
  Metric3DForDepthEstimation: () => Metric3DForDepthEstimation,
  Metric3DPreTrainedModel: () => Metric3DPreTrainedModel,
  Metric3Dv2ForDepthEstimation: () => Metric3Dv2ForDepthEstimation,
  Metric3Dv2PreTrainedModel: () => Metric3Dv2PreTrainedModel,
  MgpstrForSceneTextRecognition: () => MgpstrForSceneTextRecognition,
  MgpstrModelOutput: () => MgpstrModelOutput,
  MgpstrPreTrainedModel: () => MgpstrPreTrainedModel,
  MimiDecoderModel: () => MimiDecoderModel,
  MimiDecoderOutput: () => MimiDecoderOutput,
  MimiEncoderModel: () => MimiEncoderModel,
  MimiEncoderOutput: () => MimiEncoderOutput,
  MimiModel: () => MimiModel,
  MimiPreTrainedModel: () => MimiPreTrainedModel,
  Mistral4ForCausalLM: () => Mistral4ForCausalLM,
  Mistral4Model: () => Mistral4Model,
  Mistral4PreTrainedModel: () => Mistral4PreTrainedModel,
  MistralForCausalLM: () => MistralForCausalLM,
  MistralModel: () => MistralModel,
  MistralPreTrainedModel: () => MistralPreTrainedModel,
  MobileBertForMaskedLM: () => MobileBertForMaskedLM,
  MobileBertForQuestionAnswering: () => MobileBertForQuestionAnswering,
  MobileBertForSequenceClassification: () => MobileBertForSequenceClassification,
  MobileBertModel: () => MobileBertModel,
  MobileBertPreTrainedModel: () => MobileBertPreTrainedModel,
  MobileLLMForCausalLM: () => MobileLLMForCausalLM,
  MobileLLMModel: () => MobileLLMModel,
  MobileLLMPreTrainedModel: () => MobileLLMPreTrainedModel,
  MobileNetV1ForImageClassification: () => MobileNetV1ForImageClassification,
  MobileNetV1ForSemanticSegmentation: () => MobileNetV1ForSemanticSegmentation,
  MobileNetV1Model: () => MobileNetV1Model,
  MobileNetV1PreTrainedModel: () => MobileNetV1PreTrainedModel,
  MobileNetV2ForImageClassification: () => MobileNetV2ForImageClassification,
  MobileNetV2ForSemanticSegmentation: () => MobileNetV2ForSemanticSegmentation,
  MobileNetV2Model: () => MobileNetV2Model,
  MobileNetV2PreTrainedModel: () => MobileNetV2PreTrainedModel,
  MobileNetV3ForImageClassification: () => MobileNetV3ForImageClassification,
  MobileNetV3ForSemanticSegmentation: () => MobileNetV3ForSemanticSegmentation,
  MobileNetV3Model: () => MobileNetV3Model,
  MobileNetV3PreTrainedModel: () => MobileNetV3PreTrainedModel,
  MobileNetV4ForImageClassification: () => MobileNetV4ForImageClassification,
  MobileNetV4ForSemanticSegmentation: () => MobileNetV4ForSemanticSegmentation,
  MobileNetV4Model: () => MobileNetV4Model,
  MobileNetV4PreTrainedModel: () => MobileNetV4PreTrainedModel,
  MobileViTForImageClassification: () => MobileViTForImageClassification,
  MobileViTModel: () => MobileViTModel,
  MobileViTPreTrainedModel: () => MobileViTPreTrainedModel,
  MobileViTV2ForImageClassification: () => MobileViTV2ForImageClassification,
  MobileViTV2Model: () => MobileViTV2Model,
  MobileViTV2PreTrainedModel: () => MobileViTV2PreTrainedModel,
  ModernBertDecoderForCausalLM: () => ModernBertDecoderForCausalLM,
  ModernBertDecoderModel: () => ModernBertDecoderModel,
  ModernBertDecoderPreTrainedModel: () => ModernBertDecoderPreTrainedModel,
  ModernBertForMaskedLM: () => ModernBertForMaskedLM,
  ModernBertForSequenceClassification: () => ModernBertForSequenceClassification,
  ModernBertForTokenClassification: () => ModernBertForTokenClassification,
  ModernBertModel: () => ModernBertModel,
  ModernBertPreTrainedModel: () => ModernBertPreTrainedModel,
  Moondream1ForConditionalGeneration: () => Moondream1ForConditionalGeneration,
  MoonshineForConditionalGeneration: () => MoonshineForConditionalGeneration,
  MoonshineModel: () => MoonshineModel,
  MoonshinePreTrainedModel: () => MoonshinePreTrainedModel,
  MptForCausalLM: () => MptForCausalLM,
  MptModel: () => MptModel,
  MptPreTrainedModel: () => MptPreTrainedModel,
  MultiModalityCausalLM: () => MultiModalityCausalLM,
  MultiModalityPreTrainedModel: () => MultiModalityPreTrainedModel,
  MusicgenForCausalLM: () => MusicgenForCausalLM,
  MusicgenForConditionalGeneration: () => MusicgenForConditionalGeneration,
  MusicgenModel: () => MusicgenModel,
  MusicgenPreTrainedModel: () => MusicgenPreTrainedModel,
  NanoChatForCausalLM: () => NanoChatForCausalLM,
  NanoChatModel: () => NanoChatModel,
  NanoChatPreTrainedModel: () => NanoChatPreTrainedModel,
  NemotronHForCausalLM: () => NemotronHForCausalLM,
  NemotronHModel: () => NemotronHModel,
  NemotronHPreTrainedModel: () => NemotronHPreTrainedModel,
  NeoBertForMaskedLM: () => NeoBertForMaskedLM,
  NeoBertForQuestionAnswering: () => NeoBertForQuestionAnswering,
  NeoBertForSequenceClassification: () => NeoBertForSequenceClassification,
  NeoBertForTokenClassification: () => NeoBertForTokenClassification,
  NeoBertModel: () => NeoBertModel,
  NeoBertPreTrainedModel: () => NeoBertPreTrainedModel,
  NomicBertModel: () => NomicBertModel,
  NomicBertPreTrainedModel: () => NomicBertPreTrainedModel,
  OPTForCausalLM: () => OPTForCausalLM,
  OPTModel: () => OPTModel,
  OPTPreTrainedModel: () => OPTPreTrainedModel,
  Olmo2ForCausalLM: () => Olmo2ForCausalLM,
  Olmo2Model: () => Olmo2Model,
  Olmo2PreTrainedModel: () => Olmo2PreTrainedModel,
  Olmo3ForCausalLM: () => Olmo3ForCausalLM,
  Olmo3Model: () => Olmo3Model,
  Olmo3PreTrainedModel: () => Olmo3PreTrainedModel,
  OlmoForCausalLM: () => OlmoForCausalLM,
  OlmoHybridForCausalLM: () => OlmoHybridForCausalLM,
  OlmoHybridModel: () => OlmoHybridModel,
  OlmoHybridPreTrainedModel: () => OlmoHybridPreTrainedModel,
  OlmoModel: () => OlmoModel,
  OlmoPreTrainedModel: () => OlmoPreTrainedModel,
  OpenAIPrivacyFilterForTokenClassification: () => OpenAIPrivacyFilterForTokenClassification,
  OpenAIPrivacyFilterModel: () => OpenAIPrivacyFilterModel,
  OpenAIPrivacyFilterPreTrainedModel: () => OpenAIPrivacyFilterPreTrainedModel,
  OpenELMForCausalLM: () => OpenELMForCausalLM,
  OpenELMModel: () => OpenELMModel,
  OpenELMPreTrainedModel: () => OpenELMPreTrainedModel,
  OwlViTForObjectDetection: () => OwlViTForObjectDetection,
  OwlViTModel: () => OwlViTModel,
  OwlViTPreTrainedModel: () => OwlViTPreTrainedModel,
  Owlv2ForObjectDetection: () => Owlv2ForObjectDetection,
  Owlv2Model: () => Owlv2Model,
  Owlv2PreTrainedModel: () => Owlv2PreTrainedModel,
  PaliGemmaForConditionalGeneration: () => PaliGemmaForConditionalGeneration,
  ParakeetForCTC: () => ParakeetForCTC,
  ParakeetPreTrainedModel: () => ParakeetPreTrainedModel,
  PatchTSMixerForPrediction: () => PatchTSMixerForPrediction,
  PatchTSMixerModel: () => PatchTSMixerModel,
  PatchTSMixerPreTrainedModel: () => PatchTSMixerPreTrainedModel,
  PatchTSTForPrediction: () => PatchTSTForPrediction,
  PatchTSTModel: () => PatchTSTModel,
  PatchTSTPreTrainedModel: () => PatchTSTPreTrainedModel,
  Phi3ForCausalLM: () => Phi3ForCausalLM,
  Phi3Model: () => Phi3Model,
  Phi3PreTrainedModel: () => Phi3PreTrainedModel,
  Phi3VForCausalLM: () => Phi3VForCausalLM,
  Phi3VPreTrainedModel: () => Phi3VPreTrainedModel,
  PhiForCausalLM: () => PhiForCausalLM,
  PhiModel: () => PhiModel,
  PhiPreTrainedModel: () => PhiPreTrainedModel,
  PreTrainedModel: () => PreTrainedModel,
  PvtForImageClassification: () => PvtForImageClassification,
  PvtModel: () => PvtModel,
  PvtPreTrainedModel: () => PvtPreTrainedModel,
  PyAnnoteForAudioFrameClassification: () => PyAnnoteForAudioFrameClassification,
  PyAnnoteModel: () => PyAnnoteModel,
  PyAnnotePreTrainedModel: () => PyAnnotePreTrainedModel,
  Qwen2ForCausalLM: () => Qwen2ForCausalLM,
  Qwen2Model: () => Qwen2Model,
  Qwen2MoeForCausalLM: () => Qwen2MoeForCausalLM,
  Qwen2MoeModel: () => Qwen2MoeModel,
  Qwen2MoePreTrainedModel: () => Qwen2MoePreTrainedModel,
  Qwen2PreTrainedModel: () => Qwen2PreTrainedModel,
  Qwen2VLForCausalLM: () => Qwen2VLForCausalLM,
  Qwen2VLForConditionalGeneration: () => Qwen2VLForConditionalGeneration,
  Qwen2VLPreTrainedModel: () => Qwen2VLPreTrainedModel,
  Qwen2_5_VLForCausalLM: () => Qwen2_5_VLForCausalLM,
  Qwen2_5_VLForConditionalGeneration: () => Qwen2_5_VLForConditionalGeneration,
  Qwen3ForCausalLM: () => Qwen3ForCausalLM,
  Qwen3Model: () => Qwen3Model,
  Qwen3MoeForCausalLM: () => Qwen3MoeForCausalLM,
  Qwen3MoeModel: () => Qwen3MoeModel,
  Qwen3MoePreTrainedModel: () => Qwen3MoePreTrainedModel,
  Qwen3NextForCausalLM: () => Qwen3NextForCausalLM,
  Qwen3NextModel: () => Qwen3NextModel,
  Qwen3NextPreTrainedModel: () => Qwen3NextPreTrainedModel,
  Qwen3PreTrainedModel: () => Qwen3PreTrainedModel,
  Qwen3VLForCausalLM: () => Qwen3VLForCausalLM,
  Qwen3VLForConditionalGeneration: () => Qwen3VLForConditionalGeneration,
  Qwen3VLMoeForCausalLM: () => Qwen3VLMoeForCausalLM,
  Qwen3VLMoeForConditionalGeneration: () => Qwen3VLMoeForConditionalGeneration,
  Qwen3_5ForCausalLM: () => Qwen3_5ForCausalLM,
  Qwen3_5ForConditionalGeneration: () => Qwen3_5ForConditionalGeneration,
  Qwen3_5MoeForCausalLM: () => Qwen3_5MoeForCausalLM,
  Qwen3_5MoeForConditionalGeneration: () => Qwen3_5MoeForConditionalGeneration,
  RFDetrForObjectDetection: () => RFDetrForObjectDetection,
  RFDetrModel: () => RFDetrModel,
  RFDetrObjectDetectionOutput: () => RFDetrObjectDetectionOutput,
  RFDetrPreTrainedModel: () => RFDetrPreTrainedModel,
  RTDetrForObjectDetection: () => RTDetrForObjectDetection,
  RTDetrModel: () => RTDetrModel,
  RTDetrObjectDetectionOutput: () => RTDetrObjectDetectionOutput,
  RTDetrPreTrainedModel: () => RTDetrPreTrainedModel,
  RTDetrV2ForObjectDetection: () => RTDetrV2ForObjectDetection,
  RTDetrV2Model: () => RTDetrV2Model,
  RTDetrV2ObjectDetectionOutput: () => RTDetrV2ObjectDetectionOutput,
  RTDetrV2PreTrainedModel: () => RTDetrV2PreTrainedModel,
  ResNetForImageClassification: () => ResNetForImageClassification,
  ResNetModel: () => ResNetModel,
  ResNetPreTrainedModel: () => ResNetPreTrainedModel,
  RoFormerForMaskedLM: () => RoFormerForMaskedLM,
  RoFormerForQuestionAnswering: () => RoFormerForQuestionAnswering,
  RoFormerForSequenceClassification: () => RoFormerForSequenceClassification,
  RoFormerForTokenClassification: () => RoFormerForTokenClassification,
  RoFormerModel: () => RoFormerModel,
  RoFormerPreTrainedModel: () => RoFormerPreTrainedModel,
  RobertaForMaskedLM: () => RobertaForMaskedLM,
  RobertaForQuestionAnswering: () => RobertaForQuestionAnswering,
  RobertaForSequenceClassification: () => RobertaForSequenceClassification,
  RobertaForTokenClassification: () => RobertaForTokenClassification,
  RobertaModel: () => RobertaModel,
  RobertaPreTrainedModel: () => RobertaPreTrainedModel,
  Sam2ImageSegmentationOutput: () => Sam2ImageSegmentationOutput,
  Sam2Model: () => Sam2Model,
  Sam2PreTrainedModel: () => Sam2PreTrainedModel,
  Sam3TrackerModel: () => Sam3TrackerModel,
  SamImageSegmentationOutput: () => SamImageSegmentationOutput,
  SamModel: () => SamModel,
  SamPreTrainedModel: () => SamPreTrainedModel,
  SapiensForDepthEstimation: () => SapiensForDepthEstimation,
  SapiensForNormalEstimation: () => SapiensForNormalEstimation,
  SapiensForSemanticSegmentation: () => SapiensForSemanticSegmentation,
  SapiensPreTrainedModel: () => SapiensPreTrainedModel,
  SegformerForImageClassification: () => SegformerForImageClassification,
  SegformerForSemanticSegmentation: () => SegformerForSemanticSegmentation,
  SegformerModel: () => SegformerModel,
  SegformerPreTrainedModel: () => SegformerPreTrainedModel,
  SiglipModel: () => SiglipModel,
  SiglipPreTrainedModel: () => SiglipPreTrainedModel,
  SiglipTextModel: () => SiglipTextModel,
  SiglipVisionModel: () => SiglipVisionModel,
  SmolLM3ForCausalLM: () => SmolLM3ForCausalLM,
  SmolLM3Model: () => SmolLM3Model,
  SmolLM3PreTrainedModel: () => SmolLM3PreTrainedModel,
  SmolVLMForConditionalGeneration: () => SmolVLMForConditionalGeneration,
  SnacDecoderModel: () => SnacDecoderModel,
  SnacEncoderModel: () => SnacEncoderModel,
  SnacModel: () => SnacModel,
  SnacPreTrainedModel: () => SnacPreTrainedModel,
  SolarOpenForCausalLM: () => SolarOpenForCausalLM,
  SolarOpenModel: () => SolarOpenModel,
  SolarOpenPreTrainedModel: () => SolarOpenPreTrainedModel,
  SpeechT5ForSpeechToText: () => SpeechT5ForSpeechToText,
  SpeechT5ForTextToSpeech: () => SpeechT5ForTextToSpeech,
  SpeechT5HifiGan: () => SpeechT5HifiGan,
  SpeechT5Model: () => SpeechT5Model,
  SpeechT5PreTrainedModel: () => SpeechT5PreTrainedModel,
  SqueezeBertForMaskedLM: () => SqueezeBertForMaskedLM,
  SqueezeBertForQuestionAnswering: () => SqueezeBertForQuestionAnswering,
  SqueezeBertForSequenceClassification: () => SqueezeBertForSequenceClassification,
  SqueezeBertModel: () => SqueezeBertModel,
  SqueezeBertPreTrainedModel: () => SqueezeBertPreTrainedModel,
  StableLmForCausalLM: () => StableLmForCausalLM,
  StableLmModel: () => StableLmModel,
  StableLmPreTrainedModel: () => StableLmPreTrainedModel,
  Starcoder2ForCausalLM: () => Starcoder2ForCausalLM,
  Starcoder2Model: () => Starcoder2Model,
  Starcoder2PreTrainedModel: () => Starcoder2PreTrainedModel,
  StyleTextToSpeech2Model: () => StyleTextToSpeech2Model,
  StyleTextToSpeech2PreTrainedModel: () => StyleTextToSpeech2PreTrainedModel,
  SupertonicForConditionalGeneration: () => SupertonicForConditionalGeneration,
  SupertonicPreTrainedModel: () => SupertonicPreTrainedModel,
  Swin2SRForImageSuperResolution: () => Swin2SRForImageSuperResolution,
  Swin2SRModel: () => Swin2SRModel,
  Swin2SRPreTrainedModel: () => Swin2SRPreTrainedModel,
  SwinForImageClassification: () => SwinForImageClassification,
  SwinForSemanticSegmentation: () => SwinForSemanticSegmentation,
  SwinModel: () => SwinModel,
  SwinPreTrainedModel: () => SwinPreTrainedModel,
  T5ForConditionalGeneration: () => T5ForConditionalGeneration,
  T5Model: () => T5Model,
  T5PreTrainedModel: () => T5PreTrainedModel,
  TableTransformerForObjectDetection: () => TableTransformerForObjectDetection,
  TableTransformerModel: () => TableTransformerModel,
  TableTransformerObjectDetectionOutput: () => TableTransformerObjectDetectionOutput,
  TableTransformerPreTrainedModel: () => TableTransformerPreTrainedModel,
  TrOCRForCausalLM: () => TrOCRForCausalLM,
  TrOCRPreTrainedModel: () => TrOCRPreTrainedModel,
  UltravoxModel: () => UltravoxModel,
  UltravoxPreTrainedModel: () => UltravoxPreTrainedModel,
  UniSpeechForCTC: () => UniSpeechForCTC,
  UniSpeechForSequenceClassification: () => UniSpeechForSequenceClassification,
  UniSpeechModel: () => UniSpeechModel,
  UniSpeechPreTrainedModel: () => UniSpeechPreTrainedModel,
  UniSpeechSatForAudioFrameClassification: () => UniSpeechSatForAudioFrameClassification,
  UniSpeechSatForCTC: () => UniSpeechSatForCTC,
  UniSpeechSatForSequenceClassification: () => UniSpeechSatForSequenceClassification,
  UniSpeechSatModel: () => UniSpeechSatModel,
  UniSpeechSatPreTrainedModel: () => UniSpeechSatPreTrainedModel,
  VaultGemmaForCausalLM: () => VaultGemmaForCausalLM,
  VaultGemmaModel: () => VaultGemmaModel,
  VaultGemmaPreTrainedModel: () => VaultGemmaPreTrainedModel,
  ViTForImageClassification: () => ViTForImageClassification,
  ViTMAEModel: () => ViTMAEModel,
  ViTMAEPreTrainedModel: () => ViTMAEPreTrainedModel,
  ViTMSNForImageClassification: () => ViTMSNForImageClassification,
  ViTMSNModel: () => ViTMSNModel,
  ViTMSNPreTrainedModel: () => ViTMSNPreTrainedModel,
  ViTModel: () => ViTModel,
  ViTPreTrainedModel: () => ViTPreTrainedModel,
  VisionEncoderDecoderModel: () => VisionEncoderDecoderModel,
  VitMatteForImageMatting: () => VitMatteForImageMatting,
  VitMattePreTrainedModel: () => VitMattePreTrainedModel,
  VitPoseForPoseEstimation: () => VitPoseForPoseEstimation,
  VitPosePreTrainedModel: () => VitPosePreTrainedModel,
  VitsModel: () => VitsModel,
  VitsModelOutput: () => VitsModelOutput,
  VitsPreTrainedModel: () => VitsPreTrainedModel,
  VoxtralForConditionalGeneration: () => VoxtralForConditionalGeneration,
  VoxtralRealtimeForConditionalGeneration: () => VoxtralRealtimeForConditionalGeneration,
  VoxtralRealtimePreTrainedModel: () => VoxtralRealtimePreTrainedModel,
  Wav2Vec2BertForCTC: () => Wav2Vec2BertForCTC,
  Wav2Vec2BertForSequenceClassification: () => Wav2Vec2BertForSequenceClassification,
  Wav2Vec2BertModel: () => Wav2Vec2BertModel,
  Wav2Vec2BertPreTrainedModel: () => Wav2Vec2BertPreTrainedModel,
  Wav2Vec2ForAudioFrameClassification: () => Wav2Vec2ForAudioFrameClassification,
  Wav2Vec2ForCTC: () => Wav2Vec2ForCTC,
  Wav2Vec2ForSequenceClassification: () => Wav2Vec2ForSequenceClassification,
  Wav2Vec2Model: () => Wav2Vec2Model,
  Wav2Vec2PreTrainedModel: () => Wav2Vec2PreTrainedModel,
  WavLMForAudioFrameClassification: () => WavLMForAudioFrameClassification,
  WavLMForCTC: () => WavLMForCTC,
  WavLMForSequenceClassification: () => WavLMForSequenceClassification,
  WavLMForXVector: () => WavLMForXVector,
  WavLMModel: () => WavLMModel,
  WavLMPreTrainedModel: () => WavLMPreTrainedModel,
  WeSpeakerResNetModel: () => WeSpeakerResNetModel,
  WeSpeakerResNetPreTrainedModel: () => WeSpeakerResNetPreTrainedModel,
  WhisperForConditionalGeneration: () => WhisperForConditionalGeneration,
  WhisperModel: () => WhisperModel,
  WhisperPreTrainedModel: () => WhisperPreTrainedModel,
  XLMForQuestionAnswering: () => XLMForQuestionAnswering,
  XLMForSequenceClassification: () => XLMForSequenceClassification,
  XLMForTokenClassification: () => XLMForTokenClassification,
  XLMModel: () => XLMModel,
  XLMPreTrainedModel: () => XLMPreTrainedModel,
  XLMRobertaForMaskedLM: () => XLMRobertaForMaskedLM,
  XLMRobertaForQuestionAnswering: () => XLMRobertaForQuestionAnswering,
  XLMRobertaForSequenceClassification: () => XLMRobertaForSequenceClassification,
  XLMRobertaForTokenClassification: () => XLMRobertaForTokenClassification,
  XLMRobertaModel: () => XLMRobertaModel,
  XLMRobertaPreTrainedModel: () => XLMRobertaPreTrainedModel,
  XLMWithLMHeadModel: () => XLMWithLMHeadModel,
  XVectorOutput: () => XVectorOutput,
  YolosForObjectDetection: () => YolosForObjectDetection,
  YolosModel: () => YolosModel,
  YolosObjectDetectionOutput: () => YolosObjectDetectionOutput,
  YolosPreTrainedModel: () => YolosPreTrainedModel,
  YoutuForCausalLM: () => YoutuForCausalLM,
  YoutuModel: () => YoutuModel,
  YoutuPreTrainedModel: () => YoutuPreTrainedModel
});

// src/models/albert/modeling_albert.js
var AlbertPreTrainedModel = class extends PreTrainedModel {
};
var AlbertModel = class extends AlbertPreTrainedModel {
};
var AlbertForSequenceClassification = class extends AlbertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<SequenceClassifierOutput>} returned object
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};
var AlbertForQuestionAnswering = class extends AlbertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<QuestionAnsweringModelOutput>} returned object
   */
  async _call(model_inputs) {
    return new QuestionAnsweringModelOutput(await super._call(model_inputs));
  }
};
var AlbertForMaskedLM = class extends AlbertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<MaskedLMOutput>} returned object
   */
  async _call(model_inputs) {
    return new MaskedLMOutput(await super._call(model_inputs));
  }
};

// src/models/apertus/modeling_apertus.js
var ApertusPreTrainedModel = class extends PreTrainedModel {
};
var ApertusModel = class extends ApertusPreTrainedModel {
};
var ApertusForCausalLM = class extends ApertusPreTrainedModel {
};

// src/models/afmoe/modeling_afmoe.js
var AfmoePreTrainedModel = class extends PreTrainedModel {
};
var AfmoeModel = class extends AfmoePreTrainedModel {
};
var AfmoeForCausalLM = class extends AfmoePreTrainedModel {
};

// src/models/arcee/modeling_arcee.js
var ArceePreTrainedModel = class extends PreTrainedModel {
};
var ArceeModel = class extends ArceePreTrainedModel {
};
var ArceeForCausalLM = class extends ArceePreTrainedModel {
};

// src/models/audio_spectrogram_transformer/modeling_audio_spectrogram_transformer.js
var ASTPreTrainedModel = class extends PreTrainedModel {
};
var ASTModel = class extends ASTPreTrainedModel {
};
var ASTForAudioClassification = class extends ASTPreTrainedModel {
};

// src/models/bart/modeling_bart.js
var BartPretrainedModel = class extends PreTrainedModel {
};
var BartModel = class extends BartPretrainedModel {
};
var BartForConditionalGeneration = class extends BartPretrainedModel {
};
var BartForSequenceClassification = class extends BartPretrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<SequenceClassifierOutput>} An object containing the model's output logits for sequence classification.
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};

// src/models/beit/modeling_beit.js
var BeitPreTrainedModel = class extends PreTrainedModel {
};
var BeitModel = class extends BeitPreTrainedModel {
};
var BeitForImageClassification = class extends BeitPreTrainedModel {
  /**
   * @param {any} model_inputs
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};

// src/models/bert/modeling_bert.js
var BertPreTrainedModel = class extends PreTrainedModel {
};
var BertModel = class extends BertPreTrainedModel {
};
var BertForMaskedLM = class extends BertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<MaskedLMOutput>} An object containing the model's output logits for masked language modeling.
   */
  async _call(model_inputs) {
    return new MaskedLMOutput(await super._call(model_inputs));
  }
};
var BertForSequenceClassification = class extends BertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<SequenceClassifierOutput>} An object containing the model's output logits for sequence classification.
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};
var BertForTokenClassification = class extends BertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<TokenClassifierOutput>} An object containing the model's output logits for token classification.
   */
  async _call(model_inputs) {
    return new TokenClassifierOutput(await super._call(model_inputs));
  }
};
var BertForQuestionAnswering = class extends BertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<QuestionAnsweringModelOutput>} An object containing the model's output logits for question answering.
   */
  async _call(model_inputs) {
    return new QuestionAnsweringModelOutput(await super._call(model_inputs));
  }
};

// src/models/blenderbot/modeling_blenderbot.js
var BlenderbotPreTrainedModel = class extends PreTrainedModel {
};
var BlenderbotModel = class extends BlenderbotPreTrainedModel {
};
var BlenderbotForConditionalGeneration = class extends BlenderbotPreTrainedModel {
};

// src/models/blenderbot_small/modeling_blenderbot_small.js
var BlenderbotSmallPreTrainedModel = class extends PreTrainedModel {
};
var BlenderbotSmallModel = class extends BlenderbotSmallPreTrainedModel {
};
var BlenderbotSmallForConditionalGeneration = class extends BlenderbotSmallPreTrainedModel {
};

// src/models/bloom/modeling_bloom.js
var BloomPreTrainedModel = class extends PreTrainedModel {
};
var BloomModel = class extends BloomPreTrainedModel {
};
var BloomForCausalLM = class extends BloomPreTrainedModel {
};

// src/models/camembert/modeling_camembert.js
var CamembertPreTrainedModel = class extends PreTrainedModel {
};
var CamembertModel = class extends CamembertPreTrainedModel {
};
var CamembertForMaskedLM = class extends CamembertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<MaskedLMOutput>} An object containing the model's output logits for masked language modeling.
   */
  async _call(model_inputs) {
    return new MaskedLMOutput(await super._call(model_inputs));
  }
};
var CamembertForSequenceClassification = class extends CamembertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<SequenceClassifierOutput>} An object containing the model's output logits for sequence classification.
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};
var CamembertForTokenClassification = class extends CamembertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<TokenClassifierOutput>} An object containing the model's output logits for token classification.
   */
  async _call(model_inputs) {
    return new TokenClassifierOutput(await super._call(model_inputs));
  }
};
var CamembertForQuestionAnswering = class extends CamembertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<QuestionAnsweringModelOutput>} An object containing the model's output logits for question answering.
   */
  async _call(model_inputs) {
    return new QuestionAnsweringModelOutput(await super._call(model_inputs));
  }
};

// src/models/chatterbox/modeling_chatterbox.js
var SILENCE_TOKEN = 4299n;
var START_SPEECH_TOKEN = 6561n;
var ChatterboxPreTrainedModel = class extends PreTrainedModel {
  forward_params = [
    "input_ids",
    "inputs_embeds",
    "attention_mask",
    "position_ids",
    "audio_values",
    "exaggeration",
    "audio_features",
    "audio_tokens",
    "speaker_embeddings",
    "speaker_features",
    "past_key_values"
  ];
  main_input_name = "input_ids";
  _return_dict_in_generate_keys = ["audio_tokens", "speaker_embeddings", "speaker_features"];
};
var ChatterboxModel = class extends ChatterboxPreTrainedModel {
  /**
   * @param {Tensor} audio_values
   * @returns {Promise<{audio_features: Tensor, audio_tokens: Tensor, speaker_embeddings: Tensor, speaker_features: Tensor}>}
   */
  async encode_speech(audio_values) {
    return sessionRun(this.sessions["speech_encoder"], {
      audio_values
    });
  }
  async forward({
    // Produced by the tokenizer/processor:
    input_ids = null,
    attention_mask = null,
    audio_values = null,
    exaggeration = null,
    // Used during generation:
    position_ids = null,
    inputs_embeds = null,
    past_key_values = null,
    // Generic generation parameters
    generation_config = null,
    logits_processor = null,
    // Speaker embeddings/features (useful for re-using pre-computed speaker data)
    audio_features = null,
    // float32[batch_size,sequence_length,1024]
    audio_tokens = null,
    // int64[batch_size,audio_sequence_length]
    speaker_embeddings = null,
    // float32[batch_size,192]
    speaker_features = null,
    // float32[batch_size,feature_dim,80]
    // TODO: needed?
    ...kwargs
  }) {
    let speech_encoder_outputs;
    if (!inputs_embeds) {
      const expected_inputs = this.sessions["embed_tokens"].inputNames;
      const embed_model_inputs = { input_ids };
      if (expected_inputs.includes("exaggeration")) {
        if (!(exaggeration instanceof Tensor2)) {
          const batch_size = input_ids.dims[0];
          if (exaggeration == null) {
            exaggeration = full([batch_size], 0.5);
          } else if (typeof exaggeration === "number") {
            exaggeration = full([batch_size], exaggeration);
          } else if (Array.isArray(exaggeration)) {
            exaggeration = new Tensor2("float32", exaggeration, [batch_size]);
          } else {
            throw new Error("Unsupported type for `exaggeration` input");
          }
        }
        embed_model_inputs.exaggeration = exaggeration;
      }
      if (expected_inputs.includes("position_ids")) {
        embed_model_inputs.position_ids = position_ids;
      }
      ({ inputs_embeds } = await sessionRun(this.sessions["embed_tokens"], embed_model_inputs));
      if (audio_features && audio_tokens && speaker_embeddings && speaker_features) {
        speech_encoder_outputs = { audio_features, audio_tokens, speaker_embeddings, speaker_features };
      }
      if (speech_encoder_outputs || audio_values) {
        speech_encoder_outputs ??= await this.encode_speech(audio_values);
        inputs_embeds = cat([speech_encoder_outputs.audio_features, inputs_embeds], 1);
        attention_mask = ones([inputs_embeds.dims[0], inputs_embeds.dims[1]]);
      } else {
        const target_length = inputs_embeds.dims[1];
        if (!past_key_values || target_length !== 1) {
          throw new Error("Incorrect state encountered during generation.");
        }
        const past_length = past_key_values.get_seq_length();
        attention_mask = ones([inputs_embeds.dims[0], past_length + target_length]);
      }
    }
    const outputs = await decoder_forward(
      this,
      {
        inputs_embeds,
        past_key_values,
        attention_mask,
        generation_config,
        logits_processor
      },
      false
    );
    return {
      ...outputs,
      ...speech_encoder_outputs
    };
  }
  prepare_inputs_for_generation(input_ids, model_inputs, generation_config) {
    if (!model_inputs.position_ids && this.sessions["embed_tokens"].inputNames.includes("position_ids")) {
      if (model_inputs.input_ids.dims[1] === 1) {
        const position_ids = Array.from(
          {
            length: input_ids.length
          },
          (_, i) => input_ids[i].length - input_ids[i].findLastIndex((x) => x == START_SPEECH_TOKEN) - 1
        );
        model_inputs.position_ids = new Tensor2("int64", position_ids, [input_ids.length, 1]);
      } else {
        const batched_input_ids = model_inputs.input_ids.tolist();
        const position_ids_list = batched_input_ids.map((ids) => {
          let position = 0;
          return ids.map((id) => id >= START_SPEECH_TOKEN ? 0 : position++);
        });
        model_inputs.position_ids = new Tensor2("int64", position_ids_list.flat(), model_inputs.input_ids.dims);
      }
    }
    if (model_inputs.input_ids.dims[1] === 1) {
      delete model_inputs.audio_values;
      delete model_inputs.audio_features;
      delete model_inputs.audio_tokens;
      delete model_inputs.speaker_embeddings;
      delete model_inputs.speaker_features;
    }
    return decoder_prepare_inputs_for_generation(this, input_ids, model_inputs, generation_config);
  }
  /** @type {PreTrainedModel['generate']} */
  async generate(params) {
    const { sequences, audio_tokens, speaker_embeddings, speaker_features } = (
      /** @type {any} */
      await super.generate({
        ...params,
        return_dict_in_generate: true
      })
    );
    const new_tokens = sequences.slice(null, [
      /** @type {Tensor} */
      params.input_ids.dims[1],
      // Exclude start of speech token
      -1
      // Exclude end of speech token
    ]);
    const silence_tokens = full([new_tokens.dims[0], 3], SILENCE_TOKEN);
    const speech_tokens = cat([audio_tokens, new_tokens, silence_tokens], 1);
    const { waveform } = await sessionRun(this.sessions["conditional_decoder"], {
      speech_tokens,
      speaker_features,
      speaker_embeddings
    });
    return waveform;
  }
};

// src/models/chinese_clip/modeling_chinese_clip.js
var ChineseCLIPPreTrainedModel = class extends PreTrainedModel {
};
var ChineseCLIPModel = class extends ChineseCLIPPreTrainedModel {
};

// src/models/chmv2/modeling_chmv2.js
var CHMv2PreTrainedModel = class extends PreTrainedModel {
};
var CHMv2ForDepthEstimation = class extends CHMv2PreTrainedModel {
};

// src/models/clap/modeling_clap.js
var ClapPreTrainedModel = class extends PreTrainedModel {
};
var ClapModel = class extends ClapPreTrainedModel {
};
var ClapTextModelWithProjection = class extends ClapPreTrainedModel {
  /** @type {typeof PreTrainedModel.from_pretrained} */
  static async from_pretrained(pretrained_model_name_or_path, options = {}) {
    return super.from_pretrained(pretrained_model_name_or_path, {
      ...options,
      // Update default model file name if not provided
      model_file_name: options.model_file_name ?? "text_model"
    });
  }
};
var ClapAudioModelWithProjection = class extends ClapPreTrainedModel {
  /** @type {typeof PreTrainedModel.from_pretrained} */
  static async from_pretrained(pretrained_model_name_or_path, options = {}) {
    return super.from_pretrained(pretrained_model_name_or_path, {
      ...options,
      // Update default model file name if not provided
      model_file_name: options.model_file_name ?? "audio_model"
    });
  }
};

// src/models/clip/modeling_clip.js
var CLIPPreTrainedModel = class extends PreTrainedModel {
};
var CLIPModel = class extends CLIPPreTrainedModel {
};
var CLIPTextModel = class extends CLIPPreTrainedModel {
  /** @type {typeof PreTrainedModel.from_pretrained} */
  static async from_pretrained(pretrained_model_name_or_path, options = {}) {
    return super.from_pretrained(pretrained_model_name_or_path, {
      ...options,
      // Update default model file name if not provided
      model_file_name: options.model_file_name ?? "text_model"
    });
  }
};
var CLIPTextModelWithProjection = class extends CLIPPreTrainedModel {
  /** @type {typeof PreTrainedModel.from_pretrained} */
  static async from_pretrained(pretrained_model_name_or_path, options = {}) {
    return super.from_pretrained(pretrained_model_name_or_path, {
      ...options,
      // Update default model file name if not provided
      model_file_name: options.model_file_name ?? "text_model"
    });
  }
};
var CLIPVisionModel = class extends CLIPPreTrainedModel {
  /** @type {typeof PreTrainedModel.from_pretrained} */
  static async from_pretrained(pretrained_model_name_or_path, options = {}) {
    return super.from_pretrained(pretrained_model_name_or_path, {
      ...options,
      // Update default model file name if not provided
      model_file_name: options.model_file_name ?? "vision_model"
    });
  }
};
var CLIPVisionModelWithProjection = class extends CLIPPreTrainedModel {
  /** @type {typeof PreTrainedModel.from_pretrained} */
  static async from_pretrained(pretrained_model_name_or_path, options = {}) {
    return super.from_pretrained(pretrained_model_name_or_path, {
      ...options,
      // Update default model file name if not provided
      model_file_name: options.model_file_name ?? "vision_model"
    });
  }
};

// src/models/clipseg/modeling_clipseg.js
var CLIPSegPreTrainedModel = class extends PreTrainedModel {
};
var CLIPSegModel = class extends CLIPSegPreTrainedModel {
};
var CLIPSegForImageSegmentation = class extends CLIPSegPreTrainedModel {
};

// src/models/codegen/modeling_codegen.js
var CodeGenPreTrainedModel = class extends PreTrainedModel {
};
var CodeGenModel = class extends CodeGenPreTrainedModel {
};
var CodeGenForCausalLM = class extends CodeGenPreTrainedModel {
};

// src/models/cohere/modeling_cohere.js
var CoherePreTrainedModel = class extends PreTrainedModel {
};
var CohereModel = class extends CoherePreTrainedModel {
};
var CohereForCausalLM = class extends CoherePreTrainedModel {
};

// src/models/cohere2/modeling_cohere2.js
var Cohere2PreTrainedModel = class extends PreTrainedModel {
};
var Cohere2Model = class extends Cohere2PreTrainedModel {
};
var Cohere2ForCausalLM = class extends Cohere2PreTrainedModel {
};

// src/models/cohere_asr/modeling_cohere_asr.js
var CohereAsrPreTrainedModel = class extends PreTrainedModel {
  requires_attention_mask = false;
  main_input_name = "input_features";
  forward_params = ["input_features", "decoder_input_ids", "decoder_attention_mask", "past_key_values"];
};
var CohereAsrModel = class extends CohereAsrPreTrainedModel {
};
var CohereAsrForConditionalGeneration = class extends CohereAsrPreTrainedModel {
};

// src/models/convbert/modeling_convbert.js
var ConvBertPreTrainedModel = class extends PreTrainedModel {
};
var ConvBertModel = class extends ConvBertPreTrainedModel {
};
var ConvBertForMaskedLM = class extends ConvBertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<MaskedLMOutput>} An object containing the model's output logits for masked language modeling.
   */
  async _call(model_inputs) {
    return new MaskedLMOutput(await super._call(model_inputs));
  }
};
var ConvBertForSequenceClassification = class extends ConvBertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<SequenceClassifierOutput>} An object containing the model's output logits for sequence classification.
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};
var ConvBertForTokenClassification = class extends ConvBertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<TokenClassifierOutput>} An object containing the model's output logits for token classification.
   */
  async _call(model_inputs) {
    return new TokenClassifierOutput(await super._call(model_inputs));
  }
};
var ConvBertForQuestionAnswering = class extends ConvBertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<QuestionAnsweringModelOutput>} An object containing the model's output logits for question answering.
   */
  async _call(model_inputs) {
    return new QuestionAnsweringModelOutput(await super._call(model_inputs));
  }
};

// src/models/convnext/modeling_convnext.js
var ConvNextPreTrainedModel = class extends PreTrainedModel {
};
var ConvNextModel = class extends ConvNextPreTrainedModel {
};
var ConvNextForImageClassification = class extends ConvNextPreTrainedModel {
  /**
   * @param {any} model_inputs
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};

// src/models/convnextv2/modeling_convnextv2.js
var ConvNextV2PreTrainedModel = class extends PreTrainedModel {
};
var ConvNextV2Model = class extends ConvNextV2PreTrainedModel {
};
var ConvNextV2ForImageClassification = class extends ConvNextV2PreTrainedModel {
  /**
   * @param {any} model_inputs
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};

// src/models/rt_detr/modeling_rt_detr.js
var RTDetrPreTrainedModel = class extends PreTrainedModel {
};
var RTDetrModel = class extends RTDetrPreTrainedModel {
};
var RTDetrForObjectDetection = class extends RTDetrPreTrainedModel {
  /**
   * @param {any} model_inputs
   */
  async _call(model_inputs) {
    return new RTDetrObjectDetectionOutput(await super._call(model_inputs));
  }
};
var RTDetrObjectDetectionOutput = class extends ModelOutput {
  /**
   * @param {Object} output The output of the model.
   * @param {Tensor} output.logits Classification logits (including no-object) for all queries.
   * @param {Tensor} output.pred_boxes Normalized boxes coordinates for all queries, represented as (center_x, center_y, width, height).
   * These values are normalized in [0, 1], relative to the size of each individual image in the batch (disregarding possible padding).
   */
  constructor({ logits, pred_boxes }) {
    super();
    this.logits = logits;
    this.pred_boxes = pred_boxes;
  }
};

// src/models/d_fine/modeling_d_fine.js
var DFinePreTrainedModel = class extends PreTrainedModel {
};
var DFineModel = class extends DFinePreTrainedModel {
};
var DFineForObjectDetection = class extends DFinePreTrainedModel {
  /**
   * @param {any} model_inputs
   */
  async _call(model_inputs) {
    return new RTDetrObjectDetectionOutput(await super._call(model_inputs));
  }
};

// src/models/dac/modeling_dac.js
var DacEncoderOutput = class extends ModelOutput {
  /**
   * @param {Object} output The output of the model.
   * @param {Tensor} output.audio_codes Discrete code embeddings, of shape `(batch_size, num_quantizers, codes_length)`.
   */
  constructor({ audio_codes }) {
    super();
    this.audio_codes = audio_codes;
  }
};
var DacDecoderOutput = class extends ModelOutput {
  /**
   * @param {Object} output The output of the model.
   * @param {Tensor} output.audio_values Decoded audio values, of shape `(batch_size, num_channels, sequence_length)`.
   */
  constructor({ audio_values }) {
    super();
    this.audio_values = audio_values;
  }
};
var DacPreTrainedModel = class extends PreTrainedModel {
  main_input_name = "input_values";
  forward_params = ["input_values"];
};
var DacModel = class extends DacPreTrainedModel {
  /**
   * Encodes the input audio waveform into discrete codes.
   * @param {Object} inputs Model inputs
   * @param {Tensor} [inputs.input_values] Float values of the input audio waveform, of shape `(batch_size, channels, sequence_length)`).
   * @returns {Promise<DacEncoderOutput>} The output tensor of shape `(batch_size, num_codebooks, sequence_length)`.
   */
  async encode(inputs) {
    return new DacEncoderOutput(await sessionRun(this.sessions["encoder_model"], inputs));
  }
  /**
   * Decodes the given frames into an output audio waveform.
   * @param {DacEncoderOutput} inputs The encoded audio codes.
   * @returns {Promise<DacDecoderOutput>} The output tensor of shape `(batch_size, num_channels, sequence_length)`.
   */
  async decode(inputs) {
    return new DacDecoderOutput(await sessionRun(this.sessions["decoder_model"], inputs));
  }
};
var DacEncoderModel = class extends DacPreTrainedModel {
  /** @type {typeof PreTrainedModel.from_pretrained} */
  static async from_pretrained(pretrained_model_name_or_path, options = {}) {
    return super.from_pretrained(pretrained_model_name_or_path, {
      ...options,
      // Update default model file name if not provided
      model_file_name: options.model_file_name ?? "encoder_model"
    });
  }
};
var DacDecoderModel = class extends DacPreTrainedModel {
  /** @type {typeof PreTrainedModel.from_pretrained} */
  static async from_pretrained(pretrained_model_name_or_path, options = {}) {
    return super.from_pretrained(pretrained_model_name_or_path, {
      ...options,
      // Update default model file name if not provided
      model_file_name: options.model_file_name ?? "decoder_model"
    });
  }
};

// src/models/deberta/modeling_deberta.js
var DebertaPreTrainedModel = class extends PreTrainedModel {
};
var DebertaModel = class extends DebertaPreTrainedModel {
};
var DebertaForMaskedLM = class extends DebertaPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<MaskedLMOutput>} An object containing the model's output logits for masked language modeling.
   */
  async _call(model_inputs) {
    return new MaskedLMOutput(await super._call(model_inputs));
  }
};
var DebertaForSequenceClassification = class extends DebertaPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<SequenceClassifierOutput>} An object containing the model's output logits for sequence classification.
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};
var DebertaForTokenClassification = class extends DebertaPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<TokenClassifierOutput>} An object containing the model's output logits for token classification.
   */
  async _call(model_inputs) {
    return new TokenClassifierOutput(await super._call(model_inputs));
  }
};
var DebertaForQuestionAnswering = class extends DebertaPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<QuestionAnsweringModelOutput>} An object containing the model's output logits for question answering.
   */
  async _call(model_inputs) {
    return new QuestionAnsweringModelOutput(await super._call(model_inputs));
  }
};

// src/models/deepseek_v3/modeling_deepseek_v3.js
var DeepseekV3PreTrainedModel = class extends PreTrainedModel {
};
var DeepseekV3Model = class extends DeepseekV3PreTrainedModel {
};
var DeepseekV3ForCausalLM = class extends DeepseekV3PreTrainedModel {
};

// src/models/deberta_v2/modeling_deberta_v2.js
var DebertaV2PreTrainedModel = class extends PreTrainedModel {
};
var DebertaV2Model = class extends DebertaV2PreTrainedModel {
};
var DebertaV2ForMaskedLM = class extends DebertaV2PreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<MaskedLMOutput>} An object containing the model's output logits for masked language modeling.
   */
  async _call(model_inputs) {
    return new MaskedLMOutput(await super._call(model_inputs));
  }
};
var DebertaV2ForSequenceClassification = class extends DebertaV2PreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<SequenceClassifierOutput>} An object containing the model's output logits for sequence classification.
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};
var DebertaV2ForTokenClassification = class extends DebertaV2PreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<TokenClassifierOutput>} An object containing the model's output logits for token classification.
   */
  async _call(model_inputs) {
    return new TokenClassifierOutput(await super._call(model_inputs));
  }
};
var DebertaV2ForQuestionAnswering = class extends DebertaV2PreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<QuestionAnsweringModelOutput>} An object containing the model's output logits for question answering.
   */
  async _call(model_inputs) {
    return new QuestionAnsweringModelOutput(await super._call(model_inputs));
  }
};

// src/models/decision_transformer/modeling_decision_transformer.js
var DecisionTransformerPreTrainedModel = class extends PreTrainedModel {
};
var DecisionTransformerModel = class extends DecisionTransformerPreTrainedModel {
};

// src/models/deit/modeling_deit.js
var DeiTPreTrainedModel = class extends PreTrainedModel {
};
var DeiTModel = class extends DeiTPreTrainedModel {
};
var DeiTForImageClassification = class extends DeiTPreTrainedModel {
  /**
   * @param {any} model_inputs
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};

// src/models/depth_anything/modeling_depth_anything.js
var DepthAnythingPreTrainedModel = class extends PreTrainedModel {
};
var DepthAnythingForDepthEstimation = class extends DepthAnythingPreTrainedModel {
};

// src/models/depth_pro/modeling_depth_pro.js
var DepthProPreTrainedModel = class extends PreTrainedModel {
};
var DepthProForDepthEstimation = class extends DepthProPreTrainedModel {
};

// src/models/detr/modeling_detr.js
var DetrPreTrainedModel = class extends PreTrainedModel {
};
var DetrModel = class extends DetrPreTrainedModel {
};
var DetrForObjectDetection = class extends DetrPreTrainedModel {
  /**
   * @param {any} model_inputs
   */
  async _call(model_inputs) {
    return new DetrObjectDetectionOutput(await super._call(model_inputs));
  }
};
var DetrForSegmentation = class extends DetrPreTrainedModel {
  /**
   * Runs the model with the provided inputs
   * @param {Object} model_inputs Model inputs
   * @returns {Promise<DetrSegmentationOutput>} Object containing segmentation outputs
   */
  async _call(model_inputs) {
    return new DetrSegmentationOutput(await super._call(model_inputs));
  }
};
var DetrObjectDetectionOutput = class extends ModelOutput {
  /**
   * @param {Object} output The output of the model.
   * @param {Tensor} output.logits Classification logits (including no-object) for all queries.
   * @param {Tensor} output.pred_boxes Normalized boxes coordinates for all queries, represented as (center_x, center_y, width, height).
   * These values are normalized in [0, 1], relative to the size of each individual image in the batch (disregarding possible padding).
   */
  constructor({ logits, pred_boxes }) {
    super();
    this.logits = logits;
    this.pred_boxes = pred_boxes;
  }
};
var DetrSegmentationOutput = class extends ModelOutput {
  /**
   * @param {Object} output The output of the model.
   * @param {Tensor} output.logits The output logits of the model.
   * @param {Tensor} output.pred_boxes Predicted boxes.
   * @param {Tensor} output.pred_masks Predicted masks.
   */
  constructor({ logits, pred_boxes, pred_masks }) {
    super();
    this.logits = logits;
    this.pred_boxes = pred_boxes;
    this.pred_masks = pred_masks;
  }
};

// src/models/dinov2/modeling_dinov2.js
var Dinov2PreTrainedModel = class extends PreTrainedModel {
};
var Dinov2Model = class extends Dinov2PreTrainedModel {
};
var Dinov2ForImageClassification = class extends Dinov2PreTrainedModel {
  /**
   * @param {any} model_inputs
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};

// src/models/dinov2_with_registers/modeling_dinov2_with_registers.js
var Dinov2WithRegistersPreTrainedModel = class extends PreTrainedModel {
};
var Dinov2WithRegistersModel = class extends Dinov2WithRegistersPreTrainedModel {
};
var Dinov2WithRegistersForImageClassification = class extends Dinov2WithRegistersPreTrainedModel {
  /**
   * @param {any} model_inputs
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};

// src/models/dinov3_convnext/modeling_dinov3_convnext.js
var DINOv3ConvNextPreTrainedModel = class extends PreTrainedModel {
};
var DINOv3ConvNextModel = class extends DINOv3ConvNextPreTrainedModel {
};

// src/models/dinov3_vit/modeling_dinov3_vit.js
var DINOv3ViTPreTrainedModel = class extends PreTrainedModel {
};
var DINOv3ViTModel = class extends DINOv3ViTPreTrainedModel {
};

// src/models/distilbert/modeling_distilbert.js
var DistilBertPreTrainedModel = class extends PreTrainedModel {
};
var DistilBertModel = class extends DistilBertPreTrainedModel {
};
var DistilBertForSequenceClassification = class extends DistilBertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<SequenceClassifierOutput>} An object containing the model's output logits for sequence classification.
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};
var DistilBertForTokenClassification = class extends DistilBertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<TokenClassifierOutput>} An object containing the model's output logits for token classification.
   */
  async _call(model_inputs) {
    return new TokenClassifierOutput(await super._call(model_inputs));
  }
};
var DistilBertForQuestionAnswering = class extends DistilBertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<QuestionAnsweringModelOutput>} An object containing the model's output logits for question answering.
   */
  async _call(model_inputs) {
    return new QuestionAnsweringModelOutput(await super._call(model_inputs));
  }
};
var DistilBertForMaskedLM = class extends DistilBertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<MaskedLMOutput>} returned object
   */
  async _call(model_inputs) {
    return new MaskedLMOutput(await super._call(model_inputs));
  }
};

// src/models/donut_swin/modeling_donut_swin.js
var DonutSwinPreTrainedModel = class extends PreTrainedModel {
};
var DonutSwinModel = class extends DonutSwinPreTrainedModel {
};

// src/models/dpt/modeling_dpt.js
var DPTPreTrainedModel = class extends PreTrainedModel {
};
var DPTModel = class extends DPTPreTrainedModel {
};
var DPTForDepthEstimation = class extends DPTPreTrainedModel {
};

// src/models/efficientnet/modeling_efficientnet.js
var EfficientNetPreTrainedModel = class extends PreTrainedModel {
};
var EfficientNetModel = class extends EfficientNetPreTrainedModel {
};
var EfficientNetForImageClassification = class extends EfficientNetPreTrainedModel {
  /**
   * @param {any} model_inputs
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};

// src/models/electra/modeling_electra.js
var ElectraPreTrainedModel = class extends PreTrainedModel {
};
var ElectraModel = class extends ElectraPreTrainedModel {
};
var ElectraForMaskedLM = class extends ElectraPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<MaskedLMOutput>} An object containing the model's output logits for masked language modeling.
   */
  async _call(model_inputs) {
    return new MaskedLMOutput(await super._call(model_inputs));
  }
};
var ElectraForSequenceClassification = class extends ElectraPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<SequenceClassifierOutput>} An object containing the model's output logits for sequence classification.
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};
var ElectraForTokenClassification = class extends ElectraPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<TokenClassifierOutput>} An object containing the model's output logits for token classification.
   */
  async _call(model_inputs) {
    return new TokenClassifierOutput(await super._call(model_inputs));
  }
};
var ElectraForQuestionAnswering = class extends ElectraPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<QuestionAnsweringModelOutput>} An object containing the model's output logits for question answering.
   */
  async _call(model_inputs) {
    return new QuestionAnsweringModelOutput(await super._call(model_inputs));
  }
};

// src/models/ernie4_5/modeling_ernie4_5.js
var Ernie4_5PretrainedModel = class extends PreTrainedModel {
};
var Ernie4_5Model = class extends Ernie4_5PretrainedModel {
};
var Ernie4_5ForCausalLM = class extends Ernie4_5PretrainedModel {
};

// src/models/esm/modeling_esm.js
var EsmPreTrainedModel = class extends PreTrainedModel {
};
var EsmModel = class extends EsmPreTrainedModel {
};
var EsmForMaskedLM = class extends EsmPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<MaskedLMOutput>} An object containing the model's output logits for masked language modeling.
   */
  async _call(model_inputs) {
    return new MaskedLMOutput(await super._call(model_inputs));
  }
};
var EsmForSequenceClassification = class extends EsmPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<SequenceClassifierOutput>} An object containing the model's output logits for sequence classification.
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};
var EsmForTokenClassification = class extends EsmPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<TokenClassifierOutput>} An object containing the model's output logits for token classification.
   */
  async _call(model_inputs) {
    return new TokenClassifierOutput(await super._call(model_inputs));
  }
};

// src/models/eurobert/modeling_eurobert.js
var EuroBertPreTrainedModel = class extends PreTrainedModel {
};
var EuroBertModel = class extends EuroBertPreTrainedModel {
};
var EuroBertForMaskedLM = class extends EuroBertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<MaskedLMOutput>} An object containing the model's output logits for masked language modeling.
   */
  async _call(model_inputs) {
    return new MaskedLMOutput(await super._call(model_inputs));
  }
};
var EuroBertForSequenceClassification = class extends EuroBertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<SequenceClassifierOutput>} An object containing the model's output logits for sequence classification.
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};
var EuroBertForTokenClassification = class extends EuroBertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<TokenClassifierOutput>} An object containing the model's output logits for token classification.
   */
  async _call(model_inputs) {
    return new TokenClassifierOutput(await super._call(model_inputs));
  }
};

// src/models/exaone/modeling_exaone.js
var ExaonePreTrainedModel = class extends PreTrainedModel {
};
var ExaoneModel = class extends ExaonePreTrainedModel {
};
var ExaoneForCausalLM = class extends ExaonePreTrainedModel {
};

// src/models/falcon/modeling_falcon.js
var FalconPreTrainedModel = class extends PreTrainedModel {
};
var FalconModel = class extends FalconPreTrainedModel {
};
var FalconForCausalLM = class extends FalconPreTrainedModel {
};

// src/models/falcon_h1/modeling_falcon_h1.js
var FalconH1PreTrainedModel = class extends PreTrainedModel {
};
var FalconH1Model = class extends FalconH1PreTrainedModel {
};
var FalconH1ForCausalLM = class extends FalconH1PreTrainedModel {
};

// src/models/fastvit/modeling_fastvit.js
var FastViTPreTrainedModel = class extends PreTrainedModel {
};
var FastViTModel = class extends FastViTPreTrainedModel {
};
var FastViTForImageClassification = class extends FastViTPreTrainedModel {
  /**
   * @param {any} model_inputs
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};

// src/models/florence2/modeling_florence2.js
var Florence2PreTrainedModel = class extends PreTrainedModel {
  forward_params = [
    // Encoder inputs
    "input_ids",
    "inputs_embeds",
    "attention_mask",
    "pixel_values",
    // Decoder inputs
    "encoder_outputs",
    "decoder_input_ids",
    "decoder_inputs_embeds",
    "decoder_attention_mask",
    "past_key_values"
  ];
  main_input_name = "inputs_embeds";
};
var Florence2ForConditionalGeneration = class extends Florence2PreTrainedModel {
  _merge_input_ids_with_image_features({ inputs_embeds, image_features, input_ids, attention_mask }) {
    return {
      inputs_embeds: cat(
        [
          image_features,
          // image embeds
          inputs_embeds
          // task prefix embeds
        ],
        1
      ),
      attention_mask: cat(
        [
          ones(image_features.dims.slice(0, 2)),
          // image attention mask
          attention_mask
          // task prefix attention mask
        ],
        1
      )
    };
  }
  async _prepare_inputs_embeds({ input_ids, pixel_values, inputs_embeds, attention_mask }) {
    if (!input_ids && !pixel_values) {
      throw new Error("Either `input_ids` or `pixel_values` should be provided.");
    }
    let text_features, image_features;
    if (input_ids) {
      text_features = await this.encode_text({ input_ids });
    }
    if (pixel_values) {
      image_features = await this.encode_image({ pixel_values });
    }
    if (text_features && image_features) {
      ({ inputs_embeds, attention_mask } = this._merge_input_ids_with_image_features({
        inputs_embeds: text_features,
        image_features,
        input_ids,
        attention_mask
      }));
    } else {
      inputs_embeds = text_features || image_features;
    }
    return { inputs_embeds, attention_mask };
  }
  async forward({
    input_ids,
    pixel_values,
    attention_mask,
    decoder_input_ids,
    decoder_attention_mask,
    encoder_outputs,
    past_key_values,
    inputs_embeds,
    decoder_inputs_embeds
  }) {
    if (!inputs_embeds) {
      ({ inputs_embeds, attention_mask } = await this._prepare_inputs_embeds({
        input_ids,
        pixel_values,
        inputs_embeds,
        attention_mask
      }));
    }
    if (!encoder_outputs) {
      let { last_hidden_state } = await encoder_forward(this, { inputs_embeds, attention_mask });
      encoder_outputs = last_hidden_state;
    }
    if (!decoder_inputs_embeds) {
      if (!decoder_input_ids) {
        throw new Error("Either `decoder_input_ids` or `decoder_inputs_embeds` should be provided.");
      }
      decoder_inputs_embeds = await this.encode_text({ input_ids: decoder_input_ids });
    }
    const decoderFeeds = {
      inputs_embeds: decoder_inputs_embeds,
      attention_mask: decoder_attention_mask,
      encoder_attention_mask: attention_mask,
      encoder_hidden_states: encoder_outputs,
      past_key_values
    };
    return await decoder_forward(this, decoderFeeds, true);
  }
};

// src/models/gemma/modeling_gemma.js
var GemmaPreTrainedModel = class extends PreTrainedModel {
};
var GemmaModel = class extends GemmaPreTrainedModel {
};
var GemmaForCausalLM = class extends GemmaPreTrainedModel {
};

// src/models/gemma2/modeling_gemma2.js
var Gemma2PreTrainedModel = class extends PreTrainedModel {
};
var Gemma2Model = class extends Gemma2PreTrainedModel {
};
var Gemma2ForCausalLM = class extends Gemma2PreTrainedModel {
};

// src/models/llava/modeling_llava.js
var LlavaPreTrainedModel = class extends PreTrainedModel {
  forward_params = ["input_ids", "attention_mask", "pixel_values", "position_ids", "past_key_values"];
};
var LlavaForConditionalGeneration = class extends LlavaPreTrainedModel {
  _merge_input_ids_with_image_features(kwargs) {
    const vision_hidden_size = kwargs.image_features.dims.at(-1);
    const reshaped_image_hidden_states = kwargs.image_features.view(-1, vision_hidden_size);
    return default_merge_input_ids_with_image_features({
      // @ts-ignore
      image_token_id: this.config.image_token_index ?? this.config.image_token_id,
      ...kwargs,
      image_features: reshaped_image_hidden_states
    });
  }
};
var Moondream1ForConditionalGeneration = class extends LlavaForConditionalGeneration {
};
var LlavaQwen2ForCausalLM = class extends LlavaForConditionalGeneration {
};

// src/models/gemma3/modeling_gemma3.js
var Gemma3PreTrainedModel = class extends PreTrainedModel {
};
var Gemma3Model = class extends Gemma3PreTrainedModel {
};
var Gemma3ForConditionalGeneration = class extends LlavaForConditionalGeneration {
};
var Gemma3ForCausalLM = class extends Gemma3ForConditionalGeneration {
};

// src/models/gemma3n/modeling_gemma3n.js
var Gemma3nPreTrainedModel = class extends PreTrainedModel {
  forward_params = [
    "input_ids",
    "attention_mask",
    "inputs_embeds",
    "per_layer_inputs",
    "position_ids",
    "pixel_values",
    "input_features",
    "input_features_mask",
    "past_key_values"
  ];
};
var Gemma3nForConditionalGeneration = class extends Gemma3nPreTrainedModel {
  async forward({
    // Produced by the tokenizer/processor:
    input_ids = null,
    attention_mask = null,
    pixel_values = null,
    input_features = null,
    input_features_mask = null,
    // Used during generation:
    position_ids = null,
    inputs_embeds = null,
    per_layer_inputs = null,
    past_key_values = null,
    // Generic generation parameters
    generation_config = null,
    logits_processor = null,
    // TODO: needed?
    ...kwargs
  }) {
    if (!inputs_embeds || !per_layer_inputs) {
      ({ inputs_embeds, per_layer_inputs } = await sessionRun(this.sessions["embed_tokens"], {
        input_ids
      }));
      if (input_ids.dims[1] !== 1) {
        if (pixel_values) {
          const { image_features } = await this._encode_vision({ pixel_values, ...kwargs });
          ({ inputs_embeds, attention_mask } = this._merge_input_ids_with_image_features({
            image_features,
            inputs_embeds,
            input_ids,
            attention_mask
          }));
        }
        if (input_features) {
          const { audio_features } = await sessionRun(this.sessions["audio_encoder"], {
            input_features,
            input_features_mask
          });
          ({ inputs_embeds, attention_mask } = this._merge_input_ids_with_audio_features({
            audio_features,
            inputs_embeds,
            input_ids,
            attention_mask
          }));
        }
      }
    }
    const outputs = await decoder_forward(
      this,
      {
        inputs_embeds,
        per_layer_inputs,
        past_key_values,
        attention_mask,
        position_ids,
        generation_config,
        logits_processor
      },
      true
    );
    return outputs;
  }
  _encode_vision(kwargs) {
    return sessionRun(this.sessions["vision_encoder"], { pixel_values: kwargs.pixel_values });
  }
  _merge_input_ids_with_image_features(kwargs) {
    const vision_hidden_size = kwargs.image_features.dims.at(-1);
    const reshaped_image_hidden_states = kwargs.image_features.view(-1, vision_hidden_size);
    return default_merge_input_ids_with_image_features({
      // @ts-ignore
      image_token_id: this.config.image_token_id,
      ...kwargs,
      image_features: reshaped_image_hidden_states
    });
  }
  _merge_input_ids_with_audio_features(kwargs) {
    const audio_hidden_size = kwargs.audio_features.dims.at(-1);
    const reshaped_audio_features = kwargs.audio_features.view(-1, audio_hidden_size);
    return default_merge_input_ids_with_audio_features({
      // @ts-ignore
      audio_token_id: this.config.audio_token_id,
      ...kwargs,
      audio_features: reshaped_audio_features
    });
  }
};
var Gemma3nForCausalLM = class extends Gemma3nForConditionalGeneration {
};

// src/models/gemma4/modeling_gemma4.js
var Gemma4ForConditionalGeneration = class extends Gemma3nForConditionalGeneration {
  forward_params = [
    "input_ids",
    "attention_mask",
    "inputs_embeds",
    "per_layer_inputs",
    "position_ids",
    "pixel_values",
    "image_position_ids",
    "input_features",
    "input_features_mask",
    "past_key_values"
  ];
  _encode_vision(kwargs) {
    return sessionRun(this.sessions["vision_encoder"], {
      pixel_values: kwargs.pixel_values,
      pixel_position_ids: kwargs.image_position_ids
    });
  }
};
var Gemma4ForCausalLM = class extends Gemma4ForConditionalGeneration {
};

// src/models/glm/modeling_glm.js
var GlmPreTrainedModel = class extends PreTrainedModel {
};
var GlmModel = class extends GlmPreTrainedModel {
};
var GlmForCausalLM = class extends GlmPreTrainedModel {
};

// src/models/glm_moe_dsa/modeling_glm_moe_dsa.js
var GlmMoeDsaPreTrainedModel = class extends PreTrainedModel {
};
var GlmMoeDsaModel = class extends GlmMoeDsaPreTrainedModel {
};
var GlmMoeDsaForCausalLM = class extends GlmMoeDsaPreTrainedModel {
};

// src/models/qwen2_vl/modeling_qwen2_vl.js
var Qwen2VLPreTrainedModel = class extends PreTrainedModel {
  forward_params = [
    // Text inputs
    "input_ids",
    "attention_mask",
    "position_ids",
    "past_key_values",
    // Vision inputs
    "pixel_values",
    "image_grid_thw"
  ];
};
var Qwen2VLForConditionalGeneration = class extends Qwen2VLPreTrainedModel {
  // NOTE: This is used as the base class for all Qwen VL models and their CausalLM variants.
  // CausalLM variants (e.g., Qwen2VLForCausalLM) extend this class but load only
  // embed_tokens + decoder_model_merged (no vision_encoder) via MultimodalLanguageModelOnly type.
  image_grid_thw_name = "grid_thw";
  /**
   * Compute text-only 3D rope position IDs (all 3 dims get the same 1D positions).
   * @param {Tensor} input_ids
   * @param {Tensor} attention_mask
   * @returns {[Tensor, Tensor]} [position_ids, mrope_position_deltas]
   */
  _get_text_only_rope_index(input_ids, attention_mask) {
    if (attention_mask) {
      const { data, dims } = cumsum_masked_fill(attention_mask);
      const position_ids = BigInt64Array.from({ length: 3 * data.length }, (_, i) => data[i % data.length]);
      const mrope_position_deltas = Array.from(
        { length: dims[0] },
        (_, i) => max(data.subarray(dims[1] * i, dims[1] * (i + 1)))[0] + 1n + BigInt(dims[1])
      );
      return [
        new Tensor2("int64", position_ids, [3, ...dims]),
        new Tensor2("int64", mrope_position_deltas, [mrope_position_deltas.length, 1])
      ];
    } else {
      const [batch_size, seq_length] = input_ids.dims;
      const position_ids = BigInt64Array.from(
        { length: 3 * batch_size * seq_length },
        (_, i) => BigInt(Math.floor(i % seq_length / batch_size))
      );
      return [new Tensor2("int64", position_ids, [3, ...input_ids.dims]), zeros([batch_size, 1])];
    }
  }
  /**
   * Reorder per-segment position ID lists from [seg1[t,h,w], seg2[t,h,w], ...] into
   * global [all_t, all_h, all_w] order, then write back into the position_ids array
   * respecting attention mask.
   * @param {number[][]} llm_pos_ids_list List of per-segment position arrays, each of length 3*seg_len
   * @param {number[]} attn_mask Attention mask for this batch element
   * @param {number[][][]} position_ids_list [3][batch][seq] output array to write into
   * @param {number} batch_idx Current batch index
   * @returns {number[]} Flat reordered positions of length total_len
   */
  _reorder_and_write_positions(llm_pos_ids_list, attn_mask, position_ids_list, batch_idx) {
    const total_len = llm_pos_ids_list.reduce((acc, x) => acc + x.length, 0);
    const llm_positions = new Array(total_len);
    let index = 0;
    for (let x = 0; x < 3; ++x) {
      for (const val of llm_pos_ids_list) {
        const seg_len = val.length / 3;
        for (let z2 = x * seg_len; z2 < (x + 1) * seg_len; ++z2) {
          llm_positions[index++] = val[z2];
        }
      }
    }
    let count2 = 0;
    for (let y = 0; y < attn_mask.length; ++y) {
      if (attn_mask[y] == 1) {
        for (let x = 0; x < 3; ++x) {
          position_ids_list[x][batch_idx][y] = llm_positions[x * total_len / 3 + count2];
        }
        ++count2;
      }
    }
    return llm_positions;
  }
  /**
   * Build per-batch position ID segments for multimodal rope.
   * Override this in subclasses to change how vision/text segments are identified and positioned.
   * @param {object} params
   * @param {any[]} params.filtered_ids - attention-masked token IDs for this batch element
   * @param {any[][]} params.image_grid_thw_list - all image grid dimensions
   * @param {any[][]} params.video_grid_thw_list - all video grid dimensions
   * @param {number} params.spatial_merge_size
   * @param {{image_index: number, video_index: number}} params.state - mutable counters shared across batches
   * @returns {number[][]} llm_pos_ids_list - segments of [t..., h..., w...] positions
   */
  _get_multimodal_rope_positions({
    filtered_ids,
    image_grid_thw_list,
    video_grid_thw_list,
    spatial_merge_size,
    state
  }) {
    const { image_token_id, video_token_id, vision_start_token_id } = this.config;
    const ids = filtered_ids;
    const vision_start_indices = ids.reduce((acc, x, idx) => {
      if (x == vision_start_token_id) acc.push(idx);
      return acc;
    }, []);
    const vision_tokens = vision_start_indices.map((x) => ids[x + 1]);
    const image_nums = vision_tokens.filter((x) => x == image_token_id).length;
    const video_nums = vision_tokens.filter((x) => x == video_token_id).length;
    const llm_pos_ids_list = [];
    let st2 = 0;
    let remain_images = image_nums;
    let remain_videos = video_nums;
    for (let j = 0; j < vision_tokens.length; ++j) {
      const next_image_token = ids.findIndex((x, i) => i > st2 && x == image_token_id);
      const next_video_token = ids.findIndex((x, i) => i > st2 && x == video_token_id);
      const ed_image = remain_images > 0 && next_image_token !== -1 ? next_image_token : ids.length + 1;
      const ed_video = remain_videos > 0 && next_video_token !== -1 ? next_video_token : ids.length + 1;
      let ed;
      let t, h, w;
      if (ed_image < ed_video) {
        [t, h, w] = image_grid_thw_list[state.image_index];
        ++state.image_index;
        --remain_images;
        ed = ed_image;
      } else {
        [t, h, w] = video_grid_thw_list[state.video_index];
        ++state.video_index;
        --remain_videos;
        ed = ed_video;
      }
      const [llm_grid_t, llm_grid_h, llm_grid_w] = [
        Number(t),
        Math.floor(Number(h) / spatial_merge_size),
        Math.floor(Number(w) / spatial_merge_size)
      ];
      const text_len = ed - st2;
      const st_idx = llm_pos_ids_list.length > 0 ? max(llm_pos_ids_list.at(-1))[0] + 1 : 0;
      llm_pos_ids_list.push(Array.from({ length: 3 * text_len }, (_, i) => st_idx + i % text_len));
      const offset = text_len + st_idx;
      const grid_size = llm_grid_t * llm_grid_h * llm_grid_w;
      const t_index = Array.from(
        { length: grid_size },
        (_, i) => offset + Math.floor(i / (llm_grid_h * llm_grid_w))
      );
      const h_index = Array.from(
        { length: grid_size },
        (_, i) => offset + Math.floor(i / llm_grid_w) % llm_grid_h
      );
      const w_index = Array.from({ length: grid_size }, (_, i) => offset + i % llm_grid_w);
      llm_pos_ids_list.push([t_index, h_index, w_index].flat());
      st2 = ed + grid_size;
    }
    if (st2 < ids.length) {
      const st_idx = llm_pos_ids_list.length > 0 ? max(llm_pos_ids_list.at(-1))[0] + 1 : 0;
      const text_len = ids.length - st2;
      llm_pos_ids_list.push(Array.from({ length: 3 * text_len }, (_, i) => st_idx + i % text_len));
    }
    return llm_pos_ids_list;
  }
  /**
   * Calculate the 3D rope index based on image and video's temporal, height and width in LLM.
   *
   * Explanation:
   *     Each embedding sequence contains vision embedding and text embedding or just contains text embedding.
   *
   *     For pure text embedding sequence, the rotary position embedding has no difference with mordern LLMs.
   *     Examples:
   *         input_ids: [T T T T T], here T is for text.
   *         temporal position_ids: [0, 1, 2, 3, 4]
   *         height position_ids: [0, 1, 2, 3, 4]
   *         width position_ids: [0, 1, 2, 3, 4]
   *
   *     For vision and text embedding sequence, we calculate 3D rotary position embedding for vision part
   *     and 1D rotary position embeddin for text part.
   *     Examples:
   *         Assume we have a video input with 3 temporal patches, 2 height patches and 2 width patches.
   *         input_ids: [V V V V V V V V V V V V T T T T T], here V is for vision.
   *         vision temporal position_ids: [0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2]
   *         vision height position_ids: [0, 0, 1, 1, 0, 0, 1, 1, 0, 0, 1, 1]
   *         vision width position_ids: [0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1]
   *         text temporal position_ids: [3, 4, 5, 6, 7]
   *         text height position_ids: [3, 4, 5, 6, 7]
   *         text width position_ids: [3, 4, 5, 6, 7]
   *         Here we calculate the text start position_ids as the max vision position_ids plus 1.
   *
   * @param {Tensor} input_ids Indices of input sequence tokens in the vocabulary. Tensor of shape `(batch_size, sequence_length)`.
   * @param {Tensor} image_grid_thw (Optional) The temporal, height and width of feature shape of each image in LLM. Tensor of shape `(num_images, 3)`.
   * @param {Tensor} video_grid_thw (Optional) The temporal, height and width of feature shape of each video in LLM. Tensor of shape `(num_videos, 3)`.
   * @param {Tensor} attention_mask (Optional) Mask to avoid performing attention on padding token indices. Tensor of shape `(batch_size, sequence_length)`.
   * @returns {[Tensor, Tensor]} [position_ids, mrope_position_deltas]
   */
  get_rope_index(input_ids, image_grid_thw, video_grid_thw, attention_mask) {
    const { vision_config } = this.config;
    const spatial_merge_size = vision_config.spatial_merge_size ?? 2;
    if (image_grid_thw || video_grid_thw) {
      const total_input_ids = input_ids.tolist();
      if (!attention_mask) {
        attention_mask = ones_like(input_ids);
      }
      const attention_mask_list = attention_mask.tolist();
      const position_ids_list = Array.from(
        { length: 3 },
        () => Array.from({ length: input_ids.dims[0] }, () => Array.from({ length: input_ids.dims[1] }, () => 0))
      );
      const image_grid_thw_list = image_grid_thw ? image_grid_thw.tolist() : [];
      const video_grid_thw_list = video_grid_thw ? video_grid_thw.tolist() : [];
      const state = { image_index: 0, video_index: 0 };
      const mrope_position_deltas = [];
      for (let i = 0; i < total_input_ids.length; ++i) {
        const filtered_ids = total_input_ids[i].filter((_, j) => attention_mask_list[i][j] == 1);
        const llm_pos_ids_list = this._get_multimodal_rope_positions({
          filtered_ids,
          image_grid_thw_list,
          video_grid_thw_list,
          spatial_merge_size,
          state
        });
        const llm_positions = this._reorder_and_write_positions(
          llm_pos_ids_list,
          attention_mask_list[i],
          position_ids_list,
          i
        );
        mrope_position_deltas.push(max(llm_positions)[0] + 1 - total_input_ids[i].length);
      }
      return [
        new Tensor2("int64", position_ids_list.flat(Infinity), [3, input_ids.dims[0], input_ids.dims[1]]),
        new Tensor2("int64", mrope_position_deltas, [mrope_position_deltas.length, 1])
      ];
    } else {
      return this._get_text_only_rope_index(input_ids, attention_mask);
    }
  }
  async encode_image({ pixel_values, image_grid_thw }) {
    const features = (await sessionRun(this.sessions["vision_encoder"], {
      pixel_values,
      [this.image_grid_thw_name]: image_grid_thw
    })).image_features;
    return features;
  }
  _merge_input_ids_with_image_features(kwargs) {
    return default_merge_input_ids_with_image_features({
      // @ts-ignore
      image_token_id: this.config.image_token_id,
      ...kwargs
    });
  }
  prepare_inputs_for_generation(input_ids, model_inputs, generation_config) {
    if (!model_inputs.attention_mask || model_inputs.position_ids) {
      return model_inputs;
    }
    const session = this.sessions["decoder_model_merged"] ?? this.sessions["model"];
    if (!session.inputNames.includes("position_ids")) {
      return model_inputs;
    }
    if (!model_inputs.past_key_values) {
      [model_inputs.position_ids, model_inputs.rope_deltas] = this.get_rope_index(
        model_inputs.input_ids,
        model_inputs.image_grid_thw,
        model_inputs.video_grid_thw,
        model_inputs.attention_mask
      );
    } else {
      model_inputs.pixel_values = null;
      const past_length = model_inputs.past_key_values.get_seq_length();
      if (past_length < model_inputs.input_ids.dims[1]) {
        const [full_position_ids, rope_deltas] = this.get_rope_index(
          model_inputs.input_ids,
          model_inputs.image_grid_thw,
          model_inputs.video_grid_thw,
          model_inputs.attention_mask
        );
        model_inputs.rope_deltas = rope_deltas;
        model_inputs.position_ids = full_position_ids.slice(null, null, [past_length, null]);
        model_inputs.input_ids = model_inputs.input_ids.slice(null, [past_length, null]);
      } else {
        if (!model_inputs.rope_deltas) {
          [, model_inputs.rope_deltas] = this.get_rope_index(
            model_inputs.input_ids,
            model_inputs.image_grid_thw,
            model_inputs.video_grid_thw,
            model_inputs.attention_mask
          );
        }
        const delta = BigInt(past_length);
        const rope_deltas_list = model_inputs.rope_deltas.map((x) => delta + x);
        model_inputs.position_ids = stack([rope_deltas_list, rope_deltas_list, rope_deltas_list], 0);
      }
    }
    return model_inputs;
  }
};
var Qwen2VLForCausalLM = class extends Qwen2VLForConditionalGeneration {
};

// src/models/qwen2_5_vl/modeling_qwen2_5_vl.js
var Qwen2_5_VLForConditionalGeneration = class extends Qwen2VLForConditionalGeneration {
  image_grid_thw_name = "image_grid_thw";
};
var Qwen2_5_VLForCausalLM = class extends Qwen2VLForCausalLM {
  image_grid_thw_name = "image_grid_thw";
};

// src/models/glm_ocr/modeling_glm_ocr.js
var GlmOcrForConditionalGeneration = class extends Qwen2_5_VLForConditionalGeneration {
  /**
   * Compute 3D positional indices for vision tokens.
   * Temporal is constant, height is repeat-interleaved, width tiles.
   * @param {number} start_position
   * @param {number[]} grid_thw [T, H, W]
   * @param {number} temp_merge_size
   * @param {number} spatial_merge_size
   * @returns {number[]} Flat array of length 3 * seq_len: [temporal..., height..., width...]
   */
  get_vision_position_ids(start_position, grid_thw, temp_merge_size, spatial_merge_size) {
    const llm_grid_t = Math.floor(grid_thw[0] / temp_merge_size);
    const llm_grid_h = Math.floor(grid_thw[1] / spatial_merge_size);
    const llm_grid_w = Math.floor(grid_thw[2] / spatial_merge_size);
    const seq_len = llm_grid_h * llm_grid_w * llm_grid_t;
    const t_pos = Array.from({ length: seq_len }, () => start_position);
    const h_pos = Array.from(
      { length: seq_len },
      (_, i) => start_position + Math.floor(i / (llm_grid_w * llm_grid_t))
    );
    const w_pos = Array.from({ length: seq_len }, (_, i) => start_position + i % llm_grid_w);
    return [...t_pos, ...h_pos, ...w_pos];
  }
  /**
   * GlmOcr uses mm_token_type_ids-style grouping (image tokens identified by image_token_id)
   * instead of vision_start_token_id scanning used by Qwen2VL.
   * After a vision segment, position advances by max(h, w) / spatial_merge_size.
   */
  _get_multimodal_rope_positions({
    filtered_ids,
    image_grid_thw_list,
    video_grid_thw_list,
    spatial_merge_size,
    state
  }) {
    const { image_token_id } = this.config;
    const groups = [];
    let group_start = 0;
    let current_type = filtered_ids[0] == image_token_id ? 1 : 0;
    for (let j = 1; j <= filtered_ids.length; ++j) {
      const t = j < filtered_ids.length ? filtered_ids[j] == image_token_id ? 1 : 0 : -1;
      if (t !== current_type) {
        groups.push([current_type, group_start, j]);
        group_start = j;
        current_type = t;
      }
    }
    let current_pos = 0;
    const llm_pos_ids_list = [];
    for (const [modality_type, start_idx, end_idx] of groups) {
      if (modality_type === 0) {
        const text_len = end_idx - start_idx;
        llm_pos_ids_list.push(Array.from({ length: 3 * text_len }, (_, i) => current_pos + i % text_len));
        current_pos += text_len;
      } else {
        const grid_thw = image_grid_thw_list[state.image_index++].map(Number);
        const temp_merge_size = grid_thw[0];
        llm_pos_ids_list.push(
          this.get_vision_position_ids(current_pos, grid_thw, temp_merge_size, spatial_merge_size)
        );
        current_pos += Math.max(grid_thw[1], grid_thw[2]) / spatial_merge_size;
      }
    }
    return llm_pos_ids_list;
  }
};

// src/models/glpn/modeling_glpn.js
var GLPNPreTrainedModel = class extends PreTrainedModel {
};
var GLPNModel = class extends GLPNPreTrainedModel {
};
var GLPNForDepthEstimation = class extends GLPNPreTrainedModel {
};

// src/models/gpt_bigcode/modeling_gpt_bigcode.js
var GPTBigCodePreTrainedModel = class extends PreTrainedModel {
};
var GPTBigCodeModel = class extends GPTBigCodePreTrainedModel {
};
var GPTBigCodeForCausalLM = class extends GPTBigCodePreTrainedModel {
};

// src/models/gpt_neo/modeling_gpt_neo.js
var GPTNeoPreTrainedModel = class extends PreTrainedModel {
};
var GPTNeoModel = class extends GPTNeoPreTrainedModel {
};
var GPTNeoForCausalLM = class extends GPTNeoPreTrainedModel {
};

// src/models/gpt_neox/modeling_gpt_neox.js
var GPTNeoXPreTrainedModel = class extends PreTrainedModel {
};
var GPTNeoXModel = class extends GPTNeoXPreTrainedModel {
};
var GPTNeoXForCausalLM = class extends GPTNeoXPreTrainedModel {
};

// src/models/gpt_oss/modeling_gpt_oss.js
var GptOssPreTrainedModel = class extends PreTrainedModel {
};
var GptOssModel = class extends GptOssPreTrainedModel {
};
var GptOssForCausalLM = class extends GptOssPreTrainedModel {
};

// src/models/gpt2/modeling_gpt2.js
var GPT2PreTrainedModel = class extends PreTrainedModel {
};
var GPT2Model = class extends GPT2PreTrainedModel {
};
var GPT2LMHeadModel = class extends GPT2PreTrainedModel {
};

// src/models/gptj/modeling_gptj.js
var GPTJPreTrainedModel = class extends PreTrainedModel {
};
var GPTJModel = class extends GPTJPreTrainedModel {
};
var GPTJForCausalLM = class extends GPTJPreTrainedModel {
};

// src/models/granite/modeling_granite.js
var GranitePreTrainedModel = class extends PreTrainedModel {
};
var GraniteModel = class extends GranitePreTrainedModel {
};
var GraniteForCausalLM = class extends GranitePreTrainedModel {
};

// src/models/granitemoehybrid/modeling_granitemoehybrid.js
var GraniteMoeHybridPreTrainedModel = class extends PreTrainedModel {
};
var GraniteMoeHybridModel = class extends GraniteMoeHybridPreTrainedModel {
};
var GraniteMoeHybridForCausalLM = class extends GraniteMoeHybridPreTrainedModel {
};

// src/models/ultravox/modeling_ultravox.js
var UltravoxPreTrainedModel = class extends PreTrainedModel {
  forward_params = ["input_ids", "attention_mask", "position_ids", "audio_values", "past_key_values"];
};
var UltravoxModel = class extends UltravoxPreTrainedModel {
  _merge_input_ids_with_audio_features(kwargs) {
    const audio_hidden_size = kwargs.audio_features.dims.at(-1);
    const reshaped_audio_features = kwargs.audio_features.view(-1, audio_hidden_size);
    return default_merge_input_ids_with_audio_features({
      // @ts-ignore
      audio_token_id: this.config.ignore_index ?? this.config.audio_token_id ?? this.config.audio_token_index,
      ...kwargs,
      audio_features: reshaped_audio_features
    });
  }
};

// src/models/granite_speech/modeling_granite_speech.js
var GraniteSpeechForConditionalGeneration = class extends UltravoxModel {
  forward_params = ["input_ids", "attention_mask", "input_features", "past_key_values"];
};

// src/models/grounding_dino/modeling_grounding_dino.js
var GroundingDinoPreTrainedModel = class extends PreTrainedModel {
};
var GroundingDinoForObjectDetection = class extends GroundingDinoPreTrainedModel {
};

// src/models/groupvit/modeling_groupvit.js
var GroupViTPreTrainedModel = class extends PreTrainedModel {
};
var GroupViTModel = class extends GroupViTPreTrainedModel {
};

// src/models/helium/modeling_helium.js
var HeliumPreTrainedModel = class extends PreTrainedModel {
};
var HeliumModel = class extends HeliumPreTrainedModel {
};
var HeliumForCausalLM = class extends HeliumPreTrainedModel {
};

// src/models/hiera/modeling_hiera.js
var HieraPreTrainedModel = class extends PreTrainedModel {
};
var HieraModel = class extends HieraPreTrainedModel {
};
var HieraForImageClassification = class extends HieraPreTrainedModel {
  /**
   * @param {any} model_inputs
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};

// src/models/wav2vec2/modeling_wav2vec2.js
var Wav2Vec2PreTrainedModel = class extends PreTrainedModel {
};
var Wav2Vec2Model = class extends Wav2Vec2PreTrainedModel {
};
var Wav2Vec2ForCTC = class extends Wav2Vec2PreTrainedModel {
  /**
   * @param {Object} model_inputs
   * @param {Tensor} model_inputs.input_values Float values of input raw speech waveform.
   * @param {Tensor} model_inputs.attention_mask Mask to avoid performing convolution and attention on padding token indices. Mask values selected in [0, 1]
   */
  async _call(model_inputs) {
    return new CausalLMOutput(await super._call(model_inputs));
  }
};
var Wav2Vec2ForSequenceClassification = class extends Wav2Vec2PreTrainedModel {
  /**
   * Calls the model on new inputs.
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<SequenceClassifierOutput>} An object containing the model's output logits for sequence classification.
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};
var Wav2Vec2ForAudioFrameClassification = class extends Wav2Vec2PreTrainedModel {
  /**
   * Calls the model on new inputs.
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<TokenClassifierOutput>} An object containing the model's output logits for sequence classification.
   */
  async _call(model_inputs) {
    return new TokenClassifierOutput(await super._call(model_inputs));
  }
};

// src/models/hubert/modeling_hubert.js
var HubertPreTrainedModel = class extends PreTrainedModel {
};
var HubertModel = class extends Wav2Vec2PreTrainedModel {
};
var HubertForCTC = class extends Wav2Vec2PreTrainedModel {
  /**
   * @param {Object} model_inputs
   * @param {Tensor} model_inputs.input_values Float values of input raw speech waveform.
   * @param {Tensor} model_inputs.attention_mask Mask to avoid performing convolution and attention on padding token indices. Mask values selected in [0, 1]
   */
  async _call(model_inputs) {
    return new CausalLMOutput(await super._call(model_inputs));
  }
};
var HubertForSequenceClassification = class extends Wav2Vec2PreTrainedModel {
  /**
   * Calls the model on new inputs.
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<SequenceClassifierOutput>} An object containing the model's output logits for sequence classification.
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};

// src/models/hunyuan_v1_dense/modeling_hunyuan_v1_dense.js
var HunYuanDenseV1PreTrainedModel = class extends PreTrainedModel {
};
var HunYuanDenseV1Model = class extends HunYuanDenseV1PreTrainedModel {
};
var HunYuanDenseV1ForCausalLM = class extends HunYuanDenseV1PreTrainedModel {
};

// src/models/idefics3/modeling_idefics3.js
var Idefics3ForConditionalGeneration = class extends LlavaForConditionalGeneration {
  forward_params = [
    "input_ids",
    "attention_mask",
    "pixel_values",
    "pixel_attention_mask",
    "position_ids",
    "past_key_values"
  ];
};

// src/models/ijepa/modeling_ijepa.js
var IJepaPreTrainedModel = class extends PreTrainedModel {
};
var IJepaModel = class extends IJepaPreTrainedModel {
};
var IJepaForImageClassification = class extends IJepaPreTrainedModel {
  /**
   * @param {any} model_inputs
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};

// src/models/jais/modeling_jais.js
var JAISPreTrainedModel = class extends PreTrainedModel {
};
var JAISModel = class extends JAISPreTrainedModel {
};
var JAISLMHeadModel = class extends JAISPreTrainedModel {
};

// src/models/jina_clip/modeling_jina_clip.js
var JinaCLIPPreTrainedModel = class extends PreTrainedModel {
};
var JinaCLIPModel = class extends JinaCLIPPreTrainedModel {
  async forward(model_inputs) {
    const missing_text_inputs = !model_inputs.input_ids;
    const missing_image_inputs = !model_inputs.pixel_values;
    if (missing_text_inputs && missing_image_inputs) {
      throw new Error("Either `input_ids` or `pixel_values` should be provided.");
    }
    if (missing_text_inputs) {
      model_inputs.input_ids = ones([model_inputs.pixel_values.dims[0], 1]);
    }
    if (missing_image_inputs) {
      const { image_size } = this.config.vision_config;
      model_inputs.pixel_values = full([0, 3, image_size, image_size], 0);
    }
    const { text_embeddings, image_embeddings, l2norm_text_embeddings, l2norm_image_embeddings } = await super.forward(model_inputs);
    const result = {};
    if (!missing_text_inputs) {
      result.text_embeddings = text_embeddings;
      result.l2norm_text_embeddings = l2norm_text_embeddings;
    }
    if (!missing_image_inputs) {
      result.image_embeddings = image_embeddings;
      result.l2norm_image_embeddings = l2norm_image_embeddings;
    }
    return result;
  }
};
var JinaCLIPTextModel = class extends JinaCLIPPreTrainedModel {
  /** @type {typeof PreTrainedModel.from_pretrained} */
  static async from_pretrained(pretrained_model_name_or_path, options = {}) {
    return super.from_pretrained(pretrained_model_name_or_path, {
      ...options,
      // Update default model file name if not provided
      model_file_name: options.model_file_name ?? "text_model"
    });
  }
};
var JinaCLIPVisionModel = class extends JinaCLIPPreTrainedModel {
  /** @type {typeof PreTrainedModel.from_pretrained} */
  static async from_pretrained(pretrained_model_name_or_path, options = {}) {
    return super.from_pretrained(pretrained_model_name_or_path, {
      ...options,
      // Update default model file name if not provided
      model_file_name: options.model_file_name ?? "vision_model"
    });
  }
};

// src/models/lfm2/modeling_lfm2.js
var Lfm2PreTrainedModel = class extends PreTrainedModel {
};
var Lfm2Model = class extends Lfm2PreTrainedModel {
};
var Lfm2ForCausalLM = class extends Lfm2PreTrainedModel {
};

// src/models/lighton_ocr/modeling_lighton_ocr.js
var LightOnOcrForConditionalGeneration = class extends LlavaForConditionalGeneration {
};

// src/models/lfm2_moe/modeling_lfm2_moe.js
var Lfm2MoePreTrainedModel = class extends PreTrainedModel {
};
var Lfm2MoeModel = class extends Lfm2MoePreTrainedModel {
};
var Lfm2MoeForCausalLM = class extends Lfm2MoePreTrainedModel {
};

// src/models/lfm2_vl/modeling_lfm2_vl.js
var Lfm2VlForConditionalGeneration = class extends LlavaForConditionalGeneration {
  forward_params = [
    "input_ids",
    "attention_mask",
    "pixel_values",
    "pixel_attention_mask",
    "spatial_shapes",
    "position_ids",
    "past_key_values"
  ];
};

// src/models/llama/modeling_llama.js
var LlamaPreTrainedModel = class extends PreTrainedModel {
};
var LlamaModel = class extends LlamaPreTrainedModel {
};
var LlamaForCausalLM = class extends LlamaPreTrainedModel {
};

// src/models/llama4/modeling_llama4.js
var Llama4PreTrainedModel = class extends PreTrainedModel {
};
var Llama4ForCausalLM = class extends Llama4PreTrainedModel {
};

// src/models/longt5/modeling_longt5.js
var LongT5PreTrainedModel = class extends PreTrainedModel {
};
var LongT5Model = class extends LongT5PreTrainedModel {
};
var LongT5ForConditionalGeneration = class extends LongT5PreTrainedModel {
};

// src/models/m2m_100/modeling_m2m_100.js
var M2M100PreTrainedModel = class extends PreTrainedModel {
};
var M2M100Model = class extends M2M100PreTrainedModel {
};
var M2M100ForConditionalGeneration = class extends M2M100PreTrainedModel {
};

// src/models/marian/modeling_marian.js
var MarianPreTrainedModel = class extends PreTrainedModel {
};
var MarianModel = class extends MarianPreTrainedModel {
};
var MarianMTModel = class extends MarianPreTrainedModel {
};

// src/models/maskformer/modeling_maskformer.js
var MaskFormerPreTrainedModel = class extends PreTrainedModel {
};
var MaskFormerModel = class extends MaskFormerPreTrainedModel {
};
var MaskFormerForInstanceSegmentation = class extends MaskFormerPreTrainedModel {
};

// src/models/mbart/modeling_mbart.js
var MBartPreTrainedModel = class extends PreTrainedModel {
};
var MBartModel = class extends MBartPreTrainedModel {
};
var MBartForConditionalGeneration = class extends MBartPreTrainedModel {
};
var MBartForSequenceClassification = class extends MBartPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<SequenceClassifierOutput>} An object containing the model's output logits for sequence classification.
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};
var MBartForCausalLM = class extends MBartPreTrainedModel {
};

// src/models/metric3d/modeling_metric3d.js
var Metric3DPreTrainedModel = class extends PreTrainedModel {
};
var Metric3DForDepthEstimation = class extends Metric3DPreTrainedModel {
};

// src/models/metric3dv2/modeling_metric3dv2.js
var Metric3Dv2PreTrainedModel = class extends PreTrainedModel {
};
var Metric3Dv2ForDepthEstimation = class extends Metric3Dv2PreTrainedModel {
};

// src/models/mgp_str/modeling_mgp_str.js
var MgpstrModelOutput = class extends ModelOutput {
  constructor({ char_logits, bpe_logits, wp_logits }) {
    super();
    this.char_logits = char_logits;
    this.bpe_logits = bpe_logits;
    this.wp_logits = wp_logits;
  }
  get logits() {
    return [this.char_logits, this.bpe_logits, this.wp_logits];
  }
};
var MgpstrPreTrainedModel = class extends PreTrainedModel {
};
var MgpstrForSceneTextRecognition = class extends MgpstrPreTrainedModel {
  /**
   * @param {any} model_inputs
   */
  async _call(model_inputs) {
    return new MgpstrModelOutput(await super._call(model_inputs));
  }
};

// src/models/mimi/modeling_mimi.js
var MimiEncoderOutput = class extends ModelOutput {
  /**
   * @param {Object} output The output of the model.
   * @param {Tensor} output.audio_codes Discrete code embeddings, of shape `(batch_size, num_quantizers, codes_length)`.
   */
  constructor({ audio_codes }) {
    super();
    this.audio_codes = audio_codes;
  }
};
var MimiDecoderOutput = class extends ModelOutput {
  /**
   * @param {Object} output The output of the model.
   * @param {Tensor} output.audio_values Decoded audio values, of shape `(batch_size, num_channels, sequence_length)`.
   */
  constructor({ audio_values }) {
    super();
    this.audio_values = audio_values;
  }
};
var MimiPreTrainedModel = class extends PreTrainedModel {
  main_input_name = "input_values";
  forward_params = ["input_values"];
};
var MimiModel = class extends MimiPreTrainedModel {
  /**
   * Encodes the input audio waveform into discrete codes.
   * @param {Object} inputs Model inputs
   * @param {Tensor} [inputs.input_values] Float values of the input audio waveform, of shape `(batch_size, channels, sequence_length)`).
   * @returns {Promise<MimiEncoderOutput>} The output tensor of shape `(batch_size, num_codebooks, sequence_length)`.
   */
  async encode(inputs) {
    return new MimiEncoderOutput(await sessionRun(this.sessions["encoder_model"], inputs));
  }
  /**
   * Decodes the given frames into an output audio waveform.
   * @param {MimiEncoderOutput} inputs The encoded audio codes.
   * @returns {Promise<MimiDecoderOutput>} The output tensor of shape `(batch_size, num_channels, sequence_length)`.
   */
  async decode(inputs) {
    return new MimiDecoderOutput(await sessionRun(this.sessions["decoder_model"], inputs));
  }
};
var MimiEncoderModel = class extends MimiPreTrainedModel {
  /** @type {typeof PreTrainedModel.from_pretrained} */
  static async from_pretrained(pretrained_model_name_or_path, options = {}) {
    return super.from_pretrained(pretrained_model_name_or_path, {
      ...options,
      // Update default model file name if not provided
      model_file_name: options.model_file_name ?? "encoder_model"
    });
  }
};
var MimiDecoderModel = class extends MimiPreTrainedModel {
  /** @type {typeof PreTrainedModel.from_pretrained} */
  static async from_pretrained(pretrained_model_name_or_path, options = {}) {
    return super.from_pretrained(pretrained_model_name_or_path, {
      ...options,
      // Update default model file name if not provided
      model_file_name: options.model_file_name ?? "decoder_model"
    });
  }
};

// src/models/mistral/modeling_mistral.js
var MistralPreTrainedModel = class extends PreTrainedModel {
};
var MistralModel = class extends MistralPreTrainedModel {
};
var MistralForCausalLM = class extends MistralPreTrainedModel {
};

// src/models/mistral4/modeling_mistral4.js
var Mistral4PreTrainedModel = class extends PreTrainedModel {
};
var Mistral4Model = class extends Mistral4PreTrainedModel {
};
var Mistral4ForCausalLM = class extends Mistral4PreTrainedModel {
};

// src/models/mobilebert/modeling_mobilebert.js
var MobileBertPreTrainedModel = class extends PreTrainedModel {
};
var MobileBertModel = class extends MobileBertPreTrainedModel {
};
var MobileBertForMaskedLM = class extends MobileBertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<MaskedLMOutput>} returned object
   */
  async _call(model_inputs) {
    return new MaskedLMOutput(await super._call(model_inputs));
  }
};
var MobileBertForSequenceClassification = class extends MobileBertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<SequenceClassifierOutput>} returned object
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};
var MobileBertForQuestionAnswering = class extends MobileBertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<QuestionAnsweringModelOutput>} returned object
   */
  async _call(model_inputs) {
    return new QuestionAnsweringModelOutput(await super._call(model_inputs));
  }
};

// src/models/mobilellm/modeling_mobilellm.js
var MobileLLMPreTrainedModel = class extends PreTrainedModel {
};
var MobileLLMModel = class extends MobileLLMPreTrainedModel {
};
var MobileLLMForCausalLM = class extends MobileLLMPreTrainedModel {
};

// src/models/mobilenet_v1/modeling_mobilenet_v1.js
var MobileNetV1PreTrainedModel = class extends PreTrainedModel {
};
var MobileNetV1Model = class extends MobileNetV1PreTrainedModel {
};
var MobileNetV1ForImageClassification = class extends MobileNetV1PreTrainedModel {
  /**
   * @param {any} model_inputs
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};
var MobileNetV1ForSemanticSegmentation = class extends MobileNetV1PreTrainedModel {
};

// src/models/mobilenet_v2/modeling_mobilenet_v2.js
var MobileNetV2PreTrainedModel = class extends PreTrainedModel {
};
var MobileNetV2Model = class extends MobileNetV2PreTrainedModel {
};
var MobileNetV2ForImageClassification = class extends MobileNetV2PreTrainedModel {
  /**
   * @param {any} model_inputs
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};
var MobileNetV2ForSemanticSegmentation = class extends MobileNetV2PreTrainedModel {
};

// src/models/mobilenet_v3/modeling_mobilenet_v3.js
var MobileNetV3PreTrainedModel = class extends PreTrainedModel {
};
var MobileNetV3Model = class extends MobileNetV3PreTrainedModel {
};
var MobileNetV3ForImageClassification = class extends MobileNetV3PreTrainedModel {
  /**
   * @param {any} model_inputs
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};
var MobileNetV3ForSemanticSegmentation = class extends MobileNetV3PreTrainedModel {
};

// src/models/mobilenet_v4/modeling_mobilenet_v4.js
var MobileNetV4PreTrainedModel = class extends PreTrainedModel {
};
var MobileNetV4Model = class extends MobileNetV4PreTrainedModel {
};
var MobileNetV4ForImageClassification = class extends MobileNetV4PreTrainedModel {
  /**
   * @param {any} model_inputs
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};
var MobileNetV4ForSemanticSegmentation = class extends MobileNetV4PreTrainedModel {
};

// src/models/mobilevit/modeling_mobilevit.js
var MobileViTPreTrainedModel = class extends PreTrainedModel {
};
var MobileViTModel = class extends MobileViTPreTrainedModel {
};
var MobileViTForImageClassification = class extends MobileViTPreTrainedModel {
  /**
   * @param {any} model_inputs
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};

// src/models/mobilevitv2/modeling_mobilevitv2.js
var MobileViTV2PreTrainedModel = class extends PreTrainedModel {
};
var MobileViTV2Model = class extends MobileViTV2PreTrainedModel {
};
var MobileViTV2ForImageClassification = class extends MobileViTV2PreTrainedModel {
  /**
   * @param {any} model_inputs
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};

// src/models/modernbert/modeling_modernbert.js
var ModernBertPreTrainedModel = class extends PreTrainedModel {
};
var ModernBertModel = class extends ModernBertPreTrainedModel {
};
var ModernBertForMaskedLM = class extends ModernBertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<MaskedLMOutput>} An object containing the model's output logits for masked language modeling.
   */
  async _call(model_inputs) {
    return new MaskedLMOutput(await super._call(model_inputs));
  }
};
var ModernBertForSequenceClassification = class extends ModernBertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<SequenceClassifierOutput>} An object containing the model's output logits for sequence classification.
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};
var ModernBertForTokenClassification = class extends ModernBertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<TokenClassifierOutput>} An object containing the model's output logits for token classification.
   */
  async _call(model_inputs) {
    return new TokenClassifierOutput(await super._call(model_inputs));
  }
};

// src/models/modernbert_decoder/modeling_modernbert_decoder.js
var ModernBertDecoderPreTrainedModel = class extends PreTrainedModel {
};
var ModernBertDecoderModel = class extends ModernBertDecoderPreTrainedModel {
};
var ModernBertDecoderForCausalLM = class extends ModernBertDecoderPreTrainedModel {
};

// src/models/moonshine/modeling_moonshine.js
var MoonshinePreTrainedModel = class extends PreTrainedModel {
  requires_attention_mask = false;
  main_input_name = "input_values";
  forward_params = ["input_values", "decoder_input_ids", "past_key_values"];
};
var MoonshineModel = class extends MoonshinePreTrainedModel {
};
var MoonshineForConditionalGeneration = class extends MoonshinePreTrainedModel {
};

// src/models/mpnet/modeling_mpnet.js
var MPNetPreTrainedModel = class extends PreTrainedModel {
};
var MPNetModel = class extends MPNetPreTrainedModel {
};
var MPNetForMaskedLM = class extends MPNetPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<MaskedLMOutput>} An object containing the model's output logits for masked language modeling.
   */
  async _call(model_inputs) {
    return new MaskedLMOutput(await super._call(model_inputs));
  }
};
var MPNetForSequenceClassification = class extends MPNetPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<SequenceClassifierOutput>} An object containing the model's output logits for sequence classification.
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};
var MPNetForTokenClassification = class extends MPNetPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<TokenClassifierOutput>} An object containing the model's output logits for token classification.
   */
  async _call(model_inputs) {
    return new TokenClassifierOutput(await super._call(model_inputs));
  }
};
var MPNetForQuestionAnswering = class extends MPNetPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<QuestionAnsweringModelOutput>} An object containing the model's output logits for question answering.
   */
  async _call(model_inputs) {
    return new QuestionAnsweringModelOutput(await super._call(model_inputs));
  }
};

// src/models/mpt/modeling_mpt.js
var MptPreTrainedModel = class extends PreTrainedModel {
};
var MptModel = class extends MptPreTrainedModel {
};
var MptForCausalLM = class extends MptPreTrainedModel {
};

// src/models/mt5/modeling_mt5.js
var MT5PreTrainedModel = class extends PreTrainedModel {
};
var MT5Model = class extends MT5PreTrainedModel {
};
var MT5ForConditionalGeneration = class extends MT5PreTrainedModel {
};

// src/models/multi_modality/modeling_multi_modality.js
var MultiModalityPreTrainedModel = class extends PreTrainedModel {
};
var MultiModalityCausalLM = class extends MultiModalityPreTrainedModel {
  forward_params = [
    // prepare_inputs_embeds
    "input_ids",
    "pixel_values",
    "images_seq_mask",
    "images_emb_mask",
    // language_model
    "attention_mask",
    "position_ids",
    "past_key_values"
  ];
  /**
   * @param {ConstructorParameters<typeof MultiModalityPreTrainedModel>} args
   */
  constructor(...args) {
    super(...args);
    this._generation_mode = "text";
  }
  async forward(model_inputs) {
    const mode = this._generation_mode ?? "text";
    let output_1;
    if (mode === "text" || !model_inputs.past_key_values) {
      const session = this.sessions["prepare_inputs_embeds"];
      const prep_inputs = pick(model_inputs, session.inputNames);
      output_1 = await sessionRun(session, prep_inputs);
    } else {
      const session = this.sessions["gen_img_embeds"];
      const prep_inputs = pick(
        {
          image_ids: model_inputs.input_ids
        },
        session.inputNames
      );
      output_1 = await sessionRun(session, prep_inputs);
    }
    const input_2 = { ...model_inputs, ...output_1 };
    const output_2 = await decoder_forward(this, input_2);
    const head = this.sessions[mode === "text" ? "lm_head" : "gen_head"];
    if (!head) {
      throw new Error(`Unable to find "${head}" generation head`);
    }
    const output_3 = await sessionRun(head, pick(output_2, head.inputNames));
    return {
      ...output_1,
      ...output_2,
      ...output_3
    };
  }
  prepare_inputs_for_generation(input_ids, model_inputs, generation_config) {
    const has_past_key_values = !!model_inputs.past_key_values;
    if (generation_config.guidance_scale !== null && generation_config.guidance_scale > 1) {
      if (has_past_key_values) {
        model_inputs.input_ids = cat([model_inputs.input_ids, model_inputs.input_ids], 0);
      } else {
        model_inputs.input_ids = cat(
          [model_inputs.input_ids, full_like(model_inputs.input_ids, BigInt(generation_config.pad_token_id))],
          0
        );
        model_inputs.attention_mask = cat(
          [model_inputs.attention_mask, full_like(model_inputs.attention_mask, 0n)],
          0
        );
      }
    }
    if (has_past_key_values || !model_inputs.pixel_values) {
      model_inputs.pixel_values = full([0, 0, 3, 384, 384], 1);
    }
    if (has_past_key_values) {
      const num_img_tokens = 0;
      const num_text_tokens = 1;
      const has_image = num_img_tokens > 0 ? 1 : 0;
      const batch_size = 1;
      model_inputs.images_seq_mask = new Tensor2(
        "bool",
        new Array(num_img_tokens + num_text_tokens).fill(true).fill(false, 0, num_text_tokens),
        [batch_size, num_img_tokens + num_text_tokens]
      );
      model_inputs.images_emb_mask = new Tensor2("bool", new Array(num_img_tokens).fill(!!has_image), [
        batch_size,
        1,
        num_img_tokens
      ]);
    }
    return model_inputs;
  }
  /**
   * @param {import('../../generation/parameters.js').GenerationFunctionParameters} options
   */
  async generate(options) {
    this._generation_mode = "text";
    return super.generate(options);
  }
  /**
   * @param {import('../../generation/parameters.js').GenerationFunctionParameters} options
   */
  async generate_images(options) {
    this._generation_mode = "image";
    const start_num_tokens = (
      /** @type {Tensor} */
      (options.inputs ?? options[this.main_input_name]).dims[1]
    );
    const all_tokens = await super.generate(options);
    const generated_tokens = (
      /** @type {Tensor} */
      all_tokens.slice(null, [start_num_tokens, null])
    );
    const image_decode = this.sessions["image_decode"];
    const { decoded_image } = await sessionRun(image_decode, {
      generated_tokens
    });
    const clamped = decoded_image.add_(1).mul_(255 / 2).clamp_(0, 255).to("uint8");
    const images = [];
    for (const tensor of clamped) {
      const img = RawImage.fromTensor(tensor);
      images.push(img);
    }
    return images;
  }
};

// src/models/musicgen/modeling_musicgen.js
var MusicgenPreTrainedModel = class extends PreTrainedModel {
};
var MusicgenModel = class extends MusicgenPreTrainedModel {
};
var MusicgenForCausalLM = class extends MusicgenPreTrainedModel {
};
var MusicgenForConditionalGeneration = class extends PreTrainedModel {
  // NOTE: not MusicgenPreTrainedModel
  forward_params = [
    "input_ids",
    "attention_mask",
    "encoder_outputs",
    "decoder_input_ids",
    "decoder_attention_mask",
    "past_key_values"
  ];
  /**
   * Apply the pattern mask to the final ids,
   * then revert the pattern delay mask by filtering the pad token id in a single step.
   * @param {Tensor} outputs The output tensor from the model.
   * @returns {Tensor} The filtered output tensor.
   */
  _apply_and_filter_by_delay_pattern_mask(outputs) {
    const [bs_x_codebooks, seqLength] = outputs.dims;
    const num_codebooks = this.config.decoder.num_codebooks;
    const upperBound = seqLength - num_codebooks;
    let newDataSize = 0;
    for (let i = 0; i < outputs.size; ++i) {
      if (outputs.data[i] == this.config.decoder.pad_token_id) {
        continue;
      }
      const row = i % seqLength;
      const col = Math.floor(i / seqLength) % num_codebooks;
      const diff = row - col;
      if (diff > 0 && diff <= upperBound) {
        outputs.data[newDataSize++] = outputs.data[i];
      }
    }
    const batch_size = Math.floor(bs_x_codebooks / num_codebooks);
    const inferred = newDataSize / (batch_size * num_codebooks);
    return new Tensor2(outputs.type, outputs.data.slice(0, newDataSize), [batch_size, num_codebooks, inferred]);
  }
  prepare_inputs_for_generation(input_ids, model_inputs, generation_config) {
    const pad_token_id = BigInt(this.config.decoder.pad_token_id);
    let clonedInputIds = structuredClone(input_ids);
    for (let i = 0; i < clonedInputIds.length; ++i) {
      for (let j = 0; j < clonedInputIds[i].length; ++j) {
        if (i % this.config.decoder.num_codebooks >= j) {
          clonedInputIds[i][j] = pad_token_id;
        }
      }
    }
    if (generation_config.guidance_scale !== null && generation_config.guidance_scale > 1) {
      clonedInputIds = clonedInputIds.concat(clonedInputIds);
    }
    return encoder_decoder_prepare_inputs_for_generation(this, clonedInputIds, model_inputs, generation_config);
  }
  /**
   * Generates sequences of token ids for models with a language modeling head.
   * @param {import('../../generation/parameters.js').GenerationFunctionParameters} options
   * @returns {Promise<ModelOutput|Tensor>} The output of the model, which can contain the generated token ids, attentions, and scores.
   */
  async generate(options) {
    const output_ids = await super.generate(options);
    const audio_codes = this._apply_and_filter_by_delay_pattern_mask(
      /** @type {Tensor} */
      output_ids
    ).unsqueeze_(
      0
    );
    const { audio_values } = await sessionRun(this.sessions["encodec_decode"], { audio_codes });
    return audio_values;
  }
};

// src/models/nanochat/modeling_nanochat.js
var NanoChatPreTrainedModel = class extends PreTrainedModel {
};
var NanoChatModel = class extends NanoChatPreTrainedModel {
};
var NanoChatForCausalLM = class extends NanoChatPreTrainedModel {
};

// src/models/nemotron_h/modeling_nemotron_h.js
var NemotronHPreTrainedModel = class extends PreTrainedModel {
};
var NemotronHModel = class extends NemotronHPreTrainedModel {
};
var NemotronHForCausalLM = class extends NemotronHPreTrainedModel {
};

// src/models/neobert/modeling_neobert.js
var NeoBertPreTrainedModel = class extends PreTrainedModel {
};
var NeoBertModel = class extends NeoBertPreTrainedModel {
};
var NeoBertForMaskedLM = class extends NeoBertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<MaskedLMOutput>} An object containing the model's output logits for masked language modeling.
   */
  async _call(model_inputs) {
    return new MaskedLMOutput(await super._call(model_inputs));
  }
};
var NeoBertForSequenceClassification = class extends NeoBertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<SequenceClassifierOutput>} An object containing the model's output logits for sequence classification.
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};
var NeoBertForTokenClassification = class extends NeoBertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<TokenClassifierOutput>} An object containing the model's output logits for token classification.
   */
  async _call(model_inputs) {
    return new TokenClassifierOutput(await super._call(model_inputs));
  }
};
var NeoBertForQuestionAnswering = class extends NeoBertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<QuestionAnsweringModelOutput>} An object containing the model's output logits for question answering.
   */
  async _call(model_inputs) {
    return new QuestionAnsweringModelOutput(await super._call(model_inputs));
  }
};

// src/models/nomic_bert/modeling_nomic_bert.js
var NomicBertPreTrainedModel = class extends PreTrainedModel {
};
var NomicBertModel = class extends NomicBertPreTrainedModel {
};

// src/models/olmo/modeling_olmo.js
var OlmoPreTrainedModel = class extends PreTrainedModel {
};
var OlmoModel = class extends OlmoPreTrainedModel {
};
var OlmoForCausalLM = class extends OlmoPreTrainedModel {
};

// src/models/olmo2/modeling_olmo2.js
var Olmo2PreTrainedModel = class extends PreTrainedModel {
};
var Olmo2Model = class extends Olmo2PreTrainedModel {
};
var Olmo2ForCausalLM = class extends Olmo2PreTrainedModel {
};

// src/models/olmo3/modeling_olmo3.js
var Olmo3PreTrainedModel = class extends PreTrainedModel {
};
var Olmo3Model = class extends Olmo3PreTrainedModel {
};
var Olmo3ForCausalLM = class extends Olmo3PreTrainedModel {
};

// src/models/olmo_hybrid/modeling_olmo_hybrid.js
var OlmoHybridPreTrainedModel = class extends PreTrainedModel {
};
var OlmoHybridModel = class extends OlmoHybridPreTrainedModel {
};
var OlmoHybridForCausalLM = class extends OlmoHybridPreTrainedModel {
};

// src/models/openai_privacy_filter/modeling_openai_privacy_filter.js
var OpenAIPrivacyFilterPreTrainedModel = class extends PreTrainedModel {
};
var OpenAIPrivacyFilterModel = class extends OpenAIPrivacyFilterPreTrainedModel {
};
var OpenAIPrivacyFilterForTokenClassification = class extends OpenAIPrivacyFilterPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<SequenceClassifierOutput>} An object containing the model's output logits for sequence classification.
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};

// src/models/openelm/modeling_openelm.js
var OpenELMPreTrainedModel = class extends PreTrainedModel {
};
var OpenELMModel = class extends OpenELMPreTrainedModel {
};
var OpenELMForCausalLM = class extends OpenELMPreTrainedModel {
};

// src/models/opt/modeling_opt.js
var OPTPreTrainedModel = class extends PreTrainedModel {
};
var OPTModel = class extends OPTPreTrainedModel {
};
var OPTForCausalLM = class extends OPTPreTrainedModel {
};

// src/models/owlv2/modeling_owlv2.js
var Owlv2PreTrainedModel = class extends PreTrainedModel {
};
var Owlv2Model = class extends Owlv2PreTrainedModel {
};
var Owlv2ForObjectDetection = class extends Owlv2PreTrainedModel {
};

// src/models/owlvit/modeling_owlvit.js
var OwlViTPreTrainedModel = class extends PreTrainedModel {
};
var OwlViTModel = class extends OwlViTPreTrainedModel {
};
var OwlViTForObjectDetection = class extends OwlViTPreTrainedModel {
};

// src/models/paligemma/modeling_paligemma.js
var PaliGemmaForConditionalGeneration = class extends LlavaForConditionalGeneration {
};

// src/models/parakeet/modeling_parakeet.js
var ParakeetPreTrainedModel = class extends PreTrainedModel {
};
var ParakeetForCTC = class extends ParakeetPreTrainedModel {
  /**
   * @param {Object} model_inputs
   * @param {Tensor} model_inputs.input_values Float values of input raw speech waveform.
   * @param {Tensor} model_inputs.attention_mask Mask to avoid performing convolution and attention on padding token indices. Mask values selected in [0, 1]
   */
  async _call(model_inputs) {
    return new CausalLMOutput(await super._call(model_inputs));
  }
};

// src/models/patchtsmixer/modeling_patchtsmixer.js
var PatchTSMixerPreTrainedModel = class extends PreTrainedModel {
};
var PatchTSMixerModel = class extends PatchTSMixerPreTrainedModel {
};
var PatchTSMixerForPrediction = class extends PatchTSMixerPreTrainedModel {
};

// src/models/patchtst/modeling_patchtst.js
var PatchTSTPreTrainedModel = class extends PreTrainedModel {
};
var PatchTSTModel = class extends PatchTSTPreTrainedModel {
};
var PatchTSTForPrediction = class extends PatchTSTPreTrainedModel {
};

// src/models/phi/modeling_phi.js
var PhiPreTrainedModel = class extends PreTrainedModel {
};
var PhiModel = class extends PhiPreTrainedModel {
};
var PhiForCausalLM = class extends PhiPreTrainedModel {
};

// src/models/phi3/modeling_phi3.js
var Phi3PreTrainedModel = class extends PreTrainedModel {
};
var Phi3Model = class extends Phi3PreTrainedModel {
};
var Phi3ForCausalLM = class extends Phi3PreTrainedModel {
};

// src/models/phi3_v/modeling_phi3_v.js
var Phi3VPreTrainedModel = class extends PreTrainedModel {
  forward_params = [
    "input_ids",
    "inputs_embeds",
    "attention_mask",
    "position_ids",
    "pixel_values",
    "image_sizes",
    "past_key_values"
  ];
};
var Phi3VForCausalLM = class extends Phi3VPreTrainedModel {
  async forward({
    // Produced by the tokenizer/processor:
    input_ids = null,
    attention_mask = null,
    pixel_values = null,
    image_sizes = null,
    // Used during generation:
    position_ids = null,
    inputs_embeds = null,
    past_key_values = null,
    // Generic generation parameters
    generation_config = null,
    logits_processor = null,
    // TODO: needed?
    ...kwargs
  }) {
    if (!inputs_embeds) {
      let image_features;
      if (pixel_values && input_ids.dims[1] !== 1) {
        if (!image_sizes) {
          throw new Error("`image_sizes` must be provided when `pixel_values` is provided.");
        }
        ({ image_features } = await sessionRun(this.sessions["vision_encoder"], {
          pixel_values,
          image_sizes
        }));
      } else {
        const hidden_size = this.config.normalized_config.hidden_size;
        image_features = new Tensor2("float32", [], [0, hidden_size]);
      }
      ({ inputs_embeds } = await sessionRun(this.sessions["prepare_inputs_embeds"], {
        input_ids,
        image_features
      }));
    }
    const outputs = await decoder_forward(
      this,
      {
        inputs_embeds,
        past_key_values,
        attention_mask,
        position_ids,
        generation_config,
        logits_processor
      },
      false
    );
    return outputs;
  }
};

// src/models/pvt/modeling_pvt.js
var PvtPreTrainedModel = class extends PreTrainedModel {
};
var PvtModel = class extends PvtPreTrainedModel {
};
var PvtForImageClassification = class extends PvtPreTrainedModel {
  /**
   * @param {any} model_inputs
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};

// src/models/pyannote/modeling_pyannote.js
var PyAnnotePreTrainedModel = class extends PreTrainedModel {
};
var PyAnnoteModel = class extends PyAnnotePreTrainedModel {
};
var PyAnnoteForAudioFrameClassification = class extends PyAnnotePreTrainedModel {
  /**
   * Calls the model on new inputs.
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<TokenClassifierOutput>} An object containing the model's output logits for sequence classification.
   */
  async _call(model_inputs) {
    return new TokenClassifierOutput(await super._call(model_inputs));
  }
};

// src/models/qwen2/modeling_qwen2.js
var Qwen2PreTrainedModel = class extends PreTrainedModel {
};
var Qwen2Model = class extends Qwen2PreTrainedModel {
};
var Qwen2ForCausalLM = class extends Qwen2PreTrainedModel {
};

// src/models/qwen2_moe/modeling_qwen2_moe.js
var Qwen2MoePreTrainedModel = class extends PreTrainedModel {
};
var Qwen2MoeModel = class extends Qwen2MoePreTrainedModel {
};
var Qwen2MoeForCausalLM = class extends Qwen2MoePreTrainedModel {
};

// src/models/qwen3/modeling_qwen3.js
var Qwen3PreTrainedModel = class extends PreTrainedModel {
};
var Qwen3Model = class extends Qwen3PreTrainedModel {
};
var Qwen3ForCausalLM = class extends Qwen3PreTrainedModel {
};

// src/models/qwen3_moe/modeling_qwen3_moe.js
var Qwen3MoePreTrainedModel = class extends PreTrainedModel {
};
var Qwen3MoeModel = class extends Qwen3MoePreTrainedModel {
};
var Qwen3MoeForCausalLM = class extends Qwen3MoePreTrainedModel {
};

// src/models/qwen3_next/modeling_qwen3_next.js
var Qwen3NextPreTrainedModel = class extends PreTrainedModel {
};
var Qwen3NextModel = class extends Qwen3NextPreTrainedModel {
};
var Qwen3NextForCausalLM = class extends Qwen3NextPreTrainedModel {
};

// src/models/qwen3_vl/modeling_qwen3_vl.js
var Qwen3VLForConditionalGeneration = class extends Qwen2_5_VLForConditionalGeneration {
};
var Qwen3VLForCausalLM = class extends Qwen2_5_VLForCausalLM {
};

// src/models/qwen3_vl_moe/modeling_qwen3_vl_moe.js
var Qwen3VLMoeForConditionalGeneration = class extends Qwen3VLForConditionalGeneration {
};
var Qwen3VLMoeForCausalLM = class extends Qwen3VLForCausalLM {
};

// src/models/qwen3_5/modeling_qwen3_5.js
var Qwen3_5ForConditionalGeneration = class extends Qwen3VLForConditionalGeneration {
};
var Qwen3_5ForCausalLM = class extends Qwen3_5ForConditionalGeneration {
};

// src/models/qwen3_5_moe/modeling_qwen3_5_moe.js
var Qwen3_5MoeForConditionalGeneration = class extends Qwen3_5ForConditionalGeneration {
};
var Qwen3_5MoeForCausalLM = class extends Qwen3_5ForCausalLM {
};

// src/models/resnet/modeling_resnet.js
var ResNetPreTrainedModel = class extends PreTrainedModel {
};
var ResNetModel = class extends ResNetPreTrainedModel {
};
var ResNetForImageClassification = class extends ResNetPreTrainedModel {
  /**
   * @param {any} model_inputs
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};

// src/models/rf_detr/modeling_rf_detr.js
var RFDetrPreTrainedModel = class extends PreTrainedModel {
};
var RFDetrModel = class extends RFDetrPreTrainedModel {
};
var RFDetrForObjectDetection = class extends RFDetrPreTrainedModel {
  /**
   * @param {any} model_inputs
   */
  async _call(model_inputs) {
    return new RFDetrObjectDetectionOutput(await super._call(model_inputs));
  }
};
var RFDetrObjectDetectionOutput = class extends RTDetrObjectDetectionOutput {
};

// src/models/roberta/modeling_roberta.js
var RobertaPreTrainedModel = class extends PreTrainedModel {
};
var RobertaModel = class extends RobertaPreTrainedModel {
};
var RobertaForMaskedLM = class extends RobertaPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<MaskedLMOutput>} returned object
   */
  async _call(model_inputs) {
    return new MaskedLMOutput(await super._call(model_inputs));
  }
};
var RobertaForSequenceClassification = class extends RobertaPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<SequenceClassifierOutput>} returned object
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};
var RobertaForTokenClassification = class extends RobertaPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<TokenClassifierOutput>} An object containing the model's output logits for token classification.
   */
  async _call(model_inputs) {
    return new TokenClassifierOutput(await super._call(model_inputs));
  }
};
var RobertaForQuestionAnswering = class extends RobertaPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<QuestionAnsweringModelOutput>} returned object
   */
  async _call(model_inputs) {
    return new QuestionAnsweringModelOutput(await super._call(model_inputs));
  }
};

// src/models/roformer/modeling_roformer.js
var RoFormerPreTrainedModel = class extends PreTrainedModel {
};
var RoFormerModel = class extends RoFormerPreTrainedModel {
};
var RoFormerForMaskedLM = class extends RoFormerPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<MaskedLMOutput>} An object containing the model's output logits for masked language modeling.
   */
  async _call(model_inputs) {
    return new MaskedLMOutput(await super._call(model_inputs));
  }
};
var RoFormerForSequenceClassification = class extends RoFormerPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<SequenceClassifierOutput>} An object containing the model's output logits for sequence classification.
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};
var RoFormerForTokenClassification = class extends RoFormerPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<TokenClassifierOutput>} An object containing the model's output logits for token classification.
   */
  async _call(model_inputs) {
    return new TokenClassifierOutput(await super._call(model_inputs));
  }
};
var RoFormerForQuestionAnswering = class extends RoFormerPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<QuestionAnsweringModelOutput>} An object containing the model's output logits for question answering.
   */
  async _call(model_inputs) {
    return new QuestionAnsweringModelOutput(await super._call(model_inputs));
  }
};

// src/models/rt_detr_v2/modeling_rt_detr_v2.js
var RTDetrV2PreTrainedModel = class extends PreTrainedModel {
};
var RTDetrV2Model = class extends RTDetrV2PreTrainedModel {
};
var RTDetrV2ForObjectDetection = class extends RTDetrV2PreTrainedModel {
  /**
   * @param {any} model_inputs
   */
  async _call(model_inputs) {
    return new RTDetrV2ObjectDetectionOutput(await super._call(model_inputs));
  }
};
var RTDetrV2ObjectDetectionOutput = class extends RTDetrObjectDetectionOutput {
};

// src/models/sam/modeling_sam.js
var SamImageSegmentationOutput = class extends ModelOutput {
  /**
   * @param {Object} output The output of the model.
   * @param {Tensor} output.iou_scores The output logits of the model.
   * @param {Tensor} output.pred_masks Predicted boxes.
   */
  constructor({ iou_scores, pred_masks }) {
    super();
    this.iou_scores = iou_scores;
    this.pred_masks = pred_masks;
  }
};
var SamPreTrainedModel = class extends PreTrainedModel {
};
var SamModel = class extends SamPreTrainedModel {
  /**
   * Compute image embeddings and positional image embeddings, given the pixel values of an image.
   * @param {Object} model_inputs Object containing the model inputs.
   * @param {Tensor} model_inputs.pixel_values Pixel values obtained using a `SamProcessor`.
   * @returns {Promise<{ image_embeddings: Tensor, image_positional_embeddings: Tensor }>} The image embeddings and positional image embeddings.
   */
  async get_image_embeddings({ pixel_values }) {
    return await encoder_forward(this, { pixel_values });
  }
  /**
   * @typedef {Object} SamModelInputs Object containing the model inputs.
   * @property {Tensor} pixel_values Pixel values as a Tensor with shape `(batch_size, num_channels, height, width)`.
   * These can be obtained using a `SamProcessor`.
   * @property {Tensor} [input_points] Input 2D spatial points with shape `(batch_size, num_points, 2)`.
   * This is used by the prompt encoder to encode the prompt.
   * @property {Tensor} [input_labels] Input labels for the points, as a Tensor of shape `(batch_size, point_batch_size, num_points)`.
   * This is used by the prompt encoder to encode the prompt. There are 4 types of labels:
   *  - `1`: the point is a point that contains the object of interest
   *  - `0`: the point is a point that does not contain the object of interest
   *  - `-1`: the point corresponds to the background
   *  - `-10`: the point is a padding point, thus should be ignored by the prompt encoder
   * @property {Tensor} [input_boxes] Input bounding boxes with shape `(batch_size, num_boxes, 4)`.
   * @property {Tensor} [image_embeddings] Image embeddings used by the mask decoder.
   * @property {Tensor} [image_positional_embeddings] Image positional embeddings used by the mask decoder.
   */
  /**
   * @param {SamModelInputs} model_inputs Object containing the model inputs.
   * @returns {Promise<Object>} The output of the model.
   */
  async forward(model_inputs) {
    if (!model_inputs.image_embeddings || !model_inputs.image_positional_embeddings) {
      model_inputs = {
        ...model_inputs,
        ...await this.get_image_embeddings(model_inputs)
      };
    } else {
      model_inputs = { ...model_inputs };
    }
    model_inputs.input_labels ??= ones(model_inputs.input_points.dims.slice(0, -1));
    const decoder_inputs = {
      image_embeddings: model_inputs.image_embeddings,
      image_positional_embeddings: model_inputs.image_positional_embeddings
    };
    if (model_inputs.input_points) {
      decoder_inputs.input_points = model_inputs.input_points;
    }
    if (model_inputs.input_labels) {
      decoder_inputs.input_labels = model_inputs.input_labels;
    }
    if (model_inputs.input_boxes) {
      decoder_inputs.input_boxes = model_inputs.input_boxes;
    }
    return await sessionRun(this.sessions["prompt_encoder_mask_decoder"], decoder_inputs);
  }
  /**
   * Runs the model with the provided inputs
   * @param {Object} model_inputs Model inputs
   * @returns {Promise<SamImageSegmentationOutput>} Object containing segmentation outputs
   */
  async _call(model_inputs) {
    return new SamImageSegmentationOutput(await super._call(model_inputs));
  }
};

// src/models/sam2/modeling_sam2.js
var Sam2ImageSegmentationOutput = class extends ModelOutput {
  /**
   * @param {Object} output The output of the model.
   * @param {Tensor} output.iou_scores The output logits of the model.
   * @param {Tensor} output.pred_masks Predicted boxes.
   * @param {Tensor} output.object_score_logits Logits for the object score, indicating if an object is present.
   */
  constructor({ iou_scores, pred_masks, object_score_logits }) {
    super();
    this.iou_scores = iou_scores;
    this.pred_masks = pred_masks;
    this.object_score_logits = object_score_logits;
  }
};
var Sam2PreTrainedModel = class extends PreTrainedModel {
};
var Sam2Model = class extends Sam2PreTrainedModel {
  /**
   * Compute image embeddings and positional image embeddings, given the pixel values of an image.
   * @param {Object} model_inputs Object containing the model inputs.
   * @param {Tensor} model_inputs.pixel_values Pixel values obtained using a `Sam2Processor`.
   * @returns {Promise<Record<String, Tensor>>} The image embeddings.
   */
  async get_image_embeddings({ pixel_values }) {
    return await encoder_forward(this, { pixel_values });
  }
  async forward(model_inputs) {
    const { num_feature_levels } = this.config.vision_config;
    const image_embeddings_name = Array.from({ length: num_feature_levels }, (_, i) => `image_embeddings.${i}`);
    if (image_embeddings_name.some((name) => !model_inputs[name])) {
      model_inputs = {
        ...model_inputs,
        ...await this.get_image_embeddings(model_inputs)
      };
    } else {
      model_inputs = { ...model_inputs };
    }
    if (model_inputs.input_points) {
      if (model_inputs.input_boxes && model_inputs.input_boxes.dims[1] !== 1) {
        throw new Error(
          "When both `input_points` and `input_boxes` are provided, the number of boxes per image must be 1."
        );
      }
      const shape = model_inputs.input_points.dims;
      model_inputs.input_labels ??= ones(shape.slice(0, -1));
      model_inputs.input_boxes ??= full([shape[0], 0, 4], 0);
    } else if (model_inputs.input_boxes) {
      const shape = model_inputs.input_boxes.dims;
      model_inputs.input_labels = full([shape[0], shape[1], 0], -1n);
      model_inputs.input_points = full([shape[0], 1, 0, 2], 0);
    } else {
      throw new Error("At least one of `input_points` or `input_boxes` must be provided.");
    }
    const prompt_encoder_mask_decoder_session = this.sessions["prompt_encoder_mask_decoder"];
    const decoder_inputs = pick(model_inputs, prompt_encoder_mask_decoder_session.inputNames);
    return await sessionRun(prompt_encoder_mask_decoder_session, decoder_inputs);
  }
  /**
   * Runs the model with the provided inputs
   * @param {Object} model_inputs Model inputs
   * @returns {Promise<Sam2ImageSegmentationOutput>} Object containing segmentation outputs
   */
  async _call(model_inputs) {
    return new Sam2ImageSegmentationOutput(await super._call(model_inputs));
  }
};
var EdgeTamModel = class extends Sam2Model {
};
var Sam3TrackerModel = class extends Sam2Model {
};

// src/models/sapiens/modeling_sapiens.js
var SapiensPreTrainedModel = class extends PreTrainedModel {
};
var SapiensForSemanticSegmentation = class extends SapiensPreTrainedModel {
};
var SapiensForDepthEstimation = class extends SapiensPreTrainedModel {
};
var SapiensForNormalEstimation = class extends SapiensPreTrainedModel {
};

// src/models/segformer/modeling_segformer.js
var SegformerPreTrainedModel = class extends PreTrainedModel {
};
var SegformerModel = class extends SegformerPreTrainedModel {
};
var SegformerForImageClassification = class extends SegformerPreTrainedModel {
};
var SegformerForSemanticSegmentation = class extends SegformerPreTrainedModel {
};

// src/models/siglip/modeling_siglip.js
var SiglipPreTrainedModel = class extends PreTrainedModel {
};
var SiglipModel = class extends SiglipPreTrainedModel {
};
var SiglipTextModel = class extends SiglipPreTrainedModel {
  /** @type {typeof PreTrainedModel.from_pretrained} */
  static async from_pretrained(pretrained_model_name_or_path, options = {}) {
    return super.from_pretrained(pretrained_model_name_or_path, {
      ...options,
      // Update default model file name if not provided
      model_file_name: options.model_file_name ?? "text_model"
    });
  }
};
var SiglipVisionModel = class extends CLIPPreTrainedModel {
  /** @type {typeof PreTrainedModel.from_pretrained} */
  static async from_pretrained(pretrained_model_name_or_path, options = {}) {
    return super.from_pretrained(pretrained_model_name_or_path, {
      ...options,
      // Update default model file name if not provided
      model_file_name: options.model_file_name ?? "vision_model"
    });
  }
};

// src/models/smollm3/modeling_smollm3.js
var SmolLM3PreTrainedModel = class extends PreTrainedModel {
};
var SmolLM3Model = class extends SmolLM3PreTrainedModel {
};
var SmolLM3ForCausalLM = class extends SmolLM3PreTrainedModel {
};

// src/models/smolvlm/modeling_smolvlm.js
var SmolVLMForConditionalGeneration = class extends Idefics3ForConditionalGeneration {
};

// src/models/snac/modeling_snac.js
var SnacPreTrainedModel = class extends PreTrainedModel {
  main_input_name = "input_values";
  forward_params = ["input_values"];
};
var SnacModel = class extends SnacPreTrainedModel {
  /**
   * Encodes the input audio waveform into discrete codes.
   * @param {Object} inputs Model inputs
   * @param {Tensor} [inputs.input_values] Float values of the input audio waveform, of shape `(batch_size, channels, sequence_length)`).
   * @returns {Promise<Record<string, Tensor>>} The output tensors of shape `(batch_size, num_codebooks, sequence_length)`.
   */
  async encode(inputs) {
    return await sessionRun(this.sessions["encoder_model"], inputs);
  }
  /**
   * Decodes the given frames into an output audio waveform.
   * @param {Record<string, Tensor>} inputs The encoded audio codes.
   * @returns {Promise<{audio_values: Tensor}>} The output tensor of shape `(batch_size, num_channels, sequence_length)`.
   */
  async decode(inputs) {
    return await sessionRun(this.sessions["decoder_model"], inputs);
  }
};
var SnacEncoderModel = class extends SnacPreTrainedModel {
  /** @type {typeof PreTrainedModel.from_pretrained} */
  static async from_pretrained(pretrained_model_name_or_path, options = {}) {
    return super.from_pretrained(pretrained_model_name_or_path, {
      ...options,
      // Update default model file name if not provided
      model_file_name: options.model_file_name ?? "encoder_model"
    });
  }
};
var SnacDecoderModel = class extends SnacPreTrainedModel {
  /** @type {typeof PreTrainedModel.from_pretrained} */
  static async from_pretrained(pretrained_model_name_or_path, options = {}) {
    return super.from_pretrained(pretrained_model_name_or_path, {
      ...options,
      // Update default model file name if not provided
      model_file_name: options.model_file_name ?? "decoder_model"
    });
  }
};

// src/models/solar_open/modeling_solar_open.js
var SolarOpenPreTrainedModel = class extends PreTrainedModel {
};
var SolarOpenModel = class extends SolarOpenPreTrainedModel {
};
var SolarOpenForCausalLM = class extends SolarOpenPreTrainedModel {
};

// src/models/speecht5/modeling_speecht5.js
var SpeechT5PreTrainedModel = class extends PreTrainedModel {
};
var SpeechT5Model = class extends SpeechT5PreTrainedModel {
};
var SpeechT5ForSpeechToText = class extends SpeechT5PreTrainedModel {
};
var SpeechT5ForTextToSpeech = class extends SpeechT5PreTrainedModel {
  /**
   * @typedef {Object} SpeechOutput
   * @property {Tensor} [spectrogram] The predicted log-mel spectrogram of shape
   * `(output_sequence_length, config.num_mel_bins)`. Returned when no `vocoder` is provided
   * @property {Tensor} [waveform] The predicted waveform of shape `(num_frames,)`. Returned when a `vocoder` is provided.
   * @property {Tensor} [cross_attentions] The outputs of the decoder's cross-attention layers of shape
   * `(config.decoder_layers, config.decoder_attention_heads, output_sequence_length, input_sequence_length)`. returned when `output_cross_attentions` is `true`.
   */
  /**
   * Converts a sequence of input tokens into a sequence of mel spectrograms, which are subsequently turned into a speech waveform using a vocoder.
   * @param {Tensor} input_values Indices of input sequence tokens in the vocabulary.
   * @param {Tensor} speaker_embeddings Tensor containing the speaker embeddings.
   * @param {Object} options Optional parameters for generating speech.
   * @param {number} [options.threshold=0.5] The generated sequence ends when the predicted stop token probability exceeds this value.
   * @param {number} [options.minlenratio=0.0] Used to calculate the minimum required length for the output sequence.
   * @param {number} [options.maxlenratio=20.0] Used to calculate the maximum allowed length for the output sequence.
   * @param {Object} [options.vocoder=null] The vocoder that converts the mel spectrogram into a speech waveform. If `null`, the output is the mel spectrogram.
   * @param {boolean} [options.output_cross_attentions=false] Whether or not to return the attentions tensors of the decoder's cross-attention layers.
   * @returns {Promise<SpeechOutput>} A promise which resolves to an object containing the spectrogram, waveform, and cross-attention tensors.
   */
  async generate_speech(input_values, speaker_embeddings, {
    threshold = 0.5,
    minlenratio = 0,
    maxlenratio = 20,
    vocoder = null
    // output_cross_attentions = false, // TODO add
  } = {}) {
    const model_inputs = {
      input_ids: input_values
    };
    const { encoder_outputs, encoder_attention_mask } = await encoder_forward(this, model_inputs);
    const r = encoder_outputs.dims[1] / this.config.reduction_factor;
    const maxlen = Math.floor(r * maxlenratio);
    const minlen = Math.floor(r * minlenratio);
    const num_mel_bins = this.config.num_mel_bins;
    let spectrogramParts = [];
    let past_key_values = null;
    let decoder_outputs = null;
    let idx = 0;
    while (true) {
      ++idx;
      const use_cache_branch = boolTensor(!!decoder_outputs);
      let output_sequence;
      if (decoder_outputs) {
        output_sequence = decoder_outputs.output_sequence_out;
      } else {
        output_sequence = new Tensor2("float32", new Float32Array(num_mel_bins), [1, 1, num_mel_bins]);
      }
      let decoderFeeds = {
        use_cache_branch,
        output_sequence,
        encoder_attention_mask,
        speaker_embeddings,
        encoder_hidden_states: encoder_outputs
      };
      addPastKeyValues(this, decoderFeeds, past_key_values);
      decoder_outputs = await sessionRun(this.sessions["decoder_model_merged"], decoderFeeds);
      past_key_values = getPastKeyValues(decoder_outputs, past_key_values);
      const { prob, spectrum } = decoder_outputs;
      spectrogramParts.push(spectrum);
      if (idx >= minlen && // Finished when stop token or maximum length is reached.
      (Array.from(prob.data).filter((p) => p >= threshold).length > 0 || idx >= maxlen)) {
        break;
      }
    }
    const spectrogram2 = cat(spectrogramParts);
    const { waveform } = await sessionRun(vocoder.sessions["model"], { spectrogram: spectrogram2 });
    return {
      spectrogram: spectrogram2,
      waveform
      // cross_attentions: null, // TODO add
    };
  }
};
var SpeechT5HifiGan = class extends PreTrainedModel {
  main_input_name = "spectrogram";
};

// src/models/squeezebert/modeling_squeezebert.js
var SqueezeBertPreTrainedModel = class extends PreTrainedModel {
};
var SqueezeBertModel = class extends SqueezeBertPreTrainedModel {
};
var SqueezeBertForMaskedLM = class extends SqueezeBertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<MaskedLMOutput>} returned object
   */
  async _call(model_inputs) {
    return new MaskedLMOutput(await super._call(model_inputs));
  }
};
var SqueezeBertForSequenceClassification = class extends SqueezeBertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<SequenceClassifierOutput>} returned object
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};
var SqueezeBertForQuestionAnswering = class extends SqueezeBertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<QuestionAnsweringModelOutput>} returned object
   */
  async _call(model_inputs) {
    return new QuestionAnsweringModelOutput(await super._call(model_inputs));
  }
};

// src/models/stablelm/modeling_stablelm.js
var StableLmPreTrainedModel = class extends PreTrainedModel {
};
var StableLmModel = class extends StableLmPreTrainedModel {
};
var StableLmForCausalLM = class extends StableLmPreTrainedModel {
};

// src/models/starcoder2/modeling_starcoder2.js
var Starcoder2PreTrainedModel = class extends PreTrainedModel {
};
var Starcoder2Model = class extends Starcoder2PreTrainedModel {
};
var Starcoder2ForCausalLM = class extends Starcoder2PreTrainedModel {
};

// src/models/style_text_to_speech_2/modeling_style_text_to_speech_2.js
var StyleTextToSpeech2PreTrainedModel = class extends PreTrainedModel {
};
var StyleTextToSpeech2Model = class extends StyleTextToSpeech2PreTrainedModel {
};

// src/models/supertonic/modeling_supertonic.js
var SupertonicPreTrainedModel = class extends PreTrainedModel {
};
var SupertonicForConditionalGeneration = class extends SupertonicPreTrainedModel {
  async generate_speech({
    // Required inputs
    input_ids,
    attention_mask,
    style,
    // Optional inputs
    num_inference_steps = 5,
    speed = 1.05
  }) {
    const { sampling_rate, chunk_compress_factor, base_chunk_size, latent_dim } = this.config;
    const { last_hidden_state, durations: raw_durations } = await sessionRun(this.sessions["text_encoder"], {
      input_ids,
      attention_mask,
      style
    });
    const durations = raw_durations.div(speed).mul_(sampling_rate);
    const latent_size = base_chunk_size * chunk_compress_factor;
    const durationsData = (
      /** @type {Float32Array} */
      durations.data
    );
    const latentLengths = Int32Array.from(durationsData, (d) => Math.ceil(d / latent_size));
    const maxLatentLen = Math.max(...latentLengths);
    const batch_size = input_ids.dims[0];
    const latentMaskData = new BigInt64Array(batch_size * maxLatentLen);
    for (let i = 0; i < batch_size; ++i) {
      latentMaskData.fill(1n, i * maxLatentLen, i * maxLatentLen + latentLengths[i]);
    }
    const latent_mask = new Tensor2("int64", latentMaskData, [batch_size, maxLatentLen]);
    const latentChannels = latent_dim * chunk_compress_factor;
    const latentStride = latentChannels * maxLatentLen;
    let noisy_latents = randn([batch_size, latentChannels, maxLatentLen]);
    const latentsData = (
      /** @type {Float32Array} */
      noisy_latents.data
    );
    for (let i = 0; i < batch_size; ++i) {
      if (latentLengths[i] === maxLatentLen) continue;
      for (let c = 0; c < latentChannels; ++c) {
        latentsData.fill(
          0,
          i * latentStride + c * maxLatentLen + latentLengths[i],
          i * latentStride + (c + 1) * maxLatentLen
        );
      }
    }
    const num_steps = full([batch_size], num_inference_steps);
    for (let step = 0; step < num_inference_steps; ++step) {
      const timestep = full([batch_size], step);
      ({ denoised_latents: noisy_latents } = await sessionRun(this.sessions["latent_denoiser"], {
        style,
        noisy_latents,
        latent_mask,
        encoder_outputs: last_hidden_state,
        attention_mask,
        timestep,
        num_inference_steps: num_steps
      }));
    }
    const { waveform } = await sessionRun(this.sessions["voice_decoder"], {
      latents: noisy_latents
    });
    return {
      waveform,
      durations
    };
  }
};

// src/models/swin/modeling_swin.js
var SwinPreTrainedModel = class extends PreTrainedModel {
};
var SwinModel = class extends SwinPreTrainedModel {
};
var SwinForImageClassification = class extends SwinPreTrainedModel {
  /**
   * @param {any} model_inputs
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};
var SwinForSemanticSegmentation = class extends SwinPreTrainedModel {
};

// src/models/swin2sr/modeling_swin2sr.js
var Swin2SRPreTrainedModel = class extends PreTrainedModel {
};
var Swin2SRModel = class extends Swin2SRPreTrainedModel {
};
var Swin2SRForImageSuperResolution = class extends Swin2SRPreTrainedModel {
};

// src/models/t5/modeling_t5.js
var T5PreTrainedModel = class extends PreTrainedModel {
  forward_params = [
    "input_ids",
    "attention_mask",
    "encoder_outputs",
    "decoder_input_ids",
    "decoder_attention_mask",
    "past_key_values"
  ];
};
var T5Model = class extends T5PreTrainedModel {
};
var T5ForConditionalGeneration = class extends T5PreTrainedModel {
};

// src/models/table_transformer/modeling_table_transformer.js
var TableTransformerPreTrainedModel = class extends PreTrainedModel {
};
var TableTransformerModel = class extends TableTransformerPreTrainedModel {
};
var TableTransformerForObjectDetection = class extends TableTransformerPreTrainedModel {
  /**
   * @param {any} model_inputs
   */
  async _call(model_inputs) {
    return new TableTransformerObjectDetectionOutput(await super._call(model_inputs));
  }
};
var TableTransformerObjectDetectionOutput = class extends DetrObjectDetectionOutput {
};

// src/models/trocr/modeling_trocr.js
var TrOCRPreTrainedModel = class extends PreTrainedModel {
};
var TrOCRForCausalLM = class extends TrOCRPreTrainedModel {
};

// src/models/unispeech/modeling_unispeech.js
var UniSpeechPreTrainedModel = class extends PreTrainedModel {
};
var UniSpeechModel = class extends UniSpeechPreTrainedModel {
};
var UniSpeechForCTC = class extends UniSpeechPreTrainedModel {
  /**
   * @param {Object} model_inputs
   * @param {Tensor} model_inputs.input_values Float values of input raw speech waveform.
   * @param {Tensor} model_inputs.attention_mask Mask to avoid performing convolution and attention on padding token indices. Mask values selected in [0, 1]
   */
  async _call(model_inputs) {
    return new CausalLMOutput(await super._call(model_inputs));
  }
};
var UniSpeechForSequenceClassification = class extends UniSpeechPreTrainedModel {
  /**
   * Calls the model on new inputs.
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<SequenceClassifierOutput>} An object containing the model's output logits for sequence classification.
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};

// src/models/unispeech_sat/modeling_unispeech_sat.js
var UniSpeechSatPreTrainedModel = class extends PreTrainedModel {
};
var UniSpeechSatModel = class extends UniSpeechSatPreTrainedModel {
};
var UniSpeechSatForCTC = class extends UniSpeechSatPreTrainedModel {
  /**
   * @param {Object} model_inputs
   * @param {Tensor} model_inputs.input_values Float values of input raw speech waveform.
   * @param {Tensor} model_inputs.attention_mask Mask to avoid performing convolution and attention on padding token indices. Mask values selected in [0, 1]
   */
  async _call(model_inputs) {
    return new CausalLMOutput(await super._call(model_inputs));
  }
};
var UniSpeechSatForSequenceClassification = class extends UniSpeechSatPreTrainedModel {
  /**
   * Calls the model on new inputs.
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<SequenceClassifierOutput>} An object containing the model's output logits for sequence classification.
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};
var UniSpeechSatForAudioFrameClassification = class extends UniSpeechSatPreTrainedModel {
  /**
   * Calls the model on new inputs.
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<TokenClassifierOutput>} An object containing the model's output logits for sequence classification.
   */
  async _call(model_inputs) {
    return new TokenClassifierOutput(await super._call(model_inputs));
  }
};

// src/models/vaultgemma/modeling_vaultgemma.js
var VaultGemmaPreTrainedModel = class extends PreTrainedModel {
};
var VaultGemmaModel = class extends VaultGemmaPreTrainedModel {
};
var VaultGemmaForCausalLM = class extends VaultGemmaPreTrainedModel {
};

// src/models/vision_encoder_decoder/modeling_vision_encoder_decoder.js
var VisionEncoderDecoderModel = class extends PreTrainedModel {
  main_input_name = "pixel_values";
  forward_params = [
    // Encoder inputs
    "pixel_values",
    // Decoder inpputs
    "decoder_input_ids",
    "encoder_hidden_states",
    "past_key_values"
  ];
};

// src/models/vit/modeling_vit.js
var ViTPreTrainedModel = class extends PreTrainedModel {
};
var ViTModel = class extends ViTPreTrainedModel {
};
var ViTForImageClassification = class extends ViTPreTrainedModel {
  /**
   * @param {any} model_inputs
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};

// src/models/vit_mae/modeling_vit_mae.js
var ViTMAEPreTrainedModel = class extends PreTrainedModel {
};
var ViTMAEModel = class extends ViTMAEPreTrainedModel {
};

// src/models/vit_msn/modeling_vit_msn.js
var ViTMSNPreTrainedModel = class extends PreTrainedModel {
};
var ViTMSNModel = class extends ViTMSNPreTrainedModel {
};
var ViTMSNForImageClassification = class extends ViTMSNPreTrainedModel {
  /**
   * @param {any} model_inputs
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};

// src/models/vitmatte/modeling_vitmatte.js
var VitMattePreTrainedModel = class extends PreTrainedModel {
};
var VitMatteForImageMatting = class extends VitMattePreTrainedModel {
  /**
   * @param {any} model_inputs
   */
  async _call(model_inputs) {
    return new ImageMattingOutput(await super._call(model_inputs));
  }
};

// src/models/vitpose/modeling_vitpose.js
var VitPosePreTrainedModel = class extends PreTrainedModel {
};
var VitPoseForPoseEstimation = class extends VitPosePreTrainedModel {
};

// src/models/vits/modeling_vits.js
var VitsModelOutput = class extends ModelOutput {
  /**
   * @typedef {import('../../utils/tensor.js').Tensor} Tensor
   */
  /**
   * @param {Object} output The output of the model.
   * @param {Tensor} output.waveform The final audio waveform predicted by the model, of shape `(batch_size, sequence_length)`.
   * @param {Tensor} output.spectrogram The log-mel spectrogram predicted at the output of the flow model.
   * This spectrogram is passed to the Hi-Fi GAN decoder model to obtain the final audio waveform.
   */
  constructor({ waveform, spectrogram: spectrogram2 }) {
    super();
    this.waveform = waveform;
    this.spectrogram = spectrogram2;
  }
};
var VitsPreTrainedModel = class extends PreTrainedModel {
};
var VitsModel = class extends VitsPreTrainedModel {
  /**
   * Calls the model on new inputs.
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<VitsModelOutput>} The outputs for the VITS model.
   */
  async _call(model_inputs) {
    return new VitsModelOutput(await super._call(model_inputs));
  }
};

// src/models/voxtral/modeling_voxtral.js
var VoxtralForConditionalGeneration = class extends UltravoxModel {
};

// src/models/voxtral_realtime/modeling_voxtral_realtime.js
var CONV1_LEFT_PAD = 2;
var CONV2_LEFT_PAD = 1;
var states = /* @__PURE__ */ new WeakMap();
function createEncoderState(model, input_features) {
  const { text_config, audio_config } = (
    /** @type {any} */
    model.config
  );
  const encoder_session = model.sessions["audio_encoder"];
  const { num_mel_bins, hidden_size: enc_hidden_size } = audio_config;
  const PADDING_CACHE_CHANNELS = num_mel_bins + enc_hidden_size;
  const enc_kv_cache = new DynamicCache();
  const enc_names = getCacheNames(audio_config);
  const enc_symbols = { batch_size: 1 };
  let padding_type = "float32";
  for (const meta of encoder_session.inputMetadata) {
    if (meta.name === "past_padding_cache") {
      padding_type = meta.type;
      continue;
    }
    if (!enc_names.has(meta.name)) continue;
    const shape = resolveCacheShape(meta.shape, enc_symbols);
    const size = shape.reduce((a, b) => a * b, 1);
    const cls = DataTypeMap[meta.type];
    enc_kv_cache[meta.name] = new Tensor2(meta.type, new cls(size), shape);
  }
  const padding_cls = DataTypeMap[padding_type];
  const enc_padding_cache = new Tensor2(
    padding_type,
    new padding_cls(PADDING_CACHE_CHANNELS * CONV1_LEFT_PAD),
    [1, PADDING_CACHE_CHANNELS, CONV1_LEFT_PAD]
  );
  const chunks_iter = input_features[Symbol.asyncIterator]?.() ?? input_features[Symbol.iterator]?.();
  if (!chunks_iter) {
    throw new Error("input_features must be iterable or async iterable");
  }
  return {
    encoder_session,
    enc_kv_cache,
    enc_padding_cache,
    enc_past_seq_len: 0,
    audio_embed_queue: [],
    audio_embed_total_tokens: 0,
    audio_queue_offset: 0,
    audio_consumed: 0,
    stream_exhausted: false,
    chunks_iter,
    text_hidden_size: text_config.hidden_size
  };
}
async function encodeChunk(s, chunk_features) {
  const audio_seq_len = chunk_features.dims[2];
  const conv2_output_len = Math.floor((CONV2_LEFT_PAD + audio_seq_len - 3) / 2) + 1;
  const position_ids = new Tensor2(
    "int64",
    BigInt64Array.from({ length: conv2_output_len }, (_, i) => BigInt(s.enc_past_seq_len + i)),
    [1, conv2_output_len]
  );
  const total_seq_len = s.enc_past_seq_len + conv2_output_len;
  const attention_mask = ones([1, total_seq_len]);
  const { audio_embeds, present_padding_cache, ...present_cache } = await sessionRun(s.encoder_session, {
    input_features: chunk_features,
    attention_mask,
    position_ids,
    past_padding_cache: s.enc_padding_cache,
    ...s.enc_kv_cache
  });
  if (s.enc_padding_cache.location === "gpu-buffer") {
    s.enc_padding_cache.dispose();
  }
  s.enc_padding_cache = present_padding_cache;
  for (const name in present_cache) {
    if (name.startsWith("present.")) {
      const pastName = name.replace("present", "past_key_values");
      const prev = s.enc_kv_cache[pastName];
      if (prev?.location === "gpu-buffer") {
        prev.dispose();
      }
      s.enc_kv_cache[pastName] = present_cache[name];
    }
  }
  s.enc_past_seq_len = total_seq_len;
  return audio_embeds;
}
async function fillAudioBuffer(s, needed) {
  while (s.audio_embed_total_tokens < needed && !s.stream_exhausted) {
    const result = await s.chunks_iter.next();
    if (result.done) {
      s.stream_exhausted = true;
      break;
    }
    const new_embeds = await encodeChunk(s, result.value);
    s.audio_embed_queue.push({ data: new_embeds.data, tokens: new_embeds.dims[1] });
    s.audio_embed_total_tokens += new_embeds.dims[1];
  }
}
function addAudioEmbeddings(s, inputs_embeds, current_len) {
  if (s.audio_embed_queue.length === 0) return;
  const embed_data = inputs_embeds.data;
  let embed_write_pos = 0;
  let remaining = current_len;
  while (remaining > 0 && s.audio_embed_queue.length > 0) {
    const front = s.audio_embed_queue[0];
    const available = front.tokens - s.audio_queue_offset;
    const n = Math.min(remaining, available);
    const src_offset = s.audio_queue_offset * s.text_hidden_size;
    for (let i = 0; i < n * s.text_hidden_size; ++i) {
      embed_data[embed_write_pos * s.text_hidden_size + i] += front.data[src_offset + i];
    }
    embed_write_pos += n;
    remaining -= n;
    s.audio_queue_offset += n;
    if (s.audio_queue_offset >= front.tokens) {
      s.audio_embed_queue.shift();
      s.audio_queue_offset = 0;
    }
  }
  s.audio_consumed += current_len - remaining;
}
var AudioExhaustedCriteria = class extends StoppingCriteria {
  constructor(enc_state) {
    super();
    this._s = enc_state;
  }
  _call(input_ids) {
    const done = this._s.stream_exhausted && this._s.audio_embed_queue.length === 0;
    return input_ids.map(() => done);
  }
};
var VoxtralRealtimePreTrainedModel = class extends PreTrainedModel {
  forward_params = ["input_ids", "attention_mask", "position_ids", "past_key_values"];
};
var VoxtralRealtimeForConditionalGeneration = class extends VoxtralRealtimePreTrainedModel {
  async forward({ input_ids, past_key_values, ...kwargs }) {
    const current_len = input_ids.dims[1];
    const enc = states.get(this);
    if (enc) {
      await fillAudioBuffer(enc, enc.audio_consumed + current_len);
    }
    const { inputs_embeds } = await sessionRun(this.sessions["embed_tokens"], { input_ids });
    if (enc) {
      addAudioEmbeddings(enc, inputs_embeds, current_len);
    }
    const decoder_feeds = { inputs_embeds, ...kwargs };
    addPastKeyValues(this, decoder_feeds, past_key_values);
    const session = this.sessions["decoder_model_merged"];
    const fixed = pick(decoder_feeds, session.inputNames);
    return await sessionRun(session, fixed);
  }
  async generate({ input_features, stopping_criteria: userStoppingCriteria, ...kwargs }) {
    if (!input_features) {
      throw new Error("input_features (generator/iterable) must be provided");
    }
    const enc_state = createEncoderState(this, input_features);
    states.set(this, enc_state);
    const stopping_criteria = new StoppingCriteriaList();
    stopping_criteria.push(new AudioExhaustedCriteria(enc_state));
    if (userStoppingCriteria) stopping_criteria.extend(userStoppingCriteria);
    try {
      return await super.generate({ ...kwargs, stopping_criteria });
    } finally {
      enc_state.enc_kv_cache.dispose();
      states.delete(this);
    }
  }
};

// src/models/wav2vec2_bert/modeling_wav2vec2_bert.js
var Wav2Vec2BertPreTrainedModel = class extends PreTrainedModel {
};
var Wav2Vec2BertModel = class extends Wav2Vec2BertPreTrainedModel {
};
var Wav2Vec2BertForCTC = class extends Wav2Vec2BertPreTrainedModel {
  /**
   * @param {Object} model_inputs
   * @param {Tensor} model_inputs.input_features Float values of input mel-spectrogram.
   * @param {Tensor} model_inputs.attention_mask Mask to avoid performing convolution and attention on padding token indices. Mask values selected in [0, 1]
   */
  async _call(model_inputs) {
    return new CausalLMOutput(await super._call(model_inputs));
  }
};
var Wav2Vec2BertForSequenceClassification = class extends Wav2Vec2BertPreTrainedModel {
  /**
   * Calls the model on new inputs.
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<SequenceClassifierOutput>} An object containing the model's output logits for sequence classification.
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};

// src/models/wavlm/modeling_wavlm.js
var XVectorOutput = class extends ModelOutput {
  /**
   * @param {Object} output The output of the model.
   * @param {Tensor} output.logits Classification hidden states before AMSoftmax, of shape `(batch_size, config.xvector_output_dim)`.
   * @param {Tensor} output.embeddings Utterance embeddings used for vector similarity-based retrieval, of shape `(batch_size, config.xvector_output_dim)`.
   */
  constructor({ logits, embeddings }) {
    super();
    this.logits = logits;
    this.embeddings = embeddings;
  }
};
var WavLMPreTrainedModel = class extends PreTrainedModel {
};
var WavLMModel = class extends WavLMPreTrainedModel {
};
var WavLMForCTC = class extends WavLMPreTrainedModel {
  /**
   * @param {Object} model_inputs
   * @param {Tensor} model_inputs.input_values Float values of input raw speech waveform.
   * @param {Tensor} model_inputs.attention_mask Mask to avoid performing convolution and attention on padding token indices. Mask values selected in [0, 1]
   */
  async _call(model_inputs) {
    return new CausalLMOutput(await super._call(model_inputs));
  }
};
var WavLMForSequenceClassification = class extends WavLMPreTrainedModel {
  /**
   * Calls the model on new inputs.
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<SequenceClassifierOutput>} An object containing the model's output logits for sequence classification.
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};
var WavLMForXVector = class extends WavLMPreTrainedModel {
  /**
   * Calls the model on new inputs.
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<XVectorOutput>} An object containing the model's output logits and speaker embeddings.
   */
  async _call(model_inputs) {
    return new XVectorOutput(await super._call(model_inputs));
  }
};
var WavLMForAudioFrameClassification = class extends WavLMPreTrainedModel {
  /**
   * Calls the model on new inputs.
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<TokenClassifierOutput>} An object containing the model's output logits for sequence classification.
   */
  async _call(model_inputs) {
    return new TokenClassifierOutput(await super._call(model_inputs));
  }
};

// src/models/wespeaker_resnet/modeling_wespeaker_resnet.js
var WeSpeakerResNetPreTrainedModel = class extends PreTrainedModel {
};
var WeSpeakerResNetModel = class extends WeSpeakerResNetPreTrainedModel {
};

// src/models/whisper/generation_whisper.js
var WhisperGenerationConfig = class extends GenerationConfig {
  /**
   * Whether to return the timestamps with the text. This enables the `WhisperTimestampsLogitsProcessor`.
   * @type {boolean}
   */
  return_timestamps = null;
  /**
   * Whether to return token-level timestamps
   * with the text. This can be used with or without the `return_timestamps` option. To get word-level
   * timestamps, use the tokenizer to group the tokens into words.
   * @type {boolean}
   */
  return_token_timestamps = null;
  /**
   * The number of audio frames available in this chunk. This is only used generating word-level timestamps.
   * @type {number}
   */
  num_frames = null;
  /**
   * Alignment heads to predict word-level timestamps. This is a list of [layer, head] pairs that
   * select the cross-attention heads that are highly correlated to word-level timing.
   * @type {[number, number][]}
   */
  alignment_heads = null;
  /**
   * Task to use for generation, either "translate" or "transcribe".
   * @type {string}
   */
  task = null;
  /**
   * Language token to use for generation, can be either in the form of `<|en|>`, `en` or `english`.
   * You can find all the possible language tokens in the `model.generation_config.lang_to_id` dictionary.
   * @type {string}
   */
  language = null;
  /**
   * The id of the `"<|notimestamps|>"` token.
   * @type {number}
   */
  no_timestamps_token_id = null;
  /**
   * Rank-1 list of token IDs created by passing text to [`~WhisperProcessor.get_prompt_ids`] that is
   * provided as a prompt to each chunk. This can be used to provide or "prompt-engineer" a context for
   * transcription, e.g. custom vocabularies or proper nouns to make it more likely to predict those words
   * correctly. It cannot be used in conjunction with `decoder_start_token_id` as it overwrites this value.
   * @type {number[]}
   */
  prompt_ids = null;
  /**
   * Whether the model is multilingual or not.
   * @type {boolean}
   */
  is_multilingual = null;
  /**
   * (Optional) A mapping from language tokens to their corresponding IDs.
   * Only required if the model is multilingual.
   * @type {Record<string, number>|null}
   */
  lang_to_id = null;
  /**
   * (Optional) A mapping from task tokens to their corresponding IDs.
   * @type {Record<string, number>|null}
   */
  task_to_id = null;
  /**
   * Used to set the maximum value of the initial timestamp. This is used to prevent the model from
   * predicting timestamps that are too far in the future.
   * @type {number}
   */
  max_initial_timestamp_index = 1;
};

// src/models/whisper/modeling_whisper.js
var WhisperPreTrainedModel = class extends PreTrainedModel {
  requires_attention_mask = false;
  main_input_name = "input_features";
  forward_params = [
    "input_features",
    "attention_mask",
    "decoder_input_ids",
    "decoder_attention_mask",
    "past_key_values"
  ];
};
var WhisperModel = class extends WhisperPreTrainedModel {
};
var WhisperForConditionalGeneration = class extends WhisperPreTrainedModel {
  _prepare_generation_config(generation_config, kwargs) {
    return (
      /** @type {WhisperGenerationConfig} */
      super._prepare_generation_config(generation_config, kwargs, WhisperGenerationConfig)
    );
  }
  /**
   *
   * @param {WhisperGenerationConfig} generation_config
   */
  _retrieve_init_tokens(generation_config) {
    const init_tokens = [generation_config.decoder_start_token_id];
    let language = generation_config.language;
    const task = generation_config.task;
    if (generation_config.is_multilingual) {
      if (!language) {
        logger.warn("No language specified - defaulting to English (en).");
        language = "en";
      }
      const language_code = whisper_language_to_code(language);
      const language_token = `<|${language_code}|>`;
      init_tokens.push(generation_config.lang_to_id[language_token]);
      init_tokens.push(generation_config.task_to_id[task ?? "transcribe"]);
    } else if (language || task) {
      throw new Error(
        "Cannot specify `task` or `language` for an English-only model. If the model is intended to be multilingual, pass `is_multilingual=true` to generate, or update the generation config."
      );
    }
    if (!generation_config.return_timestamps && generation_config.no_timestamps_token_id && init_tokens.at(-1) !== generation_config.no_timestamps_token_id) {
      init_tokens.push(generation_config.no_timestamps_token_id);
    } else if (generation_config.return_timestamps && init_tokens.at(-1) === generation_config.no_timestamps_token_id) {
      logger.warn(
        "<|notimestamps|> prompt token is removed from generation_config since `return_timestamps` is set to `true`."
      );
      init_tokens.pop();
    }
    return init_tokens.filter((token) => token != null);
  }
  /**
   * Transcribes or translates log-mel input features to a sequence of auto-regressively generated token ids.
   * @param {import('./generation_whisper.js').WhisperGenerationFunctionParameters} options
   * @returns {Promise<ModelOutput|Tensor>} The output of the model, which can contain the generated token ids, attentions, and scores.
   */
  async generate({
    inputs = null,
    generation_config = null,
    logits_processor = null,
    stopping_criteria = null,
    // Whisper-specific options (passed to kwargs)
    // prompt_ids = null,
    // language = null,
    // task = null,
    ...kwargs
  }) {
    generation_config = this._prepare_generation_config(generation_config, kwargs);
    const init_tokens = kwargs.decoder_input_ids instanceof Tensor2 ? prepareTensorForDecode(kwargs.decoder_input_ids) : kwargs.decoder_input_ids ?? this._retrieve_init_tokens(generation_config);
    if (generation_config.return_timestamps) {
      logits_processor ??= new LogitsProcessorList();
      logits_processor.push(new WhisperTimeStampLogitsProcessor(generation_config, init_tokens));
    }
    if (generation_config.begin_suppress_tokens) {
      logits_processor ??= new LogitsProcessorList();
      logits_processor.push(
        new SuppressTokensAtBeginLogitsProcessor(generation_config.begin_suppress_tokens, init_tokens.length)
      );
    }
    if (generation_config.return_token_timestamps) {
      if (!generation_config.alignment_heads) {
        throw new Error(
          "Model generation config has no `alignment_heads`, token-level timestamps not available. See https://gist.github.com/hollance/42e32852f24243b748ae6bc1f985b13a on how to add this property to the generation config."
        );
      }
      if (generation_config.task === "translate") {
        logger.warn("Token-level timestamps may not be reliable for task 'translate'.");
      }
      generation_config.output_attentions = true;
      generation_config.return_dict_in_generate = true;
    }
    if (generation_config.return_timestamps && !kwargs.max_new_tokens) {
      return this._generate_with_seek({
        inputs,
        generation_config,
        logits_processor,
        init_tokens,
        kwargs
      });
    }
    const outputs = await super.generate({
      inputs,
      generation_config,
      logits_processor,
      decoder_input_ids: init_tokens,
      ...kwargs
    });
    if (generation_config.return_token_timestamps) {
      outputs["token_timestamps"] = this._extract_token_timestamps(
        // @ts-expect-error TS2345
        outputs,
        generation_config.alignment_heads,
        generation_config.num_frames,
        0.02,
        init_tokens.length
      );
    }
    return outputs;
  }
  /**
   * Generates with a seek loop for timestamp mode, re-encoding and generating
   * for each segment until all audio frames are consumed.
   * This matches Python's WhisperForConditionalGeneration.generate() behavior.
   * @private
   */
  async _generate_with_seek({ inputs, generation_config, logits_processor, init_tokens, kwargs }) {
    const timestamp_begin = generation_config.no_timestamps_token_id + 1;
    const eos_token_id = Array.isArray(generation_config.eos_token_id) ? generation_config.eos_token_id[0] : generation_config.eos_token_id;
    const return_token_timestamps = generation_config.return_token_timestamps;
    const input_features = inputs;
    const total_frames = input_features.dims[2];
    const input_stride = 2;
    const max_source_positions = (
      /** @type {number} */
      this.config.max_source_positions
    );
    const num_segment_frames = input_stride * max_source_positions;
    let seek = 0;
    const allTokens = [];
    const allTokenTimestamps = [];
    while (seek < total_frames) {
      const seek_end = Math.min(seek + num_segment_frames, total_frames);
      const segment_input = input_features.slice(null, null, [seek, seek_end]);
      let segment_features;
      const segment_frames = segment_input.dims[2];
      if (segment_frames < num_segment_frames) {
        const n_mels = input_features.dims[1];
        const padded_data = new Float32Array(n_mels * num_segment_frames);
        const src = (
          /** @type {Float32Array} */
          segment_input.data
        );
        for (let m = 0; m < n_mels; ++m) {
          padded_data.set(src.subarray(m * segment_frames, (m + 1) * segment_frames), m * num_segment_frames);
        }
        segment_features = new Tensor2("float32", padded_data, [1, n_mels, num_segment_frames]);
      } else {
        segment_features = segment_input;
      }
      if (logits_processor) {
        for (const proc of logits_processor) {
          if ("begin_index" in proc) {
            proc.begin_index = init_tokens.length;
          }
        }
      }
      const outputs = (
        /** @type {any} */
        await super.generate({
          inputs: segment_features,
          generation_config,
          logits_processor,
          decoder_input_ids: init_tokens,
          ...kwargs
        })
      );
      const raw_sequence = return_token_timestamps ? outputs.sequences : (
        /** @type {Tensor} */
        outputs
      );
      const generated_tokens = raw_sequence[0].tolist().map(Number).slice(init_tokens.length);
      let seek_token_timestamps;
      if (return_token_timestamps) {
        outputs["token_timestamps"] = this._extract_token_timestamps(
          outputs,
          generation_config.alignment_heads,
          Math.floor((seek_end - seek) / input_stride),
          0.02,
          init_tokens.length
        );
        const time_offset = seek / input_stride * 0.02;
        seek_token_timestamps = outputs.token_timestamps[0].tolist().slice(init_tokens.length).map((t) => t + time_offset);
      }
      if (generated_tokens.length > 0 && generated_tokens.at(-1) === eos_token_id) {
        generated_tokens.pop();
      }
      if (generated_tokens.length === 0) {
        break;
      }
      const is_timestamp = generated_tokens.map((t) => t >= timestamp_begin);
      const single_timestamp_ending = generated_tokens.length >= 2 && is_timestamp[generated_tokens.length - 1] && !is_timestamp[generated_tokens.length - 2];
      const segment_boundary_indices = [];
      for (let i = 0; i < generated_tokens.length - 1; ++i) {
        if (is_timestamp[i] && is_timestamp[i + 1]) {
          segment_boundary_indices.push(i + 1);
        }
      }
      let segment_offset;
      let tokens_to_keep = generated_tokens.length;
      if (segment_boundary_indices.length > 0) {
        if (single_timestamp_ending) {
          segment_offset = seek_end - seek;
        } else {
          const last_boundary = segment_boundary_indices.at(-1);
          const last_ts_pos = generated_tokens[last_boundary - 1] - timestamp_begin;
          segment_offset = last_ts_pos * input_stride;
          tokens_to_keep = last_boundary;
        }
      } else {
        segment_offset = seek_end - seek;
      }
      const timestamp_offset = Math.floor(seek / input_stride);
      const max_timestamp_token = timestamp_begin + 1500;
      for (let i = 0; i < tokens_to_keep; ++i) {
        if (generated_tokens[i] >= timestamp_begin) {
          generated_tokens[i] = Math.min(generated_tokens[i] + timestamp_offset, max_timestamp_token);
        }
      }
      allTokens.push(...generated_tokens.slice(0, tokens_to_keep));
      if (seek_token_timestamps) {
        allTokenTimestamps.push(...seek_token_timestamps.slice(0, tokens_to_keep));
      }
      seek += segment_offset;
    }
    allTokens.push(eos_token_id);
    const full_sequence = [...init_tokens, ...allTokens];
    if (return_token_timestamps) {
      const sequences = new Tensor2("int64", full_sequence.map(BigInt), [1, full_sequence.length]);
      const full_timestamps = [...new Array(init_tokens.length).fill(0), ...allTokenTimestamps, 0];
      const token_timestamps = new Tensor2("float32", new Float32Array(full_timestamps), [
        1,
        full_timestamps.length
      ]);
      return { sequences, token_timestamps };
    }
    return new Tensor2("int64", full_sequence.map(BigInt), [1, full_sequence.length]);
  }
  /**
   * Calculates token-level timestamps using the encoder-decoder cross-attentions and
   * dynamic time-warping (DTW) to map each output token to a position in the input audio.
   * If `num_frames` is specified, the encoder-decoder cross-attentions will be cropped before applying DTW.
   * @param {Object} generate_outputs Outputs generated by the model
   * @param {Tensor[][]} generate_outputs.cross_attentions The cross attentions output by the model
   * @param {Tensor} generate_outputs.sequences The sequences output by the model
   * @param {number[][]} alignment_heads Alignment heads of the model
   * @param {number} [num_frames=null] Number of frames in the input audio.
   * @param {number} [time_precision=0.02] Precision of the timestamps in seconds
   * @param {number} [num_input_ids=0] Number of decoder input ids (prefix tokens) to skip in DTW
   * @returns {Tensor} tensor containing the timestamps in seconds for each predicted token
   */
  _extract_token_timestamps(generate_outputs, alignment_heads, num_frames = null, time_precision = 0.02, num_input_ids = 0) {
    if (!generate_outputs.cross_attentions) {
      throw new Error(
        "Model outputs must contain cross attentions to extract timestamps. This is most likely because the model was not exported with `output_attentions=True`."
      );
    }
    if (num_frames == null) {
      logger.warn(
        "`num_frames` has not been set, meaning the entire audio will be analyzed. This may lead to inaccurate token-level timestamps for short audios (< 30 seconds)."
      );
    }
    let median_filter_width = this.config.median_filter_width;
    if (median_filter_width === void 0) {
      logger.warn("Model config has no `median_filter_width`, using default value of 7.");
      median_filter_width = 7;
    }
    const batch = generate_outputs.cross_attentions;
    const cross_attentions = Array.from(
      // @ts-expect-error TS2339
      { length: this.config.decoder_layers },
      // Concatenate the cross attentions for each layer across sequence length dimension.
      (_, i) => cat(
        batch.map((x) => x[i]),
        2
      )
    );
    const weights = stack(
      alignment_heads.map(([l, h]) => {
        if (l >= cross_attentions.length) {
          throw new Error(
            `Layer index ${l} is out of bounds for cross attentions (length ${cross_attentions.length}).`
          );
        }
        return num_frames ? cross_attentions[l].slice(null, h, null, [0, num_frames]) : cross_attentions[l].slice(null, h);
      })
    ).transpose(1, 0, 2, 3);
    const [std, calculatedMean] = std_mean(weights, -2, 0, true);
    const smoothedWeights = weights.clone();
    for (let a = 0; a < smoothedWeights.dims[0]; ++a) {
      const aTensor = smoothedWeights[a];
      for (let b = 0; b < aTensor.dims[0]; ++b) {
        const bTensor = aTensor[b];
        const stdTensorData = std[a][b][0].data;
        const meanTensorData = calculatedMean[a][b][0].data;
        for (let c = 0; c < bTensor.dims[0]; ++c) {
          let cTensorData = bTensor[c].data;
          for (let d = 0; d < cTensorData.length; ++d) {
            cTensorData[d] = (cTensorData[d] - meanTensorData[d]) / stdTensorData[d];
          }
          cTensorData.set(medianFilter(cTensorData, median_filter_width));
        }
      }
    }
    const croppedWeights = num_input_ids > 0 ? smoothedWeights.slice(null, null, [num_input_ids, smoothedWeights.dims[2]], null) : smoothedWeights;
    const batchedMatrices = [mean(croppedWeights, 1)];
    const timestampsShape = generate_outputs.sequences.dims;
    const timestamps = new Tensor2(
      "float32",
      new Float32Array(timestampsShape[0] * timestampsShape[1]),
      timestampsShape
    );
    for (let batch_idx = 0; batch_idx < timestampsShape[0]; ++batch_idx) {
      const matrix = batchedMatrices[batch_idx].neg().squeeze_(0);
      const [text_indices, time_indices] = dynamic_time_warping(matrix.tolist());
      const diffs = Array.from(
        { length: text_indices.length - 1 },
        (v, i) => text_indices[i + 1] - text_indices[i]
      );
      const jumps = mergeArrays([1], diffs).map((x) => !!x);
      const jump_times = [];
      for (let i = 0; i < jumps.length; ++i) {
        if (jumps[i]) {
          jump_times.push(time_indices[i] * time_precision);
        }
      }
      const padded = new Array(num_input_ids).fill(0);
      padded.push(...jump_times);
      if (jump_times.length > 0) {
        padded.push(jump_times.at(-1));
      }
      timestamps[batch_idx].data.set(padded);
    }
    return timestamps;
  }
};
var LiteWhisperForConditionalGeneration = class extends WhisperForConditionalGeneration {
};

// src/models/xlm/modeling_xlm.js
var XLMPreTrainedModel = class extends PreTrainedModel {
};
var XLMModel = class extends XLMPreTrainedModel {
};
var XLMWithLMHeadModel = class extends XLMPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<MaskedLMOutput>} returned object
   */
  async _call(model_inputs) {
    return new MaskedLMOutput(await super._call(model_inputs));
  }
};
var XLMForSequenceClassification = class extends XLMPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<SequenceClassifierOutput>} returned object
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};
var XLMForTokenClassification = class extends XLMPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<TokenClassifierOutput>} An object containing the model's output logits for token classification.
   */
  async _call(model_inputs) {
    return new TokenClassifierOutput(await super._call(model_inputs));
  }
};
var XLMForQuestionAnswering = class extends XLMPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<QuestionAnsweringModelOutput>} returned object
   */
  async _call(model_inputs) {
    return new QuestionAnsweringModelOutput(await super._call(model_inputs));
  }
};

// src/models/xlm_roberta/modeling_xlm_roberta.js
var XLMRobertaPreTrainedModel = class extends PreTrainedModel {
};
var XLMRobertaModel = class extends XLMRobertaPreTrainedModel {
};
var XLMRobertaForMaskedLM = class extends XLMRobertaPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<MaskedLMOutput>} returned object
   */
  async _call(model_inputs) {
    return new MaskedLMOutput(await super._call(model_inputs));
  }
};
var XLMRobertaForSequenceClassification = class extends XLMRobertaPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<SequenceClassifierOutput>} returned object
   */
  async _call(model_inputs) {
    return new SequenceClassifierOutput(await super._call(model_inputs));
  }
};
var XLMRobertaForTokenClassification = class extends XLMRobertaPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<TokenClassifierOutput>} An object containing the model's output logits for token classification.
   */
  async _call(model_inputs) {
    return new TokenClassifierOutput(await super._call(model_inputs));
  }
};
var XLMRobertaForQuestionAnswering = class extends XLMRobertaPreTrainedModel {
  /**
   * Calls the model on new inputs.
   *
   * @param {Object} model_inputs The inputs to the model.
   * @returns {Promise<QuestionAnsweringModelOutput>} returned object
   */
  async _call(model_inputs) {
    return new QuestionAnsweringModelOutput(await super._call(model_inputs));
  }
};

// src/models/yolos/modeling_yolos.js
var YolosPreTrainedModel = class extends PreTrainedModel {
};
var YolosModel = class extends YolosPreTrainedModel {
};
var YolosForObjectDetection = class extends YolosPreTrainedModel {
  /**
   * @param {any} model_inputs
   */
  async _call(model_inputs) {
    return new YolosObjectDetectionOutput(await super._call(model_inputs));
  }
};
var YolosObjectDetectionOutput = class extends ModelOutput {
  /**
   * @param {Object} output The output of the model.
   * @param {Tensor} output.logits Classification logits (including no-object) for all queries.
   * @param {Tensor} output.pred_boxes Normalized boxes coordinates for all queries, represented as (center_x, center_y, width, height).
   * These values are normalized in [0, 1], relative to the size of each individual image in the batch (disregarding possible padding).
   */
  constructor({ logits, pred_boxes }) {
    super();
    this.logits = logits;
    this.pred_boxes = pred_boxes;
  }
};

// src/models/youtu/modeling_youtu.js
var YoutuPreTrainedModel = class extends PreTrainedModel {
};
var YoutuModel = class extends YoutuPreTrainedModel {
};
var YoutuForCausalLM = class extends YoutuPreTrainedModel {
};

// src/models/registry.js
var MODEL_MAPPING_NAMES_ENCODER_ONLY = /* @__PURE__ */ new Map([
  ["bert", "BertModel"],
  ["eurobert", "EuroBertModel"],
  ["neobert", "NeoBertModel"],
  ["modernbert", "ModernBertModel"],
  ["nomic_bert", "NomicBertModel"],
  ["roformer", "RoFormerModel"],
  ["electra", "ElectraModel"],
  ["esm", "EsmModel"],
  ["convbert", "ConvBertModel"],
  ["camembert", "CamembertModel"],
  ["deberta", "DebertaModel"],
  ["deberta-v2", "DebertaV2Model"],
  ["mpnet", "MPNetModel"],
  ["albert", "AlbertModel"],
  ["distilbert", "DistilBertModel"],
  ["roberta", "RobertaModel"],
  ["xlm", "XLMModel"],
  ["xlm-roberta", "XLMRobertaModel"],
  ["clap", "ClapModel"],
  ["clip", "CLIPModel"],
  ["clipseg", "CLIPSegModel"],
  ["chinese_clip", "ChineseCLIPModel"],
  ["siglip", "SiglipModel"],
  ["jina_clip", "JinaCLIPModel"],
  ["mobilebert", "MobileBertModel"],
  ["squeezebert", "SqueezeBertModel"],
  ["wav2vec2", "Wav2Vec2Model"],
  ["wav2vec2-bert", "Wav2Vec2BertModel"],
  ["unispeech", "UniSpeechModel"],
  ["unispeech-sat", "UniSpeechSatModel"],
  ["hubert", "HubertModel"],
  ["wavlm", "WavLMModel"],
  ["audio-spectrogram-transformer", "ASTModel"],
  ["vits", "VitsModel"],
  ["pyannote", "PyAnnoteModel"],
  ["wespeaker-resnet", "WeSpeakerResNetModel"],
  ["detr", "DetrModel"],
  ["rt_detr", "RTDetrModel"],
  ["rt_detr_v2", "RTDetrV2Model"],
  ["rf_detr", "RFDetrModel"],
  ["d_fine", "DFineModel"],
  ["table-transformer", "TableTransformerModel"],
  ["vit", "ViTModel"],
  ["ijepa", "IJepaModel"],
  ["pvt", "PvtModel"],
  ["vit_msn", "ViTMSNModel"],
  ["vit_mae", "ViTMAEModel"],
  ["groupvit", "GroupViTModel"],
  ["fastvit", "FastViTModel"],
  ["mobilevit", "MobileViTModel"],
  ["mobilevitv2", "MobileViTV2Model"],
  ["owlvit", "OwlViTModel"],
  ["owlv2", "Owlv2Model"],
  ["beit", "BeitModel"],
  ["deit", "DeiTModel"],
  ["hiera", "HieraModel"],
  ["convnext", "ConvNextModel"],
  ["convnextv2", "ConvNextV2Model"],
  ["dinov2", "Dinov2Model"],
  ["dinov2_with_registers", "Dinov2WithRegistersModel"],
  ["dinov3_vit", "DINOv3ViTModel"],
  ["dinov3_convnext", "DINOv3ConvNextModel"],
  ["resnet", "ResNetModel"],
  ["swin", "SwinModel"],
  ["swin2sr", "Swin2SRModel"],
  ["donut-swin", "DonutSwinModel"],
  ["yolos", "YolosModel"],
  ["dpt", "DPTModel"],
  ["glpn", "GLPNModel"],
  ["hifigan", "SpeechT5HifiGan"],
  ["efficientnet", "EfficientNetModel"],
  ["decision_transformer", "DecisionTransformerModel"],
  ["patchtst", "PatchTSTModel"],
  ["patchtsmixer", "PatchTSMixerModel"],
  ["mobilenet_v1", "MobileNetV1Model"],
  ["mobilenet_v2", "MobileNetV2Model"],
  ["mobilenet_v3", "MobileNetV3Model"],
  ["mobilenet_v4", "MobileNetV4Model"],
  ["maskformer", "MaskFormerModel"],
  ["mgp-str", "MgpstrForSceneTextRecognition"],
  ["style_text_to_speech_2", "StyleTextToSpeech2Model"],
  ["openai_privacy_filter", "OpenAIPrivacyFilterModel"]
]);
var MODEL_MAPPING_NAMES_ENCODER_DECODER = /* @__PURE__ */ new Map([
  ["t5", "T5Model"],
  ["longt5", "LongT5Model"],
  ["mt5", "MT5Model"],
  ["bart", "BartModel"],
  ["mbart", "MBartModel"],
  ["marian", "MarianModel"],
  ["whisper", "WhisperModel"],
  ["cohere_asr", "CohereAsrModel"],
  ["m2m_100", "M2M100Model"],
  ["blenderbot", "BlenderbotModel"],
  ["blenderbot-small", "BlenderbotSmallModel"]
]);
var MODEL_MAPPING_NAMES_AUTO_ENCODER = /* @__PURE__ */ new Map([
  ["mimi", "MimiModel"],
  ["dac", "DacModel"],
  ["snac", "SnacModel"]
]);
var MODEL_MAPPING_NAMES_DECODER_ONLY = /* @__PURE__ */ new Map([
  ["bloom", "BloomModel"],
  ["jais", "JAISModel"],
  ["gpt2", "GPT2Model"],
  ["gpt_oss", "GptOssModel"],
  ["gptj", "GPTJModel"],
  ["gpt_bigcode", "GPTBigCodeModel"],
  ["gpt_neo", "GPTNeoModel"],
  ["gpt_neox", "GPTNeoXModel"],
  ["codegen", "CodeGenModel"],
  ["llama", "LlamaModel"],
  ["apertus", "ApertusModel"],
  ["nanochat", "NanoChatModel"],
  ["arcee", "ArceeModel"],
  ["afmoe", "AfmoeModel"],
  ["lfm2", "Lfm2Model"],
  ["lfm2_moe", "Lfm2MoeModel"],
  ["smollm3", "SmolLM3Model"],
  ["exaone", "ExaoneModel"],
  ["olmo", "OlmoModel"],
  ["olmo2", "Olmo2Model"],
  ["olmo3", "Olmo3Model"],
  ["olmo_hybrid", "OlmoHybridModel"],
  ["mobilellm", "MobileLLMModel"],
  ["granite", "GraniteModel"],
  ["granitemoehybrid", "GraniteMoeHybridModel"],
  ["cohere", "CohereModel"],
  ["cohere2", "Cohere2Model"],
  ["gemma", "GemmaModel"],
  ["gemma2", "Gemma2Model"],
  ["vaultgemma", "VaultGemmaModel"],
  ["gemma3_text", "Gemma3Model"],
  ["helium", "HeliumModel"],
  ["glm", "GlmModel"],
  ["glm_moe_dsa", "GlmMoeDsaModel"],
  ["openelm", "OpenELMModel"],
  ["qwen2", "Qwen2Model"],
  ["qwen2_moe", "Qwen2MoeModel"],
  ["qwen3", "Qwen3Model"],
  ["qwen3_moe", "Qwen3MoeModel"],
  ["qwen3_next", "Qwen3NextModel"],
  ["phi", "PhiModel"],
  ["phi3", "Phi3Model"],
  ["mpt", "MptModel"],
  ["opt", "OPTModel"],
  ["mistral", "MistralModel"],
  ["mistral4", "Mistral4Model"],
  ["ministral", "MinistralModel"],
  ["ministral3", "Ministral3Model"],
  ["ernie4_5", "Ernie4_5ForCausalLM"],
  ["starcoder2", "Starcoder2Model"],
  ["deepseek_v3", "DeepseekV3Model"],
  ["falcon", "FalconModel"],
  ["falcon_h1", "FalconH1Model"],
  ["nemotron_h", "NemotronHModel"],
  ["solar_open", "SolarOpenModel"],
  ["stablelm", "StableLmModel"],
  ["modernbert-decoder", "ModernBertDecoderModel"],
  ["hunyuan_v1_dense", "HunYuanDenseV1Model"],
  ["youtu", "YoutuModel"]
]);
var MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING_NAMES = /* @__PURE__ */ new Map([
  ["speecht5", "SpeechT5ForSpeechToText"],
  ["whisper", "WhisperForConditionalGeneration"],
  ["lite-whisper", "LiteWhisperForConditionalGeneration"],
  ["moonshine", "MoonshineForConditionalGeneration"],
  ["cohere_asr", "CohereAsrForConditionalGeneration"]
]);
var MODEL_FOR_TEXT_TO_SPECTROGRAM_MAPPING_NAMES = /* @__PURE__ */ new Map([["speecht5", "SpeechT5ForTextToSpeech"]]);
var MODEL_FOR_TEXT_TO_WAVEFORM_MAPPING_NAMES = /* @__PURE__ */ new Map([
  ["vits", "VitsModel"],
  ["musicgen", "MusicgenForConditionalGeneration"],
  ["supertonic", "SupertonicForConditionalGeneration"]
]);
var MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES = /* @__PURE__ */ new Map([
  ["bert", "BertForSequenceClassification"],
  ["eurobert", "EuroBertForSequenceClassification"],
  ["neobert", "NeoBertForSequenceClassification"],
  ["modernbert", "ModernBertForSequenceClassification"],
  ["roformer", "RoFormerForSequenceClassification"],
  ["electra", "ElectraForSequenceClassification"],
  ["esm", "EsmForSequenceClassification"],
  ["convbert", "ConvBertForSequenceClassification"],
  ["camembert", "CamembertForSequenceClassification"],
  ["deberta", "DebertaForSequenceClassification"],
  ["deberta-v2", "DebertaV2ForSequenceClassification"],
  ["mpnet", "MPNetForSequenceClassification"],
  ["albert", "AlbertForSequenceClassification"],
  ["distilbert", "DistilBertForSequenceClassification"],
  ["roberta", "RobertaForSequenceClassification"],
  ["xlm", "XLMForSequenceClassification"],
  ["xlm-roberta", "XLMRobertaForSequenceClassification"],
  ["bart", "BartForSequenceClassification"],
  ["mbart", "MBartForSequenceClassification"],
  ["mobilebert", "MobileBertForSequenceClassification"],
  ["squeezebert", "SqueezeBertForSequenceClassification"]
]);
var MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES = /* @__PURE__ */ new Map([
  ["bert", "BertForTokenClassification"],
  ["eurobert", "EuroBertForTokenClassification"],
  ["neobert", "NeoBertForTokenClassification"],
  ["modernbert", "ModernBertForTokenClassification"],
  ["roformer", "RoFormerForTokenClassification"],
  ["electra", "ElectraForTokenClassification"],
  ["esm", "EsmForTokenClassification"],
  ["convbert", "ConvBertForTokenClassification"],
  ["camembert", "CamembertForTokenClassification"],
  ["deberta", "DebertaForTokenClassification"],
  ["deberta-v2", "DebertaV2ForTokenClassification"],
  ["mpnet", "MPNetForTokenClassification"],
  ["distilbert", "DistilBertForTokenClassification"],
  ["roberta", "RobertaForTokenClassification"],
  ["xlm", "XLMForTokenClassification"],
  ["xlm-roberta", "XLMRobertaForTokenClassification"],
  ["openai_privacy_filter", "OpenAIPrivacyFilterForTokenClassification"]
]);
var MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES = /* @__PURE__ */ new Map([
  ["t5", "T5ForConditionalGeneration"],
  ["longt5", "LongT5ForConditionalGeneration"],
  ["mt5", "MT5ForConditionalGeneration"],
  ["bart", "BartForConditionalGeneration"],
  ["mbart", "MBartForConditionalGeneration"],
  ["marian", "MarianMTModel"],
  ["m2m_100", "M2M100ForConditionalGeneration"],
  ["blenderbot", "BlenderbotForConditionalGeneration"],
  ["blenderbot-small", "BlenderbotSmallForConditionalGeneration"]
]);
var MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = /* @__PURE__ */ new Map([
  ["bloom", "BloomForCausalLM"],
  ["gpt2", "GPT2LMHeadModel"],
  ["gpt_oss", "GptOssForCausalLM"],
  ["jais", "JAISLMHeadModel"],
  ["gptj", "GPTJForCausalLM"],
  ["gpt_bigcode", "GPTBigCodeForCausalLM"],
  ["gpt_neo", "GPTNeoForCausalLM"],
  ["gpt_neox", "GPTNeoXForCausalLM"],
  ["codegen", "CodeGenForCausalLM"],
  ["llama", "LlamaForCausalLM"],
  ["nanochat", "NanoChatForCausalLM"],
  ["apertus", "ApertusForCausalLM"],
  ["llama4_text", "Llama4ForCausalLM"],
  ["arcee", "ArceeForCausalLM"],
  ["afmoe", "AfmoeForCausalLM"],
  ["lfm2", "Lfm2ForCausalLM"],
  ["lfm2_moe", "Lfm2MoeForCausalLM"],
  ["smollm3", "SmolLM3ForCausalLM"],
  ["exaone", "ExaoneForCausalLM"],
  ["olmo", "OlmoForCausalLM"],
  ["olmo2", "Olmo2ForCausalLM"],
  ["olmo3", "Olmo3ForCausalLM"],
  ["olmo_hybrid", "OlmoHybridForCausalLM"],
  ["mobilellm", "MobileLLMForCausalLM"],
  ["granite", "GraniteForCausalLM"],
  ["granitemoehybrid", "GraniteMoeHybridForCausalLM"],
  ["cohere", "CohereForCausalLM"],
  ["cohere2", "Cohere2ForCausalLM"],
  ["gemma", "GemmaForCausalLM"],
  ["gemma2", "Gemma2ForCausalLM"],
  ["vaultgemma", "VaultGemmaForCausalLM"],
  ["gemma3_text", "Gemma3ForCausalLM"],
  ["gemma3", "Gemma3ForCausalLM"],
  ["helium", "HeliumForCausalLM"],
  ["glm", "GlmForCausalLM"],
  ["glm_moe_dsa", "GlmMoeDsaForCausalLM"],
  ["openelm", "OpenELMForCausalLM"],
  ["qwen2", "Qwen2ForCausalLM"],
  ["qwen2_moe", "Qwen2MoeForCausalLM"],
  ["qwen3", "Qwen3ForCausalLM"],
  ["qwen3_moe", "Qwen3MoeForCausalLM"],
  ["qwen3_next", "Qwen3NextForCausalLM"],
  ["qwen2_vl", "Qwen2VLForCausalLM"],
  ["qwen2_5_vl", "Qwen2_5_VLForCausalLM"],
  ["qwen3_vl", "Qwen3VLForCausalLM"],
  ["qwen3_vl_moe", "Qwen3VLMoeForCausalLM"],
  ["qwen3_5", "Qwen3_5ForCausalLM"],
  ["qwen3_5_text", "Qwen3_5ForCausalLM"],
  ["qwen3_5_moe", "Qwen3_5MoeForCausalLM"],
  ["gemma3n", "Gemma3nForCausalLM"],
  ["gemma4", "Gemma4ForCausalLM"],
  ["phi", "PhiForCausalLM"],
  ["phi3", "Phi3ForCausalLM"],
  ["mpt", "MptForCausalLM"],
  ["opt", "OPTForCausalLM"],
  ["mbart", "MBartForCausalLM"],
  ["mistral", "MistralForCausalLM"],
  ["mistral4", "Mistral4ForCausalLM"],
  ["ministral", "MinistralForCausalLM"],
  ["ministral3", "Ministral3ForCausalLM"],
  ["ernie4_5", "Ernie4_5ForCausalLM"],
  ["starcoder2", "Starcoder2ForCausalLM"],
  ["deepseek_v3", "DeepseekV3ForCausalLM"],
  ["falcon", "FalconForCausalLM"],
  ["falcon_h1", "FalconH1ForCausalLM"],
  ["nemotron_h", "NemotronHForCausalLM"],
  ["trocr", "TrOCRForCausalLM"],
  ["solar_open", "SolarOpenForCausalLM"],
  ["stablelm", "StableLmForCausalLM"],
  ["modernbert-decoder", "ModernBertDecoderForCausalLM"],
  ["hunyuan_v1_dense", "HunYuanDenseV1ForCausalLM"],
  ["youtu", "YoutuForCausalLM"],
  // Also image-text-to-text
  ["phi3_v", "Phi3VForCausalLM"]
]);
var MODEL_FOR_MULTIMODALITY_MAPPING_NAMES = /* @__PURE__ */ new Map([["multi_modality", "MultiModalityCausalLM"]]);
var MODEL_FOR_MASKED_LM_MAPPING_NAMES = /* @__PURE__ */ new Map([
  ["bert", "BertForMaskedLM"],
  ["eurobert", "EuroBertForMaskedLM"],
  ["neobert", "NeoBertForMaskedLM"],
  ["modernbert", "ModernBertForMaskedLM"],
  ["roformer", "RoFormerForMaskedLM"],
  ["electra", "ElectraForMaskedLM"],
  ["esm", "EsmForMaskedLM"],
  ["convbert", "ConvBertForMaskedLM"],
  ["camembert", "CamembertForMaskedLM"],
  ["deberta", "DebertaForMaskedLM"],
  ["deberta-v2", "DebertaV2ForMaskedLM"],
  ["mpnet", "MPNetForMaskedLM"],
  ["albert", "AlbertForMaskedLM"],
  ["distilbert", "DistilBertForMaskedLM"],
  ["roberta", "RobertaForMaskedLM"],
  ["xlm", "XLMWithLMHeadModel"],
  ["xlm-roberta", "XLMRobertaForMaskedLM"],
  ["mobilebert", "MobileBertForMaskedLM"],
  ["squeezebert", "SqueezeBertForMaskedLM"]
]);
var MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES = /* @__PURE__ */ new Map([
  ["bert", "BertForQuestionAnswering"],
  ["neobert", "NeoBertForQuestionAnswering"],
  ["roformer", "RoFormerForQuestionAnswering"],
  ["electra", "ElectraForQuestionAnswering"],
  ["convbert", "ConvBertForQuestionAnswering"],
  ["camembert", "CamembertForQuestionAnswering"],
  ["deberta", "DebertaForQuestionAnswering"],
  ["deberta-v2", "DebertaV2ForQuestionAnswering"],
  ["mpnet", "MPNetForQuestionAnswering"],
  ["albert", "AlbertForQuestionAnswering"],
  ["distilbert", "DistilBertForQuestionAnswering"],
  ["roberta", "RobertaForQuestionAnswering"],
  ["xlm", "XLMForQuestionAnswering"],
  ["xlm-roberta", "XLMRobertaForQuestionAnswering"],
  ["mobilebert", "MobileBertForQuestionAnswering"],
  ["squeezebert", "SqueezeBertForQuestionAnswering"]
]);
var MODEL_FOR_VISION_2_SEQ_MAPPING_NAMES = /* @__PURE__ */ new Map([
  ["vision-encoder-decoder", "VisionEncoderDecoderModel"],
  ["idefics3", "Idefics3ForConditionalGeneration"],
  ["smolvlm", "SmolVLMForConditionalGeneration"]
]);
var MODEL_FOR_IMAGE_TEXT_TO_TEXT_MAPPING_NAMES = /* @__PURE__ */ new Map([
  ["llava", "LlavaForConditionalGeneration"],
  ["llava_onevision", "LlavaOnevisionForConditionalGeneration"],
  ["moondream1", "Moondream1ForConditionalGeneration"],
  ["florence2", "Florence2ForConditionalGeneration"],
  ["qwen2_vl", "Qwen2VLForConditionalGeneration"],
  ["qwen2_5_vl", "Qwen2_5_VLForConditionalGeneration"],
  ["qwen3_vl", "Qwen3VLForConditionalGeneration"],
  ["qwen3_vl_moe", "Qwen3VLMoeForConditionalGeneration"],
  ["qwen3_5", "Qwen3_5ForConditionalGeneration"],
  ["qwen3_5_moe", "Qwen3_5MoeForConditionalGeneration"],
  ["lfm2_vl", "Lfm2VlForConditionalGeneration"],
  ["idefics3", "Idefics3ForConditionalGeneration"],
  ["smolvlm", "SmolVLMForConditionalGeneration"],
  ["paligemma", "PaliGemmaForConditionalGeneration"],
  ["llava_qwen2", "LlavaQwen2ForCausalLM"],
  ["gemma3", "Gemma3ForConditionalGeneration"],
  ["gemma3n", "Gemma3nForConditionalGeneration"],
  ["gemma4", "Gemma4ForConditionalGeneration"],
  ["mistral3", "Mistral3ForConditionalGeneration"],
  ["lighton_ocr", "LightOnOcrForConditionalGeneration"],
  ["glm_ocr", "GlmOcrForConditionalGeneration"]
]);
var MODEL_FOR_AUDIO_TEXT_TO_TEXT_MAPPING_NAMES = /* @__PURE__ */ new Map([
  ["granite_speech", "GraniteSpeechForConditionalGeneration"],
  ["ultravox", "UltravoxModel"],
  ["voxtral", "VoxtralForConditionalGeneration"],
  ["voxtral_realtime", "VoxtralRealtimeForConditionalGeneration"]
]);
var MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING_NAMES = /* @__PURE__ */ new Map([
  ["vision-encoder-decoder", "VisionEncoderDecoderModel"]
]);
var MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES = /* @__PURE__ */ new Map([
  ["vit", "ViTForImageClassification"],
  ["ijepa", "IJepaForImageClassification"],
  ["pvt", "PvtForImageClassification"],
  ["vit_msn", "ViTMSNForImageClassification"],
  ["fastvit", "FastViTForImageClassification"],
  ["mobilevit", "MobileViTForImageClassification"],
  ["mobilevitv2", "MobileViTV2ForImageClassification"],
  ["beit", "BeitForImageClassification"],
  ["deit", "DeiTForImageClassification"],
  ["hiera", "HieraForImageClassification"],
  ["convnext", "ConvNextForImageClassification"],
  ["convnextv2", "ConvNextV2ForImageClassification"],
  ["dinov2", "Dinov2ForImageClassification"],
  ["dinov2_with_registers", "Dinov2WithRegistersForImageClassification"],
  ["resnet", "ResNetForImageClassification"],
  ["swin", "SwinForImageClassification"],
  ["segformer", "SegformerForImageClassification"],
  ["efficientnet", "EfficientNetForImageClassification"],
  ["mobilenet_v1", "MobileNetV1ForImageClassification"],
  ["mobilenet_v2", "MobileNetV2ForImageClassification"],
  ["mobilenet_v3", "MobileNetV3ForImageClassification"],
  ["mobilenet_v4", "MobileNetV4ForImageClassification"]
]);
var MODEL_FOR_OBJECT_DETECTION_MAPPING_NAMES = /* @__PURE__ */ new Map([
  ["detr", "DetrForObjectDetection"],
  ["rt_detr", "RTDetrForObjectDetection"],
  ["rt_detr_v2", "RTDetrV2ForObjectDetection"],
  ["rf_detr", "RFDetrForObjectDetection"],
  ["d_fine", "DFineForObjectDetection"],
  ["table-transformer", "TableTransformerForObjectDetection"],
  ["yolos", "YolosForObjectDetection"]
]);
var MODEL_FOR_ZERO_SHOT_OBJECT_DETECTION_MAPPING_NAMES = /* @__PURE__ */ new Map([
  ["owlvit", "OwlViTForObjectDetection"],
  ["owlv2", "Owlv2ForObjectDetection"],
  ["grounding-dino", "GroundingDinoForObjectDetection"]
]);
var MODEL_FOR_IMAGE_SEGMENTATION_MAPPING_NAMES = /* @__PURE__ */ new Map([
  // TODO: Do not add new models here
  ["detr", "DetrForSegmentation"],
  ["clipseg", "CLIPSegForImageSegmentation"]
]);
var MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING_NAMES = /* @__PURE__ */ new Map([
  ["segformer", "SegformerForSemanticSegmentation"],
  ["sapiens", "SapiensForSemanticSegmentation"],
  ["swin", "SwinForSemanticSegmentation"],
  ["mobilenet_v1", "MobileNetV1ForSemanticSegmentation"],
  ["mobilenet_v2", "MobileNetV2ForSemanticSegmentation"],
  ["mobilenet_v3", "MobileNetV3ForSemanticSegmentation"],
  ["mobilenet_v4", "MobileNetV4ForSemanticSegmentation"]
]);
var MODEL_FOR_UNIVERSAL_SEGMENTATION_MAPPING_NAMES = /* @__PURE__ */ new Map([
  ["detr", "DetrForSegmentation"],
  ["maskformer", "MaskFormerForInstanceSegmentation"]
]);
var MODEL_FOR_MASK_GENERATION_MAPPING_NAMES = /* @__PURE__ */ new Map([
  ["sam", "SamModel"],
  ["sam2", "Sam2Model"],
  ["edgetam", "EdgeTamModel"],
  ["sam3_tracker", "Sam3TrackerModel"]
]);
var MODEL_FOR_CTC_MAPPING_NAMES = /* @__PURE__ */ new Map([
  ["wav2vec2", "Wav2Vec2ForCTC"],
  ["wav2vec2-bert", "Wav2Vec2BertForCTC"],
  ["unispeech", "UniSpeechForCTC"],
  ["unispeech-sat", "UniSpeechSatForCTC"],
  ["wavlm", "WavLMForCTC"],
  ["hubert", "HubertForCTC"],
  ["parakeet_ctc", "ParakeetForCTC"]
]);
var MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING_NAMES = /* @__PURE__ */ new Map([
  ["wav2vec2", "Wav2Vec2ForSequenceClassification"],
  ["wav2vec2-bert", "Wav2Vec2BertForSequenceClassification"],
  ["unispeech", "UniSpeechForSequenceClassification"],
  ["unispeech-sat", "UniSpeechSatForSequenceClassification"],
  ["wavlm", "WavLMForSequenceClassification"],
  ["hubert", "HubertForSequenceClassification"],
  ["audio-spectrogram-transformer", "ASTForAudioClassification"]
]);
var MODEL_FOR_AUDIO_XVECTOR_MAPPING_NAMES = /* @__PURE__ */ new Map([["wavlm", "WavLMForXVector"]]);
var MODEL_FOR_AUDIO_FRAME_CLASSIFICATION_MAPPING_NAMES = /* @__PURE__ */ new Map([
  ["unispeech-sat", "UniSpeechSatForAudioFrameClassification"],
  ["wavlm", "WavLMForAudioFrameClassification"],
  ["wav2vec2", "Wav2Vec2ForAudioFrameClassification"],
  ["pyannote", "PyAnnoteForAudioFrameClassification"]
]);
var MODEL_FOR_IMAGE_MATTING_MAPPING_NAMES = /* @__PURE__ */ new Map([["vitmatte", "VitMatteForImageMatting"]]);
var MODEL_FOR_TIME_SERIES_PREDICTION_MAPPING_NAMES = /* @__PURE__ */ new Map([
  ["patchtst", "PatchTSTForPrediction"],
  ["patchtsmixer", "PatchTSMixerForPrediction"]
]);
var MODEL_FOR_IMAGE_TO_IMAGE_MAPPING_NAMES = /* @__PURE__ */ new Map([["swin2sr", "Swin2SRForImageSuperResolution"]]);
var MODEL_FOR_DEPTH_ESTIMATION_MAPPING_NAMES = /* @__PURE__ */ new Map([
  ["chmv2", "CHMv2ForDepthEstimation"],
  ["dpt", "DPTForDepthEstimation"],
  ["depth_anything", "DepthAnythingForDepthEstimation"],
  ["glpn", "GLPNForDepthEstimation"],
  ["sapiens", "SapiensForDepthEstimation"],
  ["depth_pro", "DepthProForDepthEstimation"],
  ["metric3d", "Metric3DForDepthEstimation"],
  ["metric3dv2", "Metric3Dv2ForDepthEstimation"]
]);
var MODEL_FOR_NORMAL_ESTIMATION_MAPPING_NAMES = /* @__PURE__ */ new Map([["sapiens", "SapiensForNormalEstimation"]]);
var MODEL_FOR_POSE_ESTIMATION_MAPPING_NAMES = /* @__PURE__ */ new Map([["vitpose", "VitPoseForPoseEstimation"]]);
var MODEL_FOR_IMAGE_FEATURE_EXTRACTION_MAPPING_NAMES = /* @__PURE__ */ new Map([
  ["clip", "CLIPVisionModelWithProjection"],
  ["siglip", "SiglipVisionModel"],
  ["jina_clip", "JinaCLIPVisionModel"]
]);
var MODEL_CLASS_TYPE_MAPPING = [
  // MODEL_MAPPING_NAMES:
  [MODEL_MAPPING_NAMES_ENCODER_ONLY, MODEL_TYPES.EncoderOnly],
  [MODEL_MAPPING_NAMES_ENCODER_DECODER, MODEL_TYPES.EncoderDecoder],
  [MODEL_MAPPING_NAMES_DECODER_ONLY, MODEL_TYPES.DecoderOnlyWithoutHead],
  [MODEL_MAPPING_NAMES_AUTO_ENCODER, MODEL_TYPES.AutoEncoder],
  [MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES, MODEL_TYPES.EncoderOnly],
  [MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES, MODEL_TYPES.EncoderOnly],
  [MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES, MODEL_TYPES.Seq2Seq],
  [MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING_NAMES, MODEL_TYPES.Seq2Seq],
  [MODEL_FOR_CAUSAL_LM_MAPPING_NAMES, MODEL_TYPES.DecoderOnly],
  [MODEL_FOR_MULTIMODALITY_MAPPING_NAMES, MODEL_TYPES.MultiModality],
  [MODEL_FOR_MASKED_LM_MAPPING_NAMES, MODEL_TYPES.EncoderOnly],
  [MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES, MODEL_TYPES.EncoderOnly],
  [MODEL_FOR_VISION_2_SEQ_MAPPING_NAMES, MODEL_TYPES.Vision2Seq],
  [MODEL_FOR_IMAGE_TEXT_TO_TEXT_MAPPING_NAMES, MODEL_TYPES.ImageTextToText],
  [MODEL_FOR_AUDIO_TEXT_TO_TEXT_MAPPING_NAMES, MODEL_TYPES.AudioTextToText],
  [MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES, MODEL_TYPES.EncoderOnly],
  [MODEL_FOR_IMAGE_SEGMENTATION_MAPPING_NAMES, MODEL_TYPES.EncoderOnly],
  [MODEL_FOR_UNIVERSAL_SEGMENTATION_MAPPING_NAMES, MODEL_TYPES.EncoderOnly],
  [MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING_NAMES, MODEL_TYPES.EncoderOnly],
  [MODEL_FOR_IMAGE_MATTING_MAPPING_NAMES, MODEL_TYPES.EncoderOnly],
  [MODEL_FOR_TIME_SERIES_PREDICTION_MAPPING_NAMES, MODEL_TYPES.EncoderOnly],
  [MODEL_FOR_IMAGE_TO_IMAGE_MAPPING_NAMES, MODEL_TYPES.EncoderOnly],
  [MODEL_FOR_DEPTH_ESTIMATION_MAPPING_NAMES, MODEL_TYPES.EncoderOnly],
  [MODEL_FOR_NORMAL_ESTIMATION_MAPPING_NAMES, MODEL_TYPES.EncoderOnly],
  [MODEL_FOR_POSE_ESTIMATION_MAPPING_NAMES, MODEL_TYPES.EncoderOnly],
  [MODEL_FOR_OBJECT_DETECTION_MAPPING_NAMES, MODEL_TYPES.EncoderOnly],
  [MODEL_FOR_ZERO_SHOT_OBJECT_DETECTION_MAPPING_NAMES, MODEL_TYPES.EncoderOnly],
  [MODEL_FOR_MASK_GENERATION_MAPPING_NAMES, MODEL_TYPES.MaskGeneration],
  [MODEL_FOR_CTC_MAPPING_NAMES, MODEL_TYPES.EncoderOnly],
  [MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING_NAMES, MODEL_TYPES.EncoderOnly],
  [MODEL_FOR_TEXT_TO_SPECTROGRAM_MAPPING_NAMES, MODEL_TYPES.Seq2Seq],
  [MODEL_FOR_TEXT_TO_WAVEFORM_MAPPING_NAMES, MODEL_TYPES.EncoderOnly],
  [MODEL_FOR_AUDIO_XVECTOR_MAPPING_NAMES, MODEL_TYPES.EncoderOnly],
  [MODEL_FOR_AUDIO_FRAME_CLASSIFICATION_MAPPING_NAMES, MODEL_TYPES.EncoderOnly],
  // Custom:
  [MODEL_FOR_IMAGE_FEATURE_EXTRACTION_MAPPING_NAMES, MODEL_TYPES.EncoderOnly]
];
for (const [mappings, type] of MODEL_CLASS_TYPE_MAPPING) {
  for (const name of mappings.values()) {
    MODEL_TYPE_MAPPING.set(name, type);
    const model = models_exports[name];
    MODEL_CLASS_TO_NAME_MAPPING.set(model, name);
    MODEL_NAME_TO_CLASS_MAPPING.set(name, model);
  }
}
var CUSTOM_MAPPING = [
  // OVERRIDE:
  // TODO: Refactor to allow class to specify model
  ["MusicgenForConditionalGeneration", MusicgenForConditionalGeneration, MODEL_TYPES.Musicgen],
  ["Phi3VForCausalLM", Phi3VForCausalLM, MODEL_TYPES.Phi3V],
  ["CLIPTextModelWithProjection", CLIPTextModelWithProjection, MODEL_TYPES.EncoderOnly],
  ["SiglipTextModel", SiglipTextModel, MODEL_TYPES.EncoderOnly],
  ["JinaCLIPTextModel", JinaCLIPTextModel, MODEL_TYPES.EncoderOnly],
  ["ClapTextModelWithProjection", ClapTextModelWithProjection, MODEL_TYPES.EncoderOnly],
  ["ClapAudioModelWithProjection", ClapAudioModelWithProjection, MODEL_TYPES.EncoderOnly],
  ["DacEncoderModel", DacEncoderModel, MODEL_TYPES.EncoderOnly],
  ["DacDecoderModel", DacDecoderModel, MODEL_TYPES.EncoderOnly],
  ["MimiEncoderModel", MimiEncoderModel, MODEL_TYPES.EncoderOnly],
  ["MimiDecoderModel", MimiDecoderModel, MODEL_TYPES.EncoderOnly],
  ["SnacEncoderModel", SnacEncoderModel, MODEL_TYPES.EncoderOnly],
  ["SnacDecoderModel", SnacDecoderModel, MODEL_TYPES.EncoderOnly],
  [
    "Gemma3nForConditionalGeneration",
    Gemma3nForConditionalGeneration,
    MODEL_TYPES.ImageAudioTextToText
  ],
  [
    "Gemma4ForConditionalGeneration",
    Gemma4ForConditionalGeneration,
    MODEL_TYPES.ImageAudioTextToText
  ],
  ["SupertonicForConditionalGeneration", SupertonicForConditionalGeneration, MODEL_TYPES.Supertonic],
  ["ChatterboxModel", ChatterboxModel, MODEL_TYPES.Chatterbox],
  [
    "VoxtralRealtimeForConditionalGeneration",
    VoxtralRealtimeForConditionalGeneration,
    MODEL_TYPES.VoxtralRealtime
  ]
];
for (const [name, model, type] of CUSTOM_MAPPING) {
  MODEL_TYPE_MAPPING.set(name, type);
  MODEL_CLASS_TO_NAME_MAPPING.set(model, name);
  MODEL_NAME_TO_CLASS_MAPPING.set(name, model);
}
var CUSTOM_ARCHITECTURES_MAPPING = /* @__PURE__ */ new Map([
  ["modnet", MODEL_FOR_IMAGE_SEGMENTATION_MAPPING_NAMES],
  ["birefnet", MODEL_FOR_IMAGE_SEGMENTATION_MAPPING_NAMES],
  ["isnet", MODEL_FOR_IMAGE_SEGMENTATION_MAPPING_NAMES],
  ["ben", MODEL_FOR_IMAGE_SEGMENTATION_MAPPING_NAMES]
]);
for (const [name, mapping] of CUSTOM_ARCHITECTURES_MAPPING.entries()) {
  mapping.set(name, "PreTrainedModel");
  MODEL_TYPE_MAPPING.set(name, MODEL_TYPES.EncoderOnly);
  MODEL_NAME_TO_CLASS_MAPPING.set(name, PreTrainedModel);
}
var CUSTOM_ARCHITECTURES = new Set(CUSTOM_ARCHITECTURES_MAPPING.keys());
MODEL_TYPE_MAPPING.set("PreTrainedModel", MODEL_TYPES.EncoderOnly);
MODEL_CLASS_TO_NAME_MAPPING.set(PreTrainedModel, "PreTrainedModel");
var MODEL_MAPPINGS = {
  MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES,
  MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES,
  MODEL_FOR_TEXT_TO_SPECTROGRAM_MAPPING_NAMES,
  MODEL_FOR_TEXT_TO_WAVEFORM_MAPPING_NAMES,
  MODEL_FOR_MASKED_LM_MAPPING_NAMES,
  MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES,
  MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES,
  MODEL_FOR_IMAGE_SEGMENTATION_MAPPING_NAMES,
  MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING_NAMES,
  MODEL_FOR_UNIVERSAL_SEGMENTATION_MAPPING_NAMES,
  MODEL_FOR_OBJECT_DETECTION_MAPPING_NAMES,
  MODEL_FOR_ZERO_SHOT_OBJECT_DETECTION_MAPPING_NAMES,
  MODEL_FOR_MASK_GENERATION_MAPPING_NAMES,
  MODEL_FOR_CTC_MAPPING_NAMES,
  MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING_NAMES,
  MODEL_FOR_AUDIO_XVECTOR_MAPPING_NAMES,
  MODEL_FOR_AUDIO_FRAME_CLASSIFICATION_MAPPING_NAMES,
  MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING_NAMES,
  MODEL_FOR_IMAGE_MATTING_MAPPING_NAMES,
  MODEL_FOR_IMAGE_TO_IMAGE_MAPPING_NAMES,
  MODEL_FOR_DEPTH_ESTIMATION_MAPPING_NAMES,
  MODEL_FOR_NORMAL_ESTIMATION_MAPPING_NAMES,
  MODEL_FOR_POSE_ESTIMATION_MAPPING_NAMES,
  MODEL_FOR_IMAGE_FEATURE_EXTRACTION_MAPPING_NAMES,
  MODEL_FOR_IMAGE_TEXT_TO_TEXT_MAPPING_NAMES,
  MODEL_FOR_AUDIO_TEXT_TO_TEXT_MAPPING_NAMES,
  MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES,
  MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING_NAMES,
  MODEL_FOR_CAUSAL_LM_MAPPING_NAMES,
  MODEL_FOR_VISION_2_SEQ_MAPPING_NAMES
};
registerTaskMappings(MODEL_MAPPINGS);

// src/models/auto/modeling_auto.js
var PretrainedMixin = class {
  /**
   * Mapping from model type to model class.
   * @type {Map<string, Object>[]}
   */
  static MODEL_CLASS_MAPPINGS = null;
  /**
   * Whether to attempt to instantiate the base class (`PretrainedModel`) if
   * the model type is not found in the mapping.
   */
  static BASE_IF_FAIL = false;
  /**
   * Check whether this AutoModel class supports a given model type.
   * @param {string} model_type The model type from config (e.g., 'bert', 'whisper').
   * @returns {boolean} Whether this class can handle the given model type.
   */
  static supports(model_type) {
    if (!this.MODEL_CLASS_MAPPINGS) return false;
    for (const mapping of this.MODEL_CLASS_MAPPINGS) {
      if (mapping.has(model_type)) return true;
    }
    return this.BASE_IF_FAIL;
  }
  /** @type {typeof PreTrainedModel.from_pretrained} */
  static async from_pretrained(pretrained_model_name_or_path, {
    progress_callback = null,
    config = null,
    cache_dir = null,
    local_files_only = false,
    revision = "main",
    model_file_name = null,
    subfolder = "onnx",
    device = null,
    dtype = null,
    use_external_data_format = null,
    session_options = {}
  } = {}) {
    const options = {
      progress_callback,
      config,
      cache_dir,
      local_files_only,
      revision,
      model_file_name,
      subfolder,
      device,
      dtype,
      use_external_data_format,
      session_options
    };
    options.config = await AutoConfig.from_pretrained(pretrained_model_name_or_path, options);
    if (!this.MODEL_CLASS_MAPPINGS) {
      throw new Error("`MODEL_CLASS_MAPPINGS` not implemented for this type of `AutoClass`: " + this.name);
    }
    const { model_type } = options.config;
    for (const MODEL_CLASS_MAPPING of this.MODEL_CLASS_MAPPINGS) {
      let modelInfo = MODEL_CLASS_MAPPING.get(model_type);
      if (!modelInfo) {
        for (const cls of MODEL_CLASS_MAPPING.values()) {
          if (cls[0] === model_type) {
            modelInfo = cls;
            break;
          }
        }
        if (!modelInfo) continue;
      }
      return await models_exports[modelInfo].from_pretrained(pretrained_model_name_or_path, options);
    }
    if (this.BASE_IF_FAIL) {
      if (!CUSTOM_ARCHITECTURES.has(model_type)) {
        logger.warn(`Unknown model class "${model_type}", attempting to construct from base class.`);
      }
      return await PreTrainedModel.from_pretrained(pretrained_model_name_or_path, options);
    } else {
      throw Error(`Unsupported model type: ${model_type}`);
    }
  }
};
var AutoModel = class extends PretrainedMixin {
  /** @type {Map<string, Object>[]} */
  // @ts-ignore
  static MODEL_CLASS_MAPPINGS = MODEL_CLASS_TYPE_MAPPING.map((x) => x[0]);
  static BASE_IF_FAIL = true;
};
var AutoModelForSequenceClassification = class extends PretrainedMixin {
  static MODEL_CLASS_MAPPINGS = [MODEL_MAPPINGS.MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES];
};
var AutoModelForTokenClassification = class extends PretrainedMixin {
  static MODEL_CLASS_MAPPINGS = [MODEL_MAPPINGS.MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES];
};
var AutoModelForSeq2SeqLM = class extends PretrainedMixin {
  static MODEL_CLASS_MAPPINGS = [MODEL_MAPPINGS.MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES];
};
var AutoModelForSpeechSeq2Seq = class extends PretrainedMixin {
  static MODEL_CLASS_MAPPINGS = [MODEL_MAPPINGS.MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING_NAMES];
};
var AutoModelForTextToSpectrogram = class extends PretrainedMixin {
  static MODEL_CLASS_MAPPINGS = [MODEL_MAPPINGS.MODEL_FOR_TEXT_TO_SPECTROGRAM_MAPPING_NAMES];
};
var AutoModelForTextToWaveform = class extends PretrainedMixin {
  static MODEL_CLASS_MAPPINGS = [MODEL_MAPPINGS.MODEL_FOR_TEXT_TO_WAVEFORM_MAPPING_NAMES];
};
var AutoModelForCausalLM = class extends PretrainedMixin {
  static MODEL_CLASS_MAPPINGS = [MODEL_MAPPINGS.MODEL_FOR_CAUSAL_LM_MAPPING_NAMES];
};
var AutoModelForMaskedLM = class extends PretrainedMixin {
  static MODEL_CLASS_MAPPINGS = [MODEL_MAPPINGS.MODEL_FOR_MASKED_LM_MAPPING_NAMES];
};
var AutoModelForQuestionAnswering = class extends PretrainedMixin {
  static MODEL_CLASS_MAPPINGS = [MODEL_MAPPINGS.MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES];
};
var AutoModelForVision2Seq = class extends PretrainedMixin {
  static MODEL_CLASS_MAPPINGS = [MODEL_MAPPINGS.MODEL_FOR_VISION_2_SEQ_MAPPING_NAMES];
};
var AutoModelForImageClassification = class extends PretrainedMixin {
  static MODEL_CLASS_MAPPINGS = [MODEL_MAPPINGS.MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES];
};
var AutoModelForImageSegmentation = class extends PretrainedMixin {
  static MODEL_CLASS_MAPPINGS = [MODEL_MAPPINGS.MODEL_FOR_IMAGE_SEGMENTATION_MAPPING_NAMES];
};
var AutoModelForSemanticSegmentation = class extends PretrainedMixin {
  static MODEL_CLASS_MAPPINGS = [MODEL_MAPPINGS.MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING_NAMES];
};
var AutoModelForUniversalSegmentation = class extends PretrainedMixin {
  static MODEL_CLASS_MAPPINGS = [MODEL_MAPPINGS.MODEL_FOR_UNIVERSAL_SEGMENTATION_MAPPING_NAMES];
};
var AutoModelForObjectDetection = class extends PretrainedMixin {
  static MODEL_CLASS_MAPPINGS = [MODEL_MAPPINGS.MODEL_FOR_OBJECT_DETECTION_MAPPING_NAMES];
};
var AutoModelForZeroShotObjectDetection = class extends PretrainedMixin {
  static MODEL_CLASS_MAPPINGS = [MODEL_MAPPINGS.MODEL_FOR_ZERO_SHOT_OBJECT_DETECTION_MAPPING_NAMES];
};
var AutoModelForMaskGeneration = class extends PretrainedMixin {
  static MODEL_CLASS_MAPPINGS = [MODEL_MAPPINGS.MODEL_FOR_MASK_GENERATION_MAPPING_NAMES];
};
var AutoModelForCTC = class extends PretrainedMixin {
  static MODEL_CLASS_MAPPINGS = [MODEL_MAPPINGS.MODEL_FOR_CTC_MAPPING_NAMES];
};
var AutoModelForAudioClassification = class extends PretrainedMixin {
  static MODEL_CLASS_MAPPINGS = [MODEL_MAPPINGS.MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING_NAMES];
};
var AutoModelForXVector = class extends PretrainedMixin {
  static MODEL_CLASS_MAPPINGS = [MODEL_MAPPINGS.MODEL_FOR_AUDIO_XVECTOR_MAPPING_NAMES];
};
var AutoModelForAudioFrameClassification = class extends PretrainedMixin {
  static MODEL_CLASS_MAPPINGS = [MODEL_MAPPINGS.MODEL_FOR_AUDIO_FRAME_CLASSIFICATION_MAPPING_NAMES];
};
var AutoModelForDocumentQuestionAnswering = class extends PretrainedMixin {
  static MODEL_CLASS_MAPPINGS = [MODEL_MAPPINGS.MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING_NAMES];
};
var AutoModelForImageMatting = class extends PretrainedMixin {
  static MODEL_CLASS_MAPPINGS = [MODEL_MAPPINGS.MODEL_FOR_IMAGE_MATTING_MAPPING_NAMES];
};
var AutoModelForImageToImage = class extends PretrainedMixin {
  static MODEL_CLASS_MAPPINGS = [MODEL_MAPPINGS.MODEL_FOR_IMAGE_TO_IMAGE_MAPPING_NAMES];
};
var AutoModelForDepthEstimation = class extends PretrainedMixin {
  static MODEL_CLASS_MAPPINGS = [MODEL_MAPPINGS.MODEL_FOR_DEPTH_ESTIMATION_MAPPING_NAMES];
};
var AutoModelForNormalEstimation = class extends PretrainedMixin {
  static MODEL_CLASS_MAPPINGS = [MODEL_MAPPINGS.MODEL_FOR_NORMAL_ESTIMATION_MAPPING_NAMES];
};
var AutoModelForPoseEstimation = class extends PretrainedMixin {
  static MODEL_CLASS_MAPPINGS = [MODEL_MAPPINGS.MODEL_FOR_POSE_ESTIMATION_MAPPING_NAMES];
};
var AutoModelForImageFeatureExtraction = class extends PretrainedMixin {
  static MODEL_CLASS_MAPPINGS = [MODEL_MAPPINGS.MODEL_FOR_IMAGE_FEATURE_EXTRACTION_MAPPING_NAMES];
};
var AutoModelForImageTextToText = class extends PretrainedMixin {
  static MODEL_CLASS_MAPPINGS = [MODEL_MAPPINGS.MODEL_FOR_IMAGE_TEXT_TO_TEXT_MAPPING_NAMES];
};
var AutoModelForAudioTextToText = class extends PretrainedMixin {
  static MODEL_CLASS_MAPPINGS = [MODEL_MAPPINGS.MODEL_FOR_AUDIO_TEXT_TO_TEXT_MAPPING_NAMES];
};

// src/pipelines/_base.js
async function prepareImages(images) {
  if (!Array.isArray(images)) {
    images = [images];
  }
  return await Promise.all(images.map((x) => RawImage.read(x)));
}
async function prepareAudios(audios, sampling_rate) {
  if (!Array.isArray(audios)) {
    audios = [audios];
  }
  return await Promise.all(
    audios.map((x) => {
      if (typeof x === "string" || x instanceof URL) {
        return read_audio(x, sampling_rate);
      } else if (x instanceof Float64Array) {
        return new Float32Array(x);
      }
      return x;
    })
  );
}
function get_bounding_box(box, asInteger) {
  if (asInteger) {
    box = box.map((x) => x | 0);
  }
  const [xmin, ymin, xmax, ymax] = box;
  return { xmin, ymin, xmax, ymax };
}
var Pipeline = class extends Callable {
  /**
   * Create a new Pipeline.
   * @param {Object} options An object containing the following properties:
   * @param {string} [options.task] The task of the pipeline. Useful for specifying subtasks.
   * @param {PreTrainedModel} [options.model] The model used by the pipeline.
   * @param {PreTrainedTokenizer} [options.tokenizer=null] The tokenizer used by the pipeline (if any).
   * @param {Processor} [options.processor=null] The processor used by the pipeline (if any).
   */
  constructor({ task, model, tokenizer = null, processor = null }) {
    super();
    this.task = task;
    this.model = model;
    this.tokenizer = tokenizer;
    this.processor = processor;
  }
  /** @type {DisposeType} */
  async dispose() {
    await this.model.dispose();
  }
};

// src/pipelines/text-classification.js
var TextClassificationPipeline = class extends /** @type {new (options: TextPipelineConstructorArgs) => TextClassificationPipelineType} */
Pipeline {
  async _call(texts, { top_k = 1 } = {}) {
    const model_inputs = this.tokenizer(texts, {
      padding: true,
      truncation: true
    });
    const outputs = await this.model(model_inputs);
    const { problem_type, id2label } = this.model.config;
    const function_to_apply = problem_type === "multi_label_classification" ? (batch) => batch.sigmoid() : (batch) => new Tensor2("float32", softmax(batch.data), batch.dims);
    const toReturn = [];
    for (const batch of outputs.logits) {
      const output = function_to_apply(batch);
      const scores = await topk(output, top_k);
      const values = scores[0].tolist();
      const indices = scores[1].tolist();
      const vals = indices.map((x, i) => ({
        label: id2label ? id2label[x] : `LABEL_${x}`,
        score: values[i]
      }));
      if (top_k === 1) {
        toReturn.push(...vals);
      } else {
        toReturn.push(vals);
      }
    }
    return Array.isArray(texts) || top_k === 1 ? toReturn : toReturn[0];
  }
};

// src/pipelines/token-classification.js
var TokenClassificationPipeline = class extends /** @type {new (options: TextPipelineConstructorArgs) => TokenClassificationPipelineType} */
Pipeline {
  async _call(texts, { ignore_labels = ["O"], aggregation_strategy = "none" } = {}) {
    if (aggregation_strategy !== "none" && aggregation_strategy !== "simple") {
      throw new Error(
        `Invalid aggregation_strategy: "${aggregation_strategy}". Must be one of "none" or "simple".`
      );
    }
    const isBatched = Array.isArray(texts);
    const model_inputs = this.tokenizer(isBatched ? texts : [texts], {
      padding: true,
      truncation: true
    });
    const outputs = await this.model(model_inputs);
    const logits = outputs.logits;
    const id2label = this.model.config.id2label;
    const toReturn = [];
    for (let i = 0; i < logits.dims[0]; ++i) {
      const ids = model_inputs.input_ids[i].tolist();
      const batch = logits[i];
      const tokens = [];
      for (let j = 0; j < batch.dims[0]; ++j) {
        const tokenData = batch[j];
        const topScoreIndex = max(tokenData.data)[1];
        const entity = id2label ? id2label[topScoreIndex] : `LABEL_${topScoreIndex}`;
        if (ignore_labels.includes(entity)) continue;
        const word = this.tokenizer.decode([ids[j]], { skip_special_tokens: true });
        if (word === "") continue;
        const scores = softmax(tokenData.data);
        tokens.push({
          entity,
          score: scores[topScoreIndex],
          index: j,
          word
          // TODO: Add support for start and end
        });
      }
      toReturn.push(aggregation_strategy === "simple" ? groupEntities(tokens, ids, this.tokenizer) : tokens);
    }
    return isBatched ? toReturn : toReturn[0];
  }
};
function getTag(entity) {
  const p = entity[0];
  return entity[1] === "-" && (p === "B" || p === "I" || p === "E" || p === "S") ? [p, entity.slice(2)] : ["I", entity];
}
function groupEntities(tokens, ids, tokenizer) {
  const groups = [];
  let openTag = null;
  for (let i = 0; i < tokens.length; ++i) {
    const [prefix, tag] = getTag(tokens[i].entity);
    const extend = openTag === tag && prefix !== "B" && prefix !== "S";
    if (extend) {
      groups[groups.length - 1].end = i + 1;
      if (prefix === "E") openTag = null;
    } else {
      groups.push({ tag, start: i, end: i + 1 });
      openTag = prefix === "S" ? null : tag;
    }
  }
  return groups.map(({ tag, start, end }) => {
    let scoreSum = 0;
    const groupIds = [];
    for (let i = start; i < end; ++i) {
      scoreSum += tokens[i].score;
      groupIds.push(ids[tokens[i].index]);
    }
    return {
      entity_group: tag,
      score: scoreSum / (end - start),
      word: tokenizer.decode(groupIds, { skip_special_tokens: true })
    };
  });
}

// src/pipelines/question-answering.js
var QuestionAnsweringPipeline = class extends /** @type {new (options: TextPipelineConstructorArgs) => QuestionAnsweringPipelineType} */
Pipeline {
  async _call(question, context, { top_k = 1 } = {}) {
    const inputs = this.tokenizer(question, {
      text_pair: context,
      padding: true,
      truncation: true
    });
    const isBatched = Array.isArray(question);
    const { start_logits, end_logits } = await this.model(inputs);
    const input_ids = inputs.input_ids.tolist();
    const attention_mask = inputs.attention_mask.tolist();
    const { all_special_ids, sep_token_id } = this.tokenizer;
    const batchedResults = [];
    for (let j = 0; j < start_logits.dims[0]; ++j) {
      const ids = input_ids[j];
      const sepIndex = ids.findIndex(
        (x) => (
          // We use == to match bigint with number
          // @ts-ignore
          x == sep_token_id
        )
      );
      const start = start_logits[j].tolist();
      const end = end_logits[j].tolist();
      for (let i = 1; i < start.length; ++i) {
        if (attention_mask[j] == 0 || // is part of padding
        i <= sepIndex || // is before the sep_token
        all_special_ids.findIndex((x) => x == ids[i]) !== -1) {
          start[i] = -Infinity;
          end[i] = -Infinity;
        }
      }
      const start_scores = softmax(start).map((x, i) => [x, i]);
      const end_scores = softmax(end).map((x, i) => [x, i]);
      start_scores[0][0] = 0;
      end_scores[0][0] = 0;
      const options = product(start_scores, end_scores).filter((x) => x[0][1] <= x[1][1]).map((x) => [x[0][1], x[1][1], x[0][0] * x[1][0]]).sort((a, b) => b[2] - a[2]);
      const sampleResults = [];
      for (let k2 = 0; k2 < Math.min(options.length, top_k); ++k2) {
        const [start2, end2, score] = options[k2];
        const answer_tokens = ids.slice(start2, end2 + 1);
        const answer = this.tokenizer.decode(answer_tokens, {
          skip_special_tokens: true
        });
        sampleResults.push({
          answer,
          score
        });
      }
      if (top_k === 1) {
        batchedResults.push(...sampleResults);
      } else {
        batchedResults.push(sampleResults);
      }
    }
    return isBatched ? batchedResults : batchedResults[0];
  }
};

// src/pipelines/fill-mask.js
var FillMaskPipeline = class extends /** @type {new (options: TextPipelineConstructorArgs) => FillMaskPipelineType} */
Pipeline {
  async _call(texts, { top_k = 5 } = {}) {
    const { mask_token_id, mask_token } = this.tokenizer;
    const model_inputs = this.tokenizer(texts, {
      padding: true,
      truncation: true
    });
    const { logits } = await this.model(model_inputs);
    const toReturn = [];
    const input_ids = model_inputs.input_ids.tolist();
    for (let i = 0; i < input_ids.length; ++i) {
      const ids = input_ids[i];
      const mask_token_index = ids.findIndex(
        (x) => (
          // We use == to match bigint with number
          // @ts-ignore - TS2367: Intentional loose equality for bigint/number comparison
          x == mask_token_id
        )
      );
      if (mask_token_index === -1) {
        throw Error(`Mask token (${mask_token}) not found in text.`);
      }
      const itemLogits = logits[i][mask_token_index];
      const scores = await topk(new Tensor2("float32", softmax(itemLogits.data), itemLogits.dims), top_k);
      const values = scores[0].tolist();
      const indices = scores[1].tolist();
      toReturn.push(
        indices.map((x, i2) => {
          const sequence = ids.slice();
          sequence[mask_token_index] = x;
          return {
            score: values[i2],
            token: Number(x),
            token_str: this.tokenizer.decode([x]),
            sequence: this.tokenizer.decode(sequence, {
              skip_special_tokens: true
            })
          };
        })
      );
    }
    return Array.isArray(texts) ? toReturn : toReturn[0];
  }
};

// src/pipelines/text2text-generation.js
var Text2TextGenerationPipeline = class extends /** @type {new (options: TextPipelineConstructorArgs) => Text2TextGenerationPipelineType} */
Pipeline {
  _default_generation_config = {
    max_new_tokens: 256
    // do_sample: true,
    // temperature: 0.7,
  };
  /** @type {'generated_text'} */
  _key = "generated_text";
  /** @type {Text2TextGenerationPipelineCallback} */
  async _call(texts, generate_kwargs = {}) {
    if (!Array.isArray(texts)) {
      texts = [texts];
    }
    if (this.model.config.prefix) {
      texts = texts.map((x) => this.model.config.prefix + x);
    }
    const task_specific_params = this.model.config.task_specific_params;
    if (task_specific_params && task_specific_params[this.task]) {
      if (task_specific_params[this.task].prefix) {
        texts = texts.map((x) => task_specific_params[this.task].prefix + x);
      }
    }
    const tokenizer = this.tokenizer;
    const tokenizer_options = {
      padding: true,
      truncation: true
    };
    let inputs;
    if (this.task === "translation" && "_build_translation_inputs" in tokenizer) {
      inputs = tokenizer._build_translation_inputs(texts, tokenizer_options, generate_kwargs);
    } else {
      inputs = tokenizer(texts, tokenizer_options);
    }
    const outputTokenIds = await this.model.generate({
      ...inputs,
      ...this._default_generation_config,
      ...generate_kwargs
    });
    return tokenizer.batch_decode(
      /** @type {Tensor} */
      outputTokenIds,
      {
        skip_special_tokens: true
      }
    ).map((text) => ({ [this._key]: text }));
  }
};

// src/pipelines/summarization.js
var SummarizationPipeline = class extends /** @type {new (options: TextPipelineConstructorArgs) => SummarizationPipelineType} */
/** @type {any} */
Text2TextGenerationPipeline {
  /** @type {'summary_text'} */
  _key = "summary_text";
};

// src/pipelines/translation.js
var TranslationPipeline = class extends /** @type {new (options: TextPipelineConstructorArgs) => TranslationPipelineType} */
/** @type {any} */
Text2TextGenerationPipeline {
  /** @type {'translation_text'} */
  _key = "translation_text";
};

// src/pipelines/text-generation.js
function isChat(x) {
  return Array.isArray(x) && x.every((x2) => "role" in x2 && "content" in x2);
}
var TextGenerationPipeline = class extends /** @type {new (options: TextPipelineConstructorArgs) => TextGenerationPipelineType} */
Pipeline {
  _default_generation_config = {
    max_new_tokens: 256
    // do_sample: true,
    // temperature: 0.7,
  };
  /**
   * @param {string | string[] | import('../tokenization_utils.js').Message[] | import('../tokenization_utils.js').Message[][]} texts
   * @param {Partial<TextGenerationConfig>} generate_kwargs
   */
  async _call(texts, generate_kwargs = {}) {
    const {
      add_special_tokens: add_special_tokens_arg,
      return_full_text: return_full_text_arg,
      tools,
      documents,
      chat_template,
      tokenizer_encode_kwargs,
      ...generation_kwargs
    } = generate_kwargs;
    let isBatched = false;
    let isChatInput = false;
    let add_special_tokens = add_special_tokens_arg ?? (this.tokenizer.add_bos_token || this.tokenizer.add_eos_token) ?? false;
    let tokenizer_kwargs = tokenizer_encode_kwargs;
    let inputs;
    if (typeof texts === "string") {
      inputs = texts = [texts];
    } else if (Array.isArray(texts) && texts.every((x) => typeof x === "string")) {
      isBatched = true;
      inputs = /** @type {string[]} */
      texts;
    } else {
      if (isChat(texts)) {
        texts = [
          /** @type {Chat} */
          texts
        ];
      } else if (Array.isArray(texts) && texts.every(isChat)) {
        isBatched = true;
      } else {
        throw new Error("Input must be a string, an array of strings, a Chat, or an array of Chats");
      }
      isChatInput = true;
      const chat_template_kwargs = {
        tokenize: false,
        add_generation_prompt: true,
        ...pick({ tools, documents, chat_template }, ["tools", "documents", "chat_template"]),
        ...tokenizer_kwargs
      };
      inputs = /** @type {string[]} */
      /** @type {Chat[]} */
      texts.map(
        (x) => (
          /** @type {string} */
          /** @type {unknown} */
          this.tokenizer.apply_chat_template(x, chat_template_kwargs)
        )
      );
      add_special_tokens = false;
      tokenizer_kwargs = void 0;
    }
    const return_full_text = isChatInput ? false : return_full_text_arg ?? true;
    this.tokenizer.padding_side = "left";
    const text_inputs = this.tokenizer(inputs, {
      add_special_tokens,
      padding: true,
      truncation: true,
      ...tokenizer_kwargs
    });
    const outputTokenIds = (
      /** @type {Tensor} */
      await this.model.generate({
        ...text_inputs,
        ...this._default_generation_config,
        ...generation_kwargs
      })
    );
    const decoded = this.tokenizer.batch_decode(outputTokenIds, {
      skip_special_tokens: true
    });
    let promptLengths;
    if (!return_full_text && text_inputs.input_ids.dims.at(-1) > 0) {
      promptLengths = this.tokenizer.batch_decode(text_inputs.input_ids, {
        skip_special_tokens: true
      }).map((x) => x.length);
    }
    const toReturn = Array.from({ length: texts.length }, (_) => []);
    for (let i = 0; i < decoded.length; ++i) {
      const textIndex = Math.floor(i / outputTokenIds.dims[0] * texts.length);
      if (promptLengths) {
        decoded[i] = decoded[i].slice(promptLengths[textIndex]);
      }
      toReturn[textIndex].push(
        /** @type {TextGenerationSingle} */
        {
          generated_text: isChatInput ? [.../** @type {Chat[]} */
          texts[textIndex], { role: "assistant", content: decoded[i] }] : decoded[i]
        }
      );
    }
    return !isBatched && toReturn.length === 1 ? toReturn[0] : toReturn;
  }
};

// src/pipelines/zero-shot-classification.js
var ZeroShotClassificationPipeline = class extends /** @type {new (options: TextPipelineConstructorArgs) => ZeroShotClassificationPipelineType} */
Pipeline {
  /**
   * Create a new ZeroShotClassificationPipeline.
   * @param {TextPipelineConstructorArgs} options An object used to instantiate the pipeline.
   */
  constructor(options) {
    super(options);
    this.label2id = Object.fromEntries(
      Object.entries(
        /** @type {any} */
        this.model.config.label2id
      ).map(([k2, v]) => [k2.toLowerCase(), v])
    );
    this.entailment_id = this.label2id["entailment"];
    if (this.entailment_id === void 0) {
      logger.warn("Could not find 'entailment' in label2id mapping. Using 2 as entailment_id.");
      this.entailment_id = 2;
    }
    this.contradiction_id = this.label2id["contradiction"] ?? this.label2id["not_entailment"];
    if (this.contradiction_id === void 0) {
      logger.warn("Could not find 'contradiction' in label2id mapping. Using 0 as contradiction_id.");
      this.contradiction_id = 0;
    }
  }
  async _call(texts, candidate_labels, { hypothesis_template = "This example is {}.", multi_label = false } = {}) {
    const isBatched = Array.isArray(texts);
    if (!isBatched) {
      texts = [
        /** @type {string} */
        texts
      ];
    }
    if (!Array.isArray(candidate_labels)) {
      candidate_labels = [candidate_labels];
    }
    const hypotheses = candidate_labels.map((x) => hypothesis_template.replace("{}", x));
    const softmaxEach = multi_label || candidate_labels.length === 1;
    const toReturn = [];
    for (const premise of texts) {
      const entails_logits = [];
      for (const hypothesis of hypotheses) {
        const inputs = this.tokenizer(premise, {
          text_pair: hypothesis,
          padding: true,
          truncation: true
        });
        const outputs = await this.model(inputs);
        if (softmaxEach) {
          entails_logits.push([
            outputs.logits.data[this.contradiction_id],
            outputs.logits.data[this.entailment_id]
          ]);
        } else {
          entails_logits.push(outputs.logits.data[this.entailment_id]);
        }
      }
      const scores = softmaxEach ? entails_logits.map((x) => softmax(x)[1]) : softmax(entails_logits);
      const scores_sorted = scores.map((x, i) => [x, i]).sort((a, b) => b[0] - a[0]);
      toReturn.push({
        sequence: premise,
        labels: scores_sorted.map((x) => candidate_labels[x[1]]),
        scores: scores_sorted.map((x) => x[0])
      });
    }
    return isBatched ? toReturn : toReturn[0];
  }
};

// src/pipelines/audio-classification.js
var AudioClassificationPipeline = class extends /** @type {new (options: AudioPipelineConstructorArgs) => AudioClassificationPipelineType} */
Pipeline {
  async _call(audio, { top_k = 5 } = {}) {
    const sampling_rate = this.processor.feature_extractor.config.sampling_rate;
    const preparedAudios = await prepareAudios(audio, sampling_rate);
    const id2label = this.model.config.id2label;
    const toReturn = [];
    for (const aud of preparedAudios) {
      const inputs = await this.processor(aud);
      const output = await this.model(inputs);
      const logits = output.logits[0];
      const scores = await topk(new Tensor2("float32", softmax(logits.data), logits.dims), top_k);
      const values = scores[0].tolist();
      const indices = scores[1].tolist();
      const vals = indices.map((x, i) => ({
        label: (
          /** @type {string} */
          id2label ? id2label[x] : `LABEL_${x}`
        ),
        score: (
          /** @type {number} */
          values[i]
        )
      }));
      toReturn.push(vals);
    }
    return Array.isArray(audio) ? toReturn : toReturn[0];
  }
};

// src/pipelines/zero-shot-audio-classification.js
var ZeroShotAudioClassificationPipeline = class extends /** @type {new (options: TextAudioPipelineConstructorArgs) => ZeroShotAudioClassificationPipelineType} */
Pipeline {
  async _call(audio, candidate_labels, { hypothesis_template = "This is a sound of {}." } = {}) {
    const single = !Array.isArray(audio);
    if (single) {
      audio = [
        /** @type {AudioInput} */
        audio
      ];
    }
    const texts = candidate_labels.map((x) => hypothesis_template.replace("{}", x));
    const text_inputs = this.tokenizer(texts, {
      padding: true,
      truncation: true
    });
    const sampling_rate = this.processor.feature_extractor.config.sampling_rate;
    const preparedAudios = await prepareAudios(audio, sampling_rate);
    const toReturn = [];
    for (const aud of preparedAudios) {
      const audio_inputs = await this.processor(aud);
      const output = await this.model({ ...text_inputs, ...audio_inputs });
      const probs = softmax(output.logits_per_audio.data);
      toReturn.push(
        [...probs].map((x, i) => ({
          score: x,
          label: candidate_labels[i]
        }))
      );
    }
    return single ? toReturn[0] : toReturn;
  }
};

// src/pipelines/automatic-speech-recognition.js
var AutomaticSpeechRecognitionPipeline = class extends /** @type {new (options: TextAudioPipelineConstructorArgs) => AutomaticSpeechRecognitionPipelineType} */
Pipeline {
  _default_generation_config = {
    // TODO: figure out good defaults for ASR generation parameters
    // max_new_tokens: 256,
    // num_beams: 5,
  };
  async _call(audio, kwargs = {}) {
    kwargs = {
      ...this._default_generation_config,
      ...kwargs
    };
    switch (this.model.config.model_type) {
      case "whisper":
      case "lite-whisper":
        return this._call_whisper(audio, kwargs);
      case "wav2vec2":
      case "wav2vec2-bert":
      case "unispeech":
      case "unispeech-sat":
      case "hubert":
      case "parakeet_ctc":
        return this._call_wav2vec2(audio, kwargs);
      case "moonshine":
        return this._call_moonshine(audio, kwargs);
      case "cohere_asr":
        return this._call_cohere_asr(audio, kwargs);
      default:
        throw new Error(
          `AutomaticSpeechRecognitionPipeline does not support model type '${this.model.config.model_type}'.`
        );
    }
  }
  async _call_wav2vec2(audio, kwargs) {
    if (kwargs.language) {
      logger.warn('`language` parameter is not yet supported for `wav2vec2` models, defaulting to "English".');
    }
    if (kwargs.task) {
      logger.warn('`task` parameter is not yet supported for `wav2vec2` models, defaulting to "transcribe".');
    }
    const single = !Array.isArray(audio);
    const batchedAudio = single ? [audio] : audio;
    const sampling_rate = this.processor.feature_extractor.config.sampling_rate;
    const preparedAudios = await prepareAudios(batchedAudio, sampling_rate);
    const toReturn = [];
    for (const aud of preparedAudios) {
      const inputs = await this.processor(aud);
      const output = await this.model(inputs);
      const logits = output.logits[0];
      const predicted_ids = [];
      for (const item of logits) {
        predicted_ids.push(max(item.data)[1]);
      }
      const predicted_sentences = this.tokenizer.decode(predicted_ids, { skip_special_tokens: true }).trim();
      toReturn.push({ text: predicted_sentences });
    }
    return single ? toReturn[0] : toReturn;
  }
  async _call_whisper(audio, kwargs) {
    const return_timestamps = kwargs.return_timestamps ?? false;
    const chunk_length_s = kwargs.chunk_length_s ?? 0;
    const force_full_sequences = kwargs.force_full_sequences ?? false;
    let stride_length_s = kwargs.stride_length_s ?? null;
    const generation_config = { ...kwargs };
    if (return_timestamps === "word") {
      generation_config["return_token_timestamps"] = true;
      generation_config["return_timestamps"] = true;
    }
    const single = !Array.isArray(audio);
    const batchedAudio = single ? [audio] : audio;
    const feature_extractor_config = this.processor.feature_extractor.config;
    const time_precision = feature_extractor_config.chunk_length / this.model.config.max_source_positions;
    const hop_length = feature_extractor_config.hop_length;
    const sampling_rate = feature_extractor_config.sampling_rate;
    const preparedAudios = await prepareAudios(batchedAudio, sampling_rate);
    const toReturn = [];
    for (const aud of preparedAudios) {
      let chunks = [];
      if (chunk_length_s > 0) {
        if (stride_length_s === null) {
          stride_length_s = chunk_length_s / 6;
        } else if (chunk_length_s <= stride_length_s) {
          throw Error("`chunk_length_s` must be larger than `stride_length_s`.");
        }
        const window2 = sampling_rate * chunk_length_s;
        const stride = sampling_rate * stride_length_s;
        const jump = window2 - 2 * stride;
        let offset = 0;
        while (true) {
          const offset_end = offset + window2;
          const subarr = aud.subarray(offset, offset_end);
          const feature = await this.processor(subarr);
          const is_first = offset === 0;
          const is_last = offset_end >= aud.length;
          chunks.push({
            stride: [subarr.length, is_first ? 0 : stride, is_last ? 0 : stride],
            input_features: feature.input_features,
            is_last
          });
          if (is_last) break;
          offset += jump;
        }
      } else {
        chunks = [
          {
            stride: [aud.length, 0, 0],
            input_features: (await this.processor(aud)).input_features,
            is_last: true
          }
        ];
      }
      for (const chunk2 of chunks) {
        generation_config.num_frames = Math.floor(chunk2.stride[0] / hop_length);
        const data = await this.model.generate({
          inputs: chunk2.input_features,
          ...generation_config
        });
        if (return_timestamps === "word") {
          const sequences = data.sequences.tolist()[0];
          const token_ts = data.token_timestamps.tolist()[0];
          const timestamp_begin = this.tokenizer.timestamp_begin;
          const prefixLength = Math.max(
            sequences.findIndex((t) => Number(t) >= timestamp_begin),
            0
          );
          chunk2.tokens = sequences.slice(prefixLength);
          chunk2.token_timestamps = token_ts.slice(prefixLength).map((x) => round(x, 2));
        } else {
          chunk2.tokens = /** @type {Tensor} */
          data[0].tolist();
        }
        chunk2.stride = chunk2.stride.map((x) => x / sampling_rate);
      }
      const [full_text, optional] = this.tokenizer._decode_asr(chunks, {
        time_precision,
        return_timestamps,
        force_full_sequences
      });
      toReturn.push({ text: full_text, ...optional });
    }
    return single ? toReturn[0] : toReturn;
  }
  async _call_moonshine(audio, kwargs) {
    const single = !Array.isArray(audio);
    const batchedAudio = single ? [audio] : audio;
    const sampling_rate = this.processor.feature_extractor.config.sampling_rate;
    const preparedAudios = await prepareAudios(batchedAudio, sampling_rate);
    const toReturn = [];
    for (const aud of preparedAudios) {
      const inputs = await this.processor(aud);
      const max_new_tokens = Math.floor(aud.length / sampling_rate) * 6;
      const outputs = await this.model.generate({ max_new_tokens, ...kwargs, ...inputs });
      const text = this.processor.batch_decode(
        /** @type {Tensor} */
        outputs,
        { skip_special_tokens: true }
      )[0];
      toReturn.push({ text });
    }
    return single ? toReturn[0] : toReturn;
  }
  async _call_cohere_asr(audio, kwargs) {
    const single = !Array.isArray(audio);
    const batchedAudio = single ? [audio] : audio;
    const feature_extractor = this.processor.feature_extractor;
    const sampling_rate = feature_extractor.config.sampling_rate;
    const preparedAudios = await prepareAudios(batchedAudio, sampling_rate);
    const language = kwargs.language ?? "en";
    const decoder_input_ids = this.processor.get_decoder_prompt_ids(language);
    const toReturn = [];
    for (const aud of preparedAudios) {
      const audioChunks = feature_extractor.split_audio(aud);
      const chunk_texts = [];
      for (const chunk2 of audioChunks) {
        const inputs = await this.processor(chunk2);
        const outputs = await this.model.generate({
          ...inputs,
          decoder_input_ids,
          ...kwargs
        });
        const text = this.tokenizer.decode(
          /** @type {Tensor} */
          outputs[0].tolist(),
          { skip_special_tokens: true }
        ).trim();
        chunk_texts.push(text);
      }
      const full_text = this.processor.constructor.join_chunks(chunk_texts, language);
      toReturn.push({ text: full_text });
    }
    return single ? toReturn[0] : toReturn;
  }
};

// src/pipelines/text-to-audio.js
var TextToAudioPipeline = class extends /** @type {new (options: TextToAudioPipelineConstructorArgs) => TextToAudioPipelineType} */
Pipeline {
  DEFAULT_VOCODER_ID = "Xenova/speecht5_hifigan";
  /**
   * Create a new TextToAudioPipeline.
   * @param {TextToAudioPipelineConstructorArgs} options An object used to instantiate the pipeline.
   */
  constructor(options) {
    super(options);
    this.vocoder = options.vocoder ?? null;
  }
  async _prepare_speaker_embeddings(speaker_embeddings, batch_size) {
    if (typeof speaker_embeddings === "string" || speaker_embeddings instanceof URL) {
      speaker_embeddings = new Float32Array(await (await env.fetch(speaker_embeddings)).arrayBuffer());
    }
    if (speaker_embeddings instanceof Float32Array) {
      speaker_embeddings = new Tensor2("float32", speaker_embeddings, [speaker_embeddings.length]);
    } else if (!(speaker_embeddings instanceof Tensor2)) {
      throw new Error("Speaker embeddings must be a `Tensor`, `Float32Array`, `string`, or `URL`.");
    }
    if (batch_size > 1) {
      if (speaker_embeddings.dims[0] === 1) {
        speaker_embeddings = speaker_embeddings.repeat(batch_size, 1);
      } else if (speaker_embeddings.dims[0] !== batch_size) {
        throw new Error(
          `Expected speaker embeddings batch size to be 1 or ${batch_size}, but got ${speaker_embeddings.dims[0]}.`
        );
      }
    }
    return speaker_embeddings;
  }
  /**
   * Helper to convert batched waveform tensor to RawAudio output(s).
   * @param {string|string[]} text_inputs Original text input(s) to determine return type.
   * @param {Tensor} waveform The waveform tensor of shape [batch_size, waveform_length].
   * @param {number} sampling_rate The audio sampling rate.
   * @param {Tensor} [durations] Optional durations tensor for trimming (used by Supertonic).
   * @returns {RawAudio|RawAudio[]} Single RawAudio or array based on input type.
   * @private
   */
  _postprocess_waveform(text_inputs, waveform, sampling_rate, durations = null) {
    const waveformData = (
      /** @type {Float32Array} */
      waveform.data
    );
    const [batch_size, waveformLength] = waveform.dims;
    const durationsData = durations ? (
      /** @type {Float32Array} */
      durations.data
    ) : null;
    const results = [];
    for (let i = 0; i < batch_size; ++i) {
      const length = durationsData ? Math.min(Math.ceil(durationsData[i]), waveformLength) : waveformLength;
      const start = i * waveformLength;
      results.push(new RawAudio(waveformData.slice(start, start + length), sampling_rate));
    }
    return Array.isArray(text_inputs) ? results : results[0];
  }
  async _call(text_inputs, options) {
    if (this.processor) {
      return this._call_text_to_spectrogram(text_inputs, options);
    } else if (this.model.config.model_type === "supertonic") {
      return this._call_supertonic(text_inputs, options);
    } else {
      return this._call_text_to_waveform(text_inputs);
    }
  }
  async _call_supertonic(text_inputs, { speaker_embeddings, num_inference_steps, speed }) {
    if (!speaker_embeddings) {
      throw new Error("Speaker embeddings must be provided for Supertonic models.");
    }
    const { sampling_rate, style_dim } = this.model.config;
    const inputs = this.tokenizer(text_inputs, {
      padding: true,
      truncation: true
    });
    const batch_size = inputs.input_ids.dims[0];
    speaker_embeddings = await this._prepare_speaker_embeddings(speaker_embeddings, batch_size);
    speaker_embeddings = /** @type {Tensor} */
    speaker_embeddings.view(batch_size, -1, style_dim);
    const { waveform, durations } = await this.model.generate_speech({
      ...inputs,
      style: speaker_embeddings,
      num_inference_steps,
      speed
    });
    return this._postprocess_waveform(text_inputs, waveform, sampling_rate, durations);
  }
  async _call_text_to_waveform(text_inputs) {
    const inputs = this.tokenizer(text_inputs, {
      padding: true,
      truncation: true
    });
    const { waveform } = await this.model(inputs);
    const sampling_rate = this.model.config.sampling_rate;
    return this._postprocess_waveform(text_inputs, waveform, sampling_rate);
  }
  async _call_text_to_spectrogram(text_inputs, { speaker_embeddings }) {
    if (!this.vocoder) {
      logger.info("No vocoder specified, using default HifiGan vocoder.");
      this.vocoder = await AutoModel.from_pretrained(this.DEFAULT_VOCODER_ID, { dtype: "fp32" });
    }
    const { input_ids } = this.tokenizer(text_inputs, {
      padding: true,
      truncation: true
    });
    const batch_size = input_ids.dims[0];
    speaker_embeddings = await this._prepare_speaker_embeddings(speaker_embeddings, batch_size);
    speaker_embeddings = speaker_embeddings.view(batch_size, -1);
    const { waveform } = await this.model.generate_speech(input_ids, speaker_embeddings, { vocoder: this.vocoder });
    const sampling_rate = this.processor.feature_extractor.config.sampling_rate;
    return this._postprocess_waveform(text_inputs, waveform, sampling_rate);
  }
};

// src/pipelines/image-to-text.js
var ImageToTextPipeline = class extends /** @type {new (options: TextImagePipelineConstructorArgs) => ImageToTextPipelineType} */
Pipeline {
  async _call(images, generate_kwargs = {}) {
    const isBatched = Array.isArray(images);
    const preparedImages = await prepareImages(images);
    const { pixel_values } = await this.processor(preparedImages);
    const toReturn = [];
    for (const batch of pixel_values) {
      batch.dims = [1, ...batch.dims];
      const output = await this.model.generate({ inputs: batch, ...generate_kwargs });
      const decoded = this.tokenizer.batch_decode(
        /** @type {Tensor} */
        output,
        {
          skip_special_tokens: true
        }
      ).map((x) => ({ generated_text: x.trim() }));
      toReturn.push(decoded);
    }
    return isBatched ? toReturn : toReturn[0];
  }
};

// src/pipelines/image-classification.js
var ImageClassificationPipeline = class extends /** @type {new (options: ImagePipelineConstructorArgs) => ImageClassificationPipelineType} */
Pipeline {
  async _call(images, { top_k = 5 } = {}) {
    const preparedImages = await prepareImages(images);
    const { pixel_values } = await this.processor(preparedImages);
    const output = await this.model({ pixel_values });
    const { id2label } = this.model.config;
    const toReturn = [];
    for (const batch of output.logits) {
      const scores = await topk(new Tensor2("float32", softmax(batch.data), batch.dims), top_k);
      const values = scores[0].tolist();
      const indices = scores[1].tolist();
      const vals = indices.map((x, i) => ({
        label: (
          /** @type {string} */
          id2label ? id2label[x] : `LABEL_${x}`
        ),
        score: (
          /** @type {number} */
          values[i]
        )
      }));
      toReturn.push(vals);
    }
    return Array.isArray(images) ? toReturn : toReturn[0];
  }
};

// src/pipelines/image-segmentation.js
var SUBTASKS_MAPPING = {
  // Mapping of subtasks to their corresponding post-processing function names.
  panoptic: "post_process_panoptic_segmentation",
  instance: "post_process_instance_segmentation",
  semantic: "post_process_semantic_segmentation"
};
var ImageSegmentationPipeline = class extends /** @type {new (options: ImagePipelineConstructorArgs) => ImageSegmentationPipelineType} */
Pipeline {
  /** @type {ImageSegmentationPipelineCallback} */
  async _call(images, {
    threshold = 0.5,
    mask_threshold = 0.5,
    overlap_mask_area_threshold = 0.8,
    label_ids_to_fuse = null,
    target_sizes = null,
    subtask = null
  } = {}) {
    const isBatched = Array.isArray(images);
    if (isBatched && images.length !== 1) {
      throw Error("Image segmentation pipeline currently only supports a batch size of 1.");
    }
    const preparedImages = await prepareImages(images);
    const imageSizes = preparedImages.map((x) => [x.height, x.width]);
    const inputs = await this.processor(preparedImages);
    const { inputNames, outputNames } = this.model.sessions["model"];
    if (!inputNames.includes("pixel_values")) {
      if (inputNames.length !== 1) {
        throw Error(`Expected a single input name, but got ${inputNames.length} inputs: ${inputNames}.`);
      }
      const newName = inputNames[0];
      if (newName in inputs) {
        throw Error(`Input name ${newName} already exists in the inputs.`);
      }
      inputs[newName] = inputs.pixel_values;
    }
    const output = await this.model(inputs);
    let fn2 = null;
    if (subtask !== null) {
      fn2 = SUBTASKS_MAPPING[subtask];
    } else if (this.processor.image_processor) {
      for (const [task, func] of Object.entries(SUBTASKS_MAPPING)) {
        if (func in this.processor.image_processor) {
          fn2 = this.processor.image_processor[func].bind(this.processor.image_processor);
          subtask = task;
          break;
        }
      }
    }
    const id2label = this.model.config.id2label;
    const annotation = [];
    if (!subtask) {
      const epsilon = 1e-5;
      const result = output[outputNames[0]];
      for (let i = 0; i < imageSizes.length; ++i) {
        const size = imageSizes[i];
        const item = result[i];
        if (item.data.some((x) => x < -epsilon || x > 1 + epsilon)) {
          item.sigmoid_();
        }
        const mask = await RawImage.fromTensor(item.mul_(255).to("uint8")).resize(size[1], size[0]);
        annotation.push({
          label: null,
          score: null,
          mask
        });
      }
    } else if (subtask === "panoptic" || subtask === "instance") {
      const processed = fn2(
        output,
        threshold,
        mask_threshold,
        overlap_mask_area_threshold,
        label_ids_to_fuse,
        target_sizes ?? imageSizes
        // TODO FIX?
      )[0];
      const segmentation = processed.segmentation;
      for (const segment of processed.segments_info) {
        const maskData = new Uint8ClampedArray(segmentation.data.length);
        for (let i = 0; i < segmentation.data.length; ++i) {
          if (segmentation.data[i] === segment.id) {
            maskData[i] = 255;
          }
        }
        const mask = new RawImage(maskData, segmentation.dims[1], segmentation.dims[0], 1);
        annotation.push({
          score: segment.score,
          label: id2label[segment.label_id],
          mask
        });
      }
    } else if (subtask === "semantic") {
      const { segmentation, labels } = fn2(output, target_sizes ?? imageSizes)[0];
      for (const label of labels) {
        const maskData = new Uint8ClampedArray(segmentation.data.length);
        for (let i = 0; i < segmentation.data.length; ++i) {
          if (segmentation.data[i] === label) {
            maskData[i] = 255;
          }
        }
        const mask = new RawImage(maskData, segmentation.dims[1], segmentation.dims[0], 1);
        annotation.push({
          score: null,
          label: id2label[label],
          mask
        });
      }
    } else {
      throw Error(`Subtask ${subtask} not supported.`);
    }
    return annotation;
  }
};

// src/pipelines/background-removal.js
var BackgroundRemovalPipeline = class extends /** @type {new (options: ImagePipelineConstructorArgs) => BackgroundRemovalPipelineType} */
/** @type {any} */
ImageSegmentationPipeline {
  async _call(images, options = {}) {
    const preparedImages = await prepareImages(images);
    const masks = await super._call(images, options);
    const result = preparedImages.map((img, i) => {
      const cloned = img.clone();
      cloned.putAlpha(masks[i].mask);
      return cloned;
    });
    return Array.isArray(images) ? result : result[0];
  }
};

// src/pipelines/zero-shot-image-classification.js
var ZeroShotImageClassificationPipeline = class extends /** @type {new (options: TextImagePipelineConstructorArgs) => ZeroShotImageClassificationPipelineType} */
Pipeline {
  async _call(images, candidate_labels, { hypothesis_template = "This is a photo of {}" } = {}) {
    const isBatched = Array.isArray(images);
    const preparedImages = await prepareImages(images);
    const texts = candidate_labels.map((x) => hypothesis_template.replace("{}", x));
    const text_inputs = this.tokenizer(texts, {
      padding: this.model.config.model_type === "siglip" ? "max_length" : true,
      truncation: true
    });
    const { pixel_values } = await this.processor(preparedImages);
    const output = await this.model({ ...text_inputs, pixel_values });
    const function_to_apply = this.model.config.model_type === "siglip" ? (batch) => batch.sigmoid().data : (batch) => softmax(batch.data);
    const toReturn = [];
    for (const batch of output.logits_per_image) {
      const probs = function_to_apply(batch);
      const result = [...probs].map((x, i) => ({
        score: x,
        label: candidate_labels[i]
      }));
      result.sort((a, b) => b.score - a.score);
      toReturn.push(result);
    }
    return isBatched ? toReturn : toReturn[0];
  }
};

// src/pipelines/object-detection.js
var ObjectDetectionPipeline = class extends /** @type {new (options: ImagePipelineConstructorArgs) => ObjectDetectionPipelineType} */
Pipeline {
  async _call(images, { threshold = 0.9, percentage = false } = {}) {
    const isBatched = Array.isArray(images);
    if (isBatched && images.length !== 1) {
      throw Error("Object detection pipeline currently only supports a batch size of 1.");
    }
    const preparedImages = await prepareImages(images);
    const imageSizes = percentage ? null : preparedImages.map((x) => [x.height, x.width]);
    const { pixel_values, pixel_mask } = await this.processor(preparedImages);
    const output = await this.model({ pixel_values, pixel_mask });
    const processed = this.processor.image_processor.post_process_object_detection(output, threshold, imageSizes);
    const { id2label } = this.model.config;
    const result = processed.map(
      (batch) => batch.boxes.map((box, i) => ({
        score: batch.scores[i],
        label: id2label[batch.classes[i]],
        box: get_bounding_box(box, !percentage)
      }))
    );
    return isBatched ? result : result[0];
  }
};

// src/pipelines/zero-shot-object-detection.js
var ZeroShotObjectDetectionPipeline = class extends /** @type {new (options: TextImagePipelineConstructorArgs) => ZeroShotObjectDetectionPipelineType} */
Pipeline {
  async _call(images, candidate_labels, { threshold = 0.1, top_k = null, percentage = false } = {}) {
    const isBatched = Array.isArray(images);
    const preparedImages = await prepareImages(images);
    const text_inputs = this.tokenizer(candidate_labels, {
      padding: true,
      truncation: true
    });
    const model_inputs = await this.processor(preparedImages);
    const toReturn = [];
    for (let i = 0; i < preparedImages.length; ++i) {
      const image = preparedImages[i];
      const imageSize = percentage ? null : [[image.height, image.width]];
      const pixel_values = model_inputs.pixel_values[i].unsqueeze_(0);
      const output = await this.model({ ...text_inputs, pixel_values });
      let result;
      if ("post_process_grounded_object_detection" in this.processor) {
        const processed = this.processor.post_process_grounded_object_detection(output, text_inputs.input_ids, {
          // TODO: support separate threshold values
          box_threshold: threshold,
          text_threshold: threshold,
          target_sizes: imageSize
        })[0];
        result = processed.boxes.map((box, i2) => ({
          score: processed.scores[i2],
          label: processed.labels[i2],
          box: get_bounding_box(box, !percentage)
        }));
      } else {
        const processed = this.processor.image_processor.post_process_object_detection(
          output,
          threshold,
          imageSize,
          true
        )[0];
        result = processed.boxes.map((box, i2) => ({
          score: processed.scores[i2],
          label: candidate_labels[processed.classes[i2]],
          box: get_bounding_box(box, !percentage)
        }));
      }
      result.sort((a, b) => b.score - a.score);
      if (top_k !== null) {
        result = result.slice(0, top_k);
      }
      toReturn.push(result);
    }
    return isBatched ? toReturn : toReturn[0];
  }
};

// src/pipelines/document-question-answering.js
var DocumentQuestionAnsweringPipeline = class extends /** @type {new (options: TextImagePipelineConstructorArgs) => DocumentQuestionAnsweringPipelineType} */
Pipeline {
  _default_generation_config = {
    max_new_tokens: 256
  };
  async _call(image, question, generate_kwargs = {}) {
    if (Array.isArray(image)) {
      if (image.length !== 1) {
        throw Error("Document Question Answering pipeline currently only supports a batch size of 1.");
      }
      image = image[0];
    }
    const preparedImage = (await prepareImages(image))[0];
    const { pixel_values } = await this.processor(preparedImage);
    const task_prompt = `<s_docvqa><s_question>${question}</s_question><s_answer>`;
    const decoder_input_ids = this.tokenizer(task_prompt, {
      add_special_tokens: false,
      padding: true,
      truncation: true
    }).input_ids;
    const output = await this.model.generate({
      inputs: pixel_values,
      // @ts-expect-error Ts2339
      max_length: this.model.config.decoder.max_position_embeddings,
      decoder_input_ids,
      ...this._default_generation_config,
      ...generate_kwargs
    });
    const decoded = this.tokenizer.batch_decode(
      /** @type {Tensor} */
      output
    )[0];
    const match = decoded.match(/<s_answer>(.*?)<\/s_answer>/);
    let answer = null;
    if (match && match.length >= 2) {
      answer = match[1].trim();
    }
    return [{ answer }];
  }
};

// src/pipelines/image-to-image.js
var ImageToImagePipeline = class extends /** @type {new (options: ImagePipelineConstructorArgs) => ImageToImagePipelineType} */
Pipeline {
  async _call(images) {
    const preparedImages = await prepareImages(images);
    const inputs = await this.processor(preparedImages);
    const outputs = await this.model(inputs);
    const toReturn = [];
    for (const batch of outputs.reconstruction) {
      const output = batch.squeeze().clamp_(0, 1).mul_(255).round_().to("uint8");
      toReturn.push(RawImage.fromTensor(output));
    }
    return Array.isArray(images) ? toReturn : toReturn[0];
  }
};

// src/pipelines/depth-estimation.js
var DepthEstimationPipeline = class extends /** @type {new (options: ImagePipelineConstructorArgs) => DepthEstimationPipelineType} */
Pipeline {
  async _call(images) {
    const preparedImages = await prepareImages(images);
    const inputs = await this.processor(preparedImages);
    const { predicted_depth } = await this.model(inputs);
    const toReturn = [];
    for (let i = 0; i < preparedImages.length; ++i) {
      const batch = predicted_depth[i];
      const [height, width] = batch.dims.slice(-2);
      const [new_width, new_height] = preparedImages[i].size;
      const prediction = (await interpolate_4d(batch.view(1, 1, height, width), {
        size: [new_height, new_width],
        mode: "bilinear"
      })).view(new_height, new_width);
      const minval = (
        /** @type {number} */
        prediction.min().item()
      );
      const maxval = (
        /** @type {number} */
        prediction.max().item()
      );
      const formatted = prediction.sub(minval).div_(maxval - minval).mul_(255).to("uint8").unsqueeze(0);
      const depth = RawImage.fromTensor(formatted);
      toReturn.push({
        predicted_depth: prediction,
        depth
      });
    }
    return Array.isArray(images) ? toReturn : toReturn[0];
  }
};

// src/pipelines/feature-extraction.js
var FeatureExtractionPipeline = class extends /** @type {new (options: TextPipelineConstructorArgs) => FeatureExtractionPipelineType} */
Pipeline {
  /** @type {FeatureExtractionPipelineCallback} */
  async _call(texts, {
    pooling = (
      /** @type {'none'} */
      "none"
    ),
    normalize = false,
    quantize = false,
    precision = (
      /** @type {'binary'} */
      "binary"
    )
  } = {}) {
    const model_inputs = this.tokenizer(texts, {
      padding: true,
      truncation: true
    });
    const outputs = await this.model(model_inputs);
    let result = outputs.last_hidden_state ?? outputs.logits ?? outputs.token_embeddings;
    switch (pooling) {
      case "none":
        break;
      case "mean":
        result = mean_pooling(result, model_inputs.attention_mask);
        break;
      case "first_token":
      case "cls":
        result = result.slice(null, 0);
        break;
      case "last_token":
      case "eos":
        result = result.slice(null, -1);
        break;
      default:
        throw Error(`Pooling method '${pooling}' not supported.`);
    }
    if (normalize) {
      result = result.normalize(2, -1);
    }
    if (quantize) {
      result = quantize_embeddings(result, precision);
    }
    return result;
  }
};

// src/pipelines/image-feature-extraction.js
var ImageFeatureExtractionPipeline = class extends /** @type {new (options: ImagePipelineConstructorArgs) => ImageFeatureExtractionPipelineType} */
Pipeline {
  /** @type {ImageFeatureExtractionPipelineCallback} */
  async _call(images, { pool = null } = {}) {
    const preparedImages = await prepareImages(images);
    const { pixel_values } = await this.processor(preparedImages);
    const outputs = await this.model({ pixel_values });
    let result;
    if (pool) {
      if (!("pooler_output" in outputs)) {
        throw Error(
          `No pooled output was returned. Make sure the model has a 'pooler' layer when using the 'pool' option.`
        );
      }
      result = outputs.pooler_output;
    } else {
      result = outputs.last_hidden_state ?? outputs.logits ?? outputs.image_embeds;
    }
    return result;
  }
};

// src/pipelines/index.js
var SUPPORTED_TASKS = Object.freeze({
  "text-classification": {
    pipeline: TextClassificationPipeline,
    model: AutoModelForSequenceClassification,
    default: {
      model: "Xenova/distilbert-base-uncased-finetuned-sst-2-english"
    },
    type: "text"
  },
  "token-classification": {
    pipeline: TokenClassificationPipeline,
    model: AutoModelForTokenClassification,
    default: {
      model: "Xenova/bert-base-multilingual-cased-ner-hrl"
    },
    type: "text"
  },
  "question-answering": {
    pipeline: QuestionAnsweringPipeline,
    model: AutoModelForQuestionAnswering,
    default: {
      model: "Xenova/distilbert-base-cased-distilled-squad"
    },
    type: "text"
  },
  "fill-mask": {
    pipeline: FillMaskPipeline,
    model: AutoModelForMaskedLM,
    default: {
      model: "onnx-community/ettin-encoder-32m-ONNX",
      dtype: "fp32"
    },
    type: "text"
  },
  summarization: {
    pipeline: SummarizationPipeline,
    model: AutoModelForSeq2SeqLM,
    default: {
      model: "Xenova/distilbart-cnn-6-6"
    },
    type: "text"
  },
  translation: {
    pipeline: TranslationPipeline,
    model: AutoModelForSeq2SeqLM,
    default: {
      model: "Xenova/t5-small"
    },
    type: "text"
  },
  "text2text-generation": {
    pipeline: Text2TextGenerationPipeline,
    model: AutoModelForSeq2SeqLM,
    default: {
      model: "Xenova/flan-t5-small"
    },
    type: "text"
  },
  "text-generation": {
    pipeline: TextGenerationPipeline,
    model: AutoModelForCausalLM,
    default: {
      model: "onnx-community/Qwen3-0.6B-ONNX",
      dtype: "q4"
    },
    type: "text"
  },
  "zero-shot-classification": {
    pipeline: ZeroShotClassificationPipeline,
    model: AutoModelForSequenceClassification,
    default: {
      model: "Xenova/distilbert-base-uncased-mnli"
    },
    type: "text"
  },
  "audio-classification": {
    pipeline: AudioClassificationPipeline,
    model: AutoModelForAudioClassification,
    default: {
      model: "Xenova/wav2vec2-base-superb-ks"
    },
    type: "audio"
  },
  "zero-shot-audio-classification": {
    pipeline: ZeroShotAudioClassificationPipeline,
    model: AutoModel,
    default: {
      model: "Xenova/clap-htsat-unfused"
    },
    type: "multimodal"
  },
  "automatic-speech-recognition": {
    pipeline: AutomaticSpeechRecognitionPipeline,
    model: [AutoModelForSpeechSeq2Seq, AutoModelForCTC],
    default: {
      model: "Xenova/whisper-tiny.en"
    },
    type: "multimodal"
  },
  "text-to-audio": {
    pipeline: TextToAudioPipeline,
    model: [AutoModelForTextToWaveform, AutoModelForTextToSpectrogram],
    default: {
      model: "onnx-community/Supertonic-TTS-ONNX",
      dtype: "fp32"
    },
    type: "text"
  },
  "image-to-text": {
    pipeline: ImageToTextPipeline,
    model: AutoModelForVision2Seq,
    default: {
      model: "Xenova/vit-gpt2-image-captioning"
    },
    type: "multimodal"
  },
  "image-classification": {
    pipeline: ImageClassificationPipeline,
    model: AutoModelForImageClassification,
    default: {
      model: "Xenova/vit-base-patch16-224"
    },
    type: "multimodal"
  },
  "image-segmentation": {
    pipeline: ImageSegmentationPipeline,
    model: [AutoModelForImageSegmentation, AutoModelForSemanticSegmentation, AutoModelForUniversalSegmentation],
    default: {
      model: "Xenova/detr-resnet-50-panoptic"
    },
    type: "multimodal"
  },
  "background-removal": {
    pipeline: BackgroundRemovalPipeline,
    model: [AutoModelForImageSegmentation, AutoModelForSemanticSegmentation, AutoModelForUniversalSegmentation],
    default: {
      model: "Xenova/modnet"
    },
    type: "image"
  },
  "zero-shot-image-classification": {
    pipeline: ZeroShotImageClassificationPipeline,
    model: AutoModel,
    default: {
      model: "Xenova/clip-vit-base-patch32"
    },
    type: "multimodal"
  },
  "object-detection": {
    pipeline: ObjectDetectionPipeline,
    model: AutoModelForObjectDetection,
    default: {
      model: "Xenova/detr-resnet-50"
    },
    type: "multimodal"
  },
  "zero-shot-object-detection": {
    pipeline: ZeroShotObjectDetectionPipeline,
    model: AutoModelForZeroShotObjectDetection,
    default: {
      model: "Xenova/owlvit-base-patch32"
    },
    type: "multimodal"
  },
  "document-question-answering": {
    pipeline: DocumentQuestionAnsweringPipeline,
    model: AutoModelForDocumentQuestionAnswering,
    default: {
      model: "Xenova/donut-base-finetuned-docvqa"
    },
    type: "multimodal"
  },
  "image-to-image": {
    pipeline: ImageToImagePipeline,
    model: AutoModelForImageToImage,
    default: {
      model: "Xenova/swin2SR-classical-sr-x2-64"
    },
    type: "image"
  },
  "depth-estimation": {
    pipeline: DepthEstimationPipeline,
    model: AutoModelForDepthEstimation,
    default: {
      model: "onnx-community/depth-anything-v2-small"
    },
    type: "image"
  },
  "feature-extraction": {
    pipeline: FeatureExtractionPipeline,
    model: AutoModel,
    default: {
      model: "onnx-community/all-MiniLM-L6-v2-ONNX",
      dtype: "fp32"
    },
    type: "text"
  },
  "image-feature-extraction": {
    pipeline: ImageFeatureExtractionPipeline,
    model: [AutoModelForImageFeatureExtraction, AutoModel],
    default: {
      model: "onnx-community/dinov3-vits16-pretrain-lvd1689m-ONNX",
      dtype: "fp32"
    },
    type: "image"
  }
});
var TASK_ALIASES = Object.freeze({
  "sentiment-analysis": "text-classification",
  ner: "token-classification",
  // "vqa": "visual-question-answering", // TODO: Add
  asr: "automatic-speech-recognition",
  "text-to-speech": "text-to-audio",
  // Add for backwards compatibility
  embeddings: "feature-extraction"
});

// src/utils/model_registry/get_processor_files.js
async function get_processor_files(modelId) {
  if (!modelId) {
    throw new Error("modelId is required");
  }
  const metadata = await get_file_metadata(modelId, IMAGE_PROCESSOR_NAME, {});
  return metadata.exists ? [IMAGE_PROCESSOR_NAME] : [];
}

// src/utils/model_registry/get_files.js
async function get_files(modelId, {
  config = null,
  dtype = null,
  device = null,
  model_file_name = null,
  include_tokenizer = true,
  include_processor = true
} = {}) {
  const files = await get_model_files(modelId, { config, dtype, device, model_file_name });
  if (include_tokenizer) {
    const tokenizerFiles = await get_tokenizer_files(modelId);
    files.push(...tokenizerFiles);
  }
  if (include_processor) {
    const processorFiles = await get_processor_files(modelId);
    files.push(...processorFiles);
  }
  return files;
}

// src/utils/model_registry/get_pipeline_files.js
async function get_pipeline_files(task, modelId, options = {}) {
  task = TASK_ALIASES[task] ?? task;
  const taskConfig = SUPPORTED_TASKS[task];
  if (!taskConfig) {
    throw new Error(
      `Unsupported pipeline task: ${task}. Must be one of [${Object.keys(SUPPORTED_TASKS).join(", ")}]`
    );
  }
  const { type } = taskConfig;
  const include_tokenizer = type !== "audio" && type !== "image";
  const include_processor = type !== "text";
  const files = await get_files(modelId, {
    ...options,
    include_tokenizer,
    include_processor
  });
  if (task === "text-generation") {
    const config = await get_config(modelId, options);
    const modelType = resolve_model_type(config);
    const textOnlySessions = getTextOnlySessions(modelType);
    if (textOnlySessions) {
      const allowedPrefixes = Object.values(textOnlySessions).map((s) => `onnx/${s}`);
      return files.filter((f) => !f.startsWith("onnx/") || allowedPrefixes.some((p) => f.startsWith(p)));
    }
  }
  return files;
}

// src/pipelines.js
async function pipeline(task, model = null, {
  progress_callback = null,
  config = null,
  cache_dir = null,
  local_files_only = false,
  revision = "main",
  device = null,
  dtype = null,
  subfolder = "onnx",
  use_external_data_format = null,
  model_file_name = null,
  session_options = {}
} = {}) {
  task = TASK_ALIASES[task] ?? task;
  const pipelineInfo = SUPPORTED_TASKS[task.split("_", 1)[0]];
  if (!pipelineInfo) {
    throw Error(`Unsupported pipeline: ${task}. Must be one of [${Object.keys(SUPPORTED_TASKS)}]`);
  }
  if (!model) {
    model = pipelineInfo.default.model;
    logger.info(`No model specified. Using default model: "${model}".`);
    if (!dtype && pipelineInfo.default.dtype) {
      dtype = pipelineInfo.default.dtype;
    }
  }
  const expected_files = await get_pipeline_files(task, model, {
    device,
    dtype
  });
  let files_loading = {};
  if (progress_callback) {
    const metadata = await Promise.all(expected_files.map(async (file) => get_file_metadata(model, file)));
    metadata.forEach((m, i) => {
      if (m.exists) {
        files_loading[expected_files[i]] = {
          loaded: 0,
          total: m.size ?? 0
        };
      }
    });
  }
  const pretrainedOptions = {
    progress_callback: progress_callback ? new DefaultProgressCallback(progress_callback, files_loading) : void 0,
    config,
    cache_dir,
    local_files_only,
    revision,
    device,
    dtype,
    subfolder,
    use_external_data_format,
    model_file_name,
    session_options
  };
  const hasTokenizer = expected_files.includes("tokenizer.json");
  const hasProcessor = expected_files.includes("preprocessor_config.json");
  const modelClasses = pipelineInfo.model;
  let modelPromise;
  if (Array.isArray(modelClasses)) {
    const resolvedConfig = config ?? await AutoConfig.from_pretrained(model, pretrainedOptions);
    const { model_type } = resolvedConfig;
    const matchedClass = modelClasses.find((cls) => cls.supports(model_type));
    if (!matchedClass) {
      throw Error(
        `Unsupported model type "${model_type}" for task "${task}". None of the candidate model classes support this type.`
      );
    }
    modelPromise = matchedClass.from_pretrained(model, { ...pretrainedOptions, config: resolvedConfig });
  } else {
    modelPromise = modelClasses.from_pretrained(model, pretrainedOptions);
  }
  const [tokenizer, processor, model_loaded] = await Promise.all([
    hasTokenizer ? AutoTokenizer.from_pretrained(model, pretrainedOptions) : null,
    hasProcessor ? AutoProcessor.from_pretrained(model, pretrainedOptions) : null,
    modelPromise
  ]);
  const results = { task, model: model_loaded };
  if (tokenizer) results.tokenizer = tokenizer;
  if (processor) results.processor = processor;
  dispatchCallback(progress_callback, {
    status: "ready",
    task,
    model
  });
  const pipelineClass = pipelineInfo.pipeline;
  return new pipelineClass(results);
}

// src/generation/streamers.js
var is_chinese_char2 = (cp) => cp >= 19968 && cp <= 40959 || cp >= 13312 && cp <= 19903 || cp >= 131072 && cp <= 173791 || cp >= 173824 && cp <= 177983 || cp >= 177984 && cp <= 178207 || cp >= 178208 && cp <= 183983 || cp >= 63744 && cp <= 64255 || cp >= 194560 && cp <= 195103;
var BaseStreamer = class {
  /**
   * Function that is called by `.generate()` to push new tokens
   * @param {bigint[][]} value
   */
  put(value) {
    throw Error("Not implemented");
  }
  /**
   * Function that is called by `.generate()` to signal the end of generation
   */
  end() {
    throw Error("Not implemented");
  }
};
var stdout_write = apis.IS_PROCESS_AVAILABLE ? (x) => process.stdout.write(x) : (x) => console.log(x);
var TextStreamer = class extends BaseStreamer {
  /**
   *
   * @param {import('../tokenization_utils.js').PreTrainedTokenizer} tokenizer
   * @param {Object} options
   * @param {boolean} [options.skip_prompt=false] Whether to skip the prompt tokens
   * @param {boolean} [options.skip_special_tokens=true] Whether to skip special tokens when decoding
   * @param {function(string): void} [options.callback_function=null] Function to call when a piece of text is ready to display
   * @param {function(bigint[]): void} [options.token_callback_function=null] Function to call when a new token is generated
   * @param {Object} [options.decode_kwargs={}] Additional keyword arguments to pass to the tokenizer's decode method
   */
  constructor(tokenizer, {
    skip_prompt = false,
    callback_function = null,
    token_callback_function = null,
    skip_special_tokens = true,
    decode_kwargs = {},
    ...kwargs
  } = {}) {
    super();
    this.tokenizer = tokenizer;
    this.skip_prompt = skip_prompt;
    this.callback_function = callback_function ?? stdout_write;
    this.token_callback_function = token_callback_function;
    this.decode_kwargs = { skip_special_tokens, ...decode_kwargs, ...kwargs };
    this.token_cache = [];
    this.print_len = 0;
    this.next_tokens_are_prompt = true;
    this.special_ids = new Set(this.tokenizer.all_special_ids.map(BigInt));
  }
  /**
   * Receives tokens, decodes them, and prints them to stdout as soon as they form entire words.
   * @param {bigint[][]} value
   */
  put(value) {
    if (value.length > 1) {
      throw Error("TextStreamer only supports batch size of 1");
    }
    const is_prompt = this.next_tokens_are_prompt;
    if (is_prompt) {
      this.next_tokens_are_prompt = false;
      if (this.skip_prompt) return;
    }
    const tokens = value[0];
    this.token_callback_function?.(tokens);
    if (tokens.length === 1 && this.special_ids.has(tokens[0])) {
      if (this.decode_kwargs.skip_special_tokens) return;
      if (this.token_cache.length > 0) {
        const text2 = this.tokenizer.decode(this.token_cache, this.decode_kwargs);
        const printable_text2 = text2.slice(this.print_len);
        this.on_finalized_text(printable_text2, false);
        this.token_cache = [];
        this.print_len = 0;
      }
      const special_text = this.tokenizer.decode(tokens, this.decode_kwargs);
      this.on_finalized_text(special_text, false);
      return;
    }
    this.token_cache = mergeArrays(this.token_cache, tokens);
    const text = this.tokenizer.decode(this.token_cache, this.decode_kwargs);
    let printable_text;
    if (is_prompt || text.endsWith("\n")) {
      printable_text = text.slice(this.print_len);
      this.token_cache = [];
      this.print_len = 0;
    } else if (text.length > 0 && is_chinese_char2(text.charCodeAt(text.length - 1))) {
      printable_text = text.slice(this.print_len);
      this.print_len += printable_text.length;
    } else {
      printable_text = text.slice(this.print_len, text.lastIndexOf(" ") + 1);
      this.print_len += printable_text.length;
    }
    this.on_finalized_text(printable_text, false);
  }
  /**
   * Flushes any remaining cache and prints a newline to stdout.
   */
  end() {
    let printable_text;
    if (this.token_cache.length > 0) {
      const text = this.tokenizer.decode(this.token_cache, this.decode_kwargs);
      printable_text = text.slice(this.print_len);
      this.token_cache = [];
      this.print_len = 0;
    } else {
      printable_text = "";
    }
    this.next_tokens_are_prompt = true;
    this.on_finalized_text(printable_text, true);
  }
  /**
   * Prints the new text to stdout. If the stream is ending, also prints a newline.
   * @param {string} text
   * @param {boolean} stream_end
   */
  on_finalized_text(text, stream_end) {
    if (text.length > 0) {
      this.callback_function?.(text);
    }
    if (stream_end && this.callback_function === stdout_write && apis.IS_PROCESS_AVAILABLE) {
      this.callback_function?.("\n");
    }
  }
};
var WhisperTextStreamer = class extends TextStreamer {
  /**
   * @param {import('../models/whisper/tokenization_whisper.js').WhisperTokenizer} tokenizer
   * @param {Object} options
   * @param {boolean} [options.skip_prompt=false] Whether to skip the prompt tokens
   * @param {function(string): void} [options.callback_function=null] Function to call when a piece of text is ready to display
   * @param {function(bigint[]): void} [options.token_callback_function=null] Function to call when a new token is generated
   * @param {function(number): void} [options.on_chunk_start=null] Function to call when a new chunk starts
   * @param {function(number): void} [options.on_chunk_end=null] Function to call when a chunk ends
   * @param {function(): void} [options.on_finalize=null] Function to call when the stream is finalized
   * @param {number} [options.time_precision=0.02] Precision of the timestamps
   * @param {boolean} [options.skip_special_tokens=true] Whether to skip special tokens when decoding
   * @param {Object} [options.decode_kwargs={}] Additional keyword arguments to pass to the tokenizer's decode method
   */
  constructor(tokenizer, {
    skip_prompt = false,
    callback_function = null,
    token_callback_function = null,
    on_chunk_start = null,
    on_chunk_end = null,
    on_finalize = null,
    time_precision = 0.02,
    skip_special_tokens = true,
    decode_kwargs = {}
  } = {}) {
    super(tokenizer, {
      skip_prompt,
      skip_special_tokens,
      callback_function,
      token_callback_function,
      decode_kwargs
    });
    this.timestamp_begin = tokenizer.timestamp_begin;
    this.on_chunk_start = on_chunk_start;
    this.on_chunk_end = on_chunk_end;
    this.on_finalize = on_finalize;
    this.time_precision = time_precision;
    this.waiting_for_timestamp = false;
  }
  /**
   * @param {bigint[][]} value
   */
  put(value) {
    if (value.length > 1) {
      throw Error("WhisperTextStreamer only supports batch size of 1");
    }
    const tokens = value[0];
    if (tokens.length === 1) {
      const offset = Number(tokens[0]) - this.timestamp_begin;
      if (offset >= 0) {
        const time = offset * this.time_precision;
        if (this.waiting_for_timestamp) {
          this.on_chunk_end?.(time);
        } else {
          this.on_chunk_start?.(time);
        }
        this.waiting_for_timestamp = !this.waiting_for_timestamp;
        this.token_callback_function?.(tokens);
        return;
      }
    }
    return super.put(value);
  }
  end() {
    super.end();
    this.on_finalize?.();
  }
};

// src/utils/video.js
var RawVideoFrame = class {
  /**
   * @param {RawImage} image
   * @param {number} timestamp
   */
  constructor(image, timestamp) {
    this.image = image;
    this.timestamp = timestamp;
  }
};
var RawVideo = class {
  /**
   * @param {RawVideoFrame[]|RawImage[]} frames
   * @param {number} duration
   */
  constructor(frames, duration) {
    if (frames.length > 0 && frames[0] instanceof RawImage) {
      frames = frames.map((image, i) => new RawVideoFrame(image, (i + 1) / (frames.length + 1) * duration));
    }
    this.frames = /** @type {RawVideoFrame[]} */
    frames;
    this.duration = duration;
  }
  get width() {
    return this.frames[0].image.width;
  }
  get height() {
    return this.frames[0].image.height;
  }
  get fps() {
    return this.frames.length / this.duration;
  }
};
async function load_video(src, { num_frames = null, fps = null } = {}) {
  if (!apis.IS_BROWSER_ENV) {
    throw new Error("`load_video` is currently only supported in browser environments.");
  }
  if (num_frames == null && fps == null) {
    throw new Error("Either num_frames or fps must be provided.");
  }
  const frames = [];
  const video = document.createElement("video");
  video.crossOrigin = "anonymous";
  video.muted = true;
  if (typeof src === "string") {
    video.src = src;
  } else if (src instanceof Blob) {
    video.src = URL.createObjectURL(src);
  } else if (src instanceof HTMLVideoElement) {
    video.src = src.src;
  } else {
    throw new Error("Invalid URL or video element provided.");
  }
  await new Promise((resolve) => video.onloadedmetadata = resolve);
  if (video.seekable.start(0) === video.seekable.end(0)) {
    const response = await env.fetch(video.src);
    const blob = await response.blob();
    video.src = URL.createObjectURL(blob);
    await new Promise((resolve) => video.onloadedmetadata = resolve);
  }
  const duration = video.duration;
  let count2, step;
  if (num_frames != null) {
    count2 = num_frames;
    step = num_frames === 1 ? 0 : duration / (num_frames - 1);
  } else {
    step = 1 / fps;
    count2 = Math.floor(duration / step);
  }
  let sampleTimes = [];
  for (let i = 0; i < count2; ++i) {
    sampleTimes.push(num_frames === 1 ? duration / 2 : i * step);
  }
  const canvas = document.createElement("canvas");
  canvas.width = video.videoWidth;
  canvas.height = video.videoHeight;
  const ctx = canvas.getContext("2d", { willReadFrequently: true });
  for (const t of sampleTimes) {
    video.currentTime = t;
    await new Promise((resolve) => {
      video.onseeked = resolve;
    });
    ctx.drawImage(video, 0, 0, canvas.width, canvas.height);
    const imageData = ctx.getImageData(0, 0, canvas.width, canvas.height);
    const frameData = new RawImage(imageData.data, canvas.width, canvas.height, 4);
    const frame = new RawVideoFrame(frameData, t);
    frames.push(frame);
  }
  video.remove();
  return new RawVideo(frames, duration);
}

// src/utils/model_registry/is_cached.js
async function check_files_cache(modelId, files, options = {}) {
  const cache2 = await getCache(options?.cache_dir);
  if (!cache2) {
    const fileStatuses2 = files.map((filename) => ({ file: filename, cached: false }));
    return { allCached: false, files: fileStatuses2 };
  }
  const fileStatuses = await Promise.all(
    files.map(async (filename) => {
      const { localPath, proposedCacheKey } = buildResourcePaths(modelId, filename, options, cache2);
      const cached = await checkCachedResource(cache2, localPath, proposedCacheKey);
      return { file: filename, cached: !!cached };
    })
  );
  return { allCached: fileStatuses.every((f) => f.cached), files: fileStatuses };
}
async function is_file_cached(modelId, filename, options = {}) {
  const cache2 = await getCache(options?.cache_dir);
  if (!cache2) return false;
  const { localPath, proposedCacheKey } = buildResourcePaths(modelId, filename, options, cache2);
  return !!await checkCachedResource(cache2, localPath, proposedCacheKey);
}
async function is_cached(modelId, options = {}) {
  if (!modelId) {
    throw new Error("modelId is required");
  }
  if (!await is_file_cached(modelId, "config.json", options)) {
    return false;
  }
  const files = await get_files(modelId, options);
  const result = await check_files_cache(modelId, files, options);
  return result.allCached;
}
async function is_cached_files(modelId, options = {}) {
  if (!modelId) {
    throw new Error("modelId is required");
  }
  const files = await get_files(modelId, options);
  return await check_files_cache(modelId, files, options);
}
async function is_pipeline_cached(task, modelId, options = {}) {
  if (!task) {
    throw new Error("task is required");
  }
  if (!modelId) {
    throw new Error("modelId is required");
  }
  if (!await is_file_cached(modelId, "config.json", options)) {
    return false;
  }
  const files = await get_pipeline_files(task, modelId, options);
  const result = await check_files_cache(modelId, files, options);
  return result.allCached;
}
async function is_pipeline_cached_files(task, modelId, options = {}) {
  if (!task) {
    throw new Error("task is required");
  }
  if (!modelId) {
    throw new Error("modelId is required");
  }
  const files = await get_pipeline_files(task, modelId, options);
  return await check_files_cache(modelId, files, options);
}

// src/utils/model_registry/clear_cache.js
async function clear_files_from_cache(modelId, files, options = {}) {
  const cache2 = await getCache(options?.cache_dir);
  if (!cache2) {
    return {
      filesDeleted: 0,
      filesCached: 0,
      files: files.map((filename) => ({ file: filename, deleted: false, wasCached: false }))
    };
  }
  if (!cache2.delete) {
    throw new Error("Cache does not support delete operation");
  }
  const results = await Promise.all(
    files.map(async (filename) => {
      const { localPath, proposedCacheKey } = buildResourcePaths(modelId, filename, options, cache2);
      const cached = await checkCachedResource(cache2, localPath, proposedCacheKey);
      const wasCached = !!cached;
      let deleted = false;
      if (wasCached) {
        const deletedWithProposed = await cache2.delete(proposedCacheKey);
        const deletedWithLocal = !deletedWithProposed && proposedCacheKey !== localPath ? await cache2.delete(localPath) : false;
        deleted = deletedWithProposed || deletedWithLocal;
      }
      return { file: filename, deleted, wasCached };
    })
  );
  return {
    filesDeleted: results.filter((r) => r.deleted).length,
    filesCached: results.filter((r) => r.wasCached).length,
    files: results
  };
}
async function clear_cache(modelId, options = {}) {
  if (!modelId) {
    throw new Error("modelId is required");
  }
  const files = await get_files(modelId, options);
  return await clear_files_from_cache(modelId, files, options);
}
async function clear_pipeline_cache(task, modelId, options = {}) {
  if (!task) {
    throw new Error("task is required");
  }
  if (!modelId) {
    throw new Error("modelId is required");
  }
  const files = await get_pipeline_files(task, modelId, options);
  return await clear_files_from_cache(modelId, files, options);
}

// src/utils/model_registry/get_available_dtypes.js
var CONCRETE_DTYPES = Object.keys(DEFAULT_DTYPE_SUFFIX_MAPPING);
async function get_available_dtypes(modelId, { config = null, model_file_name = null, revision = "main", cache_dir = null, local_files_only = false } = {}) {
  config = await get_config(modelId, { config, cache_dir, local_files_only, revision });
  const subfolder = "onnx";
  const modelType = resolve_model_type(config);
  const { sessions } = getSessionsConfig(modelType, config, { model_file_name });
  const baseNames = Object.values(sessions);
  const metadataOptions = { revision, cache_dir, local_files_only };
  const probeResults = await Promise.all(
    CONCRETE_DTYPES.map(async (dtype) => {
      const suffix = DEFAULT_DTYPE_SUFFIX_MAPPING[dtype] ?? "";
      const allExist = await Promise.all(
        baseNames.map(async (baseName) => {
          const filename = `${subfolder}/${baseName}${suffix}.onnx`;
          const metadata = await get_file_metadata(modelId, filename, metadataOptions);
          return metadata.exists;
        })
      );
      return { dtype, available: allExist.every(Boolean) };
    })
  );
  return probeResults.filter((r) => r.available).map((r) => r.dtype);
}

// src/utils/model_registry/ModelRegistry.js
var ModelRegistry = class {
  /**
   * Get all files (model, tokenizer, processor) needed for a model.
   *
   * @param {string} modelId - The model id (e.g., "onnx-community/bert-base-uncased-ONNX")
   * @param {Object} [options] - Optional parameters
   * @param {import('../../configs.js').PretrainedConfig} [options.config=null] - Pre-loaded config
   * @param {import('../dtypes.js').DataType|Record<string, import('../dtypes.js').DataType>} [options.dtype=null] - Override dtype
   * @param {import('../devices.js').DeviceType|Record<string, import('../devices.js').DeviceType>} [options.device=null] - Override device
   * @param {string} [options.model_file_name=null] - Override the model file name (excluding .onnx suffix)
   * @param {boolean} [options.include_tokenizer=true] - Whether to check for tokenizer files
   * @param {boolean} [options.include_processor=true] - Whether to check for processor files
   * @returns {Promise<string[]>} Array of file paths
   *
   * @example
   * const files = await ModelRegistry.get_files('onnx-community/gpt2-ONNX');
   * console.log(files); // ['config.json', 'tokenizer.json', 'onnx/model_q4.onnx', ...]
   */
  static async get_files(modelId, options = {}) {
    return get_files(modelId, options);
  }
  /**
   * Get all files needed for a specific pipeline task.
   * Automatically determines which components are needed based on the task.
   *
   * @param {string} task - The pipeline task (e.g., "text-generation", "background-removal")
   * @param {string} modelId - The model id (e.g., "onnx-community/bert-base-uncased-ONNX")
   * @param {Object} [options] - Optional parameters
   * @param {import('../../configs.js').PretrainedConfig} [options.config=null] - Pre-loaded config
   * @param {import('../dtypes.js').DataType|Record<string, import('../dtypes.js').DataType>} [options.dtype=null] - Override dtype
   * @param {import('../devices.js').DeviceType|Record<string, import('../devices.js').DeviceType>} [options.device=null] - Override device
   * @param {string} [options.model_file_name=null] - Override the model file name (excluding .onnx suffix)
   * @returns {Promise<string[]>} Array of file paths
   *
   * @example
   * const files = await ModelRegistry.get_pipeline_files('text-generation', 'onnx-community/gpt2-ONNX');
   * console.log(files); // ['config.json', 'tokenizer.json', 'onnx/model_q4.onnx', ...]
   */
  static async get_pipeline_files(task, modelId, options = {}) {
    return get_pipeline_files(task, modelId, options);
  }
  /**
   * Get model files needed for a specific model.
   *
   * @param {string} modelId - The model id
   * @param {Object} [options] - Optional parameters
   * @param {import('../../configs.js').PretrainedConfig} [options.config=null] - Pre-loaded config
   * @param {import('../dtypes.js').DataType|Record<string, import('../dtypes.js').DataType>} [options.dtype=null] - Override dtype
   * @param {import('../devices.js').DeviceType|Record<string, import('../devices.js').DeviceType>} [options.device=null] - Override device
   * @param {string} [options.model_file_name=null] - Override the model file name (excluding .onnx suffix)
   * @returns {Promise<string[]>} Array of model file paths
   *
   * @example
   * const files = await ModelRegistry.get_model_files('onnx-community/bert-base-uncased-ONNX');
   * console.log(files); // ['config.json', 'onnx/model_q4.onnx', 'generation_config.json']
   */
  static async get_model_files(modelId, options = {}) {
    return get_model_files(modelId, options);
  }
  /**
   * Get tokenizer files needed for a specific model.
   *
   * @param {string} modelId - The model id
   * @returns {Promise<string[]>} Array of tokenizer file paths
   *
   * @example
   * const files = await ModelRegistry.get_tokenizer_files('onnx-community/gpt2-ONNX');
   * console.log(files); // ['tokenizer.json', 'tokenizer_config.json']
   */
  static async get_tokenizer_files(modelId) {
    return get_tokenizer_files(modelId);
  }
  /**
   * Get processor files needed for a specific model.
   *
   * @param {string} modelId - The model id
   * @returns {Promise<string[]>} Array of processor file paths
   *
   * @example
   * const files = await ModelRegistry.get_processor_files('onnx-community/vit-base-patch16-224-ONNX');
   * console.log(files); // ['preprocessor_config.json']
   */
  static async get_processor_files(modelId) {
    return get_processor_files(modelId);
  }
  /**
   * Detects which quantization levels (dtypes) are available for a model
   * by checking which ONNX files exist on the hub or locally.
   *
   * A dtype is considered available if all required model session files
   * exist for that dtype.
   *
   * @param {string} modelId - The model id (e.g., "onnx-community/all-MiniLM-L6-v2-ONNX")
   * @param {Object} [options] - Optional parameters
   * @param {import('../../configs.js').PretrainedConfig} [options.config=null] - Pre-loaded config
   * @param {string} [options.model_file_name=null] - Override the model file name (excluding .onnx suffix)
   * @param {string} [options.revision='main'] - Model revision
   * @param {string} [options.cache_dir=null] - Custom cache directory
   * @param {boolean} [options.local_files_only=false] - Only check local files
   * @returns {Promise<string[]>} Array of available dtype strings (e.g., ['fp32', 'fp16', 'q4', 'q8'])
   *
   * @example
   * const dtypes = await ModelRegistry.get_available_dtypes('onnx-community/all-MiniLM-L6-v2-ONNX');
   * console.log(dtypes); // ['fp32', 'fp16', 'int8', 'uint8', 'q8', 'q4']
   */
  static async get_available_dtypes(modelId, options = {}) {
    return get_available_dtypes(modelId, options);
  }
  /**
   * Quickly checks if a model is fully cached by verifying `config.json` is present,
   * then confirming all required files are cached.
   * Returns a plain boolean — use `is_cached_files` if you need per-file detail.
   *
   * @param {string} modelId - The model id
   * @param {Object} [options] - Optional parameters
   * @param {string} [options.cache_dir] - Custom cache directory
   * @param {string} [options.revision] - Model revision (default: 'main')
   * @param {import('../../configs.js').PretrainedConfig} [options.config] - Pre-loaded config
   * @param {import('../dtypes.js').DataType|Record<string, import('../dtypes.js').DataType>} [options.dtype=null] - Override dtype
   * @param {import('../devices.js').DeviceType|Record<string, import('../devices.js').DeviceType>} [options.device=null] - Override device
   * @returns {Promise<boolean>} Whether all required files are cached
   *
   * @example
   * const cached = await ModelRegistry.is_cached('onnx-community/bert-base-uncased-ONNX');
   * console.log(cached); // true or false
   */
  static async is_cached(modelId, options = {}) {
    return is_cached(modelId, options);
  }
  /**
   * Checks if all files for a given model are already cached, with per-file detail.
   * Automatically determines which files are needed using get_files().
   *
   * @param {string} modelId - The model id
   * @param {Object} [options] - Optional parameters
   * @param {string} [options.cache_dir] - Custom cache directory
   * @param {string} [options.revision] - Model revision (default: 'main')
   * @param {import('../../configs.js').PretrainedConfig} [options.config] - Pre-loaded config
   * @param {import('../dtypes.js').DataType|Record<string, import('../dtypes.js').DataType>} [options.dtype=null] - Override dtype
   * @param {import('../devices.js').DeviceType|Record<string, import('../devices.js').DeviceType>} [options.device=null] - Override device
   * @returns {Promise<import('./is_cached.js').CacheCheckResult>} Object with allCached boolean and files array with cache status
   *
   * @example
   * const status = await ModelRegistry.is_cached_files('onnx-community/bert-base-uncased-ONNX');
   * console.log(status.allCached); // true or false
   * console.log(status.files); // [{ file: 'config.json', cached: true }, ...]
   */
  static async is_cached_files(modelId, options = {}) {
    return is_cached_files(modelId, options);
  }
  /**
   * Quickly checks if all files for a specific pipeline task are cached by verifying
   * `config.json` is present, then confirming all required files are cached.
   * Returns a plain boolean — use `is_pipeline_cached_files` if you need per-file detail.
   *
   * @param {string} task - The pipeline task (e.g., "text-generation", "background-removal")
   * @param {string} modelId - The model id
   * @param {Object} [options] - Optional parameters
   * @param {string} [options.cache_dir] - Custom cache directory
   * @param {string} [options.revision] - Model revision (default: 'main')
   * @param {import('../../configs.js').PretrainedConfig} [options.config] - Pre-loaded config
   * @param {import('../dtypes.js').DataType|Record<string, import('../dtypes.js').DataType>} [options.dtype=null] - Override dtype
   * @param {import('../devices.js').DeviceType|Record<string, import('../devices.js').DeviceType>} [options.device=null] - Override device
   * @returns {Promise<boolean>} Whether all required files are cached
   *
   * @example
   * const cached = await ModelRegistry.is_pipeline_cached('text-generation', 'onnx-community/gpt2-ONNX');
   * console.log(cached); // true or false
   */
  static async is_pipeline_cached(task, modelId, options = {}) {
    return is_pipeline_cached(task, modelId, options);
  }
  /**
   * Checks if all files for a specific pipeline task are already cached, with per-file detail.
   * Automatically determines which components are needed based on the task.
   *
   * @param {string} task - The pipeline task (e.g., "text-generation", "background-removal")
   * @param {string} modelId - The model id
   * @param {Object} [options] - Optional parameters
   * @param {string} [options.cache_dir] - Custom cache directory
   * @param {string} [options.revision] - Model revision (default: 'main')
   * @param {import('../../configs.js').PretrainedConfig} [options.config] - Pre-loaded config
   * @param {import('../dtypes.js').DataType|Record<string, import('../dtypes.js').DataType>} [options.dtype=null] - Override dtype
   * @param {import('../devices.js').DeviceType|Record<string, import('../devices.js').DeviceType>} [options.device=null] - Override device
   * @returns {Promise<import('./is_cached.js').CacheCheckResult>} Object with allCached boolean and files array with cache status
   *
   * @example
   * const status = await ModelRegistry.is_pipeline_cached_files('text-generation', 'onnx-community/gpt2-ONNX');
   * console.log(status.allCached); // true or false
   * console.log(status.files); // [{ file: 'config.json', cached: true }, ...]
   */
  static async is_pipeline_cached_files(task, modelId, options = {}) {
    return is_pipeline_cached_files(task, modelId, options);
  }
  /**
   * Get metadata for a specific file without downloading it.
   *
   * @param {string} path_or_repo_id - Model id or path
   * @param {string} filename - The file name
   * @param {import('../hub.js').PretrainedOptions} [options] - Optional parameters
   * @returns {Promise<{exists: boolean, size?: number, contentType?: string, fromCache?: boolean}>} File metadata
   *
   * @example
   * const metadata = await ModelRegistry.get_file_metadata('onnx-community/gpt2-ONNX', 'config.json');
   * console.log(metadata.exists, metadata.size); // true, 665
   */
  static async get_file_metadata(path_or_repo_id, filename, options = {}) {
    return get_file_metadata(path_or_repo_id, filename, options);
  }
  /**
   * Clears all cached files for a given model.
   * Automatically determines which files are needed and removes them from the cache.
   *
   * @param {string} modelId - The model id (e.g., "onnx-community/gpt2-ONNX")
   * @param {Object} [options] - Optional parameters
   * @param {string} [options.cache_dir] - Custom cache directory
   * @param {string} [options.revision] - Model revision (default: 'main')
   * @param {import('../../configs.js').PretrainedConfig} [options.config] - Pre-loaded config
   * @param {import('../dtypes.js').DataType|Record<string, import('../dtypes.js').DataType>} [options.dtype] - Override dtype
   * @param {import('../devices.js').DeviceType|Record<string, import('../devices.js').DeviceType>} [options.device] - Override device
   * @param {boolean} [options.include_tokenizer=true] - Whether to clear tokenizer files
   * @param {boolean} [options.include_processor=true] - Whether to clear processor files
   * @returns {Promise<import('./clear_cache.js').CacheClearResult>} Object with deletion statistics and file status
   *
   * @example
   * const result = await ModelRegistry.clear_cache('onnx-community/bert-base-uncased-ONNX');
   * console.log(`Deleted ${result.filesDeleted} of ${result.filesCached} cached files`);
   */
  static async clear_cache(modelId, options = {}) {
    return clear_cache(modelId, options);
  }
  /**
   * Clears all cached files for a specific pipeline task.
   * Automatically determines which components are needed based on the task.
   *
   * @param {string} task - The pipeline task (e.g., "text-generation", "image-classification")
   * @param {string} modelId - The model id (e.g., "onnx-community/gpt2-ONNX")
   * @param {Object} [options] - Optional parameters
   * @param {string} [options.cache_dir] - Custom cache directory
   * @param {string} [options.revision] - Model revision (default: 'main')
   * @param {import('../../configs.js').PretrainedConfig} [options.config] - Pre-loaded config
   * @param {import('../dtypes.js').DataType|Record<string, import('../dtypes.js').DataType>} [options.dtype] - Override dtype
   * @param {import('../devices.js').DeviceType|Record<string, import('../devices.js').DeviceType>} [options.device] - Override device
   * @returns {Promise<import('./clear_cache.js').CacheClearResult>} Object with deletion statistics and file status
   *
   * @example
   * const result = await ModelRegistry.clear_pipeline_cache('text-generation', 'onnx-community/gpt2-ONNX');
   * console.log(`Deleted ${result.filesDeleted} of ${result.filesCached} cached files`);
   */
  static async clear_pipeline_cache(task, modelId, options = {}) {
    return clear_pipeline_cache(task, modelId, options);
  }
};
export {
  ASTFeatureExtractor,
  ASTForAudioClassification,
  ASTModel,
  ASTPreTrainedModel,
  AfmoeForCausalLM,
  AfmoeModel,
  AfmoePreTrainedModel,
  AlbertForMaskedLM,
  AlbertForQuestionAnswering,
  AlbertForSequenceClassification,
  AlbertModel,
  AlbertPreTrainedModel,
  AlbertTokenizer,
  ApertusForCausalLM,
  ApertusModel,
  ApertusPreTrainedModel,
  ArceeForCausalLM,
  ArceeModel,
  ArceePreTrainedModel,
  AudioClassificationPipeline,
  AutoConfig,
  AutoFeatureExtractor,
  AutoImageProcessor,
  AutoModel,
  AutoModelForAudioClassification,
  AutoModelForAudioFrameClassification,
  AutoModelForAudioTextToText,
  AutoModelForCTC,
  AutoModelForCausalLM,
  AutoModelForDepthEstimation,
  AutoModelForDocumentQuestionAnswering,
  AutoModelForImageClassification,
  AutoModelForImageFeatureExtraction,
  AutoModelForImageMatting,
  AutoModelForImageSegmentation,
  AutoModelForImageTextToText,
  AutoModelForImageToImage,
  AutoModelForMaskGeneration,
  AutoModelForMaskedLM,
  AutoModelForNormalEstimation,
  AutoModelForObjectDetection,
  AutoModelForPoseEstimation,
  AutoModelForQuestionAnswering,
  AutoModelForSemanticSegmentation,
  AutoModelForSeq2SeqLM,
  AutoModelForSequenceClassification,
  AutoModelForSpeechSeq2Seq,
  AutoModelForTextToSpectrogram,
  AutoModelForTextToWaveform,
  AutoModelForTokenClassification,
  AutoModelForUniversalSegmentation,
  AutoModelForVision2Seq,
  AutoModelForXVector,
  AutoModelForZeroShotObjectDetection,
  AutoProcessor,
  AutoTokenizer,
  AutomaticSpeechRecognitionPipeline,
  BackgroundRemovalPipeline,
  BartForConditionalGeneration,
  BartForSequenceClassification,
  BartModel,
  BartPretrainedModel,
  BartTokenizer,
  BaseStreamer,
  BeitFeatureExtractor,
  BeitForImageClassification,
  BeitModel,
  BeitPreTrainedModel,
  BertForMaskedLM,
  BertForQuestionAnswering,
  BertForSequenceClassification,
  BertForTokenClassification,
  BertModel,
  BertPreTrainedModel,
  BertTokenizer,
  BitImageProcessor,
  BlenderbotForConditionalGeneration,
  BlenderbotModel,
  BlenderbotPreTrainedModel,
  BlenderbotSmallForConditionalGeneration,
  BlenderbotSmallModel,
  BlenderbotSmallPreTrainedModel,
  BlenderbotSmallTokenizer,
  BlenderbotTokenizer,
  BloomForCausalLM,
  BloomModel,
  BloomPreTrainedModel,
  BloomTokenizer,
  CHMv2ForDepthEstimation,
  CHMv2ImageProcessor,
  CHMv2PreTrainedModel,
  CLIPFeatureExtractor,
  CLIPImageProcessor,
  CLIPModel,
  CLIPPreTrainedModel,
  CLIPSegForImageSegmentation,
  CLIPSegModel,
  CLIPSegPreTrainedModel,
  CLIPTextModel,
  CLIPTextModelWithProjection,
  CLIPTokenizer,
  CLIPVisionModel,
  CLIPVisionModelWithProjection,
  CamembertForMaskedLM,
  CamembertForQuestionAnswering,
  CamembertForSequenceClassification,
  CamembertForTokenClassification,
  CamembertModel,
  CamembertPreTrainedModel,
  CamembertTokenizer,
  ChatterboxFeatureExtractor,
  ChatterboxModel,
  ChatterboxPreTrainedModel,
  ChatterboxProcessor,
  ChineseCLIPFeatureExtractor,
  ChineseCLIPModel,
  ChineseCLIPPreTrainedModel,
  ClapAudioModelWithProjection,
  ClapFeatureExtractor,
  ClapModel,
  ClapPreTrainedModel,
  ClapTextModelWithProjection,
  ClassifierFreeGuidanceLogitsProcessor,
  CodeGenForCausalLM,
  CodeGenModel,
  CodeGenPreTrainedModel,
  CodeGenTokenizer,
  CodeLlamaTokenizer,
  Cohere2ForCausalLM,
  Cohere2Model,
  Cohere2PreTrainedModel,
  CohereAsrFeatureExtractor,
  CohereAsrForConditionalGeneration,
  CohereAsrModel,
  CohereAsrPreTrainedModel,
  CohereAsrProcessor,
  CohereAsrTokenizer,
  CohereForCausalLM,
  CohereModel,
  CoherePreTrainedModel,
  CohereTokenizer,
  ConvBertForMaskedLM,
  ConvBertForQuestionAnswering,
  ConvBertForSequenceClassification,
  ConvBertForTokenClassification,
  ConvBertModel,
  ConvBertPreTrainedModel,
  ConvBertTokenizer,
  ConvNextFeatureExtractor,
  ConvNextForImageClassification,
  ConvNextImageProcessor,
  ConvNextModel,
  ConvNextPreTrainedModel,
  ConvNextV2ForImageClassification,
  ConvNextV2Model,
  ConvNextV2PreTrainedModel,
  DFineForObjectDetection,
  DFineModel,
  DFinePreTrainedModel,
  DINOv3ConvNextModel,
  DINOv3ConvNextPreTrainedModel,
  DINOv3ViTImageProcessor,
  DINOv3ViTModel,
  DINOv3ViTPreTrainedModel,
  DPTFeatureExtractor,
  DPTForDepthEstimation,
  DPTImageProcessor,
  DPTModel,
  DPTPreTrainedModel,
  DacDecoderModel,
  DacDecoderOutput,
  DacEncoderModel,
  DacEncoderOutput,
  DacFeatureExtractor,
  DacModel,
  DacPreTrainedModel,
  DebertaForMaskedLM,
  DebertaForQuestionAnswering,
  DebertaForSequenceClassification,
  DebertaForTokenClassification,
  DebertaModel,
  DebertaPreTrainedModel,
  DebertaTokenizer,
  DebertaV2ForMaskedLM,
  DebertaV2ForQuestionAnswering,
  DebertaV2ForSequenceClassification,
  DebertaV2ForTokenClassification,
  DebertaV2Model,
  DebertaV2PreTrainedModel,
  DebertaV2Tokenizer,
  DecisionTransformerModel,
  DecisionTransformerPreTrainedModel,
  DeepseekV3ForCausalLM,
  DeepseekV3Model,
  DeepseekV3PreTrainedModel,
  DeiTFeatureExtractor,
  DeiTForImageClassification,
  DeiTImageProcessor,
  DeiTModel,
  DeiTPreTrainedModel,
  DepthAnythingForDepthEstimation,
  DepthAnythingPreTrainedModel,
  DepthEstimationPipeline,
  DepthProForDepthEstimation,
  DepthProPreTrainedModel,
  DetrFeatureExtractor,
  DetrForObjectDetection,
  DetrForSegmentation,
  DetrImageProcessor,
  DetrModel,
  DetrObjectDetectionOutput,
  DetrPreTrainedModel,
  DetrSegmentationOutput,
  Dinov2ForImageClassification,
  Dinov2Model,
  Dinov2PreTrainedModel,
  Dinov2WithRegistersForImageClassification,
  Dinov2WithRegistersModel,
  Dinov2WithRegistersPreTrainedModel,
  DistilBertForMaskedLM,
  DistilBertForQuestionAnswering,
  DistilBertForSequenceClassification,
  DistilBertForTokenClassification,
  DistilBertModel,
  DistilBertPreTrainedModel,
  DistilBertTokenizer,
  DocumentQuestionAnsweringPipeline,
  DonutFeatureExtractor,
  DonutImageProcessor,
  DonutSwinModel,
  DonutSwinPreTrainedModel,
  DynamicCache,
  EdgeTamModel,
  EfficientNetForImageClassification,
  EfficientNetImageProcessor,
  EfficientNetModel,
  EfficientNetPreTrainedModel,
  ElectraForMaskedLM,
  ElectraForQuestionAnswering,
  ElectraForSequenceClassification,
  ElectraForTokenClassification,
  ElectraModel,
  ElectraPreTrainedModel,
  ElectraTokenizer,
  EncodecFeatureExtractor,
  EosTokenCriteria,
  Ernie4_5ForCausalLM,
  Ernie4_5Model,
  Ernie4_5PretrainedModel,
  EsmForMaskedLM,
  EsmForSequenceClassification,
  EsmForTokenClassification,
  EsmModel,
  EsmPreTrainedModel,
  EsmTokenizer,
  EuroBertForMaskedLM,
  EuroBertForSequenceClassification,
  EuroBertForTokenClassification,
  EuroBertModel,
  EuroBertPreTrainedModel,
  ExaoneForCausalLM,
  ExaoneModel,
  ExaonePreTrainedModel,
  FalconForCausalLM,
  FalconH1ForCausalLM,
  FalconH1Model,
  FalconH1PreTrainedModel,
  FalconModel,
  FalconPreTrainedModel,
  FalconTokenizer,
  FastViTForImageClassification,
  FastViTModel,
  FastViTPreTrainedModel,
  FeatureExtractionPipeline,
  FeatureExtractor,
  FillMaskPipeline,
  Florence2ForConditionalGeneration,
  Florence2PreTrainedModel,
  Florence2Processor,
  ForcedBOSTokenLogitsProcessor,
  ForcedEOSTokenLogitsProcessor,
  GLPNFeatureExtractor,
  GLPNForDepthEstimation,
  GLPNModel,
  GLPNPreTrainedModel,
  GPT2LMHeadModel,
  GPT2Model,
  GPT2PreTrainedModel,
  GPT2Tokenizer,
  GPTBigCodeForCausalLM,
  GPTBigCodeModel,
  GPTBigCodePreTrainedModel,
  GPTJForCausalLM,
  GPTJModel,
  GPTJPreTrainedModel,
  GPTNeoForCausalLM,
  GPTNeoModel,
  GPTNeoPreTrainedModel,
  GPTNeoXForCausalLM,
  GPTNeoXModel,
  GPTNeoXPreTrainedModel,
  GPTNeoXTokenizer,
  Gemma2ForCausalLM,
  Gemma2Model,
  Gemma2PreTrainedModel,
  Gemma3ForCausalLM,
  Gemma3ForConditionalGeneration,
  Gemma3ImageProcessor,
  Gemma3Model,
  Gemma3PreTrainedModel,
  Gemma3Processor,
  Gemma3nAudioFeatureExtractor,
  Gemma3nForCausalLM,
  Gemma3nForConditionalGeneration,
  Gemma3nPreTrainedModel,
  Gemma3nProcessor,
  Gemma4AudioFeatureExtractor,
  Gemma4ForCausalLM,
  Gemma4ForConditionalGeneration,
  Gemma4ImageProcessor,
  Gemma4Processor,
  GemmaForCausalLM,
  GemmaModel,
  GemmaPreTrainedModel,
  GemmaTokenizer,
  Glm46VImageProcessor,
  Glm46VProcessor,
  GlmForCausalLM,
  GlmModel,
  GlmMoeDsaForCausalLM,
  GlmMoeDsaModel,
  GlmMoeDsaPreTrainedModel,
  GlmOcrForConditionalGeneration,
  GlmPreTrainedModel,
  GptOssForCausalLM,
  GptOssModel,
  GptOssPreTrainedModel,
  GraniteForCausalLM,
  GraniteModel,
  GraniteMoeHybridForCausalLM,
  GraniteMoeHybridModel,
  GraniteMoeHybridPreTrainedModel,
  GranitePreTrainedModel,
  GraniteSpeechFeatureExtractor,
  GraniteSpeechForConditionalGeneration,
  GraniteSpeechProcessor,
  GroundingDinoForObjectDetection,
  GroundingDinoImageProcessor,
  GroundingDinoPreTrainedModel,
  GroundingDinoProcessor,
  GroupViTModel,
  GroupViTPreTrainedModel,
  HeliumForCausalLM,
  HeliumModel,
  HeliumPreTrainedModel,
  HerbertTokenizer,
  HieraForImageClassification,
  HieraModel,
  HieraPreTrainedModel,
  HubertForCTC,
  HubertForSequenceClassification,
  HubertModel,
  HubertPreTrainedModel,
  HunYuanDenseV1ForCausalLM,
  HunYuanDenseV1Model,
  HunYuanDenseV1PreTrainedModel,
  IJepaForImageClassification,
  IJepaModel,
  IJepaPreTrainedModel,
  Idefics3ForConditionalGeneration,
  Idefics3ImageProcessor,
  Idefics3Processor,
  ImageClassificationPipeline,
  ImageFeatureExtractionPipeline,
  ImageProcessor as ImageFeatureExtractor,
  ImageProcessor,
  ImageSegmentationPipeline,
  ImageToImagePipeline,
  ImageToTextPipeline,
  InterruptableStoppingCriteria,
  JAISLMHeadModel,
  JAISModel,
  JAISPreTrainedModel,
  JinaCLIPImageProcessor,
  JinaCLIPModel,
  JinaCLIPPreTrainedModel,
  JinaCLIPProcessor,
  JinaCLIPTextModel,
  JinaCLIPVisionModel,
  Lfm2ForCausalLM,
  Lfm2Model,
  Lfm2MoeForCausalLM,
  Lfm2MoeModel,
  Lfm2MoePreTrainedModel,
  Lfm2PreTrainedModel,
  Lfm2VlForConditionalGeneration,
  Lfm2VlImageProcessor,
  Lfm2VlProcessor,
  LightOnOcrForConditionalGeneration,
  LiteWhisperForConditionalGeneration,
  Llama4ForCausalLM,
  Llama4PreTrainedModel,
  LlamaForCausalLM,
  LlamaModel,
  LlamaPreTrainedModel,
  LlamaTokenizer,
  LlavaForConditionalGeneration,
  LlavaForConditionalGeneration as LlavaOnevisionForConditionalGeneration,
  LlavaOnevisionImageProcessor,
  LlavaPreTrainedModel,
  LlavaProcessor,
  LlavaQwen2ForCausalLM,
  LogLevel,
  LogitsProcessor,
  LogitsProcessorList,
  LogitsWarper,
  LongT5ForConditionalGeneration,
  LongT5Model,
  LongT5PreTrainedModel,
  M2M100ForConditionalGeneration,
  M2M100Model,
  M2M100PreTrainedModel,
  M2M100Tokenizer,
  MBart50Tokenizer,
  MBartForCausalLM,
  MBartForConditionalGeneration,
  MBartForSequenceClassification,
  MBartModel,
  MBartPreTrainedModel,
  MBartTokenizer,
  MPNetForMaskedLM,
  MPNetForQuestionAnswering,
  MPNetForSequenceClassification,
  MPNetForTokenClassification,
  MPNetModel,
  MPNetPreTrainedModel,
  MPNetTokenizer,
  MT5ForConditionalGeneration,
  MT5Model,
  MT5PreTrainedModel,
  MarianMTModel,
  MarianModel,
  MarianPreTrainedModel,
  MarianTokenizer,
  Mask2FormerImageProcessor,
  MaskFormerFeatureExtractor,
  MaskFormerForInstanceSegmentation,
  MaskFormerImageProcessor,
  MaskFormerModel,
  MaskFormerPreTrainedModel,
  MaxLengthCriteria,
  Metric3DForDepthEstimation,
  Metric3DPreTrainedModel,
  Metric3Dv2ForDepthEstimation,
  Metric3Dv2PreTrainedModel,
  MgpstrForSceneTextRecognition,
  MgpstrModelOutput,
  MgpstrPreTrainedModel,
  MgpstrProcessor,
  MgpstrTokenizer,
  MimiDecoderModel,
  MimiDecoderOutput,
  MimiEncoderModel,
  MimiEncoderOutput,
  MimiModel,
  MimiPreTrainedModel,
  MinLengthLogitsProcessor,
  MinNewTokensLengthLogitsProcessor,
  Mistral4ForCausalLM,
  Mistral4Model,
  Mistral4PreTrainedModel,
  MistralForCausalLM,
  MistralModel,
  MistralPreTrainedModel,
  MobileBertForMaskedLM,
  MobileBertForQuestionAnswering,
  MobileBertForSequenceClassification,
  MobileBertModel,
  MobileBertPreTrainedModel,
  MobileBertTokenizer,
  MobileLLMForCausalLM,
  MobileLLMModel,
  MobileLLMPreTrainedModel,
  MobileNetV1FeatureExtractor,
  MobileNetV1ForImageClassification,
  MobileNetV1ForSemanticSegmentation,
  MobileNetV1ImageProcessor,
  MobileNetV1Model,
  MobileNetV1PreTrainedModel,
  MobileNetV2FeatureExtractor,
  MobileNetV2ForImageClassification,
  MobileNetV2ForSemanticSegmentation,
  MobileNetV2ImageProcessor,
  MobileNetV2Model,
  MobileNetV2PreTrainedModel,
  MobileNetV3FeatureExtractor,
  MobileNetV3ForImageClassification,
  MobileNetV3ForSemanticSegmentation,
  MobileNetV3ImageProcessor,
  MobileNetV3Model,
  MobileNetV3PreTrainedModel,
  MobileNetV4FeatureExtractor,
  MobileNetV4ForImageClassification,
  MobileNetV4ForSemanticSegmentation,
  MobileNetV4ImageProcessor,
  MobileNetV4Model,
  MobileNetV4PreTrainedModel,
  MobileViTFeatureExtractor,
  MobileViTForImageClassification,
  MobileViTImageProcessor,
  MobileViTModel,
  MobileViTPreTrainedModel,
  MobileViTV2ForImageClassification,
  MobileViTV2Model,
  MobileViTV2PreTrainedModel,
  ModelRegistry,
  ModernBertDecoderForCausalLM,
  ModernBertDecoderModel,
  ModernBertDecoderPreTrainedModel,
  ModernBertForMaskedLM,
  ModernBertForSequenceClassification,
  ModernBertForTokenClassification,
  ModernBertModel,
  ModernBertPreTrainedModel,
  Moondream1ForConditionalGeneration,
  MoonshineFeatureExtractor,
  MoonshineForConditionalGeneration,
  MoonshineModel,
  MoonshinePreTrainedModel,
  MoonshineProcessor,
  MptForCausalLM,
  MptModel,
  MptPreTrainedModel,
  MultiModalityCausalLM,
  MultiModalityPreTrainedModel,
  MusicgenForCausalLM,
  MusicgenForConditionalGeneration,
  MusicgenModel,
  MusicgenPreTrainedModel,
  NanoChatForCausalLM,
  NanoChatModel,
  NanoChatPreTrainedModel,
  NemotronHForCausalLM,
  NemotronHModel,
  NemotronHPreTrainedModel,
  NeoBertForMaskedLM,
  NeoBertForQuestionAnswering,
  NeoBertForSequenceClassification,
  NeoBertForTokenClassification,
  NeoBertModel,
  NeoBertPreTrainedModel,
  NllbTokenizer,
  NoBadWordsLogitsProcessor,
  NoRepeatNGramLogitsProcessor,
  NomicBertModel,
  NomicBertPreTrainedModel,
  NougatImageProcessor,
  NougatTokenizer,
  OPTForCausalLM,
  OPTModel,
  OPTPreTrainedModel,
  ObjectDetectionPipeline,
  Olmo2ForCausalLM,
  Olmo2Model,
  Olmo2PreTrainedModel,
  Olmo3ForCausalLM,
  Olmo3Model,
  Olmo3PreTrainedModel,
  OlmoForCausalLM,
  OlmoHybridForCausalLM,
  OlmoHybridModel,
  OlmoHybridPreTrainedModel,
  OlmoModel,
  OlmoPreTrainedModel,
  OpenAIPrivacyFilterForTokenClassification,
  OpenAIPrivacyFilterModel,
  OpenAIPrivacyFilterPreTrainedModel,
  OpenELMForCausalLM,
  OpenELMModel,
  OpenELMPreTrainedModel,
  OwlViTFeatureExtractor,
  OwlViTForObjectDetection,
  OwlViTImageProcessor,
  OwlViTModel,
  OwlViTPreTrainedModel,
  OwlViTProcessor,
  Owlv2ForObjectDetection,
  Owlv2ImageProcessor,
  Owlv2Model,
  Owlv2PreTrainedModel,
  PaliGemmaForConditionalGeneration,
  PaliGemmaProcessor,
  ParakeetFeatureExtractor,
  ParakeetForCTC,
  ParakeetPreTrainedModel,
  PatchTSMixerForPrediction,
  PatchTSMixerModel,
  PatchTSMixerPreTrainedModel,
  PatchTSTForPrediction,
  PatchTSTModel,
  PatchTSTPreTrainedModel,
  Phi3ForCausalLM,
  Phi3Model,
  Phi3PreTrainedModel,
  Phi3VForCausalLM,
  Phi3VImageProcessor,
  Phi3VPreTrainedModel,
  Phi3VProcessor,
  PhiForCausalLM,
  PhiModel,
  PhiPreTrainedModel,
  PixtralImageProcessor,
  PixtralProcessor,
  PreTrainedModel,
  PreTrainedTokenizer,
  PretrainedConfig,
  Processor,
  PvtForImageClassification,
  PvtImageProcessor,
  PvtModel,
  PvtPreTrainedModel,
  PyAnnoteFeatureExtractor,
  PyAnnoteForAudioFrameClassification,
  PyAnnoteModel,
  PyAnnotePreTrainedModel,
  PyAnnoteProcessor,
  QuestionAnsweringPipeline,
  Qwen2ForCausalLM,
  Qwen2Model,
  Qwen2MoeForCausalLM,
  Qwen2MoeModel,
  Qwen2MoePreTrainedModel,
  Qwen2PreTrainedModel,
  Qwen2Tokenizer,
  Qwen2VLForCausalLM,
  Qwen2VLForConditionalGeneration,
  Qwen2VLImageProcessor,
  Qwen2VLPreTrainedModel,
  Qwen2VLProcessor,
  Qwen2_5_VLForCausalLM,
  Qwen2_5_VLForConditionalGeneration,
  Qwen2_5_VLProcessor,
  Qwen3ForCausalLM,
  Qwen3Model,
  Qwen3MoeForCausalLM,
  Qwen3MoeModel,
  Qwen3MoePreTrainedModel,
  Qwen3NextForCausalLM,
  Qwen3NextModel,
  Qwen3NextPreTrainedModel,
  Qwen3PreTrainedModel,
  Qwen3VLForCausalLM,
  Qwen3VLForConditionalGeneration,
  Qwen3VLMoeForCausalLM,
  Qwen3VLMoeForConditionalGeneration,
  Qwen3VLProcessor,
  Qwen3_5ForCausalLM,
  Qwen3_5ForConditionalGeneration,
  Qwen3_5MoeForCausalLM,
  Qwen3_5MoeForConditionalGeneration,
  RFDetrForObjectDetection,
  RFDetrModel,
  RFDetrObjectDetectionOutput,
  RFDetrPreTrainedModel,
  RTDetrForObjectDetection,
  RTDetrImageProcessor,
  RTDetrModel,
  RTDetrObjectDetectionOutput,
  RTDetrPreTrainedModel,
  RTDetrV2ForObjectDetection,
  RTDetrV2Model,
  RTDetrV2ObjectDetectionOutput,
  RTDetrV2PreTrainedModel,
  RawAudio,
  RawImage,
  RawVideo,
  RawVideoFrame,
  RepetitionPenaltyLogitsProcessor,
  ResNetForImageClassification,
  ResNetModel,
  ResNetPreTrainedModel,
  RoFormerForMaskedLM,
  RoFormerForQuestionAnswering,
  RoFormerForSequenceClassification,
  RoFormerForTokenClassification,
  RoFormerModel,
  RoFormerPreTrainedModel,
  RoFormerTokenizer,
  RobertaForMaskedLM,
  RobertaForQuestionAnswering,
  RobertaForSequenceClassification,
  RobertaForTokenClassification,
  RobertaModel,
  RobertaPreTrainedModel,
  RobertaTokenizer,
  SamImageProcessor as Sam2ImageProcessor,
  Sam2ImageSegmentationOutput,
  Sam2Model,
  Sam2PreTrainedModel,
  Sam2Processor,
  Sam2VideoProcessor,
  SamImageProcessor as Sam3ImageProcessor,
  Sam3TrackerModel,
  SamImageProcessor,
  SamImageSegmentationOutput,
  SamModel,
  SamPreTrainedModel,
  SamProcessor,
  SapiensFeatureExtractor,
  SapiensForDepthEstimation,
  SapiensForNormalEstimation,
  SapiensForSemanticSegmentation,
  SapiensImageProcessor,
  SapiensPreTrainedModel,
  SeamlessM4TFeatureExtractor,
  SegformerFeatureExtractor,
  SegformerForImageClassification,
  SegformerForSemanticSegmentation,
  SegformerImageProcessor,
  SegformerModel,
  SegformerPreTrainedModel,
  SiglipImageProcessor,
  SiglipModel,
  SiglipPreTrainedModel,
  SiglipTextModel,
  SiglipTokenizer,
  SiglipVisionModel,
  SmolLM3ForCausalLM,
  SmolLM3Model,
  SmolLM3PreTrainedModel,
  SmolVLMForConditionalGeneration,
  Idefics3ImageProcessor as SmolVLMImageProcessor,
  Idefics3Processor as SmolVLMProcessor,
  SnacDecoderModel,
  SnacEncoderModel,
  SnacFeatureExtractor,
  SnacModel,
  SnacPreTrainedModel,
  SolarOpenForCausalLM,
  SolarOpenModel,
  SolarOpenPreTrainedModel,
  SpeechT5FeatureExtractor,
  SpeechT5ForSpeechToText,
  SpeechT5ForTextToSpeech,
  SpeechT5HifiGan,
  SpeechT5Model,
  SpeechT5PreTrainedModel,
  SpeechT5Processor,
  SpeechT5Tokenizer,
  SqueezeBertForMaskedLM,
  SqueezeBertForQuestionAnswering,
  SqueezeBertForSequenceClassification,
  SqueezeBertModel,
  SqueezeBertPreTrainedModel,
  SqueezeBertTokenizer,
  StableLmForCausalLM,
  StableLmModel,
  StableLmPreTrainedModel,
  Starcoder2ForCausalLM,
  Starcoder2Model,
  Starcoder2PreTrainedModel,
  StoppingCriteria,
  StoppingCriteriaList,
  StyleTextToSpeech2Model,
  StyleTextToSpeech2PreTrainedModel,
  SummarizationPipeline,
  SupertonicForConditionalGeneration,
  SupertonicPreTrainedModel,
  SuppressTokensAtBeginLogitsProcessor,
  SuppressTokensLogitsProcessor,
  Swin2SRForImageSuperResolution,
  Swin2SRImageProcessor,
  Swin2SRModel,
  Swin2SRPreTrainedModel,
  SwinForImageClassification,
  SwinForSemanticSegmentation,
  SwinModel,
  SwinPreTrainedModel,
  T5ForConditionalGeneration,
  T5Model,
  T5PreTrainedModel,
  T5Tokenizer,
  TableTransformerForObjectDetection,
  TableTransformerModel,
  TableTransformerObjectDetectionOutput,
  TableTransformerPreTrainedModel,
  TemperatureLogitsWarper,
  Tensor2 as Tensor,
  Text2TextGenerationPipeline,
  TextClassificationPipeline,
  TextGenerationPipeline,
  TextStreamer,
  TextToAudioPipeline,
  TokenClassificationPipeline,
  PreTrainedTokenizer as TokenizersBackend,
  TopKLogitsWarper,
  TopPLogitsWarper,
  TrOCRForCausalLM,
  TrOCRPreTrainedModel,
  TranslationPipeline,
  UltravoxModel,
  UltravoxPreTrainedModel,
  UltravoxProcessor,
  UniSpeechForCTC,
  UniSpeechForSequenceClassification,
  UniSpeechModel,
  UniSpeechPreTrainedModel,
  UniSpeechSatForAudioFrameClassification,
  UniSpeechSatForCTC,
  UniSpeechSatForSequenceClassification,
  UniSpeechSatModel,
  UniSpeechSatPreTrainedModel,
  VLChatProcessor,
  VLMImageProcessor,
  VaultGemmaForCausalLM,
  VaultGemmaModel,
  VaultGemmaPreTrainedModel,
  ViTFeatureExtractor,
  ViTForImageClassification,
  ViTImageProcessor,
  ViTMAEModel,
  ViTMAEPreTrainedModel,
  ViTMSNForImageClassification,
  ViTMSNModel,
  ViTMSNPreTrainedModel,
  ViTModel,
  ViTPreTrainedModel,
  VisionEncoderDecoderModel,
  VitMatteForImageMatting,
  VitMatteImageProcessor,
  VitMattePreTrainedModel,
  VitPoseForPoseEstimation,
  VitPoseImageProcessor,
  VitPosePreTrainedModel,
  VitsModel,
  VitsModelOutput,
  VitsPreTrainedModel,
  VitsTokenizer,
  VoxtralForConditionalGeneration,
  VoxtralProcessor,
  VoxtralRealtimeFeatureExtractor,
  VoxtralRealtimeForConditionalGeneration,
  VoxtralRealtimePreTrainedModel,
  VoxtralRealtimeProcessor,
  Wav2Vec2BertForCTC,
  Wav2Vec2BertForSequenceClassification,
  Wav2Vec2BertModel,
  Wav2Vec2BertPreTrainedModel,
  Wav2Vec2CTCTokenizer,
  Wav2Vec2FeatureExtractor,
  Wav2Vec2ForAudioFrameClassification,
  Wav2Vec2ForCTC,
  Wav2Vec2ForSequenceClassification,
  Wav2Vec2Model,
  Wav2Vec2PreTrainedModel,
  Wav2Vec2Processor,
  Wav2Vec2ProcessorWithLM,
  WavLMForAudioFrameClassification,
  WavLMForCTC,
  WavLMForSequenceClassification,
  WavLMForXVector,
  WavLMModel,
  WavLMPreTrainedModel,
  WeSpeakerFeatureExtractor,
  WeSpeakerResNetModel,
  WeSpeakerResNetPreTrainedModel,
  WhisperFeatureExtractor,
  WhisperForConditionalGeneration,
  WhisperModel,
  WhisperPreTrainedModel,
  WhisperProcessor,
  WhisperTextStreamer,
  WhisperTimeStampLogitsProcessor,
  WhisperTokenizer,
  XLMForQuestionAnswering,
  XLMForSequenceClassification,
  XLMForTokenClassification,
  XLMModel,
  XLMPreTrainedModel,
  XLMRobertaForMaskedLM,
  XLMRobertaForQuestionAnswering,
  XLMRobertaForSequenceClassification,
  XLMRobertaForTokenClassification,
  XLMRobertaModel,
  XLMRobertaPreTrainedModel,
  XLMRobertaTokenizer,
  XLMTokenizer,
  XLMWithLMHeadModel,
  XVectorOutput,
  YolosFeatureExtractor,
  YolosForObjectDetection,
  YolosImageProcessor,
  YolosModel,
  YolosObjectDetectionOutput,
  YolosPreTrainedModel,
  YoutuForCausalLM,
  YoutuModel,
  YoutuPreTrainedModel,
  ZeroShotAudioClassificationPipeline,
  ZeroShotClassificationPipeline,
  ZeroShotImageClassificationPipeline,
  ZeroShotObjectDetectionPipeline,
  cat,
  cos_sim,
  dot,
  env,
  full,
  full_like,
  interpolate,
  interpolate_4d,
  layer_norm,
  load_audio,
  load_image,
  load_video,
  log_softmax,
  matmul,
  mean,
  mean_pooling,
  ones,
  ones_like,
  permute,
  pipeline,
  quantize_embeddings,
  rand,
  randn,
  random,
  read_audio,
  rfft,
  slice2 as slice,
  softmax,
  stack,
  std_mean,
  topk,
  zeros,
  zeros_like
};
/*! Bundled license information:

onnxruntime-web/dist/ort.webgpu.bundle.min.mjs:
  (*!
   * ONNX Runtime Web v1.26.0-dev.20260416-b7804b056c
   * Copyright (c) Microsoft Corporation. All rights reserved.
   * Licensed under the MIT License.
   *)
*/