webgpu_bridge/llama_webgpu_bridge.js

// Generated by scripts/build_js_bridge.mjs from js/src/llama_webgpu_bridge.js.
// Do not edit this file directly; update js/src/ and run npm run build:js.

// js/src/internal/abort.ts
function createAbortError(message) {
  if (typeof DOMException === "function") {
    return new DOMException(message, "AbortError");
  }
  const error = new Error(message);
  error.name = "AbortError";
  return error;
}
function throwIfAborted(signal, message = "Bridge operation was cancelled.") {
  if (signal?.aborted) {
    throw createAbortError(message);
  }
}

// js/src/internal/embedding.ts
var EmbeddingError = class extends Error {
  constructor(code, message) {
    super(message);
    this.name = "EmbeddingError";
    this.code = code;
  }
};
var EMPTY_EMBEDDING_CAPABILITIES = {
  apiVersion: 1,
  text: false,
  image: false,
  audio: false
};
var EMBEDDING_MEDIA_MARKER = "<__media__>";
function embeddingAssetVersion(...sources) {
  for (const source of [...sources, import.meta.url]) {
    try {
      const version = new URL(source || "", import.meta.url).pathname.match(/(?:^|[/@])(v?\d+\.\d+\.\d+(?:-\d+)?)(?:[/]|$)/)?.[1];
      if (version) return version;
    } catch (_) {
    }
  }
  return "unversioned/local";
}
function missingEmbeddingCapability(version = embeddingAssetVersion()) {
  return new EmbeddingError(
    "EMBEDDING_CAPABILITY_UNAVAILABLE",
    `Loaded bridge/assets version ${version} has no known media-embedding capability API version 1. Upgrade the loaded bridge assets.`
  );
}
function embeddingCapabilitiesFrom(value, version = embeddingAssetVersion()) {
  let parsed = value;
  if (typeof value === "string") {
    try {
      parsed = JSON.parse(value);
    } catch (_) {
      throw missingEmbeddingCapability(version);
    }
  }
  const caps = parsed;
  if (!caps || caps.apiVersion !== 1 || typeof caps.text !== "boolean" || typeof caps.image !== "boolean" || typeof caps.audio !== "boolean") {
    throw missingEmbeddingCapability(version);
  }
  return { apiVersion: 1, text: caps.text, image: caps.image, audio: caps.audio };
}
function snapshotEmbedInput(input) {
  if (!Array.isArray(input)) return String(input);
  if (!Array.isArray(input) || input.length === 0) {
    throw new EmbeddingError("INVALID_EMBEDDING_INPUT", "Embedding parts must be a nonempty array.");
  }
  const parts = Array.from(input, (part) => {
    if (!part || typeof part !== "object") {
      throw new EmbeddingError("INVALID_EMBEDDING_INPUT", "Invalid embedding part.");
    }
    if (part.type === "text") {
      if (typeof part.text !== "string") {
        throw new EmbeddingError("INVALID_EMBEDDING_INPUT", "Embedding text part must contain a string.");
      }
      return { type: "text", text: part.text };
    }
    if (part.type !== "image" && part.type !== "audio") {
      throw new EmbeddingError("EMBEDDING_MODALITY_UNSUPPORTED", "Embedding part modality is unsupported.");
    }
    if ("path" in part) {
      throw new EmbeddingError("EMBEDDING_MODALITY_UNSUPPORTED", "Browser embeddings require media bytes, URL or PCM; filesystem paths are unsupported.");
    }
    const sourceCount = Number(part.bytes != null) + Number(part.url != null) + Number("samples" in part && part.samples != null);
    if (sourceCount !== 1 || part.bytes != null && (!(part.bytes instanceof Uint8Array) || part.bytes.length === 0) || part.url != null && (typeof part.url !== "string" || part.url.length === 0)) {
      throw new EmbeddingError("INVALID_EMBEDDING_INPUT", "Embedding media requires exactly one nonempty supported source.");
    }
    if ("samples" in part && part.samples != null) {
      if (part.type !== "audio" || !(part.samples instanceof Float32Array) || part.samples.length === 0 || !part.samples.every(Number.isFinite)) {
        throw new EmbeddingError("INVALID_EMBEDDING_INPUT", "Embedding PCM must be finite mono 16 kHz Float32 samples.");
      }
    }
    return { ...part };
  });
  const merged = [];
  for (const part of parts) {
    const previous = merged.at(-1);
    if (part.type === "text" && previous?.type === "text") previous.text += part.text;
    else merged.push(part);
  }
  if (merged.some((part) => part.type === "text" && part.text.includes(EMBEDDING_MEDIA_MARKER))) {
    throw new EmbeddingError("INVALID_EMBEDDING_INPUT", "Reserved media marker in embedding text.");
  }
  return merged;
}
function requireEmbeddingModalities(input, caps, hasProjector) {
  if (typeof input === "string") return;
  if (!caps.text) {
    throw new EmbeddingError("MODEL_NOT_EMBEDDING", "Loaded model does not support pooled embeddings.");
  }
  for (const part of input) {
    if (part.type === "text") continue;
    if (!hasProjector) {
      throw new EmbeddingError("EMBEDDING_PROJECTOR_MISSING", "Media embeddings require a loaded projector.");
    }
    if (!caps[part.type]) {
      throw new EmbeddingError("EMBEDDING_MODALITY_UNSUPPORTED", "Loaded projector/model does not support the requested embedding modality.");
    }
  }
}
function embeddingCoreError(rc, version = embeddingAssetVersion()) {
  if (rc === -20) return missingEmbeddingCapability(version);
  const errors = {
    [-21]: ["MODEL_NOT_EMBEDDING", "Loaded model does not support this pooled embedding operation."],
    [-22]: ["EMBEDDING_PROJECTOR_MISSING", "Media embeddings require a loaded projector."],
    [-23]: ["EMBEDDING_INPUT_TOO_LARGE", "Expanded embedding input exceeds the allocated context, batch, sequence or single microbatch capacity."],
    [-24]: ["EMBEDDING_MODALITY_UNSUPPORTED", "Loaded projector does not support the requested embedding modality or PCM rate."]
  };
  const [code, message] = errors[rc] || ["EMBEDDING_FAILED", "Embedding execution failed."];
  return new EmbeddingError(code, message);
}

// js/src/internal/completion_options.ts
var SPECULATIVE_DECODING_STRATEGIES = Object.freeze([
  "draft-simple",
  "draft-eagle3",
  "draft-mtp",
  "draft-dflash",
  "draft-dspark",
  "ngram-simple",
  "ngram-map-k",
  "ngram-map-k4v",
  "ngram-mod",
  "ngram-cache"
]);
function speculativeFlags(isSupported) {
  return Object.fromEntries(
    SPECULATIVE_DECODING_STRATEGIES.map((strategy) => [strategy, isSupported(strategy)])
  );
}
var NO_COMPLETION_CAPABILITIES = Object.freeze({
  presencePenalty: false,
  minP: false,
  thinkingBudget: false,
  speculativeDecoding: Object.freeze(speculativeFlags(() => false))
});
var MAX_NGRAM_SIZE = 65535;
var MAX_INT32 = 2147483647;
var MAX_THINKING_BUDGET_TOKENS = 2147483647;
function optionalNumber(value, name, isValid, range) {
  if (value == null) {
    return null;
  }
  if (typeof value !== "number" || !Number.isFinite(Math.fround(value)) || !isValid(value)) {
    throw new RangeError(`CompletionOptions.${name} must be a finite number${range}; got ${String(value)}.`);
  }
  return value;
}
function tag(budget, name) {
  const value = budget[name];
  if (typeof value !== "string" || value.trim().length === 0) {
    throw new TypeError(`CompletionOptions.thinkingBudget.${name} must be a non-empty string.`);
  }
  return value;
}
function thinkingBudget(value, hasMediaParts) {
  if (value == null) {
    return null;
  }
  if (typeof value !== "object") {
    throw new TypeError("CompletionOptions.thinkingBudget must be an object.");
  }
  const budget = value;
  const maxTokens = budget.maxTokens;
  if (!Number.isInteger(maxTokens) || maxTokens < 0 || maxTokens > MAX_THINKING_BUDGET_TOKENS) {
    throw new RangeError(
      `CompletionOptions.thinkingBudget.maxTokens must be an integer from 0 to ${MAX_THINKING_BUDGET_TOKENS}; got ${String(maxTokens)}.`
    );
  }
  const startTag = tag(budget, "startTag");
  const endTag = tag(budget, "endTag");
  const forcedMessage = budget.forcedMessage ?? "";
  if (typeof forcedMessage !== "string") {
    throw new TypeError("CompletionOptions.thinkingBudget.forcedMessage must be a string.");
  }
  if (hasMediaParts) {
    throw new Error("CompletionOptions.thinkingBudget supports text-only prompts; remove the media parts.");
  }
  return { maxTokens, startTag, endTag, forcedMessage };
}
function speculativeInteger(config, name, min, max) {
  const value = config[name];
  if (value == null) {
    return null;
  }
  if (!Number.isInteger(value) || value < min || value > max) {
    throw new RangeError(
      `CompletionOptions.speculativeDecoding.${name} must be an integer from ${min} to ${max}; got ${String(value)}.`
    );
  }
  return value;
}
function speculativeProbability(config, name) {
  const value = config[name];
  if (value == null) {
    return null;
  }
  if (typeof value !== "number" || !(value >= 0 && value <= 1)) {
    throw new RangeError(
      `CompletionOptions.speculativeDecoding.${name} must be a number from 0 to 1; got ${String(value)}.`
    );
  }
  return value;
}
function ngramCacheSource(config, name) {
  const value = config[name];
  if (value == null) {
    return null;
  }
  const isBytes = value instanceof ArrayBuffer || ArrayBuffer.isView(value);
  if (!(typeof value === "string" || isBytes)) {
    throw new TypeError(
      `CompletionOptions.speculativeDecoding.${name} must be a URL string, an ArrayBuffer or a typed array.`
    );
  }
  if ((typeof value === "string" ? value.length : value.byteLength) === 0) {
    throw new TypeError(`CompletionOptions.speculativeDecoding.${name} is empty.`);
  }
  return value;
}
function isDraftStrategy(strategy) {
  return strategy.startsWith("draft-");
}
function speculativeDecoding(value, conflicts) {
  if (value == null) {
    return null;
  }
  if (typeof value !== "object") {
    throw new TypeError("CompletionOptions.speculativeDecoding must be an object.");
  }
  const config = value;
  const requested = config.strategies;
  if (!Array.isArray(requested) || requested.length === 0) {
    throw new TypeError("CompletionOptions.speculativeDecoding.strategies must be a non-empty array.");
  }
  const strategies = [];
  for (const strategy of requested) {
    if (!SPECULATIVE_DECODING_STRATEGIES.includes(strategy)) {
      throw new TypeError(
        `CompletionOptions.speculativeDecoding.strategies has an unknown strategy ${JSON.stringify(strategy)}; expected one of ${SPECULATIVE_DECODING_STRATEGIES.join(", ")}.`
      );
    }
    if (!strategies.includes(strategy)) {
      strategies.push(strategy);
    }
  }
  const draftTokenMax = speculativeInteger(config, "draftTokenMax", 0, MAX_INT32);
  const draftTokenMin = speculativeInteger(config, "draftTokenMin", 0, MAX_INT32);
  const minProbability = speculativeProbability(config, "minProbability");
  const draftSplitProbability = speculativeProbability(config, "draftSplitProbability");
  const ngramSizeN = speculativeInteger(config, "ngramSizeN", 1, MAX_NGRAM_SIZE);
  const ngramSizeM = speculativeInteger(config, "ngramSizeM", 1, MAX_NGRAM_SIZE);
  const ngramMinHits = speculativeInteger(config, "ngramMinHits", 1, MAX_NGRAM_SIZE);
  const ngramMatch = speculativeInteger(config, "ngramMatch", 1, MAX_NGRAM_SIZE);
  const ngramTokenMin = speculativeInteger(config, "ngramTokenMin", 0, MAX_INT32);
  const ngramTokenMax = speculativeInteger(config, "ngramTokenMax", 0, MAX_INT32);
  const ngramCacheStatic = ngramCacheSource(config, "ngramCacheStatic");
  const ngramCacheDynamic = ngramCacheSource(config, "ngramCacheDynamic");
  if (strategies.filter(isDraftStrategy).length > 1) {
    throw new Error(
      "CompletionOptions.speculativeDecoding can mix n-gram strategies with at most one draft-* strategy."
    );
  }
  if (!strategies.some(isDraftStrategy) && (draftTokenMin != null || minProbability != null || draftSplitProbability != null)) {
    throw new Error(
      "CompletionOptions.speculativeDecoding: n-gram strategies use token history and do not support draftTokenMin, minProbability or draftSplitProbability unless a draft-* strategy is also enabled."
    );
  }
  if ((ngramCacheStatic != null || ngramCacheDynamic != null) && !strategies.includes("ngram-cache")) {
    throw new Error(
      "CompletionOptions.speculativeDecoding: ngramCacheStatic and ngramCacheDynamic need the ngram-cache strategy."
    );
  }
  if (ngramTokenMin != null && ngramTokenMax != null && ngramTokenMin > ngramTokenMax) {
    throw new RangeError(
      `CompletionOptions.speculativeDecoding.ngramTokenMin (${ngramTokenMin}) must not exceed ngramTokenMax (${ngramTokenMax}).`
    );
  }
  let resolvedDraftTokenMax = 0;
  for (const strategy of strategies) {
    let strategyMax;
    if (isDraftStrategy(strategy)) {
      strategyMax = draftTokenMax ?? 3;
    } else if (strategy === "ngram-mod") {
      strategyMax = ngramTokenMax ?? draftTokenMax ?? 64;
    } else if (strategy === "ngram-cache") {
      strategyMax = draftTokenMax ?? 8;
    } else {
      strategyMax = ngramSizeM ?? 48;
    }
    resolvedDraftTokenMax = Math.max(resolvedDraftTokenMax, strategyMax);
  }
  if (resolvedDraftTokenMax === 0) {
    resolvedDraftTokenMax = 64;
  }
  if ((draftTokenMin ?? 0) > resolvedDraftTokenMax) {
    throw new RangeError(
      `CompletionOptions.speculativeDecoding.draftTokenMin (${draftTokenMin}) must not exceed the draft token maximum (${resolvedDraftTokenMax}).`
    );
  }
  if (conflicts.hasMediaParts) {
    throw new Error("CompletionOptions.speculativeDecoding supports text-only prompts; remove the media parts.");
  }
  if (conflicts.hasThinkingBudget) {
    throw new Error("CompletionOptions.speculativeDecoding cannot be combined with CompletionOptions.thinkingBudget.");
  }
  if (conflicts.hasGrammar) {
    throw new Error("CompletionOptions.speculativeDecoding does not support CompletionOptions.grammar.");
  }
  return {
    strategies,
    draftTokenMax: resolvedDraftTokenMax,
    draftTokenMin: draftTokenMin ?? 0,
    minProbability: minProbability ?? -1,
    draftSplitProbability: draftSplitProbability ?? -1,
    ngramSizeN: ngramSizeN ?? 0,
    ngramSizeM: ngramSizeM ?? 0,
    ngramMinHits: ngramMinHits ?? 0,
    ngramMatch: ngramMatch ?? 0,
    ngramTokenMin: ngramTokenMin ?? -1,
    ngramTokenMax: ngramTokenMax ?? (strategies.includes("ngram-mod") ? draftTokenMax ?? 0 : 0),
    ngramCacheStatic,
    ngramCacheDynamic
  };
}
function resolveCompletionSamplingOptions(options) {
  const hasMediaParts = Array.isArray(options?.parts) && options.parts.length > 0;
  const budget = thinkingBudget(options?.thinkingBudget, hasMediaParts);
  return {
    minP: optionalNumber(options?.minP, "minP", (value) => value >= 0 && value <= 1, " from 0 to 1") ?? 0,
    presencePenalty: optionalNumber(options?.presencePenalty, "presencePenalty", () => true, "") ?? 0,
    thinkingBudget: budget,
    speculative: speculativeDecoding(options?.speculativeDecoding, {
      hasMediaParts,
      hasGrammar: typeof options?.grammar === "string" && options.grammar.length > 0,
      hasThinkingBudget: budget != null
    })
  };
}
function completionCapabilitiesFrom(raw) {
  let parsed = null;
  try {
    parsed = JSON.parse(raw || "{}");
  } catch (_) {
    return { ...NO_COMPLETION_CAPABILITIES };
  }
  const speculative = parsed?.speculativeDecoding;
  return {
    presencePenalty: parsed?.presencePenalty === true,
    minP: parsed?.minP === true,
    thinkingBudget: parsed?.thinkingBudget === true,
    speculativeDecoding: speculativeFlags((strategy) => speculative?.[strategy] === true)
  };
}
function requireCompletionCapabilities(sampling, capabilities) {
  const missing = [
    sampling.minP !== 0 && !capabilities.minP ? "minP" : null,
    sampling.presencePenalty !== 0 && !capabilities.presencePenalty ? "presencePenalty" : null,
    sampling.thinkingBudget != null && !capabilities.thinkingBudget ? "thinkingBudget" : null,
    sampling.speculative != null && !Object.values(capabilities.speculativeDecoding).includes(true) ? "speculativeDecoding" : null
  ].filter((name) => name != null);
  if (missing.length > 0) {
    throw new Error(
      `The loaded WebGPU core does not support CompletionOptions.${missing.join(" or CompletionOptions.")}.`
    );
  }
}

// js/src/internal/constants.ts
var defaultModelCacheName = "llamadart-webgpu-model-cache-v1";
var BRIDGE_DISPOSED_MESSAGE = "Bridge has been disposed.";
var GENERATION_ALREADY_ACTIVE_RC = -7;
var GENERATION_ALREADY_ACTIVE_MESSAGE = "Generation is already active on this bridge runtime.";
var INVALID_GRAMMAR_ERROR_TEXT = "Failed to initialize sampler chain (invalid grammar)";

// js/src/internal/typed_values.ts
function toUint8Array(value) {
  if (!value) {
    return null;
  }
  if (value instanceof Uint8Array) {
    return value;
  }
  if (ArrayBuffer.isView(value)) {
    return new Uint8Array(value.buffer, value.byteOffset, value.byteLength);
  }
  if (value instanceof ArrayBuffer) {
    return new Uint8Array(value);
  }
  if (Array.isArray(value)) {
    return Uint8Array.from(value.map((v) => Number(v) & 255));
  }
  return null;
}
function bufferSourceBytes(source) {
  if (source instanceof ArrayBuffer) {
    return new Uint8Array(source);
  }
  if (ArrayBuffer.isView(source)) {
    return new Uint8Array(source.buffer, source.byteOffset, source.byteLength);
  }
  return null;
}
function toFloat32Array(value) {
  if (!value) {
    return null;
  }
  if (value instanceof Float32Array) {
    return value;
  }
  if (ArrayBuffer.isView(value)) {
    return new Float32Array(
      value.buffer,
      value.byteOffset,
      Math.floor(value.byteLength / Float32Array.BYTES_PER_ELEMENT)
    );
  }
  if (value instanceof ArrayBuffer) {
    return new Float32Array(value);
  }
  if (Array.isArray(value)) {
    return Float32Array.from(value.map((v) => Number(v) || 0));
  }
  return null;
}
function isInt32(value) {
  return typeof value === "number" && Number.isInteger(value) && value >= -2147483648 && value <= 2147483647;
}

// js/src/internal/decision.ts
var DECISION_API_VERSION = 1;
var DECISION_WORKER_TIMEOUT_PER_SEQUENCE_MS = 60 * 1e3;
function decisionHandleFrom(handle) {
  if (!isInt32(handle) || handle <= 0) {
    throw new TypeError(`Decision head handle must be a positive integer, got ${String(handle)}.`);
  }
  return handle;
}
function decisionIntegerList(value, label) {
  const isList = Array.isArray(value) || ArrayBuffer.isView(value) && !(value instanceof DataView);
  if (!isList) {
    throw new TypeError(`${label} must be an array or typed array of integers.`);
  }
  const list = value;
  const out = new Int32Array(list.length);
  for (let index = 0; index < list.length; index += 1) {
    const item = list[index];
    if (!isInt32(item)) {
      throw new TypeError(`${label}[${index}] is ${String(item)}; expected a 32-bit integer.`);
    }
    out[index] = item;
  }
  return out;
}
function normalizeDecisionSequences(sequences) {
  if (!Array.isArray(sequences)) {
    throw new TypeError("Decision sequences must be an array.");
  }
  return sequences.map((sequence, index) => {
    const label = `Decision sequence ${index}`;
    if (!sequence || typeof sequence !== "object") {
      throw new TypeError(`${label} must be an object with tokens, markers and questionType.`);
    }
    const tokens = decisionIntegerList(sequence.tokens, `${label} tokens`);
    const markers = decisionIntegerList(sequence.markers, `${label} markers`);
    const questionType = sequence.questionType;
    if (!isInt32(questionType)) {
      throw new TypeError(
        `${label} questionType is ${String(questionType)}; expected 0 (choice), 1 (score) or 2 (noul).`
      );
    }
    return { tokens, markers, questionType };
  });
}
function encodeDecisionSequences(sequences) {
  const encoded = normalizeDecisionSequences(sequences);
  let words = 1;
  for (const sequence of encoded) {
    words += 3 + sequence.tokens.length + sequence.markers.length;
  }
  const bytes = new Uint8Array(words * 4);
  const view = new DataView(bytes.buffer);
  let offset = 0;
  const put = (value) => {
    view.setInt32(offset, value, true);
    offset += 4;
  };
  put(encoded.length);
  for (const sequence of encoded) {
    put(sequence.questionType);
    put(sequence.tokens.length);
    put(sequence.markers.length);
    sequence.tokens.forEach(put);
    sequence.markers.forEach(put);
  }
  return bytes;
}
function decodeDecisionOutputs(bytes, count) {
  const malformed = () => new Error("Decision output from the WebGPU core is malformed.");
  const view = new DataView(bytes.buffer, bytes.byteOffset, bytes.byteLength);
  let offset = 0;
  const floats = (length) => {
    const values = new Float32Array(length);
    for (let index = 0; index < length; index += 1) {
      values[index] = view.getFloat32(offset, true);
      offset += 4;
    }
    return values;
  };
  const outputs = [];
  for (let index = 0; index < count; index += 1) {
    if (bytes.byteLength - offset < 8) {
      throw malformed();
    }
    const logitCount = view.getInt32(offset, true);
    const actCount = view.getInt32(offset + 4, true);
    offset += 8;
    if (logitCount < 0 || actCount < 0 || (logitCount + actCount) * 4 > bytes.byteLength - offset) {
      throw malformed();
    }
    outputs.push({ logits: floats(logitCount), actLogits: floats(actCount) });
  }
  if (offset !== bytes.byteLength) {
    throw malformed();
  }
  return outputs;
}

// js/src/internal/logging.ts
function logLevelForName(level) {
  switch (level) {
    case "debug":
      return 0;
    case "log":
    case "info":
      return 1;
    case "warn":
      return 2;
    case "error":
      return 3;
    default:
      return 1;
  }
}
function logThresholdForConfiguredLevel(level) {
  switch (level) {
    case 0:
      return 99;
    case 1:
      return 0;
    case 2:
      return 1;
    case 3:
      return 2;
    case 4:
      return 3;
    default:
      return 1;
  }
}

// js/src/internal/lora.ts
var LORA_API_VERSION = 1;
var LORA_LOAD_ABORT_MESSAGE = "LoRA adapter load was cancelled.";
function loraHandleFrom(handle) {
  if (!isInt32(handle) || handle <= 0) {
    throw new TypeError(`LoRA adapter handle must be a positive integer, got ${String(handle)}.`);
  }
  return handle;
}
function loraScaleFrom(scale) {
  if (typeof scale !== "number" || !Number.isFinite(scale)) {
    throw new TypeError(`LoRA adapter scale must be a finite number, got ${String(scale)}.`);
  }
  if (!Number.isFinite(Math.fround(scale))) {
    throw new RangeError(`LoRA adapter scale ${String(scale)} is outside the 32-bit float range.`);
  }
  return scale;
}
function staleLoraAdapterError(handle) {
  return new Error(
    `LoRA adapter ${handle} is not loaded; its model was unloaded or replaced. Load the adapter again.`
  );
}

// js/src/internal/model_source.ts
function cloneModelSource(source) {
  if (Array.isArray(source)) {
    return source.slice();
  }
  return String(source || "").trim();
}
function hasModelSource(source) {
  return Array.isArray(source) ? source.length > 0 : typeof source === "string" && source.length > 0;
}
function basenameFromUrl(url) {
  try {
    const parsed = new URL(url, typeof window !== "undefined" ? window.location.href : void 0);
    const pathname = parsed.pathname || "";
    const name = pathname.split("/").pop() || "model.gguf";
    return name.includes("?") ? name.split("?")[0] : name;
  } catch (_) {
    const parts = String(url).split("/");
    return parts[parts.length - 1] || "model.gguf";
  }
}
function normalizeAbsoluteUrl(url) {
  try {
    return new URL(url, typeof window !== "undefined" ? window.location.href : void 0).toString();
  } catch (_) {
    return String(url);
  }
}
function parseSplitShardPattern(fileName) {
  if (typeof fileName !== "string" || fileName.length === 0) {
    return null;
  }
  const match = fileName.match(/^(.*)-(\d{4,6})-of-(\d{4,6})\.gguf$/i);
  if (!match) {
    return null;
  }
  const total = Number(match[3]);
  if (!Number.isInteger(total) || total < 2 || total > 512) {
    return null;
  }
  return {
    prefix: match[1],
    width: Math.max(match[2].length, match[3].length),
    total
  };
}
function expandModelShardUrls(modelUrlOrUrls) {
  if (Array.isArray(modelUrlOrUrls)) {
    return modelUrlOrUrls.map((value) => String(value || "").trim()).filter((value) => value.length > 0).map((value) => normalizeAbsoluteUrl(value));
  }
  const source = String(modelUrlOrUrls || "").trim();
  if (source.length === 0) {
    return [];
  }
  try {
    const parsed = new URL(source, typeof window !== "undefined" ? window.location.href : void 0);
    const pathname = parsed.pathname || "";
    const slash = pathname.lastIndexOf("/");
    const dirPath = slash >= 0 ? pathname.slice(0, slash + 1) : "";
    const fileName = slash >= 0 ? pathname.slice(slash + 1) : pathname;
    const split = parseSplitShardPattern(fileName);
    if (!split) {
      return [parsed.toString()];
    }
    const totalShardId = String(split.total).padStart(split.width, "0");
    const urls = [];
    for (let shardIndex = 1; shardIndex <= split.total; shardIndex += 1) {
      const shardId = String(shardIndex).padStart(split.width, "0");
      parsed.pathname = `${dirPath}${split.prefix}-${shardId}-of-${totalShardId}.gguf`;
      urls.push(parsed.toString());
    }
    return urls;
  } catch (_) {
    return [source];
  }
}

// js/src/internal/core_loader.ts
function normalizeFactory(moduleExport) {
  if (typeof moduleExport === "function") {
    return moduleExport;
  }
  if (moduleExport && typeof moduleExport.default === "function") {
    return moduleExport.default;
  }
  if (moduleExport && typeof moduleExport.createLlamaWebGpuCoreModule === "function") {
    return moduleExport.createLlamaWebGpuCoreModule;
  }
  throw new Error("Unable to resolve llama_webgpu_core factory function");
}
async function importCoreFactory(moduleUrl) {
  const exportedModule = await import(moduleUrl);
  return normalizeFactory(exportedModule);
}

// js/src/internal/parse.ts
function parsePositiveInteger(value) {
  const numeric = Number(value);
  if (!Number.isFinite(numeric) || numeric <= 0) {
    return 0;
  }
  return Math.trunc(numeric);
}
function parseInteger(value, fallback = 0) {
  const numeric = Number(value);
  if (!Number.isFinite(numeric)) {
    return fallback;
  }
  return Math.trunc(numeric);
}
function parseBooleanFlag(value, fallback = false) {
  if (typeof value === "boolean") {
    return value;
  }
  if (typeof value === "number" && Number.isFinite(value)) {
    return value !== 0;
  }
  return fallback;
}
function parseOptionalBooleanFlag(value) {
  if (typeof value === "boolean") {
    return value ? 1 : 0;
  }
  if (typeof value === "number" && Number.isFinite(value)) {
    return value !== 0 ? 1 : 0;
  }
  return -1;
}
function parseEnumValue(value, allowed, fallback) {
  const parsed = parseInteger(value, fallback);
  return allowed.includes(parsed) ? parsed : fallback;
}
function parsePositiveNumber(value) {
  const numeric = Number(value);
  return Number.isFinite(numeric) && numeric > 0 ? numeric : 0;
}

// js/src/internal/download.ts
function hasReadableResponseStream(response) {
  return !!(response && response.body && typeof response.body.getReader === "function");
}
function sumProgressValues(values) {
  let total = 0;
  for (const value of values || []) {
    const numeric = Number(value);
    if (Number.isFinite(numeric) && numeric > 0) {
      total += numeric;
    }
  }
  return total;
}
function parseTotalFromContentRangeHeader(contentRangeHeader) {
  if (typeof contentRangeHeader !== "string" || contentRangeHeader.length === 0) {
    return 0;
  }
  const slash = contentRangeHeader.lastIndexOf("/");
  if (slash < 0 || slash + 1 >= contentRangeHeader.length) {
    return 0;
  }
  return parsePositiveInteger(contentRangeHeader.slice(slash + 1));
}
function inferResponseTotalBytes(response, loadedFallback = 0) {
  if (!response || !response.headers) {
    return parsePositiveInteger(loadedFallback);
  }
  const linkedSize = parsePositiveInteger(response.headers.get("x-linked-size"));
  if (linkedSize > 0) {
    return linkedSize;
  }
  const contentRangeTotal = parseTotalFromContentRangeHeader(
    response.headers.get("content-range")
  );
  if (contentRangeTotal > 0) {
    return contentRangeTotal;
  }
  const contentLength = parsePositiveInteger(response.headers.get("content-length"));
  if (contentLength > 0) {
    return contentLength;
  }
  return parsePositiveInteger(loadedFallback);
}
function isRetryableStreamNetworkError(error) {
  const text = String(error || "").toLowerCase();
  return text.includes("network error") || text.includes("failed to fetch") || text.includes("networkerror") || text.includes("err_network_io_suspended") || text.includes("the network connection was lost") || text.includes("connection reset") || text.includes("timeout") || text.includes("timed out");
}
async function readStreamChunkWithTimeout(reader, timeoutMs, label = "stream read") {
  const resolvedTimeout = Number(timeoutMs);
  if (!Number.isFinite(resolvedTimeout) || resolvedTimeout <= 0) {
    return reader.read();
  }
  let timeoutHandle = null;
  try {
    return await Promise.race([
      reader.read(),
      new Promise((_, reject) => {
        timeoutHandle = globalThis.setTimeout(() => {
          reject(new Error(`${label} timeout (${resolvedTimeout}ms)`));
        }, resolvedTimeout);
      })
    ]);
  } finally {
    if (timeoutHandle != null) {
      globalThis.clearTimeout(timeoutHandle);
    }
  }
}
async function drainResponseWithProgress(response, progressCallback, options = {}) {
  const total = Number(response.headers.get("content-length")) || 0;
  const chunkTimeoutMs = parsePositiveInteger(options.chunkTimeoutMs);
  if (!response.body || typeof response.body.getReader !== "function") {
    const bytes = new Uint8Array(await response.arrayBuffer());
    if (typeof progressCallback === "function") {
      progressCallback({ loaded: bytes.byteLength, total: total || bytes.byteLength });
    }
    return bytes.byteLength;
  }
  const reader = response.body.getReader();
  let loaded = 0;
  let lastBucket = -1;
  while (true) {
    const { done, value } = await readStreamChunkWithTimeout(
      reader,
      chunkTimeoutMs,
      "response drain read"
    );
    if (done) {
      break;
    }
    if (!value || value.length === 0) {
      continue;
    }
    loaded += value.length;
    if (typeof progressCallback === "function") {
      const effectiveTotal = total || loaded;
      const bucket = effectiveTotal > 0 ? Math.floor(loaded / effectiveTotal * 100) : -1;
      if (bucket > lastBucket) {
        lastBucket = bucket;
        progressCallback({ loaded, total: effectiveTotal });
      }
    }
  }
  if (typeof progressCallback === "function") {
    progressCallback({ loaded, total: total || loaded });
  }
  return loaded;
}
function ensureFsDirectory(fs, dirPath) {
  const slash = dirPath.lastIndexOf("/");
  const parent = slash > 0 ? dirPath.slice(0, slash) : "/";
  const name = dirPath.slice(slash + 1);
  const exists = () => {
    try {
      const entries = fs.readdir(parent);
      return Array.isArray(entries) && entries.includes(name);
    } catch (_) {
      return false;
    }
  };
  if (exists()) {
    return;
  }
  try {
    fs.mkdir(dirPath);
  } catch (error) {
    if (!exists()) {
      throw error;
    }
  }
}
function unlinkFsFile(fs, filePath) {
  try {
    fs.unlink(filePath);
    return true;
  } catch (_) {
    return false;
  }
}
async function writeResponseToFsFileWithProgress(response, fs, filePath, progressCallback, writeOptions = {}) {
  const total = parsePositiveInteger(writeOptions.totalBytes) || inferResponseTotalBytes(response, 0);
  const useBigIntPosition = writeOptions.useBigIntPosition === true;
  const startOffset = parsePositiveInteger(writeOptions.startOffset);
  const preservePartialOnError = writeOptions.preservePartialOnError === true;
  const appendMode = startOffset > 0;
  const chunkTimeoutMs = parsePositiveInteger(writeOptions.chunkTimeoutMs);
  const signal = writeOptions.signal || null;
  const abortMessage = writeOptions.abortMessage || "Model transfer was cancelled.";
  throwIfAborted(signal, abortMessage);
  if (!appendMode) {
    unlinkFsFile(fs, filePath);
  }
  if (!response.body || typeof response.body.getReader !== "function") {
    throwIfAborted(signal, abortMessage);
    const bytes = new Uint8Array(await response.arrayBuffer());
    throwIfAborted(signal, abortMessage);
    if (appendMode) {
      const stream2 = fs.open(filePath, "r+");
      try {
        throwIfAborted(signal, abortMessage);
        const position = useBigIntPosition ? BigInt(startOffset) : startOffset;
        fs.write(stream2, bytes, 0, bytes.length, position);
      } finally {
        fs.close(stream2);
      }
    } else {
      fs.writeFile(filePath, bytes);
    }
    const finalLoaded2 = startOffset + bytes.byteLength;
    if (typeof progressCallback === "function") {
      progressCallback({ loaded: finalLoaded2, total: total || finalLoaded2 });
    }
    return finalLoaded2;
  }
  const reader = response.body.getReader();
  const openMode = appendMode ? "r+" : "w";
  const stream = fs.open(filePath, openMode);
  let loaded = 0;
  let lastBucket = -1;
  let writePosition = startOffset;
  let writePositionBigInt = null;
  try {
    while (true) {
      throwIfAborted(signal, abortMessage);
      const { done, value } = await readStreamChunkWithTimeout(
        reader,
        chunkTimeoutMs,
        "response file read"
      );
      if (done) {
        throwIfAborted(signal, abortMessage);
        break;
      }
      if (!value || value.length === 0) {
        continue;
      }
      const chunk = value.slice ? value.slice() : new Uint8Array(value);
      throwIfAborted(signal, abortMessage);
      if (useBigIntPosition) {
        if (writePositionBigInt == null) {
          writePositionBigInt = BigInt(startOffset);
        }
        fs.write(stream, chunk, 0, chunk.length, writePositionBigInt);
        writePositionBigInt += BigInt(chunk.length);
      } else {
        fs.write(stream, chunk, 0, chunk.length, writePosition);
        writePosition += chunk.length;
      }
      loaded += chunk.length;
      if (typeof progressCallback === "function") {
        const effectiveLoaded = startOffset + loaded;
        const effectiveTotal = total || effectiveLoaded;
        const bucket = effectiveTotal > 0 ? Math.floor(effectiveLoaded / effectiveTotal * 100) : -1;
        if (bucket > lastBucket) {
          lastBucket = bucket;
          progressCallback({ loaded: effectiveLoaded, total: effectiveTotal });
        }
      }
    }
  } catch (error) {
    if (signal?.aborted && error?.name !== "AbortError") {
      error = createAbortError(abortMessage);
    }
    try {
      if (error && typeof error === "object") {
        error.llamadartLoadedBytes = startOffset + loaded;
        error.llamadartFilePath = filePath;
      }
    } catch (_) {
    }
    try {
      fs.close(stream);
    } catch (_) {
    }
    try {
      await reader.cancel?.();
    } catch (_) {
    }
    if (!preservePartialOnError || signal?.aborted || error?.name === "AbortError") {
      try {
        fs.unlink(filePath);
      } catch (_) {
      }
    }
    throw error;
  }
  fs.close(stream);
  const finalLoaded = startOffset + loaded;
  if (typeof progressCallback === "function") {
    progressCallback({ loaded: finalLoaded, total: total || finalLoaded });
  }
  return finalLoaded;
}

// js/src/internal/environment.ts
function isSafariUserAgent(userAgent) {
  if (typeof userAgent !== "string" || userAgent.length === 0) {
    return false;
  }
  const hasSafariToken = /Safari\//.test(userAgent);
  const hasOtherBrowserToken = /(Chrome|Chromium|CriOS|Edg|OPR|Firefox|FxiOS)\//.test(userAgent);
  return hasSafariToken && !hasOtherBrowserToken;
}
function isCrossOriginIsolatedRuntime() {
  try {
    if (typeof globalThis.crossOriginIsolated === "boolean") {
      return globalThis.crossOriginIsolated;
    }
  } catch (_) {
  }
  return true;
}

// js/src/internal/image.ts
async function decodeImageBytesToRgb(bytes, options = {}) {
  const sourceBytes = toUint8Array(bytes);
  if (!sourceBytes || sourceBytes.length === 0) {
    return null;
  }
  if (typeof createImageBitmap !== "function" || typeof Blob !== "function") {
    return null;
  }
  const maxPixelsCandidate = Number(options.maxPixels);
  const maxPixels = Number.isFinite(maxPixelsCandidate) && maxPixelsCandidate > 0 ? Math.max(65536, Math.min(33554432, Math.trunc(maxPixelsCandidate))) : 0;
  const maxEdgeCandidate = Number(options.maxEdge);
  const maxEdge = Number.isFinite(maxEdgeCandidate) && maxEdgeCandidate > 0 ? Math.max(64, Math.min(16384, Math.trunc(maxEdgeCandidate))) : 0;
  if (maxPixels <= 0 && maxEdge <= 0) {
    return null;
  }
  let bitmap = null;
  try {
    const mimeType = typeof options.mimeType === "string" && options.mimeType.length > 0 ? options.mimeType : "image/png";
    const blob = new Blob([sourceBytes], { type: mimeType });
    bitmap = await createImageBitmap(blob);
    const sourceWidth = Math.max(1, Math.trunc(Number(bitmap.width) || 0));
    const sourceHeight = Math.max(1, Math.trunc(Number(bitmap.height) || 0));
    let scale = 1;
    if (maxPixels > 0) {
      const sourcePixels = sourceWidth * sourceHeight;
      if (sourcePixels > maxPixels) {
        scale = Math.min(scale, Math.sqrt(maxPixels / sourcePixels));
      }
    }
    if (maxEdge > 0) {
      const sourceLongest = Math.max(sourceWidth, sourceHeight);
      if (sourceLongest > maxEdge) {
        scale = Math.min(scale, maxEdge / sourceLongest);
      }
    }
    const width = Math.max(1, Math.round(sourceWidth * scale));
    const height = Math.max(1, Math.round(sourceHeight * scale));
    let canvas = null;
    let context = null;
    if (typeof OffscreenCanvas === "function") {
      canvas = new OffscreenCanvas(width, height);
      context = canvas.getContext("2d", {
        alpha: false,
        willReadFrequently: true
      });
    }
    if (!context && typeof document !== "undefined" && typeof document.createElement === "function") {
      canvas = document.createElement("canvas");
      canvas.width = width;
      canvas.height = height;
      context = canvas.getContext("2d", {
        alpha: false,
        willReadFrequently: true
      });
    }
    if (!context) {
      return null;
    }
    context.drawImage(bitmap, 0, 0, width, height);
    let encodedBytes = null;
    if (canvas && "convertToBlob" in canvas && typeof canvas.convertToBlob === "function") {
      const encodedBlob = await canvas.convertToBlob({ type: "image/png" });
      if (encodedBlob) {
        encodedBytes = new Uint8Array(await encodedBlob.arrayBuffer());
      }
    }
    if (!encodedBytes && canvas && "toBlob" in canvas && typeof canvas.toBlob === "function" && typeof Promise === "function") {
      const encodedBlob = await new Promise((resolve) => {
        canvas.toBlob((value) => {
          resolve(value || null);
        }, "image/png");
      });
      if (encodedBlob) {
        encodedBytes = new Uint8Array(await encodedBlob.arrayBuffer());
      }
    }
    if (!encodedBytes || encodedBytes.length === 0) {
      return null;
    }
    return {
      bytes: encodedBytes,
      width,
      height,
      sourceWidth,
      sourceHeight,
      resized: width !== sourceWidth || height !== sourceHeight
    };
  } catch (_) {
    return null;
  } finally {
    try {
      bitmap?.close?.();
    } catch (_) {
    }
  }
}

// js/src/internal/next_token_scores.ts
function nextTokenCandidateIds(value) {
  if (value == null) {
    return [];
  }
  const isList = Array.isArray(value) || ArrayBuffer.isView(value) && !(value instanceof DataView);
  if (!isList) {
    throw new TypeError("Next-token candidates must be an array or typed array of token ids.");
  }
  return Array.from(value, (item, index) => {
    if (!isInt32(item)) {
      throw new TypeError(`Next-token candidates[${index}] is ${String(item)}; expected a 32-bit integer.`);
    }
    return item;
  });
}
function scoredTokensFrom(entries) {
  if (!Array.isArray(entries)) {
    return [];
  }
  return entries.map((entry) => ({
    token: Number(entry?.token),
    bytes: Uint8Array.from(Array.isArray(entry?.bytes) ? entry.bytes : []),
    logprob: typeof entry?.logprob === "number" ? entry.logprob : -Infinity
  }));
}

// js/src/internal/text.ts
function buildPromptFromMessages(messages, addAssistant) {
  const lines = [];
  for (const msg of messages || []) {
    const role = String(msg?.role ?? "user");
    const content = String(msg?.content ?? "");
    lines.push(`${role}: ${content}`);
  }
  if (addAssistant) {
    lines.push("assistant: ");
  }
  return lines.join("\n");
}
function looksLikeCorruptedGeneration(text) {
  if (typeof text !== "string" || text.length === 0) {
    return false;
  }
  const normalized = text.trim();
  if (normalized.length === 0) {
    return false;
  }
  const unusedTokens = text.match(/<unused\d+>/g) || [];
  if (unusedTokens.length >= 4) {
    return true;
  }
  const tokenLikeTags = text.match(/<[^>]{1,40}>/g) || [];
  if (tokenLikeTags.length >= 8) {
    return true;
  }
  const compact = text.replace(/\s+/g, "");
  if (compact.length === 0) {
    return false;
  }
  const tagRun = compact.match(/(?:<[^>]{2,32}>){6,}/);
  if (tagRun) {
    return true;
  }
  const alphaNum = (normalized.match(/[A-Za-z0-9]/g) || []).length;
  const printable = (normalized.match(/[\x20-\x7E]/g) || []).length;
  const angleBrackets = (normalized.match(/[<>]/g) || []).length;
  const alphaNumRatio = alphaNum / normalized.length;
  const printableRatio = printable / normalized.length;
  const bracketRatio = angleBrackets / normalized.length;
  if (normalized.length >= 24 && printableRatio > 0.95 && alphaNumRatio < 0.18) {
    return true;
  }
  if (normalized.length >= 24 && bracketRatio > 0.25) {
    return true;
  }
  return false;
}
function trimUnstableUtf8Tail(text) {
  if (typeof text !== "string" || text.length === 0) {
    return "";
  }
  let end = text.length;
  while (end > 0 && text.charCodeAt(end - 1) === 65533) {
    end -= 1;
  }
  if (end > 0) {
    const tail = text.charCodeAt(end - 1);
    if (tail >= 55296 && tail <= 56319) {
      end -= 1;
    }
  }
  return end === text.length ? text : text.slice(0, end);
}

// js/src/runtime.ts
var textEncoder = new TextEncoder();
var LlamaWebGpuBridgeRuntime = class {
  constructor(config = {}) {
    this._config = config;
    this._core = null;
    this._backendLabels = [];
    this._gpuActive = false;
    this._modelPath = null;
    this._modelPaths = [];
    this._modelBytes = 0;
    this._mmProjPath = null;
    this._mmSupportsVision = false;
    this._mmSupportsAudio = false;
    this._mediaFileCounter = 0;
    this._stateFileCounter = 0;
    this._decisionFileCounter = 0;
    this._loraFileCounter = 0;
    this._loraAdaptersLoaded = false;
    this._stagedMediaPaths = [];
    this._nCtx = 4096;
    this._abortRequested = false;
    this._textToSpeechActive = false;
    this._textToSpeechDone = null;
    this._resolveTextToSpeechDone = null;
    this._runtimeNotes = [];
    this._threadPoolSizeHint = Number(config.threadPoolSize) > 0 ? Math.max(1, Math.trunc(Number(config.threadPoolSize))) : null;
    const requestedThreads = Number(config.threads) > 0 ? Number(config.threads) : this._resolveAutoThreadCount();
    this._threads = this._capThreadsToPool(requestedThreads);
    const requestedThreadsBatch = Number(config.threadsBatch) > 0 ? Number(config.threadsBatch) : this._threads;
    this._threadsBatch = this._capThreadsToPool(
      requestedThreadsBatch,
      { noteTag: "threads_batch_capped_pool" }
    );
    this._nBatch = Number(config.nBatch) > 0 ? Math.max(32, Math.trunc(Number(config.nBatch))) : 0;
    this._nUbatch = Number(config.nUbatch) > 0 ? Math.max(32, Math.trunc(Number(config.nUbatch))) : 0;
    this._nGpuLayers = Number.isFinite(config.nGpuLayers) ? Number(config.nGpuLayers) : -1;
    this._nSeqMax = 0;
    this._useMmap = false;
    this._useMlock = false;
    this._flashAttention = -1;
    this._cacheTypeK = 1;
    this._cacheTypeV = 1;
    this._kvUnified = -1;
    this._ropeFrequencyBase = 0;
    this._ropeFrequencyScale = 0;
    this._splitMode = -1;
    this._mainGpu = -1;
    this._loadMtp = false;
    this._speculativeRollbackTokenMax = 0;
    this._isSafari = isSafariUserAgent(this._config.userAgent ?? globalThis.navigator?.userAgent ?? "");
    this._coreVariant = "uninitialized";
    this._preferMemory64 = this._config.preferMemory64 !== false;
    this._modelSource = "network";
    this._modelCacheState = "disabled";
    this._modelCacheName = defaultModelCacheName;
    this._loadedModelUrl = null;
    this._mmProjSourceUrl = null;
    this._draftModel = null;
    this._speculativeFileCounter = 0;
    this._suppressedWarmupWarningCount = 0;
    this._didReportWarmupWarningSuppression = false;
    this._remoteFetchThresholdBytes = Number(config.remoteFetchThresholdBytes) > 0 ? Number(config.remoteFetchThresholdBytes) : 1900 * 1024 * 1024;
    this._remoteFetchChunkBytes = Number(config.remoteFetchChunkBytes) > 0 ? Number(config.remoteFetchChunkBytes) : 16 * 1024 * 1024;
    this._mediaMaxImagePixels = Number(config.mediaMaxImagePixels) > 0 ? Math.max(65536, Math.min(33554432, Math.trunc(Number(config.mediaMaxImagePixels)))) : 1024 * 1024;
    this._mediaMaxImageEdge = Number(config.mediaMaxImageEdge) > 0 ? Math.max(64, Math.min(16384, Math.trunc(Number(config.mediaMaxImageEdge)))) : 1280;
    this._disableImageDownscale = config.disableImageDownscale === true;
    this._activeTransferAbortController = null;
    this._lastCoreErrorText = "";
    this._lastCoreErrorHint = "";
    this._logLevel = Number.isFinite(config.logLevel) ? Math.max(0, Math.min(4, Math.trunc(config.logLevel))) : 2;
  }
  static supportsSafariAdaptiveGpu = true;
  _pushRuntimeNote(note) {
    if (typeof note !== "string" || note.length === 0) {
      return;
    }
    if (!Array.isArray(this._runtimeNotes)) {
      this._runtimeNotes = [];
    }
    if (!this._runtimeNotes.includes(note)) {
      this._runtimeNotes.push(note);
    }
  }
  _detectThreadPoolSizeFromCore() {
    if (!this._coreSupportsPthreads()) {
      return 1;
    }
    const core = this._core;
    if (!core || typeof core !== "object") {
      return null;
    }
    try {
      if (typeof core.ccall === "function") {
        const compiledPoolSize = Number(
          core.ccall("llamadart_webgpu_pthread_pool_size", "number", [], [])
        );
        if (Number.isFinite(compiledPoolSize) && compiledPoolSize > 0) {
          return Math.max(1, Math.trunc(compiledPoolSize));
        }
      }
    } catch (_) {
    }
    try {
      const pThread = core.PThread;
      if (!pThread || typeof pThread !== "object") {
        return null;
      }
      const unused = Array.isArray(pThread.unusedWorkers) ? pThread.unusedWorkers.length : 0;
      const running = Array.isArray(pThread.runningWorkers) ? pThread.runningWorkers.length : 0;
      const total = unused + running;
      return total > 0 ? total : null;
    } catch (_) {
      return null;
    }
  }
  _coreSupportsPthreads() {
    const core = this._core;
    if (!core || typeof core !== "object") {
      return false;
    }
    try {
      if (typeof core.ccall === "function") {
        const compiledWithPthreads = Number(
          core.ccall("llamadart_webgpu_supports_pthreads", "number", [], [])
        );
        if (Number.isFinite(compiledWithPthreads)) {
          return compiledWithPthreads !== 0;
        }
      }
    } catch (_) {
    }
    try {
      const wasmBuffer = core.wasmMemory?.buffer;
      if (typeof SharedArrayBuffer === "function" && wasmBuffer instanceof SharedArrayBuffer) {
        return true;
      }
    } catch (_) {
    }
    try {
      const heapBuffer = core.HEAP8?.buffer || core.HEAPU8?.buffer;
      if (typeof SharedArrayBuffer === "function" && heapBuffer instanceof SharedArrayBuffer) {
        return true;
      }
    } catch (_) {
    }
    try {
      const pThread = core.PThread;
      if (!pThread || typeof pThread !== "object") {
        return false;
      }
      return Array.isArray(pThread.unusedWorkers) || Array.isArray(pThread.runningWorkers) || typeof pThread.allocateUnusedWorker === "function";
    } catch (_) {
      return false;
    }
  }
  _syncThreadPoolSizeHintFromCore() {
    const detected = this._detectThreadPoolSizeFromCore();
    if (!Number.isFinite(detected) || detected <= 0) {
      return;
    }
    this._threadPoolSizeHint = Math.max(1, Math.trunc(detected));
  }
  _resolveAutoThreadCount() {
    const hardwareThreads = Number(globalThis.navigator?.hardwareConcurrency);
    if (Number.isFinite(hardwareThreads) && hardwareThreads > 0) {
      return Math.max(1, Math.min(8, Math.trunc(hardwareThreads)));
    }
    return 4;
  }
  _capThreadsToPool(candidate, { noteTag = "threads_capped_pool" } = {}) {
    let resolved = Number(candidate);
    if (!Number.isFinite(resolved) || resolved <= 0) {
      resolved = 1;
    }
    resolved = Math.max(1, Math.trunc(resolved));
    const poolSize = Number(this._threadPoolSizeHint);
    if (Number.isFinite(poolSize) && poolSize > 0 && resolved > poolSize) {
      if (noteTag) {
        this._pushRuntimeNote(`${noteTag}:${poolSize}`);
      }
      return poolSize;
    }
    return resolved;
  }
  _isVerboseWarmupWarning(text) {
    const lowered = String(text || "").toLowerCase();
    if (lowered.length === 0) {
      return false;
    }
    if (lowered.includes("warmup:")) {
      return true;
    }
    if (lowered.includes("please report this on github as an issue")) {
      return true;
    }
    if (lowered.includes("warning: ref:")) {
      return true;
    }
    if (lowered.includes("github.com/ggml-org/llama.cpp/pull/")) {
      return true;
    }
    if (lowered.includes("****************")) {
      return true;
    }
    return false;
  }
  _emitSuppressedWarmupWarningSummaryIfNeeded() {
    if (this._suppressedWarmupWarningCount <= 0) {
      return;
    }
    if (this._logLevel <= 2) {
      this._emitLogger(
        "log",
        `info: suppressed ${this._suppressedWarmupWarningCount} verbose warmup log lines (set bridge/runtime log level to Debug to inspect full warmup trace).`
      );
    }
    this._suppressedWarmupWarningCount = 0;
  }
  _shouldAttemptGenerationRecovery(errorText, options = {}, generated = 0) {
    if (generated > 0) {
      return false;
    }
    if (this._nGpuLayers <= 0) {
      return false;
    }
    if (this._coreVariant !== "wasm64") {
      return false;
    }
    if (!hasModelSource(this._loadedModelUrl)) {
      return false;
    }
    if (options._llamadartGenerationRecoveryAttempted === true) {
      return false;
    }
    if (Array.isArray(options.parts) && options.parts.length > 0) {
      return false;
    }
    const lowered = String(errorText || "").toLowerCase();
    if (lowered.includes("failed to decode")) {
      return true;
    }
    if (lowered.includes("failed to compute graph")) {
      return true;
    }
    if (lowered.includes("ggml_backend_sched_graph_compute_async failed")) {
      return true;
    }
    return false;
  }
  _isContextLimitGenerationError(errorText = "") {
    const lowered = [
      String(errorText || ""),
      String(this._lastCoreErrorText || ""),
      String(this._lastCoreErrorHint || "")
    ].join(" ").toLowerCase();
    return lowered.includes("failed to find a memory slot for batch") || lowered.includes("failed to prepare attention batches") || lowered.includes("context overflow") || lowered.includes("context full") || lowered.includes("kv cache full") || lowered.includes("insufficient kv") || lowered.includes("no kv slot");
  }
  async _recoverGenerationWithCpuFallback(options = {}) {
    const modelUrl = cloneModelSource(this._loadedModelUrl);
    if (!hasModelSource(modelUrl)) {
      return false;
    }
    if (this._loraAdaptersLoaded) {
      this._runtimeNotes.push("generation_recovery_cpu_skipped_lora");
      return false;
    }
    this._runtimeNotes.push("generation_recovery_cpu_attempt");
    this._emitLogger(
      "warn",
      "warning: generation failed on wasm64/WebGPU; retrying by reloading model with CPU fallback for stability."
    );
    const previousPreferMemory64 = this._preferMemory64;
    const previousMMProjSourceUrl = this._mmProjSourceUrl;
    const previousDraftModel = this._draftModel;
    const resolvedLoadState = this._resolvedLoadState();
    try {
      this._preferMemory64 = false;
      await this.loadModelFromUrl(modelUrl, {
        nCtx: this._nCtx,
        nThreads: this._threads,
        nGpuLayers: 0,
        useCache: true,
        forceRemoteFetchBackend: false,
        remoteFetchChunkBytes: this._resolveRemoteFetchChunkBytes(options),
        safariGpuProbe: false,
        signal: options.signal
      }, resolvedLoadState);
      if (typeof previousMMProjSourceUrl === "string" && previousMMProjSourceUrl.length > 0) {
        try {
          await this.loadMultimodalProjector(previousMMProjSourceUrl);
        } catch (_) {
          this._runtimeNotes.push("generation_recovery_mmproj_reload_failed");
        }
      }
      if (previousDraftModel) {
        try {
          await this.loadDraftModel(previousDraftModel.url, previousDraftModel.options);
        } catch (_) {
          this._runtimeNotes.push("generation_recovery_draft_reload_failed");
        }
      }
      this._runtimeNotes.push("generation_recovery_cpu_applied");
      return true;
    } catch (_) {
      this._runtimeNotes.push("generation_recovery_cpu_failed");
      return false;
    } finally {
      this._preferMemory64 = previousPreferMemory64;
    }
  }
  _loggerFor(level) {
    const logger = this._config?.logger;
    const fallback = typeof console !== "undefined" ? console : null;
    if (logger && typeof logger[level] === "function") {
      return logger[level].bind(logger);
    }
    if (!fallback) {
      return () => {
      };
    }
    if (typeof fallback[level] === "function") {
      return fallback[level].bind(fallback);
    }
    if (typeof fallback.log === "function") {
      return fallback.log.bind(fallback);
    }
    return () => {
    };
  }
  _shouldEmitLoggerLevel(level) {
    const current = Number(this._logLevel);
    if (!Number.isFinite(current) || current < 0) {
      return true;
    }
    const threshold = logThresholdForConfiguredLevel(
      Math.max(0, Math.min(4, Math.trunc(current)))
    );
    if (threshold > 3) {
      return false;
    }
    return logLevelForName(level) >= threshold;
  }
  _emitLogger(level, message) {
    if (!this._shouldEmitLoggerLevel(level)) {
      return;
    }
    try {
      this._loggerFor(level)(message);
    } catch (_) {
    }
  }
  _classifyCoreErrorLine(text) {
    const trimmed = String(text ?? "").trim();
    if (trimmed.length === 0) {
      return "ignore";
    }
    if (this._isVerboseWarmupWarning(trimmed)) {
      return "warmup";
    }
    const lowered = trimmed.toLowerCase();
    if (lowered.startsWith("warning") || lowered.startsWith("warn:") || lowered.includes(" warning:")) {
      return "warn";
    }
    if (lowered.startsWith("error") || lowered.startsWith("err:") || lowered.includes(" error:") || lowered.includes("failed") || lowered.includes("exception") || lowered.includes("abort") || lowered.includes("fatal") || lowered.includes("out of memory") || lowered.includes("invalid")) {
      return "error";
    }
    return "info";
  }
  _applyCoreLogLevel() {
    if (!this._core) {
      return;
    }
    try {
      this._core.ccall(
        "llamadart_webgpu_set_log_level",
        null,
        ["number"],
        [this._logLevel]
      );
    } catch (_) {
    }
  }
  _coreErrorMessage(prefix, fallbackCode = 0) {
    try {
      const err = this._core?.ccall("llamadart_webgpu_last_error", "string", [], []);
      if (err) {
        return `${prefix}: ${err}`;
      }
    } catch (_) {
    }
    return `${prefix} (code=${fallbackCode})`;
  }
  _resolveCacheName(options = {}) {
    if (typeof options.cacheName === "string" && options.cacheName.trim().length > 0) {
      return options.cacheName.trim();
    }
    if (typeof this._config.cacheName === "string" && this._config.cacheName.trim().length > 0) {
      return this._config.cacheName.trim();
    }
    return defaultModelCacheName;
  }
  _isWorkerRuntime() {
    return typeof WorkerGlobalScope !== "undefined" && globalThis instanceof WorkerGlobalScope;
  }
  _resolveRemoteFetchThresholdBytes(options = {}) {
    const candidate = Number(options.remoteFetchThresholdBytes);
    if (Number.isFinite(candidate) && candidate > 0) {
      return Math.trunc(candidate);
    }
    return Math.trunc(this._remoteFetchThresholdBytes);
  }
  _resolveRemoteFetchChunkBytes(options = {}) {
    const candidate = Number(options.remoteFetchChunkBytes);
    if (Number.isFinite(candidate) && candidate > 0) {
      return Math.max(16 * 1024, Math.trunc(candidate));
    }
    return Math.max(16 * 1024, Math.trunc(this._remoteFetchChunkBytes));
  }
  _canUseRemoteFetchBackend(options = {}) {
    if (options.forceRemoteFetchBackend === false) {
      return false;
    }
    if (!this._isWorkerRuntime()) {
      return false;
    }
    if (options.forceRemoteFetchBackend === true) {
      return true;
    }
    return this._config.allowAutoRemoteFetchBackend === true;
  }
  async _tryHeadContentLength(url) {
    function parseSizeFromHeaders(headers) {
      if (!headers) {
        return 0;
      }
      const length = Number(headers.get("content-length")) || 0;
      if (length > 0) {
        return length;
      }
      const linkedSize = Number(headers.get("x-linked-size")) || 0;
      if (linkedSize > 0) {
        return linkedSize;
      }
      const contentRange = String(headers.get("content-range") || "");
      const slash = contentRange.lastIndexOf("/");
      if (slash >= 0 && slash + 1 < contentRange.length) {
        const total = Number(contentRange.slice(slash + 1)) || 0;
        if (total > 0) {
          return total;
        }
      }
      return 0;
    }
    try {
      const response = await this._fetchWithTimeout(url, {
        method: "HEAD",
        cache: "no-store"
      }, 45e3);
      if (!response.ok) {
        throw new Error("HEAD request failed");
      }
      const size = parseSizeFromHeaders(response.headers);
      if (size > 0) {
        return size;
      }
    } catch (_) {
    }
    try {
      const probe = await this._fetchWithTimeout(url, {
        headers: {
          Range: "bytes=0-0"
        },
        cache: "no-store"
      }, 45e3);
      if (!(probe.ok || probe.status === 206)) {
        return null;
      }
      const size = parseSizeFromHeaders(probe.headers);
      if (size > 0) {
        this._runtimeNotes.push("model_fetch_size_probe");
        return size;
      }
    } catch (_) {
    }
    return null;
  }
  async _resolveRemoteFetchUrl(url) {
    function parseSizeFromHeaders(headers) {
      if (!headers) {
        return 0;
      }
      const contentRange = String(headers.get("content-range") || "");
      const slash = contentRange.lastIndexOf("/");
      if (slash >= 0 && slash + 1 < contentRange.length) {
        const total = Number(contentRange.slice(slash + 1)) || 0;
        if (total > 0) {
          return total;
        }
      }
      const linkedSize = Number(headers.get("x-linked-size")) || 0;
      if (linkedSize > 0) {
        return linkedSize;
      }
      const length = Number(headers.get("content-length")) || 0;
      if (length > 0) {
        return length;
      }
      return 0;
    }
    try {
      const probe = await this._fetchWithTimeout(url, {
        headers: {
          Range: "bytes=0-0"
        },
        cache: "no-store"
      }, 45e3);
      if (!(probe.ok || probe.status === 206)) {
        return null;
      }
      const resolvedUrl = typeof probe.url === "string" && probe.url.length > 0 ? probe.url : null;
      const sizeBytes = parseSizeFromHeaders(probe.headers);
      return {
        resolvedUrl,
        sizeBytes: sizeBytes > 0 ? sizeBytes : null
      };
    } catch (_) {
      return null;
    }
  }
  _resolveNativeLoadOptions(options = {}) {
    this._nSeqMax = parsePositiveInteger(options.nSeqMax);
    this._useMmap = parseBooleanFlag(options.useMmap, false);
    this._useMlock = parseBooleanFlag(options.useMlock, false);
    this._flashAttention = parseEnumValue(options.flashAttention, [-1, 0, 1], -1);
    this._cacheTypeK = parseEnumValue(options.cacheTypeK, [1, 2, 8], 1);
    this._cacheTypeV = parseEnumValue(options.cacheTypeV, [1, 2, 8], 1);
    this._kvUnified = parseOptionalBooleanFlag(options.kvUnified);
    this._ropeFrequencyBase = parsePositiveNumber(options.ropeFrequencyBase);
    this._ropeFrequencyScale = parsePositiveNumber(options.ropeFrequencyScale);
    this._splitMode = parseEnumValue(options.splitMode, [0, 1, 2, 3], -1);
    this._mainGpu = parseInteger(options.mainGpu, -1);
    if (this._mainGpu < 0) {
      this._mainGpu = -1;
    }
    this._loadMtp = parseBooleanFlag(options.loadMtp, false);
    this._speculativeRollbackTokenMax = parsePositiveInteger(options.speculativeRollbackTokenMax);
    const wantsQuantizedKvCache = this._cacheTypeK !== 1 || this._cacheTypeV !== 1;
    if (this._flashAttention === 0 && wantsQuantizedKvCache) {
      throw new Error(
        "Non-F16 KV cache requires flashAttention to be auto or enabled."
      );
    }
    if (this._flashAttention === -1 && wantsQuantizedKvCache) {
      this._flashAttention = 1;
      this._runtimeNotes.push("flash_attention:auto_enabled_for_kv_cache");
    }
    if (this._kvUnified < 0 && this._nSeqMax > 1) {
      this._kvUnified = 1;
      this._runtimeNotes.push("kv_unified:auto_enabled_for_sequences");
    }
  }
  _nativeLoadOptionValues() {
    return [
      this._nSeqMax,
      this._useMmap ? 1 : 0,
      this._useMlock ? 1 : 0,
      this._flashAttention,
      this._cacheTypeK,
      this._cacheTypeV,
      this._kvUnified,
      this._ropeFrequencyBase,
      this._ropeFrequencyScale,
      this._splitMode,
      this._mainGpu,
      this._loadMtp ? 1 : 0,
      this._speculativeRollbackTokenMax
    ];
  }
  _nativeLoadOptionTypes() {
    return this._nativeLoadOptionValues().map(() => "number");
  }
  _resolvedLoadState() {
    return {
      _nCtx: this._nCtx,
      _threads: this._threads,
      _threadsBatch: this._threadsBatch,
      _nBatch: this._nBatch,
      _nUbatch: this._nUbatch,
      _nSeqMax: this._nSeqMax,
      _useMmap: this._useMmap,
      _useMlock: this._useMlock,
      _flashAttention: this._flashAttention,
      _cacheTypeK: this._cacheTypeK,
      _cacheTypeV: this._cacheTypeV,
      _kvUnified: this._kvUnified,
      _ropeFrequencyBase: this._ropeFrequencyBase,
      _ropeFrequencyScale: this._ropeFrequencyScale,
      _splitMode: this._splitMode,
      _mainGpu: this._mainGpu,
      _loadMtp: this._loadMtp,
      _speculativeRollbackTokenMax: this._speculativeRollbackTokenMax
    };
  }
  async _tryLoadModelFromRemoteFetchBackend(core, url, options = {}) {
    if (!this._canUseRemoteFetchBackend(options)) {
      return { loaded: false, sizeBytes: null };
    }
    const thresholdBytes = this._resolveRemoteFetchThresholdBytes(options);
    const chunkBytes = this._resolveRemoteFetchChunkBytes(options);
    const forceRemote = options.forceRemoteFetchBackend === true;
    let sizeBytes = Number(options.modelBytesHint);
    if (!Number.isFinite(sizeBytes) || sizeBytes <= 0) {
      sizeBytes = await this._tryHeadContentLength(url);
    }
    if (!forceRemote) {
      if (Number.isFinite(sizeBytes) && sizeBytes > 0 && sizeBytes < thresholdBytes) {
        this._runtimeNotes.push("model_fetch_backend_skipped_small");
        return { loaded: false, sizeBytes };
      }
      if (!Number.isFinite(sizeBytes) || sizeBytes <= 0) {
        this._runtimeNotes.push("model_fetch_backend_size_unknown");
        this._runtimeNotes.push("model_fetch_backend_unknown_size_attempt");
      }
    }
    this._runtimeNotes.push("model_fetch_backend_attempt");
    this._runtimeNotes.push(`model_fetch_chunk:${chunkBytes}`);
    this._lastCoreErrorHint = "";
    let remoteFetchUrl = url;
    const resolvedProbe = await this._resolveRemoteFetchUrl(url);
    if (resolvedProbe?.resolvedUrl) {
      remoteFetchUrl = resolvedProbe.resolvedUrl;
      if (remoteFetchUrl !== url) {
        this._runtimeNotes.push("model_fetch_backend_resolved_url");
      }
    }
    if ((!Number.isFinite(sizeBytes) || sizeBytes <= 0) && Number.isFinite(resolvedProbe?.sizeBytes) && resolvedProbe.sizeBytes > 0) {
      sizeBytes = resolvedProbe.sizeBytes;
    }
    try {
      globalThis.__llamadartFetchBackendLastError = null;
    } catch (_) {
    }
    if (typeof options.progressCallback === "function") {
      options.progressCallback({ loaded: 0, total: Number.isFinite(sizeBytes) ? sizeBytes : 0 });
    }
    try {
      const rc = Number(
        await core.ccall(
          "llamadart_webgpu_load_model_from_url",
          "number",
          [
            "string",
            "number",
            "number",
            "number",
            "number",
            "number",
            "number",
            "number",
            ...this._nativeLoadOptionTypes()
          ],
          [
            remoteFetchUrl,
            this._nCtx,
            this._threads,
            this._threadsBatch,
            this._nBatch,
            this._nUbatch,
            this._nGpuLayers,
            chunkBytes,
            ...this._nativeLoadOptionValues()
          ],
          { async: true }
        )
      );
      if (rc !== 0) {
        throw new Error(this._coreErrorMessage("Fetch-backed model load failed", rc));
      }
      this._modelSource = "network-fetch";
      this._modelPath = null;
      this._modelBytes = Number.isFinite(sizeBytes) && sizeBytes > 0 ? Math.trunc(sizeBytes) : 1;
      this._runtimeNotes.push("model_source_fetch_backend");
      if (typeof options.progressCallback === "function") {
        const resolved = Number.isFinite(sizeBytes) && sizeBytes > 0 ? sizeBytes : 1;
        options.progressCallback({ loaded: resolved, total: resolved });
      }
      return {
        loaded: true,
        sizeBytes: Number.isFinite(sizeBytes) ? sizeBytes : null
      };
    } catch (error) {
      const text = String(error || "").toLowerCase();
      if (text.includes("aborted(native code called abort())")) {
        try {
          const stderrText = String(
            this._lastCoreErrorHint || this._lastCoreErrorText || ""
          ).trim();
          if (stderrText.length > 0) {
            const token = stderrText.slice(0, 120).replace(/[\s;=]+/g, "_").replace(/[^a-zA-Z0-9._:-]/g, "");
            if (token.length > 0) {
              this._runtimeNotes.push(`model_core_stderr:${token}`);
            }
          }
        } catch (_) {
        }
        try {
          const fetchError = String(
            globalThis.__llamadartFetchBackendLastError || ""
          ).trim();
          if (fetchError.length > 0) {
            const token = fetchError.slice(0, 120).replace(/[\s;=]+/g, "_").replace(/[^a-zA-Z0-9._:-]/g, "");
            if (token.length > 0) {
              this._runtimeNotes.push(`model_fetch_js_error:${token}`);
            }
          }
        } catch (_) {
        }
        try {
          const stats = globalThis.__llamadartFetchBackendStats;
          if (stats && typeof stats === "object") {
            const reads = Number(stats.reads) || 0;
            const getSize = Number(stats.getSize) || 0;
            const ranges = Number(stats.ranges) || 0;
            const fallbacks = Number(stats.wholeFileFallbacks) || 0;
            const errors = Number(stats.errors) || 0;
            this._runtimeNotes.push(
              `model_fetch_stats:r${reads}_s${getSize}_q${ranges}_f${fallbacks}_e${errors}`
            );
          }
        } catch (_) {
        }
        try {
          const coreError = String(
            this._core?.ccall("llamadart_webgpu_last_error", "string", [], []) || ""
          ).trim();
          if (coreError.length > 0) {
            const token = coreError.slice(0, 120).replace(/[\s;=]+/g, "_").replace(/[^a-zA-Z0-9._:-]/g, "");
            if (token.length > 0) {
              this._runtimeNotes.push(`model_fetch_core_error:${token}`);
            }
          }
        } catch (_) {
        }
        this._runtimeNotes.push("model_fetch_backend_abort");
        throw error;
      }
      if (text.includes("fetch-backed model load failed")) {
        this._runtimeNotes.push("model_fetch_backend_failed");
        return {
          loaded: false,
          sizeBytes: Number.isFinite(sizeBytes) ? sizeBytes : null
        };
      }
      if (text.includes("load_model_from_url") && (text.includes("not found") || text.includes("undefined symbol") || text.includes("missing function"))) {
        this._runtimeNotes.push("model_fetch_backend_unavailable");
        return {
          loaded: false,
          sizeBytes: Number.isFinite(sizeBytes) ? sizeBytes : null
        };
      }
      if (text.includes("worker-thread bridge runtime")) {
        this._runtimeNotes.push("model_fetch_backend_requires_worker");
        return {
          loaded: false,
          sizeBytes: Number.isFinite(sizeBytes) ? sizeBytes : null
        };
      }
      throw error;
    }
  }
  _beginTransferAbortController() {
    if (typeof AbortController !== "function") {
      this._activeTransferAbortController = null;
      return null;
    }
    if (this._activeTransferAbortController) {
      try {
        this._activeTransferAbortController.abort();
      } catch (_) {
      }
    }
    const controller = new AbortController();
    this._activeTransferAbortController = controller;
    return controller;
  }
  _clearTransferAbortController(controller) {
    if (this._activeTransferAbortController === controller) {
      this._activeTransferAbortController = null;
    }
  }
  _resolveFetchTimeoutMs(options = {}, defaultTimeoutMs = 18e4) {
    const configured = Number(options.fetchTimeoutMs);
    if (Number.isFinite(configured) && configured > 0) {
      return Math.max(1e4, Math.min(18e5, Math.trunc(configured)));
    }
    return defaultTimeoutMs;
  }
  _resolveStreamChunkTimeoutMs(options = {}, defaultTimeoutMs = 9e4) {
    const configured = Number(options.streamChunkTimeoutMs);
    if (Number.isFinite(configured) && configured > 0) {
      return Math.max(5e3, Math.min(3e5, Math.trunc(configured)));
    }
    return defaultTimeoutMs;
  }
  _resolveCoreInitTimeoutMs() {
    const configured = Number(this._config.coreInitTimeoutMs);
    if (!Number.isFinite(configured) || configured <= 0) {
      return 9e4;
    }
    return Math.max(1e4, Math.min(6e5, Math.trunc(configured)));
  }
  _resolveMediaImageMaxPixels(options = {}) {
    if (options.disableImageDownscale === true || this._disableImageDownscale) {
      return 0;
    }
    const configured = Number(options.mediaMaxImagePixels);
    if (Number.isFinite(configured) && configured > 0) {
      return Math.max(65536, Math.min(33554432, Math.trunc(configured)));
    }
    return this._mediaMaxImagePixels;
  }
  _resolveMediaImageMaxEdge(options = {}) {
    if (options.disableImageDownscale === true || this._disableImageDownscale) {
      return 0;
    }
    const configured = Number(options.mediaMaxImageEdge);
    if (Number.isFinite(configured) && configured > 0) {
      return Math.max(64, Math.min(16384, Math.trunc(configured)));
    }
    return this._mediaMaxImageEdge;
  }
  async _fetchWithTimeout(url, init = {}, timeoutMs = 0) {
    const resolvedTimeout = Number(timeoutMs);
    if (!Number.isFinite(resolvedTimeout) || resolvedTimeout <= 0) {
      return fetch(url, init);
    }
    if (typeof AbortController !== "function") {
      return Promise.race([
        fetch(url, init),
        new Promise((_, reject) => {
          globalThis.setTimeout(
            () => reject(new Error(`fetch timeout (${resolvedTimeout}ms)`)),
            resolvedTimeout
          );
        })
      ]);
    }
    const timeoutController = new AbortController();
    const externalSignal = init.signal;
    let didTimeout = false;
    let timeoutHandle = null;
    const onExternalAbort = () => {
      try {
        timeoutController.abort();
      } catch (_) {
      }
    };
    if (externalSignal && typeof externalSignal.addEventListener === "function") {
      externalSignal.addEventListener("abort", onExternalAbort, { once: true });
      if (externalSignal.aborted) {
        onExternalAbort();
      }
    }
    timeoutHandle = globalThis.setTimeout(() => {
      didTimeout = true;
      try {
        timeoutController.abort();
      } catch (_) {
      }
    }, resolvedTimeout);
    try {
      return await fetch(url, {
        ...init,
        signal: timeoutController.signal
      });
    } catch (error) {
      if (didTimeout) {
        throw new Error(`fetch timeout (${resolvedTimeout}ms)`);
      }
      throw error;
    } finally {
      if (timeoutHandle != null) {
        globalThis.clearTimeout(timeoutHandle);
      }
      if (externalSignal && typeof externalSignal.removeEventListener === "function") {
        externalSignal.removeEventListener("abort", onExternalAbort);
      }
    }
  }
  _unlinkDecisionFile(path) {
    if (!this._core || typeof path !== "string" || path.length === 0) {
      return;
    }
    try {
      this._core.FS.unlink(path);
    } catch (_) {
    }
  }
  _deleteFsFile(path) {
    if (!this._core || typeof path !== "string" || path.length === 0) {
      return false;
    }
    return unlinkFsFile(this._core.FS, path);
  }
  _releaseModelFiles() {
    const paths = Array.isArray(this._modelPaths) && this._modelPaths.length > 0 ? [...this._modelPaths] : this._modelPath ? [this._modelPath] : [];
    let removed = 0;
    for (const path of paths) {
      if (this._deleteFsFile(path)) {
        removed += 1;
      }
    }
    this._modelPaths = [];
    return removed;
  }
  // A load replaces the current model. Free it, its projector and their FS files
  // before downloading the next model so both never occupy the wasm heap at once.
  // The core refuses while a generation or speech synthesis is still active,
  // which leaves the current model untouched.
  _releaseLoadedModel(core) {
    const projectorPath = this._mmProjPath;
    const hasModelFiles = Array.isArray(this._modelPaths) && this._modelPaths.length > 0;
    if (this._modelBytes <= 0 && !projectorPath && !hasModelFiles) {
      return false;
    }
    const rc = Number(core.ccall("llamadart_webgpu_free_model", "number", [], []));
    if (rc !== 0) {
      throw new Error(this._coreErrorMessage("Failed to release the loaded model", rc));
    }
    this._loraAdaptersLoaded = false;
    this._clearStagedMediaFiles();
    this._deleteFsFile(projectorPath);
    this._releaseModelFiles();
    this._modelPath = null;
    this._modelBytes = 0;
    this._modelSource = "network";
    this._mmProjPath = null;
    this._mmSupportsVision = false;
    this._mmSupportsAudio = false;
    this._draftModel = null;
    this._runtimeNotes.push("previous_model_released");
    return true;
  }
  async _getCachedModelResponse(url, options = {}) {
    let useCache = options.useCache !== false;
    const forceRefresh = options.force === true;
    const requireReadableStream = options.requireReadableStream === true;
    const requestHeaders = options.requestHeaders && typeof options.requestHeaders === "object" ? options.requestHeaders : null;
    if (requestHeaders && Object.keys(requestHeaders).length > 0) {
      useCache = false;
    }
    const fetchOptions = {
      ...options.signal ? { signal: options.signal } : {},
      ...requestHeaders ? { headers: requestHeaders } : {}
    };
    const fetchTimeoutMs = this._resolveFetchTimeoutMs(options, 18e4);
    this._modelSource = "network";
    this._modelCacheState = useCache ? "unavailable" : "disabled";
    this._modelCacheName = this._resolveCacheName(options);
    if (!useCache) {
      const response = await this._fetchWithTimeout(url, {
        cache: "no-store",
        ...fetchOptions
      }, fetchTimeoutMs);
      this._modelCacheState = "disabled";
      if (requireReadableStream) {
        this._runtimeNotes.push(
          hasReadableResponseStream(response) ? "model_network_stream" : "model_network_no_stream"
        );
      }
      return response;
    }
    if (!globalThis.caches || typeof globalThis.caches.open !== "function") {
      this._modelCacheState = "unavailable";
      const response = await this._fetchWithTimeout(url, fetchOptions, fetchTimeoutMs);
      if (requireReadableStream) {
        this._runtimeNotes.push(
          hasReadableResponseStream(response) ? "model_network_stream" : "model_network_no_stream"
        );
      }
      return response;
    }
    const cacheKey = normalizeAbsoluteUrl(url);
    try {
      const cache = await globalThis.caches.open(this._modelCacheName);
      const cached = forceRefresh ? null : await cache.match(cacheKey);
      if (cached) {
        const cacheHasStream = hasReadableResponseStream(cached);
        if (!requireReadableStream || cacheHasStream) {
          this._modelSource = "cache";
          this._modelCacheState = "hit";
          this._runtimeNotes.push("model_cache_hit");
          if (requireReadableStream) {
            this._runtimeNotes.push("model_cache_stream");
          }
          return cached;
        }
        this._runtimeNotes.push("model_cache_hit_no_stream");
        this._modelSource = "network";
        this._modelCacheState = "refresh";
        const refreshed = await this._fetchWithTimeout(url, {
          cache: "no-store",
          ...fetchOptions
        }, fetchTimeoutMs);
        if (refreshed.ok) {
          try {
            await cache.put(cacheKey, refreshed.clone());
            this._modelCacheState = "stored";
            this._runtimeNotes.push("model_cache_stored");
          } catch (_) {
            this._modelCacheState = "store_failed";
            this._runtimeNotes.push("model_cache_store_failed");
          }
        }
        this._runtimeNotes.push(
          hasReadableResponseStream(refreshed) ? "model_network_stream" : "model_network_no_stream"
        );
        return refreshed;
      }
      this._modelCacheState = forceRefresh ? "refresh" : "miss";
      const response = await this._fetchWithTimeout(url, fetchOptions, fetchTimeoutMs);
      if (response.ok) {
        try {
          await cache.put(cacheKey, response.clone());
          this._modelCacheState = "stored";
          this._runtimeNotes.push("model_cache_stored");
        } catch (_) {
          this._modelCacheState = "store_failed";
          this._runtimeNotes.push("model_cache_store_failed");
        }
      }
      if (requireReadableStream) {
        this._runtimeNotes.push(
          hasReadableResponseStream(response) ? "model_network_stream" : "model_network_no_stream"
        );
      }
      return response;
    } catch (_) {
      this._modelCacheState = "error";
      this._runtimeNotes.push("model_cache_error");
      const response = await this._fetchWithTimeout(url, fetchOptions, fetchTimeoutMs);
      if (requireReadableStream) {
        this._runtimeNotes.push(
          hasReadableResponseStream(response) ? "model_network_stream" : "model_network_no_stream"
        );
      }
      return response;
    }
  }
  async prefetchModelToCache(url, options = {}) {
    const useCache = options.useCache !== false;
    this._modelSource = "network";
    this._modelCacheState = useCache ? "unavailable" : "disabled";
    this._modelCacheName = this._resolveCacheName(options);
    const progressCallback = typeof options.progressCallback === "function" ? options.progressCallback : null;
    const modelUrls = expandModelShardUrls(url);
    if (modelUrls.length === 0) {
      throw new Error("Model URL is empty.");
    }
    if (modelUrls.length > 1) {
      this._runtimeNotes.push(`model_split_cache_prefetch:${modelUrls.length}`);
    }
    const shardLoaded = new Array(modelUrls.length).fill(0);
    const shardTotals = new Array(modelUrls.length).fill(0);
    const emitAggregateProgress = () => {
      if (!progressCallback) {
        return;
      }
      const loaded = sumProgressValues(shardLoaded);
      const total = sumProgressValues(shardTotals);
      progressCallback({
        loaded,
        total: total > 0 ? total : loaded
      });
    };
    const controller = this._beginTransferAbortController();
    try {
      const fetchOptions = {
        cache: "no-store",
        ...controller?.signal ? { signal: controller.signal } : {}
      };
      const fetchTimeoutMs = this._resolveFetchTimeoutMs(options, 18e4);
      const chunkTimeoutMs = this._resolveStreamChunkTimeoutMs(options, 9e4);
      const cache = useCache && globalThis.caches && typeof globalThis.caches.open === "function" ? await globalThis.caches.open(this._modelCacheName) : null;
      if (!useCache || !cache) {
        this._modelCacheState = useCache ? "unavailable" : "disabled";
      }
      for (let shardIndex = 0; shardIndex < modelUrls.length; shardIndex += 1) {
        const shardUrl = modelUrls[shardIndex];
        const cacheKey = normalizeAbsoluteUrl(shardUrl);
        let response;
        let headerTotal = 0;
        if (cache) {
          const cached = options.force === true ? null : await cache.match(cacheKey);
          if (cached) {
            this._modelSource = "cache";
            this._modelCacheState = "hit";
            this._runtimeNotes.push("model_cache_hit");
            headerTotal = Number(cached.headers.get("content-length")) || 0;
            const resolved = headerTotal > 0 ? headerTotal : 1;
            shardLoaded[shardIndex] = resolved;
            shardTotals[shardIndex] = resolved;
            emitAggregateProgress();
            continue;
          }
          this._modelSource = "network";
          this._modelCacheState = options.force === true ? "refresh" : "miss";
        }
        response = await this._fetchWithTimeout(shardUrl, fetchOptions, fetchTimeoutMs);
        if (!response.ok) {
          throw new Error(
            `Failed to prefetch model shard: ${response.status} ${response.statusText}`
          );
        }
        headerTotal = Number(response.headers.get("content-length")) || 0;
        if (headerTotal > 0) {
          shardTotals[shardIndex] = headerTotal;
        }
        const putPromise = cache ? cache.put(cacheKey, response.clone()).then(() => ({ ok: true, error: null })).catch((error) => ({ ok: false, error })) : null;
        await drainResponseWithProgress(
          response,
          progressCallback ? (progress) => {
            const loaded = Number(progress?.loaded) || 0;
            const total = Number(progress?.total) || 0;
            shardLoaded[shardIndex] = loaded;
            if (total > 0) {
              shardTotals[shardIndex] = total;
            }
            emitAggregateProgress();
          } : null,
          { chunkTimeoutMs }
        );
        const finalLoaded = shardLoaded[shardIndex] > 0 ? shardLoaded[shardIndex] : shardTotals[shardIndex] > 0 ? shardTotals[shardIndex] : headerTotal;
        shardLoaded[shardIndex] = finalLoaded;
        if (finalLoaded > 0 && shardTotals[shardIndex] < finalLoaded) {
          shardTotals[shardIndex] = finalLoaded;
        }
        emitAggregateProgress();
        if (putPromise) {
          const putResult = await putPromise;
          if (putResult.ok) {
            this._modelCacheState = "stored";
            this._runtimeNotes.push("model_cache_stored");
          } else {
            const putErrorText = String(putResult.error || "").toLowerCase();
            if (putErrorText.includes("abort") || putErrorText.includes("cancel")) {
              throw putResult.error || new Error("Model cache prefetch was aborted.");
            }
            this._modelCacheState = "store_failed";
            this._runtimeNotes.push("model_cache_store_failed");
            throw new Error("Failed to store prefetched model in browser cache.");
          }
        }
      }
      if (modelUrls.length > 1) {
        this._runtimeNotes.push("model_split_cache_prefetched");
      }
      return 1;
    } catch (error) {
      this._modelCacheState = "error";
      const text = String(error || "").toLowerCase();
      if (text.includes("abort")) {
        this._runtimeNotes.push("model_cache_prefetch_aborted");
      } else if (text.includes("quota") || text.includes("storage")) {
        this._runtimeNotes.push("model_cache_quota_exceeded");
      } else {
        this._runtimeNotes.push("model_cache_error");
      }
      throw error;
    } finally {
      this._clearTransferAbortController(controller);
    }
  }
  async evictModelFromCache(url, options = {}) {
    this._modelCacheName = this._resolveCacheName(options);
    const modelUrls = expandModelShardUrls(url);
    if (modelUrls.length === 0) {
      return false;
    }
    if (!globalThis.caches || typeof globalThis.caches.open !== "function") {
      this._modelCacheState = "unavailable";
      return false;
    }
    try {
      const cache = await globalThis.caches.open(this._modelCacheName);
      let removedCount = 0;
      for (const modelUrl of modelUrls) {
        const cacheKey = normalizeAbsoluteUrl(modelUrl);
        const removed = await cache.delete(cacheKey);
        if (removed) {
          removedCount += 1;
        }
      }
      const removedAny = removedCount > 0;
      this._modelCacheState = removedAny ? "evicted" : "miss";
      if (removedAny) {
        this._runtimeNotes.push("model_cache_evicted");
      }
      if (modelUrls.length > 1) {
        this._runtimeNotes.push(`model_split_cache_evicted:${removedCount}/${modelUrls.length}`);
      }
      return removedAny;
    } catch (_) {
      this._modelCacheState = "error";
      this._runtimeNotes.push("model_cache_error");
      return false;
    }
  }
  async _ensureCore() {
    if (this._core) {
      this._applyCoreLogLevel();
      return this._core;
    }
    const candidates = [];
    if (this._config.coreModuleFactory) {
      candidates.push({
        variant: "custom",
        factoryPromise: Promise.resolve(this._config.coreModuleFactory),
        wasmUrl: this._config.wasmUrl
      });
    } else {
      if (this._preferMemory64 && typeof this._config.coreModuleUrlMem64 === "string" && this._config.coreModuleUrlMem64.length > 0) {
        candidates.push({
          variant: "wasm64",
          factoryPromise: importCoreFactory(this._config.coreModuleUrlMem64),
          wasmUrl: this._config.wasmUrlMem64 || this._config.wasmUrl
        });
      }
      candidates.push({
        variant: "wasm32",
        factoryPromise: importCoreFactory(this._config.coreModuleUrl ?? "./llama_webgpu_core.js"),
        wasmUrl: this._config.wasmUrl
      });
    }
    let lastError = null;
    for (const candidate of candidates) {
      if (candidate.variant === "wasm64") {
        this._runtimeNotes.push("core_mem64_attempt");
      }
      try {
        const moduleFactory = await candidate.factoryPromise;
        const initTimeoutMs = this._resolveCoreInitTimeoutMs();
        this._core = await Promise.race([
          moduleFactory({
            locateFile: (path, prefix) => {
              if (path.endsWith(".wasm") && candidate.wasmUrl) {
                return candidate.wasmUrl;
              }
              return `${prefix}${path}`;
            },
            print: (msg) => {
              this._emitLogger("log", msg);
            },
            printErr: (msg) => {
              const text = String(msg ?? "");
              this._lastCoreErrorText = text;
              const trimmed = text.trim();
              if (trimmed.length > 0) {
                const loweredTrimmed = trimmed.toLowerCase();
                const isGenericAbort = loweredTrimmed === "aborted(native code called abort())" || loweredTrimmed === "native code called abort()" || loweredTrimmed === "aborted";
                if (!isGenericAbort) {
                  this._lastCoreErrorHint = trimmed;
                }
              }
              const classification = this._classifyCoreErrorLine(text);
              if (classification === "ignore") {
                return;
              }
              if (classification === "warmup") {
                if (this._logLevel >= 2) {
                  this._suppressedWarmupWarningCount += 1;
                  if (!this._didReportWarmupWarningSuppression) {
                    this._didReportWarmupWarningSuppression = true;
                    if (this._logLevel <= 2) {
                      this._emitLogger(
                        "log",
                        "info: suppressing verbose warmup op logs; set bridge/runtime log level to Debug to inspect all warmup details."
                      );
                    }
                  }
                  this._pushRuntimeNote("warmup_warning_suppressed");
                  return;
                }
                this._emitLogger("log", trimmed.length > 0 ? trimmed : text);
                return;
              }
              if (classification === "warn") {
                this._emitLogger("warn", trimmed.length > 0 ? trimmed : text);
                return;
              }
              if (classification === "info") {
                this._emitLogger("log", trimmed.length > 0 ? trimmed : text);
                return;
              }
              this._emitLogger("error", trimmed.length > 0 ? trimmed : text);
            },
            onAbort: (reason) => {
              const text = String(reason ?? "").trim();
              if (text.length > 0) {
                const token = text.slice(0, 120).replace(/[\s;=]+/g, "_").replace(/[^a-zA-Z0-9._:-]/g, "");
                if (token.length > 0) {
                  this._runtimeNotes.push(`core_abort:${token}`);
                }
                this._emitLogger("error", `core abort: ${text}`);
              } else {
                this._runtimeNotes.push("core_abort");
                this._emitLogger("error", "core abort");
              }
            }
          }),
          new Promise((_, reject) => {
            globalThis.setTimeout(() => {
              reject(new Error(`Bridge core init timeout (${initTimeoutMs}ms)`));
            }, initTimeoutMs);
          })
        ]);
        this._coreVariant = candidate.variant === "wasm64" ? "wasm64" : "wasm32";
        if (candidate.variant === "wasm64") {
          this._runtimeNotes.push("core_mem64_active");
        } else if (candidate.variant === "wasm32") {
          this._runtimeNotes.push("core_wasm32_active");
        }
        break;
      } catch (error) {
        lastError = error;
        if (candidate.variant === "wasm64") {
          this._runtimeNotes.push("core_mem64_unavailable");
          continue;
        }
        throw error;
      }
    }
    if (!this._core) {
      throw lastError || new Error("Failed to initialize bridge core module");
    }
    this._applyCoreLogLevel();
    return this._core;
  }
  async _probeBackends() {
    try {
      const core = await this._ensureCore();
      const probeResult = Number(
        await core.ccall("llamadart_webgpu_probe", "number", [], [], { async: true })
      );
      const json = core.ccall("llamadart_webgpu_backends_json", "string", [], []);
      let parsed = [];
      try {
        parsed = JSON.parse(json || "[]");
      } catch (_) {
        parsed = [];
      }
      this._backendLabels = Array.isArray(parsed) ? parsed.map((v) => String(v)) : [];
      this._gpuActive = probeResult === 1;
    } catch (_err) {
      this._backendLabels = [];
      this._gpuActive = false;
    }
    return this._gpuActive;
  }
  async loadModelFromUrl(url, options = {}, resolvedLoadState = null) {
    const operationSignal = options?.signal || null;
    const abortMessage = "Model load was cancelled.";
    throwIfAborted(operationSignal, abortMessage);
    this._abortRequested = false;
    this._runtimeNotes = [];
    this._suppressedWarmupWarningCount = 0;
    this._didReportWarmupWarningSuppression = false;
    await this._probeBackends();
    throwIfAborted(operationSignal, abortMessage);
    const core = await this._ensureCore();
    throwIfAborted(operationSignal, abortMessage);
    const modelUrls = expandModelShardUrls(url);
    if (modelUrls.length === 0) {
      throw new Error("Model URL is empty.");
    }
    this._releaseLoadedModel(core);
    this._mmProjSourceUrl = null;
    const configuredPoolHint = Number(this._threadPoolSizeHint);
    this._syncThreadPoolSizeHintFromCore();
    const coreSupportsPthreads = this._coreSupportsPthreads();
    this._pushRuntimeNote(`core_pthreads:${coreSupportsPthreads ? 1 : 0}`);
    if (!coreSupportsPthreads && Number.isFinite(configuredPoolHint) && configuredPoolHint > 1) {
      this._runtimeNotes.push("threads_capped_no_pthread");
    }
    this._nCtx = Number(options.nCtx) > 0 ? Number(options.nCtx) : this._nCtx;
    const requestedThreads = Number(options.nThreads);
    if (Number.isFinite(requestedThreads) && requestedThreads > 0) {
      this._threads = this._capThreadsToPool(requestedThreads);
    } else {
      this._threads = this._capThreadsToPool(this._resolveAutoThreadCount());
    }
    const requestedThreadsBatch = Number(options.nThreadsBatch);
    if (Number.isFinite(requestedThreadsBatch) && requestedThreadsBatch > 0) {
      this._threadsBatch = this._capThreadsToPool(
        requestedThreadsBatch,
        { noteTag: "threads_batch_capped_pool" }
      );
    } else {
      this._threadsBatch = this._threads;
    }
    const requestedGpuLayers = Number(options.nGpuLayers);
    if (Number.isFinite(requestedGpuLayers)) {
      this._nGpuLayers = Math.trunc(requestedGpuLayers);
    }
    const isCpuModelMode = this._nGpuLayers === 0;
    const requestedBatch = Number(options.nBatch);
    this._nBatch = Number.isFinite(requestedBatch) && requestedBatch > 0 ? Math.max(32, Math.trunc(requestedBatch)) : isCpuModelMode ? Math.min(this._nCtx, 512) : 0;
    const requestedUbatch = Number(options.nUbatch);
    this._nUbatch = Number.isFinite(requestedUbatch) && requestedUbatch > 0 ? Math.max(32, Math.trunc(requestedUbatch)) : isCpuModelMode ? Math.min(this._nBatch || 256, 256) : 0;
    if (this._nBatch > 0 && this._nBatch > this._nCtx) {
      this._nBatch = this._nCtx;
    }
    if (this._nUbatch > 0 && this._nBatch > 0 && this._nUbatch > this._nBatch) {
      this._nUbatch = this._nBatch;
    }
    if (resolvedLoadState) {
      Object.assign(this, resolvedLoadState);
      this._threads = this._capThreadsToPool(this._threads);
      this._threadsBatch = this._capThreadsToPool(this._threadsBatch, { noteTag: "threads_batch_capped_pool" });
    } else {
      this._resolveNativeLoadOptions(options);
    }
    if (Number.isFinite(this._threadPoolSizeHint) && this._threadPoolSizeHint > 0) {
      this._pushRuntimeNote(`thread_pool_size:${this._threadPoolSizeHint}`);
    }
    if (!isCrossOriginIsolatedRuntime()) {
      this._runtimeNotes.push("threads_capped_no_coi");
      if (this._threads > 1) {
        this._threads = 1;
      }
      if (this._threadsBatch > 1) {
        this._threadsBatch = 1;
      }
    }
    this._pushRuntimeNote(`threads_batch:${this._threadsBatch}`);
    if (this._nBatch > 0) {
      this._pushRuntimeNote(`n_batch:${this._nBatch}`);
    }
    if (this._nUbatch > 0) {
      this._pushRuntimeNote(`n_ubatch:${this._nUbatch}`);
    }
    if (this._nSeqMax > 0) {
      this._pushRuntimeNote(`n_seq_max:${this._nSeqMax}`);
    }
    if (!resolvedLoadState && isCpuModelMode && !Number.isFinite(requestedBatch) && !Number.isFinite(requestedUbatch)) {
      this._runtimeNotes.push("cpu_batch_tuned_default");
    }
    if (this._isSafari && this._nGpuLayers > 0) {
      const requestedSafariMaxLayers = Number(options.safariMaxGpuLayers);
      const safariMaxGpuLayers = Number.isFinite(requestedSafariMaxLayers) ? Math.max(1, Math.trunc(requestedSafariMaxLayers)) : 1;
      if (this._nGpuLayers > safariMaxGpuLayers) {
        this._nGpuLayers = safariMaxGpuLayers;
        this._runtimeNotes.push(`safari_gpu_layers_capped:${safariMaxGpuLayers}`);
      }
    }
    this._loadedModelUrl = cloneModelSource(url);
    if (modelUrls.length > 1) {
      this._runtimeNotes.push(`model_split_detected:${modelUrls.length}`);
    }
    let loadedViaRemoteFetch = false;
    let remoteFetchReloadUrl = null;
    const remoteFetchReloadChunkBytes = this._resolveRemoteFetchChunkBytes(options);
    if (modelUrls.length === 1) {
      const remoteResult = await this._tryLoadModelFromRemoteFetchBackend(
        core,
        modelUrls[0],
        options
      );
      loadedViaRemoteFetch = remoteResult.loaded === true;
      if (loadedViaRemoteFetch) {
        remoteFetchReloadUrl = modelUrls[0];
      }
    } else {
      this._runtimeNotes.push("model_fetch_backend_skipped_split");
    }
    if (!loadedViaRemoteFetch) {
      ensureFsDirectory(core.FS, "/models");
      const progressCallback = typeof options.progressCallback === "function" ? options.progressCallback : null;
      const shardLoaded = new Array(modelUrls.length).fill(0);
      const shardTotals = new Array(modelUrls.length).fill(0);
      const emitAggregateProgress = () => {
        if (!progressCallback) {
          return;
        }
        const loaded = sumProgressValues(shardLoaded);
        const total = sumProgressValues(shardTotals);
        progressCallback({
          loaded,
          total: total > 0 ? total : loaded
        });
      };
      const modelPaths = [];
      let totalModelBytes = 0;
      const maxStreamResumeRetries = Number.isFinite(options.streamResumeRetries) ? Math.max(0, Math.trunc(options.streamResumeRetries)) : 8;
      const streamChunkTimeoutMs = this._resolveStreamChunkTimeoutMs(
        options,
        9e4
      );
      try {
        for (let shardIndex = 0; shardIndex < modelUrls.length; shardIndex += 1) {
          const shardUrl = modelUrls[shardIndex];
          const fileName = basenameFromUrl(shardUrl);
          const modelPath = `/models/${fileName}`;
          modelPaths.push(modelPath);
          let shardBytes = 0;
          let resumeOffset = 0;
          let resumeAttempt = 0;
          let activeShardUrl = shardUrl;
          let knownTotalBytes = 0;
          while (true) {
            throwIfAborted(operationSignal, abortMessage);
            const requestHeaders = resumeOffset > 0 ? { Range: `bytes=${resumeOffset}-` } : null;
            const response = await this._getCachedModelResponse(activeShardUrl, {
              ...options,
              requireReadableStream: true,
              useCache: requestHeaders ? false : options.useCache,
              force: requestHeaders ? true : options.force,
              requestHeaders
            });
            if (!response.ok) {
              throw new Error(
                `Failed to fetch model shard: ${response.status} ${response.statusText}`
              );
            }
            const responseUrl = typeof response.url === "string" && response.url.length > 0 ? response.url : activeShardUrl;
            if (responseUrl !== activeShardUrl) {
              activeShardUrl = responseUrl;
              if (resumeOffset > 0) {
                this._runtimeNotes.push("model_stream_resume_redirect");
              }
            }
            if (resumeOffset > 0 && response.status !== 206) {
              this._runtimeNotes.push(`model_stream_resume_status:${response.status}`);
              throw new Error(
                `Range resume not honored for model shard: ${response.status} ${response.statusText}`
              );
            }
            if (!hasReadableResponseStream(response)) {
              this._runtimeNotes.push("model_response_nostream");
              const declaredBytes = inferResponseTotalBytes(response, 0);
              const declaredText = declaredBytes > 0 ? `${declaredBytes} bytes` : "unknown size";
              throw new Error(
                `Model response did not expose a readable stream (${declaredText}).`
              );
            }
            this._runtimeNotes.push("model_response_stream");
            const responseTotal = inferResponseTotalBytes(response, knownTotalBytes);
            if (responseTotal > 0) {
              knownTotalBytes = responseTotal;
              shardTotals[shardIndex] = responseTotal;
            }
            try {
              shardBytes = await writeResponseToFsFileWithProgress(
                response,
                core.FS,
                modelPath,
                progressCallback ? (progress) => {
                  const loaded = Number(progress?.loaded) || 0;
                  const total = Number(progress?.total) || 0;
                  shardLoaded[shardIndex] = loaded;
                  if (total > 0) {
                    shardTotals[shardIndex] = total;
                  }
                  emitAggregateProgress();
                } : null,
                {
                  useBigIntPosition: this._coreVariant === "wasm64",
                  startOffset: resumeOffset,
                  preservePartialOnError: true,
                  totalBytes: knownTotalBytes,
                  chunkTimeoutMs: streamChunkTimeoutMs,
                  signal: operationSignal,
                  abortMessage
                }
              );
              break;
            } catch (error) {
              const text = String(error || "").toLowerCase();
              const loadedBytes = Number(
                error?.llamadartLoadedBytes
              );
              if (Number.isFinite(loadedBytes) && loadedBytes >= 0) {
                const normalizedLoaded = Math.trunc(loadedBytes);
                this._runtimeNotes.push(`model_fs_write_loaded:${normalizedLoaded}`);
                if (normalizedLoaded > resumeOffset) {
                  resumeOffset = normalizedLoaded;
                  shardLoaded[shardIndex] = normalizedLoaded;
                  if (knownTotalBytes > 0 && shardTotals[shardIndex] < knownTotalBytes) {
                    shardTotals[shardIndex] = knownTotalBytes;
                  }
                  emitAggregateProgress();
                }
              }
              const shouldRetryResume = isRetryableStreamNetworkError(error) && resumeOffset > 0 && resumeAttempt < maxStreamResumeRetries;
              if (shouldRetryResume) {
                resumeAttempt += 1;
                this._runtimeNotes.push(`model_stream_resume_retry:${resumeAttempt}`);
                this._runtimeNotes.push(`model_stream_resume_offset:${resumeOffset}`);
                continue;
              }
              if (text.includes("bigint")) {
                this._runtimeNotes.push("model_fs_write_bigint_error");
              } else if (text.includes("abort")) {
                this._runtimeNotes.push("model_fs_write_abort");
              } else if (text.includes("array buffer allocation failed")) {
                this._runtimeNotes.push("model_fs_write_arraybuffer_oom");
              } else if (isRetryableStreamNetworkError(error)) {
                this._runtimeNotes.push("model_fs_write_network_error");
                this._runtimeNotes.push("model_fs_write_failed");
              } else {
                this._runtimeNotes.push("model_fs_write_failed");
              }
              throw error;
            }
          }
          shardLoaded[shardIndex] = shardBytes;
          if (shardTotals[shardIndex] < shardBytes) {
            shardTotals[shardIndex] = shardBytes;
          }
          totalModelBytes += shardBytes;
          emitAggregateProgress();
        }
        this._modelPaths = modelPaths;
        this._modelPath = modelPaths[0] || null;
        this._modelBytes = totalModelBytes;
        let rc = 0;
        let nativeLoadStarted = false;
        let nativeAbortCleanupDone = false;
        try {
          throwIfAborted(operationSignal, abortMessage);
          nativeLoadStarted = true;
          rc = Number(
            await core.ccall(
              "llamadart_webgpu_load_model",
              "number",
              [
                "string",
                "number",
                "number",
                "number",
                "number",
                "number",
                "number",
                ...this._nativeLoadOptionTypes()
              ],
              [
                this._modelPath,
                this._nCtx,
                this._threads,
                this._threadsBatch,
                this._nBatch,
                this._nUbatch,
                this._nGpuLayers,
                ...this._nativeLoadOptionValues()
              ],
              { async: true }
            )
          );
          if (operationSignal?.aborted) {
            try {
              core.ccall("llamadart_webgpu_shutdown", null, [], []);
              nativeAbortCleanupDone = true;
            } catch (_) {
            }
            throw createAbortError(abortMessage);
          }
        } catch (error) {
          if (operationSignal?.aborted && nativeLoadStarted && !nativeAbortCleanupDone) {
            try {
              core.ccall("llamadart_webgpu_shutdown", null, [], []);
              nativeAbortCleanupDone = true;
            } catch (_) {
            }
            error = createAbortError(abortMessage);
          }
          const text = String(error || "").toLowerCase();
          if (text.includes("bigint")) {
            this._runtimeNotes.push("model_load_ccall_bigint_error");
          } else if (text.includes("abort")) {
            this._runtimeNotes.push("model_load_ccall_abort");
          } else {
            this._runtimeNotes.push("model_load_ccall_failed");
          }
          throw error;
        }
        if (rc !== 0) {
          throw new Error(this._coreErrorMessage("Failed to load GGUF model", rc));
        }
        if (modelUrls.length > 1) {
          this._runtimeNotes.push(`model_split_loaded:${modelUrls.length}`);
        }
      } catch (error) {
        this._modelBytes = 0;
        this._modelPath = null;
        this._modelPaths = [];
        if (operationSignal?.aborted) {
          this._loadedModelUrl = null;
        }
        for (const modelPath of modelPaths) {
          this._deleteFsFile(modelPath);
        }
        throw error;
      }
    }
    const shouldProbeSafariGpu = this._isSafari && this._nGpuLayers > 0 && options.safariGpuProbe !== false;
    if (shouldProbeSafariGpu) {
      if (loadedViaRemoteFetch) {
        this._runtimeNotes.push("safari_probe_on_fetch_backend");
      }
      const defaultProbePrompts = [
        "user: hi\nassistant:",
        "user: say hello in one short sentence\nassistant:"
      ];
      const probePrompts = Array.isArray(options.safariProbePrompts) ? options.safariProbePrompts.map((v) => String(v || "").trim()).filter((v) => v.length > 0) : typeof options.safariProbePrompt === "string" && options.safariProbePrompt.trim().length > 0 ? [options.safariProbePrompt.trim()] : defaultProbePrompts;
      const probeTokensRaw = Number(options.safariProbeTokens);
      const probeTokens = Number.isFinite(probeTokensRaw) && probeTokensRaw > 0 ? Math.min(Math.trunc(probeTokensRaw), 96) : 48;
      const runProbe = async (probePrompt, probeSeed) => {
        try {
          const probeOutput = await this.createCompletion(probePrompt, {
            nPredict: probeTokens,
            temp: 0,
            topK: 1,
            topP: 1,
            penalty: 1,
            seed: probeSeed
          });
          return !looksLikeCorruptedGeneration(probeOutput);
        } catch (_) {
          return false;
        }
      };
      let initialProbePassed = true;
      for (let i = 0; i < probePrompts.length; i += 1) {
        const ok = await runProbe(probePrompts[i], i + 1);
        if (!ok) {
          initialProbePassed = false;
          break;
        }
      }
      if (!initialProbePassed) {
        this._runtimeNotes.push("safari_gpu_probe_failed");
        const retryCandidates = [];
        if (this._nGpuLayers > 1) {
          retryCandidates.push(1);
        }
        retryCandidates.push(0);
        let stabilized = false;
        for (const candidateLayers of retryCandidates) {
          try {
            core.ccall("llamadart_webgpu_shutdown", null, [], []);
          } catch (_) {
          }
          let retryRc = 0;
          if (loadedViaRemoteFetch) {
            const reloadUrl = remoteFetchReloadUrl || modelUrls[0] || null;
            if (!reloadUrl) {
              continue;
            }
            this._runtimeNotes.push(`safari_probe_remote_fetch_retry:${candidateLayers}`);
            retryRc = Number(
              await core.ccall(
                "llamadart_webgpu_load_model_from_url",
                "number",
                [
                  "string",
                  "number",
                  "number",
                  "number",
                  "number",
                  "number",
                  "number",
                  "number",
                  ...this._nativeLoadOptionTypes()
                ],
                [
                  reloadUrl,
                  this._nCtx,
                  this._threads,
                  this._threadsBatch,
                  this._nBatch,
                  this._nUbatch,
                  candidateLayers,
                  remoteFetchReloadChunkBytes,
                  ...this._nativeLoadOptionValues()
                ],
                { async: true }
              )
            );
          } else {
            retryRc = Number(
              await core.ccall(
                "llamadart_webgpu_load_model",
                "number",
                [
                  "string",
                  "number",
                  "number",
                  "number",
                  "number",
                  "number",
                  "number",
                  ...this._nativeLoadOptionTypes()
                ],
                [
                  this._modelPath,
                  this._nCtx,
                  this._threads,
                  this._threadsBatch,
                  this._nBatch,
                  this._nUbatch,
                  candidateLayers,
                  ...this._nativeLoadOptionValues()
                ],
                { async: true }
              )
            );
          }
          if (retryRc !== 0) {
            continue;
          }
          this._nGpuLayers = candidateLayers;
          if (candidateLayers === 0) {
            this._runtimeNotes.push("safari_fallback_cpu");
            stabilized = true;
            break;
          }
          let retryProbePassed = true;
          for (let i = 0; i < probePrompts.length; i += 1) {
            const ok = await runProbe(probePrompts[i], i + 11);
            if (!ok) {
              retryProbePassed = false;
              break;
            }
          }
          if (retryProbePassed) {
            this._runtimeNotes.push(`safari_gpu_layers_capped:${candidateLayers}`);
            stabilized = true;
            break;
          }
        }
        if (!stabilized) {
          throw new Error("Safari GPU probe failed and fallback attempts were unsuccessful.");
        }
      } else {
        this._runtimeNotes.push("safari_gpu_probe_passed");
      }
    }
    try {
      const effectiveNctx = Number(core.ccall("llamadart_webgpu_get_context_size", "number", [], []));
      if (effectiveNctx > 0) {
        this._nCtx = effectiveNctx;
      }
    } catch (_) {
    }
    this._mmProjPath = null;
    this._mmSupportsVision = false;
    this._mmSupportsAudio = false;
    this._mediaFileCounter = 0;
    this._stateFileCounter = 0;
    this._stagedMediaPaths = [];
    this._gpuActive = this._gpuActive && this._nGpuLayers > 0;
    if (this._releaseModelFiles() > 0) {
      this._runtimeNotes.push("model_file_released");
    }
    this._emitSuppressedWarmupWarningSummaryIfNeeded();
    return 1;
  }
  async loadMultimodalProjector(url) {
    if (!this._core || this._modelBytes <= 0) {
      throw new Error("No model loaded. Call loadModelFromUrl first.");
    }
    if (typeof url !== "string" || url.length === 0) {
      throw new Error("Multimodal projector URL/path is empty.");
    }
    const core = await this._ensureCore();
    ensureFsDirectory(core.FS, "/mmproj");
    const fileName = basenameFromUrl(url);
    const mmprojPath = `/mmproj/${fileName}`;
    const fetchTimeoutMs = this._resolveFetchTimeoutMs({}, 18e4);
    const chunkTimeoutMs = this._resolveStreamChunkTimeoutMs({}, 9e4);
    let lastError = null;
    for (let attempt = 0; attempt < 2; attempt += 1) {
      try {
        const response = await this._fetchWithTimeout(
          url,
          { cache: "no-store" },
          fetchTimeoutMs
        );
        if (!response.ok) {
          throw new Error(
            `Failed to fetch multimodal projector: ${response.status} ${response.statusText}`
          );
        }
        this._mmProjPath = mmprojPath;
        await writeResponseToFsFileWithProgress(
          response,
          core.FS,
          this._mmProjPath,
          null,
          {
            useBigIntPosition: this._coreVariant === "wasm64",
            chunkTimeoutMs
          }
        );
        const rc = Number(
          await core.ccall(
            "llamadart_webgpu_mmproj_load",
            "number",
            ["string"],
            [this._mmProjPath],
            { async: true }
          )
        );
        if (rc !== 0) {
          throw new Error(this._coreErrorMessage("Failed to load multimodal projector", rc));
        }
        this._mmSupportsVision = Number(
          core.ccall("llamadart_webgpu_mmproj_supports_vision", "number", [], [])
        ) === 1;
        this._mmSupportsAudio = Number(
          core.ccall("llamadart_webgpu_mmproj_supports_audio", "number", [], [])
        ) === 1;
        this._mmProjSourceUrl = url;
        return 1;
      } catch (error) {
        lastError = error;
        this._mmProjPath = null;
        this._mmSupportsVision = false;
        this._mmSupportsAudio = false;
        this._deleteFsFile(mmprojPath);
        const retryable = isRetryableStreamNetworkError(error);
        if (!retryable || attempt >= 1) {
          throw error;
        }
        this._runtimeNotes.push(`mmproj_load_retry:${attempt + 1}`);
      }
    }
    if (lastError) {
      throw lastError;
    }
    throw new Error("Failed to load multimodal projector");
  }
  async unloadMultimodalProjector() {
    if (!this._core) {
      this._mmProjPath = null;
      this._mmSupportsVision = false;
      this._mmSupportsAudio = false;
      this._mmProjSourceUrl = null;
      return;
    }
    const mmprojPath = this._mmProjPath;
    const rc = Number(
      this._core.ccall("llamadart_webgpu_mmproj_free", "number", [], [])
    );
    if (rc !== 0) {
      throw new Error(this._coreErrorMessage("Failed to unload multimodal projector", rc));
    }
    this._deleteFsFile(mmprojPath);
    this._mmProjPath = null;
    this._mmSupportsVision = false;
    this._mmSupportsAudio = false;
    this._mmProjSourceUrl = null;
  }
  // Bytes the core heap can still allocate, or null for a core without the
  // probe.
  _heapHeadroomBytes(core) {
    try {
      const value = Number(core.ccall("llamadart_webgpu_heap_headroom_bytes", "number", [], []));
      return Number.isFinite(value) && value >= 0 ? value : null;
    } catch (_) {
      return null;
    }
  }
  // Throws when `bytes` more would not fit in the core heap, naming the
  // memory64 core when the wasm32 core's 4 GiB limit is the reason.
  _requireHeapHeadroom(core, bytes, what) {
    const headroom = this._heapHeadroomBytes(core);
    if (headroom == null || !(bytes > headroom)) {
      return;
    }
    const mib = (value, round) => `${round(value / (1024 * 1024))} MiB`;
    const detail = this._coreVariant === "wasm32" ? "the wasm32 core addresses at most 4 GiB; memory64 is required: load the bridge with coreModuleUrlMem64 and preferMemory64 in a browser with WebAssembly memory64 support" : "free memory by unloading other models or use a smaller draft model";
    throw new Error(
      `${what} needs about ${mib(bytes, Math.ceil)} but only ${mib(headroom, Math.floor)} of WebAssembly memory is left; ${detail}.`
    );
  }
  async loadDraftModel(url, options = {}) {
    if (!this._core || this._modelBytes <= 0) {
      throw new Error("No model loaded. Call loadModelFromUrl first.");
    }
    if (typeof url !== "string" || url.length === 0) {
      throw new Error("Draft model URL is empty.");
    }
    const core = this._core;
    const abortMessage = "Draft model load was cancelled.";
    throwIfAborted(options.signal || null, abortMessage);
    const rc = Number(await core.ccall("llamadart_webgpu_draft_model_free", "number", [], [], { async: true }));
    if (rc !== 0) {
      throw new Error(this._coreErrorMessage("Failed to release the draft model", rc));
    }
    this._draftModel = null;
    ensureFsDirectory(core.FS, "/draft");
    const draftPath = `/draft/${basenameFromUrl(url)}`;
    const controller = this._beginTransferAbortController();
    const onCallerAbort = () => controller?.abort();
    options.signal?.addEventListener("abort", onCallerAbort, { once: true });
    const signal = controller?.signal ?? options.signal ?? null;
    const targetCacheFields = [this._modelSource, this._modelCacheState, this._modelCacheName];
    try {
      let response;
      try {
        response = await this._getCachedModelResponse(url, {
          useCache: options.useCache,
          force: options.force,
          signal,
          requireReadableStream: true
        });
      } finally {
        [this._modelSource, this._modelCacheState, this._modelCacheName] = targetCacheFields;
      }
      if (!response.ok) {
        throw new Error(`Failed to fetch draft model: ${response.status} ${response.statusText}`);
      }
      const totalBytes = inferResponseTotalBytes(response, 0);
      if (totalBytes > 0) {
        this._requireHeapHeadroom(
          core,
          this._nGpuLayers === 0 ? totalBytes * 2 : totalBytes,
          "The draft model"
        );
      }
      const progressCallback = typeof options.progressCallback === "function" ? options.progressCallback : null;
      await writeResponseToFsFileWithProgress(
        response,
        core.FS,
        draftPath,
        progressCallback,
        {
          useBigIntPosition: this._coreVariant === "wasm64",
          totalBytes,
          chunkTimeoutMs: this._resolveStreamChunkTimeoutMs({}, 9e4),
          signal,
          abortMessage
        }
      );
      throwIfAborted(signal, abortMessage);
      const loadRc = Number(
        await core.ccall(
          "llamadart_webgpu_draft_model_load",
          "number",
          ["string"],
          [draftPath],
          { async: true }
        )
      );
      if (loadRc !== 0) {
        let message = this._coreErrorMessage("Failed to load draft model", loadRc);
        if (this._coreVariant === "wasm32" && /memory|alloc/i.test(message)) {
          message += "; the wasm32 core addresses at most 4 GiB, so a larger draft model needs the memory64 core";
        }
        throw new Error(message);
      }
      const info = JSON.parse(
        core.ccall("llamadart_webgpu_draft_model_info_json", "string", [], []) || "{}"
      );
      this._draftModel = {
        url,
        options: { useCache: options.useCache, force: false },
        info
      };
      return { ...info };
    } finally {
      options.signal?.removeEventListener("abort", onCallerAbort);
      this._clearTransferAbortController(controller);
      this._deleteFsFile(draftPath);
    }
  }
  async unloadDraftModel() {
    if (!this._core) {
      this._draftModel = null;
      return;
    }
    const rc = Number(
      await this._core.ccall("llamadart_webgpu_draft_model_free", "number", [], [], { async: true })
    );
    if (rc !== 0) {
      throw new Error(this._coreErrorMessage("Failed to unload the draft model", rc));
    }
    this._draftModel = null;
  }
  // Writes one n-gram cache source to the core filesystem and returns its
  // path; the caller deletes it.
  async _stageNgramCache(source, name) {
    const core = this._core;
    ensureFsDirectory(core.FS, "/speculative");
    this._speculativeFileCounter += 1;
    const path = `/speculative/${name}_${Date.now()}_${this._speculativeFileCounter}.lcs`;
    if (typeof source === "string") {
      const response = await this._fetchWithTimeout(
        source,
        { cache: "no-store" },
        this._resolveFetchTimeoutMs({}, 18e4)
      );
      if (!response.ok) {
        throw new Error(`Failed to fetch the ${name} n-gram cache: ${response.status} ${response.statusText}`);
      }
      core.FS.writeFile(path, new Uint8Array(await response.arrayBuffer()));
    } else {
      const bytes = source instanceof ArrayBuffer ? new Uint8Array(source) : new Uint8Array(source.buffer, source.byteOffset, source.byteLength);
      core.FS.writeFile(path, bytes);
    }
    return path;
  }
  // Passes the speculative settings to the core for the next
  // begin_generation, which consumes them. Returns the staged cache files.
  async _setNextSpeculative(speculative, maxTokens, stagedPaths) {
    const core = this._core;
    const staticPath = speculative.ngramCacheStatic == null ? "" : await this._stageNgramCache(speculative.ngramCacheStatic, "static");
    if (staticPath) {
      stagedPaths.push(staticPath);
    }
    const dynamicPath = speculative.ngramCacheDynamic == null ? "" : await this._stageNgramCache(speculative.ngramCacheDynamic, "dynamic");
    if (dynamicPath) {
      stagedPaths.push(dynamicPath);
    }
    const rc = Number(core.ccall(
      "llamadart_webgpu_set_next_speculative",
      "number",
      [
        "string",
        "number",
        "number",
        "number",
        "number",
        "number",
        "number",
        "number",
        "number",
        "number",
        "number",
        "string",
        "string",
        "number"
      ],
      [
        speculative.strategies.join(","),
        speculative.draftTokenMax,
        speculative.draftTokenMin,
        speculative.minProbability,
        speculative.draftSplitProbability,
        speculative.ngramSizeN,
        speculative.ngramSizeM,
        speculative.ngramMinHits,
        speculative.ngramMatch,
        speculative.ngramTokenMin,
        speculative.ngramTokenMax,
        staticPath,
        dynamicPath,
        maxTokens
      ]
    ));
    if (rc !== 0) {
      throw new Error(this._coreErrorMessage("Failed to configure speculative decoding", rc));
    }
  }
  supportsVision() {
    return this._mmSupportsVision;
  }
  supportsAudio() {
    return this._mmSupportsAudio;
  }
  getTextToSpeechCapabilities() {
    if (!this._core) {
      return {
        apiVersion: 0,
        supported: false,
        modelType: 0,
        capabilities: 0,
        supportsLanguage: false,
        supportsSpeakerReference: false,
        sampleRate: 0,
        channels: 0,
        reason: "WebGPU core is not initialized"
      };
    }
    const raw = this._core.ccall(
      "llamadart_webgpu_tts_info_json",
      "string",
      [],
      []
    ) || "{}";
    try {
      return JSON.parse(raw);
    } catch (_) {
      return {
        apiVersion: 0,
        supported: false,
        modelType: 0,
        capabilities: 0,
        supportsLanguage: false,
        supportsSpeakerReference: false,
        sampleRate: 0,
        channels: 0,
        reason: "WebGPU TTS capability response is invalid"
      };
    }
  }
  _ensureTextToSpeechDir() {
    try {
      this._core.FS.mkdir("/tts");
    } catch (_) {
    }
  }
  async synthesizeSpeech(options = {}) {
    if (this._modelBytes <= 0) {
      throw new Error("No model loaded. Call loadModelFromUrl first.");
    }
    if (options.signal?.aborted) {
      throw new DOMException("Text-to-speech synthesis was cancelled.", "AbortError");
    }
    const capabilities = this.getTextToSpeechCapabilities();
    if (capabilities.supported !== true) {
      throw new Error(
        capabilities.reason || "Loaded model/projector does not support text-to-speech."
      );
    }
    const text = typeof options.text === "string" ? options.text : "";
    if (text.length === 0) {
      throw new Error("Text-to-speech input text is empty.");
    }
    if (this._textToSpeechActive) {
      throw new Error("Text-to-speech synthesis is already active.");
    }
    this._ensureTextToSpeechDir();
    const taskId = `${Date.now()}_${++this._mediaFileCounter}`;
    const speakerPath = `/tts/speaker_${taskId}.bin`;
    const outputPath = `/tts/output_${taskId}.pcm`;
    const speakerBytes = toUint8Array(options.speakerAudio);
    if (speakerBytes && speakerBytes.length > 0) {
      this._core.FS.writeFile(speakerPath, speakerBytes);
    }
    const promptBatchSize = Number(options.promptBatchSize) > 0 ? Math.max(1, Math.trunc(Number(options.promptBatchSize))) : 512;
    const maxFrames = Number(options.maxFrames) > 0 ? Math.max(1, Math.trunc(Number(options.maxFrames))) : 512;
    const topK = Number.isFinite(options.topK) ? Math.max(0, Math.trunc(Number(options.topK))) : 40;
    const topP = Number.isFinite(options.topP) ? Number(options.topP) : 0.95;
    const minP = Number.isFinite(options.minP) ? Number(options.minP) : 0;
    const temperature = Number.isFinite(options.temperature) ? Number(options.temperature) : 0.8;
    const seed = Number.isInteger(options.seed) ? Number(options.seed) >>> 0 : Math.floor(Math.random() * 4294967295) >>> 0;
    const language = typeof options.language === "string" ? options.language : "";
    let started = false;
    this._textToSpeechActive = true;
    this._textToSpeechDone = new Promise((resolve) => {
      this._resolveTextToSpeechDone = resolve;
    });
    try {
      this._abortRequested = false;
      const startRc = Number(
        await this._core.ccall(
          "llamadart_webgpu_tts_start",
          "number",
          ["string", "string", "string", "number", "number", "number", "number", "number", "number", "number"],
          [
            text,
            language,
            speakerBytes && speakerBytes.length > 0 ? speakerPath : null,
            promptBatchSize,
            maxFrames,
            topK,
            topP,
            minP,
            temperature,
            seed
          ],
          { async: true }
        )
      );
      if (startRc !== 0) {
        throw new Error(this._coreErrorMessage("Failed to start text-to-speech synthesis", startRc));
      }
      started = true;
      let progress = {};
      for (; ; ) {
        if (this._abortRequested || options.signal?.aborted) {
          this.cancel();
        }
        const stepRc = Number(
          await this._core.ccall(
            "llamadart_webgpu_tts_step",
            "number",
            [],
            [],
            { async: true }
          )
        );
        const rawProgress = this._core.ccall(
          "llamadart_webgpu_tts_progress_json",
          "string",
          [],
          []
        ) || "{}";
        progress = JSON.parse(rawProgress);
        options.onProgress?.(progress);
        if (stepRc === -6 || progress.state === 4) {
          throw new DOMException("Text-to-speech synthesis was cancelled.", "AbortError");
        }
        if (stepRc !== 0) {
          throw new Error(this._coreErrorMessage("Text-to-speech synthesis failed", stepRc));
        }
        if (progress.state === 3) {
          break;
        }
        await new Promise((resolve) => setTimeout(resolve, 0));
      }
      const outputRc = Number(
        this._core.ccall(
          "llamadart_webgpu_tts_write_pcm",
          "number",
          ["string"],
          [outputPath]
        )
      );
      if (outputRc !== 0) {
        throw new Error(this._coreErrorMessage("Failed to read synthesized speech", outputRc));
      }
      const bytes = this._core.FS.readFile(outputPath);
      if (!(bytes instanceof Uint8Array) || bytes.byteLength === 0 || bytes.byteLength % 4 !== 0) {
        throw new Error("Synthesized PCM output is empty or malformed.");
      }
      const copy = bytes.buffer.slice(bytes.byteOffset, bytes.byteOffset + bytes.byteLength);
      const pcm = new Float32Array(copy);
      return {
        pcm,
        sampleRate: Number(capabilities.sampleRate) || 0,
        channels: Number(capabilities.channels) || 1,
        sampleCount: pcm.length,
        framesGenerated: Number(progress.framesGenerated) || 0,
        truncated: progress.truncated === true
      };
    } finally {
      if (started) {
        try {
          this._core.ccall("llamadart_webgpu_tts_reset", "number", [], []);
        } catch (_) {
        }
      }
      this._deleteFsFile(speakerPath);
      this._deleteFsFile(outputPath);
      this._textToSpeechActive = false;
      this._resolveTextToSpeechDone?.();
      this._resolveTextToSpeechDone = null;
      this._textToSpeechDone = null;
    }
  }
  getDecisionCapabilities() {
    const unsupported = (reason) => ({
      apiVersion: DECISION_API_VERSION,
      supported: false,
      reason
    });
    const core = this._core;
    if (!core) {
      return unsupported("WebGPU core is not initialized");
    }
    if (typeof core._llamadart_webgpu_decision_capabilities_json !== "function") {
      return unsupported("This WebGPU core build does not include decision heads.");
    }
    const coreVersion = Number(
      core.ccall("llamadart_webgpu_decision_api_version", "number", [], [])
    );
    if (coreVersion !== DECISION_API_VERSION) {
      return unsupported(
        `The WebGPU core implements decision API version ${coreVersion}; this bridge needs version ${DECISION_API_VERSION}.`
      );
    }
    const raw = core.ccall(
      "llamadart_webgpu_decision_capabilities_json",
      "string",
      [],
      []
    ) || "{}";
    try {
      return JSON.parse(raw);
    } catch (_) {
      return unsupported("WebGPU decision capability response is invalid");
    }
  }
  _ensureDecisionDir() {
    try {
      this._core.FS.mkdir("/decision");
    } catch (_) {
    }
  }
  async _stageDecisionHeadFromUrl(url, headPath, onProgress) {
    const fetchTimeoutMs = this._resolveFetchTimeoutMs({}, 18e4);
    const chunkTimeoutMs = this._resolveStreamChunkTimeoutMs({}, 9e4);
    for (let attempt = 0; ; attempt += 1) {
      try {
        const response = await this._fetchWithTimeout(
          url,
          { cache: "no-store" },
          fetchTimeoutMs
        );
        if (!response.ok) {
          throw new Error(
            `Failed to fetch decision head: ${response.status} ${response.statusText}`
          );
        }
        await writeResponseToFsFileWithProgress(
          response,
          this._core.FS,
          headPath,
          typeof onProgress === "function" ? onProgress : null,
          {
            useBigIntPosition: this._coreVariant === "wasm64",
            chunkTimeoutMs
          }
        );
        return;
      } catch (error) {
        this._unlinkDecisionFile(headPath);
        if (!isRetryableStreamNetworkError(error) || attempt >= 1) {
          throw error;
        }
        this._runtimeNotes.push(`decision_head_fetch_retry:${attempt + 1}`);
      }
    }
  }
  async loadDecisionHead(source, options = {}) {
    if (!this._core || this._modelBytes <= 0) {
      throw new Error("No model loaded. Call loadModelFromUrl first.");
    }
    const configJson = options?.configJson ?? null;
    if (configJson !== null && typeof configJson !== "string") {
      throw new TypeError("Decision head configJson must be a string.");
    }
    let bytes = null;
    if (typeof source === "string") {
      if (source.length === 0) {
        throw new Error("Decision head URL is empty.");
      }
    } else {
      bytes = bufferSourceBytes(source);
      if (!bytes) {
        throw new TypeError("Decision head source must be a URL string, an ArrayBuffer or a typed array.");
      }
      if (bytes.byteLength === 0) {
        throw new Error("Decision head bytes are empty.");
      }
    }
    const capabilities = this.getDecisionCapabilities();
    if (capabilities.supported !== true) {
      throw new Error(
        capabilities.reason || "The loaded model does not support decision heads."
      );
    }
    const core = this._core;
    this._ensureDecisionDir();
    const taskId = `${Date.now()}_${++this._decisionFileCounter}`;
    const headPath = `/decision/head_${taskId}.safetensors`;
    const configPath = configJson === null ? null : `/decision/config_${taskId}.json`;
    try {
      let label = "decision head bytes";
      if (bytes) {
        core.FS.writeFile(headPath, bytes);
      } else {
        label = basenameFromUrl(source).slice(0, 256);
        await this._stageDecisionHeadFromUrl(source, headPath, options?.onProgress);
      }
      if (configPath !== null) {
        core.FS.writeFile(configPath, textEncoder.encode(configJson));
      }
      const handle = Number(
        await core.ccall(
          "llamadart_webgpu_decision_load",
          "number",
          ["string", "string", "string"],
          [headPath, label, configPath],
          { async: true }
        )
      );
      if (handle <= 0) {
        throw new Error(this._coreErrorMessage("Failed to load decision head", handle));
      }
      const raw = core.ccall(
        "llamadart_webgpu_decision_head_info_json",
        "string",
        ["number"],
        [handle]
      ) || "{}";
      return JSON.parse(raw);
    } finally {
      this._unlinkDecisionFile(headPath);
      this._unlinkDecisionFile(configPath);
    }
  }
  async runDecision(handle, sequences) {
    const nativeHandle = decisionHandleFrom(handle);
    const input = encodeDecisionSequences(sequences);
    if (!this._core || this._modelBytes <= 0) {
      throw new Error("No model loaded. Call loadModelFromUrl first.");
    }
    const core = this._core;
    this._ensureDecisionDir();
    const taskId = `${Date.now()}_${++this._decisionFileCounter}`;
    const inputPath = `/decision/input_${taskId}.bin`;
    const outputPath = `/decision/output_${taskId}.bin`;
    try {
      core.FS.writeFile(inputPath, input);
      const count = Number(
        await core.ccall(
          "llamadart_webgpu_decision_run",
          "number",
          ["number", "string", "string"],
          [nativeHandle, inputPath, outputPath],
          { async: true }
        )
      );
      if (count < 0) {
        throw new Error(this._coreErrorMessage("Decision run failed", count));
      }
      return decodeDecisionOutputs(core.FS.readFile(outputPath), count);
    } finally {
      this._unlinkDecisionFile(inputPath);
      this._unlinkDecisionFile(outputPath);
    }
  }
  async freeDecisionHead(handle) {
    const nativeHandle = decisionHandleFrom(handle);
    if (!this._core) {
      return;
    }
    await this._core.ccall(
      "llamadart_webgpu_decision_free",
      "number",
      ["number"],
      [nativeHandle],
      { async: true }
    );
  }
  getLoraAdapterCapabilities() {
    const unsupported = (reason) => ({
      apiVersion: LORA_API_VERSION,
      supported: false,
      reason
    });
    const core = this._core;
    if (!core) {
      return unsupported("WebGPU core is not initialized");
    }
    if (typeof core._llamadart_webgpu_lora_api_version !== "function") {
      return unsupported("This WebGPU core build does not include LoRA adapters.");
    }
    const coreVersion = Number(
      core.ccall("llamadart_webgpu_lora_api_version", "number", [], [])
    );
    if (coreVersion !== LORA_API_VERSION) {
      return unsupported(
        `The WebGPU core implements LoRA API version ${coreVersion}; this bridge needs version ${LORA_API_VERSION}.`
      );
    }
    return { apiVersion: LORA_API_VERSION, supported: true };
  }
  _requireLoraCore() {
    if (!this._core || this._modelBytes <= 0) {
      throw new Error("No model loaded. Call loadModelFromUrl first.");
    }
    const capabilities = this.getLoraAdapterCapabilities();
    if (capabilities.supported !== true) {
      throw new Error(capabilities.reason || "This bridge cannot load LoRA adapters.");
    }
    return this._core;
  }
  /**
   * Downloads a URL adapter into `adapterPath`, through the Cache API unless
   * `options.useCache` is false. A cache miss stores the response while it is
   * written, so progress starts with the first chunk.
   */
  async _stageLoraAdapterFromUrl(url, adapterPath, options) {
    const core = this._core;
    const signal = options.signal || null;
    const progressCallback = typeof options.progressCallback === "function" ? options.progressCallback : null;
    const fetchTimeoutMs = this._resolveFetchTimeoutMs(options, 18e4);
    const chunkTimeoutMs = this._resolveStreamChunkTimeoutMs(options, 9e4);
    const fetchInit = signal ? { signal } : {};
    const cacheKey = normalizeAbsoluteUrl(url);
    let cache = null;
    if (options.useCache !== false && typeof globalThis.caches?.open === "function") {
      try {
        cache = await globalThis.caches.open(this._resolveCacheName(options));
      } catch (_) {
        this._runtimeNotes.push("lora_cache_error");
      }
    }
    for (let attempt = 0; ; attempt += 1) {
      throwIfAborted(signal, LORA_LOAD_ABORT_MESSAGE);
      try {
        let response = null;
        if (cache) {
          try {
            response = await cache.match(cacheKey) || null;
          } catch (_) {
            this._runtimeNotes.push("lora_cache_error");
          }
        }
        let stored = null;
        if (!response) {
          response = await this._fetchWithTimeout(
            url,
            cache ? fetchInit : { cache: "no-store", ...fetchInit },
            fetchTimeoutMs
          );
          if (!response.ok) {
            throw new Error(
              `Failed to fetch LoRA adapter: ${response.status} ${response.statusText}`
            );
          }
          if (cache) {
            stored = cache.put(cacheKey, response.clone()).then(
              () => {
                this._runtimeNotes.push("lora_cache_stored");
              },
              () => {
                this._runtimeNotes.push("lora_cache_store_failed");
              }
            );
          }
        }
        await writeResponseToFsFileWithProgress(
          response,
          core.FS,
          adapterPath,
          progressCallback,
          {
            useBigIntPosition: this._coreVariant === "wasm64",
            chunkTimeoutMs,
            signal,
            abortMessage: LORA_LOAD_ABORT_MESSAGE
          }
        );
        await stored;
        return;
      } catch (error) {
        this._deleteFsFile(adapterPath);
        throwIfAborted(signal, LORA_LOAD_ABORT_MESSAGE);
        if (!isRetryableStreamNetworkError(error) || attempt >= 1) {
          throw error;
        }
        this._runtimeNotes.push(`lora_fetch_retry:${attempt + 1}`);
      }
    }
  }
  async loadLoraAdapter(source, options = {}) {
    let bytes = null;
    if (typeof source === "string") {
      if (source.length === 0) {
        throw new Error("LoRA adapter URL is empty.");
      }
    } else {
      bytes = bufferSourceBytes(source);
      if (!bytes) {
        throw new TypeError("LoRA adapter source must be a URL string, an ArrayBuffer or a typed array.");
      }
      if (bytes.byteLength === 0) {
        throw new Error("LoRA adapter bytes are empty.");
      }
    }
    const core = this._requireLoraCore();
    const signal = options.signal || null;
    throwIfAborted(signal, LORA_LOAD_ABORT_MESSAGE);
    ensureFsDirectory(core.FS, "/lora");
    const adapterPath = `/lora/adapter_${++this._loraFileCounter}.gguf`;
    try {
      if (bytes) {
        core.FS.writeFile(adapterPath, bytes);
      } else {
        await this._stageLoraAdapterFromUrl(source, adapterPath, options);
      }
      throwIfAborted(signal, LORA_LOAD_ABORT_MESSAGE);
      const handle = Number(
        await core.ccall(
          "llamadart_webgpu_lora_load",
          "number",
          ["string"],
          [adapterPath],
          { async: true }
        )
      );
      if (handle <= 0) {
        throw new Error(this._coreErrorMessage("Failed to load LoRA adapter", handle));
      }
      this._loraAdaptersLoaded = true;
      return { handle };
    } finally {
      this._deleteFsFile(adapterPath);
    }
  }
  async _callLoraCore(name, argTypes, args, failure) {
    const core = this._requireLoraCore();
    const rc = Number(await core.ccall(name, "number", argTypes, args, { async: true }));
    if (rc !== 0) {
      throw new Error(this._coreErrorMessage(failure, rc));
    }
  }
  async setLoraAdapter(handle, scale = 1) {
    await this._callLoraCore(
      "llamadart_webgpu_lora_set",
      ["number", "number"],
      [loraHandleFrom(handle), loraScaleFrom(scale)],
      "Failed to apply LoRA adapter"
    );
  }
  async removeLoraAdapter(handle) {
    await this._callLoraCore(
      "llamadart_webgpu_lora_remove",
      ["number"],
      [loraHandleFrom(handle)],
      "Failed to remove LoRA adapter"
    );
  }
  async clearLoraAdapters() {
    await this._callLoraCore(
      "llamadart_webgpu_lora_clear",
      [],
      [],
      "Failed to clear LoRA adapters"
    );
  }
  _clearStagedMediaFiles() {
    if (!this._core || this._stagedMediaPaths.length === 0) {
      this._stagedMediaPaths = [];
      return;
    }
    for (const mediaPath of this._stagedMediaPaths) {
      try {
        this._core.FS.unlink(mediaPath);
      } catch (_) {
      }
    }
    this._stagedMediaPaths = [];
  }
  _clearPendingMedia() {
    this._core?.ccall("llamadart_webgpu_media_clear_pending", null, [], []);
    this._clearStagedMediaFiles();
  }
  _persistMediaBytes(bytes, extension = ".bin") {
    if (!this._core) {
      throw new Error("WebGPU core is not initialized.");
    }
    try {
      this._core.FS.mkdir("/media");
    } catch (_) {
    }
    this._mediaFileCounter += 1;
    const suffix = typeof extension === "string" && extension.startsWith(".") ? extension : ".bin";
    const mediaPath = `/media/input_${Date.now()}_${this._mediaFileCounter}${suffix}`;
    this._core.FS.writeFile(mediaPath, bytes);
    this._stagedMediaPaths.push(mediaPath);
    return mediaPath;
  }
  _addMediaFile(mediaPath) {
    const rc = Number(
      this._core.ccall(
        "llamadart_webgpu_media_add_file",
        "number",
        ["string"],
        [mediaPath]
      )
    );
    if (rc !== 0) {
      throw new Error(this._coreErrorMessage("Failed to add media file", rc));
    }
  }
  _addRawRgbMediaBytes(bytes, width, height) {
    const useHeapBuffer = this._core && typeof this._core._malloc === "function" && typeof this._core._free === "function" && this._core.HEAPU8 && typeof this._core.HEAPU8.set === "function";
    let rc = 0;
    if (useHeapBuffer) {
      const ptr = this._core._malloc(bytes.length);
      if (!Number.isFinite(ptr) || ptr <= 0) {
        throw new Error("Failed to allocate core heap buffer for RGB media bytes");
      }
      try {
        this._core.HEAPU8.set(bytes, ptr);
        rc = Number(
          this._core.ccall(
            "llamadart_webgpu_media_add_rgb",
            "number",
            ["number", "number", "number", "number"],
            [width, height, ptr, bytes.length]
          )
        );
      } finally {
        this._core._free(ptr);
      }
    } else {
      rc = Number(
        this._core.ccall(
          "llamadart_webgpu_media_add_rgb",
          "number",
          ["number", "number", "array", "number"],
          [width, height, bytes, bytes.length]
        )
      );
    }
    if (rc !== 0) {
      throw new Error(this._coreErrorMessage("Failed to add raw RGB media bytes", rc));
    }
  }
  _addAudioSamples(samples) {
    const sampleBytes = new Uint8Array(samples.buffer, samples.byteOffset, samples.byteLength);
    const useHeapBuffer = this._core && typeof this._core._malloc === "function" && typeof this._core._free === "function" && this._core.HEAPU8 && typeof this._core.HEAPU8.set === "function";
    let rc = 0;
    if (useHeapBuffer) {
      const ptr = this._core._malloc(sampleBytes.length);
      if (!Number.isFinite(ptr) || ptr <= 0) {
        throw new Error("Failed to allocate core heap buffer for audio samples");
      }
      try {
        this._core.HEAPU8.set(sampleBytes, ptr);
        rc = Number(
          this._core.ccall(
            "llamadart_webgpu_media_add_audio_f32",
            "number",
            ["number", "number"],
            [ptr, samples.length]
          )
        );
      } finally {
        this._core._free(ptr);
      }
    } else {
      rc = Number(
        this._core.ccall(
          "llamadart_webgpu_media_add_audio_f32",
          "number",
          ["array", "number"],
          [sampleBytes, samples.length]
        )
      );
    }
    if (rc !== 0) {
      throw new Error(this._coreErrorMessage("Failed to add audio samples", rc));
    }
  }
  async _fetchMediaBytes(url) {
    const response = await this._fetchWithTimeout(
      url,
      { cache: "no-store" },
      this._resolveFetchTimeoutMs({}, 12e4)
    );
    if (!response.ok) {
      throw new Error(`Failed to fetch media: ${response.status} ${response.statusText}`);
    }
    return new Uint8Array(await response.arrayBuffer());
  }
  async _prepareImageBytesForMultimodal(bytes, options = {}) {
    const maxPixels = this._resolveMediaImageMaxPixels(options);
    const maxEdge = this._resolveMediaImageMaxEdge(options);
    if (maxPixels <= 0 && maxEdge <= 0) {
      return null;
    }
    return decodeImageBytesToRgb(bytes, {
      maxPixels,
      maxEdge
    });
  }
  async _stageMultimodalParts(parts, options = {}) {
    this._clearPendingMedia();
    const mediaParts = Array.isArray(parts) ? parts : [];
    if (mediaParts.length === 0) {
      return;
    }
    if (!this._mmProjPath) {
      throw new Error(
        "Multimodal input requires a loaded projector. Call loadMultimodalProjector first."
      );
    }
    for (const rawPart of mediaParts) {
      const part = rawPart && typeof rawPart === "object" ? rawPart : {};
      const type = String(part.type || "").toLowerCase();
      if (type === "image") {
        const bytes = toUint8Array(part.bytes);
        if (bytes && bytes.length > 0) {
          const width = Number(part.width);
          const height = Number(part.height);
          const isRawRgb = Number.isInteger(width) && Number.isInteger(height) && width > 0 && height > 0 && bytes.length === width * height * 3;
          if (isRawRgb) {
            this._addRawRgbMediaBytes(bytes, width, height);
          } else {
            const prepared2 = await this._prepareImageBytesForMultimodal(bytes, options);
            if (prepared2 && prepared2.bytes && prepared2.bytes.length > 0) {
              const mediaPath = this._persistMediaBytes(prepared2.bytes, ".img");
              this._addMediaFile(mediaPath);
              if (prepared2.resized) {
                this._runtimeNotes.push(
                  `media_image_resized:${prepared2.sourceWidth}x${prepared2.sourceHeight}->${prepared2.width}x${prepared2.height}`
                );
              }
            } else {
              const mediaPath = this._persistMediaBytes(bytes, ".img");
              this._addMediaFile(mediaPath);
            }
          }
          continue;
        }
        if (typeof part.url !== "string" || part.url.length === 0) {
          throw new Error("Image part must provide bytes or url.");
        }
        const fetched = await this._fetchMediaBytes(part.url);
        const prepared = await this._prepareImageBytesForMultimodal(fetched, options);
        if (prepared && prepared.bytes && prepared.bytes.length > 0) {
          const mediaPath = this._persistMediaBytes(prepared.bytes, ".img");
          this._addMediaFile(mediaPath);
          if (prepared.resized) {
            this._runtimeNotes.push(
              `media_image_resized:${prepared.sourceWidth}x${prepared.sourceHeight}->${prepared.width}x${prepared.height}`
            );
          }
        } else {
          const mediaPath = this._persistMediaBytes(fetched, ".img");
          this._addMediaFile(mediaPath);
        }
        continue;
      }
      if (type === "audio") {
        const samples = toFloat32Array(part.samples);
        if (samples && samples.length > 0) {
          this._addAudioSamples(samples);
          continue;
        }
        const bytes = toUint8Array(part.bytes);
        if (bytes && bytes.length > 0) {
          const mediaPath2 = this._persistMediaBytes(bytes, ".aud");
          this._addMediaFile(mediaPath2);
          continue;
        }
        if (typeof part.url !== "string" || part.url.length === 0) {
          throw new Error("Audio part must provide samples, bytes, or url.");
        }
        const fetched = await this._fetchMediaBytes(part.url);
        const mediaPath = this._persistMediaBytes(fetched, ".aud");
        this._addMediaFile(mediaPath);
      }
    }
  }
  getCompletionCapabilities() {
    const core = this._core;
    if (!core || typeof core._llamadart_webgpu_completion_capabilities_json !== "function") {
      return { ...NO_COMPLETION_CAPABILITIES };
    }
    return completionCapabilitiesFrom(
      core.ccall("llamadart_webgpu_completion_capabilities_json", "string", [], [])
    );
  }
  async createCompletion(prompt, options = {}) {
    if (this._modelBytes <= 0) {
      throw new Error("No model loaded. Call loadModelFromUrl first.");
    }
    const sampling = resolveCompletionSamplingOptions(options);
    requireCompletionCapabilities(sampling, this.getCompletionCapabilities());
    this._abortRequested = false;
    const startedAt = performance.now();
    let firstTokenAt = null;
    let nPredict = Number(options.nPredict) > 0 ? Number(options.nPredict) : 256;
    const hasMediaParts = Array.isArray(options.parts) && options.parts.length > 0;
    if (hasMediaParts) {
      const requestedMediaCap = Number(options.mediaMaxPredict);
      const mediaCap = Number.isFinite(requestedMediaCap) && requestedMediaCap > 0 ? Math.max(32, Math.trunc(requestedMediaCap)) : 256;
      if (nPredict > mediaCap) {
        nPredict = mediaCap;
        this._runtimeNotes.push(`media_n_predict_capped:${mediaCap}`);
      }
    }
    const temp = Number.isFinite(options.temp) ? Number(options.temp) : 0.8;
    const topK = Number.isFinite(options.topK) ? Number(options.topK) : 40;
    const topP = Number.isFinite(options.topP) ? Number(options.topP) : 0.95;
    const penalty = Number.isFinite(options.penalty) ? Number(options.penalty) : 1.1;
    const grammar = typeof options.grammar === "string" && options.grammar.length > 0 ? options.grammar : null;
    const seed = Number.isInteger(options.seed) ? Number(options.seed) : Math.floor(Math.random() * 4294967295);
    await this._stageMultimodalParts(options.parts, options);
    let generationStarted = false;
    const speculativePaths = [];
    try {
      if (sampling.speculative) {
        await this._setNextSpeculative(sampling.speculative, nPredict, speculativePaths);
      }
      const beginRc = Number(
        await this._core.ccall(
          "llamadart_webgpu_begin_generation",
          "number",
          [
            "string",
            "number",
            "number",
            "number",
            "number",
            "string",
            "number",
            "number",
            "number",
            "number",
            "string",
            "string",
            "string"
          ],
          [
            String(prompt),
            temp,
            topK,
            topP,
            penalty,
            grammar,
            seed >>> 0,
            sampling.minP,
            sampling.presencePenalty,
            sampling.thinkingBudget?.maxTokens ?? 0,
            sampling.thinkingBudget?.startTag ?? null,
            sampling.thinkingBudget?.endTag ?? null,
            sampling.thinkingBudget?.forcedMessage ?? null
          ],
          { async: true }
        )
      );
      if (beginRc === GENERATION_ALREADY_ACTIVE_RC) {
        throw new Error(GENERATION_ALREADY_ACTIVE_MESSAGE);
      }
      if (beginRc !== 0) {
        throw new Error(this._coreErrorMessage("Failed to start generation", beginRc));
      }
      generationStarted = true;
      for (const path of speculativePaths.splice(0)) {
        this._deleteFsFile(path);
      }
      let generated = 0;
      const shouldEmitCurrentText = options.emitCurrentTextOnToken !== false;
      const tokenEventEncoding = typeof options.tokenEventEncoding === "string" ? String(options.tokenEventEncoding || "").toLowerCase() : "bytes";
      const emitTokenText = tokenEventEncoding === "text";
      const yieldInterval = 4;
      let streamed = "";
      let emittedStableText = "";
      let finishReason = "length";
      while (generated < nPredict) {
        if (this._abortRequested || options.signal?.aborted) {
          finishReason = "cancelled";
          break;
        }
        const stepRc = Number(
          await this._core.ccall(
            "llamadart_webgpu_next_token",
            "number",
            [],
            [],
            { async: true }
          )
        );
        if (stepRc === 0) {
          finishReason = this._abortRequested || options.signal?.aborted ? "cancelled" : "stop";
          break;
        }
        if (stepRc < 0) {
          const stepErrorText = this._coreErrorMessage("Generation step failed", stepRc);
          if (generated > 0 && this._isContextLimitGenerationError(stepErrorText)) {
            this._runtimeNotes.push("generation_stopped_context_limit");
            this._emitLogger(
              "warn",
              "warning: generation reached context/memory limit; returning partial output."
            );
            break;
          }
          if (sampling.speculative == null && this._shouldAttemptGenerationRecovery(stepErrorText, options, generated)) {
            const recovered = await this._recoverGenerationWithCpuFallback(options);
            if (recovered) {
              if (generationStarted) {
                try {
                  this._core.ccall("llamadart_webgpu_end_generation", null, [], []);
                } catch (_) {
                }
                generationStarted = false;
              }
              this._clearPendingMedia();
              const retryOptions = {
                ...options,
                _llamadartGenerationRecoveryAttempted: true
              };
              return await this.createCompletion(prompt, retryOptions);
            }
          }
          throw new Error(stepErrorText);
        }
        generated += 1;
        const fullText = this._core.ccall("llamadart_webgpu_last_output", "string", [], []) || "";
        streamed = fullText;
        const stableText = trimUnstableUtf8Tail(fullText);
        if (!stableText.startsWith(emittedStableText)) {
          emittedStableText = "";
        }
        const deltaText = stableText.slice(emittedStableText.length);
        if (deltaText.length === 0) {
          continue;
        }
        emittedStableText = stableText;
        firstTokenAt ??= performance.now();
        if (typeof options.onToken === "function") {
          const piecePayload = emitTokenText ? deltaText : textEncoder.encode(deltaText);
          options.onToken(piecePayload, shouldEmitCurrentText ? fullText : null);
        }
        if (yieldInterval > 0 && generated % yieldInterval === 0) {
          await new Promise((resolve) => setTimeout(resolve, 0));
        }
      }
      const text = this._core.ccall("llamadart_webgpu_last_output", "string", [], []) || streamed || "";
      const tailText = text.startsWith(emittedStableText) ? text.slice(emittedStableText.length) : "";
      if (tailText.length > 0) {
        firstTokenAt ??= performance.now();
        if (typeof options.onToken === "function") {
          const piecePayload = emitTokenText ? tailText : textEncoder.encode(tailText);
          options.onToken(piecePayload, shouldEmitCurrentText ? text : null);
        }
      }
      if (typeof options.onUsage === "function") {
        const counts = JSON.parse(
          this._core.ccall("llamadart_webgpu_last_generation_usage_json", "string", [], []) || "{}"
        );
        const speculative = counts?.speculative;
        options.onUsage({
          promptTokens: Number(counts?.promptTokens) || 0,
          cachedPromptTokens: Number(counts?.cachedPromptTokens) || 0,
          completionTokens: Number(counts?.completionTokens) || 0,
          timeToFirstTokenMs: firstTokenAt == null ? null : firstTokenAt - startedAt,
          durationMs: performance.now() - startedAt,
          finishReason,
          ...speculative && typeof speculative === "object" ? {
            speculative: {
              draftTokens: Number(speculative.draftTokens) || 0,
              acceptedDraftTokens: Number(speculative.acceptedDraftTokens) || 0,
              draftAttempts: Number(speculative.draftAttempts) || 0,
              verifyTokens: Number(speculative.verifyTokens) || 0,
              replayTokens: Number(speculative.replayTokens) || 0
            }
          } : {}
        });
      }
      return text;
    } finally {
      if (generationStarted) {
        await this._core.ccall("llamadart_webgpu_end_generation", null, [], [], { async: true });
      }
      for (const path of speculativePaths) {
        this._deleteFsFile(path);
      }
      this._clearPendingMedia();
    }
  }
  async tokenize(text, _addSpecial = true) {
    if (this._modelBytes <= 0) {
      throw new Error("No model loaded. Call loadModelFromUrl first.");
    }
    const rc = Number(
      await this._core.ccall(
        "llamadart_webgpu_tokenize_to_json",
        "number",
        ["string", "number"],
        [String(text), _addSpecial ? 1 : 0],
        { async: true }
      )
    );
    if (rc < 0) {
      throw new Error(this._coreErrorMessage("Tokenization failed", rc));
    }
    const raw = this._core.ccall("llamadart_webgpu_last_tokens_json", "string", [], []) || "[]";
    const parsed = JSON.parse(raw);
    return Array.isArray(parsed) ? parsed.map((v) => Number(v) | 0) : [];
  }
  _ensureStateDir() {
    const core = this._core;
    if (!core?.FS) {
      throw new Error("Bridge filesystem is not initialized");
    }
    ensureFsDirectory(core.FS, "/states");
  }
  _normalizeStateTokens(tokens) {
    const normalized = Array.isArray(tokens) ? tokens : Array.from(tokens || []);
    return normalized.map((value) => Number(value) | 0);
  }
  _nextStateTempPath() {
    this._ensureStateDir();
    this._stateFileCounter += 1;
    return `/states/state_${Date.now()}_${this._stateFileCounter}.bin`;
  }
  async stateSaveFile(path, tokens = []) {
    if (this._modelBytes <= 0) {
      throw new Error("No model loaded. Call loadModelFromUrl first.");
    }
    if (typeof path !== "string" || path.length === 0) {
      throw new Error("State file path is empty.");
    }
    const normalized = this._normalizeStateTokens(tokens);
    const tokenText = JSON.stringify(normalized);
    const rc = Number(
      await this._core.ccall(
        "llamadart_webgpu_state_save_file",
        "number",
        ["string", "string"],
        [path, tokenText],
        { async: true }
      )
    );
    if (rc < 0) {
      throw new Error(this._coreErrorMessage("State save failed", rc));
    }
    return true;
  }
  async stateLoadFile(path, tokenCapacity = this.getContextSize()) {
    if (this._modelBytes <= 0) {
      throw new Error("No model loaded. Call loadModelFromUrl first.");
    }
    if (typeof path !== "string" || path.length === 0) {
      throw new Error("State file path is empty.");
    }
    const capacity = Number(tokenCapacity) > 0 ? Math.trunc(Number(tokenCapacity)) : this.getContextSize();
    const rc = Number(
      await this._core.ccall(
        "llamadart_webgpu_state_load_file",
        "number",
        ["string", "number"],
        [path, capacity],
        { async: true }
      )
    );
    if (rc < 0) {
      throw new Error(this._coreErrorMessage("State load failed", rc));
    }
    const raw = this._core.ccall("llamadart_webgpu_last_tokens_json", "string", [], []) || "[]";
    const parsed = JSON.parse(raw);
    const restoredTokens = Array.isArray(parsed) ? parsed.map((v) => Number(v) | 0) : [];
    return { tokens: restoredTokens };
  }
  async stateSaveBytes(tokens = []) {
    if (this._modelBytes <= 0) {
      throw new Error("No model loaded. Call loadModelFromUrl first.");
    }
    const tempPath = this._nextStateTempPath();
    try {
      await this.stateSaveFile(tempPath, tokens);
      const bytes = this._core.FS.readFile(tempPath);
      return bytes instanceof Uint8Array ? bytes : new Uint8Array(bytes || []);
    } finally {
      this._deleteFsFile(tempPath);
    }
  }
  async stateLoadBytes(bytes, tokenCapacity = this.getContextSize()) {
    if (this._modelBytes <= 0) {
      throw new Error("No model loaded. Call loadModelFromUrl first.");
    }
    const normalizedBytes = toUint8Array(bytes);
    if (!normalizedBytes || normalizedBytes.length === 0) {
      throw new Error("State bytes are empty.");
    }
    const tempPath = this._nextStateTempPath();
    try {
      this._core.FS.writeFile(tempPath, normalizedBytes);
      return await this.stateLoadFile(tempPath, tokenCapacity);
    } finally {
      this._deleteFsFile(tempPath);
    }
  }
  async detokenize(tokens, _special = false) {
    if (this._modelBytes <= 0) {
      throw new Error("No model loaded. Call loadModelFromUrl first.");
    }
    const normalized = Array.isArray(tokens) ? tokens : Array.from(tokens || []);
    const tokenText = JSON.stringify(normalized.map((v) => Number(v) | 0));
    const rc = Number(
      await this._core.ccall(
        "llamadart_webgpu_detokenize_from_json",
        "number",
        ["string", "number"],
        [tokenText, _special ? 1 : 0],
        { async: true }
      )
    );
    if (rc < 0) {
      throw new Error(this._coreErrorMessage("Detokenization failed", rc));
    }
    return this._core.ccall("llamadart_webgpu_last_detokenized", "string", [], []) || "";
  }
  _embeddingAssetVersion() {
    return embeddingAssetVersion(this._coreVariant === "wasm64" ? this._config.coreModuleUrlMem64 : this._config.coreModuleUrl);
  }
  async getEmbeddingCapabilities() {
    if (this._modelBytes <= 0) return { ...EMPTY_EMBEDDING_CAPABILITIES };
    if (typeof this._core?._llamadart_webgpu_embedding_capabilities_json !== "function" || typeof this._core?._llamadart_webgpu_embed_parts_to_json !== "function") {
      throw missingEmbeddingCapability(this._embeddingAssetVersion());
    }
    return embeddingCapabilitiesFrom(
      this._core.ccall("llamadart_webgpu_embedding_capabilities_json", "string", [], []),
      this._embeddingAssetVersion()
    );
  }
  async _preflightEmbeddingInputs(inputs) {
    if (!inputs.some(Array.isArray)) return;
    const caps = await this.getEmbeddingCapabilities();
    for (const input of inputs) requireEmbeddingModalities(input, caps, !!this._mmProjPath);
  }
  async _embedParts(parts, options) {
    const prompt = parts.map((part) => part.type === "text" ? part.text : EMBEDDING_MEDIA_MARKER).join("");
    const promptBytes = textEncoder.encode(prompt);
    const media = parts.filter((part) => part.type !== "text");
    try {
      await this._stageMultimodalParts(media);
      if (this._abortRequested) throw createAbortError("Embedding was cancelled.");
      const rc = Number(await this._core.ccall(
        "llamadart_webgpu_embed_parts_to_json",
        "number",
        ["array", "number", "number"],
        [promptBytes, options?.normalize !== false ? 1 : 0, promptBytes.length],
        { async: true }
      ));
      if (this._abortRequested) throw createAbortError("Embedding was cancelled.");
      if (rc < 0) throw embeddingCoreError(rc, this._embeddingAssetVersion());
      const raw = this._core.ccall("llamadart_webgpu_last_embedding_json", "string", [], []) || "[]";
      const parsed = JSON.parse(raw);
      return Array.isArray(parsed) ? parsed.map((v) => Number.isFinite(Number(v)) ? Number(v) : 0) : [];
    } catch (error) {
      if (error instanceof WebAssembly.RuntimeError) throw error;
      if (error instanceof EmbeddingError || error?.name === "AbortError") throw error;
      throw new EmbeddingError("EMBEDDING_FAILED", "Embedding media preparation or execution failed.");
    } finally {
      this._clearPendingMedia();
    }
  }
  async embed(text, options = {}) {
    if (Array.isArray(text)) {
      const parts = snapshotEmbedInput(text);
      this._abortRequested = false;
      await this._preflightEmbeddingInputs([parts]);
      if (this._abortRequested) throw createAbortError("Embedding was cancelled.");
      return this._embedParts(parts, options);
    }
    if (this._modelBytes <= 0) {
      throw new Error("No model loaded. Call loadModelFromUrl first.");
    }
    const normalize = options?.normalize !== false;
    const rc = Number(
      await this._core.ccall(
        "llamadart_webgpu_embed_to_json",
        "number",
        ["string", "number"],
        [String(text), normalize ? 1 : 0],
        { async: true }
      )
    );
    if (rc < 0) {
      throw new Error(this._coreErrorMessage("Embedding generation failed", rc));
    }
    const raw = this._core.ccall("llamadart_webgpu_last_embedding_json", "string", [], []) || "[]";
    const parsed = JSON.parse(raw);
    return Array.isArray(parsed) ? parsed.map((v) => {
      const numeric = Number(v);
      return Number.isFinite(numeric) ? numeric : 0;
    }) : [];
  }
  async embedBatch(texts, options = {}) {
    const normalized = Array.isArray(texts) ? texts : Array.from(texts || []);
    if (normalized.length === 0) {
      return [];
    }
    const inputs = normalized.map((input) => typeof input === "string" ? input : snapshotEmbedInput(input));
    this._abortRequested = false;
    await this._preflightEmbeddingInputs(inputs);
    const normalize = options?.normalize !== false;
    const vectors = [];
    for (const text of inputs) {
      if (this._abortRequested) throw createAbortError("Batch embedding was cancelled.");
      vectors.push(await (typeof text === "string" ? this.embed(text, { normalize }) : this._embedParts(text, { normalize })));
    }
    return vectors;
  }
  async scoreNextToken(prompt, options = {}) {
    if (this._modelBytes <= 0) {
      throw new Error("No model loaded. Call loadModelFromUrl first.");
    }
    if (typeof prompt !== "string") {
      throw new TypeError("Next-token scoring prompt must be a string.");
    }
    const candidates = nextTokenCandidateIds(options?.candidates);
    const topK = options?.topK ?? 0;
    if (!isInt32(topK)) {
      throw new TypeError(`Next-token topK is ${String(topK)}; expected a 32-bit integer.`);
    }
    const reusePromptPrefix = options?.reusePromptPrefix !== false;
    const rc = Number(
      await this._core.ccall(
        "llamadart_webgpu_score_next_token_to_json",
        "number",
        ["string", "string", "number", "number"],
        [prompt, JSON.stringify(candidates), topK, reusePromptPrefix ? 1 : 0],
        { async: true }
      )
    );
    if (rc < 0) {
      throw new Error(this._coreErrorMessage("Next-token scoring failed", rc));
    }
    const raw = this._core.ccall("llamadart_webgpu_last_next_token_scores_json", "string", [], []) || "{}";
    const parsed = JSON.parse(raw);
    return {
      candidates: scoredTokensFrom(parsed?.candidates),
      top: scoredTokensFrom(parsed?.top),
      promptTokens: Number(parsed?.promptTokens) || 0
    };
  }
  getModelMetadata() {
    let modelMetadata = {};
    try {
      const raw = this._core?.ccall("llamadart_webgpu_model_meta_json", "string", [], []);
      if (raw) {
        const parsed = JSON.parse(raw);
        if (parsed && typeof parsed === "object") {
          modelMetadata = parsed;
        }
      }
    } catch (_) {
    }
    return {
      ...modelMetadata,
      "llamadart.webgpu.prototype": "1",
      "llamadart.webgpu.backends": this._backendLabels.join(","),
      "llamadart.webgpu.model_bytes": String(this._modelBytes),
      "llamadart.webgpu.n_threads": String(this._threads),
      "llamadart.webgpu.n_threads_batch": String(this._threadsBatch),
      "llamadart.webgpu.n_batch": this._nBatch > 0 ? String(this._nBatch) : "",
      "llamadart.webgpu.n_ubatch": this._nUbatch > 0 ? String(this._nUbatch) : "",
      "llamadart.webgpu.n_seq_max": this._nSeqMax > 0 ? String(this._nSeqMax) : "",
      "llamadart.webgpu.flash_attention": String(this._flashAttention),
      "llamadart.webgpu.cache_type_k": String(this._cacheTypeK),
      "llamadart.webgpu.cache_type_v": String(this._cacheTypeV),
      "llamadart.webgpu.kv_unified": this._kvUnified >= 0 ? String(this._kvUnified) : "",
      "llamadart.webgpu.rope_freq_base": this._ropeFrequencyBase > 0 ? String(this._ropeFrequencyBase) : "",
      "llamadart.webgpu.rope_freq_scale": this._ropeFrequencyScale > 0 ? String(this._ropeFrequencyScale) : "",
      "llamadart.webgpu.split_mode": this._splitMode >= 0 ? String(this._splitMode) : "",
      "llamadart.webgpu.main_gpu": this._mainGpu >= 0 ? String(this._mainGpu) : "",
      "llamadart.webgpu.thread_pool_size": Number.isFinite(this._threadPoolSizeHint) && this._threadPoolSizeHint > 0 ? String(this._threadPoolSizeHint) : "",
      "llamadart.webgpu.n_gpu_layers": String(this._nGpuLayers),
      "llamadart.webgpu.core_variant": this._coreVariant,
      "llamadart.webgpu.model_source": this._modelSource,
      "llamadart.webgpu.model_cache_state": this._modelCacheState,
      "llamadart.webgpu.model_cache_name": this._modelCacheName,
      "llamadart.webgpu.runtime_notes": this._runtimeNotes.join(";"),
      "llamadart.webgpu.mmproj_loaded": this._mmProjPath ? "1" : "0",
      "llamadart.webgpu.supports_vision": this._mmSupportsVision ? "1" : "0",
      "llamadart.webgpu.supports_audio": this._mmSupportsAudio ? "1" : "0"
    };
  }
  getContextSize() {
    try {
      const nctx = Number(this._core?.ccall("llamadart_webgpu_get_context_size", "number", [], []));
      if (nctx > 0) {
        return nctx;
      }
    } catch (_) {
    }
    return this._nCtx;
  }
  isGpuActive() {
    return this._gpuActive;
  }
  getBackendName() {
    if (this._nGpuLayers === 0) {
      return "WASM (Prototype bridge)";
    }
    if (this._backendLabels.length > 0) {
      return this._backendLabels.join(", ");
    }
    return this._gpuActive ? "WebGPU (Prototype bridge)" : "WASM (Prototype bridge)";
  }
  setLogLevel(level) {
    if (Number.isFinite(level)) {
      this._logLevel = Math.max(0, Math.min(4, Math.trunc(level)));
    }
    this._applyCoreLogLevel();
  }
  cancel() {
    this._abortRequested = true;
    if (this._activeTransferAbortController) {
      try {
        this._activeTransferAbortController.abort();
      } catch (_) {
      }
      this._activeTransferAbortController = null;
    }
    try {
      this._core?.ccall("llamadart_webgpu_request_cancel", null, [], []);
    } catch (_) {
    }
  }
  async dispose() {
    if (this._textToSpeechActive && this._textToSpeechDone) {
      this.cancel();
      await this._textToSpeechDone;
    }
    if (this._activeTransferAbortController) {
      try {
        this._activeTransferAbortController.abort();
      } catch (_) {
      }
      this._activeTransferAbortController = null;
    }
    if (this._core) {
      this._clearPendingMedia();
      const mmprojPath = this._mmProjPath;
      const mmprojFreeRc = Number(
        this._core.ccall("llamadart_webgpu_mmproj_free", "number", [], [])
      );
      this._core.ccall("llamadart_webgpu_shutdown", null, [], []);
      this._deleteFsFile(mmprojPath);
      if (mmprojFreeRc !== 0) {
        this._emitLogger(
          "warn",
          this._coreErrorMessage("Failed to unload multimodal projector during dispose", mmprojFreeRc)
        );
      }
    }
    this._modelPath = null;
    this._modelPaths = [];
    this._modelBytes = 0;
    this._modelSource = "network";
    this._modelCacheState = "disabled";
    this._loadedModelUrl = null;
    this._mmProjPath = null;
    this._mmProjSourceUrl = null;
    this._mmSupportsVision = false;
    this._mmSupportsAudio = false;
    this._loraAdaptersLoaded = false;
    this._abortRequested = false;
    this._textToSpeechActive = false;
    this._textToSpeechDone = null;
    this._resolveTextToSpeechDone = null;
    this._activeTransferAbortController = null;
    this._suppressedWarmupWarningCount = 0;
    this._didReportWarmupWarningSuppression = false;
  }
  async applyChatTemplate(messages, addAssistant = true, _customTemplate = null) {
    return buildPromptFromMessages(messages, addAssistant);
  }
};

// js/src/worker_protocol.ts
function serializeWorkerError(error) {
  if (!error) {
    return "Unknown worker error";
  }
  if (typeof error === "string") {
    return error;
  }
  if (typeof error.message === "string" && error.message.length > 0) {
    return error.message;
  }
  try {
    return JSON.stringify(error);
  } catch (_) {
    return String(error);
  }
}
var bridgeWorkerModeParam = "__llamadartBridgeWorker";
function emptyBridgeState() {
  return {
    metadata: {},
    contextSize: 0,
    gpuActive: false,
    backendName: "WASM (Prototype bridge)",
    supportsVision: false,
    supportsAudio: false
  };
}
function snapshotBridgeState(target) {
  return {
    metadata: target.getModelMetadata(),
    contextSize: target.getContextSize(),
    gpuActive: target.isGpuActive(),
    backendName: target.getBackendName(),
    supportsVision: target.supportsVision(),
    supportsAudio: target.supportsAudio()
  };
}

// js/src/worker_proxy.ts
function createBridgeWorkerSource(moduleUrl) {
  return `import * as workerModule from ${JSON.stringify(moduleUrl)};
if (workerModule && typeof workerModule.enableBridgeWorkerHost === 'function') { workerModule.enableBridgeWorkerHost(); }
`;
}
function resolveWorkerEntryUrl(moduleUrl) {
  if (typeof moduleUrl !== "string" || moduleUrl.length === 0) {
    return null;
  }
  try {
    const base = typeof window !== "undefined" && window.location?.href ? window.location.href : void 0;
    const url = new URL(moduleUrl, base);
    const path = url.pathname || "";
    const usesDedicatedWorkerEntry = path.endsWith("_worker.js");
    if (!usesDedicatedWorkerEntry) {
      url.searchParams.set(bridgeWorkerModeParam, "1");
    }
    return url.toString();
  } catch (_) {
    return null;
  }
}
function deriveBridgeModuleUrlFromWorkerEntry(moduleUrl) {
  if (typeof moduleUrl !== "string" || moduleUrl.length === 0) {
    return null;
  }
  try {
    const base = typeof window !== "undefined" && window.location?.href ? window.location.href : void 0;
    const url = new URL(moduleUrl, base);
    const path = url.pathname || "";
    if (!path.endsWith("_worker.js")) {
      return null;
    }
    url.pathname = path.replace(/_worker\.js$/, ".js");
    url.searchParams.delete(bridgeWorkerModeParam);
    return url.toString();
  } catch (_) {
    return null;
  }
}
var BridgeWorkerProxy = class {
  constructor({ moduleUrl, config }) {
    this._config = config && typeof config === "object" ? config : {};
    this._nextId = 1;
    this._pending = /* @__PURE__ */ new Map();
    this._workerBlobUrl = null;
    let workerInitError = null;
    const moduleCandidates = [moduleUrl];
    const bridgeModuleFallback = deriveBridgeModuleUrlFromWorkerEntry(moduleUrl);
    if (bridgeModuleFallback && bridgeModuleFallback !== moduleUrl) {
      moduleCandidates.push(bridgeModuleFallback);
    }
    for (const candidate of moduleCandidates) {
      if (this._worker) {
        break;
      }
      const directWorkerUrl = resolveWorkerEntryUrl(candidate);
      if (!directWorkerUrl) {
        continue;
      }
      try {
        this._worker = new Worker(directWorkerUrl, { type: "module" });
      } catch (error) {
        workerInitError = error;
      }
    }
    for (const candidate of moduleCandidates) {
      if (this._worker) {
        break;
      }
      const source = createBridgeWorkerSource(candidate);
      this._workerBlobUrl = URL.createObjectURL(
        new Blob([source], { type: "text/javascript" })
      );
      try {
        this._worker = new Worker(this._workerBlobUrl, { type: "module" });
      } catch (error) {
        workerInitError = error;
        URL.revokeObjectURL(this._workerBlobUrl);
        this._workerBlobUrl = null;
      }
    }
    if (!this._worker) {
      throw workerInitError || new Error("Failed to initialize bridge worker");
    }
    this._ready = new Promise((resolve, reject) => {
      this._readyResolve = resolve;
      this._readyReject = reject;
    });
    this._readyTimeoutHandle = null;
    this._armReadyTimeout();
    this._worker.onmessage = (event) => {
      const message = event.data || {};
      const type = message.type;
      if (type === "ready") {
        this._clearReadyTimeout();
        this._readyResolve?.();
        return;
      }
      const id = Number(message.id || 0);
      const pending = this._pending.get(id);
      if (!pending) {
        return;
      }
      if (type === "event") {
        pending.onEvent?.(message);
        return;
      }
      this._pending.delete(id);
      if (type === "error") {
        const errorMessage = String(message.message || "Worker request failed");
        const workerError = message.embeddingErrorCode ? new EmbeddingError(message.embeddingErrorCode, errorMessage) : new Error(errorMessage);
        if (message.state && typeof message.state === "object") {
          workerError.state = message.state;
        }
        pending.reject(workerError);
        return;
      }
      pending.resolve(message);
    };
    this._worker.onerror = (event) => {
      const message = event?.message || "Bridge worker crashed";
      const error = new Error(String(message));
      error.llamadartWorkerCrash = true;
      this._clearReadyTimeout();
      this._readyReject?.(error);
      for (const pending of this._pending.values()) {
        pending.reject(error);
      }
      this._pending.clear();
    };
    this._worker.postMessage({ type: "init", config });
  }
  async call(method, args, onEvent, transferList = [], operationMeta = null) {
    await this._ready;
    const id = this._nextId++;
    const timeoutMs = this._resolveRequestTimeoutMs(method, args);
    const transfers = Array.isArray(transferList) ? transferList.filter((item) => item != null) : [];
    return new Promise((resolve, reject) => {
      let timeoutHandle = null;
      const clearTimer = () => {
        if (timeoutHandle != null) {
          globalThis.clearTimeout(timeoutHandle);
          timeoutHandle = null;
        }
      };
      const armTimer = () => {
        clearTimer();
        if (!Number.isFinite(timeoutMs) || timeoutMs <= 0) {
          return;
        }
        timeoutHandle = globalThis.setTimeout(() => {
          this._pending.delete(id);
          reject(new Error(`Worker request timeout (${method}, ${timeoutMs}ms)`));
        }, timeoutMs);
      };
      armTimer();
      this._pending.set(id, {
        resolve: (value) => {
          clearTimer();
          resolve(value);
        },
        reject: (error) => {
          clearTimer();
          reject(error);
        },
        onEvent: (event) => {
          armTimer();
          onEvent?.(event);
        }
      });
      const message = { type: "call", id, method, args };
      if (operationMeta && typeof operationMeta === "object") {
        Object.assign(message, operationMeta);
      }
      this._worker.postMessage(message, transfers);
    });
  }
  _resolveWorkerReadyTimeoutMs() {
    const configured = Number(this._config.workerInitTimeoutMs);
    if (!Number.isFinite(configured) || configured <= 0) {
      return 2e4;
    }
    return Math.max(3e3, Math.min(12e4, Math.trunc(configured)));
  }
  _clearReadyTimeout() {
    if (this._readyTimeoutHandle != null) {
      globalThis.clearTimeout(this._readyTimeoutHandle);
      this._readyTimeoutHandle = null;
    }
  }
  _armReadyTimeout() {
    this._clearReadyTimeout();
    const timeoutMs = this._resolveWorkerReadyTimeoutMs();
    this._readyTimeoutHandle = globalThis.setTimeout(() => {
      this._readyTimeoutHandle = null;
      const timeoutError = new Error(`Bridge worker init timeout (${timeoutMs}ms)`);
      this._readyReject?.(timeoutError);
      try {
        this._worker?.terminate();
      } catch (_) {
      }
    }, timeoutMs);
  }
  _resolveRequestTimeoutMs(method, args = []) {
    const explicitGlobal = Number(this._config.workerRequestTimeoutMs);
    const clamp = (value, fallback) => {
      if (!Number.isFinite(value) || value <= 0) {
        return fallback;
      }
      return Math.max(5e3, Math.min(36e5, Math.trunc(value)));
    };
    if (method === "loadModelFromUrl" || method === "loadDraftModel") {
      return clamp(Number(this._config.workerModelLoadTimeoutMs), clamp(explicitGlobal, 3 * 60 * 1e3));
    }
    if (method === "loadMultimodalProjector") {
      return clamp(Number(this._config.workerMmprojLoadTimeoutMs), clamp(explicitGlobal, 8 * 60 * 1e3));
    }
    if (method === "createCompletion") {
      return clamp(Number(this._config.workerCompletionTimeoutMs), clamp(explicitGlobal, 6 * 60 * 1e3));
    }
    if (method === "synthesizeSpeech") {
      return clamp(Number(this._config.workerTextToSpeechTimeoutMs), clamp(explicitGlobal, 20 * 60 * 1e3));
    }
    if (method === "loadDecisionHead" || method === "loadLoraAdapter") {
      return clamp(explicitGlobal, 10 * 60 * 1e3);
    }
    if (method === "runDecision") {
      const sequences = Array.isArray(args?.[1]) ? args[1].length : 0;
      return clamp(explicitGlobal, 10 * 60 * 1e3 + sequences * DECISION_WORKER_TIMEOUT_PER_SEQUENCE_MS);
    }
    return clamp(explicitGlobal, 12e4);
  }
  async dispose() {
    this._clearReadyTimeout();
    let didTimeout = false;
    try {
      await Promise.race([
        this.call("dispose", []),
        new Promise((resolve) => {
          globalThis.setTimeout(() => {
            didTimeout = true;
            resolve(null);
          }, 800);
        })
      ]);
    } catch (_) {
    }
    if (didTimeout) {
    }
    for (const pending of this._pending.values()) {
      pending.reject(new Error("Bridge worker disposed"));
    }
    this._pending.clear();
    this._worker.terminate();
    if (this._workerBlobUrl) {
      URL.revokeObjectURL(this._workerBlobUrl);
      this._workerBlobUrl = null;
    }
  }
};

// js/src/bridge.ts
var LlamaWebGpuBridge = class {
  static supportsSafariAdaptiveGpu = LlamaWebGpuBridgeRuntime.supportsSafariAdaptiveGpu === true;
  static supportsCompletionUsage = true;
  constructor(config = {}) {
    this._config = config;
    this._runtime = null;
    this._workerProxy = null;
    this._workerGeneration = 0;
    this._workerDisposePromise = null;
    this._retiringWorkerDisposals = /* @__PURE__ */ new Set();
    this._retiredWorkerProxies = /* @__PURE__ */ new WeakSet();
    this._workerFallbackReason = null;
    this._metadata = {};
    this._contextSize = 0;
    this._gpuActive = false;
    this._backendName = "WASM (Prototype bridge)";
    this._supportsVision = false;
    this._supportsAudio = false;
    this._loadedModelUrl = null;
    this._loadedModelOptions = null;
    this._loadedMmProjUrl = null;
    this._loadedDraftModel = null;
    this._multimodalWorkerCpuMode = false;
    this._workerModelMissing = false;
    this._bridgeWarnRecent = /* @__PURE__ */ new Map();
    this._operationQueueTail = null;
    this._activeOperation = null;
    this._nextOperationId = 0;
    this._lifecycleState = "open";
    this._shadowStateTransactionDepth = 0;
    this._deferredShadowState = null;
    this._disposed = false;
    this._disposePromise = null;
    this._disposalWaiters = /* @__PURE__ */ new Set();
    this._decisionHeads = /* @__PURE__ */ new Map();
    this._nextDecisionHandle = 1;
    this._loraAdapters = /* @__PURE__ */ new Map();
    this._activeLoraScales = /* @__PURE__ */ new Map();
    this._nextLoraHandle = 1;
    if (this._shouldUseWorker()) {
      try {
        this._workerProxy = this._createWorkerProxy();
      } catch (error) {
        this._disableWorkerFallback(error);
      }
    }
    if (this._workerProxy) {
      this._workerGeneration = 1;
    }
    if (!this._workerProxy) {
      this._runtime = this._createRuntime();
      this._captureDirectRuntimeState();
    }
  }
  _createRuntime() {
    return new LlamaWebGpuBridgeRuntime({
      ...this._config,
      disableWorker: true
    });
  }
  _operationSignal(operation = this._activeOperation) {
    return operation?.abortController?.signal || null;
  }
  _recordWorkerGeneration(operation = this._activeOperation, proxy = this._workerProxy) {
    if (!operation) {
      return;
    }
    const generation = this._workerGeneration || 0;
    operation.workerGeneration = generation;
    operation.workerGenerations?.add(generation);
    if (proxy) {
      operation.workerProxy = proxy;
      operation.currentWorkerProxy = proxy;
      operation.workerProxies?.add(proxy);
    }
  }
  _advanceWorkerGeneration() {
    this._workerGeneration = (this._workerGeneration || 0) + 1;
    this._recordWorkerGeneration();
    return this._workerGeneration;
  }
  _isOperationCurrent(operation) {
    return !!(operation && this._activeOperation === operation && this._lifecycleState !== "disposed" && operation.state !== "cancelling" && operation.state !== "cancelled" && operation.state !== "disposed");
  }
  _isOperationCallbackCurrent(operation, proxy) {
    return this._isOperationCurrent(operation) && proxy === this._workerProxy;
  }
  _sendWorkerCancel(proxy) {
    if (!proxy) {
      return;
    }
    if (typeof proxy.call !== "function") {
      if (proxy === this._workerProxy && typeof this._callWorker === "function") {
        try {
          Promise.resolve(this._callWorker("cancel", [])).catch(() => {
          });
        } catch (_) {
        }
      }
      return;
    }
    try {
      Promise.resolve(proxy.call("cancel", [], null)).catch(() => {
      });
    } catch (_) {
    }
  }
  _cancelOperation(operation, reason = "user") {
    if (!operation) {
      return;
    }
    if (reason === "worker-timeout") {
      if (operation.cancelRequested || operation.workerTimeoutCancellationSent) {
        return;
      }
      operation.workerTimeoutCancellationSent = true;
      operation.cancelReason = reason;
      operation.state = "recovering";
      const proxies2 = new Set(operation.workerProxies || []);
      if (operation.workerProxy) {
        proxies2.add(operation.workerProxy);
      }
      for (const proxy of proxies2) {
        this._sendWorkerCancel(proxy);
      }
      return;
    }
    if (operation.cancelRequested) {
      return;
    }
    operation.cancelRequested = true;
    operation.cancelReason = reason;
    operation.state = "cancelling";
    try {
      operation.abortController?.abort();
    } catch (_) {
    }
    const proxies = new Set(operation.workerProxies || []);
    if (operation.workerProxy) {
      proxies.add(operation.workerProxy);
    }
    for (const proxy of proxies) {
      this._sendWorkerCancel(proxy);
    }
    const runtimes = new Set(operation.runtimes || []);
    if (operation.runtime) {
      runtimes.add(operation.runtime);
    }
    for (const runtime of runtimes) {
      try {
        runtime?.cancel?.();
      } catch (_) {
      }
    }
  }
  _isOperationCancellation(error, operation = this._activeOperation) {
    if (operation?.callerSignal?.aborted) {
      return true;
    }
    if (operation?.cancelRequested && operation.cancelReason !== "worker-timeout") {
      return true;
    }
    return error?.name === "AbortError" && !this._isWorkerTimeoutError(error);
  }
  _throwIfOperationCancelled(error, message = "Bridge operation was cancelled.", operation = this._activeOperation) {
    if (this._isOperationCancellation(error, operation)) {
      throw createAbortError(message);
    }
  }
  /**
   * Guards the success path: a worker that ignored cancellation and settled
   * with a full result must still reject once the caller's signal was aborted,
   * and must not commit that result. Only the caller's signal counts here.
   * A bare `cancel()` stays the documented best-effort control signal, so a
   * result that still arrived is delivered rather than discarded.
   */
  _throwIfCallerCancelled(message = "Bridge operation was cancelled.", operation = this._activeOperation) {
    if (operation?.callerSignal?.aborted) {
      throw createAbortError(message);
    }
  }
  _retireWorkerProxy(proxy) {
    if (!proxy || typeof proxy !== "object") {
      return Promise.resolve();
    }
    if (this._retiredWorkerProxies?.has(proxy)) {
      return Promise.resolve();
    }
    this._retiredWorkerProxies?.add(proxy);
    const operation = this._activeOperation;
    operation?.workerProxies?.add(proxy);
    let disposal;
    try {
      disposal = Promise.resolve(proxy.dispose?.()).catch(() => {
      });
    } catch (_) {
      disposal = Promise.resolve();
    }
    this._retiringWorkerDisposals ??= /* @__PURE__ */ new Set();
    this._retiringWorkerDisposals.add(disposal);
    this._workerDisposePromise = disposal;
    operation?.retirements?.add(disposal);
    disposal.then(
      () => {
        this._retiringWorkerDisposals.delete(disposal);
        if (this._workerDisposePromise === disposal && this._retiringWorkerDisposals.size === 0) {
          this._workerDisposePromise = null;
        }
      },
      () => {
        this._retiringWorkerDisposals.delete(disposal);
        if (this._workerDisposePromise === disposal && this._retiringWorkerDisposals.size === 0) {
          this._workerDisposePromise = null;
        }
      }
    );
    return disposal;
  }
  async _waitForWorkerDisposal() {
    this._retiringWorkerDisposals ??= /* @__PURE__ */ new Set();
    while (this._retiringWorkerDisposals.size > 0) {
      const pending = [...this._retiringWorkerDisposals];
      await Promise.all(pending);
    }
    this._workerDisposePromise = null;
  }
  /**
   * llama.cpp exposes one model/context per bridge instance and publishes
   * results through process-global C buffers (`last_output`, `last_tokens_json`,
   * `last_detokenized`, `last_embedding_json`, ...). Every async call that reads
   * or mutates that runtime therefore has to own it exclusively for its whole
   * duration: a second caller interleaving between the work ccall and the buffer
   * read would otherwise observe the other caller's result. This FIFO queue is
   * the single-writer lock, shared by the worker and direct-runtime paths.
   *
   * A caller whose signal is already aborted, or that aborts while still queued,
   * never runs and never dispatches; its slot is skipped in turn without
   * touching the unrelated operation that currently owns the runtime. Once an
   * operation has started, its signal is routed to `_cancelOperation` so the
   * cancel reaches the worker generation and runtime that operation captured
   * rather than whatever topology happens to be current.
   */
  _runExclusive(run, options = {}) {
    const signal = options.signal || null;
    const abortMessage = options.abortMessage || "Bridge operation was cancelled.";
    const allowDisposed = options.allowDisposed === true;
    const abortError = () => createAbortError(abortMessage);
    const lifecycleState = this._lifecycleState || (this._disposed ? "disposed" : "open");
    if (!allowDisposed && (this._disposed || lifecycleState !== "open")) {
      return Promise.reject(new Error(BRIDGE_DISPOSED_MESSAGE));
    }
    if (signal?.aborted) {
      return Promise.reject(abortError());
    }
    const operation = {
      id: ++this._nextOperationId,
      kind: options.kind || "bridge-operation",
      state: "queued",
      callerSignal: signal,
      abortController: typeof AbortController === "function" ? new AbortController() : null,
      cancelRequested: false,
      cancelReason: null,
      workerTimeoutCancellationSent: false,
      workerGeneration: this._workerGeneration || 0,
      workerGenerations: /* @__PURE__ */ new Set([this._workerGeneration || 0]),
      workerProxy: this._workerProxy || null,
      runtime: this._runtime || null,
      workerProxies: new Set(this._workerProxy ? [this._workerProxy] : []),
      runtimes: new Set(this._runtime ? [this._runtime] : []),
      modelSource: null,
      modelOptions: null,
      projectorSource: null,
      recoveryAttempted: false,
      generationStarted: false,
      retirements: /* @__PURE__ */ new Set(),
      deferredShadowState: null
    };
    const predecessor = this._operationQueueTail;
    let releaseSuccessor = () => {
    };
    const queueSlot = new Promise((resolve) => {
      releaseSuccessor = () => resolve(void 0);
    });
    this._operationQueueTail = queueSlot;
    let started = false;
    let removeAbortListener = null;
    let removeDisposalWaiter = null;
    const detachWatchers = () => {
      removeAbortListener?.();
      removeDisposalWaiter?.();
      removeAbortListener = null;
      removeDisposalWaiter = null;
    };
    const watchesAbort = signal != null && typeof signal.addEventListener === "function";
    const watchesDisposal = !allowDisposed;
    const ordered = (async () => {
      if (predecessor) {
        await predecessor;
      }
      if (signal?.aborted) {
        operation.state = "skipped";
        throw abortError();
      }
      if (!allowDisposed && (this._disposed || this._lifecycleState === "disposing")) {
        operation.state = "disposed";
        throw new Error(BRIDGE_DISPOSED_MESSAGE);
      }
      started = true;
      removeDisposalWaiter?.();
      removeDisposalWaiter = null;
      operation.state = "starting";
      operation.workerProxy = this._workerProxy || null;
      operation.workerGeneration = this._workerGeneration || 0;
      operation.workerGenerations.add(operation.workerGeneration);
      operation.runtime = this._runtime || null;
      if (operation.workerProxy) {
        operation.workerProxies.add(operation.workerProxy);
        operation.currentWorkerProxy = operation.workerProxy;
      }
      if (operation.runtime) {
        operation.runtimes.add(operation.runtime);
      }
      this._activeOperation = operation;
      operation.state = "running";
      let operationFailed = false;
      try {
        return await run(operation);
      } catch (error) {
        operationFailed = true;
        throw error;
      } finally {
        let terminalState = operation.cancelRequested ? "cancelled" : operationFailed ? "failed" : "completed";
        operation.state = terminalState === "completed" ? "settling" : terminalState;
        try {
          this._captureDirectRuntimeState();
        } catch (error) {
          terminalState = "failed";
          throw error;
        } finally {
          if (this._activeOperation === operation) {
            this._activeOperation = null;
          }
          operation.state = terminalState;
        }
      }
    })();
    const release = () => {
      if (this._operationQueueTail === queueSlot) {
        this._operationQueueTail = null;
      }
      releaseSuccessor();
    };
    ordered.then(release, release);
    if (!watchesAbort && !watchesDisposal) {
      return ordered;
    }
    const abandonedWhileQueued = new Promise((_, reject) => {
      if (watchesAbort) {
        const onAbort = () => {
          if (!started) {
            operation.state = "skipped";
            reject(abortError());
            return;
          }
          this._cancelOperation(operation, "signal");
        };
        signal.addEventListener("abort", onAbort, { once: true });
        removeAbortListener = () => signal.removeEventListener("abort", onAbort);
        if (signal.aborted) {
          onAbort();
        }
      }
      if (watchesDisposal && !started) {
        removeDisposalWaiter = this._addDisposalWaiter(() => {
          if (!started) {
            operation.state = "disposed";
            reject(new Error(BRIDGE_DISPOSED_MESSAGE));
          }
        });
      }
    });
    const raced = Promise.race([ordered, abandonedWhileQueued]);
    raced.then(detachWatchers, detachWatchers);
    return raced;
  }
  /**
   * Registers a disposal listener for the lifetime of one pending operation and
   * returns its remover. A listener set is used rather than a long-lived promise
   * because a promise reaction cannot be detached.
   */
  _addDisposalWaiter(onDisposed) {
    const waiters = this._disposalWaiters;
    waiters.add(onDisposed);
    return () => {
      waiters.delete(onDisposed);
    };
  }
  /**
   * True while `operation` is the non-dispose owner of the queue slot. Teardown
   * is queued behind it, so it is allowed to finish the work it already started
   * even once disposal has been requested.
   */
  _isSettleFirstOwner(operation = this._activeOperation) {
    return !!(operation && this._activeOperation === operation && operation.kind !== "dispose" && operation.state !== "cancelled" && operation.state !== "disposed");
  }
  /**
   * Lifecycle gate for the helpers that stay off the queue. They must not
   * dispatch to a torn-down worker, and must not lazily recreate the runtime
   * disposal just released. `disposing` still admits the operation that already
   * owns the slot, because disposal settles it first rather than interrupting
   * it; once teardown has run, every caller is rejected.
   */
  _throwIfDisposed() {
    if (this._lifecycleState === "disposed") {
      throw new Error(BRIDGE_DISPOSED_MESSAGE);
    }
    if ((this._disposed || this._lifecycleState === "disposing") && !this._isSettleFirstOwner()) {
      throw new Error(BRIDGE_DISPOSED_MESSAGE);
    }
  }
  _notifyDisposalWaiters() {
    const waiters = this._disposalWaiters;
    if (waiters.size === 0) {
      return;
    }
    this._disposalWaiters = /* @__PURE__ */ new Set();
    for (const waiter of waiters) {
      waiter();
    }
  }
  /**
   * Mirror the direct runtime's model/context state into facade fields so the
   * synchronous getters can answer from cache. Reading them straight from the
   * runtime would ccall into the singleton core outside the queue, racing
   * whichever async operation currently owns it.
   */
  _captureDirectRuntimeState() {
    if (this._workerProxy || !this._runtime) {
      return;
    }
    try {
      this._metadata = this._runtime.getModelMetadata();
      this._contextSize = this._runtime.getContextSize();
      this._gpuActive = this._runtime.isGpuActive();
      this._backendName = this._runtime.getBackendName();
      this._supportsVision = this._runtime.supportsVision();
      this._supportsAudio = this._runtime.supportsAudio();
    } catch (_) {
    }
  }
  /**
   * prefetch/evict stay off the queue because they only touch Cache Storage and
   * the network. In direct mode their plain-JS cache bookkeeping is copied into
   * the facade snapshot. A worker-backed snapshot remains wholly worker-owned so
   * a separate cache-only runtime cannot misdescribe the active worker model.
   */
  _refreshCacheMetadataSnapshot() {
    const runtime = this._runtime;
    if (this._workerProxy || !runtime) {
      return;
    }
    this._metadata = {
      ...this._metadata || {},
      "llamadart.webgpu.model_source": String(runtime._modelSource ?? ""),
      "llamadart.webgpu.model_cache_state": String(runtime._modelCacheState ?? ""),
      "llamadart.webgpu.model_cache_name": String(runtime._modelCacheName ?? ""),
      "llamadart.webgpu.runtime_notes": Array.isArray(runtime._runtimeNotes) ? runtime._runtimeNotes.join(";") : ""
    };
  }
  _sanitizeModelLoadOptions(options = {}) {
    const source = options && typeof options === "object" ? options : {};
    const sanitized = { ...source };
    delete sanitized.progressCallback;
    delete sanitized.signal;
    return sanitized;
  }
  _createCpuSafeMultimodalLoadOptions(options = {}) {
    const sanitized = this._sanitizeModelLoadOptions(options);
    sanitized.nGpuLayers = 0;
    if (Number.isFinite(Number(sanitized.nCtx)) && Number(sanitized.nCtx) > 4096) {
      sanitized.nCtx = 4096;
    }
    if (!Number.isFinite(Number(sanitized.nThreads)) || Number(sanitized.nThreads) <= 0) {
      sanitized.nThreads = 4;
    } else {
      sanitized.nThreads = Math.min(4, Math.max(1, Math.trunc(Number(sanitized.nThreads))));
    }
    sanitized.nThreadsBatch = sanitized.nThreads;
    if (!Number.isFinite(Number(sanitized.nBatch)) || Number(sanitized.nBatch) <= 0) {
      sanitized.nBatch = 128;
    } else {
      sanitized.nBatch = Math.min(128, Math.max(32, Math.trunc(Number(sanitized.nBatch))));
    }
    if (!Number.isFinite(Number(sanitized.nUbatch)) || Number(sanitized.nUbatch) <= 0) {
      sanitized.nUbatch = Math.min(64, sanitized.nBatch);
    } else {
      sanitized.nUbatch = Math.min(
        sanitized.nBatch,
        Math.min(64, Math.max(1, Math.trunc(Number(sanitized.nUbatch))))
      );
    }
    return sanitized;
  }
  _rememberLoadedModel(url, options = {}) {
    const normalizedUrl = cloneModelSource(url);
    if (!hasModelSource(normalizedUrl)) {
      return;
    }
    this._loadedModelUrl = normalizedUrl;
    this._loadedModelOptions = this._sanitizeModelLoadOptions(options);
    this._loadedMmProjUrl = null;
    this._forgetLoraAdapters();
    this._loadedDraftModel = null;
    this._multimodalWorkerCpuMode = this._workerProxy != null;
    this._workerModelMissing = false;
    if (this._activeOperation) {
      this._activeOperation.modelSource = cloneModelSource(normalizedUrl);
      this._activeOperation.modelOptions = { ...this._loadedModelOptions };
      this._activeOperation.projectorSource = null;
    }
  }
  _forgetLoadedModel() {
    this._loadedModelUrl = null;
    this._loadedModelOptions = null;
    this._loadedMmProjUrl = null;
    this._forgetLoraAdapters();
    this._loadedDraftModel = null;
    this._multimodalWorkerCpuMode = false;
    this._workerModelMissing = false;
  }
  /**
   * A failed load can leave its target without any model: the runtime drops
   * the previous model once the new download starts. Keep the facade's model
   * source only while the target still reports a loaded model, so recovery
   * never reloads a model the target no longer holds.
   *
   * A worker's reply is read from `attempt`, not the facade snapshot: a
   * cancelled operation no longer accepts worker state, so the snapshot can
   * still describe a model the worker already dropped. Only the bookkeeping is
   * corrected here; a cancelled load never publishes the worker's snapshot.
   */
  _syncLoadedModelAfterFailedLoad(attempt) {
    const proxy = this._workerProxy;
    if (!proxy) {
      if (!(Number(this._runtime?._modelBytes) > 0)) {
        this._forgetLoadedModel();
      }
      return;
    }
    const fromCurrentWorker = attempt.proxy === proxy;
    const replacedModel = fromCurrentWorker && attempt.loaded;
    const state = fromCurrentWorker ? attempt.state : null;
    const modelBytes = Number(
      (state ? state.metadata : this._metadata)?.["llamadart.webgpu.model_bytes"]
    );
    if (replacedModel || !(modelBytes > 0)) {
      this._forgetLoadedModel();
    }
  }
  _rememberLoadedMmProj(url) {
    const normalizedUrl = String(url || "").trim();
    if (normalizedUrl.length === 0) {
      return;
    }
    this._loadedMmProjUrl = normalizedUrl;
    if (this._activeOperation) {
      this._activeOperation.projectorSource = normalizedUrl;
    }
  }
  _hasMediaParts(options = {}) {
    return Array.isArray(options?.parts) && options.parts.length > 0;
  }
  async _replaceWorkerProxyForMultimodalCpuMode() {
    this._throwIfDisposed();
    const hadWorkerProxy = this._workerProxy != null;
    if (hadWorkerProxy) {
      const staleProxy = this._workerProxy;
      this._advanceWorkerGeneration();
      this._workerProxy = null;
      this._retireWorkerProxy(staleProxy);
      await this._waitForWorkerDisposal();
      this._throwIfDisposed();
    }
    let replacement;
    try {
      replacement = this._createWorkerProxy();
    } catch (error) {
      this._disableWorkerFallback(error);
      throw error;
    }
    try {
      this._throwIfDisposed();
    } catch (error) {
      this._retireWorkerProxy(replacement);
      throw error;
    }
    if (!hadWorkerProxy) {
      this._advanceWorkerGeneration();
    }
    this._workerProxy = replacement;
    this._recordWorkerGeneration(this._activeOperation, replacement);
    this._multimodalWorkerCpuMode = false;
    this._workerModelMissing = true;
  }
  _isRecoverableWorkerFsError(error) {
    const text = serializeWorkerError(error).toLowerCase();
    if (text.includes("model shard:")) {
      return false;
    }
    return text.includes("fs error") || text.includes("no such file") || text.includes("not found") || text.includes("invalid argument") || text.includes("timed out");
  }
  _isWorkerRequestTimeoutError(error) {
    const text = serializeWorkerError(error).toLowerCase();
    return text.includes("worker request timeout") || text.includes("worker init timeout") || text.includes("worker timed out");
  }
  async _ensureWorkerMultimodalCpuMode() {
    this._throwIfDisposed();
    if (!this._workerProxy) {
      await this._replaceWorkerProxyForMultimodalCpuMode();
    }
    this._throwIfDisposed();
    if (this._multimodalWorkerCpuMode) {
      return true;
    }
    if (!hasModelSource(this._loadedModelUrl)) {
      return false;
    }
    const selectedOptions = this._sanitizeModelLoadOptions(
      this._loadedModelOptions || {}
    );
    const operation = this._activeOperation;
    if (operation) {
      operation.recoveryAttempted = true;
      operation.state = "recovering";
    }
    const previousDeferredState = this._deferredShadowState;
    this._shadowStateTransactionDepth += 1;
    this._deferredShadowState = null;
    let recoverySucceeded = false;
    const applyWorkerSafeMode = () => this._loadRememberedModelIntoWorker(selectedOptions);
    try {
      await applyWorkerSafeMode();
      recoverySucceeded = true;
      this._emitBridgeWarn(
        "llamadart: multimodal worker prepared in selected backend mode."
      );
      return true;
    } catch (error) {
      this._throwIfDisposed();
      this._emitBridgeWarn(
        `llamadart: multimodal worker setup failed once; restarting worker (${serializeWorkerError(error)}).`
      );
      await this._replaceWorkerProxyForMultimodalCpuMode();
      await applyWorkerSafeMode();
      recoverySucceeded = true;
      this._emitBridgeWarn(
        "llamadart: multimodal worker recovered after restart."
      );
      return true;
    } finally {
      this._shadowStateTransactionDepth = Math.max(0, this._shadowStateTransactionDepth - 1);
      if (this._shadowStateTransactionDepth === 0) {
        if (recoverySucceeded && this._deferredShadowState && this._workerProxy) {
          this._applyShadowStateFrom(this._workerProxy, this._deferredShadowState);
        }
        this._deferredShadowState = previousDeferredState || null;
      }
      if (operation && this._activeOperation === operation && operation.state === "recovering") {
        operation.state = operation.cancelRequested ? "cancelling" : "running";
      }
    }
  }
  async _loadRememberedModelIntoWorker(selectedOptions) {
    this._throwIfDisposed();
    this._dropLoraAdaptersOf(this._workerProxy);
    await this._callWorker("loadModelFromUrl", [this._loadedModelUrl, selectedOptions]);
    this._throwIfDisposed();
    if (typeof this._loadedMmProjUrl === "string" && this._loadedMmProjUrl.length > 0) {
      await this._callWorker("loadMultimodalProjector", [this._loadedMmProjUrl]);
    }
    const draft = this._loadedDraftModel;
    if (draft) {
      try {
        await this._callWorker("loadDraftModel", [draft.url, draft.options]);
      } catch (error) {
        this._throwIfOperationCancelled(error, "Draft model reload was cancelled.");
        this._forgetDraftModelAfterReloadFailure(error);
      }
    }
    this._throwIfDisposed();
    await this._restoreLoraAdapters();
    this._throwIfDisposed();
    this._workerModelMissing = false;
    this._loadedModelOptions = selectedOptions;
    this._multimodalWorkerCpuMode = true;
  }
  /**
   * Multimodal recovery can replace the worker and then fail to reload the
   * model, its projector or its applied LoRA adapters into it. The facade
   * still holds that model, but the fresh worker would answer every request
   * with "No model loaded" or without the adapters. Reload them first, and
   * again on the next call if this fails. A failure here reaches the caller's
   * own worker-error handling.
   */
  async _restoreWorkerModelIfMissing() {
    if (!this._workerProxy || this._workerModelMissing !== true || !hasModelSource(this._loadedModelUrl)) {
      return;
    }
    this._emitBridgeWarn(
      "llamadart: bridge worker was restarted without its model; reloading it."
    );
    await this._loadRememberedModelIntoWorker(
      this._sanitizeModelLoadOptions(this._loadedModelOptions || {})
    );
  }
  _isDispatchWorkgroupLimitError(error) {
    const text = serializeWorkerError(error).toLowerCase();
    return text.includes("dispatch workgroup count") || text.includes("max compute workgroups per dimension") || text.includes("invalid commandbuffer") || text.includes("ggml_webgpu: device error") || text.includes("runtimeerror: aborted") || text.includes("aborted()");
  }
  // An aborted or trapped Wasm core stays dead. The worker host posts only
  // the message text, so a worker-side RuntimeError arrives without its name.
  _isWasmCoreAbortError(error) {
    if (typeof WebAssembly !== "undefined" && error instanceof WebAssembly.RuntimeError) {
      return true;
    }
    const text = serializeWorkerError(error).toLowerCase();
    return text.includes("aborted(") || text.includes("runtimeerror") || text.includes("unreachable") || text.includes("memory access out of bounds") || text.includes("function signature mismatch") || text.includes("program terminated with exit");
  }
  // A worker request falls back only when the worker can no longer serve
  // requests: it crashed, stalled, timed out, never initialized, or its core
  // aborted. A core error from a healthy worker (invalid grammar, a context
  // limit, an unsupported template, a model without embeddings) is
  // deterministic; retrying on the main thread fails the same way and strands
  // the session there.
  _isWorkerUnusableError(error) {
    if (error && typeof error === "object" && error.llamadartWorkerCrash === true) {
      return true;
    }
    return this._isWorkerTimeoutError(error) || this._isWasmCoreAbortError(error) || this._shouldFallbackToMainThread(error);
  }
  _isWorkerTimeoutError(error) {
    if (error && typeof error === "object" && error.llamadartWorkerTimeout === true) {
      return true;
    }
    const text = serializeWorkerError(error).toLowerCase();
    return text.includes("worker completion stalled") || text.includes("worker createcompletion stalled") || text.includes("worker timed out") || text.includes("worker timeout");
  }
  _isCpuModelMode() {
    const requestedLayers = Number(this._loadedModelOptions?.nGpuLayers);
    if (Number.isFinite(requestedLayers)) {
      return requestedLayers === 0;
    }
    const metadataLayers = Number(this._metadata?.["llamadart.webgpu.n_gpu_layers"]);
    if (Number.isFinite(metadataLayers)) {
      return metadataLayers === 0;
    }
    return false;
  }
  _workerCompletionStallTimeoutMs(options = {}) {
    const configured = Number(this._config?.workerGenerationStallTimeoutMs);
    if (Number.isFinite(configured) && configured > 0) {
      return Math.max(5e3, Math.min(3e5, Math.trunc(configured)));
    }
    if (!this._hasMediaParts(options)) {
      return 9e4;
    }
    if (this._isCpuModelMode()) {
      return 0;
    }
    return 18e4;
  }
  async _ensureRuntimeReadyAfterWorkerFallback(options = {}, fallbackError = null) {
    this._throwIfDisposed();
    await this._waitForWorkerDisposal();
    this._throwIfDisposed();
    if (!this._runtime) {
      this._runtime = this._createRuntime();
    }
    const operation = this._activeOperation;
    operation?.runtimes?.add(this._runtime);
    if (operation && operation.state === "running") {
      operation.state = "recovering";
    }
    const forceReloadRequested = options?._llamadartForceRuntimeReload === true;
    const mediaPartsRequested = this._hasMediaParts(options);
    const textToSpeechRequested = options?._llamadartTextToSpeech === true;
    const multimodalRuntimeRequired = mediaPartsRequested || textToSpeechRequested;
    const shouldEnsureMultimodalInRuntime = multimodalRuntimeRequired && typeof this._loadedMmProjUrl === "string" && this._loadedMmProjUrl.length > 0;
    const workerTimedOut = this._isWorkerTimeoutError(fallbackError);
    const dispatchWorkgroupFallback = this._isDispatchWorkgroupLimitError(fallbackError);
    const loadedGpuLayers = Number(this._loadedModelOptions?.nGpuLayers);
    const metadataGpuLayers = Number(this._metadata?.["llamadart.webgpu.n_gpu_layers"]);
    const modelLoadedWithGpu = Number.isFinite(loadedGpuLayers) ? loadedGpuLayers !== 0 : Number.isFinite(metadataGpuLayers) ? metadataGpuLayers !== 0 : true;
    const shouldUseCpuMultimodalFallback = multimodalRuntimeRequired && modelLoadedWithGpu && (dispatchWorkgroupFallback || workerTimedOut);
    try {
      if (Number(this._runtime?._modelBytes) > 0 && !forceReloadRequested && !shouldUseCpuMultimodalFallback) {
        await this._ensureRuntimeDraftModel();
        if (shouldEnsureMultimodalInRuntime) {
          const runtimeSupportsMedia = typeof this._runtime.supportsVision === "function" && this._runtime.supportsVision() || typeof this._runtime.supportsAudio === "function" && this._runtime.supportsAudio();
          if (!runtimeSupportsMedia) {
            this._throwIfDisposed();
            await this._runtime.loadMultimodalProjector(this._loadedMmProjUrl);
            this._throwIfDisposed();
            if (Array.isArray(this._runtime._runtimeNotes)) {
              this._runtime._runtimeNotes.push("worker_fallback_reload_mmproj");
            }
          }
        }
        await this._restoreLoraAdapters();
        this._throwIfDisposed();
        return;
      }
      if (!hasModelSource(this._loadedModelUrl)) {
        return;
      }
      const loadOptions = shouldUseCpuMultimodalFallback ? this._createCpuSafeMultimodalLoadOptions(this._loadedModelOptions || {}) : this._sanitizeModelLoadOptions(this._loadedModelOptions || {});
      const recoverySignal = this._operationSignal(operation) || options.signal || null;
      if (recoverySignal) {
        loadOptions.signal = recoverySignal;
      }
      if (shouldUseCpuMultimodalFallback) {
        if (textToSpeechRequested) {
          this._emitBridgeWarn(
            "llamadart: retrying text-to-speech once with CPU fallback after WebGPU failure."
          );
        } else if (workerTimedOut) {
          this._emitBridgeWarn(
            "llamadart: retrying multimodal generation with CPU fallback after worker timeout."
          );
        } else {
          this._emitBridgeWarn(
            "llamadart: retrying multimodal generation with CPU fallback after WebGPU workgroup limit failure."
          );
        }
      }
      if (workerTimedOut) {
        this._emitBridgeWarn(
          "llamadart: bridge worker completion stalled; restarting generation path on main-thread runtime."
        );
      }
      this._throwIfDisposed();
      this._dropLoraAdaptersOf(this._runtime);
      await this._runtime.loadModelFromUrl(this._loadedModelUrl, loadOptions);
      this._throwIfDisposed();
      if (Array.isArray(this._runtime._runtimeNotes)) {
        this._runtime._runtimeNotes.push("worker_fallback_reload_model");
        if (forceReloadRequested) {
          this._runtime._runtimeNotes.push("worker_fallback_reload_forced");
        }
        if (workerTimedOut) {
          this._runtime._runtimeNotes.push("worker_fallback_timeout");
        }
        if (shouldUseCpuMultimodalFallback) {
          this._runtime._runtimeNotes.push("worker_fallback_cpu_multimodal");
          if (textToSpeechRequested) {
            this._runtime._runtimeNotes.push("worker_fallback_cpu_text_to_speech");
          }
        }
      }
      if (shouldEnsureMultimodalInRuntime) {
        this._throwIfDisposed();
        await this._runtime.loadMultimodalProjector(this._loadedMmProjUrl);
        this._throwIfDisposed();
      }
      await this._restoreLoraAdapters();
      this._throwIfDisposed();
      await this._ensureRuntimeDraftModel();
    } catch (error) {
      if (this._disposed || this._lifecycleState === "disposing") {
        throw new Error(BRIDGE_DISPOSED_MESSAGE);
      }
      if (error?.name === "AbortError") {
        throw error;
      }
      try {
        await this._runtime?.dispose?.();
      } catch (_) {
      }
      this._runtime = null;
      this._metadata = {};
      this._contextSize = 0;
      this._gpuActive = false;
      this._backendName = "";
      this._supportsVision = false;
      this._supportsAudio = false;
      this._loadedModelUrl = null;
      this._loadedModelOptions = null;
      this._loadedMmProjUrl = null;
      this._forgetLoraAdapters();
      this._loadedDraftModel = null;
      if (!this._disposed && this._lifecycleState === "open" && !this._workerProxy) {
        this._runtime = this._createRuntime();
        operation?.runtimes?.add(this._runtime);
      }
      throw error;
    } finally {
      if (operation && this._activeOperation === operation && operation.state === "recovering") {
        operation.state = operation.cancelRequested ? "cancelling" : "running";
      }
    }
  }
  _createWorkerProxy() {
    return new BridgeWorkerProxy({
      moduleUrl: this._workerModuleUrl(),
      config: this._workerConfig()
    });
  }
  _shouldUseWorker() {
    if (this._config?.disableWorker === true) {
      return false;
    }
    if (typeof Worker === "undefined" || typeof Blob === "undefined" || typeof URL === "undefined" || typeof URL.createObjectURL !== "function") {
      return false;
    }
    if (this._config?.coreModuleFactory != null) {
      return false;
    }
    return true;
  }
  _workerModuleUrl() {
    const candidate = this._config?.workerUrl;
    if (typeof candidate === "string" && candidate.trim().length > 0) {
      return candidate.trim();
    }
    try {
      return new URL("./llama_webgpu_bridge_worker.js", import.meta.url).toString();
    } catch (_) {
      return import.meta.url;
    }
  }
  _workerConfig() {
    const config = this._config || {};
    return {
      workerRequestTimeoutMs: config.workerRequestTimeoutMs,
      workerInitTimeoutMs: config.workerInitTimeoutMs,
      workerModelLoadTimeoutMs: config.workerModelLoadTimeoutMs,
      workerMmprojLoadTimeoutMs: config.workerMmprojLoadTimeoutMs,
      workerCompletionTimeoutMs: config.workerCompletionTimeoutMs,
      workerTextToSpeechTimeoutMs: config.workerTextToSpeechTimeoutMs,
      coreInitTimeoutMs: config.coreInitTimeoutMs,
      wasmUrl: typeof config.wasmUrl === "string" ? config.wasmUrl : void 0,
      wasmUrlMem64: typeof config.wasmUrlMem64 === "string" ? config.wasmUrlMem64 : void 0,
      coreModuleUrl: typeof config.coreModuleUrl === "string" ? config.coreModuleUrl : void 0,
      coreModuleUrlMem64: typeof config.coreModuleUrlMem64 === "string" ? config.coreModuleUrlMem64 : void 0,
      preferMemory64: typeof config.preferMemory64 === "boolean" ? config.preferMemory64 : void 0,
      threads: Number(config.threads) > 0 ? Number(config.threads) : void 0,
      threadsBatch: Number(config.threadsBatch) > 0 ? Number(config.threadsBatch) : void 0,
      threadPoolSize: Number(config.threadPoolSize) > 0 ? Number(config.threadPoolSize) : void 0,
      nBatch: Number(config.nBatch) > 0 ? Number(config.nBatch) : void 0,
      nUbatch: Number(config.nUbatch) > 0 ? Number(config.nUbatch) : void 0,
      nGpuLayers: Number.isFinite(config.nGpuLayers) ? Number(config.nGpuLayers) : void 0,
      userAgent: typeof config.userAgent === "string" ? config.userAgent : void 0,
      cacheName: typeof config.cacheName === "string" ? config.cacheName : void 0,
      remoteFetchThresholdBytes: Number(config.remoteFetchThresholdBytes) > 0 ? Number(config.remoteFetchThresholdBytes) : void 0,
      remoteFetchChunkBytes: Number(config.remoteFetchChunkBytes) > 0 ? Number(config.remoteFetchChunkBytes) : void 0,
      mediaMaxImagePixels: Number(config.mediaMaxImagePixels) > 0 ? Number(config.mediaMaxImagePixels) : void 0,
      mediaMaxImageEdge: Number(config.mediaMaxImageEdge) > 0 ? Number(config.mediaMaxImageEdge) : void 0,
      disableImageDownscale: config.disableImageDownscale === true,
      allowAutoRemoteFetchBackend: config.allowAutoRemoteFetchBackend === true,
      logLevel: Number.isFinite(config.logLevel) ? Number(config.logLevel) : 2
    };
  }
  _applyShadowState(state) {
    if (!state || typeof state !== "object") {
      return;
    }
    if (state.metadata && typeof state.metadata === "object") {
      this._metadata = state.metadata;
    }
    if (Number.isFinite(state.contextSize)) {
      this._contextSize = Number(state.contextSize);
    }
    if (typeof state.gpuActive === "boolean") {
      this._gpuActive = state.gpuActive;
    }
    if (typeof state.backendName === "string" && state.backendName.length > 0) {
      this._backendName = state.backendName;
    }
    if (typeof state.supportsVision === "boolean") {
      this._supportsVision = state.supportsVision;
    }
    if (typeof state.supportsAudio === "boolean") {
      this._supportsAudio = state.supportsAudio;
    }
  }
  _shouldFallbackToMainThread(error) {
    const text = serializeWorkerError(error).toLowerCase();
    if (text.includes("aborted(native code called abort())")) {
      return false;
    }
    if (text.includes("array buffer allocation failed")) {
      return false;
    }
    if (text.includes("bad_alloc")) {
      return false;
    }
    if (text.includes("out of memory")) {
      return false;
    }
    if (text.includes("memory access out of bounds")) {
      return false;
    }
    if (text.includes("bridge worker")) {
      return true;
    }
    if (text.includes("worker request failed")) {
      return true;
    }
    if (text.includes("worker request timeout")) {
      return true;
    }
    if (text.includes("worker init timeout")) {
      return true;
    }
    if (text.includes("timed out")) {
      return true;
    }
    if (text.includes("worker proxy is not available")) {
      return true;
    }
    if (text.includes("worker is not initialized")) {
      return true;
    }
    if (text.includes("failed to initialize bridge worker")) {
      return true;
    }
    if (text.includes("script error")) {
      return true;
    }
    return false;
  }
  _resolvedBridgeLogLevel() {
    const configured = Number(this._config?.logLevel);
    if (Number.isFinite(configured)) {
      return Math.max(0, Math.min(4, Math.trunc(configured)));
    }
    const runtimeLevel = Number(this._runtime?._logLevel);
    if (Number.isFinite(runtimeLevel)) {
      return Math.max(0, Math.min(4, Math.trunc(runtimeLevel)));
    }
    return 2;
  }
  _shouldEmitBridgeLevel(level) {
    const configured = this._resolvedBridgeLogLevel();
    const threshold = logThresholdForConfiguredLevel(configured);
    if (threshold > 3) {
      return false;
    }
    return logLevelForName(level) >= threshold;
  }
  _shouldSuppressBridgeWarn(message) {
    const text = String(message || "").trim();
    if (text.length === 0) {
      return false;
    }
    const configuredWindow = Number(this._config?.warnDedupWindowMs);
    const dedupWindowMs = Number.isFinite(configuredWindow) && configuredWindow > 0 ? Math.max(500, Math.min(6e4, Math.trunc(configuredWindow))) : 5e3;
    const now = Date.now();
    const last = Number(this._bridgeWarnRecent.get(text) || 0);
    this._bridgeWarnRecent.set(text, now);
    if (this._bridgeWarnRecent.size > 80) {
      const staleThreshold = now - dedupWindowMs * 2;
      for (const [key, atMs] of this._bridgeWarnRecent.entries()) {
        if (Number(atMs) < staleThreshold) {
          this._bridgeWarnRecent.delete(key);
        }
      }
    }
    return last > 0 && now - last < dedupWindowMs;
  }
  _emitBridgeWarn(message) {
    if (!this._shouldEmitBridgeLevel("warn")) {
      return;
    }
    if (this._shouldSuppressBridgeWarn(message)) {
      return;
    }
    if (this._runtime && typeof this._runtime._emitLogger === "function") {
      this._runtime._emitLogger("warn", message);
      return;
    }
    if (typeof console !== "undefined" && typeof console.warn === "function") {
      console.warn(message);
    }
  }
  _disableWorkerFallback(error) {
    const reason = serializeWorkerError(error);
    this._workerFallbackReason = reason;
    if (typeof globalThis !== "undefined") {
      globalThis.__llamadartBridgeWorkerFallbackReason = reason;
    }
    this._emitBridgeWarn(
      `llamadart: bridge worker unavailable, falling back to main thread (${reason})`
    );
    if (this._workerProxy) {
      const workerProxy = this._workerProxy;
      this._advanceWorkerGeneration();
      this._workerProxy = null;
      this._retireWorkerProxy(workerProxy);
    }
    this._multimodalWorkerCpuMode = false;
    this._workerModelMissing = false;
    const activeOperationMayFinish = this._activeOperation && this._activeOperation.kind !== "dispose" && this._activeOperation.state !== "cancelled" && this._activeOperation.state !== "disposed";
    if ((this._disposed || this._lifecycleState === "disposing") && !activeOperationMayFinish) {
      return;
    }
    if (!this._runtime) {
      this._runtime = this._createRuntime();
      this._activeOperation?.runtimes?.add(this._runtime);
    }
    const note = `worker_fallback:${reason}`;
    if (this._runtime && Array.isArray(this._runtime._runtimeNotes) && typeof reason === "string" && reason.length > 0 && !this._runtime._runtimeNotes.includes(note)) {
      this._runtime._runtimeNotes.push(note);
    }
  }
  async _callWorker(method, args, onEvent, transferList = []) {
    const proxy = this._workerProxy;
    if (!proxy) {
      throw new Error("Bridge worker proxy is not available");
    }
    const operation = this._activeOperation;
    if (operation) {
      this._recordWorkerGeneration(operation, proxy);
    }
    const guardedEvent = typeof onEvent === "function" ? (event) => {
      if (!operation || this._isOperationCallbackCurrent(operation, proxy) || event.event === "usage" && proxy === this._workerProxy) {
        onEvent(event);
      }
    } : void 0;
    try {
      const response = await proxy.call(
        method,
        args,
        guardedEvent,
        transferList,
        operation ? { operationId: operation.id, workerGeneration: operation.workerGeneration } : void 0
      );
      if (response?.state) {
        this._acceptWorkerState(operation, proxy, response.state);
      }
      return response?.value;
    } catch (error) {
      if (error && typeof error === "object" && error.state) {
        this._acceptWorkerState(operation, proxy, error.state);
      }
      throw error;
    }
  }
  /**
   * A worker response may only reach facade state while its proxy is still the
   * operation's current owner. Staging it for an open recovery transaction is
   * the same commit deferred, so a retired generation must be rejected before
   * the transaction branch, not after it.
   */
  _acceptWorkerState(operation, proxy, state) {
    if (operation && !this._isOperationCallbackCurrent(operation, proxy)) {
      return;
    }
    if (this._shadowStateTransactionDepth > 0) {
      this._deferredShadowState = state;
      if (operation) {
        operation.deferredShadowState = state;
      }
      return;
    }
    this._applyShadowStateFrom(proxy, state);
  }
  /**
   * The getters now answer from the facade snapshot, so a response from a proxy
   * that has since been replaced must not write into the current worker's state.
   */
  _applyShadowStateFrom(proxy, state) {
    if (proxy !== this._workerProxy) {
      return;
    }
    this._applyShadowState(state);
  }
  async loadModelFromUrl(url, options = {}) {
    return this._runExclusive(
      () => this._loadModelFromUrlUnlocked(url, options),
      {
        signal: options?.signal,
        abortMessage: "Model load was cancelled.",
        kind: "load-model"
      }
    );
  }
  async _loadModelFromUrlUnlocked(url, options = {}) {
    const attempt = { proxy: null, state: null, loaded: false };
    try {
      return await this._loadModelFromUrlOnTarget(url, options, attempt);
    } catch (error) {
      this._syncLoadedModelAfterFailedLoad(attempt);
      throw error;
    }
  }
  /**
   * @param attempt Records the last worker load's proxy and reply for
   *   `_syncLoadedModelAfterFailedLoad`.
   */
  async _loadModelFromUrlOnTarget(url, options, attempt) {
    if (!this._workerProxy) {
      const runtimeOptions = {
        ...options,
        signal: this._operationSignal() || options.signal || null
      };
      const result = await this._runtime.loadModelFromUrl(url, runtimeOptions);
      this._rememberLoadedModel(url, options);
      return result;
    }
    const invokeWorkerLoad = async () => {
      const workerOptions = { ...options };
      delete workerOptions.progressCallback;
      delete workerOptions.signal;
      const proxy = this._workerProxy;
      attempt.proxy = proxy;
      attempt.state = null;
      attempt.loaded = false;
      let result;
      try {
        result = await this._callWorker(
          "loadModelFromUrl",
          [url, workerOptions],
          (event) => {
            if (event.event !== "progress") {
              return;
            }
            if (typeof options.progressCallback !== "function") {
              return;
            }
            options.progressCallback(event.payload || {});
          }
        );
      } catch (error) {
        attempt.state = error && typeof error === "object" ? error.state || null : null;
        throw error;
      }
      attempt.loaded = true;
      this._throwIfCallerCancelled("Model load was cancelled.");
      this._rememberLoadedModel(url, workerOptions);
      return result;
    };
    try {
      return await invokeWorkerLoad();
    } catch (error) {
      this._throwIfOperationCancelled(error, "Model load was cancelled.");
      if (this._isRecoverableWorkerFsError(error) && !this._isWorkerRequestTimeoutError(error)) {
        this._emitBridgeWarn(
          `llamadart: worker model-load FS error detected; restarting worker (${serializeWorkerError(error)}).`
        );
        try {
          await this._replaceWorkerProxyForMultimodalCpuMode();
          this._forgetLoadedModel();
          this._applyShadowState(emptyBridgeState());
          return await invokeWorkerLoad();
        } catch (retryError) {
          this._emitBridgeWarn(
            `llamadart: worker model-load retry failed (${serializeWorkerError(retryError)}).`
          );
          error = retryError;
        }
      }
      if (!this._shouldFallbackToMainThread(error)) {
        throw error;
      }
      this._disableWorkerFallback(error);
      await this._waitForWorkerDisposal();
      this._throwIfDisposed();
      const result = await this._runtime.loadModelFromUrl(url, {
        ...options,
        signal: this._operationSignal() || options.signal || null
      });
      this._rememberLoadedModel(url, options);
      return result;
    }
  }
  async prefetchModelToCache(url, options = {}) {
    this._throwIfDisposed();
    if (!this._runtime) {
      this._runtime = this._createRuntime();
    }
    try {
      return await this._runtime.prefetchModelToCache(url, options);
    } finally {
      this._refreshCacheMetadataSnapshot();
    }
  }
  async evictModelFromCache(url, options = {}) {
    this._throwIfDisposed();
    if (!this._runtime) {
      this._runtime = this._createRuntime();
    }
    try {
      return await this._runtime.evictModelFromCache(url, options);
    } finally {
      this._refreshCacheMetadataSnapshot();
    }
  }
  async createCompletion(prompt, options = {}) {
    resolveCompletionSamplingOptions(options);
    const onUsage = options?.onUsage;
    if (typeof onUsage !== "function") {
      return this._runExclusive(
        () => this._createCompletionUnlocked(prompt, options),
        {
          signal: options?.signal,
          abortMessage: "Generation was cancelled.",
          kind: "generation"
        }
      );
    }
    let usage = null;
    let text;
    try {
      text = await this._runExclusive(
        () => this._createCompletionUnlocked(prompt, {
          ...options,
          onUsage: (value) => {
            usage = value;
          }
        }),
        {
          signal: options.signal,
          abortMessage: "Generation was cancelled.",
          kind: "generation"
        }
      );
    } catch (error) {
      if (usage != null && error?.name === "AbortError") {
        onUsage(usage);
      }
      throw error;
    }
    if (usage != null) {
      onUsage(usage);
    }
    return text;
  }
  async getCompletionCapabilities() {
    return this._runExclusive(
      () => this._getCompletionCapabilitiesUnlocked(),
      { kind: "completion-capabilities" }
    );
  }
  async _getCompletionCapabilitiesUnlocked() {
    if (!this._workerProxy) {
      return this._runtime?.getCompletionCapabilities() ?? { ...NO_COMPLETION_CAPABILITIES };
    }
    try {
      await this._restoreWorkerModelIfMissing();
      return await this._callWorker("getCompletionCapabilities", []);
    } catch (error) {
      this._throwIfOperationCancelled(error, "Completion capability probe was cancelled.");
      if (!this._isWorkerUnusableError(error)) {
        throw error;
      }
      this._disableWorkerFallback(error);
      await this._waitForWorkerDisposal();
      await this._ensureRuntimeReadyAfterWorkerFallback({}, error);
      return this._runtime.getCompletionCapabilities();
    }
  }
  async _createCompletionUnlocked(prompt, options = {}) {
    const isWarmup = options?.warmup === true;
    const hasRetriedEmptyMultimodal = options?.__llamadartEmptyRetryAttempted === true;
    const workerAllowed = this._config?.disableWorker !== true;
    if (this._hasMediaParts(options) && workerAllowed) {
      const hasWorkerFallback = typeof this._workerFallbackReason === "string" && this._workerFallbackReason.length > 0;
      if (hasWorkerFallback && !this._workerProxy && !this._isCpuModelMode()) {
        await this._ensureRuntimeReadyAfterWorkerFallback(options, null);
        return this._runtime.createCompletion(prompt, options);
      }
      try {
        if (!this._workerProxy) {
          await this._replaceWorkerProxyForMultimodalCpuMode();
        }
        await this._ensureWorkerMultimodalCpuMode();
      } catch (error) {
        this._throwIfOperationCancelled(error, "Generation was cancelled.");
        const reason = serializeWorkerError(error);
        if (isWarmup) {
          this._emitBridgeWarn(
            `llamadart: multimodal warmup skipped after worker setup issue (${reason}).`
          );
          return "";
        }
        this._emitBridgeWarn(
          `llamadart: unable to prepare multimodal worker CPU mode (${reason}).`
        );
        if (this._isCpuModelMode()) {
          throw new Error(
            `CPU multimodal worker setup failed (${reason}). Reload model and retry with a smaller image.`
          );
        }
        this._disableWorkerFallback(error);
        await this._waitForWorkerDisposal();
        await this._ensureRuntimeReadyAfterWorkerFallback(options, error);
        return this._runtime.createCompletion(prompt, options);
      }
    }
    if (!this._workerProxy) {
      return this._runtime.createCompletion(prompt, options);
    }
    try {
      await this._restoreWorkerModelIfMissing();
      const workerOptions = { ...options };
      delete workerOptions.onToken;
      delete workerOptions.onUsage;
      delete workerOptions.signal;
      delete workerOptions.__llamadartEmptyRetryAttempted;
      const speculative = options.speculativeDecoding;
      if (speculative != null && typeof speculative === "object" && !Array.isArray(speculative)) {
        const resolveCache = (source) => typeof source === "string" && source.length > 0 ? normalizeAbsoluteUrl(source) : source;
        workerOptions.speculativeDecoding = {
          ...speculative,
          ngramCacheStatic: resolveCache(speculative.ngramCacheStatic),
          ngramCacheDynamic: resolveCache(speculative.ngramCacheDynamic)
        };
      }
      const stallTimeoutMs = this._workerCompletionStallTimeoutMs(options);
      let timeoutHandle = null;
      let rejectOnStall = null;
      const clearStallTimer = () => {
        if (timeoutHandle != null) {
          globalThis.clearTimeout(timeoutHandle);
          timeoutHandle = null;
        }
      };
      const armStallTimer = () => {
        clearStallTimer();
        if (!Number.isFinite(stallTimeoutMs) || stallTimeoutMs <= 0) {
          return;
        }
        timeoutHandle = globalThis.setTimeout(() => {
          const timeoutError = new Error(
            `Bridge worker completion stalled for ${stallTimeoutMs}ms.`
          );
          timeoutError.llamadartWorkerTimeout = true;
          this._cancelOperation(this._activeOperation, "worker-timeout");
          rejectOnStall?.(timeoutError);
        }, stallTimeoutMs);
      };
      const stallPromise = new Promise((_, reject) => {
        rejectOnStall = reject;
      });
      armStallTimer();
      let sawWorkerTokenEvent = false;
      try {
        const workerResult = await Promise.race([
          this._callWorker(
            "createCompletion",
            [prompt, workerOptions],
            (event) => {
              if (event.event === "usage") {
                if (typeof options.onUsage === "function") {
                  options.onUsage(event.payload);
                }
                return;
              }
              if (event.event !== "token") {
                return;
              }
              armStallTimer();
              sawWorkerTokenEvent = true;
              if (typeof options.onToken !== "function") {
                return;
              }
              const payload = event.payload || {};
              const piece = typeof payload.pieceText === "string" ? payload.pieceText : Uint8Array.from(Array.isArray(payload.piece) ? payload.piece : []);
              options.onToken(piece, String(payload.currentText || ""));
            }
          ),
          stallPromise
        ]);
        if (this._hasMediaParts(options) && !isWarmup && !sawWorkerTokenEvent && String(workerResult || "").trim().length == 0) {
          this._emitBridgeWarn(
            "llamadart: multimodal worker produced empty response without token events."
          );
          if (!hasRetriedEmptyMultimodal) {
            this._emitBridgeWarn(
              "llamadart: retrying multimodal worker once after empty response."
            );
            try {
              await this._replaceWorkerProxyForMultimodalCpuMode();
              await this._ensureWorkerMultimodalCpuMode();
            } catch (retrySetupError) {
              this._emitBridgeWarn(
                `llamadart: multimodal empty-response retry setup failed (${serializeWorkerError(retrySetupError)}).`
              );
            }
            return this._createCompletionUnlocked(prompt, {
              ...options,
              __llamadartEmptyRetryAttempted: true
            });
          }
        }
        this._throwIfCallerCancelled("Generation was cancelled.");
        return workerResult;
      } finally {
        clearStallTimer();
      }
    } catch (error) {
      this._throwIfOperationCancelled(error, "Generation was cancelled.");
      if (serializeWorkerError(error).includes(INVALID_GRAMMAR_ERROR_TEXT)) {
        throw error;
      }
      if (this._hasMediaParts(options)) {
        const reason = serializeWorkerError(error);
        if (isWarmup) {
          this._emitBridgeWarn(
            `llamadart: multimodal warmup skipped after worker request issue (${reason}).`
          );
          return "";
        }
        if (this._isCpuModelMode()) {
          this._emitBridgeWarn(
            `llamadart: CPU multimodal worker request failed (${reason}); skipping main-thread fallback.`
          );
          throw new Error(
            `CPU multimodal request failed (${reason}). Reload model and retry with a smaller image.`
          );
        }
        this._emitBridgeWarn(
          `llamadart: multimodal worker request failed (${reason}); falling back to main-thread runtime.`
        );
        this._disableWorkerFallback(error);
        await this._waitForWorkerDisposal();
        await this._ensureRuntimeReadyAfterWorkerFallback(options, error);
        return this._runtime.createCompletion(prompt, options);
      }
      if (!this._isWorkerUnusableError(error)) {
        throw error;
      }
      this._disableWorkerFallback(error);
      await this._waitForWorkerDisposal();
      await this._ensureRuntimeReadyAfterWorkerFallback(options, error);
      return this._runtime.createCompletion(prompt, options);
    }
  }
  async loadMultimodalProjector(url) {
    return this._runExclusive(
      () => this._loadMultimodalProjectorUnlocked(url),
      { kind: "projector-load" }
    );
  }
  async _loadMultimodalProjectorUnlocked(url) {
    const startedWithWorker = this._workerProxy != null;
    const invokeRuntimeLoad = async () => {
      if (!this._runtime) {
        this._runtime = this._createRuntime();
      }
      await this._ensureRuntimeReadyAfterWorkerFallback({}, null);
      const result = await this._runtime.loadMultimodalProjector(url);
      this._rememberLoadedMmProj(url);
      this._supportsVision = this._runtime.supportsVision();
      this._supportsAudio = this._runtime.supportsAudio();
      return result;
    };
    try {
      if (!this._workerProxy) {
        return await invokeRuntimeLoad();
      }
      await this._ensureWorkerMultimodalCpuMode();
      const result = await this._callWorker("loadMultimodalProjector", [url]);
      this._rememberLoadedMmProj(url);
      return result;
    } catch (error) {
      this._throwIfOperationCancelled(error, "Projector operation was cancelled.");
      if (!startedWithWorker) {
        throw error;
      }
      const reason = serializeWorkerError(error);
      this._emitBridgeWarn(
        `llamadart: multimodal worker setup failed (${reason}).`
      );
      if (this._isCpuModelMode()) {
        try {
          await this._replaceWorkerProxyForMultimodalCpuMode();
          await this._ensureWorkerMultimodalCpuMode();
          const retryResult = await this._callWorker("loadMultimodalProjector", [url]);
          this._rememberLoadedMmProj(url);
          this._emitBridgeWarn(
            "llamadart: CPU multimodal worker setup recovered after worker restart."
          );
          return retryResult;
        } catch (retryError) {
          const retryReason = serializeWorkerError(retryError);
          throw new Error(
            `CPU multimodal projector setup failed (${retryReason}). Reload model and retry with a smaller image.`
          );
        }
      }
      this._disableWorkerFallback(error);
      await this._waitForWorkerDisposal();
      return invokeRuntimeLoad();
    }
  }
  async unloadMultimodalProjector() {
    return this._runExclusive(
      () => this._unloadMultimodalProjectorUnlocked(),
      { kind: "projector-unload" }
    );
  }
  async _unloadMultimodalProjectorUnlocked() {
    if (!this._workerProxy) {
      const result = await this._runtime.unloadMultimodalProjector();
      this._loadedMmProjUrl = null;
      this._supportsVision = this._runtime.supportsVision();
      this._supportsAudio = this._runtime.supportsAudio();
      return result;
    }
    try {
      const result = await this._callWorker("unloadMultimodalProjector", []);
      this._loadedMmProjUrl = null;
      return result;
    } catch (error) {
      this._throwIfOperationCancelled(error, "Projector operation was cancelled.");
      this._disableWorkerFallback(error);
      await this._waitForWorkerDisposal();
      const result = await this._runtime.unloadMultimodalProjector();
      this._loadedMmProjUrl = null;
      this._supportsVision = this._runtime.supportsVision();
      this._supportsAudio = this._runtime.supportsAudio();
      return result;
    }
  }
  // Loads the remembered draft model into the direct runtime when that
  // runtime does not hold it.
  async _ensureRuntimeDraftModel() {
    const draft = this._loadedDraftModel;
    const runtime = this._runtime;
    if (!draft || !runtime || runtime._draftModel?.url === draft.url) {
      return;
    }
    this._throwIfDisposed();
    try {
      await runtime.loadDraftModel(draft.url, {
        ...draft.options,
        signal: this._operationSignal() || void 0
      });
    } catch (error) {
      this._throwIfOperationCancelled(error, "Draft model reload was cancelled.");
      this._forgetDraftModelAfterReloadFailure(error);
    }
    this._throwIfDisposed();
  }
  // A reloaded target starts without its draft model. A failed draft reload
  // leaves none, so draft-* strategies then reject until loadDraftModel runs.
  _forgetDraftModelAfterReloadFailure(error) {
    this._loadedDraftModel = null;
    this._emitBridgeWarn(
      `llamadart: draft model reload failed after the model was reloaded (${serializeWorkerError(error)}); call loadDraftModel again.`
    );
  }
  async loadDraftModel(url, options = {}) {
    return this._runExclusive(
      () => this._loadDraftModelUnlocked(url, options),
      {
        signal: options?.signal,
        abortMessage: "Draft model load was cancelled.",
        kind: "draft-model-load"
      }
    );
  }
  async _loadDraftModelUnlocked(url, options = {}) {
    const absoluteUrl = typeof url === "string" && url.length > 0 ? normalizeAbsoluteUrl(url) : url;
    const remembered = {
      url: absoluteUrl,
      options: typeof options?.useCache === "boolean" ? { useCache: options.useCache } : {}
    };
    const loadInRuntime = async () => {
      this._loadedDraftModel = null;
      const info = await this._runtime.loadDraftModel(absoluteUrl, {
        ...options,
        signal: this._operationSignal() || options.signal || void 0
      });
      this._loadedDraftModel = remembered;
      return info;
    };
    if (!this._workerProxy) {
      return loadInRuntime();
    }
    const workerOptions = { ...options };
    delete workerOptions.progressCallback;
    delete workerOptions.signal;
    try {
      await this._restoreWorkerModelIfMissing();
      this._loadedDraftModel = null;
      const info = await this._callWorker(
        "loadDraftModel",
        [absoluteUrl, workerOptions],
        (event) => {
          if (event.event === "progress" && typeof options.progressCallback === "function") {
            options.progressCallback(event.payload || {});
          }
        }
      );
      this._throwIfCallerCancelled("Draft model load was cancelled.");
      this._loadedDraftModel = remembered;
      return info;
    } catch (error) {
      this._throwIfOperationCancelled(error, "Draft model load was cancelled.");
      if (!this._isWorkerUnusableError(error)) {
        throw error;
      }
      this._disableWorkerFallback(error);
      await this._waitForWorkerDisposal();
      await this._ensureRuntimeReadyAfterWorkerFallback({}, error);
      return loadInRuntime();
    }
  }
  async unloadDraftModel() {
    return this._runExclusive(
      () => this._unloadDraftModelUnlocked(),
      { kind: "draft-model-unload" }
    );
  }
  async _unloadDraftModelUnlocked() {
    if (!this._workerProxy) {
      await this._runtime?.unloadDraftModel();
      this._loadedDraftModel = null;
      return;
    }
    try {
      await this._callWorker("unloadDraftModel", []);
      this._loadedDraftModel = null;
    } catch (error) {
      this._throwIfOperationCancelled(error, "Draft model unload was cancelled.");
      if (!this._isWorkerUnusableError(error)) {
        throw error;
      }
      this._loadedDraftModel = null;
      this._disableWorkerFallback(error);
      await this._waitForWorkerDisposal();
      await this._runtime?.unloadDraftModel();
    }
  }
  supportsVision() {
    return this._supportsVision;
  }
  supportsAudio() {
    return this._supportsAudio;
  }
  async getTextToSpeechCapabilities() {
    return this._runExclusive(
      () => this._getTextToSpeechCapabilitiesUnlocked(),
      { kind: "tts-capabilities" }
    );
  }
  async _getTextToSpeechCapabilitiesUnlocked() {
    if (!this._workerProxy) {
      return this._runtime.getTextToSpeechCapabilities();
    }
    try {
      await this._restoreWorkerModelIfMissing();
      return await this._callWorker("getTextToSpeechCapabilities", []);
    } catch (error) {
      this._throwIfOperationCancelled(error, "Text-to-speech capability probe was cancelled.");
      this._disableWorkerFallback(error);
      await this._waitForWorkerDisposal();
      await this._ensureRuntimeReadyAfterWorkerFallback({}, error);
      return this._runtime.getTextToSpeechCapabilities();
    }
  }
  async synthesizeSpeech(options = {}) {
    return this._runExclusive(
      () => this._synthesizeSpeechUnlocked(options),
      {
        signal: options?.signal,
        abortMessage: "Text-to-speech synthesis was cancelled.",
        kind: "text-to-speech"
      }
    );
  }
  async _synthesizeSpeechUnlocked(options = {}) {
    if (options.signal?.aborted) {
      throw new DOMException("Text-to-speech synthesis was cancelled.", "AbortError");
    }
    if (!this._workerProxy) {
      return this._runtime.synthesizeSpeech(options);
    }
    try {
      await this._restoreWorkerModelIfMissing();
      const workerOptions = { ...options };
      delete workerOptions.onProgress;
      delete workerOptions.signal;
      const transferList = [];
      const speakerBytes = toUint8Array(options.speakerAudio);
      if (speakerBytes && speakerBytes.length > 0) {
        const copy = new Uint8Array(speakerBytes);
        workerOptions.speakerAudio = copy;
        transferList.push(copy.buffer);
      }
      const result = await this._callWorker(
        "synthesizeSpeech",
        [workerOptions],
        (event) => {
          if (event.event === "progress") {
            options.onProgress?.(event.payload || {});
          }
        },
        transferList
      );
      this._throwIfCallerCancelled("Text-to-speech synthesis was cancelled.");
      return result;
    } catch (error) {
      this._throwIfOperationCancelled(error, "Text-to-speech synthesis was cancelled.");
      if (error?.name === "AbortError" || options.signal?.aborted || serializeWorkerError(error).toLowerCase().includes("cancel")) {
        throw new DOMException("Text-to-speech synthesis was cancelled.", "AbortError");
      }
      const canRetryOnCpu = !this._isCpuModelMode() && (this._isDispatchWorkgroupLimitError(error) || this._isWorkerRequestTimeoutError(error));
      if (canRetryOnCpu) {
        this._disableWorkerFallback(error);
        await this._waitForWorkerDisposal();
        await this._ensureRuntimeReadyAfterWorkerFallback(
          {
            _llamadartForceRuntimeReload: true,
            _llamadartTextToSpeech: true
          },
          error
        );
        if (options.signal?.aborted) {
          throw new DOMException("Text-to-speech synthesis was cancelled.", "AbortError");
        }
        return this._runtime.synthesizeSpeech(options);
      }
      if (this._isWorkerRequestTimeoutError(error)) {
        this._disableWorkerFallback(error);
        await this._waitForWorkerDisposal();
      }
      throw error;
    }
  }
  _decisionHeadMap() {
    if (!(this._decisionHeads instanceof Map)) {
      this._decisionHeads = /* @__PURE__ */ new Map();
      this._nextDecisionHandle = 1;
    }
    return this._decisionHeads;
  }
  _registerDecisionHead(owner, info) {
    const heads = this._decisionHeadMap();
    const handle = this._nextDecisionHandle++;
    heads.set(handle, { owner, handle: Number(info?.handle) });
    return { ...info, handle };
  }
  _decisionOwner() {
    return this._workerProxy || this._runtime || null;
  }
  _staleDecisionHeadError(handle) {
    return new Error(
      `Decision head ${handle} is not loaded; it was freed, its model was unloaded, or the bridge runtime restarted. Load the decision head again.`
    );
  }
  _resolveDecisionHead(handle) {
    decisionHandleFrom(handle);
    const heads = this._decisionHeadMap();
    const entry = heads.get(handle);
    if (!entry) {
      return null;
    }
    if (entry.owner !== this._decisionOwner()) {
      heads.delete(handle);
      return null;
    }
    return entry;
  }
  _requireDecisionRuntime() {
    if (!this._runtime) {
      throw new Error("No model loaded. Call loadModelFromUrl first.");
    }
    return this._runtime;
  }
  /**
   * Moves a failed worker's work to the main-thread runtime. Decision heads
   * the worker held are gone with it; the model, its projector and the applied
   * LoRA adapters are restored so later calls keep working.
   */
  async _recoverWorkerOnMainThread(error) {
    this._disableWorkerFallback(error);
    await this._waitForWorkerDisposal();
    await this._ensureRuntimeReadyAfterWorkerFallback({}, error);
  }
  async getDecisionCapabilities() {
    return this._runExclusive(
      () => this._getDecisionCapabilitiesUnlocked(),
      { kind: "decision-capabilities" }
    );
  }
  async _getDecisionCapabilitiesUnlocked() {
    if (!this._workerProxy) {
      return this._runtime ? this._runtime.getDecisionCapabilities() : { apiVersion: DECISION_API_VERSION, supported: false, reason: "WebGPU core is not initialized" };
    }
    try {
      return await this._callWorker("getDecisionCapabilities", []);
    } catch (error) {
      this._throwIfOperationCancelled(error, "Decision capability probe was cancelled.");
      if (!this._shouldFallbackToMainThread(error)) {
        throw error;
      }
      await this._recoverWorkerOnMainThread(error);
      return this._requireDecisionRuntime().getDecisionCapabilities();
    }
  }
  async loadDecisionHead(source, options = {}) {
    return this._runExclusive(
      () => this._loadDecisionHeadUnlocked(source, options),
      { kind: "decision-head-load" }
    );
  }
  async _loadDecisionHeadUnlocked(source, options = {}) {
    const loadInRuntime = async () => {
      const runtime = this._requireDecisionRuntime();
      return this._registerDecisionHead(
        runtime,
        await runtime.loadDecisionHead(source, options)
      );
    };
    if (!this._workerProxy) {
      return loadInRuntime();
    }
    const workerOptions = { configJson: options?.configJson ?? null };
    let workerSource = source;
    const transferList = [];
    if (typeof source === "string") {
      workerSource = source.length > 0 ? normalizeAbsoluteUrl(source) : source;
    } else {
      const bytes = bufferSourceBytes(source);
      if (bytes) {
        const copy = new Uint8Array(bytes);
        workerSource = copy;
        transferList.push(copy.buffer);
      }
    }
    try {
      await this._restoreWorkerModelIfMissing();
      const proxy = this._workerProxy;
      const info = await this._callWorker(
        "loadDecisionHead",
        [workerSource, workerOptions],
        (event) => {
          if (event.event === "progress") {
            options?.onProgress?.(event.payload || {});
          }
        },
        transferList
      );
      return this._registerDecisionHead(proxy, info);
    } catch (error) {
      this._throwIfOperationCancelled(error, "Decision head load was cancelled.");
      if (!this._shouldFallbackToMainThread(error)) {
        throw error;
      }
      await this._recoverWorkerOnMainThread(error);
      return loadInRuntime();
    }
  }
  async runDecision(handle, sequences) {
    return this._runExclusive(
      () => this._runDecisionUnlocked(handle, sequences),
      { kind: "decision-run" }
    );
  }
  async _runDecisionUnlocked(handle, sequences) {
    const entry = this._resolveDecisionHead(handle);
    if (!entry) {
      throw this._staleDecisionHeadError(handle);
    }
    const normalized = normalizeDecisionSequences(sequences);
    if (!this._workerProxy) {
      return this._requireDecisionRuntime().runDecision(entry.handle, normalized);
    }
    try {
      return await this._callWorker("runDecision", [entry.handle, normalized]);
    } catch (error) {
      this._throwIfOperationCancelled(error, "Decision run was cancelled.");
      if (!this._shouldFallbackToMainThread(error)) {
        throw error;
      }
      this._decisionHeadMap().delete(handle);
      await this._recoverWorkerOnMainThread(error);
      throw new Error(
        `Decision head ${handle} was lost when the bridge worker failed (${serializeWorkerError(error)}). Load the decision head again.`
      );
    }
  }
  async freeDecisionHead(handle) {
    return this._runExclusive(
      () => this._freeDecisionHeadUnlocked(handle),
      { kind: "decision-head-free" }
    );
  }
  async _freeDecisionHeadUnlocked(handle) {
    const entry = this._resolveDecisionHead(handle);
    if (!entry) {
      return;
    }
    this._decisionHeadMap().delete(handle);
    if (!this._workerProxy) {
      await this._requireDecisionRuntime().freeDecisionHead(entry.handle);
      return;
    }
    try {
      await this._callWorker("freeDecisionHead", [entry.handle]);
    } catch (error) {
      this._throwIfOperationCancelled(error, "Decision head free was cancelled.");
      if (!this._shouldFallbackToMainThread(error)) {
        throw error;
      }
      await this._recoverWorkerOnMainThread(error);
    }
  }
  // Test doubles skip the constructor, so the LoRA state starts on first use.
  _loraAdapterMap() {
    if (!(this._loraAdapters instanceof Map)) {
      this._loraAdapters = /* @__PURE__ */ new Map();
      this._activeLoraScales = /* @__PURE__ */ new Map();
      this._nextLoraHandle = 1;
    }
    return this._loraAdapters;
  }
  _activeLoraScaleMap() {
    this._loraAdapterMap();
    return this._activeLoraScales;
  }
  _forgetLoraAdapters() {
    this._loraAdapterMap().clear();
    this._activeLoraScaleMap().clear();
  }
  /**
   * Marks the adapters `owner` holds as lost before it loads the model again,
   * which frees them. An adapter with a retained source reloads on the next
   * restore or set; one without is forgotten.
   */
  _dropLoraAdaptersOf(owner) {
    for (const [handle, entry] of this._loraAdapterMap()) {
      if (owner === null || entry.owner !== owner) {
        continue;
      }
      if (entry.source === null) {
        this._loraAdapterMap().delete(handle);
        this._activeLoraScaleMap().delete(handle);
      } else {
        entry.owner = null;
      }
    }
  }
  _loraOwner() {
    return this._workerProxy || this._runtime || null;
  }
  _requireLoraRuntime() {
    if (!this._runtime) {
      throw new Error("No model loaded. Call loadModelFromUrl first.");
    }
    return this._runtime;
  }
  _loraAdapterEntry(handle) {
    const entry = this._loraAdapterMap().get(handle);
    if (!entry) {
      throw staleLoraAdapterError(handle);
    }
    return entry;
  }
  // Loads an adapter into the current owner: the worker, or the direct runtime.
  async _loadLoraAdapterIntoOwner(source, options) {
    if (!this._workerProxy) {
      const runtime = this._requireLoraRuntime();
      const info2 = await runtime.loadLoraAdapter(source, {
        ...options,
        signal: this._operationSignal() || options.signal || null
      });
      return { owner: runtime, handle: info2.handle };
    }
    const workerOptions = {};
    if (options.useCache !== void 0) {
      workerOptions.useCache = options.useCache;
    }
    if (options.cacheName !== void 0) {
      workerOptions.cacheName = options.cacheName;
    }
    let workerSource = source;
    const transferList = [];
    if (typeof source !== "string") {
      workerSource = new Uint8Array(source);
      transferList.push(workerSource.buffer);
    }
    const proxy = this._workerProxy;
    const info = await this._callWorker(
      "loadLoraAdapter",
      [workerSource, workerOptions],
      (event) => {
        if (event.event === "progress") {
          options.progressCallback?.(event.payload || {});
        }
      },
      transferList
    );
    return { owner: proxy, handle: Number(info?.handle) };
  }
  // Reloads an adapter its owner lost, into the current owner.
  async _ensureLoraAdapterOnOwner(handle, entry) {
    if (entry.owner === this._loraOwner()) {
      return;
    }
    if (entry.source === null) {
      this._loraAdapterMap().delete(handle);
      this._activeLoraScaleMap().delete(handle);
      throw staleLoraAdapterError(handle);
    }
    const loaded = await this._loadLoraAdapterIntoOwner(entry.source, entry.cacheOptions);
    entry.owner = loaded.owner;
    entry.handle = loaded.handle;
  }
  async _callLoraOwner(method, args) {
    if (this._workerProxy) {
      await this._callWorker(method, args);
      return;
    }
    const runtime = this._requireLoraRuntime();
    if (method === "setLoraAdapter") {
      await runtime.setLoraAdapter(args[0], args[1]);
    } else if (method === "removeLoraAdapter") {
      await runtime.removeLoraAdapter(args[0]);
    } else {
      await runtime.clearLoraAdapters();
    }
  }
  /**
   * Makes the current owner apply exactly the facade's active adapters after a
   * worker restart or a fallback to the main thread. The owner's own set is
   * cleared first, since a runtime that held the model earlier may still apply
   * adapters from then. Inactive adapters reload when next set.
   */
  async _restoreLoraAdapters() {
    const owner = this._loraOwner();
    if (![...this._loraAdapterMap().values()].some((entry) => entry.owner !== owner)) {
      return;
    }
    await this._callLoraOwner("clearLoraAdapters", []);
    for (const [handle, scale] of [...this._activeLoraScaleMap()]) {
      const entry = this._loraAdapterEntry(handle);
      await this._ensureLoraAdapterOnOwner(handle, entry);
      await this._callLoraOwner("setLoraAdapter", [entry.handle, scale]);
    }
  }
  // Runs a LoRA call on the current owner, moving to the main-thread runtime
  // and running it there once if the worker fails.
  async _runLoraCall(call, cancelMessage) {
    if (!this._workerProxy) {
      return call();
    }
    try {
      await this._restoreWorkerModelIfMissing();
      return await call();
    } catch (error) {
      this._throwIfOperationCancelled(error, cancelMessage);
      if (!this._shouldFallbackToMainThread(error)) {
        throw error;
      }
      await this._recoverWorkerOnMainThread(error);
      return call();
    }
  }
  async getLoraAdapterCapabilities() {
    return this._runExclusive(
      () => this._getLoraAdapterCapabilitiesUnlocked(),
      { kind: "lora-capabilities" }
    );
  }
  async _getLoraAdapterCapabilitiesUnlocked() {
    if (!this._workerProxy && !this._runtime) {
      return { apiVersion: LORA_API_VERSION, supported: false, reason: "WebGPU core is not initialized" };
    }
    return this._runLoraCall(
      () => this._workerProxy ? this._callWorker("getLoraAdapterCapabilities", []) : Promise.resolve(this._requireLoraRuntime().getLoraAdapterCapabilities()),
      "LoRA capability probe was cancelled."
    );
  }
  async loadLoraAdapter(source, options = {}) {
    return this._runExclusive(
      () => this._loadLoraAdapterUnlocked(source, options || {}),
      {
        signal: options?.signal,
        abortMessage: LORA_LOAD_ABORT_MESSAGE,
        kind: "lora-adapter-load"
      }
    );
  }
  async _loadLoraAdapterUnlocked(source, options) {
    let normalized;
    if (typeof source === "string") {
      if (source.length === 0) {
        throw new Error("LoRA adapter URL is empty.");
      }
      normalized = normalizeAbsoluteUrl(source);
    } else {
      const bytes = bufferSourceBytes(source);
      if (!bytes) {
        throw new TypeError("LoRA adapter source must be a URL string, an ArrayBuffer or a typed array.");
      }
      if (bytes.byteLength === 0) {
        throw new Error("LoRA adapter bytes are empty.");
      }
      normalized = bytes;
    }
    let retained = null;
    if (this._config?.disableWorker !== true) {
      retained = typeof normalized === "string" ? normalized : new Uint8Array(normalized);
    }
    const loaded = await this._runLoraCall(
      () => this._loadLoraAdapterIntoOwner(retained ?? normalized, options),
      LORA_LOAD_ABORT_MESSAGE
    );
    this._loraAdapterMap();
    const handle = this._nextLoraHandle++;
    this._loraAdapterMap().set(handle, {
      owner: loaded.owner,
      handle: loaded.handle,
      source: retained,
      cacheOptions: { useCache: options.useCache, cacheName: options.cacheName }
    });
    return { handle };
  }
  async setLoraAdapter(handle, scale = 1) {
    const facadeHandle = loraHandleFrom(handle);
    const checkedScale = loraScaleFrom(scale);
    return this._runExclusive(
      () => this._setLoraAdapterUnlocked(facadeHandle, checkedScale),
      { kind: "lora-adapter-set" }
    );
  }
  async _setLoraAdapterUnlocked(handle, scale) {
    const entry = this._loraAdapterEntry(handle);
    await this._runLoraCall(async () => {
      await this._ensureLoraAdapterOnOwner(handle, entry);
      await this._callLoraOwner("setLoraAdapter", [entry.handle, scale]);
    }, "LoRA adapter update was cancelled.");
    this._activeLoraScaleMap().set(handle, scale);
  }
  async removeLoraAdapter(handle) {
    const facadeHandle = loraHandleFrom(handle);
    return this._runExclusive(
      () => this._removeLoraAdapterUnlocked(facadeHandle),
      { kind: "lora-adapter-remove" }
    );
  }
  async _removeLoraAdapterUnlocked(handle) {
    const entry = this._loraAdapterEntry(handle);
    await this._runLoraCall(async () => {
      if (entry.owner === this._loraOwner()) {
        await this._callLoraOwner("removeLoraAdapter", [entry.handle]);
      }
    }, "LoRA adapter update was cancelled.");
    this._activeLoraScaleMap().delete(handle);
  }
  async clearLoraAdapters() {
    return this._runExclusive(
      () => this._clearLoraAdaptersUnlocked(),
      { kind: "lora-adapters-clear" }
    );
  }
  async _clearLoraAdaptersUnlocked() {
    await this._runLoraCall(
      () => this._callLoraOwner("clearLoraAdapters", []),
      "LoRA adapter update was cancelled."
    );
    this._activeLoraScaleMap().clear();
  }
  async tokenize(text, addSpecial = true) {
    return this._runExclusive(
      () => this._tokenizeUnlocked(text, addSpecial),
      { kind: "tokenize" }
    );
  }
  async _tokenizeUnlocked(text, addSpecial = true) {
    if (!this._workerProxy) {
      return this._runtime.tokenize(text, addSpecial);
    }
    try {
      await this._restoreWorkerModelIfMissing();
      return await this._callWorker("tokenize", [text, addSpecial]);
    } catch (error) {
      this._throwIfOperationCancelled(error, "Tokenization was cancelled.");
      if (!this._isWorkerUnusableError(error)) {
        throw error;
      }
      this._disableWorkerFallback(error);
      await this._waitForWorkerDisposal();
      await this._ensureRuntimeReadyAfterWorkerFallback({}, error);
      return this._runtime.tokenize(text, addSpecial);
    }
  }
  async stateSaveFile(path, tokens = []) {
    return this._runExclusive(
      () => this._stateSaveFileUnlocked(path, tokens),
      { kind: "state-save-file" }
    );
  }
  async _stateSaveFileUnlocked(path, tokens = []) {
    if (!this._workerProxy) {
      return this._runtime.stateSaveFile(path, tokens);
    }
    const normalized = Array.isArray(tokens) ? tokens : Array.from(tokens || []);
    await this._restoreWorkerModelIfMissing();
    return this._callWorker("stateSaveFile", [path, normalized]);
  }
  // The capacity default is resolved inside the slot: evaluating it as a
  // parameter default would read runtime state before the queue was taken.
  async stateLoadFile(path, tokenCapacity = void 0) {
    return this._runExclusive(
      () => this._stateLoadFileUnlocked(path, tokenCapacity),
      { kind: "state-load-file" }
    );
  }
  _resolveTokenCapacity(tokenCapacity) {
    let numericCapacity;
    try {
      numericCapacity = Number(tokenCapacity);
    } catch {
      return this.getContextSize();
    }
    return numericCapacity > 0 ? Math.trunc(numericCapacity) : this.getContextSize();
  }
  async _stateLoadFileUnlocked(path, tokenCapacity) {
    tokenCapacity = this._resolveTokenCapacity(tokenCapacity);
    if (!this._workerProxy) {
      return this._runtime.stateLoadFile(path, tokenCapacity);
    }
    await this._restoreWorkerModelIfMissing();
    return this._callWorker("stateLoadFile", [path, tokenCapacity]);
  }
  async stateSaveBytes(tokens = []) {
    return this._runExclusive(
      () => this._stateSaveBytesUnlocked(tokens),
      { kind: "state-save-bytes" }
    );
  }
  async _stateSaveBytesUnlocked(tokens = []) {
    if (!this._workerProxy) {
      return this._runtime.stateSaveBytes(tokens);
    }
    const normalized = Array.isArray(tokens) ? tokens : Array.from(tokens || []);
    await this._restoreWorkerModelIfMissing();
    return this._callWorker("stateSaveBytes", [normalized]);
  }
  async stateLoadBytes(bytes, tokenCapacity = void 0) {
    return this._runExclusive(
      () => this._stateLoadBytesUnlocked(bytes, tokenCapacity),
      { kind: "state-load-bytes" }
    );
  }
  async _stateLoadBytesUnlocked(bytes, tokenCapacity) {
    tokenCapacity = this._resolveTokenCapacity(tokenCapacity);
    if (!this._workerProxy) {
      return this._runtime.stateLoadBytes(bytes, tokenCapacity);
    }
    const normalizedBytes = toUint8Array(bytes);
    if (!normalizedBytes || normalizedBytes.length === 0) {
      throw new Error("State bytes are empty.");
    }
    await this._restoreWorkerModelIfMissing();
    const transferableBytes = new Uint8Array(normalizedBytes);
    return this._callWorker(
      "stateLoadBytes",
      [transferableBytes, tokenCapacity],
      null,
      [transferableBytes.buffer]
    );
  }
  async detokenize(tokens, special = false) {
    return this._runExclusive(
      () => this._detokenizeUnlocked(tokens, special),
      { kind: "detokenize" }
    );
  }
  async _detokenizeUnlocked(tokens, special = false) {
    if (!this._workerProxy) {
      return this._runtime.detokenize(tokens, special);
    }
    const normalized = Array.isArray(tokens) ? tokens : Array.from(tokens || []);
    try {
      await this._restoreWorkerModelIfMissing();
      return await this._callWorker("detokenize", [normalized, special]);
    } catch (error) {
      this._throwIfOperationCancelled(error, "Detokenization was cancelled.");
      if (!this._isWorkerUnusableError(error)) {
        throw error;
      }
      this._disableWorkerFallback(error);
      await this._waitForWorkerDisposal();
      await this._ensureRuntimeReadyAfterWorkerFallback({}, error);
      return this._runtime.detokenize(normalized, special);
    }
  }
  async getEmbeddingCapabilities() {
    return this._runExclusive(async () => {
      try {
        return await this._getEmbeddingCapabilitiesUnlocked();
      } catch (error) {
        this._throwIfOperationCancelled(error, "Embedding capability probe was cancelled.");
        if (!this._workerProxy || !this._isWorkerUnusableError(error)) throw error;
        await this._recoverEmbeddingWorker(error);
        return this._runtime.getEmbeddingCapabilities();
      }
    }, { kind: "embedding-capabilities" });
  }
  async _getEmbeddingCapabilitiesUnlocked() {
    if (!this._workerProxy) {
      await this._ensureRememberedEmbeddingProjector();
      return this._runtime.getEmbeddingCapabilities();
    }
    try {
      await this._restoreWorkerModelIfMissing();
      return embeddingCapabilitiesFrom(
        await this._callWorker("getEmbeddingCapabilities", []),
        embeddingAssetVersion(this._config.workerUrl)
      );
    } catch (error) {
      this._throwIfOperationCancelled(error, "Embedding capability probe was cancelled.");
      if (error instanceof Error && /bridge\[method\] is not a function/.test(error.message)) {
        throw missingEmbeddingCapability(embeddingAssetVersion(this._config.workerUrl));
      }
      throw error;
    }
  }
  async embed(text, options = {}) {
    const input = typeof text === "string" ? text : snapshotEmbedInput(text);
    return this._runExclusive(
      () => this._embedUnlocked(input, options),
      { kind: "embedding" }
    );
  }
  _hasEmbeddingMediaParts(input) {
    return Array.isArray(input) && input.some((part) => part.type !== "text");
  }
  async _recoverEmbeddingWorker(error) {
    this._disableWorkerFallback(error);
    await this._waitForWorkerDisposal();
    await this._ensureRuntimeReadyAfterWorkerFallback({}, error);
    await this._ensureRememberedEmbeddingProjector();
  }
  async _ensureRememberedEmbeddingProjector() {
    if (!this._loadedMmProjUrl || !this._runtime?._modelBytes || this._runtime._mmProjPath) return;
    this._throwIfDisposed();
    try {
      await this._runtime.loadMultimodalProjector(this._loadedMmProjUrl);
    } catch (error) {
      this._throwIfOperationCancelled(error, "Embedding projector restore was cancelled.");
      this._emitBridgeWarn("llamadart: embedding projector restore failed; keeping the loaded model.");
    }
    this._throwIfDisposed();
  }
  async _embedUnlocked(text, options = {}) {
    if (!this._workerProxy) {
      if (this._hasEmbeddingMediaParts(text)) await this._ensureRememberedEmbeddingProjector();
      return this._runtime.embed(text, options);
    }
    try {
      await this._restoreWorkerModelIfMissing();
      if (Array.isArray(text)) await this._getEmbeddingCapabilitiesUnlocked();
      return await this._callWorker("embed", [text, options]);
    } catch (error) {
      this._throwIfOperationCancelled(error, "Embedding was cancelled.");
      if (!this._isWorkerUnusableError(error)) {
        throw error;
      }
      await this._recoverEmbeddingWorker(error);
      return this._runtime.embed(text, options);
    }
  }
  async embedBatch(texts, options = {}) {
    const inputs = Array.from(texts || [], (input) => typeof input === "string" ? input : snapshotEmbedInput(input));
    return this._runExclusive(
      () => this._embedBatchUnlocked(inputs, options),
      { kind: "embedding-batch" }
    );
  }
  async _embedBatchUnlocked(texts, options = {}) {
    const normalized = Array.isArray(texts) ? texts : Array.from(texts || []);
    if (!this._workerProxy) {
      if (normalized.some((input) => this._hasEmbeddingMediaParts(input))) await this._ensureRememberedEmbeddingProjector();
      return this._runtime.embedBatch(normalized, options);
    }
    try {
      await this._restoreWorkerModelIfMissing();
      if (normalized.some(Array.isArray)) await this._getEmbeddingCapabilitiesUnlocked();
      return await this._callWorker("embedBatch", [normalized, options]);
    } catch (error) {
      this._throwIfOperationCancelled(error, "Batch embedding was cancelled.");
      if (!this._isWorkerUnusableError(error)) {
        throw error;
      }
      await this._recoverEmbeddingWorker(error);
      return this._runtime.embedBatch(normalized, options);
    }
  }
  async scoreNextToken(prompt, options = {}) {
    return this._runExclusive(
      () => this._scoreNextTokenUnlocked(prompt, options),
      { kind: "next-token-scoring" }
    );
  }
  async _scoreNextTokenUnlocked(prompt, options = {}) {
    if (!this._workerProxy) {
      return this._runtime.scoreNextToken(prompt, options);
    }
    try {
      await this._restoreWorkerModelIfMissing();
      return await this._callWorker("scoreNextToken", [prompt, options]);
    } catch (error) {
      this._throwIfOperationCancelled(error, "Next-token scoring was cancelled.");
      if (!this._isWorkerUnusableError(error)) {
        throw error;
      }
      this._disableWorkerFallback(error);
      await this._waitForWorkerDisposal();
      await this._ensureRuntimeReadyAfterWorkerFallback({}, error);
      return this._runtime.scoreNextToken(prompt, options);
    }
  }
  getModelMetadata() {
    if (this._workerProxy) {
      return {
        ...this._metadata || {},
        "llamadart.webgpu.execution": "worker"
      };
    }
    const workerReason = typeof this._workerFallbackReason === "string" && this._workerFallbackReason.length > 0 ? this._workerFallbackReason : null;
    return {
      ...this._metadata || {},
      "llamadart.webgpu.execution": "main-thread",
      ...workerReason == null ? {} : { "llamadart.webgpu.worker_fallback_reason": workerReason }
    };
  }
  // The synchronous getters answer from the facade snapshot in both modes.
  // Reading the direct runtime here would ccall into the singleton core outside
  // the queue and race whichever async operation currently owns it; the snapshot
  // is refreshed by _captureDirectRuntimeState() while the slot is held.
  getContextSize() {
    return this._contextSize || 0;
  }
  isGpuActive() {
    return this._gpuActive;
  }
  getBackendName() {
    return this._backendName;
  }
  setLogLevel(level) {
    this._throwIfDisposed();
    if (Number.isFinite(level)) {
      this._config.logLevel = Math.max(0, Math.min(4, Math.trunc(level)));
    }
    if (this._workerProxy) {
      this._callWorker("setLogLevel", [level]).catch(() => {
      });
      if (this._runtime) {
        this._runtime.setLogLevel(level);
      }
      return;
    }
    this._runtime.setLogLevel(level);
  }
  cancel() {
    if (this._lifecycleState === "disposing" || this._lifecycleState === "disposed") {
      return;
    }
    if (this._activeOperation) {
      this._cancelOperation(this._activeOperation);
      return;
    }
    if (this._workerProxy) {
      this._sendWorkerCancel(this._workerProxy);
    }
    if (this._runtime) {
      try {
        this._runtime.cancel();
      } catch (_) {
      }
    }
  }
  /**
   * Marks the bridge disposed synchronously, then tears down on the queue so
   * teardown never runs underneath an operation that still owns the runtime.
   * Work queued before disposal is rejected with a stable lifecycle error rather
   * than dereferencing a torn-down runtime. Not `async`, so repeated calls hand
   * back the identical teardown promise.
   */
  dispose() {
    if (this._disposed) {
      return this._disposePromise;
    }
    this._disposed = true;
    this._lifecycleState = "disposing";
    let resolveDispose = () => {
    };
    let rejectDispose = () => {
    };
    this._disposePromise = new Promise((resolve, reject) => {
      resolveDispose = resolve;
      rejectDispose = reject;
    });
    this._notifyDisposalWaiters();
    const teardown = this._runExclusive(
      () => this._disposeUnlocked(),
      { allowDisposed: true, kind: "dispose" }
    );
    teardown.then(
      () => resolveDispose(),
      (error) => rejectDispose(error)
    );
    return this._disposePromise;
  }
  async _disposeUnlocked() {
    const workerProxy = this._workerProxy;
    this._workerProxy = null;
    if (workerProxy) {
      this._retireWorkerProxy(workerProxy);
    }
    try {
      if (this._retiringWorkerDisposals?.size > 0) {
        await this._waitForWorkerDisposal();
      }
      if (this._runtime) {
        const runtime = this._runtime;
        this._runtime = null;
        await runtime.dispose();
      }
    } finally {
      this._metadata = {};
      this._contextSize = 0;
      this._gpuActive = false;
      this._backendName = "WASM (Prototype bridge)";
      this._supportsVision = false;
      this._supportsAudio = false;
      this._loadedModelUrl = null;
      this._loadedModelOptions = null;
      this._loadedMmProjUrl = null;
      this._loadedDraftModel = null;
      this._workerFallbackReason = null;
      this._multimodalWorkerCpuMode = false;
      this._workerModelMissing = false;
      this._decisionHeads?.clear();
      this._forgetLoraAdapters();
      this._lifecycleState = "disposed";
    }
  }
  async applyChatTemplate(messages, addAssistant = true, customTemplate = null) {
    return this._runExclusive(
      () => this._applyChatTemplateUnlocked(messages, addAssistant, customTemplate),
      { kind: "template" }
    );
  }
  async _applyChatTemplateUnlocked(messages, addAssistant = true, customTemplate = null) {
    if (!this._workerProxy) {
      return this._runtime.applyChatTemplate(messages, addAssistant, customTemplate);
    }
    try {
      return await this._callWorker("applyChatTemplate", [messages, addAssistant, customTemplate]);
    } catch (error) {
      this._throwIfOperationCancelled(error, "Chat template operation was cancelled.");
      if (!this._isWorkerUnusableError(error)) {
        throw error;
      }
      this._disableWorkerFallback(error);
      return this._runtime.applyChatTemplate(messages, addAssistant, customTemplate);
    }
  }
};

// js/src/worker_host.ts
var textDecoder = new TextDecoder();
var bridgeWorkerHostInstalled = false;
function installBridgeWorkerHost() {
  if (bridgeWorkerHostInstalled) {
    return;
  }
  if (typeof self === "undefined" || typeof self.postMessage !== "function") {
    throw new Error("Bridge worker host can only run inside a worker context");
  }
  bridgeWorkerHostInstalled = true;
  let bridge = null;
  const postError = (id, error) => {
    let state;
    try {
      state = bridge ? snapshotBridgeState(bridge) : void 0;
    } catch (_) {
      state = void 0;
    }
    self.postMessage({
      type: "error",
      id,
      message: serializeWorkerError(error),
      ...error instanceof EmbeddingError ? { embeddingErrorCode: error.code } : {},
      state
    });
  };
  self.onmessage = async (event) => {
    const message = event.data || {};
    const type = message.type;
    const id = message.id ?? 0;
    try {
      if (type === "init") {
        bridge = new LlamaWebGpuBridge({
          ...message.config || {},
          disableWorker: true
        });
        self.postMessage({ type: "ready" });
        return;
      }
      if (type !== "call") {
        return;
      }
      if (!bridge) {
        throw new Error("Bridge worker is not initialized");
      }
      const method = String(message.method || "");
      const args = Array.isArray(message.args) ? message.args : [];
      if (method === "loadModelFromUrl") {
        const url = args[0];
        const options = args[1] && typeof args[1] === "object" ? { ...args[1] } : {};
        options.progressCallback = (progress) => {
          self.postMessage({ type: "event", id, event: "progress", payload: progress || {} });
        };
        const value2 = await bridge.loadModelFromUrl(url, options);
        self.postMessage({ type: "result", id, value: value2, state: snapshotBridgeState(bridge) });
        return;
      }
      if (method === "loadDraftModel") {
        const options = args[1] && typeof args[1] === "object" ? { ...args[1] } : {};
        delete options.signal;
        options.progressCallback = (progress) => {
          self.postMessage({ type: "event", id, event: "progress", payload: progress || {} });
        };
        const value2 = await bridge.loadDraftModel(args[0], options);
        self.postMessage({ type: "result", id, value: value2 });
        return;
      }
      if (method === "createCompletion") {
        const prompt = args[0];
        const options = args[1] && typeof args[1] === "object" ? { ...args[1] } : {};
        delete options.signal;
        let usage = null;
        options.onUsage = (value3) => {
          usage = value3;
        };
        const tokenEventEncoding = typeof options.tokenEventEncoding === "string" ? String(options.tokenEventEncoding || "").toLowerCase() : "bytes";
        const flushMsRaw = Number(options.tokenEventFlushMs);
        const tokenEventFlushMs = Number.isFinite(flushMsRaw) && flushMsRaw >= 0 ? Math.max(0, Math.min(200, Math.trunc(flushMsRaw))) : 0;
        const flushCharsRaw = Number(options.tokenEventFlushChars);
        const tokenEventFlushChars = Number.isFinite(flushCharsRaw) && flushCharsRaw > 0 ? Math.max(1, Math.min(1024, Math.trunc(flushCharsRaw))) : 0;
        const shouldEmitCurrentText = options.emitCurrentTextOnToken === true;
        let pendingPieceText = "";
        let pendingPieceBytes = [];
        let pendingPieceByteLength = 0;
        let pendingPieceCharLength = 0;
        let pendingPieceCharDecoder = tokenEventFlushChars > 0 ? new TextDecoder() : null;
        let pendingCurrentText = "";
        let flushTimer = null;
        const flushTokenTextPayload = () => {
          if (pendingPieceText.length === 0) {
            return;
          }
          self.postMessage({
            type: "event",
            id,
            event: "token",
            payload: {
              pieceText: pendingPieceText,
              currentText: shouldEmitCurrentText ? pendingCurrentText : ""
            }
          });
          pendingPieceText = "";
          pendingCurrentText = "";
        };
        const flushTokenBytePayload = () => {
          if (pendingPieceByteLength === 0) {
            return;
          }
          const piece = new Uint8Array(pendingPieceByteLength);
          let offset = 0;
          for (const chunk of pendingPieceBytes) {
            piece.set(chunk, offset);
            offset += chunk.byteLength;
          }
          self.postMessage({
            type: "event",
            id,
            event: "token",
            payload: {
              piece: Array.from(piece),
              currentText: shouldEmitCurrentText ? pendingCurrentText : ""
            }
          });
          pendingPieceBytes = [];
          pendingPieceByteLength = 0;
          pendingPieceCharLength = 0;
          pendingCurrentText = "";
        };
        const flushTokenPayload = () => {
          if (tokenEventEncoding === "text") {
            flushTokenTextPayload();
            return;
          }
          flushTokenBytePayload();
        };
        const scheduleTokenFlush = () => {
          if (tokenEventFlushMs <= 0 || flushTimer != null) {
            return;
          }
          flushTimer = globalThis.setTimeout(() => {
            flushTimer = null;
            flushTokenPayload();
          }, tokenEventFlushMs);
        };
        options.onToken = (piece, currentText) => {
          if (tokenEventEncoding === "text") {
            const pieceText = typeof piece === "string" ? piece : textDecoder.decode(toUint8Array(piece) || new Uint8Array());
            if (pieceText.length === 0) {
              return;
            }
            if (tokenEventFlushMs > 0) {
              pendingPieceText += pieceText;
              if (shouldEmitCurrentText) {
                pendingCurrentText = String(currentText || "");
              }
              if (tokenEventFlushChars > 0 && pendingPieceText.length >= tokenEventFlushChars) {
                if (flushTimer != null) {
                  globalThis.clearTimeout(flushTimer);
                  flushTimer = null;
                }
                flushTokenTextPayload();
                return;
              }
              scheduleTokenFlush();
              return;
            }
            self.postMessage({
              type: "event",
              id,
              event: "token",
              payload: {
                pieceText,
                currentText: shouldEmitCurrentText ? String(currentText || "") : ""
              }
            });
            return;
          }
          const normalizedPieceBytes = toUint8Array(piece) || new Uint8Array();
          if (normalizedPieceBytes.byteLength === 0) {
            return;
          }
          if (tokenEventFlushMs > 0) {
            const pieceBytes = Array.isArray(piece) ? normalizedPieceBytes : Uint8Array.from(normalizedPieceBytes);
            pendingPieceBytes.push(pieceBytes);
            pendingPieceByteLength += pieceBytes.byteLength;
            if (pendingPieceCharDecoder != null) {
              pendingPieceCharLength += pendingPieceCharDecoder.decode(
                pieceBytes,
                { stream: true }
              ).length;
            }
            if (shouldEmitCurrentText) {
              pendingCurrentText = String(currentText || "");
            }
            if (tokenEventFlushChars > 0 && pendingPieceCharLength >= tokenEventFlushChars) {
              if (flushTimer != null) {
                globalThis.clearTimeout(flushTimer);
                flushTimer = null;
              }
              flushTokenBytePayload();
              return;
            }
            scheduleTokenFlush();
            return;
          }
          self.postMessage({
            type: "event",
            id,
            event: "token",
            payload: {
              piece: Array.from(normalizedPieceBytes),
              currentText: shouldEmitCurrentText ? String(currentText || "") : ""
            }
          });
        };
        let value2;
        try {
          value2 = await bridge.createCompletion(prompt, options);
        } finally {
          if (flushTimer != null) {
            globalThis.clearTimeout(flushTimer);
            flushTimer = null;
          }
          flushTokenPayload();
        }
        flushTokenTextPayload();
        if (usage != null) {
          self.postMessage({ type: "event", id, event: "usage", payload: usage });
        }
        self.postMessage({ type: "result", id, value: value2, state: snapshotBridgeState(bridge) });
        return;
      }
      if (method === "loadMultimodalProjector") {
        const value2 = await bridge.loadMultimodalProjector(args[0]);
        self.postMessage({ type: "result", id, value: value2, state: snapshotBridgeState(bridge) });
        return;
      }
      if (method === "synthesizeSpeech") {
        const options = args[0] && typeof args[0] === "object" ? { ...args[0] } : {};
        delete options.signal;
        options.onProgress = (progress) => {
          self.postMessage({ type: "event", id, event: "progress", payload: progress || {} });
        };
        const value2 = await bridge.synthesizeSpeech(options);
        const transfers = value2?.pcm?.buffer instanceof ArrayBuffer ? [value2.pcm.buffer] : [];
        self.postMessage(
          { type: "result", id, value: value2, state: snapshotBridgeState(bridge) },
          transfers
        );
        return;
      }
      if (method === "loadDecisionHead") {
        const options = args[1] && typeof args[1] === "object" ? { ...args[1] } : {};
        options.onProgress = (progress) => {
          self.postMessage({ type: "event", id, event: "progress", payload: progress || {} });
        };
        const value2 = await bridge.loadDecisionHead(args[0], options);
        self.postMessage({ type: "result", id, value: value2 });
        return;
      }
      if (method === "loadLoraAdapter") {
        const options = args[1] && typeof args[1] === "object" ? { ...args[1] } : {};
        delete options.signal;
        options.progressCallback = (progress) => {
          self.postMessage({ type: "event", id, event: "progress", payload: progress || {} });
        };
        const value2 = await bridge.loadLoraAdapter(args[0], options);
        self.postMessage({ type: "result", id, value: value2 });
        return;
      }
      if (method === "runDecision") {
        const value2 = await bridge.runDecision(args[0], args[1]);
        const transfers = [];
        for (const output of Array.isArray(value2) ? value2 : []) {
          for (const values of [output?.logits, output?.actLogits]) {
            const buffer = values?.buffer;
            if (buffer instanceof ArrayBuffer && !transfers.includes(buffer)) {
              transfers.push(buffer);
            }
          }
        }
        self.postMessage({ type: "result", id, value: value2 }, transfers);
        return;
      }
      if (method === "unloadMultimodalProjector") {
        const value2 = await bridge.unloadMultimodalProjector();
        self.postMessage({ type: "result", id, value: value2, state: snapshotBridgeState(bridge) });
        return;
      }
      if (method === "stateSaveBytes") {
        const value2 = await bridge.stateSaveBytes(args[0]);
        const transfers = value2 && value2.buffer instanceof ArrayBuffer ? [value2.buffer] : [];
        self.postMessage({ type: "result", id, value: value2 }, transfers);
        return;
      }
      if (method === "stateLoadBytes") {
        const value2 = await bridge.stateLoadBytes(args[0], args[1]);
        self.postMessage({ type: "result", id, value: value2 });
        return;
      }
      if (method === "dispose") {
        const value2 = await bridge.dispose();
        self.postMessage({
          type: "result",
          id,
          value: value2,
          state: emptyBridgeState()
        });
        return;
      }
      const value = await bridge[method](
        ...args || []
      );
      self.postMessage({ type: "result", id, value });
    } catch (error) {
      postError(id, error);
    }
  };
}
function shouldAutoBootBridgeWorkerHost() {
  if (typeof WorkerGlobalScope === "undefined" || !(globalThis instanceof WorkerGlobalScope)) {
    return false;
  }
  try {
    const href = String(globalThis.location?.href || "");
    if (!href) {
      return false;
    }
    const url = new URL(href);
    return url.searchParams.get(bridgeWorkerModeParam) === "1";
  } catch (_) {
    return false;
  }
}
function enableBridgeWorkerHost() {
  installBridgeWorkerHost();
}

// js/src/llama_webgpu_bridge.js
if (shouldAutoBootBridgeWorkerHost()) {
  installBridgeWorkerHost();
}
if (typeof window !== "undefined") {
  const browserWindow = (
    /** @type {Window & typeof globalThis & { LlamaWebGpuBridge?: typeof LlamaWebGpuBridge }} */
    window
  );
  if (!browserWindow.LlamaWebGpuBridge) {
    browserWindow.LlamaWebGpuBridge = LlamaWebGpuBridge;
  }
}
export {
  EmbeddingError,
  LlamaWebGpuBridge,
  enableBridgeWorkerHost
};