@missionsquad/browserai 0.1.0 → 0.2.0

package/src/multimodal.ts

@@ -1,3 +1,4 @@
+import type { TdtModel } from "./tdt/index.js";
 import { type ModelPreset, type MultimodalRuntime, SUPERTONIC_2_MODEL_ID } from "./models.js";
 import { AUDIO_SR } from "./audio/pcm.js";
 
@@ -73,6 +74,7 @@ export type ProgressCallback = (info: { status?: string; progress?: number; file
 export type GenStats = { completionTokens: number; firstTokenMs?: number; startedMs: number; endedMs: number };
 /** VLM / Gemma generation result: text plus telemetry. */
 export type VlmResult = { text: string } & GenStats;
+export type ImageClassificationResult = { scores: Array<{ label: string; score: number }>; startedMs: number; endedMs: number };
 export type SttResult = { text: string; segments: Array<{ start: number; end: number; text: string }>; gen?: GenStats };
 
 /** Count tokens from a TextStreamer token_callback payload (array / typed array / tensor). */
@@ -93,10 +95,11 @@ function makeCountingStreamer(
   tokenizer: unknown,
   onDelta: ((text: string) => void) | undefined,
   sink: { text: string; completionTokens: number; firstTokenMs?: number },
+  preserveSpecialTokens = false,
 ): unknown {
   return new Streamer(tokenizer, {
     skip_prompt: true,
-    skip_special_tokens: true,
+    skip_special_tokens: !preserveSpecialTokens,
     callback_function: (chunk: unknown) => {
       if (typeof chunk === "string") {
         if (sink.firstTokenMs === undefined) sink.firstTokenMs = performance.now();
@@ -112,6 +115,8 @@ function makeCountingStreamer(
 
 /** A loaded multimodal runtime, discriminated by kind. */
 export type MultimodalHandle =
+  | { kind: "tdt"; model: TdtModel }
+  | { kind: "image-classification"; classifier: ((src: string, opts: AnyRecord) => Promise<unknown>) & { dispose?: () => Promise<void> } }
   | { kind: "stt"; transcriber: (input: unknown, opts?: AnyRecord) => Promise<any> }
   | { kind: "tts-kokoro"; tts: KokoroInstance; TextSplitterStream: new () => TtsTextSplitter }
   | { kind: "tts-pipeline"; synth: (text: string, opts: AnyRecord) => Promise<RawAudio>; modelId: string }
@@ -126,6 +131,7 @@ export type MultimodalHandle =
       /** SmolVLM passes the image as [image] + a trailing {}; others pass a bare image. */
       imageArray: boolean;
       maxNewTokens: number;
+      visionOptions?: ModelPreset["visionOptions"];
     }
   | {
       kind: "gemma";
@@ -169,7 +175,10 @@ export type LoadMultimodalOptions = {
 };
 
 function vlmDtype(preset: ModelPreset, runtime: MultimodalRuntime): unknown {
-  if (runtime === "vlm-vision2seq") return { embed_tokens: "fp16", vision_encoder: "q4", decoder_model_merged: "q4" };
+  if (runtime === "vlm-vision2seq") {
+    if (preset.visionOptions?.outputFormat === "doctags") return preset.transformersDtype ?? "fp32";
+    return { embed_tokens: "fp16", vision_encoder: "q4", decoder_model_merged: "q4" };
+  }
   // imagetext family: GLM-OCR ships a single q4f16 dtype; Qwen3-VL uses a per-module mixed map.
   if (preset.transformersDtype) return preset.transformersDtype;
   return { embed_tokens: "fp16", vision_encoder: "fp16", decoder_model_merged: "q4f16" };
@@ -204,6 +213,14 @@ export async function loadMultimodal(
   if (!runtime) throw new Error(`Model ${preset.shortName} has no multimodal runtime.`);
 
   switch (runtime) {
+    case "stt-tdt": {
+      const { loadTdtModel } = await import("./tdt/index.js");
+      return { kind: "tdt", model: await loadTdtModel(preset.id, { device, onProgress: progress_callback }) };
+    }
+    case "image-classification": {
+      const classifier = await transformers.pipeline("image-classification", preset.id, { device, dtype: preset.transformersDtype ?? "q4f16", progress_callback });
+      return { kind: "image-classification", classifier };
+    }
     case "stt": {
       // Whisper/Moonshine load an fp32 encoder + a quantized decoder (encoder quantization hurts
       // accuracy). Some ASR models document a single flat dtype instead — e.g. Cohere's cohere_asr,
@@ -214,7 +231,16 @@ export async function loadMultimodal(
         (device === "webgpu"
           ? { encoder_model: "fp32", decoder_model_merged: "q4" }
           : { encoder_model: "fp32", decoder_model_merged: "q8" });
-      const transcriber = await transformers.pipeline("automatic-speech-recognition", preset.id, { device, dtype, progress_callback });
+      const transcriber = await transformers.pipeline("automatic-speech-recognition", preset.id, {
+        device,
+        dtype,
+        progress_callback,
+        // onnxruntime-web 1.26-dev's extended graph optimizations fail at session creation for the
+        // quantized ASR decoders on the wasm EP (qdq_actions.cc: "Missing required scale …
+        // TransposeDQWeightsForMatMulNBits"). Capping at "basic" skips the broken QDQ rewrites —
+        // verified to load and transcribe. The WebGPU path keeps the default level.
+        ...(device === "wasm" ? { session_options: { graphOptimizationLevel: "basic" as const } } : {}),
+      });
       return { kind: "stt", transcriber };
     }
     case "tts-pipeline": {
@@ -235,7 +261,14 @@ export async function loadMultimodal(
     case "vlm-imagetext": {
       const processor = await transformers.AutoProcessor.from_pretrained(preset.id, { progress_callback });
       const ModelClass = runtime === "vlm-vision2seq" ? transformers.AutoModelForVision2Seq : transformers.AutoModelForImageTextToText;
-      const model = await ModelClass.from_pretrained(preset.id, { device: "webgpu", dtype: vlmDtype(preset, runtime), progress_callback });
+      let model;
+      try {
+        model = await ModelClass.from_pretrained(preset.id, { device: "webgpu", dtype: vlmDtype(preset, runtime), progress_callback });
+      } catch (error) {
+        // Release any processor-owned resources if the subsequent model load fails.
+        try { await processor.dispose?.(); } catch { /* preserve the loading error */ }
+        throw error;
+      }
       return {
         kind: "vlm",
         processor,
@@ -245,7 +278,8 @@ export async function loadMultimodal(
         maxPixels: runtime === "vlm-imagetext" ? MAX_PIXELS : 0,
         imageArray: runtime === "vlm-vision2seq",
         // Generous caps so descriptions/OCR are not truncated mid-output (any OCR model → 2048).
-        maxNewTokens: preset.id.toLowerCase().includes("ocr") ? 2048 : 1024,
+        maxNewTokens: preset.visionOptions?.maxNewTokens ?? (preset.id.toLowerCase().includes("ocr") ? 2048 : 1024),
+        visionOptions: preset.visionOptions,
       };
     }
     case "gemma-mm": {
@@ -274,6 +308,10 @@ async function loadKokoroModule(options?: LoadMultimodalOptions["kokoro"]): Prom
 
 export async function disposeMultimodal(handle: MultimodalHandle | null): Promise<void> {
   if (!handle) return;
+  if (handle.kind === "image-classification") {
+    try { await handle.classifier.dispose?.(); } catch { /* best-effort cleanup */ }
+    return;
+  }
   const model = (handle as { model?: { dispose?: () => unknown } }).model;
   try {
     if (typeof model?.dispose === "function") await model.dispose();
@@ -320,10 +358,12 @@ export function planGemmaAudioWindows(totalSamples: number): AudioWindow[] {
  * the generative STT recipes (Gemma, Granite Speech, Voxtral) require decoded PCM.
  * @throws {Error} when a URL is passed to a model that cannot fetch audio itself.
  */
-export async function transcribe(handle: MultimodalHandle, audio: Float32Array | string, onDelta?: (text: string) => void): Promise<SttResult> {
+export async function transcribe(handle: MultimodalHandle, audio: Float32Array | string, onDelta?: (text: string) => void, signal?: AbortSignal): Promise<SttResult> {
   if (typeof audio === "string" && handle.kind !== "stt") {
     throw new Error("URL audio input is only supported by ASR-pipeline models (Whisper/Moonshine). Decode to 16 kHz mono PCM first (decodeAudioTo16kMono).");
   }
+  signal?.throwIfAborted();
+  if (handle.kind === "tdt") return handle.model.transcribe(audio as Float32Array, onDelta, signal);
   if (handle.kind === "gemma") {
     // Callers supply decoded 16 kHz samples; chunk anything past Gemma's ~30s window.
     return transcribeGemmaChunked(handle, audio as Float32Array, onDelta);
@@ -566,22 +606,36 @@ export function synthesizeStream(
   return { splitter, chunks: chunks() };
 }
 
+/** Classify one image, returning all labels and inference timing. */
+export async function classifyImage(handle: MultimodalHandle, imageSrc: string): Promise<ImageClassificationResult> {
+  if (handle.kind !== "image-classification") throw new Error("Loaded model cannot classify images; select an image-classification model.");
+  if (!imageSrc.trim()) throw new Error("An image source is required.");
+  const startedMs = performance.now();
+  const output = await handle.classifier(imageSrc, { top_k: null });
+  if (!Array.isArray(output) || !output.every((item: unknown) => typeof item === "object" && item !== null && "label" in item && typeof item.label === "string" && "score" in item && typeof item.score === "number" && Number.isFinite(item.score))) {
+    throw new Error("Image classifier returned invalid label scores.");
+  }
+  return { scores: output.map((item: { label: string; score: number }) => ({ label: item.label, score: item.score })), startedMs, endedMs: performance.now() };
+}
+
 /** Answer a question about an image (VLM). Streams partial text via onDelta; returns text + telemetry. */
 export async function describeImage(handle: MultimodalHandle, imageSrc: string, prompt: string, onDelta?: (text: string) => void): Promise<VlmResult> {
+  if (handle.kind === "image-classification") throw new Error("Loaded model classifies images; use classifyImage() to obtain label scores.");
   if (handle.kind === "gemma") return runGemma(handle, { imageSrc, prompt }, onDelta);
   if (handle.kind !== "vlm") throw new Error("Loaded model cannot analyze images.");
 
-  const messages = [{ role: "user", content: [{ type: "image" }, { type: "text", text: prompt }] }];
+  const messages = [{ role: "user", content: [{ type: "image" }, { type: "text", text: prompt.trim() || handle.visionOptions?.defaultPrompt || prompt }] }];
   const text = handle.processor.apply_chat_template(messages, { add_generation_prompt: true });
   const src = handle.maxPixels ? await capSrc(imageSrc, handle.maxPixels) : imageSrc;
   const image = await handle.loadImage(src);
-  const inputs = handle.imageArray ? await handle.processor(text, [image], {}) : await handle.processor(text, image);
+  const inputs = handle.imageArray ? await handle.processor(text, [image], handle.visionOptions?.doImageSplitting === undefined ? {} : { do_image_splitting: handle.visionOptions.doImageSplitting }) : await handle.processor(text, image);
 
   const sink = { text: "", completionTokens: 0, firstTokenMs: undefined as number | undefined };
   const startedMs = performance.now();
-  const streamer = makeCountingStreamer(handle.Streamer, handle.processor.tokenizer, onDelta, sink);
+  const cleanOutput = (value: string): string => handle.visionOptions?.outputFormat === "doctags" ? value.replace(/<\|end_of_text\|>\s*$/, "").trimEnd() : value;
+  const streamer = makeCountingStreamer(handle.Streamer, handle.processor.tokenizer, onDelta ? (value) => onDelta(cleanOutput(value)) : undefined, sink, handle.visionOptions?.preserveSpecialTokens);
   await handle.model.generate({ ...inputs, max_new_tokens: handle.maxNewTokens, do_sample: false, streamer });
-  return { text: sink.text.trim(), completionTokens: sink.completionTokens, firstTokenMs: sink.firstTokenMs, startedMs, endedMs: performance.now() };
+  return { text: cleanOutput(sink.text).trim(), completionTokens: sink.completionTokens, firstTokenMs: sink.firstTokenMs, startedMs, endedMs: performance.now() };
 }
 
 /** Gemma-4 text + image + audio. image/audio placeholders go before the text; audio is a 16 kHz Float32Array. */

package/src/tdt/frontend.ts

@@ -0,0 +1,61 @@
+/** Phonon reference front end: NeMo log-mel, without dither. */
+export function phononFeatures(audio: Float32Array): { data: Float32Array; frames: number } {
+  const frames = Math.floor(audio.length / 160);
+  if (frames < 2) throw new Error('TDT audio must contain at least 320 samples (20 ms).');
+  const emphasized = new Float32Array(audio.length);
+  emphasized[0] = audio[0];
+  for (let i = 1; i < audio.length; i++) emphasized[i] = audio[i] - .97 * audio[i - 1];
+  const hzToMel = (hz: number): number => hz < 1000 ? hz / (200 / 3) : 15 + Math.log(hz / 1000) / (Math.log(6.4) / 27);
+  const melToHz = (mel: number): number => mel < 15 ? mel * (200 / 3) : 1000 * Math.exp((mel - 15) * Math.log(6.4) / 27);
+  const edges = Array.from({ length: 130 }, (_, i) => melToHz(hzToMel(8000) * i / 129));
+  const filters = Array.from({ length: 128 }, (_, m) => Float64Array.from({ length: 257 }, (_, k) => {
+    const hz = k * 16000 / 512;
+    return Math.max(0, Math.min((hz - edges[m]) / (edges[m + 1] - edges[m]), (edges[m + 2] - hz) / (edges[m + 2] - edges[m + 1]))) * 2 / (edges[m + 2] - edges[m]);
+  }));
+  const window = Float64Array.from({ length: 400 }, (_, i) => .5 - .5 * Math.cos(2 * Math.PI * i / 399));
+  const data = new Float32Array(frames * 128);
+  const real = new Float64Array(512), imag = new Float64Array(512);
+  for (let f = 0; f < frames; f++) {
+    real.fill(0); imag.fill(0);
+    for (let i = 0; i < 400; i++) {
+      const sample = f * 160 + i - 200;
+      real[i + 56] = (sample >= 0 && sample < audio.length ? emphasized[sample] : 0) * window[i];
+    }
+    fft(real, imag);
+    for (let m = 0; m < 128; m++) {
+      let power = 0;
+      for (let k = 0; k <= 256; k++) power += (real[k] ** 2 + imag[k] ** 2) * filters[m][k];
+      data[f * 128 + m] = Math.log(power + 2 ** -24);
+    }
+  }
+  for (let m = 0; m < 128; m++) {
+    let mean = 0;
+    for (let f = 0; f < frames; f++) mean += data[f * 128 + m];
+    mean /= frames;
+    let variance = 0;
+    for (let f = 0; f < frames; f++) variance += (data[f * 128 + m] - mean) ** 2;
+    const std = Math.sqrt(variance / (frames - 1)) + 1e-5;
+    for (let f = 0; f < frames; f++) data[f * 128 + m] = (data[f * 128 + m] - mean) / std;
+  }
+  return { data, frames };
+}
+
+export function fft(real: Float64Array, imag: Float64Array): void {
+  for (let i = 1, j = 0; i < 512; i++) {
+    let bit = 256;
+    for (; j & bit; bit >>= 1) j ^= bit;
+    j ^= bit;
+    if (i < j) { [real[i], real[j]] = [real[j], real[i]]; }
+  }
+  for (let size = 2; size <= 512; size *= 2) {
+    for (let start = 0; start < 512; start += size) {
+      for (let j = 0; j < size / 2; j++) {
+        const angle = -2 * Math.PI * j / size, c = Math.cos(angle), s = Math.sin(angle);
+        const a = start + j, b = a + size / 2;
+        const r = c * real[b] - s * imag[b], im = s * real[b] + c * imag[b];
+        real[b] = real[a] - r; imag[b] = imag[a] - im;
+        real[a] += r; imag[a] += im;
+      }
+    }
+  }
+}

package/src/tdt/index.ts

@@ -0,0 +1,230 @@
+/// <reference path="./runtime-types.d.ts" />
+import * as ort from 'onnxruntime-tdt/webgpu';
+import { phononFeatures } from './frontend.js';
+import { nemoFeatures } from './nemo-frontend.js';
+
+export interface TdtSegment { start: number; end: number; text: string }
+export interface TdtModel {
+  transcribe(audio: Float32Array, onDelta?: (text: string) => void, signal?: AbortSignal): Promise<{ text: string; segments: TdtSegment[] }>;
+  dispose(): Promise<void>;
+}
+export interface TdtLoadOptions {
+  device: 'webgpu' | 'wasm'; baseUrl?: string; signal?: AbortSignal;
+  wasmPaths?: string | { mjs: string; wasm: string };
+  onProgress?: (info: { status?: string; progress?: number; file?: string }) => void;
+}
+export function checkAbort(signal?: AbortSignal): void { signal?.throwIfAborted(); }
+/** Let queued input/cancellation events run even when WASM resolves only microtasks. */
+async function yieldToBrowser(signal?: AbortSignal): Promise<void> {
+  checkAbort(signal);
+  await new Promise<void>(resolve => setTimeout(resolve, 0));
+  checkAbort(signal);
+}
+export function parseVocabulary(text: string): string[] {
+  const vocab: string[] = [];
+  for (const line of text.split(/\r?\n/)) {
+    if (!line) continue;
+    const match = /^(.*) (\d+)$/.exec(line);
+    if (!match) throw new Error('Invalid TDT vocabulary line.');
+    const id = Number(match[2]);
+    if (vocab[id] !== undefined) throw new Error('Duplicate TDT vocabulary ID.');
+    vocab[id] = match[1];
+  }
+  if (!vocab.length || Array.from({ length: vocab.length }, (_, i) => vocab[i]).some(x => x === undefined)) throw new Error('Incomplete TDT vocabulary.');
+  return vocab;
+}
+export function decodePieces(ids: number[], vocab: string[]): string {
+  return ids.map(id => vocab[id] ?? '').filter(piece => piece !== '<unk>' && piece !== '<pad>' && piece !== '<blk>' && !/^<\|.*\|>$/.test(piece)).join('').replace(/▁/g, ' ').trim();
+}
+/** Split near the quietest 100 ms interval in the final five seconds, bounded at 30 seconds. */
+export function splitAudio(audio: Float32Array): Array<{ start: number; end: number }> {
+  const chunks: Array<{ start: number; end: number }> = [];
+  let start = 0;
+  while (audio.length - start > 480000) {
+    let end = start + 480000, lowest = Infinity;
+    for (let p = start + 400000; p <= start + 478400; p += 1600) {
+      let energy = 0;
+      for (let i = p; i < p + 1600; i++) energy += audio[i] ** 2;
+      if (energy < lowest) { lowest = energy; end = p + 800; }
+    }
+    chunks.push({ start, end }); start = end;
+  }
+  if (start < audio.length) chunks.push({ start, end: audio.length });
+  return chunks;
+}
+async function artifact(url: string, options: TdtLoadOptions): Promise<Uint8Array> {
+  checkAbort(options.signal);
+  const cacheWarning = (): void => options.onProgress?.({ status: 'cache-warning', file: url });
+  let cache: Cache | undefined;
+  let response: Response | undefined;
+  try { cache = typeof caches !== 'undefined' ? await caches.open('transformers-cache') : undefined; }
+  catch { cacheWarning(); }
+  try { response = await cache?.match(url); }
+  catch { cacheWarning(); }
+  const cachePut = async (value: Response): Promise<void> => {
+    if (!cache) return;
+    try { await cache.put(url, value); }
+    catch { cacheWarning(); }
+  };
+  if (!response) {
+    response = await fetch(url, { signal: options.signal });
+    if (!response.ok) throw new Error(`TDT download failed (${response.status}): ${url}`);
+    // Consume incrementally for meaningful download progress, then cache only a complete artifact.
+    const total = Number(response.headers.get('content-length'));
+    const reader = response.body?.getReader();
+    if (reader) {
+      const chunks: Uint8Array[] = []; let received = 0;
+      try {
+        while (true) {
+          checkAbort(options.signal);
+          const next = await reader.read(); if (next.done) break;
+          chunks.push(next.value); received += next.value.length;
+          options.onProgress?.({ status: 'progress', file: url, ...(total > 0 ? { progress: Math.min(100, received / total * 100) } : {}) });
+        }
+      } finally { reader.releaseLock(); }
+      const bytes = new Uint8Array(received); let offset = 0;
+      for (const chunk of chunks) { bytes.set(chunk, offset); offset += chunk.length; }
+      checkAbort(options.signal);
+      await cachePut(new Response(bytes, { headers: response.headers }));
+      options.onProgress?.({ status: 'done', file: url, progress: 100 });
+      return bytes;
+    }
+    await cachePut(response.clone());
+  }
+  if (!response.ok) throw new Error(`Invalid cached TDT artifact (${response.status}): ${url}`);
+  const bytes = new Uint8Array(await response.arrayBuffer());
+  checkAbort(options.signal);
+  options.onProgress?.({ status: 'done', file: url, progress: 100 });
+  return bytes;
+}
+function floats(tensor: ort.Tensor | undefined, name: string): Float32Array {
+  if (!tensor || tensor.type !== 'float32' || !(tensor.data instanceof Float32Array)) throw new Error(`Invalid TDT ${name} tensor.`);
+  return tensor.data;
+}
+function releaseTensors(values: Record<string, ort.Tensor>): void { for (const tensor of new Set(Object.values(values))) tensor.dispose(); }
+function argmax(values: Float32Array, start: number, end: number): number {
+  let best = start; for (let i = start + 1; i < end; i++) if (values[i] > values[best]) best = i;
+  return best - start;
+}
+
+/** Load the two explicitly supported ONNX TDT exports. Sessions are released on any partial load failure. */
+export async function loadTdtModel(modelId: string, options: TdtLoadOptions): Promise<TdtModel> {
+  const phonon = modelId === 'enriqueag/phonon-2.5-onnx';
+  if (!phonon && modelId !== 'istupakov/parakeet-tdt-0.6b-v3-onnx') throw new Error(`Unsupported TDT model: ${modelId}`);
+  if (phonon && options.device !== 'webgpu') throw new Error('Phonon exact2x2 requires WebGPU.');
+  const base = options.baseUrl ?? `https://huggingface.co/${modelId}/resolve/main/`;
+  const url = (file: string): string => new URL(file, base.endsWith('/') ? base : `${base}/`).href;
+  ort.env.wasm.numThreads = 1;
+  ort.env.wasm.wasmPaths = options.wasmPaths ?? {
+    mjs: 'https://cdn.jsdelivr.net/npm/onnxruntime-web@1.31.1/dist/ort-wasm-simd-threaded.asyncify.mjs',
+    wasm: 'https://cdn.jsdelivr.net/npm/onnxruntime-web@1.31.1/dist/ort-wasm-simd-threaded.asyncify.wasm',
+  };
+  const sessions: ort.InferenceSession[] = [];
+  const create = async (file: string, device: 'webgpu' | 'wasm', decoder = false): Promise<ort.InferenceSession> => {
+    const bytes = await artifact(url(file), options); checkAbort(options.signal);
+    const session = await ort.InferenceSession.create(bytes, { executionProviders: [device], ...(decoder && phonon ? { extra: { session: { disable_quant_qdq: '1' } } } : {}) });
+    sessions.push(session); checkAbort(options.signal); return session;
+  };
+  try {
+    const vocab = parseVocabulary(new TextDecoder().decode(await artifact(url('vocab.txt'), options)));
+    const blank = phonon ? 8192 : vocab.indexOf('<blk>');
+    if (blank < 0 || (phonon && vocab.length !== 8192 && vocab.length !== 8193)) throw new Error('Invalid TDT blank/vocabulary configuration.');
+    const encoder = await create(phonon ? 'encoder.exact2x2.onnx' : 'encoder-model.int8.onnx', phonon ? 'webgpu' : 'wasm');
+    const decoder = await create(phonon ? 'decoder_joint.int8.onnx' : 'decoder_joint-model.int8.onnx', 'wasm', true);
+    let disposed = false, busy = false;
+    const stateShape = [2, 1, 640];
+    return {
+      async dispose() {
+        if (busy) throw new Error('Cannot dispose TDT model during transcription.');
+        if (!disposed) { disposed = true; await Promise.all(sessions.map(s => s.release())); }
+      },
+      async transcribe(audio, onDelta, signal) {
+        if (disposed) throw new Error('TDT model is disposed.');
+        if (busy) throw new Error('TDT model is already transcribing.');
+        if (!(audio instanceof Float32Array) || audio.some(x => !Number.isFinite(x))) throw new Error('TDT requires finite mono 16 kHz Float32 audio.');
+        checkAbort(signal); busy = true;
+        const segments: TdtSegment[] = [];
+        let runsSinceYield = 0, lastYield = performance.now();
+        try {
+          for (const chunk of splitAudio(audio)) {
+            checkAbort(signal);
+            const wave = audio.slice(chunk.start, chunk.end);
+            if (wave.length < 320) continue;
+            const resources = new Set<ort.Tensor>();
+            const own = <T extends ort.Tensor>(t: T): T => { resources.add(t); return t; };
+            const run = async (session: ort.InferenceSession, feeds: Record<string, ort.Tensor>): Promise<Record<string, ort.Tensor>> => {
+              checkAbort(signal);
+              if (++runsSinceYield >= 32 || performance.now() - lastYield >= 20) {
+                await yieldToBrowser(signal); runsSinceYield = 0; lastYield = performance.now();
+              }
+              const outputs = await session.run(feeds);
+              for (const tensor of Object.values(outputs)) own(tensor);
+              checkAbort(signal); return outputs;
+            };
+            try {
+              let encoded: Record<string, ort.Tensor>;
+              if (phonon) {
+                await yieldToBrowser(signal);
+                const features = phononFeatures(wave);
+                await yieldToBrowser(signal);
+                runsSinceYield = 0; lastYield = performance.now();
+                encoded = await run(encoder, { features: own(new ort.Tensor('float32', features.data, [1, features.frames, 128])) });
+              } else {
+                // The published nemo128 graph uses float64 STFT, unsupported by ORT Web.
+                // This front end reproduces that graph's reflect padding and statistics.
+                await yieldToBrowser(signal);
+                const features = nemoFeatures(wave);
+                await yieldToBrowser(signal);
+                runsSinceYield = 0; lastYield = performance.now();
+                encoded = await run(encoder, {
+                  audio_signal: own(new ort.Tensor('float32', features.data, [1, 128, features.frames])),
+                  length: own(new ort.Tensor('int64', BigInt64Array.of(BigInt(features.frames)), [1])),
+                });
+              }
+              // Large encoder runs are synchronous on WASM. Give input handlers a turn
+              // before entering the otherwise microtask-only decoder loop.
+              await yieldToBrowser(signal); runsSinceYield = 0; lastYield = performance.now();
+              const output = encoded[phonon ? 'encoded' : 'outputs'];
+              const data = floats(output, 'encoder');
+              const width = output.dims[phonon ? 2 : 1], paddedSteps = output.dims[phonon ? 1 : 2];
+              const lengths = encoded.encoded_lengths;
+              const steps = phonon ? paddedSteps : Number(lengths?.data[0]);
+              if (!Number.isInteger(steps) || steps < 0 || steps > paddedSteps) throw new Error('Invalid TDT encoder length.');
+              let h: ort.Tensor = own(new ort.Tensor('float32', new Float32Array(stateShape.reduce((a, b) => a * b, 1)), stateShape));
+              let c: ort.Tensor = own(new ort.Tensor('float32', new Float32Array(stateShape.reduce((a, b) => a * b, 1)), stateShape));
+              let token = blank, t = 0, emitted = 0;
+              const ids: number[] = [];
+              while (t < steps) {
+                const frame = new Float32Array(width);
+                for (let i = 0; i < width; i++) frame[i] = data[phonon ? t * width + i : i * paddedSteps + t];
+                const ephemeral: Record<string, ort.Tensor> = phonon ? {
+                  encoded: new ort.Tensor('float32', frame, [1, width]), token: new ort.Tensor('int32', Int32Array.of(token), [1]),
+                } : {
+                  encoder_outputs: new ort.Tensor('float32', frame, [1, width, 1]), targets: new ort.Tensor('int32', Int32Array.of(token), [1, 1]), target_length: new ort.Tensor('int32', Int32Array.of(1), [1]),
+                };
+                let result: Record<string, ort.Tensor>;
+                try { result = await run(decoder, { ...ephemeral, ...(phonon ? { h, c } : { input_states_1: h, input_states_2: c }) }); }
+                finally { releaseTensors(ephemeral); }
+                const logits = floats(result[phonon ? 'logits' : 'outputs'], 'decoder');
+                if (logits.length !== blank + 1 + 5) throw new Error('Invalid TDT token/duration logits.');
+                const next = argmax(logits, 0, blank + 1); let duration = argmax(logits, blank + 1, logits.length);
+                if (next !== blank) {
+                  ids.push(next); token = next; emitted++;
+                  resources.delete(h); h.dispose(); resources.delete(c); c.dispose();
+                  h = result[phonon ? 'h_out' : 'output_states_1']; c = result[phonon ? 'c_out' : 'output_states_2'];
+                }
+                for (const tensor of Object.values(result)) if (tensor !== h && tensor !== c) { resources.delete(tensor); tensor.dispose(); }
+                if ((next === blank || emitted >= 10) && duration === 0) duration = 1;
+                if (duration > 0) emitted = 0;
+                t += duration;
+              }
+              const text = decodePieces(ids, vocab);
+              if (text) { segments.push({ start: chunk.start / 16000, end: chunk.end / 16000, text }); onDelta?.(segments.map(segment => segment.text).join(' ')); }
+            } finally { for (const tensor of resources) tensor.dispose(); }
+          }
+          return { text: segments.map(s => s.text).join(' '), segments };
+        } finally { busy = false; }
+      },
+    };
+  } catch (error) { await Promise.allSettled(sessions.map(s => s.release())); throw error; }
+}

package/src/tdt/nemo-frontend.ts

@@ -0,0 +1,43 @@
+import { fft } from './frontend.js';
+
+// Exact sparse mel matrix from istupakov/parakeet-tdt-0.6b-v3-onnx/nemo128.onnx.
+// Source graph SHA-256: a9fde1486ebfcc08f328d75ad4610c67835fea58c73ba57e3209a6f6cf019e9f.
+// Keeping these small constants avoids rounding differences in the exported filter edges.
+const MEL: number[][] = [[1,0,0.02837754227221012],[1,1,0.014389011077582836],[2,1,0.01398853026330471],[2,2,0.028778022155165672],[3,3,0.04236607998609543],[3,4,0.0004004797956440598],[4,4,0.027977053076028824],[4,5,0.014789499342441559],[5,5,0.013588042929768562],[5,6,0.02917851135134697],[6,7,0.04196559265255928],[6,8,0.0008009598823264241],[7,8,0.027576573193073273],[7,9,0.01518999319523573],[8,9,0.013187538832426071],[8,10,0.029579007998108864],[9,11,0.041565097868442535],[9,12,0.0012014397652819753],[10,12,0.02717609517276287],[10,13,0.015590474009513855],[11,13,0.01278705894947052],[11,14,0.029979487881064415],[12,15,0.041164617985486984],[12,16,0.0016019197646528482],[13,16,0.02677561342716217],[13,17,0.015990953892469406],[14,17,0.012386579066514969],[14,18,0.03037998639047146],[15,19,0.040764085948467255],[15,20,0.002002452965825796],[16,20,0.026375113055109978],[16,21,0.016391444951295853],[17,21,0.011986115016043186],[17,22,0.030780402943491936],[18,23,0.04036368802189827],[18,24,0.0024028795305639505],[19,24,0.025974653661251068],[19,25,0.01679191365838051],[20,25,0.011585619300603867],[20,26,0.03118094801902771],[21,27,0.03996315225958824],[21,28,0.0028034152928739786],[22,28,0.02557411789894104],[22,29,0.017192449420690536],[23,29,0.011185083538293839],[23,30,0.031581442803144455],[24,31,0.03956267610192299],[24,32,0.0032038395293056965],[25,32,0.025173693895339966],[25,33,0.01759287342429161],[26,33,0.010784659534692764],[26,34,0.03198190778493881],[27,35,0.03916219249367714],[27,36,0.0036043752916157246],[28,36,0.024773158133029938],[28,37,0.01799340918660164],[29,37,0.010384122841060162],[29,38,0.03238244354724884],[30,39,0.03876161575317383],[30,40,0.004004895221441984],[31,40,0.024372702464461327],[31,41,0.018320715054869652],[32,41,0.010202586650848389],[32,42,0.03161151707172394],[33,43,0.03855384513735771],[33,44,0.001656117383390665],[34,44,0.028360538184642792],[34,45,0.010670339688658714],[35,45,0.01995559222996235],[35,46,0.01796126179397106],[36,46,0.013171199709177017],[36,47,0.023691896349191666],[37,47,0.007852383889257908],[37,48,0.0280131958425045],[38,48,0.0038550780154764652],[38,49,0.031065288931131363],[39,49,0.001046702265739441],[39,50,0.03297678381204605],[40,51,0.0325256809592247],[40,52,0.0006470278021879494],[41,52,0.03098229691386223],[41,53,0.0013818024890497327],[42,53,0.03026564046740532],[42,54,0.0013279326958581805],[43,54,0.03028314933180809],[44,54,0.000853460980579257],[43,55,0.0005752422730438411],[44,55,0.029303690418601036],[45,55,0.002906062640249729],[45,56,0.026582011952996254],[46,56,0.00546437269076705],[46,57,0.023383773863315582],[47,57,0.00845953170210123],[47,58,0.019775748252868652],[48,58,0.011828471906483173],[48,59,0.015819108113646507],[49,59,0.015513437800109386],[49,60,0.01156991720199585],[50,60,0.01946169137954712],[50,61,0.007079293020069599],[51,61,0.02362503670156002],[52,61,0.0026273371186107397],[51,62,0.002393868751823902],[52,62,0.022891240194439888],[53,62,0.007950658909976482],[53,63,0.017088213935494423],[54,63,0.013358916155993938],[54,64,0.011216753162443638],[55,64,0.01881822757422924],[56,64,0.0006491197855211794],[55,65,0.005309573374688625],[56,65,0.023045793175697327],[57,65,0.006984363775700331],[57,66,0.016297349706292152],[58,66,0.013272317126393318],[58,67,0.009608481079339981],[59,67,0.019490249454975128],[60,67,0.003768591908738017],[59,68,0.0030010610353201628],[60,68,0.018348077312111855],[61,68,0.010636676102876663],[61,69,0.011118832044303417],[62,69,0.01736198551952839],[63,69,0.00308591709472239],[62,70,0.004041498526930809],[63,70,0.017977513372898102],[64,70,0.010326532647013664],[64,71,0.010409270413219929],[65,71,0.017366580665111542],[66,71,0.00440334714949131],[65,72,0.0030487067997455597],[66,72,0.015703044831752777],[67,72,0.011844555847346783],[67,73,0.007962611503899097],[68,73,0.019040562212467194],[69,73,0.007269133348017931],[68,74,0.000472926622023806],[69,74,0.011963930912315845],[70,74,0.01476955134421587],[71,74,0.003552281530573964],[70,75,0.00418809661641717],[71,75,0.01513804029673338],[72,75,0.0113017363473773],[73,75,0.0006128569366410375],[72,76,0.007129882462322712],[73,76,0.01756405457854271],[74,76,0.00856065098196268],[74,77,0.009372946806252003],[75,77,0.01618116907775402],[76,77,0.00647500017657876],[75,78,0.0015116231516003609],[76,78,0.010986563749611378],[77,78,0.014228261075913906],[78,78,0.004979042336344719],[77,79,0.003005674807354808],[78,79,0.012034479528665543],[79,79,0.012825818732380867],[80,79,0.004012242890894413],[79,80,0.003972089849412441],[80,80,0.01257565151900053],[81,80,0.01191700343042612],[82,80,0.0035183706786483526],[81,81,0.00446630734950304],[82,81,0.012664892710745335],[83,81,0.011448914185166359],[84,81,0.0034456099383533],[83,82,0.004539744462817907],[84,82,0.012352367863059044],[85,82,0.011372808367013931],[86,82,0.003746368922293186],[85,83,0.004239588975906372],[86,83,0.011684300377964973],[87,83,0.011644029058516026],[88,83,0.004376695025712252],[87,84,0.0036092570517212152],[88,84,0.010703402571380138],[89,84,0.012221302837133408],[90,84,0.005296255461871624],[89,85,0.0026888747233897448],[90,85,0.009448954835534096],[91,85,0.013066906481981277],[92,85,0.006467771250754595],[91,86,0.0015151641564443707],[92,86,0.007957103662192822],[93,86,0.014145628549158573],[94,86,0.007857220247387886],[95,86,0.0015688107814639807],[93,87,0.0001221956918016076],[94,87,0.006260755006223917],[95,87,0.012399313971400261],[96,87,0.009433379396796227],[97,87,0.003441078355535865],[96,88,0.00439020711928606],[97,88,0.010239733383059502],[98,88,0.01116772647947073],[99,88,0.005457570310682058],[98,89,0.0023731838446110487],[99,89,0.007947309873998165],[100,89,0.013034172356128693],[101,89,0.007592816837131977],[102,89,0.0021514608524739742],[100,90,0.00023496054927818477],[101,90,0.005546679254621267],[102,90,0.010858396999537945],[103,90,0.009823615662753582],[104,90,0.004638473968952894],[103,91,0.0030602705664932728],[104,91,0.008121836930513382],[105,91,0.012128844857215881],[106,91,0.0071879224851727486],[107,91,0.002247000113129616],[105,92,0.0005085079465061426],[106,92,0.005331702530384064],[107,92,0.010154897347092628],[108,92,0.009781261906027794],[109,92,0.0050729489885270596],[110,92,0.0003646367113105953],[108,93,0.002505941316485405],[109,93,0.007102088537067175],[110,93,0.011698235757648945],[111,93,0.007915019057691097],[112,93,0.0034283846616744995],[111,94,0.004040570929646492],[112,94,0.008420334197580814],[113,94,0.010758612304925919],[114,94,0.00648314505815506],[115,94,0.0022076775785535574],[113,95,0.0009844168089330196],[114,95,0.005158022977411747],[115,95,0.009331629611551762],[116,95,0.009516296908259392],[117,95,0.005442172288894653],[118,95,0.0013680467382073402],[116,96,0.0019234465435147285],[117,96,0.005900462623685598],[118,96,0.009877478703856468],[119,96,0.008634302765130997],[120,96,0.004752081818878651],[121,96,0.0008698602323420346],[119,97,0.0025171886663883924],[120,97,0.006306931842118502],[121,97,0.010096673853695393],[122,97,0.00807521678507328],[123,97,0.004375695250928402],[124,97,0.0006761741824448109],[122,98,0.002802216447889805],[123,98,0.006413611583411694],[124,98,0.010025006718933582],[125,98,0.007804089691489935],[126,98,0.004278772044926882],[127,98,0.0007534541073255241],[125,99,0.002812325721606612],[126,99,0.006253630854189396],[127,99,0.009694935753941536],[128,99,0.007789100985974073],[129,99,0.004429795313626528],[130,99,0.0010704898741096258],[128,100,0.0025784021709114313],[129,100,0.005857671145349741],[130,100,0.009136940352618694],[131,100,0.008000868372619152],[132,100,0.004799720831215382],[133,100,0.0015985730569809675],[131,101,0.0021290841978043318],[132,101,0.0052539510652422905],[133,101,0.00837881863117218],[134,101,0.008412200026214123],[135,101,0.005361814517527819],[136,101,0.002311429474502802],[134,102,0.0014907211298123002],[135,102,0.004468441475182772],[136,102,0.007446161471307278],[137,102,0.008998350240290165],[138,102,0.006091524846851826],[139,102,0.00318469968624413],[140,102,0.0002778743510134518],[137,103,0.0006874035461805761],[138,103,0.003524982836097479],[139,103,0.006362562067806721],[140,103,0.009200140833854675],[141,103,0.006966239772737026],[142,103,0.004196294117718935],[143,103,0.0014263488119468093],[141,104,0.002445453079417348],[142,104,0.005149385891854763],[143,104,0.007853318005800247],[144,104,0.007965351454913616],[145,104,0.005325851496309042],[146,104,0.002686351304873824],[147,104,4.6851124352542683e-05],[144,105,0.0012498258147388697],[145,105,0.003826454281806946],[146,105,0.006403082516044378],[147,105,0.00897971075028181],[148,105,0.006554814055562019],[149,105,0.004039536230266094],[150,105,0.001524258521385491],[148,106,0.002411770634353161],[149,106,0.004867123905569315],[150,106,0.007322477176785469],[151,106,0.00786607339978218],[152,106,0.005469250492751598],[153,106,0.0030724273528903723],[154,106,0.0006756045040674508],[151,107,0.0009218427003361285],[152,107,0.003261544741690159],[153,107,0.00560124684125185],[154,107,0.007940948940813541],[155,107,0.006960286758840084],[156,107,0.004676329903304577],[157,107,0.0023923732805997133],[158,107,0.00010841657058335841],[155,108,0.001601085183210671],[156,108,0.003830620553344488],[157,108,0.006060156039893627],[158,108,0.008289691060781479],[159,108,0.006322428584098816],[160,108,0.004146021790802479],[161,108,0.0019696145318448544],[159,109,0.0020237017888575792],[160,109,0.004148253705352545],[161,109,0.006272805389016867],[162,109,0.007998434826731682],[163,109,0.005924496799707413],[164,109,0.0038505580741912127],[165,109,0.0017766198143363],[162,110,0.00019236317893955857],[163,110,0.0022169058211147785],[164,110,0.004241448361426592],[165,110,0.0062659913673996925],[166,110,0.007716970518231392],[167,110,0.005740635562688112],[168,110,0.003764300374314189],[169,110,0.0017879655351862311],[166,111,0.0002765751851256937],[167,111,0.002205819124355912],[168,111,0.0041350629180669785],[169,111,0.006064306478947401],[170,111,0.0076301731169223785],[171,111,0.005746927578002214],[172,111,0.003863682970404625],[173,111,0.001980437897145748],[174,111,9.719309309730306e-05],[170,112,0.0001752191747073084],[171,112,0.002013584366068244],[172,112,0.003851949702948332],[173,112,0.005690314806997776],[174,112,0.0075286803767085075],[175,112,0.005921597592532635],[176,112,0.004127021413296461],[177,112,0.002332444768399],[178,112,0.0005378681817092001],[175,113,0.0016614228952676058],[176,113,0.0034132569562643766],[177,113,0.005165090784430504],[178,113,0.006916924845427275],[179,113,0.006244422402232885],[180,113,0.004534309729933739],[181,113,0.0028241975232958794],[182,113,0.0011140849674120545],[179,114,0.0011690219398587942],[180,114,0.0028383824974298477],[181,114,0.004507742822170258],[182,114,0.006177103146910667],[183,114,0.00669668847694993],[184,114,0.005067131482064724],[185,114,0.0034375751856714487],[186,114,0.001808018540032208],[187,114,0.00017846190894488245],[183,115,0.0005544161540456116],[184,115,0.0021451376378536224],[185,115,0.0037358589470386505],[186,115,0.005326580721884966],[187,115,0.006917302496731281],[188,115,0.005708739627152681],[189,115,0.004155908711254597],[190,115,0.0026030780281871557],[191,115,0.0010502473451197147],[188,116,0.0013498314656317234],[189,116,0.002865670947358012],[190,116,0.004381510429084301],[191,116,0.0058973501436412334],[192,116,0.006443656049668789],[193,116,0.004963928833603859],[194,116,0.0034842020832002163],[195,116,0.00200447509996593],[196,116,0.0005247481749393046],[192,117,0.00046751255285926163],[193,117,0.0019119851058349013],[194,117,0.003356457455083728],[195,117,0.00480093015357852],[196,117,0.006245402619242668],[197,117,0.005847633816301823],[198,117,0.004437577910721302],[199,117,0.003027522237971425],[200,117,0.0016174663323909044],[201,117,0.00020741041225846857],[197,118,0.0008883357513695955],[198,118,0.0022648002486675978],[199,118,0.0036412647459656],[200,118,0.005017729010432959],[201,118,0.006394193507730961],[202,118,0.005450631491839886],[203,118,0.0041069528087973595],[204,118,0.0027632745914161205],[205,118,0.0014195962576195598],[206,118,7.591782923555002e-05],[202,119,0.0011187248164787889],[203,119,0.0024303868412971497],[204,119,0.0037420487497001886],[205,119,0.005053711123764515],[206,119,0.00636537279933691],[207,119,0.0052314260974526405],[208,119,0.003951027989387512],[209,119,0.002670630346983671],[210,119,0.001390232122503221],[211,119,0.00010983415268128738],[207,120,0.0011792670702561736],[208,120,0.002429152838885784],[209,120,0.003679038258269429],[210,120,0.004928923677653074],[211,120,0.00617880979552865],[212,120,0.005170582327991724],[213,120,0.0039504761807620525],[214,120,0.00273036933504045],[215,120,0.001510262954980135],[216,120,0.0002901565458159894],[212,121,0.001088875811547041],[213,121,0.002279920969158411],[214,121,0.003470966126769781],[215,121,0.004662011284381151],[216,121,0.005853056441992521],[217,121,0.005250118672847748],[218,121,0.004087434150278568],[219,121,0.002924750093370676],[220,121,0.001762065920047462],[221,121,0.0005993818631395698],[217,122,0.000865069217979908],[218,122,0.00200005155056715],[219,122,0.0031350336503237486],[220,122,0.004270015750080347],[221,122,0.005404998082667589],[222,122,0.005453326739370823],[223,122,0.004345402121543884],[224,122,0.0032374775037169456],[225,122,0.0021295526530593634],[226,122,0.001021627802401781],[222,123,0.0005239806487224996],[223,123,0.0016055030282586813],[224,123,0.0026870255824178457],[225,123,0.003768547670915723],[226,123,0.004850070457905531],[227,123,0.005765118170529604],[228,123,0.004709359724074602],[229,123,0.0036536010447889566],[230,123,0.002597842598333955],[231,123,0.0015420839190483093],[232,123,0.00048632529797032475],[227,124,8.027452713577077e-05],[228,124,0.0011108804028481245],[229,124,0.002141486154869199],[230,124,0.003172091906890273],[231,124,0.004202697891741991],[232,124,0.005233303643763065],[233,124,0.005165420472621918],[234,124,0.0041593704372644424],[235,124,0.003153321100398898],[236,124,0.00214727153070271],[237,124,0.0011412217281758785],[238,124,0.00013517220213543624],[233,125,0.000529693381395191],[234,125,0.0015117720467969775],[235,125,0.002493850886821747],[236,125,0.0034759296104311943],[237,125,0.004458008334040642],[238,125,0.005440087523311377],[239,125,0.004742166958749294],[240,125,0.0037834877148270607],[241,125,0.002824808470904827],[242,125,0.0018661292269825935],[243,125,0.0009074498666450381],[239,126,0.0008101033745333552],[240,126,0.00174594612326473],[241,126,0.002681788755580783],[242,126,0.0036176315043121576],[243,126,0.004553474020212889],[244,126,0.005390489473938942],[245,126,0.00447693420574069],[246,126,0.003563379170373082],[247,126,0.0026498239021748304],[248,126,0.0017362686339765787],[249,126,0.0008227135404013097],[244,127,4.7655015805503353e-05],[245,127,0.0009394447552040219],[246,127,0.0018312345491722226],[247,127,0.0027230242267251015],[248,127,0.003614814253523946],[249,127,0.004506603814661503],[250,127,0.005223152693361044],[251,127,0.004352618008852005],[252,127,0.0034820837900042534],[253,127,0.0026115491054952145],[254,127,0.0017410146538168192],[255,127,0.0008704800857231021]];
+const WINDOW = Float64Array.from({ length: 400 }, (_, i) => .5 - .5 * Math.cos(2 * Math.PI * i / 399));
+
+/** Matches nemo128.onnx: reflected centered STFT, includes its final frame. */
+export function nemoFeatures(audio: Float32Array): { data: Float32Array; frames: number } {
+  if (audio.length < 320) throw new Error('TDT audio must contain at least 320 samples (20 ms).');
+  const frames = Math.floor(audio.length / 160) + 1;
+  const emphasized = new Float32Array(audio.length);
+  emphasized[0] = audio[0];
+  for (let i = 1; i < audio.length; i++) emphasized[i] = audio[i] - Math.fround(Math.fround(.97) * audio[i - 1]);
+  const real = new Float64Array(512), imag = new Float64Array(512);
+  const power = new Float32Array(257), data = new Float32Array(128 * frames);
+  for (let f = 0; f < frames; f++) {
+    real.fill(0); imag.fill(0);
+    for (let i = 0; i < 400; i++) {
+      let sample = f * 160 + i - 200;
+      if (sample < 0) sample = -sample;
+      if (sample >= audio.length) sample = 2 * audio.length - 2 - sample;
+      real[i + 56] = emphasized[sample] * WINDOW[i];
+    }
+    fft(real, imag);
+    for (let k = 0; k < 257; k++) power[k] = real[k] ** 2 + imag[k] ** 2;
+    const bands = new Float64Array(128);
+    for (const [k, m, weight] of MEL) bands[m] += power[k] * weight;
+    for (let m = 0; m < 128; m++) data[m * frames + f] = Math.log(Math.fround(Math.fround(bands[m]) + 2 ** -24));
+  }
+  for (let m = 0; m < 128; m++) {
+    let mean = 0;
+    for (let f = 0; f < frames; f++) mean += data[m * frames + f];
+    mean = Math.fround(Math.fround(mean) / frames);
+    let variance = 0;
+    for (let f = 0; f < frames; f++) variance += Math.fround(data[m * frames + f] - mean) ** 2;
+    const std = Math.fround(Math.sqrt(Math.fround(Math.fround(variance) / (frames - 1))));
+    const denominator = Math.fround(std + Math.fround(1e-5));
+    for (let f = 0; f < frames; f++) data[m * frames + f] = Math.fround(data[m * frames + f] - mean) / denominator;
+  }
+  return { data, frames };
+}

package/src/tdt/runtime-types.d.ts

@@ -0,0 +1,5 @@
+// The npm alias declares only its original package name. These shared public
+// session/tensor types also describe the API used by the isolated 1.31.1 runtime.
+declare module 'onnxruntime-tdt/webgpu' {
+  export * from 'onnxruntime-common';
+}

package/worker-template/cloudflare-worker.ts

@@ -50,7 +50,7 @@ const ALLOWED_HF_REPOS_BY_OWNER = new Map<string, Set<string>>([
 // through /hf-transformers/*. Derive those allowlist entries from the SDK's model catalog so the
 // proxy can never drift out of sync with it — a missing entry 403s hosted model loads.
 for (const preset of MODEL_PRESETS) {
-  if (preset.backend !== "transformers-js") continue;
+  if (preset.backend === "webllm") continue;
   const [owner, ...repoParts] = preset.id.split("/");
   if (!owner || repoParts.length === 0) continue;
   const repos = ALLOWED_HF_REPOS_BY_OWNER.get(owner) ?? new Set<string>();