File size: 27,030 Bytes
24b9788
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
// Main worker: DiT + encoders + VAE on WebGPU. Spawns a dedicated LM worker
// (isolated WASM heap) for autoregressive generation.
import { AutoTokenizer } from "@huggingface/transformers";
import * as ort from "onnxruntime-web/webgpu";

const MODEL_REPO = "shreyask/ACE-Step-v1.5-ONNX";
const MODEL_REVISION = "bdabfb5684fd70fcc76f98cbb51bb9ebc47ee342";
const ONNX_BASE = `https://huggingface.co/${MODEL_REPO}/resolve/${MODEL_REVISION}/onnx`;
const TEXT_TOKENIZER_REPO = "Qwen/Qwen3-Embedding-0.6B";

const SAMPLE_RATE = 48000;
const LATENT_RATE = 25;
const LATENT_CHANNELS = 64;
const HIDDEN_SIZE = 2048;
const POOL_WINDOW = 5;
const FSQ_DIM = 6;
const NUM_CODES = 64000;

// 8-step turbo schedules (from ACE-Step)
const SHIFT_TIMESTEPS_8 = {
  1.0: [1.0, 0.875, 0.75, 0.625, 0.5, 0.375, 0.25, 0.125],
  2.0: [1.0, 0.9333, 0.8571, 0.7692, 0.6667, 0.5455, 0.4, 0.2222],
  3.0: [1.0, 0.9545, 0.9, 0.8333, 0.75, 0.6429, 0.5, 0.3],
};

// Generate N-step shifted schedule matching MLX port:
//   timesteps = linspace(1.0, 0.001, N)
//   sigmas = shift * t / (1 + (shift-1) * t)
function buildSchedule(numSteps, shift) {
  if (numSteps === 8 && SHIFT_TIMESTEPS_8[shift]) return SHIFT_TIMESTEPS_8[shift];
  const sigmaMax = 1.0;
  const sigmaMin = 0.001;
  const schedule = [];
  for (let i = 0; i < numSteps; i++) {
    // linspace inclusive of both endpoints
    const t = sigmaMax + (sigmaMin - sigmaMax) * (i / (numSteps - 1));
    const tShifted = (shift * t) / (1.0 + (shift - 1.0) * t);
    schedule.push(tShifted);
  }
  return schedule;
}

const CACHE_NAME = "ace-step-onnx-v12";

let textTokenizer = null;
let sessions = {};
let silenceLatent = null;
let fsqCodebooks = null;
let fsqScales = null;
let fsqProjectOutW = null;
let fsqProjectOutB = null;
let lmWorker = null;
let lmLoaded = false;

function post(type, data = {}) {
  self.postMessage({ type, ...data });
}

async function fetchBuffer(url, label) {
  const cache = await caches.open(CACHE_NAME);
  const cached = await cache.match(url);
  if (cached) {
    post("progress", { label, loaded: 1, total: 1, percent: 100 });
    return await cached.arrayBuffer();
  }

  const response = await fetch(url);
  const total = parseInt(response.headers.get("content-length") || "0");
  const reader = response.body.getReader();
  const chunks = [];
  let loaded = 0;

  while (true) {
    const { done, value } = await reader.read();
    if (done) break;
    chunks.push(value);
    loaded += value.length;
    if (total > 0) post("progress", { label, loaded, total, percent: (loaded / total) * 100 });
  }

  const buffer = new Uint8Array(loaded);
  let offset = 0;
  for (const chunk of chunks) { buffer.set(chunk, offset); offset += chunk.length; }

  try {
    await cache.put(url, new Response(buffer.buffer.slice(0), {
      headers: { "Content-Type": "application/octet-stream" },
    }));
  } catch (_) {}

  return buffer.buffer;
}

async function loadSession(name, filename, useUrlData = false, providers = ["webgpu"]) {
  post("status", { message: `Loading ${name}...` });
  try {
    const modelBuffer = await fetchBuffer(`${ONNX_BASE}/${filename}`, `${name} graph`);
    if (useUrlData) {
      return await ort.InferenceSession.create(modelBuffer, {
        executionProviders: providers,
        externalData: [{ path: `${filename}.data`, data: `${ONNX_BASE}/${filename}.data` }],
      });
    }
    const weightsBuffer = await fetchBuffer(`${ONNX_BASE}/${filename}.data`, `${name} weights`);
    return await ort.InferenceSession.create(modelBuffer, {
      executionProviders: providers,
      externalData: [{ path: `${filename}.data`, data: weightsBuffer }],
    });
  } catch (err) {
    throw new Error(`Failed loading ${name}: ${err.message}`);
  }
}

function tensor(data, dims, type = "float32") {
  return new ort.Tensor(type, data, dims);
}

function tensorStats(name, data) {
  const arr = data instanceof Float32Array ? data : new Float32Array(data);
  let min = Infinity, max = -Infinity, sum = 0;
  for (let i = 0; i < arr.length; i++) {
    if (arr[i] < min) min = arr[i];
    if (arr[i] > max) max = arr[i];
    sum += arr[i];
  }
  console.log(`[stats] ${name}: len=${arr.length} min=${min.toFixed(4)} max=${max.toFixed(4)} mean=${(sum / arr.length).toFixed(4)}`);
}

function randn(shape) {
  const size = shape.reduce((a, b) => a * b, 1);
  const data = new Float32Array(size);
  for (let i = 0; i < size; i += 2) {
    const u1 = Math.random();
    const u2 = Math.random();
    const r = Math.sqrt(-2 * Math.log(u1));
    data[i] = r * Math.cos(2 * Math.PI * u2);
    if (i + 1 < size) data[i + 1] = r * Math.sin(2 * Math.PI * u2);
  }
  return data;
}

function packSequences(hidden1, mask1, hidden2, mask2, batchSize, dim) {
  const l1 = hidden1.length / (batchSize * dim);
  const l2 = hidden2.length / (batchSize * dim);
  const totalLen = l1 + l2;
  const packedHidden = new Float32Array(batchSize * totalLen * dim);
  const packedMask = new Float32Array(batchSize * totalLen);

  for (let b = 0; b < batchSize; b++) {
    const indices = [];
    for (let i = 0; i < l1; i++) indices.push({ src: 1, idx: i, mask: mask1[b * l1 + i] });
    for (let i = 0; i < l2; i++) indices.push({ src: 2, idx: i, mask: mask2[b * l2 + i] });
    indices.sort((a, c) => c.mask - a.mask);

    for (let pos = 0; pos < totalLen; pos++) {
      const entry = indices[pos];
      const srcArray = entry.src === 1 ? hidden1 : hidden2;
      const srcLen = entry.src === 1 ? l1 : l2;
      const srcOffset = (b * srcLen + entry.idx) * dim;
      const dstOffset = (b * totalLen + pos) * dim;
      packedHidden.set(srcArray.slice(srcOffset, srcOffset + dim), dstOffset);
      packedMask[b * totalLen + pos] = entry.mask > 0 ? 1 : 0;
    }
  }
  return { hidden: packedHidden, mask: packedMask, seqLen: totalLen };
}

function fsqLookup(indices, batchSize, seqLen) {
  const out = new Float32Array(batchSize * seqLen * HIDDEN_SIZE);
  for (let b = 0; b < batchSize; b++) {
    for (let t = 0; t < seqLen; t++) {
      const idx = indices[b * seqLen + t];
      const codeOffset = idx * FSQ_DIM;
      const scaledCode = new Float32Array(FSQ_DIM);
      for (let d = 0; d < FSQ_DIM; d++) scaledCode[d] = fsqCodebooks[codeOffset + d] * fsqScales[d];
      const outOffset = (b * seqLen + t) * HIDDEN_SIZE;
      for (let h = 0; h < HIDDEN_SIZE; h++) {
        let val = fsqProjectOutB[h];
        for (let d = 0; d < FSQ_DIM; d++) val += scaledCode[d] * fsqProjectOutW[h * FSQ_DIM + d];
        out[outOffset + h] = val;
      }
    }
  }
  return out;
}

// Spawn the LM worker and forward its status/progress messages up to the main thread
function spawnLMWorker() {
  const worker = new Worker(new URL("./lm-worker.js", import.meta.url), { type: "module" });
  worker.onmessage = (e) => {
    const { type, ...data } = e.data;
    if (type === "status" || type === "progress" || type === "error") {
      self.postMessage(e.data);  // forward as-is
    }
    // "loaded" and "audio_codes" are handled by the promise-based callers below
  };
  return worker;
}

function loadLMWorker() {
  return new Promise((resolve, reject) => {
    if (!lmWorker) lmWorker = spawnLMWorker();
    const onMsg = (e) => {
      if (e.data.type === "loaded") {
        lmWorker.removeEventListener("message", onMsg);
        lmLoaded = true;
        resolve();
      } else if (e.data.type === "error") {
        lmWorker.removeEventListener("message", onMsg);
        reject(new Error(e.data.message));
      }
    };
    lmWorker.addEventListener("message", onMsg);
    lmWorker.postMessage({ type: "load" });
  });
}

function generateAudioCodesViaLM({ caption, lyrics, duration, numLatentFrames }) {
  return new Promise((resolve, reject) => {
    const onMsg = (e) => {
      if (e.data.type === "audio_codes") {
        lmWorker.removeEventListener("message", onMsg);
        resolve(e.data);
      } else if (e.data.type === "error") {
        lmWorker.removeEventListener("message", onMsg);
        reject(new Error(e.data.message));
      }
    };
    lmWorker.addEventListener("message", onMsg);
    lmWorker.postMessage({ type: "generate", caption, lyrics, duration, numLatentFrames });
  });
}

async function loadModels() {
  ort.env.wasm.numThreads = 1;
  ort.env.wasm.simd = true;
  ort.env.wasm.proxy = false;

  console.log(`[models] ONNX revision ${MODEL_REVISION}`);
  post("status", { message: `Using ONNX revision ${MODEL_REVISION.slice(0, 7)}` });

  post("status", { message: "Spawning LM worker..." });
  // Kick off LM loading in parallel with main-worker model loads
  const lmLoadPromise = loadLMWorker();

  post("status", { message: "Loading text tokenizer..." });
  textTokenizer = await AutoTokenizer.from_pretrained(TEXT_TOKENIZER_REPO);

  sessions.embedTokens = await loadSession("Embed Tokens", "text_embed_tokens_fp16.onnx");
  sessions.detokenizer = await loadSession("Detokenizer", "detokenizer.onnx");
  // VAE on WASM — WebGPU produces constant output past ~1.5s for conv1d upsample chain
  sessions.vaeDecoder = await loadSession("VAE Decoder (CPU)", "vae_decoder_fp16.onnx", false, ["wasm"]);
  sessions.textEncoder = await loadSession("Text Encoder", "text_encoder_fp16.onnx", true);
  // FP32 condition_encoder — q4v2 had max_diff=13.92 vs PyTorch with real inputs,
  // degrading conditioning so badly that DiT output was garbled. FP32 is 2.4GB via URL.
  sessions.conditionEncoder = await loadSession("Condition Encoder (fp32)", "condition_encoder.onnx", true);
  // DEBUG: dit_decoder_fp16_v2 is the quality baseline (max_diff=0.021 per step).
  // dit_cached trades quality for speed (max_diff=0.074). Reverting while we diagnose
  // the ONNX-vs-MLX spectral gap — compounded drift over 8 steps matters here.
  sessions.ditDecoder = await loadSession("DiT Decoder (uncached)", "dit_decoder_fp16_v2.onnx", true);

  post("status", { message: "Loading auxiliary data..." });
  const [cbBuf, scBuf, powBuf, pobBuf, silBuf] = await Promise.all([
    fetchBuffer(`${ONNX_BASE}/fsq_codebooks.bin`, "codebooks"),
    fetchBuffer(`${ONNX_BASE}/fsq_scales.bin`, "scales"),
    fetchBuffer(`${ONNX_BASE}/fsq_project_out_weight.bin`, "proj_out_w"),
    fetchBuffer(`${ONNX_BASE}/fsq_project_out_bias.bin`, "proj_out_b"),
    fetchBuffer("/silence_latent.bin", "silence latent"),
  ]);
  fsqCodebooks = new Float32Array(cbBuf);
  fsqScales = new Float32Array(scBuf);
  fsqProjectOutW = new Float32Array(powBuf);
  fsqProjectOutB = new Float32Array(pobBuf);
  silenceLatent = new Float32Array(silBuf);

  post("status", { message: "Waiting for LM worker..." });
  await lmLoadPromise;

  post("status", { message: "All models loaded!" });
  post("loaded");
}

function buildSFTPrompt(caption, metas) {
  const instruction = "Fill the audio semantic mask based on the given conditions:";
  return `# Instruction\n${instruction}\n\n# Caption\n${caption}\n\n# Metas\n${metas}<|endoftext|>`;
}

async function encodeText(caption, metas) {
  const prompt = buildSFTPrompt(caption, metas);
  const encoded = textTokenizer(prompt, { padding: "max_length", max_length: 256, truncation: true });
  const idsRaw = encoded.input_ids.data;
  const inputIds = idsRaw instanceof BigInt64Array ? idsRaw : new BigInt64Array(Array.from(idsRaw, BigInt));

  const result = await sessions.textEncoder.run({ input_ids: tensor(inputIds, [1, 256], "int64") });
  const projected = await sessions.textProjector.run({ text_hidden_states: result.hidden_states });

  const maskRaw = encoded.attention_mask.data;
  const attentionMask = new Float32Array(maskRaw.length);
  for (let i = 0; i < maskRaw.length; i++) attentionMask[i] = Number(maskRaw[i]);
  return { hidden: projected.projected.data, mask: attentionMask, seqLen: 256 };
}

async function encodeLyrics(lyrics, language = "en") {
  const fullText = `# Languages\n${language}\n\n# Lyric\n${lyrics}`;
  // max_length=2048 matches the original handler (conditioning_text.py)
  const encoded = textTokenizer(fullText, { padding: "max_length", max_length: 2048, truncation: true });
  const idsRaw = encoded.input_ids.data;
  const inputIds = idsRaw instanceof BigInt64Array ? idsRaw : new BigInt64Array(Array.from(idsRaw, BigInt));
  const seqLen = inputIds.length;

  const embedResult = await sessions.embedTokens.run({ input_ids: tensor(inputIds, [1, seqLen], "int64") });
  const maskRaw = encoded.attention_mask.data;
  const attentionMask = new Float32Array(maskRaw.length);
  for (let i = 0; i < maskRaw.length; i++) attentionMask[i] = Number(maskRaw[i]);

  const lyricResult = await sessions.lyricEncoder.run({
    inputs_embeds: embedResult.hidden_states,
    attention_mask: tensor(attentionMask, [1, seqLen]),
  });
  return { hidden: lyricResult.hidden_states.data, mask: attentionMask, seqLen };
}

async function encodeTimbre() {
  const silenceRef = silenceLatent.slice(0, 750 * LATENT_CHANNELS);
  const result = await sessions.timbreEncoder.run({
    refer_audio: tensor(silenceRef, [1, 750, LATENT_CHANNELS]),
  });
  const timbreHidden = new Float32Array(HIDDEN_SIZE);
  timbreHidden.set(result.timbre_embedding.data);
  return { hidden: timbreHidden, mask: new Float32Array([1.0]), seqLen: 1 };
}

async function generateLMHints(caption, lyrics, numLatentFrames, duration) {
  const { codes, elapsed, tokenCount } = await generateAudioCodesViaLM({ caption, lyrics, duration, numLatentFrames });
  post("status", { message: `LM: ${codes.length} codes from ${tokenCount} tokens in ${elapsed}s` });

  if (codes.length === 0) {
    console.warn("[lm] No audio codes generated, returning silence");
    return new Float32Array(numLatentFrames * LATENT_CHANNELS);
  }

  const numCodes5Hz = codes.length;
  post("status", { message: "FSQ codebook lookup..." });
  const lmHints5Hz = fsqLookup(codes, 1, numCodes5Hz);
  tensorStats("lm_hints_5hz", lmHints5Hz);

  post("status", { message: "Detokenizing 5Hz → 25Hz..." });
  const detokResult = await sessions.detokenizer.run({
    quantized: tensor(lmHints5Hz, [1, numCodes5Hz, HIDDEN_SIZE]),
  });
  const lmHints25HzRaw = detokResult.lm_hints_25hz.data;
  const rawLen = lmHints25HzRaw.length / LATENT_CHANNELS;
  tensorStats("lm_hints_25hz_raw", lmHints25HzRaw);

  // Pad with last frame (MLX port behavior) or truncate
  const lmHints25Hz = new Float32Array(numLatentFrames * LATENT_CHANNELS);
  if (rawLen >= numLatentFrames) {
    lmHints25Hz.set(lmHints25HzRaw.slice(0, numLatentFrames * LATENT_CHANNELS));
  } else {
    lmHints25Hz.set(lmHints25HzRaw);
    // Repeat last frame to fill remaining
    const lastFrameStart = (rawLen - 1) * LATENT_CHANNELS;
    const lastFrame = lmHints25HzRaw.slice(lastFrameStart, lastFrameStart + LATENT_CHANNELS);
    for (let t = rawLen; t < numLatentFrames; t++) {
      lmHints25Hz.set(lastFrame, t * LATENT_CHANNELS);
    }
    console.log(`[hints] padded ${rawLen}${numLatentFrames} frames with last-frame replication`);
  }
  tensorStats("lm_hints_25hz_final", lmHints25Hz);
  return lmHints25Hz;
}

async function generateAudio({ caption, lyrics, duration, shift, numSteps = 8 }) {
  const totalStartTime = performance.now();
  const filenameStamp = Date.now();
  const batchSize = 1;
  const numLatentFrames = Math.round(duration * LATENT_RATE);
  const tSchedule = buildSchedule(numSteps, shift);
  const metas = `duration: ${duration}s`;

  // 1. Text → Qwen3 embedding (1024-dim hidden states, BEFORE projection)
  post("status", { message: "Encoding text..." });
  const sftPrompt = buildSFTPrompt(caption, metas);
  const textEnc = textTokenizer(sftPrompt, { padding: "max_length", max_length: 256, truncation: true });
  const textIdsRaw = textEnc.input_ids.data;
  const textIds = textIdsRaw instanceof BigInt64Array ? textIdsRaw : new BigInt64Array(Array.from(textIdsRaw, BigInt));
  const textHiddenRes = await sessions.textEncoder.run({ input_ids: tensor(textIds, [1, 256], "int64") });
  const textHidden = textHiddenRes.hidden_states;
  const textMaskRaw = textEnc.attention_mask.data;
  const textMask = new Float32Array(textMaskRaw.length);
  for (let i = 0; i < textMaskRaw.length; i++) textMask[i] = Number(textMaskRaw[i]);

  // 2. Lyric tokens → embed_tokens (1024-dim, passed into condition_encoder's lyric_encoder)
  post("status", { message: "Embedding lyrics..." });
  const lyricFullText = `# Languages\nen\n\n# Lyric\n${lyrics}`;
  const lyricEnc = textTokenizer(lyricFullText, { padding: "max_length", max_length: 2048, truncation: true });
  const lyricIdsRaw = lyricEnc.input_ids.data;
  const lyricIds = lyricIdsRaw instanceof BigInt64Array ? lyricIdsRaw : new BigInt64Array(Array.from(lyricIdsRaw, BigInt));
  const lyricEmbRes = await sessions.embedTokens.run({ input_ids: tensor(lyricIds, [1, 2048], "int64") });
  const lyricEmb = lyricEmbRes.hidden_states;
  const lyricMaskRaw = lyricEnc.attention_mask.data;
  const lyricMask = new Float32Array(lyricMaskRaw.length);
  for (let i = 0; i < lyricMaskRaw.length; i++) lyricMask[i] = Number(lyricMaskRaw[i]);

  // 3. LM hints (mandatory for turbo model)
  const lmHints25Hz = await generateLMHints(caption, lyrics, numLatentFrames, duration);

  // 4. Silence for ref audio (timbre) and src_latents
  const silenceRef = silenceLatent.slice(0, 750 * LATENT_CHANNELS);
  const srcLatents = new Float32Array(numLatentFrames * LATENT_CHANNELS);
  const chunkMasks = new Float32Array(numLatentFrames * LATENT_CHANNELS).fill(1.0);
  const isCovers = new Float32Array([1.0]);  // force use of LM hints

  // 5. condition_encoder: does text_projector + lyric_encoder + timbre_encoder + pack_sequences + context_latents
  post("status", { message: "Running condition encoder..." });
  const condResult = await sessions.conditionEncoder.run({
    text_hidden_states: textHidden,
    text_attention_mask: tensor(textMask, [1, 256]),
    lyric_hidden_states: lyricEmb,
    lyric_attention_mask: tensor(lyricMask, [1, 2048]),
    refer_audio_acoustic_hidden_states_packed: tensor(silenceRef, [1, 750, LATENT_CHANNELS]),
    refer_audio_order_mask: tensor(new BigInt64Array([0n]), [1], "int64"),
    src_latents: tensor(srcLatents, [1, numLatentFrames, LATENT_CHANNELS]),
    chunk_masks: tensor(chunkMasks, [1, numLatentFrames, LATENT_CHANNELS]),
    is_covers: tensor(isCovers, [1]),
    precomputed_lm_hints_25hz: tensor(lmHints25Hz, [1, numLatentFrames, LATENT_CHANNELS]),
  });
  const encoderHiddenStates = condResult.encoder_hidden_states;
  const contextLatentsTensor = condResult.context_latents;
  tensorStats("encoder_hidden_states", encoderHiddenStates.data);
  tensorStats("context_latents", contextLatentsTensor.data);

  post("status", { message: "Starting denoising..." });
  let xt = randn([batchSize, numLatentFrames, LATENT_CHANNELS]);
  const startTime = performance.now();

  for (let step = 0; step < tSchedule.length; step++) {
    const tCurr = tSchedule[step];
    post("status", { message: `Denoising step ${step + 1}/${tSchedule.length}...` });

    const timestepData = new Float32Array(batchSize).fill(tCurr);
    const result = await sessions.ditDecoder.run({
      hidden_states: tensor(xt, [batchSize, numLatentFrames, LATENT_CHANNELS]),
      timestep: tensor(timestepData, [batchSize]),
      encoder_hidden_states: encoderHiddenStates,
      context_latents: contextLatentsTensor,
    });

    const vt = result.velocity.data;
    if (step === tSchedule.length - 1) {
      for (let i = 0; i < xt.length; i++) xt[i] = xt[i] - vt[i] * tCurr;
    } else {
      const dt = tCurr - tSchedule[step + 1];
      for (let i = 0; i < xt.length; i++) xt[i] = xt[i] - vt[i] * dt;
    }
  }

  const diffusionTime = ((performance.now() - startTime) / 1000).toFixed(2);
  tensorStats("final_latent", xt);

  // Per-frame variance check — detects if later frames are constant
  const perFrameVariance = new Float32Array(numLatentFrames);
  for (let t = 0; t < numLatentFrames; t++) {
    let mean = 0;
    for (let c = 0; c < LATENT_CHANNELS; c++) mean += xt[t * LATENT_CHANNELS + c];
    mean /= LATENT_CHANNELS;
    let varSum = 0;
    for (let c = 0; c < LATENT_CHANNELS; c++) {
      const d = xt[t * LATENT_CHANNELS + c] - mean;
      varSum += d * d;
    }
    perFrameVariance[t] = varSum / LATENT_CHANNELS;
  }
  console.log("[perframe] variance samples:", Array.from(perFrameVariance.filter((_, i) => i % 25 === 0)).map(v => v.toFixed(3)));

  // Also check LM hints per-frame variance
  const hintsVar = new Float32Array(numLatentFrames);
  for (let t = 0; t < numLatentFrames; t++) {
    let mean = 0;
    for (let c = 0; c < LATENT_CHANNELS; c++) mean += lmHints25Hz[t * LATENT_CHANNELS + c];
    mean /= LATENT_CHANNELS;
    let varSum = 0;
    for (let c = 0; c < LATENT_CHANNELS; c++) {
      const d = lmHints25Hz[t * LATENT_CHANNELS + c] - mean;
      varSum += d * d;
    }
    hintsVar[t] = varSum / LATENT_CHANNELS;
  }
  console.log("[hints var] samples:", Array.from(hintsVar.filter((_, i) => i % 25 === 0)).map(v => v.toFixed(3)));

  post("status", { message: "Decoding audio..." });
  const latentsForVae = new Float32Array(batchSize * LATENT_CHANNELS * numLatentFrames);
  for (let t = 0; t < numLatentFrames; t++) {
    for (let c = 0; c < LATENT_CHANNELS; c++) {
      latentsForVae[c * numLatentFrames + t] = xt[t * LATENT_CHANNELS + c];
    }
  }

  const vaeResult = await sessions.vaeDecoder.run({
    latents: tensor(latentsForVae, [batchSize, LATENT_CHANNELS, numLatentFrames]),
  });

  const waveform = vaeResult.waveform.data;
  tensorStats("waveform", waveform);

  masterWaveform(waveform, SAMPLE_RATE, 2);

  const wavBuffer = float32ToWav(waveform, SAMPLE_RATE, 2);
  // totalTime measures the whole pipeline (LM + encoders + diffusion + VAE),
  // not just the diffusion loop. diffusionTime is reported separately below.
  const totalTime = ((performance.now() - totalStartTime) / 1000).toFixed(2);

  post("audio", { wavBuffer, duration, diffusionTime, totalTime, filenameStamp }, [wavBuffer]);
}

function measureAudio(samples) {
  let peak = 0;
  let sumSq = 0;
  for (let i = 0; i < samples.length; i++) {
    const v = samples[i];
    const abs = Math.abs(v);
    if (abs > peak) peak = abs;
    sumSq += v * v;
  }
  return { peak, rms: Math.sqrt(sumSq / Math.max(1, samples.length)) };
}

function goertzelPower(data, sampleRate, freq) {
  const omega = 2 * Math.PI * freq / sampleRate;
  const coeff = 2 * Math.cos(omega);
  let s0 = 0, s1 = 0, s2 = 0;
  for (let i = 0; i < data.length; i++) {
    s0 = data[i] + coeff * s1 - s2;
    s2 = s1;
    s1 = s0;
  }
  return s1 * s1 + s2 * s2 - coeff * s1 * s2;
}

function detectDronePeaks(samples, sampleRate, channels) {
  const numSamples = samples.length / channels;
  const step = Math.max(1, Math.floor(sampleRate / 4000));
  const downsampleRate = sampleRate / step;
  const downsampledLength = Math.floor(numSamples / step);
  if (downsampledLength < 1024) return [];

  const mono = new Float32Array(downsampledLength);
  let mean = 0;
  for (let i = 0; i < downsampledLength; i++) {
    const src = i * step;
    let v = 0;
    for (let ch = 0; ch < channels; ch++) v += samples[ch * numSamples + src];
    v /= channels;
    mono[i] = v;
    mean += v;
  }
  mean /= downsampledLength;
  for (let i = 0; i < mono.length; i++) mono[i] -= mean;

  const bins = [];
  for (let freq = 250; freq <= 950; freq += 12.5) {
    bins.push({ freq, power: goertzelPower(mono, downsampleRate, freq) });
  }
  const sortedPowers = bins.map((bin) => bin.power).sort((a, b) => a - b);
  const median = sortedPowers[Math.floor(sortedPowers.length / 2)] + 1e-12;
  bins.sort((a, b) => b.power - a.power);

  const peaks = [];
  for (const bin of bins) {
    const score = bin.power / median;
    if (score < 12) break;
    if (peaks.every((peak) => Math.abs(peak.freq - bin.freq) >= 50)) {
      peaks.push({ freq: bin.freq, score });
      if (peaks.length >= 2) break;
    }
  }
  return peaks;
}

function applyNotch(samples, sampleRate, channels, freq, q = 20, depth = 0.45) {
  const numSamples = samples.length / channels;
  const w0 = 2 * Math.PI * freq / sampleRate;
  const cos = Math.cos(w0);
  const alpha = Math.sin(w0) / (2 * q);
  const a0 = 1 + alpha;
  const b0 = 1 / a0;
  const b1 = (-2 * cos) / a0;
  const b2 = 1 / a0;
  const a1 = (-2 * cos) / a0;
  const a2 = (1 - alpha) / a0;

  for (let ch = 0; ch < channels; ch++) {
    const offset = ch * numSamples;
    let x1 = 0, x2 = 0, y1 = 0, y2 = 0;
    for (let i = 0; i < numSamples; i++) {
      const x0 = samples[offset + i];
      const y0 = b0 * x0 + b1 * x1 + b2 * x2 - a1 * y1 - a2 * y2;
      samples[offset + i] = x0 * (1 - depth) + y0 * depth;
      x2 = x1; x1 = x0;
      y2 = y1; y1 = y0;
    }
  }
}

function masterWaveform(samples, sampleRate, channels) {
  const before = measureAudio(samples);
  if (before.peak <= 0.001) return;

  const dronePeaks = detectDronePeaks(samples, sampleRate, channels);
  for (const peak of dronePeaks) applyNotch(samples, sampleRate, channels, peak.freq);

  const afterEq = measureAudio(samples);
  const targetRms = 0.085;
  const maxPeak = 0.891;
  const maxGain = 12.0;
  const gain = Math.min(
    maxGain,
    targetRms / Math.max(afterEq.rms, 1e-6),
    maxPeak / Math.max(afterEq.peak, 1e-6),
  );
  for (let i = 0; i < samples.length; i++) samples[i] *= gain;

  const after = measureAudio(samples);
  const peakText = dronePeaks.map((peak) => `${peak.freq.toFixed(1)}Hz/${peak.score.toFixed(0)}x`).join(", ") || "none";
  console.log(
    `[master] rawPeak=${before.peak.toFixed(4)} rawRms=${before.rms.toFixed(4)} ` +
    `dronePeaks=${peakText} gain=${gain.toFixed(2)}x peak=${after.peak.toFixed(4)} rms=${after.rms.toFixed(4)}`,
  );
}

function float32ToWav(samples, sampleRate, channels = 2) {
  const numSamples = samples.length / channels;
  const bitsPerSample = 16;
  const blockAlign = channels * (bitsPerSample / 8);
  const byteRate = sampleRate * blockAlign;
  const dataSize = numSamples * blockAlign;
  const buffer = new ArrayBuffer(44 + dataSize);
  const view = new DataView(buffer);
  const w = (o, s) => { for (let i = 0; i < s.length; i++) view.setUint8(o + i, s.charCodeAt(i)); };
  w(0, "RIFF"); view.setUint32(4, 36 + dataSize, true);
  w(8, "WAVE"); w(12, "fmt "); view.setUint32(16, 16, true);
  view.setUint16(20, 1, true); view.setUint16(22, channels, true);
  view.setUint32(24, sampleRate, true); view.setUint32(28, byteRate, true);
  view.setUint16(32, blockAlign, true); view.setUint16(34, bitsPerSample, true);
  w(36, "data"); view.setUint32(40, dataSize, true);
  let offset = 44;
  for (let i = 0; i < numSamples; i++) {
    for (let ch = 0; ch < channels; ch++) {
      const sample = Math.max(-1, Math.min(1, samples[ch * numSamples + i]));
      view.setInt16(offset, sample * 32767, true);
      offset += 2;
    }
  }
  return buffer;
}

self.onmessage = async (e) => {
  const { type, ...data } = e.data;
  try {
    if (type === "load") await loadModels();
    else if (type === "generate") await generateAudio(data);
  } catch (err) {
    post("error", { message: err.message, stack: err.stack });
  }
};