File size: 24,518 Bytes
c971a45
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
// Full GPU-resident greedy decode: tokenize → encoder (one submit) → decode
// loop in groups of K=8 steps per submit. Within a group all 25·K dispatches
// are recorded into one compute pass — the argmax kernel writes each step's
// token into the ring and the next step's embed reads it, so the CPU never
// sits between steps. After each group ONE readback (ring slice + done flags)
// tells the CPU which tokens were produced and whether every row has hit eos.
//
// Decode-ahead (default): group g+1 is encoded and submitted BEFORE awaiting
// group g's readback, so the GPU never drains while the CPU maps/collects.
// GPU-side correctness is free — queue order guarantees group g's argmax
// writes the ring before group g+1's embed reads it. The cost is that EOS
// early-exit lags one group: when group g's flags show all rows done, group
// g+1 has already been submitted (≤ (inFlight-1)·GROUP_STEPS wasted steps per
// run; done rows just produce PAD, which never enters a sequence). inFlight
// staging buffers rotate: a buffer is copied into, mapped, read, unmapped —
// and only then reused (N groups in flight ⇒ N buffers is exactly enough).
// Pass {decodeAhead: false} for the sequential A/B path.
//
// Adaptive submit budget (submitBudgetMs, default 1000): group readback
// waits over the budget halve the steps-per-submit for the rest of the
// batch — slow mobile GPUs converge to submits the OS watchdog tolerates
// instead of dying with VK_ERROR_DEVICE_LOST; fast GPUs never trip it.
//
// Timing semantics (spike-grade, documented not perfect):
//   encoderMs    wall time from the runEncoder() call until
//                queue.onSubmittedWorkDone() resolves for its submit —
//                includes tokenized-batch upload + command recording overhead.
//   decodeMs     wall time of the whole decode loop (GPU work + per-group
//                mapAsync readbacks + CPU collection).
//   cpuEncodeMs  Σ per-group CPU time recording commands + submitting.
//   awaitMs      Σ per-group time awaiting mapAsync + copying the readback.
import { tokenizeBatch } from './tokenizer.js';
import { runEncoder } from './encoder.js';
import {
  createDecodeState, compactDecodeState, compactDecodeStateInPlace,
  encodeDecodeStep, growDecodeKV,
} from './decoder.js';
import {
  createUniformParamPool, getDispatchStats, shouldUseUniformParamPool,
} from './pipelines.js';
import { maxNewTokensFor, maxBatchForLimits } from './shapes.js';
import { EOS, DECODER_START, DECODE_CAP } from './constants.js';

const GROUP_STEPS = 8; // decode steps recorded per submit
const SUBMIT_BUDGET_MS = 1000; // adaptive per-submit ceiling (submitBudgetMs)

// One-way ratchet for the adaptive submit budget: halve steps-per-submit
// whenever a group's readback wait exceeded the budget. With the pipeline
// primed the CPU parks on mapAsync for almost exactly the GPU tail of that
// group, so the wait is a lower bound on the group's GPU time — overshooting
// means the submit kept the GPU busy past the budget, the regime where
// Android compositor fences start missing and the driver eventually kills
// the context (measured on Adreno 710: 3000ms fence misses, then
// vkQueueSubmit VK_ERROR_DEVICE_LOST once submits reach ~4-5s; ~1.5s
// submits survive a 286k-char run). Never grows back within a batch: file
// mode sorts long rows last so pressure only rises, a false shrink costs a
// few % submit overhead, a missed shrink costs the device.
export function nextGroupSteps(cur, groupAwaitMs, budgetMs) {
  if (!budgetMs || cur <= 1 || !(groupAwaitMs > budgetMs)) return cur;
  return Math.max(1, cur >> 1);
}

// translateBatch(ctx, weights, tok, sources,
//                {maxNewTokens?, onProgress?, onPrimed?, decodeAhead?,
//                 inFlight?, compact?, overlapEnc?})
//   ctx      {device, limits?, ...} from initDevice()
//   sources  array of source strings (one batch)
//   onPrimed called ONCE, right after the encoder and the first decode
//            groups are submitted and the loop is about to park on the GPU —
//            the spot where caller CPU work (pre-tokenizing the next batch)
//            overlaps GPU execution instead of delaying submits
//   overlapEnc (default true) skip the queue drain between encoder and
//            decode: queue order already sequences crossKV before its
//            readers, so the first decode groups are recorded while the GPU
//            still encodes. encoderMs is then measured via a non-blocking
//            onSubmittedWorkDone.then() and overlaps decodeMs by a few ms
//            (stage sums read slightly high; wall is what drops). false
//            restores the drained baseline (A/B arm).
//   inFlight decode-ahead depth: 2 (default, double-buffered staging) or 3
//            (triple-buffered). Measured on 64 file23k chunks at B=64
//            (m5_inflight_ab, 2026-07-06): median 28737 vs 28724 tok/s — a
//            wash; the readback gap is already hidden at depth 2, so 2 stays
//            the default.
//   compact  EOS row compaction (default true): when ≥ max(8, B·compactFrac)
//            rows of the current batch have emitted eos, stop submitting,
//            drain the in-flight groups, and rebuild the decode context with
//            only the live rows (compactDecodeState) — finished rows
//            otherwise keep burning GEMM rows and attention workgroups until
//            the whole group hits eos. Kernel routing is pinned across
//            compactions, so output is token-exact vs {compact: false}
//            (compact_equiv gate).
//   compactFrac  dead fraction of the CURRENT batch that triggers a compaction
//            (default 0.25 — the original B>>2). Lower = more compactions:
//            each costs a drain of the in-flight groups plus the live-row
//            copies, each saves dead-row GEMM/attention work for every
//            remaining step. Compaction count and step timings shift but
//            tokens stay exact at ANY value. MEASURED INSENSITIVE on sorted
//            file batches (dec_compact_sweep 2026-07-09: 0.25→0.03 identical
//            wall/compactions — uniform rows die in one synchronized wave, so
//            every threshold fires at the same group boundary).
//   groupSteps  decode steps recorded per submit (default 8). Smaller halves
//            the EOS-detection latency (dead rows compute until their group's
//            readback lands) at the cost of more submits; token-exact at any
//            value — compaction timing shifts, row math doesn't
//            (dec_group_sweep A/Bs this).
//   submitBudgetMs  adaptive submit ceiling (default 1000, null/0 disables):
//            whenever one group's readback wait exceeds this, steps-per-submit
//            halve for the REST of the batch (8→4→2→1, never back up). Fast
//            GPUs never trip it (group waits are tens of ms); slow mobile
//            GPUs converge within a few groups to submits the OS watchdog
//            tolerates instead of VK_ERROR_DEVICE_LOST. Token-exact like any
//            groupSteps value. Callers can seed the next batch with this
//            batch's landing point via metrics.groupStepsFinal.
// Returns:
//   rows     [{ids, text, forcedEos, steps}] — ids = [0, ...tokens] trimmed at
//            the first eos inclusive; rows that never emitted eos within the
//            cap get one appended (forcedEos: true — HF max_length semantics)
//   metrics  {tokenizeMs, encoderMs, decodeMs, detokMs, cpuEncodeMs, awaitMs,
//             submits, steps, tokensGenerated, tokPerSec}
//            tokenizeMs/detokMs bracket the CPU tokenizer calls — the two
//            stages the GPU timings can't see (the app-stage breakdown needs
//            them to locate wall-clock loss on large files)
//            tokensGenerated counts argmax-produced tokens across ALL rows
//            (incl. an emitted eos, excl. a force-appended one); tokPerSec is
//            total tokens over decodeMs.
export async function translateBatch(ctx, weights, tok, sources, {
  maxNewTokens, onProgress, onPrimed = null, decodeAhead = true, inFlight = 2, compact = true, compactFrac = 0.25, inPlaceCompact = false, groupSteps = GROUP_STEPS, submitBudgetMs = SUBMIT_BUDGET_MS, overlapEnc = true,
  kvCapacity = null,
  // runEncoder passthrough (row-packing A/B; encSplitSubmits cuts the
  // encoder into per-layer submits — the watchdog guard for the encoder
  // side, see runEncoder splitSubmits)
  encPacked = 'auto',
  encSplitSubmits = false,
  // createDecodeState passthrough ('q8' / fusion / layout A/B tests, and the
  // per-batch options tunedOptions(tuned, B) resolves after an autotune run)
  lmHead = 'auto', ffn = 'auto', lmHeadFuse = 'auto', fuseLn = 'auto', proj = 'auto', ffnSplitK = 'auto', projSplitK = 'auto', decodeMega = 'auto', sg = 'auto', encAttnSafe = false,
  immediates = 'auto',
  uniformPool = false,
  tiledProj = 'auto',
} = {}) {
  const { device } = ctx;
  if (typeof inPlaceCompact !== 'boolean') {
    throw new Error(`translateBatch: inPlaceCompact must be boolean, got ${inPlaceCompact}`);
  }
  const dispatchBefore = getDispatchStats(device);
  if (inFlight !== 2 && inFlight !== 3) {
    throw new Error(`translateBatch: inFlight must be 2 or 3, got ${inFlight}`);
  }
  if (!Number.isInteger(groupSteps) || groupSteps < 1) {
    throw new Error(`translateBatch: groupSteps must be a positive integer, got ${groupSteps}`);
  }
  const tTok0 = performance.now();
  const batch = await tokenizeBatch(tok, sources);
  const tokenizeMs = performance.now() - tTok0;
  const { B } = batch;
  const srcTruncated = new Set(batch.truncated ?? []);

  // Memory guard: the encoder ffnTmp [B·S, 1792] binding is the ceiling. No
  // silent sub-batching — the bench/app layer owns batch-size policy, the
  // engine stays explicit.
  const maxB = maxBatchForLimits(ctx, batch.S, weights.dtype === 'f16' ? 2 : 4);
  if (B > maxB) {
    throw new Error(
      `translateBatch: batch ${B} at S=${batch.S} exceeds ` +
      `maxStorageBufferBindingSize=${ctx?.limits?.maxStorageBufferBindingSize ?? 'default 134217728'} ` +
      `(encoder ffnTmp [B·S, 1792]); split into batches of ≤ ${maxB}`,
    );
  }

  const cap = Math.min(
    DECODE_CAP,
    maxNewTokens ?? Math.max(...sources.map((s) => maxNewTokensFor(s.length))),
  );

  const tEnc0 = performance.now();
  const encRun = await runEncoder(ctx, weights, batch, {
    packed: encPacked, sg: sg === 'on', attnQbAlign8: encAttnSafe,
    retainEncOut: false, splitSubmits: encSplitSubmits,
  });
  let cur = null;
  const stagings = [];
  let paramPool = null;
  let cleaned = false;
  const cleanup = () => {
    if (cleaned) return;
    cleaned = true;
    // Cleanup is best effort so an allocation/decode failure is never masked
    // by a secondary destroy error. Every owner is idempotent.
    for (const staging of stagings) {
      try { staging.destroy(); } catch {}
    }
    try { paramPool?.destroy(); } catch {}
    try { cur?.state?.destroy(); } catch {}
    try { cur?.arena?.destroy(); } catch {}
    try { encRun.arena.destroy(); } catch {}
  };

  try {
  // Don't drain the queue between encoder and decode (overlapEnc, default):
  // queue order already guarantees the decode dispatches see the finished
  // crossKV, so the CPU can record/submit the first decode groups WHILE the
  // GPU is still encoding — the old blocking await left the GPU idle for
  // exactly the CPU-side recording time of those groups every batch.
  // encoderMs still brackets submit → GPU-done via a non-blocking then()
  // (±one macrotask); with overlap on, the decode loop starts inside that
  // window, so encoderMs and decodeMs overlap by up to a few ms — the stage
  // sums in app_stage/translateText read slightly high, wall clock is what
  // dropped. {overlapEnc: false} restores the drained A/B baseline.
  let encoderMs = 0;
  const encDone = (encRun.submittedDone ?? device.queue.onSubmittedWorkDone()).then(
    () => { encoderMs = performance.now() - tEnc0; },
    () => {}, // measurement only — a lost device surfaces via the decode loop
  );
  if (!overlapEnc) await encDone;

  // The mutable decode context: state + the cross-attention view. Replaced
  // wholesale by each compaction; curMap maps current row -> original row.
  cur = {
    state: createDecodeState(ctx, weights, {
      B, S: encRun.S, maxSteps: cap, lmHead, ffn, lmHeadFuse, fuseLn, proj,
      tiledProj, ffnSplitK, projSplitK, decodeMega, sg, immediates, inPlaceCompact,
      kvCapacity,
    }),
    crossKV: encRun.crossKV, lensBuf: encRun.lensBuf, S: encRun.S, arena: null,
  };
  let curB = B;
  let curMap = Array.from({ length: B }, (_, i) => i);
  const rowTokens = Array.from({ length: B }, () => []);
  const rowDone = new Array(B).fill(false); // CPU mirror: saw eos in the ring
  let submits = 0;
  let steps = 0;
  let compactions = 0;
  const initialKvCapacity = cur.state.kvCapacity;
  const kvCapacitySequence = [initialKvCapacity];
  let kvGrows = 0;
  let kvGrowsWithPending = 0;
  let kvGrowBindGroupsPurged = 0;
  let cpuEncodeMs = 0;
  let awaitMs = 0;
  // Adaptive submit budget state: curGroupSteps only ever shrinks (see
  // nextGroupSteps); groupIndex rotates staging/pool banks by SUBMIT order,
  // which g/groupSteps no longer encodes once the group size changes.
  let curGroupSteps = groupSteps;
  let groupIndex = 0;
  let submitShrinks = 0;
  let maxGroupAwaitMs = 0;

  // inFlight staging buffers, rotated by submit index. Sized for a full
  // group at the STARTING groupSteps (the adaptive path only shrinks);
  // shorter groups copy less and read accordingly.
  const stagingBytes = (groupSteps * B + B) * 4;
  for (let i = 0; i < inFlight; i++) {
    stagings.push(device.createBuffer({
      label: `decode group staging ${i}`,
      size: stagingBytes,
      usage: GPUBufferUsage.COPY_DST | GPUBufferUsage.MAP_READ,
    }));
  }
  // Safari/WebKit currently lacks immediate_address_space. The selected pool
  // keeps ordinary uniform bindings but gives each in-flight group a stable
  // buffer bank, so parameter uploads collapse to one write and bind groups
  // become reusable. Immediate-mode states keep their existing path.
  paramPool = shouldUseUniformParamPool(uniformPool, cur.state)
    ? createUniformParamPool(device, { banks: inFlight })
    : null;

  // Record + submit one group [g, tEnd). Returns the in-flight descriptor —
  // it carries the batch geometry (B, row map) it was submitted with, since a
  // compaction may swap the context before its readback is processed.
  const encodeGroup = (g, pendingGroups = 0) => {
    const t0 = performance.now();
    const tEnd = Math.min(g + curGroupSteps, cap);
    if (tEnd > cur.state.kvCapacity) {
      const grow = growDecodeKV(ctx, weights, cur.state, {
        requiredCapacity: tEnd,
        submittedSteps: g,
        groupSteps: curGroupSteps,
      });
      if (grow.grown) {
        kvGrows++;
        if (pendingGroups > 0) kvGrowsWithPending++;
        kvGrowBindGroupsPurged += grow.bindGroupsPurged;
        kvCapacitySequence.push(grow.newCapacity);
      }
    }
    const bank = groupIndex % stagings.length;
    groupIndex++;
    const staging = stagings[bank];
    if (staging.mapState !== 'unmapped') {
      throw new Error(`decode staging reused while ${staging.mapState}`);
    }
    const scratch = [];
    let poolFrameActive = false;
    try {
      if (paramPool) {
        paramPool.begin(bank);
        poolFrameActive = true;
      }
      const encoder = device.createCommandEncoder({ label: `decode ${g}..${tEnd}` });
      const pass = encoder.beginComputePass({ label: `decode ${g}..${tEnd}` });
      for (let t = g; t < tEnd; t++) {
        scratch.push(...encodeDecodeStep(ctx, weights, cur, cur.state, t, pass).scratch);
      }
      pass.end();
      if (paramPool) {
        paramPool.flush();
        poolFrameActive = false;
      }

      // One readback per group: ring slice [g·B, tEnd·B) + done[B].
      const nTok = (tEnd - g) * curB;
      encoder.copyBufferToBuffer(cur.state.tokenRing, g * curB * 4, staging, 0, nTok * 4);
      encoder.copyBufferToBuffer(cur.state.done, 0, staging, nTok * 4, curB * 4);
      device.queue.submit([encoder.finish()]);
      submits++;
      for (const buf of scratch) buf.destroy(); // safe post-submit
      cpuEncodeMs += performance.now() - t0;
      return { g, tEnd, nTok, staging, B: curB, map: curMap, poolBank: paramPool ? bank : null };
    } catch (err) {
      if (poolFrameActive) paramPool.abort();
      for (const buf of scratch) buf.destroy();
      throw err;
    }
  };

  // Compact away the finished rows: pays when a decent slice of the batch is
  // dead AND there are steps left. The max(8, ·) floor keeps small batches
  // (b1–b8 gates, latency runs) permanently on the no-compaction path.
  const compactWanted = () => {
    if (!compact || steps === 0) return false;
    const live = curMap.reduce((n, orig) => n + (rowDone[orig] ? 0 : 1), 0);
    return live > 0 && curB - live >= Math.max(8, Math.ceil(curB * compactFrac));
  };
  const doCompact = () => {
    const liveIdx = [];
    for (let i = 0; i < curB; i++) if (!rowDone[curMap[i]]) liveIdx.push(i);
    const newMap = liveIdx.map((i) => curMap[i]);
    const lens = new Uint32Array(newMap.map((orig) => batch.lens[orig]));
    const lastTok = new Uint32Array(newMap.map((orig) => rowTokens[orig].at(-1)));
    if (inPlaceCompact) {
      compactDecodeStateInPlace(ctx, weights, cur, {
        liveIdx, t0: steps, lens, lastTok,
      });
    } else {
      const prev = cur;
      // pending is empty here, so every pool bank has completed its readback
      // and been released. Drop bind groups for the old resource generation
      // BEFORE allocating the compacted state: on WebKit those bind groups
      // keep the old KV/crossKV/arena backing memory alive after destroy().
      paramPool?.invalidateBindings();
      cur = compactDecodeState(ctx, weights, prev, {
        liveIdx, t0: steps, cap, lens, lastTok,
      });
      // Old buffers are queue-retained by the just-submitted copies; destroy()
      // only blocks future submissions. First compaction: the encoder arena
      // (encOut + old crossKV/lens) is no longer referenced either.
      prev.state.destroy();
      if (prev.arena) prev.arena.destroy();
      else encRun.arena.destroy();
    }
    curB = liveIdx.length;
    curMap = newMap;
    compactions++;
  };

  const tDec0 = performance.now();
  try {
    const depth = decodeAhead ? inFlight : 1; // groups in flight (submitted, unread)
    const pending = []; // submitted-but-unread groups, oldest first
    let nextG = 0;
    let allDone = false;
    for (;;) {
      // A wanted compaction stalls new submits so the in-flight groups (built
      // against the OLD layout) drain first — one pipeline bubble per compact.
      const wantCompact = !allDone && nextG < cap && compactWanted();
      if (wantCompact && pending.length === 0) {
        doCompact();
        continue;
      }
      while (!allDone && !wantCompact && nextG < cap && pending.length < depth) {
        const grp = encodeGroup(nextG, pending.length);
        pending.push(grp);
        nextG = grp.tEnd; // curGroupSteps may shrink between submits
      }
      // The pipeline is primed: encoder + the first decode groups are all
      // submitted, and the next await parks on the GPU. This is the one spot
      // where a caller can burn CPU for free (e.g. pre-tokenizing the NEXT
      // batch into the SPM LRU) — earlier would delay these submits, later
      // (onProgress) the GPU is already half done.
      if (onPrimed) {
        const cb = onPrimed;
        onPrimed = null;
        cb();
      }
      const grp = pending.shift();
      if (!grp) break;
      if (allDone) continue; // submitted before all-done was seen; results are
      //                        PAD-only for done rows — safe to ignore unread.

      const tA0 = performance.now();
      await grp.staging.mapAsync(GPUMapMode.READ, 0, (grp.nTok + grp.B) * 4);
      const data = new Uint32Array(grp.staging.getMappedRange(0, (grp.nTok + grp.B) * 4).slice(0));
      grp.staging.unmap(); // staging is now free for group g+2
      if (grp.poolBank !== null) paramPool.release(grp.poolBank);
      const groupAwait = performance.now() - tA0;
      awaitMs += groupAwait;
      if (groupAwait > maxGroupAwaitMs) maxGroupAwaitMs = groupAwait;
      // Adaptive submit budget: a long wait here means the GPU chewed on one
      // submit past the budget — shrink the groups still to be submitted.
      const shrunk = nextGroupSteps(curGroupSteps, groupAwait, submitBudgetMs);
      if (shrunk !== curGroupSteps) {
        curGroupSteps = shrunk;
        submitShrinks++;
      }

      // Collect per row (map current -> original), stopping at its first eos —
      // done rows produce PAD afterwards, which must NOT enter the sequence.
      for (let t = grp.g; t < grp.tEnd; t++) {
        for (let b = 0; b < grp.B; b++) {
          const orig = grp.map[b];
          if (rowDone[orig]) continue;
          const id = data[(t - grp.g) * grp.B + b];
          rowTokens[orig].push(id);
          if (id === EOS) rowDone[orig] = true;
        }
      }
      steps = grp.tEnd;
      onProgress?.({ step: grp.tEnd, cap, done: rowDone.filter(Boolean).length, B });

      // GPU done flags (queue-ordered snapshot after step tEnd-1).
      const doneFlags = data.subarray(grp.nTok, grp.nTok + grp.B);
      if (doneFlags.every((d) => d === 1)) allDone = true;
    }
  } finally {
    cleanup();
  }
  const decodeMs = performance.now() - tDec0;
  // The decode loop's readbacks are queue-ordered after the encoder submit,
  // so encDone has long resolved — this await only pins encoderMs before the
  // metrics object is built.
  await encDone;

  const tDet0 = performance.now();
  const rows = [];
  let tokensGenerated = 0;
  for (let b = 0; b < B; b++) {
    const forcedEos = !rowDone[b];
    tokensGenerated += rowTokens[b].length; // argmax-produced (incl. emitted eos)
    const ids = [DECODER_START, ...rowTokens[b]];
    if (forcedEos) ids.push(EOS);
    const text = tok.decode(ids, { skip_special_tokens: true }).trim();
    // srcTruncated: the ENCODER saw a cut-off source (tokenizeBatch hit
    // SRC_CAP) — the decode itself is fine, but `text` only translates the
    // prefix. Distinct from forcedEos, which is the TARGET ring cap.
    rows.push({
      ids, text, forcedEos, steps: rowTokens[b].length,
      srcTruncated: srcTruncated.has(b),
    });
  }
  const detokMs = performance.now() - tDet0;
  const dispatchAfter = getDispatchStats(device);
  const dispatch = {};
  for (const field of [
    'uniformBuffersCreated', 'uniformPoolBuffersCreated', 'uniformPoolBuffersDestroyed',
    'uniformPoolFramesBegun', 'uniformPoolFramesFlushed', 'uniformPoolBlocks',
    'uniformPoolBytes', 'uniformPoolBindGroupCacheHits',
    'uniformPoolWarmBindGroupLookups', 'uniformPoolWarmBindGroupCacheHits',
    'uniformPoolWarmBindGroupResets', 'uniformPoolGenerationInvalidations',
    'uniformPoolCachePurges',
    'dummyBuffersCreated', 'bindGroupsCreated', 'bindGroupCacheHits',
    'bindGroupEvictions', 'bindGroupTargetedPurgeCalls',
    'bindGroupTargetedPurges', 'immediateSets',
  ]) {
    dispatch[field] = dispatchAfter[field] - dispatchBefore[field];
  }
  dispatch.bindGroupCacheSize = dispatchAfter.bindGroupCacheSize;
  dispatch.bindGroupCacheLimit = dispatchAfter.bindGroupCacheLimit;

  return {
    rows,
    metrics: {
      tokenizeMs, encoderMs, decodeMs, detokMs, cpuEncodeMs, awaitMs,
      B, S: encRun.S, cap,
      submits, steps, compactions,
      groupStepsStart: groupSteps, groupStepsFinal: curGroupSteps,
      submitShrinks, maxGroupAwaitMs,
      initialKvCapacity, finalKvCapacity: cur.state.kvCapacity,
      kvGrows, kvGrowsWithPending, kvCapacitySequence, kvGrowBindGroupsPurged,
      srcTruncated: srcTruncated.size,
      compactMode: inPlaceCompact ? 'inplace' : 'realloc', tokensGenerated,
      dispatch, uniformPool: paramPool?.snapshot() ?? null,
      tokPerSec: decodeMs > 0 ? (tokensGenerated / decodeMs) * 1000 : 0,
    },
  };
  } finally {
    // Covers initialization failures before the decode loop's narrower
    // finally is entered (decode-state, partial staging, or pool setup).
    cleanup();
  }
}