File size: 16,960 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 | // APE A2 WebGPU AR ship engine — guarded decode loop over the vendored webMT
// engine. Design: docs/plans/2026-08-10-ape-a2-webgpu-ar-ship-engine-design.md
// (§3.4). Zero vendor modifications:
// - encoder: stock runEncoder over caller-tokenized ids (the Python
// contract: add_special_tokens=False — NO trailing </s>).
// - decode: stock encodeDecodeStep with lmHeadFuse 'off' so the f32 logits
// buffer materializes (COPY_SRC); one step per submit, logits copied to a
// staging buffer and mapped. The vendor argmax_penalty still runs but its
// ring write is OVERWRITTEN each step by the CPU-guarded pick
// (queue-ordered writeBuffer — the compactDecodeState precedent); its
// done/bitmask side effects are never read on this path.
// - Python-semantics note: the reference project() ADDS final_logits_bias
// to the logits; the engine leaves the bias to the argmax kernel — so
// this loop adds the bias on CPU after readback (readTensorF32 once).
// The per-step token decision is decode-guard.js's GuardedRow — the exact
// stack P1 proved equivalent (21,348/21,348 steps) to the frozen Python loop.
import { runEncoder } from './engine/encoder.js';
import { createDecodeState, encodeDecodeStep } from './engine/decoder.js';
import { VOCAB, PAD, SRC_CAP } from './engine/constants.js';
import { GuardedRow, EOS_ID, MAX_NEW_TOKENS, NEG, hasNewRepeat } from './decode-guard.js';
import { createGuardedLogitRanker } from './guarded-logits.js';
import { createGuardedTopK, createGuardedTopKWindow } from './guarded-topk.js';
// Read one tensor back from the uploaded weights mega-buffer as f32.
export async function readTensorF32(device, weights, name) {
const meta = weights.tensors.get(name);
if (!meta) throw new Error(`readTensorF32: unknown tensor ${name}`);
if (meta.dtype !== 'f32') throw new Error(`readTensorF32: ${name} is ${meta.dtype}, not f32`);
const bind = weights.bindingFor(name);
const buffer = bind.buffer ?? bind;
const offset = bind.offset ?? 0;
const staging = device.createBuffer({
size: meta.byteLength,
usage: GPUBufferUsage.MAP_READ | GPUBufferUsage.COPY_DST,
});
const enc = device.createCommandEncoder({ label: `read ${name}` });
enc.copyBufferToBuffer(buffer, offset, staging, 0, meta.byteLength);
device.queue.submit([enc.finish()]);
await staging.mapAsync(GPUMapMode.READ);
const out = new Float32Array(staging.getMappedRange().slice(0));
staging.unmap();
staging.destroy();
return out;
}
function tierHasAcceptable(guard, ids) {
for (const token of ids) {
if (token === EOS_ID) {
if (guard.name.broken || guard.name.eosOk()) return true;
continue;
}
const text = guard.pieceTable.pieceText[token];
if (!text) continue;
if (!(guard.name.broken || guard.name.trial(text))) continue;
if (hasNewRepeat(guard.sim + text, guard.draftPatterns)) continue;
return true;
}
return false;
}
// Decode one batch with the guarded ship stack.
// ctx initDevice() result
// weights loadWeights() result (f32 oracle or f16 ship build)
// pieceTable {pieceText, digitIds, bannedBytes} (piece_table.json)
// bias Float32Array[VOCAB] — final_logits_bias (readTensorF32)
// rows [{id, draft, srcIds: number[] (NO trailing eos)}]
// opts {specWindow}: >= 2 selects the speculative-window executor
// (design addendum 2026-08-11); default/1 keeps the per-step
// path. Both must produce bit-identical outputs (receipt _v2).
// Returns {rows: [{id, outputs, sim, stats, finished, steps}], timing}.
export async function guardedDecodeBatch(ctx, weights, pieceTable, bias, rows, opts = {}) {
const { device } = ctx;
const B = rows.length;
if (!B) throw new Error('guardedDecodeBatch: empty batch');
let S = 0;
for (const row of rows) {
if (!row.srcIds.length || row.srcIds.length > SRC_CAP) {
throw new Error(`guardedDecodeBatch: row ${row.id} has ${row.srcIds.length} src tokens (cap ${SRC_CAP})`);
}
if (row.srcIds[row.srcIds.length - 1] === EOS_ID) {
throw new Error(`guardedDecodeBatch: row ${row.id} ends with EOS — the contract is add_special_tokens=False`);
}
S = Math.max(S, row.srcIds.length);
}
if (bias.length !== VOCAB) throw new Error('guardedDecodeBatch: bias length != VOCAB');
const ids = new Uint32Array(B * S).fill(PAD);
const lens = new Uint32Array(B);
rows.forEach((row, i) => {
ids.set(row.srcIds, i * S);
lens[i] = row.srcIds.length;
});
const specWindow = Math.trunc(opts.specWindow ?? 0);
const useWindow = specWindow >= 2;
const t0 = performance.now();
const encRun = await runEncoder(ctx, weights, { ids, lens, B, S });
const state = createDecodeState(ctx, weights, { B, S, maxSteps: MAX_NEW_TOKENS, lmHeadFuse: 'off' });
const gpuTopK = useWindow ? null : await createGuardedTopK(device, weights, state.logits, B, VOCAB);
const win = useWindow
? await createGuardedTopKWindow(device, weights, state.logits, state.tokenRing, B, VOCAB, specWindow)
: null;
const t1 = performance.now();
const memo = new Map();
const guards = rows.map((row) => new GuardedRow(row.draft, pieceTable, memo));
const stepsPerRow = new Array(B).fill(0);
const ranker = createGuardedLogitRanker(VOCAB, NEG);
let stepsRun = 0;
const diag = { windows: 0, corrections: 0, deferrals: 0, rescues: 0 };
try {
if (useWindow) {
// Speculative-window executor. Per-row frontier = accepted steps; the
// GuardedRow is always exactly at its frontier (stepped forward once
// per accepted step, never re-stepped, never rolled back).
const FINISHED = 0xffffffff;
const frontier = new Uint32Array(B);
const needsFull = new Array(B).fill(false);
while (true) {
let t0w = Infinity;
for (let i = 0; i < B; i += 1) {
if (!guards[i].finished) t0w = Math.min(t0w, frontier[i]);
}
if (t0w === Infinity || t0w >= MAX_NEW_TOKENS) break;
// A row whose frontier step needs full logits forces a 1-step window
// once it defines t0w — state.logits then holds exactly that step.
const forceSingle = guards.some((g, i) => !g.finished && needsFull[i] && frontier[i] === t0w);
// Scheduling-only tail clamp (round-2 design §2.2): any Kp yields the
// same outputs; this just trims window overshoot past row finishes.
// Polish outputs track source length, so a source-token estimate with
// slack sizes the tail window; when a row outruns its estimate the
// clamp simply stops applying to it. Never used for token decisions.
let maxSpan = 1;
for (let i = 0; i < B; i += 1) {
if (guards[i].finished) continue;
const est = Math.ceil(1.2 * rows[i].srcIds.length) + 4;
const remaining = est > frontier[i] ? est - frontier[i] : specWindow;
maxSpan = Math.max(maxSpan, frontier[i] - t0w + remaining);
}
const Kp = forceSingle ? 1 : Math.min(specWindow, MAX_NEW_TOKENS - t0w, maxSpan);
const requests = guards.map((guard) => (guard.finished
? { mode: 'none', ids: [] }
: guard.maskRequest()));
const frontCpu = new Uint32Array(B);
for (let i = 0; i < B; i += 1) frontCpu[i] = guards[i].finished ? FINISHED : frontier[i];
const hist = new Uint32Array(Kp * B).fill(EOS_ID);
for (let i = 0; i < B; i += 1) {
for (let k = 0; k < Kp; k += 1) {
if (t0w + k < frontCpu[i]) {
const tok = guards[i].outputs[t0w + k];
if (tok !== undefined) hist[k * B + i] = tok;
}
}
}
win.prepare(requests, frontCpu, hist, t0w, Kp);
const cmd = device.createCommandEncoder({ label: `guarded window ${t0w}+${Kp}` });
const scratchAll = [];
for (let k = 0; k < Kp; k += 1) {
const pass = cmd.beginComputePass({ label: `guarded window step ${t0w + k}` });
const { scratch } = encodeDecodeStep(ctx, weights, encRun, state, t0w + k, pass);
pass.end();
win.record(cmd, k);
scratchAll.push(...scratch);
}
win.copyOut(cmd, Kp);
device.queue.submit([cmd.finish()]);
for (const buf of scratchAll) buf.destroy();
stepsRun += Kp;
diag.windows += 1;
const top = await win.read(Kp);
const corrections = [];
const fullRows = [];
for (let i = 0; i < B; i += 1) {
const guard = guards[i];
if (guard.finished) continue;
let f = frontier[i];
for (let j = Math.max(t0w, f); j < t0w + Kp; j += 1) {
const base = ((j - t0w) * B + i) * 18;
const atFrontier = j === f;
if (atFrontier && needsFull[i]) {
if (Kp === 1 && f === t0w) fullRows.push(i);
break;
}
const visible = [];
for (let k = 0; k < 16; k += 1) {
const id = top[base + k];
if (id === win.emptyId) break;
visible.push(id);
}
const rawTop = top[base + 17];
let tier;
let maskedTop;
if (atFrontier) {
// The kernel applied the true mask here — per-step semantics.
maskedTop = top[base + 16];
if (!tierHasAcceptable(guard, visible)) {
needsFull[i] = true;
diag.rescues += 1;
break;
}
tier = visible;
} else {
// Free-run step: the readback is unmasked. The permitted
// sub-sequence of the unmasked top-16 is an exact prefix of
// the masked descending order (unbanned/allowed values keep
// their raw value; everything hidden ranks below them), so
// the guard decision is exact whenever it lands inside it.
const req = guard.maskRequest();
let prefix = visible;
if (req.mode === 'ban') {
const banned = new Set(req.ids);
prefix = visible.filter((id) => !banned.has(id));
} else if (req.mode === 'allow') {
const allowed = new Set(req.ids);
prefix = visible.filter((id) => allowed.has(id));
}
if (!prefix.length || !tierHasAcceptable(guard, prefix)) {
f = j; // next window applies the true mask at j
diag.deferrals += 1;
break;
}
maskedTop = prefix[0];
tier = prefix;
}
const ringVal = atFrontier ? top[base + 16] : rawTop;
const chosen = guard.step(rawTop, maskedTop, (tierIndex) => {
if (tierIndex !== 0) throw new Error('guarded window tier escalation past preflight');
return { ids: tier, validCount: tier.length };
});
stepsPerRow[i] += 1;
f = j + 1;
if (chosen !== ringVal) {
// The free-run trajectory diverged: pin the true token and
// recompute later steps next window (unaffected rows below
// their own frontier restore identical history/KV).
corrections.push([j, i, chosen]);
break;
}
if (guard.finished) break;
}
frontier[i] = f;
}
if (fullRows.length) {
// Rescue: tier-16 had no acceptable candidate at a frontier step
// (P1 frequency: 0). Kp == 1, so state.logits holds this step.
const rowBytes = VOCAB * 4;
const rescueStaging = device.createBuffer({
label: 'guarded window rescue logits',
size: fullRows.length * rowBytes,
usage: GPUBufferUsage.MAP_READ | GPUBufferUsage.COPY_DST,
});
const rescueCmd = device.createCommandEncoder({ label: 'guarded window rescue readback' });
fullRows.forEach((row, index) => {
rescueCmd.copyBufferToBuffer(state.logits, row * rowBytes, rescueStaging, index * rowBytes, rowBytes);
});
device.queue.submit([rescueCmd.finish()]);
await rescueStaging.mapAsync(GPUMapMode.READ);
const view = new Float32Array(rescueStaging.getMappedRange());
fullRows.forEach((row, index) => {
const guard = guards[row];
const ranked = ranker.rank(view, index * VOCAB, bias, requests[row]);
const chosen = guard.step(ranked.rawTop, ranked.maskedTop, ranked.getTier);
stepsPerRow[row] += 1;
needsFull[row] = false;
frontier[row] = t0w + 1;
const ringVal = top[row * 18 + 16];
if (chosen !== ringVal) corrections.push([t0w, row, chosen]);
});
rescueStaging.unmap();
rescueStaging.destroy();
}
for (const [t, rowIdx, token] of corrections) {
device.queue.writeBuffer(state.tokenRing, (t * B + rowIdx) * 4, new Uint32Array([token]));
}
diag.corrections += corrections.length;
}
} else {
for (let t = 0; t < MAX_NEW_TOKENS; t += 1) {
const requests = guards.map((guard) => (guard.finished
? { mode: 'none', ids: [] }
: guard.maskRequest()));
gpuTopK.prepare(requests);
const cmd = device.createCommandEncoder({ label: `guarded step ${t}` });
const pass = cmd.beginComputePass({ label: `guarded step ${t}` });
const { scratch } = encodeDecodeStep(ctx, weights, encRun, state, t, pass);
pass.end();
gpuTopK.record(cmd);
device.queue.submit([cmd.finish()]);
for (const buf of scratch) buf.destroy();
stepsRun += 1;
const top = await gpuTopK.read();
const picks = new Uint32Array(B).fill(EOS_ID);
const fallbackRows = [];
const gpuTiers = new Array(B);
for (let i = 0; i < B; i += 1) {
if (guards[i].finished) continue;
const base = i * gpuTopK.outputStride;
const ids = [];
for (let k = 0; k < 16; k += 1) {
if (top[base + k] === gpuTopK.emptyId) break;
ids.push(top[base + k]);
}
gpuTiers[i] = ids;
if (!tierHasAcceptable(guards[i], ids)) fallbackRows.push(i);
}
let fallbackStaging = null;
let fallbackView = null;
if (fallbackRows.length) {
const rowBytes = VOCAB * 4;
fallbackStaging = device.createBuffer({
label: 'guarded fallback logits',
size: fallbackRows.length * rowBytes,
usage: GPUBufferUsage.MAP_READ | GPUBufferUsage.COPY_DST,
});
const fallbackCmd = device.createCommandEncoder({ label: 'guarded fallback readback' });
fallbackRows.forEach((row, index) => {
fallbackCmd.copyBufferToBuffer(state.logits, row * rowBytes, fallbackStaging, index * rowBytes, rowBytes);
});
device.queue.submit([fallbackCmd.finish()]);
await fallbackStaging.mapAsync(GPUMapMode.READ);
fallbackView = new Float32Array(fallbackStaging.getMappedRange());
}
const fallbackIndex = new Map(fallbackRows.map((row, index) => [row, index]));
let live = 0;
for (let i = 0; i < B; i += 1) {
const guard = guards[i];
if (guard.finished) continue;
const compactIndex = fallbackIndex.get(i);
let chosen;
if (compactIndex !== undefined) {
const ranked = ranker.rank(fallbackView, compactIndex * VOCAB, bias, requests[i]);
chosen = guard.step(ranked.rawTop, ranked.maskedTop, ranked.getTier);
} else {
const base = i * gpuTopK.outputStride;
const tier = { ids: gpuTiers[i], validCount: gpuTiers[i].length };
chosen = guard.step(top[base + 17], top[base + 16], (tierIndex) => {
if (tierIndex !== 0) throw new Error('guarded top-16 preflight disagreement');
return tier;
});
}
picks[i] = chosen;
stepsPerRow[i] += 1;
if (!guard.finished) live += 1;
}
if (fallbackStaging) {
fallbackStaging.unmap();
fallbackStaging.destroy();
}
device.queue.writeBuffer(state.tokenRing, t * B * 4, picks);
if (live === 0) break;
}
}
} finally {
(useWindow ? win : gpuTopK).destroy();
state.destroy();
encRun.arena.destroy();
}
const t2 = performance.now();
return {
rows: rows.map((row, i) => ({
id: row.id,
outputs: guards[i].outputs,
sim: guards[i].sim,
stats: guards[i].stats,
finished: guards[i].finished,
steps: stepsPerRow[i],
})),
timing: {
encoderMs: Math.round(t1 - t0),
decodeMs: Math.round(t2 - t1),
stepsRun,
spec: useWindow ? specWindow : 0,
...(useWindow ? diag : {}),
},
};
}
|