doppler-gpu 0.4.8 → 0.4.9

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (60) hide show
  1. package/CHANGELOG.md +10 -0
  2. package/models/catalog.json +21 -7
  3. package/package.json +8 -1
  4. package/src/config/conversion/qwen3/qwen-3-5-9b-f16-af32.json +474 -0
  5. package/src/config/conversion/qwen3/qwen-3-5-9b-q4k-ehaf16.json +509 -0
  6. package/src/config/kernels/backward-registry.json +13 -2
  7. package/src/config/kernels/kernel-ref-digests.js +8 -4
  8. package/src/config/kernels/registry.json +1035 -2677
  9. package/src/config/runtime/profiles/qwen-3-5-9b-metal-correctness.json +46 -0
  10. package/src/config/schema/training-metrics.schema.js +4 -0
  11. package/src/experimental/bridge/trainer-artifact-bridge.d.ts +161 -0
  12. package/src/experimental/bridge/trainer-artifact-bridge.js +573 -0
  13. package/src/experimental/training/attention-backward.js +123 -55
  14. package/src/experimental/training/autograd.js +101 -5
  15. package/src/experimental/training/clip.js +6 -1
  16. package/src/experimental/training/datasets/text-pairs.js +80 -8
  17. package/src/experimental/training/distillation/student-fixture.js +470 -114
  18. package/src/experimental/training/lora-pipeline.js +202 -27
  19. package/src/experimental/training/lora.js +10 -4
  20. package/src/experimental/training/optimizer.js +9 -1
  21. package/src/experimental/training/runner.js +18 -4
  22. package/src/experimental/training/workloads.js +11 -0
  23. package/src/gpu/kernels/backward/adam.js +6 -4
  24. package/src/gpu/kernels/backward/causal_conv1d_silu_backward.d.ts +24 -0
  25. package/src/gpu/kernels/backward/causal_conv1d_silu_backward.js +74 -0
  26. package/src/gpu/kernels/backward/causal_conv1d_silu_backward.wgsl +59 -0
  27. package/src/gpu/kernels/backward/cross_entropy_backward.wgsl +4 -0
  28. package/src/gpu/kernels/backward/gated_delta_recurrent_backward.d.ts +39 -0
  29. package/src/gpu/kernels/backward/gated_delta_recurrent_backward.js +124 -0
  30. package/src/gpu/kernels/backward/gated_delta_recurrent_backward.wgsl +177 -0
  31. package/src/gpu/kernels/backward/gated_rmsnorm_backward.d.ts +32 -0
  32. package/src/gpu/kernels/backward/gated_rmsnorm_backward.js +90 -0
  33. package/src/gpu/kernels/backward/gated_rmsnorm_backward.wgsl +86 -0
  34. package/src/gpu/kernels/backward/index.d.ts +3 -0
  35. package/src/gpu/kernels/backward/index.js +3 -0
  36. package/src/gpu/kernels/backward/matmul_backward.wgsl +6 -6
  37. package/src/gpu/kernels/backward/matmul_backward_f16w.wgsl +82 -0
  38. package/src/gpu/kernels/backward/rmsnorm_backward.js +26 -4
  39. package/src/gpu/kernels/backward/rmsnorm_backward.wgsl +6 -3
  40. package/src/gpu/kernels/backward/utils.d.ts +12 -2
  41. package/src/gpu/kernels/backward/utils.js +9 -2
  42. package/src/gpu/kernels/cast_f16_to_f32.wgsl +6 -2
  43. package/src/gpu/kernels/index.d.ts +5 -0
  44. package/src/gpu/kernels/index.js +5 -0
  45. package/src/gpu/kernels/matmul-selection.js +5 -0
  46. package/src/gpu/kernels/silu.d.ts +1 -0
  47. package/src/gpu/kernels/silu.js +19 -4
  48. package/src/inference/pipelines/text/embed.js +1 -1
  49. package/src/inference/pipelines/text/execution-v1.js +8 -2
  50. package/src/inference/pipelines/text/ffn/dense.js +2 -0
  51. package/src/inference/pipelines/text/generator.js +29 -0
  52. package/src/inference/pipelines/text/lora-apply.js +4 -4
  53. package/src/inference/pipelines/text/moe-cpu.js +1 -0
  54. package/src/inference/pipelines/text/moe-gpu.js +1 -0
  55. package/src/inference/tokenizers/bundled.js +105 -32
  56. package/src/rules/kernels/matmul.rules.json +1 -0
  57. package/src/tooling/node-converter.js +20 -5
  58. package/src/tooling-experimental-exports.shared.d.ts +25 -0
  59. package/src/tooling-experimental-exports.shared.js +18 -0
  60. package/src/version.js +1 -1
@@ -95,6 +95,67 @@ function resolveByteLevelPretokenizerConfig(preTokenizer) {
95
95
  };
96
96
  }
97
97
 
98
+ function normalizeInlineCaseInsensitiveGroups(pattern) {
99
+ let source = String(pattern || '');
100
+ let caseInsensitive = false;
101
+ let start = source.indexOf('(?i:');
102
+ while (start !== -1) {
103
+ let depth = 1;
104
+ let escaped = false;
105
+ let end = start + 4;
106
+ for (; end < source.length; end++) {
107
+ const char = source[end];
108
+ if (escaped) {
109
+ escaped = false;
110
+ continue;
111
+ }
112
+ if (char === '\\') {
113
+ escaped = true;
114
+ continue;
115
+ }
116
+ if (char === '(') depth += 1;
117
+ else if (char === ')') {
118
+ depth -= 1;
119
+ if (depth === 0) break;
120
+ }
121
+ }
122
+ if (depth !== 0) {
123
+ throw new Error('[Tokenizer] Unterminated inline case-insensitive regex group.');
124
+ }
125
+ source = `${source.slice(0, start)}(?:${source.slice(start + 4, end)})${source.slice(end + 1)}`;
126
+ caseInsensitive = true;
127
+ start = source.indexOf('(?i:', start + 3);
128
+ }
129
+ return { source, caseInsensitive };
130
+ }
131
+
132
+ function resolveIsolatedSplitPretokenizer(preTokenizer) {
133
+ if (!preTokenizer || typeof preTokenizer !== 'object') return null;
134
+ if (preTokenizer.type === 'Sequence' && Array.isArray(preTokenizer.pretokenizers)) {
135
+ for (const entry of preTokenizer.pretokenizers) {
136
+ const resolved = resolveIsolatedSplitPretokenizer(entry);
137
+ if (resolved) return resolved;
138
+ }
139
+ return null;
140
+ }
141
+ if (
142
+ preTokenizer.type !== 'Split'
143
+ || preTokenizer.behavior !== 'Isolated'
144
+ || preTokenizer.invert === true
145
+ || typeof preTokenizer.pattern?.Regex !== 'string'
146
+ ) {
147
+ return null;
148
+ }
149
+ const normalized = normalizeInlineCaseInsensitiveGroups(preTokenizer.pattern.Regex);
150
+ const flags = `gu${normalized.caseInsensitive ? 'i' : ''}`;
151
+ try {
152
+ new RegExp(normalized.source, flags);
153
+ } catch (error) {
154
+ throw new Error(`[Tokenizer] Unsupported split pre-tokenizer regex: ${error.message}`);
155
+ }
156
+ return { source: normalized.source, flags };
157
+ }
158
+
98
159
  function hexNibble(code) {
99
160
  if (code >= 48 && code <= 57) return code - 48;
100
161
  if (code >= 65 && code <= 70) return code - 55;
@@ -388,6 +449,8 @@ export class BundledTokenizer extends BaseTokenizer {
388
449
 
389
450
  #useByteLevelEncoding = false;
390
451
 
452
+ #splitPretokenizer = null;
453
+
391
454
 
392
455
  constructor(config = {}) {
393
456
  // BundledTokenizer gets vocabSize from load(), so defer validation
@@ -412,6 +475,7 @@ export class BundledTokenizer extends BaseTokenizer {
412
475
  this.#byteDecoder = null;
413
476
  this.#byteEncoder = null;
414
477
  this.#useByteLevelEncoding = false;
478
+ this.#splitPretokenizer = null;
415
479
  this.vocabSize = 0;
416
480
  }
417
481
 
@@ -574,6 +638,7 @@ export class BundledTokenizer extends BaseTokenizer {
574
638
  // Handle behavior flags (use HF config if present, else runtime defaults)
575
639
  const runtimeDefaults = getRuntimeConfig().inference.tokenizer;
576
640
  const byteLevelPretokenizer = resolveByteLevelPretokenizerConfig(hf.pre_tokenizer);
641
+ this.#splitPretokenizer = resolveIsolatedSplitPretokenizer(hf.pre_tokenizer);
577
642
  const configuredAddBosToken = this.addBosToken;
578
643
  const configuredAddEosToken = this.addEosToken;
579
644
  const inferredFlags = inferBundledTokenizerBehaviorFlags(hf, this.specialTokens);
@@ -741,6 +806,7 @@ export class BundledTokenizer extends BaseTokenizer {
741
806
 
742
807
  const runtimeDefaults = getRuntimeConfig().inference.tokenizer;
743
808
  const byteLevelPretokenizer = resolveByteLevelPretokenizerConfig(tokenizerJson.pre_tokenizer);
809
+ this.#splitPretokenizer = resolveIsolatedSplitPretokenizer(tokenizerJson.pre_tokenizer);
744
810
  const configuredAddBosToken = this.addBosToken;
745
811
  const configuredAddEosToken = this.addEosToken;
746
812
  const inferredFlags = inferBundledTokenizerBehaviorFlags(tokenizerJson, this.specialTokens);
@@ -953,48 +1019,55 @@ export class BundledTokenizer extends BaseTokenizer {
953
1019
  if (text.length === 0) return [];
954
1020
 
955
1021
  let normalized = text;
956
- let prefixed;
957
- if (this.#useByteLevelEncoding) {
958
- if (this.#addSpacePrefix && !normalized.startsWith(' ')) {
959
- normalized = ` ${normalized}`;
960
- }
961
- prefixed = this.#encodeByteLevelText(normalized);
962
- } else {
963
- if (this.#addSpacePrefix && !normalized.startsWith(' ')) {
964
- normalized = ` ${normalized}`;
965
- }
966
- const sp = this.#spacePrefixChar;
967
- prefixed = normalized.replace(/ /g, sp);
968
- }
969
-
970
- if (this.#mergeRanks.size === 0) {
971
- return this.#encodeBPEGreedy(prefixed);
972
- }
973
-
974
- const tokens = this.#bpeTokenize(prefixed);
975
-
976
1022
  const ids = [];
977
- for (const token of tokens) {
978
- const id = this.#vocab.get(token);
979
- if (id !== undefined) {
980
- ids.push(id);
1023
+ if (this.#addSpacePrefix && !normalized.startsWith(' ')) {
1024
+ normalized = ` ${normalized}`;
1025
+ }
1026
+ const pretokens = this.#splitBpePretokens(normalized);
1027
+ for (const pretoken of pretokens) {
1028
+ const encoded = this.#useByteLevelEncoding
1029
+ ? this.#encodeByteLevelText(pretoken)
1030
+ : pretoken.replace(/ /g, this.#spacePrefixChar);
1031
+ if (this.#mergeRanks.size === 0) {
1032
+ ids.push(...this.#encodeBPEGreedy(encoded));
981
1033
  continue;
982
1034
  }
983
- const bytes = new TextEncoder().encode(token);
984
- for (const b of bytes) {
985
- const byteId = this.#byteTokens.get(b);
986
- if (byteId !== undefined) {
987
- ids.push(byteId);
1035
+ for (const token of this.#bpeTokenize(encoded)) {
1036
+ const id = this.#vocab.get(token);
1037
+ if (id !== undefined) {
1038
+ ids.push(id);
988
1039
  continue;
989
1040
  }
990
- const byteToken = `<0x${b.toString(16).padStart(2, '0').toUpperCase()}>`;
991
- ids.push(this.#vocab.get(byteToken) ?? this.#getUnkTokenId());
1041
+ const bytes = new TextEncoder().encode(token);
1042
+ for (const b of bytes) {
1043
+ const byteId = this.#byteTokens.get(b);
1044
+ if (byteId !== undefined) {
1045
+ ids.push(byteId);
1046
+ continue;
1047
+ }
1048
+ const byteToken = `<0x${b.toString(16).padStart(2, '0').toUpperCase()}>`;
1049
+ ids.push(this.#vocab.get(byteToken) ?? this.#getUnkTokenId());
1050
+ }
992
1051
  }
993
1052
  }
994
-
995
1053
  return ids;
996
1054
  }
997
1055
 
1056
+ #splitBpePretokens(text) {
1057
+ if (!this.#useByteLevelEncoding || !this.#splitPretokenizer) return [text];
1058
+ const regex = new RegExp(this.#splitPretokenizer.source, this.#splitPretokenizer.flags);
1059
+ const pretokens = [];
1060
+ let cursor = 0;
1061
+ for (const match of text.matchAll(regex)) {
1062
+ const index = match.index ?? cursor;
1063
+ if (index > cursor) pretokens.push(text.slice(cursor, index));
1064
+ if (match[0]) pretokens.push(match[0]);
1065
+ cursor = index + match[0].length;
1066
+ }
1067
+ if (cursor < text.length) pretokens.push(text.slice(cursor));
1068
+ return pretokens.length > 0 ? pretokens : [text];
1069
+ }
1070
+
998
1071
 
999
1072
  #encodeBPEGreedy(text) {
1000
1073
 
@@ -74,6 +74,7 @@
74
74
  { "match": {}, "value": { "context": "variantWeightDtype" } }
75
75
  ],
76
76
  "weightDtypeMismatchPolicy": [
77
+ { "match": { "actualWeightDtype": "f16", "requiredWeightDtype": "q4k" }, "value": "fallthrough" },
77
78
  { "match": { "actualWeightDtype": "q4k", "requiredWeightDtype": "f16", "hasSubgroups": true }, "value": "fallthrough" },
78
79
  { "match": { "actualWeightDtype": "q4k", "requiredWeightDtype": "f32", "hasSubgroups": true }, "value": "fallthrough" },
79
80
  { "match": { "actualWeightDtype": "litert_int4", "requiredWeightDtype": "f16" }, "value": "fallthrough" },
@@ -275,10 +275,10 @@ async function loadNodeGpuCastRuntime() {
275
275
  gpuCastRuntimePromise = (async () => {
276
276
  await bootstrapNodeWebGPU();
277
277
  const [
278
- { initDevice, getDevice },
278
+ { initDevice, getDevice, destroyDevice, resetDeviceState },
279
279
  { castF32ToF16, runBF16ToF16 },
280
280
  { createTensor },
281
- { acquireBuffer, releaseBuffer, getBufferPool },
281
+ { acquireBuffer, releaseBuffer, getBufferPool, destroyBufferPool },
282
282
  ] = await Promise.all([
283
283
  import('../gpu/device.js'),
284
284
  import('../gpu/kernel-selector.js'),
@@ -293,12 +293,15 @@ async function loadNodeGpuCastRuntime() {
293
293
  }
294
294
  return {
295
295
  getDevice,
296
+ destroyDevice,
297
+ resetDeviceState,
296
298
  castF32ToF16,
297
299
  runBF16ToF16,
298
300
  createTensor,
299
301
  acquireBuffer,
300
302
  releaseBuffer,
301
303
  getBufferPool,
304
+ destroyBufferPool,
302
305
  };
303
306
  })();
304
307
  }
@@ -1056,6 +1059,7 @@ export async function convertSafetensorsDirectory(options) {
1056
1059
  installNodeFileFetchShim();
1057
1060
  const fileRangeReader = createFileRangeReader();
1058
1061
  const totalTimer = createStageTimer('Total');
1062
+ let gpuRuntimeForCleanup = null;
1059
1063
  try {
1060
1064
 
1061
1065
  const [
@@ -1457,9 +1461,8 @@ export async function convertSafetensorsDirectory(options) {
1457
1461
  let result = null;
1458
1462
  try {
1459
1463
  if (executionPlan.useGpuCast) {
1460
- let gpuRuntime;
1461
1464
  try {
1462
- gpuRuntime = await loadNodeGpuCastRuntime();
1465
+ gpuRuntimeForCleanup = await loadNodeGpuCastRuntime();
1463
1466
  } catch (error) {
1464
1467
  const message = error instanceof Error ? error.message : String(error);
1465
1468
  throw new Error(
@@ -1467,7 +1470,7 @@ export async function convertSafetensorsDirectory(options) {
1467
1470
  );
1468
1471
  }
1469
1472
  gpuTensorTransformer = createNodeGpuTensorTransformer({
1470
- runtime: gpuRuntime,
1473
+ runtime: gpuRuntimeForCleanup,
1471
1474
  gpuCastMinTensorBytes: executionPlan.gpuCastMinTensorBytes,
1472
1475
  requireGpuCast: executionPlan.gpuCastRequestedExplicitly === true,
1473
1476
  resolveTensorTargetQuant,
@@ -1609,6 +1612,18 @@ export async function convertSafetensorsDirectory(options) {
1609
1612
  };
1610
1613
  } finally {
1611
1614
  await fileRangeReader.closeAll();
1615
+ if (gpuRuntimeForCleanup) {
1616
+ try {
1617
+ gpuRuntimeForCleanup.destroyBufferPool();
1618
+ } finally {
1619
+ try {
1620
+ gpuRuntimeForCleanup.destroyDevice();
1621
+ } finally {
1622
+ gpuRuntimeForCleanup.resetDeviceState();
1623
+ gpuCastRuntimePromise = null;
1624
+ }
1625
+ }
1626
+ }
1612
1627
  totalTimer.stop();
1613
1628
  }
1614
1629
  }
@@ -6,6 +6,31 @@
6
6
 
7
7
  export { convertModel, createRemoteModelSources, isConversionSupported } from './experimental/browser/browser-converter.js';
8
8
  export { pickModelDirectory, pickModelFiles } from './experimental/browser/file-picker.js';
9
+ export {
10
+ TRAINER_ARTIFACT_BRIDGE_SCHEMA_ID,
11
+ TRAINER_ARTIFACT_IMPORT_PLAN_SCHEMA_ID,
12
+ TRAINER_ARTIFACT_PARITY_EVIDENCE_SCHEMA_ID,
13
+ TRAINER_ARTIFACT_PARITY_RECEIPT_SCHEMA_ID,
14
+ TRAINER_ARTIFACT_KIND_FULL_CHECKPOINT,
15
+ TRAINER_ARTIFACT_KIND_PEFT_ADAPTER,
16
+ TRANSLATION_FULL_CHECKPOINT_PARITY_CHECKS,
17
+ COLUMBO_QWEN_ADAPTER_PARITY_CHECKS,
18
+ normalizeGammaTrainerArtifactHandoff,
19
+ validateTrainerArtifactBridgeDescriptor,
20
+ assertTrainerArtifactCandidateEntry,
21
+ buildTrainerArtifactImportPlan,
22
+ buildTrainerArtifactParityTemplate,
23
+ verifyTrainerArtifactParityEvidence,
24
+ } from './experimental/bridge/trainer-artifact-bridge.js';
25
+ export type {
26
+ TrainerArtifactArchitecture,
27
+ TrainerArtifactBridgeDescriptor,
28
+ TrainerArtifactFileIdentity,
29
+ TrainerArtifactImportPlan,
30
+ TrainerArtifactParityEvidence,
31
+ TrainerArtifactParityEvidenceCheck,
32
+ TrainerArtifactVerificationReceipt,
33
+ } from './experimental/bridge/trainer-artifact-bridge.js';
9
34
  export {
10
35
  P2P_WEBRTC_DATA_PLANE_CONTRACT_VERSION,
11
36
  isBrowserWebRTCAvailable,
@@ -8,6 +8,24 @@
8
8
  export { convertModel, createRemoteModelSources, isConversionSupported } from './experimental/browser/browser-converter.js';
9
9
  export { pickModelDirectory, pickModelFiles } from './experimental/browser/file-picker.js';
10
10
 
11
+ // Trainer artifact bridge contract shared by Gamma handoffs and Columbo adapters.
12
+ export {
13
+ TRAINER_ARTIFACT_BRIDGE_SCHEMA_ID,
14
+ TRAINER_ARTIFACT_IMPORT_PLAN_SCHEMA_ID,
15
+ TRAINER_ARTIFACT_PARITY_EVIDENCE_SCHEMA_ID,
16
+ TRAINER_ARTIFACT_PARITY_RECEIPT_SCHEMA_ID,
17
+ TRAINER_ARTIFACT_KIND_FULL_CHECKPOINT,
18
+ TRAINER_ARTIFACT_KIND_PEFT_ADAPTER,
19
+ TRANSLATION_FULL_CHECKPOINT_PARITY_CHECKS,
20
+ COLUMBO_QWEN_ADAPTER_PARITY_CHECKS,
21
+ normalizeGammaTrainerArtifactHandoff,
22
+ validateTrainerArtifactBridgeDescriptor,
23
+ assertTrainerArtifactCandidateEntry,
24
+ buildTrainerArtifactImportPlan,
25
+ buildTrainerArtifactParityTemplate,
26
+ verifyTrainerArtifactParityEvidence,
27
+ } from './experimental/bridge/trainer-artifact-bridge.js';
28
+
11
29
  // Distribution tooling
12
30
  export {
13
31
  P2P_WEBRTC_DATA_PLANE_CONTRACT_VERSION,
package/src/version.js CHANGED
@@ -1,2 +1,2 @@
1
- export const DOPPLER_VERSION = '0.4.8';
1
+ export const DOPPLER_VERSION = '0.4.9';
2
2
  export const DOPPLER_PROVIDER_VERSION = DOPPLER_VERSION;