doppler-gpu 0.4.10 → 0.4.14
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.
- package/CHANGELOG.md +95 -0
- package/README.md +40 -1
- package/models/catalog.json +637 -32
- package/package.json +24 -3
- package/src/client/doppler-registry.js +6 -0
- package/src/client/doppler-registry.json +135 -13
- package/src/client/runtime/lora.d.ts +10 -0
- package/src/client/runtime/lora.js +69 -0
- package/src/client/runtime/model-manager.js +6 -167
- package/src/client/runtime/model-session.d.ts +3 -1
- package/src/client/runtime/model-session.js +6 -0
- package/src/client/wrap-pipeline-handle.js +7 -0
- package/src/config/conversion/amplify/amplify-120m-f16-af32.json +516 -0
- package/src/config/conversion/esm/esm2-t12-35m-ur50d-f32-af32.json +571 -0
- package/src/config/conversion/esmc/esmc-300m-f32-af32.json +342 -0
- package/src/config/conversion/nucleotide-transformer/nucleotide-transformer-v2-50m-f32-af32.json +388 -0
- package/src/config/conversion/qwen3/qwen-3-5-0-8b-q4k-ehaf16.json +3 -2
- package/src/config/kernels/kernel-ref-digests.js +2 -1
- package/src/config/kernels/registry.json +73 -11
- package/src/config/merge.js +24 -0
- package/src/config/required-inference-fields-contract-check.js +2 -0
- package/src/config/schema/adapter.schema.d.ts +6 -1
- package/src/config/schema/adapter.schema.js +5 -0
- package/src/config/schema/converter.schema.d.ts +7 -0
- package/src/config/schema/converter.schema.js +15 -0
- package/src/config/schema/inference.schema.d.ts +8 -0
- package/src/config/schema/manifest.schema.d.ts +32 -1
- package/src/config/schema/manifest.schema.js +26 -0
- package/src/converter/conversion-plan.js +15 -1
- package/src/converter/core.js +5 -4
- package/src/converter/source-tensor-rules.d.ts +50 -0
- package/src/converter/source-tensor-rules.js +180 -0
- package/src/converter/source-tokenizer.d.ts +45 -0
- package/src/converter/source-tokenizer.js +155 -0
- package/src/experimental/adapters/artifact-contract.d.ts +83 -0
- package/src/experimental/adapters/artifact-contract.js +254 -0
- package/src/experimental/adapters/index.d.ts +25 -0
- package/src/experimental/adapters/index.js +12 -0
- package/src/experimental/adapters/lora-loader.js +24 -4
- package/src/experimental/bridge/trainer-artifact-bridge.d.ts +2 -1
- package/src/experimental/bridge/trainer-artifact-bridge.js +35 -13
- package/src/experimental/training/artifacts.d.ts +160 -0
- package/src/experimental/training/attention-backward.d.ts +64 -0
- package/src/experimental/training/attention-forward.d.ts +22 -0
- package/src/experimental/training/attention-forward.js +82 -0
- package/src/experimental/training/autograd.d.ts +76 -0
- package/src/experimental/training/checkpoint.d.ts +36 -0
- package/src/experimental/training/clip.d.ts +9 -0
- package/src/experimental/training/dataloader.d.ts +8 -0
- package/src/experimental/training/distillation/student-fixture.js +1 -0
- package/src/experimental/training/export.d.ts +41 -0
- package/src/experimental/training/lora-pipeline.d.ts +186 -0
- package/src/experimental/training/lora-pipeline.js +5 -0
- package/src/experimental/training/lora.d.ts +20 -0
- package/src/experimental/training/lora.js +4 -3
- package/src/experimental/training/loss-scaling.d.ts +21 -0
- package/src/experimental/training/loss.d.ts +10 -0
- package/src/experimental/training/native-qwen-lora.d.ts +106 -0
- package/src/experimental/training/native-qwen-lora.js +449 -0
- package/src/experimental/training/objectives/base.d.ts +58 -0
- package/src/experimental/training/optimizer.d.ts +22 -0
- package/src/experimental/training/optimizer.js +1 -0
- package/src/experimental/training/runner.d.ts +249 -0
- package/src/experimental/training/trainer.d.ts +89 -0
- package/src/experimental/training/workloads.d.ts +205 -0
- package/src/gpu/kernels/backward/adam.d.ts +1 -0
- package/src/gpu/kernels/backward/adam.js +6 -2
- package/src/gpu/kernels/backward/adam.wgsl +3 -2
- package/src/gpu/kernels/backward/matmul_backward_q4k.wgsl +90 -0
- package/src/gpu/kernels/backward/utils.js +18 -4
- package/src/inference/pipelines/text/attention/attn-config.js +2 -0
- package/src/inference/pipelines/text/attention/projections.js +161 -24
- package/src/inference/pipelines/text/attention/record.js +67 -9
- package/src/inference/pipelines/text/attention/run.js +60 -7
- package/src/inference/pipelines/text/attention/types.js +18 -9
- package/src/inference/pipelines/text/config.d.ts +5 -0
- package/src/inference/pipelines/text/config.js +91 -0
- package/src/inference/pipelines/text/ffn/dense.js +205 -2
- package/src/inference/pipelines/text/ffn/standard.js +50 -3
- package/src/inference/pipelines/text/generator-helpers.js +11 -1
- package/src/inference/pipelines/text/generator-runtime.js +63 -15
- package/src/inference/pipelines/text/generator-steps.js +3 -1
- package/src/inference/pipelines/text/generator.d.ts +5 -1
- package/src/inference/pipelines/text/generator.js +196 -3
- package/src/inference/pipelines/text/init-qkv-fusion.js +85 -1
- package/src/inference/pipelines/text/init.d.ts +2 -0
- package/src/inference/pipelines/text/init.js +2 -0
- package/src/inference/pipelines/text/layer.js +6 -0
- package/src/inference/pipelines/text/linear-attention.d.ts +2 -0
- package/src/inference/pipelines/text/linear-attention.js +77 -6
- package/src/inference/pipelines/text/logits/cpu.d.ts +6 -0
- package/src/inference/pipelines/text/logits/cpu.js +22 -0
- package/src/inference/pipelines/text/logits/index.d.ts +1 -1
- package/src/inference/pipelines/text/logits/index.js +85 -11
- package/src/inference/pipelines/text/logits/types.d.ts +4 -0
- package/src/inference/pipelines/text/lora-types.d.ts +6 -1
- package/src/inference/pipelines/text/lora-types.js +5 -0
- package/src/inference/pipelines/text/ops.js +33 -0
- package/src/inference/pipelines/text/types.d.ts +66 -0
- package/src/inference/pipelines/text/weights.d.ts +10 -0
- package/src/inference/pipelines/text/weights.js +33 -1
- package/src/inference/pipelines/text.d.ts +9 -2
- package/src/inference/pipelines/text.js +116 -0
- package/src/inference/tokenizers/behavior-flags.js +5 -1
- package/src/inference/tokenizers/bundled.js +99 -8
- package/src/loader/doppler-loader.d.ts +2 -0
- package/src/loader/doppler-loader.js +9 -0
- package/src/loader/final-weights-loader.d.ts +8 -0
- package/src/loader/final-weights-loader.js +42 -0
- package/src/loader/layer-loader.d.ts +6 -0
- package/src/loader/layer-loader.js +102 -3
- package/src/loader/loader-types.d.ts +15 -2
- package/src/models/diffusiongemma.js +1 -1
- package/src/models/embeddinggemma.js +1 -1
- package/src/models/gemma3.js +1 -1
- package/src/models/gemma4.js +1 -1
- package/src/models/qwen3.js +1 -1
- package/src/tooling/hf-registry-utils.js +1 -1
- package/src/tooling/node-browser-command-runner.d.ts +16 -0
- package/src/tooling/node-browser-command-runner.js +1 -1
- package/src/tooling/node-converter.js +33 -1
- package/src/tooling/node-webgpu.js +85 -2
- package/src/training-browser.d.ts +19 -0
- package/src/training-browser.js +13 -0
- package/src/training.d.ts +137 -0
- package/src/training.js +219 -0
- package/src/version.js +1 -1
- package/tests/kernels/browser/kernel-suite.js +9 -1
- package/tests/kernels/browser/test-page.js +4 -2
- package/tests/kernels/reference/adam.js +4 -2
- package/NOTICE +0 -5
|
@@ -29,6 +29,19 @@ export const COLUMBO_QWEN_ADAPTER_PARITY_CHECKS = Object.freeze([
|
|
|
29
29
|
'browser_capability_evaluation',
|
|
30
30
|
]);
|
|
31
31
|
|
|
32
|
+
export const TINKER_PEFT_BROWSER_ADAPTER_PARITY_CHECKS = Object.freeze([
|
|
33
|
+
'source_artifact_byte_identity',
|
|
34
|
+
'tokenizer_and_prompt_identity',
|
|
35
|
+
'architecture_and_conversion_lineage',
|
|
36
|
+
'evaluation_input_identity',
|
|
37
|
+
'tinker_checkpoint_vs_peft_reference',
|
|
38
|
+
'peft_reference_vs_doppler_f16',
|
|
39
|
+
'doppler_f16_vs_quantized_browser',
|
|
40
|
+
'gamma_sealed_task_gain',
|
|
41
|
+
'gamma_retention_floor',
|
|
42
|
+
'browser_capability_evaluation',
|
|
43
|
+
]);
|
|
44
|
+
|
|
32
45
|
const SHA256_PATTERN = /^[0-9a-f]{64}$/;
|
|
33
46
|
const FILE_ROLES_BY_KIND = Object.freeze({
|
|
34
47
|
[TRAINER_ARTIFACT_KIND_FULL_CHECKPOINT]: Object.freeze([
|
|
@@ -44,9 +57,22 @@ const FILE_ROLES_BY_KIND = Object.freeze({
|
|
|
44
57
|
'training_report',
|
|
45
58
|
]),
|
|
46
59
|
});
|
|
47
|
-
const
|
|
48
|
-
translation_full_checkpoint:
|
|
49
|
-
|
|
60
|
+
const PARITY_PROFILES = Object.freeze({
|
|
61
|
+
translation_full_checkpoint: Object.freeze({
|
|
62
|
+
artifactKind: TRAINER_ARTIFACT_KIND_FULL_CHECKPOINT,
|
|
63
|
+
selectionAuthority: 'clocksmith/gamma',
|
|
64
|
+
requiredChecks: TRANSLATION_FULL_CHECKPOINT_PARITY_CHECKS,
|
|
65
|
+
}),
|
|
66
|
+
columbo_qwen_adapter: Object.freeze({
|
|
67
|
+
artifactKind: TRAINER_ARTIFACT_KIND_PEFT_ADAPTER,
|
|
68
|
+
selectionAuthority: 'clocksmith/columbo',
|
|
69
|
+
requiredChecks: COLUMBO_QWEN_ADAPTER_PARITY_CHECKS,
|
|
70
|
+
}),
|
|
71
|
+
tinker_peft_browser_adapter: Object.freeze({
|
|
72
|
+
artifactKind: TRAINER_ARTIFACT_KIND_PEFT_ADAPTER,
|
|
73
|
+
selectionAuthority: 'clocksmith/gamma',
|
|
74
|
+
requiredChecks: TINKER_PEFT_BROWSER_ADAPTER_PARITY_CHECKS,
|
|
75
|
+
}),
|
|
50
76
|
});
|
|
51
77
|
|
|
52
78
|
function stableJson(value) {
|
|
@@ -181,18 +207,16 @@ function normalizeParity(parity, artifactKind) {
|
|
|
181
207
|
throw new Error('trainer artifact bridge: parity must be an object.');
|
|
182
208
|
}
|
|
183
209
|
const profile = requireText(parity.profile, 'parity.profile');
|
|
184
|
-
const
|
|
185
|
-
if (!
|
|
210
|
+
const profileContract = PARITY_PROFILES[profile];
|
|
211
|
+
if (!profileContract) {
|
|
186
212
|
throw new Error(`trainer artifact bridge: unsupported parity profile "${profile}".`);
|
|
187
213
|
}
|
|
188
|
-
|
|
189
|
-
? 'translation_full_checkpoint'
|
|
190
|
-
: 'columbo_qwen_adapter';
|
|
191
|
-
if (profile !== expectedProfile) {
|
|
214
|
+
if (profileContract.artifactKind !== artifactKind) {
|
|
192
215
|
throw new Error(
|
|
193
|
-
`trainer artifact bridge:
|
|
216
|
+
`trainer artifact bridge: parity profile "${profile}" does not support artifact kind "${artifactKind}".`
|
|
194
217
|
);
|
|
195
218
|
}
|
|
219
|
+
const canonicalChecks = profileContract.requiredChecks;
|
|
196
220
|
const requiredChecks = Array.isArray(parity.requiredChecks) ? parity.requiredChecks : [];
|
|
197
221
|
if (stableJson(requiredChecks) !== stableJson(canonicalChecks)) {
|
|
198
222
|
throw new Error(`trainer artifact bridge: parity.requiredChecks must match profile "${profile}".`);
|
|
@@ -435,9 +459,7 @@ export function assertTrainerArtifactCandidateEntry(descriptor) {
|
|
|
435
459
|
if (normalized.selection.status !== 'selected' || !normalized.selection.receipt) {
|
|
436
460
|
throw new Error('trainer artifact bridge: candidate entry denied; selection receipt is required.');
|
|
437
461
|
}
|
|
438
|
-
const expectedAuthority = normalized.
|
|
439
|
-
? 'clocksmith/gamma'
|
|
440
|
-
: 'clocksmith/columbo';
|
|
462
|
+
const expectedAuthority = PARITY_PROFILES[normalized.parity.profile].selectionAuthority;
|
|
441
463
|
if (normalized.selection.authority !== expectedAuthority) {
|
|
442
464
|
throw new Error(
|
|
443
465
|
`trainer artifact bridge: candidate entry denied; selection authority must be "${expectedAuthority}".`
|
|
@@ -0,0 +1,160 @@
|
|
|
1
|
+
import type { TrainingConfigSchema } from '../../config/training-defaults.d.ts';
|
|
2
|
+
import type { DistillTrainingStage } from '../../config/schema/distill-training.schema.js';
|
|
3
|
+
import type { UlTrainingStage } from '../../config/schema/ul-training.schema.js';
|
|
4
|
+
|
|
5
|
+
export interface UlTrainingContract {
|
|
6
|
+
enabled: boolean;
|
|
7
|
+
stage: UlTrainingStage | null;
|
|
8
|
+
artifactDir: string | null;
|
|
9
|
+
stage1Artifact: string | null;
|
|
10
|
+
stage1ArtifactHash: string | null;
|
|
11
|
+
}
|
|
12
|
+
|
|
13
|
+
export interface DistillTrainingContract {
|
|
14
|
+
enabled: boolean;
|
|
15
|
+
stage: DistillTrainingStage | null;
|
|
16
|
+
artifactDir: string | null;
|
|
17
|
+
stageAArtifact: string | null;
|
|
18
|
+
stageAArtifactHash: string | null;
|
|
19
|
+
teacherModelId: string | null;
|
|
20
|
+
studentModelId: string | null;
|
|
21
|
+
datasetId: string | null;
|
|
22
|
+
datasetPath: string | null;
|
|
23
|
+
languagePair: string | null;
|
|
24
|
+
allowHintFallback: boolean;
|
|
25
|
+
}
|
|
26
|
+
|
|
27
|
+
export interface UlArtifactFinalizeResult {
|
|
28
|
+
kind: 'ul';
|
|
29
|
+
stage: UlTrainingStage;
|
|
30
|
+
runDir: string;
|
|
31
|
+
metricsPath: string;
|
|
32
|
+
manifestPath: string;
|
|
33
|
+
manifestHash: string;
|
|
34
|
+
manifestContentHash: string;
|
|
35
|
+
manifestFileHash: string;
|
|
36
|
+
stage1Dependency: {
|
|
37
|
+
path: string;
|
|
38
|
+
hash: string;
|
|
39
|
+
manifestHash: string | null;
|
|
40
|
+
} | null;
|
|
41
|
+
}
|
|
42
|
+
|
|
43
|
+
export interface DistillArtifactFinalizeResult {
|
|
44
|
+
kind: 'distill';
|
|
45
|
+
stage: DistillTrainingStage;
|
|
46
|
+
runDir: string;
|
|
47
|
+
metricsPath: string;
|
|
48
|
+
manifestPath: string;
|
|
49
|
+
manifestHash: string;
|
|
50
|
+
manifestContentHash: string;
|
|
51
|
+
manifestFileHash: string;
|
|
52
|
+
stageADependency: {
|
|
53
|
+
path: string;
|
|
54
|
+
hash: string;
|
|
55
|
+
manifestHash: string | null;
|
|
56
|
+
metricsHash: string;
|
|
57
|
+
metricsSummary: {
|
|
58
|
+
stepCount: number;
|
|
59
|
+
kdCount: number;
|
|
60
|
+
tripletCount: number;
|
|
61
|
+
kdMean: number | null;
|
|
62
|
+
tripletMean: number | null;
|
|
63
|
+
totalLossMean: number | null;
|
|
64
|
+
};
|
|
65
|
+
} | null;
|
|
66
|
+
}
|
|
67
|
+
|
|
68
|
+
export interface UlArtifactSession {
|
|
69
|
+
appendStep(entry: Record<string, unknown>): Promise<void>;
|
|
70
|
+
finalize(stepMetrics: Record<string, unknown>[]): Promise<UlArtifactFinalizeResult>;
|
|
71
|
+
}
|
|
72
|
+
|
|
73
|
+
export interface DistillArtifactSession {
|
|
74
|
+
appendStep(entry: Record<string, unknown>): Promise<void>;
|
|
75
|
+
finalize(stepMetrics: Record<string, unknown>[]): Promise<DistillArtifactFinalizeResult>;
|
|
76
|
+
}
|
|
77
|
+
|
|
78
|
+
export interface CreateUlArtifactSessionOptions {
|
|
79
|
+
config: TrainingConfigSchema;
|
|
80
|
+
stage: UlTrainingStage;
|
|
81
|
+
runOptions?: Record<string, unknown>;
|
|
82
|
+
}
|
|
83
|
+
|
|
84
|
+
export interface CreateDistillArtifactSessionOptions {
|
|
85
|
+
config: TrainingConfigSchema;
|
|
86
|
+
stage: DistillTrainingStage;
|
|
87
|
+
runOptions?: Record<string, unknown>;
|
|
88
|
+
}
|
|
89
|
+
|
|
90
|
+
export interface Stage1ArtifactContext {
|
|
91
|
+
manifestPath: string;
|
|
92
|
+
manifestHash: string;
|
|
93
|
+
ulContractHash: string | null;
|
|
94
|
+
latentDataset: {
|
|
95
|
+
path: string;
|
|
96
|
+
hash: string;
|
|
97
|
+
count: number;
|
|
98
|
+
summary: {
|
|
99
|
+
lambdaMean: number;
|
|
100
|
+
noisyStdMean: number;
|
|
101
|
+
cleanStdMean: number;
|
|
102
|
+
noiseStdMean: number;
|
|
103
|
+
scheduleMaxStep: number;
|
|
104
|
+
vectorCount: number;
|
|
105
|
+
};
|
|
106
|
+
entries: Record<string, unknown>[];
|
|
107
|
+
};
|
|
108
|
+
}
|
|
109
|
+
|
|
110
|
+
export interface StageAArtifactContext {
|
|
111
|
+
manifestPath: string;
|
|
112
|
+
manifestHash: string;
|
|
113
|
+
distillContractHash: string | null;
|
|
114
|
+
metrics: {
|
|
115
|
+
path: string;
|
|
116
|
+
hash: string;
|
|
117
|
+
count: number;
|
|
118
|
+
summary: {
|
|
119
|
+
stepCount: number;
|
|
120
|
+
kdCount: number;
|
|
121
|
+
tripletCount: number;
|
|
122
|
+
kdMean: number | null;
|
|
123
|
+
tripletMean: number | null;
|
|
124
|
+
totalLossMean: number | null;
|
|
125
|
+
};
|
|
126
|
+
entries: Record<string, unknown>[];
|
|
127
|
+
};
|
|
128
|
+
metricsSummary: {
|
|
129
|
+
stepCount: number;
|
|
130
|
+
kdCount: number;
|
|
131
|
+
tripletCount: number;
|
|
132
|
+
kdMean: number | null;
|
|
133
|
+
tripletMean: number | null;
|
|
134
|
+
totalLossMean: number | null;
|
|
135
|
+
};
|
|
136
|
+
}
|
|
137
|
+
|
|
138
|
+
export declare function resolveUlTrainingContract(
|
|
139
|
+
ulConfig: TrainingConfigSchema['training']['ul'] | null | undefined
|
|
140
|
+
): UlTrainingContract;
|
|
141
|
+
|
|
142
|
+
export declare function resolveDistillTrainingContract(
|
|
143
|
+
distillConfig: TrainingConfigSchema['training']['distill'] | null | undefined
|
|
144
|
+
): DistillTrainingContract;
|
|
145
|
+
|
|
146
|
+
export declare function createUlArtifactSession(
|
|
147
|
+
options: CreateUlArtifactSessionOptions
|
|
148
|
+
): Promise<UlArtifactSession | null>;
|
|
149
|
+
|
|
150
|
+
export declare function createDistillArtifactSession(
|
|
151
|
+
options: CreateDistillArtifactSessionOptions
|
|
152
|
+
): Promise<DistillArtifactSession | null>;
|
|
153
|
+
|
|
154
|
+
export declare function resolveStage1ArtifactContext(
|
|
155
|
+
config: TrainingConfigSchema
|
|
156
|
+
): Promise<Stage1ArtifactContext | null>;
|
|
157
|
+
|
|
158
|
+
export declare function resolveStageAArtifactContext(
|
|
159
|
+
config: TrainingConfigSchema
|
|
160
|
+
): Promise<StageAArtifactContext | null>;
|
|
@@ -0,0 +1,64 @@
|
|
|
1
|
+
import type { Tensor } from '../../gpu/tensor.js';
|
|
2
|
+
|
|
3
|
+
export interface AttentionBackwardOptions {
|
|
4
|
+
seqLen: number;
|
|
5
|
+
numHeads: number;
|
|
6
|
+
numKVHeads?: number;
|
|
7
|
+
headDim: number;
|
|
8
|
+
scale?: number;
|
|
9
|
+
causal?: boolean;
|
|
10
|
+
}
|
|
11
|
+
|
|
12
|
+
export interface AttentionGeometry {
|
|
13
|
+
seqLen: number;
|
|
14
|
+
numHeads: number;
|
|
15
|
+
numKVHeads: number;
|
|
16
|
+
headDim: number;
|
|
17
|
+
headsPerKV: number;
|
|
18
|
+
scale: number;
|
|
19
|
+
causal: boolean;
|
|
20
|
+
}
|
|
21
|
+
|
|
22
|
+
export interface AttentionBackwardResult {
|
|
23
|
+
gradQ: Tensor;
|
|
24
|
+
gradK: Tensor;
|
|
25
|
+
gradV: Tensor;
|
|
26
|
+
geometry: AttentionGeometry;
|
|
27
|
+
}
|
|
28
|
+
|
|
29
|
+
export interface AttentionBackwardDataResult {
|
|
30
|
+
dQ: Float32Array;
|
|
31
|
+
dK: Float32Array;
|
|
32
|
+
dV: Float32Array;
|
|
33
|
+
geometry: AttentionGeometry;
|
|
34
|
+
}
|
|
35
|
+
|
|
36
|
+
export declare function computeAttentionSoftmaxData(
|
|
37
|
+
qData: Float32Array,
|
|
38
|
+
kData: Float32Array,
|
|
39
|
+
options: AttentionBackwardOptions
|
|
40
|
+
): Float32Array;
|
|
41
|
+
|
|
42
|
+
export declare function computeAttentionBackwardData(
|
|
43
|
+
qData: Float32Array,
|
|
44
|
+
kData: Float32Array,
|
|
45
|
+
vData: Float32Array,
|
|
46
|
+
softmaxData: Float32Array,
|
|
47
|
+
gradOutputData: Float32Array,
|
|
48
|
+
options: AttentionBackwardOptions
|
|
49
|
+
): AttentionBackwardDataResult;
|
|
50
|
+
|
|
51
|
+
export declare function buildAttentionSoftmaxCache(
|
|
52
|
+
q: Tensor,
|
|
53
|
+
k: Tensor,
|
|
54
|
+
options: AttentionBackwardOptions
|
|
55
|
+
): Promise<Tensor>;
|
|
56
|
+
|
|
57
|
+
export declare function attentionBackwardCpu(
|
|
58
|
+
q: Tensor,
|
|
59
|
+
k: Tensor,
|
|
60
|
+
v: Tensor,
|
|
61
|
+
softmax: Tensor | null,
|
|
62
|
+
gradOutput: Tensor,
|
|
63
|
+
options: AttentionBackwardOptions
|
|
64
|
+
): Promise<AttentionBackwardResult>;
|
|
@@ -0,0 +1,22 @@
|
|
|
1
|
+
import type { TrainingConfigSchema } from '../../config/training-defaults.d.ts';
|
|
2
|
+
import type { Tensor } from '../../gpu/tensor.js';
|
|
3
|
+
|
|
4
|
+
export declare function recordAttentionForward(
|
|
5
|
+
q: Tensor,
|
|
6
|
+
k: Tensor,
|
|
7
|
+
v: Tensor,
|
|
8
|
+
config: TrainingConfigSchema,
|
|
9
|
+
tape: unknown,
|
|
10
|
+
options?: {
|
|
11
|
+
seqLen?: number;
|
|
12
|
+
numHeads?: number;
|
|
13
|
+
numKVHeads?: number;
|
|
14
|
+
headDim?: number;
|
|
15
|
+
scale?: number;
|
|
16
|
+
causal?: boolean;
|
|
17
|
+
startPos?: number;
|
|
18
|
+
attnSoftcap?: number;
|
|
19
|
+
slidingWindow?: number;
|
|
20
|
+
kvLen?: number;
|
|
21
|
+
}
|
|
22
|
+
): Promise<{ output: Tensor; softmax: Tensor | null }>;
|
|
@@ -0,0 +1,82 @@
|
|
|
1
|
+
import { runAttention } from '../../gpu/kernels/index.js';
|
|
2
|
+
import { buildAttentionSoftmaxCache } from './attention-backward.js';
|
|
3
|
+
import { OpType } from './autograd.js';
|
|
4
|
+
|
|
5
|
+
export async function recordAttentionForward(
|
|
6
|
+
q,
|
|
7
|
+
k,
|
|
8
|
+
v,
|
|
9
|
+
config,
|
|
10
|
+
tape,
|
|
11
|
+
options = {}
|
|
12
|
+
) {
|
|
13
|
+
if (!tape || typeof tape.record !== 'function') {
|
|
14
|
+
throw new Error('recordAttentionForward requires an autograd tape');
|
|
15
|
+
}
|
|
16
|
+
|
|
17
|
+
const {
|
|
18
|
+
seqLen,
|
|
19
|
+
numHeads,
|
|
20
|
+
numKVHeads = numHeads,
|
|
21
|
+
headDim,
|
|
22
|
+
scale,
|
|
23
|
+
causal = true,
|
|
24
|
+
startPos = 0,
|
|
25
|
+
attnSoftcap = 0,
|
|
26
|
+
slidingWindow = 0,
|
|
27
|
+
kvLen = seqLen,
|
|
28
|
+
} = options;
|
|
29
|
+
|
|
30
|
+
if (!seqLen || !numHeads || !headDim) {
|
|
31
|
+
throw new Error('recordAttentionForward requires seqLen, numHeads, and headDim');
|
|
32
|
+
}
|
|
33
|
+
|
|
34
|
+
const resolvedScale = scale ?? 1.0 / Math.sqrt(headDim);
|
|
35
|
+
const recomputeForward = Boolean(config?.training?.attention?.recomputeForward);
|
|
36
|
+
const softmax = recomputeForward
|
|
37
|
+
? null
|
|
38
|
+
: await buildAttentionSoftmaxCache(q, k, {
|
|
39
|
+
seqLen,
|
|
40
|
+
numHeads,
|
|
41
|
+
headDim,
|
|
42
|
+
scale: resolvedScale,
|
|
43
|
+
causal,
|
|
44
|
+
});
|
|
45
|
+
|
|
46
|
+
const attentionOptions = {
|
|
47
|
+
seqLen,
|
|
48
|
+
kvLen,
|
|
49
|
+
numHeads,
|
|
50
|
+
numKVHeads,
|
|
51
|
+
headDim,
|
|
52
|
+
scale: resolvedScale,
|
|
53
|
+
causal,
|
|
54
|
+
startPos,
|
|
55
|
+
attnSoftcap,
|
|
56
|
+
slidingWindow,
|
|
57
|
+
};
|
|
58
|
+
|
|
59
|
+
const output = await tape.record(
|
|
60
|
+
OpType.ATTENTION,
|
|
61
|
+
(inputQ, inputK, inputV) => runAttention(
|
|
62
|
+
inputQ,
|
|
63
|
+
inputK,
|
|
64
|
+
inputV,
|
|
65
|
+
null,
|
|
66
|
+
numHeads,
|
|
67
|
+
headDim,
|
|
68
|
+
attentionOptions
|
|
69
|
+
),
|
|
70
|
+
[q, k, v, softmax],
|
|
71
|
+
{
|
|
72
|
+
seqLen,
|
|
73
|
+
numHeads,
|
|
74
|
+
headDim,
|
|
75
|
+
scale: resolvedScale,
|
|
76
|
+
causal,
|
|
77
|
+
recomputeForward,
|
|
78
|
+
}
|
|
79
|
+
);
|
|
80
|
+
|
|
81
|
+
return { output, softmax };
|
|
82
|
+
}
|
|
@@ -0,0 +1,76 @@
|
|
|
1
|
+
import type { Tensor } from '../../gpu/tensor.js';
|
|
2
|
+
import type { BackwardRegistrySchema } from '../../config/schema/backward-registry.schema.js';
|
|
3
|
+
|
|
4
|
+
export const OpType: {
|
|
5
|
+
EMBED: 'embed';
|
|
6
|
+
MATMUL: 'matmul';
|
|
7
|
+
RMSNORM: 'rmsnorm';
|
|
8
|
+
RESIDUAL_ADD: 'residual_add';
|
|
9
|
+
RESHAPE: 'reshape';
|
|
10
|
+
ROW_SLICE: 'row_slice';
|
|
11
|
+
LAYERNORM: 'layernorm';
|
|
12
|
+
ATTENTION: 'attention';
|
|
13
|
+
SOFTMAX: 'softmax';
|
|
14
|
+
ROPE: 'rope';
|
|
15
|
+
SILU: 'silu';
|
|
16
|
+
SILU_ROWSPLIT: 'silu_rowsplit';
|
|
17
|
+
SILU_GATED: 'silu_gated';
|
|
18
|
+
GELU: 'gelu';
|
|
19
|
+
SCALE: 'scale';
|
|
20
|
+
CROSS_ENTROPY: 'cross_entropy';
|
|
21
|
+
BIAS_ADD: 'bias_add';
|
|
22
|
+
UPSAMPLE2D: 'upsample2d';
|
|
23
|
+
PIXEL_SHUFFLE: 'pixel_shuffle';
|
|
24
|
+
GROUPNORM: 'groupnorm';
|
|
25
|
+
CONV2D: 'conv2d';
|
|
26
|
+
};
|
|
27
|
+
|
|
28
|
+
export declare function resolveMatmulBackwardOptions<T extends Record<string, unknown> = Record<string, unknown>>(
|
|
29
|
+
options?: T
|
|
30
|
+
): T & {
|
|
31
|
+
computeGradInput: boolean;
|
|
32
|
+
computeGradWeight: boolean;
|
|
33
|
+
};
|
|
34
|
+
|
|
35
|
+
export declare function computeSiluGatedBackwardValues(
|
|
36
|
+
gate: ArrayLike<number>,
|
|
37
|
+
up: ArrayLike<number>,
|
|
38
|
+
gradOutput: ArrayLike<number>,
|
|
39
|
+
swigluLimit?: number
|
|
40
|
+
): {
|
|
41
|
+
gradGate: Float32Array;
|
|
42
|
+
gradUp: Float32Array;
|
|
43
|
+
};
|
|
44
|
+
|
|
45
|
+
export interface AutogradRecord {
|
|
46
|
+
op: string;
|
|
47
|
+
inputs: Tensor[];
|
|
48
|
+
output: Tensor;
|
|
49
|
+
options?: Record<string, unknown>;
|
|
50
|
+
}
|
|
51
|
+
|
|
52
|
+
export interface BackwardSeed {
|
|
53
|
+
tensor: Tensor;
|
|
54
|
+
grad: Tensor;
|
|
55
|
+
}
|
|
56
|
+
|
|
57
|
+
export type BackwardSeedInput =
|
|
58
|
+
| Tensor
|
|
59
|
+
| Map<Tensor, Tensor>
|
|
60
|
+
| BackwardSeed[]
|
|
61
|
+
| { seeds: BackwardSeed[] };
|
|
62
|
+
|
|
63
|
+
export declare class AutogradTape {
|
|
64
|
+
constructor(registry: BackwardRegistrySchema);
|
|
65
|
+
registry: BackwardRegistrySchema;
|
|
66
|
+
records: AutogradRecord[];
|
|
67
|
+
watch<T extends Tensor>(tensor: T): T;
|
|
68
|
+
record(
|
|
69
|
+
op: string,
|
|
70
|
+
fn: (...args: Tensor[]) => Promise<Tensor>,
|
|
71
|
+
inputs: Tensor[],
|
|
72
|
+
options?: Record<string, unknown>
|
|
73
|
+
): Promise<Tensor>;
|
|
74
|
+
backward(gradOutput: BackwardSeedInput): Promise<Map<Tensor, Tensor>>;
|
|
75
|
+
reset(): void;
|
|
76
|
+
}
|
|
@@ -0,0 +1,36 @@
|
|
|
1
|
+
export interface CheckpointStoreOptions {
|
|
2
|
+
dbName?: string;
|
|
3
|
+
storeName?: string;
|
|
4
|
+
version?: number;
|
|
5
|
+
nodeDir?: string;
|
|
6
|
+
configHash?: string;
|
|
7
|
+
datasetHash?: string;
|
|
8
|
+
tokenizerHash?: string;
|
|
9
|
+
optimizerHash?: string;
|
|
10
|
+
runtimeProfileId?: string;
|
|
11
|
+
kernelPathId?: string;
|
|
12
|
+
environmentMetadata?: unknown;
|
|
13
|
+
buildProvenance?: Record<string, unknown> | null;
|
|
14
|
+
priorCheckpointHash?: string;
|
|
15
|
+
expectedMetadata?: Record<string, unknown>;
|
|
16
|
+
forceResume?: boolean;
|
|
17
|
+
forceResumeReason?: string;
|
|
18
|
+
forceResumeSource?: string;
|
|
19
|
+
forceResumeOperator?: string | null;
|
|
20
|
+
}
|
|
21
|
+
|
|
22
|
+
export declare function saveCheckpoint(
|
|
23
|
+
key: string,
|
|
24
|
+
data: unknown,
|
|
25
|
+
options?: CheckpointStoreOptions
|
|
26
|
+
): Promise<{
|
|
27
|
+
key: string;
|
|
28
|
+
path: string | null;
|
|
29
|
+
metadata: Record<string, unknown>;
|
|
30
|
+
data: unknown;
|
|
31
|
+
}>;
|
|
32
|
+
|
|
33
|
+
export declare function loadCheckpoint(
|
|
34
|
+
key: string,
|
|
35
|
+
options?: CheckpointStoreOptions
|
|
36
|
+
): Promise<unknown | null>;
|
|
@@ -0,0 +1,9 @@
|
|
|
1
|
+
import type { TrainingConfigSchema } from '../../config/training-defaults.d.ts';
|
|
2
|
+
import type { Tensor } from '../../gpu/tensor.js';
|
|
3
|
+
|
|
4
|
+
import type { ClipMetrics } from './trainer.d.ts';
|
|
5
|
+
|
|
6
|
+
export declare function clipGradients(
|
|
7
|
+
grads: Map<Tensor, Tensor>,
|
|
8
|
+
config: TrainingConfigSchema
|
|
9
|
+
): Promise<ClipMetrics>;
|
|
@@ -0,0 +1,41 @@
|
|
|
1
|
+
import type { AdapterManifest } from '../adapters/adapter-manifest.js';
|
|
2
|
+
import type { LoRAModuleName } from '../../inference/pipelines/text/lora-types.js';
|
|
3
|
+
import type { Tensor } from '../../gpu/tensor.js';
|
|
4
|
+
|
|
5
|
+
export interface LoRAExportTensor {
|
|
6
|
+
name: string;
|
|
7
|
+
tensor: Tensor | Float32Array | GPUBuffer;
|
|
8
|
+
shape?: [number, number];
|
|
9
|
+
dtype?: 'f16' | 'f32';
|
|
10
|
+
}
|
|
11
|
+
|
|
12
|
+
export interface LoRAExportOptions {
|
|
13
|
+
id: string;
|
|
14
|
+
name: string;
|
|
15
|
+
baseModel: string;
|
|
16
|
+
rank: number;
|
|
17
|
+
alpha: number;
|
|
18
|
+
targetModules: LoRAModuleName[];
|
|
19
|
+
version?: string;
|
|
20
|
+
description?: string;
|
|
21
|
+
metadata?: Record<string, unknown>;
|
|
22
|
+
tensors: LoRAExportTensor[];
|
|
23
|
+
format?: 'base64' | 'array';
|
|
24
|
+
weightsFormat?: 'json' | 'safetensors';
|
|
25
|
+
weightsPath?: string;
|
|
26
|
+
pretty?: boolean;
|
|
27
|
+
}
|
|
28
|
+
|
|
29
|
+
export interface LoRAExportResult {
|
|
30
|
+
manifest: AdapterManifest;
|
|
31
|
+
json: string;
|
|
32
|
+
weights?: ArrayBuffer;
|
|
33
|
+
weightsSha256?: string;
|
|
34
|
+
weightsPath?: string;
|
|
35
|
+
}
|
|
36
|
+
|
|
37
|
+
export declare function serializeLoRASafetensors(
|
|
38
|
+
tensors: Array<{ name: string; shape: [number, number]; data: Float32Array | number[] }>
|
|
39
|
+
): ArrayBuffer;
|
|
40
|
+
|
|
41
|
+
export declare function exportLoRAAdapter(options: LoRAExportOptions): Promise<LoRAExportResult>;
|