doppler-gpu 0.4.15 → 0.5.1
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/README.md +9 -0
- package/package.json +25 -2
- package/src/cli/doppler-cli.js +253 -94
- package/src/cli/runtime-profiles.js +3 -7
- package/src/client/doppler-api.d.ts +28 -0
- package/src/client/doppler-api.js +2 -0
- package/src/client/inspection.d.ts +122 -0
- package/src/client/inspection.js +343 -0
- package/src/client/runtime/index.d.ts +54 -0
- package/src/client/runtime/index.js +190 -18
- package/src/client/runtime/model-manager.d.ts +0 -4
- package/src/client/runtime/model-manager.js +0 -28
- package/src/client/runtime/model-session.d.ts +48 -1
- package/src/client/runtime/model-session.js +149 -30
- package/src/client/runtime/model-source.d.ts +10 -1
- package/src/client/runtime/scoped-session.d.ts +117 -0
- package/src/client/runtime/scoped-session.js +257 -0
- package/src/client/runtime/types.d.ts +0 -1
- package/src/client/runtime/types.js +0 -1
- package/src/config/evidence/boundary-evidence-policy.json +54 -0
- package/src/config/evidence/source-intake-policy.json +122 -0
- package/src/config/inspection/observation-policies.json +80 -0
- package/src/config/param-validator.js +21 -36
- package/src/config/runtime/compute/f16-activations.json +1 -0
- package/src/config/runtime/compute/f16-batched.json +1 -0
- package/src/config/runtime/default.json +1 -0
- package/src/config/runtime/diagnostics/debug-logits.json +1 -0
- package/src/config/runtime/experiments/bench/gemma3-bench-q4k-alt.json +1 -0
- package/src/config/runtime/experiments/bench/gemma3-bench-q4k-batched.json +1 -0
- package/src/config/runtime/experiments/bench/gemma3-bench-q4k.json +1 -0
- package/src/config/runtime/experiments/bench/gemma4-bench-q4k.json +1 -0
- package/src/config/runtime/experiments/bench/qwen3-bench-q4k-alt.json +1 -0
- package/src/config/runtime/experiments/bench/qwen3-bench-q4k-batched.json +1 -0
- package/src/config/runtime/experiments/debug/gemma3-browser-diagnose-q4k.json +1 -1
- package/src/config/runtime/experiments/debug/gemma3-debug-q4k.json +1 -3
- package/src/config/runtime/experiments/debug/gemma4-e2b-browser-diagnose-q4k.json +1 -1
- package/src/config/runtime/experiments/debug/qwen35-browser-diagnose-q4k.json +1 -1
- package/src/config/runtime/experiments/verify/gemma3-verify.json +1 -3
- package/src/config/runtime/experiments/verify/gemma4-e2b-verify.json +1 -3
- package/src/config/runtime/experiments/verify/lfm2-verify.json +1 -3
- package/src/config/runtime/experiments/verify/translategemma-verify.json +1 -3
- package/src/config/runtime/model/gemma2-debug.json +1 -0
- package/src/config/runtime/model/gemma2-pipeline-debug.json +1 -0
- package/src/config/runtime/model/gemma2-pipeline.json +1 -0
- package/src/config/runtime/model/gemma3-layer-probe.json +1 -0
- package/src/config/runtime/model/qwen3-5-layer-probe.json +1 -0
- package/src/config/runtime/model/qwen3-5-layer3-boundary-probe.json +1 -0
- package/src/config/runtime/model/qwen3-5-linear-attn-debug.json +1 -0
- package/src/config/runtime/optimization-candidates/registry.json +5 -0
- package/src/config/runtime/platform/metal-apple-q4k.json +1 -0
- package/src/config/runtime/profiles/default.json +1 -0
- package/src/config/runtime/profiles/diffusiongemma-26b-a4b-throughput.json +1 -0
- package/src/config/runtime/profiles/gemma3-270m-q4k-throughput-overlapped-probe.json +1 -0
- package/src/config/runtime/profiles/gemma3-270m-q4k-throughput-probe.json +1 -0
- package/src/config/runtime/profiles/gemma4-12b-rdrr-safe-b4.json +1 -3
- package/src/config/runtime/profiles/gemma4-12b-rdrr-safe-single.json +1 -3
- package/src/config/runtime/profiles/gemma4-31b-f16-activations-probe.json +1 -5
- package/src/config/runtime/profiles/gemma4-e2b-baseline-probe.json +1 -3
- package/src/config/runtime/profiles/gemma4-e2b-fused-rmsnorm-probe.json +1 -3
- package/src/config/runtime/profiles/gemma4-e2b-fused-rmsnorm-test.json +1 -0
- package/src/config/runtime/profiles/gemma4-e2b-int4ple-baseline-probe.json +1 -5
- package/src/config/runtime/profiles/gemma4-e2b-int4ple-f16a-probe.json +1 -5
- package/src/config/runtime/profiles/gemma4-e2b-int4ple-reference-export.json +1 -3
- package/src/config/runtime/profiles/gemma4-e2b-int4ple-retain-q4k-probe.json +1 -5
- package/src/config/runtime/profiles/gemma4-e2b-low-memory.json +1 -0
- package/src/config/runtime/profiles/gemma4-e2b-prefill-profile.json +1 -3
- package/src/config/runtime/profiles/gemma4-e2b-throughput-probe.json +1 -0
- package/src/config/runtime/profiles/low-memory.json +1 -0
- package/src/config/runtime/profiles/production.json +1 -3
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-attn128-candidate.json +1 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-maxbatch64-candidate.json +1 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-normed-ffn-candidate.json +1 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-overlap-ringtokens1-candidate.json +1 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-parity.json +1 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-sequential-candidate.json +1 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-throughput.json +1 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-prefill-profile.json +1 -3
- package/src/config/runtime/profiles/qwen-3-5-0-8b-throughput-probe.json +1 -0
- package/src/config/runtime/profiles/qwen-3-5-2b-f16-primary-throughput.json +1 -5
- package/src/config/runtime/profiles/qwen-3-5-2b-metal-parity.json +1 -0
- package/src/config/runtime/profiles/qwen-3-5-2b-metal-throughput.json +1 -0
- package/src/config/runtime/profiles/qwen-3-5-9b-metal-correctness.json +1 -0
- package/src/config/runtime/profiles/qwen-3-embedding-0-6b-metal-stability.json +1 -3
- package/src/config/runtime/profiles/qwen-3-embedding-0-6b-metal-throughput.json +1 -0
- package/src/config/runtime/profiles/qwen-3-reranker-0-6b-metal-stability.json +1 -0
- package/src/config/runtime/profiles/qwen-3-reranker-0-6b-metal-throughput.json +1 -0
- package/src/config/runtime/profiles/qwen3-6-27b-f16-activations-probe.json +1 -5
- package/src/config/runtime/profiles/rerank-stability.json +1 -3
- package/src/config/runtime/profiles/rerank-throughput.json +1 -3
- package/src/config/runtime/profiles/simulation.json +1 -3
- package/src/config/runtime/profiles/throughput.json +1 -3
- package/src/config/runtime/profiles/trace-layers.json +1 -3
- package/src/config/runtime/profiles/translategemma-1b-nativekd2-q4k-rdna3-throughput-probe.json +1 -0
- package/src/config/runtime/profiles/turboquant-contiguous-prod.json +1 -0
- package/src/config/runtime/profiles/turboquant-contiguous.json +1 -0
- package/src/config/runtime/profiles/turboquant.json +1 -0
- package/src/config/runtime/profiles/vector-stability.json +1 -3
- package/src/config/runtime/profiles/vector-throughput.json +1 -0
- package/src/config/runtime/profiles/verbose-trace.json +1 -3
- package/src/config/runtime/tiers/gemma4-128gb.json +1 -0
- package/src/config/runtime/tiers/gemma4-16gb.json +1 -0
- package/src/config/runtime/tiers/gemma4-24gb.json +1 -0
- package/src/config/runtime/tiers/gemma4-32gb.json +1 -0
- package/src/config/schema/benchmark.schema.d.ts +7 -0
- package/src/config/schema/benchmark.schema.js +5 -0
- package/src/config/schema/index.d.ts +8 -16
- package/src/config/schema/index.js +5 -8
- package/src/config/schema/kernel-thresholds.schema.d.ts +0 -12
- package/src/config/schema/kernel-thresholds.schema.js +0 -12
- package/src/config/schema/kernel-warmup.schema.d.ts +1 -3
- package/src/config/schema/kernel-warmup.schema.js +0 -1
- package/src/config/schema/runtime-profile.schema.d.ts +24 -0
- package/src/config/schema/runtime-profile.schema.js +92 -0
- package/src/config/schema/shared-runtime.schema.d.ts +1 -4
- package/src/config/schema/shared-runtime.schema.js +0 -2
- package/src/config/schema/tooling.schema.d.ts +8 -5
- package/src/config/schema/tooling.schema.js +3 -2
- package/src/debug/tensor.d.ts +2 -0
- package/src/debug/tensor.js +21 -2
- package/src/gpu/command-recorder.d.ts +5 -0
- package/src/gpu/command-recorder.js +17 -0
- package/src/gpu/kernel-runtime.js +2 -10
- package/src/gpu/kernels/index.d.ts +0 -2
- package/src/gpu/kernels/index.js +0 -2
- package/src/gpu/kernels/kernel-prewarm.d.ts +3 -0
- package/src/gpu/kernels/kernel-prewarm.js +60 -0
- package/src/gpu/kernels/utils.d.ts +2 -4
- package/src/gpu/kernels/utils.js +2 -4
- package/src/index.d.ts +13 -1
- package/src/index.js +7 -1
- package/src/inference/browser-harness-model-helpers.js +4 -0
- package/src/inference/browser-harness-runtime-helpers.js +25 -10
- package/src/inference/browser-harness-text-helpers.js +4 -1
- package/src/inference/browser-harness.d.ts +4 -0
- package/src/inference/browser-harness.js +91 -4
- package/src/inference/observation-context.d.ts +22 -0
- package/src/inference/observation-context.js +41 -0
- package/src/inference/pipelines/text/attention/attn-config.js +1 -1
- package/src/inference/pipelines/text/attention/executor-immediate.js +42 -0
- package/src/inference/pipelines/text/attention/executor-recorded.js +29 -0
- package/src/inference/pipelines/text/attention/index.js +7 -0
- package/src/inference/pipelines/text/attention/interpreter.js +1334 -0
- package/src/inference/pipelines/text/attention/plan.d.ts +36 -0
- package/src/inference/pipelines/text/attention/plan.js +328 -0
- package/src/inference/pipelines/text/attention/receipt.js +60 -0
- package/src/inference/pipelines/text/attention/record.js +8 -1278
- package/src/inference/pipelines/text/attention/run.js +8 -1386
- package/src/inference/pipelines/text/generator-decode-policy.js +4 -1
- package/src/inference/pipelines/text/generator-helpers.js +2 -0
- package/src/inference/pipelines/text/generator-steps.js +32 -15
- package/src/inference/pipelines/text/generator.js +10 -5
- package/src/inference/pipelines/text/init.d.ts +3 -0
- package/src/inference/pipelines/text/layer-plan-gpu.js +2 -0
- package/src/inference/pipelines/text/layer.js +2 -0
- package/src/inference/pipelines/text/model-load.d.ts +0 -1
- package/src/inference/pipelines/text/model-load.js +2 -11
- package/src/inference/pipelines/text/probes.js +10 -0
- package/src/inference/pipelines/text/resolved-runtime-session.d.ts +56 -0
- package/src/inference/pipelines/text/resolved-runtime-session.js +175 -0
- package/src/inference/pipelines/text/state.d.ts +6 -0
- package/src/inference/pipelines/text/state.js +3 -0
- package/src/inference/pipelines/text/types.d.ts +6 -0
- package/src/inference/pipelines/text.js +16 -1
- package/src/inference/refactor-receipt.d.ts +49 -0
- package/src/inference/refactor-receipt.js +188 -0
- package/src/inference/resource-scope.d.ts +33 -0
- package/src/inference/resource-scope.js +168 -0
- package/src/inference/test-harness.d.ts +3 -0
- package/src/inference/test-harness.js +1 -0
- package/src/storage/model-cache.d.ts +53 -0
- package/src/storage/model-cache.js +550 -0
- package/src/tooling/boundary-evidence.d.ts +51 -0
- package/src/tooling/boundary-evidence.js +408 -0
- package/src/tooling/browser-command-runner.js +4 -4
- package/src/tooling/command-context.d.ts +25 -0
- package/src/tooling/command-context.js +53 -0
- package/src/tooling/command-runner-shared.js +22 -5
- package/src/tooling/demo-receipts.d.ts +56 -0
- package/src/tooling/demo-receipts.js +113 -0
- package/src/tooling/diagnose-runner.js +2 -2
- package/src/tooling/execution-cost-ledger.d.ts +21 -0
- package/src/tooling/execution-cost-ledger.js +199 -0
- package/src/tooling/node-command-runner.js +2 -2
- package/src/tooling/opfs-cache.d.ts +11 -49
- package/src/tooling/opfs-cache.js +4 -510
- package/src/tooling/precision-replay-math.d.ts +42 -0
- package/src/tooling/precision-replay-math.js +174 -0
- package/src/tooling/program-bundle-parity.d.ts +3 -0
- package/src/tooling/program-bundle-parity.js +8 -0
- package/src/tooling/program-bundle.d.ts +1 -52
- package/src/tooling/program-bundle.js +0 -139
- package/src/tooling/registered-variant-calibration.d.ts +24 -0
- package/src/tooling/registered-variant-calibration.js +286 -0
- package/src/tooling/runtime-input-composition.js +105 -53
- package/src/tooling/runtime-intent.js +41 -0
- package/src/tooling/runtime-optimization-index.d.ts +6 -0
- package/src/tooling/runtime-optimization-index.js +72 -0
- package/src/tooling/runtime-optimization.d.ts +47 -6
- package/src/tooling/runtime-optimization.js +559 -40
- package/src/tooling/source-intake.d.ts +49 -0
- package/src/tooling/source-intake.js +414 -0
- package/src/tooling-exports/evidence.d.ts +34 -0
- package/src/tooling-exports/evidence.js +34 -0
- package/src/tooling-exports/runtime.d.ts +7 -0
- package/src/tooling-exports/runtime.js +7 -0
- package/src/tooling-exports.d.ts +11 -0
- package/src/tooling-exports.js +6 -3
- package/src/tooling-exports.shared.d.ts +32 -0
- package/src/tooling-exports.shared.js +31 -0
- package/src/version.js +1 -1
- package/src/config/schema/tuner.schema.d.ts +0 -39
- package/src/config/schema/tuner.schema.js +0 -13
- package/src/gpu/kernel-tuner/benchmarks.js +0 -902
- package/src/gpu/kernel-tuner/cache.js +0 -145
- package/src/gpu/kernel-tuner/index.js +0 -38
- package/src/gpu/kernel-tuner/tuner.js +0 -247
- package/src/gpu/kernel-tuner.js +0 -12
- package/src/gpu/kernels/kernel-tuning.d.ts +0 -42
- package/src/gpu/kernels/kernel-tuning.js +0 -158
- package/src/gpu/profiler.js +0 -382
- package/src/inference/pipelines/text/attention/output-gate-fusion.js +0 -37
|
@@ -0,0 +1,36 @@
|
|
|
1
|
+
export const ATTENTION_PLAN_SCHEMA: 'doppler.attention-plan/v1';
|
|
2
|
+
|
|
3
|
+
export interface SemanticAttentionPlan {
|
|
4
|
+
readonly schema: typeof ATTENTION_PLAN_SCHEMA;
|
|
5
|
+
readonly id: string;
|
|
6
|
+
readonly phase: 'prefill' | 'decode';
|
|
7
|
+
readonly geometry: Readonly<Record<string, number>>;
|
|
8
|
+
readonly dtypes: Readonly<Record<string, string>>;
|
|
9
|
+
readonly transitions: Readonly<Record<string, string | null>>;
|
|
10
|
+
readonly fusion: Readonly<Record<string, boolean>>;
|
|
11
|
+
readonly kv: Readonly<Record<string, unknown>>;
|
|
12
|
+
readonly attention: Readonly<Record<string, unknown>>;
|
|
13
|
+
readonly outputGate: Readonly<Record<string, unknown>>;
|
|
14
|
+
readonly outputProjection: Readonly<Record<string, unknown>>;
|
|
15
|
+
readonly observation: Readonly<Record<string, unknown>>;
|
|
16
|
+
readonly lifetimes: Readonly<Record<string, boolean>>;
|
|
17
|
+
readonly stages: readonly string[];
|
|
18
|
+
}
|
|
19
|
+
|
|
20
|
+
export function resolveAttentionPlan(input: Record<string, unknown>): SemanticAttentionPlan;
|
|
21
|
+
export function resolveAttentionPlanForDispatch(options: Record<string, unknown>): SemanticAttentionPlan;
|
|
22
|
+
export function kernelPathSupportsOutputGateFusion(kernelPath: Record<string, unknown> | null): boolean;
|
|
23
|
+
export function bindAttentionPlan<T extends Record<string, unknown>>(
|
|
24
|
+
plan: SemanticAttentionPlan,
|
|
25
|
+
resources: T
|
|
26
|
+
): Readonly<{ plan: SemanticAttentionPlan; resources: T }>;
|
|
27
|
+
export function executeBoundAttentionPlan<T>(
|
|
28
|
+
boundPlan: Readonly<{ plan: SemanticAttentionPlan; resources: Record<string, unknown> }>,
|
|
29
|
+
executor: {
|
|
30
|
+
executeStage(stage: string, boundPlan: unknown, previous: unknown): Promise<unknown>;
|
|
31
|
+
}
|
|
32
|
+
): Promise<T>;
|
|
33
|
+
export function createAttentionExecutor(
|
|
34
|
+
mode: 'immediate' | 'recorded',
|
|
35
|
+
stageRunners?: Record<string, (boundPlan: unknown, previous: unknown) => unknown>
|
|
36
|
+
): Readonly<Record<string, unknown>>;
|
|
@@ -0,0 +1,328 @@
|
|
|
1
|
+
import { selectRuleValue } from '../../../../rules/rule-registry.js';
|
|
2
|
+
import { hashRefactorReceiptValue } from '../../../refactor-receipt.js';
|
|
3
|
+
|
|
4
|
+
export const ATTENTION_PLAN_SCHEMA = 'doppler.attention-plan/v1';
|
|
5
|
+
|
|
6
|
+
function deepFreeze(value) {
|
|
7
|
+
if (!value || typeof value !== 'object' || Object.isFrozen(value)) {
|
|
8
|
+
return value;
|
|
9
|
+
}
|
|
10
|
+
Object.freeze(value);
|
|
11
|
+
for (const entry of Object.values(value)) {
|
|
12
|
+
deepFreeze(entry);
|
|
13
|
+
}
|
|
14
|
+
return value;
|
|
15
|
+
}
|
|
16
|
+
|
|
17
|
+
function requirePositiveInteger(value, label) {
|
|
18
|
+
if (!Number.isInteger(value) || value < 1) {
|
|
19
|
+
throw new Error(`[AttentionPlan] ${label} must be a positive integer.`);
|
|
20
|
+
}
|
|
21
|
+
return value;
|
|
22
|
+
}
|
|
23
|
+
|
|
24
|
+
function requireDtype(value, label) {
|
|
25
|
+
if (value !== 'f16' && value !== 'f32') {
|
|
26
|
+
throw new Error(`[AttentionPlan] ${label} must be "f16" or "f32".`);
|
|
27
|
+
}
|
|
28
|
+
return value;
|
|
29
|
+
}
|
|
30
|
+
|
|
31
|
+
function resolveAttentionImplementation(input) {
|
|
32
|
+
return selectRuleValue('inference', 'attention', 'attentionKernelVariant', {
|
|
33
|
+
kvLayout: input.kv.layout,
|
|
34
|
+
numTokens: input.geometry.numTokens,
|
|
35
|
+
coldQuantMode: input.kv.coldQuantMode,
|
|
36
|
+
});
|
|
37
|
+
}
|
|
38
|
+
|
|
39
|
+
function resolveOutputResidualMode(input) {
|
|
40
|
+
const hasResidual = input.outputProjection.hasResidual === true;
|
|
41
|
+
const residualMatches = input.outputProjection.residualDtype === input.dtypes.outputProjectionInput;
|
|
42
|
+
const standardFused = selectRuleValue('inference', 'attention', 'useFusedOProjResidual', {
|
|
43
|
+
allowFusedResidual: input.geometry.numTokens === 1,
|
|
44
|
+
hasResidual,
|
|
45
|
+
residualMatches,
|
|
46
|
+
attnIsF32: input.dtypes.outputProjectionInput === 'f32',
|
|
47
|
+
attnIsF16: input.dtypes.outputProjectionInput === 'f16',
|
|
48
|
+
hasLoRA: input.outputProjection.hasLoRA === true,
|
|
49
|
+
oProjIsF16: input.outputProjection.weightDtype === 'f16',
|
|
50
|
+
});
|
|
51
|
+
if (standardFused) {
|
|
52
|
+
return 'matmul-residual';
|
|
53
|
+
}
|
|
54
|
+
const wideTileFused = hasResidual
|
|
55
|
+
&& residualMatches
|
|
56
|
+
&& input.outputProjection.weightDtype === 'q4k'
|
|
57
|
+
&& input.dtypes.outputProjectionInput === 'f32'
|
|
58
|
+
&& input.dtypes.output === 'f32'
|
|
59
|
+
&& input.outputProjection.hasLoRA !== true
|
|
60
|
+
&& input.capabilities.hasF16 === true
|
|
61
|
+
&& input.capabilities.hasSubgroups === true
|
|
62
|
+
&& input.session.useWideTileResidualFusion === true
|
|
63
|
+
&& input.session.retainQ4KMaterialization === true
|
|
64
|
+
&& (
|
|
65
|
+
input.phase === 'decode'
|
|
66
|
+
? input.session.useWideTileQ4KDecode === true
|
|
67
|
+
: input.session.useWideTileQ4KPrefill === true
|
|
68
|
+
);
|
|
69
|
+
return wideTileFused ? 'wide-tile-epilogue' : 'separate';
|
|
70
|
+
}
|
|
71
|
+
|
|
72
|
+
function resolveOutputGateMode(input) {
|
|
73
|
+
if (input.outputGate.enabled !== true) {
|
|
74
|
+
return 'none';
|
|
75
|
+
}
|
|
76
|
+
const kernelPathSupportsFusion = input.outputGate.kernelPathSupportsFusion === true;
|
|
77
|
+
const fused = input.phase === 'decode'
|
|
78
|
+
&& input.geometry.numTokens === 1
|
|
79
|
+
&& input.geometry.headDim === 256
|
|
80
|
+
&& input.dtypes.cachedK === 'f16'
|
|
81
|
+
&& input.dtypes.cachedV === 'f16'
|
|
82
|
+
&& input.outputGate.tensorDtype === 'f32'
|
|
83
|
+
&& input.outputGate.tensorElementCount >= (
|
|
84
|
+
input.geometry.numTokens * input.geometry.numHeads * input.geometry.headDim
|
|
85
|
+
)
|
|
86
|
+
&& input.session.attentionDecodeOnline?.useOutputGateFusion === true
|
|
87
|
+
&& kernelPathSupportsFusion
|
|
88
|
+
&& input.diffusionGemmaDecoder !== true;
|
|
89
|
+
return fused ? 'attention-epilogue' : 'sigmoid-post-attention';
|
|
90
|
+
}
|
|
91
|
+
|
|
92
|
+
export function kernelPathSupportsOutputGateFusion(kernelPath) {
|
|
93
|
+
const steps = kernelPath?.decode?.steps;
|
|
94
|
+
if (!Array.isArray(steps)) {
|
|
95
|
+
return false;
|
|
96
|
+
}
|
|
97
|
+
return steps.some((step) => (
|
|
98
|
+
step?.op === 'attention'
|
|
99
|
+
&& step?.kernel === 'attention_decode_online_head256_f16kv_output_gate.wgsl'
|
|
100
|
+
&& (step?.entry === undefined || step.entry === 'main')
|
|
101
|
+
));
|
|
102
|
+
}
|
|
103
|
+
|
|
104
|
+
export function resolveAttentionPlan(input = {}) {
|
|
105
|
+
if (input.phase !== 'prefill' && input.phase !== 'decode') {
|
|
106
|
+
throw new Error('[AttentionPlan] phase must be "prefill" or "decode".');
|
|
107
|
+
}
|
|
108
|
+
const geometry = {
|
|
109
|
+
layerIdx: Number.isInteger(input.geometry?.layerIdx) ? input.geometry.layerIdx : 0,
|
|
110
|
+
numTokens: requirePositiveInteger(input.geometry?.numTokens, 'geometry.numTokens'),
|
|
111
|
+
numHeads: requirePositiveInteger(input.geometry?.numHeads, 'geometry.numHeads'),
|
|
112
|
+
numKVHeads: requirePositiveInteger(input.geometry?.numKVHeads, 'geometry.numKVHeads'),
|
|
113
|
+
headDim: requirePositiveInteger(input.geometry?.headDim, 'geometry.headDim'),
|
|
114
|
+
hiddenSize: requirePositiveInteger(input.geometry?.hiddenSize, 'geometry.hiddenSize'),
|
|
115
|
+
currentSeqLen: Number.isInteger(input.geometry?.currentSeqLen) && input.geometry.currentSeqLen >= 0
|
|
116
|
+
? input.geometry.currentSeqLen
|
|
117
|
+
: 0,
|
|
118
|
+
};
|
|
119
|
+
const normalized = {
|
|
120
|
+
phase: input.phase,
|
|
121
|
+
geometry,
|
|
122
|
+
dtypes: {
|
|
123
|
+
input: requireDtype(input.dtypes?.input, 'dtypes.input'),
|
|
124
|
+
activation: requireDtype(input.dtypes?.activation, 'dtypes.activation'),
|
|
125
|
+
projection: requireDtype(input.dtypes?.projection, 'dtypes.projection'),
|
|
126
|
+
kv: requireDtype(input.dtypes?.kv, 'dtypes.kv'),
|
|
127
|
+
cachedK: requireDtype(input.dtypes?.cachedK, 'dtypes.cachedK'),
|
|
128
|
+
cachedV: requireDtype(input.dtypes?.cachedV, 'dtypes.cachedV'),
|
|
129
|
+
outputProjectionInput: requireDtype(
|
|
130
|
+
input.dtypes?.outputProjectionInput,
|
|
131
|
+
'dtypes.outputProjectionInput'
|
|
132
|
+
),
|
|
133
|
+
output: requireDtype(input.dtypes?.output, 'dtypes.output'),
|
|
134
|
+
},
|
|
135
|
+
kv: {
|
|
136
|
+
layout: String(input.kv?.layout ?? 'contiguous'),
|
|
137
|
+
coldQuantMode: String(input.kv?.coldQuantMode ?? 'none'),
|
|
138
|
+
length: requirePositiveInteger(input.kv?.length, 'kv.length'),
|
|
139
|
+
pageSize: Number.isInteger(input.kv?.pageSize) ? input.kv.pageSize : 0,
|
|
140
|
+
},
|
|
141
|
+
session: input.session ?? {},
|
|
142
|
+
capabilities: {
|
|
143
|
+
hasF16: input.capabilities?.hasF16 === true,
|
|
144
|
+
hasSubgroups: input.capabilities?.hasSubgroups === true,
|
|
145
|
+
},
|
|
146
|
+
fusion: {
|
|
147
|
+
inputNormProjection: input.fusion?.inputNormProjection === true,
|
|
148
|
+
qkvProjection: input.fusion?.qkvProjection === true,
|
|
149
|
+
qkNorm: input.fusion?.qkNorm === true,
|
|
150
|
+
qkNormRoPE: input.fusion?.qkNormRoPE === true,
|
|
151
|
+
},
|
|
152
|
+
outputGate: {
|
|
153
|
+
enabled: input.outputGate?.enabled === true,
|
|
154
|
+
tensorDtype: input.outputGate?.tensorDtype ?? null,
|
|
155
|
+
tensorElementCount: Number.isInteger(input.outputGate?.tensorElementCount)
|
|
156
|
+
? input.outputGate.tensorElementCount
|
|
157
|
+
: 0,
|
|
158
|
+
kernelPathSupportsFusion: input.outputGate?.kernelPathSupportsFusion === true,
|
|
159
|
+
},
|
|
160
|
+
outputProjection: {
|
|
161
|
+
weightDtype: input.outputProjection?.weightDtype ?? null,
|
|
162
|
+
hasResidual: input.outputProjection?.hasResidual === true,
|
|
163
|
+
residualDtype: input.outputProjection?.residualDtype ?? null,
|
|
164
|
+
hasLoRA: input.outputProjection?.hasLoRA === true,
|
|
165
|
+
},
|
|
166
|
+
diagnosticsEligible: input.diagnosticsEligible === true,
|
|
167
|
+
sharedKV: {
|
|
168
|
+
reuses: input.sharedKV?.reuses === true,
|
|
169
|
+
stores: input.sharedKV?.stores === true,
|
|
170
|
+
},
|
|
171
|
+
diffusionGemmaDecoder: input.diffusionGemmaDecoder === true,
|
|
172
|
+
};
|
|
173
|
+
const implementation = resolveAttentionImplementation(normalized);
|
|
174
|
+
const planCore = {
|
|
175
|
+
phase: normalized.phase,
|
|
176
|
+
geometry: normalized.geometry,
|
|
177
|
+
dtypes: normalized.dtypes,
|
|
178
|
+
transitions: {
|
|
179
|
+
inputCast: normalized.dtypes.input === normalized.dtypes.activation
|
|
180
|
+
? null
|
|
181
|
+
: `${normalized.dtypes.input}->${normalized.dtypes.activation}`,
|
|
182
|
+
projectionCast: normalized.dtypes.projection === normalized.dtypes.outputProjectionInput
|
|
183
|
+
? null
|
|
184
|
+
: `${normalized.dtypes.projection}->${normalized.dtypes.outputProjectionInput}`,
|
|
185
|
+
},
|
|
186
|
+
fusion: normalized.fusion,
|
|
187
|
+
kv: {
|
|
188
|
+
...normalized.kv,
|
|
189
|
+
implementation,
|
|
190
|
+
},
|
|
191
|
+
attention: {
|
|
192
|
+
implementation,
|
|
193
|
+
flashPrefill: normalized.phase === 'prefill'
|
|
194
|
+
&& normalized.diffusionGemmaDecoder !== true
|
|
195
|
+
&& normalized.session.useFlashPrefillAttention === true,
|
|
196
|
+
ortFlashPrefill: normalized.phase === 'prefill'
|
|
197
|
+
&& normalized.diffusionGemmaDecoder !== true
|
|
198
|
+
&& normalized.session.useOrtFlashPrefillAttention === true,
|
|
199
|
+
},
|
|
200
|
+
outputGate: {
|
|
201
|
+
enabled: normalized.outputGate.enabled,
|
|
202
|
+
semantics: normalized.outputGate.enabled ? 'sigmoid' : 'none',
|
|
203
|
+
mode: resolveOutputGateMode(normalized),
|
|
204
|
+
},
|
|
205
|
+
outputProjection: {
|
|
206
|
+
inputDtype: normalized.dtypes.outputProjectionInput,
|
|
207
|
+
outputDtype: normalized.dtypes.output,
|
|
208
|
+
residualMode: resolveOutputResidualMode(normalized),
|
|
209
|
+
hasLoRA: normalized.outputProjection.hasLoRA,
|
|
210
|
+
},
|
|
211
|
+
observation: {
|
|
212
|
+
diagnosticsEligible: normalized.diagnosticsEligible,
|
|
213
|
+
},
|
|
214
|
+
lifetimes: {
|
|
215
|
+
retainSharedKV: normalized.sharedKV.reuses || normalized.sharedKV.stores,
|
|
216
|
+
valueMayAliasKey: true,
|
|
217
|
+
submitRetentionRequired: true,
|
|
218
|
+
},
|
|
219
|
+
stages: [
|
|
220
|
+
'input-normalization',
|
|
221
|
+
'qkv-projection',
|
|
222
|
+
'kv-update',
|
|
223
|
+
'attention',
|
|
224
|
+
'output-gate',
|
|
225
|
+
'output-projection',
|
|
226
|
+
],
|
|
227
|
+
};
|
|
228
|
+
return deepFreeze({
|
|
229
|
+
schema: ATTENTION_PLAN_SCHEMA,
|
|
230
|
+
id: hashRefactorReceiptValue(planCore),
|
|
231
|
+
...planCore,
|
|
232
|
+
});
|
|
233
|
+
}
|
|
234
|
+
|
|
235
|
+
function tensorElementCount(tensor) {
|
|
236
|
+
if (!Array.isArray(tensor?.shape)) {
|
|
237
|
+
return 0;
|
|
238
|
+
}
|
|
239
|
+
return tensor.shape.reduce((total, value) => total * value, 1);
|
|
240
|
+
}
|
|
241
|
+
|
|
242
|
+
export function resolveAttentionPlanForDispatch(options = {}) {
|
|
243
|
+
const config = options.config ?? {};
|
|
244
|
+
const dispatchParams = options.dispatchParams ?? {};
|
|
245
|
+
const kvState = options.kvState ?? {};
|
|
246
|
+
return resolveAttentionPlan({
|
|
247
|
+
phase: config.isPrefill ? 'prefill' : 'decode',
|
|
248
|
+
geometry: {
|
|
249
|
+
layerIdx: config.layerIdx,
|
|
250
|
+
numTokens: config.numTokens,
|
|
251
|
+
numHeads: config.numHeads,
|
|
252
|
+
numKVHeads: config.numKVHeads,
|
|
253
|
+
headDim: config.headDim,
|
|
254
|
+
hiddenSize: config.hiddenSize,
|
|
255
|
+
currentSeqLen: config.currentSeqLen,
|
|
256
|
+
},
|
|
257
|
+
dtypes: {
|
|
258
|
+
input: options.inputDtype,
|
|
259
|
+
activation: options.activationDtype,
|
|
260
|
+
projection: options.projectionDtype,
|
|
261
|
+
kv: options.kvDtype,
|
|
262
|
+
cachedK: dispatchParams.cachedKDtype,
|
|
263
|
+
cachedV: dispatchParams.cachedVDtype,
|
|
264
|
+
outputProjectionInput: options.outputProjectionInputDtype,
|
|
265
|
+
output: options.outputDtype,
|
|
266
|
+
},
|
|
267
|
+
kv: {
|
|
268
|
+
layout: kvState.kvLayout,
|
|
269
|
+
coldQuantMode: kvState.coldQuantMode,
|
|
270
|
+
length: kvState.kvLenForAttention,
|
|
271
|
+
pageSize: kvState.kvPageSize ?? kvState.coldPageSize ?? 0,
|
|
272
|
+
},
|
|
273
|
+
session: options.session,
|
|
274
|
+
capabilities: options.capabilities,
|
|
275
|
+
fusion: options.fusion,
|
|
276
|
+
outputGate: {
|
|
277
|
+
enabled: config.attentionOutputGate === true,
|
|
278
|
+
tensorDtype: options.qGateTensor?.dtype ?? null,
|
|
279
|
+
tensorElementCount: tensorElementCount(options.qGateTensor),
|
|
280
|
+
kernelPathSupportsFusion: kernelPathSupportsOutputGateFusion(config.kernelPath),
|
|
281
|
+
},
|
|
282
|
+
outputProjection: {
|
|
283
|
+
weightDtype: options.outputProjectionWeightDtype ?? null,
|
|
284
|
+
hasResidual: config.residualTensor != null,
|
|
285
|
+
residualDtype: config.residualTensor?.dtype ?? null,
|
|
286
|
+
hasLoRA: options.outputProjectionHasLoRA === true,
|
|
287
|
+
},
|
|
288
|
+
diagnosticsEligible: options.diagnosticsEligible === true,
|
|
289
|
+
sharedKV: options.sharedKV,
|
|
290
|
+
diffusionGemmaDecoder: config.diffusionGemmaDecoder === true,
|
|
291
|
+
});
|
|
292
|
+
}
|
|
293
|
+
|
|
294
|
+
export function bindAttentionPlan(plan, resources = {}) {
|
|
295
|
+
if (!plan || plan.schema !== ATTENTION_PLAN_SCHEMA) {
|
|
296
|
+
throw new Error('[AttentionPlan] bindAttentionPlan requires a resolved semantic plan.');
|
|
297
|
+
}
|
|
298
|
+
return Object.freeze({
|
|
299
|
+
plan,
|
|
300
|
+
resources,
|
|
301
|
+
});
|
|
302
|
+
}
|
|
303
|
+
|
|
304
|
+
export async function executeBoundAttentionPlan(boundPlan, executor) {
|
|
305
|
+
if (!boundPlan?.plan || !executor || typeof executor.executeStage !== 'function') {
|
|
306
|
+
throw new Error('[AttentionPlan] executeBoundAttentionPlan requires a bound plan and executor.');
|
|
307
|
+
}
|
|
308
|
+
let result;
|
|
309
|
+
for (const stage of boundPlan.plan.stages) {
|
|
310
|
+
result = await executor.executeStage(stage, boundPlan, result);
|
|
311
|
+
}
|
|
312
|
+
return result;
|
|
313
|
+
}
|
|
314
|
+
|
|
315
|
+
export function createAttentionExecutor(mode, stageRunners = {}) {
|
|
316
|
+
if (mode !== 'immediate' && mode !== 'recorded') {
|
|
317
|
+
throw new Error('[AttentionPlan] executor mode must be "immediate" or "recorded".');
|
|
318
|
+
}
|
|
319
|
+
return Object.freeze({
|
|
320
|
+
mode,
|
|
321
|
+
async executeStage(stage, boundPlan, previous) {
|
|
322
|
+
const runner = stageRunners[stage];
|
|
323
|
+
return typeof runner === 'function'
|
|
324
|
+
? runner(boundPlan, previous)
|
|
325
|
+
: previous;
|
|
326
|
+
},
|
|
327
|
+
});
|
|
328
|
+
}
|
|
@@ -0,0 +1,60 @@
|
|
|
1
|
+
import { createRefactorReceipt } from '../../../refactor-receipt.js';
|
|
2
|
+
|
|
3
|
+
function normalizeError(error, boundary) {
|
|
4
|
+
if (!error) {
|
|
5
|
+
return null;
|
|
6
|
+
}
|
|
7
|
+
return {
|
|
8
|
+
boundary,
|
|
9
|
+
name: typeof error.name === 'string' ? error.name : 'Error',
|
|
10
|
+
message: typeof error.message === 'string' ? error.message : String(error),
|
|
11
|
+
cleanup: 'completed',
|
|
12
|
+
};
|
|
13
|
+
}
|
|
14
|
+
|
|
15
|
+
export function captureAttentionRefactorReceipt(options = {}) {
|
|
16
|
+
const state = options.state ?? {};
|
|
17
|
+
const policy = state.observationContext?.receiptPolicy ?? 'off';
|
|
18
|
+
const failure = options.error
|
|
19
|
+
? normalizeError(options.error, options.failureBoundary ?? 'attention')
|
|
20
|
+
: null;
|
|
21
|
+
if (policy === 'off' || (policy === 'on_failure' && !failure)) {
|
|
22
|
+
return null;
|
|
23
|
+
}
|
|
24
|
+
const plan = options.plan ?? null;
|
|
25
|
+
const receipt = createRefactorReceipt({
|
|
26
|
+
commandContext: state.observationContext?.commandContext ?? null,
|
|
27
|
+
resolvedSession: state.resolvedRuntimeSession ?? null,
|
|
28
|
+
observationContext: state.observationContext ?? null,
|
|
29
|
+
operationPlan: plan,
|
|
30
|
+
operations: (plan?.stages ?? []).map((stage, sequence) => ({
|
|
31
|
+
sequence,
|
|
32
|
+
stage,
|
|
33
|
+
})),
|
|
34
|
+
dtypeTransitions: Object.entries(plan?.transitions ?? {})
|
|
35
|
+
.filter(([, transition]) => transition != null)
|
|
36
|
+
.map(([boundary, transition]) => ({
|
|
37
|
+
boundary,
|
|
38
|
+
transition,
|
|
39
|
+
})),
|
|
40
|
+
resourceEvents: options.resourceEvents ?? [],
|
|
41
|
+
failure,
|
|
42
|
+
});
|
|
43
|
+
state.stats ??= {};
|
|
44
|
+
state.stats.refactorReceipts ??= [];
|
|
45
|
+
state.stats.refactorReceipts.push(receipt);
|
|
46
|
+
if (
|
|
47
|
+
failure
|
|
48
|
+
&& options.error
|
|
49
|
+
&& typeof options.error === 'object'
|
|
50
|
+
&& Object.isExtensible(options.error)
|
|
51
|
+
) {
|
|
52
|
+
options.error.details = {
|
|
53
|
+
...(options.error.details && typeof options.error.details === 'object'
|
|
54
|
+
? options.error.details
|
|
55
|
+
: {}),
|
|
56
|
+
refactorReceipt: receipt,
|
|
57
|
+
};
|
|
58
|
+
}
|
|
59
|
+
return receipt;
|
|
60
|
+
}
|