doppler-gpu 0.4.14 → 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/CHANGELOG.md +9 -0
- package/README.md +83 -52
- 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 +116 -1
- package/src/client/runtime/model-session.js +351 -1
- 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 +1,257 @@
|
|
|
1
|
+
const SESSION_SCHEMA = 'doppler.scoped-session/v1';
|
|
2
|
+
const RESULT_SCHEMA = 'doppler.generation-result/v1';
|
|
3
|
+
const EVENT_SCHEMA = 'doppler.generation-event/v1';
|
|
4
|
+
|
|
5
|
+
const OBSERVATION_POLICIES = Object.freeze({
|
|
6
|
+
always: Object.freeze({
|
|
7
|
+
id: 'demo/always-on',
|
|
8
|
+
tier: 'always',
|
|
9
|
+
executionClassification: 'representative',
|
|
10
|
+
}),
|
|
11
|
+
'guided-quality': Object.freeze({
|
|
12
|
+
id: 'demo/guided-quality',
|
|
13
|
+
tier: 'guided-quality',
|
|
14
|
+
executionClassification: 'observed',
|
|
15
|
+
}),
|
|
16
|
+
'deep-xray': Object.freeze({
|
|
17
|
+
id: 'demo/deep-xray',
|
|
18
|
+
tier: 'deep-xray',
|
|
19
|
+
executionClassification: 'deep-diagnostic',
|
|
20
|
+
}),
|
|
21
|
+
});
|
|
22
|
+
|
|
23
|
+
function resolveObservationPolicy(value = 'always') {
|
|
24
|
+
const key = String(value || 'always').trim();
|
|
25
|
+
const policy = OBSERVATION_POLICIES[key]
|
|
26
|
+
|| Object.values(OBSERVATION_POLICIES).find((entry) => entry.id === key);
|
|
27
|
+
if (!policy) {
|
|
28
|
+
throw new Error(
|
|
29
|
+
`Unsupported Doppler observation policy "${key}". `
|
|
30
|
+
+ `Expected one of: ${Object.keys(OBSERVATION_POLICIES).join(', ')}.`
|
|
31
|
+
);
|
|
32
|
+
}
|
|
33
|
+
return policy;
|
|
34
|
+
}
|
|
35
|
+
|
|
36
|
+
function capabilityMap(handle) {
|
|
37
|
+
const specializedModel = handle?.supportsEmbedding === true || handle?.supportsSequence === true;
|
|
38
|
+
return Object.freeze({
|
|
39
|
+
generate: !specializedModel && typeof handle?.generateWithEvidence === 'function',
|
|
40
|
+
stream: !specializedModel && typeof handle?.generate === 'function',
|
|
41
|
+
embed: handle?.supportsEmbedding === true && typeof handle?.embed === 'function',
|
|
42
|
+
sequence: handle?.supportsSequence === true && typeof handle?.encodeSequence === 'function',
|
|
43
|
+
inspect: typeof handle?.inspect?.generate === 'function',
|
|
44
|
+
lora: typeof handle?.loadLoRA === 'function' && typeof handle?.unloadLoRA === 'function',
|
|
45
|
+
advanced: handle?.advanced && typeof handle.advanced === 'object',
|
|
46
|
+
});
|
|
47
|
+
}
|
|
48
|
+
|
|
49
|
+
function buildObservation(policy, evidence, inspectionReceipt = null) {
|
|
50
|
+
const statsAvailable = Boolean(evidence?.stats && typeof evidence.stats === 'object');
|
|
51
|
+
return {
|
|
52
|
+
policyId: policy.id,
|
|
53
|
+
tier: policy.tier,
|
|
54
|
+
executionClassification: policy.executionClassification,
|
|
55
|
+
executionChanged: inspectionReceipt?.policy?.modifiesExecution === true,
|
|
56
|
+
unavailable: policy.tier === 'guided-quality' && !inspectionReceipt?.quality
|
|
57
|
+
? ['token-surprisal', 'word-surprisal', 'rolling-perplexity']
|
|
58
|
+
: [],
|
|
59
|
+
deepEvidenceAvailable: policy.tier === 'deep-xray' ? statsAvailable : null,
|
|
60
|
+
};
|
|
61
|
+
}
|
|
62
|
+
|
|
63
|
+
function buildGenerationResult(evidence, policy, inspectionReceipt = null) {
|
|
64
|
+
const outputText = String(evidence?.outputText ?? '');
|
|
65
|
+
const tokenIds = Array.from(evidence?.tokenIds || [], Number);
|
|
66
|
+
return {
|
|
67
|
+
schema: RESULT_SCHEMA,
|
|
68
|
+
outputText,
|
|
69
|
+
content: outputText,
|
|
70
|
+
tokenIds,
|
|
71
|
+
usage: {
|
|
72
|
+
promptTokens: null,
|
|
73
|
+
completionTokens: tokenIds.length,
|
|
74
|
+
totalTokens: null,
|
|
75
|
+
},
|
|
76
|
+
observation: buildObservation(policy, evidence, inspectionReceipt),
|
|
77
|
+
fingerprint: inspectionReceipt?.fingerprint || {
|
|
78
|
+
modelId: evidence?.runtimeProfile?.model?.modelId ?? null,
|
|
79
|
+
manifestHash: evidence?.runtimeProfile?.model?.manifestHash ?? null,
|
|
80
|
+
tokenizerHash: evidence?.stats?.tokenizerHash ?? null,
|
|
81
|
+
executionPlanId: evidence?.backendIdentity?.executionPlanId ?? null,
|
|
82
|
+
kernelPathId: evidence?.backendIdentity?.kernelPathId ?? null,
|
|
83
|
+
runtimeProfileHash: evidence?.runtimeProfileHash ?? null,
|
|
84
|
+
backendIdentityHash: evidence?.backendIdentityHash ?? null,
|
|
85
|
+
observationPolicyId: policy.id,
|
|
86
|
+
},
|
|
87
|
+
inspectionReceipt,
|
|
88
|
+
evidence,
|
|
89
|
+
};
|
|
90
|
+
}
|
|
91
|
+
|
|
92
|
+
export function createScopedModelSession(handle) {
|
|
93
|
+
if (!handle || typeof handle !== 'object') {
|
|
94
|
+
throw new Error('createScopedModelSession requires a Doppler model handle.');
|
|
95
|
+
}
|
|
96
|
+
let closed = false;
|
|
97
|
+
const capabilities = capabilityMap(handle);
|
|
98
|
+
|
|
99
|
+
function requireOpen() {
|
|
100
|
+
if (closed) {
|
|
101
|
+
throw new Error('Doppler scoped session is closed.');
|
|
102
|
+
}
|
|
103
|
+
}
|
|
104
|
+
|
|
105
|
+
function supports(capability) {
|
|
106
|
+
return capabilities[String(capability || '')] === true;
|
|
107
|
+
}
|
|
108
|
+
|
|
109
|
+
function requireCapability(capability) {
|
|
110
|
+
requireOpen();
|
|
111
|
+
if (!supports(capability)) {
|
|
112
|
+
throw new Error(
|
|
113
|
+
`Doppler model "${handle.modelId || 'unknown'}" does not support "${capability}".`
|
|
114
|
+
);
|
|
115
|
+
}
|
|
116
|
+
return true;
|
|
117
|
+
}
|
|
118
|
+
|
|
119
|
+
const session = {
|
|
120
|
+
schema: SESSION_SCHEMA,
|
|
121
|
+
capabilities,
|
|
122
|
+
supports,
|
|
123
|
+
require: requireCapability,
|
|
124
|
+
async generate(input, options = {}) {
|
|
125
|
+
requireCapability('generate');
|
|
126
|
+
const policy = resolveObservationPolicy(options.observe);
|
|
127
|
+
const {
|
|
128
|
+
observe: _observe,
|
|
129
|
+
observationPolicy: _observationPolicy,
|
|
130
|
+
...generationOptions
|
|
131
|
+
} = options;
|
|
132
|
+
if (
|
|
133
|
+
typeof input === 'string'
|
|
134
|
+
&& policy.tier === 'always'
|
|
135
|
+
&& typeof handle.inspect?.generate === 'function'
|
|
136
|
+
) {
|
|
137
|
+
const inspectionReceipt = await handle.inspect.generate(input, {
|
|
138
|
+
policyId: policy.id,
|
|
139
|
+
generation: generationOptions,
|
|
140
|
+
});
|
|
141
|
+
return buildGenerationResult(
|
|
142
|
+
inspectionReceipt.generationEvidence,
|
|
143
|
+
policy,
|
|
144
|
+
inspectionReceipt
|
|
145
|
+
);
|
|
146
|
+
}
|
|
147
|
+
const evidence = await handle.generateWithEvidence(input, generationOptions);
|
|
148
|
+
return buildGenerationResult(evidence, policy);
|
|
149
|
+
},
|
|
150
|
+
async *stream(input, options = {}) {
|
|
151
|
+
requireCapability('stream');
|
|
152
|
+
const policy = resolveObservationPolicy(options.observe);
|
|
153
|
+
const {
|
|
154
|
+
observe: _observe,
|
|
155
|
+
observationPolicy: _observationPolicy,
|
|
156
|
+
...generationOptions
|
|
157
|
+
} = options;
|
|
158
|
+
let outputText = '';
|
|
159
|
+
for await (const text of handle.generate(input, generationOptions)) {
|
|
160
|
+
const delta = String(text ?? '');
|
|
161
|
+
outputText += delta;
|
|
162
|
+
yield {
|
|
163
|
+
schema: EVENT_SCHEMA,
|
|
164
|
+
type: 'text-delta',
|
|
165
|
+
text: delta,
|
|
166
|
+
observationPolicyId: policy.id,
|
|
167
|
+
};
|
|
168
|
+
}
|
|
169
|
+
yield {
|
|
170
|
+
schema: EVENT_SCHEMA,
|
|
171
|
+
type: 'complete',
|
|
172
|
+
outputText,
|
|
173
|
+
observationPolicyId: policy.id,
|
|
174
|
+
};
|
|
175
|
+
},
|
|
176
|
+
async inspect(input, options = {}) {
|
|
177
|
+
requireCapability('inspect');
|
|
178
|
+
const observationPolicy = options.observationPolicy ?? options.observe ?? 'guided-quality';
|
|
179
|
+
const policy = resolveObservationPolicy(observationPolicy);
|
|
180
|
+
if (typeof handle.inspect?.generate !== 'function') {
|
|
181
|
+
throw new Error('Loaded Doppler handle does not expose semantic inspection.');
|
|
182
|
+
}
|
|
183
|
+
const {
|
|
184
|
+
observe: _observe,
|
|
185
|
+
observationPolicy: _observationPolicy,
|
|
186
|
+
topKSize,
|
|
187
|
+
onEvent,
|
|
188
|
+
...generation
|
|
189
|
+
} = options;
|
|
190
|
+
return handle.inspect.generate(input, {
|
|
191
|
+
policyId: policy.id,
|
|
192
|
+
generation,
|
|
193
|
+
topKSize,
|
|
194
|
+
onEvent,
|
|
195
|
+
});
|
|
196
|
+
},
|
|
197
|
+
async embed(input, options = {}) {
|
|
198
|
+
requireCapability('embed');
|
|
199
|
+
return handle.embed(input, options);
|
|
200
|
+
},
|
|
201
|
+
async encodeSequence(sequence, options = {}) {
|
|
202
|
+
requireCapability('sequence');
|
|
203
|
+
return handle.encodeSequence(sequence, options);
|
|
204
|
+
},
|
|
205
|
+
async loadLoRA(adapter, options = {}) {
|
|
206
|
+
requireCapability('lora');
|
|
207
|
+
return handle.loadLoRA(adapter, options);
|
|
208
|
+
},
|
|
209
|
+
async unloadLoRA() {
|
|
210
|
+
requireCapability('lora');
|
|
211
|
+
return handle.unloadLoRA();
|
|
212
|
+
},
|
|
213
|
+
resetGenerationState() {
|
|
214
|
+
requireOpen();
|
|
215
|
+
return handle.resetGenerationState();
|
|
216
|
+
},
|
|
217
|
+
async close() {
|
|
218
|
+
if (closed) return;
|
|
219
|
+
closed = true;
|
|
220
|
+
await handle.unload();
|
|
221
|
+
},
|
|
222
|
+
async [Symbol.asyncDispose]() {
|
|
223
|
+
await session.close();
|
|
224
|
+
},
|
|
225
|
+
get closed() {
|
|
226
|
+
return closed;
|
|
227
|
+
},
|
|
228
|
+
get loaded() {
|
|
229
|
+
return !closed && handle.loaded === true;
|
|
230
|
+
},
|
|
231
|
+
get modelId() {
|
|
232
|
+
return handle.modelId;
|
|
233
|
+
},
|
|
234
|
+
get manifestHash() {
|
|
235
|
+
return handle.manifestHash ?? null;
|
|
236
|
+
},
|
|
237
|
+
get persistentCache() {
|
|
238
|
+
return handle.persistentCache ?? null;
|
|
239
|
+
},
|
|
240
|
+
get manifest() {
|
|
241
|
+
return handle.manifest ?? null;
|
|
242
|
+
},
|
|
243
|
+
get activeLoRA() {
|
|
244
|
+
return handle.activeLoRA ?? null;
|
|
245
|
+
},
|
|
246
|
+
get deviceInfo() {
|
|
247
|
+
return handle.deviceInfo ?? null;
|
|
248
|
+
},
|
|
249
|
+
get advanced() {
|
|
250
|
+
requireCapability('advanced');
|
|
251
|
+
return handle.advanced;
|
|
252
|
+
},
|
|
253
|
+
};
|
|
254
|
+
return session;
|
|
255
|
+
}
|
|
256
|
+
|
|
257
|
+
export { OBSERVATION_POLICIES };
|
|
@@ -78,7 +78,6 @@ export interface DopplerCapabilitiesType {
|
|
|
78
78
|
initialized: boolean;
|
|
79
79
|
currentModelId: string | null;
|
|
80
80
|
kernelsWarmed: boolean;
|
|
81
|
-
kernelsTuned: boolean;
|
|
82
81
|
lastModelEstimate: ModelEstimate | null;
|
|
83
82
|
bridgeClient?: ExtensionBridgeClient | null;
|
|
84
83
|
localPath?: string | null;
|
|
@@ -0,0 +1,54 @@
|
|
|
1
|
+
{
|
|
2
|
+
"schema": "doppler.boundary-evidence-policy/v1",
|
|
3
|
+
"boundaries": [
|
|
4
|
+
{
|
|
5
|
+
"idTemplate": "model.embedding.output",
|
|
6
|
+
"stages": ["embed.out"]
|
|
7
|
+
},
|
|
8
|
+
{
|
|
9
|
+
"idTemplate": "layer.{L}.attention.q.pre_rope",
|
|
10
|
+
"stages": ["attn.q_norm", "attn.q_proj"]
|
|
11
|
+
},
|
|
12
|
+
{
|
|
13
|
+
"idTemplate": "layer.{L}.attention.k.pre_rope",
|
|
14
|
+
"stages": ["attn.k_norm", "attn.k_proj"]
|
|
15
|
+
},
|
|
16
|
+
{
|
|
17
|
+
"idTemplate": "layer.{L}.attention.q.post_rope",
|
|
18
|
+
"stages": ["attn.q_rope"]
|
|
19
|
+
},
|
|
20
|
+
{
|
|
21
|
+
"idTemplate": "layer.{L}.attention.k.post_rope",
|
|
22
|
+
"stages": ["attn.k_rope"]
|
|
23
|
+
},
|
|
24
|
+
{
|
|
25
|
+
"idTemplate": "layer.{L}.attention.output",
|
|
26
|
+
"stages": ["attn.out", "attn.o_proj", "attn.core_out"]
|
|
27
|
+
},
|
|
28
|
+
{
|
|
29
|
+
"idTemplate": "layer.{L}.ffn.output",
|
|
30
|
+
"stages": ["ffn.out"]
|
|
31
|
+
},
|
|
32
|
+
{
|
|
33
|
+
"idTemplate": "model.logits",
|
|
34
|
+
"stages": ["logits.final", "logits.out"]
|
|
35
|
+
}
|
|
36
|
+
],
|
|
37
|
+
"tolerancePolicies": {
|
|
38
|
+
"doppler.boundary-tolerance/source-f32-v1": {
|
|
39
|
+
"absolute": 0.00001,
|
|
40
|
+
"relative": 0.00001,
|
|
41
|
+
"requireDtypeMatch": true
|
|
42
|
+
},
|
|
43
|
+
"doppler.boundary-tolerance/source-f16-v1": {
|
|
44
|
+
"absolute": 0.002,
|
|
45
|
+
"relative": 0.002,
|
|
46
|
+
"requireDtypeMatch": false
|
|
47
|
+
},
|
|
48
|
+
"doppler.boundary-tolerance/quantized-v1": {
|
|
49
|
+
"absolute": 0.05,
|
|
50
|
+
"relative": 0.05,
|
|
51
|
+
"requireDtypeMatch": false
|
|
52
|
+
}
|
|
53
|
+
}
|
|
54
|
+
}
|
|
@@ -0,0 +1,122 @@
|
|
|
1
|
+
{
|
|
2
|
+
"schema": "doppler.source-intake-policy/v1",
|
|
3
|
+
"facts": [
|
|
4
|
+
{
|
|
5
|
+
"factId": "architecture.model_type",
|
|
6
|
+
"sourcePointers": ["/model_type", "/text_config/model_type"],
|
|
7
|
+
"owner": "conversion.modelType"
|
|
8
|
+
},
|
|
9
|
+
{
|
|
10
|
+
"factId": "architecture.hidden_size",
|
|
11
|
+
"sourcePointers": ["/hidden_size", "/text_config/hidden_size"],
|
|
12
|
+
"owner": "manifest.architecture.hiddenSize"
|
|
13
|
+
},
|
|
14
|
+
{
|
|
15
|
+
"factId": "architecture.intermediate_size",
|
|
16
|
+
"sourcePointers": ["/intermediate_size", "/text_config/intermediate_size"],
|
|
17
|
+
"owner": "manifest.architecture.intermediateSize"
|
|
18
|
+
},
|
|
19
|
+
{
|
|
20
|
+
"factId": "architecture.layer_count",
|
|
21
|
+
"sourcePointers": ["/num_hidden_layers", "/text_config/num_hidden_layers"],
|
|
22
|
+
"owner": "manifest.architecture.numLayers"
|
|
23
|
+
},
|
|
24
|
+
{
|
|
25
|
+
"factId": "architecture.attention.query_heads",
|
|
26
|
+
"sourcePointers": ["/num_attention_heads", "/text_config/num_attention_heads"],
|
|
27
|
+
"owner": "manifest.architecture.numAttentionHeads"
|
|
28
|
+
},
|
|
29
|
+
{
|
|
30
|
+
"factId": "architecture.attention.kv_heads",
|
|
31
|
+
"sourcePointers": ["/num_key_value_heads", "/text_config/num_key_value_heads"],
|
|
32
|
+
"owner": "manifest.architecture.numKeyValueHeads"
|
|
33
|
+
},
|
|
34
|
+
{
|
|
35
|
+
"factId": "architecture.attention.head_dim",
|
|
36
|
+
"sourcePointers": ["/head_dim", "/text_config/head_dim"],
|
|
37
|
+
"owner": "manifest.architecture.headDim"
|
|
38
|
+
},
|
|
39
|
+
{
|
|
40
|
+
"factId": "architecture.vocab_size",
|
|
41
|
+
"sourcePointers": ["/vocab_size", "/text_config/vocab_size"],
|
|
42
|
+
"owner": "manifest.architecture.vocabSize"
|
|
43
|
+
},
|
|
44
|
+
{
|
|
45
|
+
"factId": "architecture.max_position_embeddings",
|
|
46
|
+
"sourcePointers": ["/max_position_embeddings", "/text_config/max_position_embeddings"],
|
|
47
|
+
"owner": "manifest.architecture.maxSeqLen"
|
|
48
|
+
},
|
|
49
|
+
{
|
|
50
|
+
"factId": "architecture.activation",
|
|
51
|
+
"sourcePointers": ["/hidden_act", "/hidden_activation", "/text_config/hidden_act", "/text_config/hidden_activation"],
|
|
52
|
+
"owner": "manifest.inference.ffn.activation",
|
|
53
|
+
"verificationBoundary": "layer.0.ffn.output"
|
|
54
|
+
},
|
|
55
|
+
{
|
|
56
|
+
"factId": "architecture.normalization.epsilon",
|
|
57
|
+
"sourcePointers": ["/rms_norm_eps", "/layer_norm_eps", "/text_config/rms_norm_eps", "/text_config/layer_norm_eps"],
|
|
58
|
+
"owner": "manifest.inference.normalization.epsilon",
|
|
59
|
+
"verificationBoundary": "layer.0.attention.q.pre_rope"
|
|
60
|
+
},
|
|
61
|
+
{
|
|
62
|
+
"factId": "architecture.rope.theta",
|
|
63
|
+
"sourcePointers": ["/rope_theta", "/text_config/rope_theta"],
|
|
64
|
+
"owner": "manifest.inference.rope.ropeTheta",
|
|
65
|
+
"verificationBoundary": "layer.0.attention.q.post_rope"
|
|
66
|
+
},
|
|
67
|
+
{
|
|
68
|
+
"factId": "architecture.rope.scaling",
|
|
69
|
+
"sourcePointers": ["/rope_scaling", "/text_config/rope_scaling"],
|
|
70
|
+
"owner": "manifest.inference.rope.scaling",
|
|
71
|
+
"verificationBoundary": "layer.0.attention.q.post_rope"
|
|
72
|
+
},
|
|
73
|
+
{
|
|
74
|
+
"factId": "architecture.attention.sliding_window",
|
|
75
|
+
"sourcePointers": ["/sliding_window", "/text_config/sliding_window"],
|
|
76
|
+
"owner": "manifest.inference.attention.slidingWindow",
|
|
77
|
+
"verificationBoundary": "layer.0.attention.output"
|
|
78
|
+
},
|
|
79
|
+
{
|
|
80
|
+
"factId": "architecture.attention.bias",
|
|
81
|
+
"sourcePointers": ["/attention_bias", "/text_config/attention_bias"],
|
|
82
|
+
"owner": "manifest.inference.attention.attentionBias",
|
|
83
|
+
"verificationBoundary": "layer.0.attention.output"
|
|
84
|
+
},
|
|
85
|
+
{
|
|
86
|
+
"factId": "architecture.layer_pattern",
|
|
87
|
+
"sourcePointers": ["/layer_types", "/text_config/layer_types"],
|
|
88
|
+
"owner": "manifest.inference.layerPattern.layerTypes",
|
|
89
|
+
"verificationBoundary": "layer.0.attention.output"
|
|
90
|
+
},
|
|
91
|
+
{
|
|
92
|
+
"factId": "architecture.tie_word_embeddings",
|
|
93
|
+
"sourcePointers": ["/tie_word_embeddings", "/text_config/tie_word_embeddings"],
|
|
94
|
+
"owner": "manifest.inference.output.tieWordEmbeddings",
|
|
95
|
+
"verificationBoundary": "model.logits"
|
|
96
|
+
},
|
|
97
|
+
{
|
|
98
|
+
"factId": "tokenizer.bos_token_id",
|
|
99
|
+
"sourcePointers": ["/bos_token_id"],
|
|
100
|
+
"owner": "manifest.tokenizer.bosTokenId"
|
|
101
|
+
},
|
|
102
|
+
{
|
|
103
|
+
"factId": "tokenizer.eos_token_id",
|
|
104
|
+
"sourcePointers": ["/eos_token_id"],
|
|
105
|
+
"owner": "manifest.tokenizer.eosTokenId"
|
|
106
|
+
},
|
|
107
|
+
{
|
|
108
|
+
"factId": "tokenizer.pad_token_id",
|
|
109
|
+
"sourcePointers": ["/pad_token_id"],
|
|
110
|
+
"owner": "manifest.tokenizer.padTokenId"
|
|
111
|
+
}
|
|
112
|
+
],
|
|
113
|
+
"requiredFacts": [
|
|
114
|
+
"architecture.model_type",
|
|
115
|
+
"architecture.hidden_size",
|
|
116
|
+
"architecture.intermediate_size",
|
|
117
|
+
"architecture.layer_count",
|
|
118
|
+
"architecture.attention.query_heads",
|
|
119
|
+
"architecture.vocab_size"
|
|
120
|
+
],
|
|
121
|
+
"architectureKeyPattern": "(attention|head|hidden|layer|norm|rope|rotary|position|sliding|activation|expert|moe|vocab|embedding|architecture|model_type)"
|
|
122
|
+
}
|
|
@@ -0,0 +1,80 @@
|
|
|
1
|
+
{
|
|
2
|
+
"$schema": "../schema/observation-policy-registry.schema.json",
|
|
3
|
+
"schema": "doppler.observation-policy-registry/v1",
|
|
4
|
+
"defaultPolicyId": "demo/always-on",
|
|
5
|
+
"policies": [
|
|
6
|
+
{
|
|
7
|
+
"id": "demo/always-on",
|
|
8
|
+
"label": "Always-on evidence",
|
|
9
|
+
"modifiesExecution": false,
|
|
10
|
+
"performanceRepresentative": true,
|
|
11
|
+
"requiredCaptures": [
|
|
12
|
+
"artifact-identity",
|
|
13
|
+
"prompt-token-ids",
|
|
14
|
+
"generated-token-ids",
|
|
15
|
+
"wall-timing"
|
|
16
|
+
],
|
|
17
|
+
"allowedClaimTypes": [
|
|
18
|
+
"execution-identity",
|
|
19
|
+
"coarse-performance"
|
|
20
|
+
],
|
|
21
|
+
"gpuTimestampQueries": false,
|
|
22
|
+
"perplexity": null
|
|
23
|
+
},
|
|
24
|
+
{
|
|
25
|
+
"id": "demo/guided-quality",
|
|
26
|
+
"label": "Guided quality inspection",
|
|
27
|
+
"modifiesExecution": true,
|
|
28
|
+
"performanceRepresentative": false,
|
|
29
|
+
"requiredCaptures": [
|
|
30
|
+
"artifact-identity",
|
|
31
|
+
"prompt-token-ids",
|
|
32
|
+
"generated-token-ids",
|
|
33
|
+
"selected-token-probabilities",
|
|
34
|
+
"top-candidates",
|
|
35
|
+
"word-quality"
|
|
36
|
+
],
|
|
37
|
+
"allowedClaimTypes": [
|
|
38
|
+
"quality-inspection"
|
|
39
|
+
],
|
|
40
|
+
"gpuTimestampQueries": false,
|
|
41
|
+
"perplexity": {
|
|
42
|
+
"wordSegmentation": "doppler.word-segmentation/unicode-whitespace-v1",
|
|
43
|
+
"aggregation": "doppler.perplexity/summed-word-surprisal-v1",
|
|
44
|
+
"rollingWindow": {
|
|
45
|
+
"unit": "words",
|
|
46
|
+
"size": 8
|
|
47
|
+
}
|
|
48
|
+
}
|
|
49
|
+
},
|
|
50
|
+
{
|
|
51
|
+
"id": "demo/deep-xray",
|
|
52
|
+
"label": "Deep X-Ray",
|
|
53
|
+
"modifiesExecution": true,
|
|
54
|
+
"performanceRepresentative": false,
|
|
55
|
+
"requiredCaptures": [
|
|
56
|
+
"artifact-identity",
|
|
57
|
+
"prompt-token-ids",
|
|
58
|
+
"generated-token-ids",
|
|
59
|
+
"selected-token-probabilities",
|
|
60
|
+
"top-candidates",
|
|
61
|
+
"word-quality",
|
|
62
|
+
"execution-plan",
|
|
63
|
+
"operator-diagnostics",
|
|
64
|
+
"gpu-timestamps"
|
|
65
|
+
],
|
|
66
|
+
"allowedClaimTypes": [
|
|
67
|
+
"diagnostic"
|
|
68
|
+
],
|
|
69
|
+
"gpuTimestampQueries": true,
|
|
70
|
+
"perplexity": {
|
|
71
|
+
"wordSegmentation": "doppler.word-segmentation/unicode-whitespace-v1",
|
|
72
|
+
"aggregation": "doppler.perplexity/summed-word-surprisal-v1",
|
|
73
|
+
"rollingWindow": {
|
|
74
|
+
"unit": "words",
|
|
75
|
+
"size": 8
|
|
76
|
+
}
|
|
77
|
+
}
|
|
78
|
+
}
|
|
79
|
+
]
|
|
80
|
+
}
|
|
@@ -1,6 +1,9 @@
|
|
|
1
1
|
import { log } from '../debug/index.js';
|
|
2
2
|
import { PARAM_CATEGORIES, CategoryRules } from './param-categories.js';
|
|
3
|
-
import {
|
|
3
|
+
import {
|
|
4
|
+
REFACTOR_RECEIPT_POLICIES,
|
|
5
|
+
TOOLING_DIAGNOSTICS,
|
|
6
|
+
} from './schema/tooling.schema.js';
|
|
4
7
|
import { validateEcosystemConfig } from './schema/ecosystem.schema.js';
|
|
5
8
|
import { isPlainObject } from '../utils/plain-object.js';
|
|
6
9
|
|
|
@@ -50,6 +53,12 @@ export function validateRuntimeOverrides(overrides) {
|
|
|
50
53
|
assertRuntimeOverrideObject(overrides?.inference, key, 'runtime.inference');
|
|
51
54
|
}
|
|
52
55
|
validateRuntimeKernelPath('runtime.inference.kernelPath', overrides?.inference?.kernelPath);
|
|
56
|
+
if (overrides?.shared?.tooling?.intent !== undefined) {
|
|
57
|
+
throw new Error(
|
|
58
|
+
'DopplerConfigError: runtime.shared.tooling.intent is removed. ' +
|
|
59
|
+
'Command intent is normalized from the command request.'
|
|
60
|
+
);
|
|
61
|
+
}
|
|
53
62
|
|
|
54
63
|
const modelOverrides = overrides?.inference?.modelOverrides;
|
|
55
64
|
validateModelOverrides(modelOverrides, 'runtime.inference.modelOverrides');
|
|
@@ -155,7 +164,7 @@ export function validateRuntimeConfig(runtimeConfig) {
|
|
|
155
164
|
);
|
|
156
165
|
}
|
|
157
166
|
|
|
158
|
-
|
|
167
|
+
validateToolingDiagnostics(runtimeConfig);
|
|
159
168
|
validateEcosystemConfig(runtimeConfig.shared?.ecosystem);
|
|
160
169
|
|
|
161
170
|
const debug = runtimeConfig.shared?.debug;
|
|
@@ -188,50 +197,26 @@ function validateRuntimeKernelPath(label, value) {
|
|
|
188
197
|
}
|
|
189
198
|
}
|
|
190
199
|
|
|
191
|
-
function
|
|
200
|
+
function validateToolingDiagnostics(runtimeConfig) {
|
|
192
201
|
const tooling = runtimeConfig.shared?.tooling;
|
|
193
|
-
const intent = tooling?.intent ?? null;
|
|
194
202
|
const diagnostics = tooling?.diagnostics ?? null;
|
|
195
203
|
|
|
196
|
-
if (intent !== null && !TOOLING_INTENTS.includes(intent)) {
|
|
197
|
-
throw new Error(
|
|
198
|
-
`DopplerConfigError: runtime.shared.tooling.intent must be one of ` +
|
|
199
|
-
`${TOOLING_INTENTS.join(', ')} or null.`
|
|
200
|
-
);
|
|
201
|
-
}
|
|
202
|
-
|
|
203
204
|
if (diagnostics !== null && !TOOLING_DIAGNOSTICS.includes(diagnostics)) {
|
|
204
205
|
throw new Error(
|
|
205
206
|
`DopplerConfigError: runtime.shared.tooling.diagnostics must be one of ` +
|
|
206
207
|
`${TOOLING_DIAGNOSTICS.join(', ')}.`
|
|
207
208
|
);
|
|
208
209
|
}
|
|
209
|
-
|
|
210
|
-
if (
|
|
211
|
-
|
|
212
|
-
|
|
213
|
-
|
|
214
|
-
|
|
215
|
-
|
|
216
|
-
|
|
217
|
-
|
|
218
|
-
if (debug?.probes?.length) violations.push('runtime.shared.debug.probes');
|
|
219
|
-
if (debug?.profiler?.enabled) violations.push('runtime.shared.debug.profiler.enabled');
|
|
220
|
-
if (benchmarkRun?.debug) violations.push('runtime.shared.benchmark.run.debug');
|
|
221
|
-
if (benchmarkRun?.profile) violations.push('runtime.shared.benchmark.run.profile');
|
|
222
|
-
if (benchmarkRun?.captureMemoryTimeSeries) {
|
|
223
|
-
violations.push('runtime.shared.benchmark.run.captureMemoryTimeSeries');
|
|
210
|
+
const refactorReceipts = tooling?.refactorReceipts ?? null;
|
|
211
|
+
if (
|
|
212
|
+
refactorReceipts !== null
|
|
213
|
+
&& !REFACTOR_RECEIPT_POLICIES.includes(refactorReceipts)
|
|
214
|
+
) {
|
|
215
|
+
throw new Error(
|
|
216
|
+
'DopplerConfigError: runtime.shared.tooling.refactorReceipts must be one of ' +
|
|
217
|
+
`${REFACTOR_RECEIPT_POLICIES.join(', ')}.`
|
|
218
|
+
);
|
|
224
219
|
}
|
|
225
|
-
|
|
226
|
-
if (violations.length === 0) return;
|
|
227
|
-
|
|
228
|
-
throw new Error(
|
|
229
|
-
'DopplerConfigError: runtime.shared.tooling.intent="calibrate" forbids ' +
|
|
230
|
-
'investigation instrumentation.\n' +
|
|
231
|
-
`Disable ${violations.join(', ')} or set runtime.shared.tooling.intent="investigate".\n` +
|
|
232
|
-
'If this run is launched via "bench", use "debug" for profiling/trace profiles because ' +
|
|
233
|
-
'"bench" enforces intent="calibrate".'
|
|
234
|
-
);
|
|
235
220
|
}
|
|
236
221
|
|
|
237
222
|
function flattenObject(obj, prefix = '') {
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "f16-activations",
|
|
4
4
|
"description": "Experimental F16 activations for reduced memory bandwidth",
|
|
5
5
|
"intent": "investigate",
|
|
6
|
+
"compatibleIntents": ["investigate"],
|
|
6
7
|
"stability": "canonical",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-02-25T00:00:00Z",
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "f16-batched",
|
|
4
4
|
"description": "F16 activations with multi-token decode batching for lower submit/readback overhead.",
|
|
5
5
|
"intent": "investigate",
|
|
6
|
+
"compatibleIntents": ["investigate"],
|
|
6
7
|
"stability": "canonical",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-02-25T00:00:00Z",
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "debug-logits",
|
|
4
4
|
"description": "Trace sample + logits with a top-k probe for a single token.",
|
|
5
5
|
"intent": "investigate",
|
|
6
|
+
"compatibleIntents": ["investigate"],
|
|
6
7
|
"stability": "canonical",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-02-25T00:00:00Z",
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "gemma3-bench-q4k-alt",
|
|
4
4
|
"description": "Gemma 3 1B Q4K alternate throughput profile for batched decode tuning.",
|
|
5
5
|
"intent": "calibrate",
|
|
6
|
+
"compatibleIntents": ["calibrate"],
|
|
6
7
|
"stability": "experimental",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-03-25T01:45:00Z",
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "gemma3-bench-q4k-batched",
|
|
4
4
|
"description": "Gemma 3 1B Q4K benchmark profile with batch decode enabled.",
|
|
5
5
|
"intent": "calibrate",
|
|
6
|
+
"compatibleIntents": ["calibrate"],
|
|
6
7
|
"stability": "experimental",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-03-25T00:00:00Z",
|