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
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "gemma3-bench-q4k",
|
|
4
4
|
"description": "Benchmark run for Gemma 3 1B Q4K — calibration-mode throughput measurement.",
|
|
5
5
|
"intent": "calibrate",
|
|
6
|
+
"compatibleIntents": ["calibrate"],
|
|
6
7
|
"stability": "experimental",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-02-25T00:00:00Z",
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "gemma4-bench-q4k",
|
|
4
4
|
"description": "Benchmark run for Gemma 4 E2B Q4K with deterministic generation settings.",
|
|
5
5
|
"intent": "calibrate",
|
|
6
|
+
"compatibleIntents": ["calibrate"],
|
|
6
7
|
"stability": "experimental",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-04-14T00:00:00Z",
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "qwen3-bench-q4k-alt",
|
|
4
4
|
"description": "Qwen 3.5 0.8B 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": "qwen3-bench-q4k-batched",
|
|
4
4
|
"description": "Qwen 3.5 0.8B 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",
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "gemma3-browser-diagnose-q4k",
|
|
4
4
|
"description": "Browser/Chromium investigation profile for Gemma 3 1B Q4K that preserves compare-like runtime conditions while enabling GPU profiling and low-overhead operator diagnostics.",
|
|
5
5
|
"intent": "investigate",
|
|
6
|
+
"compatibleIntents": ["investigate"],
|
|
6
7
|
"stability": "experimental",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-03-28T00:00:00Z",
|
|
@@ -11,7 +12,6 @@
|
|
|
11
12
|
"runtime": {
|
|
12
13
|
"shared": {
|
|
13
14
|
"tooling": {
|
|
14
|
-
"intent": "investigate",
|
|
15
15
|
"diagnostics": "always"
|
|
16
16
|
},
|
|
17
17
|
"debug": {
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "gemma3-q4k-debug",
|
|
4
4
|
"description": "Debug Gemma 3 1B Q4K inference - check kernel selection and alignment",
|
|
5
5
|
"intent": "investigate",
|
|
6
|
+
"compatibleIntents": ["investigate"],
|
|
6
7
|
"stability": "experimental",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-02-25T00:00:00Z",
|
|
@@ -10,9 +11,6 @@
|
|
|
10
11
|
"model": "gemma-3-1b-it-q4k-ehf16-af32",
|
|
11
12
|
"runtime": {
|
|
12
13
|
"shared": {
|
|
13
|
-
"tooling": {
|
|
14
|
-
"intent": "investigate"
|
|
15
|
-
},
|
|
16
14
|
"debug": {
|
|
17
15
|
"logLevel": {
|
|
18
16
|
"defaultLogLevel": "verbose"
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "gemma4-e2b-browser-diagnose-q4k",
|
|
4
4
|
"description": "Browser/Chromium investigation profile for Gemma 4 E2B Q4K that preserves compare-like runtime conditions while enabling GPU profiling and low-overhead operator diagnostics. Mirrors gemma3-browser-diagnose-q4k. Use for classifying the prefill wall (GPU compute vs submit/readback/orchestration) before any kernel edit.",
|
|
5
5
|
"intent": "investigate",
|
|
6
|
+
"compatibleIntents": ["investigate"],
|
|
6
7
|
"stability": "experimental",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-04-17T00:00:00Z",
|
|
@@ -11,7 +12,6 @@
|
|
|
11
12
|
"runtime": {
|
|
12
13
|
"shared": {
|
|
13
14
|
"tooling": {
|
|
14
|
-
"intent": "investigate",
|
|
15
15
|
"diagnostics": "always"
|
|
16
16
|
},
|
|
17
17
|
"debug": {
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "qwen35-browser-diagnose-q4k",
|
|
4
4
|
"description": "Browser/Chromium investigation profile for Qwen 3.5 0.8B Q4K that preserves compare-like prefill conditions while enabling GPU profiling and low-overhead operator diagnostics.",
|
|
5
5
|
"intent": "investigate",
|
|
6
|
+
"compatibleIntents": ["investigate"],
|
|
6
7
|
"stability": "experimental",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-03-29T00:00:00Z",
|
|
@@ -11,7 +12,6 @@
|
|
|
11
12
|
"runtime": {
|
|
12
13
|
"shared": {
|
|
13
14
|
"tooling": {
|
|
14
|
-
"intent": "investigate",
|
|
15
15
|
"diagnostics": "always"
|
|
16
16
|
},
|
|
17
17
|
"debug": {
|
|
@@ -2,6 +2,7 @@
|
|
|
2
2
|
"id": "experiments/verify/gemma3-verify",
|
|
3
3
|
"name": "gemma3-verify",
|
|
4
4
|
"intent": "verify",
|
|
5
|
+
"compatibleIntents": ["verify"],
|
|
5
6
|
"stability": "experimental",
|
|
6
7
|
"owner": "doppler-core",
|
|
7
8
|
"createdAtUtc": "2026-02-25T00:00:00Z",
|
|
@@ -9,9 +10,6 @@
|
|
|
9
10
|
"model": "gemma-3-1b-it-f16-af32",
|
|
10
11
|
"runtime": {
|
|
11
12
|
"shared": {
|
|
12
|
-
"tooling": {
|
|
13
|
-
"intent": "verify"
|
|
14
|
-
},
|
|
15
13
|
"debug": {
|
|
16
14
|
"logLevel": {
|
|
17
15
|
"defaultLogLevel": "warn"
|
|
@@ -2,6 +2,7 @@
|
|
|
2
2
|
"id": "experiments/verify/gemma4-e2b-verify",
|
|
3
3
|
"name": "gemma4-e2b-verify",
|
|
4
4
|
"intent": "verify",
|
|
5
|
+
"compatibleIntents": ["verify"],
|
|
5
6
|
"stability": "experimental",
|
|
6
7
|
"owner": "doppler-core",
|
|
7
8
|
"createdAtUtc": "2026-04-04T00:00:00Z",
|
|
@@ -9,9 +10,6 @@
|
|
|
9
10
|
"model": "gemma-4-e2b-it-q4k-ehf16-af32",
|
|
10
11
|
"runtime": {
|
|
11
12
|
"shared": {
|
|
12
|
-
"tooling": {
|
|
13
|
-
"intent": "verify"
|
|
14
|
-
},
|
|
15
13
|
"debug": {
|
|
16
14
|
"logLevel": {
|
|
17
15
|
"defaultLogLevel": "warn"
|
|
@@ -2,6 +2,7 @@
|
|
|
2
2
|
"id": "experiments/verify/lfm2-verify",
|
|
3
3
|
"name": "lfm2-verify",
|
|
4
4
|
"intent": "verify",
|
|
5
|
+
"compatibleIntents": ["verify"],
|
|
5
6
|
"stability": "experimental",
|
|
6
7
|
"owner": "doppler-core",
|
|
7
8
|
"createdAtUtc": "2026-03-16T00:00:00Z",
|
|
@@ -9,9 +10,6 @@
|
|
|
9
10
|
"model": "lfm2-5-1-2b-instruct-q4k-ehf16-af32",
|
|
10
11
|
"runtime": {
|
|
11
12
|
"shared": {
|
|
12
|
-
"tooling": {
|
|
13
|
-
"intent": "verify"
|
|
14
|
-
},
|
|
15
13
|
"debug": {
|
|
16
14
|
"logLevel": {
|
|
17
15
|
"defaultLogLevel": "warn"
|
|
@@ -2,6 +2,7 @@
|
|
|
2
2
|
"id": "experiments/verify/translategemma-verify",
|
|
3
3
|
"name": "translategemma-verify",
|
|
4
4
|
"intent": "verify",
|
|
5
|
+
"compatibleIntents": ["verify"],
|
|
5
6
|
"stability": "experimental",
|
|
6
7
|
"owner": "doppler-core",
|
|
7
8
|
"createdAtUtc": "2026-03-16T00:00:00Z",
|
|
@@ -9,9 +10,6 @@
|
|
|
9
10
|
"model": "translategemma-4b-it-q4k-ehf16-af32",
|
|
10
11
|
"runtime": {
|
|
11
12
|
"shared": {
|
|
12
|
-
"tooling": {
|
|
13
|
-
"intent": "verify"
|
|
14
|
-
},
|
|
15
13
|
"debug": {
|
|
16
14
|
"logLevel": {
|
|
17
15
|
"defaultLogLevel": "warn"
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "gemma2-debug",
|
|
4
4
|
"description": "Gemma 2 probe-focused debugging with targeted trace categories",
|
|
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": "gemma2-pipeline-debug",
|
|
4
4
|
"description": "Gemma 2 explicit pipeline with focused trace/probes",
|
|
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": "gemma2-pipeline",
|
|
4
4
|
"description": "Gemma 2 explicit layer pipeline (all norms as separate steps)",
|
|
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-layer-probe",
|
|
4
4
|
"description": "Probe layer outputs to compare against HuggingFace reference",
|
|
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": "qwen3-5-layer-probe",
|
|
4
4
|
"description": "Probe all 24 layer outputs in Qwen 3.5 to isolate where the hidden state distribution collapses.",
|
|
5
5
|
"intent": "investigate",
|
|
6
|
+
"compatibleIntents": ["investigate"],
|
|
6
7
|
"stability": "canonical",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-03-13T00:00:00Z",
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "qwen3-5-layer3-boundary-probe",
|
|
4
4
|
"description": "Capture full-attention boundary checkpoints around layer 3 to isolate Q/K norm and partial RoPE behavior.",
|
|
5
5
|
"intent": "investigate",
|
|
6
|
+
"compatibleIntents": ["investigate"],
|
|
6
7
|
"stability": "canonical",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-03-23T00:00:00Z",
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "qwen3-5-linear-attn-debug",
|
|
4
4
|
"description": "Probe linear attention intermediates in Qwen 3.5 layer 0 for comparison with HF reference.",
|
|
5
5
|
"intent": "investigate",
|
|
6
|
+
"compatibleIntents": ["investigate"],
|
|
6
7
|
"stability": "canonical",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-03-13T00:00:00Z",
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "metal-apple-q4k",
|
|
4
4
|
"description": "Metal (Apple GPU) tuned default for Q4K F16 activation/cache runtime policy.",
|
|
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": "diffusiongemma-26b-a4b-throughput",
|
|
4
4
|
"description": "DiffusionGemma calibration profile that selects the GPU route-expert scheduler for the Gemma 4 A4B MoE block-diffusion artifact.",
|
|
5
5
|
"intent": "calibrate",
|
|
6
|
+
"compatibleIntents": ["calibrate"],
|
|
6
7
|
"stability": "experimental",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-06-12T00:00:00Z",
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "gemma3-270m-q4k-throughput-overlapped-probe",
|
|
4
4
|
"description": "Gemma 3 270M calibration profile that uses the Q4K 8x8 throughput cadence, overlapped decode readback, fused packed-QKV QK-norm/RoPE decode path, and sandwich RMSNorm-pair fusion for readback-bound browser diagnostics. Used to reduce decode readback submissions and dispatch count without changing shared prompt or sampling semantics.",
|
|
5
5
|
"intent": "calibrate",
|
|
6
|
+
"compatibleIntents": ["calibrate"],
|
|
6
7
|
"stability": "experimental",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-06-26T00:00:00Z",
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "gemma3-270m-q4k-throughput-probe",
|
|
4
4
|
"description": "Gemma 3 270M calibration profile that keeps the Q4K 8x4 throughput cadence and opts the Q4K lane into sync prefill, retained Q4K materialization, the wide-tile Q4K prefill path, and sequential decode readback. Used to measure reusable decode/prefill bottlenecks without changing shared prompt or sampling semantics.",
|
|
5
5
|
"intent": "calibrate",
|
|
6
|
+
"compatibleIntents": ["calibrate"],
|
|
6
7
|
"stability": "experimental",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-06-26T00:00:00Z",
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "gemma4-12b-rdrr-safe-b4",
|
|
4
4
|
"description": "Deprecated Gemma 4 12B RDRR safe profile alias. Use profiles/gemma4-12b-rdrr-safe-single for coherent local QAT runs.",
|
|
5
5
|
"intent": "investigate",
|
|
6
|
+
"compatibleIntents": ["investigate"],
|
|
6
7
|
"stability": "deprecated",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-06-07T00:00:00Z",
|
|
@@ -11,9 +12,6 @@
|
|
|
11
12
|
"extends": "profiles/default",
|
|
12
13
|
"runtime": {
|
|
13
14
|
"shared": {
|
|
14
|
-
"tooling": {
|
|
15
|
-
"intent": "investigate"
|
|
16
|
-
},
|
|
17
15
|
"debug": {
|
|
18
16
|
"logLevel": {
|
|
19
17
|
"defaultLogLevel": "info"
|
|
@@ -3,15 +3,13 @@
|
|
|
3
3
|
"name": "gemma4-12b-rdrr-safe-single",
|
|
4
4
|
"description": "Gemma 4 12B RDRR safe single-token decode cadence for coherent local QAT runs on consumer hardware.",
|
|
5
5
|
"intent": "investigate",
|
|
6
|
+
"compatibleIntents": ["investigate"],
|
|
6
7
|
"stability": "experimental",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-06-08T00:00:00Z",
|
|
9
10
|
"extends": "profiles/default",
|
|
10
11
|
"runtime": {
|
|
11
12
|
"shared": {
|
|
12
|
-
"tooling": {
|
|
13
|
-
"intent": "investigate"
|
|
14
|
-
},
|
|
15
13
|
"debug": {
|
|
16
14
|
"logLevel": {
|
|
17
15
|
"defaultLogLevel": "info"
|
|
@@ -3,17 +3,13 @@
|
|
|
3
3
|
"name": "gemma4-31b-f16-activations-probe",
|
|
4
4
|
"description": "Experimental Gemma 4 31B text profile that requests the execution-v1 f16 compute lane against the f16 activation manifest variant sharing the existing Q4K weight pack. Use for browser/WebGPU proof only after transcript parity is captured.",
|
|
5
5
|
"intent": "investigate",
|
|
6
|
+
"compatibleIntents": ["investigate"],
|
|
6
7
|
"stability": "experimental",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-04-28T00:00:00Z",
|
|
9
10
|
"extends": "profiles/default",
|
|
10
11
|
"model": "gemma-4-31b-it-text-q4k-ehf16-af16",
|
|
11
12
|
"runtime": {
|
|
12
|
-
"shared": {
|
|
13
|
-
"tooling": {
|
|
14
|
-
"intent": "investigate"
|
|
15
|
-
}
|
|
16
|
-
},
|
|
17
13
|
"inference": {
|
|
18
14
|
"compute": {
|
|
19
15
|
"activationDtype": "f16"
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "gemma4-e2b-baseline-probe",
|
|
4
4
|
"description": "One-off diagnostic profile with layer-0 kernel-debug dumps. Baseline (no RMSNorm+Q4K fusion) for comparison against gemma4-e2b-fused-rmsnorm-probe. Not for perf measurement — only correctness probe.",
|
|
5
5
|
"intent": "investigate",
|
|
6
|
+
"compatibleIntents": ["investigate"],
|
|
6
7
|
"stability": "experimental",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-04-18T00:00:00Z",
|
|
@@ -10,9 +11,6 @@
|
|
|
10
11
|
"model": "gemma-4-e2b-it-q4k-ehf16-af32",
|
|
11
12
|
"runtime": {
|
|
12
13
|
"shared": {
|
|
13
|
-
"tooling": {
|
|
14
|
-
"intent": "investigate"
|
|
15
|
-
},
|
|
16
14
|
"debug": {
|
|
17
15
|
"logLevel": {
|
|
18
16
|
"defaultLogLevel": "debug"
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "gemma4-e2b-fused-rmsnorm-probe",
|
|
4
4
|
"description": "One-off diagnostic profile with useFusedRmsnormWideTile=true and layer-0 kernel-debug dumps. Compares fused RMSNorm+Q4K WideTile output to baseline rmsnorm+matmul chain at q/k/v_proj call sites. Not for perf measurement — only correctness probe.",
|
|
5
5
|
"intent": "investigate",
|
|
6
|
+
"compatibleIntents": ["investigate"],
|
|
6
7
|
"stability": "experimental",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-04-18T00:00:00Z",
|
|
@@ -10,9 +11,6 @@
|
|
|
10
11
|
"model": "gemma-4-e2b-it-q4k-ehf16-af32",
|
|
11
12
|
"runtime": {
|
|
12
13
|
"shared": {
|
|
13
|
-
"tooling": {
|
|
14
|
-
"intent": "investigate"
|
|
15
|
-
},
|
|
16
14
|
"debug": {
|
|
17
15
|
"logLevel": {
|
|
18
16
|
"defaultLogLevel": "debug"
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "gemma4-e2b-fused-rmsnorm-test",
|
|
4
4
|
"description": "Test profile with useFusedRmsnormWideTile=true for end-to-end correctness + perf measurement. Model fast-path settings come from the converted manifest.",
|
|
5
5
|
"intent": "calibrate",
|
|
6
|
+
"compatibleIntents": ["calibrate"],
|
|
6
7
|
"stability": "experimental",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-04-18T00:00:00Z",
|
|
@@ -3,17 +3,13 @@
|
|
|
3
3
|
"name": "gemma4-e2b-int4ple-baseline-probe",
|
|
4
4
|
"description": "Apples-to-apples baseline counterpart for profiles/gemma4-e2b-int4ple-retain-q4k-probe. Same prompt/maxTokens/sampling but runs against the shipped manifest defaults (retainQ4KMaterialization=false, activationDtype=f32). Used to attribute perf deltas to the retain=true switch only.",
|
|
5
5
|
"intent": "investigate",
|
|
6
|
+
"compatibleIntents": ["investigate"],
|
|
6
7
|
"stability": "experimental",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-05-01T00:00:00Z",
|
|
9
10
|
"extends": "profiles/default",
|
|
10
11
|
"model": "gemma-4-e2b-it-q4k-ehf16-af32-int4ple",
|
|
11
12
|
"runtime": {
|
|
12
|
-
"shared": {
|
|
13
|
-
"tooling": {
|
|
14
|
-
"intent": "investigate"
|
|
15
|
-
}
|
|
16
|
-
},
|
|
17
13
|
"inference": {
|
|
18
14
|
"generation": {
|
|
19
15
|
"maxTokens": 16
|
|
@@ -3,17 +3,13 @@
|
|
|
3
3
|
"name": "gemma4-e2b-int4ple-f16a-probe",
|
|
4
4
|
"description": "Experimental Gemma 4 E2B INT4 PLE profile that requests f16 activation narrowing against the existing execution-v1 manifest. Stage-1 probe only; do not use for release claims until correctness and perf gates pass.",
|
|
5
5
|
"intent": "investigate",
|
|
6
|
+
"compatibleIntents": ["investigate"],
|
|
6
7
|
"stability": "experimental",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-04-20T00:00:00Z",
|
|
9
10
|
"extends": "profiles/throughput",
|
|
10
11
|
"model": "gemma-4-e2b-it-q4k-ehf16-af32-int4ple",
|
|
11
12
|
"runtime": {
|
|
12
|
-
"shared": {
|
|
13
|
-
"tooling": {
|
|
14
|
-
"intent": "investigate"
|
|
15
|
-
}
|
|
16
|
-
},
|
|
17
13
|
"inference": {
|
|
18
14
|
"compute": {
|
|
19
15
|
"activationDtype": "f16"
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "gemma4-e2b-int4ple-reference-export",
|
|
4
4
|
"description": "Gemma 4 E2B INT4 PLE reference-export profile for constrained Node WebGPU hosts. Keeps low-memory loading caps while preserving plain contiguous KV cache required by the execution-v1 INT4 PLE path.",
|
|
5
5
|
"intent": "verify",
|
|
6
|
+
"compatibleIntents": ["verify"],
|
|
6
7
|
"stability": "experimental",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-04-25T00:00:00Z",
|
|
@@ -10,9 +11,6 @@
|
|
|
10
11
|
"model": "gemma-4-e2b-it-q4k-ehf16-af32-int4ple",
|
|
11
12
|
"runtime": {
|
|
12
13
|
"shared": {
|
|
13
|
-
"tooling": {
|
|
14
|
-
"intent": "verify"
|
|
15
|
-
},
|
|
16
14
|
"bufferPool": {
|
|
17
15
|
"budget": {
|
|
18
16
|
"maxTotalBytes": 13958643712,
|
|
@@ -3,17 +3,13 @@
|
|
|
3
3
|
"name": "gemma4-e2b-int4ple-retain-q4k-probe",
|
|
4
4
|
"description": "Investigation probe that opts gemma-4-e2b INT4-PLE into retainQ4KMaterialization=true plus f16 activations so the fused-q4k+f16 kernel pool (q4k_f16a fused FFN, q4_fused_widetile_f16a matmul) can be exercised end-to-end. Existing capability-transforms.rules.json rule force-disables retain on Apple Metal because of the documented L0.ffn_down NaN; AMD/RDNA3 passes through. Use this profile to capture the AMD transcript before any kernel is promoted into a release-eligible runtime profile.",
|
|
5
5
|
"intent": "investigate",
|
|
6
|
+
"compatibleIntents": ["investigate"],
|
|
6
7
|
"stability": "experimental",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-05-01T00:00:00Z",
|
|
9
10
|
"extends": "profiles/default",
|
|
10
11
|
"model": "gemma-4-e2b-it-q4k-ehf16-af32-int4ple",
|
|
11
12
|
"runtime": {
|
|
12
|
-
"shared": {
|
|
13
|
-
"tooling": {
|
|
14
|
-
"intent": "investigate"
|
|
15
|
-
}
|
|
16
|
-
},
|
|
17
13
|
"inference": {
|
|
18
14
|
"compute": {
|
|
19
15
|
"activationDtype": "f16"
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "gemma4-e2b-low-memory",
|
|
4
4
|
"description": "Model-scoped Gemma 4 E2B low-memory profile with explicit contiguous TurboQuant KV compression and shorter context.",
|
|
5
5
|
"intent": "investigate",
|
|
6
|
+
"compatibleIntents": ["investigate"],
|
|
6
7
|
"stability": "experimental",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-04-06T00:00:00Z",
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "gemma4-e2b-prefill-profile",
|
|
4
4
|
"description": "Investigation profile with GPU timestamp-query profiling enabled, longer prompt (prefill≥81), and debug log level so per-kernel timings surface via log.warn('Profile', ...). Used to identify the actual prefill wall.",
|
|
5
5
|
"intent": "investigate",
|
|
6
|
+
"compatibleIntents": ["investigate"],
|
|
6
7
|
"stability": "experimental",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-04-18T00:00:00Z",
|
|
@@ -10,9 +11,6 @@
|
|
|
10
11
|
"model": "gemma-4-e2b-it-q4k-ehf16-af32",
|
|
11
12
|
"runtime": {
|
|
12
13
|
"shared": {
|
|
13
|
-
"tooling": {
|
|
14
|
-
"intent": "investigate"
|
|
15
|
-
},
|
|
16
14
|
"debug": {
|
|
17
15
|
"logLevel": {
|
|
18
16
|
"defaultLogLevel": "warn"
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "gemma4-e2b-throughput-probe",
|
|
4
4
|
"description": "Throughput probe with a longer-prompt override (prefill≥64 tokens) for apples-to-apples baseline measurement against gemma4-e2b-fused-rmsnorm-test. Model fast-path settings come from the converted manifest.",
|
|
5
5
|
"intent": "calibrate",
|
|
6
|
+
"compatibleIntents": ["calibrate"],
|
|
6
7
|
"stability": "experimental",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-04-18T00:00:00Z",
|
|
@@ -3,15 +3,13 @@
|
|
|
3
3
|
"name": "production",
|
|
4
4
|
"description": "Production configuration with optimized settings",
|
|
5
5
|
"intent": "verify",
|
|
6
|
+
"compatibleIntents": ["verify"],
|
|
6
7
|
"stability": "canonical",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-02-25T00:00:00Z",
|
|
9
10
|
"extends": "profiles/default",
|
|
10
11
|
"runtime": {
|
|
11
12
|
"shared": {
|
|
12
|
-
"tooling": {
|
|
13
|
-
"intent": "verify"
|
|
14
|
-
},
|
|
15
13
|
"debug": {
|
|
16
14
|
"logOutput": {
|
|
17
15
|
"stdout": false,
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "qwen-3-5-0-8b-metal-attn128-candidate",
|
|
4
4
|
"description": "Investigation candidate for Qwen 3.5 0.8B browser/Metal parity. Copies the Metal parity profile and changes only online decode attention WORKGROUP_SIZE from the shader default 256 to 128.",
|
|
5
5
|
"intent": "calibrate",
|
|
6
|
+
"compatibleIntents": ["calibrate"],
|
|
6
7
|
"stability": "experimental",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-07-08T00:00:00Z",
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "qwen-3-5-0-8b-metal-maxbatch64-candidate",
|
|
4
4
|
"description": "Investigation candidate for Qwen 3.5 0.8B browser/Metal throughput. Copies the baseline Metal profile and changes only maxBatchDecodeTokens from 128 to 64.",
|
|
5
5
|
"intent": "investigate",
|
|
6
|
+
"compatibleIntents": ["investigate"],
|
|
6
7
|
"stability": "experimental",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-07-08T00:00:00Z",
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "qwen-3-5-0-8b-metal-normed-ffn-candidate",
|
|
4
4
|
"description": "Qwen 3.5 0.8B browser/Metal diagnostic profile copied from the Metal throughput profile with only the post-attention-norm fused gate/up prelude enabled. This is tuning evidence only unless receipts prove exact parity and faster decode.",
|
|
5
5
|
"intent": "calibrate",
|
|
6
|
+
"compatibleIntents": ["calibrate"],
|
|
6
7
|
"stability": "experimental",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-07-08T00:00:00Z",
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "qwen-3-5-0-8b-metal-overlap-ringtokens1-candidate",
|
|
4
4
|
"description": "Investigation candidate for Qwen 3.5 0.8B browser/Metal throughput. Copies the baseline Metal profile and changes only overlapped token ring depth while keeping ringStaging valid for overlapped readback.",
|
|
5
5
|
"intent": "investigate",
|
|
6
|
+
"compatibleIntents": ["investigate"],
|
|
6
7
|
"stability": "experimental",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-07-08T00:00:00Z",
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "qwen-3-5-0-8b-metal-parity",
|
|
4
4
|
"description": "Qwen 3.5 0.8B browser/Metal parity profile. Preserves the release 4x4 decode cadence while enabling exact-safe decode fusions and overlapped readback on Apple Metal.",
|
|
5
5
|
"intent": "calibrate",
|
|
6
|
+
"compatibleIntents": ["calibrate"],
|
|
6
7
|
"stability": "experimental",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-07-08T00:00:00Z",
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "qwen-3-5-0-8b-metal-sequential-candidate",
|
|
4
4
|
"description": "Investigation candidate for Qwen 3.5 0.8B browser/Metal throughput. Copies the baseline Metal profile and changes only readback mode/ring sizing from overlapped to sequential.",
|
|
5
5
|
"intent": "investigate",
|
|
6
|
+
"compatibleIntents": ["investigate"],
|
|
6
7
|
"stability": "experimental",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-07-08T00:00:00Z",
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "qwen-3-5-0-8b-metal-throughput",
|
|
4
4
|
"description": "Qwen 3.5 0.8B browser/Metal throughput profile. Keeps the shared Qwen Q4K execution graph and uses overlapped decode readback on Apple Metal so submit/readback-bound decode can reuse ring staging without changing Vulkan cadence.",
|
|
5
5
|
"intent": "calibrate",
|
|
6
|
+
"compatibleIntents": ["calibrate"],
|
|
6
7
|
"stability": "experimental",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-07-08T00:00:00Z",
|
|
@@ -3,6 +3,7 @@
|
|
|
3
3
|
"name": "qwen-3-5-0-8b-prefill-profile",
|
|
4
4
|
"description": "Qwen 3.5 0.8B investigation profile with GPU timestamp-query profiling enabled and longer prompt (~80 tokens). Used to identify the actual prefill/decode wall via log.warn('Profile', ...).",
|
|
5
5
|
"intent": "investigate",
|
|
6
|
+
"compatibleIntents": ["investigate"],
|
|
6
7
|
"stability": "experimental",
|
|
7
8
|
"owner": "doppler-core",
|
|
8
9
|
"createdAtUtc": "2026-04-18T00:00:00Z",
|
|
@@ -10,9 +11,6 @@
|
|
|
10
11
|
"model": "qwen-3-5-0-8b-q4k-ehaf16",
|
|
11
12
|
"runtime": {
|
|
12
13
|
"shared": {
|
|
13
|
-
"tooling": {
|
|
14
|
-
"intent": "investigate"
|
|
15
|
-
},
|
|
16
14
|
"debug": {
|
|
17
15
|
"logLevel": {
|
|
18
16
|
"defaultLogLevel": "info"
|