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.
Files changed (221) hide show
  1. package/README.md +9 -0
  2. package/package.json +25 -2
  3. package/src/cli/doppler-cli.js +253 -94
  4. package/src/cli/runtime-profiles.js +3 -7
  5. package/src/client/doppler-api.d.ts +28 -0
  6. package/src/client/doppler-api.js +2 -0
  7. package/src/client/inspection.d.ts +122 -0
  8. package/src/client/inspection.js +343 -0
  9. package/src/client/runtime/index.d.ts +54 -0
  10. package/src/client/runtime/index.js +190 -18
  11. package/src/client/runtime/model-manager.d.ts +0 -4
  12. package/src/client/runtime/model-manager.js +0 -28
  13. package/src/client/runtime/model-session.d.ts +48 -1
  14. package/src/client/runtime/model-session.js +149 -30
  15. package/src/client/runtime/model-source.d.ts +10 -1
  16. package/src/client/runtime/scoped-session.d.ts +117 -0
  17. package/src/client/runtime/scoped-session.js +257 -0
  18. package/src/client/runtime/types.d.ts +0 -1
  19. package/src/client/runtime/types.js +0 -1
  20. package/src/config/evidence/boundary-evidence-policy.json +54 -0
  21. package/src/config/evidence/source-intake-policy.json +122 -0
  22. package/src/config/inspection/observation-policies.json +80 -0
  23. package/src/config/param-validator.js +21 -36
  24. package/src/config/runtime/compute/f16-activations.json +1 -0
  25. package/src/config/runtime/compute/f16-batched.json +1 -0
  26. package/src/config/runtime/default.json +1 -0
  27. package/src/config/runtime/diagnostics/debug-logits.json +1 -0
  28. package/src/config/runtime/experiments/bench/gemma3-bench-q4k-alt.json +1 -0
  29. package/src/config/runtime/experiments/bench/gemma3-bench-q4k-batched.json +1 -0
  30. package/src/config/runtime/experiments/bench/gemma3-bench-q4k.json +1 -0
  31. package/src/config/runtime/experiments/bench/gemma4-bench-q4k.json +1 -0
  32. package/src/config/runtime/experiments/bench/qwen3-bench-q4k-alt.json +1 -0
  33. package/src/config/runtime/experiments/bench/qwen3-bench-q4k-batched.json +1 -0
  34. package/src/config/runtime/experiments/debug/gemma3-browser-diagnose-q4k.json +1 -1
  35. package/src/config/runtime/experiments/debug/gemma3-debug-q4k.json +1 -3
  36. package/src/config/runtime/experiments/debug/gemma4-e2b-browser-diagnose-q4k.json +1 -1
  37. package/src/config/runtime/experiments/debug/qwen35-browser-diagnose-q4k.json +1 -1
  38. package/src/config/runtime/experiments/verify/gemma3-verify.json +1 -3
  39. package/src/config/runtime/experiments/verify/gemma4-e2b-verify.json +1 -3
  40. package/src/config/runtime/experiments/verify/lfm2-verify.json +1 -3
  41. package/src/config/runtime/experiments/verify/translategemma-verify.json +1 -3
  42. package/src/config/runtime/model/gemma2-debug.json +1 -0
  43. package/src/config/runtime/model/gemma2-pipeline-debug.json +1 -0
  44. package/src/config/runtime/model/gemma2-pipeline.json +1 -0
  45. package/src/config/runtime/model/gemma3-layer-probe.json +1 -0
  46. package/src/config/runtime/model/qwen3-5-layer-probe.json +1 -0
  47. package/src/config/runtime/model/qwen3-5-layer3-boundary-probe.json +1 -0
  48. package/src/config/runtime/model/qwen3-5-linear-attn-debug.json +1 -0
  49. package/src/config/runtime/optimization-candidates/registry.json +5 -0
  50. package/src/config/runtime/platform/metal-apple-q4k.json +1 -0
  51. package/src/config/runtime/profiles/default.json +1 -0
  52. package/src/config/runtime/profiles/diffusiongemma-26b-a4b-throughput.json +1 -0
  53. package/src/config/runtime/profiles/gemma3-270m-q4k-throughput-overlapped-probe.json +1 -0
  54. package/src/config/runtime/profiles/gemma3-270m-q4k-throughput-probe.json +1 -0
  55. package/src/config/runtime/profiles/gemma4-12b-rdrr-safe-b4.json +1 -3
  56. package/src/config/runtime/profiles/gemma4-12b-rdrr-safe-single.json +1 -3
  57. package/src/config/runtime/profiles/gemma4-31b-f16-activations-probe.json +1 -5
  58. package/src/config/runtime/profiles/gemma4-e2b-baseline-probe.json +1 -3
  59. package/src/config/runtime/profiles/gemma4-e2b-fused-rmsnorm-probe.json +1 -3
  60. package/src/config/runtime/profiles/gemma4-e2b-fused-rmsnorm-test.json +1 -0
  61. package/src/config/runtime/profiles/gemma4-e2b-int4ple-baseline-probe.json +1 -5
  62. package/src/config/runtime/profiles/gemma4-e2b-int4ple-f16a-probe.json +1 -5
  63. package/src/config/runtime/profiles/gemma4-e2b-int4ple-reference-export.json +1 -3
  64. package/src/config/runtime/profiles/gemma4-e2b-int4ple-retain-q4k-probe.json +1 -5
  65. package/src/config/runtime/profiles/gemma4-e2b-low-memory.json +1 -0
  66. package/src/config/runtime/profiles/gemma4-e2b-prefill-profile.json +1 -3
  67. package/src/config/runtime/profiles/gemma4-e2b-throughput-probe.json +1 -0
  68. package/src/config/runtime/profiles/low-memory.json +1 -0
  69. package/src/config/runtime/profiles/production.json +1 -3
  70. package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-attn128-candidate.json +1 -0
  71. package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-maxbatch64-candidate.json +1 -0
  72. package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-normed-ffn-candidate.json +1 -0
  73. package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-overlap-ringtokens1-candidate.json +1 -0
  74. package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-parity.json +1 -0
  75. package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-sequential-candidate.json +1 -0
  76. package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-throughput.json +1 -0
  77. package/src/config/runtime/profiles/qwen-3-5-0-8b-prefill-profile.json +1 -3
  78. package/src/config/runtime/profiles/qwen-3-5-0-8b-throughput-probe.json +1 -0
  79. package/src/config/runtime/profiles/qwen-3-5-2b-f16-primary-throughput.json +1 -5
  80. package/src/config/runtime/profiles/qwen-3-5-2b-metal-parity.json +1 -0
  81. package/src/config/runtime/profiles/qwen-3-5-2b-metal-throughput.json +1 -0
  82. package/src/config/runtime/profiles/qwen-3-5-9b-metal-correctness.json +1 -0
  83. package/src/config/runtime/profiles/qwen-3-embedding-0-6b-metal-stability.json +1 -3
  84. package/src/config/runtime/profiles/qwen-3-embedding-0-6b-metal-throughput.json +1 -0
  85. package/src/config/runtime/profiles/qwen-3-reranker-0-6b-metal-stability.json +1 -0
  86. package/src/config/runtime/profiles/qwen-3-reranker-0-6b-metal-throughput.json +1 -0
  87. package/src/config/runtime/profiles/qwen3-6-27b-f16-activations-probe.json +1 -5
  88. package/src/config/runtime/profiles/rerank-stability.json +1 -3
  89. package/src/config/runtime/profiles/rerank-throughput.json +1 -3
  90. package/src/config/runtime/profiles/simulation.json +1 -3
  91. package/src/config/runtime/profiles/throughput.json +1 -3
  92. package/src/config/runtime/profiles/trace-layers.json +1 -3
  93. package/src/config/runtime/profiles/translategemma-1b-nativekd2-q4k-rdna3-throughput-probe.json +1 -0
  94. package/src/config/runtime/profiles/turboquant-contiguous-prod.json +1 -0
  95. package/src/config/runtime/profiles/turboquant-contiguous.json +1 -0
  96. package/src/config/runtime/profiles/turboquant.json +1 -0
  97. package/src/config/runtime/profiles/vector-stability.json +1 -3
  98. package/src/config/runtime/profiles/vector-throughput.json +1 -0
  99. package/src/config/runtime/profiles/verbose-trace.json +1 -3
  100. package/src/config/runtime/tiers/gemma4-128gb.json +1 -0
  101. package/src/config/runtime/tiers/gemma4-16gb.json +1 -0
  102. package/src/config/runtime/tiers/gemma4-24gb.json +1 -0
  103. package/src/config/runtime/tiers/gemma4-32gb.json +1 -0
  104. package/src/config/schema/benchmark.schema.d.ts +7 -0
  105. package/src/config/schema/benchmark.schema.js +5 -0
  106. package/src/config/schema/index.d.ts +8 -16
  107. package/src/config/schema/index.js +5 -8
  108. package/src/config/schema/kernel-thresholds.schema.d.ts +0 -12
  109. package/src/config/schema/kernel-thresholds.schema.js +0 -12
  110. package/src/config/schema/kernel-warmup.schema.d.ts +1 -3
  111. package/src/config/schema/kernel-warmup.schema.js +0 -1
  112. package/src/config/schema/runtime-profile.schema.d.ts +24 -0
  113. package/src/config/schema/runtime-profile.schema.js +92 -0
  114. package/src/config/schema/shared-runtime.schema.d.ts +1 -4
  115. package/src/config/schema/shared-runtime.schema.js +0 -2
  116. package/src/config/schema/tooling.schema.d.ts +8 -5
  117. package/src/config/schema/tooling.schema.js +3 -2
  118. package/src/debug/tensor.d.ts +2 -0
  119. package/src/debug/tensor.js +21 -2
  120. package/src/gpu/command-recorder.d.ts +5 -0
  121. package/src/gpu/command-recorder.js +17 -0
  122. package/src/gpu/kernel-runtime.js +2 -10
  123. package/src/gpu/kernels/index.d.ts +0 -2
  124. package/src/gpu/kernels/index.js +0 -2
  125. package/src/gpu/kernels/kernel-prewarm.d.ts +3 -0
  126. package/src/gpu/kernels/kernel-prewarm.js +60 -0
  127. package/src/gpu/kernels/utils.d.ts +2 -4
  128. package/src/gpu/kernels/utils.js +2 -4
  129. package/src/index.d.ts +13 -1
  130. package/src/index.js +7 -1
  131. package/src/inference/browser-harness-model-helpers.js +4 -0
  132. package/src/inference/browser-harness-runtime-helpers.js +25 -10
  133. package/src/inference/browser-harness-text-helpers.js +4 -1
  134. package/src/inference/browser-harness.d.ts +4 -0
  135. package/src/inference/browser-harness.js +91 -4
  136. package/src/inference/observation-context.d.ts +22 -0
  137. package/src/inference/observation-context.js +41 -0
  138. package/src/inference/pipelines/text/attention/attn-config.js +1 -1
  139. package/src/inference/pipelines/text/attention/executor-immediate.js +42 -0
  140. package/src/inference/pipelines/text/attention/executor-recorded.js +29 -0
  141. package/src/inference/pipelines/text/attention/index.js +7 -0
  142. package/src/inference/pipelines/text/attention/interpreter.js +1334 -0
  143. package/src/inference/pipelines/text/attention/plan.d.ts +36 -0
  144. package/src/inference/pipelines/text/attention/plan.js +328 -0
  145. package/src/inference/pipelines/text/attention/receipt.js +60 -0
  146. package/src/inference/pipelines/text/attention/record.js +8 -1278
  147. package/src/inference/pipelines/text/attention/run.js +8 -1386
  148. package/src/inference/pipelines/text/generator-decode-policy.js +4 -1
  149. package/src/inference/pipelines/text/generator-helpers.js +2 -0
  150. package/src/inference/pipelines/text/generator-steps.js +32 -15
  151. package/src/inference/pipelines/text/generator.js +10 -5
  152. package/src/inference/pipelines/text/init.d.ts +3 -0
  153. package/src/inference/pipelines/text/layer-plan-gpu.js +2 -0
  154. package/src/inference/pipelines/text/layer.js +2 -0
  155. package/src/inference/pipelines/text/model-load.d.ts +0 -1
  156. package/src/inference/pipelines/text/model-load.js +2 -11
  157. package/src/inference/pipelines/text/probes.js +10 -0
  158. package/src/inference/pipelines/text/resolved-runtime-session.d.ts +56 -0
  159. package/src/inference/pipelines/text/resolved-runtime-session.js +175 -0
  160. package/src/inference/pipelines/text/state.d.ts +6 -0
  161. package/src/inference/pipelines/text/state.js +3 -0
  162. package/src/inference/pipelines/text/types.d.ts +6 -0
  163. package/src/inference/pipelines/text.js +16 -1
  164. package/src/inference/refactor-receipt.d.ts +49 -0
  165. package/src/inference/refactor-receipt.js +188 -0
  166. package/src/inference/resource-scope.d.ts +33 -0
  167. package/src/inference/resource-scope.js +168 -0
  168. package/src/inference/test-harness.d.ts +3 -0
  169. package/src/inference/test-harness.js +1 -0
  170. package/src/storage/model-cache.d.ts +53 -0
  171. package/src/storage/model-cache.js +550 -0
  172. package/src/tooling/boundary-evidence.d.ts +51 -0
  173. package/src/tooling/boundary-evidence.js +408 -0
  174. package/src/tooling/browser-command-runner.js +4 -4
  175. package/src/tooling/command-context.d.ts +25 -0
  176. package/src/tooling/command-context.js +53 -0
  177. package/src/tooling/command-runner-shared.js +22 -5
  178. package/src/tooling/demo-receipts.d.ts +56 -0
  179. package/src/tooling/demo-receipts.js +113 -0
  180. package/src/tooling/diagnose-runner.js +2 -2
  181. package/src/tooling/execution-cost-ledger.d.ts +21 -0
  182. package/src/tooling/execution-cost-ledger.js +199 -0
  183. package/src/tooling/node-command-runner.js +2 -2
  184. package/src/tooling/opfs-cache.d.ts +11 -49
  185. package/src/tooling/opfs-cache.js +4 -510
  186. package/src/tooling/precision-replay-math.d.ts +42 -0
  187. package/src/tooling/precision-replay-math.js +174 -0
  188. package/src/tooling/program-bundle-parity.d.ts +3 -0
  189. package/src/tooling/program-bundle-parity.js +8 -0
  190. package/src/tooling/program-bundle.d.ts +1 -52
  191. package/src/tooling/program-bundle.js +0 -139
  192. package/src/tooling/registered-variant-calibration.d.ts +24 -0
  193. package/src/tooling/registered-variant-calibration.js +286 -0
  194. package/src/tooling/runtime-input-composition.js +105 -53
  195. package/src/tooling/runtime-intent.js +41 -0
  196. package/src/tooling/runtime-optimization-index.d.ts +6 -0
  197. package/src/tooling/runtime-optimization-index.js +72 -0
  198. package/src/tooling/runtime-optimization.d.ts +47 -6
  199. package/src/tooling/runtime-optimization.js +559 -40
  200. package/src/tooling/source-intake.d.ts +49 -0
  201. package/src/tooling/source-intake.js +414 -0
  202. package/src/tooling-exports/evidence.d.ts +34 -0
  203. package/src/tooling-exports/evidence.js +34 -0
  204. package/src/tooling-exports/runtime.d.ts +7 -0
  205. package/src/tooling-exports/runtime.js +7 -0
  206. package/src/tooling-exports.d.ts +11 -0
  207. package/src/tooling-exports.js +6 -3
  208. package/src/tooling-exports.shared.d.ts +32 -0
  209. package/src/tooling-exports.shared.js +31 -0
  210. package/src/version.js +1 -1
  211. package/src/config/schema/tuner.schema.d.ts +0 -39
  212. package/src/config/schema/tuner.schema.js +0 -13
  213. package/src/gpu/kernel-tuner/benchmarks.js +0 -902
  214. package/src/gpu/kernel-tuner/cache.js +0 -145
  215. package/src/gpu/kernel-tuner/index.js +0 -38
  216. package/src/gpu/kernel-tuner/tuner.js +0 -247
  217. package/src/gpu/kernel-tuner.js +0 -12
  218. package/src/gpu/kernels/kernel-tuning.d.ts +0 -42
  219. package/src/gpu/kernels/kernel-tuning.js +0 -158
  220. package/src/gpu/profiler.js +0 -382
  221. package/src/inference/pipelines/text/attention/output-gate-fusion.js +0 -37
@@ -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",
@@ -3,6 +3,7 @@
3
3
  "name": "low-memory",
4
4
  "description": "Low memory configuration for constrained environments",
5
5
  "intent": "investigate",
6
+ "compatibleIntents": ["investigate"],
6
7
  "stability": "canonical",
7
8
  "owner": "doppler-core",
8
9
  "createdAtUtc": "2026-02-25T00:00:00Z",
@@ -3,15 +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"
@@ -3,6 +3,7 @@
3
3
  "name": "qwen-3-5-0-8b-throughput-probe",
4
4
  "description": "Qwen 3.5 0.8B throughput probe with a longer-prompt override (~80 tokens) for apples-to-apples 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": "qwen-3-5-2b-f16-primary-throughput",
4
4
  "description": "Qwen 3.5 2B throughput profile that explicitly requests the selective f16 primary projection lane. The compare harness still owns prompt, sampling, token budget, and decode cadence.",
5
5
  "intent": "calibrate",
6
+ "compatibleIntents": ["calibrate"],
6
7
  "stability": "experimental",
7
8
  "owner": "doppler-core",
8
9
  "createdAtUtc": "2026-07-07T00:00:00Z",
9
10
  "extends": "profiles/throughput",
10
11
  "model": "qwen-3-5-2b-q4k-ehaf16",
11
12
  "runtime": {
12
- "shared": {
13
- "tooling": {
14
- "intent": "calibrate"
15
- }
16
- },
17
13
  "inference": {
18
14
  "compute": {
19
15
  "activationDtype": "f16"
@@ -3,6 +3,7 @@
3
3
  "name": "qwen-3-5-2b-metal-parity",
4
4
  "description": "Qwen 3.5 2B browser/Metal parity profile. Preserves the release 4x4 decode cadence while inheriting the exact Metal Q4 GEMV and decode fusions from the throughput profile.",
5
5
  "intent": "calibrate",
6
+ "compatibleIntents": ["calibrate"],
6
7
  "stability": "experimental",
7
8
  "owner": "doppler-core",
8
9
  "createdAtUtc": "2026-07-10T00:00:00Z",
@@ -3,6 +3,7 @@
3
3
  "name": "qwen-3-5-2b-metal-throughput",
4
4
  "description": "Qwen 3.5 2B browser/Metal throughput profile. Uses 16-lane Q4_K decode projections, fused gate/up, and exact linear-attention fusions while preserving the manifest f32 compute lane.",
5
5
  "intent": "calibrate",
6
+ "compatibleIntents": ["calibrate"],
6
7
  "stability": "experimental",
7
8
  "owner": "doppler-core",
8
9
  "createdAtUtc": "2026-07-07T00:00:00Z",
@@ -3,6 +3,7 @@
3
3
  "name": "qwen-3-5-9b-metal-correctness",
4
4
  "description": "Shared single-token decode contract for F16 and mixed-Q4_K correctness comparison on Apple Metal.",
5
5
  "intent": "investigate",
6
+ "compatibleIntents": ["investigate"],
6
7
  "stability": "experimental",
7
8
  "owner": "doppler-core",
8
9
  "createdAtUtc": "2026-07-12T00:00:00Z",
@@ -3,15 +3,13 @@
3
3
  "name": "qwen-3-embedding-0-6b-metal-stability",
4
4
  "description": "Qwen 3 Embedding 0.6B Metal validation profile preserving manifest compute precision.",
5
5
  "intent": "verify",
6
+ "compatibleIntents": ["verify"],
6
7
  "stability": "canonical",
7
8
  "owner": "doppler-core",
8
9
  "createdAtUtc": "2026-07-09T00:00:00Z",
9
10
  "extends": "profiles/default",
10
11
  "runtime": {
11
12
  "shared": {
12
- "tooling": {
13
- "intent": "verify"
14
- },
15
13
  "debug": {
16
14
  "logLevel": {
17
15
  "defaultLogLevel": "silent"
@@ -3,6 +3,7 @@
3
3
  "name": "qwen-3-embedding-0-6b-metal-throughput",
4
4
  "description": "Qwen 3 Embedding 0.6B Metal throughput profile preserving manifest compute precision.",
5
5
  "intent": "calibrate",
6
+ "compatibleIntents": ["calibrate"],
6
7
  "stability": "canonical",
7
8
  "owner": "doppler-core",
8
9
  "createdAtUtc": "2026-07-09T00:00:00Z",
@@ -3,6 +3,7 @@
3
3
  "name": "qwen-3-reranker-0-6b-metal-stability",
4
4
  "description": "Qwen 3 Reranker 0.6B browser/Metal validation profile preserving manifest compute precision while exercising the fixed head_dim=128 f16-KV prefill attention path.",
5
5
  "intent": "verify",
6
+ "compatibleIntents": ["verify"],
6
7
  "stability": "canonical",
7
8
  "owner": "doppler-core",
8
9
  "createdAtUtc": "2026-07-09T00:00:00Z",
@@ -3,6 +3,7 @@
3
3
  "name": "qwen-3-reranker-0-6b-metal-throughput",
4
4
  "description": "Qwen 3 Reranker 0.6B browser/Metal throughput profile preserving manifest compute precision while using fixed head_dim=128 f16-KV prefill attention.",
5
5
  "intent": "calibrate",
6
+ "compatibleIntents": ["calibrate"],
6
7
  "stability": "canonical",
7
8
  "owner": "doppler-core",
8
9
  "createdAtUtc": "2026-07-09T00:00:00Z",
@@ -3,17 +3,13 @@
3
3
  "name": "qwen3-6-27b-f16-activations-probe",
4
4
  "description": "Experimental Qwen 3.6 27B profile that requests the additive execution-v1 f16 compute lane against the f16 manifest sibling sharing the existing Q4K weight pack.",
5
5
  "intent": "investigate",
6
+ "compatibleIntents": ["investigate"],
6
7
  "stability": "experimental",
7
8
  "owner": "doppler-core",
8
9
  "createdAtUtc": "2026-04-29T00:00:00Z",
9
10
  "extends": "profiles/default",
10
11
  "model": "qwen-3-6-27b-q4k-eaf16",
11
12
  "runtime": {
12
- "shared": {
13
- "tooling": {
14
- "intent": "investigate"
15
- }
16
- },
17
13
  "inference": {
18
14
  "compute": {
19
15
  "activationDtype": "f16"
@@ -3,15 +3,13 @@
3
3
  "name": "rerank-stability",
4
4
  "description": "Stable rerank validation profile with deterministic scoring and quiet tracing.",
5
5
  "intent": "verify",
6
+ "compatibleIntents": ["verify"],
6
7
  "stability": "canonical",
7
8
  "owner": "doppler-core",
8
9
  "createdAtUtc": "2026-07-06T00:00:00Z",
9
10
  "extends": "profiles/default",
10
11
  "runtime": {
11
12
  "shared": {
12
- "tooling": {
13
- "intent": "verify"
14
- },
15
13
  "debug": {
16
14
  "logLevel": {
17
15
  "defaultLogLevel": "silent"
@@ -3,15 +3,13 @@
3
3
  "name": "rerank-throughput",
4
4
  "description": "Rerank benchmark profile for repeated yes/no-logit scoring throughput runs.",
5
5
  "intent": "calibrate",
6
+ "compatibleIntents": ["calibrate"],
6
7
  "stability": "canonical",
7
8
  "owner": "doppler-core",
8
9
  "createdAtUtc": "2026-07-06T00:00:00Z",
9
10
  "extends": "profiles/rerank-stability",
10
11
  "runtime": {
11
12
  "shared": {
12
- "tooling": {
13
- "intent": "calibrate"
14
- },
15
13
  "debug": {
16
14
  "logOutput": {
17
15
  "stdout": false
@@ -3,15 +3,13 @@
3
3
  "name": "simulation",
4
4
  "description": "Simulation profile for GH200/GB200 pod emulation",
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
  "harness": {
16
14
  "mode": "simulation",
17
15
  "workload": "inference",
@@ -3,15 +3,13 @@
3
3
  "name": "throughput",
4
4
  "description": "Calibration profile with quiet logging and throughput-oriented sampling defaults.",
5
5
  "intent": "calibrate",
6
+ "compatibleIntents": ["calibrate"],
6
7
  "stability": "canonical",
7
8
  "owner": "doppler-core",
8
9
  "createdAtUtc": "2026-03-20T00:00:00Z",
9
10
  "extends": "profiles/default",
10
11
  "runtime": {
11
12
  "shared": {
12
- "tooling": {
13
- "intent": "calibrate"
14
- },
15
13
  "debug": {
16
14
  "logOutput": {
17
15
  "stdout": true,
@@ -3,15 +3,13 @@
3
3
  "name": "trace-layers",
4
4
  "description": "Trace layer outputs for F16 vs Q4K comparison",
5
5
  "intent": "investigate",
6
+ "compatibleIntents": ["investigate"],
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": "investigate"
14
- },
15
13
  "debug": {
16
14
  "trace": {
17
15
  "enabled": true,
@@ -3,6 +3,7 @@
3
3
  "name": "translategemma-1b-nativekd2-q4k-rdna3-throughput-probe",
4
4
  "description": "Experimental TranslateGemma 1B NativeKD2 EN/ES calibration profile for AMD RDNA3/Vulkan evidence. Preserves the manifest 4x1 sequential decode cadence while retaining Q4K materialization, selecting wide-tile Q4K prefill, and enabling exact-safe packed-QKV and RMSNorm-pair fusions.",
5
5
  "intent": "calibrate",
6
+ "compatibleIntents": ["calibrate"],
6
7
  "stability": "experimental",
7
8
  "owner": "doppler-core",
8
9
  "createdAtUtc": "2026-07-11T00:00:00Z",
@@ -3,6 +3,7 @@
3
3
  "name": "turboquant-contiguous-prod",
4
4
  "description": "Experimental contiguous TurboQuant prod profile with execution-v1-native session wiring and QJL residual buffers.",
5
5
  "intent": "calibrate",
6
+ "compatibleIntents": ["calibrate"],
6
7
  "stability": "experimental",
7
8
  "owner": "inference",
8
9
  "createdAtUtc": "2026-03-25T00:00:00Z",
@@ -3,6 +3,7 @@
3
3
  "name": "turboquant-contiguous",
4
4
  "description": "Experimental contiguous TurboQuant profile for full-attention decode with execution-v1-native session wiring.",
5
5
  "intent": "calibrate",
6
+ "compatibleIntents": ["calibrate"],
6
7
  "stability": "experimental",
7
8
  "owner": "inference",
8
9
  "createdAtUtc": "2026-03-25T00:00:00Z",
@@ -3,6 +3,7 @@
3
3
  "name": "turboquant",
4
4
  "description": "Experimental tiered TurboQuant profile with execution-v1-native session KV cache settings.",
5
5
  "intent": "calibrate",
6
+ "compatibleIntents": ["calibrate"],
6
7
  "stability": "experimental",
7
8
  "owner": "inference",
8
9
  "createdAtUtc": "2026-03-25T00:00:00Z",
@@ -3,15 +3,13 @@
3
3
  "name": "vector-stability",
4
4
  "description": "Numerically stable single-vector profile for embedding validation runs.",
5
5
  "intent": "investigate",
6
+ "compatibleIntents": ["investigate"],
6
7
  "stability": "canonical",
7
8
  "owner": "doppler-core",
8
9
  "createdAtUtc": "2026-03-20T00:00:00Z",
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": "debug"
@@ -3,6 +3,7 @@
3
3
  "name": "vector-throughput",
4
4
  "description": "Repeated embedding profile tuned for stable latency measurements.",
5
5
  "intent": "investigate",
6
+ "compatibleIntents": ["investigate"],
6
7
  "stability": "canonical",
7
8
  "owner": "doppler-core",
8
9
  "createdAtUtc": "2026-03-20T00:00:00Z",
@@ -3,15 +3,13 @@
3
3
  "name": "verbose-trace",
4
4
  "description": "Investigation profile with verbose logging and broad tracing enabled.",
5
5
  "intent": "investigate",
6
+ "compatibleIntents": ["investigate"],
6
7
  "stability": "canonical",
7
8
  "owner": "doppler-core",
8
9
  "createdAtUtc": "2026-03-20T00:00:00Z",
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": "verbose"
@@ -3,6 +3,7 @@
3
3
  "name": "Gemma 4 — 128 GB tier (unified memory APU)",
4
4
  "description": "Gemma 4 MoE runtime tier for 128 GB unified memory systems (e.g., Ryzen AI MAX+ 395). Generous expert cache, contiguous KV, extended context. Budget leaves headroom for system and replay-prefill intermediates.",
5
5
  "intent": "investigate",
6
+ "compatibleIntents": ["investigate"],
6
7
  "stability": "experimental",
7
8
  "owner": "doppler-core",
8
9
  "createdAtUtc": "2026-04-04T00:00:00Z",
@@ -3,6 +3,7 @@
3
3
  "name": "Gemma 4 — 16 GB tier (constrained)",
4
4
  "description": "Gemma 4 MoE runtime tier for 16 GB GPU memory. Aggressively constrained: short context, minimal expert cache, hard budget enforcement. Fail-closed if budget is not met.",
5
5
  "intent": "investigate",
6
+ "compatibleIntents": ["investigate"],
6
7
  "stability": "experimental",
7
8
  "owner": "doppler-core",
8
9
  "createdAtUtc": "2026-03-17T00:00:00Z",
@@ -3,6 +3,7 @@
3
3
  "name": "Gemma 4 — 24 GB tier",
4
4
  "description": "Gemma 4 MoE runtime tier for 24 GB GPU memory. Moderate expert cache, contiguous KV, reduced context length.",
5
5
  "intent": "investigate",
6
+ "compatibleIntents": ["investigate"],
6
7
  "stability": "experimental",
7
8
  "owner": "doppler-core",
8
9
  "createdAtUtc": "2026-03-17T00:00:00Z",
@@ -3,6 +3,7 @@
3
3
  "name": "Gemma 4 — 32 GB tier",
4
4
  "description": "Gemma 4 MoE runtime tier for 32 GB GPU memory. Generous expert cache, contiguous KV, full-length context.",
5
5
  "intent": "investigate",
6
+ "compatibleIntents": ["investigate"],
6
7
  "stability": "experimental",
7
8
  "owner": "doppler-core",
8
9
  "createdAtUtc": "2026-03-17T00:00:00Z",
@@ -18,7 +18,14 @@ export interface BenchmarkRunConfig {
18
18
  };
19
19
  debug: boolean;
20
20
  profile: boolean;
21
+ executionObserver: {
22
+ enabled: boolean;
23
+ includeDispatchGeometry: boolean;
24
+ estimateBytesMoved: boolean;
25
+ };
21
26
  useChatTemplate?: boolean;
27
+ captureMemoryTimeSeries: boolean;
28
+ memoryTimeSeriesIntervalMs: number;
22
29
  }
23
30
 
24
31
  export interface BenchmarkStatsConfig {
@@ -25,6 +25,11 @@ export const DEFAULT_BENCHMARK_RUN_CONFIG = {
25
25
  },
26
26
  debug: false,
27
27
  profile: false,
28
+ executionObserver: {
29
+ enabled: false,
30
+ includeDispatchGeometry: true,
31
+ estimateBytesMoved: false,
32
+ },
28
33
  useChatTemplate: undefined,
29
34
  // Memory time series capture (VRAM over time)
30
35
  captureMemoryTimeSeries: false,
@@ -449,17 +449,6 @@ export {
449
449
  DEFAULT_GPU_CACHE_CONFIG,
450
450
  } from './gpu-cache.schema.js';
451
451
 
452
- // =============================================================================
453
- // Tuner Schema
454
- // =============================================================================
455
- export {
456
- // Types
457
- type TunerConfigSchema,
458
-
459
- // Defaults
460
- DEFAULT_TUNER_CONFIG,
461
- } from './tuner.schema.js';
462
-
463
452
  // =============================================================================
464
453
  // Debug Schema
465
454
  // =============================================================================
@@ -768,14 +757,22 @@ export {
768
757
  type ToolingConfigSchema,
769
758
  type ToolingIntent,
770
759
  type ToolingDiagnosticsMode,
760
+ type RefactorReceiptPolicy,
771
761
 
772
762
  // Constants
773
763
  TOOLING_INTENTS,
774
764
  TOOLING_DIAGNOSTICS,
765
+ REFACTOR_RECEIPT_POLICIES,
775
766
 
776
767
  // Defaults
777
768
  DEFAULT_TOOLING_CONFIG,
778
769
  } from './tooling.schema.js';
770
+ export {
771
+ type RuntimeProfileMetadata,
772
+ type RuntimeProfileStability,
773
+ validateRuntimeProfileMetadata,
774
+ assertRuntimeProfileIntentCompatibility,
775
+ } from './runtime-profile.schema.js';
779
776
 
780
777
  // =============================================================================
781
778
  // Ecosystem Schema
@@ -901,11 +898,6 @@ export {
901
898
  DEFAULT_ADAPTER_CONFIG,
902
899
  } from './adapter.schema.js';
903
900
 
904
- // =============================================================================
905
- // Tuner Limits
906
- // =============================================================================
907
- export { DEFAULT_TUNER_LIMITS } from './kernel-thresholds.schema.js';
908
-
909
901
  // =============================================================================
910
902
  // Self-Speculation Schema
911
903
  // =============================================================================
@@ -208,13 +208,6 @@ export {
208
208
  DEFAULT_GPU_CACHE_CONFIG,
209
209
  } from './gpu-cache.schema.js';
210
210
 
211
- // =============================================================================
212
- // Tuner Schema
213
- // =============================================================================
214
- export {
215
- DEFAULT_TUNER_CONFIG,
216
- } from './tuner.schema.js';
217
-
218
211
  // =============================================================================
219
212
  // Debug Schema
220
213
  // =============================================================================
@@ -251,8 +244,13 @@ export {
251
244
  export {
252
245
  TOOLING_INTENTS,
253
246
  TOOLING_DIAGNOSTICS,
247
+ REFACTOR_RECEIPT_POLICIES,
254
248
  DEFAULT_TOOLING_CONFIG,
255
249
  } from './tooling.schema.js';
250
+ export {
251
+ validateRuntimeProfileMetadata,
252
+ assertRuntimeProfileIntentCompatibility,
253
+ } from './runtime-profile.schema.js';
256
254
 
257
255
  // =============================================================================
258
256
  // Ecosystem Schema
@@ -439,7 +437,6 @@ export {
439
437
  DEFAULT_ROPE_DEFAULTS,
440
438
  DEFAULT_ATTENTION_THRESHOLDS,
441
439
  DEFAULT_CAST_THRESHOLDS,
442
- DEFAULT_TUNER_LIMITS,
443
440
  DEFAULT_KERNEL_THRESHOLDS,
444
441
 
445
442
  // Functions
@@ -117,17 +117,6 @@ export interface RopeDefaultsSchema {
117
117
 
118
118
  export declare const DEFAULT_ROPE_DEFAULTS: RopeDefaultsSchema;
119
119
 
120
- /**
121
- * Defaults for kernel tuner limits when device limits are unavailable.
122
- */
123
- export interface TunerLimitsSchema {
124
- maxComputeWorkgroupSizeX: number;
125
- maxComputeWorkgroupSizeY: number;
126
- maxComputeInvocationsPerWorkgroup: number;
127
- }
128
-
129
- export declare const DEFAULT_TUNER_LIMITS: TunerLimitsSchema;
130
-
131
120
  /**
132
121
  * Thresholds for attention kernel variant selection.
133
122
  */
@@ -279,7 +268,6 @@ export interface KernelThresholdsConfigSchema {
279
268
  attention: AttentionThresholdsSchema;
280
269
  fusedMatmul: FusedMatmulThresholdsSchema;
281
270
  cast: CastThresholdsSchema;
282
- tuner: TunerLimitsSchema;
283
271
  }
284
272
 
285
273
  export declare const DEFAULT_KERNEL_THRESHOLDS: KernelThresholdsConfigSchema;