doppler-gpu 0.4.15 → 0.6.0

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 (1063) hide show
  1. package/CHANGELOG.md +33 -577
  2. package/LICENSE +21 -201
  3. package/README.md +165 -213
  4. package/models/catalog.json +241 -33
  5. package/package.json +122 -86
  6. package/src/capsule-runtime.d.ts +23 -0
  7. package/src/capsule-runtime.js +23 -0
  8. package/src/capsule.d.ts +8 -0
  9. package/src/capsule.js +4 -0
  10. package/src/cli/argument-contract.js +200 -0
  11. package/src/cli/browser-report-output.js +24 -0
  12. package/src/cli/cli-model-resolution.js +2 -2
  13. package/src/cli/cli-output.js +10 -0
  14. package/src/cli/commands/bundle.js +828 -0
  15. package/src/cli/doppler-cli.js +101 -1349
  16. package/src/cli/doppler-serve.js +18 -244
  17. package/src/cli/output/manifest.js +368 -0
  18. package/src/cli/runtime-profiles.js +4 -8
  19. package/src/cli/serve/capsule-handler.d.ts +33 -0
  20. package/src/cli/serve/capsule-handler.js +105 -0
  21. package/src/cli/serve/capsule-http.js +102 -0
  22. package/src/cli/serve/receipt.js +269 -0
  23. package/src/client/capsule-host.browser.js +1 -0
  24. package/src/client/capsule-host.d.ts +3 -0
  25. package/src/client/capsule-host.js +1 -0
  26. package/src/client/doppler-api.browser.js +42 -0
  27. package/src/client/doppler-api.d.ts +64 -1
  28. package/src/client/doppler-api.js +42 -1
  29. package/src/client/doppler-registry.js +15 -7
  30. package/src/client/electron/index.d.ts +4 -0
  31. package/src/client/electron/index.js +13 -0
  32. package/src/client/electron/ipc-contract.d.ts +9 -0
  33. package/src/client/electron/ipc-contract.js +56 -0
  34. package/src/client/electron/release-state.d.ts +53 -0
  35. package/src/client/electron/release-state.js +425 -0
  36. package/src/client/electron/renderer-runtime.d.ts +19 -0
  37. package/src/client/electron/renderer-runtime.js +104 -0
  38. package/src/client/failure-taxonomy.js +1 -28
  39. package/src/client/fault-injection.js +0 -8
  40. package/src/client/inspection.d.ts +122 -0
  41. package/src/client/inspection.js +343 -0
  42. package/src/client/model-host/index.d.ts +140 -0
  43. package/src/client/model-host/index.js +548 -0
  44. package/src/client/model-host/model-evidence.js +255 -0
  45. package/src/client/model-host/model-session.d.ts +322 -0
  46. package/src/client/model-host/model-session.js +593 -0
  47. package/src/client/provider.d.ts +18 -2
  48. package/src/client/provider.js +26 -3
  49. package/src/client/receipt.js +30 -72
  50. package/src/client/runtime/capsule-acquisition.d.ts +18 -0
  51. package/src/client/runtime/capsule-acquisition.js +103 -0
  52. package/src/client/runtime/capsule-adapter-execution.d.ts +17 -0
  53. package/src/client/runtime/capsule-adapter-execution.js +63 -0
  54. package/src/client/runtime/capsule-artifact-source.js +62 -0
  55. package/src/client/runtime/capsule-embedding.d.ts +29 -0
  56. package/src/client/runtime/capsule-embedding.js +76 -0
  57. package/src/client/runtime/capsule-forecast-program.d.ts +2 -0
  58. package/src/client/runtime/capsule-forecast-program.js +85 -0
  59. package/src/client/runtime/capsule-forecast.d.ts +17 -0
  60. package/src/client/runtime/capsule-forecast.js +31 -0
  61. package/src/client/runtime/capsule-operation-adapters.d.ts +12 -0
  62. package/src/client/runtime/capsule-operation-adapters.js +68 -0
  63. package/src/client/runtime/capsule-operation-executor.d.ts +25 -0
  64. package/src/client/runtime/capsule-operation-executor.js +100 -0
  65. package/src/client/runtime/capsule-program-adapter.js +151 -0
  66. package/src/client/runtime/capsule-release-authorization.js +25 -0
  67. package/src/client/runtime/capsule-rerank.d.ts +50 -0
  68. package/src/client/runtime/capsule-rerank.js +105 -0
  69. package/src/client/runtime/capsule-session-execution.js +85 -0
  70. package/src/client/runtime/command-executor.d.ts +15 -0
  71. package/src/client/runtime/command-executor.js +128 -0
  72. package/src/client/runtime/composition-root.d.ts +74 -0
  73. package/src/client/runtime/composition-root.js +344 -0
  74. package/src/client/runtime/fetch-capsule-artifact-store.d.ts +10 -0
  75. package/src/client/runtime/fetch-capsule-artifact-store.js +25 -0
  76. package/src/client/runtime/index.d.ts +1 -63
  77. package/src/client/runtime/index.js +1 -261
  78. package/src/client/runtime/lora.d.ts +14 -0
  79. package/src/client/runtime/lora.js +19 -1
  80. package/src/client/runtime/model-manager.d.ts +0 -4
  81. package/src/client/runtime/model-manager.js +12 -28
  82. package/src/client/runtime/model-rerank.js +30 -0
  83. package/src/client/runtime/model-session.d.ts +1 -174
  84. package/src/client/runtime/model-source.d.ts +19 -1
  85. package/src/client/runtime/model-source.js +46 -17
  86. package/src/client/runtime/node-quickstart-cache.js +1 -1
  87. package/src/client/runtime/resolution-policy.d.ts +20 -0
  88. package/src/client/runtime/resolution-policy.js +84 -0
  89. package/src/client/runtime/resource-binder.d.ts +24 -0
  90. package/src/client/runtime/resource-binder.js +135 -0
  91. package/src/client/runtime/scoped-session.d.ts +126 -0
  92. package/src/client/runtime/scoped-session.js +276 -0
  93. package/src/client/runtime/session-controller.d.ts +30 -0
  94. package/src/client/runtime/session-controller.js +170 -0
  95. package/src/client/runtime/source-runtime.js +1 -1
  96. package/src/client/runtime/target-selector.d.ts +9 -0
  97. package/src/client/runtime/target-selector.js +2 -0
  98. package/src/client/runtime/types.d.ts +1 -1
  99. package/src/client/runtime/types.js +0 -1
  100. package/src/client/runtime/verified-capsule-artifact-store.js +72 -0
  101. package/src/client/wrap-pipeline-handle.js +3 -9
  102. package/src/config/backward-registry-loader.js +1 -1
  103. package/src/config/capsule-adapter-policy.js +30 -0
  104. package/src/config/capsule-adapters.d.ts +21 -0
  105. package/src/config/capsule-adapters.js +40 -0
  106. package/src/config/capsule-adapters.json +9 -0
  107. package/src/config/capsule-loading.d.ts +2 -0
  108. package/src/config/capsule-loading.js +14 -0
  109. package/src/config/capsule-loading.schema.json +11 -0
  110. package/src/config/capsule-operation.d.ts +19 -0
  111. package/src/config/capsule-operation.js +58 -0
  112. package/src/config/capsule-operations.json +18 -0
  113. package/src/config/capsule-release-contract.d.ts +66 -0
  114. package/src/config/capsule-release-contract.js +197 -0
  115. package/src/config/capsule-release-events.d.ts +53 -0
  116. package/src/config/capsule-release-events.js +172 -0
  117. package/src/config/capsule-sequence-receipt.js +19 -0
  118. package/src/config/capsule-serve.d.ts +9 -0
  119. package/src/config/capsule-serve.js +28 -0
  120. package/src/config/capsule-serve.schema.json +14 -0
  121. package/src/config/capsule-signature.d.ts +6 -0
  122. package/src/config/capsule-signature.js +50 -0
  123. package/src/config/capsule-v2.d.ts +70 -0
  124. package/src/config/capsule-v2.js +431 -0
  125. package/src/config/capsule-v3.d.ts +19 -0
  126. package/src/config/capsule-v3.js +72 -0
  127. package/src/config/capsule.d.ts +22 -0
  128. package/src/config/capsule.js +52 -0
  129. package/src/config/conversion/amplify/amplify-120m-f16-af32.json +1 -1
  130. package/src/config/conversion/embeddinggemma/google-embeddinggemma-300m-q4k-ehf16-af32.json +1 -1
  131. package/src/config/conversion/esm/esm2-t12-35m-ur50d-f32-af32.json +41 -1
  132. package/src/config/conversion/esmc/esmc-300m-f32-af32.json +1 -1
  133. package/src/config/conversion/gemma3/gemma-3-1b-it-q4k-ehf16-af32.json +1 -1
  134. package/src/config/conversion/gemma3/gemma-3-270m-it-f16-af32.json +1 -1
  135. package/src/config/conversion/gemma3/gemma-3-270m-it-q4k-ehf16-af16.json +1 -1
  136. package/src/config/conversion/gemma3/gemma-3-270m-it-q4k-ehf16-af32.json +1 -1
  137. package/src/config/conversion/gemma3/translategemma-4b-1b-enes-q4k-ehf16-af32.json +1 -1
  138. package/src/config/conversion/gemma3/translategemma-4b-it-q4k-ehf16-af32.json +1 -1
  139. package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-af16.json +2 -2
  140. package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-af32.json +1 -1
  141. package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-hq4k-af16.json +1 -1
  142. package/src/config/conversion/gemma4/gemma-4-12b-it-text-w4a16-ct-ehf16-af16.json +1 -1
  143. package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af16.json +2 -2
  144. package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af32.json +1 -1
  145. package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af16-int4ple.json +3 -3
  146. package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32-int4ple.json +1 -1
  147. package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32.json +1 -1
  148. package/src/config/conversion/glmocr/glm-ocr-f16-af32.json +442 -0
  149. package/src/config/conversion/minicpm/minicpm4-0-5b-f16-af32.json +1 -1
  150. package/src/config/conversion/muse-glimmer/muse-glimmer-30b-text-bf16-af32.json +529 -0
  151. package/src/config/conversion/muse-glimmer/muse-glimmer-30b-text-f16-af32.json +529 -0
  152. package/src/config/conversion/nucleotide-transformer/nucleotide-transformer-v2-50m-f32-af32.json +1 -1
  153. package/src/config/conversion/qwen3/qwen-3-5-0-8b-q4k-ehaf16.json +1 -1
  154. package/src/config/conversion/qwen3/qwen-3-5-2b-q4k-ehaf16.json +1 -1
  155. package/src/config/conversion/qwen3/qwen-3-5-9b-f16-af32.json +1 -1
  156. package/src/config/conversion/qwen3/qwen-3-5-9b-q4k-ehaf16.json +1 -1
  157. package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-eaf16.json +2 -2
  158. package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-ehaf16.json +1 -1
  159. package/src/config/conversion/qwen3/qwen-3-embedding-0-6b-f16-af32.json +463 -0
  160. package/src/config/conversion/qwen3/qwen-3-embedding-0-6b-q4k-ehf16-af32.json +1 -1
  161. package/src/config/conversion/qwen3/qwen-3-reranker-0-6b-f16-af32.json +1 -1
  162. package/src/config/conversion/qwen3/qwen-3-reranker-0-6b-f16-true-logit-af32.json +481 -0
  163. package/src/config/conversion/qwen3/qwen-3-reranker-0-6b-q4k-ehf16-af32.json +59 -7
  164. package/src/config/conversion/qwen3/qwen3.8-27b-text-f16-af32.json +561 -0
  165. package/src/config/conversion/qwen3/qwen3.8-27b-text-q4k-ehaf16.json +558 -0
  166. package/src/config/dtype-transition-contract.js +40 -0
  167. package/src/config/embedding-contract.js +26 -0
  168. package/src/config/embedding-reference.js +85 -0
  169. package/src/config/evidence/boundary-evidence-policy.json +54 -0
  170. package/src/config/evidence/source-intake-policy.json +122 -0
  171. package/src/config/execution-contract-check.d.ts +16 -73
  172. package/src/config/execution-contract-check.js +7 -429
  173. package/src/config/forecast-manifest.js +56 -0
  174. package/src/config/heterogeneous-attention-contract.js +46 -0
  175. package/src/config/initial-execution-identity.d.ts +58 -0
  176. package/src/config/initial-execution-identity.js +331 -0
  177. package/src/config/inspection/observation-policies.json +80 -0
  178. package/src/config/kernel-path-loader.d.ts +10 -0
  179. package/src/config/kernel-path-loader.js +23 -8
  180. package/src/config/kernel-registry-contract.d.ts +25 -0
  181. package/src/config/kernel-registry-contract.js +46 -0
  182. package/src/config/kernels/kernel-ref-digests.js +50 -2
  183. package/src/config/kernels/registry.js +1 -1
  184. package/src/config/kernels/registry.json +13592 -11377
  185. package/src/config/lora-layouts.d.ts +10 -0
  186. package/src/config/lora-layouts.js +17 -0
  187. package/src/config/lora-layouts.json +9 -0
  188. package/src/config/manifest-inference-validation.js +575 -0
  189. package/src/config/merge/execution-patches.js +12 -0
  190. package/src/config/{merge-helpers.js → merge/kernel-path-policy.js} +0 -52
  191. package/src/config/merge/value-selection.js +38 -0
  192. package/src/config/merge.js +43 -11
  193. package/src/config/model-ir-v2.d.ts +57 -0
  194. package/src/config/model-ir-v2.js +258 -0
  195. package/src/config/model-ir.d.ts +49 -0
  196. package/src/config/model-ir.js +180 -0
  197. package/src/config/param-validator.js +22 -37
  198. package/src/config/platforms/amd-strix-halo-gfx1151.json +17 -0
  199. package/src/config/platforms/loader.js +2 -1
  200. package/src/config/production-release-evidence.d.ts +23 -0
  201. package/src/config/production-release-evidence.js +423 -0
  202. package/src/config/production-release.js +321 -0
  203. package/src/{client/doppler-registry.json → config/quickstart-registry.json} +7 -7
  204. package/src/config/required-inference-fields-contract-check.js +2 -2
  205. package/src/config/rerank-reference.d.ts +28 -0
  206. package/src/config/rerank-reference.js +89 -0
  207. package/src/config/revocation-policy.d.ts +65 -0
  208. package/src/config/revocation-policy.js +290 -0
  209. package/src/config/revocation-registry.json +11 -0
  210. package/src/config/revocation-updates.d.ts +68 -0
  211. package/src/config/revocation-updates.js +440 -0
  212. package/src/config/runtime/compute/f16-activations.json +1 -0
  213. package/src/config/runtime/compute/f16-batched.json +1 -0
  214. package/src/config/runtime/default.json +1 -0
  215. package/src/config/runtime/diagnostics/debug-logits.json +1 -0
  216. package/src/config/runtime/experiments/bench/gemma3-bench-q4k-alt.json +1 -0
  217. package/src/config/runtime/experiments/bench/gemma3-bench-q4k-batched.json +1 -0
  218. package/src/config/runtime/experiments/bench/gemma3-bench-q4k.json +1 -0
  219. package/src/config/runtime/experiments/bench/gemma4-bench-q4k.json +1 -0
  220. package/src/config/runtime/experiments/bench/qwen3-bench-q4k-alt.json +1 -0
  221. package/src/config/runtime/experiments/bench/qwen3-bench-q4k-batched.json +1 -0
  222. package/src/config/runtime/experiments/debug/gemma3-browser-diagnose-q4k.json +1 -1
  223. package/src/config/runtime/experiments/debug/gemma3-debug-q4k.json +1 -3
  224. package/src/config/runtime/experiments/debug/gemma4-e2b-browser-diagnose-q4k.json +1 -1
  225. package/src/config/runtime/experiments/debug/qwen35-browser-diagnose-q4k.json +1 -1
  226. package/src/config/runtime/experiments/verify/gemma3-verify.json +1 -3
  227. package/src/config/runtime/experiments/verify/gemma4-e2b-verify.json +1 -3
  228. package/src/config/runtime/experiments/verify/lfm2-verify.json +1 -3
  229. package/src/config/runtime/experiments/verify/translategemma-verify.json +1 -3
  230. package/src/config/runtime/model/gemma2-debug.json +1 -0
  231. package/src/config/runtime/model/gemma2-pipeline-debug.json +1 -0
  232. package/src/config/runtime/model/gemma2-pipeline.json +1 -0
  233. package/src/config/runtime/model/gemma3-layer-probe.json +1 -0
  234. package/src/config/runtime/model/qwen3-5-layer-probe.json +1 -0
  235. package/src/config/runtime/model/qwen3-5-layer3-boundary-probe.json +1 -0
  236. package/src/config/runtime/model/qwen3-5-linear-attn-debug.json +1 -0
  237. package/src/config/runtime/optimization-candidates/registry.json +5 -0
  238. package/src/config/runtime/platform/metal-apple-q4k.json +1 -0
  239. package/src/config/runtime/profiles/default.json +1 -0
  240. package/src/config/runtime/profiles/diffusiongemma-26b-a4b-throughput.json +1 -0
  241. package/src/config/runtime/profiles/gemma3-270m-q4k-throughput-overlapped-probe.json +1 -0
  242. package/src/config/runtime/profiles/gemma3-270m-q4k-throughput-probe.json +1 -0
  243. package/src/config/runtime/profiles/gemma4-12b-rdrr-safe-b4.json +1 -3
  244. package/src/config/runtime/profiles/gemma4-12b-rdrr-safe-single.json +1 -3
  245. package/src/config/runtime/profiles/gemma4-31b-f16-activations-probe.json +2 -6
  246. package/src/config/runtime/profiles/gemma4-e2b-baseline-probe.json +1 -3
  247. package/src/config/runtime/profiles/gemma4-e2b-fused-rmsnorm-probe.json +1 -3
  248. package/src/config/runtime/profiles/gemma4-e2b-fused-rmsnorm-test.json +1 -0
  249. package/src/config/runtime/profiles/gemma4-e2b-int4ple-baseline-probe.json +1 -5
  250. package/src/config/runtime/profiles/gemma4-e2b-int4ple-f16a-probe.json +1 -5
  251. package/src/config/runtime/profiles/gemma4-e2b-int4ple-reference-export.json +1 -3
  252. package/src/config/runtime/profiles/gemma4-e2b-int4ple-retain-q4k-probe.json +1 -5
  253. package/src/config/runtime/profiles/gemma4-e2b-low-memory.json +1 -0
  254. package/src/config/runtime/profiles/gemma4-e2b-prefill-profile.json +1 -3
  255. package/src/config/runtime/profiles/gemma4-e2b-throughput-probe.json +1 -0
  256. package/src/config/runtime/profiles/low-memory.json +1 -0
  257. package/src/config/runtime/profiles/muse-glimmer-30b-source-parity-f32kv.json +30 -0
  258. package/src/config/runtime/profiles/muse-glimmer-30b-source-parity.json +110 -0
  259. package/src/config/runtime/profiles/production.json +1 -3
  260. package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-attn128-candidate.json +1 -0
  261. package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-maxbatch64-candidate.json +1 -0
  262. package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-normed-ffn-candidate.json +1 -0
  263. package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-overlap-ringtokens1-candidate.json +1 -0
  264. package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-parity.json +1 -0
  265. package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-sequential-candidate.json +1 -0
  266. package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-throughput.json +1 -0
  267. package/src/config/runtime/profiles/qwen-3-5-0-8b-prefill-profile.json +1 -3
  268. package/src/config/runtime/profiles/qwen-3-5-0-8b-throughput-probe.json +1 -0
  269. package/src/config/runtime/profiles/qwen-3-5-2b-f16-primary-throughput.json +1 -5
  270. package/src/config/runtime/profiles/qwen-3-5-2b-metal-parity.json +1 -0
  271. package/src/config/runtime/profiles/qwen-3-5-2b-metal-throughput.json +1 -0
  272. package/src/config/runtime/profiles/qwen-3-5-9b-metal-correctness.json +1 -0
  273. package/src/config/runtime/profiles/qwen-3-embedding-0-6b-metal-stability.json +1 -3
  274. package/src/config/runtime/profiles/qwen-3-embedding-0-6b-metal-throughput.json +1 -0
  275. package/src/config/runtime/profiles/qwen-3-reranker-0-6b-metal-stability.json +1 -0
  276. package/src/config/runtime/profiles/qwen-3-reranker-0-6b-metal-throughput.json +1 -0
  277. package/src/config/runtime/profiles/qwen3-6-27b-f16-activations-probe.json +2 -6
  278. package/src/config/runtime/profiles/qwen3-8-27b-source-parity.json +87 -0
  279. package/src/config/runtime/profiles/rerank-stability.json +1 -3
  280. package/src/config/runtime/profiles/rerank-throughput.json +1 -3
  281. package/src/config/runtime/profiles/simulation.json +1 -3
  282. package/src/config/runtime/profiles/throughput.json +1 -3
  283. package/src/config/runtime/profiles/trace-layers.json +1 -3
  284. package/src/config/runtime/profiles/translategemma-1b-nativekd2-q4k-rdna3-throughput-probe.json +1 -0
  285. package/src/config/runtime/profiles/turboquant-contiguous-prod.json +1 -0
  286. package/src/config/runtime/profiles/turboquant-contiguous.json +1 -0
  287. package/src/config/runtime/profiles/turboquant.json +1 -0
  288. package/src/config/runtime/profiles/vector-stability.json +1 -3
  289. package/src/config/runtime/profiles/vector-throughput.json +1 -0
  290. package/src/config/runtime/profiles/verbose-trace.json +1 -3
  291. package/src/config/runtime/tiers/gemma4-128gb.json +1 -0
  292. package/src/config/runtime/tiers/gemma4-16gb.json +1 -0
  293. package/src/config/runtime/tiers/gemma4-24gb.json +1 -0
  294. package/src/config/runtime/tiers/gemma4-32gb.json +1 -0
  295. package/src/config/runtime-merge.js +1 -1
  296. package/src/config/runtime.js +1 -1
  297. package/src/config/schema/benchmark.schema.d.ts +7 -0
  298. package/src/config/schema/benchmark.schema.js +5 -0
  299. package/src/config/schema/browser-suite-metrics.schema.d.ts +1 -0
  300. package/src/config/schema/browser-suite-metrics.schema.js +2 -0
  301. package/src/config/schema/conversion-report.schema.d.ts +3 -0
  302. package/src/config/schema/conversion-report.schema.js +29 -0
  303. package/src/config/schema/converter.schema.d.ts +1 -0
  304. package/src/config/schema/debug.schema.d.ts +1 -0
  305. package/src/config/schema/doppler.schema.js +5 -5
  306. package/src/config/schema/emulation.schema.js +1 -1
  307. package/src/config/schema/execution-v1.schema.d.ts +3 -0
  308. package/src/config/schema/execution-v1.schema.js +19 -0
  309. package/src/config/schema/harness.schema.d.ts +13 -0
  310. package/src/config/schema/harness.schema.js +6 -0
  311. package/src/config/schema/index.d.ts +8 -16
  312. package/src/config/schema/index.js +5 -8
  313. package/src/config/schema/inference.schema.d.ts +21 -2
  314. package/src/config/schema/kernel-thresholds.schema.d.ts +0 -12
  315. package/src/config/schema/kernel-thresholds.schema.js +1 -13
  316. package/src/config/schema/kernel-warmup.schema.d.ts +1 -3
  317. package/src/config/schema/kernel-warmup.schema.js +0 -1
  318. package/src/config/schema/manifest.schema.d.ts +26 -7
  319. package/src/config/schema/manifest.schema.js +10 -1
  320. package/src/config/schema/program-bundle.schema.d.ts +52 -6
  321. package/src/config/schema/program-bundle.schema.js +145 -10
  322. package/src/config/schema/runtime-profile.schema.d.ts +24 -0
  323. package/src/config/schema/runtime-profile.schema.js +92 -0
  324. package/src/config/schema/shared-runtime.schema.d.ts +1 -4
  325. package/src/config/schema/shared-runtime.schema.js +0 -2
  326. package/src/config/schema/tooling.schema.d.ts +8 -5
  327. package/src/config/schema/tooling.schema.js +3 -2
  328. package/src/config/sequence-reference.d.ts +26 -0
  329. package/src/config/sequence-reference.js +100 -0
  330. package/src/config/target-plan.d.ts +108 -0
  331. package/src/config/target-plan.js +411 -0
  332. package/src/config/training-defaults.js +1 -1
  333. package/src/config/transforms/capability-transform-resolver.js +4 -105
  334. package/src/config/transforms/execution-graph/dtype.js +591 -0
  335. package/src/config/transforms/execution-graph/fusion.js +436 -0
  336. package/src/config/transforms/execution-graph/validation.js +414 -0
  337. package/src/config/transforms/execution-graph-transforms.js +13 -1419
  338. package/src/converter/artifact-identity.js +475 -0
  339. package/src/converter/conversion-plan.js +1 -1
  340. package/src/converter/core.js +94 -2096
  341. package/src/converter/forge-candidate-evaluation.js +271 -0
  342. package/src/converter/forge-qualification.js +162 -0
  343. package/src/converter/forge-stages.js +798 -0
  344. package/src/converter/manifest-builder.js +683 -0
  345. package/src/converter/quantization-info.js +4 -1
  346. package/src/converter/quantizer.js +58 -9
  347. package/src/converter/shard-packer.d.ts +1 -1
  348. package/src/converter/shard-packer.js +2 -2
  349. package/src/converter/tensor-transform.js +640 -0
  350. package/src/converter/tokenizer-contract.js +243 -0
  351. package/src/converter/tokenizer-manifest.js +243 -0
  352. package/src/debug/history.js +0 -8
  353. package/src/debug/index.js +0 -10
  354. package/src/debug/tensor.d.ts +2 -0
  355. package/src/debug/tensor.js +21 -2
  356. package/src/debug/trace.js +0 -7
  357. package/src/experimental/adapters/adapter-manager.js +2 -0
  358. package/src/experimental/adapters/adapter-manifest.d.ts +6 -0
  359. package/src/experimental/adapters/adapter-manifest.js +11 -0
  360. package/src/experimental/adapters/artifact-contract.d.ts +2 -2
  361. package/src/experimental/adapters/artifact-contract.js +5 -2
  362. package/src/experimental/adapters/lora-loader.d.ts +10 -1
  363. package/src/experimental/adapters/lora-loader.js +110 -48
  364. package/src/{inference/pipelines → experimental}/audio/gemma4.js +9 -31
  365. package/src/experimental/bdpa/basis-decomposed-paged.d.ts +50 -0
  366. package/src/experimental/bdpa/basis-decomposed-paged.js +276 -0
  367. package/src/experimental/bridge/trainer-artifact-bridge.js +3 -3
  368. package/src/experimental/browser/browser-converter.js +11 -11
  369. package/src/experimental/diffusion/index.d.ts +1 -1
  370. package/src/experimental/diffusion/index.js +1 -1
  371. package/src/{inference/pipelines → experimental}/diffusion/pipeline.d.ts +2 -2
  372. package/src/{inference/pipelines → experimental}/diffusion/pipeline.js +17 -17
  373. package/src/{inference/pipelines → experimental}/diffusion/text-encoder-gpu.js +21 -19
  374. package/src/experimental/diffusion/vae.js +596 -0
  375. package/src/experimental/distribution/shard-delivery/receipt.js +588 -0
  376. package/src/experimental/distribution/shard-delivery/retry.js +392 -0
  377. package/src/experimental/distribution/shard-delivery/transport.js +831 -0
  378. package/src/experimental/distribution/shard-delivery.js +4 -1709
  379. package/src/experimental/energy/index.d.ts +2 -2
  380. package/src/experimental/energy/index.js +2 -2
  381. package/src/{inference/pipelines → experimental}/energy/pipeline.js +13 -13
  382. package/src/{inference/pipelines → experimental}/energy/types.d.ts +2 -2
  383. package/src/experimental/logits/cpu-output.d.ts +27 -0
  384. package/src/{inference/pipelines/text/logits/utils.js → experimental/logits/cpu-output.js} +19 -27
  385. package/src/experimental/logits/cpu.d.ts +87 -0
  386. package/src/experimental/logits/cpu.js +95 -0
  387. package/src/experimental/orchestration/energy-row-head-pipeline.d.ts +103 -0
  388. package/src/{inference/pipelines/energy-head/row-head-pipeline.js → experimental/orchestration/energy-row-head-pipeline.js} +13 -13
  389. package/src/experimental/orchestration/index.d.ts +1 -1
  390. package/src/experimental/orchestration/index.js +1 -1
  391. package/src/experimental/training/artifacts.js +2 -2
  392. package/src/experimental/training/checkpoint.js +2 -2
  393. package/src/experimental/training/distillation/artifacts.js +1 -1
  394. package/src/experimental/training/distillation/dataset.js +2 -2
  395. package/src/experimental/training/distillation/eval.js +1 -1
  396. package/src/experimental/training/distillation/runtime.js +2 -2
  397. package/src/experimental/training/distillation/student-fixture/execution.js +166 -0
  398. package/src/experimental/training/distillation/student-fixture/model.js +353 -0
  399. package/src/experimental/training/distillation/student-fixture.js +2 -469
  400. package/src/experimental/training/lora/checkpoint.js +746 -0
  401. package/src/experimental/training/lora/execution.js +550 -0
  402. package/src/experimental/training/lora/plan.js +387 -0
  403. package/src/experimental/training/lora/recovery.js +418 -0
  404. package/src/experimental/training/lora-pipeline.js +7 -1949
  405. package/src/experimental/training/operator-artifacts.js +2 -2
  406. package/src/experimental/training/operator-command.js +4 -4
  407. package/src/experimental/training/runner/evaluation.js +62 -0
  408. package/src/experimental/training/runner/lifecycle.js +257 -0
  409. package/src/experimental/training/runner.js +5 -268
  410. package/src/experimental/training/suite/execution.js +879 -0
  411. package/src/experimental/training/suite/plan.js +855 -0
  412. package/src/experimental/training/suite.js +5 -1659
  413. package/src/experimental/training/workloads.d.ts +5 -5
  414. package/src/experimental/training/workloads.js +6 -6
  415. package/src/experimental/turboquant/codebook.js +470 -0
  416. package/src/formats/quantization-constants.d.ts +11 -0
  417. package/src/formats/quantization-constants.js +11 -0
  418. package/src/formats/rdrr/integrity.js +1 -1
  419. package/src/formats/rdrr/lowerings.js +0 -29
  420. package/src/formats/rdrr/merkle.js +1 -1
  421. package/src/formats/rdrr/tensor-config-validator.js +25 -5
  422. package/src/formats/rdrr/validation.js +3 -3
  423. package/src/gpu/command-recorder.d.ts +15 -1
  424. package/src/gpu/command-recorder.js +63 -2
  425. package/src/gpu/constant-buffer.js +29 -0
  426. package/src/gpu/device-state.js +36 -0
  427. package/src/gpu/device.js +6 -41
  428. package/src/gpu/kernel-runtime.js +3 -10
  429. package/src/gpu/kernels/activation-static-qdq.js +1 -1
  430. package/src/gpu/kernels/append_embedding.wgsl +14 -0
  431. package/src/gpu/kernels/attention/executors.js +960 -0
  432. package/src/gpu/kernels/attention/plan.js +472 -0
  433. package/src/gpu/kernels/attention.d.ts +4 -0
  434. package/src/gpu/kernels/attention.js +145 -1545
  435. package/src/gpu/kernels/attention_relative.wgsl +59 -0
  436. package/src/gpu/kernels/backward/adam.js +2 -1
  437. package/src/gpu/kernels/backward/attention_backward.js +1 -1
  438. package/src/gpu/kernels/backward/bias_add_backward.d.ts +1 -1
  439. package/src/gpu/kernels/backward/bias_add_backward.js +1 -1
  440. package/src/gpu/kernels/backward/causal_conv1d_silu_backward.js +2 -1
  441. package/src/gpu/kernels/backward/conv2d_backward.d.ts +1 -1
  442. package/src/gpu/kernels/backward/conv2d_backward.js +2 -1
  443. package/src/gpu/kernels/backward/cross_entropy_backward.d.ts +1 -1
  444. package/src/gpu/kernels/backward/cross_entropy_backward.js +1 -1
  445. package/src/gpu/kernels/backward/embed_backward.js +2 -1
  446. package/src/gpu/kernels/backward/executor.d.ts +34 -0
  447. package/src/gpu/kernels/backward/executor.js +158 -0
  448. package/src/gpu/kernels/backward/gated_delta_recurrent_backward.js +2 -1
  449. package/src/gpu/kernels/backward/gated_rmsnorm_backward.js +2 -1
  450. package/src/gpu/kernels/backward/gelu_backward.d.ts +1 -1
  451. package/src/gpu/kernels/backward/gelu_backward.js +1 -1
  452. package/src/gpu/kernels/backward/groupnorm_backward.d.ts +1 -1
  453. package/src/gpu/kernels/backward/groupnorm_backward.js +1 -1
  454. package/src/gpu/kernels/backward/layernorm_backward.d.ts +1 -1
  455. package/src/gpu/kernels/backward/layernorm_backward.js +2 -1
  456. package/src/gpu/kernels/backward/matmul.js +133 -0
  457. package/src/gpu/kernels/backward/matmul_backward.d.ts +0 -1
  458. package/src/gpu/kernels/backward/matmul_backward.js +1 -1
  459. package/src/gpu/kernels/backward/pixel_shuffle_backward.d.ts +1 -1
  460. package/src/gpu/kernels/backward/pixel_shuffle_backward.js +1 -1
  461. package/src/gpu/kernels/backward/rmsnorm_backward.d.ts +1 -1
  462. package/src/gpu/kernels/backward/rmsnorm_backward.js +2 -1
  463. package/src/gpu/kernels/backward/rope_backward.d.ts +1 -1
  464. package/src/gpu/kernels/backward/rope_backward.js +2 -1
  465. package/src/gpu/kernels/backward/scale_backward.d.ts +1 -1
  466. package/src/gpu/kernels/backward/scale_backward.js +2 -1
  467. package/src/gpu/kernels/backward/silu_backward.d.ts +1 -1
  468. package/src/gpu/kernels/backward/silu_backward.js +1 -1
  469. package/src/gpu/kernels/backward/softmax_backward.d.ts +1 -1
  470. package/src/gpu/kernels/backward/softmax_backward.js +1 -1
  471. package/src/gpu/kernels/backward/upsample2d_backward.d.ts +1 -1
  472. package/src/gpu/kernels/backward/upsample2d_backward.js +1 -1
  473. package/src/gpu/kernels/backward/utils.js +2 -392
  474. package/src/gpu/kernels/bind-group-validation.js +16 -0
  475. package/src/gpu/kernels/cast.js +2 -1
  476. package/src/gpu/kernels/check-finiteness.js +13 -49
  477. package/src/gpu/kernels/check-hot-vocab-stop.js +13 -39
  478. package/src/gpu/kernels/check-stop.js +14 -33
  479. package/src/gpu/kernels/check_finiteness.wgsl +33 -0
  480. package/src/gpu/kernels/check_hot_vocab_stop.wgsl +29 -0
  481. package/src/gpu/kernels/check_stop.wgsl +20 -0
  482. package/src/gpu/kernels/clamp.js +1 -1
  483. package/src/gpu/kernels/conv2d.js +1 -1
  484. package/src/gpu/kernels/cross_entropy_loss.js +1 -1
  485. package/src/gpu/kernels/depthwise_conv1d.js +1 -1
  486. package/src/gpu/kernels/depthwise_conv2d.js +1 -1
  487. package/src/gpu/kernels/dequant.js +7 -2
  488. package/src/gpu/kernels/diffusion-gemma-sampling.js +3 -5
  489. package/src/gpu/kernels/dispatch.js +12 -1
  490. package/src/gpu/kernels/embedding-pool.js +72 -0
  491. package/src/gpu/kernels/embedding_pool_f16.wgsl +38 -0
  492. package/src/gpu/kernels/embedding_pool_f32.wgsl +36 -0
  493. package/src/gpu/kernels/energy.js +2 -1
  494. package/src/gpu/kernels/forecast_output.wgsl +17 -0
  495. package/src/gpu/kernels/fused-gate-up-gelu.js +1 -1
  496. package/src/gpu/kernels/fused_ffn.js +10 -220
  497. package/src/gpu/kernels/fused_matmul_residual.js +2 -1
  498. package/src/gpu/kernels/fused_matmul_rmsnorm.js +2 -1
  499. package/src/gpu/kernels/fused_normed_ffn_q4k.wgsl +208 -0
  500. package/src/gpu/kernels/gated-short-conv.js +11 -69
  501. package/src/gpu/kernels/gated_delta_conv.wgsl +66 -0
  502. package/src/gpu/kernels/gated_delta_conv_f16.wgsl +68 -0
  503. package/src/gpu/kernels/gated_delta_fused_decode.wgsl +173 -0
  504. package/src/gpu/kernels/gated_delta_fused_decode_f16.wgsl +175 -0
  505. package/src/gpu/kernels/gated_delta_recurrent.wgsl +165 -0
  506. package/src/gpu/kernels/gated_delta_recurrent_f16.wgsl +167 -0
  507. package/src/gpu/kernels/gated_short_conv.wgsl +47 -0
  508. package/src/gpu/kernels/gather.d.ts +1 -1
  509. package/src/gpu/kernels/gather.js +30 -6
  510. package/src/gpu/kernels/gather_bf16.wgsl +69 -0
  511. package/src/gpu/kernels/gelu.js +1 -1
  512. package/src/gpu/kernels/gemma4-route-expert.js +1 -1
  513. package/src/gpu/kernels/grouped_pointwise_conv2d.js +1 -1
  514. package/src/gpu/kernels/groupnorm.js +1 -1
  515. package/src/gpu/kernels/index.d.ts +38 -7
  516. package/src/gpu/kernels/index.js +18 -6
  517. package/src/gpu/kernels/instance_stats.wgsl +26 -0
  518. package/src/gpu/kernels/kernel-base.js +1 -1
  519. package/src/gpu/kernels/kernel-cache-lifecycle.d.ts +9 -0
  520. package/src/gpu/kernels/kernel-cache-lifecycle.js +26 -0
  521. package/src/gpu/kernels/kernel-configs.d.ts +10 -61
  522. package/src/gpu/kernels/kernel-configs.js +26 -54
  523. package/src/gpu/kernels/kernel-execution.js +141 -0
  524. package/src/gpu/kernels/kernel-prewarm.d.ts +3 -0
  525. package/src/gpu/kernels/kernel-prewarm.js +60 -0
  526. package/src/gpu/kernels/kv-cache-write.js +4 -4
  527. package/src/gpu/kernels/kv-quantize.js +4 -1
  528. package/src/gpu/kernels/l2-normalize.js +36 -0
  529. package/src/gpu/kernels/l2_normalize.wgsl +53 -0
  530. package/src/gpu/kernels/layernorm.d.ts +4 -4
  531. package/src/gpu/kernels/layernorm.js +44 -7
  532. package/src/gpu/kernels/linear-attention-core.js +29 -449
  533. package/src/gpu/kernels/linear_attention.js +1 -1
  534. package/src/gpu/kernels/lm-head-argmax.js +2 -2
  535. package/src/gpu/kernels/lm-head-select-logits.js +2 -2
  536. package/src/gpu/kernels/logit-finalize.js +103 -0
  537. package/src/gpu/kernels/logit-merge.js +13 -125
  538. package/src/gpu/kernels/logit-scatter.js +40 -0
  539. package/src/gpu/kernels/logit_finalize_f16.wgsl +43 -0
  540. package/src/gpu/kernels/logit_finalize_f32.wgsl +41 -0
  541. package/src/gpu/kernels/logit_merge_geometric.wgsl +23 -0
  542. package/src/gpu/kernels/logit_merge_max.wgsl +23 -0
  543. package/src/gpu/kernels/logit_merge_weighted.wgsl +23 -0
  544. package/src/gpu/kernels/logit_scatter_f16.wgsl +26 -0
  545. package/src/gpu/kernels/logit_scatter_f32.wgsl +24 -0
  546. package/src/gpu/kernels/matmul-dispatch.js +6 -1
  547. package/src/gpu/kernels/matmul-selection.js +3 -2
  548. package/src/gpu/kernels/matmul.js +4 -15
  549. package/src/gpu/kernels/matmul_bf16w_f32a.wgsl +93 -0
  550. package/src/gpu/kernels/matmul_gemv_subgroup_bf16w.wgsl +433 -0
  551. package/src/gpu/kernels/modulate.js +1 -1
  552. package/src/gpu/kernels/moe.js +3 -1
  553. package/src/gpu/kernels/moe_router_bias_add_f16_f16.wgsl +23 -0
  554. package/src/gpu/kernels/moe_router_bias_add_f16_f32.wgsl +23 -0
  555. package/src/gpu/kernels/moe_router_bias_add_f32_f16.wgsl +23 -0
  556. package/src/gpu/kernels/moe_router_bias_add_f32_f32.wgsl +21 -0
  557. package/src/gpu/kernels/patch_context.wgsl +29 -0
  558. package/src/gpu/kernels/pipeline-cache.js +3 -1
  559. package/src/gpu/kernels/pixel_shuffle.js +1 -1
  560. package/src/gpu/kernels/probe_shader_f16.wgsl +7 -0
  561. package/src/gpu/kernels/relu.js +1 -1
  562. package/src/gpu/kernels/repeat_channels.js +1 -1
  563. package/src/gpu/kernels/residual.js +2 -2
  564. package/src/gpu/kernels/rmsnorm-pair.js +1 -1
  565. package/src/gpu/kernels/rmsnorm-qk.js +1 -1
  566. package/src/gpu/kernels/rmsnorm-stats.js +14 -113
  567. package/src/gpu/kernels/rmsnorm.js +1 -1
  568. package/src/gpu/kernels/rmsnorm_serial.wgsl +19 -0
  569. package/src/gpu/kernels/rmsnorm_stats.wgsl +59 -0
  570. package/src/gpu/kernels/rmsnorm_stats_subgroups.wgsl +76 -0
  571. package/src/gpu/kernels/rope-precompute.js +160 -0
  572. package/src/gpu/kernels/rope-qk.js +1 -1
  573. package/src/gpu/kernels/rope.js +2 -2
  574. package/src/gpu/kernels/rope_precompute.wgsl +74 -0
  575. package/src/gpu/kernels/sample.js +5 -1
  576. package/src/gpu/kernels/scale.js +1 -1
  577. package/src/gpu/kernels/shader-cache.js +8 -4
  578. package/src/gpu/kernels/shader-source-scope.js +68 -0
  579. package/src/gpu/kernels/silu.js +2 -1
  580. package/src/gpu/kernels/soft-embedding.js +3 -1
  581. package/src/gpu/kernels/softmax.js +4 -2
  582. package/src/gpu/kernels/split-qkv-rmsnorm-qk.js +1 -1
  583. package/src/gpu/kernels/split-qkv-rmsnorm-rope-qk.js +1 -1
  584. package/src/gpu/kernels/split_qg.js +1 -1
  585. package/src/gpu/kernels/split_qkv.d.ts +2 -0
  586. package/src/gpu/kernels/split_qkv.js +18 -3
  587. package/src/gpu/kernels/split_qkv.wgsl +5 -2
  588. package/src/gpu/kernels/split_qkv_f16.wgsl +5 -2
  589. package/src/gpu/kernels/submit_probe.wgsl +6 -0
  590. package/src/gpu/kernels/transpose.js +1 -1
  591. package/src/gpu/kernels/turboquant-codebook.js +4 -535
  592. package/src/gpu/kernels/uniform-utils.js +7 -2
  593. package/src/gpu/kernels/upsample2d.js +1 -1
  594. package/src/gpu/kernels/utils.js +9 -332
  595. package/src/gpu/kernels/vision-average-pool.d.ts +13 -0
  596. package/src/gpu/kernels/vision-average-pool.js +68 -0
  597. package/src/gpu/kernels/vision-patch-embed.d.ts +17 -0
  598. package/src/gpu/kernels/vision-patch-embed.js +117 -0
  599. package/src/gpu/kernels/vision-position-embedding.d.ts +13 -0
  600. package/src/gpu/kernels/vision-position-embedding.js +79 -0
  601. package/src/gpu/kernels/vision-rope-2d.d.ts +16 -0
  602. package/src/gpu/kernels/vision-rope-2d.js +73 -0
  603. package/src/gpu/kernels/vision-spatial-merge.d.ts +19 -0
  604. package/src/gpu/kernels/vision-spatial-merge.js +100 -0
  605. package/src/gpu/kernels/vision_average_pool.wgsl +42 -0
  606. package/src/gpu/kernels/vision_patch_embed.wgsl +59 -0
  607. package/src/gpu/kernels/vision_patch_embed_f16w.wgsl +61 -0
  608. package/src/gpu/kernels/vision_position_embedding.wgsl +33 -0
  609. package/src/gpu/kernels/vision_position_embedding_f16.wgsl +35 -0
  610. package/src/gpu/kernels/vision_rope_2d.wgsl +60 -0
  611. package/src/gpu/kernels/vision_spatial_merge.wgsl +48 -0
  612. package/src/gpu/readback-buffer.js +20 -0
  613. package/src/gpu/submit-probe.js +2 -26
  614. package/src/gpu/weight-buffer.d.ts +1 -0
  615. package/src/gpu/weight-buffer.js +7 -0
  616. package/src/index.d.ts +23 -1
  617. package/src/index.js +8 -1
  618. package/src/inference/browser-harness/benchmark-execution.js +872 -0
  619. package/src/inference/browser-harness/benchmark-rerank.js +284 -0
  620. package/src/inference/browser-harness/decode-diagnostics.js +241 -0
  621. package/src/inference/browser-harness/diffusion-performance.js +102 -0
  622. package/src/inference/browser-harness/inference-execution.js +543 -0
  623. package/src/inference/{browser-harness-model-helpers.js → browser-harness/model-initialization.js} +18 -145
  624. package/src/inference/browser-harness/model-resolution.js +137 -0
  625. package/src/inference/browser-harness/report.js +338 -0
  626. package/src/inference/browser-harness/request.js +263 -0
  627. package/src/inference/browser-harness/runtime-config.d.ts +25 -0
  628. package/src/inference/browser-harness/runtime-config.js +255 -0
  629. package/src/inference/browser-harness/runtime-isolation.d.ts +13 -0
  630. package/src/inference/browser-harness/runtime-isolation.js +45 -0
  631. package/src/inference/browser-harness/runtime-report.d.ts +26 -0
  632. package/src/inference/browser-harness/runtime-report.js +117 -0
  633. package/src/inference/browser-harness/suite-summary.js +39 -0
  634. package/src/inference/browser-harness/text-evidence.js +694 -0
  635. package/src/inference/browser-harness/text-execution.js +757 -0
  636. package/src/inference/browser-harness/text-input.js +725 -0
  637. package/src/inference/{browser-harness-suite-helpers.js → browser-harness/timing-diagnostics.js} +1 -379
  638. package/src/inference/browser-harness-contract-helpers.js +1 -1
  639. package/src/inference/browser-harness-diffusion-energy-suites.js +8 -3
  640. package/src/inference/browser-harness-runtime-helpers.d.ts +3 -56
  641. package/src/inference/browser-harness-runtime-helpers.js +3 -408
  642. package/src/inference/browser-harness-text-helpers.js +31 -2196
  643. package/src/inference/browser-harness.d.ts +19 -2
  644. package/src/inference/browser-harness.js +46 -1873
  645. package/src/inference/kv-cache/base.js +67 -381
  646. package/src/inference/kv-cache/basis-decomposed-paged.d.ts +2 -50
  647. package/src/inference/kv-cache/basis-decomposed-paged.js +1 -276
  648. package/src/inference/kv-cache/quantized.d.ts +2 -2
  649. package/src/inference/kv-cache/quantized.js +4 -8
  650. package/src/inference/kv-cache/update.js +309 -0
  651. package/src/inference/model-checkpoint-evidence.d.ts +75 -0
  652. package/src/inference/model-checkpoint-evidence.js +264 -0
  653. package/src/inference/moe-router.js +36 -216
  654. package/src/inference/multi-model-network.js +5 -341
  655. package/src/inference/observation-context.d.ts +22 -0
  656. package/src/inference/observation-context.js +41 -0
  657. package/src/inference/pipelines/audio/index.js +1 -20
  658. package/src/inference/pipelines/audio/mel.js +0 -38
  659. package/src/inference/pipelines/diffusion/profile-timing.js +10 -0
  660. package/src/inference/pipelines/diffusion/runtime-resources.js +54 -0
  661. package/src/inference/pipelines/diffusion/sd3/execution.js +805 -0
  662. package/src/inference/pipelines/diffusion/sd3/plan.js +125 -0
  663. package/src/inference/pipelines/diffusion/sd3-transformer.js +44 -895
  664. package/src/inference/pipelines/diffusion/{vae.js → vae/execution.js} +46 -618
  665. package/src/inference/pipelines/diffusion/{helpers.js → weight-contract.js} +1 -68
  666. package/src/inference/pipelines/energy-head/row-head-pipeline.d.ts +1 -103
  667. package/src/inference/pipelines/factory.js +14 -3
  668. package/src/inference/pipelines/shader-scoped-pipeline.js +29 -0
  669. package/src/inference/pipelines/structured/json-grammar-mask.js +0 -6
  670. package/src/inference/pipelines/text/abort-contract.js +22 -0
  671. package/src/inference/pipelines/text/attention/attn-config.js +1 -16
  672. package/src/inference/pipelines/text/attention/executor-immediate.js +42 -0
  673. package/src/inference/pipelines/text/attention/executor-recorded.js +29 -0
  674. package/src/inference/pipelines/text/attention/gate-projection.js +30 -0
  675. package/src/inference/pipelines/text/attention/heterogeneous-contract.js +8 -0
  676. package/src/inference/pipelines/text/attention/index.js +7 -0
  677. package/src/inference/pipelines/text/attention/interpreter/immediate.js +211 -0
  678. package/src/inference/pipelines/text/attention/interpreter/recorded.js +451 -0
  679. package/src/inference/pipelines/text/attention/interpreter.js +984 -0
  680. package/src/inference/pipelines/text/attention/plan.d.ts +37 -0
  681. package/src/inference/pipelines/text/attention/plan.js +345 -0
  682. package/src/inference/pipelines/text/attention/projection-executor.js +331 -0
  683. package/src/inference/pipelines/text/attention/projection-plan.js +259 -0
  684. package/src/inference/pipelines/text/attention/projections.js +20 -537
  685. package/src/inference/pipelines/text/attention/query-transform.js +11 -0
  686. package/src/inference/pipelines/text/attention/receipt.js +60 -0
  687. package/src/inference/pipelines/text/attention/record.js +1 -1277
  688. package/src/inference/pipelines/text/attention/recorded-health.js +21 -0
  689. package/src/inference/pipelines/text/attention/rope-observation.js +40 -0
  690. package/src/inference/pipelines/text/attention/run.js +1 -1385
  691. package/src/inference/pipelines/text/attention/types.js +3 -31
  692. package/src/inference/pipelines/text/chat-format.d.ts +3 -0
  693. package/src/inference/pipelines/text/chat-format.js +42 -0
  694. package/src/inference/pipelines/text/config/normalization.js +729 -0
  695. package/src/inference/pipelines/text/config/validation.js +9 -0
  696. package/src/inference/pipelines/text/config.d.ts +12 -1
  697. package/src/inference/pipelines/text/config.js +14 -1228
  698. package/src/inference/pipelines/text/debug-utils/health-metrics.js +57 -0
  699. package/src/inference/pipelines/text/debug-utils/index.js +3 -8
  700. package/src/inference/pipelines/text/debug-utils/profiles.js +8 -0
  701. package/src/inference/pipelines/text/debug-utils/readback-decoding.js +24 -0
  702. package/src/inference/pipelines/text/debug-utils/tensor.js +1 -1
  703. package/src/inference/pipelines/text/dtype-contract.js +1 -68
  704. package/src/inference/pipelines/text/embed.js +178 -241
  705. package/src/inference/pipelines/text/embedding-contract.d.ts +7 -0
  706. package/src/inference/pipelines/text/embedding-contract.js +21 -0
  707. package/src/inference/pipelines/text/embedding-extraction.js +225 -0
  708. package/src/inference/pipelines/text/embedding-normalization.js +39 -0
  709. package/src/inference/pipelines/text/execution-runtime-builders.js +0 -32
  710. package/src/inference/pipelines/text/execution-v1/plan.js +69 -0
  711. package/src/inference/pipelines/text/execution-v1/validation.js +233 -0
  712. package/src/inference/pipelines/text/execution-v1.js +7 -283
  713. package/src/inference/pipelines/text/execution.js +517 -0
  714. package/src/inference/pipelines/text/ffn/dense-executor.js +718 -0
  715. package/src/inference/pipelines/text/ffn/dense-plan.js +327 -0
  716. package/src/inference/pipelines/text/ffn/dense.js +4 -968
  717. package/src/inference/pipelines/text/ffn/sandwich.js +10 -9
  718. package/src/inference/pipelines/text/ffn/standard.js +2 -1
  719. package/src/inference/pipelines/text/generator/attention-lifecycle.js +24 -0
  720. package/src/inference/pipelines/text/generator/decode-runtime.js +991 -0
  721. package/src/inference/pipelines/text/generator/decode.js +952 -0
  722. package/src/inference/pipelines/text/generator/diffusion.js +607 -0
  723. package/src/inference/pipelines/text/generator/logits-config.js +63 -0
  724. package/src/inference/pipelines/text/generator/prefill-observation.js +29 -0
  725. package/src/inference/pipelines/text/generator/prefill-policy.js +56 -0
  726. package/src/inference/pipelines/text/generator/prefill-runtime.js +896 -0
  727. package/src/inference/pipelines/text/{generator-prefill-helpers.js → generator/prefix-embedding.js} +16 -121
  728. package/src/inference/pipelines/text/generator/prompt-input.js +47 -0
  729. package/src/inference/pipelines/text/generator/recovery.js +310 -0
  730. package/src/inference/pipelines/text/{generator-helpers.js → generator/session-context.js} +44 -110
  731. package/src/inference/pipelines/text/generator/text.js +448 -0
  732. package/src/inference/pipelines/text/generator/training.js +752 -0
  733. package/src/inference/pipelines/text/generator-decode-policy.js +5 -2
  734. package/src/inference/pipelines/text/generator-logits-observation.js +45 -0
  735. package/src/inference/pipelines/text/generator-runtime.js +5 -135
  736. package/src/inference/pipelines/text/generator-steps.js +37 -962
  737. package/src/inference/pipelines/text/generator.js +165 -3539
  738. package/src/inference/pipelines/text/image-transcription.js +304 -0
  739. package/src/inference/pipelines/text/init-qkv-fusion.js +18 -9
  740. package/src/inference/pipelines/text/init-rope.js +37 -189
  741. package/src/inference/pipelines/text/init.d.ts +11 -3
  742. package/src/inference/pipelines/text/init.js +6 -216
  743. package/src/inference/pipelines/text/layer-execution-contract.js +97 -0
  744. package/src/inference/pipelines/text/layer-input-execution.js +435 -0
  745. package/src/inference/pipelines/text/layer-plan-gpu.js +15 -6
  746. package/src/inference/pipelines/text/layer-plan.d.ts +21 -5
  747. package/src/inference/pipelines/text/layer-plan.js +45 -3
  748. package/src/inference/pipelines/text/layer.js +54 -522
  749. package/src/inference/pipelines/text/lifecycle.js +839 -0
  750. package/src/inference/pipelines/text/linear-attention/plan.js +554 -0
  751. package/src/inference/pipelines/text/linear-attention-ab-fusion.js +2 -2
  752. package/src/inference/pipelines/text/linear-attention.d.ts +2 -2
  753. package/src/inference/pipelines/text/linear-attention.js +7 -537
  754. package/src/inference/pipelines/text/logits/cpu-output.d.ts +1 -0
  755. package/src/inference/pipelines/text/logits/cpu-output.js +5 -0
  756. package/src/inference/pipelines/text/logits/cpu.d.ts +1 -87
  757. package/src/inference/pipelines/text/logits/cpu.js +8 -95
  758. package/src/inference/pipelines/text/logits/gpu-executor.js +368 -0
  759. package/src/inference/pipelines/text/logits/gpu.d.ts +27 -0
  760. package/src/inference/pipelines/text/logits/gpu.js +116 -579
  761. package/src/inference/pipelines/text/logits/index.d.ts +2 -1
  762. package/src/inference/pipelines/text/logits/index.js +62 -102
  763. package/src/inference/pipelines/text/logits/output-transform.js +47 -0
  764. package/src/inference/pipelines/text/logits/plan.js +184 -0
  765. package/src/inference/pipelines/text/logits/readback.d.ts +6 -0
  766. package/src/inference/pipelines/text/logits/readback.js +9 -0
  767. package/src/inference/pipelines/text/logits/scale-policy.js +20 -0
  768. package/src/inference/pipelines/text/logits/types.d.ts +1 -0
  769. package/src/inference/pipelines/text/lora-apply.js +12 -4
  770. package/src/inference/pipelines/text/lora-types.d.ts +12 -0
  771. package/src/inference/pipelines/text/modality-token-contract.js +115 -0
  772. package/src/inference/pipelines/text/model-load.d.ts +0 -1
  773. package/src/inference/pipelines/text/model-load.js +2 -31
  774. package/src/inference/pipelines/text/moe/expert-loading.js +9 -0
  775. package/src/inference/pipelines/text/moe/gpu-executor.js +513 -0
  776. package/src/inference/pipelines/text/moe/plan.js +47 -0
  777. package/src/inference/pipelines/text/moe-cpu.js +9 -135
  778. package/src/inference/pipelines/text/moe-gpu.js +21 -545
  779. package/src/inference/pipelines/text/normalization-contract.js +17 -0
  780. package/src/inference/pipelines/text/per-layer/materialize.js +736 -0
  781. package/src/inference/pipelines/text/per-layer/plan.js +362 -0
  782. package/src/inference/pipelines/text/per-layer-inputs.js +4 -1037
  783. package/src/inference/pipelines/text/pipeline-load-timing.js +77 -0
  784. package/src/inference/pipelines/text/probes.js +30 -0
  785. package/src/inference/pipelines/text/resolved-runtime-session.d.ts +56 -0
  786. package/src/inference/pipelines/text/resolved-runtime-session.js +175 -0
  787. package/src/inference/pipelines/text/state.d.ts +14 -0
  788. package/src/inference/pipelines/text/state.js +4 -0
  789. package/src/inference/pipelines/text/tsir-fixture-writer.js +88 -16
  790. package/src/inference/pipelines/text/types.d.ts +10 -0
  791. package/src/inference/pipelines/text/weight-loading-config.js +194 -0
  792. package/src/inference/pipelines/text.d.ts +2 -2
  793. package/src/inference/pipelines/text.js +48 -1620
  794. package/src/inference/pipelines/video/frame-sampler.js +0 -13
  795. package/src/inference/pipelines/video/index.js +0 -21
  796. package/src/inference/pipelines/vision/attention.js +59 -0
  797. package/src/inference/pipelines/vision/encoder.js +254 -369
  798. package/src/inference/pipelines/vision/gemma4.js +204 -318
  799. package/src/inference/pipelines/vision/glmocr-preprocess.js +273 -0
  800. package/src/inference/pipelines/vision/glmocr-rope.js +106 -0
  801. package/src/inference/pipelines/vision/glmocr.js +557 -0
  802. package/src/inference/pipelines/vision/image-preprocess.js +0 -27
  803. package/src/inference/pipelines/vision/index.js +8 -27
  804. package/src/inference/pipelines/vision/ops.js +11 -33
  805. package/src/inference/pipelines/vision/patch-embed.js +51 -145
  806. package/src/inference/pipelines/vision/spatial-merge.js +78 -0
  807. package/src/inference/pipelines/vision/weight-loading.js +79 -0
  808. package/src/inference/refactor-receipt.d.ts +49 -0
  809. package/src/inference/refactor-receipt.js +188 -0
  810. package/src/inference/rerank.js +182 -0
  811. package/src/inference/resource-scope.d.ts +33 -0
  812. package/src/inference/resource-scope.js +168 -0
  813. package/src/inference/test-harness.d.ts +3 -0
  814. package/src/inference/test-harness.js +1 -0
  815. package/src/inference/tokenizer.js +1 -1
  816. package/src/inference/tokenizers/behavior-flags.js +6 -3
  817. package/src/inference/tokenizers/bundled/decoders.js +133 -0
  818. package/src/inference/tokenizers/bundled/registry.js +216 -0
  819. package/src/inference/tokenizers/bundled.js +23 -353
  820. package/src/loader/doppler-loader.d.ts +4 -0
  821. package/src/loader/doppler-loader.js +119 -930
  822. package/src/loader/embedding-limit-preflight.js +11 -4
  823. package/src/loader/embedding-loader.js +1 -3
  824. package/src/loader/gpu-tensor-materialization.js +20 -0
  825. package/src/loader/layer-loader.js +15 -119
  826. package/src/loader/layer-tensor-aliases.js +99 -0
  827. package/src/loader/layer-weight-shape.js +29 -0
  828. package/src/loader/load-timing.js +64 -0
  829. package/src/loader/loader-types.d.ts +3 -0
  830. package/src/loader/model-executor.js +819 -0
  831. package/src/loader/model-load-validation.js +15 -0
  832. package/src/loader/quantization-constants.d.ts +11 -29
  833. package/src/loader/quantization-constants.js +1 -2
  834. package/src/loader/shard-resolver.js +0 -14
  835. package/src/loader/tensor-load-error.js +14 -0
  836. package/src/loader/tensor-shard-indices.js +8 -0
  837. package/src/loader/tensors/bf16-loader.js +125 -0
  838. package/src/loader/tensors/functional-descriptor-loader.js +3 -375
  839. package/src/loader/tensors/load-executor.js +577 -0
  840. package/src/loader/tensors/source-transform/executor.js +123 -0
  841. package/src/loader/tensors/source-transform/plan.js +107 -0
  842. package/src/loader/tensors/source-transform.js +15 -264
  843. package/src/loader/tensors/tensor-loader.d.ts +4 -4
  844. package/src/loader/tensors/tensor-loader.js +7 -653
  845. package/src/memory/buffer-pool.js +7 -29
  846. package/src/rules/capability-transform-resolver.js +80 -0
  847. package/src/rules/execution-contract.d.ts +73 -0
  848. package/src/rules/execution-contract.js +429 -0
  849. package/src/rules/execution-rules-contract-check.js +1 -1
  850. package/src/rules/generated/rule-bundle.json +4305 -0
  851. package/src/rules/inference/capability-transforms.rules.json +3 -3
  852. package/src/rules/layer-pattern-contract-check.js +1 -1
  853. package/src/rules/rule-registry.js +70 -57
  854. package/src/storage/artifact-storage-context.js +2 -2
  855. package/src/storage/backend-selection.js +43 -0
  856. package/src/storage/distribution-transport.js +5 -0
  857. package/src/storage/download/integrity.d.ts +33 -0
  858. package/src/storage/download/integrity.js +130 -0
  859. package/src/storage/download/resume-inspection.d.ts +22 -0
  860. package/src/storage/download/resume-inspection.js +134 -0
  861. package/src/storage/download/retry.js +57 -0
  862. package/src/storage/download/state.js +131 -0
  863. package/src/storage/download/transport.js +99 -0
  864. package/src/storage/downloader.js +74 -453
  865. package/src/storage/model-cache.d.ts +54 -0
  866. package/src/storage/model-cache.js +590 -0
  867. package/src/storage/shard-manager.d.ts +5 -0
  868. package/src/storage/shard-manager.js +50 -369
  869. package/src/storage/shards/index.js +60 -0
  870. package/src/storage/shards/integrity.js +225 -0
  871. package/src/storage/shards/lifecycle.js +22 -0
  872. package/src/storage/source-artifact-store.js +14 -4
  873. package/src/tooling/boundary-evidence.d.ts +51 -0
  874. package/src/tooling/boundary-evidence.js +408 -0
  875. package/src/tooling/browser-command-runner.js +17 -10
  876. package/src/tooling/capsule-v2.d.ts +11 -0
  877. package/src/tooling/capsule-v2.js +58 -0
  878. package/src/tooling/capsule.js +20 -0
  879. package/src/tooling/command-api-constants.js +1 -1
  880. package/src/tooling/command-api-family-normalizers.js +163 -5
  881. package/src/tooling/command-api.d.ts +42 -3
  882. package/src/tooling/command-api.js +7 -2
  883. package/src/tooling/command-context.d.ts +25 -0
  884. package/src/tooling/command-context.js +53 -0
  885. package/src/tooling/command-envelope.js +1 -1
  886. package/src/tooling/{command-api-helpers.js → command-request-normalization.js} +1 -1
  887. package/src/tooling/command-runner.html +49 -1
  888. package/src/tooling/{command-runner-shared.js → command-runtime-execution.js} +22 -5
  889. package/src/tooling/command-validation.js +1 -24
  890. package/src/tooling/conversion-run-timing.js +24 -0
  891. package/src/tooling/demo-receipts.d.ts +56 -0
  892. package/src/tooling/demo-receipts.js +113 -0
  893. package/src/tooling/diagnose-runner.js +13 -11
  894. package/src/tooling/distribution-shard-transport.js +11 -0
  895. package/src/tooling/execution-cost-ledger.d.ts +95 -0
  896. package/src/tooling/execution-cost-ledger.js +381 -0
  897. package/src/tooling/hf-registry/catalog-io.js +63 -0
  898. package/src/tooling/litert/package.js +863 -0
  899. package/src/tooling/litert-package-runtime.js +3 -849
  900. package/src/tooling/model-capsule-forge.js +478 -0
  901. package/src/tooling/node-browser/execution.d.ts +40 -0
  902. package/src/tooling/node-browser/execution.js +308 -0
  903. package/src/tooling/node-browser/static-server.js +188 -0
  904. package/src/tooling/node-browser/transport.d.ts +80 -0
  905. package/src/tooling/node-browser/transport.js +622 -0
  906. package/src/tooling/node-browser-command-runner.d.ts +18 -78
  907. package/src/tooling/node-browser-command-runner.js +10 -1112
  908. package/src/tooling/node-capsule-artifact-store.js +44 -0
  909. package/src/tooling/node-command-execution.js +237 -0
  910. package/src/tooling/node-command-runner.js +26 -205
  911. package/src/tooling/node-converter/input.js +521 -0
  912. package/src/tooling/node-converter/output.js +244 -0
  913. package/src/tooling/node-converter-chunk-batches.js +21 -0
  914. package/src/tooling/node-converter.js +7 -720
  915. package/src/tooling/node-source-runtime.js +16 -276
  916. package/src/tooling/node-webgpu-provider.v1.json +33 -0
  917. package/src/tooling/node-webgpu.js +178 -453
  918. package/src/tooling/opfs-cache.d.ts +15 -48
  919. package/src/tooling/opfs-cache.js +8 -510
  920. package/src/tooling/precision-replay-math.d.ts +42 -0
  921. package/src/tooling/precision-replay-math.js +174 -0
  922. package/src/tooling/production-release.d.ts +19 -0
  923. package/src/tooling/production-release.js +647 -0
  924. package/src/tooling/program-bundle/embedding-reference.js +19 -0
  925. package/src/tooling/program-bundle/materialize.js +324 -0
  926. package/src/tooling/program-bundle/rerank-reference.js +22 -0
  927. package/src/tooling/program-bundle/sequence-reference.js +42 -0
  928. package/src/tooling/program-bundle/validation.js +31 -0
  929. package/src/tooling/program-bundle/wgsl-closure.js +254 -0
  930. package/src/tooling/program-bundle-parity.d.ts +14 -4
  931. package/src/tooling/program-bundle-parity.js +267 -106
  932. package/src/tooling/program-bundle.d.ts +8 -52
  933. package/src/tooling/program-bundle.js +186 -730
  934. package/src/tooling/provider-v1-contract.d.ts +61 -0
  935. package/src/tooling/provider-v1-contract.js +192 -0
  936. package/src/tooling/registered-variant-calibration-job.d.ts +98 -0
  937. package/src/tooling/registered-variant-calibration-job.js +210 -0
  938. package/src/tooling/registered-variant-calibration.d.ts +184 -0
  939. package/src/tooling/registered-variant-calibration.js +336 -0
  940. package/src/tooling/runtime-input-composition.js +105 -53
  941. package/src/tooling/runtime-intent.js +41 -0
  942. package/src/tooling/runtime-optimization/candidates.js +664 -0
  943. package/src/tooling/runtime-optimization-campaign.js +166 -0
  944. package/src/tooling/runtime-optimization-index.d.ts +6 -0
  945. package/src/tooling/runtime-optimization-index.js +87 -0
  946. package/src/tooling/runtime-optimization.d.ts +91 -6
  947. package/src/tooling/runtime-optimization.js +290 -414
  948. package/src/tooling/source-intake.d.ts +49 -0
  949. package/src/tooling/source-intake.js +431 -0
  950. package/src/tooling/source-package-profiles.js +1 -1
  951. package/src/tooling/source-runtime/bundle-materialize.js +298 -0
  952. package/src/tooling/source-runtime/bundle-plan.js +121 -0
  953. package/src/tooling/source-runtime/node-loader.js +258 -0
  954. package/src/tooling/source-runtime-bundle.js +26 -411
  955. package/src/tooling/source-runtime-converter-config.js +1 -1
  956. package/src/tooling-experimental-exports.js +0 -5
  957. package/src/tooling-exports/evidence.d.ts +45 -0
  958. package/src/tooling-exports/evidence.js +45 -0
  959. package/src/tooling-exports/runtime.d.ts +7 -0
  960. package/src/tooling-exports/runtime.js +7 -0
  961. package/src/tooling-exports/storage.d.ts +5 -0
  962. package/src/tooling-exports/storage.js +4 -0
  963. package/src/tooling-exports.d.ts +14 -0
  964. package/src/tooling-exports.js +8 -3
  965. package/src/tooling-exports.shared.d.ts +46 -0
  966. package/src/tooling-exports.shared.js +37 -0
  967. package/src/training.d.ts +6 -6
  968. package/src/training.js +4 -4
  969. package/src/version.js +1 -1
  970. package/tests/kernels/browser/kernel-suite.js +512 -0
  971. package/tests/kernels/browser/test-page.js +169 -14
  972. package/tests/kernels/harness/tolerance.js +1 -0
  973. package/tests/kernels/reference/index.js +13 -1
  974. package/tests/kernels/reference/vision.js +160 -0
  975. package/src/client/runtime/model-session.js +0 -432
  976. package/src/config/schema/tuner.schema.d.ts +0 -39
  977. package/src/config/schema/tuner.schema.js +0 -13
  978. package/src/converter/quantizer.d.ts +0 -121
  979. package/src/gpu/kernel-tuner/benchmarks.js +0 -902
  980. package/src/gpu/kernel-tuner/cache.js +0 -145
  981. package/src/gpu/kernel-tuner/index.js +0 -38
  982. package/src/gpu/kernel-tuner/tuner.js +0 -247
  983. package/src/gpu/kernel-tuner.js +0 -12
  984. package/src/gpu/kernels/backward/utils.d.ts +0 -90
  985. package/src/gpu/kernels/kernel-tuning.d.ts +0 -42
  986. package/src/gpu/kernels/kernel-tuning.js +0 -158
  987. package/src/gpu/kernels/uniform-utils.d.ts +0 -66
  988. package/src/gpu/kernels/utils.d.ts +0 -149
  989. package/src/gpu/profiler.js +0 -382
  990. package/src/inference/pipelines/text/attention/output-gate-fusion.js +0 -37
  991. package/src/inference/pipelines/text/debug-utils/utils.js +0 -118
  992. package/src/inference/pipelines/text/logits/utils.d.ts +0 -58
  993. package/src/inference/pipelines/text/moe-cpu-gptoss.js +0 -115
  994. package/src/inference/pipelines/text/moe-helpers.js +0 -21
  995. package/src/rules/converter/execution.rules.json +0 -20
  996. package/src/rules/converter/tensor-roles.rules.json +0 -13
  997. package/src/rules/converter/tokenizer.rules.json +0 -12
  998. package/src/rules/inference/attention.rules.json +0 -69
  999. package/src/rules/inference/config.rules.json +0 -82
  1000. package/src/rules/inference/dtype.rules.json +0 -113
  1001. package/src/rules/inference/execution.rules.json +0 -195
  1002. package/src/rules/inference/ffn.rules.json +0 -123
  1003. package/src/rules/inference/layer-pattern.rules.json +0 -16
  1004. package/src/rules/inference/layer.rules.json +0 -7
  1005. package/src/rules/inference/moe.rules.json +0 -145
  1006. package/src/rules/kernels/attention.rules.json +0 -68
  1007. package/src/rules/kernels/conv2d.rules.json +0 -6
  1008. package/src/rules/kernels/depthwise-conv2d.rules.json +0 -6
  1009. package/src/rules/kernels/dequant.rules.json +0 -58
  1010. package/src/rules/kernels/energy.rules.json +0 -22
  1011. package/src/rules/kernels/fused-ffn.rules.json +0 -15
  1012. package/src/rules/kernels/fused-matmul-residual.rules.json +0 -6
  1013. package/src/rules/kernels/fused-matmul-rmsnorm.rules.json +0 -8
  1014. package/src/rules/kernels/gather.rules.json +0 -93
  1015. package/src/rules/kernels/gelu.rules.json +0 -11
  1016. package/src/rules/kernels/grouped-pointwise-conv2d.rules.json +0 -6
  1017. package/src/rules/kernels/groupnorm.rules.json +0 -10
  1018. package/src/rules/kernels/kv_quantize.rules.json +0 -9
  1019. package/src/rules/kernels/layernorm.rules.json +0 -6
  1020. package/src/rules/kernels/linear-attention.rules.json +0 -6
  1021. package/src/rules/kernels/lm-head-argmax.rules.json +0 -10
  1022. package/src/rules/kernels/matmul.rules.json +0 -95
  1023. package/src/rules/kernels/modulate.rules.json +0 -6
  1024. package/src/rules/kernels/moe.rules.gptoss.json +0 -105
  1025. package/src/rules/kernels/moe.rules.json +0 -14
  1026. package/src/rules/kernels/moe.rules.mixtral.json +0 -75
  1027. package/src/rules/kernels/pixel_shuffle.rules.json +0 -6
  1028. package/src/rules/kernels/relu.rules.json +0 -6
  1029. package/src/rules/kernels/rep-penalty.rules.json +0 -6
  1030. package/src/rules/kernels/repeat-channels.rules.json +0 -6
  1031. package/src/rules/kernels/residual.rules.json +0 -12
  1032. package/src/rules/kernels/rmsnorm-qk.rules.json +0 -6
  1033. package/src/rules/kernels/rmsnorm.rules.json +0 -46
  1034. package/src/rules/kernels/rope-qk.rules.json +0 -6
  1035. package/src/rules/kernels/rope.rules.json +0 -6
  1036. package/src/rules/kernels/sample.rules.json +0 -6
  1037. package/src/rules/kernels/scale.rules.json +0 -8
  1038. package/src/rules/kernels/silu.rules.json +0 -22
  1039. package/src/rules/kernels/softmax.rules.json +0 -25
  1040. package/src/rules/kernels/split-qg.rules.json +0 -6
  1041. package/src/rules/kernels/split-qkv.rules.json +0 -6
  1042. package/src/rules/kernels/upsample2d.rules.json +0 -6
  1043. package/src/rules/loader/tensor-loader.rules.json +0 -42
  1044. package/src/rules/loader/weights.rules.json +0 -42
  1045. package/src/rules/tooling/command-runtime.rules.json +0 -65
  1046. package/src/tooling/hf-registry-utils.js +0 -395
  1047. /package/src/converter/{tokenizer-utils.js → tokenizer-special-tokens.js} +0 -0
  1048. /package/src/{inference/pipelines/text/bdpa-steamroller.js → experimental/bdpa/steamroller.js} +0 -0
  1049. /package/src/{inference/pipelines → experimental}/diffusion/scheduler.js +0 -0
  1050. /package/src/{inference/pipelines → experimental}/energy/pipeline.d.ts +0 -0
  1051. /package/src/{inference/pipelines → experimental}/energy/quintel.d.ts +0 -0
  1052. /package/src/{inference/pipelines → experimental}/energy/quintel.js +0 -0
  1053. /package/src/{utils → formats}/array-buffer.js +0 -0
  1054. /package/src/{utils → formats}/canonical-hash.js +0 -0
  1055. /package/src/{utils → formats}/clone-json.js +0 -0
  1056. /package/src/{utils → formats}/encode-utf8.js +0 -0
  1057. /package/src/{utils → formats}/load-json.js +0 -0
  1058. /package/src/{utils → formats}/plain-object.js +0 -0
  1059. /package/src/{utils → formats}/sha256.js +0 -0
  1060. /package/src/{utils → formats}/stable-sort-object.js +0 -0
  1061. /package/src/{gpu/kernels → rules}/rule-matcher.js +0 -0
  1062. /package/src/{utils → storage}/hf-resolve-url.js +0 -0
  1063. /package/src/{utils → storage}/runtime-env.js +0 -0
@@ -11,148 +11,91 @@ import {
11
11
  getActiveKernelPathPolicy,
12
12
  } from '../config/kernel-path-loader.js';
13
13
  import { validateTrainingMetricsReport } from '../config/schema/training-metrics.schema.js';
14
- import { modelSupportsEmbedding, modelSupportsRerank } from '../config/schema/manifest.schema.js';
15
14
  import {
16
- resolveReportTimestamp,
15
+ modelSupportsEmbedding,
16
+ modelSupportsRerank,
17
+ modelSupportsSequence,
18
+ } from '../config/schema/manifest.schema.js';
19
+ import {
17
20
  resolveRuntime,
18
- cloneRuntimeConfig,
19
- runWithRuntimeIsolationForSuite,
20
- sanitizeReportOutput,
21
21
  loadRuntimeConfigFromUrl,
22
22
  applyRuntimeConfigFromUrl,
23
23
  loadRuntimeProfile,
24
24
  applyRuntimeProfile,
25
25
  applyRuntimeForRun,
26
+ } from './browser-harness/runtime-config.js';
27
+ import {
28
+ cloneRuntimeConfig,
29
+ runWithRuntimeIsolationForSuite,
30
+ } from './browser-harness/runtime-isolation.js';
31
+ import {
32
+ resolveReportTimestamp,
33
+ sanitizeReportOutput,
26
34
  normalizeManifest,
27
35
  mergeRunDefaults,
28
36
  summarizeManifestRuns,
29
- } from './browser-harness-runtime-helpers.js';
37
+ } from './browser-harness/runtime-report.js';
30
38
  import {
31
39
  buildSuiteSummary,
32
40
  normalizeCacheMode,
33
41
  normalizeLoadMode,
34
42
  normalizeWorkloadType,
43
+ } from './browser-harness/suite-summary.js';
44
+ import {
35
45
  assertDiffusionPerformanceArtifact,
46
+ } from './browser-harness/diffusion-performance.js';
47
+ import {
36
48
  toTimingNumber,
37
49
  buildFirstLoadComposition,
38
50
  safeToFixed,
39
51
  sampleTimingNumber,
40
52
  buildCanonicalTiming,
41
53
  buildLoadTimingDiagnostics,
42
- buildDecodeBottleneckDiagnostics,
43
54
  buildTimingDiagnostics,
44
- } from './browser-harness-suite-helpers.js';
55
+ } from './browser-harness/timing-diagnostics.js';
56
+ import { buildDecodeBottleneckDiagnostics } from './browser-harness/decode-diagnostics.js';
45
57
  import {
46
58
  resolveDeviceInfo,
47
59
  resolveKernelPathForModel,
48
- initializeSuiteModel,
49
- } from './browser-harness-model-helpers.js';
60
+ } from './browser-harness/model-resolution.js';
61
+ import { initializeSuiteModel } from './browser-harness/model-initialization.js';
50
62
  import {
51
63
  resolveBenchmarkRunSettings,
64
+ } from './browser-harness/text-input.js';
65
+ import {
52
66
  normalizeDecodeRecordOpLabels,
53
67
  normalizeUniformCacheStats,
54
68
  buildDecodeRecordTopOps,
55
69
  buildDecodeRecordTopOpGroups,
70
+ isCoherentOutput,
71
+ } from './browser-harness/text-evidence.js';
72
+ import {
56
73
  runEmbeddingSemanticChecks,
57
74
  runRerank,
58
75
  runRerankSemanticChecks,
59
- isCoherentOutput,
76
+ runSequenceEncoding,
60
77
  runTextInference,
61
78
  runEmbedding,
62
- } from './browser-harness-text-helpers.js';
79
+ } from './browser-harness/text-execution.js';
63
80
  import { buildSuiteContractMetrics } from './browser-harness-contract-helpers.js';
64
81
  import {
65
82
  runDiffusionSuite,
66
83
  runEnergySuite,
67
84
  } from './browser-harness-diffusion-energy-suites.js';
68
85
  import { collectTrainingArtifactsFromSuiteResult } from './browser-harness-report-helpers.js';
69
- import { sha256Hex } from '../utils/sha256.js';
70
- import { stableSortObject } from '../utils/stable-sort-object.js';
71
-
72
- const TRAINING_SUITE_MODULE_PATH = '../experimental/training/suite.js';
73
- let trainingSuiteModulePromise = null;
74
-
75
- function resolvePipelineLoadTimings(pipeline) {
76
- if (!pipeline || typeof pipeline.getStats !== 'function') {
77
- return { loadTiming: null, pipelineLoadTiming: null };
78
- }
79
- try {
80
- const stats = pipeline.getStats() ?? {};
81
- return {
82
- loadTiming: stats.loadTiming ?? null,
83
- pipelineLoadTiming: stats.pipelineLoadTiming ?? null,
84
- };
85
- } catch {
86
- return { loadTiming: null, pipelineLoadTiming: null };
87
- }
88
- }
89
-
90
- function getNestedSampleValue(sample, key) {
91
- if (!sample || typeof sample !== 'object' || typeof key !== 'string' || key.length === 0) {
92
- return null;
93
- }
94
- let current = sample;
95
- for (const part of key.split('.')) {
96
- if (!current || typeof current !== 'object') {
97
- return null;
98
- }
99
- current = current[part];
100
- }
101
- return current;
102
- }
103
-
104
- function getNestedPhaseValue(sample, key) {
105
- const value = getNestedSampleValue(sample, key);
106
- return Number.isFinite(value) ? value : null;
107
- }
108
-
109
- function summarizeTimingPhaseSamples(samples, keys) {
110
- const rows = Array.isArray(samples) ? samples.filter((sample) => sample && typeof sample === 'object') : [];
111
- const summary = {};
112
- for (const key of keys) {
113
- const values = rows
114
- .map((sample) => getNestedPhaseValue(sample, key))
115
- .filter((value) => Number.isFinite(value));
116
- summary[key] = computeSampleStats(values);
117
- }
118
- return summary;
119
- }
120
-
121
- function summarizePrefillRecordOps(samples) {
122
- const rows = Array.isArray(samples) ? samples.filter((sample) => sample && typeof sample === 'object') : [];
123
- const labelSums = {};
124
- let labelSampleCount = 0;
125
- for (const sample of rows) {
126
- const labelCounts = normalizeDecodeRecordOpLabels(getNestedSampleValue(sample, 'prefillRecordOpLabels'));
127
- if (!labelCounts) {
128
- continue;
129
- }
130
- labelSampleCount += 1;
131
- for (const [label, count] of Object.entries(labelCounts)) {
132
- labelSums[label] = (labelSums[label] ?? 0) + count;
133
- }
134
- }
135
- if (labelSampleCount === 0) {
136
- return {};
137
- }
138
- const meanLabels = {};
139
- for (const [label, count] of Object.entries(labelSums)) {
140
- meanLabels[label] = count / labelSampleCount;
141
- }
142
- const prefillRecordOps = summarizeTimingPhaseSamples(rows, ['prefillRecordOps']).prefillRecordOps;
143
- return {
144
- prefillRecordUniqueOpLabels: Object.keys(meanLabels).length,
145
- prefillRecordTopOps: buildDecodeRecordTopOps(meanLabels, prefillRecordOps?.mean),
146
- prefillRecordTopOpGroups: buildDecodeRecordTopOpGroups(meanLabels, prefillRecordOps?.mean),
147
- };
148
- }
149
-
150
- async function loadTrainingSuiteModule() {
151
- if (!trainingSuiteModulePromise) {
152
- trainingSuiteModulePromise = import(TRAINING_SUITE_MODULE_PATH);
153
- }
154
- return trainingSuiteModulePromise;
155
- }
86
+ import { sha256Hex } from '../formats/sha256.js';
87
+ import { stableSortObject } from '../formats/stable-sort-object.js';
88
+ import { assertCommandContextMatchesOptions } from '../tooling/command-context.js';
89
+ import {
90
+ buildTokenCostLedger,
91
+ isExecutionObservationRequested,
92
+ } from '../tooling/execution-cost-ledger.js';
93
+ import { createUnsupportedWorkloadError, hashStableJson, resolveDecodeCadence, resolveDispatchSuite, resolveExecutionGraphHash, resolveHarnessContext, resolveHarnessMode, resolveWorkload } from './browser-harness/request.js';
94
+ import { attachExecutionCostLedger, buildReferenceTranscriptSeed, resolvePipelineLoadTimings, serializeSequenceProbeRows, summarizePrefillRecordOps, summarizeTimingPhaseSamples } from './browser-harness/report.js';
95
+ import { runInferenceSuite, runKernelSuite, withHarnessPhase } from './browser-harness/inference-execution.js';
96
+ import { loadTrainingSuiteModule, runBenchSuite } from './browser-harness/benchmark-execution.js';
97
+ export { buildReferenceTranscriptSeed } from './browser-harness/report.js';
98
+ export { getBrowserSuiteDispatchMap, getBrowserSupportedSuites, resolveExecutionGraphHash } from './browser-harness/request.js';
156
99
 
157
100
  export async function runTrainingSuite(options = {}) {
158
101
  const module = await loadTrainingSuiteModule();
@@ -168,1750 +111,6 @@ export {
168
111
  buildSuiteSummary,
169
112
  };
170
113
 
171
- async function runTrainingBenchSuite(options = {}) {
172
- const module = await loadTrainingSuiteModule();
173
- return module.runTrainingBenchSuite(options);
174
- }
175
-
176
- const BROWSER_WORKLOAD_SET = Object.freeze([
177
- 'kernels',
178
- 'inference',
179
- 'embedding',
180
- 'rerank',
181
- 'training',
182
- 'diffusion',
183
- 'energy',
184
- ]);
185
-
186
- const BROWSER_MODE_SET = Object.freeze(['verify', 'debug', 'bench', 'diagnose']);
187
-
188
- const BROWSER_WORKLOAD_DISPATCH_MAP = Object.freeze({
189
- verify: Object.freeze({
190
- kernels: 'runKernelSuite',
191
- inference: 'runInferenceSuite',
192
- embedding: 'runEmbeddingSuite',
193
- rerank: 'runRerankSuite',
194
- training: 'runTrainingSuite',
195
- diffusion: 'runDiffusionSuite',
196
- energy: 'runEnergySuite',
197
- }),
198
- debug: Object.freeze({
199
- inference: 'runInferenceSuite(debug)',
200
- embedding: 'runEmbeddingSuite(debug)',
201
- }),
202
- diagnose: Object.freeze({
203
- inference: 'runInferenceSuite(diagnose)',
204
- }),
205
- bench: Object.freeze({
206
- inference: 'runBenchSuite',
207
- embedding: 'runBenchSuite',
208
- rerank: 'runBenchSuite',
209
- training: 'runBenchSuite(training)',
210
- diffusion: 'runBenchSuite(diffusion)',
211
- }),
212
- });
213
-
214
- function isPlainObject(value) {
215
- return Boolean(value) && typeof value === 'object' && !Array.isArray(value);
216
- }
217
-
218
- function resolveExecutionPlanCadence(executionPlan) {
219
- if (!isPlainObject(executionPlan)) {
220
- return null;
221
- }
222
- const finalActivePlanId = typeof executionPlan.finalActivePlanId === 'string'
223
- ? executionPlan.finalActivePlanId
224
- : null;
225
- for (const candidate of [executionPlan.primary, executionPlan.fallback]) {
226
- if (!isPlainObject(candidate)) {
227
- continue;
228
- }
229
- if (finalActivePlanId == null || candidate.id === finalActivePlanId) {
230
- return candidate;
231
- }
232
- }
233
- return isPlainObject(executionPlan.primary) ? executionPlan.primary : null;
234
- }
235
-
236
- function resolveDecodeCadence(runtimeConfig, executionPlan = null) {
237
- const inference = runtimeConfig?.inference;
238
- const batching = inference?.batching;
239
- const session = inference?.session;
240
- const decodeLoop = session?.decodeLoop;
241
- if (!isPlainObject(batching) || !isPlainObject(decodeLoop)) {
242
- return null;
243
- }
244
- const executionPlanCadence = resolveExecutionPlanCadence(executionPlan);
245
- const batchSize = executionPlanCadence?.batchSize ?? decodeLoop.batchSize ?? batching.batchSize ?? null;
246
- const readbackInterval = executionPlanCadence?.readbackInterval ?? decodeLoop.readbackInterval ?? batching.readbackInterval ?? null;
247
- const maxBatchDecodeTokens = executionPlanCadence?.maxBatchDecodeTokens
248
- ?? decodeLoop.maxBatchDecodeTokens
249
- ?? null;
250
- return {
251
- batchSize,
252
- readbackInterval,
253
- maxBatchDecodeTokens,
254
- stopCheckMode: executionPlanCadence?.stopCheckMode ?? decodeLoop.stopCheckMode ?? batching.stopCheckMode ?? null,
255
- readbackMode: executionPlanCadence?.readbackMode ?? decodeLoop.readbackMode ?? batching.readbackMode ?? null,
256
- disableCommandBatching: executionPlanCadence?.disableCommandBatching ?? decodeLoop.disableCommandBatching ?? null,
257
- disableMultiTokenDecode: inference?.generation?.disableMultiTokenDecode === true,
258
- speculationMode: session?.speculation?.mode ?? null,
259
- tokensPerReadback: Number.isFinite(batchSize) && Number.isFinite(readbackInterval)
260
- ? batchSize * readbackInterval
261
- : null,
262
- runtimeMirror: {
263
- batching: {
264
- batchSize: batching.batchSize ?? null,
265
- readbackInterval: batching.readbackInterval ?? null,
266
- stopCheckMode: batching.stopCheckMode ?? null,
267
- readbackMode: batching.readbackMode ?? null,
268
- },
269
- decodeLoop: {
270
- batchSize: decodeLoop.batchSize ?? null,
271
- readbackInterval: decodeLoop.readbackInterval ?? null,
272
- maxBatchDecodeTokens: decodeLoop.maxBatchDecodeTokens ?? null,
273
- stopCheckMode: decodeLoop.stopCheckMode ?? null,
274
- readbackMode: decodeLoop.readbackMode ?? null,
275
- ringTokens: decodeLoop.ringTokens ?? null,
276
- ringStop: decodeLoop.ringStop ?? null,
277
- ringStaging: decodeLoop.ringStaging ?? null,
278
- },
279
- },
280
- executionPlan: executionPlanCadence
281
- ? {
282
- id: executionPlanCadence.id ?? null,
283
- batchSize: executionPlanCadence.batchSize ?? null,
284
- readbackInterval: executionPlanCadence.readbackInterval ?? null,
285
- maxBatchDecodeTokens: executionPlanCadence.maxBatchDecodeTokens ?? null,
286
- stopCheckMode: executionPlanCadence.stopCheckMode ?? null,
287
- readbackMode: executionPlanCadence.readbackMode ?? null,
288
- disableCommandBatching: executionPlanCadence.disableCommandBatching ?? null,
289
- ringTokens: executionPlanCadence.ringTokens ?? null,
290
- ringStop: executionPlanCadence.ringStop ?? null,
291
- ringStaging: executionPlanCadence.ringStaging ?? null,
292
- }
293
- : null,
294
- };
295
- }
296
-
297
- export function getBrowserSupportedSuites() {
298
- return [...BROWSER_WORKLOAD_SET];
299
- }
300
-
301
- export function getBrowserSuiteDispatchMap() {
302
- return {
303
- verify: { ...BROWSER_WORKLOAD_DISPATCH_MAP.verify },
304
- debug: { ...BROWSER_WORKLOAD_DISPATCH_MAP.debug },
305
- bench: { ...BROWSER_WORKLOAD_DISPATCH_MAP.bench },
306
- };
307
- }
308
-
309
- function getAllowedWorkloadsForMode(mode) {
310
- return Object.keys(BROWSER_WORKLOAD_DISPATCH_MAP[mode] || {});
311
- }
312
-
313
- function createUnsupportedWorkloadError(requestedWorkload, context = {}) {
314
- const command = typeof context.command === 'string' && context.command.trim()
315
- ? context.command.trim()
316
- : 'run-browser-suite';
317
- const surface = typeof context.surface === 'string' && context.surface.trim()
318
- ? context.surface.trim()
319
- : 'browser';
320
- const mode = typeof context.mode === 'string' && context.mode.trim()
321
- ? context.mode.trim()
322
- : 'verify';
323
- const allowedWorkloads = getAllowedWorkloadsForMode(mode);
324
- const error = new Error(
325
- `Unsupported workload "${requestedWorkload}". Allowed workloads: ${allowedWorkloads.join(', ')}. ` +
326
- `command="${command}" mode="${mode}" surface="${surface}".`
327
- );
328
- error.code = 'unsupported_workload';
329
- error.requestedWorkload = requestedWorkload;
330
- error.allowedWorkloads = allowedWorkloads;
331
- error.command = command;
332
- error.mode = mode;
333
- error.surface = surface;
334
- error.details = {
335
- requestedWorkload,
336
- allowedWorkloads,
337
- command,
338
- mode,
339
- surface,
340
- };
341
- return error;
342
- }
343
-
344
- function resolveHarnessContext(options = {}) {
345
- const command = typeof options.command === 'string' ? options.command : null;
346
- const surface = typeof options.surface === 'string' ? options.surface : null;
347
- const mode = typeof options.mode === 'string' ? options.mode : null;
348
- return {
349
- command: command ?? 'run-browser-suite',
350
- mode: mode ?? command ?? 'verify',
351
- surface: surface ?? 'browser',
352
- };
353
- }
354
-
355
- function normalizeLegacySuite(value) {
356
- const suite = String(value || '').trim().toLowerCase();
357
- if (!suite) {
358
- return null;
359
- }
360
- return suite === 'benchmark' ? 'bench' : suite;
361
- }
362
-
363
- function normalizeMode(value, context = {}) {
364
- const mode = String(value || '').trim().toLowerCase();
365
- if (!mode) {
366
- return 'verify';
367
- }
368
- if (!BROWSER_MODE_SET.includes(mode)) {
369
- throw new Error(`Unsupported browser harness mode "${mode}" for command "${context.command || 'run-browser-suite'}".`);
370
- }
371
- return mode;
372
- }
373
-
374
- function resolveHarnessMode(options = {}, context = {}) {
375
- const explicitMode = options.mode ?? options.command ?? null;
376
- if (explicitMode) {
377
- return normalizeMode(explicitMode, context);
378
- }
379
- const legacySuite = normalizeLegacySuite(options.suite);
380
- if (legacySuite === 'debug' || legacySuite === 'bench') {
381
- return legacySuite;
382
- }
383
- return 'verify';
384
- }
385
-
386
- function normalizeWorkload(value, mode, context = {}) {
387
- const workload = String(value || '').trim().toLowerCase();
388
- if (!workload) {
389
- throw createUnsupportedWorkloadError(workload, { ...context, mode });
390
- }
391
- if (!getAllowedWorkloadsForMode(mode).includes(workload)) {
392
- throw createUnsupportedWorkloadError(workload, { ...context, mode });
393
- }
394
- return workload;
395
- }
396
-
397
- function resolveWorkload(options = {}, mode, context = {}) {
398
- if (options.workload) {
399
- return normalizeWorkload(options.workload, mode, context);
400
- }
401
- const legacySuite = normalizeLegacySuite(options.suite);
402
- if (legacySuite && legacySuite !== 'debug' && legacySuite !== 'bench') {
403
- return normalizeWorkload(legacySuite, mode, context);
404
- }
405
- if (mode === 'debug' || mode === 'bench') {
406
- return 'inference';
407
- }
408
- return normalizeWorkload('', mode, context);
409
- }
410
-
411
- function resolveDispatchSuite(mode, workload) {
412
- if (mode === 'debug') {
413
- return 'debug';
414
- }
415
- if (mode === 'bench') {
416
- return 'bench';
417
- }
418
- return workload;
419
- }
420
-
421
- function stableJson(value) {
422
- return JSON.stringify(stableSortObject(value)) ?? 'null';
423
- }
424
-
425
- function hashStableJson(value) {
426
- return `sha256:${sha256Hex(stableJson(value))}`;
427
- }
428
-
429
- export function resolveExecutionGraphHash(manifest) {
430
- const execution = manifest?.inference?.execution;
431
- if (!execution || typeof execution !== 'object') {
432
- return null;
433
- }
434
- return hashStableJson(execution);
435
- }
436
-
437
- function buildPerStepTokenProof(tokenIds) {
438
- return tokenIds.map((tokenId, index) => ({
439
- index,
440
- tokenId,
441
- tokenHash: hashStableJson({ index, tokenId }),
442
- }));
443
- }
444
-
445
- function buildKvCacheTranscriptSeed(kvCache, byteProof = null) {
446
- const source = kvCache && typeof kvCache === 'object' ? kvCache : null;
447
- const proof = byteProof && typeof byteProof === 'object' ? byteProof : null;
448
- const seed = {
449
- mode: proof ? 'stats+sha256-layer-kv-bytes' : (source ? 'stats' : 'not-captured'),
450
- layout: typeof source?.layout === 'string' ? source.layout : null,
451
- kvDtype: typeof source?.kvDtype === 'string' ? source.kvDtype : null,
452
- seqLen: Number.isFinite(source?.seqLen) ? source.seqLen : null,
453
- maxSeqLen: Number.isFinite(source?.maxSeqLen) ? source.maxSeqLen : null,
454
- usedBytes: Number.isFinite(source?.usedBytes) ? source.usedBytes : null,
455
- allocatedBytes: Number.isFinite(source?.allocatedBytes) ? source.allocatedBytes : null,
456
- counters: source?.counters ?? null,
457
- byteDigestMode: typeof proof?.mode === 'string' ? proof.mode : null,
458
- byteDigest: typeof proof?.digest === 'string' ? proof.digest : null,
459
- byteDigests: Array.isArray(proof?.layers) ? proof.layers : null,
460
- };
461
- return {
462
- ...seed,
463
- stateHash: hashStableJson(seed),
464
- };
465
- }
466
-
467
- export function buildReferenceTranscriptSeed(run, context = {}) {
468
- const promptPayload = run.promptInput ?? run.prompt ?? null;
469
- const outputText = typeof run.output === 'string' ? run.output : '';
470
- const tokenIds = Array.isArray(run.tokenIds)
471
- ? run.tokenIds.map((value) => Number(value)).filter((value) => Number.isInteger(value))
472
- : [];
473
- const promptTokenIds = Array.isArray(run.promptTokenIds)
474
- ? run.promptTokenIds.map((value) => Number(value)).filter((value) => Number.isInteger(value))
475
- : null;
476
- const logitsDigests = Array.isArray(run.logitsDigests)
477
- ? run.logitsDigests
478
- : [];
479
- const hasCompleteLogitsDigests = logitsDigests.length === tokenIds.length && tokenIds.length > 0;
480
- const transcript = {
481
- schema: 'doppler.reference-transcript/v1',
482
- source: {
483
- kind: 'inline-browser-suite',
484
- path: 'inline',
485
- hash: 'sha256:' + '0'.repeat(64),
486
- },
487
- executionGraphHash: context.executionGraphHash ?? null,
488
- surface: 'browser-webgpu',
489
- prompt: {
490
- identity: typeof run.prompt === 'string' && run.prompt.trim() ? run.prompt : 'promptInput',
491
- hash: hashStableJson(promptPayload),
492
- tokenIdsHash: promptTokenIds ? hashStableJson(promptTokenIds) : null,
493
- tokenCount: promptTokenIds ? promptTokenIds.length : null,
494
- },
495
- output: {
496
- textHash: `sha256:${sha256Hex(outputText)}`,
497
- tokensGenerated: tokenIds.length,
498
- stopReason: typeof run.phase?.stopReason === 'string' ? run.phase.stopReason : 'unknown',
499
- stopTokenId: Number.isInteger(run.phase?.stopTokenId) ? run.phase.stopTokenId : null,
500
- },
501
- tokens: {
502
- ids: tokenIds,
503
- generatedTokenIdsHash: hashStableJson(tokenIds),
504
- generatedTextHash: `sha256:${sha256Hex(outputText)}`,
505
- preview: Array.isArray(run.tokenDiagnostics?.preview) ? run.tokenDiagnostics.preview : [],
506
- perStep: buildPerStepTokenProof(tokenIds),
507
- coverage: {
508
- mode: 'full-token-ids',
509
- omitted: 0,
510
- },
511
- },
512
- phase: {
513
- prefillMs: Number.isFinite(run.phase?.prefillMs) ? run.phase.prefillMs : null,
514
- decodeMs: Number.isFinite(run.phase?.decodeMs) ? run.phase.decodeMs : null,
515
- prefillTokens: Number.isFinite(run.phase?.prefillTokens) ? run.phase.prefillTokens : null,
516
- decodeTokens: Number.isFinite(run.phase?.decodeTokens) ? run.phase.decodeTokens : null,
517
- },
518
- kvCache: buildKvCacheTranscriptSeed(
519
- run.phase?.kvCache ?? context.kvCache ?? null,
520
- run.kvCacheByteProof ?? null
521
- ),
522
- logits: hasCompleteLogitsDigests ? {
523
- mode: 'sha256-per-step',
524
- perStepDigests: logitsDigests.map((entry) => entry.digest),
525
- steps: logitsDigests,
526
- } : {
527
- mode: 'not-captured',
528
- reason: logitsDigests.length > 0
529
- ? 'Per-step logits digest count did not match generated token count.'
530
- : 'Per-step logits digests were not requested for this browser harness run.',
531
- perStepDigests: null,
532
- },
533
- tolerance: {
534
- tokenPolicy: 'exact generated token IDs',
535
- logitsPolicy: hasCompleteLogitsDigests
536
- ? 'exact sha256 digest per generated step over finalized f32 logits before sampling'
537
- : 'not captured',
538
- kvPolicy: run.kvCacheByteProof
539
- ? 'exact sha256 digest over used KV cache bytes by layer/key/value'
540
- : 'metadata hash only; KV tensor bytes are not read back by default',
541
- },
542
- };
543
- return {
544
- ...transcript,
545
- source: {
546
- ...transcript.source,
547
- hash: hashStableJson({
548
- prompt: transcript.prompt,
549
- output: transcript.output,
550
- tokens: {
551
- generatedTokenIdsHash: transcript.tokens.generatedTokenIdsHash,
552
- generatedTextHash: transcript.tokens.generatedTextHash,
553
- },
554
- phase: transcript.phase,
555
- kvCache: transcript.kvCache,
556
- logits: transcript.logits,
557
- }),
558
- },
559
- };
560
- }
561
-
562
- async function runKernelSuite(options = {}) {
563
- const startTime = performance.now();
564
- const { testHarness, initGPU } = await import('../../tests/kernels/browser/test-page.js');
565
- const { runKernelSuite: runAllKernelTests } = await import('../../tests/kernels/browser/kernel-suite.js');
566
- await initGPU();
567
-
568
- const previousKernelPath = getActiveKernelPath();
569
- const previousKernelSource = getActiveKernelPathSource();
570
- const previousKernelPathPolicy = getActiveKernelPathPolicy();
571
- if (options.modelId) {
572
- await resolveKernelPathForModel(options);
573
- }
574
- let results = [];
575
- try {
576
- results = await runAllKernelTests(testHarness);
577
- } finally {
578
- setActiveKernelPath(previousKernelPath, previousKernelSource, previousKernelPathPolicy);
579
- }
580
-
581
- const summary = buildSuiteSummary('kernels', results, startTime);
582
- return {
583
- ...summary,
584
- deviceInfo: resolveDeviceInfo(),
585
- };
586
- }
587
-
588
- async function runInferenceSuite(options = {}) {
589
- const startTime = performance.now();
590
- const cacheMode = normalizeCacheMode(options.cacheMode);
591
- const loadMode = normalizeLoadMode(options.loadMode, !!options.modelUrl, options.modelUrl);
592
- const harness = await withHarnessPhase(
593
- 'inference.initializeSuiteModel',
594
- {
595
- modelId: options.modelId ?? null,
596
- loadMode,
597
- cacheMode,
598
- },
599
- () => initializeSuiteModel(options)
600
- );
601
- const runtimeConfig = getRuntimeConfig();
602
- const modelType = harness.manifest?.modelType || 'transformer';
603
- const supportsEmbedding = modelSupportsEmbedding(harness.manifest);
604
- const supportsRerank = modelSupportsRerank(harness.manifest);
605
- if (options.expectedModelType === 'embedding' && !supportsEmbedding) {
606
- throw new Error(
607
- `Expected an embedding-capable model for workload "${options.workload || 'inference'}", got modelType="${modelType}". ` +
608
- `Set inference.supportsEmbedding=true in the manifest for text-generation models that should expose pipeline.embed().`
609
- );
610
- }
611
- if (options.expectedModelType === 'rerank' && !supportsRerank) {
612
- throw new Error(
613
- `Expected a rerank-capable model for workload "${options.workload || 'inference'}", got modelType="${modelType}". ` +
614
- 'Set inference.supportsRerank=true and inference.rerank in the manifest for models that should expose rerank scoring.'
615
- );
616
- }
617
- const safeModelLoadMs = toTimingNumber(harness.modelLoadMs, 0);
618
-
619
- let results;
620
- let output = null;
621
- let metrics;
622
-
623
- if (options.workload === 'rerank' && supportsRerank) {
624
- const run = await runRerank(harness.pipeline, runtimeConfig);
625
- const semantic = await runRerankSemanticChecks(harness.pipeline, options);
626
- const allScoresFinite = run.scores.every((entry) => (
627
- Number.isFinite(entry.score)
628
- && Number.isFinite(entry.probability)
629
- && Number.isFinite(entry.trueLogit)
630
- && Number.isFinite(entry.falseLogit)
631
- ));
632
- const hasRanking = Array.isArray(run.ranking) && run.ranking.length === run.documentCount;
633
- const isValidRerank = allScoresFinite && hasRanking && run.documentCount > 0;
634
- const isSemanticValid = semantic.passed;
635
- output = {
636
- mode: 'rerank',
637
- query: run.query,
638
- documentCount: run.documentCount,
639
- topDocument: run.topDocument,
640
- ranking: run.ranking,
641
- semantic: {
642
- passed: isSemanticValid,
643
- pairAcc: Number(semantic.pairAcc.toFixed(4)),
644
- failedCaseIds: semantic.failedCaseIds,
645
- details: {
646
- pairs: semantic.pairs,
647
- },
648
- },
649
- };
650
- results = [
651
- {
652
- name: 'rerank',
653
- passed: isValidRerank,
654
- duration: run.durationMs,
655
- error: isValidRerank
656
- ? undefined
657
- : 'Rerank scores must be finite and produce a full ranking.',
658
- },
659
- {
660
- name: 'rerank-semantic',
661
- passed: isSemanticValid,
662
- duration: semantic.durationMs,
663
- error: isSemanticValid
664
- ? undefined
665
- : (
666
- `Rerank semantic checks below threshold: pairs=${(semantic.pairAcc * 100).toFixed(1)}% `
667
- + `(min ${(semantic.minPairAcc * 100).toFixed(1)}%). `
668
- + (semantic.failedCaseIds.length > 0 ? `Failed: ${semantic.failedCaseIds.join(', ')}` : '')
669
- ),
670
- },
671
- ];
672
- metrics = {
673
- query: run.query,
674
- documentCount: run.documentCount,
675
- topDocumentIndex: run.topDocument?.index ?? null,
676
- topDocumentScore: run.topDocument?.score == null ? null : Number(run.topDocument.score.toFixed(6)),
677
- topDocumentProbability: run.topDocument?.probability == null ? null : Number(run.topDocument.probability.toFixed(6)),
678
- rerankMs: Number(run.durationMs.toFixed(2)),
679
- rerankRanking: run.ranking,
680
- semanticPassed: isSemanticValid,
681
- semanticDurationMs: Number(semantic.durationMs.toFixed(2)),
682
- semanticPairAcc: Number(semantic.pairAcc.toFixed(4)),
683
- semanticPairPassed: semantic.pairPassed,
684
- semanticPairTotal: semantic.pairTotal,
685
- semanticMinPairAcc: Number(semantic.minPairAcc.toFixed(4)),
686
- semanticMinScoreMargin: Number(semantic.minScoreMargin.toFixed(4)),
687
- semanticFailedCases: semantic.failedCaseIds,
688
- semanticDetails: {
689
- pairs: semantic.pairs,
690
- },
691
- modelLoadMs: safeModelLoadMs,
692
- endToEndMs: safeToFixed(safeModelLoadMs + run.durationMs),
693
- };
694
- } else if (modelType === 'embedding' || (options.workload === 'embedding' && supportsEmbedding)) {
695
- const run = await runEmbedding(harness.pipeline, runtimeConfig);
696
- const semantic = await runEmbeddingSemanticChecks(harness.pipeline, options);
697
- const isValidEmbedding = run.embeddingDim > 0 && run.nonFiniteCount === 0;
698
- const isSemanticValid = semantic.passed;
699
- output = {
700
- mode: 'embedding',
701
- tokens: run.tokenCount,
702
- embeddingDim: run.embeddingDim,
703
- finiteValues: run.finiteCount,
704
- nonFiniteValues: run.nonFiniteCount,
705
- finiteRatio: Number((run.finiteRatio ?? 0).toFixed(6)),
706
- min: run.min == null ? null : Number(run.min.toFixed(6)),
707
- max: run.max == null ? null : Number(run.max.toFixed(6)),
708
- maxAbs: run.maxAbs == null ? null : Number(run.maxAbs.toFixed(6)),
709
- mean: run.mean == null ? null : Number(run.mean.toFixed(6)),
710
- stdDev: run.stdDev == null ? null : Number(run.stdDev.toFixed(6)),
711
- l2Norm: run.l2Norm == null ? null : Number(run.l2Norm.toFixed(6)),
712
- preview: run.preview,
713
- semantic: {
714
- passed: isSemanticValid,
715
- style: semantic.style,
716
- retrievalTop1Acc: Number(semantic.retrievalTop1Acc.toFixed(4)),
717
- pairAcc: Number(semantic.pairAcc.toFixed(4)),
718
- failedCaseIds: semantic.failedCaseIds,
719
- details: {
720
- retrieval: semantic.retrieval,
721
- pairs: semantic.pairs,
722
- },
723
- },
724
- };
725
- results = [
726
- {
727
- name: 'embedding',
728
- passed: isValidEmbedding,
729
- duration: run.durationMs,
730
- error: isValidEmbedding
731
- ? undefined
732
- : (
733
- run.embeddingDim <= 0
734
- ? 'No embedding returned'
735
- : `Embedding contains non-finite values (${run.nonFiniteCount}/${run.embeddingDim})`
736
- ),
737
- },
738
- {
739
- name: 'embedding-semantic',
740
- passed: isSemanticValid,
741
- duration: semantic.durationMs,
742
- error: isSemanticValid
743
- ? undefined
744
- : (
745
- `Semantic checks below threshold: retrieval=${(semantic.retrievalTop1Acc * 100).toFixed(1)}% `
746
- + `(min ${(semantic.minRetrievalTop1Acc * 100).toFixed(1)}%), `
747
- + `pairs=${(semantic.pairAcc * 100).toFixed(1)}% `
748
- + `(min ${(semantic.minPairAcc * 100).toFixed(1)}%). `
749
- + (semantic.failedCaseIds.length > 0 ? `Failed: ${semantic.failedCaseIds.join(', ')}` : '')
750
- ),
751
- },
752
- ];
753
- metrics = {
754
- prompt: run.prompt,
755
- embeddingTokens: run.tokenCount,
756
- embeddingDim: run.embeddingDim,
757
- finiteValues: run.finiteCount,
758
- finiteRatio: Number((run.finiteRatio ?? 0).toFixed(6)),
759
- nonFiniteValues: run.nonFiniteCount,
760
- embeddingMin: run.min == null ? null : Number(run.min.toFixed(6)),
761
- embeddingMax: run.max == null ? null : Number(run.max.toFixed(6)),
762
- embeddingMaxAbs: run.maxAbs == null ? null : Number(run.maxAbs.toFixed(6)),
763
- embeddingMean: run.mean == null ? null : Number(run.mean.toFixed(6)),
764
- embeddingStdDev: run.stdDev == null ? null : Number(run.stdDev.toFixed(6)),
765
- embeddingL2Norm: run.l2Norm == null ? null : Number(run.l2Norm.toFixed(6)),
766
- embeddingMs: Number(run.durationMs.toFixed(2)),
767
- semanticPassed: isSemanticValid,
768
- semanticDurationMs: Number(semantic.durationMs.toFixed(2)),
769
- semanticRetrievalTop1Acc: Number(semantic.retrievalTop1Acc.toFixed(4)),
770
- semanticPairAcc: Number(semantic.pairAcc.toFixed(4)),
771
- semanticRetrievalPassed: semantic.retrievalPassed,
772
- semanticRetrievalTotal: semantic.retrievalTotal,
773
- semanticPairPassed: semantic.pairPassed,
774
- semanticPairTotal: semantic.pairTotal,
775
- semanticMinRetrievalTop1Acc: Number(semantic.minRetrievalTop1Acc.toFixed(4)),
776
- semanticMinPairAcc: Number(semantic.minPairAcc.toFixed(4)),
777
- semanticPairMarginThreshold: Number(semantic.pairMarginThreshold.toFixed(4)),
778
- semanticStyle: semantic.style,
779
- semanticFailedCases: semantic.failedCaseIds,
780
- semanticDetails: {
781
- retrieval: semantic.retrieval,
782
- pairs: semantic.pairs,
783
- },
784
- modelLoadMs: safeModelLoadMs,
785
- endToEndMs: safeToFixed(safeModelLoadMs + run.durationMs),
786
- embeddingPreview: run.preview,
787
- };
788
- } else {
789
- const run = await runTextInference(
790
- harness.pipeline,
791
- runtimeConfig,
792
- options.inferenceInput ?? null
793
- );
794
- const coherent = isCoherentOutput(run.tokens, run.output);
795
- results = [
796
- {
797
- name: 'generation',
798
- passed: run.tokens.length > 0 && coherent,
799
- duration: run.durationMs,
800
- error: run.tokens.length === 0
801
- ? 'No tokens generated'
802
- : (!coherent ? 'Output dominated by padding or special tokens' : undefined),
803
- },
804
- ];
805
- output = run.output;
806
- metrics = {
807
- prompt: run.prompt,
808
- maxTokens: run.maxTokens,
809
- tokensGenerated: run.tokens.length,
810
- tokensPerSec: safeToFixed(run.tokensPerSec),
811
- totalRunMs: safeToFixed(run.phase.totalMs),
812
- firstTokenMs: safeToFixed(run.phase.ttftMs),
813
- firstResponseMs: safeToFixed(safeModelLoadMs + run.phase.ttftMs),
814
- prefillMs: safeToFixed(run.phase.prefillMs),
815
- decodeMs: safeToFixed(run.phase.decodeMs),
816
- wallRunMs: safeToFixed(run.phase.wallMs ?? run.durationMs),
817
- prefillTokens: Math.round(run.phase.prefillTokens),
818
- decodeTokens: Math.round(run.phase.decodeTokens),
819
- stopReason: run.phase.stopReason ?? null,
820
- stopTokenId: Number.isInteger(run.phase.stopTokenId) ? run.phase.stopTokenId : null,
821
- prefillTokensPerSec: safeToFixed(run.phase.prefillTokensPerSec),
822
- prefillTokensPerSecTtft: safeToFixed(run.phase.prefillTokensPerSecTtft),
823
- decodeTokensPerSec: safeToFixed(run.phase.decodeTokensPerSec),
824
- modelLoadMs: safeModelLoadMs,
825
- gpu: run.phase.gpu,
826
- batching: run.phase.batching ?? null,
827
- plePreparedTokenCache: run.phase.plePreparedTokenCache ?? null,
828
- prefillProfileSteps: run.phase.prefillProfileSteps,
829
- decodeProfileSteps: run.phase.decodeProfileSteps,
830
- executionPlan: run.phase.executionPlan,
831
- kernelPathId: run.phase.kernelPathId,
832
- kernelPathSource: run.phase.kernelPathSource,
833
- generationDiagnostics: run.tokenDiagnostics,
834
- kvCache: run.phase.kvCache ?? null,
835
- referenceTranscript: buildReferenceTranscriptSeed(run, {
836
- executionGraphHash: resolveExecutionGraphHash(harness.manifest),
837
- kvCache: run.phase.kvCache ?? null,
838
- }),
839
- operatorDiagnostics: run.phase.operatorDiagnostics ?? null,
840
- };
841
- }
842
-
843
- const memoryStats = typeof harness.pipeline?.getMemoryStats === 'function'
844
- ? harness.pipeline.getMemoryStats()
845
- : null;
846
- const loadTimings = resolvePipelineLoadTimings(harness.pipeline);
847
- const loadDiagnostics = buildLoadTimingDiagnostics(
848
- safeModelLoadMs,
849
- loadTimings.loadTiming,
850
- loadTimings.pipelineLoadTiming
851
- );
852
- if (typeof harness.pipeline.unload === 'function' && !options.keepPipeline) {
853
- await harness.pipeline.unload();
854
- }
855
-
856
- const summary = buildSuiteSummary(options.suiteName || 'inference', results, startTime);
857
- const timing = buildCanonicalTiming({
858
- modelLoadMs: safeModelLoadMs,
859
- firstTokenMs: metrics.firstTokenMs ?? null,
860
- firstResponseMs: Number.isFinite(metrics.firstTokenMs)
861
- ? safeModelLoadMs + metrics.firstTokenMs
862
- : null,
863
- prefillMs: metrics.prefillMs ?? 0,
864
- decodeMs: metrics.decodeMs ?? 0,
865
- decodeMsPerTokenP50: metrics.decodeMsPerTokenP50 ?? null,
866
- decodeMsPerTokenP95: metrics.decodeMsPerTokenP95 ?? null,
867
- decodeMsPerTokenP99: metrics.decodeMsPerTokenP99 ?? null,
868
- totalRunMs: metrics.totalRunMs ?? metrics.decodeMs ?? 0,
869
- decodeTokensPerSec: metrics.decodeTokensPerSec,
870
- prefillTokensPerSec: metrics.prefillTokensPerSec,
871
- cacheMode,
872
- loadMode,
873
- });
874
- const timingDiagnostics = buildTimingDiagnostics(timing, {
875
- source: 'doppler',
876
- prefillSemantics: 'internal_prefill_phase',
877
- loadTiming: loadTimings.loadTiming,
878
- pipelineLoadTiming: loadTimings.pipelineLoadTiming,
879
- });
880
- const decodeBottleneck = buildDecodeBottleneckDiagnostics(metrics, timing);
881
- const metricsWithTimingDiagnostics = decodeBottleneck
882
- ? { ...metrics, decodeBottleneck }
883
- : metrics;
884
- if (decodeBottleneck) {
885
- timingDiagnostics.decodeBottleneck = decodeBottleneck;
886
- }
887
- const firstLoad = buildFirstLoadComposition({
888
- modelLoadMs: timing.modelLoadMs,
889
- firstTokenMs: timing.firstTokenMs,
890
- firstResponseMs: timing.firstResponseMs,
891
- });
892
- const metricsWithContracts = buildSuiteContractMetrics(
893
- options.suiteName || 'inference',
894
- loadDiagnostics
895
- ? { ...metricsWithTimingDiagnostics, load: loadDiagnostics }
896
- : metricsWithTimingDiagnostics,
897
- harness.manifest
898
- );
899
- return {
900
- ...summary,
901
- modelId: options.modelId || harness.manifest?.modelId || 'unknown',
902
- cacheMode,
903
- loadMode,
904
- env: {
905
- library: 'doppler',
906
- runtime: 'browser',
907
- device: 'webgpu',
908
- browserUserAgent: typeof navigator !== 'undefined' ? (navigator.userAgent || null) : null,
909
- browserPlatform: typeof navigator !== 'undefined' ? (navigator.platform || null) : null,
910
- browserLanguage: typeof navigator !== 'undefined' ? (navigator.language || null) : null,
911
- browserVendor: typeof navigator !== 'undefined' ? (navigator.vendor || null) : null,
912
- },
913
- timing,
914
- timingDiagnostics,
915
- firstLoad,
916
- output,
917
- metrics: metricsWithContracts,
918
- memoryStats,
919
- deviceInfo: resolveDeviceInfo(),
920
- pipeline: options.keepPipeline ? harness.pipeline : null,
921
- };
922
- }
923
-
924
- function resolveBenchmarkIterationSettings(runtimeConfig) {
925
- const benchConfig = runtimeConfig?.shared?.benchmark?.run || {};
926
- return {
927
- warmupRuns: Math.max(0, Math.floor(benchConfig.warmupRuns ?? 0)),
928
- timedRuns: Math.max(1, Math.floor(benchConfig.timedRuns ?? 1)),
929
- };
930
- }
931
-
932
- async function runBenchSuite(options = {}) {
933
- const startTime = performance.now();
934
- const runtimeConfig = getRuntimeConfig();
935
- const iterationSettings = resolveBenchmarkIterationSettings(runtimeConfig);
936
- const warmupRuns = iterationSettings.warmupRuns;
937
- const timedRuns = iterationSettings.timedRuns;
938
- const cacheMode = normalizeCacheMode(options.cacheMode);
939
- const loadMode = normalizeLoadMode(options.loadMode, !!options.modelUrl, options.modelUrl);
940
- const workloadType = normalizeWorkloadType(
941
- options.workloadType
942
- ?? (
943
- options.mode === 'bench' && (options.workload === 'training' || options.workload === 'diffusion')
944
- ? options.workload
945
- : null
946
- )
947
- );
948
-
949
- if (workloadType === 'training') {
950
- const trainingBench = await runTrainingBenchSuite({
951
- ...options,
952
- benchRun: iterationSettings,
953
- workloadType,
954
- });
955
- const trainingReport = trainingBench?.metrics?.trainingMetricsReport;
956
- if (Array.isArray(trainingReport) && trainingReport.length > 0) {
957
- validateTrainingMetricsReport(trainingReport);
958
- }
959
- const runStats = trainingBench?.metrics?.latency?.runMs || computeSampleStats([]);
960
- const stepStats = trainingBench?.metrics?.latency?.stepMs || computeSampleStats([]);
961
- const throughputStats = trainingBench?.metrics?.throughput?.stepsPerSec || computeSampleStats([]);
962
- const timing = buildCanonicalTiming({
963
- modelLoadMs: 0,
964
- firstTokenMs: null,
965
- firstResponseMs: null,
966
- prefillMs: null,
967
- decodeMs: stepStats.median,
968
- totalRunMs: runStats.median,
969
- decodeTokensPerSec: throughputStats.median,
970
- prefillTokensPerSec: null,
971
- cacheMode,
972
- loadMode,
973
- });
974
- const timingDiagnostics = buildTimingDiagnostics(timing, {
975
- source: 'doppler',
976
- prefillSemantics: 'not_applicable_training_workload',
977
- });
978
- const firstLoad = buildFirstLoadComposition({
979
- modelLoadMs: timing.modelLoadMs,
980
- firstTokenMs: timing.firstTokenMs,
981
- firstResponseMs: timing.firstResponseMs,
982
- });
983
- return {
984
- ...trainingBench,
985
- modelId: trainingBench.modelId || options.modelId || options.modelUrl || 'training',
986
- cacheMode,
987
- loadMode,
988
- env: {
989
- library: 'doppler',
990
- runtime: 'browser',
991
- device: 'webgpu',
992
- browserUserAgent: typeof navigator !== 'undefined' ? (navigator.userAgent || null) : null,
993
- browserPlatform: typeof navigator !== 'undefined' ? (navigator.platform || null) : null,
994
- browserLanguage: typeof navigator !== 'undefined' ? (navigator.language || null) : null,
995
- browserVendor: typeof navigator !== 'undefined' ? (navigator.vendor || null) : null,
996
- },
997
- timing,
998
- timingDiagnostics,
999
- firstLoad,
1000
- output: null,
1001
- memoryStats: null,
1002
- deviceInfo: trainingBench.deviceInfo ?? resolveDeviceInfo(),
1003
- pipeline: null,
1004
- };
1005
- }
1006
-
1007
- if (workloadType === 'diffusion' || workloadType === 'diffusion_gemma') {
1008
- const diffusionBench = await runDiffusionSuite({
1009
- ...options,
1010
- command: 'bench',
1011
- workload: 'diffusion',
1012
- captureOutput: options.captureOutput === true,
1013
- cacheMode,
1014
- loadMode,
1015
- });
1016
-
1017
- const benchResults = [
1018
- {
1019
- name: 'benchmark-diffusion',
1020
- passed: diffusionBench.passed > 0 && diffusionBench.failed === 0,
1021
- duration: diffusionBench.duration,
1022
- error: diffusionBench.failed === 0 ? undefined : 'Diffusion benchmark run failed.',
1023
- },
1024
- ];
1025
- const summary = buildSuiteSummary('bench', benchResults, startTime);
1026
-
1027
- return {
1028
- ...diffusionBench,
1029
- ...summary,
1030
- suite: 'bench',
1031
- results: benchResults,
1032
- metrics: {
1033
- ...(diffusionBench.metrics || {}),
1034
- workloadType,
1035
- },
1036
- };
1037
- }
1038
-
1039
- const harness = await withHarnessPhase(
1040
- 'bench.initializeSuiteModel',
1041
- {
1042
- modelId: options.modelId ?? null,
1043
- loadMode,
1044
- cacheMode,
1045
- },
1046
- () => initializeSuiteModel(options)
1047
- );
1048
- const benchRun = options.workload === 'rerank'
1049
- ? iterationSettings
1050
- : resolveBenchmarkRunSettings(runtimeConfig, harness.pipeline ?? harness);
1051
- const modelType = harness.manifest?.modelType || 'transformer';
1052
- const supportsEmbedding = modelSupportsEmbedding(harness.manifest);
1053
- const supportsRerank = modelSupportsRerank(harness.manifest);
1054
- if (options.expectedModelType === 'embedding' && !supportsEmbedding) {
1055
- throw new Error(
1056
- `Expected an embedding-capable model for bench workload "${options.workload || 'inference'}", got modelType="${modelType}". ` +
1057
- `Set inference.supportsEmbedding=true in the manifest for text-generation models that should expose pipeline.embed().`
1058
- );
1059
- }
1060
- if (options.expectedModelType === 'rerank' && !supportsRerank) {
1061
- throw new Error(
1062
- `Expected a rerank-capable model for bench workload "${options.workload || 'inference'}", got modelType="${modelType}". ` +
1063
- 'Set inference.supportsRerank=true and inference.rerank in the manifest for models that should expose rerank scoring.'
1064
- );
1065
- }
1066
- const safeModelLoadMs = toTimingNumber(harness.modelLoadMs, 0);
1067
-
1068
- let results;
1069
- let metrics;
1070
- let output = null;
1071
- let timing;
1072
-
1073
- if (options.workload === 'rerank' && supportsRerank) {
1074
- const durations = [];
1075
- const timedDurations = [];
1076
- const documentCounts = [];
1077
- const topDocumentScores = [];
1078
- const topDocumentProbabilities = [];
1079
- const rerankPhases = [];
1080
- const rerankDocumentPhases = [];
1081
- let invalidRuns = 0;
1082
- let nonFiniteScores = 0;
1083
- let lastRun = null;
1084
-
1085
- for (let i = 0; i < warmupRuns + timedRuns; i++) {
1086
- harness.pipeline.reset?.();
1087
- const run = await runRerank(harness.pipeline, runtimeConfig, {
1088
- ...benchRun,
1089
- benchmark: true,
1090
- });
1091
- if (i >= warmupRuns) {
1092
- timedDurations.push(run.durationMs);
1093
- const finiteScores = run.scores.filter((entry) => (
1094
- Number.isFinite(entry.score)
1095
- && Number.isFinite(entry.probability)
1096
- && Number.isFinite(entry.trueLogit)
1097
- && Number.isFinite(entry.falseLogit)
1098
- ));
1099
- nonFiniteScores += run.scores.length - finiteScores.length;
1100
- const hasRanking = Array.isArray(run.ranking) && run.ranking.length === run.documentCount;
1101
- if (finiteScores.length === run.scores.length && hasRanking && run.documentCount > 0) {
1102
- durations.push(run.durationMs);
1103
- if (run.phase && typeof run.phase === 'object') {
1104
- rerankPhases.push(run.phase);
1105
- if (Array.isArray(run.phase.documents)) {
1106
- for (const documentPhase of run.phase.documents) {
1107
- if (documentPhase?.phase && typeof documentPhase.phase === 'object') {
1108
- rerankDocumentPhases.push(documentPhase.phase);
1109
- }
1110
- }
1111
- }
1112
- }
1113
- documentCounts.push(run.documentCount);
1114
- if (Number.isFinite(run.topDocument?.score)) {
1115
- topDocumentScores.push(run.topDocument.score);
1116
- }
1117
- if (Number.isFinite(run.topDocument?.probability)) {
1118
- topDocumentProbabilities.push(run.topDocument.probability);
1119
- }
1120
- } else {
1121
- invalidRuns++;
1122
- }
1123
- lastRun = run;
1124
- }
1125
- }
1126
-
1127
- const semantic = await runRerankSemanticChecks(harness.pipeline, options);
1128
- const rerankMsStats = computeSampleStats(durations);
1129
- const timedRerankMsStats = computeSampleStats(timedDurations);
1130
- const documentCountStats = computeSampleStats(documentCounts);
1131
- const topScoreStats = computeSampleStats(topDocumentScores);
1132
- const topProbabilityStats = computeSampleStats(topDocumentProbabilities);
1133
- const avgMs = rerankMsStats.mean;
1134
- const semanticPassed = semantic.passed;
1135
- const rerankPrefixPhases = rerankPhases
1136
- .map((phase) => phase?.prefix)
1137
- .filter((phase) => phase && typeof phase === 'object');
1138
-
1139
- results = [
1140
- {
1141
- name: 'benchmark-rerank',
1142
- passed: durations.length > 0 && invalidRuns === 0,
1143
- duration: durations.reduce((sum, value) => sum + value, 0),
1144
- error: durations.length > 0
1145
- ? (
1146
- invalidRuns === 0
1147
- ? undefined
1148
- : `Invalid rerank runs: ${invalidRuns} (non-finite scores or incomplete ranking observed)`
1149
- )
1150
- : 'No valid rerank benchmark runs completed',
1151
- },
1152
- {
1153
- name: 'benchmark-rerank-semantic',
1154
- passed: semanticPassed,
1155
- duration: semantic.durationMs,
1156
- error: semanticPassed
1157
- ? undefined
1158
- : (
1159
- `Rerank semantic checks below threshold: pairs=${(semantic.pairAcc * 100).toFixed(1)}% `
1160
- + `(min ${(semantic.minPairAcc * 100).toFixed(1)}%). `
1161
- + (semantic.failedCaseIds.length > 0 ? `Failed: ${semantic.failedCaseIds.join(', ')}` : '')
1162
- ),
1163
- },
1164
- ];
1165
-
1166
- output = {
1167
- mode: 'rerank',
1168
- query: lastRun?.query ?? null,
1169
- documentCount: lastRun?.documentCount ?? null,
1170
- topDocument: lastRun?.topDocument ?? null,
1171
- ranking: lastRun?.ranking ?? [],
1172
- semantic: {
1173
- passed: semanticPassed,
1174
- pairAcc: Number(semantic.pairAcc.toFixed(4)),
1175
- failedCaseIds: semantic.failedCaseIds,
1176
- details: {
1177
- pairs: semantic.pairs,
1178
- },
1179
- },
1180
- };
1181
-
1182
- metrics = {
1183
- warmupRuns,
1184
- timedRuns,
1185
- validRuns: durations.length,
1186
- invalidRuns,
1187
- invalidRatePct: Number((timedRuns > 0 ? (invalidRuns / timedRuns) * 100 : 0).toFixed(2)),
1188
- query: lastRun?.query ?? null,
1189
- documentCount: Math.round(documentCountStats.mean),
1190
- topDocumentIndex: lastRun?.topDocument?.index ?? null,
1191
- topDocumentScore: lastRun?.topDocument?.score == null ? null : Number(lastRun.topDocument.score.toFixed(6)),
1192
- topDocumentProbability: lastRun?.topDocument?.probability == null ? null : Number(lastRun.topDocument.probability.toFixed(6)),
1193
- topDocumentScoreStats: topScoreStats,
1194
- topDocumentProbabilityStats: topProbabilityStats,
1195
- nonFiniteScores,
1196
- firstTimedRerankMs: Number((timedDurations[0] ?? 0).toFixed(2)),
1197
- minRerankMs: Number(rerankMsStats.min.toFixed(2)),
1198
- medianRerankMs: Number(rerankMsStats.median.toFixed(2)),
1199
- p95RerankMs: Number(rerankMsStats.p95.toFixed(2)),
1200
- p99RerankMs: Number(rerankMsStats.p99.toFixed(2)),
1201
- maxRerankMs: Number(rerankMsStats.max.toFixed(2)),
1202
- stdDevRerankMs: Number(rerankMsStats.stdDev.toFixed(2)),
1203
- ci95RerankMs: Number(rerankMsStats.ci95.toFixed(2)),
1204
- avgRerankMs: Number(avgMs.toFixed(2)),
1205
- avgReranksPerSec: Number((avgMs > 0 ? (1000 / avgMs) : 0).toFixed(2)),
1206
- semanticPassed,
1207
- semanticDurationMs: Number(semantic.durationMs.toFixed(2)),
1208
- semanticPairAcc: Number(semantic.pairAcc.toFixed(4)),
1209
- semanticPairPassed: semantic.pairPassed,
1210
- semanticPairTotal: semantic.pairTotal,
1211
- semanticMinPairAcc: Number(semantic.minPairAcc.toFixed(4)),
1212
- semanticMinScoreMargin: Number(semantic.minScoreMargin.toFixed(4)),
1213
- semanticFailedCases: semantic.failedCaseIds,
1214
- semanticDetails: {
1215
- pairs: semantic.pairs,
1216
- },
1217
- modelLoadMs: safeModelLoadMs,
1218
- latency: {
1219
- timedRerankMs: timedRerankMsStats,
1220
- rerankMs: rerankMsStats,
1221
- },
1222
- phase: summarizeTimingPhaseSamples(rerankPhases, [
1223
- 'totalMs',
1224
- 'prefixMs',
1225
- 'prefixTokens',
1226
- 'documentCount',
1227
- 'documentTotalMs',
1228
- 'maxDocumentMs',
1229
- 'avgDocumentMs',
1230
- 'prefix.totalMs',
1231
- 'prefix.prefillMs',
1232
- 'prefix.prefillRecordMs',
1233
- 'prefix.prefillRecordOps',
1234
- 'prefix.prefillRecordPasses',
1235
- 'prefix.prefillSubmitWaitMs',
1236
- ]),
1237
- prefixPhase: {
1238
- ...summarizeTimingPhaseSamples(rerankPrefixPhases, [
1239
- 'totalMs',
1240
- 'prefillMs',
1241
- 'prefillRecordMs',
1242
- 'prefillRecordOps',
1243
- 'prefillRecordPasses',
1244
- 'prefillSubmitWaitMs',
1245
- 'gpuPrefillMs',
1246
- 'tokens',
1247
- ]),
1248
- ...summarizePrefillRecordOps(rerankPrefixPhases),
1249
- },
1250
- documentPhase: {
1251
- ...summarizeTimingPhaseSamples(rerankDocumentPhases, [
1252
- 'totalMs',
1253
- 'prefillCallMs',
1254
- 'inputMs',
1255
- 'prefillMs',
1256
- 'prefillRecordMs',
1257
- 'prefillRecordOps',
1258
- 'prefillRecordPasses',
1259
- 'prefillSubmitWaitMs',
1260
- 'gpuPrefillMs',
1261
- 'tokens',
1262
- 'selectedTokenCount',
1263
- 'prefixTokens',
1264
- 'suffixTokens',
1265
- 'promptChars',
1266
- ]),
1267
- ...summarizePrefillRecordOps(rerankDocumentPhases),
1268
- },
1269
- documents: {
1270
- count: documentCountStats,
1271
- },
1272
- };
1273
-
1274
- timing = buildCanonicalTiming({
1275
- modelLoadMs: safeModelLoadMs,
1276
- firstTokenMs: null,
1277
- firstResponseMs: Number.isFinite(timedDurations[0])
1278
- ? safeModelLoadMs + timedDurations[0]
1279
- : null,
1280
- prefillMs: null,
1281
- decodeMs: null,
1282
- totalRunMs: rerankMsStats.median,
1283
- cacheMode,
1284
- loadMode,
1285
- });
1286
- } else if (modelType === 'embedding' || (options.workload === 'embedding' && supportsEmbedding)) {
1287
- const durations = [];
1288
- const timedDurations = [];
1289
- const embeddingDims = [];
1290
- const embeddingTokenCounts = [];
1291
- const embeddingNorms = [];
1292
- const embeddingPhases = [];
1293
- let firstTimedEmbeddingMs = null;
1294
- let invalidRuns = 0;
1295
- let totalNonFiniteValues = 0;
1296
- for (let i = 0; i < warmupRuns + timedRuns; i++) {
1297
- harness.pipeline.reset?.();
1298
- const run = await runEmbedding(harness.pipeline, runtimeConfig, {
1299
- ...benchRun,
1300
- benchmark: true,
1301
- });
1302
- if (i >= warmupRuns) {
1303
- timedDurations.push(run.durationMs);
1304
- if (firstTimedEmbeddingMs == null) {
1305
- firstTimedEmbeddingMs = run.durationMs;
1306
- }
1307
- totalNonFiniteValues += run.nonFiniteCount;
1308
- if (Number.isFinite(run.tokenCount)) {
1309
- embeddingTokenCounts.push(run.tokenCount);
1310
- }
1311
- if (Number.isFinite(run.l2Norm)) {
1312
- embeddingNorms.push(run.l2Norm);
1313
- }
1314
- if (run.embeddingDim > 0 && run.nonFiniteCount === 0) {
1315
- durations.push(run.durationMs);
1316
- embeddingDims.push(run.embeddingDim);
1317
- if (run.phase && typeof run.phase === 'object') {
1318
- embeddingPhases.push(run.phase);
1319
- }
1320
- } else {
1321
- invalidRuns++;
1322
- }
1323
- }
1324
- }
1325
-
1326
- const embeddingMsStats = computeSampleStats(durations);
1327
- const timedEmbeddingMsStats = computeSampleStats(timedDurations);
1328
- const embeddingDimStats = computeSampleStats(embeddingDims);
1329
- const embeddingTokensStats = computeSampleStats(embeddingTokenCounts);
1330
- const embeddingNormStats = computeSampleStats(embeddingNorms);
1331
- const avgMs = embeddingMsStats.mean;
1332
-
1333
- results = [
1334
- {
1335
- name: 'benchmark-embedding',
1336
- passed: durations.length > 0 && invalidRuns === 0,
1337
- duration: durations.reduce((sum, value) => sum + value, 0),
1338
- error: durations.length > 0
1339
- ? (
1340
- invalidRuns === 0
1341
- ? undefined
1342
- : `Invalid embedding runs: ${invalidRuns} (non-finite values observed)`
1343
- )
1344
- : 'No valid embedding benchmark runs completed',
1345
- },
1346
- ];
1347
-
1348
- metrics = {
1349
- warmupRuns,
1350
- timedRuns,
1351
- validRuns: durations.length,
1352
- invalidRuns,
1353
- invalidRatePct: Number((timedRuns > 0 ? (invalidRuns / timedRuns) * 100 : 0).toFixed(2)),
1354
- prompt: benchRun.promptLabel,
1355
- embeddingDim: Math.round(embeddingDims.reduce((a, b) => a + b, 0) / (embeddingDims.length || 1)),
1356
- nonFiniteValues: totalNonFiniteValues,
1357
- firstTimedEmbeddingMs: Number((firstTimedEmbeddingMs ?? 0).toFixed(2)),
1358
- minEmbeddingMs: Number(embeddingMsStats.min.toFixed(2)),
1359
- medianEmbeddingMs: Number(embeddingMsStats.median.toFixed(2)),
1360
- p95EmbeddingMs: Number(embeddingMsStats.p95.toFixed(2)),
1361
- p99EmbeddingMs: Number(embeddingMsStats.p99.toFixed(2)),
1362
- maxEmbeddingMs: Number(embeddingMsStats.max.toFixed(2)),
1363
- stdDevEmbeddingMs: Number(embeddingMsStats.stdDev.toFixed(2)),
1364
- ci95EmbeddingMs: Number(embeddingMsStats.ci95.toFixed(2)),
1365
- avgEmbeddingMs: Number(avgMs.toFixed(2)),
1366
- avgEmbeddingsPerSec: Number((avgMs > 0 ? (1000 / avgMs) : 0).toFixed(2)),
1367
- avgEmbeddingTokens: Number(embeddingTokensStats.mean.toFixed(2)),
1368
- avgEmbeddingL2Norm: Number(embeddingNormStats.mean.toFixed(4)),
1369
- modelLoadMs: safeModelLoadMs,
1370
- latency: {
1371
- timedEmbeddingMs: timedEmbeddingMsStats,
1372
- embeddingMs: embeddingMsStats,
1373
- },
1374
- phase: {
1375
- ...summarizeTimingPhaseSamples(embeddingPhases, [
1376
- 'totalMs',
1377
- 'inputMs',
1378
- 'prefillMs',
1379
- 'submitWaitMs',
1380
- 'readbackMs',
1381
- 'decodeHiddenMs',
1382
- 'finalNormMs',
1383
- 'extractMs',
1384
- 'hiddenBytes',
1385
- 'tokens',
1386
- 'prefillRecordMs',
1387
- 'prefillRecordOps',
1388
- 'prefillRecordPasses',
1389
- 'prefillSubmitWaitMs',
1390
- 'gpuPrefillMs',
1391
- ]),
1392
- ...summarizePrefillRecordOps(embeddingPhases),
1393
- },
1394
- dimensions: {
1395
- embedding: embeddingDimStats,
1396
- },
1397
- embedding: {
1398
- tokens: embeddingTokensStats,
1399
- l2Norm: embeddingNormStats,
1400
- },
1401
- };
1402
-
1403
- const timedStats = computeSampleStats(durations);
1404
- timing = buildCanonicalTiming({
1405
- modelLoadMs: safeModelLoadMs,
1406
- firstTokenMs: null,
1407
- firstResponseMs: Number.isFinite(firstTimedEmbeddingMs)
1408
- ? safeModelLoadMs + firstTimedEmbeddingMs
1409
- : null,
1410
- prefillMs: null,
1411
- decodeMs: null,
1412
- totalRunMs: timedStats.median,
1413
- cacheMode,
1414
- loadMode,
1415
- });
1416
- } else {
1417
- const tokensPerSec = [];
1418
- const durations = [];
1419
- const phaseTotals = [];
1420
- const tokensGenerated = [];
1421
- const decodeMsPerToken = [];
1422
- const ttftMs = [];
1423
- const prefillMs = [];
1424
- const decodeMs = [];
1425
- const prefillTokens = [];
1426
- const decodeTokens = [];
1427
- const decodeTokensPerSec = [];
1428
- const prefillTokensPerSec = [];
1429
- const prefillTokensPerSecTtft = [];
1430
- const gpuPrefillMs = [];
1431
- const gpuDecodeMs = [];
1432
- const gpuDecodeRecordMs = [];
1433
- const gpuDecodeRecordOps = [];
1434
- const gpuDecodeRecordPasses = [];
1435
- const gpuDecodeRecordMsPerOp = [];
1436
- const gpuDecodeRecordMsPerPass = [];
1437
- const gpuDecodeRecordPassesPerOp = [];
1438
- const gpuDecodeRecordMsPerExecutedBatchToken = [];
1439
- const gpuDecodeRecordOpsPerExecutedBatchToken = [];
1440
- const gpuDecodeRecordPassesPerExecutedBatchToken = [];
1441
- const gpuDecodeRecordOpLabels = {};
1442
- let hasGpuDecodeRecordOpLabels = false;
1443
- const gpuDecodeSubmitWaitMs = [];
1444
- const gpuDecodeReadbackWaitMs = [];
1445
- const gpuDecodeReadbackMapWaitMs = [];
1446
- const gpuDecodeReadbackCleanupMs = [];
1447
- const gpuDecodeReadbackCopyMs = [];
1448
- const gpuDecodeOrchestrationMs = [];
1449
- const gpuPrefillRecordMs = [];
1450
- const gpuPrefillRecordOps = [];
1451
- const gpuPrefillRecordPasses = [];
1452
- const gpuPrefillRecordOpLabels = {};
1453
- let hasGpuPrefillRecordOpLabels = false;
1454
- const gpuPrefillSubmitWaitMs = [];
1455
- const singleTokenSubmitWaitMs = [];
1456
- const singleTokenReadbackWaitMs = [];
1457
- const singleTokenReadbackMapWaitMs = [];
1458
- const singleTokenReadbackCleanupMs = [];
1459
- const singleTokenReadbackCopyMs = [];
1460
- const singleTokenOrchestrationMs = [];
1461
- const batchedForwardCalls = [];
1462
- const unbatchedForwardCalls = [];
1463
- const totalBatchedTimeMs = [];
1464
- const totalUnbatchedTimeMs = [];
1465
- const gpuSubmissions = [];
1466
- const requestedBatchTokens = [];
1467
- const effectiveBatchTokens = [];
1468
- const executedBatchTokens = [];
1469
- const resolvedBatchTokens = [];
1470
- const maxBatchTokenCap = [];
1471
- const batchClampCount = [];
1472
- const plePreparedTokenCacheHits = [];
1473
- const plePreparedTokenCacheMisses = [];
1474
- const plePreparedTokenCacheEntries = [];
1475
- const plePreparedTokenCacheBytes = [];
1476
-
1477
- let generatedText = null;
1478
- let generatedPromptInput = null;
1479
- let generatedReferenceTranscript = null;
1480
- let lastPromptLabel = benchRun.promptLabel;
1481
- let lastMaxTokens = benchRun.maxTokens;
1482
- let lastDecodeMode = null;
1483
- let lastBatchGuardReason = null;
1484
- let lastExecutionPlan = null;
1485
- let lastGpuUniformCache = null;
1486
- for (let i = 0; i < warmupRuns + timedRuns; i++) {
1487
- harness.pipeline.reset?.();
1488
- const run = await withHarnessPhase(
1489
- `bench.runTextInference[${i}]`,
1490
- {
1491
- modelId: options.modelId ?? harness.manifest?.modelId ?? null,
1492
- loadMode,
1493
- cacheMode,
1494
- warmupRuns,
1495
- timedRuns,
1496
- },
1497
- () => runTextInference(harness.pipeline, runtimeConfig, {
1498
- ...benchRun,
1499
- benchmark: true,
1500
- ...(options.inferenceInput ?? {}),
1501
- })
1502
- );
1503
- if (i === warmupRuns + timedRuns - 1) {
1504
- generatedText = run?.output ?? null;
1505
- generatedPromptInput = run?.promptInput ?? null;
1506
- generatedReferenceTranscript = buildReferenceTranscriptSeed(run, {
1507
- executionGraphHash: resolveExecutionGraphHash(harness.manifest),
1508
- kvCache: run?.phase?.kvCache ?? null,
1509
- });
1510
- lastPromptLabel = run?.prompt ?? benchRun.promptLabel;
1511
- lastMaxTokens = Number.isFinite(run?.maxTokens) ? run.maxTokens : benchRun.maxTokens;
1512
- lastDecodeMode = run?.phase?.decodeMode ?? null;
1513
- lastBatchGuardReason = run?.phase?.batchGuardReason ?? null;
1514
- lastExecutionPlan = run?.phase?.executionPlan ?? null;
1515
- }
1516
- if (i >= warmupRuns) {
1517
- const phase = run?.phase ?? {};
1518
- const phaseTokens = Array.isArray(run?.tokens) ? run.tokens : [];
1519
- const phaseGpu = phase.gpu;
1520
- const phaseBatching = phase.batching;
1521
- const phasePlePreparedTokenCache = phase.plePreparedTokenCache;
1522
- tokensPerSec.push(run?.tokensPerSec);
1523
- durations.push(run?.durationMs);
1524
- phaseTotals.push(phase.totalMs);
1525
- tokensGenerated.push(phaseTokens.length);
1526
- ttftMs.push(phase.ttftMs);
1527
- prefillMs.push(phase.prefillMs);
1528
- decodeMs.push(phase.decodeMs);
1529
- prefillTokens.push(phase.prefillTokens);
1530
- decodeTokens.push(phase.decodeTokens);
1531
- decodeTokensPerSec.push(phase.decodeTokensPerSec);
1532
- prefillTokensPerSec.push(phase.prefillTokensPerSec);
1533
- prefillTokensPerSecTtft.push(phase.prefillTokensPerSecTtft);
1534
- if (phase.decodeMs > 0 && phase.decodeTokens > 0) {
1535
- decodeMsPerToken.push(phase.decodeMs / phase.decodeTokens);
1536
- }
1537
- const phaseGpuUniformCache = normalizeUniformCacheStats(phaseGpu?.uniformCache);
1538
- if (phaseGpuUniformCache) {
1539
- lastGpuUniformCache = phaseGpuUniformCache;
1540
- }
1541
- if (Number.isFinite(phaseGpu?.prefillMs)) gpuPrefillMs.push(phaseGpu.prefillMs);
1542
- if (Number.isFinite(phaseGpu?.decodeMs)) gpuDecodeMs.push(phaseGpu.decodeMs);
1543
- if (Number.isFinite(phaseGpu?.decodeRecordMs)) gpuDecodeRecordMs.push(phaseGpu.decodeRecordMs);
1544
- if (Number.isFinite(phaseGpu?.decodeRecordOps)) gpuDecodeRecordOps.push(phaseGpu.decodeRecordOps);
1545
- if (Number.isFinite(phaseGpu?.decodeRecordPasses)) gpuDecodeRecordPasses.push(phaseGpu.decodeRecordPasses);
1546
- const phaseDecodeRecordOpLabels = normalizeDecodeRecordOpLabels(phaseGpu?.decodeRecordOpLabels);
1547
- if (phaseDecodeRecordOpLabels) {
1548
- hasGpuDecodeRecordOpLabels = true;
1549
- for (const [label, count] of Object.entries(phaseDecodeRecordOpLabels)) {
1550
- gpuDecodeRecordOpLabels[label] = (gpuDecodeRecordOpLabels[label] ?? 0) + count;
1551
- }
1552
- }
1553
- if (Number.isFinite(phaseGpu?.decodeRecordMsPerOp)) {
1554
- gpuDecodeRecordMsPerOp.push(phaseGpu.decodeRecordMsPerOp);
1555
- }
1556
- if (Number.isFinite(phaseGpu?.decodeRecordMsPerPass)) {
1557
- gpuDecodeRecordMsPerPass.push(phaseGpu.decodeRecordMsPerPass);
1558
- }
1559
- if (Number.isFinite(phaseGpu?.decodeRecordPassesPerOp)) {
1560
- gpuDecodeRecordPassesPerOp.push(phaseGpu.decodeRecordPassesPerOp);
1561
- }
1562
- if (Number.isFinite(phaseGpu?.decodeRecordMsPerExecutedBatchToken)) {
1563
- gpuDecodeRecordMsPerExecutedBatchToken.push(phaseGpu.decodeRecordMsPerExecutedBatchToken);
1564
- }
1565
- if (Number.isFinite(phaseGpu?.decodeRecordOpsPerExecutedBatchToken)) {
1566
- gpuDecodeRecordOpsPerExecutedBatchToken.push(phaseGpu.decodeRecordOpsPerExecutedBatchToken);
1567
- }
1568
- if (Number.isFinite(phaseGpu?.decodeRecordPassesPerExecutedBatchToken)) {
1569
- gpuDecodeRecordPassesPerExecutedBatchToken.push(phaseGpu.decodeRecordPassesPerExecutedBatchToken);
1570
- }
1571
- if (Number.isFinite(phaseGpu?.decodeSubmitWaitMs)) gpuDecodeSubmitWaitMs.push(phaseGpu.decodeSubmitWaitMs);
1572
- if (Number.isFinite(phaseGpu?.decodeReadbackWaitMs)) gpuDecodeReadbackWaitMs.push(phaseGpu.decodeReadbackWaitMs);
1573
- if (Number.isFinite(phaseGpu?.decodeReadbackMapWaitMs)) gpuDecodeReadbackMapWaitMs.push(phaseGpu.decodeReadbackMapWaitMs);
1574
- if (Number.isFinite(phaseGpu?.decodeReadbackCleanupMs)) gpuDecodeReadbackCleanupMs.push(phaseGpu.decodeReadbackCleanupMs);
1575
- if (Number.isFinite(phaseGpu?.decodeReadbackCopyMs)) gpuDecodeReadbackCopyMs.push(phaseGpu.decodeReadbackCopyMs);
1576
- if (Number.isFinite(phaseGpu?.prefillRecordMs)) gpuPrefillRecordMs.push(phaseGpu.prefillRecordMs);
1577
- if (Number.isFinite(phaseGpu?.prefillRecordOps)) gpuPrefillRecordOps.push(phaseGpu.prefillRecordOps);
1578
- if (Number.isFinite(phaseGpu?.prefillRecordPasses)) gpuPrefillRecordPasses.push(phaseGpu.prefillRecordPasses);
1579
- const phasePrefillRecordOpLabels = normalizeDecodeRecordOpLabels(phaseGpu?.prefillRecordOpLabels);
1580
- if (phasePrefillRecordOpLabels) {
1581
- hasGpuPrefillRecordOpLabels = true;
1582
- for (const [label, count] of Object.entries(phasePrefillRecordOpLabels)) {
1583
- gpuPrefillRecordOpLabels[label] = (gpuPrefillRecordOpLabels[label] ?? 0) + count;
1584
- }
1585
- }
1586
- if (Number.isFinite(phaseGpu?.prefillSubmitWaitMs)) gpuPrefillSubmitWaitMs.push(phaseGpu.prefillSubmitWaitMs);
1587
- if (Number.isFinite(phaseGpu?.decodeOrchestrationMs)) {
1588
- gpuDecodeOrchestrationMs.push(phaseGpu.decodeOrchestrationMs);
1589
- }
1590
- if (Number.isFinite(phaseGpu?.singleTokenSubmitWaitMs)) singleTokenSubmitWaitMs.push(phaseGpu.singleTokenSubmitWaitMs);
1591
- if (Number.isFinite(phaseGpu?.singleTokenReadbackWaitMs)) singleTokenReadbackWaitMs.push(phaseGpu.singleTokenReadbackWaitMs);
1592
- if (Number.isFinite(phaseGpu?.singleTokenReadbackMapWaitMs)) singleTokenReadbackMapWaitMs.push(phaseGpu.singleTokenReadbackMapWaitMs);
1593
- if (Number.isFinite(phaseGpu?.singleTokenReadbackCleanupMs)) singleTokenReadbackCleanupMs.push(phaseGpu.singleTokenReadbackCleanupMs);
1594
- if (Number.isFinite(phaseGpu?.singleTokenReadbackCopyMs)) singleTokenReadbackCopyMs.push(phaseGpu.singleTokenReadbackCopyMs);
1595
- if (Number.isFinite(phaseGpu?.singleTokenOrchestrationMs)) singleTokenOrchestrationMs.push(phaseGpu.singleTokenOrchestrationMs);
1596
- if (Number.isFinite(phaseBatching?.batchedForwardCalls)) batchedForwardCalls.push(phaseBatching.batchedForwardCalls);
1597
- if (Number.isFinite(phaseBatching?.unbatchedForwardCalls)) unbatchedForwardCalls.push(phaseBatching.unbatchedForwardCalls);
1598
- if (Number.isFinite(phaseBatching?.totalBatchedTimeMs)) totalBatchedTimeMs.push(phaseBatching.totalBatchedTimeMs);
1599
- if (Number.isFinite(phaseBatching?.totalUnbatchedTimeMs)) totalUnbatchedTimeMs.push(phaseBatching.totalUnbatchedTimeMs);
1600
- if (Number.isFinite(phaseBatching?.gpuSubmissions)) gpuSubmissions.push(phaseBatching.gpuSubmissions);
1601
- if (Number.isFinite(phaseBatching?.requestedBatchTokens)) requestedBatchTokens.push(phaseBatching.requestedBatchTokens);
1602
- if (Number.isFinite(phaseBatching?.effectiveBatchTokens)) effectiveBatchTokens.push(phaseBatching.effectiveBatchTokens);
1603
- if (Number.isFinite(phaseBatching?.executedBatchTokens)) executedBatchTokens.push(phaseBatching.executedBatchTokens);
1604
- if (Number.isFinite(phaseBatching?.resolvedBatchTokens)) resolvedBatchTokens.push(phaseBatching.resolvedBatchTokens);
1605
- if (Number.isFinite(phaseBatching?.maxBatchTokenCap)) maxBatchTokenCap.push(phaseBatching.maxBatchTokenCap);
1606
- if (Number.isFinite(phaseBatching?.batchClampCount)) batchClampCount.push(phaseBatching.batchClampCount);
1607
- if (Number.isFinite(phasePlePreparedTokenCache?.hits)) plePreparedTokenCacheHits.push(phasePlePreparedTokenCache.hits);
1608
- if (Number.isFinite(phasePlePreparedTokenCache?.misses)) plePreparedTokenCacheMisses.push(phasePlePreparedTokenCache.misses);
1609
- if (Number.isFinite(phasePlePreparedTokenCache?.entries)) plePreparedTokenCacheEntries.push(phasePlePreparedTokenCache.entries);
1610
- if (Number.isFinite(phasePlePreparedTokenCache?.bytes)) plePreparedTokenCacheBytes.push(phasePlePreparedTokenCache.bytes);
1611
- }
1612
- }
1613
-
1614
- const totalMsStats = computeSampleStats(phaseTotals);
1615
- const wallRunMsStats = computeSampleStats(durations);
1616
- const tokensPerSecStats = computeSampleStats(tokensPerSec);
1617
- const decodeTokensPerSecStats = computeSampleStats(decodeTokensPerSec);
1618
- const prefillTokensPerSecStats = computeSampleStats(prefillTokensPerSec);
1619
- const prefillTokensPerSecTtftStats = computeSampleStats(prefillTokensPerSecTtft);
1620
- const decodeMsPerTokenStats = computeSampleStats(decodeMsPerToken);
1621
- const ttftMsStats = computeSampleStats(ttftMs);
1622
- const prefillMsStats = computeSampleStats(prefillMs);
1623
- const decodeMsStats = computeSampleStats(decodeMs);
1624
- const tokensGeneratedStats = computeSampleStats(tokensGenerated);
1625
- const prefillTokensStats = computeSampleStats(prefillTokens);
1626
- const decodeTokensStats = computeSampleStats(decodeTokens);
1627
- const gpuDecodeRecordOpsStats = computeSampleStats(gpuDecodeRecordOps);
1628
- const gpuDecodeRecordPassesStats = computeSampleStats(gpuDecodeRecordPasses);
1629
- const gpuPrefillRecordOpsStats = computeSampleStats(gpuPrefillRecordOps);
1630
- const gpuPrefillRecordPassesStats = computeSampleStats(gpuPrefillRecordPasses);
1631
- const gpuDecodeRecordOpLabelSampleCount = gpuDecodeRecordOps.length > 0
1632
- ? gpuDecodeRecordOps.length
1633
- : 1;
1634
- const gpuDecodeRecordMeanOpLabels = {};
1635
- if (hasGpuDecodeRecordOpLabels) {
1636
- for (const [label, count] of Object.entries(gpuDecodeRecordOpLabels)) {
1637
- gpuDecodeRecordMeanOpLabels[label] = count / gpuDecodeRecordOpLabelSampleCount;
1638
- }
1639
- }
1640
- const gpuPrefillRecordOpLabelSampleCount = gpuPrefillRecordOps.length > 0
1641
- ? gpuPrefillRecordOps.length
1642
- : 1;
1643
- const gpuPrefillRecordMeanOpLabels = {};
1644
- if (hasGpuPrefillRecordOpLabels) {
1645
- for (const [label, count] of Object.entries(gpuPrefillRecordOpLabels)) {
1646
- gpuPrefillRecordMeanOpLabels[label] = count / gpuPrefillRecordOpLabelSampleCount;
1647
- }
1648
- }
1649
- const hasGpuStats = gpuPrefillMs.length > 0 || gpuDecodeMs.length > 0 || gpuDecodeRecordMs.length > 0
1650
- || gpuDecodeRecordOps.length > 0 || gpuDecodeRecordPasses.length > 0
1651
- || gpuPrefillRecordOps.length > 0 || gpuPrefillRecordPasses.length > 0
1652
- || gpuDecodeRecordMsPerOp.length > 0 || gpuDecodeRecordMsPerPass.length > 0
1653
- || gpuDecodeRecordPassesPerOp.length > 0
1654
- || gpuDecodeRecordMsPerExecutedBatchToken.length > 0
1655
- || gpuDecodeRecordOpsPerExecutedBatchToken.length > 0
1656
- || gpuDecodeRecordPassesPerExecutedBatchToken.length > 0
1657
- || hasGpuDecodeRecordOpLabels
1658
- || hasGpuPrefillRecordOpLabels
1659
- || gpuDecodeSubmitWaitMs.length > 0 || gpuDecodeReadbackWaitMs.length > 0
1660
- || gpuDecodeReadbackMapWaitMs.length > 0 || gpuDecodeReadbackCleanupMs.length > 0
1661
- || gpuDecodeReadbackCopyMs.length > 0
1662
- || gpuDecodeOrchestrationMs.length > 0
1663
- || lastGpuUniformCache
1664
- || singleTokenSubmitWaitMs.length > 0 || singleTokenReadbackWaitMs.length > 0
1665
- || singleTokenReadbackMapWaitMs.length > 0 || singleTokenReadbackCleanupMs.length > 0
1666
- || singleTokenReadbackCopyMs.length > 0
1667
- || singleTokenOrchestrationMs.length > 0;
1668
- const gpuPhaseStats = hasGpuStats
1669
- ? {
1670
- prefillMs: computeSampleStats(gpuPrefillMs),
1671
- decodeMs: computeSampleStats(gpuDecodeMs),
1672
- decodeRecordMs: computeSampleStats(gpuDecodeRecordMs),
1673
- decodeRecordOps: gpuDecodeRecordOpsStats,
1674
- decodeRecordPasses: gpuDecodeRecordPassesStats,
1675
- decodeRecordMsPerOp: computeSampleStats(gpuDecodeRecordMsPerOp),
1676
- decodeRecordMsPerPass: computeSampleStats(gpuDecodeRecordMsPerPass),
1677
- decodeRecordPassesPerOp: computeSampleStats(gpuDecodeRecordPassesPerOp),
1678
- decodeRecordMsPerExecutedBatchToken: computeSampleStats(gpuDecodeRecordMsPerExecutedBatchToken),
1679
- decodeRecordOpsPerExecutedBatchToken: computeSampleStats(gpuDecodeRecordOpsPerExecutedBatchToken),
1680
- decodeRecordPassesPerExecutedBatchToken: computeSampleStats(gpuDecodeRecordPassesPerExecutedBatchToken),
1681
- decodeRecordUniqueOpLabels: hasGpuDecodeRecordOpLabels ? Object.keys(gpuDecodeRecordOpLabels).length : null,
1682
- decodeRecordTopOps: hasGpuDecodeRecordOpLabels
1683
- ? buildDecodeRecordTopOps(
1684
- gpuDecodeRecordMeanOpLabels,
1685
- gpuDecodeRecordOpsStats?.mean
1686
- )
1687
- : [],
1688
- decodeRecordTopOpGroups: hasGpuDecodeRecordOpLabels
1689
- ? buildDecodeRecordTopOpGroups(
1690
- gpuDecodeRecordMeanOpLabels,
1691
- gpuDecodeRecordOpsStats?.mean
1692
- )
1693
- : [],
1694
- decodeSubmitWaitMs: computeSampleStats(gpuDecodeSubmitWaitMs),
1695
- decodeReadbackWaitMs: computeSampleStats(gpuDecodeReadbackWaitMs),
1696
- decodeReadbackMapWaitMs: computeSampleStats(gpuDecodeReadbackMapWaitMs),
1697
- decodeReadbackCleanupMs: computeSampleStats(gpuDecodeReadbackCleanupMs),
1698
- decodeReadbackCopyMs: computeSampleStats(gpuDecodeReadbackCopyMs),
1699
- decodeOrchestrationMs: computeSampleStats(gpuDecodeOrchestrationMs),
1700
- prefillRecordMs: computeSampleStats(gpuPrefillRecordMs),
1701
- prefillRecordOps: gpuPrefillRecordOpsStats,
1702
- prefillRecordPasses: gpuPrefillRecordPassesStats,
1703
- prefillRecordUniqueOpLabels: hasGpuPrefillRecordOpLabels ? Object.keys(gpuPrefillRecordOpLabels).length : null,
1704
- prefillRecordTopOps: hasGpuPrefillRecordOpLabels
1705
- ? buildDecodeRecordTopOps(
1706
- gpuPrefillRecordMeanOpLabels,
1707
- gpuPrefillRecordOpsStats?.mean
1708
- )
1709
- : [],
1710
- prefillRecordTopOpGroups: hasGpuPrefillRecordOpLabels
1711
- ? buildDecodeRecordTopOpGroups(
1712
- gpuPrefillRecordMeanOpLabels,
1713
- gpuPrefillRecordOpsStats?.mean
1714
- )
1715
- : [],
1716
- prefillSubmitWaitMs: computeSampleStats(gpuPrefillSubmitWaitMs),
1717
- uniformCache: lastGpuUniformCache,
1718
- singleTokenSubmitWaitMs: computeSampleStats(singleTokenSubmitWaitMs),
1719
- singleTokenReadbackWaitMs: computeSampleStats(singleTokenReadbackWaitMs),
1720
- singleTokenReadbackMapWaitMs: computeSampleStats(singleTokenReadbackMapWaitMs),
1721
- singleTokenReadbackCleanupMs: computeSampleStats(singleTokenReadbackCleanupMs),
1722
- singleTokenReadbackCopyMs: computeSampleStats(singleTokenReadbackCopyMs),
1723
- singleTokenOrchestrationMs: computeSampleStats(singleTokenOrchestrationMs),
1724
- }
1725
- : null;
1726
- const hasBatchingStats = batchedForwardCalls.length > 0
1727
- || unbatchedForwardCalls.length > 0
1728
- || totalBatchedTimeMs.length > 0
1729
- || totalUnbatchedTimeMs.length > 0
1730
- || gpuSubmissions.length > 0
1731
- || requestedBatchTokens.length > 0
1732
- || effectiveBatchTokens.length > 0
1733
- || executedBatchTokens.length > 0
1734
- || resolvedBatchTokens.length > 0
1735
- || maxBatchTokenCap.length > 0
1736
- || batchClampCount.length > 0;
1737
- const batchingPhaseStats = hasBatchingStats
1738
- ? {
1739
- batchedForwardCalls: computeSampleStats(batchedForwardCalls),
1740
- unbatchedForwardCalls: computeSampleStats(unbatchedForwardCalls),
1741
- totalBatchedTimeMs: computeSampleStats(totalBatchedTimeMs),
1742
- totalUnbatchedTimeMs: computeSampleStats(totalUnbatchedTimeMs),
1743
- gpuSubmissions: computeSampleStats(gpuSubmissions),
1744
- requestedBatchTokens: computeSampleStats(requestedBatchTokens),
1745
- effectiveBatchTokens: computeSampleStats(effectiveBatchTokens),
1746
- executedBatchTokens: computeSampleStats(executedBatchTokens),
1747
- resolvedBatchTokens: computeSampleStats(resolvedBatchTokens),
1748
- maxBatchTokenCap: computeSampleStats(maxBatchTokenCap),
1749
- batchClampCount: computeSampleStats(batchClampCount),
1750
- }
1751
- : null;
1752
- const hasPlePreparedTokenCacheStats = plePreparedTokenCacheHits.length > 0
1753
- || plePreparedTokenCacheMisses.length > 0
1754
- || plePreparedTokenCacheEntries.length > 0
1755
- || plePreparedTokenCacheBytes.length > 0;
1756
- const plePreparedTokenCacheStats = hasPlePreparedTokenCacheStats
1757
- ? {
1758
- hits: computeSampleStats(plePreparedTokenCacheHits),
1759
- misses: computeSampleStats(plePreparedTokenCacheMisses),
1760
- entries: computeSampleStats(plePreparedTokenCacheEntries),
1761
- bytes: computeSampleStats(plePreparedTokenCacheBytes),
1762
- }
1763
- : null;
1764
-
1765
- results = [
1766
- {
1767
- name: 'benchmark',
1768
- passed: tokensPerSec.length > 0,
1769
- duration: durations.reduce((sum, value) => sum + value, 0),
1770
- error: tokensPerSec.length > 0 ? undefined : 'No benchmark runs completed',
1771
- },
1772
- ];
1773
-
1774
- const normalizedFirstTokenMs = sampleTimingNumber(ttftMsStats, 'median', null);
1775
-
1776
- metrics = {
1777
- warmupRuns,
1778
- timedRuns,
1779
- ...(Number.isFinite(benchRun?.seed) ? { seed: benchRun.seed } : {}),
1780
- prompt: lastPromptLabel,
1781
- maxTokens: lastMaxTokens,
1782
- decodeTokensPerSec: sampleTimingNumber(decodeTokensPerSecStats, 'median'),
1783
- avgTokensGenerated: Math.round(tokensGeneratedStats.mean),
1784
- avgPrefillTokens: Math.round(prefillTokensStats.mean),
1785
- avgDecodeTokens: Math.round(decodeTokensStats.mean),
1786
- medianPrefillTokensPerSec: sampleTimingNumber(prefillTokensPerSecStats, 'median'),
1787
- avgPrefillTokensPerSec: sampleTimingNumber(prefillTokensPerSecStats, 'mean'),
1788
- medianPrefillTokensPerSecTtft: sampleTimingNumber(prefillTokensPerSecTtftStats, 'median'),
1789
- avgPrefillTokensPerSecTtft: sampleTimingNumber(prefillTokensPerSecTtftStats, 'mean'),
1790
- avgDecodeTokensPerSec: sampleTimingNumber(decodeTokensPerSecStats, 'mean'),
1791
- firstTokenMs: normalizedFirstTokenMs,
1792
- firstResponseMs: safeToFixed(safeModelLoadMs + normalizedFirstTokenMs, null),
1793
- prefillMs: sampleTimingNumber(prefillMsStats, 'median'),
1794
- decodeMs: sampleTimingNumber(decodeMsStats, 'median'),
1795
- totalRunMs: sampleTimingNumber(totalMsStats, 'median'),
1796
- decodeMsPerTokenP50: sampleTimingNumber(decodeMsPerTokenStats, 'median'),
1797
- decodeMsPerTokenP95: sampleTimingNumber(decodeMsPerTokenStats, 'p95'),
1798
- decodeMsPerTokenP99: sampleTimingNumber(decodeMsPerTokenStats, 'p99'),
1799
- avgPrefillMs: sampleTimingNumber(prefillMsStats, 'mean'),
1800
- wallRunMs: sampleTimingNumber(wallRunMsStats, 'median'),
1801
- modelLoadMs: safeModelLoadMs,
1802
- throughput: {
1803
- tokensPerSec: tokensPerSecStats,
1804
- prefillTokensPerSec: prefillTokensPerSecStats,
1805
- prefillTokensPerSecTtft: prefillTokensPerSecTtftStats,
1806
- decodeTokensPerSec: decodeTokensPerSecStats,
1807
- },
1808
- latency: {
1809
- totalMs: totalMsStats,
1810
- wallRunMs: wallRunMsStats,
1811
- prefillMs: prefillMsStats,
1812
- decodeMs: decodeMsStats,
1813
- firstTokenMs: ttftMsStats,
1814
- },
1815
- tokens: {
1816
- generated: tokensGeneratedStats,
1817
- prefill: prefillTokensStats,
1818
- decode: decodeTokensStats,
1819
- },
1820
- gpu: gpuPhaseStats,
1821
- batching: batchingPhaseStats,
1822
- decodeCadence: resolveDecodeCadence(getRuntimeConfig(), lastExecutionPlan),
1823
- plePreparedTokenCache: plePreparedTokenCacheStats,
1824
- decodeMode: lastDecodeMode,
1825
- batchGuardReason: lastBatchGuardReason,
1826
- executionPlan: lastExecutionPlan,
1827
- generatedText,
1828
- referenceTranscript: generatedReferenceTranscript,
1829
- promptInput: generatedPromptInput,
1830
- };
1831
-
1832
- timing = buildCanonicalTiming({
1833
- modelLoadMs: safeModelLoadMs,
1834
- firstTokenMs: normalizedFirstTokenMs,
1835
- firstResponseMs: Number.isFinite(normalizedFirstTokenMs)
1836
- ? safeModelLoadMs + normalizedFirstTokenMs
1837
- : null,
1838
- prefillMs: prefillMsStats?.median ?? null,
1839
- decodeMs: decodeMsStats?.median ?? null,
1840
- decodeMsPerTokenP50: decodeMsPerTokenStats?.median ?? null,
1841
- decodeMsPerTokenP95: decodeMsPerTokenStats?.p95 ?? null,
1842
- decodeMsPerTokenP99: decodeMsPerTokenStats?.p99 ?? null,
1843
- totalRunMs: totalMsStats.median,
1844
- decodeTokensPerSec: decodeTokensPerSecStats?.median,
1845
- prefillTokensPerSec: prefillTokensPerSecStats?.median,
1846
- prefillTokensPerSecTtft: prefillTokensPerSecTtftStats?.median,
1847
- cacheMode,
1848
- loadMode,
1849
- });
1850
- }
1851
-
1852
- const memoryStats = typeof harness.pipeline?.getMemoryStats === 'function'
1853
- ? harness.pipeline.getMemoryStats()
1854
- : null;
1855
- const loadTimings = resolvePipelineLoadTimings(harness.pipeline);
1856
- const loadDiagnostics = buildLoadTimingDiagnostics(
1857
- safeModelLoadMs,
1858
- loadTimings.loadTiming,
1859
- loadTimings.pipelineLoadTiming
1860
- );
1861
- const decodeBottleneck = buildDecodeBottleneckDiagnostics(metrics, timing);
1862
- if (decodeBottleneck) {
1863
- metrics.decodeBottleneck = decodeBottleneck;
1864
- }
1865
-
1866
- if (typeof harness.pipeline.unload === 'function' && !options.keepPipeline) {
1867
- await harness.pipeline.unload();
1868
- }
1869
-
1870
- const summary = buildSuiteSummary('bench', results, startTime);
1871
- const timingDiagnostics = buildTimingDiagnostics(timing, {
1872
- source: 'doppler',
1873
- prefillSemantics: 'internal_prefill_phase',
1874
- loadTiming: loadTimings.loadTiming,
1875
- pipelineLoadTiming: loadTimings.pipelineLoadTiming,
1876
- });
1877
- if (decodeBottleneck) {
1878
- timingDiagnostics.decodeBottleneck = decodeBottleneck;
1879
- }
1880
- const firstLoad = buildFirstLoadComposition({
1881
- modelLoadMs: timing.modelLoadMs,
1882
- firstTokenMs: timing.firstTokenMs,
1883
- firstResponseMs: timing.firstResponseMs,
1884
- });
1885
- const metricsWithContracts = buildSuiteContractMetrics(
1886
- 'bench',
1887
- loadDiagnostics ? { ...metrics, load: loadDiagnostics } : metrics,
1888
- harness.manifest
1889
- );
1890
- return {
1891
- ...summary,
1892
- modelId: options.modelId || harness.manifest?.modelId || 'unknown',
1893
- cacheMode,
1894
- loadMode,
1895
- env: {
1896
- library: 'doppler',
1897
- runtime: 'browser',
1898
- device: 'webgpu',
1899
- browserUserAgent: typeof navigator !== 'undefined' ? (navigator.userAgent || null) : null,
1900
- browserPlatform: typeof navigator !== 'undefined' ? (navigator.platform || null) : null,
1901
- browserLanguage: typeof navigator !== 'undefined' ? (navigator.language || null) : null,
1902
- browserVendor: typeof navigator !== 'undefined' ? (navigator.vendor || null) : null,
1903
- },
1904
- timing,
1905
- timingDiagnostics,
1906
- firstLoad,
1907
- output,
1908
- metrics: metricsWithContracts,
1909
- memoryStats,
1910
- deviceInfo: resolveDeviceInfo(),
1911
- pipeline: options.keepPipeline ? harness.pipeline : null,
1912
- };
1913
- }
1914
-
1915
114
  async function dispatchBrowserSuite(mode, workload, options) {
1916
115
  if (mode === 'verify' && workload === 'kernels') {
1917
116
  return runKernelSuite(options);
@@ -1966,37 +165,11 @@ function shouldCaptureDebugSnapshot(mode, runtimeConfig) {
1966
165
  || logLevel === 'verbose';
1967
166
  }
1968
167
 
1969
- function createHarnessPhaseError(error, phase, context = {}) {
1970
- const message = error?.message || String(error);
1971
- const wrapped = new Error(
1972
- `Browser harness phase "${phase}" failed: ${message}`,
1973
- error instanceof Error ? { cause: error } : undefined
1974
- );
1975
- wrapped.name = error?.name || 'Error';
1976
- if (error?.code !== undefined) {
1977
- wrapped.code = error.code;
1978
- }
1979
- wrapped.details = {
1980
- ...(error?.details && typeof error.details === 'object' ? error.details : {}),
1981
- harnessPhase: phase,
1982
- ...context,
1983
- };
1984
- return wrapped;
1985
- }
1986
-
1987
- async function withHarnessPhase(phase, context, run) {
1988
- try {
1989
- return await run();
1990
- } catch (error) {
1991
- if (error?.details?.harnessPhase) {
1992
- throw error;
1993
- }
1994
- throw createHarnessPhaseError(error, phase, context);
1995
- }
1996
- }
1997
-
1998
168
  export async function runBrowserSuite(options = {}) {
1999
169
  return runWithRuntimeIsolationForSuite(async () => {
170
+ if (options.commandContext) {
171
+ assertCommandContextMatchesOptions(options.commandContext, options);
172
+ }
2000
173
  const suiteTimestamp = resolveReportTimestamp(options.timestamp, 'runBrowserSuite timestamp');
2001
174
  const harnessContext = resolveHarnessContext(options);
2002
175
  const mode = resolveHarnessMode(options, harnessContext);
@@ -2049,7 +222,7 @@ export async function runBrowserSuite(options = {}) {
2049
222
  mode,
2050
223
  workload,
2051
224
  suite,
2052
- modelId,
225
+ modelId, surface: options.surface ?? 'browser',
2053
226
  runtimeProfile: options.runtimeProfile ?? null,
2054
227
  deviceInfo: suiteResult.deviceInfo ?? null,
2055
228
  results: suiteResult.results,