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
@@ -1,1386 +1,2 @@
1
1
 
2
-
3
- import { isGpuBufferInstance, isWeightBuffer, getWeightDtype } from '../../../../gpu/weight-buffer.js';
4
- import { getDevice, getKernelCapabilities } from '../../../../gpu/device.js';
5
- import { getRuntimeConfig } from '../../../../config/runtime.js';
6
- import { acquireBuffer, releaseBuffer } from '../../../../memory/buffer-pool.js';
7
- import {
8
- runMatmul,
9
- runRMSNorm,
10
- runLayerNorm,
11
- runBiasAdd,
12
- runRoPE,
13
- canUseRoPEQK,
14
- runRoPEQK,
15
- runAttention,
16
- runAttentionBDPA,
17
- runAttentionTieredQuant,
18
- runAttentionTiered,
19
- runAttentionContiguousQuant,
20
- runSiLU,
21
- castF16ToF32,
22
- castF32ToF16,
23
- runMatmulResidualFused,
24
- shouldUseFusedMatmulResidual,
25
- } from '../../../../gpu/kernel-selector.js';
26
- import { createTensor } from '../../../../gpu/tensor.js';
27
- import { isKernelDebugEnabled, dumpTokenVector, dumpKVCache, logKernelStep } from '../debug-utils.js';
28
- import { applyLoRA } from '../lora-apply.js';
29
- import { getLoRAModule } from '../lora.js';
30
- import { kernelTrace, traceStep } from '../kernel-trace.js';
31
- import { log, trace } from '../../../../debug/index.js';
32
- import { selectRuleValue } from '../../../../rules/rule-registry.js';
33
- import { runProbes } from '../probes.js';
34
- import {
35
- recordAttentionInputs,
36
- shouldForceF32AttentionProjectionForRoPE,
37
- resolveAttentionProjectionOutputDtype,
38
- projectAttentionQKV,
39
- applyAttentionQKNorm,
40
- applyAttentionValueNorm,
41
- hasAttentionProjectionDiagnostics,
42
- hasAttentionStageDiagnostics,
43
- resolveAttentionQKNormState,
44
- } from './projections.js';
45
- import { prepareAttentionProjectionInput } from './output-projection.js';
46
-
47
- import {
48
- shouldDebugLayer,
49
- markStageLogged,
50
- } from './types.js';
51
- import {
52
- getKernelPathMatmulPrecision,
53
- getKernelPathMatmulVariant,
54
- } from '../../../../config/kernel-path-loader.js';
55
- import {
56
- resolveKVCacheState,
57
- createDiffusionGemmaDecoderKVState,
58
- buildAttentionDispatchParams,
59
- buildAttentionInputsData,
60
- } from './dispatch-params.js';
61
- import {
62
- buildTieredQuantAttentionOptions,
63
- buildContiguousQuantAttentionOptions,
64
- } from './quant-options.js';
65
- import { assertImplicitDtypeTransitionAllowed } from '../dtype-contract.js';
66
- import {
67
- resolveAttentionPrecisionContract,
68
- isAttentionKvDtypeExplicit,
69
- } from './precision-contract.js';
70
- import { canUseRmsNormWideTileProjectionFusion } from './rmsnorm-fusion-gate.js';
71
- import { canUseAttentionOutputGateFusion } from './output-gate-fusion.js';
72
- import { getVectorTensor } from '../weights.js';
73
-
74
- const ATTENTION_DTYPE_LOGGED = new Set();
75
-
76
- function isWideTileQ4KPhaseEnabled(session, phase) {
77
- return phase === 'decode'
78
- ? session?.useWideTileQ4KDecode === true
79
- : session?.useWideTileQ4KPrefill === true;
80
- }
81
-
82
- function canUseQ4KWideTileResidualFusion(options = {}) {
83
- return options.hasResidual === true
84
- && options.oProjDtype === 'q4k'
85
- && options.inputDtype === 'f32'
86
- && options.outputDtype === 'f32'
87
- && options.residualMatches === true
88
- && options.hasLoRA !== true
89
- && getKernelCapabilities().hasF16 === true
90
- && getKernelCapabilities().hasSubgroups === true
91
- && options.session?.useWideTileResidualFusion === true
92
- && options.session?.retainQ4KMaterialization === true
93
- && isWideTileQ4KPhaseEnabled(options.session, options.phase);
94
- }
95
-
96
- function assertAttentionDtypeTransitionAllowed(state, fromDtype, toDtype, detail, transitionDeclaredBy = null) {
97
- assertImplicitDtypeTransitionAllowed({
98
- executionPolicies: state?.executionPolicies ?? null,
99
- fromDtype,
100
- toDtype,
101
- op: 'attention',
102
- detail,
103
- transitionDeclaredBy,
104
- });
105
- }
106
-
107
- export async function runLayerAttentionGPU(
108
- input,
109
- layerWeights,
110
- config,
111
- state,
112
- debug = false,
113
- debugFlags = {},
114
- getWeightBuffer,
115
- getNormWeightBuffer,
116
- debugCheckBuffer,
117
- lora
118
- ) {
119
- const {
120
- layerIdx,
121
- numTokens,
122
- isPrefill,
123
- numHeads,
124
- numKVHeads,
125
- headDim,
126
- hiddenSize,
127
- rmsNormEps,
128
- currentSeqLen,
129
- slidingWindow,
130
- layerType,
131
- residualTensor,
132
- attnSoftcap,
133
- queryPreAttnScalar,
134
- skipInputNorm = false,
135
- tokenIds = null,
136
- kernelPath = null,
137
- disableRoPE = false,
138
- multimodalBidirectionalSpan = null,
139
- sharedKVSourceLayerIdx = null,
140
- storeSharedKV = false,
141
- diffusionGemmaDecoder = false,
142
- precomputedInputNorm = null,
143
- } = config;
144
-
145
- const device = getDevice();
146
- const phase = isPrefill ? 'prefill' : 'decode';
147
- const attentionPrecisionContract = resolveAttentionPrecisionContract(config, state);
148
- const attentionActivationDtype = selectRuleValue('shared', 'dtype', 'f16OrF32FromDtype', {
149
- dtype: attentionPrecisionContract.resolvedActivationDtype ?? config.activationDtype,
150
- });
151
- const oProjPrecision = getKernelPathMatmulPrecision('o_proj', phase, layerIdx, kernelPath);
152
- const oProjInputDtype = selectRuleValue('shared', 'dtype', 'f16OrF32FromDtype', {
153
- dtype: oProjPrecision?.inputDtype ?? attentionActivationDtype,
154
- });
155
- const oProjOutputDtype = selectRuleValue('shared', 'dtype', 'f16OrF32FromDtype', {
156
- dtype: oProjPrecision?.outputDtype
157
- ?? attentionPrecisionContract.resolvedOutputDtype
158
- ?? config.activationDtype,
159
- });
160
- const wantsF16Output = oProjOutputDtype === 'f16';
161
- const useF16Activations = attentionActivationDtype === 'f16';
162
- const kvCacheFallback = selectRuleValue('inference', 'dtype', 'f16OrF32', { useF16: useF16Activations });
163
- const kvCacheDtype = attentionPrecisionContract.resolvedKvCacheDtype ?? state.kvCache?.kvDtype ?? kvCacheFallback;
164
- const allowF16Attention = useF16Activations && kvCacheDtype === 'f16';
165
- const attentionInputTransitionDeclaredBy = attentionPrecisionContract.explicitInputDtype
166
- ? 'step_precision'
167
- : null;
168
- let attentionInput = input;
169
- let attentionInputTemp = false;
170
- let normed = attentionInput;
171
- let qTensor = null;
172
- let qGateTensor = null;
173
- let kTensor = null;
174
- let vTensor = null;
175
- let attnOutput = null;
176
- let attnForProjection = null;
177
- let output = null;
178
- let finalOutput = null;
179
- let oProjInputTemp = null;
180
- let valueAliasesKey = false;
181
- let decoderKVState = null;
182
- let decoderKTemp = null;
183
- let decoderVTemp = null;
184
- if (!allowF16Attention && input.dtype !== attentionActivationDtype) {
185
- assertAttentionDtypeTransitionAllowed(
186
- state,
187
- input.dtype,
188
- attentionActivationDtype,
189
- 'The attention kernel selection would widen the input implicitly.',
190
- attentionInputTransitionDeclaredBy
191
- );
192
- attentionInput = attentionActivationDtype === 'f16'
193
- ? await castF32ToF16(input)
194
- : await castF16ToF32(input);
195
- attentionInputTemp = true;
196
- normed = attentionInput;
197
- }
198
-
199
- // Debug: attention input for configured layers
200
- if (isPrefill && shouldDebugLayer(layerIdx, debugFlags.debugLayers) && !markStageLogged(layerIdx, 'attn_input', debugFlags) && debugCheckBuffer) {
201
- await debugCheckBuffer(attentionInput.buffer, `L${layerIdx} attention input (GPU)`, numTokens, hiddenSize);
202
- }
203
-
204
- // Debug logging moved to debug-utils.ts (enable via setDebugConfig)
205
-
206
- if (!layerWeights) {
207
- const bytesPerElement = wantsF16Output ? 2 : 4;
208
- const outputBuf = acquireBuffer(numTokens * hiddenSize * bytesPerElement, undefined, 'attn_output');
209
- const output = createTensor(outputBuf, oProjOutputDtype, [numTokens, hiddenSize], 'attn_output');
210
- return { output, residualFused: false };
211
- }
212
-
213
- const qSize = numTokens * numHeads * headDim;
214
- const kvSize = numTokens * numKVHeads * headDim;
215
- const sharedKVEntry = sharedKVSourceLayerIdx == null
216
- ? null
217
- : (state.sharedAttentionState?.get(sharedKVSourceLayerIdx) ?? null);
218
- if (sharedKVSourceLayerIdx != null && !sharedKVEntry) {
219
- throw new Error(
220
- `Layer ${layerIdx} requires shared K/V from layer ${sharedKVSourceLayerIdx}, ` +
221
- 'but no shared K/V state was stored for that source layer.'
222
- );
223
- }
224
- if (sharedKVEntry && (
225
- sharedKVEntry.headDim !== headDim
226
- || sharedKVEntry.numKVHeads !== numKVHeads
227
- )) {
228
- throw new Error(
229
- `Layer ${layerIdx} shared K/V geometry mismatch. ` +
230
- `Expected numKVHeads=${numKVHeads}, headDim=${headDim}; ` +
231
- `got numKVHeads=${sharedKVEntry.numKVHeads}, headDim=${sharedKVEntry.headDim}.`
232
- );
233
- }
234
- const reusesSharedKV = sharedKVEntry != null;
235
-
236
- // 1. Input norm
237
- // Opt-in fusion: when useFusedRmsnormWideTile is set AND preconditions match,
238
- // defer the rmsnorm into each q/k/v_proj matmul via a dedicated WideTile
239
- // variant. Saves one standalone rmsnorm dispatch per layer.
240
- let fusedNormWeight = null;
241
- let fusedNormEps = null;
242
- let fusedNormOffset = false;
243
- let fusedNormWeightOwned = false;
244
- // state.runtimeConfig.inference.session is a stale snapshot ({}); the live
245
- // module-level runtime config carries the merged profile/override session.
246
- const rmsNormFusionFlag = getRuntimeConfig()?.inference?.session?.useFusedRmsnormWideTile === true;
247
- const hasInputNormWeight = layerWeights?.inputNorm != null && getNormWeightBuffer != null;
248
- const canSelectFusedRmsNormProjection = canUseRmsNormWideTileProjectionFusion(layerWeights, reusesSharedKV);
249
- const canFuseInputNormIntoProjections = rmsNormFusionFlag
250
- && config.normalizationType === 'rmsnorm'
251
- && hasInputNormWeight
252
- && canSelectFusedRmsNormProjection
253
- && !skipInputNorm
254
- && isPrefill
255
- && numTokens > 1
256
- && attentionInput.dtype === 'f32';
257
-
258
- try {
259
- if (precomputedInputNorm) {
260
- if (
261
- isPrefill
262
- || numTokens !== 1
263
- || skipInputNorm
264
- || !layerWeights.inputNorm
265
- || !getNormWeightBuffer
266
- || precomputedInputNorm.dtype !== attentionInput.dtype
267
- || precomputedInputNorm.dtype !== 'f32'
268
- ) {
269
- throw new Error(`Layer ${layerIdx} received an incompatible precomputed input norm tensor.`);
270
- }
271
- normed = precomputedInputNorm;
272
- await runProbes('post_input_norm', normed.buffer, {
273
- layerIdx,
274
- numTokens,
275
- hiddenSize,
276
- probes: state.debugProbes,
277
- operatorDiagnostics: state.operatorDiagnostics,
278
- dtype: normed.dtype,
279
- });
280
- } else if (canFuseInputNormIntoProjections) {
281
- const normWeightBuf = getNormWeightBuffer(layerWeights.inputNorm, 'input_norm');
282
- fusedNormWeight = normWeightBuf;
283
- fusedNormEps = rmsNormEps;
284
- fusedNormOffset = config.rmsNormWeightOffset === true;
285
- fusedNormWeightOwned = !isGpuBufferInstance(layerWeights.inputNorm) && !isWeightBuffer(layerWeights.inputNorm);
286
- // Keep normed = attentionInput (raw). Each q/k/v_proj matmul runs the
287
- // norm internally via the fused kernel variant.
288
- } else if (!skipInputNorm && layerWeights.inputNorm && getNormWeightBuffer) {
289
- const normWeightBuf = getNormWeightBuffer(layerWeights.inputNorm, 'input_norm');
290
- let normBiasTensor = null;
291
- let normBiasOwned = false;
292
- try {
293
- // Debug: norm weights for configured layers
294
- if (isPrefill && shouldDebugLayer(layerIdx, debugFlags.debugLayers) && !markStageLogged(layerIdx, 'norm_weights', debugFlags) && debugCheckBuffer) {
295
- await debugCheckBuffer(normWeightBuf, `L${layerIdx} input norm weights (GPU)`, 1, hiddenSize);
296
- }
297
-
298
- if (config.normalizationType === 'layernorm') {
299
- if (!layerWeights.inputNormBias) {
300
- throw new Error(`Layer ${layerIdx} requires input_layernorm.bias for LayerNorm execution.`);
301
- }
302
- ({ tensor: normBiasTensor, owned: normBiasOwned } = getVectorTensor(
303
- layerWeights.inputNormBias,
304
- `L${layerIdx}.input_norm_bias`,
305
- hiddenSize,
306
- {},
307
- debugFlags
308
- ));
309
- normed = await runLayerNorm(
310
- attentionInput,
311
- normWeightBuf,
312
- normBiasTensor.buffer,
313
- rmsNormEps,
314
- { batchSize: numTokens, hiddenSize }
315
- );
316
- } else {
317
- normed = await runRMSNorm(attentionInput, normWeightBuf, rmsNormEps, {
318
- batchSize: numTokens,
319
- hiddenSize,
320
- rmsNormWeightOffset: config.rmsNormWeightOffset,
321
- });
322
- }
323
- } finally {
324
- if (normBiasOwned && normBiasTensor) releaseBuffer(normBiasTensor.buffer);
325
- if (!isGpuBufferInstance(layerWeights.inputNorm) && !isWeightBuffer(layerWeights.inputNorm)) releaseBuffer(normWeightBuf);
326
- }
327
-
328
- // Trace input norm output
329
- if (kernelTrace.enabled) {
330
- await traceStep('rmsnorm', `L${layerIdx}.input_norm`, layerIdx, normed.buffer, [numTokens, hiddenSize]);
331
- }
332
-
333
- if (isKernelDebugEnabled(layerIdx)) {
334
- logKernelStep('rmsnorm', { layerIdx, label: 'input_norm', size: numTokens * hiddenSize });
335
- await dumpTokenVector(normed.buffer, 'input_norm_out', {
336
- layerIdx,
337
- tokenIdx: Math.max(0, numTokens - 1),
338
- rowSize: hiddenSize,
339
- dtype: normed.dtype,
340
- });
341
- }
342
-
343
- await runProbes('post_input_norm', normed.buffer, {
344
- layerIdx,
345
- numTokens,
346
- hiddenSize,
347
- probes: state.debugProbes,
348
- operatorDiagnostics: state.operatorDiagnostics,
349
- dtype: normed.dtype,
350
- });
351
- }
352
-
353
- // Debug: Check normed input for L0 prefill
354
- // Debug: normed input for configured layers
355
- if (isPrefill && shouldDebugLayer(layerIdx, debugFlags.debugLayers) && !markStageLogged(layerIdx, 'attn_normed', debugFlags) && debugCheckBuffer) {
356
- await debugCheckBuffer(normed.buffer, `L${layerIdx} normed input (GPU)`, numTokens);
357
- }
358
-
359
- const debugLayers = debugFlags.debugLayers;
360
- const shouldLogLayer = debugLayers === null ? layerIdx === 0 : shouldDebugLayer(layerIdx, debugLayers);
361
- if (shouldLogLayer) {
362
- const phase = selectRuleValue('kernels', 'attention', 'phase', { isDecode: !isPrefill });
363
- const logKey = `L${layerIdx}_${phase}_dtypes`;
364
- if (!ATTENTION_DTYPE_LOGGED.has(logKey)) {
365
- ATTENTION_DTYPE_LOGGED.add(logKey);
366
- trace.attn(layerIdx, `dtypes: activation=${attentionActivationDtype}, input=${input.dtype}, normed=${normed.dtype}`);
367
- }
368
- }
369
-
370
- if (isKernelDebugEnabled(layerIdx)) {
371
- await dumpTokenVector(normed.buffer, 'attn_in', {
372
- layerIdx,
373
- tokenIdx: Math.max(0, numTokens - 1),
374
- rowSize: hiddenSize,
375
- dtype: normed.dtype,
376
- });
377
- }
378
-
379
- // 2. Q/K/V projections
380
- const qProjVariant = getKernelPathMatmulVariant('q_proj', phase, layerIdx, kernelPath);
381
- const kernelPathIsF16 = qProjVariant != null && qProjVariant.includes('f16') && !qProjVariant.includes('f32');
382
- const matmulOutputDtype = resolveAttentionProjectionOutputDtype(attentionActivationDtype, {
383
- forceF32: shouldForceF32AttentionProjectionForRoPE({
384
- attentionInputDtype: attentionActivationDtype,
385
- headDim,
386
- rotaryDim: config.ropeRotaryDim,
387
- interleaved: config.ropeInterleaved,
388
- kernelPathIsF16,
389
- }),
390
- });
391
- let usedFusedQKV = false;
392
- let qkNormApplied = false;
393
- let ropeApplied = false;
394
- const qkNormState = resolveAttentionQKNormState({ config, layerWeights, layerIdx, reusesSharedKV });
395
- const qkNormProjectionDiagnostics = hasAttentionProjectionDiagnostics(state)
396
- || kernelTrace.enabled
397
- || isKernelDebugEnabled(layerIdx)
398
- || (isPrefill && shouldDebugLayer(layerIdx, debugFlags.debugLayers) && !!debugCheckBuffer);
399
- const qkNormRoPEDiagnostics = hasAttentionStageDiagnostics(
400
- state,
401
- ['q_proj', 'k_proj', 'v_proj', 'q_norm', 'k_norm']
402
- )
403
- || kernelTrace.enabled
404
- || isKernelDebugEnabled(layerIdx)
405
- || (isPrefill && shouldDebugLayer(layerIdx, debugFlags.debugLayers) && !!debugCheckBuffer);
406
- const runtimeSession = getRuntimeConfig()?.inference?.session;
407
- const qkNormFusionFlag = runtimeSession?.useFusedQKVSplitQKNorm === true;
408
- const qkNormRoPEFusionFlag = runtimeSession?.useFusedQKVSplitQKNormRoPE === true;
409
- ({ qTensor, qGateTensor, kTensor, vTensor, usedFusedQKV, valueAliasesKey, qkNormApplied, ropeApplied } = await projectAttentionQKV({
410
- recorder: null,
411
- normed,
412
- layerWeights,
413
- numTokens,
414
- numHeads,
415
- numKVHeads,
416
- headDim,
417
- hiddenSize,
418
- layerIdx,
419
- kernelPath,
420
- matmulOutputDtype,
421
- getWeightBuffer,
422
- lora,
423
- matmulDebug: state.runtimeConfig?.shared?.debug?.matmul ?? null,
424
- attentionOutputGate: config.attentionOutputGate === true,
425
- sharedKTensor: sharedKVEntry?.kTensor ?? null,
426
- sharedVTensor: sharedKVEntry?.vTensor ?? null,
427
- executionPolicies: state.executionPolicies ?? null,
428
- releaseTemporary: (buffer) => releaseBuffer(buffer),
429
- onFusedQKV: layerIdx === 0 && isPrefill
430
- ? ({ qSize: qSizeFused, kSize: kSizeFused, vSize: vSizeFused, totalSize }) => {
431
- trace.attn(layerIdx, `Using fused QKV path: ${qSizeFused}+${kSizeFused}+${vSizeFused}=${totalSize}`);
432
- }
433
- : null,
434
- fusedNormWeight,
435
- fusedNormEps,
436
- fusedNormOffset,
437
- qkNormFusion: {
438
- enabled: qkNormFusionFlag
439
- && qkNormState.wantsQKNorm
440
- && config.queryKeyNormType === 'rmsnorm'
441
- && config.queryKeyNormAxis === 'head',
442
- getNormWeightBuffer,
443
- rmsNormEps,
444
- rmsNormWeightOffset: config.rmsNormWeightOffset,
445
- skipKNorm: qkNormState.skipKNorm,
446
- allowUnitQKNorm: qkNormState.allowUnitQKNorm,
447
- projectionDiagnosticsEnabled: qkNormProjectionDiagnostics,
448
- },
449
- qkNormRoPEFusion: {
450
- enabled: qkNormRoPEFusionFlag
451
- && qkNormState.wantsQKNorm
452
- && config.queryKeyNormType === 'rmsnorm'
453
- && config.queryKeyNormAxis === 'head'
454
- && !disableRoPE
455
- && !!state.ropeFreqsCos
456
- && !!state.ropeFreqsSin,
457
- getNormWeightBuffer,
458
- rmsNormEps,
459
- rmsNormWeightOffset: config.rmsNormWeightOffset,
460
- skipKNorm: qkNormState.skipKNorm,
461
- allowUnitQKNorm: qkNormState.allowUnitQKNorm,
462
- projectionDiagnosticsEnabled: qkNormRoPEDiagnostics,
463
- freqsCos: state.ropeFreqsCos,
464
- freqsSin: state.ropeFreqsSin,
465
- startPos: currentSeqLen,
466
- rotaryDim: config.ropeRotaryDim,
467
- pairSpanDim: config.ropeFrequencyBaseDim ?? config.ropeRotaryDim,
468
- interleaved: config.ropeInterleaved,
469
- reusesSharedKV,
470
- },
471
- }));
472
- // Release the in-flight norm weight buffer used by the fused projections.
473
- if (fusedNormWeight && fusedNormWeightOwned) {
474
- releaseBuffer(fusedNormWeight);
475
- }
476
-
477
- // Trace Q/K/V projections
478
- if (kernelTrace.enabled) {
479
- await traceStep('matmul', `L${layerIdx}.q_proj`, layerIdx, qTensor.buffer, [numTokens, numHeads * headDim]);
480
- await traceStep('matmul', `L${layerIdx}.k_proj`, layerIdx, kTensor.buffer, [numTokens, numKVHeads * headDim]);
481
- await traceStep('matmul', `L${layerIdx}.v_proj`, layerIdx, vTensor.buffer, [numTokens, numKVHeads * headDim]);
482
- }
483
- await runProbes('q_proj', qTensor.buffer, {
484
- layerIdx,
485
- numTokens,
486
- hiddenSize: numHeads * headDim,
487
- probes: state.debugProbes,
488
- operatorDiagnostics: state.operatorDiagnostics,
489
- dtype: qTensor.dtype,
490
- });
491
- if (!reusesSharedKV) {
492
- await runProbes('k_proj', kTensor.buffer, {
493
- layerIdx,
494
- numTokens,
495
- hiddenSize: numKVHeads * headDim,
496
- probes: state.debugProbes,
497
- operatorDiagnostics: state.operatorDiagnostics,
498
- dtype: kTensor.dtype,
499
- });
500
- await runProbes('v_proj', vTensor.buffer, {
501
- layerIdx,
502
- numTokens,
503
- hiddenSize: numKVHeads * headDim,
504
- probes: state.debugProbes,
505
- operatorDiagnostics: state.operatorDiagnostics,
506
- dtype: vTensor.dtype,
507
- });
508
- }
509
-
510
- // Kernel step debug: Q/K/V projections
511
- if (isKernelDebugEnabled(layerIdx)) {
512
- logKernelStep('matmul', { layerIdx, label: 'Q_proj', M: numTokens, N: numHeads * headDim, K: hiddenSize });
513
- await dumpTokenVector(qTensor.buffer, 'Q_proj', {
514
- layerIdx,
515
- tokenIdx: Math.max(0, numTokens - 1),
516
- rowSize: numHeads * headDim,
517
- dtype: qTensor.dtype,
518
- });
519
- logKernelStep('matmul', { layerIdx, label: 'K_proj', M: numTokens, N: numKVHeads * headDim, K: hiddenSize });
520
- await dumpTokenVector(kTensor.buffer, 'K_proj', {
521
- layerIdx,
522
- tokenIdx: Math.max(0, numTokens - 1),
523
- rowSize: numKVHeads * headDim,
524
- dtype: kTensor.dtype,
525
- });
526
- logKernelStep('matmul', { layerIdx, label: 'V_proj', M: numTokens, N: numKVHeads * headDim, K: hiddenSize });
527
- await dumpTokenVector(vTensor.buffer, 'V_proj', {
528
- layerIdx,
529
- tokenIdx: Math.max(0, numTokens - 1),
530
- rowSize: numKVHeads * headDim,
531
- dtype: vTensor.dtype,
532
- });
533
- }
534
-
535
- // Debug: Check Q/K/V after projections for L0 prefill
536
- // Debug: Q/K/V projections for configured layers
537
- if (isPrefill && shouldDebugLayer(layerIdx, debugFlags.debugLayers) && !markStageLogged(layerIdx, 'qkv_proj', debugFlags) && debugCheckBuffer) {
538
- await debugCheckBuffer(qTensor.buffer, `L${layerIdx} Q after proj (GPU)`, numTokens, numHeads * headDim);
539
- await debugCheckBuffer(kTensor.buffer, `L${layerIdx} K after proj (GPU)`, numTokens, numKVHeads * headDim);
540
- await debugCheckBuffer(vTensor.buffer, `L${layerIdx} V after proj (GPU)`, numTokens, numKVHeads * headDim);
541
- }
542
-
543
- // Optional Q/K normalization; the manifest owns the axis and norm family.
544
- if (isKernelDebugEnabled(layerIdx)) {
545
- logKernelStep(
546
- 'qk_norm',
547
- {
548
- layerIdx,
549
- label: `hasQ=${qkNormState.hasQNorm} hasK=${qkNormState.hasKNorm} ` +
550
- `wants=${qkNormState.wantsQKNorm} unit=${qkNormState.allowUnitQKNorm}`,
551
- }
552
- );
553
- }
554
-
555
- if (qkNormState.wantsQKNorm && !qkNormApplied) {
556
- ({ qTensor, kTensor } = await applyAttentionQKNorm({
557
- recorder: null,
558
- qTensor,
559
- kTensor,
560
- layerWeights,
561
- getNormWeightBuffer,
562
- rmsNormEps,
563
- numTokens,
564
- numHeads,
565
- numKVHeads,
566
- headDim,
567
- rmsNormWeightOffset: config.rmsNormWeightOffset,
568
- normalizationType: config.queryKeyNormType,
569
- normalizationAxis: config.queryKeyNormAxis,
570
- releaseTemporary: (buffer) => releaseBuffer(buffer),
571
- skipKNorm: qkNormState.skipKNorm,
572
- onQNormApplied: isKernelDebugEnabled(layerIdx)
573
- ? async (tensor) => {
574
- await dumpTokenVector(tensor.buffer, 'Q_norm', {
575
- layerIdx,
576
- tokenIdx: Math.max(0, numTokens - 1),
577
- rowSize: numHeads * headDim,
578
- dtype: tensor.dtype,
579
- });
580
- }
581
- : null,
582
- onKNormApplied: isKernelDebugEnabled(layerIdx)
583
- ? async (tensor) => {
584
- await dumpTokenVector(tensor.buffer, 'K_norm', {
585
- layerIdx,
586
- tokenIdx: Math.max(0, numTokens - 1),
587
- rowSize: numKVHeads * headDim,
588
- dtype: tensor.dtype,
589
- });
590
- }
591
- : null,
592
- retainKInput: valueAliasesKey,
593
- allowUnitQKNorm: qkNormState.allowUnitQKNorm,
594
- }));
595
- }
596
-
597
- if (config.valueNorm === true && !reusesSharedKV) {
598
- const valueNormInputAliasesKey = vTensor.buffer === kTensor.buffer;
599
- vTensor = await applyAttentionValueNorm({
600
- recorder: null,
601
- vTensor,
602
- rmsNormEps,
603
- numTokens,
604
- numKVHeads,
605
- headDim,
606
- releaseTemporary: (buffer) => {
607
- if (!valueNormInputAliasesKey) {
608
- releaseBuffer(buffer);
609
- }
610
- },
611
- });
612
- await runProbes('v_norm', vTensor.buffer, {
613
- layerIdx,
614
- numTokens,
615
- hiddenSize: numKVHeads * headDim,
616
- probes: state.debugProbes,
617
- operatorDiagnostics: state.operatorDiagnostics,
618
- dtype: vTensor.dtype,
619
- });
620
- }
621
-
622
- if (qkNormState.wantsQKNorm) {
623
- await runProbes('q_norm', qTensor.buffer, {
624
- layerIdx,
625
- numTokens,
626
- hiddenSize: numHeads * headDim,
627
- probes: state.debugProbes,
628
- operatorDiagnostics: state.operatorDiagnostics,
629
- dtype: qTensor.dtype,
630
- });
631
- if (!reusesSharedKV) {
632
- await runProbes('k_norm', kTensor.buffer, {
633
- layerIdx,
634
- numTokens,
635
- hiddenSize: numKVHeads * headDim,
636
- probes: state.debugProbes,
637
- operatorDiagnostics: state.operatorDiagnostics,
638
- dtype: kTensor.dtype,
639
- });
640
- }
641
- }
642
-
643
- if (normed !== attentionInput) releaseBuffer(normed.buffer);
644
- if (attentionInputTemp) releaseBuffer(attentionInput.buffer);
645
-
646
- // 3. RoPE (modifies tensor in-place)
647
-
648
- if (!ropeApplied && !disableRoPE && state.ropeFreqsCos && state.ropeFreqsSin) {
649
- const ropeOptions = {
650
- headDim,
651
- rotaryDim: config.ropeRotaryDim,
652
- pairSpanDim: config.ropeFrequencyBaseDim ?? config.ropeRotaryDim,
653
- interleaved: config.ropeInterleaved,
654
- startPos: currentSeqLen,
655
- executionPolicies: state.executionPolicies ?? null,
656
- };
657
- if (canUseRoPEQK(qTensor, kTensor, { reusesSharedKV })) {
658
- await runRoPEQK(qTensor, kTensor, state.ropeFreqsCos, state.ropeFreqsSin, numTokens, {
659
- ...ropeOptions,
660
- numQHeads: numHeads,
661
- numKVHeads,
662
- });
663
- } else {
664
- await runRoPE(qTensor, state.ropeFreqsCos, state.ropeFreqsSin, numTokens, {
665
- ...ropeOptions,
666
- numHeads,
667
- });
668
- if (!reusesSharedKV) {
669
- await runRoPE(kTensor, state.ropeFreqsCos, state.ropeFreqsSin, numTokens, {
670
- ...ropeOptions,
671
- numHeads: numKVHeads,
672
- });
673
- }
674
- }
675
-
676
- // Trace RoPE outputs
677
- if (kernelTrace.enabled) {
678
- await traceStep('rope', `L${layerIdx}.q_rope`, layerIdx, qTensor.buffer, [numTokens, numHeads * headDim]);
679
- if (!reusesSharedKV) {
680
- await traceStep('rope', `L${layerIdx}.k_rope`, layerIdx, kTensor.buffer, [numTokens, numKVHeads * headDim]);
681
- }
682
- }
683
- }
684
- await runProbes('q_rope', qTensor.buffer, {
685
- layerIdx,
686
- numTokens,
687
- hiddenSize: numHeads * headDim,
688
- probes: state.debugProbes,
689
- operatorDiagnostics: state.operatorDiagnostics,
690
- dtype: qTensor.dtype,
691
- });
692
- await runProbes('k_rope', kTensor.buffer, {
693
- layerIdx,
694
- numTokens,
695
- hiddenSize: numKVHeads * headDim,
696
- probes: state.debugProbes,
697
- operatorDiagnostics: state.operatorDiagnostics,
698
- dtype: kTensor.dtype,
699
- });
700
- if (isKernelDebugEnabled(layerIdx)) {
701
- logKernelStep('rope', { layerIdx, label: `startPos=${currentSeqLen}` });
702
- await dumpTokenVector(qTensor.buffer, 'Q_rope', {
703
- layerIdx,
704
- tokenIdx: Math.max(0, numTokens - 1),
705
- rowSize: numHeads * headDim,
706
- dtype: qTensor.dtype,
707
- });
708
- await dumpTokenVector(kTensor.buffer, 'K_rope', {
709
- layerIdx,
710
- tokenIdx: Math.max(0, numTokens - 1),
711
- rowSize: numKVHeads * headDim,
712
- dtype: kTensor.dtype,
713
- });
714
- }
715
-
716
- if (storeSharedKV && state.sharedAttentionState) {
717
- state.sharedAttentionState.set(layerIdx, {
718
- kTensor,
719
- vTensor,
720
- headDim,
721
- numKVHeads,
722
- });
723
- }
724
-
725
- // Debug: Check Q/K after RoPE for L0 prefill
726
- // Debug: Q/K after RoPE for configured layers
727
- if (isPrefill && shouldDebugLayer(layerIdx, debugFlags.debugLayers) && !markStageLogged(layerIdx, 'qk_rope', debugFlags) && debugCheckBuffer) {
728
- await debugCheckBuffer(qTensor.buffer, `L${layerIdx} Q after RoPE (GPU)`, numTokens, numHeads * headDim);
729
- await debugCheckBuffer(kTensor.buffer, `L${layerIdx} K after RoPE (GPU)`, numTokens, numKVHeads * headDim);
730
- }
731
-
732
- // 4. Update KV cache (cache stores raw GPUBuffers for memory efficiency)
733
-
734
- if (!diffusionGemmaDecoder && state.kvCache?.hasGPUCache?.()) {
735
- if (state.kvCache.kvDtype === 'f16') {
736
- const hasExplicitF16KvContract = isAttentionKvDtypeExplicit(attentionPrecisionContract, 'f16');
737
- if (kTensor.dtype !== 'f16' && !hasExplicitF16KvContract) {
738
- assertAttentionDtypeTransitionAllowed(state, kTensor.dtype, 'f16', 'K would be narrowed implicitly for KV cache storage.');
739
- }
740
- if (vTensor.dtype !== 'f16' && !hasExplicitF16KvContract) {
741
- assertAttentionDtypeTransitionAllowed(state, vTensor.dtype, 'f16', 'V would be narrowed implicitly for KV cache storage.');
742
- }
743
- const kCasted = kTensor.dtype === 'f16' ? kTensor : await castF32ToF16(kTensor);
744
- const vCasted = vTensor.dtype === 'f16' ? vTensor : await castF32ToF16(vTensor);
745
-
746
- await state.kvCache.updateFromGPU(layerIdx, kCasted.buffer, vCasted.buffer, currentSeqLen, numTokens, tokenIds);
747
-
748
- if (kTensor.dtype !== 'f16') releaseBuffer(kCasted.buffer);
749
- if (vTensor.dtype !== 'f16') releaseBuffer(vCasted.buffer);
750
- } else {
751
- await state.kvCache.updateFromGPU(layerIdx, kTensor.buffer, vTensor.buffer, currentSeqLen, numTokens, tokenIds);
752
- }
753
-
754
- // Kernel step debug: KV cache state after update
755
- if (isKernelDebugEnabled(layerIdx)) {
756
- trace.kv(layerIdx, `KV cache updated: startPos=${currentSeqLen}, numTokens=${numTokens}`);
757
- await dumpKVCache(( (state.kvCache)), layerIdx);
758
- }
759
- }
760
-
761
- // Resolve KV cache state and build dispatch parameters (shared with record.js)
762
- let kvState;
763
- if (diffusionGemmaDecoder) {
764
- const decoderKVDtype = kvCacheDtype ?? kTensor.dtype;
765
- let decoderKTensor = kTensor;
766
- let decoderVTensor = vTensor;
767
- if (decoderKVDtype === 'f16' && decoderKTensor.dtype !== 'f16') {
768
- const hasExplicitF16KvContract = isAttentionKvDtypeExplicit(attentionPrecisionContract, 'f16');
769
- if (!hasExplicitF16KvContract) {
770
- assertAttentionDtypeTransitionAllowed(state, decoderKTensor.dtype, 'f16', 'DiffusionGemma decoder K would be narrowed implicitly.');
771
- }
772
- decoderKTensor = await castF32ToF16(decoderKTensor);
773
- decoderKTemp = decoderKTensor;
774
- } else if (decoderKVDtype === 'f32' && decoderKTensor.dtype !== 'f32') {
775
- assertAttentionDtypeTransitionAllowed(state, decoderKTensor.dtype, 'f32', 'DiffusionGemma decoder K would be widened implicitly.');
776
- decoderKTensor = await castF16ToF32(decoderKTensor);
777
- decoderKTemp = decoderKTensor;
778
- }
779
- if (decoderKVDtype === 'f16' && decoderVTensor.dtype !== 'f16') {
780
- const hasExplicitF16KvContract = isAttentionKvDtypeExplicit(attentionPrecisionContract, 'f16');
781
- if (!hasExplicitF16KvContract) {
782
- assertAttentionDtypeTransitionAllowed(state, decoderVTensor.dtype, 'f16', 'DiffusionGemma decoder V would be narrowed implicitly.');
783
- }
784
- decoderVTensor = await castF32ToF16(decoderVTensor);
785
- decoderVTemp = decoderVTensor;
786
- } else if (decoderKVDtype === 'f32' && decoderVTensor.dtype !== 'f32') {
787
- assertAttentionDtypeTransitionAllowed(state, decoderVTensor.dtype, 'f32', 'DiffusionGemma decoder V would be widened implicitly.');
788
- decoderVTensor = await castF16ToF32(decoderVTensor);
789
- decoderVTemp = decoderVTensor;
790
- }
791
- kvState = await createDiffusionGemmaDecoderKVState({
792
- state,
793
- layerIdx,
794
- kTensor: decoderKTensor,
795
- vTensor: decoderVTensor,
796
- currentSeqLen,
797
- numTokens,
798
- numKVHeads,
799
- headDim,
800
- layerType,
801
- slidingWindow,
802
- kvDtype: decoderKVDtype,
803
- });
804
- if (decoderKTemp) {
805
- releaseBuffer(decoderKTemp.buffer);
806
- decoderKTemp = null;
807
- }
808
- if (decoderVTemp && decoderVTemp.buffer !== decoderKTensor.buffer) {
809
- releaseBuffer(decoderVTemp.buffer);
810
- }
811
- decoderVTemp = null;
812
- } else {
813
- kvState = resolveKVCacheState(state, layerIdx, kTensor, vTensor, currentSeqLen, numTokens);
814
- }
815
- const dispatchConfig = {
816
- layerIdx, numTokens, isPrefill, numHeads, numKVHeads, headDim, hiddenSize,
817
- slidingWindow: diffusionGemmaDecoder ? null : slidingWindow,
818
- layerType,
819
- layerTypes: config.layerTypes,
820
- queryPreAttnScalar,
821
- causalAttention: diffusionGemmaDecoder ? false : config.causalAttention,
822
- activationDtype: attentionActivationDtype,
823
- kvCacheDtype: attentionPrecisionContract.resolvedKvCacheDtype ?? state.kvCache?.kvDtype ?? null,
824
- };
825
- const dispatchParams = buildAttentionDispatchParams(dispatchConfig, state, kTensor, vTensor, kvState);
826
- const {
827
- effectiveSlidingWindow, attentionKernelVariant, attnScale,
828
- cachedKDtype, cachedVDtype, cachedKTensor, cachedVTensor,
829
- prefillFallbackNeedsCast, causalForAttention,
830
- } = dispatchParams;
831
- const mergedSession = getRuntimeConfig()?.inference?.session;
832
- let attentionOutputGateFused = false;
833
-
834
- // 5. Attention (uses raw GPUBuffers)
835
-
836
- // Debug: log scale on layer 0
837
- if (layerIdx === 0 && isPrefill) {
838
- trace.attn(layerIdx, `Attention scale=${attnScale.toFixed(6)}, queryPreAttnScalar=${queryPreAttnScalar ?? 'undefined'}, headDim=${headDim}`);
839
- }
840
-
841
- // Kernel step debug: KV cache state after resolve
842
- if (isKernelDebugEnabled(layerIdx)) {
843
- trace.kv(layerIdx, `KV resolved: kvLen=${kvState.kvLenForAttention}, layout=${kvState.kvLayout}`);
844
- }
845
-
846
- recordAttentionInputs(state, buildAttentionInputsData(
847
- dispatchConfig, input, normed, kvState, dispatchParams,
848
- { useF16Activations, matmulOutputDtype },
849
- usedFusedQKV, qTensor, kTensor, vTensor,
850
- ));
851
-
852
- const attentionKernelRunners = {
853
- bdpa: async () => {
854
- const BDPA_MAX_HEAD_DIM = 256;
855
- const BDPA_MAX_KV_LEN = 2048;
856
- if (numTokens !== 1) {
857
- throw new Error(`BDPA attention supports decode-only seqLen=1; got seqLen=${numTokens}.`);
858
- }
859
- if (headDim > BDPA_MAX_HEAD_DIM) {
860
- throw new Error(
861
- `BDPA attention kernel supports headDim <= ${BDPA_MAX_HEAD_DIM}; got headDim=${headDim}.`
862
- );
863
- }
864
- if (kvState.kvLenForAttention > BDPA_MAX_KV_LEN) {
865
- throw new Error(
866
- `BDPA attention kernel supports kvLen <= ${BDPA_MAX_KV_LEN}; got kvLen=${kvState.kvLenForAttention}.`
867
- );
868
- }
869
- const basisKDtype = 'f16';
870
- const basisVDtype = 'f16';
871
- const basisCount = Math.max(1, kvState.bdpaBasisCount);
872
- const basisKTensor = createTensor(kvState.bdpaBasisK, basisKDtype, [basisCount, numKVHeads * headDim], 'bdpa_basis_k');
873
- const basisVTensor = createTensor(kvState.bdpaBasisV, basisVDtype, [basisCount, numKVHeads * headDim], 'bdpa_basis_v');
874
- let qForBDPA = qTensor;
875
- if (qForBDPA.dtype !== 'f16') {
876
- assertAttentionDtypeTransitionAllowed(state, qTensor.dtype, 'f16', 'Q would be narrowed implicitly for BDPA.');
877
- qForBDPA = await castF32ToF16(qTensor);
878
- }
879
- const output = await runAttentionBDPA(qForBDPA, basisKTensor, basisVTensor, kvState.bdpaPagedK, kvState.bdpaPagedV, kvState.bdpaIndex, numHeads, headDim, {
880
- seqLen: numTokens,
881
- kvLen: kvState.kvLenForAttention,
882
- numKVHeads,
883
- causal: causalForAttention,
884
- startPos: kvState.startPosForMask,
885
- layerIdx,
886
- slidingWindow: effectiveSlidingWindow,
887
- attnSoftcap,
888
- scale: attnScale,
889
- ropeCos: state.ropeFreqsCos,
890
- ropeSin: state.ropeFreqsSin,
891
- });
892
- if (qForBDPA !== qTensor) {
893
- releaseBuffer(qForBDPA.buffer);
894
- }
895
- return output;
896
- },
897
- tieredQuant: async () => {
898
- let qForAttention = qTensor;
899
- let qTemp = null;
900
- if (kvState.coldQuantMode !== 'none' && qTensor.dtype !== 'f32') {
901
- assertAttentionDtypeTransitionAllowed(state, qTensor.dtype, 'f32', 'Q would be widened implicitly for tiered quant attention.');
902
- qForAttention = await castF16ToF32(qTensor);
903
- qTemp = qForAttention;
904
- }
905
- const cachedHotKTensor = createTensor(kvState.cachedKHot, cachedKDtype, [kvState.hotLen, numKVHeads * headDim], 'cached_K_hot');
906
- const cachedHotVTensor = createTensor(kvState.cachedVHot, cachedVDtype, [kvState.hotLen, numKVHeads * headDim], 'cached_V_hot');
907
-
908
- if (kvState.coldQuantMode === 'none') {
909
- throw new Error('Tiered quant attention requires cold quant mode.');
910
- }
911
- if (!kvState.coldScalesK || !kvState.coldScalesV) {
912
- throw new Error('Tiered quant attention requires cold scale buffers.');
913
- }
914
-
915
- const output = await runAttentionTieredQuant(
916
- qForAttention,
917
- cachedHotKTensor,
918
- cachedHotVTensor,
919
- kvState.cachedKCold,
920
- kvState.cachedVCold,
921
- kvState.coldScalesK,
922
- kvState.coldScalesV,
923
- numHeads,
924
- headDim,
925
- buildTieredQuantAttentionOptions(kvState, {
926
- seqLen: numTokens,
927
- numKVHeads,
928
- causal: causalForAttention,
929
- startPos: kvState.startPosForMask,
930
- slidingWindow: effectiveSlidingWindow ?? 0,
931
- attnSoftcap,
932
- scale: attnScale,
933
- })
934
- );
935
-
936
- if (qTemp) {
937
- releaseBuffer(qTemp.buffer);
938
- }
939
- return output;
940
- },
941
- contiguousQuant: async () => {
942
- let qForAttention = qTensor;
943
- let qTemp = null;
944
- if (qTensor.dtype !== 'f32') {
945
- assertAttentionDtypeTransitionAllowed(state, qTensor.dtype, 'f32', 'Q would be widened implicitly for contiguous quant attention.');
946
- qForAttention = await castF16ToF32(qTensor);
947
- qTemp = qForAttention;
948
- }
949
-
950
- if (!kvState.coldScalesK || !kvState.coldScalesV) {
951
- throw new Error('Contiguous quant attention requires scale buffers.');
952
- }
953
- if (!kvState.rotationMatrixBuffer || !kvState.codebookCentroidsBuffer) {
954
- throw new Error('Contiguous quant attention requires TurboQuant shared buffers.');
955
- }
956
-
957
- const output = await runAttentionContiguousQuant(
958
- qForAttention,
959
- kvState.cachedKCold,
960
- kvState.cachedVCold,
961
- kvState.coldScalesK,
962
- kvState.coldScalesV,
963
- numHeads,
964
- headDim,
965
- buildContiguousQuantAttentionOptions(kvState, {
966
- seqLen: numTokens,
967
- kvLen: kvState.kvLenForAttention,
968
- numKVHeads,
969
- causal: causalForAttention,
970
- startPos: kvState.startPosForMask,
971
- slidingWindow: effectiveSlidingWindow ?? 0,
972
- attnSoftcap,
973
- scale: attnScale,
974
- })
975
- );
976
-
977
- if (qTemp) {
978
- releaseBuffer(qTemp.buffer);
979
- }
980
- return output;
981
- },
982
- tiered: async () => {
983
- const qForAttention = qTensor;
984
- const cachedHotKTensor = createTensor(kvState.cachedKHot, cachedKDtype, [kvState.hotLen, numKVHeads * headDim], 'cached_K_hot');
985
- const cachedHotVTensor = createTensor(kvState.cachedVHot, cachedVDtype, [kvState.hotLen, numKVHeads * headDim], 'cached_V_hot');
986
- const cachedColdKTensor = createTensor(kvState.cachedKCold, cachedKDtype, [kvState.coldLen, numKVHeads * headDim], 'cached_K_cold');
987
- const cachedColdVTensor = createTensor(kvState.cachedVCold, cachedVDtype, [kvState.coldLen, numKVHeads * headDim], 'cached_V_cold');
988
- return runAttentionTiered(qForAttention, cachedHotKTensor, cachedHotVTensor, cachedColdKTensor, cachedColdVTensor, numHeads, headDim, {
989
- seqLen: numTokens,
990
- coldLen: kvState.coldLen,
991
- hotLen: kvState.hotLen,
992
- numKVHeads,
993
- causal: causalForAttention,
994
- startPos: kvState.startPosForMask,
995
- slidingWindow: effectiveSlidingWindow ?? 0,
996
- attnSoftcap,
997
- scale: attnScale,
998
- hotWindow: kvState.hotWindow,
999
- hotStart: kvState.hotStart,
1000
- coldPageTable: kvState.coldPageTable,
1001
- coldPageSize: kvState.coldPageSize,
1002
- coldLayout: kvState.coldPageTable ? 2 : 0,
1003
- hotLayout: kvState.hotWindow > 0 ? 1 : 0,
1004
- });
1005
- },
1006
- contiguous: async () => {
1007
- // Prefill fallback: quantized/tiered layouts use raw K/V for prefill, cast to f16 to match kernel path
1008
- let kForAttn = cachedKTensor;
1009
- let vForAttn = cachedVTensor;
1010
- if (prefillFallbackNeedsCast) {
1011
- const hasExplicitF16KvContract = isAttentionKvDtypeExplicit(attentionPrecisionContract, 'f16');
1012
- if (cachedKDtype === 'f16' && kTensor.dtype !== 'f16' && !hasExplicitF16KvContract) {
1013
- assertAttentionDtypeTransitionAllowed(state, kTensor.dtype, 'f16', 'K would be narrowed implicitly for prefill fallback.');
1014
- }
1015
- if (cachedVDtype === 'f16' && vTensor.dtype !== 'f16' && !hasExplicitF16KvContract) {
1016
- assertAttentionDtypeTransitionAllowed(state, vTensor.dtype, 'f16', 'V would be narrowed implicitly for prefill fallback.');
1017
- }
1018
- const kCasted = cachedKDtype === 'f16' && kTensor.dtype !== 'f16'
1019
- ? await castF32ToF16(kTensor) : kTensor;
1020
- const vCasted = cachedVDtype === 'f16' && vTensor.dtype !== 'f16'
1021
- ? await castF32ToF16(vTensor) : vTensor;
1022
- kForAttn = createTensor(kCasted.buffer, kCasted.dtype, [kvState.kvLenForAttention, numKVHeads * headDim], 'cached_K');
1023
- vForAttn = createTensor(vCasted.buffer, vCasted.dtype, [kvState.kvLenForAttention, numKVHeads * headDim], 'cached_V');
1024
- }
1025
- // Session precedence is runtime-over-manifest per config-style-guide
1026
- // §Category Rules: `resolvedSession = merge(manifest.session, runtime.session)`
1027
- // with runtime fields winning. getRuntimeConfig() returns that merged
1028
- // session (state.runtimeConfig.inference.session is a stale snapshot —
1029
- // see comment above at the rmsNorm fusion check). The kernel itself
1030
- // enforces remaining preconditions (head_dim=256, etc.).
1031
- const useFlashPrefill = !diffusionGemmaDecoder && mergedSession?.useFlashPrefillAttention === true && numTokens > 1;
1032
- const useOrtFlashPrefill = !diffusionGemmaDecoder && mergedSession?.useOrtFlashPrefillAttention === true && numTokens > 1;
1033
- const useOutputGateFusion = canUseAttentionOutputGateFusion({
1034
- session: mergedSession,
1035
- qGateTensor,
1036
- numTokens,
1037
- numHeads,
1038
- headDim,
1039
- cachedKDtype,
1040
- cachedVDtype,
1041
- kernelPath,
1042
- diffusionGemmaDecoder,
1043
- });
1044
- const result = await runAttention(qTensor, kForAttn, vForAttn, null, numHeads, headDim, {
1045
- seqLen: numTokens,
1046
- kvLen: kvState.kvLenForAttention,
1047
- numKVHeads,
1048
- causal: causalForAttention,
1049
- bidirectionalSpanStart: multimodalBidirectionalSpan?.start ?? 0,
1050
- bidirectionalSpanLength: multimodalBidirectionalSpan?.length ?? 0,
1051
- startPos: kvState.startPosForMask,
1052
- layerIdx,
1053
- slidingWindow: effectiveSlidingWindow,
1054
- attnSoftcap,
1055
- scale: attnScale,
1056
- kvStart: kvState.kvStart,
1057
- kvLayout: kvState.kvLayout,
1058
- kvPageTable: kvState.kvPageTable,
1059
- kvPageSize: kvState.kvPageSize,
1060
- kernelPath,
1061
- useFlashPrefill,
1062
- useOrtFlashPrefill,
1063
- outputGate: useOutputGateFusion ? qGateTensor : null,
1064
- });
1065
- attentionOutputGateFused = result?.outputGateFused === true;
1066
- if (prefillFallbackNeedsCast) {
1067
- if (kTensor.dtype !== 'f16') releaseBuffer(kForAttn.buffer);
1068
- if (vTensor.dtype !== 'f16') releaseBuffer(vForAttn.buffer);
1069
- }
1070
- return result;
1071
- },
1072
- };
1073
-
1074
- const runAttentionKernel = attentionKernelRunners[attentionKernelVariant];
1075
- if (!runAttentionKernel) {
1076
- throw new Error(`Unsupported attention kernel variant "${attentionKernelVariant}" at layer ${layerIdx}`);
1077
- }
1078
-
1079
- try {
1080
- attnOutput = await runAttentionKernel();
1081
- } finally {
1082
- if (decoderKVState?.ownedBuffers) {
1083
- for (const buffer of decoderKVState.ownedBuffers) {
1084
- releaseBuffer(buffer);
1085
- }
1086
- decoderKVState.ownedBuffers = null;
1087
- }
1088
- }
1089
- await runProbes('attn_core_out', attnOutput.buffer, {
1090
- layerIdx,
1091
- numTokens,
1092
- hiddenSize: numHeads * headDim,
1093
- probes: state.debugProbes,
1094
- operatorDiagnostics: state.operatorDiagnostics,
1095
- dtype: attnOutput.dtype,
1096
- });
1097
-
1098
- // Trace attention output
1099
- if (kernelTrace.enabled) {
1100
- await traceStep('attention', `L${layerIdx}.attention`, layerIdx, attnOutput.buffer, [numTokens, numHeads * headDim]);
1101
- }
1102
-
1103
- // Kernel step debug: attention output
1104
- if (isKernelDebugEnabled(layerIdx)) {
1105
- logKernelStep('attention', { layerIdx, label: `seqLen=${numTokens} kvLen=${kvState.kvLenForAttention}` });
1106
- await dumpTokenVector(attnOutput.buffer, 'attn_out', {
1107
- layerIdx,
1108
- tokenIdx: Math.max(0, numTokens - 1),
1109
- rowSize: numHeads * headDim,
1110
- dtype: attnOutput.dtype,
1111
- });
1112
- }
1113
-
1114
- // Debug: attention output for configured layers
1115
- if (isPrefill && shouldDebugLayer(layerIdx, debugFlags.debugLayers) && !markStageLogged(layerIdx, 'attn_out', debugFlags) && debugCheckBuffer) {
1116
- await debugCheckBuffer(attnOutput.buffer, `L${layerIdx} attention output (before o_proj, GPU)`, numTokens, numHeads * headDim);
1117
- }
1118
-
1119
- attnForProjection = attnOutput;
1120
- if (qGateTensor && !attentionOutputGateFused) {
1121
- // The Qwen3_5 reference (HF transformers) full-attention path
1122
- // applies sigmoid(gate) — NOT silu(gate) — to the attention
1123
- // output regardless of how the architecture's gate label is
1124
- // surfaced in the HF config. See Qwen3_5Attention.forward in
1125
- // transformers/models/qwen3_5/modeling_qwen3_5.py:
1126
- // attn_output = attn_output * torch.sigmoid(gate)
1127
- // The earlier `swish` -> `silu` mapping here was incorrect and
1128
- // produced sign-flipped output for any token whose gate value
1129
- // was negative (silu is sign-preserving via x * sigmoid(x);
1130
- // sigmoid is unconditionally in (0, 1)). At Qwen 3.6 27B scale
1131
- // this drove residual-stream divergence by L=3 (HF reference vs
1132
- // upstream rel_l2≈1.37, mean per-token correlation ≈ -0.96 — a
1133
- // sign-flip pattern). config.outputGateType is retained for
1134
- // future architectures whose gate truly is silu, but Qwen 3.6's
1135
- // `output_gate_type=swish` still routes here as sigmoid because
1136
- // that is what the reference forward computes.
1137
- const gateActivation = 'sigmoid';
1138
- attnForProjection = await runSiLU(attnOutput, {
1139
- size: numTokens * numHeads * headDim,
1140
- gate: qGateTensor,
1141
- useVec4: (numTokens * numHeads * headDim) % 4 === 0,
1142
- gateActivation,
1143
- inputActivation: 'identity',
1144
- swigluLimit: null,
1145
- });
1146
- releaseBuffer(attnOutput.buffer);
1147
- }
1148
-
1149
- // 6. Output projection (with optional fused residual for decode)
1150
-
1151
- output = null;
1152
- let residualFused = false;
1153
- let oProjInput = attnForProjection;
1154
- oProjInputTemp = null;
1155
- if (layerWeights.oProj && getWeightBuffer) {
1156
- ({ oProjInput, oProjInputTemp } = await prepareAttentionProjectionInput(
1157
- attnForProjection,
1158
- oProjInputDtype,
1159
- async (tensor) => {
1160
- assertAttentionDtypeTransitionAllowed(
1161
- state,
1162
- tensor.dtype,
1163
- oProjInputDtype,
1164
- 'o_proj input would change dtype implicitly.',
1165
- 'step_precision'
1166
- );
1167
- return oProjInputDtype === 'f16'
1168
- ? castF32ToF16(tensor)
1169
- : castF16ToF32(tensor);
1170
- }
1171
- ));
1172
- const oProjBuf = getWeightBuffer(layerWeights.oProj, 'o_proj');
1173
- const loraO = getLoRAModule(lora, layerIdx, 'o_proj');
1174
-
1175
- // Use fused o_proj + residual for decode when possible
1176
- // Note: dtype from WeightBuffer metadata (buffer-dtypes WeakMap removed)
1177
- const oProjDtype = getWeightDtype(oProjBuf);
1178
- const canUseFused = selectRuleValue('inference', 'attention', 'useFusedOProjResidual', {
1179
- allowFusedResidual: shouldUseFusedMatmulResidual(numTokens),
1180
- hasResidual: Boolean(residualTensor),
1181
- residualMatches: Boolean(residualTensor && residualTensor.dtype === oProjInput.dtype),
1182
- attnIsF32: oProjInput.dtype === 'f32',
1183
- attnIsF16: oProjInput.dtype === 'f16',
1184
- hasLoRA: Boolean(loraO),
1185
- oProjIsF16: oProjDtype === 'f16',
1186
- }); // GEMV kernel expects f16 weights
1187
- const canUseWideTileResidual = canUseQ4KWideTileResidualFusion({
1188
- phase,
1189
- session: mergedSession,
1190
- hasResidual: Boolean(residualTensor),
1191
- residualMatches: Boolean(residualTensor && residualTensor.dtype === oProjInput.dtype),
1192
- inputDtype: oProjInput.dtype,
1193
- outputDtype: oProjOutputDtype,
1194
- oProjDtype,
1195
- hasLoRA: Boolean(loraO),
1196
- });
1197
- if (canUseFused && residualTensor) {
1198
- // FUSED PATH: o_proj matmul + residual add in one dispatch
1199
- output = await runMatmulResidualFused(oProjInput, oProjBuf, residualTensor, {
1200
- N: hiddenSize,
1201
- K: numHeads * headDim,
1202
- });
1203
- residualFused = true;
1204
-
1205
- if (layerIdx === 0 && !isPrefill) {
1206
- trace.attn(layerIdx, `Using fused o_proj+residual path`);
1207
- }
1208
- } else {
1209
- // STANDARD PATH: o_proj matmul only unless Q4K WideTile fuses residual in the epilogue
1210
- output = await runMatmul(oProjInput, oProjBuf, numTokens, hiddenSize, numHeads * headDim, {
1211
- transposeB: 'auto',
1212
- role: 'o_proj',
1213
- layerIdx,
1214
- kernelPath,
1215
- outputDtype: oProjOutputDtype,
1216
- executionPolicies: state.executionPolicies ?? null,
1217
- residualTensor: canUseWideTileResidual ? residualTensor : null,
1218
- });
1219
- residualFused = canUseWideTileResidual;
1220
- }
1221
- // Release temporary buffer if we created it (original was not already on GPU)
1222
- if (!isGpuBufferInstance(layerWeights.oProj) && !isWeightBuffer(layerWeights.oProj)) {
1223
- releaseBuffer(isWeightBuffer(oProjBuf) ? oProjBuf.buffer : oProjBuf);
1224
- }
1225
-
1226
- // Trace output projection
1227
- if (kernelTrace.enabled) {
1228
- await traceStep('matmul', `L${layerIdx}.o_proj${residualFused ? '+residual' : ''}`, layerIdx, output.buffer, [numTokens, hiddenSize]);
1229
- }
1230
-
1231
- // Kernel step debug: output projection
1232
- if (isKernelDebugEnabled(layerIdx)) {
1233
- logKernelStep('matmul', { layerIdx, label: residualFused ? 'O_proj+residual' : 'O_proj', M: numTokens, N: hiddenSize, K: numHeads * headDim });
1234
- await dumpTokenVector(output.buffer, 'o_proj_out', {
1235
- layerIdx,
1236
- tokenIdx: Math.max(0, numTokens - 1),
1237
- rowSize: hiddenSize,
1238
- dtype: output.dtype,
1239
- });
1240
- }
1241
-
1242
- } else {
1243
- output = oProjInput;
1244
- }
1245
-
1246
- // Apply LoRA to output projection if present (only if not using fused path)
1247
- if (!residualFused) {
1248
- const loraO = getLoRAModule(lora, layerIdx, 'o_proj');
1249
- if (loraO && getWeightBuffer) {
1250
- const combined = await applyLoRA(
1251
- oProjInput,
1252
- output,
1253
- loraO,
1254
- { M: numTokens, N: hiddenSize, K: numHeads * headDim },
1255
- getWeightBuffer,
1256
- undefined,
1257
- { kernelPath }
1258
- );
1259
- if (combined.buffer !== output.buffer) {
1260
- releaseBuffer(output.buffer);
1261
- output = combined;
1262
- }
1263
- }
1264
- }
1265
-
1266
- if (layerWeights.oProjBias) {
1267
- const { tensor: oProjBiasTensor, owned } = getVectorTensor(
1268
- layerWeights.oProjBias,
1269
- `L${layerIdx}.o_proj_bias`,
1270
- hiddenSize,
1271
- {},
1272
- debugFlags
1273
- );
1274
- try {
1275
- output = await runBiasAdd(output, oProjBiasTensor, numTokens, hiddenSize, {
1276
- label: `L${layerIdx}.o_proj_bias`,
1277
- layerIdx,
1278
- executionPolicies: state.executionPolicies ?? null,
1279
- });
1280
- } finally {
1281
- if (owned) releaseBuffer(oProjBiasTensor.buffer);
1282
- }
1283
- }
1284
-
1285
- if (oProjInputTemp) {
1286
- releaseBuffer(oProjInputTemp.buffer);
1287
- }
1288
-
1289
- // Debug: o_proj output for configured layers
1290
- if (isPrefill && shouldDebugLayer(layerIdx, debugFlags.debugLayers) && !markStageLogged(layerIdx, 'o_proj', debugFlags) && debugCheckBuffer) {
1291
- await debugCheckBuffer(output.buffer, `L${layerIdx} attention output (after o_proj, GPU)`, numTokens, hiddenSize);
1292
- }
1293
-
1294
- finalOutput = output;
1295
-
1296
- const buffersToRelease = [];
1297
- if (output.buffer !== attnForProjection.buffer) {
1298
- buffersToRelease.push(attnForProjection.buffer);
1299
- }
1300
-
1301
- if (output.dtype !== oProjOutputDtype) {
1302
- assertAttentionDtypeTransitionAllowed(
1303
- state,
1304
- output.dtype,
1305
- oProjOutputDtype,
1306
- 'Attention output would change dtype implicitly at the end of the op.',
1307
- 'step_precision'
1308
- );
1309
- const coercedOutput = oProjOutputDtype === 'f16'
1310
- ? await castF32ToF16(output)
1311
- : await castF16ToF32(output);
1312
- buffersToRelease.push(output.buffer);
1313
- finalOutput = coercedOutput;
1314
- }
1315
-
1316
- // Cleanup
1317
- releaseBuffer(qTensor.buffer);
1318
- if (qGateTensor) {
1319
- releaseBuffer(qGateTensor.buffer);
1320
- }
1321
- if (!reusesSharedKV && !storeSharedKV) {
1322
- releaseBuffer(kTensor.buffer);
1323
- if (vTensor.buffer !== kTensor.buffer) {
1324
- releaseBuffer(vTensor.buffer);
1325
- }
1326
- }
1327
- for (const buffer of buffersToRelease) {
1328
- releaseBuffer(buffer);
1329
- }
1330
-
1331
- return { output: finalOutput, residualFused };
1332
- } catch (error) {
1333
- const released = new Set();
1334
- const releaseOnce = (buffer) => {
1335
- if (!buffer || released.has(buffer)) return;
1336
- released.add(buffer);
1337
- releaseBuffer(buffer);
1338
- };
1339
- if (finalOutput?.buffer && finalOutput.buffer !== output?.buffer) {
1340
- releaseOnce(finalOutput.buffer);
1341
- }
1342
- if (output?.buffer && output.buffer !== attnForProjection?.buffer) {
1343
- releaseOnce(output.buffer);
1344
- }
1345
- if (oProjInputTemp?.buffer) {
1346
- releaseOnce(oProjInputTemp.buffer);
1347
- }
1348
- if (attnForProjection?.buffer && attnForProjection.buffer !== attnOutput?.buffer) {
1349
- releaseOnce(attnForProjection.buffer);
1350
- }
1351
- if (attnOutput?.buffer) {
1352
- releaseOnce(attnOutput.buffer);
1353
- }
1354
- if (qGateTensor?.buffer) {
1355
- releaseOnce(qGateTensor.buffer);
1356
- }
1357
- if (qTensor?.buffer) {
1358
- releaseOnce(qTensor.buffer);
1359
- }
1360
- if (kTensor?.buffer && !reusesSharedKV && !storeSharedKV) {
1361
- releaseOnce(kTensor.buffer);
1362
- }
1363
- if (vTensor?.buffer && !reusesSharedKV && !storeSharedKV) {
1364
- releaseOnce(vTensor.buffer);
1365
- }
1366
- if (normed?.buffer && normed.buffer !== attentionInput?.buffer) {
1367
- releaseOnce(normed.buffer);
1368
- }
1369
- if (attentionInputTemp && attentionInput?.buffer) {
1370
- releaseOnce(attentionInput.buffer);
1371
- }
1372
- if (decoderKTemp?.buffer) {
1373
- releaseOnce(decoderKTemp.buffer);
1374
- }
1375
- if (decoderVTemp?.buffer) {
1376
- releaseOnce(decoderVTemp.buffer);
1377
- }
1378
- if (decoderKVState?.ownedBuffers) {
1379
- for (const buffer of decoderKVState.ownedBuffers) {
1380
- releaseOnce(buffer);
1381
- }
1382
- decoderKVState.ownedBuffers = null;
1383
- }
1384
- throw error;
1385
- }
1386
- }
2
+ export { runLayerAttentionGPU } from './executor-immediate.js';