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
@@ -30,225 +30,17 @@ import {
30
30
  import { watchFinalizedCheckpoints } from './checkpoint-watch.js';
31
31
  import { loadLoRAFromManifest } from '../adapters/lora-loader.js';
32
32
  import { createUploadedTensor } from './tensor-factory.js';
33
- import { stableSortObject } from '../../utils/stable-sort-object.js';
33
+ import { stableSortObject } from '../../formats/stable-sort-object.js';
34
34
  import { LORA_MODULE_ALIASES } from '../../inference/pipelines/text/lora.js';
35
35
  import { loadDistillModelHandle } from './suite.js';
36
36
  import { createDistillStudentRuntimeModelFixture } from './distillation/student-fixture.js';
37
37
  import { f16ToF32Array } from '../../inference/kv-cache/types.js';
38
- import { sha256BytesHex, sha256Hex } from '../../utils/sha256.js';
39
-
40
- const CAUSAL_LM_TEXT_PAIR_RUNNER_KEYS = Object.freeze([
41
- 'gemma-3-270m-it-f16-af32::text-pairs::text_generation',
42
- 'gemma-3-270m-it-q4k-ehf16-af32::text-pairs::text_generation',
43
- 'gemma4-e2b-it::text-pairs::text_generation',
44
- 'gemma-4-e2b-it-q4k-ehf16-af32::text-pairs::text_generation',
45
- 'gemma-4-e2b-it-q4k-ehf16-af32-int4ple::text-pairs::text_generation',
46
- 'qwen-3-5-0-8b-q4k-ehaf16::text-pairs::text_generation',
47
- 'qwen-3-5-2b-q4k-ehaf16::text-pairs::text_generation',
48
- 'qwen-3-5-9b-hf-bf16::text-pairs::text_generation',
49
- 'qwen-3-6-27b-q4k-ehaf16::text-pairs::text_generation',
50
- 'qwen-3-6-27b-q4k-eaf16::text-pairs::text_generation',
51
- ]);
52
-
53
- export const LORA_RUNNER_SUPPORT_CONTRACT = Object.freeze({
54
- supportedBaseModelId: 'training-toy',
55
- supportedDatasetFormat: 'toy_linear_classification_jsonl',
56
- registeredBaseModelIds: Object.freeze([
57
- 'training-toy',
58
- 'gemma-3-270m-it-f16-af32',
59
- 'gemma-3-270m-it-q4k-ehf16-af32',
60
- 'gemma4-e2b-it',
61
- 'gemma-4-e2b-it-q4k-ehf16-af32',
62
- 'gemma-4-e2b-it-q4k-ehf16-af32-int4ple',
63
- 'qwen-3-5-0-8b-q4k-ehaf16',
64
- 'qwen-3-5-2b-q4k-ehaf16',
65
- 'qwen-3-5-9b-hf-bf16',
66
- 'qwen-3-6-27b-q4k-ehaf16',
67
- 'qwen-3-6-27b-q4k-eaf16',
68
- ]),
69
- registeredDatasetFormats: Object.freeze([
70
- 'toy_linear_classification_jsonl',
71
- 'text-pairs',
72
- ]),
73
- implementedRunnerKeys: Object.freeze([
74
- 'training-toy::toy_linear_classification_jsonl::classification',
75
- ...CAUSAL_LM_TEXT_PAIR_RUNNER_KEYS,
76
- ]),
77
- });
78
-
79
- export const LORA_RUNNER_BASE_MODEL_REGISTRY = Object.freeze({
80
- 'training-toy': Object.freeze({
81
- baseModelId: 'training-toy',
82
- family: 'training_fixture',
83
- runnerKind: 'toy_linear_classification',
84
- }),
85
- 'gemma-3-270m-it-f16-af32': Object.freeze({
86
- baseModelId: 'gemma-3-270m-it-f16-af32',
87
- modelRef: 'gemma-3-270m-it-f16-af32',
88
- family: 'gemma3',
89
- runnerKind: 'causal_lm_text_generation',
90
- }),
91
- 'gemma-3-270m-it-q4k-ehf16-af32': Object.freeze({
92
- baseModelId: 'gemma-3-270m-it-q4k-ehf16-af32',
93
- modelRef: 'gemma-3-270m-it-q4k-ehf16-af32',
94
- family: 'gemma3',
95
- runnerKind: 'causal_lm_text_generation',
96
- requiresExternalTrainer: true,
97
- }),
98
- 'gemma4-e2b-it': Object.freeze({
99
- baseModelId: 'gemma4-e2b-it',
100
- modelRef: 'gemma-4-e2b-it-q4k-ehf16-af32',
101
- family: 'gemma4',
102
- runnerKind: 'causal_lm_text_generation',
103
- requiresExternalTrainer: true,
104
- }),
105
- 'gemma-4-e2b-it-q4k-ehf16-af32': Object.freeze({
106
- baseModelId: 'gemma-4-e2b-it-q4k-ehf16-af32',
107
- modelRef: 'gemma-4-e2b-it-q4k-ehf16-af32',
108
- family: 'gemma4',
109
- runnerKind: 'causal_lm_text_generation',
110
- requiresExternalTrainer: true,
111
- }),
112
- 'gemma-4-e2b-it-q4k-ehf16-af32-int4ple': Object.freeze({
113
- baseModelId: 'gemma-4-e2b-it-q4k-ehf16-af32-int4ple',
114
- modelRef: 'gemma-4-e2b-it-q4k-ehf16-af32-int4ple',
115
- family: 'gemma4',
116
- runnerKind: 'causal_lm_text_generation',
117
- requiresExternalTrainer: true,
118
- }),
119
- 'qwen-3-5-0-8b-q4k-ehaf16': Object.freeze({
120
- baseModelId: 'qwen-3-5-0-8b-q4k-ehaf16',
121
- modelRef: 'qwen-3-5-0-8b-q4k-ehaf16',
122
- family: 'qwen3',
123
- runnerKind: 'causal_lm_text_generation',
124
- requiresExternalTrainer: true,
125
- nativeLoraTargets: Object.freeze([
126
- Object.freeze({ module: 'down_proj', layer: 'last' }),
127
- ]),
128
- }),
129
- 'qwen-3-5-2b-q4k-ehaf16': Object.freeze({
130
- baseModelId: 'qwen-3-5-2b-q4k-ehaf16',
131
- modelRef: 'qwen-3-5-2b-q4k-ehaf16',
132
- family: 'qwen3',
133
- runnerKind: 'causal_lm_text_generation',
134
- requiresExternalTrainer: true,
135
- }),
136
- 'qwen-3-5-9b-hf-bf16': Object.freeze({
137
- baseModelId: 'qwen-3-5-9b-hf-bf16',
138
- modelRef: 'Qwen/Qwen3.5-9B',
139
- family: 'qwen3',
140
- runnerKind: 'causal_lm_text_generation',
141
- requiresExternalTrainer: true,
142
- }),
143
- 'qwen-3-6-27b-q4k-ehaf16': Object.freeze({
144
- baseModelId: 'qwen-3-6-27b-q4k-ehaf16',
145
- modelRef: 'qwen-3-6-27b-q4k-ehaf16',
146
- family: 'qwen3',
147
- runnerKind: 'causal_lm_text_generation',
148
- requiresExternalTrainer: true,
149
- }),
150
- 'qwen-3-6-27b-q4k-eaf16': Object.freeze({
151
- baseModelId: 'qwen-3-6-27b-q4k-eaf16',
152
- modelRef: 'qwen-3-6-27b-q4k-eaf16',
153
- family: 'qwen3',
154
- runnerKind: 'causal_lm_text_generation',
155
- requiresExternalTrainer: true,
156
- }),
157
- });
158
-
159
- export const LORA_RUNNER_DATASET_FORMAT_REGISTRY = Object.freeze({
160
- toy_linear_classification_jsonl: Object.freeze({
161
- datasetFormat: 'toy_linear_classification_jsonl',
162
- datasetKind: 'toy_linear_classification',
163
- }),
164
- 'text-pairs': Object.freeze({
165
- datasetFormat: 'text-pairs',
166
- datasetKind: 'causal_lm_text_pairs',
167
- }),
168
- });
169
-
170
- function getPipelineConfig(workload) {
171
- return workload?.pipeline || workload?.lora || {};
172
- }
173
-
174
- function normalizeLoraTargetModules(adapter) {
175
- const rawModules = Array.isArray(adapter?.targetModules) ? adapter.targetModules : [];
176
- const modules = [];
177
- for (const rawModule of rawModules) {
178
- const normalized = String(rawModule || '').trim();
179
- if (!normalized) continue;
180
- const moduleName = LORA_MODULE_ALIASES[normalized] || normalized;
181
- if (!modules.includes(moduleName)) {
182
- modules.push(moduleName);
183
- }
184
- }
185
- if (modules.length === 0) {
186
- throw new Error('Causal-LM LoRA workload requires adapter.targetModules.');
187
- }
188
- return modules;
189
- }
190
-
191
- function getCausalLmBaseModelRef(workload) {
192
- const pipeline = getPipelineConfig(workload);
193
- const baseModel = LORA_RUNNER_BASE_MODEL_REGISTRY[String(workload?.baseModelId || '')] || null;
194
- return String(
195
- pipeline.baseModelRef
196
- || workload.baseModelRef
197
- || workload.studentModelId
198
- || baseModel?.modelRef
199
- || workload.baseModelId
200
- || ''
201
- );
202
- }
203
-
204
- function getRunnerKey(baseModelId, datasetFormat, taskType) {
205
- return `${baseModelId}::${datasetFormat}::${taskType}`;
206
- }
207
-
208
- function isCausalLmLoraWorkload(workload, compatibility = getLoraRunnerCompatibility(workload)) {
209
- return compatibility.observed.baseModelRunnerKind === 'causal_lm_text_generation'
210
- || compatibility.observed.datasetKind === 'causal_lm_text_pairs'
211
- || compatibility.observed.taskType === 'text_generation';
212
- }
213
-
214
- export function getLoraRunnerCompatibility(workload) {
215
- const baseModelId = String(workload?.baseModelId || '');
216
- const pipeline = getPipelineConfig(workload);
217
- const datasetFormat = String(pipeline?.datasetFormat || '');
218
- const taskType = String(pipeline?.taskType || '');
219
- const baseModel = LORA_RUNNER_BASE_MODEL_REGISTRY[baseModelId] || null;
220
- const dataset = LORA_RUNNER_DATASET_FORMAT_REGISTRY[datasetFormat] || null;
221
- const runnerKey = getRunnerKey(baseModelId, datasetFormat, taskType);
222
- const blockedReasons = [];
223
- if (!baseModel) {
224
- blockedReasons.push('base_model_not_registered_for_current_lora_runner');
225
- }
226
- if (!dataset) {
227
- blockedReasons.push('dataset_format_not_supported_by_current_lora_runner');
228
- }
229
- if (baseModel && dataset && !LORA_RUNNER_SUPPORT_CONTRACT.implementedRunnerKeys.includes(runnerKey)) {
230
- blockedReasons.push('runner_combination_not_supported_by_current_lora_runner');
231
- }
232
- return {
233
- schemaVersion: 1,
234
- supported: blockedReasons.length === 0,
235
- runnerContract: LORA_RUNNER_SUPPORT_CONTRACT,
236
- observed: {
237
- baseModelId,
238
- datasetFormat,
239
- taskType,
240
- runnerKey,
241
- baseModelFamily: baseModel?.family || null,
242
- baseModelRunnerKind: baseModel?.runnerKind || null,
243
- requiresExternalTrainer: baseModel?.requiresExternalTrainer === true,
244
- nativeLoraTargets: baseModel?.nativeLoraTargets || Object.freeze([]),
245
- datasetKind: dataset?.datasetKind || null,
246
- registeredBaseModel: Boolean(baseModel),
247
- registeredDatasetFormat: Boolean(dataset),
248
- },
249
- blockedReasons,
250
- };
251
- }
38
+ import { sha256BytesHex, sha256Hex } from '../../formats/sha256.js';
39
+ import { LORA_RUNNER_BASE_MODEL_REGISTRY, finiteMetric, getLoraRunnerCompatibility, getPipelineConfig, isCausalLmLoraWorkload, isObjectRecord, normalizeLoraTargetModules, normalizeProviderEvalReport, preflightCausalLmLoraWorkload } from './lora/plan.js';
40
+ import { buildArtifact, createCausalLmDatasetBatches, evaluateCausalLmLoraModel, evaluateToyLoraModel, exportCausalLmLoraModel, exportProviderCausalLmLoraModel, exportToyLoraModel, loadCausalLmTextPairSamples, loadToyLoraDataset, makeTensorFromFloat32, makeTensorFromUint32, releaseTensor, resolveDatasetPathForLoadedWorkload } from './lora/checkpoint.js';
41
+ import { runInternalCausalLmLoraPipeline } from './lora/execution.js';
42
+ import { buildRunContract, createLoraRunLayout, runProviderCausalLmLoraPipeline } from './lora/recovery.js';
43
+ export { LORA_RUNNER_BASE_MODEL_REGISTRY, LORA_RUNNER_DATASET_FORMAT_REGISTRY, LORA_RUNNER_SUPPORT_CONTRACT, getLoraRunnerCompatibility, preflightCausalLmLoraWorkload } from './lora/plan.js';
252
44
 
253
45
  export function assertLoraRunnerCompatibility(workload) {
254
46
  const compatibility = getLoraRunnerCompatibility(workload);
@@ -266,130 +58,10 @@ export function assertLoraRunnerCompatibility(workload) {
266
58
  ].join(' '));
267
59
  }
268
60
 
269
- function summarizeTextPairLengths(rows) {
270
- let minPromptChars = Number.POSITIVE_INFINITY;
271
- let maxPromptChars = 0;
272
- let minCompletionChars = Number.POSITIVE_INFINITY;
273
- let maxCompletionChars = 0;
274
- for (const row of rows) {
275
- const promptChars = row.prompt.length;
276
- const completionChars = row.completion.length;
277
- minPromptChars = Math.min(minPromptChars, promptChars);
278
- maxPromptChars = Math.max(maxPromptChars, promptChars);
279
- minCompletionChars = Math.min(minCompletionChars, completionChars);
280
- maxCompletionChars = Math.max(maxCompletionChars, completionChars);
281
- }
282
- if (!rows.length) {
283
- minPromptChars = 0;
284
- minCompletionChars = 0;
285
- }
286
- return {
287
- minPromptChars,
288
- maxPromptChars,
289
- minCompletionChars,
290
- maxCompletionChars,
291
- };
292
- }
293
-
294
- async function pathIsReadable(filePath) {
295
- try {
296
- await access(filePath);
297
- return true;
298
- } catch {
299
- return false;
300
- }
301
- }
302
-
303
- async function resolveDatasetPathForLoadedWorkload(datasetPath, loadedWorkload) {
304
- const source = String(datasetPath || '');
305
- if (!source || isAbsolute(source) || /^https?:\/\//i.test(source)) {
306
- return source;
307
- }
308
- const candidates = [];
309
- const pushCandidate = (candidate) => {
310
- if (candidate && !candidates.includes(candidate)) {
311
- candidates.push(candidate);
312
- }
313
- };
314
- pushCandidate(resolve(source));
315
- let cursor = loadedWorkload?.absolutePath ? dirname(resolve(String(loadedWorkload.absolutePath))) : null;
316
- while (cursor) {
317
- pushCandidate(join(cursor, source));
318
- const parent = dirname(cursor);
319
- if (parent === cursor) break;
320
- cursor = parent;
321
- }
322
- for (const candidate of candidates) {
323
- if (await pathIsReadable(candidate)) {
324
- return candidate;
325
- }
326
- }
327
- return resolve(source);
328
- }
329
-
330
- export async function preflightCausalLmLoraWorkload(workload, options = {}) {
331
- const compatibility = getLoraRunnerCompatibility(workload);
332
- if (!isCausalLmLoraWorkload(workload, compatibility)) {
333
- throw new Error('preflightCausalLmLoraWorkload requires a causal-LM LoRA workload.');
334
- }
335
- if (!workload?.datasetPath) {
336
- throw new Error('preflightCausalLmLoraWorkload requires workload.datasetPath.');
337
- }
338
- const datasetPath = options.datasetPath || workload.datasetPath;
339
- const dataset = await loadTextPairsDataset(datasetPath, {
340
- fetch: options.fetch,
341
- readFile: options.readFile,
342
- });
343
- if (dataset.rowCount < 1) {
344
- throw new Error(`Causal-LM LoRA dataset ${workload.datasetPath} has no rows.`);
345
- }
346
- const pipeline = getPipelineConfig(workload);
347
- const adapter = pipeline.adapter || {};
348
- return {
349
- schemaVersion: 1,
350
- supported: compatibility.supported,
351
- runnerKey: compatibility.observed.runnerKey,
352
- baseModelId: compatibility.observed.baseModelId,
353
- baseModelFamily: compatibility.observed.baseModelFamily,
354
- datasetPath: dataset.absolutePath,
355
- datasetFormat: compatibility.observed.datasetFormat,
356
- taskType: compatibility.observed.taskType,
357
- rowCount: dataset.rowCount,
358
- firstRowId: dataset.rows[0]?.id || null,
359
- lastRowId: dataset.rows[dataset.rows.length - 1]?.id || null,
360
- textPairFields: {
361
- prompt: dataset.rows[0]?.promptField || null,
362
- completion: dataset.rows[0]?.completionField || null,
363
- },
364
- textPairLengths: summarizeTextPairLengths(dataset.rows),
365
- adapter: {
366
- rank: Number(adapter.rank ?? 0),
367
- alpha: Number(adapter.alpha ?? 0),
368
- targetModules: normalizeLoraTargetModules(adapter),
369
- },
370
- blockedReasons: compatibility.blockedReasons.slice(),
371
- };
372
- }
373
-
374
61
  function stableJson(value) {
375
62
  return JSON.stringify(stableSortObject(value));
376
63
  }
377
64
 
378
- function makeTensorFromFloat32(values, shape, label) {
379
- const data = values instanceof Float32Array ? values : new Float32Array(values);
380
- return createUploadedTensor(data, 'f32', shape, label);
381
- }
382
-
383
- function makeTensorFromUint32(values, shape, label) {
384
- const data = values instanceof Uint32Array ? values : new Uint32Array(values);
385
- return createUploadedTensor(data, 'u32', shape, label);
386
- }
387
-
388
- function releaseTensor(tensor) {
389
- if (!tensor?.buffer) return;
390
- releaseBuffer(tensor.buffer);
391
- }
392
-
393
65
  function createToyLoraModel(workload) {
394
66
  const targetModule = workload.pipeline.adapter.targetModules[0];
395
67
  if (!targetModule) {
@@ -449,52 +121,6 @@ function createToyLoraModel(workload) {
449
121
  };
450
122
  }
451
123
 
452
- function normalizeToyRow(record, index) {
453
- if (!record || typeof record !== 'object' || Array.isArray(record)) {
454
- throw new Error(`LoRA toy dataset row ${index + 1} must be an object.`);
455
- }
456
- const values = Array.isArray(record.input)
457
- ? record.input
458
- : (Array.isArray(record.features) ? record.features : null);
459
- if (!Array.isArray(values) || values.length !== 3) {
460
- throw new Error(`LoRA toy dataset row ${index + 1} requires input[3].`);
461
- }
462
- const input = values.map((value, valueIndex) => {
463
- const parsed = Number(value);
464
- if (!Number.isFinite(parsed)) {
465
- throw new Error(`LoRA toy dataset row ${index + 1} input[${valueIndex}] must be finite.`);
466
- }
467
- return parsed;
468
- });
469
- const target = Number(record.target ?? record.label);
470
- if (!Number.isInteger(target) || target < 0 || target > 1) {
471
- throw new Error(`LoRA toy dataset row ${index + 1} requires integer target 0 or 1.`);
472
- }
473
- return {
474
- id: String(record.id || `row-${index + 1}`),
475
- input,
476
- target,
477
- };
478
- }
479
-
480
- async function loadToyLoraDataset(datasetPath) {
481
- const absolutePath = resolve(String(datasetPath));
482
- const raw = await readFile(absolutePath, 'utf8');
483
- const rows = absolutePath.endsWith('.json')
484
- ? JSON.parse(raw)
485
- : parseJsonl(raw);
486
- if (!Array.isArray(rows)) {
487
- throw new Error(`LoRA dataset "${absolutePath}" must be a JSON array or JSONL file.`);
488
- }
489
- const normalizedRows = rows.map((row, index) => normalizeToyRow(row, index));
490
- return {
491
- absolutePath,
492
- raw,
493
- rows: normalizedRows,
494
- datasetHash: hashArtifactPayload({ rows: normalizedRows }),
495
- };
496
- }
497
-
498
124
  function createToyDatasetBatches(rows, batchSize) {
499
125
  return {
500
126
  async *batches() {
@@ -533,1579 +159,11 @@ function createToyDatasetBatches(rows, batchSize) {
533
159
  };
534
160
  }
535
161
 
536
- function createLossGradient(loss, lossScale) {
537
- const lossElements = loss.shape.reduce((acc, value) => acc * value, 1);
538
- const gradData = new Float32Array(lossElements);
539
- gradData.fill(lossScale);
540
- return createUploadedTensor(gradData, 'f32', loss.shape, 'lora_causal_lm_loss_grad');
541
- }
542
-
543
- function createCausalLmTrainingObjective() {
544
- return {
545
- name: 'causal_lm_cross_entropy',
546
- async forward({ model, batch, tape }) {
547
- if (typeof model.forwardCausalLm === 'function') {
548
- return model.forwardCausalLm(batch, tape);
549
- }
550
- const logits = await model.forward(batch.input, tape);
551
- return { logits };
552
- },
553
- async computeLoss({ batch, config, tape, forwardState }) {
554
- const loss = await crossEntropyLoss(forwardState.logits, batch.targets, config, tape);
555
- return {
556
- loss,
557
- components: {
558
- supervised_token_count: batch.supervisedTokenCount,
559
- ignored_target_count: batch.ignoredTargetCount,
560
- },
561
- };
562
- },
563
- backwardTargets({ batch, loss, lossScale }) {
564
- if (!Number.isInteger(batch.supervisedTokenCount) || batch.supervisedTokenCount < 1) {
565
- throw new Error('Causal-LM training batch requires supervisedTokenCount >= 1.');
566
- }
567
- return createLossGradient(loss, lossScale / batch.supervisedTokenCount);
568
- },
569
- };
570
- }
571
-
572
- function createCausalLmDatasetBatches(samples) {
573
- return {
574
- async *batches() {
575
- for (const sample of samples) {
576
- const inputTensor = makeTensorFromUint32(
577
- sample.inputIds,
578
- [sample.inputIds.length],
579
- `lora_causal_lm_input_${sample.id}`
580
- );
581
- const targetTensor = makeTensorFromUint32(
582
- sample.targetIds,
583
- [sample.targetIds.length],
584
- `lora_causal_lm_target_${sample.id}`
585
- );
586
- try {
587
- yield {
588
- id: sample.id,
589
- input: inputTensor,
590
- targets: targetTensor,
591
- prompt: sample.prompt,
592
- completion: sample.completion,
593
- supervisedTokenCount: sample.supervisedTokenCount,
594
- ignoredTargetCount: sample.ignoredTargetCount,
595
- };
596
- } finally {
597
- releaseTensor(inputTensor);
598
- releaseTensor(targetTensor);
599
- }
600
- }
601
- },
602
- };
603
- }
604
-
605
- async function loadCausalLmTextPairSamples(workload, datasetPath, tokenizer) {
606
- const dataset = await loadTextPairsDataset(datasetPath);
607
- const pipeline = getPipelineConfig(workload);
608
- const maxLength = Math.floor(Number(
609
- pipeline.maxLength
610
- ?? pipeline.sequenceLength
611
- ?? workload.training?.maxLength
612
- ));
613
- if (!Number.isInteger(maxLength) || maxLength < 2) {
614
- throw new Error('Causal-LM LoRA workload requires lora.maxLength or lora.sequenceLength >= 2.');
615
- }
616
- if (typeof pipeline.joinWith !== 'string') {
617
- throw new Error('Causal-LM LoRA workload requires lora.joinWith.');
618
- }
619
- const joinWith = pipeline.joinWith;
620
- const samples = await tokenizeTextPairs(tokenizer, dataset.rows, {
621
- maxLength,
622
- joinWith,
623
- });
624
- if (samples.length < 1) {
625
- throw new Error(`Causal-LM LoRA dataset ${dataset.absolutePath} produced no tokenized samples.`);
626
- }
627
- const tokenizationContract = {
628
- version: 2,
629
- objective: 'completion_only_causal_lm',
630
- truncation: 'preserve_completion_head_tail_prompt',
631
- maxLength,
632
- joinWith,
633
- };
634
- const supervisedTokenCount = samples.reduce(
635
- (sum, sample) => sum + sample.supervisedTokenCount,
636
- 0
637
- );
638
- const ignoredTargetCount = samples.reduce(
639
- (sum, sample) => sum + sample.ignoredTargetCount,
640
- 0
641
- );
642
- const truncatedPromptTokenCount = samples.reduce(
643
- (sum, sample) => sum + sample.truncatedPromptTokenCount,
644
- 0
645
- );
646
- return {
647
- ...dataset,
648
- samples,
649
- datasetHash: hashArtifactPayload({
650
- rows: dataset.rows,
651
- tokenization: tokenizationContract,
652
- }),
653
- tokenization: {
654
- ...tokenizationContract,
655
- sampleCount: samples.length,
656
- supervisedTokenCount,
657
- ignoredTargetCount,
658
- truncatedPromptTokenCount,
659
- truncatedSampleCount: samples.filter(
660
- (sample) => sample.truncatedPromptTokenCount > 0
661
- ).length,
662
- },
663
- };
664
- }
665
-
666
- async function createCausalLmLoraFixture(workload) {
667
- const pipeline = getPipelineConfig(workload);
668
- const adapter = pipeline.adapter || {};
669
- const modelRef = getCausalLmBaseModelRef(workload);
670
- const handle = await loadDistillModelHandle(modelRef, 'lora base', {
671
- runtime: {
672
- shared: {
673
- debug: {
674
- logLevel: {
675
- defaultLogLevel: 'debug',
676
- },
677
- },
678
- },
679
- inference: {
680
- compute: {
681
- activationDtype: 'f32',
682
- keepF32Weights: true,
683
- },
684
- },
685
- },
686
- });
687
- let fixture = null;
688
- try {
689
- fixture = await createDistillStudentRuntimeModelFixture({
690
- training: {
691
- precision: workload.training?.precision || {},
692
- },
693
- }, {
694
- distillRuntime: {
695
- studentPipeline: handle.pipeline,
696
- studentGraphMode: 'transformer_full',
697
- },
698
- studentGraphMode: 'transformer_full',
699
- loraAdapter: {
700
- rank: adapter.rank,
701
- alpha: adapter.alpha,
702
- targetModules: normalizeLoraTargetModules(adapter),
703
- },
704
- });
705
- } catch (error) {
706
- if (handle.pipeline && typeof handle.pipeline.unload === 'function') {
707
- await handle.pipeline.unload();
708
- }
709
- throw error;
710
- }
711
- return {
712
- model: fixture.model,
713
- baseModelRef: modelRef,
714
- baseModelUrl: handle.modelUrl || null,
715
- baseManifest: handle.manifest || null,
716
- tokenizer: handle.pipeline.tokenizer,
717
- cleanup() {
718
- fixture.cleanup();
719
- if (handle.pipeline && typeof handle.pipeline.unload === 'function') {
720
- return handle.pipeline.unload();
721
- }
722
- return undefined;
723
- },
724
- };
725
- }
726
-
727
- function collectProtectedBuffers(model) {
728
- const protectedBuffers = new Set();
729
- const groups = model.paramGroups();
730
- for (const params of Object.values(groups)) {
731
- for (const tensor of params) {
732
- if (tensor?.buffer) {
733
- protectedBuffers.add(tensor.buffer);
734
- }
735
- }
736
- }
737
- return protectedBuffers;
738
- }
739
-
740
- async function captureLoraParameterState(model) {
741
- const entries = typeof model?.loraTensorEntries === 'function'
742
- ? model.loraTensorEntries()
743
- : [];
744
- if (entries.length === 0) {
745
- throw new Error('Causal-LM LoRA parameter receipt requires adapter tensors.');
746
- }
747
- const tensors = [];
748
- for (const entry of [...entries].sort((left, right) => left.name.localeCompare(right.name))) {
749
- const tensor = entry.tensor;
750
- const elementCount = tensor.shape.reduce((product, value) => product * value, 1);
751
- const bytesPerElement = tensor.dtype === 'f16' ? 2 : 4;
752
- const raw = await readBuffer(tensor.buffer, elementCount * bytesPerElement);
753
- const values = tensor.dtype === 'f16'
754
- ? f16ToF32Array(new Uint16Array(raw))
755
- : new Float32Array(raw);
756
- let nonzeroCount = 0;
757
- let sumSquares = 0;
758
- let maxAbs = 0;
759
- for (const value of values) {
760
- if (value !== 0) nonzeroCount += 1;
761
- sumSquares += value * value;
762
- maxAbs = Math.max(maxAbs, Math.abs(value));
763
- }
764
- tensors.push({
765
- name: entry.name,
766
- dtype: tensor.dtype,
767
- shape: [...tensor.shape],
768
- elementCount,
769
- hash: sha256BytesHex(new Uint8Array(raw)),
770
- nonzeroCount,
771
- maxAbs,
772
- l2Norm: Math.sqrt(sumSquares),
773
- values,
774
- });
775
- }
776
- return {
777
- aggregateHash: sha256Hex(tensors.map((tensor) => `${tensor.name}\0${tensor.hash}\n`).join('')),
778
- tensors,
779
- };
780
- }
781
-
782
- function buildLoraParameterReceipt(initial, final) {
783
- const initialByName = new Map(initial.tensors.map((tensor) => [tensor.name, tensor]));
784
- let totalSquaredDelta = 0;
785
- const tensors = final.tensors.map((finalTensor) => {
786
- const initialTensor = initialByName.get(finalTensor.name);
787
- if (!initialTensor || initialTensor.elementCount !== finalTensor.elementCount) {
788
- throw new Error(`LoRA parameter receipt shape mismatch for ${finalTensor.name}.`);
789
- }
790
- let squaredDelta = 0;
791
- let maxAbsDelta = 0;
792
- for (let index = 0; index < finalTensor.values.length; index += 1) {
793
- const delta = finalTensor.values[index] - initialTensor.values[index];
794
- squaredDelta += delta * delta;
795
- maxAbsDelta = Math.max(maxAbsDelta, Math.abs(delta));
796
- }
797
- totalSquaredDelta += squaredDelta;
798
- return {
799
- name: finalTensor.name,
800
- dtype: finalTensor.dtype,
801
- shape: finalTensor.shape,
802
- elementCount: finalTensor.elementCount,
803
- initialHash: initialTensor.hash,
804
- finalHash: finalTensor.hash,
805
- changed: initialTensor.hash !== finalTensor.hash,
806
- initialNonzeroCount: initialTensor.nonzeroCount,
807
- finalNonzeroCount: finalTensor.nonzeroCount,
808
- l2Delta: Math.sqrt(squaredDelta),
809
- maxAbsDelta,
810
- };
811
- });
812
- return {
813
- tensorCount: tensors.length,
814
- changedTensorCount: tensors.filter((tensor) => tensor.changed).length,
815
- nonzeroFinalTensorCount: tensors.filter((tensor) => tensor.finalNonzeroCount > 0).length,
816
- initialAggregateHash: initial.aggregateHash,
817
- finalAggregateHash: final.aggregateHash,
818
- aggregateChanged: initial.aggregateHash !== final.aggregateHash,
819
- l2Delta: Math.sqrt(totalSquaredDelta),
820
- tensors,
821
- };
822
- }
823
-
824
- function disposeTapeOutputs(tape, protectedBuffers = new Set()) {
825
- if (!Array.isArray(tape?.records)) return;
826
- const released = new Set();
827
- for (const record of tape.records) {
828
- const output = record?.output;
829
- if (output?.buffer && !protectedBuffers.has(output.buffer) && !released.has(output.buffer)) {
830
- released.add(output.buffer);
831
- releaseBuffer(output.buffer);
832
- }
833
- }
834
- }
835
-
836
- function argmax(values) {
837
- let bestIndex = 0;
838
- let bestValue = Number.NEGATIVE_INFINITY;
839
- for (let index = 0; index < values.length; index += 1) {
840
- const value = Number.isFinite(values[index]) ? values[index] : Number.NEGATIVE_INFINITY;
841
- if (value > bestValue) {
842
- bestValue = value;
843
- bestIndex = index;
844
- }
845
- }
846
- return bestIndex;
847
- }
848
-
849
- async function evaluateToyLoraModel(workload, model, dataset, layout = null, checkpointMeta = {}) {
850
- const protectedBuffers = collectProtectedBuffers(model);
851
- const evalReports = [];
852
- const evalDatasets = Array.isArray(workload.evalDatasets) ? workload.evalDatasets : [];
853
- for (const evalDataset of evalDatasets) {
854
- if (evalDataset.evalKind !== 'classification' && evalDataset.evalKind !== 'text_generation') {
855
- throw new Error(`LoRA eval currently supports classification/text_generation only, got "${evalDataset.evalKind}".`);
856
- }
857
- const evalDatasetMaterialized = evalDataset.datasetPath === dataset.absolutePath
858
- ? dataset
859
- : await loadToyLoraDataset(evalDataset.datasetPath);
860
- const rows = evalDatasetMaterialized.rows;
861
- const predictions = [];
862
- const labels = [];
863
- for (const row of rows) {
864
- const tape = new AutogradTape(loadBackwardRegistry());
865
- const inputTensor = makeTensorFromFloat32(row.input, [1, 3], 'lora_eval_input');
866
- let logits = null;
867
- try {
868
- logits = await model.forward(inputTensor, tape);
869
- const logitsData = new Float32Array(await readBuffer(logits.buffer));
870
- predictions.push(String(argmax(logitsData)));
871
- labels.push(String(row.target));
872
- } finally {
873
- releaseTensor(inputTensor);
874
- if (logits?.buffer && !protectedBuffers.has(logits.buffer)) {
875
- releaseBuffer(logits.buffer);
876
- }
877
- disposeTapeOutputs(tape, protectedBuffers);
878
- }
879
- }
880
- const metrics = computeEvalMetrics('classification', predictions, labels, {});
881
- const reportPayload = {
882
- artifactType: 'training_eval_report',
883
- schemaVersion: 1,
884
- generatedAt: new Date().toISOString(),
885
- workloadId: workload.id,
886
- workloadPath: checkpointMeta.workloadPath || null,
887
- workloadSha256: checkpointMeta.workloadSha256 || null,
888
- configHash: checkpointMeta.configHash || workload.configHash,
889
- datasetPath: evalDataset.datasetPath,
890
- datasetHash: evalDatasetMaterialized.datasetHash,
891
- baseModelId: workload.baseModelId,
892
- stage: 'lora',
893
- checkpointStep: checkpointMeta.checkpointStep ?? null,
894
- evalDatasetId: evalDataset.id,
895
- metrics,
896
- primaryMetric: metrics.primaryMetric,
897
- primaryScore: metrics.primaryScore,
898
- accuracy: metrics.accuracy?.score ?? null,
899
- };
900
- const reportFile = layout
901
- ? await writeJsonArtifact(
902
- join(layout.eval, `${checkpointMeta.checkpointId || 'checkpoint'}__${evalDataset.id}.json`),
903
- reportPayload
904
- )
905
- : null;
906
- evalReports.push({
907
- ...reportPayload,
908
- reportPath: reportFile?.path || null,
909
- });
910
- }
911
- return evalReports;
912
- }
913
-
914
- function buildRunContract(loadedWorkload) {
915
- return {
916
- artifactType: 'training_run_contract',
917
- schemaVersion: 1,
918
- generatedAt: new Date().toISOString(),
919
- workloadId: loadedWorkload.workload.id,
920
- workloadPath: loadedWorkload.absolutePath,
921
- workloadSha256: loadedWorkload.workloadSha256,
922
- configHash: loadedWorkload.workload.configHash,
923
- claimBoundary: loadedWorkload.workload.claimBoundary,
924
- kind: loadedWorkload.workload.kind,
925
- evalDatasets: loadedWorkload.workload.evalDatasets,
926
- };
927
- }
928
-
929
- function buildArtifact(loadedWorkload, options) {
930
- const workload = loadedWorkload.workload;
931
- const payload = buildArtifactBase({
932
- artifactType: options.artifactType,
933
- reportId: `${options.prefix}_${workload.id}_${options.id}`,
934
- workload,
935
- workloadPath: loadedWorkload.absolutePath,
936
- workloadSha256: loadedWorkload.workloadSha256,
937
- datasetPath: options.datasetPath || workload.datasetPath,
938
- datasetHash: options.datasetHash || null,
939
- baseModelId: workload.baseModelId,
940
- stage: options.stage || 'lora',
941
- checkpointStep: options.checkpointStep ?? null,
942
- parentArtifacts: options.parentArtifacts || [],
943
- runtime: 'node',
944
- surface: 'node',
945
- claimBoundary: workload.claimBoundary,
946
- configHash: options.configHash || workload.configHash,
947
- });
948
- return {
949
- ...payload,
950
- artifactHash: hashArtifactPayload(payload),
951
- };
952
- }
953
-
954
- async function exportToyLoraModel(loadedWorkload, layout, model, checkpointId, checkpointStep, datasetHash) {
955
- const workload = loadedWorkload.workload;
956
- const targetModule = model.targetModule || workload.pipeline.adapter.targetModules[0];
957
- const weightsFilename = `${checkpointId}.adapters.safetensors`;
958
- const exported = await exportLoRAAdapter({
959
- id: workload.pipeline.export?.id || `${workload.id}-${checkpointId}`,
960
- name: workload.pipeline.export?.name || `${workload.id}-${checkpointId}`,
961
- baseModel: workload.baseModelId,
962
- rank: workload.pipeline.adapter.rank,
963
- alpha: workload.pipeline.adapter.alpha,
964
- targetModules: [targetModule],
965
- tensors: [
966
- { name: `layers.0.${targetModule}.lora_a`, tensor: model.adapter.A },
967
- { name: `layers.0.${targetModule}.lora_b`, tensor: model.adapter.B },
968
- ],
969
- weightsFormat: 'safetensors',
970
- weightsPath: weightsFilename,
971
- });
972
- const manifestPath = join(layout.exports, `${checkpointId}.adapter.manifest.json`);
973
- const weightsPath = join(layout.exports, weightsFilename);
974
- if (!exported.weights) {
975
- throw new Error('LoRA safetensors export did not return weights bytes.');
976
- }
977
- await writeFile(weightsPath, new Uint8Array(exported.weights));
978
- await writeFile(manifestPath, exported.json, 'utf8');
979
- await loadLoRAFromManifest(exported.manifest, {
980
- readFile: async (filePath) => readFile(join(layout.exports, filePath)),
981
- });
982
- const artifactPayload = {
983
- ...buildArtifact(loadedWorkload, {
984
- prefix: 'lora_export',
985
- id: checkpointId,
986
- artifactType: 'lora_adapter_manifest',
987
- checkpointStep,
988
- datasetHash,
989
- }),
990
- checkpointId,
991
- manifestPath,
992
- weightsPath,
993
- weightsSha256: exported.weightsSha256 || null,
994
- manifest: exported.manifest,
995
- };
996
- const artifactFile = await writeJsonArtifact(
997
- join(layout.exports, `${checkpointId}.export.json`),
998
- artifactPayload
999
- );
1000
- return {
1001
- checkpointId,
1002
- manifestPath,
1003
- weightsPath,
1004
- weightsSha256: exported.weightsSha256 || null,
1005
- exportPath: artifactFile.path,
1006
- manifest: exported.manifest,
1007
- };
1008
- }
1009
-
1010
- async function exportCausalLmLoraModel(loadedWorkload, layout, fixture, checkpointId, checkpointStep, datasetHash) {
1011
- const workload = loadedWorkload.workload;
1012
- const pipeline = getPipelineConfig(workload);
1013
- const adapter = pipeline.adapter || {};
1014
- const weightsFilename = `${checkpointId}.adapters.safetensors`;
1015
- const tensors = typeof fixture.model.loraTensorEntries === 'function'
1016
- ? fixture.model.loraTensorEntries()
1017
- : [];
1018
- if (!tensors.length) {
1019
- throw new Error('Causal-LM LoRA export requires trained adapter tensors.');
1020
- }
1021
- const exported = await exportLoRAAdapter({
1022
- id: pipeline.export?.id || `${workload.id}-${checkpointId}`,
1023
- name: pipeline.export?.name || `${workload.id}-${checkpointId}`,
1024
- baseModel: workload.baseModelId,
1025
- rank: adapter.rank,
1026
- alpha: adapter.alpha,
1027
- targetModules: normalizeLoraTargetModules(adapter),
1028
- tensors,
1029
- weightsFormat: 'safetensors',
1030
- weightsPath: weightsFilename,
1031
- metadata: {
1032
- baseModelRef: fixture.baseModelRef,
1033
- baseModelUrl: fixture.baseModelUrl,
1034
- datasetFormat: pipeline.datasetFormat,
1035
- taskType: pipeline.taskType,
1036
- },
1037
- });
1038
- const manifestPath = join(layout.exports, `${checkpointId}.adapter.manifest.json`);
1039
- const runtimeManifestPath = join(layout.exports, 'runtime-adapter-manifest.json');
1040
- const weightsPath = join(layout.exports, weightsFilename);
1041
- if (!exported.weights) {
1042
- throw new Error('Causal-LM LoRA safetensors export did not return weights bytes.');
1043
- }
1044
- await writeFile(weightsPath, new Uint8Array(exported.weights));
1045
- await writeFile(manifestPath, exported.json, 'utf8');
1046
- await writeFile(runtimeManifestPath, exported.json, 'utf8');
1047
- await loadLoRAFromManifest(exported.manifest, {
1048
- readFile: async (filePath) => readFile(join(layout.exports, filePath)),
1049
- });
1050
- const artifactPayload = {
1051
- ...buildArtifact(loadedWorkload, {
1052
- prefix: 'lora_export',
1053
- id: checkpointId,
1054
- artifactType: 'lora_adapter_manifest',
1055
- checkpointStep,
1056
- datasetHash,
1057
- }),
1058
- checkpointId,
1059
- manifestPath,
1060
- runtimeManifestPath,
1061
- weightsPath,
1062
- weightsSha256: exported.weightsSha256 || null,
1063
- manifest: exported.manifest,
1064
- };
1065
- const artifactFile = await writeJsonArtifact(
1066
- join(layout.exports, `${checkpointId}.export.json`),
1067
- artifactPayload
1068
- );
1069
- return {
1070
- checkpointId,
1071
- manifestPath,
1072
- runtimeManifestPath,
1073
- weightsPath,
1074
- weightsSha256: exported.weightsSha256 || null,
1075
- exportPath: artifactFile.path,
1076
- manifest: exported.manifest,
1077
- };
1078
- }
1079
-
1080
- async function readSupervisedLossStats(loss, targets, vocabSize) {
1081
- const lossElementCount = loss.shape.reduce((product, value) => product * value, 1);
1082
- const targetElementCount = targets.shape.reduce((product, value) => product * value, 1);
1083
- const [rawLoss, rawTargets] = await Promise.all([
1084
- readBuffer(loss.buffer, lossElementCount * (loss.dtype === 'f16' ? 2 : 4)),
1085
- readBuffer(targets.buffer, targetElementCount * 4),
1086
- ]);
1087
- const data = loss.dtype === 'f16'
1088
- ? f16ToF32Array(new Uint16Array(rawLoss))
1089
- : new Float32Array(rawLoss);
1090
- const targetIds = new Uint32Array(rawTargets);
1091
- if (data.length !== targetIds.length) {
1092
- throw new Error(
1093
- `Causal-LM loss/target length mismatch: ${data.length} losses for ${targetIds.length} targets.`
1094
- );
1095
- }
1096
- let sum = 0;
1097
- let supervisedTokenCount = 0;
1098
- for (let index = 0; index < data.length; index += 1) {
1099
- if (targetIds[index] >= vocabSize) continue;
1100
- const value = data[index];
1101
- if (!Number.isFinite(value)) {
1102
- throw new Error(`Causal-LM loss is non-finite at supervised target ${index}.`);
1103
- }
1104
- sum += value;
1105
- supervisedTokenCount += 1;
1106
- }
1107
- if (supervisedTokenCount < 1) {
1108
- throw new Error('Causal-LM evaluation batch contains no supervised targets.');
1109
- }
1110
- return {
1111
- lossSum: sum,
1112
- meanLoss: sum / supervisedTokenCount,
1113
- supervisedTokenCount,
1114
- ignoredTargetCount: data.length - supervisedTokenCount,
1115
- };
1116
- }
1117
-
1118
- function finiteMetric(value) {
1119
- const parsed = Number(value);
1120
- return Number.isFinite(parsed) ? parsed : null;
1121
- }
1122
-
1123
- function buildLossQualityClaim(evalDataset, meanLoss, baselineReport) {
1124
- const baselineLoss = finiteMetric(baselineReport?.loss ?? baselineReport?.primaryScore);
1125
- const adapterLoss = finiteMetric(meanLoss);
1126
- if (baselineLoss === null || adapterLoss === null) {
1127
- return null;
1128
- }
1129
- const quality = evalDataset.quality || null;
1130
- const absoluteImprovement = baselineLoss - adapterLoss;
1131
- const relativeImprovement = baselineLoss === 0 ? 0 : absoluteImprovement / Math.abs(baselineLoss);
1132
- const minAbsoluteImprovement = finiteMetric(quality?.minAbsoluteImprovement) ?? 0;
1133
- const minRelativeImprovement = finiteMetric(quality?.minRelativeImprovement) ?? 0;
1134
- const improved = absoluteImprovement >= minAbsoluteImprovement
1135
- && relativeImprovement >= minRelativeImprovement;
1136
- return {
1137
- baseline: quality?.baseline || 'base_model',
1138
- metric: 'loss',
1139
- selectionGoal: 'min',
1140
- adapterScore: adapterLoss,
1141
- baselineScore: baselineLoss,
1142
- delta: adapterLoss - baselineLoss,
1143
- absoluteImprovement,
1144
- relativeImprovement,
1145
- minAbsoluteImprovement,
1146
- minRelativeImprovement,
1147
- improved,
1148
- requireImprovement: quality?.requireImprovement === true,
1149
- };
1150
- }
1151
-
1152
- async function computeCausalLmTextPairLosses(workload, fixture, samples, protectedBuffers) {
1153
- const sampleLosses = [];
1154
- let lossSum = 0;
1155
- let supervisedTokenCount = 0;
1156
- let ignoredTargetCount = 0;
1157
- for await (const resolvedBatch of createCausalLmDatasetBatches(samples).batches()) {
1158
- const tape = new AutogradTape(loadBackwardRegistry());
1159
- let loss = null;
1160
- try {
1161
- const { logits } = await fixture.model.forwardCausalLm(resolvedBatch, tape);
1162
- loss = await crossEntropyLoss(logits, resolvedBatch.targets, { training: { precision: workload.training.precision } }, tape);
1163
- const stats = await readSupervisedLossStats(
1164
- loss,
1165
- resolvedBatch.targets,
1166
- logits.shape.at(-1)
1167
- );
1168
- sampleLosses.push(stats.meanLoss);
1169
- lossSum += stats.lossSum;
1170
- supervisedTokenCount += stats.supervisedTokenCount;
1171
- ignoredTargetCount += stats.ignoredTargetCount;
1172
- } finally {
1173
- disposeTapeOutputs(tape, protectedBuffers);
1174
- }
1175
- }
1176
- return {
1177
- sampleLosses,
1178
- sampleCount: sampleLosses.length,
1179
- lossSum,
1180
- supervisedTokenCount,
1181
- ignoredTargetCount,
1182
- meanLoss: supervisedTokenCount > 0 ? lossSum / supervisedTokenCount : null,
1183
- };
1184
- }
1185
-
1186
- async function evaluateCausalLmLoraModel(workload, fixture, dataset, layout = null, checkpointMeta = {}) {
1187
- const protectedBuffers = collectProtectedBuffers(fixture.model);
1188
- const evalReports = [];
1189
- const evalDatasets = Array.isArray(workload.evalDatasets) ? workload.evalDatasets : [];
1190
- for (const evalDataset of evalDatasets) {
1191
- if (evalDataset.evalKind !== 'text_generation') {
1192
- throw new Error(`Causal-LM LoRA eval supports text_generation only, got "${evalDataset.evalKind}".`);
1193
- }
1194
- const evalDatasetPath = await resolveDatasetPathForLoadedWorkload(evalDataset.datasetPath, {
1195
- absolutePath: checkpointMeta.workloadPath || null,
1196
- });
1197
- const evalDatasetMaterialized = evalDataset.datasetPath === dataset.absolutePath
1198
- ? dataset
1199
- : await loadCausalLmTextPairSamples(workload, evalDatasetPath, fixture.tokenizer);
1200
- const lossStats = await computeCausalLmTextPairLosses(
1201
- workload,
1202
- fixture,
1203
- evalDatasetMaterialized.samples,
1204
- protectedBuffers
1205
- );
1206
- const meanLoss = lossStats.meanLoss;
1207
- const baselineReports = checkpointMeta.baselineReportsByEvalDatasetId || {};
1208
- const baselineReport = baselineReports[evalDataset.id] || null;
1209
- const qualityClaim = buildLossQualityClaim(evalDataset, meanLoss, baselineReport);
1210
- const reportPayload = {
1211
- artifactType: 'training_eval_report',
1212
- schemaVersion: 1,
1213
- generatedAt: new Date().toISOString(),
1214
- workloadId: workload.id,
1215
- workloadPath: checkpointMeta.workloadPath || null,
1216
- workloadSha256: checkpointMeta.workloadSha256 || null,
1217
- configHash: checkpointMeta.configHash || workload.configHash,
1218
- datasetPath: evalDataset.datasetPath,
1219
- datasetHash: evalDatasetMaterialized.datasetHash,
1220
- baseModelId: workload.baseModelId,
1221
- baseModelRef: fixture.baseModelRef,
1222
- stage: checkpointMeta.stage || 'lora',
1223
- checkpointStep: checkpointMeta.checkpointStep ?? null,
1224
- checkpointId: checkpointMeta.checkpointId || null,
1225
- evalDatasetId: evalDataset.id,
1226
- metrics: {
1227
- loss: {
1228
- score: meanLoss,
1229
- samples: lossStats.sampleCount,
1230
- supervisedTokens: lossStats.supervisedTokenCount,
1231
- ignoredTargets: lossStats.ignoredTargetCount,
1232
- },
1233
- },
1234
- primaryMetric: 'loss',
1235
- primaryScore: meanLoss,
1236
- loss: meanLoss,
1237
- baseline: baselineReport
1238
- ? {
1239
- checkpointId: baselineReport.checkpointId || null,
1240
- stage: baselineReport.stage || null,
1241
- primaryMetric: baselineReport.primaryMetric || null,
1242
- primaryScore: baselineReport.primaryScore ?? null,
1243
- loss: baselineReport.loss ?? null,
1244
- reportPath: baselineReport.reportPath || null,
1245
- }
1246
- : null,
1247
- qualityClaim,
1248
- };
1249
- const reportFile = layout
1250
- ? await writeJsonArtifact(
1251
- join(layout.eval, `${checkpointMeta.checkpointId || 'checkpoint'}__${evalDataset.id}.json`),
1252
- reportPayload
1253
- )
1254
- : null;
1255
- evalReports.push({
1256
- ...reportPayload,
1257
- reportPath: reportFile?.path || null,
1258
- });
1259
- }
1260
- return evalReports;
1261
- }
1262
-
1263
- function getCausalLmFreezeConfig(workload) {
1264
- const pipeline = getPipelineConfig(workload);
1265
- const freeze = pipeline.freeze;
1266
- if (!freeze || typeof freeze !== 'object' || Array.isArray(freeze)) {
1267
- throw new Error('Causal-LM LoRA workload requires lora.freeze.');
1268
- }
1269
- return {
1270
- encoder: freeze.encoder === true,
1271
- prior: freeze.prior === true,
1272
- decoder: freeze.decoder === true,
1273
- base: freeze.base === true,
1274
- lora: freeze.lora === true,
1275
- };
1276
- }
1277
-
1278
- function createLoraRunnerTrainingConfig(workload, freeze) {
1279
- return {
1280
- training: {
1281
- enabled: true,
1282
- optimizer: {
1283
- type: workload.training.optimizer.type,
1284
- lr: workload.training.optimizer.lr,
1285
- beta1: workload.training.optimizer.beta1,
1286
- beta2: workload.training.optimizer.beta2,
1287
- eps: workload.training.optimizer.eps,
1288
- weightDecay: workload.training.optimizer.weightDecay,
1289
- scheduler: workload.training.optimizer.scheduler,
1290
- },
1291
- gradient: {
1292
- maxNorm: workload.training.gradientClipping.maxNorm,
1293
- },
1294
- precision: workload.training.precision,
1295
- lossScaling: { enabled: false },
1296
- distill: {
1297
- enabled: false,
1298
- stage: 'stage_a',
1299
- teacherModelId: null,
1300
- studentModelId: null,
1301
- datasetId: null,
1302
- datasetPath: null,
1303
- languagePair: null,
1304
- sourceLangs: null,
1305
- targetLangs: null,
1306
- pairAllowlist: null,
1307
- strictPairContract: false,
1308
- shardIndex: null,
1309
- shardCount: null,
1310
- resumeFrom: null,
1311
- artifactDir: null,
1312
- stageAArtifact: null,
1313
- stageAArtifactHash: null,
1314
- temperature: 1,
1315
- alphaKd: 1,
1316
- alphaCe: 0,
1317
- allowHintFallback: false,
1318
- tripletMargin: 0.2,
1319
- studentGraphMode: 'transformer_full',
1320
- freeze,
1321
- },
1322
- ul: {
1323
- enabled: false,
1324
- stage: 'stage1_joint',
1325
- stage1Artifact: null,
1326
- stage1ArtifactHash: null,
1327
- artifactDir: null,
1328
- lambda0: 5,
1329
- seed: workload.seed,
1330
- noiseSchedule: { name: 'linear', minSigma: 0.1, maxSigma: 1, steps: 1 },
1331
- priorAlignment: { enabled: false, weight: 1 },
1332
- decoderSigmoidWeight: { enabled: false, maxWeight: 1 },
1333
- lossWeights: { prior: 1, decoder: 1, recon: 1 },
1334
- freeze: null,
1335
- },
1336
- },
1337
- };
1338
- }
1339
-
1340
- function createLoraOptimizer(workload) {
1341
- return new AdamOptimizer({
1342
- training: {
1343
- optimizer: {
1344
- type: workload.training.optimizer.type,
1345
- lr: workload.training.optimizer.lr,
1346
- beta1: workload.training.optimizer.beta1,
1347
- beta2: workload.training.optimizer.beta2,
1348
- eps: workload.training.optimizer.eps,
1349
- weightDecay: workload.training.optimizer.weightDecay,
1350
- scheduler: workload.training.optimizer.scheduler,
1351
- },
1352
- gradient: {
1353
- maxNorm: workload.training.gradientClipping.maxNorm,
1354
- },
1355
- precision: workload.training.precision,
1356
- },
1357
- });
1358
- }
1359
-
1360
- async function createLoraRunLayout(options, workload) {
1361
- const layout = options.runRoot
1362
- ? {
1363
- runRoot: resolve(String(options.runRoot)),
1364
- logs: join(resolve(String(options.runRoot)), 'logs'),
1365
- checkpoints: join(resolve(String(options.runRoot)), 'checkpoints'),
1366
- eval: join(resolve(String(options.runRoot)), 'eval'),
1367
- scoreboard: join(resolve(String(options.runRoot)), 'scoreboard'),
1368
- exports: join(resolve(String(options.runRoot)), 'exports'),
1369
- compare: join(resolve(String(options.runRoot)), 'compare'),
1370
- qualityGate: join(resolve(String(options.runRoot)), 'quality-gate'),
1371
- }
1372
- : await createTrainingRunLayout({
1373
- kind: 'lora',
1374
- workloadId: workload.id,
1375
- timestamp: options.timestamp || null,
1376
- });
1377
- await Promise.all(Object.values(layout).map((dirPath) => mkdir(dirPath, { recursive: true })));
1378
- return layout;
1379
- }
1380
-
1381
- function isObjectRecord(value) {
1382
- return value && typeof value === 'object' && !Array.isArray(value);
1383
- }
1384
-
1385
- function hashTextPairsDataset(dataset) {
1386
- return hashArtifactPayload({ rows: dataset.rows });
1387
- }
1388
-
1389
- function parseCausalLmLoraTensorName(name) {
1390
- const text = String(name || '');
1391
- const match = text.match(/(?:^|\.)layers?\.?(\d+)\.(?:[A-Za-z0-9_]+\.)*([A-Za-z0-9_]+)\.lora_([ab])(?:\.[A-Za-z0-9_]+)?$/i);
1392
- if (!match) return null;
1393
- const layer = Number.parseInt(match[1], 10);
1394
- const rawModule = match[2].toLowerCase();
1395
- const module = LORA_MODULE_ALIASES[rawModule];
1396
- if (!module) return null;
1397
- return {
1398
- layer,
1399
- module,
1400
- kind: match[3].toLowerCase() === 'a' ? 'a' : 'b',
1401
- };
1402
- }
1403
-
1404
- function normalizeTensorShape(value, label) {
1405
- const shape = Array.isArray(value) ? value.map((entry) => Number(entry)) : [];
1406
- if (shape.length !== 2 || shape.some((entry) => !Number.isInteger(entry) || entry < 1)) {
1407
- throw new Error(`${label} requires shape [rows, cols].`);
1408
- }
1409
- return shape;
1410
- }
1411
-
1412
- function normalizeTrainerTensor(entry, index) {
1413
- if (!isObjectRecord(entry)) {
1414
- throw new Error(`Causal-LM trainer tensor ${index + 1} must be an object.`);
1415
- }
1416
- const name = String(entry.name || '').trim();
1417
- if (!name) {
1418
- throw new Error(`Causal-LM trainer tensor ${index + 1} requires name.`);
1419
- }
1420
- const tensor = entry.tensor ?? entry.data ?? entry.values;
1421
- if (tensor === undefined || tensor === null) {
1422
- throw new Error(`Causal-LM trainer tensor "${name}" requires tensor data.`);
1423
- }
1424
- const shape = normalizeTensorShape(entry.shape ?? tensor?.shape, `Causal-LM trainer tensor "${name}"`);
1425
- let normalizedTensor = tensor;
1426
- if (Array.isArray(tensor)) {
1427
- normalizedTensor = new Float32Array(tensor.map((value) => Number(value)));
1428
- }
1429
- if (normalizedTensor instanceof Float32Array && normalizedTensor.length !== shape[0] * shape[1]) {
1430
- throw new Error(
1431
- `Causal-LM trainer tensor "${name}" shape mismatch: expected ${shape[0] * shape[1]}, got ${normalizedTensor.length}.`
1432
- );
1433
- }
1434
- return {
1435
- name,
1436
- shape,
1437
- dtype: entry.dtype || 'f32',
1438
- tensor: normalizedTensor,
1439
- };
1440
- }
1441
-
1442
- function assertCausalLmTensorCoverage(tensors, adapter) {
1443
- const targetModules = normalizeLoraTargetModules(adapter);
1444
- if (!targetModules.length) {
1445
- throw new Error('Causal-LM LoRA export requires at least one target module.');
1446
- }
1447
- const targetSet = new Set(targetModules);
1448
- const layerModules = new Map();
1449
- const seenModules = new Set();
1450
- for (const tensor of tensors) {
1451
- const parsed = parseCausalLmLoraTensorName(tensor.name);
1452
- if (!parsed) {
1453
- throw new Error(`Unrecognized Causal-LM LoRA tensor name: ${tensor.name}`);
1454
- }
1455
- if (!targetSet.has(parsed.module)) {
1456
- throw new Error(
1457
- `Causal-LM LoRA tensor "${tensor.name}" targets module "${parsed.module}" outside workload targetModules.`
1458
- );
1459
- }
1460
- seenModules.add(parsed.module);
1461
- if (!layerModules.has(parsed.layer)) {
1462
- layerModules.set(parsed.layer, new Map());
1463
- }
1464
- const modules = layerModules.get(parsed.layer);
1465
- if (!modules.has(parsed.module)) {
1466
- modules.set(parsed.module, new Set());
1467
- }
1468
- modules.get(parsed.module).add(parsed.kind);
1469
- }
1470
- if (layerModules.size === 0) {
1471
- throw new Error('Causal-LM trainer returned no LoRA layer tensors.');
1472
- }
1473
- for (const moduleName of targetModules) {
1474
- if (!seenModules.has(moduleName)) {
1475
- throw new Error(`Causal-LM trainer returned no tensors for target module "${moduleName}".`);
1476
- }
1477
- }
1478
- for (const [layerIndex, modules] of layerModules.entries()) {
1479
- for (const [moduleName, kinds] of modules.entries()) {
1480
- if (!kinds?.has('a') || !kinds?.has('b')) {
1481
- throw new Error(
1482
- `Causal-LM trainer layer ${layerIndex} module ${moduleName} must include both lora_a and lora_b tensors.`
1483
- );
1484
- }
1485
- }
1486
- }
1487
- }
1488
-
1489
- function normalizeCausalLmTrainerOutput(output, workload) {
1490
- if (!isObjectRecord(output)) {
1491
- throw new Error('Causal-LM trainer must return an object.');
1492
- }
1493
- const rawTensors = Array.isArray(output.tensors)
1494
- ? output.tensors
1495
- : (Array.isArray(output.weights) ? output.weights : null);
1496
- if (!rawTensors || rawTensors.length === 0) {
1497
- throw new Error('Causal-LM trainer must return non-empty tensors.');
1498
- }
1499
- const tensors = rawTensors.map((entry, index) => normalizeTrainerTensor(entry, index));
1500
- assertCausalLmTensorCoverage(tensors, getPipelineConfig(workload).adapter);
1501
- const checkpointStep = Number.isInteger(Number(output.checkpointStep))
1502
- ? Number(output.checkpointStep)
1503
- : Number(workload.training?.steps || 0);
1504
- const checkpointId = String(output.checkpointId || `checkpoint-${String(checkpointStep).padStart(6, '0')}`).trim();
1505
- if (!/^[A-Za-z0-9_-]+$/.test(checkpointId)) {
1506
- throw new Error(`Causal-LM trainer checkpointId "${checkpointId}" must use alphanumeric, underscore, or hyphen characters.`);
1507
- }
1508
- return {
1509
- checkpointId,
1510
- checkpointStep,
1511
- adapterId: String(output.adapterId || '').trim() || null,
1512
- adapterName: String(output.adapterName || '').trim() || null,
1513
- trainerId: String(output.trainerId || '').trim() || null,
1514
- runnerId: String(output.runnerId || '').trim() || null,
1515
- metrics: isObjectRecord(output.metrics) ? output.metrics : {},
1516
- receipts: Array.isArray(output.receipts) ? output.receipts.slice() : [],
1517
- evalReports: Array.isArray(output.evalReports) ? output.evalReports.slice() : [],
1518
- tensors,
1519
- };
1520
- }
1521
-
1522
- function normalizeProviderEvalReport(entry, index, context) {
1523
- if (!isObjectRecord(entry)) {
1524
- throw new Error(`Causal-LM trainer evalReports[${index}] must be an object.`);
1525
- }
1526
- const evalDatasetId = String(entry.evalDatasetId || entry.id || '').trim();
1527
- if (!evalDatasetId) {
1528
- throw new Error(`Causal-LM trainer evalReports[${index}].evalDatasetId is required.`);
1529
- }
1530
- const primaryMetric = String(entry.primaryMetric || 'loss').trim();
1531
- if (!primaryMetric) {
1532
- throw new Error(`Causal-LM trainer evalReports[${index}].primaryMetric is required.`);
1533
- }
1534
- const primaryScore = finiteMetric(entry.primaryScore ?? entry.score ?? entry.loss);
1535
- if (primaryScore === null) {
1536
- throw new Error(`Causal-LM trainer evalReports[${index}].primaryScore must be finite.`);
1537
- }
1538
- const loss = finiteMetric(entry.loss ?? (primaryMetric === 'loss' ? primaryScore : null));
1539
- const baselineScore = finiteMetric(entry.baseline?.primaryScore ?? entry.baseline?.loss);
1540
- const qualityClaim = isObjectRecord(entry.qualityClaim)
1541
- ? entry.qualityClaim
1542
- : (baselineScore === null || loss === null
1543
- ? null
1544
- : {
1545
- baseline: entry.baseline?.stage || 'base_model',
1546
- metric: 'loss',
1547
- selectionGoal: 'min',
1548
- adapterScore: loss,
1549
- baselineScore,
1550
- delta: loss - baselineScore,
1551
- absoluteImprovement: baselineScore - loss,
1552
- relativeImprovement: baselineScore === 0 ? 0 : (baselineScore - loss) / Math.abs(baselineScore),
1553
- minAbsoluteImprovement: 0,
1554
- minRelativeImprovement: 0,
1555
- improved: loss <= baselineScore,
1556
- requireImprovement: false,
1557
- });
1558
- const agentEval = isObjectRecord(entry.agentEval)
1559
- ? entry.agentEval
1560
- : (isObjectRecord(entry.heldoutGate) ? entry.heldoutGate : null);
1561
- return {
1562
- artifactType: 'training_eval_report',
1563
- schemaVersion: 1,
1564
- generatedAt: new Date().toISOString(),
1565
- workloadId: context.workload.id,
1566
- workloadPath: context.loadedWorkload.absolutePath || null,
1567
- workloadSha256: context.loadedWorkload.workloadSha256 || null,
1568
- configHash: context.workload.configHash,
1569
- datasetPath: entry.datasetPath || context.dataset.absolutePath,
1570
- datasetHash: entry.datasetHash || context.datasetHash,
1571
- baseModelId: context.workload.baseModelId,
1572
- baseModelRef: context.preflight.baseModelId,
1573
- stage: entry.stage || 'lora',
1574
- checkpointId: context.checkpointId,
1575
- checkpointStep: context.checkpointStep,
1576
- evalDatasetId,
1577
- metrics: isObjectRecord(entry.metrics)
1578
- ? entry.metrics
1579
- : {
1580
- [primaryMetric]: {
1581
- score: primaryScore,
1582
- samples: finiteMetric(entry.samples) ?? context.dataset.rowCount,
1583
- },
1584
- },
1585
- primaryMetric,
1586
- primaryScore,
1587
- loss,
1588
- baseline: isObjectRecord(entry.baseline) ? entry.baseline : null,
1589
- qualityClaim,
1590
- agentEval,
1591
- };
1592
- }
1593
-
1594
- async function writeProviderEvalReports(context) {
1595
- const reports = [];
1596
- for (let index = 0; index < context.trainerOutput.evalReports.length; index += 1) {
1597
- const report = normalizeProviderEvalReport(context.trainerOutput.evalReports[index], index, context);
1598
- const reportFile = await writeJsonArtifact(
1599
- join(context.layout.eval, `${context.checkpointId}__${report.evalDatasetId}.json`),
1600
- report
1601
- );
1602
- const materialized = {
1603
- ...report,
1604
- reportPath: reportFile.path,
1605
- };
1606
- reports.push(materialized);
1607
- await appendScoreboardRow(context.layout.scoreboard, {
1608
- artifactType: 'training_scoreboard',
1609
- schemaVersion: 1,
1610
- generatedAt: new Date().toISOString(),
1611
- checkpointId: materialized.checkpointId,
1612
- checkpointStep: materialized.checkpointStep,
1613
- stage: materialized.stage,
1614
- evalDatasetId: materialized.evalDatasetId,
1615
- primaryMetric: materialized.primaryMetric,
1616
- primaryScore: materialized.primaryScore,
1617
- loss: materialized.loss,
1618
- qualityClaim: materialized.qualityClaim || null,
1619
- agentEval: materialized.agentEval || null,
1620
- metrics: {
1621
- [materialized.primaryMetric]: materialized.primaryScore,
1622
- primaryScore: materialized.primaryScore,
1623
- agent_heldout_gate: materialized.agentEval?.passRate ?? null,
1624
- },
1625
- }, {
1626
- selectionMetric: context.workload.selectionMetric,
1627
- selectionGoal: context.workload.selectionGoal,
1628
- });
1629
- }
1630
- return reports;
1631
- }
1632
-
1633
- async function resolveCausalLmTrainer(loadedWorkload, options = {}) {
1634
- if (typeof options.causalLmTrainer === 'function') {
1635
- return {
1636
- train: options.causalLmTrainer,
1637
- runnerId: options.causalLmTrainer.runnerId || 'injected_causal_lm_lora_trainer',
1638
- source: 'runLoraPipeline.options.causalLmTrainer',
1639
- exportName: null,
1640
- };
1641
- }
1642
- const trainerConfig = getPipelineConfig(loadedWorkload.workload).trainer;
1643
- if (!trainerConfig) {
1644
- throw new Error(
1645
- 'causal_lm_trainer_not_configured: provide runLoraPipeline({ causalLmTrainer }) or lora.trainer.modulePath in the workload.'
1646
- );
1647
- }
1648
- const modulePath = String(trainerConfig.modulePath || trainerConfig.path || '').trim();
1649
- if (!modulePath) {
1650
- throw new Error('causal_lm_trainer_not_configured: lora.trainer.modulePath is required.');
1651
- }
1652
- const exportName = String(trainerConfig.exportName || 'trainCausalLmLora').trim();
1653
- if (!exportName) {
1654
- throw new Error('causal_lm_trainer_not_configured: lora.trainer.exportName is required.');
1655
- }
1656
- const workloadDir = loadedWorkload.absolutePath
1657
- ? dirname(resolve(String(loadedWorkload.absolutePath)))
1658
- : process.cwd();
1659
- const absoluteModulePath = isAbsolute(modulePath)
1660
- ? modulePath
1661
- : resolve(workloadDir, modulePath);
1662
- const trainerModule = await import(pathToFileURL(absoluteModulePath).href);
1663
- const train = trainerModule[exportName];
1664
- if (typeof train !== 'function') {
1665
- throw new Error(`causal_lm_trainer_not_configured: ${absoluteModulePath} does not export ${exportName}().`);
1666
- }
1667
- return {
1668
- train,
1669
- runnerId: String(trainerConfig.runnerId || exportName).trim(),
1670
- source: absoluteModulePath,
1671
- exportName,
1672
- };
1673
- }
1674
-
1675
- async function exportProviderCausalLmLoraModel(
1676
- loadedWorkload,
1677
- layout,
1678
- trainerOutput,
1679
- checkpointId,
1680
- checkpointStep,
1681
- datasetHash,
1682
- trainerInfo,
1683
- preflight
1684
- ) {
1685
- const workload = loadedWorkload.workload;
1686
- const pipeline = getPipelineConfig(workload);
1687
- const weightsFilename = `${checkpointId}.adapters.safetensors`;
1688
- const exported = await exportLoRAAdapter({
1689
- id: pipeline.export?.id || trainerOutput.adapterId || `${workload.id}-${checkpointId}`,
1690
- name: pipeline.export?.name || trainerOutput.adapterName || `${workload.id}-${checkpointId}`,
1691
- baseModel: workload.baseModelId,
1692
- rank: pipeline.adapter.rank,
1693
- alpha: pipeline.adapter.alpha,
1694
- targetModules: normalizeLoraTargetModules(pipeline.adapter),
1695
- tensors: trainerOutput.tensors,
1696
- weightsFormat: 'safetensors',
1697
- weightsPath: weightsFilename,
1698
- metadata: {
1699
- runnerKind: 'causal_lm_text_generation',
1700
- runnerKey: preflight.runnerKey,
1701
- runnerId: trainerOutput.runnerId || trainerInfo.runnerId,
1702
- trainerId: trainerOutput.trainerId,
1703
- trainerSource: trainerInfo.source,
1704
- datasetHash,
1705
- workloadSha256: loadedWorkload.workloadSha256,
1706
- metrics: trainerOutput.metrics,
1707
- receipts: trainerOutput.receipts,
1708
- },
1709
- });
1710
- if (!exported.weights) {
1711
- throw new Error('Causal-LM LoRA safetensors export did not return weights bytes.');
1712
- }
1713
- const manifestPath = join(layout.exports, `${checkpointId}.adapter.manifest.json`);
1714
- const runtimeManifestPath = join(layout.exports, 'runtime-adapter-manifest.json');
1715
- const weightsPath = join(layout.exports, weightsFilename);
1716
- await writeFile(weightsPath, new Uint8Array(exported.weights));
1717
- await writeFile(manifestPath, exported.json, 'utf8');
1718
- await writeFile(runtimeManifestPath, exported.json, 'utf8');
1719
- await loadLoRAFromManifest(exported.manifest, {
1720
- readFile: async (filePath) => readFile(join(layout.exports, filePath)),
1721
- });
1722
- const artifactPayload = {
1723
- ...buildArtifact(loadedWorkload, {
1724
- prefix: 'lora_export',
1725
- id: checkpointId,
1726
- artifactType: 'lora_adapter_manifest',
1727
- checkpointStep,
1728
- datasetHash,
1729
- }),
1730
- checkpointId,
1731
- manifestPath,
1732
- runtimeManifestPath,
1733
- weightsPath,
1734
- weightsSha256: exported.weightsSha256 || null,
1735
- runnerKey: preflight.runnerKey,
1736
- trainer: {
1737
- runnerId: trainerOutput.runnerId || trainerInfo.runnerId,
1738
- trainerId: trainerOutput.trainerId,
1739
- source: trainerInfo.source,
1740
- exportName: trainerInfo.exportName,
1741
- },
1742
- metrics: trainerOutput.metrics,
1743
- manifest: exported.manifest,
1744
- };
1745
- const artifactFile = await writeJsonArtifact(
1746
- join(layout.exports, `${checkpointId}.export.json`),
1747
- artifactPayload
1748
- );
1749
- return {
1750
- checkpointId,
1751
- manifestPath,
1752
- runtimeManifestPath,
1753
- weightsPath,
1754
- weightsSha256: exported.weightsSha256 || null,
1755
- exportPath: artifactFile.path,
1756
- manifest: exported.manifest,
1757
- };
1758
- }
1759
-
1760
162
  function hasExternalCausalLmTrainer(loadedWorkload, options = {}) {
1761
163
  return typeof options.causalLmTrainer === 'function'
1762
164
  || Boolean(getPipelineConfig(loadedWorkload.workload).trainer);
1763
165
  }
1764
166
 
1765
- function assertInternalCausalLmTrainerAllowed(workload, compatibility) {
1766
- if (compatibility.observed.requiresExternalTrainer !== true) return;
1767
- throw new Error(
1768
- `Causal-LM LoRA base "${workload.baseModelId}" requires lora.trainer.modulePath or runLoraPipeline({ causalLmTrainer }) ` +
1769
- 'because the internal full-graph runner does not train packed q4k base weights.'
1770
- );
1771
- }
1772
-
1773
- async function runInternalCausalLmLoraPipeline(options, layout, compatibility) {
1774
- const loadedWorkload = options.loadedWorkload;
1775
- const workload = loadedWorkload.workload;
1776
- const pipeline = getPipelineConfig(workload);
1777
- const datasetPath = await resolveDatasetPathForLoadedWorkload(workload.datasetPath, loadedWorkload);
1778
- const preflight = await preflightCausalLmLoraWorkload(workload, {
1779
- datasetPath,
1780
- fetch: options.fetch,
1781
- readFile: options.readFile,
1782
- });
1783
- if (!preflight.supported || preflight.blockedReasons.length > 0) {
1784
- throw new Error(`Causal-LM LoRA workload is blocked: ${preflight.blockedReasons.join(',')}`);
1785
- }
1786
- assertInternalCausalLmTrainerAllowed(workload, compatibility);
1787
- if (Math.floor(Number(workload.training.batchSize)) !== 1) {
1788
- throw new Error('Causal-LM LoRA workload requires training.batchSize=1.');
1789
- }
1790
- const freeze = getCausalLmFreezeConfig(workload);
1791
- const fixture = await createCausalLmLoraFixture(workload);
1792
- try {
1793
- const dataset = await loadCausalLmTextPairSamples(workload, datasetPath, fixture.tokenizer);
1794
- const initialParameterState = await captureLoraParameterState(fixture.model);
1795
- const evalReports = [];
1796
- const checkpointArtifacts = [];
1797
- const exports = [];
1798
- const baselineEvalReports = await evaluateCausalLmLoraModel(workload, fixture, dataset, layout, {
1799
- checkpointId: 'base-model',
1800
- checkpointStep: 0,
1801
- stage: 'base_model',
1802
- configHash: workload.configHash,
1803
- workloadPath: loadedWorkload.absolutePath,
1804
- workloadSha256: loadedWorkload.workloadSha256,
1805
- });
1806
- for (const report of baselineEvalReports) {
1807
- evalReports.push(report);
1808
- await appendScoreboardRow(layout.scoreboard, {
1809
- artifactType: 'training_scoreboard',
1810
- schemaVersion: 1,
1811
- generatedAt: new Date().toISOString(),
1812
- checkpointId: report.checkpointId,
1813
- checkpointStep: report.checkpointStep,
1814
- stage: report.stage,
1815
- evalDatasetId: report.evalDatasetId,
1816
- primaryMetric: report.primaryMetric,
1817
- primaryScore: report.primaryScore,
1818
- loss: report.loss,
1819
- metrics: {
1820
- loss: report.loss,
1821
- primaryScore: report.primaryScore,
1822
- },
1823
- }, {
1824
- selectionMetric: workload.selectionMetric,
1825
- selectionGoal: workload.selectionGoal,
1826
- });
1827
- }
1828
- const baselineReportsByEvalDatasetId = Object.fromEntries(
1829
- baselineEvalReports.map((report) => [report.evalDatasetId, report])
1830
- );
1831
- const finalizedCheckpointIds = new Set();
1832
- const runner = new TrainingRunner(createLoraRunnerTrainingConfig(workload, freeze), {
1833
- optimizer: createLoraOptimizer(workload),
1834
- crossEntropyLoss,
1835
- clipGradients,
1836
- trainingObjective: createCausalLmTrainingObjective(),
1837
- onCheckpoint: async (checkpoint) => {
1838
- const checkpointId = `checkpoint-${String(checkpoint.step).padStart(6, '0')}`;
1839
- if (finalizedCheckpointIds.has(checkpointId)) return;
1840
- finalizedCheckpointIds.add(checkpointId);
1841
- const checkpointPayload = {
1842
- ...buildArtifact(loadedWorkload, {
1843
- prefix: 'lora_ckpt',
1844
- id: checkpointId,
1845
- artifactType: 'training_checkpoint',
1846
- datasetPath: dataset.absolutePath,
1847
- datasetHash: dataset.datasetHash,
1848
- checkpointStep: checkpoint.step,
1849
- }),
1850
- checkpointId,
1851
- checkpointPath: checkpoint.path,
1852
- optimizerStatePresent: true,
1853
- schedulerStatePresent: workload.training.optimizer.scheduler.enabled === true,
1854
- runnerKey: compatibility.observed.runnerKey,
1855
- baseModelRef: fixture.baseModelRef,
1856
- tokenization: dataset.tokenization,
1857
- resumeLineage: checkpoint.metadata?.lineage || null,
1858
- };
1859
- await writeJsonArtifact(
1860
- join(layout.checkpoints, checkpointId, 'checkpoint.json'),
1861
- checkpointPayload
1862
- );
1863
- const checkpointArtifact = await writeJsonArtifact(
1864
- join(layout.checkpoints, checkpointId, 'checkpoint.complete.json'),
1865
- checkpointPayload
1866
- );
1867
- checkpointArtifacts.push({
1868
- checkpointId,
1869
- checkpointPath: checkpoint.path,
1870
- markerPath: checkpointArtifact.path,
1871
- checkpointStep: checkpoint.step,
1872
- });
1873
- if (pipeline.export?.enabled === true && pipeline.export.atCheckpoints === true) {
1874
- exports.push(await exportCausalLmLoraModel(
1875
- loadedWorkload,
1876
- layout,
1877
- fixture,
1878
- checkpointId,
1879
- checkpoint.step,
1880
- dataset.datasetHash
1881
- ));
1882
- }
1883
- const reports = await evaluateCausalLmLoraModel(workload, fixture, dataset, layout, {
1884
- checkpointId,
1885
- checkpointStep: checkpoint.step,
1886
- configHash: workload.configHash,
1887
- workloadPath: loadedWorkload.absolutePath,
1888
- workloadSha256: loadedWorkload.workloadSha256,
1889
- baselineReportsByEvalDatasetId,
1890
- });
1891
- for (const report of reports) {
1892
- evalReports.push(report);
1893
- await appendScoreboardRow(layout.scoreboard, {
1894
- artifactType: 'training_scoreboard',
1895
- schemaVersion: 1,
1896
- generatedAt: new Date().toISOString(),
1897
- checkpointId,
1898
- checkpointStep: checkpoint.step,
1899
- stage: report.stage,
1900
- evalDatasetId: report.evalDatasetId,
1901
- primaryMetric: report.primaryMetric,
1902
- primaryScore: report.primaryScore,
1903
- loss: report.loss,
1904
- qualityClaim: report.qualityClaim || null,
1905
- metrics: {
1906
- loss: report.loss,
1907
- primaryScore: report.primaryScore,
1908
- },
1909
- }, {
1910
- selectionMetric: workload.selectionMetric,
1911
- selectionGoal: workload.selectionGoal,
1912
- });
1913
- }
1914
- },
1915
- });
1916
- const metrics = await runner.run(
1917
- fixture.model,
1918
- createCausalLmDatasetBatches(dataset.samples),
1919
- {
1920
- epochs: 1,
1921
- batchSize: 1,
1922
- shuffle: false,
1923
- maxSteps: workload.training.steps,
1924
- checkpointEvery: workload.checkpointEvery,
1925
- checkpointKey: join(layout.checkpoints, 'latest.state.json'),
1926
- modelId: workload.baseModelId,
1927
- modelUrl: fixture.baseModelUrl,
1928
- tokenizerHash: fixture.baseManifest?.tokenizerHash || null,
1929
- }
1930
- );
1931
- const finalParameterState = await captureLoraParameterState(fixture.model);
1932
- const parameterReceipt = buildLoraParameterReceipt(
1933
- initialParameterState,
1934
- finalParameterState
1935
- );
1936
- const finalCheckpointId = runner.lastCheckpoint
1937
- ? `checkpoint-${String(runner.lastCheckpoint.step).padStart(6, '0')}`
1938
- : null;
1939
- if (
1940
- pipeline.export?.enabled !== false
1941
- && finalCheckpointId
1942
- && exports.every((entry) => entry.checkpointId !== finalCheckpointId)
1943
- ) {
1944
- exports.push(await exportCausalLmLoraModel(
1945
- loadedWorkload,
1946
- layout,
1947
- fixture,
1948
- finalCheckpointId,
1949
- runner.lastCheckpoint.step,
1950
- dataset.datasetHash
1951
- ));
1952
- }
1953
- return {
1954
- ok: true,
1955
- kind: 'lora',
1956
- action: 'run',
1957
- runnerKind: 'causal_lm_lora',
1958
- workloadId: workload.id,
1959
- runRoot: layout.runRoot,
1960
- preflight,
1961
- checkpointArtifacts,
1962
- evalReports,
1963
- exports,
1964
- metrics,
1965
- parameterReceipt,
1966
- lastCheckpoint: runner.lastCheckpoint,
1967
- dataset: {
1968
- path: dataset.absolutePath,
1969
- rowCount: dataset.rowCount,
1970
- sampleCount: dataset.samples.length,
1971
- datasetHash: dataset.datasetHash,
1972
- tokenization: dataset.tokenization,
1973
- },
1974
- baseModel: {
1975
- id: workload.baseModelId,
1976
- ref: fixture.baseModelRef,
1977
- url: fixture.baseModelUrl,
1978
- },
1979
- };
1980
- } finally {
1981
- await fixture.cleanup();
1982
- }
1983
- }
1984
-
1985
- async function runProviderCausalLmLoraPipeline(options, compatibility) {
1986
- const loadedWorkload = options.loadedWorkload;
1987
- const workload = loadedWorkload.workload;
1988
- const pipeline = getPipelineConfig(workload);
1989
- if (!workload.datasetPath) {
1990
- throw new Error('preflightCausalLmLoraWorkload requires workload.datasetPath.');
1991
- }
1992
- const layout = await createLoraRunLayout(options, workload);
1993
- await writeRunContract(layout, buildRunContract(loadedWorkload));
1994
- await writeWorkloadLock(layout, loadedWorkload);
1995
- const datasetPath = await resolveDatasetPathForLoadedWorkload(workload.datasetPath, loadedWorkload);
1996
- const dataset = await loadTextPairsDataset(datasetPath, {
1997
- fetch: options.fetch,
1998
- readFile: options.readFile,
1999
- });
2000
- if (dataset.rowCount < 1) {
2001
- throw new Error(`Causal-LM LoRA dataset ${dataset.absolutePath} has no rows.`);
2002
- }
2003
- const datasetHash = hashTextPairsDataset(dataset);
2004
- const preflight = await preflightCausalLmLoraWorkload(workload, {
2005
- datasetPath: dataset.absolutePath,
2006
- fetch: options.fetch,
2007
- readFile: options.readFile,
2008
- });
2009
- const trainerInfo = await resolveCausalLmTrainer(loadedWorkload, options);
2010
- const trainerResult = await trainerInfo.train({
2011
- schemaVersion: 1,
2012
- runnerKind: 'causal_lm_lora',
2013
- workload,
2014
- loadedWorkload,
2015
- compatibility,
2016
- preflight,
2017
- dataset: {
2018
- absolutePath: dataset.absolutePath,
2019
- rowCount: dataset.rowCount,
2020
- rows: dataset.rows,
2021
- datasetHash,
2022
- },
2023
- adapter: pipeline.adapter,
2024
- training: workload.training,
2025
- export: pipeline.export,
2026
- layout,
2027
- });
2028
- const trainerOutput = normalizeCausalLmTrainerOutput(trainerResult, workload);
2029
- const checkpointPayload = {
2030
- ...buildArtifact(loadedWorkload, {
2031
- prefix: 'lora_ckpt',
2032
- id: trainerOutput.checkpointId,
2033
- artifactType: 'training_checkpoint',
2034
- datasetPath: dataset.absolutePath,
2035
- datasetHash,
2036
- checkpointStep: trainerOutput.checkpointStep,
2037
- }),
2038
- checkpointId: trainerOutput.checkpointId,
2039
- checkpointPath: join(layout.checkpoints, trainerOutput.checkpointId, 'trainer-output.json'),
2040
- optimizerStatePresent: false,
2041
- schedulerStatePresent: false,
2042
- runnerKey: preflight.runnerKey,
2043
- trainer: {
2044
- runnerId: trainerOutput.runnerId || trainerInfo.runnerId,
2045
- trainerId: trainerOutput.trainerId,
2046
- source: trainerInfo.source,
2047
- exportName: trainerInfo.exportName,
2048
- },
2049
- tensorNames: trainerOutput.tensors.map((entry) => entry.name),
2050
- metrics: trainerOutput.metrics,
2051
- receipts: trainerOutput.receipts,
2052
- };
2053
- await writeJsonArtifact(
2054
- join(layout.checkpoints, trainerOutput.checkpointId, 'trainer-output.json'),
2055
- checkpointPayload
2056
- );
2057
- const checkpointArtifact = await writeJsonArtifact(
2058
- join(layout.checkpoints, trainerOutput.checkpointId, 'checkpoint.complete.json'),
2059
- checkpointPayload
2060
- );
2061
- const exported = pipeline.export?.enabled === false
2062
- ? null
2063
- : await exportProviderCausalLmLoraModel(
2064
- loadedWorkload,
2065
- layout,
2066
- trainerOutput,
2067
- trainerOutput.checkpointId,
2068
- trainerOutput.checkpointStep,
2069
- datasetHash,
2070
- trainerInfo,
2071
- preflight
2072
- );
2073
- const evalReports = await writeProviderEvalReports({
2074
- loadedWorkload,
2075
- workload,
2076
- layout,
2077
- dataset,
2078
- datasetHash,
2079
- preflight,
2080
- trainerOutput,
2081
- checkpointId: trainerOutput.checkpointId,
2082
- checkpointStep: trainerOutput.checkpointStep,
2083
- });
2084
- return {
2085
- ok: true,
2086
- kind: 'lora',
2087
- action: 'run',
2088
- runnerKind: 'causal_lm_lora',
2089
- workloadId: workload.id,
2090
- runRoot: layout.runRoot,
2091
- preflight,
2092
- checkpointArtifacts: [{
2093
- checkpointId: trainerOutput.checkpointId,
2094
- checkpointPath: checkpointPayload.checkpointPath,
2095
- markerPath: checkpointArtifact.path,
2096
- checkpointStep: trainerOutput.checkpointStep,
2097
- }],
2098
- evalReports,
2099
- exports: exported ? [exported] : [],
2100
- metrics: trainerOutput.metrics,
2101
- lastCheckpoint: {
2102
- id: trainerOutput.checkpointId,
2103
- step: trainerOutput.checkpointStep,
2104
- path: checkpointPayload.checkpointPath,
2105
- },
2106
- };
2107
- }
2108
-
2109
167
  async function runCausalLmLoraPipeline(options, compatibility) {
2110
168
  if (!options.loadedWorkload.workload.datasetPath) {
2111
169
  throw new Error('preflightCausalLmLoraWorkload requires workload.datasetPath.');