doppler-gpu 0.2.0

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 (1405) hide show
  1. package/BRANDING.md +14 -0
  2. package/CHANGELOG.md +158 -0
  3. package/LICENSE +201 -0
  4. package/NOTICE +5 -0
  5. package/README.md +85 -0
  6. package/SECURITY.md +19 -0
  7. package/package.json +195 -0
  8. package/src/adapters/adapter-manager.d.ts +200 -0
  9. package/src/adapters/adapter-manager.js +509 -0
  10. package/src/adapters/adapter-manifest.d.ts +290 -0
  11. package/src/adapters/adapter-manifest.js +320 -0
  12. package/src/adapters/adapter-registry.d.ts +192 -0
  13. package/src/adapters/adapter-registry.js +477 -0
  14. package/src/adapters/index.d.ts +89 -0
  15. package/src/adapters/index.js +42 -0
  16. package/src/adapters/lora-loader.d.ts +105 -0
  17. package/src/adapters/lora-loader.js +425 -0
  18. package/src/bootstrap.d.ts +1 -0
  19. package/src/bootstrap.js +30 -0
  20. package/src/bridge/extension/background.d.ts +14 -0
  21. package/src/bridge/extension/background.js +168 -0
  22. package/src/bridge/extension/manifest.json +34 -0
  23. package/src/bridge/extension-client.d.ts +114 -0
  24. package/src/bridge/extension-client.js +409 -0
  25. package/src/bridge/index.d.ts +69 -0
  26. package/src/bridge/index.js +53 -0
  27. package/src/bridge/node-bridge.js +26 -0
  28. package/src/bridge/protocol.d.ts +96 -0
  29. package/src/bridge/protocol.js +130 -0
  30. package/src/browser/browser-converter.d.ts +71 -0
  31. package/src/browser/browser-converter.js +978 -0
  32. package/src/browser/file-picker.d.ts +63 -0
  33. package/src/browser/file-picker.js +281 -0
  34. package/src/browser/gguf-importer.d.ts +136 -0
  35. package/src/browser/gguf-importer.js +532 -0
  36. package/src/browser/gguf-parser-browser.d.ts +14 -0
  37. package/src/browser/gguf-parser-browser.js +17 -0
  38. package/src/browser/quantization.d.ts +69 -0
  39. package/src/browser/quantization.js +328 -0
  40. package/src/browser/safetensors-parser-browser.d.ts +193 -0
  41. package/src/browser/safetensors-parser-browser.js +347 -0
  42. package/src/browser/shard-io-browser.d.ts +57 -0
  43. package/src/browser/shard-io-browser.js +89 -0
  44. package/src/browser/tensor-source-download.d.ts +27 -0
  45. package/src/browser/tensor-source-download.js +245 -0
  46. package/src/browser/tensor-source-file.d.ts +26 -0
  47. package/src/browser/tensor-source-file.js +53 -0
  48. package/src/browser/tensor-source-http.d.ts +29 -0
  49. package/src/browser/tensor-source-http.js +130 -0
  50. package/src/cli/config/doppler-cli-policy.json +28 -0
  51. package/src/cli/doppler-cli.d.ts +28 -0
  52. package/src/cli/doppler-cli.js +1664 -0
  53. package/src/client/doppler-api.browser.d.ts +1 -0
  54. package/src/client/doppler-api.browser.js +389 -0
  55. package/src/client/doppler-api.d.ts +83 -0
  56. package/src/client/doppler-api.js +389 -0
  57. package/src/client/doppler-provider/generation.d.ts +25 -0
  58. package/src/client/doppler-provider/generation.js +126 -0
  59. package/src/client/doppler-provider/index.d.ts +2 -0
  60. package/src/client/doppler-provider/index.js +3 -0
  61. package/src/client/doppler-provider/model-manager.d.ts +71 -0
  62. package/src/client/doppler-provider/model-manager.js +739 -0
  63. package/src/client/doppler-provider/provider.d.ts +5 -0
  64. package/src/client/doppler-provider/provider.js +102 -0
  65. package/src/client/doppler-provider/source-runtime.d.ts +23 -0
  66. package/src/client/doppler-provider/source-runtime.js +677 -0
  67. package/src/client/doppler-provider/types.d.ts +127 -0
  68. package/src/client/doppler-provider/types.js +17 -0
  69. package/src/client/doppler-provider.d.ts +46 -0
  70. package/src/client/doppler-provider.js +36 -0
  71. package/src/client/doppler-registry.d.ts +23 -0
  72. package/src/client/doppler-registry.js +86 -0
  73. package/src/client/doppler-registry.json +41 -0
  74. package/src/config/README.md +57 -0
  75. package/src/config/backward-registry-loader.d.ts +3 -0
  76. package/src/config/backward-registry-loader.js +23 -0
  77. package/src/config/conversion/README.md +147 -0
  78. package/src/config/conversion/embeddinggemma/google-embeddinggemma-300m-q4k-ehf16-af32.json +285 -0
  79. package/src/config/conversion/gemma3/gemma-3-1b-it-f16-af32.json +321 -0
  80. package/src/config/conversion/gemma3/gemma-3-1b-it-q4k-ehf16-af32.json +322 -0
  81. package/src/config/conversion/gemma3/gemma-3-270m-it-q4k-ehf16-af32.json +322 -0
  82. package/src/config/conversion/gemma3/translategemma-4b-1b-enes-q4k-ehf16-af32.json +322 -0
  83. package/src/config/conversion/gemma3/translategemma-4b-it-q4k-ehf16-af32.json +323 -0
  84. package/src/config/conversion/gemma4/gemma-4-moe-q4k-ehf16-af32.json +158 -0
  85. package/src/config/conversion/gpt-oss-20b-f16-xmxfp4.json +164 -0
  86. package/src/config/conversion/janus/janus-pro-1b-text-q4k-ehaf16.json +158 -0
  87. package/src/config/conversion/lfm2/lfm2.5-1.2b-instruct-q4k-ehf16-af32.json +339 -0
  88. package/src/config/conversion/qwen3/qwen-3-5-0-8b-q4k-ehaf16.json +385 -0
  89. package/src/config/conversion/qwen3/qwen-3-5-2b-q4k-ehaf16.json +385 -0
  90. package/src/config/conversion/sana/sana-sprint-0.6b-f16.json +219 -0
  91. package/src/config/execution-contract-check.d.ts +72 -0
  92. package/src/config/execution-contract-check.js +290 -0
  93. package/src/config/index.d.ts +52 -0
  94. package/src/config/index.js +22 -0
  95. package/src/config/kernel-path-loader.d.ts +156 -0
  96. package/src/config/kernel-path-loader.js +396 -0
  97. package/src/config/kernels/backward-registry.json +99 -0
  98. package/src/config/kernels/kernel-ref-digests.d.ts +1 -0
  99. package/src/config/kernels/kernel-ref-digests.js +233 -0
  100. package/src/config/kernels/kernel-ref.d.ts +17 -0
  101. package/src/config/kernels/kernel-ref.js +75 -0
  102. package/src/config/kernels/moe/gpt-oss.paths.json +49 -0
  103. package/src/config/kernels/moe/mixtral.paths.json +46 -0
  104. package/src/config/kernels/registry.d.ts +86 -0
  105. package/src/config/kernels/registry.js +116 -0
  106. package/src/config/kernels/registry.json +9446 -0
  107. package/src/config/merge-contract-check.d.ts +16 -0
  108. package/src/config/merge-contract-check.js +311 -0
  109. package/src/config/merge-helpers.d.ts +58 -0
  110. package/src/config/merge-helpers.js +179 -0
  111. package/src/config/merge.d.ts +153 -0
  112. package/src/config/merge.js +445 -0
  113. package/src/config/param-categories.d.ts +17 -0
  114. package/src/config/param-categories.js +72 -0
  115. package/src/config/param-validator.d.ts +26 -0
  116. package/src/config/param-validator.js +274 -0
  117. package/src/config/platforms/amd-rdna3.json +16 -0
  118. package/src/config/platforms/apple-m1.json +16 -0
  119. package/src/config/platforms/apple-m2.json +16 -0
  120. package/src/config/platforms/apple-m3.json +16 -0
  121. package/src/config/platforms/generic.json +14 -0
  122. package/src/config/platforms/loader.d.ts +65 -0
  123. package/src/config/platforms/loader.js +158 -0
  124. package/src/config/platforms/nvidia-gb200-8gpu.json +45 -0
  125. package/src/config/platforms/nvidia-gb200-nvl72.json +45 -0
  126. package/src/config/platforms/nvidia-gh200-nvl2.json +44 -0
  127. package/src/config/platforms/nvidia-gh200.json +44 -0
  128. package/src/config/platforms/nvidia-rtx30.json +16 -0
  129. package/src/config/platforms/nvidia-rtx40.json +16 -0
  130. package/src/config/quantization-contract-check.d.ts +12 -0
  131. package/src/config/quantization-contract-check.js +91 -0
  132. package/src/config/required-inference-fields-contract-check.d.ts +24 -0
  133. package/src/config/required-inference-fields-contract-check.js +238 -0
  134. package/src/config/runtime/compute/f16-activations.json +30 -0
  135. package/src/config/runtime/compute/f16-batched.json +32 -0
  136. package/src/config/runtime/default.json +100 -0
  137. package/src/config/runtime/diagnostics/debug-logits.json +53 -0
  138. package/src/config/runtime/diagnostics/drift-policies.json +147 -0
  139. package/src/config/runtime/experiments/bench/gemma3-bench-q4k.json +54 -0
  140. package/src/config/runtime/experiments/debug/gemma3-debug-q4k.json +210 -0
  141. package/src/config/runtime/experiments/verify/gemma3-verify.json +39 -0
  142. package/src/config/runtime/experiments/verify/lfm2-verify.json +46 -0
  143. package/src/config/runtime/experiments/verify/translategemma-verify.json +39 -0
  144. package/src/config/runtime/kernels/dequant-f16-q4k.json +13 -0
  145. package/src/config/runtime/kernels/dequant-f32-q4k.json +13 -0
  146. package/src/config/runtime/kernels/embeddinggemma-q4k-dequant-f32a.json +37 -0
  147. package/src/config/runtime/kernels/fused-q4k.json +13 -0
  148. package/src/config/runtime/kernels/gemma2-q4k-dequant-f16a.json +33 -0
  149. package/src/config/runtime/kernels/gemma2-q4k-dequant-f32a-nosubgroups.json +33 -0
  150. package/src/config/runtime/kernels/gemma2-q4k-fused-f32a.json +33 -0
  151. package/src/config/runtime/kernels/safe-q4k.json +13 -0
  152. package/src/config/runtime/model/gemma2-debug.json +77 -0
  153. package/src/config/runtime/model/gemma2-pipeline-debug.json +66 -0
  154. package/src/config/runtime/model/gemma2-pipeline.json +75 -0
  155. package/src/config/runtime/model/gemma3-layer-probe.json +85 -0
  156. package/src/config/runtime/model/qwen3-5-layer-probe.json +52 -0
  157. package/src/config/runtime/model/qwen3-5-layer3-boundary-probe.json +108 -0
  158. package/src/config/runtime/model/qwen3-5-linear-attn-debug.json +90 -0
  159. package/src/config/runtime/overlays/debug-f32a.json +3 -0
  160. package/src/config/runtime/overlays/trace-config.json +28 -0
  161. package/src/config/runtime/platform/metal-apple-q4k.json +11 -0
  162. package/src/config/runtime/profiles/default.json +10 -0
  163. package/src/config/runtime/profiles/low-memory.json +40 -0
  164. package/src/config/runtime/profiles/production.json +48 -0
  165. package/src/config/runtime/profiles/simulation.json +31 -0
  166. package/src/config/runtime/profiles/throughput.json +37 -0
  167. package/src/config/runtime/profiles/trace-layers.json +127 -0
  168. package/src/config/runtime/profiles/vector-stability.json +54 -0
  169. package/src/config/runtime/profiles/vector-throughput.json +28 -0
  170. package/src/config/runtime/profiles/verbose-trace.json +39 -0
  171. package/src/config/runtime/tiers/gemma4-16gb.json +69 -0
  172. package/src/config/runtime/tiers/gemma4-24gb.json +66 -0
  173. package/src/config/runtime/tiers/gemma4-32gb.json +66 -0
  174. package/src/config/runtime-merge.d.ts +5 -0
  175. package/src/config/runtime-merge.js +30 -0
  176. package/src/config/runtime.d.ts +28 -0
  177. package/src/config/runtime.js +69 -0
  178. package/src/config/schema/adapter.schema.d.ts +53 -0
  179. package/src/config/schema/adapter.schema.js +60 -0
  180. package/src/config/schema/backward-registry.schema.d.ts +14 -0
  181. package/src/config/schema/backward-registry.schema.js +46 -0
  182. package/src/config/schema/benchmark.schema.d.ts +54 -0
  183. package/src/config/schema/benchmark.schema.js +74 -0
  184. package/src/config/schema/bridge.schema.d.ts +25 -0
  185. package/src/config/schema/bridge.schema.js +22 -0
  186. package/src/config/schema/browser-suite-metrics.schema.d.ts +16 -0
  187. package/src/config/schema/browser-suite-metrics.schema.js +44 -0
  188. package/src/config/schema/buffer-pool.schema.d.ts +92 -0
  189. package/src/config/schema/buffer-pool.schema.js +50 -0
  190. package/src/config/schema/conversion-report.schema.d.ts +36 -0
  191. package/src/config/schema/conversion-report.schema.js +97 -0
  192. package/src/config/schema/conversion.schema.d.ts +191 -0
  193. package/src/config/schema/conversion.schema.js +13 -0
  194. package/src/config/schema/converter.schema.d.ts +117 -0
  195. package/src/config/schema/converter.schema.js +140 -0
  196. package/src/config/schema/debug.schema.d.ts +292 -0
  197. package/src/config/schema/debug.schema.js +134 -0
  198. package/src/config/schema/diffusion.schema.d.ts +88 -0
  199. package/src/config/schema/diffusion.schema.js +62 -0
  200. package/src/config/schema/distill-training.schema.d.ts +48 -0
  201. package/src/config/schema/distill-training.schema.js +139 -0
  202. package/src/config/schema/distribution.schema.d.ts +155 -0
  203. package/src/config/schema/distribution.schema.js +81 -0
  204. package/src/config/schema/doppler.schema.d.ts +74 -0
  205. package/src/config/schema/doppler.schema.js +341 -0
  206. package/src/config/schema/ecosystem.schema.d.ts +255 -0
  207. package/src/config/schema/ecosystem.schema.js +534 -0
  208. package/src/config/schema/emulation.schema.d.ts +351 -0
  209. package/src/config/schema/emulation.schema.js +301 -0
  210. package/src/config/schema/energy.schema.d.ts +102 -0
  211. package/src/config/schema/energy.schema.js +72 -0
  212. package/src/config/schema/execution-v1.schema.d.ts +258 -0
  213. package/src/config/schema/execution-v1.schema.js +211 -0
  214. package/src/config/schema/gpu-cache.schema.d.ts +26 -0
  215. package/src/config/schema/gpu-cache.schema.js +8 -0
  216. package/src/config/schema/harness.schema.d.ts +34 -0
  217. package/src/config/schema/harness.schema.js +21 -0
  218. package/src/config/schema/hotswap.schema.d.ts +55 -0
  219. package/src/config/schema/hotswap.schema.js +18 -0
  220. package/src/config/schema/index.d.ts +869 -0
  221. package/src/config/schema/index.js +498 -0
  222. package/src/config/schema/inference-defaults.schema.d.ts +276 -0
  223. package/src/config/schema/inference-defaults.schema.js +189 -0
  224. package/src/config/schema/inference.schema.d.ts +298 -0
  225. package/src/config/schema/inference.schema.js +39 -0
  226. package/src/config/schema/intent-bundle.schema.d.ts +28 -0
  227. package/src/config/schema/intent-bundle.schema.js +12 -0
  228. package/src/config/schema/kernel-path.schema.d.ts +184 -0
  229. package/src/config/schema/kernel-path.schema.js +9 -0
  230. package/src/config/schema/kernel-registry.schema.d.ts +199 -0
  231. package/src/config/schema/kernel-registry.schema.js +46 -0
  232. package/src/config/schema/kernel-thresholds.schema.d.ts +302 -0
  233. package/src/config/schema/kernel-thresholds.schema.js +195 -0
  234. package/src/config/schema/kernel-warmup.schema.d.ts +19 -0
  235. package/src/config/schema/kernel-warmup.schema.js +5 -0
  236. package/src/config/schema/kvcache.schema.d.ts +131 -0
  237. package/src/config/schema/kvcache.schema.js +41 -0
  238. package/src/config/schema/loading.schema.d.ts +153 -0
  239. package/src/config/schema/loading.schema.js +84 -0
  240. package/src/config/schema/lora.schema.d.ts +12 -0
  241. package/src/config/schema/lora.schema.js +12 -0
  242. package/src/config/schema/manifest.schema.d.ts +520 -0
  243. package/src/config/schema/manifest.schema.js +145 -0
  244. package/src/config/schema/memory-limits.schema.d.ts +107 -0
  245. package/src/config/schema/memory-limits.schema.js +57 -0
  246. package/src/config/schema/moe.schema.d.ts +78 -0
  247. package/src/config/schema/moe.schema.js +40 -0
  248. package/src/config/schema/platform.schema.d.ts +121 -0
  249. package/src/config/schema/platform.schema.js +1 -0
  250. package/src/config/schema/quantization-defaults.schema.d.ts +34 -0
  251. package/src/config/schema/quantization-defaults.schema.js +5 -0
  252. package/src/config/schema/quantization.schema.d.ts +10 -0
  253. package/src/config/schema/quantization.schema.js +33 -0
  254. package/src/config/schema/shared-runtime.schema.d.ts +75 -0
  255. package/src/config/schema/shared-runtime.schema.js +45 -0
  256. package/src/config/schema/speculation-self.schema.js +54 -0
  257. package/src/config/schema/speculative.schema.d.ts +21 -0
  258. package/src/config/schema/speculative.schema.js +11 -0
  259. package/src/config/schema/storage.schema.d.ts +123 -0
  260. package/src/config/schema/storage.schema.js +66 -0
  261. package/src/config/schema/tooling.schema.d.ts +29 -0
  262. package/src/config/schema/tooling.schema.js +12 -0
  263. package/src/config/schema/training-metrics.schema.d.ts +89 -0
  264. package/src/config/schema/training-metrics.schema.js +374 -0
  265. package/src/config/schema/training.schema.d.ts +88 -0
  266. package/src/config/schema/training.schema.js +106 -0
  267. package/src/config/schema/tuner.schema.d.ts +39 -0
  268. package/src/config/schema/tuner.schema.js +13 -0
  269. package/src/config/schema/ul-training.schema.d.ts +61 -0
  270. package/src/config/schema/ul-training.schema.js +140 -0
  271. package/src/config/schema/units.schema.d.ts +27 -0
  272. package/src/config/schema/units.schema.js +30 -0
  273. package/src/config/training-defaults.d.ts +24 -0
  274. package/src/config/training-defaults.js +99 -0
  275. package/src/config/transforms/capability-transform-resolver.d.ts +34 -0
  276. package/src/config/transforms/capability-transform-resolver.js +62 -0
  277. package/src/config/transforms/execution-graph-transforms.d.ts +92 -0
  278. package/src/config/transforms/execution-graph-transforms.js +512 -0
  279. package/src/converter/conversion-plan.d.ts +54 -0
  280. package/src/converter/conversion-plan.js +268 -0
  281. package/src/converter/core.d.ts +263 -0
  282. package/src/converter/core.js +1530 -0
  283. package/src/converter/index.d.ts +98 -0
  284. package/src/converter/index.js +59 -0
  285. package/src/converter/parsers/diffusion.d.ts +50 -0
  286. package/src/converter/parsers/diffusion.js +327 -0
  287. package/src/converter/parsers/gguf.d.ts +22 -0
  288. package/src/converter/parsers/gguf.js +46 -0
  289. package/src/converter/parsers/index.d.ts +21 -0
  290. package/src/converter/parsers/index.js +12 -0
  291. package/src/converter/parsers/transformer.d.ts +31 -0
  292. package/src/converter/parsers/transformer.js +198 -0
  293. package/src/converter/quantization-info.d.ts +37 -0
  294. package/src/converter/quantization-info.js +422 -0
  295. package/src/converter/quantizer.d.ts +101 -0
  296. package/src/converter/quantizer.js +444 -0
  297. package/src/converter/rope-config.d.ts +15 -0
  298. package/src/converter/rope-config.js +262 -0
  299. package/src/converter/shard-packer.d.ts +138 -0
  300. package/src/converter/shard-packer.js +425 -0
  301. package/src/converter/tokenizer-utils.d.ts +12 -0
  302. package/src/converter/tokenizer-utils.js +104 -0
  303. package/src/debug/capture-policy.d.ts +88 -0
  304. package/src/debug/capture-policy.js +196 -0
  305. package/src/debug/config.d.ts +78 -0
  306. package/src/debug/config.js +347 -0
  307. package/src/debug/history.d.ts +65 -0
  308. package/src/debug/history.js +82 -0
  309. package/src/debug/index.d.ts +268 -0
  310. package/src/debug/index.js +215 -0
  311. package/src/debug/log.d.ts +46 -0
  312. package/src/debug/log.js +132 -0
  313. package/src/debug/perf.d.ts +33 -0
  314. package/src/debug/perf.js +51 -0
  315. package/src/debug/reference/README.md +114 -0
  316. package/src/debug/reference/hf_attn_debug.py +114 -0
  317. package/src/debug/reference/hf_embed_check.py +89 -0
  318. package/src/debug/reference/hf_layer_out.py +100 -0
  319. package/src/debug/reference/hf_qwen35_linear_attn_debug.py +268 -0
  320. package/src/debug/reference/hf_rope_check.py +116 -0
  321. package/src/debug/reference/hf_weights.py +75 -0
  322. package/src/debug/signals.d.ts +63 -0
  323. package/src/debug/signals.js +39 -0
  324. package/src/debug/stats.d.ts +47 -0
  325. package/src/debug/stats.js +160 -0
  326. package/src/debug/tensor.d.ts +125 -0
  327. package/src/debug/tensor.js +268 -0
  328. package/src/debug/trace.d.ts +17 -0
  329. package/src/debug/trace.js +188 -0
  330. package/src/diffusion/config/expansion-lanes.json +29 -0
  331. package/src/diffusion/image-regression.d.ts +31 -0
  332. package/src/diffusion/image-regression.js +107 -0
  333. package/src/diffusion/index.d.ts +8 -0
  334. package/src/diffusion/index.js +8 -0
  335. package/src/distribution/p2p-control-plane.d.ts +52 -0
  336. package/src/distribution/p2p-control-plane.js +272 -0
  337. package/src/distribution/p2p-observability.d.ts +116 -0
  338. package/src/distribution/p2p-observability.js +303 -0
  339. package/src/distribution/p2p-transport-contract.d.ts +57 -0
  340. package/src/distribution/p2p-transport-contract.js +310 -0
  341. package/src/distribution/p2p-transport.js +75 -0
  342. package/src/distribution/p2p-webrtc-browser.d.ts +37 -0
  343. package/src/distribution/p2p-webrtc-browser.js +454 -0
  344. package/src/distribution/shard-delivery.d.ts +251 -0
  345. package/src/distribution/shard-delivery.js +2186 -0
  346. package/src/energy/index.d.ts +2 -0
  347. package/src/energy/index.js +2 -0
  348. package/src/errors/doppler-error.d.ts +20 -0
  349. package/src/errors/doppler-error.js +37 -0
  350. package/src/errors/index.d.ts +1 -0
  351. package/src/errors/index.js +1 -0
  352. package/src/formats/gguf/index.d.ts +8 -0
  353. package/src/formats/gguf/index.js +4 -0
  354. package/src/formats/gguf/types.d.ts +137 -0
  355. package/src/formats/gguf/types.js +460 -0
  356. package/src/formats/index.d.ts +51 -0
  357. package/src/formats/index.js +13 -0
  358. package/src/formats/rdrr/classification.d.ts +54 -0
  359. package/src/formats/rdrr/classification.js +329 -0
  360. package/src/formats/rdrr/groups.d.ts +35 -0
  361. package/src/formats/rdrr/groups.js +73 -0
  362. package/src/formats/rdrr/index.d.ts +25 -0
  363. package/src/formats/rdrr/index.js +19 -0
  364. package/src/formats/rdrr/manifest.d.ts +32 -0
  365. package/src/formats/rdrr/manifest.js +108 -0
  366. package/src/formats/rdrr/parsing.d.ts +27 -0
  367. package/src/formats/rdrr/parsing.js +167 -0
  368. package/src/formats/rdrr/tensor-config-validator.d.ts +42 -0
  369. package/src/formats/rdrr/tensor-config-validator.js +156 -0
  370. package/src/formats/rdrr/types.d.ts +201 -0
  371. package/src/formats/rdrr/types.js +16 -0
  372. package/src/formats/rdrr/validation.d.ts +9 -0
  373. package/src/formats/rdrr/validation.js +213 -0
  374. package/src/formats/safetensors/index.d.ts +8 -0
  375. package/src/formats/safetensors/index.js +4 -0
  376. package/src/formats/safetensors/types.d.ts +67 -0
  377. package/src/formats/safetensors/types.js +102 -0
  378. package/src/formats/tokenizer/index.d.ts +5 -0
  379. package/src/formats/tokenizer/index.js +3 -0
  380. package/src/formats/tokenizer/types.d.ts +9 -0
  381. package/src/formats/tokenizer/types.js +22 -0
  382. package/src/generation/index.d.ts +18 -0
  383. package/src/generation/index.js +32 -0
  384. package/src/gpu/command-recorder.d.ts +175 -0
  385. package/src/gpu/command-recorder.js +498 -0
  386. package/src/gpu/device.d.ts +142 -0
  387. package/src/gpu/device.js +532 -0
  388. package/src/gpu/kernel-runtime.d.ts +20 -0
  389. package/src/gpu/kernel-runtime.js +39 -0
  390. package/src/gpu/kernel-selection-cache.d.ts +13 -0
  391. package/src/gpu/kernel-selection-cache.js +13 -0
  392. package/src/gpu/kernel-selection-log.d.ts +12 -0
  393. package/src/gpu/kernel-selection-log.js +28 -0
  394. package/src/gpu/kernel-selector.d.ts +11 -0
  395. package/src/gpu/kernel-selector.js +10 -0
  396. package/src/gpu/kernel-tuner/benchmarks.d.ts +144 -0
  397. package/src/gpu/kernel-tuner/benchmarks.js +902 -0
  398. package/src/gpu/kernel-tuner/cache.d.ts +55 -0
  399. package/src/gpu/kernel-tuner/cache.js +152 -0
  400. package/src/gpu/kernel-tuner/index.d.ts +59 -0
  401. package/src/gpu/kernel-tuner/index.js +38 -0
  402. package/src/gpu/kernel-tuner/tuner.d.ts +82 -0
  403. package/src/gpu/kernel-tuner/tuner.js +247 -0
  404. package/src/gpu/kernel-tuner/types.d.ts +101 -0
  405. package/src/gpu/kernel-tuner/types.js +4 -0
  406. package/src/gpu/kernel-tuner.d.ts +33 -0
  407. package/src/gpu/kernel-tuner.js +12 -0
  408. package/src/gpu/kernels/README.md +127 -0
  409. package/src/gpu/kernels/attention.d.ts +236 -0
  410. package/src/gpu/kernels/attention.js +1437 -0
  411. package/src/gpu/kernels/attention.wgsl +249 -0
  412. package/src/gpu/kernels/attention_bdpa_decode_f16.wgsl +246 -0
  413. package/src/gpu/kernels/attention_decode.wgsl +233 -0
  414. package/src/gpu/kernels/attention_decode_chunked_f16.wgsl +183 -0
  415. package/src/gpu/kernels/attention_decode_chunked_f16kv.wgsl +208 -0
  416. package/src/gpu/kernels/attention_decode_f16.wgsl +202 -0
  417. package/src/gpu/kernels/attention_decode_f16kv.wgsl +224 -0
  418. package/src/gpu/kernels/attention_decode_online_f16.wgsl +223 -0
  419. package/src/gpu/kernels/attention_decode_online_f16kv.wgsl +225 -0
  420. package/src/gpu/kernels/attention_decode_optimized.wgsl +445 -0
  421. package/src/gpu/kernels/attention_decode_paged_f16.wgsl +172 -0
  422. package/src/gpu/kernels/attention_decode_paged_f16kv.wgsl +174 -0
  423. package/src/gpu/kernels/attention_decode_subgroup.wgsl +233 -0
  424. package/src/gpu/kernels/attention_decode_tiered_f16.wgsl +218 -0
  425. package/src/gpu/kernels/attention_decode_tiered_f16kv.wgsl +220 -0
  426. package/src/gpu/kernels/attention_decode_tiered_int4_f16kv.wgsl +242 -0
  427. package/src/gpu/kernels/attention_decode_tiered_int8_f16kv.wgsl +242 -0
  428. package/src/gpu/kernels/attention_f16.wgsl +214 -0
  429. package/src/gpu/kernels/attention_f16kv.wgsl +242 -0
  430. package/src/gpu/kernels/attention_small.wgsl +260 -0
  431. package/src/gpu/kernels/attention_small_f16.wgsl +240 -0
  432. package/src/gpu/kernels/attention_small_f16kv.wgsl +266 -0
  433. package/src/gpu/kernels/attention_streaming.wgsl +149 -0
  434. package/src/gpu/kernels/attention_streaming_f16.wgsl +147 -0
  435. package/src/gpu/kernels/attention_streaming_f16kv.wgsl +151 -0
  436. package/src/gpu/kernels/backward/adam.d.ts +28 -0
  437. package/src/gpu/kernels/backward/adam.js +203 -0
  438. package/src/gpu/kernels/backward/adam.wgsl +50 -0
  439. package/src/gpu/kernels/backward/attention_backward.d.ts +22 -0
  440. package/src/gpu/kernels/backward/attention_backward.js +364 -0
  441. package/src/gpu/kernels/backward/attention_backward.wgsl +49 -0
  442. package/src/gpu/kernels/backward/bias_add_backward.d.ts +17 -0
  443. package/src/gpu/kernels/backward/bias_add_backward.js +24 -0
  444. package/src/gpu/kernels/backward/bias_add_backward.wgsl +33 -0
  445. package/src/gpu/kernels/backward/conv2d_backward.d.ts +31 -0
  446. package/src/gpu/kernels/backward/conv2d_backward.js +148 -0
  447. package/src/gpu/kernels/backward/conv2d_backward_input.wgsl +83 -0
  448. package/src/gpu/kernels/backward/conv2d_backward_weight.wgsl +70 -0
  449. package/src/gpu/kernels/backward/cross_entropy_backward.d.ts +23 -0
  450. package/src/gpu/kernels/backward/cross_entropy_backward.js +29 -0
  451. package/src/gpu/kernels/backward/cross_entropy_backward.wgsl +39 -0
  452. package/src/gpu/kernels/backward/embed_backward.d.ts +29 -0
  453. package/src/gpu/kernels/backward/embed_backward.js +118 -0
  454. package/src/gpu/kernels/backward/embed_backward.wgsl +73 -0
  455. package/src/gpu/kernels/backward/gelu_backward.d.ts +16 -0
  456. package/src/gpu/kernels/backward/gelu_backward.js +39 -0
  457. package/src/gpu/kernels/backward/gelu_backward.wgsl +38 -0
  458. package/src/gpu/kernels/backward/groupnorm_backward.d.ts +24 -0
  459. package/src/gpu/kernels/backward/groupnorm_backward.js +29 -0
  460. package/src/gpu/kernels/backward/groupnorm_backward.wgsl +143 -0
  461. package/src/gpu/kernels/backward/index.d.ts +17 -0
  462. package/src/gpu/kernels/backward/index.js +23 -0
  463. package/src/gpu/kernels/backward/layernorm_backward.d.ts +22 -0
  464. package/src/gpu/kernels/backward/layernorm_backward.js +135 -0
  465. package/src/gpu/kernels/backward/layernorm_backward.wgsl +194 -0
  466. package/src/gpu/kernels/backward/matmul_backward.d.ts +32 -0
  467. package/src/gpu/kernels/backward/matmul_backward.js +124 -0
  468. package/src/gpu/kernels/backward/matmul_backward.wgsl +90 -0
  469. package/src/gpu/kernels/backward/matmul_transpose_a.wgsl +84 -0
  470. package/src/gpu/kernels/backward/pixel_shuffle_backward.d.ts +22 -0
  471. package/src/gpu/kernels/backward/pixel_shuffle_backward.js +30 -0
  472. package/src/gpu/kernels/backward/pixel_shuffle_backward.wgsl +54 -0
  473. package/src/gpu/kernels/backward/rmsnorm_backward.d.ts +24 -0
  474. package/src/gpu/kernels/backward/rmsnorm_backward.js +101 -0
  475. package/src/gpu/kernels/backward/rmsnorm_backward.wgsl +78 -0
  476. package/src/gpu/kernels/backward/rope_backward.d.ts +25 -0
  477. package/src/gpu/kernels/backward/rope_backward.js +109 -0
  478. package/src/gpu/kernels/backward/rope_backward.wgsl +59 -0
  479. package/src/gpu/kernels/backward/scale_backward.d.ts +16 -0
  480. package/src/gpu/kernels/backward/scale_backward.js +84 -0
  481. package/src/gpu/kernels/backward/scale_backward.wgsl +27 -0
  482. package/src/gpu/kernels/backward/silu_backward.d.ts +16 -0
  483. package/src/gpu/kernels/backward/silu_backward.js +39 -0
  484. package/src/gpu/kernels/backward/silu_backward.wgsl +31 -0
  485. package/src/gpu/kernels/backward/softmax_backward.d.ts +16 -0
  486. package/src/gpu/kernels/backward/softmax_backward.js +43 -0
  487. package/src/gpu/kernels/backward/softmax_backward.wgsl +44 -0
  488. package/src/gpu/kernels/backward/upsample2d_backward.d.ts +21 -0
  489. package/src/gpu/kernels/backward/upsample2d_backward.js +30 -0
  490. package/src/gpu/kernels/backward/upsample2d_backward.wgsl +59 -0
  491. package/src/gpu/kernels/backward/utils.d.ts +45 -0
  492. package/src/gpu/kernels/backward/utils.js +371 -0
  493. package/src/gpu/kernels/bf16_to_f16.wgsl +54 -0
  494. package/src/gpu/kernels/bf16_to_f32.wgsl +70 -0
  495. package/src/gpu/kernels/bias_add.wgsl +42 -0
  496. package/src/gpu/kernels/bias_add_f16.wgsl +47 -0
  497. package/src/gpu/kernels/cast.d.ts +67 -0
  498. package/src/gpu/kernels/cast.js +339 -0
  499. package/src/gpu/kernels/cast_f16_to_f32.wgsl +31 -0
  500. package/src/gpu/kernels/cast_f32_to_f16.wgsl +36 -0
  501. package/src/gpu/kernels/check-finiteness.d.ts +15 -0
  502. package/src/gpu/kernels/check-finiteness.js +149 -0
  503. package/src/gpu/kernels/check-stop.d.ts +31 -0
  504. package/src/gpu/kernels/check-stop.js +170 -0
  505. package/src/gpu/kernels/clamp.d.ts +22 -0
  506. package/src/gpu/kernels/clamp.js +42 -0
  507. package/src/gpu/kernels/clamp.wgsl +24 -0
  508. package/src/gpu/kernels/codegen/patches/attention_decode_f16.from.attention_decode.diff +213 -0
  509. package/src/gpu/kernels/codegen/patches/attention_decode_f16kv.from.attention_decode.diff +89 -0
  510. package/src/gpu/kernels/codegen/patches/attention_f16.from.attention.diff +258 -0
  511. package/src/gpu/kernels/codegen/patches/attention_f16kv.from.attention.diff +114 -0
  512. package/src/gpu/kernels/codegen/patches/attention_small_f16.from.attention_small.diff +193 -0
  513. package/src/gpu/kernels/codegen/patches/attention_small_f16kv.from.attention_small.diff +264 -0
  514. package/src/gpu/kernels/codegen/patches/attention_streaming_f16.from.attention_streaming.diff +109 -0
  515. package/src/gpu/kernels/codegen/patches/attention_streaming_f16kv.from.attention_streaming.diff +53 -0
  516. package/src/gpu/kernels/codegen/patches/bias_add_f16.from.bias_add.diff +54 -0
  517. package/src/gpu/kernels/codegen/patches/dequant_mxfp4_expert_f16.from.dequant_mxfp4_expert.diff +125 -0
  518. package/src/gpu/kernels/codegen/patches/fused_ffn_f16.from.fused_ffn.diff +428 -0
  519. package/src/gpu/kernels/codegen/patches/fused_matmul_q4_batched_f16.from.fused_matmul_q4_batched.diff +114 -0
  520. package/src/gpu/kernels/codegen/patches/fused_matmul_rmsnorm_f16.from.fused_matmul_rmsnorm.diff +477 -0
  521. package/src/gpu/kernels/codegen/patches/fused_swiglu_f16.from.fused_swiglu.diff +60 -0
  522. package/src/gpu/kernels/codegen/patches/gelu_f16.from.gelu.diff +58 -0
  523. package/src/gpu/kernels/codegen/patches/layernorm_f16.from.layernorm.diff +82 -0
  524. package/src/gpu/kernels/codegen/patches/matmul_gemv_residual_f16.from.matmul_gemv_residual.diff +87 -0
  525. package/src/gpu/kernels/codegen/patches/moe_gather_f16.from.moe_gather.diff +181 -0
  526. package/src/gpu/kernels/codegen/patches/residual_f16.from.residual.diff +62 -0
  527. package/src/gpu/kernels/codegen/patches/rmsnorm_f16.from.rmsnorm.diff +430 -0
  528. package/src/gpu/kernels/codegen/patches/rope_f16.from.rope.diff +462 -0
  529. package/src/gpu/kernels/codegen/patches/sample_f16.from.sample.diff +307 -0
  530. package/src/gpu/kernels/codegen/patches/scatter_add_dynamic_f16.from.scatter_add_dynamic.diff +65 -0
  531. package/src/gpu/kernels/codegen/patches/silu_f16.from.silu.diff +81 -0
  532. package/src/gpu/kernels/codegen/patches/topk_f16.from.topk.diff +231 -0
  533. package/src/gpu/kernels/codegen/wgsl-patch-variants.js +152 -0
  534. package/src/gpu/kernels/codegen/wgsl-variants.js +733 -0
  535. package/src/gpu/kernels/constants.d.ts +168 -0
  536. package/src/gpu/kernels/constants.js +142 -0
  537. package/src/gpu/kernels/conv2d.d.ts +34 -0
  538. package/src/gpu/kernels/conv2d.js +91 -0
  539. package/src/gpu/kernels/conv2d.wgsl +70 -0
  540. package/src/gpu/kernels/conv2d_f16.wgsl +72 -0
  541. package/src/gpu/kernels/cross_entropy_loss.d.ts +21 -0
  542. package/src/gpu/kernels/cross_entropy_loss.js +60 -0
  543. package/src/gpu/kernels/cross_entropy_loss.wgsl +39 -0
  544. package/src/gpu/kernels/depthwise_conv2d.d.ts +29 -0
  545. package/src/gpu/kernels/depthwise_conv2d.js +109 -0
  546. package/src/gpu/kernels/depthwise_conv2d.wgsl +55 -0
  547. package/src/gpu/kernels/depthwise_conv2d_f16.wgsl +59 -0
  548. package/src/gpu/kernels/dequant.d.ts +108 -0
  549. package/src/gpu/kernels/dequant.js +575 -0
  550. package/src/gpu/kernels/dequant_f16_out.wgsl +153 -0
  551. package/src/gpu/kernels/dequant_f16_out_vec4.wgsl +152 -0
  552. package/src/gpu/kernels/dequant_f16_rowwise.wgsl +139 -0
  553. package/src/gpu/kernels/dequant_f32_rowwise.wgsl +133 -0
  554. package/src/gpu/kernels/dequant_mxfp4.wgsl +120 -0
  555. package/src/gpu/kernels/dequant_mxfp4_expert.wgsl +129 -0
  556. package/src/gpu/kernels/dequant_mxfp4_expert_f16.wgsl +105 -0
  557. package/src/gpu/kernels/dequant_mxfp4_vec4.wgsl +116 -0
  558. package/src/gpu/kernels/dequant_q6k.wgsl +140 -0
  559. package/src/gpu/kernels/dequant_q8_0.wgsl +98 -0
  560. package/src/gpu/kernels/dequant_shared.wgsl +200 -0
  561. package/src/gpu/kernels/dequant_shared_vec4.wgsl +160 -0
  562. package/src/gpu/kernels/dequant_subgroup.wgsl +206 -0
  563. package/src/gpu/kernels/dispatch.d.ts +157 -0
  564. package/src/gpu/kernels/dispatch.js +235 -0
  565. package/src/gpu/kernels/energy.d.ts +113 -0
  566. package/src/gpu/kernels/energy.js +448 -0
  567. package/src/gpu/kernels/energy_eval.wgsl +26 -0
  568. package/src/gpu/kernels/energy_eval_f16.wgsl +30 -0
  569. package/src/gpu/kernels/energy_quintel_grad.wgsl +92 -0
  570. package/src/gpu/kernels/energy_quintel_grad_f16.wgsl +96 -0
  571. package/src/gpu/kernels/energy_quintel_reduce.wgsl +112 -0
  572. package/src/gpu/kernels/energy_quintel_reduce_f16.wgsl +116 -0
  573. package/src/gpu/kernels/energy_quintel_update.wgsl +92 -0
  574. package/src/gpu/kernels/energy_quintel_update_f16.wgsl +96 -0
  575. package/src/gpu/kernels/energy_update.wgsl +25 -0
  576. package/src/gpu/kernels/energy_update_f16.wgsl +30 -0
  577. package/src/gpu/kernels/feature-check.d.ts +42 -0
  578. package/src/gpu/kernels/feature-check.js +70 -0
  579. package/src/gpu/kernels/fused_ffn.d.ts +65 -0
  580. package/src/gpu/kernels/fused_ffn.js +326 -0
  581. package/src/gpu/kernels/fused_ffn.wgsl +420 -0
  582. package/src/gpu/kernels/fused_ffn_f16.wgsl +213 -0
  583. package/src/gpu/kernels/fused_ffn_q4k.wgsl +375 -0
  584. package/src/gpu/kernels/fused_matmul_q4.wgsl +404 -0
  585. package/src/gpu/kernels/fused_matmul_q4_batched.wgsl +194 -0
  586. package/src/gpu/kernels/fused_matmul_q4_batched_f16.wgsl +170 -0
  587. package/src/gpu/kernels/fused_matmul_q4_batched_f16a.wgsl +154 -0
  588. package/src/gpu/kernels/fused_matmul_q4_f16a.wgsl +219 -0
  589. package/src/gpu/kernels/fused_matmul_q4_multicol_f16.wgsl +216 -0
  590. package/src/gpu/kernels/fused_matmul_q4_multicol_f16a.wgsl +204 -0
  591. package/src/gpu/kernels/fused_matmul_residual.d.ts +46 -0
  592. package/src/gpu/kernels/fused_matmul_residual.js +175 -0
  593. package/src/gpu/kernels/fused_matmul_rmsnorm.d.ts +64 -0
  594. package/src/gpu/kernels/fused_matmul_rmsnorm.js +290 -0
  595. package/src/gpu/kernels/fused_matmul_rmsnorm.wgsl +324 -0
  596. package/src/gpu/kernels/fused_matmul_rmsnorm_f16.wgsl +303 -0
  597. package/src/gpu/kernels/fused_swiglu.wgsl +63 -0
  598. package/src/gpu/kernels/fused_swiglu_f16.wgsl +57 -0
  599. package/src/gpu/kernels/gated-short-conv.d.ts +63 -0
  600. package/src/gpu/kernels/gated-short-conv.js +284 -0
  601. package/src/gpu/kernels/gather.d.ts +64 -0
  602. package/src/gpu/kernels/gather.js +137 -0
  603. package/src/gpu/kernels/gather.wgsl +61 -0
  604. package/src/gpu/kernels/gather_f16.wgsl +65 -0
  605. package/src/gpu/kernels/gather_f16_f16_out.wgsl +55 -0
  606. package/src/gpu/kernels/gather_f16_out.wgsl +55 -0
  607. package/src/gpu/kernels/gather_f16_vec4.wgsl +76 -0
  608. package/src/gpu/kernels/gather_f16_vec4_f16_out.wgsl +68 -0
  609. package/src/gpu/kernels/gather_vec4.wgsl +74 -0
  610. package/src/gpu/kernels/gather_vec4_f16_out.wgsl +68 -0
  611. package/src/gpu/kernels/gelu.d.ts +33 -0
  612. package/src/gpu/kernels/gelu.js +55 -0
  613. package/src/gpu/kernels/gelu.wgsl +64 -0
  614. package/src/gpu/kernels/gelu_f16.wgsl +66 -0
  615. package/src/gpu/kernels/gptoss_mxfp4_expert_fused.wgsl +127 -0
  616. package/src/gpu/kernels/gptoss_router_topk.wgsl +119 -0
  617. package/src/gpu/kernels/grouped_pointwise_conv2d.d.ts +27 -0
  618. package/src/gpu/kernels/grouped_pointwise_conv2d.js +103 -0
  619. package/src/gpu/kernels/grouped_pointwise_conv2d.wgsl +44 -0
  620. package/src/gpu/kernels/grouped_pointwise_conv2d_f16.wgsl +48 -0
  621. package/src/gpu/kernels/groupnorm.d.ts +31 -0
  622. package/src/gpu/kernels/groupnorm.js +102 -0
  623. package/src/gpu/kernels/groupnorm_apply.wgsl +41 -0
  624. package/src/gpu/kernels/groupnorm_apply_f16.wgsl +46 -0
  625. package/src/gpu/kernels/groupnorm_stats.wgsl +76 -0
  626. package/src/gpu/kernels/groupnorm_stats_f16.wgsl +79 -0
  627. package/src/gpu/kernels/index.d.ts +374 -0
  628. package/src/gpu/kernels/index.js +315 -0
  629. package/src/gpu/kernels/kernel-base.d.ts +33 -0
  630. package/src/gpu/kernels/kernel-base.js +46 -0
  631. package/src/gpu/kernels/kernel-configs.d.ts +65 -0
  632. package/src/gpu/kernels/kernel-configs.js +60 -0
  633. package/src/gpu/kernels/kernel-tuning.d.ts +42 -0
  634. package/src/gpu/kernels/kernel-tuning.js +158 -0
  635. package/src/gpu/kernels/kv-quantize.d.ts +37 -0
  636. package/src/gpu/kernels/kv-quantize.js +141 -0
  637. package/src/gpu/kernels/kv_quantize_int4.wgsl +119 -0
  638. package/src/gpu/kernels/kv_quantize_int8.wgsl +119 -0
  639. package/src/gpu/kernels/layernorm.d.ts +37 -0
  640. package/src/gpu/kernels/layernorm.js +96 -0
  641. package/src/gpu/kernels/layernorm.wgsl +121 -0
  642. package/src/gpu/kernels/layernorm_f16.wgsl +103 -0
  643. package/src/gpu/kernels/linear-attention-core.d.ts +39 -0
  644. package/src/gpu/kernels/linear-attention-core.js +559 -0
  645. package/src/gpu/kernels/logit-merge.d.ts +110 -0
  646. package/src/gpu/kernels/logit-merge.js +394 -0
  647. package/src/gpu/kernels/matmul-dispatch.d.ts +38 -0
  648. package/src/gpu/kernels/matmul-dispatch.js +156 -0
  649. package/src/gpu/kernels/matmul-selection.d.ts +87 -0
  650. package/src/gpu/kernels/matmul-selection.js +518 -0
  651. package/src/gpu/kernels/matmul.d.ts +114 -0
  652. package/src/gpu/kernels/matmul.js +384 -0
  653. package/src/gpu/kernels/matmul_f16.wgsl +170 -0
  654. package/src/gpu/kernels/matmul_f16_tiled.wgsl +165 -0
  655. package/src/gpu/kernels/matmul_f16w_f32a.wgsl +89 -0
  656. package/src/gpu/kernels/matmul_f16w_f32a_tiled.wgsl +154 -0
  657. package/src/gpu/kernels/matmul_f32.wgsl +100 -0
  658. package/src/gpu/kernels/matmul_gemv.wgsl +80 -0
  659. package/src/gpu/kernels/matmul_gemv_f16a.wgsl +81 -0
  660. package/src/gpu/kernels/matmul_gemv_residual.wgsl +119 -0
  661. package/src/gpu/kernels/matmul_gemv_residual_f16.wgsl +78 -0
  662. package/src/gpu/kernels/matmul_gemv_subgroup.wgsl +429 -0
  663. package/src/gpu/kernels/matmul_gemv_subgroup_f16a.wgsl +674 -0
  664. package/src/gpu/kernels/modulate.d.ts +29 -0
  665. package/src/gpu/kernels/modulate.js +57 -0
  666. package/src/gpu/kernels/modulate.wgsl +40 -0
  667. package/src/gpu/kernels/modulate_f16.wgsl +43 -0
  668. package/src/gpu/kernels/moe.d.ts +164 -0
  669. package/src/gpu/kernels/moe.js +542 -0
  670. package/src/gpu/kernels/moe_gather.wgsl +170 -0
  671. package/src/gpu/kernels/moe_gather_f16.wgsl +82 -0
  672. package/src/gpu/kernels/moe_gather_vec4.wgsl +74 -0
  673. package/src/gpu/kernels/moe_offsets.wgsl +48 -0
  674. package/src/gpu/kernels/pipeline-cache.d.ts +88 -0
  675. package/src/gpu/kernels/pipeline-cache.js +305 -0
  676. package/src/gpu/kernels/pixel_shuffle.d.ts +27 -0
  677. package/src/gpu/kernels/pixel_shuffle.js +57 -0
  678. package/src/gpu/kernels/pixel_shuffle.wgsl +43 -0
  679. package/src/gpu/kernels/pixel_shuffle_f16.wgsl +46 -0
  680. package/src/gpu/kernels/relu.d.ts +18 -0
  681. package/src/gpu/kernels/relu.js +66 -0
  682. package/src/gpu/kernels/relu.wgsl +22 -0
  683. package/src/gpu/kernels/relu_f16.wgsl +24 -0
  684. package/src/gpu/kernels/rep-penalty.js +103 -0
  685. package/src/gpu/kernels/rep_penalty.wgsl +57 -0
  686. package/src/gpu/kernels/rep_penalty_f16.wgsl +54 -0
  687. package/src/gpu/kernels/repeat_channels.d.ts +21 -0
  688. package/src/gpu/kernels/repeat_channels.js +68 -0
  689. package/src/gpu/kernels/repeat_channels.wgsl +28 -0
  690. package/src/gpu/kernels/repeat_channels_f16.wgsl +30 -0
  691. package/src/gpu/kernels/residual.d.ts +74 -0
  692. package/src/gpu/kernels/residual.js +173 -0
  693. package/src/gpu/kernels/residual.wgsl +56 -0
  694. package/src/gpu/kernels/residual_f16.wgsl +36 -0
  695. package/src/gpu/kernels/residual_f16_vec4.wgsl +48 -0
  696. package/src/gpu/kernels/residual_vec4.wgsl +47 -0
  697. package/src/gpu/kernels/rmsnorm.d.ts +53 -0
  698. package/src/gpu/kernels/rmsnorm.js +256 -0
  699. package/src/gpu/kernels/rmsnorm.wgsl +449 -0
  700. package/src/gpu/kernels/rmsnorm_f16.wgsl +198 -0
  701. package/src/gpu/kernels/rope.d.ts +50 -0
  702. package/src/gpu/kernels/rope.js +91 -0
  703. package/src/gpu/kernels/rope.wgsl +344 -0
  704. package/src/gpu/kernels/rope_f16.wgsl +271 -0
  705. package/src/gpu/kernels/rule-matcher.d.ts +30 -0
  706. package/src/gpu/kernels/rule-matcher.js +42 -0
  707. package/src/gpu/kernels/rule-registry.d.ts +7 -0
  708. package/src/gpu/kernels/rule-registry.js +54 -0
  709. package/src/gpu/kernels/sample.d.ts +75 -0
  710. package/src/gpu/kernels/sample.js +565 -0
  711. package/src/gpu/kernels/sample.wgsl +407 -0
  712. package/src/gpu/kernels/sample_f16.wgsl +363 -0
  713. package/src/gpu/kernels/sana_linear_attention.d.ts +27 -0
  714. package/src/gpu/kernels/sana_linear_attention.js +129 -0
  715. package/src/gpu/kernels/sana_linear_attention_apply.wgsl +43 -0
  716. package/src/gpu/kernels/sana_linear_attention_apply_f16.wgsl +46 -0
  717. package/src/gpu/kernels/sana_linear_attention_summary.wgsl +51 -0
  718. package/src/gpu/kernels/sana_linear_attention_summary_f16.wgsl +53 -0
  719. package/src/gpu/kernels/scale.d.ts +35 -0
  720. package/src/gpu/kernels/scale.js +44 -0
  721. package/src/gpu/kernels/scale.wgsl +38 -0
  722. package/src/gpu/kernels/scatter_add.wgsl +88 -0
  723. package/src/gpu/kernels/scatter_add_dynamic.wgsl +59 -0
  724. package/src/gpu/kernels/scatter_add_dynamic_f16.wgsl +52 -0
  725. package/src/gpu/kernels/scatter_add_dynamic_f16_weights.wgsl +50 -0
  726. package/src/gpu/kernels/scatter_add_vec4.wgsl +70 -0
  727. package/src/gpu/kernels/shader-cache.d.ts +56 -0
  728. package/src/gpu/kernels/shader-cache.js +234 -0
  729. package/src/gpu/kernels/silu.d.ts +76 -0
  730. package/src/gpu/kernels/silu.js +406 -0
  731. package/src/gpu/kernels/silu.wgsl +109 -0
  732. package/src/gpu/kernels/silu_f16.wgsl +108 -0
  733. package/src/gpu/kernels/softmax.d.ts +57 -0
  734. package/src/gpu/kernels/softmax.js +125 -0
  735. package/src/gpu/kernels/softmax.wgsl +388 -0
  736. package/src/gpu/kernels/softmax_subgroup.wgsl +175 -0
  737. package/src/gpu/kernels/split_qg.d.ts +50 -0
  738. package/src/gpu/kernels/split_qg.js +46 -0
  739. package/src/gpu/kernels/split_qg.wgsl +58 -0
  740. package/src/gpu/kernels/split_qg_f16.wgsl +62 -0
  741. package/src/gpu/kernels/split_qkv.d.ts +51 -0
  742. package/src/gpu/kernels/split_qkv.js +51 -0
  743. package/src/gpu/kernels/split_qkv.wgsl +71 -0
  744. package/src/gpu/kernels/split_qkv_f16.wgsl +75 -0
  745. package/src/gpu/kernels/topk.wgsl +243 -0
  746. package/src/gpu/kernels/topk_f16.wgsl +108 -0
  747. package/src/gpu/kernels/topk_f16_weights.wgsl +101 -0
  748. package/src/gpu/kernels/transpose.d.ts +21 -0
  749. package/src/gpu/kernels/transpose.js +51 -0
  750. package/src/gpu/kernels/transpose.wgsl +33 -0
  751. package/src/gpu/kernels/types.d.ts +21 -0
  752. package/src/gpu/kernels/types.js +4 -0
  753. package/src/gpu/kernels/uniform-utils.d.ts +48 -0
  754. package/src/gpu/kernels/uniform-utils.js +94 -0
  755. package/src/gpu/kernels/upsample2d.d.ts +25 -0
  756. package/src/gpu/kernels/upsample2d.js +67 -0
  757. package/src/gpu/kernels/upsample2d.wgsl +34 -0
  758. package/src/gpu/kernels/upsample2d_f16.wgsl +38 -0
  759. package/src/gpu/kernels/utils.d.ts +106 -0
  760. package/src/gpu/kernels/utils.js +257 -0
  761. package/src/gpu/multi-model-recorder.d.ts +21 -0
  762. package/src/gpu/multi-model-recorder.js +31 -0
  763. package/src/gpu/partitioned-buffer-pool.d.ts +28 -0
  764. package/src/gpu/partitioned-buffer-pool.js +57 -0
  765. package/src/gpu/perf-guards.d.ts +25 -0
  766. package/src/gpu/perf-guards.js +133 -0
  767. package/src/gpu/profiler.d.ts +114 -0
  768. package/src/gpu/profiler.js +406 -0
  769. package/src/gpu/readback-utils.d.ts +16 -0
  770. package/src/gpu/readback-utils.js +41 -0
  771. package/src/gpu/submit-tracker.d.ts +111 -0
  772. package/src/gpu/submit-tracker.js +242 -0
  773. package/src/gpu/tensor.d.ts +69 -0
  774. package/src/gpu/tensor.js +75 -0
  775. package/src/gpu/uniform-cache.d.ts +109 -0
  776. package/src/gpu/uniform-cache.js +263 -0
  777. package/src/gpu/weight-buffer.d.ts +121 -0
  778. package/src/gpu/weight-buffer.js +132 -0
  779. package/src/hotswap/intent-bundle.d.ts +37 -0
  780. package/src/hotswap/intent-bundle.js +129 -0
  781. package/src/hotswap/manifest.d.ts +42 -0
  782. package/src/hotswap/manifest.js +124 -0
  783. package/src/hotswap/model-swap.js +42 -0
  784. package/src/hotswap/runtime.d.ts +31 -0
  785. package/src/hotswap/runtime.js +150 -0
  786. package/src/index-browser.d.ts +92 -0
  787. package/src/index-browser.js +68 -0
  788. package/src/index-internal.d.ts +2 -0
  789. package/src/index-internal.js +2 -0
  790. package/src/index.d.ts +103 -0
  791. package/src/index.js +76 -0
  792. package/src/inference/README.md +593 -0
  793. package/src/inference/browser-harness-contract-helpers.d.ts +5 -0
  794. package/src/inference/browser-harness-contract-helpers.js +26 -0
  795. package/src/inference/browser-harness-diffusion-energy-suites.d.ts +2 -0
  796. package/src/inference/browser-harness-diffusion-energy-suites.js +269 -0
  797. package/src/inference/browser-harness-model-helpers.d.ts +16 -0
  798. package/src/inference/browser-harness-model-helpers.js +217 -0
  799. package/src/inference/browser-harness-report-helpers.d.ts +7 -0
  800. package/src/inference/browser-harness-report-helpers.js +42 -0
  801. package/src/inference/browser-harness-runtime-helpers.d.ts +61 -0
  802. package/src/inference/browser-harness-runtime-helpers.js +416 -0
  803. package/src/inference/browser-harness-suite-helpers.d.ts +28 -0
  804. package/src/inference/browser-harness-suite-helpers.js +270 -0
  805. package/src/inference/browser-harness-text-helpers.d.ts +34 -0
  806. package/src/inference/browser-harness-text-helpers.js +879 -0
  807. package/src/inference/browser-harness.d.ts +249 -0
  808. package/src/inference/browser-harness.js +1150 -0
  809. package/src/inference/decode-buffers.d.ts +108 -0
  810. package/src/inference/decode-buffers.js +181 -0
  811. package/src/inference/decode-ring.d.ts +52 -0
  812. package/src/inference/decode-ring.js +273 -0
  813. package/src/inference/expert-router.d.ts +27 -0
  814. package/src/inference/expert-router.js +55 -0
  815. package/src/inference/fixtures/embedding-semantic-fixtures.json +382 -0
  816. package/src/inference/functiongemma.d.ts +15 -0
  817. package/src/inference/functiongemma.js +1 -0
  818. package/src/inference/kv-cache/base.d.ts +150 -0
  819. package/src/inference/kv-cache/base.js +1105 -0
  820. package/src/inference/kv-cache/basis-decomposed-paged.d.ts +50 -0
  821. package/src/inference/kv-cache/basis-decomposed-paged.js +276 -0
  822. package/src/inference/kv-cache/index.d.ts +35 -0
  823. package/src/inference/kv-cache/index.js +20 -0
  824. package/src/inference/kv-cache/sliding-window.d.ts +72 -0
  825. package/src/inference/kv-cache/sliding-window.js +243 -0
  826. package/src/inference/kv-cache/tiered.d.ts +89 -0
  827. package/src/inference/kv-cache/tiered.js +576 -0
  828. package/src/inference/kv-cache/types.d.ts +188 -0
  829. package/src/inference/kv-cache/types.js +80 -0
  830. package/src/inference/kv-cache.d.ts +36 -0
  831. package/src/inference/kv-cache.js +18 -0
  832. package/src/inference/moe-router.d.ts +212 -0
  833. package/src/inference/moe-router.js +600 -0
  834. package/src/inference/multi-model-network.d.ts +139 -0
  835. package/src/inference/multi-model-network.js +771 -0
  836. package/src/inference/multi-pipeline-pool.d.ts +62 -0
  837. package/src/inference/multi-pipeline-pool.js +161 -0
  838. package/src/inference/network-evolution.d.ts +55 -0
  839. package/src/inference/network-evolution.js +79 -0
  840. package/src/inference/pipelines/context.d.ts +21 -0
  841. package/src/inference/pipelines/context.js +184 -0
  842. package/src/inference/pipelines/diffusion/helpers.d.ts +29 -0
  843. package/src/inference/pipelines/diffusion/helpers.js +120 -0
  844. package/src/inference/pipelines/diffusion/index.d.ts +3 -0
  845. package/src/inference/pipelines/diffusion/index.js +3 -0
  846. package/src/inference/pipelines/diffusion/init.d.ts +24 -0
  847. package/src/inference/pipelines/diffusion/init.js +138 -0
  848. package/src/inference/pipelines/diffusion/pipeline.d.ts +38 -0
  849. package/src/inference/pipelines/diffusion/pipeline.js +782 -0
  850. package/src/inference/pipelines/diffusion/sana-transformer.d.ts +53 -0
  851. package/src/inference/pipelines/diffusion/sana-transformer.js +738 -0
  852. package/src/inference/pipelines/diffusion/scheduler.d.ts +35 -0
  853. package/src/inference/pipelines/diffusion/scheduler.js +153 -0
  854. package/src/inference/pipelines/diffusion/sd3-transformer.d.ts +20 -0
  855. package/src/inference/pipelines/diffusion/sd3-transformer.js +1194 -0
  856. package/src/inference/pipelines/diffusion/sd3-weights.d.ts +21 -0
  857. package/src/inference/pipelines/diffusion/sd3-weights.js +287 -0
  858. package/src/inference/pipelines/diffusion/text-encoder-gpu.d.ts +87 -0
  859. package/src/inference/pipelines/diffusion/text-encoder-gpu.js +1224 -0
  860. package/src/inference/pipelines/diffusion/text-encoder.d.ts +29 -0
  861. package/src/inference/pipelines/diffusion/text-encoder.js +195 -0
  862. package/src/inference/pipelines/diffusion/types.d.ts +116 -0
  863. package/src/inference/pipelines/diffusion/types.js +1 -0
  864. package/src/inference/pipelines/diffusion/vae.d.ts +20 -0
  865. package/src/inference/pipelines/diffusion/vae.js +1375 -0
  866. package/src/inference/pipelines/diffusion/weights.d.ts +40 -0
  867. package/src/inference/pipelines/diffusion/weights.js +155 -0
  868. package/src/inference/pipelines/dream/energy-head-pipeline.d.ts +29 -0
  869. package/src/inference/pipelines/dream/energy-head-pipeline.js +6 -0
  870. package/src/inference/pipelines/dream/pipeline.d.ts +17 -0
  871. package/src/inference/pipelines/dream/pipeline.js +8 -0
  872. package/src/inference/pipelines/energy/index.d.ts +1 -0
  873. package/src/inference/pipelines/energy/index.js +1 -0
  874. package/src/inference/pipelines/energy/pipeline.d.ts +27 -0
  875. package/src/inference/pipelines/energy/pipeline.js +697 -0
  876. package/src/inference/pipelines/energy/quintel.d.ts +92 -0
  877. package/src/inference/pipelines/energy/quintel.js +218 -0
  878. package/src/inference/pipelines/energy/types.d.ts +63 -0
  879. package/src/inference/pipelines/energy/types.js +1 -0
  880. package/src/inference/pipelines/energy-head/index.d.ts +6 -0
  881. package/src/inference/pipelines/energy-head/index.js +6 -0
  882. package/src/inference/pipelines/energy-head/row-head-pipeline.d.ts +103 -0
  883. package/src/inference/pipelines/energy-head/row-head-pipeline.js +491 -0
  884. package/src/inference/pipelines/factory.d.ts +10 -0
  885. package/src/inference/pipelines/factory.js +6 -0
  886. package/src/inference/pipelines/index.d.ts +22 -0
  887. package/src/inference/pipelines/index.js +19 -0
  888. package/src/inference/pipelines/registry.d.ts +15 -0
  889. package/src/inference/pipelines/registry.js +23 -0
  890. package/src/inference/pipelines/rng.d.ts +2 -0
  891. package/src/inference/pipelines/rng.js +17 -0
  892. package/src/inference/pipelines/structured/index.d.ts +8 -0
  893. package/src/inference/pipelines/structured/index.js +8 -0
  894. package/src/inference/pipelines/structured/json-head-pipeline.d.ts +58 -0
  895. package/src/inference/pipelines/structured/json-head-pipeline.js +196 -0
  896. package/src/inference/pipelines/text/attention/attn-config.d.ts +28 -0
  897. package/src/inference/pipelines/text/attention/attn-config.js +51 -0
  898. package/src/inference/pipelines/text/attention/dispatch-params.d.ts +148 -0
  899. package/src/inference/pipelines/text/attention/dispatch-params.js +251 -0
  900. package/src/inference/pipelines/text/attention/index.d.ts +24 -0
  901. package/src/inference/pipelines/text/attention/index.js +17 -0
  902. package/src/inference/pipelines/text/attention/output-projection.d.ts +12 -0
  903. package/src/inference/pipelines/text/attention/output-projection.js +8 -0
  904. package/src/inference/pipelines/text/attention/projections.d.ts +114 -0
  905. package/src/inference/pipelines/text/attention/projections.js +531 -0
  906. package/src/inference/pipelines/text/attention/record.d.ts +36 -0
  907. package/src/inference/pipelines/text/attention/record.js +523 -0
  908. package/src/inference/pipelines/text/attention/run.d.ts +38 -0
  909. package/src/inference/pipelines/text/attention/run.js +813 -0
  910. package/src/inference/pipelines/text/attention/types.d.ts +108 -0
  911. package/src/inference/pipelines/text/attention/types.js +67 -0
  912. package/src/inference/pipelines/text/attention.d.ts +23 -0
  913. package/src/inference/pipelines/text/attention.js +12 -0
  914. package/src/inference/pipelines/text/bdpa-steamroller.d.ts +22 -0
  915. package/src/inference/pipelines/text/bdpa-steamroller.js +158 -0
  916. package/src/inference/pipelines/text/buffer-types.d.ts +7 -0
  917. package/src/inference/pipelines/text/buffer-types.js +4 -0
  918. package/src/inference/pipelines/text/chat-format.d.ts +46 -0
  919. package/src/inference/pipelines/text/chat-format.js +365 -0
  920. package/src/inference/pipelines/text/config.d.ts +253 -0
  921. package/src/inference/pipelines/text/config.js +857 -0
  922. package/src/inference/pipelines/text/debug-utils/config.d.ts +144 -0
  923. package/src/inference/pipelines/text/debug-utils/config.js +156 -0
  924. package/src/inference/pipelines/text/debug-utils/index.d.ts +53 -0
  925. package/src/inference/pipelines/text/debug-utils/index.js +44 -0
  926. package/src/inference/pipelines/text/debug-utils/logging.d.ts +106 -0
  927. package/src/inference/pipelines/text/debug-utils/logging.js +152 -0
  928. package/src/inference/pipelines/text/debug-utils/tensor.d.ts +119 -0
  929. package/src/inference/pipelines/text/debug-utils/tensor.js +268 -0
  930. package/src/inference/pipelines/text/debug-utils/utils.d.ts +77 -0
  931. package/src/inference/pipelines/text/debug-utils/utils.js +139 -0
  932. package/src/inference/pipelines/text/debug-utils.d.ts +42 -0
  933. package/src/inference/pipelines/text/debug-utils.js +34 -0
  934. package/src/inference/pipelines/text/drift-policy.d.ts +57 -0
  935. package/src/inference/pipelines/text/drift-policy.js +133 -0
  936. package/src/inference/pipelines/text/dtype-contract.d.ts +38 -0
  937. package/src/inference/pipelines/text/dtype-contract.js +87 -0
  938. package/src/inference/pipelines/text/embed.d.ts +67 -0
  939. package/src/inference/pipelines/text/embed.js +490 -0
  940. package/src/inference/pipelines/text/execution-plan.d.ts +123 -0
  941. package/src/inference/pipelines/text/execution-plan.js +371 -0
  942. package/src/inference/pipelines/text/execution-runtime-builders.d.ts +45 -0
  943. package/src/inference/pipelines/text/execution-runtime-builders.js +382 -0
  944. package/src/inference/pipelines/text/execution-v1.d.ts +45 -0
  945. package/src/inference/pipelines/text/execution-v1.js +288 -0
  946. package/src/inference/pipelines/text/failure-reduction.d.ts +84 -0
  947. package/src/inference/pipelines/text/failure-reduction.js +180 -0
  948. package/src/inference/pipelines/text/ffn/dense.d.ts +40 -0
  949. package/src/inference/pipelines/text/ffn/dense.js +753 -0
  950. package/src/inference/pipelines/text/ffn/index.d.ts +23 -0
  951. package/src/inference/pipelines/text/ffn/index.js +16 -0
  952. package/src/inference/pipelines/text/ffn/moe.d.ts +21 -0
  953. package/src/inference/pipelines/text/ffn/moe.js +49 -0
  954. package/src/inference/pipelines/text/ffn/sandwich.d.ts +25 -0
  955. package/src/inference/pipelines/text/ffn/sandwich.js +196 -0
  956. package/src/inference/pipelines/text/ffn/standard.d.ts +23 -0
  957. package/src/inference/pipelines/text/ffn/standard.js +113 -0
  958. package/src/inference/pipelines/text/ffn/types.d.ts +30 -0
  959. package/src/inference/pipelines/text/ffn/types.js +25 -0
  960. package/src/inference/pipelines/text/ffn.d.ts +31 -0
  961. package/src/inference/pipelines/text/ffn.js +18 -0
  962. package/src/inference/pipelines/text/finiteness-guard-status.d.ts +11 -0
  963. package/src/inference/pipelines/text/finiteness-guard-status.js +21 -0
  964. package/src/inference/pipelines/text/finiteness-policy.d.ts +35 -0
  965. package/src/inference/pipelines/text/finiteness-policy.js +45 -0
  966. package/src/inference/pipelines/text/generator-helpers.d.ts +34 -0
  967. package/src/inference/pipelines/text/generator-helpers.js +178 -0
  968. package/src/inference/pipelines/text/generator-runtime.d.ts +95 -0
  969. package/src/inference/pipelines/text/generator-runtime.js +492 -0
  970. package/src/inference/pipelines/text/generator-steps.d.ts +136 -0
  971. package/src/inference/pipelines/text/generator-steps.js +1294 -0
  972. package/src/inference/pipelines/text/generator.d.ts +46 -0
  973. package/src/inference/pipelines/text/generator.js +1779 -0
  974. package/src/inference/pipelines/text/index.d.ts +5 -0
  975. package/src/inference/pipelines/text/index.js +6 -0
  976. package/src/inference/pipelines/text/init.d.ts +316 -0
  977. package/src/inference/pipelines/text/init.js +1187 -0
  978. package/src/inference/pipelines/text/kernel-trace.d.ts +152 -0
  979. package/src/inference/pipelines/text/kernel-trace.js +330 -0
  980. package/src/inference/pipelines/text/layer-plan.d.ts +65 -0
  981. package/src/inference/pipelines/text/layer-plan.js +249 -0
  982. package/src/inference/pipelines/text/layer.d.ts +56 -0
  983. package/src/inference/pipelines/text/layer.js +974 -0
  984. package/src/inference/pipelines/text/linear-attention.d.ts +109 -0
  985. package/src/inference/pipelines/text/linear-attention.js +948 -0
  986. package/src/inference/pipelines/text/logits/cpu.d.ts +81 -0
  987. package/src/inference/pipelines/text/logits/cpu.js +91 -0
  988. package/src/inference/pipelines/text/logits/gpu.d.ts +113 -0
  989. package/src/inference/pipelines/text/logits/gpu.js +419 -0
  990. package/src/inference/pipelines/text/logits/index.d.ts +62 -0
  991. package/src/inference/pipelines/text/logits/index.js +316 -0
  992. package/src/inference/pipelines/text/logits/types.d.ts +46 -0
  993. package/src/inference/pipelines/text/logits/types.js +4 -0
  994. package/src/inference/pipelines/text/logits/utils.d.ts +56 -0
  995. package/src/inference/pipelines/text/logits/utils.js +68 -0
  996. package/src/inference/pipelines/text/logits.d.ts +27 -0
  997. package/src/inference/pipelines/text/logits.js +16 -0
  998. package/src/inference/pipelines/text/lora-apply.d.ts +28 -0
  999. package/src/inference/pipelines/text/lora-apply.js +76 -0
  1000. package/src/inference/pipelines/text/lora-types.d.ts +39 -0
  1001. package/src/inference/pipelines/text/lora-types.js +18 -0
  1002. package/src/inference/pipelines/text/lora.d.ts +18 -0
  1003. package/src/inference/pipelines/text/lora.js +12 -0
  1004. package/src/inference/pipelines/text/model-load.d.ts +58 -0
  1005. package/src/inference/pipelines/text/model-load.js +778 -0
  1006. package/src/inference/pipelines/text/moe-cache.d.ts +32 -0
  1007. package/src/inference/pipelines/text/moe-cache.js +108 -0
  1008. package/src/inference/pipelines/text/moe-cpu-gptoss.d.ts +9 -0
  1009. package/src/inference/pipelines/text/moe-cpu-gptoss.js +115 -0
  1010. package/src/inference/pipelines/text/moe-cpu.d.ts +13 -0
  1011. package/src/inference/pipelines/text/moe-cpu.js +120 -0
  1012. package/src/inference/pipelines/text/moe-gpu.d.ts +13 -0
  1013. package/src/inference/pipelines/text/moe-gpu.js +653 -0
  1014. package/src/inference/pipelines/text/moe-helpers.d.ts +12 -0
  1015. package/src/inference/pipelines/text/moe-helpers.js +21 -0
  1016. package/src/inference/pipelines/text/moe-impl.d.ts +117 -0
  1017. package/src/inference/pipelines/text/moe-impl.js +9 -0
  1018. package/src/inference/pipelines/text/moe-shape-validator.d.ts +40 -0
  1019. package/src/inference/pipelines/text/moe-shape-validator.js +98 -0
  1020. package/src/inference/pipelines/text/operator-events.d.ts +125 -0
  1021. package/src/inference/pipelines/text/operator-events.js +305 -0
  1022. package/src/inference/pipelines/text/operator-identity.d.ts +60 -0
  1023. package/src/inference/pipelines/text/operator-identity.js +133 -0
  1024. package/src/inference/pipelines/text/ops.d.ts +167 -0
  1025. package/src/inference/pipelines/text/ops.js +437 -0
  1026. package/src/inference/pipelines/text/probes.d.ts +31 -0
  1027. package/src/inference/pipelines/text/probes.js +240 -0
  1028. package/src/inference/pipelines/text/sampling-config.d.ts +54 -0
  1029. package/src/inference/pipelines/text/sampling-config.js +71 -0
  1030. package/src/inference/pipelines/text/sampling.d.ts +54 -0
  1031. package/src/inference/pipelines/text/sampling.js +249 -0
  1032. package/src/inference/pipelines/text/stage-names.d.ts +88 -0
  1033. package/src/inference/pipelines/text/stage-names.js +202 -0
  1034. package/src/inference/pipelines/text/state.d.ts +114 -0
  1035. package/src/inference/pipelines/text/state.js +165 -0
  1036. package/src/inference/pipelines/text/types.d.ts +682 -0
  1037. package/src/inference/pipelines/text/types.js +4 -0
  1038. package/src/inference/pipelines/text/weights.d.ts +110 -0
  1039. package/src/inference/pipelines/text/weights.js +173 -0
  1040. package/src/inference/pipelines/text.d.ts +162 -0
  1041. package/src/inference/pipelines/text.js +852 -0
  1042. package/src/inference/pipelines/vision/encoder.js +386 -0
  1043. package/src/inference/pipelines/vision/image-preprocess.js +157 -0
  1044. package/src/inference/pipelines/vision/index.js +173 -0
  1045. package/src/inference/pipelines/vision/ops.js +78 -0
  1046. package/src/inference/pipelines/vision/patch-embed.js +151 -0
  1047. package/src/inference/speculative.d.ts +239 -0
  1048. package/src/inference/speculative.js +402 -0
  1049. package/src/inference/test-harness.d.ts +178 -0
  1050. package/src/inference/test-harness.js +361 -0
  1051. package/src/inference/tokenizer.d.ts +72 -0
  1052. package/src/inference/tokenizer.js +251 -0
  1053. package/src/inference/tokenizers/base.d.ts +39 -0
  1054. package/src/inference/tokenizers/base.js +69 -0
  1055. package/src/inference/tokenizers/bpe.d.ts +27 -0
  1056. package/src/inference/tokenizers/bpe.js +180 -0
  1057. package/src/inference/tokenizers/bundled.d.ts +63 -0
  1058. package/src/inference/tokenizers/bundled.js +1009 -0
  1059. package/src/inference/tokenizers/sentencepiece.d.ts +28 -0
  1060. package/src/inference/tokenizers/sentencepiece.js +401 -0
  1061. package/src/inference/tokenizers/types.d.ts +166 -0
  1062. package/src/inference/tokenizers/types.js +7 -0
  1063. package/src/loader/doppler-loader.d.ts +139 -0
  1064. package/src/loader/doppler-loader.js +1103 -0
  1065. package/src/loader/dtype-utils.d.ts +40 -0
  1066. package/src/loader/dtype-utils.js +61 -0
  1067. package/src/loader/embedding-loader.d.ts +56 -0
  1068. package/src/loader/embedding-loader.js +211 -0
  1069. package/src/loader/experts/expert-cache.d.ts +156 -0
  1070. package/src/loader/experts/expert-cache.js +386 -0
  1071. package/src/loader/experts/expert-loader.d.ts +108 -0
  1072. package/src/loader/experts/expert-loader.js +392 -0
  1073. package/src/loader/final-weights-loader.d.ts +88 -0
  1074. package/src/loader/final-weights-loader.js +339 -0
  1075. package/src/loader/index.d.ts +150 -0
  1076. package/src/loader/index.js +124 -0
  1077. package/src/loader/layer-loader.d.ts +63 -0
  1078. package/src/loader/layer-loader.js +474 -0
  1079. package/src/loader/loader-state.d.ts +51 -0
  1080. package/src/loader/loader-state.js +142 -0
  1081. package/src/loader/loader-types.d.ts +236 -0
  1082. package/src/loader/loader-types.js +4 -0
  1083. package/src/loader/manifest-config.d.ts +97 -0
  1084. package/src/loader/manifest-config.js +134 -0
  1085. package/src/loader/memory-monitor.d.ts +112 -0
  1086. package/src/loader/memory-monitor.js +284 -0
  1087. package/src/loader/multi-model-loader.d.ts +51 -0
  1088. package/src/loader/multi-model-loader.js +133 -0
  1089. package/src/loader/quantization-constants.d.ts +23 -0
  1090. package/src/loader/quantization-constants.js +14 -0
  1091. package/src/loader/shard-cache.d.ts +60 -0
  1092. package/src/loader/shard-cache.js +663 -0
  1093. package/src/loader/shard-resolver.d.ts +12 -0
  1094. package/src/loader/shard-resolver.js +144 -0
  1095. package/src/loader/tensors/tensor-loader.d.ts +157 -0
  1096. package/src/loader/tensors/tensor-loader.js +652 -0
  1097. package/src/loader/tensors/tensor-reader.d.ts +22 -0
  1098. package/src/loader/tensors/tensor-reader.js +113 -0
  1099. package/src/loader/tensors/tensor-role.d.ts +7 -0
  1100. package/src/loader/tensors/tensor-role.js +12 -0
  1101. package/src/loader/weight-downcast.d.ts +62 -0
  1102. package/src/loader/weight-downcast.js +213 -0
  1103. package/src/loader/weights.d.ts +22 -0
  1104. package/src/loader/weights.js +4 -0
  1105. package/src/memory/address-table.d.ts +104 -0
  1106. package/src/memory/address-table.js +114 -0
  1107. package/src/memory/buffer-pool.d.ts +204 -0
  1108. package/src/memory/buffer-pool.js +867 -0
  1109. package/src/memory/capability.d.ts +49 -0
  1110. package/src/memory/capability.js +95 -0
  1111. package/src/memory/heap-manager.d.ts +104 -0
  1112. package/src/memory/heap-manager.js +264 -0
  1113. package/src/memory/unified-detect.d.ts +59 -0
  1114. package/src/memory/unified-detect.js +192 -0
  1115. package/src/rules/converter/execution.rules.json +20 -0
  1116. package/src/rules/converter/tensor-roles.rules.json +13 -0
  1117. package/src/rules/converter/tokenizer.rules.json +7 -0
  1118. package/src/rules/execution-rules-contract-check.d.ts +17 -0
  1119. package/src/rules/execution-rules-contract-check.js +287 -0
  1120. package/src/rules/inference/attention.rules.json +54 -0
  1121. package/src/rules/inference/capability-transforms.rules.json +24 -0
  1122. package/src/rules/inference/config.rules.json +57 -0
  1123. package/src/rules/inference/dtype.rules.json +99 -0
  1124. package/src/rules/inference/execution.rules.json +56 -0
  1125. package/src/rules/inference/ffn.rules.json +35 -0
  1126. package/src/rules/inference/layer-pattern.rules.json +16 -0
  1127. package/src/rules/inference/layer.rules.json +7 -0
  1128. package/src/rules/inference/moe.rules.json +48 -0
  1129. package/src/rules/kernels/attention.rules.json +61 -0
  1130. package/src/rules/kernels/conv2d.rules.json +6 -0
  1131. package/src/rules/kernels/depthwise-conv2d.rules.json +6 -0
  1132. package/src/rules/kernels/dequant.rules.json +58 -0
  1133. package/src/rules/kernels/energy.rules.json +22 -0
  1134. package/src/rules/kernels/fused-ffn.rules.json +13 -0
  1135. package/src/rules/kernels/fused-matmul-residual.rules.json +6 -0
  1136. package/src/rules/kernels/fused-matmul-rmsnorm.rules.json +8 -0
  1137. package/src/rules/kernels/gather.rules.json +12 -0
  1138. package/src/rules/kernels/gelu.rules.json +11 -0
  1139. package/src/rules/kernels/grouped-pointwise-conv2d.rules.json +6 -0
  1140. package/src/rules/kernels/groupnorm.rules.json +10 -0
  1141. package/src/rules/kernels/kernel-validator.d.ts +24 -0
  1142. package/src/rules/kernels/kernel-validator.js +160 -0
  1143. package/src/rules/kernels/kv_quantize.rules.json +7 -0
  1144. package/src/rules/kernels/layernorm.rules.json +6 -0
  1145. package/src/rules/kernels/matmul.rules.json +60 -0
  1146. package/src/rules/kernels/modulate.rules.json +6 -0
  1147. package/src/rules/kernels/moe.rules.gptoss.json +105 -0
  1148. package/src/rules/kernels/moe.rules.json +11 -0
  1149. package/src/rules/kernels/moe.rules.mixtral.json +75 -0
  1150. package/src/rules/kernels/pixel_shuffle.rules.json +6 -0
  1151. package/src/rules/kernels/relu.rules.json +6 -0
  1152. package/src/rules/kernels/repeat-channels.rules.json +6 -0
  1153. package/src/rules/kernels/residual.rules.json +12 -0
  1154. package/src/rules/kernels/rmsnorm.rules.json +11 -0
  1155. package/src/rules/kernels/rope.rules.json +6 -0
  1156. package/src/rules/kernels/sample.rules.json +6 -0
  1157. package/src/rules/kernels/sana-linear-attention.rules.json +6 -0
  1158. package/src/rules/kernels/scale.rules.json +6 -0
  1159. package/src/rules/kernels/silu.rules.json +21 -0
  1160. package/src/rules/kernels/softmax.rules.json +25 -0
  1161. package/src/rules/kernels/split-qg.rules.json +6 -0
  1162. package/src/rules/kernels/split-qkv.rules.json +6 -0
  1163. package/src/rules/kernels/upsample2d.rules.json +6 -0
  1164. package/src/rules/layer-pattern-contract-check.d.ts +17 -0
  1165. package/src/rules/layer-pattern-contract-check.js +231 -0
  1166. package/src/rules/loader/tensor-loader.rules.json +24 -0
  1167. package/src/rules/loader/weights.rules.json +41 -0
  1168. package/src/rules/rule-registry.d.ts +77 -0
  1169. package/src/rules/rule-registry.js +264 -0
  1170. package/src/rules/tooling/command-runtime.rules.json +65 -0
  1171. package/src/storage/backends/idb-store.d.ts +52 -0
  1172. package/src/storage/backends/idb-store.js +590 -0
  1173. package/src/storage/backends/memory-store.d.ts +36 -0
  1174. package/src/storage/backends/memory-store.js +242 -0
  1175. package/src/storage/backends/opfs-store.d.ts +41 -0
  1176. package/src/storage/backends/opfs-store.js +478 -0
  1177. package/src/storage/blake3.d.ts +17 -0
  1178. package/src/storage/blake3.js +269 -0
  1179. package/src/storage/download-types.d.ts +157 -0
  1180. package/src/storage/download-types.js +48 -0
  1181. package/src/storage/downloader.d.ts +103 -0
  1182. package/src/storage/downloader.js +1121 -0
  1183. package/src/storage/emulated-vram.d.ts +264 -0
  1184. package/src/storage/emulated-vram.js +576 -0
  1185. package/src/storage/export.d.ts +20 -0
  1186. package/src/storage/export.js +159 -0
  1187. package/src/storage/index.d.ts +256 -0
  1188. package/src/storage/index.js +188 -0
  1189. package/src/storage/inventory.d.ts +26 -0
  1190. package/src/storage/inventory.js +218 -0
  1191. package/src/storage/preflight.d.ts +144 -0
  1192. package/src/storage/preflight.js +316 -0
  1193. package/src/storage/quickstart-downloader.d.ts +157 -0
  1194. package/src/storage/quickstart-downloader.js +268 -0
  1195. package/src/storage/quota.d.ts +150 -0
  1196. package/src/storage/quota.js +305 -0
  1197. package/src/storage/registry.d.ts +28 -0
  1198. package/src/storage/registry.js +131 -0
  1199. package/src/storage/reports.d.ts +20 -0
  1200. package/src/storage/reports.js +94 -0
  1201. package/src/storage/shard-manager.d.ts +151 -0
  1202. package/src/storage/shard-manager.js +861 -0
  1203. package/src/storage/source-artifact-store.d.ts +52 -0
  1204. package/src/storage/source-artifact-store.js +234 -0
  1205. package/src/sw.d.ts +1 -0
  1206. package/src/sw.js +187 -0
  1207. package/src/tooling/bench-runner.js +64 -0
  1208. package/src/tooling/browser-command-runner.d.ts +28 -0
  1209. package/src/tooling/browser-command-runner.js +85 -0
  1210. package/src/tooling/command-api-constants.d.ts +9 -0
  1211. package/src/tooling/command-api-constants.js +9 -0
  1212. package/src/tooling/command-api-family-normalizers.d.ts +9 -0
  1213. package/src/tooling/command-api-family-normalizers.js +395 -0
  1214. package/src/tooling/command-api-helpers.d.ts +25 -0
  1215. package/src/tooling/command-api-helpers.js +265 -0
  1216. package/src/tooling/command-api.d.ts +177 -0
  1217. package/src/tooling/command-api.js +82 -0
  1218. package/src/tooling/command-envelope.d.ts +84 -0
  1219. package/src/tooling/command-envelope.js +201 -0
  1220. package/src/tooling/command-runner-shared.d.ts +75 -0
  1221. package/src/tooling/command-runner-shared.js +191 -0
  1222. package/src/tooling/command-runner.html +45 -0
  1223. package/src/tooling/command-validation.js +44 -0
  1224. package/src/tooling/conversion-config-materializer.d.ts +24 -0
  1225. package/src/tooling/conversion-config-materializer.js +96 -0
  1226. package/src/tooling/diagnose-runner.d.ts +28 -0
  1227. package/src/tooling/diagnose-runner.js +333 -0
  1228. package/src/tooling/distill-runner.js +56 -0
  1229. package/src/tooling/hf-registry-utils.d.ts +59 -0
  1230. package/src/tooling/hf-registry-utils.js +324 -0
  1231. package/src/tooling/kernel-path-builder/custom-runtime-facts.d.ts +18 -0
  1232. package/src/tooling/kernel-path-builder/custom-runtime-facts.js +72 -0
  1233. package/src/tooling/kernel-path-builder/index.d.ts +65 -0
  1234. package/src/tooling/kernel-path-builder/index.js +1348 -0
  1235. package/src/tooling/kernel-path-builder/runtime-overlay.d.ts +9 -0
  1236. package/src/tooling/kernel-path-builder/runtime-overlay.js +149 -0
  1237. package/src/tooling/lean-execution-contract-runner.d.ts +43 -0
  1238. package/src/tooling/lean-execution-contract-runner.js +158 -0
  1239. package/src/tooling/lean-execution-contract.d.ts +16 -0
  1240. package/src/tooling/lean-execution-contract.js +228 -0
  1241. package/src/tooling/lora-runner.js +37 -0
  1242. package/src/tooling/node-browser-command-runner.d.ts +34 -0
  1243. package/src/tooling/node-browser-command-runner.js +827 -0
  1244. package/src/tooling/node-command-runner.d.ts +36 -0
  1245. package/src/tooling/node-command-runner.js +187 -0
  1246. package/src/tooling/node-convert-worker-pool.d.ts +16 -0
  1247. package/src/tooling/node-convert-worker-pool.js +186 -0
  1248. package/src/tooling/node-convert-worker.d.ts +1 -0
  1249. package/src/tooling/node-convert-worker.js +60 -0
  1250. package/src/tooling/node-converter.d.ts +1 -0
  1251. package/src/tooling/node-converter.js +1330 -0
  1252. package/src/tooling/node-file-fetch.d.ts +1 -0
  1253. package/src/tooling/node-file-fetch.js +38 -0
  1254. package/src/tooling/node-source-runtime.d.ts +19 -0
  1255. package/src/tooling/node-source-runtime.js +612 -0
  1256. package/src/tooling/node-webgpu.d.ts +16 -0
  1257. package/src/tooling/node-webgpu.js +421 -0
  1258. package/src/tooling/opfs-cache.d.ts +12 -0
  1259. package/src/tooling/opfs-cache.js +260 -0
  1260. package/src/tooling/runtime-input-composition.d.ts +38 -0
  1261. package/src/tooling/runtime-input-composition.js +90 -0
  1262. package/src/tooling/source-runtime-bundle.d.ts +141 -0
  1263. package/src/tooling/source-runtime-bundle.js +712 -0
  1264. package/src/tooling/source-runtime-materializer.d.ts +6 -0
  1265. package/src/tooling/source-runtime-materializer.js +93 -0
  1266. package/src/tooling/verify-runner.js +63 -0
  1267. package/src/tooling-exports.browser.d.ts +7 -0
  1268. package/src/tooling-exports.browser.js +2 -0
  1269. package/src/tooling-exports.d.ts +22 -0
  1270. package/src/tooling-exports.js +7 -0
  1271. package/src/tooling-exports.shared.d.ts +106 -0
  1272. package/src/tooling-exports.shared.js +93 -0
  1273. package/src/training/README.md +153 -0
  1274. package/src/training/artifacts.d.ts +160 -0
  1275. package/src/training/artifacts.js +896 -0
  1276. package/src/training/attention-backward.d.ts +30 -0
  1277. package/src/training/attention-backward.js +232 -0
  1278. package/src/training/attention-forward.d.ts +22 -0
  1279. package/src/training/attention-forward.js +82 -0
  1280. package/src/training/autograd.d.ts +51 -0
  1281. package/src/training/autograd.js +408 -0
  1282. package/src/training/backward-kernels.js +39 -0
  1283. package/src/training/checkpoint-watch.d.ts +8 -0
  1284. package/src/training/checkpoint-watch.js +139 -0
  1285. package/src/training/checkpoint.d.ts +36 -0
  1286. package/src/training/checkpoint.js +281 -0
  1287. package/src/training/clip.d.ts +9 -0
  1288. package/src/training/clip.js +55 -0
  1289. package/src/training/dataloader.d.ts +8 -0
  1290. package/src/training/dataloader.js +44 -0
  1291. package/src/training/datasets/index.d.ts +12 -0
  1292. package/src/training/datasets/index.js +6 -0
  1293. package/src/training/datasets/jsonl.d.ts +11 -0
  1294. package/src/training/datasets/jsonl.js +50 -0
  1295. package/src/training/datasets/reploid.d.ts +3 -0
  1296. package/src/training/datasets/reploid.js +36 -0
  1297. package/src/training/datasets/text-pairs.d.ts +21 -0
  1298. package/src/training/datasets/text-pairs.js +42 -0
  1299. package/src/training/datasets/token-batch.d.ts +21 -0
  1300. package/src/training/datasets/token-batch.js +52 -0
  1301. package/src/training/datasets/translation-pairs.d.ts +34 -0
  1302. package/src/training/datasets/translation-pairs.js +49 -0
  1303. package/src/training/distillation/artifacts.d.ts +71 -0
  1304. package/src/training/distillation/artifacts.js +132 -0
  1305. package/src/training/distillation/checkpoint-watch.d.ts +10 -0
  1306. package/src/training/distillation/checkpoint-watch.js +58 -0
  1307. package/src/training/distillation/dataset.d.ts +59 -0
  1308. package/src/training/distillation/dataset.js +337 -0
  1309. package/src/training/distillation/eval.d.ts +34 -0
  1310. package/src/training/distillation/eval.js +313 -0
  1311. package/src/training/distillation/index.d.ts +29 -0
  1312. package/src/training/distillation/index.js +29 -0
  1313. package/src/training/distillation/runtime.d.ts +20 -0
  1314. package/src/training/distillation/runtime.js +121 -0
  1315. package/src/training/distillation/scoreboard.d.ts +6 -0
  1316. package/src/training/distillation/scoreboard.js +8 -0
  1317. package/src/training/distillation/stage-a.d.ts +45 -0
  1318. package/src/training/distillation/stage-a.js +338 -0
  1319. package/src/training/distillation/stage-b.d.ts +24 -0
  1320. package/src/training/distillation/stage-b.js +20 -0
  1321. package/src/training/distillation/student-fixture.d.ts +22 -0
  1322. package/src/training/distillation/student-fixture.js +846 -0
  1323. package/src/training/distillation/suite-data.d.ts +45 -0
  1324. package/src/training/distillation/suite-data.js +189 -0
  1325. package/src/training/export.d.ts +32 -0
  1326. package/src/training/export.js +112 -0
  1327. package/src/training/index.d.ts +62 -0
  1328. package/src/training/index.js +51 -0
  1329. package/src/training/lora-pipeline.d.ts +40 -0
  1330. package/src/training/lora-pipeline.js +793 -0
  1331. package/src/training/lora.d.ts +19 -0
  1332. package/src/training/lora.js +71 -0
  1333. package/src/training/loss-scaling.d.ts +21 -0
  1334. package/src/training/loss-scaling.js +80 -0
  1335. package/src/training/loss.d.ts +10 -0
  1336. package/src/training/loss.js +40 -0
  1337. package/src/training/objectives/base.d.ts +58 -0
  1338. package/src/training/objectives/base.js +38 -0
  1339. package/src/training/objectives/cross_entropy.d.ts +18 -0
  1340. package/src/training/objectives/cross_entropy.js +34 -0
  1341. package/src/training/objectives/distill_kd.d.ts +16 -0
  1342. package/src/training/objectives/distill_kd.js +369 -0
  1343. package/src/training/objectives/distill_triplet.d.ts +16 -0
  1344. package/src/training/objectives/distill_triplet.js +414 -0
  1345. package/src/training/objectives/index.d.ts +12 -0
  1346. package/src/training/objectives/index.js +6 -0
  1347. package/src/training/objectives/ul_stage1_joint.d.ts +16 -0
  1348. package/src/training/objectives/ul_stage1_joint.js +188 -0
  1349. package/src/training/objectives/ul_stage2_base.d.ts +16 -0
  1350. package/src/training/objectives/ul_stage2_base.js +218 -0
  1351. package/src/training/operator-artifacts.d.ts +62 -0
  1352. package/src/training/operator-artifacts.js +140 -0
  1353. package/src/training/operator-command.d.ts +5 -0
  1354. package/src/training/operator-command.js +455 -0
  1355. package/src/training/operator-eval.d.ts +48 -0
  1356. package/src/training/operator-eval.js +230 -0
  1357. package/src/training/operator-scoreboard.d.ts +5 -0
  1358. package/src/training/operator-scoreboard.js +44 -0
  1359. package/src/training/optimizer.d.ts +22 -0
  1360. package/src/training/optimizer.js +127 -0
  1361. package/src/training/runner.d.ts +248 -0
  1362. package/src/training/runner.js +1236 -0
  1363. package/src/training/suite.d.ts +299 -0
  1364. package/src/training/suite.js +2196 -0
  1365. package/src/training/tensor-factory.d.ts +9 -0
  1366. package/src/training/tensor-factory.js +13 -0
  1367. package/src/training/trainer.d.ts +89 -0
  1368. package/src/training/trainer.js +299 -0
  1369. package/src/training/ul_dataset.d.ts +47 -0
  1370. package/src/training/ul_dataset.js +151 -0
  1371. package/src/training/ul_schedule.d.ts +6 -0
  1372. package/src/training/ul_schedule.js +29 -0
  1373. package/src/training/workload-packs/distill-translategemma-medium.json +103 -0
  1374. package/src/training/workload-packs/distill-translategemma-tiny.json +103 -0
  1375. package/src/training/workload-packs/lora-toy-medium.json +91 -0
  1376. package/src/training/workload-packs/lora-toy-tiny.json +91 -0
  1377. package/src/training/workload-packs/registry.json +98 -0
  1378. package/src/training/workload-packs/ul-training-medium.json +55 -0
  1379. package/src/training/workload-packs/ul-training-tiny.json +55 -0
  1380. package/src/training/workloads.d.ts +164 -0
  1381. package/src/training/workloads.js +530 -0
  1382. package/src/types/chrome.d.ts +36 -0
  1383. package/src/types/chrome.js +1 -0
  1384. package/src/types/gpu.d.ts +185 -0
  1385. package/src/types/gpu.js +5 -0
  1386. package/src/types/index.d.ts +4 -0
  1387. package/src/types/index.js +4 -0
  1388. package/src/types/inference.d.ts +197 -0
  1389. package/src/types/inference.js +5 -0
  1390. package/src/types/model.d.ts +130 -0
  1391. package/src/types/model.js +5 -0
  1392. package/src/utils/format-bytes.js +20 -0
  1393. package/src/utils/hf-resolve-url.d.ts +16 -0
  1394. package/src/utils/hf-resolve-url.js +17 -0
  1395. package/src/utils/index.d.ts +7 -0
  1396. package/src/utils/index.js +7 -0
  1397. package/src/utils/load-json.d.ts +5 -0
  1398. package/src/utils/load-json.js +33 -0
  1399. package/src/utils/plain-object.d.ts +1 -0
  1400. package/src/utils/plain-object.js +3 -0
  1401. package/src/utils/sha256.d.ts +4 -0
  1402. package/src/utils/sha256.js +135 -0
  1403. package/src/version.d.ts +2 -0
  1404. package/src/version.js +2 -0
  1405. package/tools/convert-safetensors-node.js +240 -0
@@ -0,0 +1,1530 @@
1
+
2
+
3
+ import {
4
+ // Constants
5
+ SHARD_SIZE as SCHEMA_SHARD_SIZE,
6
+ RDRR_VERSION as SCHEMA_RDRR_VERSION,
7
+ ConversionStage as SchemaConversionStage,
8
+ DEFAULT_MANIFEST_INFERENCE,
9
+ formatBytes,
10
+ } from '../config/schema/index.js';
11
+
12
+ import {
13
+ classifyTensorRole,
14
+ generateShardFilename,
15
+ resolveTensorGroup,
16
+ resolveTensorRole,
17
+ } from '../formats/rdrr/index.js';
18
+ import { log } from '../debug/index.js';
19
+ import {
20
+ getInferenceLayerPatternContractArtifact,
21
+ selectRuleValue,
22
+ } from '../rules/rule-registry.js';
23
+ import {
24
+ createConverterConfig,
25
+ } from '../config/index.js';
26
+ import { buildExecutionContractArtifact } from '../config/execution-contract-check.js';
27
+ import { buildManifestRequiredInferenceFieldsArtifact } from '../config/required-inference-fields-contract-check.js';
28
+ import { resolveEosTokenId } from './tokenizer-utils.js';
29
+ import {
30
+ normalizeQ4KLayout,
31
+ resolveManifestQuantization,
32
+ resolveEffectiveQuantizationInfo,
33
+ } from './quantization-info.js';
34
+ import {
35
+ float16ToFloat32,
36
+ float32ToFloat16,
37
+ quantizeToQ4KM,
38
+ quantizeToQ4KMRowWise,
39
+ quantizeToQ4KMColumnWise,
40
+ } from './quantizer.js';
41
+
42
+ // ============================================================================
43
+ // Re-exports for Backward Compatibility
44
+ // ============================================================================
45
+
46
+
47
+ export const ConvertStage = SchemaConversionStage;
48
+
49
+ // Re-export constants
50
+ export const SHARD_SIZE = SCHEMA_SHARD_SIZE;
51
+ export const RDRR_VERSION = SCHEMA_RDRR_VERSION;
52
+
53
+ // ============================================================================
54
+ // Embedding Output Inference
55
+ // ============================================================================
56
+
57
+ const EMBEDDING_TENSOR_NAMES = [
58
+ 'language_model.model.embed_tokens.weight',
59
+ 'model.embed_tokens.weight',
60
+ 'embed_tokens.weight',
61
+ 'token_embd.weight',
62
+ 'wte.weight',
63
+ 'transformer.wte.weight',
64
+ ];
65
+
66
+ export function inferEmbeddingOutputConfig(tensorLocations) {
67
+ // Normalize Map input to a plain object so the rest of the function
68
+ // handles a single type consistently.
69
+ const normalized = tensorLocations instanceof Map
70
+ ? Object.fromEntries(tensorLocations)
71
+ : tensorLocations ?? {};
72
+
73
+ const getLocation = (name) => normalized[name];
74
+
75
+ const entries = Object.entries(normalized);
76
+ for (const [_name, loc] of entries) {
77
+ if (loc?.role === 'embedding' && loc.shape?.length === 2) {
78
+ const [dim0, dim1] = loc.shape;
79
+ const isGGUFLayout = dim0 < dim1;
80
+ return {
81
+ embeddingTranspose: isGGUFLayout,
82
+ embeddingVocabSize: isGGUFLayout ? dim1 : dim0,
83
+ };
84
+ }
85
+ }
86
+
87
+ for (const name of EMBEDDING_TENSOR_NAMES) {
88
+ const loc = getLocation(name);
89
+ if (loc?.shape && loc.shape.length === 2) {
90
+ const [dim0, dim1] = loc.shape;
91
+ const isGGUFLayout = dim0 < dim1;
92
+ return {
93
+ embeddingTranspose: isGGUFLayout,
94
+ embeddingVocabSize: isGGUFLayout ? dim1 : dim0,
95
+ };
96
+ }
97
+ }
98
+
99
+ return null;
100
+ }
101
+
102
+ // ============================================================================
103
+ // Pure Functions (no I/O, no platform dependencies)
104
+ // ============================================================================
105
+
106
+ function resolveTokenizerId(value) {
107
+ if (typeof value === 'number') return value;
108
+ return null;
109
+ }
110
+
111
+ function resolveTokenizerIds(value) {
112
+ if (Array.isArray(value) && value.every((id) => typeof id === 'number')) {
113
+ return value;
114
+ }
115
+ if (typeof value === 'number') return [value];
116
+ return null;
117
+ }
118
+
119
+ function resolveTokenizerField(tokenizerConfig, ...keys) {
120
+ if (!tokenizerConfig) return null;
121
+ for (const key of keys) {
122
+ if (tokenizerConfig[key] != null) {
123
+ return tokenizerConfig[key];
124
+ }
125
+ }
126
+ return null;
127
+ }
128
+
129
+ function resolveConfigBoolean(rawConfig, ...keys) {
130
+ // Same lookup logic as resolveTokenizerField: return the first non-nullish
131
+ // value from the given keys. Delegates to avoid duplicating the pattern.
132
+ return resolveTokenizerField(rawConfig, ...keys);
133
+ }
134
+
135
+ function resolveTokenizerVocabSize(tokenizerConfig, rawConfig, architecture) {
136
+ const nestedTextConfig = getNestedTextConfig(rawConfig);
137
+ const configVocab = rawConfig?.vocab_size ?? nestedTextConfig?.vocab_size;
138
+ const tokenizerVocab = tokenizerConfig?.vocab_size ?? tokenizerConfig?.vocabSize;
139
+ const archVocab = architecture?.vocabSize;
140
+
141
+ // Warn if multiple sources provide vocab size and they disagree
142
+ const sources = [
143
+ tokenizerVocab != null ? { label: 'tokenizer', value: tokenizerVocab } : null,
144
+ configVocab != null ? { label: 'config', value: configVocab } : null,
145
+ archVocab != null ? { label: 'architecture', value: archVocab } : null,
146
+ ].filter(Boolean);
147
+ if (sources.length > 1) {
148
+ const distinct = new Set(sources.map((s) => s.value));
149
+ if (distinct.size > 1) {
150
+ const detail = sources.map((s) => `${s.label}=${s.value}`).join(', ');
151
+ log.error(
152
+ 'Convert',
153
+ `Vocab size sources disagree: ${detail}. Using first available (${sources[0].label}=${sources[0].value}). ` +
154
+ 'This may cause embedding size mismatches at runtime. Verify the correct vocab size in the conversion config.'
155
+ );
156
+ }
157
+ }
158
+
159
+ return tokenizerVocab ?? configVocab ?? archVocab ?? null;
160
+ }
161
+
162
+ export function normalizeStorageQuant(value) {
163
+ if (value == null) return null;
164
+ const lower = String(value).trim().toLowerCase();
165
+ if (!lower) return null;
166
+ if (lower === 'fp16' || lower === 'float16') return 'f16';
167
+ if (lower === 'fp32' || lower === 'float32') return 'f32';
168
+ if (lower === 'bfloat16') return 'bf16';
169
+ if (lower === 'q4_k_m' || lower === 'q4km') return 'q4k';
170
+ return lower;
171
+ }
172
+
173
+ export function resolveTensorTargetQuant(tensorOrName, fallbackQuant, quantizationInfo) {
174
+ const fallback = normalizeStorageQuant(fallbackQuant);
175
+ if (!quantizationInfo || typeof quantizationInfo !== 'object') {
176
+ return fallback;
177
+ }
178
+
179
+ const role = resolveTensorRole(tensorOrName);
180
+ if (role === 'embedding') {
181
+ return normalizeStorageQuant(quantizationInfo.embeddings ?? fallback) ?? fallback;
182
+ }
183
+ if (role === 'lm_head') {
184
+ const headQuant = quantizationInfo.lmHead ?? quantizationInfo.embeddings ?? fallback;
185
+ return normalizeStorageQuant(headQuant) ?? fallback;
186
+ }
187
+ if (role === 'vision') {
188
+ return normalizeStorageQuant(quantizationInfo.vision ?? fallback) ?? fallback;
189
+ }
190
+ if (role === 'projector') {
191
+ return normalizeStorageQuant(quantizationInfo.projector ?? fallback) ?? fallback;
192
+ }
193
+ if (role === 'audio') {
194
+ return normalizeStorageQuant(quantizationInfo.audio ?? fallback) ?? fallback;
195
+ }
196
+ return normalizeStorageQuant(quantizationInfo.weights ?? fallback) ?? fallback;
197
+ }
198
+
199
+ function bf16ToFloat32(value) {
200
+ const view = new DataView(new ArrayBuffer(4));
201
+ view.setUint32(0, (value & 0xffff) << 16, true);
202
+ return view.getFloat32(0, true);
203
+ }
204
+
205
+ function normalizeTensorName(tensor) {
206
+ const name = tensor?.name;
207
+ return typeof name === 'string' ? name : '';
208
+ }
209
+
210
+ function shouldExcludeTextOnlyTensor(name) {
211
+ const lower = name.toLowerCase();
212
+ return lower.startsWith('vision_tower.')
213
+ || lower.startsWith('vision_model.')
214
+ || lower.startsWith('visual.')
215
+ || lower.startsWith('model.visual.')
216
+ || lower.startsWith('vision.')
217
+ || lower.startsWith('model.vision.')
218
+ || lower.startsWith('vision_encoder.')
219
+ || lower.startsWith('image_encoder.')
220
+ || lower.startsWith('image_tower.')
221
+ || lower.startsWith('image.')
222
+ || lower.startsWith('model.image.')
223
+ || lower.startsWith('audio_tower.')
224
+ || lower.startsWith('audio_model.')
225
+ || lower.startsWith('audio.')
226
+ || lower.startsWith('model.audio.')
227
+ || lower.startsWith('audio_encoder.')
228
+ || lower.startsWith('multi_modal_projector.')
229
+ || lower.startsWith('model.multi_modal_projector.')
230
+ || lower.startsWith('mm_projector.')
231
+ || lower.startsWith('model.mm_projector.');
232
+ }
233
+
234
+ function resolveConversionTensors(model, converterConfig) {
235
+ const source = Array.isArray(model?.tensors) ? model.tensors : [];
236
+ if (source.length === 0) {
237
+ return source;
238
+ }
239
+ const textOnly = converterConfig?.output?.textOnly === true;
240
+ if (!textOnly) {
241
+ return source;
242
+ }
243
+
244
+ const hasLanguageModelNamespace = source.some((tensor) => {
245
+ const lower = normalizeTensorName(tensor).toLowerCase();
246
+ return lower.startsWith('language_model.') || lower.startsWith('model.language_model.');
247
+ });
248
+ if (hasLanguageModelNamespace) {
249
+ return source.filter((tensor) => {
250
+ const lower = normalizeTensorName(tensor).toLowerCase();
251
+ return lower.startsWith('language_model.') || lower.startsWith('model.language_model.');
252
+ });
253
+ }
254
+
255
+ return source.filter((tensor) => (
256
+ !shouldExcludeTextOnlyTensor(normalizeTensorName(tensor))
257
+ ));
258
+ }
259
+
260
+ function toFloat32ForQ4K(tensorData, sourceDtype, tensorName) {
261
+ const dtype = String(sourceDtype || '').toUpperCase();
262
+ if (dtype === 'F32') {
263
+ if (tensorData.byteLength % 4 !== 0) {
264
+ throw new Error(`Invalid F32 tensor byte length for ${tensorName}: ${tensorData.byteLength}`);
265
+ }
266
+ return new Float32Array(
267
+ tensorData.buffer,
268
+ tensorData.byteOffset,
269
+ tensorData.byteLength / 4
270
+ );
271
+ }
272
+ if (dtype === 'F16') {
273
+ if (tensorData.byteLength % 2 !== 0) {
274
+ throw new Error(`Invalid F16 tensor byte length for ${tensorName}: ${tensorData.byteLength}`);
275
+ }
276
+ const f16 = new Uint16Array(
277
+ tensorData.buffer,
278
+ tensorData.byteOffset,
279
+ tensorData.byteLength / 2
280
+ );
281
+ const f32 = new Float32Array(f16.length);
282
+ for (let i = 0; i < f16.length; i++) {
283
+ f32[i] = float16ToFloat32(f16[i]);
284
+ }
285
+ return f32;
286
+ }
287
+ if (dtype === 'BF16') {
288
+ if (tensorData.byteLength % 2 !== 0) {
289
+ throw new Error(`Invalid BF16 tensor byte length for ${tensorName}: ${tensorData.byteLength}`);
290
+ }
291
+ const bf16 = new Uint16Array(
292
+ tensorData.buffer,
293
+ tensorData.byteOffset,
294
+ tensorData.byteLength / 2
295
+ );
296
+ const f32 = new Float32Array(bf16.length);
297
+ for (let i = 0; i < bf16.length; i++) {
298
+ f32[i] = bf16ToFloat32(bf16[i]);
299
+ }
300
+ return f32;
301
+ }
302
+ throw new Error(`Cannot quantize ${tensorName} from ${dtype} to Q4_K_M`);
303
+ }
304
+
305
+ function resolveConfigTokenId(rawConfig, key) {
306
+ const direct = rawConfig?.[key];
307
+ const nested = getNestedTextConfig(rawConfig)?.[key];
308
+ return resolveTokenizerId(direct ?? nested);
309
+ }
310
+
311
+ function resolveConfigTokenIds(rawConfig, key) {
312
+ const direct = rawConfig?.[key];
313
+ const nested = getNestedTextConfig(rawConfig)?.[key];
314
+ return resolveTokenizerIds(direct ?? nested);
315
+ }
316
+
317
+ function resolveMoEConfigNumber(rawConfig, ...keys) {
318
+ const nestedTextConfig = getNestedTextConfig(rawConfig);
319
+ for (const key of keys) {
320
+ const direct = rawConfig?.[key];
321
+ if (Number.isFinite(direct) && direct > 0) return Number(direct);
322
+ const nested = nestedTextConfig?.[key];
323
+ if (Number.isFinite(nested) && nested > 0) return Number(nested);
324
+ }
325
+ return null;
326
+ }
327
+
328
+ function normalizeTensorShape(value) {
329
+ if (!Array.isArray(value) || value.length !== 2) return null;
330
+ const rows = Number(value[0]);
331
+ const cols = Number(value[1]);
332
+ if (!Number.isFinite(rows) || !Number.isFinite(cols)) return null;
333
+ if (rows <= 0 || cols <= 0) return null;
334
+ return [Math.trunc(rows), Math.trunc(cols)];
335
+ }
336
+
337
+ function isExpertTensorName(name) {
338
+ const lower = String(name || '').toLowerCase();
339
+ return lower.includes('.experts.') || lower.includes('.expert.') || lower.includes('block_sparse_moe');
340
+ }
341
+
342
+ function inferDenseIntermediateSizeFromTensorEntries(entries) {
343
+ if (!Array.isArray(entries) || entries.length === 0) return null;
344
+ const candidates = [];
345
+ for (const entry of entries) {
346
+ const name = String(entry?.name || '');
347
+ if (!name || isExpertTensorName(name)) continue;
348
+ const shape = normalizeTensorShape(entry?.shape);
349
+ if (!shape) continue;
350
+ const lower = name.toLowerCase();
351
+ if (
352
+ lower.endsWith('.feed_forward.w1.weight')
353
+ || lower.endsWith('.feed_forward.w3.weight')
354
+ || lower.endsWith('.ffn_gate.weight')
355
+ || lower.endsWith('.ffn_up.weight')
356
+ || lower.endsWith('.ffn.gate_proj.weight')
357
+ || lower.endsWith('.ffn.up_proj.weight')
358
+ || lower.endsWith('.mlp.gate_proj.weight')
359
+ || lower.endsWith('.mlp.up_proj.weight')
360
+ ) {
361
+ candidates.push(shape[0]);
362
+ continue;
363
+ }
364
+ if (
365
+ lower.endsWith('.feed_forward.w2.weight')
366
+ || lower.endsWith('.ffn_down.weight')
367
+ || lower.endsWith('.ffn.down_proj.weight')
368
+ || lower.endsWith('.mlp.down_proj.weight')
369
+ ) {
370
+ candidates.push(shape[1]);
371
+ continue;
372
+ }
373
+ if (
374
+ lower.endsWith('.feed_forward.w1_w3.weight')
375
+ || lower.endsWith('.ffn_gate_up.weight')
376
+ || lower.endsWith('.ffn.gate_up_proj.weight')
377
+ || lower.endsWith('.mlp.gate_up_proj.weight')
378
+ ) {
379
+ if (shape[0] % 2 === 0) {
380
+ candidates.push(Math.trunc(shape[0] / 2));
381
+ }
382
+ }
383
+ }
384
+ if (candidates.length === 0) return null;
385
+ const counts = new Map();
386
+ for (const value of candidates) {
387
+ counts.set(value, (counts.get(value) ?? 0) + 1);
388
+ }
389
+ return [...counts.entries()]
390
+ .sort((a, b) => {
391
+ if (b[1] !== a[1]) return b[1] - a[1];
392
+ return a[0] - b[0];
393
+ })[0]?.[0] ?? null;
394
+ }
395
+
396
+ function resolveIntermediateSizeFromTensors(architecture, model, tensorLocations, rawConfig, modelId) {
397
+ if (!architecture || typeof architecture !== 'object') return architecture;
398
+ const current = architecture.intermediateSize;
399
+ if (typeof current !== 'number' || !Number.isFinite(current) || current <= 0) {
400
+ return architecture;
401
+ }
402
+ const modelType = String(rawConfig?.model_type ?? getNestedTextConfig(rawConfig)?.model_type ?? '').toLowerCase();
403
+ if (modelType !== 'lfm2') {
404
+ return architecture;
405
+ }
406
+ const entries = Array.isArray(model?.tensors) && model.tensors.length > 0
407
+ ? model.tensors
408
+ : Object.entries(tensorLocations ?? {}).map(([name, location]) => ({ name, shape: location?.shape }));
409
+ const inferred = inferDenseIntermediateSizeFromTensorEntries(entries);
410
+ if (inferred == null || inferred === current) {
411
+ return architecture;
412
+ }
413
+ log.warn(
414
+ 'Convert',
415
+ `Adjusted architecture.intermediateSize for "${modelId}": ${current} -> ${inferred} (from FFN tensor shapes)`
416
+ );
417
+ return {
418
+ ...architecture,
419
+ intermediateSize: inferred,
420
+ };
421
+ }
422
+
423
+ function modelHasMoETensors(model) {
424
+ if (!Array.isArray(model?.tensors)) return false;
425
+ return model.tensors.some((tensor) => {
426
+ const name = String(tensor?.name || '').toLowerCase();
427
+ return (
428
+ name.includes('.experts.') ||
429
+ name.includes('.expert.') ||
430
+ name.includes('block_sparse_moe')
431
+ );
432
+ });
433
+ }
434
+
435
+ function resolveMoEExpertFormat(rawConfig, resolvedModelType, quantizationInfo, explicitFormat) {
436
+ if (explicitFormat) return explicitFormat;
437
+ const fromQuant = quantizationInfo?.expertsFormat;
438
+ if (typeof fromQuant === 'string' && fromQuant.length > 0) {
439
+ return fromQuant;
440
+ }
441
+ const modelType = String(
442
+ resolvedModelType ??
443
+ rawConfig?.model_type ??
444
+ getNestedTextConfig(rawConfig)?.model_type ??
445
+ ''
446
+ ).toLowerCase();
447
+ if (modelType.includes('gpt_oss') || modelType.includes('gpt-oss') || modelType.includes('gptoss')) {
448
+ return 'gpt-oss';
449
+ }
450
+ return 'mixtral';
451
+ }
452
+
453
+ function normalizeMoEConfig(config, contextLabel) {
454
+ if (!config) return null;
455
+ const numExperts = Number(config.numExperts);
456
+ const numExpertsPerToken = Number(config.numExpertsPerToken);
457
+ const expertFormat = String(config.expertFormat || '').trim();
458
+ const allowedExpertFormat = expertFormat === 'gpt-oss' || expertFormat === 'mixtral';
459
+ if (!Number.isFinite(numExperts) || numExperts <= 0) {
460
+ throw new Error(`Invalid moeConfig.numExperts for ${contextLabel}`);
461
+ }
462
+ if (!Number.isFinite(numExpertsPerToken) || numExpertsPerToken <= 0) {
463
+ throw new Error(`Invalid moeConfig.numExpertsPerToken for ${contextLabel}`);
464
+ }
465
+ if (numExpertsPerToken > numExperts) {
466
+ throw new Error(`Invalid moeConfig for ${contextLabel}: numExpertsPerToken cannot exceed numExperts`);
467
+ }
468
+ if (!allowedExpertFormat) {
469
+ throw new Error(`Invalid moeConfig.expertFormat for ${contextLabel}: "${expertFormat}"`);
470
+ }
471
+ return {
472
+ numExperts,
473
+ numExpertsPerToken,
474
+ expertFormat,
475
+ };
476
+ }
477
+
478
+ function resolveManifestMoEConfig(model, options, rawConfig, resolvedModelType) {
479
+ const explicit = normalizeMoEConfig(options?.moeConfig ?? null, options?.modelId ?? 'model');
480
+ if (explicit) return explicit;
481
+
482
+ const hasMoETensors = modelHasMoETensors(model);
483
+ const numExperts = resolveMoEConfigNumber(rawConfig, 'num_local_experts', 'num_experts', 'expertCount');
484
+
485
+ // If the checkpoint does not expose MoE tensors and config does not declare experts,
486
+ // this is a dense model and should not emit moeConfig.
487
+ if (!hasMoETensors && (!numExperts || numExperts <= 1)) {
488
+ return null;
489
+ }
490
+
491
+ if (!numExperts || numExperts <= 0) {
492
+ throw new Error(
493
+ `MoE tensors detected for "${options?.modelId ?? 'model'}" but expert count is missing in config`
494
+ );
495
+ }
496
+
497
+ const numExpertsPerToken = resolveMoEConfigNumber(
498
+ rawConfig,
499
+ 'num_experts_per_tok',
500
+ 'num_experts_per_token',
501
+ 'experts_per_token',
502
+ 'expertUsedCount'
503
+ );
504
+
505
+ if (!numExpertsPerToken) {
506
+ throw new Error(
507
+ `MoE model "${options?.modelId ?? 'model'}" missing experts-per-token config ` +
508
+ '(expected num_experts_per_tok/num_experts_per_token/experts_per_token)'
509
+ );
510
+ }
511
+
512
+ const expertFormat = resolveMoEExpertFormat(
513
+ rawConfig,
514
+ resolvedModelType,
515
+ options?.quantizationInfo ?? null,
516
+ null
517
+ );
518
+
519
+ return normalizeMoEConfig(
520
+ { numExperts, numExpertsPerToken, expertFormat },
521
+ options?.modelId ?? 'model'
522
+ );
523
+ }
524
+
525
+ function buildSentencepieceTokenizer(tokenizerConfig, rawConfig, architecture, modelTokenizerModel) {
526
+ if (!modelTokenizerModel) return null;
527
+
528
+ const vocabSize = resolveTokenizerVocabSize(tokenizerConfig, rawConfig, architecture);
529
+ const sentencepieceModel = typeof modelTokenizerModel === 'string'
530
+ ? modelTokenizerModel
531
+ : modelTokenizerModel?.file ?? 'tokenizer.model';
532
+
533
+ const bosTokenId = resolveTokenizerId(
534
+ resolveTokenizerField(tokenizerConfig, 'bos_token_id', 'bosTokenId')
535
+ ?? resolveConfigTokenId(rawConfig, 'bos_token_id')
536
+ );
537
+ const eosTokenId = resolveTokenizerId(
538
+ resolveTokenizerField(tokenizerConfig, 'eos_token_id', 'eosTokenId')
539
+ ?? resolveConfigTokenId(rawConfig, 'eos_token_id')
540
+ );
541
+ const eosTokens = resolveTokenizerIds(
542
+ resolveTokenizerField(tokenizerConfig, 'eos_token_ids', 'eosTokens', 'eos_token_id')
543
+ ?? resolveConfigTokenIds(rawConfig, 'eos_token_ids')
544
+ );
545
+ const padTokenId = resolveTokenizerId(
546
+ resolveTokenizerField(tokenizerConfig, 'pad_token_id', 'padTokenId')
547
+ ?? resolveConfigTokenId(rawConfig, 'pad_token_id')
548
+ );
549
+ const unkTokenId = resolveTokenizerId(
550
+ resolveTokenizerField(tokenizerConfig, 'unk_token_id', 'unkTokenId')
551
+ ?? resolveConfigTokenId(rawConfig, 'unk_token_id')
552
+ );
553
+ const addBosToken = resolveTokenizerField(tokenizerConfig, 'add_bos_token', 'addBosToken');
554
+ const addEosToken = resolveTokenizerField(tokenizerConfig, 'add_eos_token', 'addEosToken');
555
+
556
+ const tokenizer = {
557
+ type: 'sentencepiece',
558
+ sentencepieceModel,
559
+ vocabSize: vocabSize ?? 0,
560
+ };
561
+
562
+ if (bosTokenId != null) tokenizer.bosTokenId = bosTokenId;
563
+ if (eosTokenId != null) tokenizer.eosTokenId = eosTokenId;
564
+ if (eosTokens) tokenizer.eosTokens = eosTokens;
565
+ if (padTokenId != null) tokenizer.padTokenId = padTokenId;
566
+ if (unkTokenId != null) tokenizer.unkTokenId = unkTokenId;
567
+ if (addBosToken != null) tokenizer.addBosToken = addBosToken;
568
+ if (addEosToken != null) tokenizer.addEosToken = addEosToken;
569
+
570
+ return tokenizer;
571
+ }
572
+
573
+ export function resolveBundledTokenizerVocabSize(tokenizerJson) {
574
+ const vocab = tokenizerJson?.model?.vocab;
575
+ if (Array.isArray(vocab)) {
576
+ return vocab.length;
577
+ }
578
+ if (vocab && typeof vocab === 'object') {
579
+ return Object.keys(vocab).length;
580
+ }
581
+ return 0;
582
+ }
583
+
584
+ function buildBundledTokenizer(tokenizerJson, tokenizerConfig, rawConfig) {
585
+ const vocabSize = resolveBundledTokenizerVocabSize(tokenizerJson);
586
+ if (!vocabSize) {
587
+ throw new Error('Tokenizer vocab is missing or empty');
588
+ }
589
+
590
+ const tokenizer = {
591
+ type: 'bundled',
592
+ vocabSize,
593
+ file: 'tokenizer.json',
594
+ };
595
+ const addBosToken = (
596
+ resolveTokenizerField(tokenizerJson, 'add_bos_token', 'addBosToken')
597
+ ?? resolveTokenizerField(tokenizerConfig, 'add_bos_token', 'addBosToken')
598
+ ?? resolveConfigBoolean(rawConfig, 'add_bos_token', 'addBosToken')
599
+ );
600
+ const addEosToken = (
601
+ resolveTokenizerField(tokenizerJson, 'add_eos_token', 'addEosToken')
602
+ ?? resolveTokenizerField(tokenizerConfig, 'add_eos_token', 'addEosToken')
603
+ ?? resolveConfigBoolean(rawConfig, 'add_eos_token', 'addEosToken')
604
+ );
605
+
606
+ if (addBosToken != null) tokenizer.addBosToken = addBosToken;
607
+ if (addEosToken != null) tokenizer.addEosToken = addEosToken;
608
+
609
+ return tokenizer;
610
+ }
611
+
612
+
613
+ export function sanitizeModelId(name) {
614
+ const sanitized = name
615
+ .toLowerCase()
616
+ .replace(/[^a-z0-9_-]/g, '-')
617
+ .replace(/-+/g, '-')
618
+ .replace(/^-|-$/g, '')
619
+ .slice(0, 64);
620
+ return sanitized || null;
621
+ }
622
+
623
+
624
+ // Re-export formatBytes from schema for backward compatibility
625
+ export { formatBytes };
626
+
627
+ const BF16_ROUND_VIEW = new DataView(new ArrayBuffer(4));
628
+
629
+ function float32ToBFloat16(value) {
630
+ BF16_ROUND_VIEW.setFloat32(0, value, true);
631
+ const bits = BF16_ROUND_VIEW.getUint32(0, true);
632
+ const lsb = (bits >> 16) & 1;
633
+ const roundingBias = 0x7fff + lsb;
634
+ return ((bits + roundingBias) >> 16) & 0xffff;
635
+ }
636
+
637
+ function resolveQuantizeEmbeddings(quantizationInfo, explicitValue = null) {
638
+ if (typeof explicitValue === 'boolean') {
639
+ return explicitValue;
640
+ }
641
+ return (
642
+ normalizeStorageQuant(quantizationInfo?.embeddings ?? null) === 'q4k'
643
+ || normalizeStorageQuant(quantizationInfo?.lmHead ?? null) === 'q4k'
644
+ );
645
+ }
646
+
647
+ function normalizeModulesToNotConvert(modulesToNotConvert) {
648
+ if (!Array.isArray(modulesToNotConvert)) {
649
+ return null;
650
+ }
651
+ const normalized = modulesToNotConvert
652
+ .map((value) => (
653
+ typeof value === 'string' ? value.trim() : ''
654
+ ))
655
+ .filter(Boolean);
656
+ return normalized.length > 0 ? normalized : null;
657
+ }
658
+
659
+ function shouldSkipModuleQuantization(tensorName, modulesToNotConvert) {
660
+ const patterns = normalizeModulesToNotConvert(modulesToNotConvert);
661
+ if (!patterns) {
662
+ return false;
663
+ }
664
+
665
+ for (const pattern of patterns) {
666
+ const regexPattern = pattern
667
+ .replace(/\./g, '\\.')
668
+ .replace(/\*/g, '\\d+');
669
+ const matcher = new RegExp(regexPattern);
670
+ if (matcher.test(tensorName)) {
671
+ return true;
672
+ }
673
+ }
674
+ return false;
675
+ }
676
+
677
+
678
+ export function shouldQuantize(tensorName, shape, options = {}) {
679
+ const {
680
+ quantizeEmbeddings = false,
681
+ modulesToNotConvert = null,
682
+ role: explicitRole = null,
683
+ } = options;
684
+
685
+ if (!shape || !Array.isArray(shape) || shape.length === 0) {
686
+ log.warn('Convert', `Invalid shape for tensor "${tensorName}": ${JSON.stringify(shape)}`);
687
+ return false;
688
+ }
689
+ const numElements = shape.reduce((a, b) => a * b, 1);
690
+ const role = typeof explicitRole === 'string' && explicitRole.trim()
691
+ ? explicitRole.trim()
692
+ : classifyTensorRole(tensorName);
693
+ const lower = tensorName.toLowerCase();
694
+ const isBias = lower.endsWith('.bias') || lower.endsWith('_bias');
695
+
696
+ const shouldQuantizeByRole = selectRuleValue('converter', 'tensorRoles', 'shouldQuantize', {
697
+ numElements,
698
+ role,
699
+ isBias,
700
+ quantizeEmbeddings,
701
+ });
702
+
703
+ if (!shouldQuantizeByRole) {
704
+ return false;
705
+ }
706
+
707
+ if (shouldSkipModuleQuantization(tensorName, modulesToNotConvert)) {
708
+ return false;
709
+ }
710
+
711
+ return true;
712
+ }
713
+
714
+ export function transformTensorBytes(tensor, rawData, options = {}) {
715
+ const tensorDataInput = rawData instanceof Uint8Array ? rawData : new Uint8Array(rawData);
716
+ let tensorData = tensorDataInput;
717
+ let outDtype = tensor.dtype;
718
+ let outLayout = null;
719
+
720
+ const sourceDtype = String(tensor.dtype).toUpperCase();
721
+ const targetQuant = normalizeStorageQuant(options.targetQuant ?? options.quantization ?? null);
722
+ const quantizationInfo = options.quantizationInfo ?? null;
723
+ const tensorTargetQuant = resolveTensorTargetQuant(
724
+ tensor,
725
+ targetQuant,
726
+ quantizationInfo
727
+ );
728
+ const q4kLayout = normalizeQ4KLayout(options.q4kLayout ?? quantizationInfo?.layout);
729
+ const quantizeEmbeddings = resolveQuantizeEmbeddings(
730
+ quantizationInfo,
731
+ options.quantizeEmbeddings
732
+ );
733
+ const modulesToNotConvert = normalizeModulesToNotConvert(
734
+ options.modulesToNotConvert ?? null
735
+ );
736
+ const forceQuantizeDecision = (
737
+ typeof options.forceQuantizeDecision === 'boolean'
738
+ ? options.forceQuantizeDecision
739
+ : null
740
+ );
741
+
742
+ if (tensorTargetQuant === 'q4k') {
743
+ const sourceQuant = normalizeStorageQuant(sourceDtype);
744
+ if (sourceQuant === 'q4k') {
745
+ outDtype = 'Q4_K_M';
746
+ if (Array.isArray(tensor.shape) && tensor.shape.length === 2) {
747
+ outLayout = q4kLayout;
748
+ }
749
+ return {
750
+ tensorData,
751
+ outDtype,
752
+ outLayout,
753
+ sourceDtype,
754
+ tensorTargetQuant,
755
+ };
756
+ }
757
+
758
+ const shouldQuantizeTensor = (
759
+ forceQuantizeDecision ?? shouldQuantize(tensor.name, tensor.shape, {
760
+ quantizeEmbeddings,
761
+ modulesToNotConvert,
762
+ role: tensor.role ?? null,
763
+ })
764
+ );
765
+ if (shouldQuantizeTensor) {
766
+ const f32Data = toFloat32ForQ4K(tensorData, sourceDtype, tensor.name);
767
+ const quantized = (
768
+ Array.isArray(tensor.shape) && tensor.shape.length === 2
769
+ ? (q4kLayout === 'col'
770
+ ? quantizeToQ4KMColumnWise(f32Data, tensor.shape)
771
+ : quantizeToQ4KMRowWise(f32Data, tensor.shape))
772
+ : quantizeToQ4KM(f32Data, tensor.shape)
773
+ );
774
+ tensorData = quantized.quantized;
775
+ outDtype = 'Q4_K_M';
776
+ if (Array.isArray(tensor.shape) && tensor.shape.length === 2) {
777
+ outLayout = q4kLayout;
778
+ }
779
+ } else if (sourceDtype === 'BF16') {
780
+ // BF16 is not a native WebGPU dtype. When quantization is skipped
781
+ // (e.g. via modulesToNotConvert), convert BF16→F16 so the runtime
782
+ // can load the tensor without a BF16 dequant shader.
783
+ const bf16 = new Uint16Array(
784
+ tensorData.buffer,
785
+ tensorData.byteOffset,
786
+ tensorData.byteLength / 2
787
+ );
788
+ const f16 = new Uint16Array(bf16.length);
789
+ for (let j = 0; j < bf16.length; j++) {
790
+ f16[j] = float32ToFloat16(bf16ToFloat32(bf16[j]));
791
+ }
792
+ tensorData = new Uint8Array(f16.buffer, f16.byteOffset, f16.byteLength);
793
+ outDtype = 'F16';
794
+ }
795
+ } else if (tensorTargetQuant === 'f16' && sourceDtype === 'F32') {
796
+ if (tensorData.byteLength % 4 !== 0) {
797
+ throw new Error(`Invalid F32 tensor byte length for ${tensor.name}: ${tensorData.byteLength}`);
798
+ }
799
+ const f32 = new Float32Array(
800
+ tensorData.buffer,
801
+ tensorData.byteOffset,
802
+ tensorData.byteLength / 4
803
+ );
804
+ const f16 = new Uint16Array(f32.length);
805
+ for (let j = 0; j < f32.length; j++) {
806
+ f16[j] = float32ToFloat16(f32[j]);
807
+ }
808
+ tensorData = new Uint8Array(f16.buffer, f16.byteOffset, f16.byteLength);
809
+ outDtype = 'F16';
810
+ } else if (tensorTargetQuant === 'f16' && sourceDtype === 'BF16') {
811
+ if (tensorData.byteLength % 2 !== 0) {
812
+ throw new Error(`Invalid BF16 tensor byte length for ${tensor.name}: ${tensorData.byteLength}`);
813
+ }
814
+ const bf16 = new Uint16Array(
815
+ tensorData.buffer,
816
+ tensorData.byteOffset,
817
+ tensorData.byteLength / 2
818
+ );
819
+ const f16 = new Uint16Array(bf16.length);
820
+ for (let j = 0; j < bf16.length; j++) {
821
+ f16[j] = float32ToFloat16(bf16ToFloat32(bf16[j]));
822
+ }
823
+ tensorData = new Uint8Array(f16.buffer, f16.byteOffset, f16.byteLength);
824
+ outDtype = 'F16';
825
+ } else if (tensorTargetQuant === 'bf16' && sourceDtype === 'F32') {
826
+ if (tensorData.byteLength % 4 !== 0) {
827
+ throw new Error(`Invalid F32 tensor byte length for ${tensor.name}: ${tensorData.byteLength}`);
828
+ }
829
+ const f32 = new Float32Array(
830
+ tensorData.buffer,
831
+ tensorData.byteOffset,
832
+ tensorData.byteLength / 4
833
+ );
834
+ const bf16 = new Uint16Array(f32.length);
835
+ for (let j = 0; j < f32.length; j++) {
836
+ bf16[j] = float32ToBFloat16(f32[j]);
837
+ }
838
+ tensorData = new Uint8Array(bf16.buffer, bf16.byteOffset, bf16.byteLength);
839
+ outDtype = 'BF16';
840
+ } else if (tensorTargetQuant === 'f32' && sourceDtype === 'F16') {
841
+ if (tensorData.byteLength % 2 !== 0) {
842
+ throw new Error(`Invalid F16 tensor byte length for ${tensor.name}: ${tensorData.byteLength}`);
843
+ }
844
+ const f16 = new Uint16Array(
845
+ tensorData.buffer,
846
+ tensorData.byteOffset,
847
+ tensorData.byteLength / 2
848
+ );
849
+ const f32 = new Float32Array(f16.length);
850
+ for (let j = 0; j < f16.length; j++) {
851
+ f32[j] = float16ToFloat32(f16[j]);
852
+ }
853
+ tensorData = new Uint8Array(f32.buffer, f32.byteOffset, f32.byteLength);
854
+ outDtype = 'F32';
855
+ } else if (tensorTargetQuant === 'f32' && sourceDtype === 'BF16') {
856
+ if (tensorData.byteLength % 2 !== 0) {
857
+ throw new Error(`Invalid BF16 tensor byte length for ${tensor.name}: ${tensorData.byteLength}`);
858
+ }
859
+ const bf16 = new Uint16Array(
860
+ tensorData.buffer,
861
+ tensorData.byteOffset,
862
+ tensorData.byteLength / 2
863
+ );
864
+ const f32 = new Float32Array(bf16.length);
865
+ for (let j = 0; j < bf16.length; j++) {
866
+ f32[j] = bf16ToFloat32(bf16[j]);
867
+ }
868
+ tensorData = new Uint8Array(f32.buffer, f32.byteOffset, f32.byteLength);
869
+ outDtype = 'F32';
870
+ }
871
+
872
+ return {
873
+ tensorData,
874
+ outDtype,
875
+ outLayout,
876
+ sourceDtype,
877
+ tensorTargetQuant,
878
+ };
879
+ }
880
+
881
+
882
+ export function extractArchitecture(config, ggufConfig) {
883
+ const firstNumber = (...values) => {
884
+ for (const value of values) {
885
+ if (typeof value === 'number' && Number.isFinite(value)) {
886
+ return value;
887
+ }
888
+ }
889
+ return null;
890
+ };
891
+
892
+ const requireNumber = (value, label) => {
893
+ if (typeof value !== 'number' || !Number.isFinite(value)) {
894
+ throw new Error(`Missing ${label} in model config`);
895
+ }
896
+ return value;
897
+ };
898
+
899
+ const normalizeLinearNormMode = (value, sharedFlag = null) => {
900
+ if (typeof value === 'string') {
901
+ const normalized = value.trim().toLowerCase();
902
+ if (normalized === 'shared') return 'shared';
903
+ if (normalized === 'per_head' || normalized === 'per-head' || normalized === 'perhead') {
904
+ return 'per_head';
905
+ }
906
+ throw new Error(
907
+ `Unsupported linear_norm_mode="${value}" in model config. Supported values: "shared", "per_head".`
908
+ );
909
+ }
910
+ if (typeof sharedFlag === 'boolean') {
911
+ return sharedFlag ? 'shared' : 'per_head';
912
+ }
913
+ return undefined;
914
+ };
915
+
916
+ // Try HuggingFace config first
917
+ if (config && Object.keys(config).length > 0) {
918
+ const textConfig = getNestedTextConfig(config);
919
+ const fromConfig = (...keys) => {
920
+ const values = [];
921
+ for (const key of keys) {
922
+ values.push(config[key]);
923
+ }
924
+ for (const key of keys) {
925
+ values.push(textConfig?.[key]);
926
+ }
927
+ return firstNumber(...values);
928
+ };
929
+ const fromConfigValue = (...keys) => {
930
+ for (const key of keys) {
931
+ if (config[key] !== undefined) return config[key];
932
+ }
933
+ for (const key of keys) {
934
+ if (textConfig?.[key] !== undefined) return textConfig[key];
935
+ }
936
+ return undefined;
937
+ };
938
+ const numLayers = requireNumber(
939
+ fromConfig('num_hidden_layers', 'n_layer', 'num_layers'),
940
+ 'num_hidden_layers'
941
+ );
942
+ const hiddenSize = requireNumber(
943
+ fromConfig('hidden_size', 'n_embd', 'embedding_size'),
944
+ 'hidden_size'
945
+ );
946
+ const intermediateSize = requireNumber(
947
+ fromConfig('intermediate_size', 'n_inner', 'ffn_dim'),
948
+ 'intermediate_size'
949
+ );
950
+ const numHeads = requireNumber(
951
+ fromConfig('num_attention_heads', 'n_head', 'attention_heads'),
952
+ 'num_attention_heads'
953
+ );
954
+ const numKVHeads = fromConfig('num_key_value_heads', 'num_kv_heads') ?? numHeads;
955
+ const headDimFromConfig = fromConfig('head_dim') ?? Math.floor(hiddenSize / numHeads);
956
+ const vocabSize = requireNumber(
957
+ fromConfig('vocab_size', 'n_vocab'),
958
+ 'vocab_size'
959
+ );
960
+ const maxSeqLen = requireNumber(
961
+ fromConfig('max_position_embeddings', 'n_positions', 'max_seq_len'),
962
+ 'max_position_embeddings'
963
+ );
964
+ const ropeTheta = fromConfig('rope_theta') ?? undefined;
965
+ const linearNumKeyHeads = fromConfig('linear_num_key_heads');
966
+ const linearNumValueHeads = fromConfig('linear_num_value_heads');
967
+ const linearKeyHeadDim = fromConfig('linear_key_head_dim');
968
+ const linearValueHeadDim = fromConfig('linear_value_head_dim');
969
+ const linearConvKernelDim = fromConfig('linear_conv_kernel_dim');
970
+ const linearNormModeConfigured = normalizeLinearNormMode(
971
+ fromConfigValue('linear_norm_mode'),
972
+ fromConfigValue('linear_norm_shared')
973
+ );
974
+ const modelType = String(fromConfigValue('model_type') ?? '').trim().toLowerCase();
975
+ const rawLayerTypes = fromConfigValue('layer_types');
976
+ const layerTypes = Array.isArray(rawLayerTypes) ? rawLayerTypes : null;
977
+ const hasLinearLayers = Array.isArray(layerTypes)
978
+ && layerTypes.some((entry) => {
979
+ const normalized = String(entry ?? '').trim().toLowerCase();
980
+ return normalized === 'linear_attention'
981
+ || normalized === 'linear'
982
+ || normalized === 'gated_delta'
983
+ || normalized === 'gated_delta_net';
984
+ });
985
+ const linearNormMode = linearNormModeConfigured
986
+ ?? ((hasLinearLayers && modelType.startsWith('qwen')) ? 'shared' : undefined);
987
+
988
+ return {
989
+ numLayers,
990
+ hiddenSize,
991
+ intermediateSize,
992
+ numAttentionHeads: numHeads,
993
+ numKeyValueHeads: numKVHeads,
994
+ headDim: headDimFromConfig,
995
+ vocabSize,
996
+ maxSeqLen,
997
+ ropeTheta,
998
+ linearNumKeyHeads,
999
+ linearNumValueHeads,
1000
+ linearKeyHeadDim,
1001
+ linearValueHeadDim,
1002
+ linearConvKernelDim,
1003
+ linearNormMode,
1004
+ };
1005
+ }
1006
+
1007
+ // GGUF config
1008
+ if (ggufConfig) {
1009
+ const c = ggufConfig;
1010
+ const numLayers = requireNumber(
1011
+ firstNumber(c.blockCount, c.block_count),
1012
+ 'blockCount'
1013
+ );
1014
+ const hiddenSize = requireNumber(
1015
+ firstNumber(c.embeddingLength, c.embedding_length),
1016
+ 'embeddingLength'
1017
+ );
1018
+ const intermediateSize = requireNumber(
1019
+ firstNumber(c.feedForwardLength, c.feed_forward_length),
1020
+ 'feedForwardLength'
1021
+ );
1022
+ const numHeads = requireNumber(
1023
+ firstNumber(c.attentionHeadCount, c.attention_head_count),
1024
+ 'attentionHeadCount'
1025
+ );
1026
+ const numKVHeads = firstNumber(c.attentionHeadCountKV, c.attention_head_count_kv) ?? numHeads;
1027
+ const vocabSize = requireNumber(
1028
+ firstNumber(c.vocabSize, c.vocab_size),
1029
+ 'vocabSize'
1030
+ );
1031
+ const maxSeqLen = requireNumber(
1032
+ firstNumber(c.contextLength, c.context_length),
1033
+ 'contextLength'
1034
+ );
1035
+
1036
+ return {
1037
+ numLayers,
1038
+ hiddenSize,
1039
+ intermediateSize,
1040
+ numAttentionHeads: numHeads,
1041
+ numKeyValueHeads: numKVHeads,
1042
+ headDim: Math.floor(hiddenSize / numHeads),
1043
+ vocabSize,
1044
+ maxSeqLen,
1045
+ };
1046
+ }
1047
+
1048
+ throw new Error('Missing model config: cannot extract architecture');
1049
+ }
1050
+
1051
+ function getNestedTextConfig(config) {
1052
+ if (!config || typeof config !== 'object' || Array.isArray(config)) {
1053
+ return null;
1054
+ }
1055
+ if (config.text_config && typeof config.text_config === 'object' && !Array.isArray(config.text_config)) {
1056
+ return config.text_config;
1057
+ }
1058
+ if (config.language_config && typeof config.language_config === 'object' && !Array.isArray(config.language_config)) {
1059
+ return config.language_config;
1060
+ }
1061
+ return null;
1062
+ }
1063
+
1064
+
1065
+ export function buildTensorMap(tensors, shardSize) {
1066
+ if (!shardSize || shardSize <= 0) {
1067
+ throw new Error('Missing shard size for tensor map');
1068
+ }
1069
+ const tensorMap = {};
1070
+
1071
+ let globalOffset = 0;
1072
+ for (const tensor of tensors) {
1073
+ const startShard = Math.floor(globalOffset / shardSize);
1074
+ const offsetInShard = globalOffset % shardSize;
1075
+
1076
+ if (offsetInShard + tensor.size <= shardSize) {
1077
+ // Fits in single shard
1078
+ tensorMap[tensor.name] = {
1079
+ shard: startShard,
1080
+ offset: offsetInShard,
1081
+ size: tensor.size,
1082
+ shape: tensor.shape,
1083
+ dtype: tensor.dtype,
1084
+ };
1085
+ } else {
1086
+ // Spans multiple shards
1087
+ const spans = [];
1088
+ let remaining = tensor.size;
1089
+ let currentShard = startShard;
1090
+ let currentOffset = offsetInShard;
1091
+
1092
+ while (remaining > 0) {
1093
+ const available = shardSize - currentOffset;
1094
+ const chunkSize = Math.min(remaining, available);
1095
+ spans.push({
1096
+ shardIndex: currentShard,
1097
+ offset: currentOffset,
1098
+ size: chunkSize,
1099
+ });
1100
+ remaining -= chunkSize;
1101
+ currentShard++;
1102
+ currentOffset = 0;
1103
+ }
1104
+
1105
+ tensorMap[tensor.name] = {
1106
+ spans,
1107
+ size: tensor.size,
1108
+ shape: tensor.shape,
1109
+ dtype: tensor.dtype,
1110
+ };
1111
+ }
1112
+
1113
+ globalOffset += tensor.size;
1114
+ }
1115
+
1116
+ return tensorMap;
1117
+ }
1118
+
1119
+ function resolveConvertedAt(value) {
1120
+ if (value === undefined || value === null || value === '') {
1121
+ return new Date().toISOString();
1122
+ }
1123
+ if (typeof value !== 'string') {
1124
+ throw new Error('manifest convertedAt must be a string when provided.');
1125
+ }
1126
+ const parsed = Date.parse(value);
1127
+ if (Number.isNaN(parsed)) {
1128
+ throw new Error(`Invalid manifest convertedAt timestamp: "${value}"`);
1129
+ }
1130
+ return new Date(parsed).toISOString();
1131
+ }
1132
+
1133
+
1134
+ export function createManifest(
1135
+ modelId,
1136
+ model,
1137
+ shards,
1138
+ tensorLocations,
1139
+ sourceOrOptions
1140
+ ) {
1141
+ if (!sourceOrOptions) {
1142
+ throw new Error('Missing manifest options');
1143
+ }
1144
+ const options = typeof sourceOrOptions === 'string' ? { source: sourceOrOptions } : sourceOrOptions ?? {};
1145
+ const source = options.source;
1146
+ if (!source) {
1147
+ throw new Error('Missing manifest source');
1148
+ }
1149
+ const resolvedModelType =
1150
+ options.modelType ??
1151
+ model.modelType ??
1152
+ model.config?.architectures?.[0] ??
1153
+ model.architecture;
1154
+ if (!resolvedModelType) {
1155
+ throw new Error('Missing modelType for manifest');
1156
+ }
1157
+ const isDiffusion = resolvedModelType === 'diffusion';
1158
+ const architecture = options.architecture ?? model.architecture ?? (
1159
+ isDiffusion ? 'diffusion' : extractArchitecture(model.config, model.ggufConfig)
1160
+ );
1161
+ const rawConfig = model.config || {};
1162
+ const generationConfig = model.generationConfig ?? null;
1163
+ const resolvedArchitecture = isDiffusion
1164
+ ? architecture
1165
+ : resolveIntermediateSizeFromTensors(architecture, model, tensorLocations, rawConfig, modelId);
1166
+ const moeConfig = isDiffusion
1167
+ ? null
1168
+ : resolveManifestMoEConfig(model, { ...options, modelId }, rawConfig, resolvedModelType);
1169
+ let inference = options.inference;
1170
+ if (!inference) {
1171
+ throw new Error('inference config is required — use a v1 conversion config');
1172
+ }
1173
+
1174
+ const embeddingOutput = inferEmbeddingOutputConfig(tensorLocations);
1175
+ const hasExplicitEmbeddingPostprocessor = Object.prototype.hasOwnProperty.call(
1176
+ inference?.output ?? {},
1177
+ 'embeddingPostprocessor'
1178
+ );
1179
+ const embeddingPostprocessor = hasExplicitEmbeddingPostprocessor
1180
+ ? inference?.output?.embeddingPostprocessor
1181
+ : (model.embeddingPostprocessor ?? null);
1182
+ if (embeddingOutput || hasExplicitEmbeddingPostprocessor || embeddingPostprocessor) {
1183
+ inference = {
1184
+ ...inference,
1185
+ output: {
1186
+ ...inference.output,
1187
+ ...embeddingOutput,
1188
+ embeddingPostprocessor,
1189
+ },
1190
+ };
1191
+ }
1192
+
1193
+ const eosTokenId = options.eosTokenId !== undefined
1194
+ ? options.eosTokenId
1195
+ : isDiffusion
1196
+ ? null
1197
+ : resolveEosTokenId({
1198
+ config: rawConfig,
1199
+ generationConfig,
1200
+ tokenizer: model.tokenizer ?? model.tokenizerConfig ?? null,
1201
+ tokenizerJson: model.tokenizerJson ?? null,
1202
+ });
1203
+ const resolvedQuantization = options.quantization ?? model.quantization;
1204
+ if (!resolvedQuantization) {
1205
+ throw new Error('Missing quantization for manifest');
1206
+ }
1207
+ const hashAlgorithm = options.hashAlgorithm;
1208
+ if (!hashAlgorithm) {
1209
+ throw new Error('Missing hashAlgorithm for manifest');
1210
+ }
1211
+
1212
+ const manifest = {
1213
+ version: RDRR_VERSION,
1214
+ modelId,
1215
+ modelType: resolvedModelType,
1216
+ quantization: resolvedQuantization,
1217
+ quantizationInfo: options.quantizationInfo,
1218
+ architecture: resolvedArchitecture,
1219
+ moeConfig,
1220
+ inference,
1221
+ shards,
1222
+ tensors: tensorLocations,
1223
+ totalSize: shards.reduce((sum, s) => sum + s.size, 0),
1224
+ hashAlgorithm,
1225
+ eos_token_id: eosTokenId,
1226
+ config: isDiffusion ? rawConfig : (rawConfig.vision_config ? { vision_config: rawConfig.vision_config } : undefined),
1227
+ conversion: options.conversionInfo,
1228
+ metadata: {
1229
+ source,
1230
+ convertedAt: resolveConvertedAt(
1231
+ options.convertedAt
1232
+ ?? options.conversionInfo?.convertedAt
1233
+ ),
1234
+ },
1235
+ };
1236
+
1237
+ // Include tokenizer if available
1238
+ if (model.tokenizerJson) {
1239
+ manifest.tokenizer = buildBundledTokenizer(
1240
+ model.tokenizerJson,
1241
+ model.tokenizerConfig ?? null,
1242
+ rawConfig
1243
+ );
1244
+ manifest.metadata.hasTokenizer = true;
1245
+ } else {
1246
+ const tokenizer = buildSentencepieceTokenizer(
1247
+ model.tokenizerConfig ?? null,
1248
+ rawConfig,
1249
+ architecture,
1250
+ model.tokenizerModel ?? null
1251
+ );
1252
+ if (tokenizer) {
1253
+ manifest.tokenizer = tokenizer;
1254
+ manifest.metadata.hasTokenizer = true;
1255
+ }
1256
+ }
1257
+
1258
+ return manifest;
1259
+ }
1260
+
1261
+ // ============================================================================
1262
+ // Main Converter (uses I/O adapter)
1263
+ // ============================================================================
1264
+
1265
+
1266
+ export async function convertModel(model, io, options = {}) {
1267
+ const { onProgress, signal } = options;
1268
+ const converterConfig = options.converterConfig || createConverterConfig();
1269
+ const shardSize = options.shardSize ?? converterConfig.sharding.shardSizeBytes;
1270
+ if (!shardSize || shardSize <= 0) {
1271
+ throw new Error('Missing shardSize for conversion');
1272
+ }
1273
+ const modelIdInput = (
1274
+ options.modelId
1275
+ ?? converterConfig.output.modelBaseId
1276
+ ?? model.modelId
1277
+ ?? model.name
1278
+ );
1279
+ const modelId = modelIdInput ? sanitizeModelId(modelIdInput) : null;
1280
+ if (!modelId) {
1281
+ throw new Error('Missing modelId for conversion');
1282
+ }
1283
+ const tensors = resolveConversionTensors(model, converterConfig);
1284
+ if (!Array.isArray(tensors) || tensors.length === 0) {
1285
+ const textOnly = converterConfig?.output?.textOnly === true;
1286
+ if (textOnly) {
1287
+ throw new Error(
1288
+ 'No tensors selected for text-only conversion. ' +
1289
+ 'Expected language_model.* tensors or non-vision tensor names.'
1290
+ );
1291
+ }
1292
+ throw new Error('Missing tensors for conversion');
1293
+ }
1294
+ const totalTensors = tensors.length;
1295
+ const targetQuant = String(options.quantization ?? model.quantization ?? '').trim().toLowerCase();
1296
+ const tensorGroupModelType = String(options.modelType ?? model.modelType ?? 'transformer');
1297
+ const q4kLayout = normalizeQ4KLayout(options.quantizationInfo?.layout);
1298
+ const quantizeEmbeddings = resolveQuantizeEmbeddings(
1299
+ options.quantizationInfo ?? null,
1300
+ options.quantizeEmbeddings
1301
+ );
1302
+ const modulesToNotConvert = normalizeModulesToNotConvert(
1303
+ converterConfig?.quantization?.modulesToNotConvert ?? null
1304
+ );
1305
+ const shards = [];
1306
+ const tensorLocations = {};
1307
+
1308
+ // Current shard accumulator
1309
+ let currentShardIndex = 0;
1310
+ let currentShardBuffer = new Uint8Array(shardSize);
1311
+ let currentShardSize = 0;
1312
+ let totalSize = 0;
1313
+
1314
+ // Helper to flush current shard
1315
+ const flushShard = async () => {
1316
+ if (currentShardSize === 0) return;
1317
+ const shardData = currentShardBuffer.subarray(0, currentShardSize);
1318
+
1319
+ // Write shard and get hash
1320
+ const hash = await io.writeShard(currentShardIndex, shardData);
1321
+
1322
+ shards.push({
1323
+ index: currentShardIndex,
1324
+ filename: generateShardFilename(currentShardIndex),
1325
+ size: currentShardSize,
1326
+ hash,
1327
+ offset: currentShardIndex * shardSize,
1328
+ });
1329
+
1330
+ currentShardIndex++;
1331
+ currentShardSize = 0;
1332
+ };
1333
+
1334
+ // Process tensors
1335
+ for (let i = 0; i < tensors.length; i++) {
1336
+ if (signal?.aborted) {
1337
+ throw new DOMException('Conversion cancelled', 'AbortError');
1338
+ }
1339
+
1340
+ const tensor = tensors[i];
1341
+
1342
+ onProgress?.({
1343
+ stage: ConvertStage.WRITING,
1344
+ message: `Processing ${tensor.name}`,
1345
+ current: i + 1,
1346
+ total: totalTensors,
1347
+ percent: Math.round(((i + 1) / totalTensors) * 100),
1348
+ });
1349
+
1350
+ // Read tensor data
1351
+ const data = await io.readTensorData(tensor);
1352
+ const tensorDataInput = new Uint8Array(data);
1353
+ const transformContext = {
1354
+ targetQuant,
1355
+ q4kLayout,
1356
+ quantizationInfo: options.quantizationInfo ?? null,
1357
+ quantizeEmbeddings,
1358
+ modulesToNotConvert,
1359
+ };
1360
+ const transformResult = (
1361
+ typeof options.tensorTransformer === 'function'
1362
+ ? await options.tensorTransformer({
1363
+ tensor,
1364
+ tensorData: tensorDataInput,
1365
+ transformContext,
1366
+ reportProgress(currentBytes, totalBytes) {
1367
+ if (!Number.isFinite(currentBytes) || !Number.isFinite(totalBytes)) return;
1368
+ onProgress?.({
1369
+ stage: ConvertStage.WRITING,
1370
+ message: `Processing ${tensor.name}`,
1371
+ current: i + 1,
1372
+ total: totalTensors,
1373
+ percent: Math.round(((i + 1) / totalTensors) * 100),
1374
+ tensorName: tensor.name,
1375
+ tensorBytesCurrent: currentBytes,
1376
+ tensorBytesTotal: totalBytes,
1377
+ });
1378
+ },
1379
+ })
1380
+ : transformTensorBytes(tensor, tensorDataInput, transformContext)
1381
+ );
1382
+
1383
+ let tensorData = transformResult?.tensorData;
1384
+ if (!(tensorData instanceof Uint8Array)) {
1385
+ throw new Error(`Tensor transformer must return Uint8Array data for ${tensor.name}.`);
1386
+ }
1387
+ const outDtype = transformResult?.outDtype ?? tensor.dtype;
1388
+ const outLayout = transformResult?.outLayout ?? null;
1389
+
1390
+ const tensorSize = tensorData.byteLength;
1391
+
1392
+ // Track tensor location
1393
+ const tensorSpans = [];
1394
+
1395
+ // Add to current shard, splitting if necessary
1396
+ let remainingOffset = 0;
1397
+ while (remainingOffset < tensorData.length) {
1398
+ const availableInShard = shardSize - currentShardSize;
1399
+ const remainingSize = tensorData.length - remainingOffset;
1400
+ const chunkSize = Math.min(remainingSize, availableInShard);
1401
+ const chunk = tensorData.subarray(remainingOffset, remainingOffset + chunkSize);
1402
+ currentShardBuffer.set(chunk, currentShardSize);
1403
+
1404
+ const chunkOffset = currentShardSize;
1405
+ currentShardSize += chunkSize;
1406
+ totalSize += chunkSize;
1407
+
1408
+ tensorSpans.push({
1409
+ shardIndex: currentShardIndex,
1410
+ offset: chunkOffset,
1411
+ size: chunkSize,
1412
+ });
1413
+
1414
+ remainingOffset += chunkSize;
1415
+
1416
+ // Flush shard if full
1417
+ if (currentShardSize >= shardSize) {
1418
+ await flushShard();
1419
+ }
1420
+ }
1421
+
1422
+ // Record tensor location
1423
+ const role = resolveTensorRole(tensor);
1424
+ const group = resolveTensorGroup(tensor, tensorGroupModelType);
1425
+
1426
+ if (tensorSpans.length === 1) {
1427
+ tensorLocations[tensor.name] = {
1428
+ shard: tensorSpans[0].shardIndex,
1429
+ offset: tensorSpans[0].offset,
1430
+ size: tensorSize,
1431
+ shape: tensor.shape,
1432
+ dtype: outDtype,
1433
+ role,
1434
+ group,
1435
+ ...(outLayout ? { layout: outLayout } : {}),
1436
+ };
1437
+ } else {
1438
+ tensorLocations[tensor.name] = {
1439
+ spans: tensorSpans,
1440
+ size: tensorSize,
1441
+ shape: tensor.shape,
1442
+ dtype: outDtype,
1443
+ role,
1444
+ group,
1445
+ ...(outLayout ? { layout: outLayout } : {}),
1446
+ };
1447
+ }
1448
+
1449
+ }
1450
+
1451
+ // Flush final shard
1452
+ await flushShard();
1453
+
1454
+ if (signal?.aborted) {
1455
+ throw new DOMException('Conversion cancelled', 'AbortError');
1456
+ }
1457
+
1458
+ // Create manifest
1459
+ onProgress?.({
1460
+ stage: ConvertStage.MANIFEST,
1461
+ message: 'Creating manifest...',
1462
+ });
1463
+
1464
+ const tensorEntries = Object.entries(tensorLocations).map(([name, location]) => ({
1465
+ name,
1466
+ dtype: location?.dtype ?? null,
1467
+ role: location?.role ?? null,
1468
+ layout: location?.layout ?? null,
1469
+ }));
1470
+ const effectiveQuantizationInfo = resolveEffectiveQuantizationInfo(
1471
+ options.quantizationInfo ?? null,
1472
+ tensorEntries
1473
+ );
1474
+ const effectiveManifestQuantization = resolveManifestQuantization(
1475
+ effectiveQuantizationInfo.weights,
1476
+ options.quantization ?? model.quantization
1477
+ );
1478
+
1479
+ const manifest = createManifest(modelId, model, shards, tensorLocations, {
1480
+ source: 'convert-core',
1481
+ modelType: options.modelType,
1482
+ quantization: effectiveManifestQuantization,
1483
+ quantizationInfo: effectiveQuantizationInfo,
1484
+ hashAlgorithm: converterConfig.manifest.hashAlgorithm,
1485
+ architecture: options.architecture,
1486
+ inference: options.inference,
1487
+ eosTokenId: options.eosTokenId,
1488
+ convertedAt: converterConfig?.manifest?.conversion?.convertedAt ?? null,
1489
+ conversionInfo: converterConfig?.manifest?.conversion ?? null,
1490
+ });
1491
+
1492
+ // Write manifest
1493
+ await io.writeManifest(manifest);
1494
+
1495
+ onProgress?.({
1496
+ stage: ConvertStage.COMPLETE,
1497
+ message: 'Conversion complete!',
1498
+ modelId,
1499
+ shardCount: shards.length,
1500
+ totalSize: formatBytes(totalSize),
1501
+ });
1502
+
1503
+ const executionContractArtifact = buildExecutionContractArtifact(manifest);
1504
+ const layerPatternContractArtifact = getInferenceLayerPatternContractArtifact();
1505
+ const requiredInferenceFieldsArtifact = manifest?.modelType === 'transformer'
1506
+ && manifest?.inference
1507
+ && typeof manifest.inference === 'object'
1508
+ && manifest.inference.attention
1509
+ && typeof manifest.inference.attention === 'object'
1510
+ ? buildManifestRequiredInferenceFieldsArtifact(
1511
+ manifest?.inference ?? null,
1512
+ `${manifest?.modelId ?? modelId}.inference`
1513
+ )
1514
+ : null;
1515
+ return {
1516
+ manifest,
1517
+ shardCount: shards.length,
1518
+ tensorCount: tensors.length,
1519
+ totalSize,
1520
+ executionContractArtifact,
1521
+ layerPatternContractArtifact,
1522
+ requiredInferenceFieldsArtifact,
1523
+ };
1524
+ }
1525
+
1526
+ // ============================================================================
1527
+ // Utility Exports
1528
+ // ============================================================================
1529
+
1530
+ export { generateShardFilename };