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 +1,1375 @@
1
+ import { getDevice } from '../../../gpu/device.js';
2
+ import { acquireBuffer, releaseBuffer, readBuffer, isBufferActive } from '../../../memory/buffer-pool.js';
3
+ import { createTensor, dtypeBytes } from '../../../gpu/tensor.js';
4
+ import { getBuffer, getWeightDtype } from '../../../gpu/weight-buffer.js';
5
+ import { CommandRecorder } from '../../../gpu/command-recorder.js';
6
+ import { runConv2D, recordConv2D } from '../../../gpu/kernels/conv2d.js';
7
+ import { runGroupNorm, recordGroupNorm } from '../../../gpu/kernels/groupnorm.js';
8
+ import { runRMSNorm, recordRMSNorm } from '../../../gpu/kernels/rmsnorm.js';
9
+ import { runSiLU, runSiLURowSplit, recordSiLU, recordSiLURowSplit } from '../../../gpu/kernels/silu.js';
10
+ import { runMatmul, recordMatmul } from '../../../gpu/kernels/matmul.js';
11
+ import { runAttention, recordAttention } from '../../../gpu/kernels/attention.js';
12
+ import { runTranspose, recordTranspose } from '../../../gpu/kernels/transpose.js';
13
+ import { runResidualAdd, runBiasAdd, recordResidualAdd, recordBiasAdd } from '../../../gpu/kernels/residual.js';
14
+ import { runUpsample2D, recordUpsample2D } from '../../../gpu/kernels/upsample2d.js';
15
+ import { runDepthwiseConv2D, recordDepthwiseConv2D } from '../../../gpu/kernels/depthwise_conv2d.js';
16
+ import { runGroupedPointwiseConv2D, recordGroupedPointwiseConv2D } from '../../../gpu/kernels/grouped_pointwise_conv2d.js';
17
+ import { runSanaLinearAttention, recordSanaLinearAttention } from '../../../gpu/kernels/sana_linear_attention.js';
18
+ import { runPixelShuffle, recordPixelShuffle } from '../../../gpu/kernels/pixel_shuffle.js';
19
+ import { runRepeatChannels, recordRepeatChannels } from '../../../gpu/kernels/repeat_channels.js';
20
+ import { runReLU, recordReLU } from '../../../gpu/kernels/relu.js';
21
+ import { castF32ToF16, recordCastF32ToF16 } from '../../../gpu/kernels/cast.js';
22
+ import { f16ToF32 } from '../../../loader/dtype-utils.js';
23
+ import { log } from '../../../debug/index.js';
24
+
25
+ function clamp(value, min, max) {
26
+ return Math.max(min, Math.min(max, value));
27
+ }
28
+
29
+ function reshapeTensor(tensor, shape, label) {
30
+ return createTensor(tensor.buffer, tensor.dtype, shape, label ?? tensor.label);
31
+ }
32
+
33
+ function getWeight(weights, shapes, name) {
34
+ const value = weights.get(name);
35
+ if (!value) {
36
+ throw new Error(`Missing VAE weight: ${name}`);
37
+ }
38
+ const shape = shapes.get(name);
39
+ if (!shape) {
40
+ throw new Error(`Missing VAE weight shape: ${name}`);
41
+ }
42
+ return { value, shape };
43
+ }
44
+
45
+ function getWeightOptional(weights, shapes, name) {
46
+ const value = weights.get(name);
47
+ if (!value) return null;
48
+ const shape = shapes.get(name);
49
+ if (!shape) return null;
50
+ return { value, shape };
51
+ }
52
+
53
+ function getWeightByCandidates(weights, shapes, candidates, label) {
54
+ for (const name of candidates) {
55
+ const value = getWeightOptional(weights, shapes, name);
56
+ if (value) {
57
+ return { ...value, name };
58
+ }
59
+ }
60
+ throw new Error(
61
+ `Missing VAE weight: ${label}. Tried: ${candidates.join(', ')}`
62
+ );
63
+ }
64
+
65
+ function getConvShape(shape) {
66
+ if (!Array.isArray(shape) || shape.length !== 4) {
67
+ throw new Error(`Conv2D weight shape must be [out,in,h,w], got ${shape}`);
68
+ }
69
+ return {
70
+ outChannels: shape[0],
71
+ inChannels: shape[1],
72
+ kernelH: shape[2],
73
+ kernelW: shape[3],
74
+ };
75
+ }
76
+
77
+ function getLinearShape(shape, label) {
78
+ if (Array.isArray(shape) && shape.length === 2) {
79
+ return {
80
+ outFeatures: shape[0],
81
+ inFeatures: shape[1],
82
+ };
83
+ }
84
+ if (Array.isArray(shape) && shape.length === 4) {
85
+ if (shape[2] !== 1 || shape[3] !== 1) {
86
+ throw new Error(`Linear weight "${label}" with 4D shape must be 1x1, got ${shape}`);
87
+ }
88
+ return {
89
+ outFeatures: shape[0],
90
+ inFeatures: shape[1],
91
+ };
92
+ }
93
+ throw new Error(`Linear weight shape must be [out,in] or [out,in,1,1], got ${shape}`);
94
+ }
95
+
96
+ function readPositiveInteger(value) {
97
+ const parsed = Number(value);
98
+ if (!Number.isInteger(parsed) || parsed < 1) return null;
99
+ return parsed;
100
+ }
101
+
102
+ function resolveAttentionHeadShape(channels, config) {
103
+ const rawHeadDim = Array.isArray(config?.attention_head_dim)
104
+ ? config.attention_head_dim[0]
105
+ : config?.attention_head_dim;
106
+ const configuredHeadDim = readPositiveInteger(rawHeadDim);
107
+ if (configuredHeadDim && channels % configuredHeadDim === 0) {
108
+ return {
109
+ numHeads: channels / configuredHeadDim,
110
+ headDim: configuredHeadDim,
111
+ };
112
+ }
113
+
114
+ const configuredNumHeads = readPositiveInteger(config?.num_attention_heads);
115
+ if (configuredNumHeads && channels % configuredNumHeads === 0) {
116
+ return {
117
+ numHeads: configuredNumHeads,
118
+ headDim: channels / configuredNumHeads,
119
+ };
120
+ }
121
+ throw new Error(
122
+ `VAE attention requires explicit compatible attention_head_dim or num_attention_heads for channels=${channels}.`
123
+ );
124
+ }
125
+
126
+ function createBiasTensor(weight, label, fallbackDtype = 'f16') {
127
+ if (!weight) return null;
128
+ const dtype = getWeightDtype(weight.value) || fallbackDtype;
129
+ const shape = Array.isArray(weight.shape) && weight.shape.length > 0
130
+ ? weight.shape
131
+ : [0];
132
+ const size = shape.reduce((acc, value) => acc * value, 1);
133
+ if (!Number.isInteger(size) || size < 1) {
134
+ throw new Error(`Bias "${label}" has invalid shape: ${shape}`);
135
+ }
136
+ return createTensor(getBuffer(weight.value), dtype, [size], label);
137
+ }
138
+
139
+ function buildIndexList(weights, prefix) {
140
+ const indices = new Set();
141
+ for (const name of weights.keys()) {
142
+ if (!name.startsWith(prefix)) continue;
143
+ const rest = name.slice(prefix.length);
144
+ const match = rest.match(/^(\d+)\./);
145
+ if (!match) continue;
146
+ const idx = Number.parseInt(match[1], 10);
147
+ if (Number.isFinite(idx)) indices.add(idx);
148
+ }
149
+ return Array.from(indices).sort((a, b) => a - b);
150
+ }
151
+
152
+ function normalizePerBlockValue(value, count, label) {
153
+ if (Array.isArray(value)) {
154
+ if (value.length !== count) {
155
+ throw new Error(`${label} must have ${count} entries, got ${value.length}.`);
156
+ }
157
+ return value;
158
+ }
159
+ return Array.from({ length: count }, () => value);
160
+ }
161
+
162
+ function tensorElementCount(tensor) {
163
+ if (!Array.isArray(tensor?.shape) || tensor.shape.length === 0) {
164
+ throw new Error('Tensor shape is required.');
165
+ }
166
+ return tensor.shape.reduce((acc, value) => acc * value, 1);
167
+ }
168
+
169
+ function createKernelOps(recorder) {
170
+ if (!recorder) {
171
+ return {
172
+ conv2d: runConv2D,
173
+ groupNorm: runGroupNorm,
174
+ rmsNorm: runRMSNorm,
175
+ silu: runSiLU,
176
+ siluRowSplit: runSiLURowSplit,
177
+ matmul: runMatmul,
178
+ attention: runAttention,
179
+ transpose: runTranspose,
180
+ residualAdd: runResidualAdd,
181
+ biasAdd: runBiasAdd,
182
+ upsample2d: runUpsample2D,
183
+ depthwiseConv2d: runDepthwiseConv2D,
184
+ groupedPointwiseConv2d: runGroupedPointwiseConv2D,
185
+ sanaLinearAttention: runSanaLinearAttention,
186
+ pixelShuffle: runPixelShuffle,
187
+ repeatChannels: runRepeatChannels,
188
+ relu: runReLU,
189
+ castF32ToF16,
190
+ };
191
+ }
192
+ return {
193
+ conv2d: (...args) => recordConv2D(recorder, ...args),
194
+ groupNorm: (...args) => recordGroupNorm(recorder, ...args),
195
+ rmsNorm: (...args) => recordRMSNorm(recorder, ...args),
196
+ silu: (...args) => recordSiLU(recorder, ...args),
197
+ siluRowSplit: (...args) => recordSiLURowSplit(recorder, ...args),
198
+ matmul: (...args) => recordMatmul(recorder, ...args),
199
+ attention: (...args) => recordAttention(recorder, ...args),
200
+ transpose: (...args) => recordTranspose(recorder, ...args),
201
+ residualAdd: (...args) => recordResidualAdd(recorder, ...args),
202
+ biasAdd: (...args) => recordBiasAdd(recorder, ...args),
203
+ upsample2d: (...args) => recordUpsample2D(recorder, ...args),
204
+ depthwiseConv2d: (...args) => recordDepthwiseConv2D(recorder, ...args),
205
+ groupedPointwiseConv2d: (...args) => recordGroupedPointwiseConv2D(recorder, ...args),
206
+ sanaLinearAttention: (...args) => recordSanaLinearAttention(recorder, ...args),
207
+ pixelShuffle: (...args) => recordPixelShuffle(recorder, ...args),
208
+ repeatChannels: (...args) => recordRepeatChannels(recorder, ...args),
209
+ relu: (...args) => recordReLU(recorder, ...args),
210
+ castF32ToF16: (...args) => recordCastF32ToF16(recorder, ...args),
211
+ };
212
+ }
213
+
214
+ function createBufferReleaser(recorder) {
215
+ if (!recorder) {
216
+ return (buffer) => {
217
+ if (!buffer || !isBufferActive(buffer)) return;
218
+ releaseBuffer(buffer);
219
+ };
220
+ }
221
+ return (buffer) => {
222
+ if (!buffer) return;
223
+ recorder.trackTemporaryBuffer(buffer);
224
+ };
225
+ }
226
+
227
+ function sumProfileTimings(timings) {
228
+ if (!timings) return null;
229
+ return Object.values(timings).reduce((sum, value) => sum + value, 0);
230
+ }
231
+
232
+ async function applyConv2D(state, weights, shapes, namePrefix, options = {}, ops, release) {
233
+ const weightName = `${namePrefix}.weight`;
234
+ const biasName = `${namePrefix}.bias`;
235
+ const weight = getWeight(weights, shapes, weightName);
236
+ const bias = getWeightOptional(weights, shapes, biasName);
237
+ const { outChannels, inChannels, kernelH, kernelW } = getConvShape(weight.shape);
238
+
239
+ if (inChannels !== state.channels) {
240
+ log.warn('Diffusion', `VAE conv channel mismatch: ${namePrefix} in=${inChannels} state=${state.channels}`);
241
+ }
242
+
243
+ const output = await ops.conv2d(
244
+ state.tensor,
245
+ weight.value,
246
+ bias?.value ?? null,
247
+ {
248
+ inChannels,
249
+ outChannels,
250
+ height: state.height,
251
+ width: state.width,
252
+ kernelH,
253
+ kernelW,
254
+ stride: options.stride ?? 1,
255
+ pad: options.pad ?? 1,
256
+ }
257
+ );
258
+
259
+ release(state.tensor.buffer);
260
+
261
+ return {
262
+ tensor: output,
263
+ channels: outChannels,
264
+ height: Math.floor((state.height + (options.pad ?? 1) * 2 - kernelH) / (options.stride ?? 1)) + 1,
265
+ width: Math.floor((state.width + (options.pad ?? 1) * 2 - kernelW) / (options.stride ?? 1)) + 1,
266
+ };
267
+ }
268
+
269
+ async function submitCopyWork(device, recorder, encoder) {
270
+ if (recorder) {
271
+ return;
272
+ }
273
+ device.queue.submit([encoder.finish()]);
274
+ }
275
+
276
+ async function concatChannelTensors(tensors, height, width, recorder) {
277
+ if (!Array.isArray(tensors) || tensors.length === 0) {
278
+ throw new Error('concatChannelTensors requires at least one tensor.');
279
+ }
280
+ const device = getDevice();
281
+ if (!device) {
282
+ throw new Error('Channel tensor concatenation requires a WebGPU device.');
283
+ }
284
+ const dtype = tensors[0].dtype;
285
+ const bytesPerElement = dtypeBytes(dtype);
286
+ let totalChannels = 0;
287
+ for (const tensor of tensors) {
288
+ if (tensor.dtype !== dtype) {
289
+ throw new Error('concatChannelTensors requires matching dtypes.');
290
+ }
291
+ if (tensor.shape[1] !== height || tensor.shape[2] !== width) {
292
+ throw new Error('concatChannelTensors requires matching spatial dimensions.');
293
+ }
294
+ totalChannels += tensor.shape[0];
295
+ }
296
+
297
+ const output = acquireBuffer(totalChannels * height * width * bytesPerElement, undefined, 'vae_concat_channels');
298
+ const encoder = recorder ? recorder.getEncoder() : device.createCommandEncoder({ label: 'vae_concat_channels' });
299
+ let channelOffset = 0;
300
+ for (const tensor of tensors) {
301
+ const byteLength = tensor.shape[0] * height * width * bytesPerElement;
302
+ encoder.copyBufferToBuffer(
303
+ tensor.buffer,
304
+ 0,
305
+ output,
306
+ channelOffset * height * width * bytesPerElement,
307
+ byteLength
308
+ );
309
+ channelOffset += tensor.shape[0];
310
+ }
311
+ await submitCopyWork(device, recorder, encoder);
312
+ return createTensor(output, dtype, [totalChannels, height, width], 'vae_concat_channels');
313
+ }
314
+
315
+ async function sliceChannelTensor(tensor, startChannel, channelCount, height, width, recorder) {
316
+ const device = getDevice();
317
+ if (!device) {
318
+ throw new Error('Channel tensor slicing requires a WebGPU device.');
319
+ }
320
+ const bytesPerElement = dtypeBytes(tensor.dtype);
321
+ const channelSize = height * width * bytesPerElement;
322
+ const output = acquireBuffer(channelCount * channelSize, undefined, 'vae_slice_channels');
323
+ const encoder = recorder ? recorder.getEncoder() : device.createCommandEncoder({ label: 'vae_slice_channels' });
324
+ encoder.copyBufferToBuffer(
325
+ tensor.buffer,
326
+ startChannel * channelSize,
327
+ output,
328
+ 0,
329
+ channelCount * channelSize
330
+ );
331
+ await submitCopyWork(device, recorder, encoder);
332
+ return createTensor(output, tensor.dtype, [channelCount, height, width], 'vae_slice_channels');
333
+ }
334
+
335
+ async function runChannelwiseRmsNorm(state, normWeight, normBias, eps, ops, release) {
336
+ const spatial = state.height * state.width;
337
+ const channelsSpatial = reshapeTensor(state.tensor, [state.channels, spatial], 'vae_rmsnorm_channels_spatial');
338
+ const tokens = await ops.transpose(channelsSpatial, state.channels, spatial);
339
+ const normed = await ops.rmsNorm(tokens, normWeight.value, eps, {
340
+ batchSize: spatial,
341
+ hiddenSize: state.channels,
342
+ });
343
+ release(tokens.buffer);
344
+ let shifted = normed;
345
+ if (normBias) {
346
+ const biasTensor = createBiasTensor(normBias, `${normBias.name ?? 'vae_rmsnorm_bias'}`, normed.dtype);
347
+ shifted = await ops.biasAdd(normed, biasTensor, spatial, state.channels);
348
+ release(normed.buffer);
349
+ }
350
+ const channelsFirst = await ops.transpose(shifted, spatial, state.channels);
351
+ release(shifted.buffer);
352
+ return {
353
+ tensor: reshapeTensor(channelsFirst, [state.channels, state.height, state.width], 'vae_rmsnorm_output'),
354
+ channels: state.channels,
355
+ height: state.height,
356
+ width: state.width,
357
+ };
358
+ }
359
+
360
+ async function channelsToTokens(state, ops) {
361
+ const spatial = state.height * state.width;
362
+ const channelsSpatial = reshapeTensor(state.tensor, [state.channels, spatial], 'vae_channels_spatial');
363
+ const tokens = await ops.transpose(channelsSpatial, state.channels, spatial);
364
+ return {
365
+ tensor: tokens,
366
+ numTokens: spatial,
367
+ };
368
+ }
369
+
370
+ async function tokensToChannels(tokens, channels, height, width, ops) {
371
+ const channelsSpatial = await ops.transpose(tokens, height * width, channels);
372
+ return {
373
+ tensor: reshapeTensor(channelsSpatial, [channels, height, width], 'vae_tokens_channels'),
374
+ channels,
375
+ height,
376
+ width,
377
+ };
378
+ }
379
+
380
+ async function runResnetBlock(state, weights, shapes, prefix, config, ops, release) {
381
+ const numGroups = config.numGroups;
382
+ const eps = config.eps;
383
+ const channels = state.channels;
384
+
385
+ const norm1 = getWeight(weights, shapes, `${prefix}.norm1.weight`);
386
+ const norm1Bias = getWeight(weights, shapes, `${prefix}.norm1.bias`);
387
+ const normed1 = await ops.groupNorm(state.tensor, norm1.value, norm1Bias.value, {
388
+ channels,
389
+ height: state.height,
390
+ width: state.width,
391
+ numGroups,
392
+ eps,
393
+ });
394
+
395
+ const silu1 = await ops.silu(normed1, { size: channels * state.height * state.width, swigluLimit: null });
396
+ release(normed1.buffer);
397
+ const silu1View = reshapeTensor(silu1, [channels, state.height, state.width], 'vae_resnet_silu1');
398
+
399
+ const conv1 = await applyConv2D(
400
+ { tensor: silu1View, channels, height: state.height, width: state.width },
401
+ weights,
402
+ shapes,
403
+ `${prefix}.conv1`,
404
+ { pad: 1 },
405
+ ops,
406
+ release
407
+ );
408
+
409
+ const norm2 = getWeight(weights, shapes, `${prefix}.norm2.weight`);
410
+ const norm2Bias = getWeight(weights, shapes, `${prefix}.norm2.bias`);
411
+ const normed2 = await ops.groupNorm(conv1.tensor, norm2.value, norm2Bias.value, {
412
+ channels: conv1.channels,
413
+ height: conv1.height,
414
+ width: conv1.width,
415
+ numGroups,
416
+ eps,
417
+ });
418
+
419
+ release(conv1.tensor.buffer);
420
+
421
+ const silu2 = await ops.silu(normed2, { size: conv1.channels * conv1.height * conv1.width, swigluLimit: null });
422
+ release(normed2.buffer);
423
+ const silu2View = reshapeTensor(silu2, [conv1.channels, conv1.height, conv1.width], 'vae_resnet_silu2');
424
+
425
+ const conv2 = await applyConv2D(
426
+ { tensor: silu2View, channels: conv1.channels, height: conv1.height, width: conv1.width },
427
+ weights,
428
+ shapes,
429
+ `${prefix}.conv2`,
430
+ { pad: 1 },
431
+ ops,
432
+ release
433
+ );
434
+
435
+ let residualTensor = state.tensor;
436
+
437
+ if (weights.has(`${prefix}.conv_shortcut.weight`)) {
438
+ const shortcut = await applyConv2D(state, weights, shapes, `${prefix}.conv_shortcut`, { pad: 0 }, ops, release);
439
+ residualTensor = shortcut.tensor;
440
+ }
441
+
442
+ const size = conv2.channels * conv2.height * conv2.width;
443
+ const residual = reshapeTensor(residualTensor, [size], 'vae_resnet_residual');
444
+ const output = await ops.residualAdd(
445
+ reshapeTensor(conv2.tensor, [size], 'vae_resnet_main'),
446
+ residual,
447
+ size,
448
+ { useVec4: true }
449
+ );
450
+
451
+ if (residualTensor === state.tensor) {
452
+ release(state.tensor.buffer);
453
+ } else {
454
+ release(residualTensor.buffer);
455
+ }
456
+
457
+ release(conv2.tensor.buffer);
458
+
459
+ return {
460
+ tensor: reshapeTensor(output, [conv2.channels, conv2.height, conv2.width], 'vae_resnet_output'),
461
+ channels: conv2.channels,
462
+ height: conv2.height,
463
+ width: conv2.width,
464
+ };
465
+ }
466
+
467
+ async function runMidBlockAttention(state, weights, shapes, prefix, config, ops, release) {
468
+ const channels = state.channels;
469
+ const height = state.height;
470
+ const width = state.width;
471
+ const spatial = height * width;
472
+ if (!Number.isFinite(spatial) || spatial <= 0) {
473
+ throw new Error('VAE mid-block attention requires a positive spatial size.');
474
+ }
475
+
476
+ const normWeight = getWeightByCandidates(
477
+ weights,
478
+ shapes,
479
+ [`${prefix}.group_norm.weight`, `${prefix}.norm.weight`],
480
+ `${prefix}.group_norm.weight`
481
+ );
482
+ const normBias = getWeightByCandidates(
483
+ weights,
484
+ shapes,
485
+ [`${prefix}.group_norm.bias`, `${prefix}.norm.bias`],
486
+ `${prefix}.group_norm.bias`
487
+ );
488
+
489
+ const normed = await ops.groupNorm(state.tensor, normWeight.value, normBias.value, {
490
+ channels,
491
+ height,
492
+ width,
493
+ numGroups: config.numGroups,
494
+ eps: config.eps,
495
+ });
496
+ const normedChannelsSpatial = reshapeTensor(normed, [channels, spatial], 'vae_attn_norm_cs');
497
+ const normedTokens = await ops.transpose(normedChannelsSpatial, channels, spatial);
498
+ release(normed.buffer);
499
+
500
+ const residualChannelsSpatial = reshapeTensor(state.tensor, [channels, spatial], 'vae_attn_residual_cs');
501
+ const residualTokens = await ops.transpose(residualChannelsSpatial, channels, spatial);
502
+ release(state.tensor.buffer);
503
+
504
+ const qWeight = getWeightByCandidates(weights, shapes, [`${prefix}.to_q.weight`], `${prefix}.to_q.weight`);
505
+ const kWeight = getWeightByCandidates(weights, shapes, [`${prefix}.to_k.weight`], `${prefix}.to_k.weight`);
506
+ const vWeight = getWeightByCandidates(weights, shapes, [`${prefix}.to_v.weight`], `${prefix}.to_v.weight`);
507
+ const qBias = getWeightOptional(weights, shapes, `${prefix}.to_q.bias`);
508
+ const kBias = getWeightOptional(weights, shapes, `${prefix}.to_k.bias`);
509
+ const vBias = getWeightOptional(weights, shapes, `${prefix}.to_v.bias`);
510
+ const qShape = getLinearShape(qWeight.shape, qWeight.name);
511
+ const kShape = getLinearShape(kWeight.shape, kWeight.name);
512
+ const vShape = getLinearShape(vWeight.shape, vWeight.name);
513
+
514
+ if (qShape.inFeatures !== channels || kShape.inFeatures !== channels || vShape.inFeatures !== channels) {
515
+ throw new Error(
516
+ `VAE mid-block attention projection mismatch: expected inFeatures=${channels}, ` +
517
+ `got q=${qShape.inFeatures}, k=${kShape.inFeatures}, v=${vShape.inFeatures}.`
518
+ );
519
+ }
520
+ if (qShape.outFeatures !== kShape.outFeatures || qShape.outFeatures !== vShape.outFeatures) {
521
+ throw new Error(
522
+ `VAE mid-block attention projection mismatch: q/k/v outFeatures differ ` +
523
+ `(${qShape.outFeatures}, ${kShape.outFeatures}, ${vShape.outFeatures}).`
524
+ );
525
+ }
526
+
527
+ const hiddenSize = qShape.outFeatures;
528
+ const projectionDtype = normedTokens.dtype;
529
+ let q = await ops.matmul(normedTokens, qWeight.value, spatial, hiddenSize, channels, {
530
+ outputDtype: projectionDtype,
531
+ transposeB: 'auto',
532
+ });
533
+ let k = await ops.matmul(normedTokens, kWeight.value, spatial, hiddenSize, channels, {
534
+ outputDtype: projectionDtype,
535
+ transposeB: 'auto',
536
+ });
537
+ let v = await ops.matmul(normedTokens, vWeight.value, spatial, hiddenSize, channels, {
538
+ outputDtype: projectionDtype,
539
+ transposeB: 'auto',
540
+ });
541
+
542
+ const qBiasTensor = createBiasTensor(qBias, `${prefix}.to_q.bias`, projectionDtype);
543
+ const kBiasTensor = createBiasTensor(kBias, `${prefix}.to_k.bias`, projectionDtype);
544
+ const vBiasTensor = createBiasTensor(vBias, `${prefix}.to_v.bias`, projectionDtype);
545
+ if (qBiasTensor) q = await ops.biasAdd(q, qBiasTensor, spatial, hiddenSize);
546
+ if (kBiasTensor) k = await ops.biasAdd(k, kBiasTensor, spatial, hiddenSize);
547
+ if (vBiasTensor) v = await ops.biasAdd(v, vBiasTensor, spatial, hiddenSize);
548
+
549
+ const { numHeads, headDim } = resolveAttentionHeadShape(hiddenSize, config.modelConfig);
550
+ const attn = await ops.attention(
551
+ q,
552
+ k,
553
+ v,
554
+ null,
555
+ numHeads,
556
+ headDim,
557
+ {
558
+ seqLen: spatial,
559
+ kvLen: spatial,
560
+ numKVHeads: numHeads,
561
+ causal: false,
562
+ }
563
+ );
564
+ release(q.buffer);
565
+ release(k.buffer);
566
+ release(v.buffer);
567
+
568
+ const outWeight = getWeightByCandidates(
569
+ weights,
570
+ shapes,
571
+ [`${prefix}.to_out.0.weight`, `${prefix}.to_out.weight`],
572
+ `${prefix}.to_out.0.weight`
573
+ );
574
+ const outBias = getWeightOptional(weights, shapes, `${prefix}.to_out.0.bias`)
575
+ || getWeightOptional(weights, shapes, `${prefix}.to_out.bias`);
576
+ const outShape = getLinearShape(outWeight.shape, outWeight.name);
577
+ if (outShape.inFeatures !== hiddenSize) {
578
+ throw new Error(
579
+ `VAE mid-block attention output projection mismatch: expected inFeatures=${hiddenSize}, got ${outShape.inFeatures}.`
580
+ );
581
+ }
582
+ if (outShape.outFeatures !== channels) {
583
+ throw new Error(
584
+ `VAE mid-block attention output projection mismatch: expected outFeatures=${channels}, got ${outShape.outFeatures}.`
585
+ );
586
+ }
587
+
588
+ let projected = await ops.matmul(attn, outWeight.value, spatial, outShape.outFeatures, outShape.inFeatures, {
589
+ outputDtype: projectionDtype,
590
+ transposeB: 'auto',
591
+ });
592
+ release(attn.buffer);
593
+ const outBiasTensor = createBiasTensor(outBias, `${prefix}.to_out.0.bias`, projectionDtype);
594
+ if (outBiasTensor) {
595
+ projected = await ops.biasAdd(projected, outBiasTensor, spatial, outShape.outFeatures);
596
+ }
597
+
598
+ const combined = await ops.residualAdd(projected, residualTokens, spatial * outShape.outFeatures, { useVec4: true });
599
+ release(projected.buffer);
600
+ release(residualTokens.buffer);
601
+ release(normedTokens.buffer);
602
+
603
+ const combinedChannelsSpatial = await ops.transpose(combined, spatial, outShape.outFeatures);
604
+ release(combined.buffer);
605
+
606
+ return {
607
+ tensor: reshapeTensor(combinedChannelsSpatial, [outShape.outFeatures, height, width], 'vae_attn_out'),
608
+ channels: outShape.outFeatures,
609
+ height,
610
+ width,
611
+ };
612
+ }
613
+
614
+ async function runAutoencoderDCInputProjection(state, weights, shapes, config, ops, release) {
615
+ const blockOutChannels = config.decoder_block_out_channels;
616
+ if (!Array.isArray(blockOutChannels) || blockOutChannels.length === 0) {
617
+ throw new Error('AutoencoderDC decode requires decoder_block_out_channels.');
618
+ }
619
+ const outChannels = blockOutChannels[blockOutChannels.length - 1];
620
+ const repeats = outChannels / state.channels;
621
+ if (!Number.isInteger(repeats) || repeats < 1) {
622
+ throw new Error(
623
+ `AutoencoderDC input shortcut requires an integer repeat factor; got ${outChannels}/${state.channels}.`
624
+ );
625
+ }
626
+
627
+ const shortcut = await ops.repeatChannels(state.tensor, {
628
+ inChannels: state.channels,
629
+ height: state.height,
630
+ width: state.width,
631
+ repeats,
632
+ });
633
+ const projected = await applyConv2D(state, weights, shapes, 'vae.decoder.conv_in', { pad: 1 }, ops, release);
634
+ const size = projected.channels * projected.height * projected.width;
635
+ const combined = await ops.residualAdd(
636
+ reshapeTensor(projected.tensor, [size], 'vae_dc_conv_in'),
637
+ reshapeTensor(shortcut, [size], 'vae_dc_conv_in_shortcut'),
638
+ size,
639
+ { useVec4: true }
640
+ );
641
+ release(projected.tensor.buffer);
642
+ release(shortcut.buffer);
643
+ return {
644
+ tensor: reshapeTensor(combined, [projected.channels, projected.height, projected.width], 'vae_dc_conv_in_out'),
645
+ channels: projected.channels,
646
+ height: projected.height,
647
+ width: projected.width,
648
+ };
649
+ }
650
+
651
+ async function runAutoencoderDCUpBlock(state, weights, shapes, prefix, config, ops, release, recorder) {
652
+ const convWeight = getWeight(weights, shapes, `${prefix}.conv.weight`);
653
+ const convBias = getWeightOptional(weights, shapes, `${prefix}.conv.bias`);
654
+ const { outChannels, inChannels, kernelH, kernelW } = getConvShape(convWeight.shape);
655
+ if (inChannels !== state.channels) {
656
+ throw new Error(
657
+ `AutoencoderDC up block "${prefix}" expected ${inChannels} input channels, got ${state.channels}.`
658
+ );
659
+ }
660
+ const factor = 2;
661
+ const outHeight = state.height * factor;
662
+ const outWidth = state.width * factor;
663
+ const shortcutRepeats = outChannels * factor * factor / state.channels;
664
+ if (!Number.isInteger(shortcutRepeats) || shortcutRepeats < 1) {
665
+ throw new Error(
666
+ `AutoencoderDC up block "${prefix}" requires integer shortcut repeats; got ${outChannels}/${state.channels}.`
667
+ );
668
+ }
669
+
670
+ let projected;
671
+ if (config.upsample_block_type === 'interpolate') {
672
+ const upsampled = await ops.upsample2d(state.tensor, {
673
+ channels: state.channels,
674
+ height: state.height,
675
+ width: state.width,
676
+ scale: factor,
677
+ });
678
+ projected = await ops.conv2d(
679
+ reshapeTensor(upsampled, [state.channels, outHeight, outWidth], 'vae_dc_upsample'),
680
+ convWeight.value,
681
+ convBias?.value ?? null,
682
+ {
683
+ inChannels: state.channels,
684
+ outChannels,
685
+ height: outHeight,
686
+ width: outWidth,
687
+ kernelH,
688
+ kernelW,
689
+ stride: 1,
690
+ pad: 1,
691
+ }
692
+ );
693
+ release(upsampled.buffer);
694
+ } else if (config.upsample_block_type === 'pixel_shuffle') {
695
+ const conv = await ops.conv2d(state.tensor, convWeight.value, convBias?.value ?? null, {
696
+ inChannels: state.channels,
697
+ outChannels: outChannels * factor * factor,
698
+ height: state.height,
699
+ width: state.width,
700
+ kernelH,
701
+ kernelW,
702
+ stride: 1,
703
+ pad: 1,
704
+ });
705
+ projected = await ops.pixelShuffle(conv, {
706
+ outChannels,
707
+ outHeight,
708
+ outWidth,
709
+ gridWidth: state.width,
710
+ gridHeight: state.height,
711
+ patchSize: factor,
712
+ patchChannels: outChannels * factor * factor,
713
+ });
714
+ release(conv.buffer);
715
+ } else {
716
+ throw new Error(
717
+ `Unsupported AutoencoderDC upsample_block_type "${config.upsample_block_type}".`
718
+ );
719
+ }
720
+
721
+ const repeated = await ops.repeatChannels(state.tensor, {
722
+ inChannels: state.channels,
723
+ height: state.height,
724
+ width: state.width,
725
+ repeats: shortcutRepeats,
726
+ });
727
+ const shortcut = await ops.pixelShuffle(repeated, {
728
+ outChannels,
729
+ outHeight,
730
+ outWidth,
731
+ gridWidth: state.width,
732
+ gridHeight: state.height,
733
+ patchSize: factor,
734
+ patchChannels: state.channels * shortcutRepeats,
735
+ });
736
+ release(repeated.buffer);
737
+ release(state.tensor.buffer);
738
+
739
+ const size = outChannels * outHeight * outWidth;
740
+ const combined = await ops.residualAdd(
741
+ reshapeTensor(projected, [size], 'vae_dc_up_main'),
742
+ reshapeTensor(shortcut, [size], 'vae_dc_up_shortcut'),
743
+ size,
744
+ { useVec4: true }
745
+ );
746
+ release(projected.buffer);
747
+ release(shortcut.buffer);
748
+ return {
749
+ tensor: reshapeTensor(combined, [outChannels, outHeight, outWidth], 'vae_dc_up_out'),
750
+ channels: outChannels,
751
+ height: outHeight,
752
+ width: outWidth,
753
+ };
754
+ }
755
+
756
+ async function runAutoencoderDCResBlock(state, weights, shapes, prefix, eps, ops, release) {
757
+ const conv1Weight = getWeight(weights, shapes, `${prefix}.conv1.weight`);
758
+ const conv1Bias = getWeightOptional(weights, shapes, `${prefix}.conv1.bias`);
759
+ const conv1Shape = getConvShape(conv1Weight.shape);
760
+ const conv1Tensor = await ops.conv2d(state.tensor, conv1Weight.value, conv1Bias?.value ?? null, {
761
+ inChannels: conv1Shape.inChannels,
762
+ outChannels: conv1Shape.outChannels,
763
+ height: state.height,
764
+ width: state.width,
765
+ kernelH: conv1Shape.kernelH,
766
+ kernelW: conv1Shape.kernelW,
767
+ stride: 1,
768
+ pad: 1,
769
+ });
770
+ const conv1 = {
771
+ tensor: reshapeTensor(conv1Tensor, [conv1Shape.outChannels, state.height, state.width], 'vae_dc_resblock_conv1'),
772
+ channels: conv1Shape.outChannels,
773
+ height: state.height,
774
+ width: state.width,
775
+ };
776
+ const activated = await ops.silu(conv1.tensor, {
777
+ size: conv1.channels * conv1.height * conv1.width,
778
+ swigluLimit: null,
779
+ });
780
+ release(conv1.tensor.buffer);
781
+ const conv2Weight = getWeight(weights, shapes, `${prefix}.conv2.weight`);
782
+ const conv2Shape = getConvShape(conv2Weight.shape);
783
+ const conv2 = await ops.conv2d(
784
+ reshapeTensor(activated, [conv1.channels, conv1.height, conv1.width], 'vae_dc_resblock_act'),
785
+ conv2Weight.value,
786
+ null,
787
+ {
788
+ inChannels: conv1.channels,
789
+ outChannels: conv2Shape.outChannels,
790
+ height: conv1.height,
791
+ width: conv1.width,
792
+ kernelH: 3,
793
+ kernelW: 3,
794
+ stride: 1,
795
+ pad: 1,
796
+ }
797
+ );
798
+ release(activated.buffer);
799
+
800
+ const normed = await runChannelwiseRmsNorm(
801
+ {
802
+ tensor: reshapeTensor(conv2, [conv2Shape.outChannels, conv1.height, conv1.width], 'vae_dc_resblock_conv2'),
803
+ channels: conv2Shape.outChannels,
804
+ height: conv1.height,
805
+ width: conv1.width,
806
+ },
807
+ getWeight(weights, shapes, `${prefix}.norm.weight`),
808
+ getWeightOptional(weights, shapes, `${prefix}.norm.bias`),
809
+ eps,
810
+ ops,
811
+ release
812
+ );
813
+ release(conv2.buffer);
814
+
815
+ const size = normed.channels * normed.height * normed.width;
816
+ const combined = await ops.residualAdd(
817
+ reshapeTensor(normed.tensor, [size], 'vae_dc_resblock_main'),
818
+ reshapeTensor(state.tensor, [size], 'vae_dc_resblock_residual'),
819
+ size,
820
+ { useVec4: true }
821
+ );
822
+ release(normed.tensor.buffer);
823
+ release(state.tensor.buffer);
824
+ return {
825
+ tensor: reshapeTensor(combined, [normed.channels, normed.height, normed.width], 'vae_dc_resblock_out'),
826
+ channels: normed.channels,
827
+ height: normed.height,
828
+ width: normed.width,
829
+ };
830
+ }
831
+
832
+ async function runAutoencoderDCAttention(state, weights, shapes, prefix, attentionHeadDim, qkvMultiscales, eps, ops, release, recorder) {
833
+ const qWeight = getWeight(weights, shapes, `${prefix}.attn.to_q.weight`);
834
+ const kWeight = getWeight(weights, shapes, `${prefix}.attn.to_k.weight`);
835
+ const vWeight = getWeight(weights, shapes, `${prefix}.attn.to_v.weight`);
836
+ const qShape = getLinearShape(qWeight.shape, `${prefix}.attn.to_q.weight`);
837
+ const innerDim = qShape.outFeatures;
838
+ if (qShape.inFeatures !== state.channels || innerDim !== getLinearShape(kWeight.shape, `${prefix}.attn.to_k.weight`).outFeatures || innerDim !== getLinearShape(vWeight.shape, `${prefix}.attn.to_v.weight`).outFeatures) {
839
+ throw new Error(`AutoencoderDC attention "${prefix}" has incompatible q/k/v projection shapes.`);
840
+ }
841
+ if (!Number.isFinite(attentionHeadDim) || attentionHeadDim <= 0 || innerDim % attentionHeadDim !== 0) {
842
+ throw new Error(`AutoencoderDC attention "${prefix}" requires innerDim divisible by attentionHeadDim.`);
843
+ }
844
+ const numHeads = innerDim / attentionHeadDim;
845
+ const baseOptions = {
846
+ inChannels: state.channels,
847
+ outChannels: innerDim,
848
+ height: state.height,
849
+ width: state.width,
850
+ groups: 1,
851
+ };
852
+ const qBase = await ops.groupedPointwiseConv2d(state.tensor, qWeight.value, null, baseOptions);
853
+ const kBase = await ops.groupedPointwiseConv2d(state.tensor, kWeight.value, null, baseOptions);
854
+ const vBase = await ops.groupedPointwiseConv2d(state.tensor, vWeight.value, null, baseOptions);
855
+ const qVariants = [qBase];
856
+ const kVariants = [kBase];
857
+ const vVariants = [vBase];
858
+
859
+ if (Array.isArray(qkvMultiscales)) {
860
+ const qkvBase = await concatChannelTensors([qBase, kBase, vBase], state.height, state.width, recorder);
861
+ for (let scaleIdx = 0; scaleIdx < qkvMultiscales.length; scaleIdx++) {
862
+ const depthWeight = getWeight(weights, shapes, `${prefix}.attn.to_qkv_multiscale.${scaleIdx}.proj_in.weight`);
863
+ const pointWeight = getWeight(weights, shapes, `${prefix}.attn.to_qkv_multiscale.${scaleIdx}.proj_out.weight`);
864
+ const depthShape = getConvShape(depthWeight.shape);
865
+ const pointShape = getConvShape(pointWeight.shape);
866
+ const groups = pointShape.outChannels / pointShape.inChannels;
867
+ const depth = await ops.depthwiseConv2d(qkvBase, depthWeight.value, null, {
868
+ channels: qkvBase.shape[0],
869
+ height: state.height,
870
+ width: state.width,
871
+ kernelH: depthShape.kernelH,
872
+ kernelW: depthShape.kernelW,
873
+ stride: 1,
874
+ pad: Math.floor(depthShape.kernelH / 2),
875
+ });
876
+ const projected = await ops.groupedPointwiseConv2d(depth, pointWeight.value, null, {
877
+ inChannels: qkvBase.shape[0],
878
+ outChannels: pointShape.outChannels,
879
+ height: state.height,
880
+ width: state.width,
881
+ groups,
882
+ });
883
+ release(depth.buffer);
884
+
885
+ const qScale = await sliceChannelTensor(projected, 0, innerDim, state.height, state.width, recorder);
886
+ const kScale = await sliceChannelTensor(projected, innerDim, innerDim, state.height, state.width, recorder);
887
+ const vScale = await sliceChannelTensor(projected, innerDim * 2, innerDim, state.height, state.width, recorder);
888
+ release(projected.buffer);
889
+ qVariants.push(qScale);
890
+ kVariants.push(kScale);
891
+ vVariants.push(vScale);
892
+ }
893
+ release(qkvBase.buffer);
894
+ }
895
+
896
+ const qAll = await concatChannelTensors(qVariants, state.height, state.width, recorder);
897
+ const kAll = await concatChannelTensors(kVariants, state.height, state.width, recorder);
898
+ const vAll = await concatChannelTensors(vVariants, state.height, state.width, recorder);
899
+ for (const tensor of qVariants) release(tensor.buffer);
900
+ for (const tensor of kVariants) release(tensor.buffer);
901
+ for (const tensor of vVariants) release(tensor.buffer);
902
+
903
+ const qTokens = await channelsToTokens({ tensor: qAll, channels: qAll.shape[0], height: state.height, width: state.width }, ops);
904
+ const kTokens = await channelsToTokens({ tensor: kAll, channels: kAll.shape[0], height: state.height, width: state.width }, ops);
905
+ const vTokens = await channelsToTokens({ tensor: vAll, channels: vAll.shape[0], height: state.height, width: state.width }, ops);
906
+ release(qAll.buffer);
907
+ release(kAll.buffer);
908
+ release(vAll.buffer);
909
+
910
+ const qRelu = await ops.relu(qTokens.tensor, { count: tensorElementCount(qTokens.tensor) });
911
+ const kRelu = await ops.relu(kTokens.tensor, { count: tensorElementCount(kTokens.tensor) });
912
+ release(qTokens.tensor.buffer);
913
+ release(kTokens.tensor.buffer);
914
+
915
+ const allHeads = numHeads * qVariants.length;
916
+ const attention = await ops.sanaLinearAttention(qRelu, kRelu, vTokens.tensor, {
917
+ numHeads: allHeads,
918
+ headDim: attentionHeadDim,
919
+ numTokens: qTokens.numTokens,
920
+ hiddenSize: allHeads * attentionHeadDim,
921
+ eps,
922
+ });
923
+ release(qRelu.buffer);
924
+ release(kRelu.buffer);
925
+ release(vTokens.tensor.buffer);
926
+
927
+ const attended = await tokensToChannels(attention, allHeads * attentionHeadDim, state.height, state.width, ops);
928
+ release(attention.buffer);
929
+ const outWeight = getWeight(weights, shapes, `${prefix}.attn.to_out.weight`);
930
+ const outShape = getLinearShape(outWeight.shape, `${prefix}.attn.to_out.weight`);
931
+ const projected = await ops.groupedPointwiseConv2d(attended.tensor, outWeight.value, null, {
932
+ inChannels: attended.channels,
933
+ outChannels: outShape.outFeatures,
934
+ height: state.height,
935
+ width: state.width,
936
+ groups: 1,
937
+ });
938
+ release(attended.tensor.buffer);
939
+ const normed = await runChannelwiseRmsNorm(
940
+ {
941
+ tensor: reshapeTensor(projected, [outShape.outFeatures, state.height, state.width], 'vae_dc_attn_projected'),
942
+ channels: outShape.outFeatures,
943
+ height: state.height,
944
+ width: state.width,
945
+ },
946
+ getWeight(weights, shapes, `${prefix}.attn.norm_out.weight`),
947
+ getWeightOptional(weights, shapes, `${prefix}.attn.norm_out.bias`),
948
+ 1e-5,
949
+ ops,
950
+ release
951
+ );
952
+ release(projected.buffer);
953
+
954
+ const size = normed.channels * normed.height * normed.width;
955
+ const combined = await ops.residualAdd(
956
+ reshapeTensor(normed.tensor, [size], 'vae_dc_attn_main'),
957
+ reshapeTensor(state.tensor, [size], 'vae_dc_attn_residual'),
958
+ size,
959
+ { useVec4: true }
960
+ );
961
+ release(normed.tensor.buffer);
962
+ release(state.tensor.buffer);
963
+ return {
964
+ tensor: reshapeTensor(combined, [normed.channels, normed.height, normed.width], 'vae_dc_attn_out'),
965
+ channels: normed.channels,
966
+ height: normed.height,
967
+ width: normed.width,
968
+ };
969
+ }
970
+
971
+ async function runAutoencoderDCGlumbConv(state, weights, shapes, prefix, eps, ops, release) {
972
+ const invertedWeight = getWeight(weights, shapes, `${prefix}.conv_out.conv_inverted.weight`);
973
+ const invertedBias = getWeightOptional(weights, shapes, `${prefix}.conv_out.conv_inverted.bias`);
974
+ const invertedShape = getLinearShape(invertedWeight.shape, `${prefix}.conv_out.conv_inverted.weight`);
975
+ const hiddenChannels = Math.floor(invertedShape.outFeatures / 2);
976
+ const inverted = await ops.groupedPointwiseConv2d(state.tensor, invertedWeight.value, invertedBias?.value ?? null, {
977
+ inChannels: state.channels,
978
+ outChannels: invertedShape.outFeatures,
979
+ height: state.height,
980
+ width: state.width,
981
+ groups: 1,
982
+ });
983
+ const activated = await ops.silu(inverted, {
984
+ size: invertedShape.outFeatures * state.height * state.width,
985
+ swigluLimit: null,
986
+ });
987
+ release(inverted.buffer);
988
+ const depthWeight = getWeight(weights, shapes, `${prefix}.conv_out.conv_depth.weight`);
989
+ const depthBias = getWeightOptional(weights, shapes, `${prefix}.conv_out.conv_depth.bias`);
990
+ const depthShape = getConvShape(depthWeight.shape);
991
+ const depth = await ops.depthwiseConv2d(
992
+ reshapeTensor(activated, [invertedShape.outFeatures, state.height, state.width], 'vae_dc_glumb_act'),
993
+ depthWeight.value,
994
+ depthBias?.value ?? null,
995
+ {
996
+ channels: invertedShape.outFeatures,
997
+ height: state.height,
998
+ width: state.width,
999
+ kernelH: depthShape.kernelH,
1000
+ kernelW: depthShape.kernelW,
1001
+ stride: 1,
1002
+ pad: 1,
1003
+ }
1004
+ );
1005
+ release(activated.buffer);
1006
+ const depthTokens = await channelsToTokens({ tensor: depth, channels: invertedShape.outFeatures, height: state.height, width: state.width }, ops);
1007
+ release(depth.buffer);
1008
+ const gated = await ops.siluRowSplit(depthTokens.tensor, {
1009
+ numTokens: depthTokens.numTokens,
1010
+ dim: hiddenChannels,
1011
+ activation: 'silu',
1012
+ swigluLimit: null,
1013
+ });
1014
+ release(depthTokens.tensor.buffer);
1015
+ const gatedChannels = await tokensToChannels(gated, hiddenChannels, state.height, state.width, ops);
1016
+ release(gated.buffer);
1017
+
1018
+ const pointWeight = getWeight(weights, shapes, `${prefix}.conv_out.conv_point.weight`);
1019
+ const pointShape = getLinearShape(pointWeight.shape, `${prefix}.conv_out.conv_point.weight`);
1020
+ const projected = await ops.groupedPointwiseConv2d(gatedChannels.tensor, pointWeight.value, null, {
1021
+ inChannels: hiddenChannels,
1022
+ outChannels: pointShape.outFeatures,
1023
+ height: state.height,
1024
+ width: state.width,
1025
+ groups: 1,
1026
+ });
1027
+ release(gatedChannels.tensor.buffer);
1028
+ const normed = await runChannelwiseRmsNorm(
1029
+ {
1030
+ tensor: reshapeTensor(projected, [pointShape.outFeatures, state.height, state.width], 'vae_dc_glumb_projected'),
1031
+ channels: pointShape.outFeatures,
1032
+ height: state.height,
1033
+ width: state.width,
1034
+ },
1035
+ getWeight(weights, shapes, `${prefix}.conv_out.norm.weight`),
1036
+ getWeightOptional(weights, shapes, `${prefix}.conv_out.norm.bias`),
1037
+ eps,
1038
+ ops,
1039
+ release
1040
+ );
1041
+ release(projected.buffer);
1042
+
1043
+ const size = normed.channels * normed.height * normed.width;
1044
+ const combined = await ops.residualAdd(
1045
+ reshapeTensor(normed.tensor, [size], 'vae_dc_glumb_main'),
1046
+ reshapeTensor(state.tensor, [size], 'vae_dc_glumb_residual'),
1047
+ size,
1048
+ { useVec4: true }
1049
+ );
1050
+ release(normed.tensor.buffer);
1051
+ release(state.tensor.buffer);
1052
+ return {
1053
+ tensor: reshapeTensor(combined, [normed.channels, normed.height, normed.width], 'vae_dc_glumb_out'),
1054
+ channels: normed.channels,
1055
+ height: normed.height,
1056
+ width: normed.width,
1057
+ };
1058
+ }
1059
+
1060
+ async function runAutoencoderDCEfficientVitBlock(state, weights, shapes, prefix, attentionHeadDim, qkvMultiscales, eps, ops, release, recorder) {
1061
+ const attended = await runAutoencoderDCAttention(
1062
+ state,
1063
+ weights,
1064
+ shapes,
1065
+ prefix,
1066
+ attentionHeadDim,
1067
+ qkvMultiscales,
1068
+ 1e-15,
1069
+ ops,
1070
+ release,
1071
+ recorder
1072
+ );
1073
+ return runAutoencoderDCGlumbConv(attended, weights, shapes, prefix, eps, ops, release);
1074
+ }
1075
+
1076
+ async function decodeLatentsAutoencoderDC(state, config, weights, shapes, ops, release, recorder) {
1077
+ const blockTypes = normalizePerBlockValue(config.decoder_block_types, config.decoder_block_out_channels.length, 'decoder_block_types');
1078
+ const layersPerBlock = normalizePerBlockValue(config.decoder_layers_per_block, config.decoder_block_out_channels.length, 'decoder_layers_per_block');
1079
+ const qkvMultiscales = normalizePerBlockValue(config.decoder_qkv_multiscales, config.decoder_block_out_channels.length, 'decoder_qkv_multiscales');
1080
+ const normTypes = normalizePerBlockValue(config.decoder_norm_types, config.decoder_block_out_channels.length, 'decoder_norm_types');
1081
+ const actFns = normalizePerBlockValue(config.decoder_act_fns, config.decoder_block_out_channels.length, 'decoder_act_fns');
1082
+ const rmsNormEps = 1e-5;
1083
+
1084
+ state = await runAutoencoderDCInputProjection(state, weights, shapes, config, ops, release);
1085
+
1086
+ for (let blockIdx = blockTypes.length - 1; blockIdx >= 0; blockIdx--) {
1087
+ const prefix = `vae.decoder.up_blocks.${blockIdx}`;
1088
+ const hasUpsample = weights.has(`${prefix}.0.conv.weight`);
1089
+ if (hasUpsample) {
1090
+ state = await runAutoencoderDCUpBlock(state, weights, shapes, `${prefix}.0`, config, ops, release, recorder);
1091
+ }
1092
+
1093
+ if (normTypes[blockIdx] !== 'rms_norm') {
1094
+ throw new Error(
1095
+ `Unsupported AutoencoderDC norm type "${normTypes[blockIdx]}" in block ${blockIdx}.`
1096
+ );
1097
+ }
1098
+ if (actFns[blockIdx] !== 'silu') {
1099
+ throw new Error(
1100
+ `Unsupported AutoencoderDC activation "${actFns[blockIdx]}" in block ${blockIdx}.`
1101
+ );
1102
+ }
1103
+
1104
+ const startIndex = hasUpsample ? 1 : 0;
1105
+ const blockType = blockTypes[blockIdx];
1106
+ const numLayers = layersPerBlock[blockIdx];
1107
+ for (let layerOffset = 0; layerOffset < numLayers; layerOffset++) {
1108
+ const layerPrefix = `${prefix}.${startIndex + layerOffset}`;
1109
+ if (blockType === 'ResBlock') {
1110
+ state = await runAutoencoderDCResBlock(state, weights, shapes, layerPrefix, rmsNormEps, ops, release);
1111
+ continue;
1112
+ }
1113
+ if (blockType === 'EfficientViTBlock') {
1114
+ state = await runAutoencoderDCEfficientVitBlock(
1115
+ state,
1116
+ weights,
1117
+ shapes,
1118
+ layerPrefix,
1119
+ config.attention_head_dim,
1120
+ qkvMultiscales[blockIdx],
1121
+ rmsNormEps,
1122
+ ops,
1123
+ release,
1124
+ recorder
1125
+ );
1126
+ continue;
1127
+ }
1128
+ throw new Error(`Unsupported AutoencoderDC block type "${blockType}" in block ${blockIdx}.`);
1129
+ }
1130
+ }
1131
+
1132
+ const normed = await runChannelwiseRmsNorm(
1133
+ state,
1134
+ getWeight(weights, shapes, 'vae.decoder.norm_out.weight'),
1135
+ getWeightOptional(weights, shapes, 'vae.decoder.norm_out.bias'),
1136
+ rmsNormEps,
1137
+ ops,
1138
+ release
1139
+ );
1140
+ release(state.tensor.buffer);
1141
+ const activated = await ops.relu(normed.tensor, {
1142
+ count: normed.channels * normed.height * normed.width,
1143
+ });
1144
+ release(normed.tensor.buffer);
1145
+ return applyConv2D(
1146
+ {
1147
+ tensor: reshapeTensor(activated, [normed.channels, normed.height, normed.width], 'vae_dc_norm_out'),
1148
+ channels: normed.channels,
1149
+ height: normed.height,
1150
+ width: normed.width,
1151
+ },
1152
+ weights,
1153
+ shapes,
1154
+ 'vae.decoder.conv_out',
1155
+ { pad: 1 },
1156
+ ops,
1157
+ release
1158
+ );
1159
+ }
1160
+
1161
+ async function decodeLatentsGPU(latents, options) {
1162
+ const device = getDevice();
1163
+ if (!device) {
1164
+ throw new Error('VAE GPU decode requires a WebGPU device.');
1165
+ }
1166
+
1167
+ const profileTarget = options.profile ?? null;
1168
+ const wantsProfile = profileTarget === true || typeof profileTarget === 'object';
1169
+ const localRecorder = wantsProfile
1170
+ ? new CommandRecorder(device, 'vae_decode', { profile: true })
1171
+ : null;
1172
+ const recorder = localRecorder;
1173
+ const ops = createKernelOps(recorder);
1174
+ const release = createBufferReleaser(recorder);
1175
+
1176
+ const config = options.modelConfig?.components?.vae?.config || {};
1177
+ const runtime = options.runtime || {};
1178
+ const weightsEntry = options.weights;
1179
+
1180
+ if (!weightsEntry?.weights || !weightsEntry?.shapes) {
1181
+ throw new Error('VAE GPU decode requires loaded weights.');
1182
+ }
1183
+
1184
+ const weights = weightsEntry.weights;
1185
+ const shapes = weightsEntry.shapes;
1186
+
1187
+ const scalingFactor = config.scaling_factor;
1188
+ if (!Number.isFinite(scalingFactor) || scalingFactor === 0) {
1189
+ throw new Error('VAE decode requires a valid scaling_factor in config.');
1190
+ }
1191
+ const shiftFactor = Number.isFinite(config.shift_factor) ? config.shift_factor : 0.0;
1192
+ const isAutoencoderDC = config._class_name === 'AutoencoderDC' || Array.isArray(config.decoder_block_types);
1193
+
1194
+ const scaledLatents = new Float32Array(latents.length);
1195
+ for (let i = 0; i < latents.length; i++) {
1196
+ scaledLatents[i] = latents[i] / scalingFactor + shiftFactor;
1197
+ }
1198
+
1199
+ const latentBuffer = acquireBuffer(scaledLatents.byteLength, undefined, 'vae_latents');
1200
+ device.queue.writeBuffer(latentBuffer, 0, scaledLatents);
1201
+
1202
+ let state = {
1203
+ tensor: createTensor(latentBuffer, 'f32', [options.latentChannels, options.latentHeight, options.latentWidth], 'vae_latents_f32'),
1204
+ channels: options.latentChannels,
1205
+ height: options.latentHeight,
1206
+ width: options.latentWidth,
1207
+ };
1208
+
1209
+ const computeDtype = runtime.latent?.dtype;
1210
+ if (!computeDtype) {
1211
+ throw new Error('VAE decode requires runtime.latent.dtype.');
1212
+ }
1213
+ if (computeDtype !== 'f16') {
1214
+ throw new Error(
1215
+ `VAE GPU decode requires runtime.latent.dtype="f16"; got "${computeDtype}".`
1216
+ );
1217
+ }
1218
+ const casted = await ops.castF32ToF16(state.tensor);
1219
+ release(state.tensor.buffer);
1220
+ state = {
1221
+ tensor: reshapeTensor(casted, [state.channels, state.height, state.width], 'vae_latents_f16'),
1222
+ channels: state.channels,
1223
+ height: state.height,
1224
+ width: state.width,
1225
+ };
1226
+
1227
+ if (isAutoencoderDC) {
1228
+ state = await decodeLatentsAutoencoderDC(state, config, weights, shapes, ops, release, recorder);
1229
+ } else {
1230
+ const numGroups = config.norm_num_groups;
1231
+ if (!Number.isFinite(numGroups) || numGroups <= 0) {
1232
+ throw new Error('VAE decode requires norm_num_groups in config.');
1233
+ }
1234
+ const eps = runtime.decode?.groupNormEps;
1235
+ if (!Number.isFinite(eps)) {
1236
+ throw new Error('VAE decode requires runtime.decode.groupNormEps.');
1237
+ }
1238
+
1239
+ state = await applyConv2D(state, weights, shapes, 'vae.decoder.conv_in', { pad: 1 }, ops, release);
1240
+
1241
+ const midResnetPrefix = 'vae.decoder.mid_block.resnets.';
1242
+ const midResnetIds = buildIndexList(weights, midResnetPrefix);
1243
+ for (const idx of midResnetIds) {
1244
+ state = await runResnetBlock(state, weights, shapes, `${midResnetPrefix}${idx}`, { numGroups, eps }, ops, release);
1245
+ }
1246
+
1247
+ const midAttentionPrefix = 'vae.decoder.mid_block.attentions.';
1248
+ const midAttentionIds = buildIndexList(weights, midAttentionPrefix);
1249
+ for (const idx of midAttentionIds) {
1250
+ state = await runMidBlockAttention(
1251
+ state,
1252
+ weights,
1253
+ shapes,
1254
+ `${midAttentionPrefix}${idx}`,
1255
+ {
1256
+ numGroups,
1257
+ eps,
1258
+ modelConfig: config,
1259
+ },
1260
+ ops,
1261
+ release
1262
+ );
1263
+ }
1264
+
1265
+ const upBlockPrefix = 'vae.decoder.up_blocks.';
1266
+ const upBlocks = buildIndexList(weights, upBlockPrefix);
1267
+ for (const blockIdx of upBlocks) {
1268
+ const resnetPrefix = `${upBlockPrefix}${blockIdx}.resnets.`;
1269
+ const resnetIds = buildIndexList(weights, resnetPrefix);
1270
+ for (const idx of resnetIds) {
1271
+ state = await runResnetBlock(state, weights, shapes, `${resnetPrefix}${idx}`, { numGroups, eps }, ops, release);
1272
+ }
1273
+
1274
+ const upsampleWeightName = `${upBlockPrefix}${blockIdx}.upsamplers.0.conv.weight`;
1275
+ if (weights.has(upsampleWeightName)) {
1276
+ const upsample = await ops.upsample2d(state.tensor, {
1277
+ channels: state.channels,
1278
+ height: state.height,
1279
+ width: state.width,
1280
+ scale: 2,
1281
+ });
1282
+ release(state.tensor.buffer);
1283
+ state = {
1284
+ tensor: reshapeTensor(upsample, [state.channels, state.height * 2, state.width * 2], 'vae_upsample'),
1285
+ channels: state.channels,
1286
+ height: state.height * 2,
1287
+ width: state.width * 2,
1288
+ };
1289
+
1290
+ state = await applyConv2D(state, weights, shapes, `${upBlockPrefix}${blockIdx}.upsamplers.0.conv`, { pad: 1 }, ops, release);
1291
+ }
1292
+ }
1293
+
1294
+ const normOut = getWeight(weights, shapes, 'vae.decoder.conv_norm_out.weight');
1295
+ const normOutBias = getWeight(weights, shapes, 'vae.decoder.conv_norm_out.bias');
1296
+ const normed = await ops.groupNorm(state.tensor, normOut.value, normOutBias.value, {
1297
+ channels: state.channels,
1298
+ height: state.height,
1299
+ width: state.width,
1300
+ numGroups,
1301
+ eps,
1302
+ });
1303
+ release(state.tensor.buffer);
1304
+
1305
+ const siluOut = await ops.silu(normed, { size: state.channels * state.height * state.width, swigluLimit: null });
1306
+ release(normed.buffer);
1307
+ state = {
1308
+ tensor: reshapeTensor(siluOut, [state.channels, state.height, state.width], 'vae_norm_out'),
1309
+ channels: state.channels,
1310
+ height: state.height,
1311
+ width: state.width,
1312
+ };
1313
+
1314
+ state = await applyConv2D(state, weights, shapes, 'vae.decoder.conv_out', { pad: 1 }, ops, release);
1315
+ }
1316
+
1317
+ const outputSize = state.channels * state.height * state.width * dtypeBytes(state.tensor.dtype);
1318
+ if (localRecorder) {
1319
+ localRecorder.submit();
1320
+ }
1321
+ const outputRaw = await readBuffer(state.tensor.buffer, outputSize);
1322
+ releaseBuffer(state.tensor.buffer);
1323
+
1324
+ if (localRecorder) {
1325
+ const timings = await localRecorder.resolveProfileTimings();
1326
+ if (profileTarget && typeof profileTarget === 'object') {
1327
+ profileTarget.totalMs = sumProfileTimings(timings) ?? null;
1328
+ profileTarget.timings = timings ?? null;
1329
+ }
1330
+ }
1331
+
1332
+ const output = state.tensor.dtype === 'f16'
1333
+ ? new Uint16Array(outputRaw)
1334
+ : new Float32Array(outputRaw);
1335
+
1336
+ const outHeight = state.height;
1337
+ const outWidth = state.width;
1338
+ if (outHeight !== options.height || outWidth !== options.width) {
1339
+ log.warn('Diffusion', `VAE output size ${outWidth}x${outHeight} differs from request ${options.width}x${options.height}.`);
1340
+ }
1341
+ const pixels = new Uint8ClampedArray(outWidth * outHeight * 4);
1342
+ const height = outHeight;
1343
+ const width = outWidth;
1344
+
1345
+ for (let y = 0; y < height; y++) {
1346
+ for (let x = 0; x < width; x++) {
1347
+ const outIndex = (y * width + x) * 4;
1348
+ const base = (y * width + x);
1349
+ const rIdx = base;
1350
+ const gIdx = base + height * width;
1351
+ const bIdx = base + 2 * height * width;
1352
+
1353
+ const r = state.tensor.dtype === 'f16' ? f16ToF32(output[rIdx]) : output[rIdx];
1354
+ const g = state.tensor.dtype === 'f16' ? f16ToF32(output[gIdx]) : output[gIdx];
1355
+ const b = state.tensor.dtype === 'f16' ? f16ToF32(output[bIdx]) : output[bIdx];
1356
+
1357
+ pixels[outIndex] = clamp(Math.round((r * 0.5 + 0.5) * 255), 0, 255);
1358
+ pixels[outIndex + 1] = clamp(Math.round((g * 0.5 + 0.5) * 255), 0, 255);
1359
+ pixels[outIndex + 2] = clamp(Math.round((b * 0.5 + 0.5) * 255), 0, 255);
1360
+ pixels[outIndex + 3] = 255;
1361
+ }
1362
+ }
1363
+
1364
+ return pixels;
1365
+ }
1366
+
1367
+ export async function decodeLatents(latents, options) {
1368
+ if (!options?.weights || !getDevice()) {
1369
+ throw new Error(
1370
+ 'Diffusion decode requires GPU VAE weights and a WebGPU device. ' +
1371
+ 'CPU decode fallback is unsupported.'
1372
+ );
1373
+ }
1374
+ return decodeLatentsGPU(latents, options);
1375
+ }