@simulatte/doppler 0.1.0

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 (1182) hide show
  1. package/BRANDING.md +14 -0
  2. package/LICENSE +201 -0
  3. package/NOTICE +5 -0
  4. package/README.md +85 -0
  5. package/SECURITY.md +19 -0
  6. package/package.json +144 -0
  7. package/src/adapters/adapter-manager.d.ts +200 -0
  8. package/src/adapters/adapter-manager.js +509 -0
  9. package/src/adapters/adapter-manifest.d.ts +290 -0
  10. package/src/adapters/adapter-manifest.js +320 -0
  11. package/src/adapters/adapter-registry.d.ts +192 -0
  12. package/src/adapters/adapter-registry.js +466 -0
  13. package/src/adapters/index.d.ts +89 -0
  14. package/src/adapters/index.js +42 -0
  15. package/src/adapters/lora-loader.d.ts +105 -0
  16. package/src/adapters/lora-loader.js +397 -0
  17. package/src/bootstrap.d.ts +1 -0
  18. package/src/bootstrap.js +30 -0
  19. package/src/bridge/extension/background.d.ts +14 -0
  20. package/src/bridge/extension/background.js +168 -0
  21. package/src/bridge/extension/manifest.json +34 -0
  22. package/src/bridge/extension-client.d.ts +109 -0
  23. package/src/bridge/extension-client.js +369 -0
  24. package/src/bridge/index.d.ts +68 -0
  25. package/src/bridge/index.js +51 -0
  26. package/src/bridge/protocol.d.ts +96 -0
  27. package/src/bridge/protocol.js +130 -0
  28. package/src/browser/browser-converter.d.ts +71 -0
  29. package/src/browser/browser-converter.js +947 -0
  30. package/src/browser/file-picker.d.ts +63 -0
  31. package/src/browser/file-picker.js +275 -0
  32. package/src/browser/gguf-importer.d.ts +136 -0
  33. package/src/browser/gguf-importer.js +532 -0
  34. package/src/browser/gguf-parser-browser.d.ts +14 -0
  35. package/src/browser/gguf-parser-browser.js +17 -0
  36. package/src/browser/quantization.d.ts +69 -0
  37. package/src/browser/quantization.js +328 -0
  38. package/src/browser/safetensors-parser-browser.d.ts +193 -0
  39. package/src/browser/safetensors-parser-browser.js +264 -0
  40. package/src/browser/shard-io-browser.d.ts +57 -0
  41. package/src/browser/shard-io-browser.js +89 -0
  42. package/src/browser/tensor-source-download.d.ts +27 -0
  43. package/src/browser/tensor-source-download.js +239 -0
  44. package/src/browser/tensor-source-file.d.ts +26 -0
  45. package/src/browser/tensor-source-file.js +53 -0
  46. package/src/browser/tensor-source-http.d.ts +28 -0
  47. package/src/browser/tensor-source-http.js +126 -0
  48. package/src/client/doppler-provider/generation.d.ts +25 -0
  49. package/src/client/doppler-provider/generation.js +114 -0
  50. package/src/client/doppler-provider/index.d.ts +2 -0
  51. package/src/client/doppler-provider/index.js +3 -0
  52. package/src/client/doppler-provider/model-manager.d.ts +61 -0
  53. package/src/client/doppler-provider/model-manager.js +667 -0
  54. package/src/client/doppler-provider/provider.d.ts +5 -0
  55. package/src/client/doppler-provider/provider.js +102 -0
  56. package/src/client/doppler-provider/source-runtime.d.ts +22 -0
  57. package/src/client/doppler-provider/source-runtime.js +522 -0
  58. package/src/client/doppler-provider/types.d.ts +127 -0
  59. package/src/client/doppler-provider/types.js +17 -0
  60. package/src/client/doppler-provider.d.ts +46 -0
  61. package/src/client/doppler-provider.js +36 -0
  62. package/src/config/README.md +69 -0
  63. package/src/config/backward-registry-loader.d.ts +3 -0
  64. package/src/config/backward-registry-loader.js +8 -0
  65. package/src/config/index.d.ts +63 -0
  66. package/src/config/index.js +31 -0
  67. package/src/config/kernel-path-loader.d.ts +149 -0
  68. package/src/config/kernel-path-loader.js +534 -0
  69. package/src/config/kernels/backward-registry.json +99 -0
  70. package/src/config/kernels/kernel-ref-digests.d.ts +1 -0
  71. package/src/config/kernels/kernel-ref-digests.js +214 -0
  72. package/src/config/kernels/kernel-ref.d.ts +17 -0
  73. package/src/config/kernels/kernel-ref.js +75 -0
  74. package/src/config/kernels/moe/gpt-oss.paths.json +49 -0
  75. package/src/config/kernels/registry.d.ts +86 -0
  76. package/src/config/kernels/registry.js +103 -0
  77. package/src/config/kernels/registry.json +6771 -0
  78. package/src/config/loader.d.ts +57 -0
  79. package/src/config/loader.js +513 -0
  80. package/src/config/merge.d.ts +142 -0
  81. package/src/config/merge.js +389 -0
  82. package/src/config/param-categories.d.ts +17 -0
  83. package/src/config/param-categories.js +72 -0
  84. package/src/config/param-validator.d.ts +26 -0
  85. package/src/config/param-validator.js +235 -0
  86. package/src/config/platforms/amd-rdna3.json +16 -0
  87. package/src/config/platforms/apple-m1.json +16 -0
  88. package/src/config/platforms/apple-m2.json +16 -0
  89. package/src/config/platforms/apple-m3.json +16 -0
  90. package/src/config/platforms/generic.json +14 -0
  91. package/src/config/platforms/loader.d.ts +65 -0
  92. package/src/config/platforms/loader.js +153 -0
  93. package/src/config/platforms/nvidia-rtx30.json +16 -0
  94. package/src/config/platforms/nvidia-rtx40.json +16 -0
  95. package/src/config/presets/kernel-paths/embeddinggemma-f16-f32a.json +60 -0
  96. package/src/config/presets/kernel-paths/embeddinggemma-f32-f32a.json +60 -0
  97. package/src/config/presets/kernel-paths/embeddinggemma-q4k-dequant-f32a.json +60 -0
  98. package/src/config/presets/kernel-paths/gemma2-f16-f16a.json +61 -0
  99. package/src/config/presets/kernel-paths/gemma2-f16-f32a.json +60 -0
  100. package/src/config/presets/kernel-paths/gemma2-q4k-dequant-f16a.json +61 -0
  101. package/src/config/presets/kernel-paths/gemma2-q4k-dequant-f32a.json +60 -0
  102. package/src/config/presets/kernel-paths/gemma2-q4k-fused-f32a.json +57 -0
  103. package/src/config/presets/kernel-paths/gemma3-f16-fused-f16a-online.json +200 -0
  104. package/src/config/presets/kernel-paths/gemma3-f16-fused-f32a-online.json +223 -0
  105. package/src/config/presets/kernel-paths/gemma3-q4k-dequant-f16a-online.json +60 -0
  106. package/src/config/presets/kernel-paths/gemma3-q4k-dequant-f32a-online.json +61 -0
  107. package/src/config/presets/kernel-paths/gemma3-q4k-dequant-f32a.json +61 -0
  108. package/src/config/presets/kernel-paths/lfm2-q4k-dequant-f32a-online.json +61 -0
  109. package/src/config/presets/kernel-paths/registry.json +103 -0
  110. package/src/config/presets/models/deepseek.json +20 -0
  111. package/src/config/presets/models/diffusion.json +10 -0
  112. package/src/config/presets/models/embeddinggemma.json +74 -0
  113. package/src/config/presets/models/functiongemma.json +31 -0
  114. package/src/config/presets/models/gemma2.json +59 -0
  115. package/src/config/presets/models/gemma3.json +75 -0
  116. package/src/config/presets/models/gpt-oss.json +68 -0
  117. package/src/config/presets/models/kimi-k2.json +25 -0
  118. package/src/config/presets/models/lfm2.json +83 -0
  119. package/src/config/presets/models/llama3.json +40 -0
  120. package/src/config/presets/models/mamba.json +34 -0
  121. package/src/config/presets/models/mixtral.json +37 -0
  122. package/src/config/presets/models/modernbert.json +32 -0
  123. package/src/config/presets/models/qwen3.json +41 -0
  124. package/src/config/presets/models/transformer.json +73 -0
  125. package/src/config/presets/models/translategemma.json +30 -0
  126. package/src/config/presets/platforms/nvidia-gb200-8gpu.json +45 -0
  127. package/src/config/presets/platforms/nvidia-gb200-nvl72.json +45 -0
  128. package/src/config/presets/platforms/nvidia-gh200-nvl2.json +44 -0
  129. package/src/config/presets/platforms/nvidia-gh200.json +44 -0
  130. package/src/config/presets/runtime/compute/f16-activations.json +30 -0
  131. package/src/config/presets/runtime/compute/f16-batched.json +32 -0
  132. package/src/config/presets/runtime/default.json +101 -0
  133. package/src/config/presets/runtime/diagnostics/debug-logits.json +53 -0
  134. package/src/config/presets/runtime/experiments/bench/gemma3-bench-q4k.json +53 -0
  135. package/src/config/presets/runtime/experiments/debug/gemma3-debug-q4k.json +210 -0
  136. package/src/config/presets/runtime/experiments/verify/gemma3-verify.json +39 -0
  137. package/src/config/presets/runtime/kernels/dequant-f16-q4k.json +20 -0
  138. package/src/config/presets/runtime/kernels/dequant-f32-q4k.json +20 -0
  139. package/src/config/presets/runtime/kernels/fused-q4k.json +20 -0
  140. package/src/config/presets/runtime/kernels/safe-q4k.json +20 -0
  141. package/src/config/presets/runtime/model/gemma2-debug.json +77 -0
  142. package/src/config/presets/runtime/model/gemma2-pipeline-debug.json +66 -0
  143. package/src/config/presets/runtime/model/gemma2-pipeline.json +75 -0
  144. package/src/config/presets/runtime/model/gemma3-layer-probe.json +85 -0
  145. package/src/config/presets/runtime/modes/bench.json +37 -0
  146. package/src/config/presets/runtime/modes/debug.json +39 -0
  147. package/src/config/presets/runtime/modes/default.json +10 -0
  148. package/src/config/presets/runtime/modes/embedding-bench.json +28 -0
  149. package/src/config/presets/runtime/modes/embedding.json +54 -0
  150. package/src/config/presets/runtime/modes/low-memory.json +40 -0
  151. package/src/config/presets/runtime/modes/production.json +48 -0
  152. package/src/config/presets/runtime/modes/simulation.json +30 -0
  153. package/src/config/presets/runtime/modes/trace-layers.json +126 -0
  154. package/src/config/presets/runtime/platform/metal-apple-q4k.json +11 -0
  155. package/src/config/runtime-merge.d.ts +5 -0
  156. package/src/config/runtime-merge.js +21 -0
  157. package/src/config/runtime.d.ts +28 -0
  158. package/src/config/runtime.js +56 -0
  159. package/src/config/schema/adapter.schema.d.ts +53 -0
  160. package/src/config/schema/adapter.schema.js +60 -0
  161. package/src/config/schema/backward-registry.schema.d.ts +14 -0
  162. package/src/config/schema/backward-registry.schema.js +46 -0
  163. package/src/config/schema/benchmark.schema.d.ts +54 -0
  164. package/src/config/schema/benchmark.schema.js +74 -0
  165. package/src/config/schema/bridge.schema.d.ts +25 -0
  166. package/src/config/schema/bridge.schema.js +22 -0
  167. package/src/config/schema/buffer-pool.schema.d.ts +92 -0
  168. package/src/config/schema/buffer-pool.schema.js +50 -0
  169. package/src/config/schema/conversion.schema.d.ts +183 -0
  170. package/src/config/schema/conversion.schema.js +13 -0
  171. package/src/config/schema/converter.schema.d.ts +123 -0
  172. package/src/config/schema/converter.schema.js +136 -0
  173. package/src/config/schema/debug.schema.d.ts +245 -0
  174. package/src/config/schema/debug.schema.js +106 -0
  175. package/src/config/schema/diffusion.schema.d.ts +88 -0
  176. package/src/config/schema/diffusion.schema.js +62 -0
  177. package/src/config/schema/distill-training.schema.d.ts +48 -0
  178. package/src/config/schema/distill-training.schema.js +139 -0
  179. package/src/config/schema/distribution.schema.d.ts +155 -0
  180. package/src/config/schema/distribution.schema.js +81 -0
  181. package/src/config/schema/doppler.schema.d.ts +75 -0
  182. package/src/config/schema/doppler.schema.js +352 -0
  183. package/src/config/schema/ecosystem.schema.d.ts +255 -0
  184. package/src/config/schema/ecosystem.schema.js +534 -0
  185. package/src/config/schema/emulation.schema.d.ts +351 -0
  186. package/src/config/schema/emulation.schema.js +299 -0
  187. package/src/config/schema/energy.schema.d.ts +102 -0
  188. package/src/config/schema/energy.schema.js +72 -0
  189. package/src/config/schema/execution-v0.schema.d.ts +187 -0
  190. package/src/config/schema/execution-v0.schema.js +55 -0
  191. package/src/config/schema/gpu-cache.schema.d.ts +26 -0
  192. package/src/config/schema/gpu-cache.schema.js +8 -0
  193. package/src/config/schema/harness.schema.d.ts +32 -0
  194. package/src/config/schema/harness.schema.js +20 -0
  195. package/src/config/schema/hotswap.schema.d.ts +55 -0
  196. package/src/config/schema/hotswap.schema.js +18 -0
  197. package/src/config/schema/index.d.ts +863 -0
  198. package/src/config/schema/index.js +471 -0
  199. package/src/config/schema/inference-defaults.schema.d.ts +276 -0
  200. package/src/config/schema/inference-defaults.schema.js +185 -0
  201. package/src/config/schema/inference.schema.d.ts +289 -0
  202. package/src/config/schema/inference.schema.js +39 -0
  203. package/src/config/schema/intent-bundle.schema.d.ts +28 -0
  204. package/src/config/schema/intent-bundle.schema.js +12 -0
  205. package/src/config/schema/kernel-path.schema.d.ts +173 -0
  206. package/src/config/schema/kernel-path.schema.js +9 -0
  207. package/src/config/schema/kernel-registry.schema.d.ts +199 -0
  208. package/src/config/schema/kernel-registry.schema.js +46 -0
  209. package/src/config/schema/kernel-thresholds.schema.d.ts +302 -0
  210. package/src/config/schema/kernel-thresholds.schema.js +187 -0
  211. package/src/config/schema/kernel-warmup.schema.d.ts +19 -0
  212. package/src/config/schema/kernel-warmup.schema.js +5 -0
  213. package/src/config/schema/kvcache.schema.d.ts +131 -0
  214. package/src/config/schema/kvcache.schema.js +31 -0
  215. package/src/config/schema/loading.schema.d.ts +153 -0
  216. package/src/config/schema/loading.schema.js +84 -0
  217. package/src/config/schema/lora.schema.d.ts +12 -0
  218. package/src/config/schema/lora.schema.js +12 -0
  219. package/src/config/schema/manifest.schema.d.ts +500 -0
  220. package/src/config/schema/manifest.schema.js +130 -0
  221. package/src/config/schema/memory-limits.schema.d.ts +107 -0
  222. package/src/config/schema/memory-limits.schema.js +57 -0
  223. package/src/config/schema/moe.schema.d.ts +78 -0
  224. package/src/config/schema/moe.schema.js +31 -0
  225. package/src/config/schema/platform.schema.d.ts +121 -0
  226. package/src/config/schema/platform.schema.js +1 -0
  227. package/src/config/schema/preset.schema.d.ts +124 -0
  228. package/src/config/schema/preset.schema.js +1 -0
  229. package/src/config/schema/quantization-defaults.schema.d.ts +34 -0
  230. package/src/config/schema/quantization-defaults.schema.js +5 -0
  231. package/src/config/schema/quantization.schema.d.ts +10 -0
  232. package/src/config/schema/quantization.schema.js +33 -0
  233. package/src/config/schema/shared-runtime.schema.d.ts +75 -0
  234. package/src/config/schema/shared-runtime.schema.js +45 -0
  235. package/src/config/schema/speculative.schema.d.ts +21 -0
  236. package/src/config/schema/speculative.schema.js +11 -0
  237. package/src/config/schema/storage.schema.d.ts +123 -0
  238. package/src/config/schema/storage.schema.js +66 -0
  239. package/src/config/schema/tooling.schema.d.ts +29 -0
  240. package/src/config/schema/tooling.schema.js +12 -0
  241. package/src/config/schema/training-metrics.schema.d.ts +89 -0
  242. package/src/config/schema/training-metrics.schema.js +374 -0
  243. package/src/config/schema/training.schema.d.ts +88 -0
  244. package/src/config/schema/training.schema.js +106 -0
  245. package/src/config/schema/tuner.schema.d.ts +39 -0
  246. package/src/config/schema/tuner.schema.js +13 -0
  247. package/src/config/schema/ul-training.schema.d.ts +61 -0
  248. package/src/config/schema/ul-training.schema.js +140 -0
  249. package/src/config/schema/units.schema.d.ts +27 -0
  250. package/src/config/schema/units.schema.js +26 -0
  251. package/src/config/training-defaults.d.ts +24 -0
  252. package/src/config/training-defaults.js +91 -0
  253. package/src/converter/conversion-plan.d.ts +64 -0
  254. package/src/converter/conversion-plan.js +472 -0
  255. package/src/converter/core.d.ts +247 -0
  256. package/src/converter/core.js +1329 -0
  257. package/src/converter/execution-v0-manifest.d.ts +15 -0
  258. package/src/converter/execution-v0-manifest.js +146 -0
  259. package/src/converter/index.d.ts +98 -0
  260. package/src/converter/index.js +59 -0
  261. package/src/converter/manifest-inference.d.ts +20 -0
  262. package/src/converter/manifest-inference.js +492 -0
  263. package/src/converter/parsers/diffusion.d.ts +50 -0
  264. package/src/converter/parsers/diffusion.js +270 -0
  265. package/src/converter/parsers/gguf.d.ts +22 -0
  266. package/src/converter/parsers/gguf.js +46 -0
  267. package/src/converter/parsers/index.d.ts +21 -0
  268. package/src/converter/parsers/index.js +12 -0
  269. package/src/converter/parsers/transformer.d.ts +16 -0
  270. package/src/converter/parsers/transformer.js +25 -0
  271. package/src/converter/quantization-info.d.ts +37 -0
  272. package/src/converter/quantization-info.js +398 -0
  273. package/src/converter/quantizer.d.ts +96 -0
  274. package/src/converter/quantizer.js +422 -0
  275. package/src/converter/rope-config.d.ts +15 -0
  276. package/src/converter/rope-config.js +218 -0
  277. package/src/converter/shard-packer.d.ts +138 -0
  278. package/src/converter/shard-packer.js +422 -0
  279. package/src/converter/tokenizer-utils.d.ts +11 -0
  280. package/src/converter/tokenizer-utils.js +87 -0
  281. package/src/debug/config.d.ts +78 -0
  282. package/src/debug/config.js +235 -0
  283. package/src/debug/history.d.ts +65 -0
  284. package/src/debug/history.js +71 -0
  285. package/src/debug/index.d.ts +268 -0
  286. package/src/debug/index.js +192 -0
  287. package/src/debug/log.d.ts +46 -0
  288. package/src/debug/log.js +132 -0
  289. package/src/debug/perf.d.ts +33 -0
  290. package/src/debug/perf.js +51 -0
  291. package/src/debug/reference/README.md +114 -0
  292. package/src/debug/reference/hf_attn_debug.py +114 -0
  293. package/src/debug/reference/hf_embed_check.py +89 -0
  294. package/src/debug/reference/hf_layer_out.py +100 -0
  295. package/src/debug/reference/hf_rope_check.py +116 -0
  296. package/src/debug/reference/hf_weights.py +75 -0
  297. package/src/debug/signals.d.ts +63 -0
  298. package/src/debug/signals.js +33 -0
  299. package/src/debug/stats.d.ts +47 -0
  300. package/src/debug/stats.js +160 -0
  301. package/src/debug/tensor.d.ts +123 -0
  302. package/src/debug/tensor.js +257 -0
  303. package/src/debug/trace.d.ts +17 -0
  304. package/src/debug/trace.js +167 -0
  305. package/src/diffusion/image-regression.d.ts +31 -0
  306. package/src/diffusion/image-regression.js +107 -0
  307. package/src/diffusion/index.d.ts +8 -0
  308. package/src/diffusion/index.js +8 -0
  309. package/src/distribution/p2p-control-plane.d.ts +52 -0
  310. package/src/distribution/p2p-control-plane.js +232 -0
  311. package/src/distribution/p2p-observability.d.ts +116 -0
  312. package/src/distribution/p2p-observability.js +267 -0
  313. package/src/distribution/p2p-transport-contract.d.ts +57 -0
  314. package/src/distribution/p2p-transport-contract.js +310 -0
  315. package/src/distribution/p2p-webrtc-browser.d.ts +37 -0
  316. package/src/distribution/p2p-webrtc-browser.js +434 -0
  317. package/src/distribution/shard-delivery.d.ts +251 -0
  318. package/src/distribution/shard-delivery.js +2096 -0
  319. package/src/energy/index.d.ts +2 -0
  320. package/src/energy/index.js +2 -0
  321. package/src/errors/doppler-error.d.ts +21 -0
  322. package/src/errors/doppler-error.js +25 -0
  323. package/src/errors/index.d.ts +1 -0
  324. package/src/errors/index.js +1 -0
  325. package/src/formats/gguf/index.d.ts +8 -0
  326. package/src/formats/gguf/index.js +4 -0
  327. package/src/formats/gguf/types.d.ts +137 -0
  328. package/src/formats/gguf/types.js +443 -0
  329. package/src/formats/index.d.ts +51 -0
  330. package/src/formats/index.js +13 -0
  331. package/src/formats/rdrr/classification.d.ts +39 -0
  332. package/src/formats/rdrr/classification.js +275 -0
  333. package/src/formats/rdrr/groups.d.ts +27 -0
  334. package/src/formats/rdrr/groups.js +76 -0
  335. package/src/formats/rdrr/index.d.ts +25 -0
  336. package/src/formats/rdrr/index.js +19 -0
  337. package/src/formats/rdrr/manifest.d.ts +32 -0
  338. package/src/formats/rdrr/manifest.js +108 -0
  339. package/src/formats/rdrr/parsing.d.ts +23 -0
  340. package/src/formats/rdrr/parsing.js +101 -0
  341. package/src/formats/rdrr/tensor-config-validator.d.ts +42 -0
  342. package/src/formats/rdrr/tensor-config-validator.js +156 -0
  343. package/src/formats/rdrr/types.d.ts +200 -0
  344. package/src/formats/rdrr/types.js +16 -0
  345. package/src/formats/rdrr/validation.d.ts +9 -0
  346. package/src/formats/rdrr/validation.js +200 -0
  347. package/src/formats/safetensors/index.d.ts +8 -0
  348. package/src/formats/safetensors/index.js +4 -0
  349. package/src/formats/safetensors/types.d.ts +67 -0
  350. package/src/formats/safetensors/types.js +102 -0
  351. package/src/formats/tokenizer/index.d.ts +5 -0
  352. package/src/formats/tokenizer/index.js +3 -0
  353. package/src/formats/tokenizer/types.d.ts +9 -0
  354. package/src/formats/tokenizer/types.js +22 -0
  355. package/src/generation/index.d.ts +18 -0
  356. package/src/generation/index.js +12 -0
  357. package/src/gpu/command-recorder.d.ts +175 -0
  358. package/src/gpu/command-recorder.js +473 -0
  359. package/src/gpu/device.d.ts +141 -0
  360. package/src/gpu/device.js +350 -0
  361. package/src/gpu/kernel-runtime.d.ts +20 -0
  362. package/src/gpu/kernel-runtime.js +37 -0
  363. package/src/gpu/kernel-selection-cache.d.ts +13 -0
  364. package/src/gpu/kernel-selection-cache.js +13 -0
  365. package/src/gpu/kernel-selection-log.d.ts +12 -0
  366. package/src/gpu/kernel-selection-log.js +28 -0
  367. package/src/gpu/kernel-selector.d.ts +11 -0
  368. package/src/gpu/kernel-selector.js +10 -0
  369. package/src/gpu/kernel-tuner/benchmarks.d.ts +144 -0
  370. package/src/gpu/kernel-tuner/benchmarks.js +892 -0
  371. package/src/gpu/kernel-tuner/cache.d.ts +55 -0
  372. package/src/gpu/kernel-tuner/cache.js +66 -0
  373. package/src/gpu/kernel-tuner/index.d.ts +59 -0
  374. package/src/gpu/kernel-tuner/index.js +38 -0
  375. package/src/gpu/kernel-tuner/tuner.d.ts +82 -0
  376. package/src/gpu/kernel-tuner/tuner.js +229 -0
  377. package/src/gpu/kernel-tuner/types.d.ts +101 -0
  378. package/src/gpu/kernel-tuner/types.js +4 -0
  379. package/src/gpu/kernel-tuner.d.ts +33 -0
  380. package/src/gpu/kernel-tuner.js +12 -0
  381. package/src/gpu/kernels/README.md +127 -0
  382. package/src/gpu/kernels/attention.d.ts +236 -0
  383. package/src/gpu/kernels/attention.js +1359 -0
  384. package/src/gpu/kernels/attention.wgsl +249 -0
  385. package/src/gpu/kernels/attention_bdpa_decode_f16.wgsl +246 -0
  386. package/src/gpu/kernels/attention_decode.wgsl +233 -0
  387. package/src/gpu/kernels/attention_decode_chunked_f16.wgsl +183 -0
  388. package/src/gpu/kernels/attention_decode_chunked_f16kv.wgsl +208 -0
  389. package/src/gpu/kernels/attention_decode_f16.wgsl +202 -0
  390. package/src/gpu/kernels/attention_decode_f16kv.wgsl +224 -0
  391. package/src/gpu/kernels/attention_decode_online_f16.wgsl +223 -0
  392. package/src/gpu/kernels/attention_decode_online_f16kv.wgsl +225 -0
  393. package/src/gpu/kernels/attention_decode_optimized.wgsl +445 -0
  394. package/src/gpu/kernels/attention_decode_paged_f16.wgsl +172 -0
  395. package/src/gpu/kernels/attention_decode_paged_f16kv.wgsl +174 -0
  396. package/src/gpu/kernels/attention_decode_subgroup.wgsl +233 -0
  397. package/src/gpu/kernels/attention_decode_tiered_f16.wgsl +218 -0
  398. package/src/gpu/kernels/attention_decode_tiered_f16kv.wgsl +220 -0
  399. package/src/gpu/kernels/attention_decode_tiered_int4_f16kv.wgsl +242 -0
  400. package/src/gpu/kernels/attention_decode_tiered_int8_f16kv.wgsl +242 -0
  401. package/src/gpu/kernels/attention_f16.wgsl +214 -0
  402. package/src/gpu/kernels/attention_f16kv.wgsl +242 -0
  403. package/src/gpu/kernels/attention_small.wgsl +260 -0
  404. package/src/gpu/kernels/attention_small_f16.wgsl +240 -0
  405. package/src/gpu/kernels/attention_small_f16kv.wgsl +266 -0
  406. package/src/gpu/kernels/attention_streaming.wgsl +149 -0
  407. package/src/gpu/kernels/attention_streaming_f16.wgsl +147 -0
  408. package/src/gpu/kernels/attention_streaming_f16kv.wgsl +151 -0
  409. package/src/gpu/kernels/backward/adam.d.ts +28 -0
  410. package/src/gpu/kernels/backward/adam.js +199 -0
  411. package/src/gpu/kernels/backward/adam.wgsl +50 -0
  412. package/src/gpu/kernels/backward/attention_backward.d.ts +22 -0
  413. package/src/gpu/kernels/backward/attention_backward.js +276 -0
  414. package/src/gpu/kernels/backward/attention_backward.wgsl +49 -0
  415. package/src/gpu/kernels/backward/bias_add_backward.d.ts +17 -0
  416. package/src/gpu/kernels/backward/bias_add_backward.js +24 -0
  417. package/src/gpu/kernels/backward/bias_add_backward.wgsl +33 -0
  418. package/src/gpu/kernels/backward/conv2d_backward.d.ts +31 -0
  419. package/src/gpu/kernels/backward/conv2d_backward.js +135 -0
  420. package/src/gpu/kernels/backward/conv2d_backward_input.wgsl +83 -0
  421. package/src/gpu/kernels/backward/conv2d_backward_weight.wgsl +70 -0
  422. package/src/gpu/kernels/backward/cross_entropy_backward.d.ts +23 -0
  423. package/src/gpu/kernels/backward/cross_entropy_backward.js +29 -0
  424. package/src/gpu/kernels/backward/cross_entropy_backward.wgsl +39 -0
  425. package/src/gpu/kernels/backward/embed_backward.d.ts +29 -0
  426. package/src/gpu/kernels/backward/embed_backward.js +118 -0
  427. package/src/gpu/kernels/backward/embed_backward.wgsl +73 -0
  428. package/src/gpu/kernels/backward/gelu_backward.d.ts +16 -0
  429. package/src/gpu/kernels/backward/gelu_backward.js +39 -0
  430. package/src/gpu/kernels/backward/gelu_backward.wgsl +38 -0
  431. package/src/gpu/kernels/backward/groupnorm_backward.d.ts +24 -0
  432. package/src/gpu/kernels/backward/groupnorm_backward.js +29 -0
  433. package/src/gpu/kernels/backward/groupnorm_backward.wgsl +143 -0
  434. package/src/gpu/kernels/backward/index.d.ts +17 -0
  435. package/src/gpu/kernels/backward/index.js +23 -0
  436. package/src/gpu/kernels/backward/layernorm_backward.d.ts +22 -0
  437. package/src/gpu/kernels/backward/layernorm_backward.js +135 -0
  438. package/src/gpu/kernels/backward/layernorm_backward.wgsl +194 -0
  439. package/src/gpu/kernels/backward/matmul_backward.d.ts +32 -0
  440. package/src/gpu/kernels/backward/matmul_backward.js +124 -0
  441. package/src/gpu/kernels/backward/matmul_backward.wgsl +90 -0
  442. package/src/gpu/kernels/backward/matmul_transpose_a.wgsl +84 -0
  443. package/src/gpu/kernels/backward/pixel_shuffle_backward.d.ts +22 -0
  444. package/src/gpu/kernels/backward/pixel_shuffle_backward.js +30 -0
  445. package/src/gpu/kernels/backward/pixel_shuffle_backward.wgsl +54 -0
  446. package/src/gpu/kernels/backward/rmsnorm_backward.d.ts +24 -0
  447. package/src/gpu/kernels/backward/rmsnorm_backward.js +101 -0
  448. package/src/gpu/kernels/backward/rmsnorm_backward.wgsl +78 -0
  449. package/src/gpu/kernels/backward/rope_backward.d.ts +25 -0
  450. package/src/gpu/kernels/backward/rope_backward.js +109 -0
  451. package/src/gpu/kernels/backward/rope_backward.wgsl +59 -0
  452. package/src/gpu/kernels/backward/scale_backward.d.ts +16 -0
  453. package/src/gpu/kernels/backward/scale_backward.js +84 -0
  454. package/src/gpu/kernels/backward/scale_backward.wgsl +27 -0
  455. package/src/gpu/kernels/backward/silu_backward.d.ts +16 -0
  456. package/src/gpu/kernels/backward/silu_backward.js +39 -0
  457. package/src/gpu/kernels/backward/silu_backward.wgsl +31 -0
  458. package/src/gpu/kernels/backward/softmax_backward.d.ts +16 -0
  459. package/src/gpu/kernels/backward/softmax_backward.js +43 -0
  460. package/src/gpu/kernels/backward/softmax_backward.wgsl +44 -0
  461. package/src/gpu/kernels/backward/upsample2d_backward.d.ts +21 -0
  462. package/src/gpu/kernels/backward/upsample2d_backward.js +30 -0
  463. package/src/gpu/kernels/backward/upsample2d_backward.wgsl +59 -0
  464. package/src/gpu/kernels/backward/utils.d.ts +45 -0
  465. package/src/gpu/kernels/backward/utils.js +371 -0
  466. package/src/gpu/kernels/bf16_to_f16.wgsl +54 -0
  467. package/src/gpu/kernels/bf16_to_f32.wgsl +70 -0
  468. package/src/gpu/kernels/bias_add.wgsl +40 -0
  469. package/src/gpu/kernels/bias_add_f16.wgsl +44 -0
  470. package/src/gpu/kernels/cast.d.ts +67 -0
  471. package/src/gpu/kernels/cast.js +422 -0
  472. package/src/gpu/kernels/cast_f16_to_f32.wgsl +31 -0
  473. package/src/gpu/kernels/cast_f32_to_f16.wgsl +36 -0
  474. package/src/gpu/kernels/check-finiteness.d.ts +15 -0
  475. package/src/gpu/kernels/check-finiteness.js +149 -0
  476. package/src/gpu/kernels/check-stop.d.ts +31 -0
  477. package/src/gpu/kernels/check-stop.js +181 -0
  478. package/src/gpu/kernels/clamp.d.ts +22 -0
  479. package/src/gpu/kernels/clamp.js +42 -0
  480. package/src/gpu/kernels/clamp.wgsl +24 -0
  481. package/src/gpu/kernels/constants.d.ts +168 -0
  482. package/src/gpu/kernels/constants.js +129 -0
  483. package/src/gpu/kernels/conv2d.d.ts +34 -0
  484. package/src/gpu/kernels/conv2d.js +81 -0
  485. package/src/gpu/kernels/conv2d.wgsl +71 -0
  486. package/src/gpu/kernels/conv2d_f16.wgsl +73 -0
  487. package/src/gpu/kernels/cross_entropy_loss.d.ts +21 -0
  488. package/src/gpu/kernels/cross_entropy_loss.js +54 -0
  489. package/src/gpu/kernels/cross_entropy_loss.wgsl +39 -0
  490. package/src/gpu/kernels/dequant.d.ts +108 -0
  491. package/src/gpu/kernels/dequant.js +524 -0
  492. package/src/gpu/kernels/dequant_f16_out.wgsl +151 -0
  493. package/src/gpu/kernels/dequant_f16_out_vec4.wgsl +149 -0
  494. package/src/gpu/kernels/dequant_f16_rowwise.wgsl +139 -0
  495. package/src/gpu/kernels/dequant_f32_rowwise.wgsl +133 -0
  496. package/src/gpu/kernels/dequant_mxfp4.wgsl +120 -0
  497. package/src/gpu/kernels/dequant_mxfp4_expert.wgsl +129 -0
  498. package/src/gpu/kernels/dequant_mxfp4_expert_f16.wgsl +105 -0
  499. package/src/gpu/kernels/dequant_mxfp4_vec4.wgsl +116 -0
  500. package/src/gpu/kernels/dequant_q6k.wgsl +140 -0
  501. package/src/gpu/kernels/dequant_q8_0.wgsl +98 -0
  502. package/src/gpu/kernels/dequant_shared.wgsl +202 -0
  503. package/src/gpu/kernels/dequant_shared_vec4.wgsl +153 -0
  504. package/src/gpu/kernels/dequant_subgroup.wgsl +202 -0
  505. package/src/gpu/kernels/dispatch.d.ts +157 -0
  506. package/src/gpu/kernels/dispatch.js +235 -0
  507. package/src/gpu/kernels/energy.d.ts +131 -0
  508. package/src/gpu/kernels/energy.js +425 -0
  509. package/src/gpu/kernels/energy_eval.wgsl +26 -0
  510. package/src/gpu/kernels/energy_eval_f16.wgsl +30 -0
  511. package/src/gpu/kernels/energy_quintel_grad.wgsl +92 -0
  512. package/src/gpu/kernels/energy_quintel_grad_f16.wgsl +96 -0
  513. package/src/gpu/kernels/energy_quintel_reduce.wgsl +112 -0
  514. package/src/gpu/kernels/energy_quintel_reduce_f16.wgsl +116 -0
  515. package/src/gpu/kernels/energy_quintel_update.wgsl +92 -0
  516. package/src/gpu/kernels/energy_quintel_update_f16.wgsl +96 -0
  517. package/src/gpu/kernels/energy_update.wgsl +25 -0
  518. package/src/gpu/kernels/energy_update_f16.wgsl +30 -0
  519. package/src/gpu/kernels/feature-check.d.ts +42 -0
  520. package/src/gpu/kernels/feature-check.js +70 -0
  521. package/src/gpu/kernels/fused_ffn.d.ts +65 -0
  522. package/src/gpu/kernels/fused_ffn.js +318 -0
  523. package/src/gpu/kernels/fused_ffn.wgsl +420 -0
  524. package/src/gpu/kernels/fused_ffn_f16.wgsl +213 -0
  525. package/src/gpu/kernels/fused_ffn_q4k.wgsl +375 -0
  526. package/src/gpu/kernels/fused_matmul_q4.wgsl +404 -0
  527. package/src/gpu/kernels/fused_matmul_q4_batched.wgsl +194 -0
  528. package/src/gpu/kernels/fused_matmul_q4_batched_f16.wgsl +170 -0
  529. package/src/gpu/kernels/fused_matmul_q4_batched_f16a.wgsl +154 -0
  530. package/src/gpu/kernels/fused_matmul_q4_f16a.wgsl +219 -0
  531. package/src/gpu/kernels/fused_matmul_q4_multicol_f16.wgsl +216 -0
  532. package/src/gpu/kernels/fused_matmul_q4_multicol_f16a.wgsl +204 -0
  533. package/src/gpu/kernels/fused_matmul_residual.d.ts +46 -0
  534. package/src/gpu/kernels/fused_matmul_residual.js +152 -0
  535. package/src/gpu/kernels/fused_matmul_rmsnorm.d.ts +64 -0
  536. package/src/gpu/kernels/fused_matmul_rmsnorm.js +273 -0
  537. package/src/gpu/kernels/fused_matmul_rmsnorm.wgsl +324 -0
  538. package/src/gpu/kernels/fused_matmul_rmsnorm_f16.wgsl +303 -0
  539. package/src/gpu/kernels/fused_swiglu.wgsl +63 -0
  540. package/src/gpu/kernels/fused_swiglu_f16.wgsl +57 -0
  541. package/src/gpu/kernels/gather.d.ts +64 -0
  542. package/src/gpu/kernels/gather.js +119 -0
  543. package/src/gpu/kernels/gather.wgsl +61 -0
  544. package/src/gpu/kernels/gather_f16.wgsl +65 -0
  545. package/src/gpu/kernels/gather_f16_f16_out.wgsl +55 -0
  546. package/src/gpu/kernels/gather_f16_out.wgsl +55 -0
  547. package/src/gpu/kernels/gather_f16_vec4.wgsl +76 -0
  548. package/src/gpu/kernels/gather_f16_vec4_f16_out.wgsl +68 -0
  549. package/src/gpu/kernels/gather_vec4.wgsl +74 -0
  550. package/src/gpu/kernels/gather_vec4_f16_out.wgsl +68 -0
  551. package/src/gpu/kernels/gelu.d.ts +33 -0
  552. package/src/gpu/kernels/gelu.js +47 -0
  553. package/src/gpu/kernels/gelu.wgsl +64 -0
  554. package/src/gpu/kernels/gelu_f16.wgsl +66 -0
  555. package/src/gpu/kernels/gptoss_mxfp4_expert_fused.wgsl +127 -0
  556. package/src/gpu/kernels/gptoss_router_topk.wgsl +119 -0
  557. package/src/gpu/kernels/groupnorm.d.ts +31 -0
  558. package/src/gpu/kernels/groupnorm.js +91 -0
  559. package/src/gpu/kernels/groupnorm_apply.wgsl +41 -0
  560. package/src/gpu/kernels/groupnorm_apply_f16.wgsl +46 -0
  561. package/src/gpu/kernels/groupnorm_stats.wgsl +76 -0
  562. package/src/gpu/kernels/groupnorm_stats_f16.wgsl +79 -0
  563. package/src/gpu/kernels/index.d.ts +336 -0
  564. package/src/gpu/kernels/index.js +284 -0
  565. package/src/gpu/kernels/kernel-base.d.ts +33 -0
  566. package/src/gpu/kernels/kernel-base.js +46 -0
  567. package/src/gpu/kernels/kernel-configs.d.ts +65 -0
  568. package/src/gpu/kernels/kernel-configs.js +50 -0
  569. package/src/gpu/kernels/kernel-tuning.d.ts +42 -0
  570. package/src/gpu/kernels/kernel-tuning.js +149 -0
  571. package/src/gpu/kernels/kv-quantize.d.ts +37 -0
  572. package/src/gpu/kernels/kv-quantize.js +138 -0
  573. package/src/gpu/kernels/kv_quantize_int4.wgsl +119 -0
  574. package/src/gpu/kernels/kv_quantize_int8.wgsl +119 -0
  575. package/src/gpu/kernels/layernorm.d.ts +37 -0
  576. package/src/gpu/kernels/layernorm.js +80 -0
  577. package/src/gpu/kernels/layernorm.wgsl +121 -0
  578. package/src/gpu/kernels/layernorm_f16.wgsl +103 -0
  579. package/src/gpu/kernels/linear-attention-core.d.ts +39 -0
  580. package/src/gpu/kernels/linear-attention-core.js +535 -0
  581. package/src/gpu/kernels/logit-merge.d.ts +110 -0
  582. package/src/gpu/kernels/logit-merge.js +392 -0
  583. package/src/gpu/kernels/matmul-dispatch.d.ts +38 -0
  584. package/src/gpu/kernels/matmul-dispatch.js +155 -0
  585. package/src/gpu/kernels/matmul-selection.d.ts +87 -0
  586. package/src/gpu/kernels/matmul-selection.js +474 -0
  587. package/src/gpu/kernels/matmul.d.ts +109 -0
  588. package/src/gpu/kernels/matmul.js +271 -0
  589. package/src/gpu/kernels/matmul_f16.wgsl +170 -0
  590. package/src/gpu/kernels/matmul_f16_tiled.wgsl +165 -0
  591. package/src/gpu/kernels/matmul_f16w_f32a.wgsl +89 -0
  592. package/src/gpu/kernels/matmul_f16w_f32a_tiled.wgsl +154 -0
  593. package/src/gpu/kernels/matmul_f32.wgsl +100 -0
  594. package/src/gpu/kernels/matmul_gemv.wgsl +80 -0
  595. package/src/gpu/kernels/matmul_gemv_f16a.wgsl +81 -0
  596. package/src/gpu/kernels/matmul_gemv_residual.wgsl +119 -0
  597. package/src/gpu/kernels/matmul_gemv_residual_f16.wgsl +78 -0
  598. package/src/gpu/kernels/matmul_gemv_subgroup.wgsl +345 -0
  599. package/src/gpu/kernels/matmul_gemv_subgroup_f16a.wgsl +514 -0
  600. package/src/gpu/kernels/modulate.d.ts +29 -0
  601. package/src/gpu/kernels/modulate.js +49 -0
  602. package/src/gpu/kernels/modulate.wgsl +40 -0
  603. package/src/gpu/kernels/modulate_f16.wgsl +43 -0
  604. package/src/gpu/kernels/moe.d.ts +164 -0
  605. package/src/gpu/kernels/moe.js +496 -0
  606. package/src/gpu/kernels/moe_gather.wgsl +170 -0
  607. package/src/gpu/kernels/moe_gather_f16.wgsl +82 -0
  608. package/src/gpu/kernels/moe_gather_vec4.wgsl +74 -0
  609. package/src/gpu/kernels/moe_offsets.wgsl +48 -0
  610. package/src/gpu/kernels/pipeline-cache.d.ts +88 -0
  611. package/src/gpu/kernels/pipeline-cache.js +305 -0
  612. package/src/gpu/kernels/pixel_shuffle.d.ts +27 -0
  613. package/src/gpu/kernels/pixel_shuffle.js +49 -0
  614. package/src/gpu/kernels/pixel_shuffle.wgsl +44 -0
  615. package/src/gpu/kernels/pixel_shuffle_f16.wgsl +47 -0
  616. package/src/gpu/kernels/residual.d.ts +74 -0
  617. package/src/gpu/kernels/residual.js +127 -0
  618. package/src/gpu/kernels/residual.wgsl +53 -0
  619. package/src/gpu/kernels/residual_f16.wgsl +35 -0
  620. package/src/gpu/kernels/residual_f16_vec4.wgsl +47 -0
  621. package/src/gpu/kernels/residual_vec4.wgsl +46 -0
  622. package/src/gpu/kernels/rmsnorm.d.ts +53 -0
  623. package/src/gpu/kernels/rmsnorm.js +140 -0
  624. package/src/gpu/kernels/rmsnorm.wgsl +417 -0
  625. package/src/gpu/kernels/rmsnorm_f16.wgsl +164 -0
  626. package/src/gpu/kernels/rope.d.ts +48 -0
  627. package/src/gpu/kernels/rope.js +53 -0
  628. package/src/gpu/kernels/rope.wgsl +328 -0
  629. package/src/gpu/kernels/rope_f16.wgsl +271 -0
  630. package/src/gpu/kernels/rule-matcher.d.ts +30 -0
  631. package/src/gpu/kernels/rule-matcher.js +42 -0
  632. package/src/gpu/kernels/rule-registry.d.ts +7 -0
  633. package/src/gpu/kernels/rule-registry.js +41 -0
  634. package/src/gpu/kernels/sample.d.ts +75 -0
  635. package/src/gpu/kernels/sample.js +578 -0
  636. package/src/gpu/kernels/sample.wgsl +377 -0
  637. package/src/gpu/kernels/sample_f16.wgsl +331 -0
  638. package/src/gpu/kernels/scale.d.ts +35 -0
  639. package/src/gpu/kernels/scale.js +37 -0
  640. package/src/gpu/kernels/scale.wgsl +38 -0
  641. package/src/gpu/kernels/scatter_add.wgsl +88 -0
  642. package/src/gpu/kernels/scatter_add_dynamic.wgsl +59 -0
  643. package/src/gpu/kernels/scatter_add_dynamic_f16.wgsl +52 -0
  644. package/src/gpu/kernels/scatter_add_dynamic_f16_weights.wgsl +50 -0
  645. package/src/gpu/kernels/scatter_add_vec4.wgsl +70 -0
  646. package/src/gpu/kernels/shader-cache.d.ts +56 -0
  647. package/src/gpu/kernels/shader-cache.js +206 -0
  648. package/src/gpu/kernels/silu.d.ts +75 -0
  649. package/src/gpu/kernels/silu.js +340 -0
  650. package/src/gpu/kernels/silu.wgsl +99 -0
  651. package/src/gpu/kernels/silu_f16.wgsl +98 -0
  652. package/src/gpu/kernels/softmax.d.ts +57 -0
  653. package/src/gpu/kernels/softmax.js +106 -0
  654. package/src/gpu/kernels/softmax.wgsl +388 -0
  655. package/src/gpu/kernels/softmax_subgroup.wgsl +175 -0
  656. package/src/gpu/kernels/split_qkv.d.ts +51 -0
  657. package/src/gpu/kernels/split_qkv.js +41 -0
  658. package/src/gpu/kernels/split_qkv.wgsl +71 -0
  659. package/src/gpu/kernels/split_qkv_f16.wgsl +75 -0
  660. package/src/gpu/kernels/topk.wgsl +243 -0
  661. package/src/gpu/kernels/topk_f16.wgsl +108 -0
  662. package/src/gpu/kernels/topk_f16_weights.wgsl +101 -0
  663. package/src/gpu/kernels/transpose.d.ts +21 -0
  664. package/src/gpu/kernels/transpose.js +30 -0
  665. package/src/gpu/kernels/transpose.wgsl +32 -0
  666. package/src/gpu/kernels/types.d.ts +21 -0
  667. package/src/gpu/kernels/types.js +4 -0
  668. package/src/gpu/kernels/uniform-utils.d.ts +48 -0
  669. package/src/gpu/kernels/uniform-utils.js +94 -0
  670. package/src/gpu/kernels/upsample2d.d.ts +25 -0
  671. package/src/gpu/kernels/upsample2d.js +58 -0
  672. package/src/gpu/kernels/upsample2d.wgsl +37 -0
  673. package/src/gpu/kernels/upsample2d_f16.wgsl +41 -0
  674. package/src/gpu/kernels/utils.d.ts +106 -0
  675. package/src/gpu/kernels/utils.js +224 -0
  676. package/src/gpu/multi-model-recorder.d.ts +21 -0
  677. package/src/gpu/multi-model-recorder.js +31 -0
  678. package/src/gpu/partitioned-buffer-pool.d.ts +28 -0
  679. package/src/gpu/partitioned-buffer-pool.js +49 -0
  680. package/src/gpu/perf-guards.d.ts +25 -0
  681. package/src/gpu/perf-guards.js +140 -0
  682. package/src/gpu/profiler.d.ts +114 -0
  683. package/src/gpu/profiler.js +391 -0
  684. package/src/gpu/submit-tracker.d.ts +111 -0
  685. package/src/gpu/submit-tracker.js +229 -0
  686. package/src/gpu/tensor.d.ts +69 -0
  687. package/src/gpu/tensor.js +75 -0
  688. package/src/gpu/uniform-cache.d.ts +108 -0
  689. package/src/gpu/uniform-cache.js +242 -0
  690. package/src/gpu/weight-buffer.d.ts +115 -0
  691. package/src/gpu/weight-buffer.js +118 -0
  692. package/src/hotswap/intent-bundle.d.ts +37 -0
  693. package/src/hotswap/intent-bundle.js +123 -0
  694. package/src/hotswap/manifest.d.ts +33 -0
  695. package/src/hotswap/manifest.js +114 -0
  696. package/src/hotswap/runtime.d.ts +31 -0
  697. package/src/hotswap/runtime.js +128 -0
  698. package/src/index-browser.d.ts +47 -0
  699. package/src/index-browser.js +53 -0
  700. package/src/index-internal.d.ts +2 -0
  701. package/src/index-internal.js +2 -0
  702. package/src/index.d.ts +102 -0
  703. package/src/index.js +75 -0
  704. package/src/inference/README.md +593 -0
  705. package/src/inference/browser-harness.d.ts +234 -0
  706. package/src/inference/browser-harness.js +2665 -0
  707. package/src/inference/decode-buffers.d.ts +108 -0
  708. package/src/inference/decode-buffers.js +181 -0
  709. package/src/inference/decode-ring.d.ts +52 -0
  710. package/src/inference/decode-ring.js +273 -0
  711. package/src/inference/expert-router.d.ts +27 -0
  712. package/src/inference/expert-router.js +55 -0
  713. package/src/inference/functiongemma.d.ts +15 -0
  714. package/src/inference/functiongemma.js +1 -0
  715. package/src/inference/kv-cache/base.d.ts +150 -0
  716. package/src/inference/kv-cache/base.js +1037 -0
  717. package/src/inference/kv-cache/basis-decomposed-paged.d.ts +50 -0
  718. package/src/inference/kv-cache/basis-decomposed-paged.js +276 -0
  719. package/src/inference/kv-cache/index.d.ts +35 -0
  720. package/src/inference/kv-cache/index.js +20 -0
  721. package/src/inference/kv-cache/sliding-window.d.ts +72 -0
  722. package/src/inference/kv-cache/sliding-window.js +243 -0
  723. package/src/inference/kv-cache/tiered.d.ts +89 -0
  724. package/src/inference/kv-cache/tiered.js +574 -0
  725. package/src/inference/kv-cache/types.d.ts +188 -0
  726. package/src/inference/kv-cache/types.js +80 -0
  727. package/src/inference/kv-cache.d.ts +36 -0
  728. package/src/inference/kv-cache.js +18 -0
  729. package/src/inference/moe-router.d.ts +212 -0
  730. package/src/inference/moe-router.js +553 -0
  731. package/src/inference/multi-model-network.d.ts +139 -0
  732. package/src/inference/multi-model-network.js +769 -0
  733. package/src/inference/multi-pipeline-pool.d.ts +62 -0
  734. package/src/inference/multi-pipeline-pool.js +161 -0
  735. package/src/inference/network-evolution.d.ts +46 -0
  736. package/src/inference/network-evolution.js +80 -0
  737. package/src/inference/pipelines/context.d.ts +18 -0
  738. package/src/inference/pipelines/context.js +44 -0
  739. package/src/inference/pipelines/diffusion/helpers.d.ts +29 -0
  740. package/src/inference/pipelines/diffusion/helpers.js +112 -0
  741. package/src/inference/pipelines/diffusion/index.d.ts +3 -0
  742. package/src/inference/pipelines/diffusion/index.js +3 -0
  743. package/src/inference/pipelines/diffusion/init.d.ts +24 -0
  744. package/src/inference/pipelines/diffusion/init.js +124 -0
  745. package/src/inference/pipelines/diffusion/pipeline.d.ts +38 -0
  746. package/src/inference/pipelines/diffusion/pipeline.js +632 -0
  747. package/src/inference/pipelines/diffusion/scheduler.d.ts +19 -0
  748. package/src/inference/pipelines/diffusion/scheduler.js +65 -0
  749. package/src/inference/pipelines/diffusion/sd3-transformer.d.ts +20 -0
  750. package/src/inference/pipelines/diffusion/sd3-transformer.js +1194 -0
  751. package/src/inference/pipelines/diffusion/sd3-weights.d.ts +21 -0
  752. package/src/inference/pipelines/diffusion/sd3-weights.js +287 -0
  753. package/src/inference/pipelines/diffusion/text-encoder-gpu.d.ts +80 -0
  754. package/src/inference/pipelines/diffusion/text-encoder-gpu.js +935 -0
  755. package/src/inference/pipelines/diffusion/text-encoder.d.ts +29 -0
  756. package/src/inference/pipelines/diffusion/text-encoder.js +178 -0
  757. package/src/inference/pipelines/diffusion/types.d.ts +112 -0
  758. package/src/inference/pipelines/diffusion/types.js +1 -0
  759. package/src/inference/pipelines/diffusion/vae.d.ts +20 -0
  760. package/src/inference/pipelines/diffusion/vae.js +675 -0
  761. package/src/inference/pipelines/diffusion/weights.d.ts +40 -0
  762. package/src/inference/pipelines/diffusion/weights.js +150 -0
  763. package/src/inference/pipelines/dream/energy-head-pipeline.d.ts +29 -0
  764. package/src/inference/pipelines/dream/energy-head-pipeline.js +6 -0
  765. package/src/inference/pipelines/dream/pipeline.d.ts +17 -0
  766. package/src/inference/pipelines/dream/pipeline.js +8 -0
  767. package/src/inference/pipelines/energy/index.d.ts +1 -0
  768. package/src/inference/pipelines/energy/index.js +1 -0
  769. package/src/inference/pipelines/energy/pipeline.d.ts +27 -0
  770. package/src/inference/pipelines/energy/pipeline.js +680 -0
  771. package/src/inference/pipelines/energy/quintel.d.ts +87 -0
  772. package/src/inference/pipelines/energy/quintel.js +207 -0
  773. package/src/inference/pipelines/energy/types.d.ts +63 -0
  774. package/src/inference/pipelines/energy/types.js +1 -0
  775. package/src/inference/pipelines/energy-head/index.d.ts +6 -0
  776. package/src/inference/pipelines/energy-head/index.js +6 -0
  777. package/src/inference/pipelines/energy-head/row-head-pipeline.d.ts +103 -0
  778. package/src/inference/pipelines/energy-head/row-head-pipeline.js +487 -0
  779. package/src/inference/pipelines/factory.d.ts +10 -0
  780. package/src/inference/pipelines/factory.js +6 -0
  781. package/src/inference/pipelines/index.d.ts +22 -0
  782. package/src/inference/pipelines/index.js +19 -0
  783. package/src/inference/pipelines/registry.d.ts +15 -0
  784. package/src/inference/pipelines/registry.js +23 -0
  785. package/src/inference/pipelines/rng.d.ts +2 -0
  786. package/src/inference/pipelines/rng.js +17 -0
  787. package/src/inference/pipelines/structured/index.d.ts +8 -0
  788. package/src/inference/pipelines/structured/index.js +8 -0
  789. package/src/inference/pipelines/structured/json-head-pipeline.d.ts +58 -0
  790. package/src/inference/pipelines/structured/json-head-pipeline.js +181 -0
  791. package/src/inference/pipelines/text/attention/index.d.ts +24 -0
  792. package/src/inference/pipelines/text/attention/index.js +17 -0
  793. package/src/inference/pipelines/text/attention/projections.d.ts +101 -0
  794. package/src/inference/pipelines/text/attention/projections.js +435 -0
  795. package/src/inference/pipelines/text/attention/record.d.ts +36 -0
  796. package/src/inference/pipelines/text/attention/record.js +613 -0
  797. package/src/inference/pipelines/text/attention/run.d.ts +38 -0
  798. package/src/inference/pipelines/text/attention/run.js +826 -0
  799. package/src/inference/pipelines/text/attention/types.d.ts +98 -0
  800. package/src/inference/pipelines/text/attention/types.js +67 -0
  801. package/src/inference/pipelines/text/attention.d.ts +23 -0
  802. package/src/inference/pipelines/text/attention.js +12 -0
  803. package/src/inference/pipelines/text/bdpa-steamroller.d.ts +22 -0
  804. package/src/inference/pipelines/text/bdpa-steamroller.js +158 -0
  805. package/src/inference/pipelines/text/buffer-types.d.ts +7 -0
  806. package/src/inference/pipelines/text/buffer-types.js +4 -0
  807. package/src/inference/pipelines/text/chat-format.d.ts +46 -0
  808. package/src/inference/pipelines/text/chat-format.js +366 -0
  809. package/src/inference/pipelines/text/config.d.ts +235 -0
  810. package/src/inference/pipelines/text/config.js +623 -0
  811. package/src/inference/pipelines/text/debug-utils/config.d.ts +144 -0
  812. package/src/inference/pipelines/text/debug-utils/config.js +156 -0
  813. package/src/inference/pipelines/text/debug-utils/index.d.ts +53 -0
  814. package/src/inference/pipelines/text/debug-utils/index.js +44 -0
  815. package/src/inference/pipelines/text/debug-utils/logging.d.ts +106 -0
  816. package/src/inference/pipelines/text/debug-utils/logging.js +152 -0
  817. package/src/inference/pipelines/text/debug-utils/tensor.d.ts +119 -0
  818. package/src/inference/pipelines/text/debug-utils/tensor.js +268 -0
  819. package/src/inference/pipelines/text/debug-utils/utils.d.ts +77 -0
  820. package/src/inference/pipelines/text/debug-utils/utils.js +139 -0
  821. package/src/inference/pipelines/text/debug-utils.d.ts +42 -0
  822. package/src/inference/pipelines/text/debug-utils.js +34 -0
  823. package/src/inference/pipelines/text/embed.d.ts +67 -0
  824. package/src/inference/pipelines/text/embed.js +461 -0
  825. package/src/inference/pipelines/text/execution-plan.d.ts +116 -0
  826. package/src/inference/pipelines/text/execution-plan.js +314 -0
  827. package/src/inference/pipelines/text/execution-v0.d.ts +66 -0
  828. package/src/inference/pipelines/text/execution-v0.js +1139 -0
  829. package/src/inference/pipelines/text/ffn/dense.d.ts +40 -0
  830. package/src/inference/pipelines/text/ffn/dense.js +759 -0
  831. package/src/inference/pipelines/text/ffn/index.d.ts +23 -0
  832. package/src/inference/pipelines/text/ffn/index.js +16 -0
  833. package/src/inference/pipelines/text/ffn/moe.d.ts +21 -0
  834. package/src/inference/pipelines/text/ffn/moe.js +49 -0
  835. package/src/inference/pipelines/text/ffn/sandwich.d.ts +25 -0
  836. package/src/inference/pipelines/text/ffn/sandwich.js +196 -0
  837. package/src/inference/pipelines/text/ffn/standard.d.ts +23 -0
  838. package/src/inference/pipelines/text/ffn/standard.js +84 -0
  839. package/src/inference/pipelines/text/ffn/types.d.ts +30 -0
  840. package/src/inference/pipelines/text/ffn/types.js +25 -0
  841. package/src/inference/pipelines/text/ffn.d.ts +31 -0
  842. package/src/inference/pipelines/text/ffn.js +18 -0
  843. package/src/inference/pipelines/text/finiteness-guard-status.d.ts +11 -0
  844. package/src/inference/pipelines/text/finiteness-guard-status.js +21 -0
  845. package/src/inference/pipelines/text/finiteness-policy.d.ts +35 -0
  846. package/src/inference/pipelines/text/finiteness-policy.js +45 -0
  847. package/src/inference/pipelines/text/generator-helpers.d.ts +34 -0
  848. package/src/inference/pipelines/text/generator-helpers.js +175 -0
  849. package/src/inference/pipelines/text/generator-runtime.d.ts +93 -0
  850. package/src/inference/pipelines/text/generator-runtime.js +373 -0
  851. package/src/inference/pipelines/text/generator-steps.d.ts +75 -0
  852. package/src/inference/pipelines/text/generator-steps.js +1078 -0
  853. package/src/inference/pipelines/text/generator.d.ts +41 -0
  854. package/src/inference/pipelines/text/generator.js +1345 -0
  855. package/src/inference/pipelines/text/index.d.ts +5 -0
  856. package/src/inference/pipelines/text/index.js +6 -0
  857. package/src/inference/pipelines/text/init.d.ts +295 -0
  858. package/src/inference/pipelines/text/init.js +965 -0
  859. package/src/inference/pipelines/text/kernel-path-auto-select.d.ts +12 -0
  860. package/src/inference/pipelines/text/kernel-path-auto-select.js +90 -0
  861. package/src/inference/pipelines/text/kernel-trace.d.ts +150 -0
  862. package/src/inference/pipelines/text/kernel-trace.js +324 -0
  863. package/src/inference/pipelines/text/layer-plan.d.ts +65 -0
  864. package/src/inference/pipelines/text/layer-plan.js +249 -0
  865. package/src/inference/pipelines/text/layer.d.ts +56 -0
  866. package/src/inference/pipelines/text/layer.js +916 -0
  867. package/src/inference/pipelines/text/linear-attention.d.ts +94 -0
  868. package/src/inference/pipelines/text/linear-attention.js +803 -0
  869. package/src/inference/pipelines/text/logits/cpu.d.ts +81 -0
  870. package/src/inference/pipelines/text/logits/cpu.js +91 -0
  871. package/src/inference/pipelines/text/logits/gpu.d.ts +113 -0
  872. package/src/inference/pipelines/text/logits/gpu.js +406 -0
  873. package/src/inference/pipelines/text/logits/index.d.ts +57 -0
  874. package/src/inference/pipelines/text/logits/index.js +305 -0
  875. package/src/inference/pipelines/text/logits/types.d.ts +46 -0
  876. package/src/inference/pipelines/text/logits/types.js +4 -0
  877. package/src/inference/pipelines/text/logits/utils.d.ts +49 -0
  878. package/src/inference/pipelines/text/logits/utils.js +59 -0
  879. package/src/inference/pipelines/text/logits.d.ts +27 -0
  880. package/src/inference/pipelines/text/logits.js +16 -0
  881. package/src/inference/pipelines/text/lora-apply.d.ts +28 -0
  882. package/src/inference/pipelines/text/lora-apply.js +58 -0
  883. package/src/inference/pipelines/text/lora-types.d.ts +39 -0
  884. package/src/inference/pipelines/text/lora-types.js +18 -0
  885. package/src/inference/pipelines/text/lora.d.ts +18 -0
  886. package/src/inference/pipelines/text/lora.js +12 -0
  887. package/src/inference/pipelines/text/model-load.d.ts +58 -0
  888. package/src/inference/pipelines/text/model-load.js +561 -0
  889. package/src/inference/pipelines/text/moe-cache.d.ts +32 -0
  890. package/src/inference/pipelines/text/moe-cache.js +107 -0
  891. package/src/inference/pipelines/text/moe-cpu-gptoss.d.ts +9 -0
  892. package/src/inference/pipelines/text/moe-cpu-gptoss.js +110 -0
  893. package/src/inference/pipelines/text/moe-cpu.d.ts +13 -0
  894. package/src/inference/pipelines/text/moe-cpu.js +116 -0
  895. package/src/inference/pipelines/text/moe-gpu.d.ts +13 -0
  896. package/src/inference/pipelines/text/moe-gpu.js +611 -0
  897. package/src/inference/pipelines/text/moe-helpers.d.ts +12 -0
  898. package/src/inference/pipelines/text/moe-helpers.js +21 -0
  899. package/src/inference/pipelines/text/moe-impl.d.ts +117 -0
  900. package/src/inference/pipelines/text/moe-impl.js +9 -0
  901. package/src/inference/pipelines/text/moe-shape-validator.d.ts +31 -0
  902. package/src/inference/pipelines/text/moe-shape-validator.js +78 -0
  903. package/src/inference/pipelines/text/ops.d.ts +167 -0
  904. package/src/inference/pipelines/text/ops.js +367 -0
  905. package/src/inference/pipelines/text/probes.d.ts +31 -0
  906. package/src/inference/pipelines/text/probes.js +170 -0
  907. package/src/inference/pipelines/text/sampling.d.ts +54 -0
  908. package/src/inference/pipelines/text/sampling.js +203 -0
  909. package/src/inference/pipelines/text/state.d.ts +112 -0
  910. package/src/inference/pipelines/text/state.js +152 -0
  911. package/src/inference/pipelines/text/types.d.ts +627 -0
  912. package/src/inference/pipelines/text/types.js +4 -0
  913. package/src/inference/pipelines/text/weights.d.ts +110 -0
  914. package/src/inference/pipelines/text/weights.js +163 -0
  915. package/src/inference/pipelines/text.d.ts +157 -0
  916. package/src/inference/pipelines/text.js +586 -0
  917. package/src/inference/speculative.d.ts +239 -0
  918. package/src/inference/speculative.js +416 -0
  919. package/src/inference/test-harness.d.ts +178 -0
  920. package/src/inference/test-harness.js +349 -0
  921. package/src/inference/tokenizer.d.ts +77 -0
  922. package/src/inference/tokenizer.js +258 -0
  923. package/src/inference/tokenizers/base.d.ts +39 -0
  924. package/src/inference/tokenizers/base.js +69 -0
  925. package/src/inference/tokenizers/bpe.d.ts +27 -0
  926. package/src/inference/tokenizers/bpe.js +171 -0
  927. package/src/inference/tokenizers/bundled.d.ts +63 -0
  928. package/src/inference/tokenizers/bundled.js +866 -0
  929. package/src/inference/tokenizers/sentencepiece.d.ts +28 -0
  930. package/src/inference/tokenizers/sentencepiece.js +389 -0
  931. package/src/inference/tokenizers/types.d.ts +166 -0
  932. package/src/inference/tokenizers/types.js +7 -0
  933. package/src/loader/doppler-loader.d.ts +134 -0
  934. package/src/loader/doppler-loader.js +1036 -0
  935. package/src/loader/dtype-utils.d.ts +40 -0
  936. package/src/loader/dtype-utils.js +102 -0
  937. package/src/loader/embedding-loader.d.ts +56 -0
  938. package/src/loader/embedding-loader.js +207 -0
  939. package/src/loader/experts/expert-cache.d.ts +156 -0
  940. package/src/loader/experts/expert-cache.js +375 -0
  941. package/src/loader/experts/expert-loader.d.ts +108 -0
  942. package/src/loader/experts/expert-loader.js +384 -0
  943. package/src/loader/final-weights-loader.d.ts +68 -0
  944. package/src/loader/final-weights-loader.js +262 -0
  945. package/src/loader/index.d.ts +150 -0
  946. package/src/loader/index.js +124 -0
  947. package/src/loader/layer-loader.d.ts +63 -0
  948. package/src/loader/layer-loader.js +417 -0
  949. package/src/loader/loader-state.d.ts +51 -0
  950. package/src/loader/loader-state.js +142 -0
  951. package/src/loader/loader-types.d.ts +236 -0
  952. package/src/loader/loader-types.js +4 -0
  953. package/src/loader/manifest-config.d.ts +97 -0
  954. package/src/loader/manifest-config.js +132 -0
  955. package/src/loader/memory-monitor.d.ts +112 -0
  956. package/src/loader/memory-monitor.js +276 -0
  957. package/src/loader/multi-model-loader.d.ts +37 -0
  958. package/src/loader/multi-model-loader.js +87 -0
  959. package/src/loader/quantization-constants.d.ts +23 -0
  960. package/src/loader/quantization-constants.js +14 -0
  961. package/src/loader/shard-cache.d.ts +60 -0
  962. package/src/loader/shard-cache.js +568 -0
  963. package/src/loader/shard-resolver.d.ts +12 -0
  964. package/src/loader/shard-resolver.js +83 -0
  965. package/src/loader/tensors/tensor-loader.d.ts +154 -0
  966. package/src/loader/tensors/tensor-loader.js +427 -0
  967. package/src/loader/tensors/tensor-reader.d.ts +22 -0
  968. package/src/loader/tensors/tensor-reader.js +56 -0
  969. package/src/loader/tensors/tensor-role.d.ts +7 -0
  970. package/src/loader/tensors/tensor-role.js +12 -0
  971. package/src/loader/weight-downcast.d.ts +62 -0
  972. package/src/loader/weight-downcast.js +213 -0
  973. package/src/loader/weights.d.ts +22 -0
  974. package/src/loader/weights.js +4 -0
  975. package/src/memory/address-table.d.ts +104 -0
  976. package/src/memory/address-table.js +114 -0
  977. package/src/memory/buffer-pool.d.ts +196 -0
  978. package/src/memory/buffer-pool.js +756 -0
  979. package/src/memory/capability.d.ts +49 -0
  980. package/src/memory/capability.js +95 -0
  981. package/src/memory/heap-manager.d.ts +104 -0
  982. package/src/memory/heap-manager.js +264 -0
  983. package/src/memory/unified-detect.d.ts +59 -0
  984. package/src/memory/unified-detect.js +192 -0
  985. package/src/rules/converter/execution.rules.json +20 -0
  986. package/src/rules/converter/tensor-roles.rules.json +13 -0
  987. package/src/rules/converter/tokenizer.rules.json +7 -0
  988. package/src/rules/inference/attention.rules.json +54 -0
  989. package/src/rules/inference/config.rules.json +58 -0
  990. package/src/rules/inference/dtype.rules.json +94 -0
  991. package/src/rules/inference/execution.rules.json +45 -0
  992. package/src/rules/inference/ffn.rules.json +35 -0
  993. package/src/rules/inference/kernel-path.rules.json +76 -0
  994. package/src/rules/inference/layer-pattern.rules.json +16 -0
  995. package/src/rules/inference/layer.rules.json +7 -0
  996. package/src/rules/inference/moe.rules.json +48 -0
  997. package/src/rules/kernels/attention.rules.json +61 -0
  998. package/src/rules/kernels/conv2d.rules.json +6 -0
  999. package/src/rules/kernels/dequant.rules.json +58 -0
  1000. package/src/rules/kernels/energy.rules.json +22 -0
  1001. package/src/rules/kernels/fused-ffn.rules.json +13 -0
  1002. package/src/rules/kernels/fused-matmul-residual.rules.json +6 -0
  1003. package/src/rules/kernels/fused-matmul-rmsnorm.rules.json +8 -0
  1004. package/src/rules/kernels/gather.rules.json +12 -0
  1005. package/src/rules/kernels/gelu.rules.json +11 -0
  1006. package/src/rules/kernels/groupnorm.rules.json +10 -0
  1007. package/src/rules/kernels/kernel-validator.d.ts +24 -0
  1008. package/src/rules/kernels/kernel-validator.js +160 -0
  1009. package/src/rules/kernels/kv_quantize.rules.json +7 -0
  1010. package/src/rules/kernels/layernorm.rules.json +6 -0
  1011. package/src/rules/kernels/matmul.rules.json +60 -0
  1012. package/src/rules/kernels/modulate.rules.json +6 -0
  1013. package/src/rules/kernels/moe.rules.gptoss.json +105 -0
  1014. package/src/rules/kernels/moe.rules.json +11 -0
  1015. package/src/rules/kernels/pixel_shuffle.rules.json +6 -0
  1016. package/src/rules/kernels/residual.rules.json +12 -0
  1017. package/src/rules/kernels/rmsnorm.rules.json +11 -0
  1018. package/src/rules/kernels/rope.rules.json +6 -0
  1019. package/src/rules/kernels/sample.rules.json +6 -0
  1020. package/src/rules/kernels/scale.rules.json +6 -0
  1021. package/src/rules/kernels/silu.rules.json +21 -0
  1022. package/src/rules/kernels/softmax.rules.json +23 -0
  1023. package/src/rules/kernels/split-qkv.rules.json +6 -0
  1024. package/src/rules/kernels/upsample2d.rules.json +6 -0
  1025. package/src/rules/loader/tensor-loader.rules.json +15 -0
  1026. package/src/rules/loader/weights.rules.json +41 -0
  1027. package/src/rules/rule-registry.d.ts +48 -0
  1028. package/src/rules/rule-registry.js +177 -0
  1029. package/src/rules/tooling/command-runtime.rules.json +38 -0
  1030. package/src/storage/backends/idb-store.d.ts +52 -0
  1031. package/src/storage/backends/idb-store.js +590 -0
  1032. package/src/storage/backends/memory-store.d.ts +36 -0
  1033. package/src/storage/backends/memory-store.js +242 -0
  1034. package/src/storage/backends/opfs-store.d.ts +41 -0
  1035. package/src/storage/backends/opfs-store.js +429 -0
  1036. package/src/storage/blake3.d.ts +17 -0
  1037. package/src/storage/blake3.js +269 -0
  1038. package/src/storage/download-types.d.ts +157 -0
  1039. package/src/storage/download-types.js +48 -0
  1040. package/src/storage/downloader.d.ts +103 -0
  1041. package/src/storage/downloader.js +839 -0
  1042. package/src/storage/emulated-vram.d.ts +264 -0
  1043. package/src/storage/emulated-vram.js +576 -0
  1044. package/src/storage/export.d.ts +20 -0
  1045. package/src/storage/export.js +159 -0
  1046. package/src/storage/index.d.ts +253 -0
  1047. package/src/storage/index.js +185 -0
  1048. package/src/storage/inventory.d.ts +26 -0
  1049. package/src/storage/inventory.js +218 -0
  1050. package/src/storage/preflight.d.ts +144 -0
  1051. package/src/storage/preflight.js +294 -0
  1052. package/src/storage/quickstart-downloader.d.ts +154 -0
  1053. package/src/storage/quickstart-downloader.js +265 -0
  1054. package/src/storage/quota.d.ts +150 -0
  1055. package/src/storage/quota.js +304 -0
  1056. package/src/storage/registry.d.ts +28 -0
  1057. package/src/storage/registry.js +125 -0
  1058. package/src/storage/reports.d.ts +20 -0
  1059. package/src/storage/reports.js +94 -0
  1060. package/src/storage/shard-manager.d.ts +137 -0
  1061. package/src/storage/shard-manager.js +801 -0
  1062. package/src/sw.d.ts +1 -0
  1063. package/src/sw.js +187 -0
  1064. package/src/tooling/browser-command-runner.d.ts +28 -0
  1065. package/src/tooling/browser-command-runner.js +82 -0
  1066. package/src/tooling/command-api.d.ts +147 -0
  1067. package/src/tooling/command-api.js +523 -0
  1068. package/src/tooling/command-envelope.d.ts +81 -0
  1069. package/src/tooling/command-envelope.js +195 -0
  1070. package/src/tooling/command-runner-shared.d.ts +73 -0
  1071. package/src/tooling/command-runner-shared.js +146 -0
  1072. package/src/tooling/command-runner.html +45 -0
  1073. package/src/tooling/node-browser-command-runner.d.ts +30 -0
  1074. package/src/tooling/node-browser-command-runner.js +868 -0
  1075. package/src/tooling/node-command-runner.d.ts +36 -0
  1076. package/src/tooling/node-command-runner.js +127 -0
  1077. package/src/tooling/node-convert-worker-pool.d.ts +16 -0
  1078. package/src/tooling/node-convert-worker-pool.js +186 -0
  1079. package/src/tooling/node-convert-worker.d.ts +1 -0
  1080. package/src/tooling/node-convert-worker.js +60 -0
  1081. package/src/tooling/node-convert.d.ts +44 -0
  1082. package/src/tooling/node-converter.d.ts +1 -0
  1083. package/src/tooling/node-converter.js +1227 -0
  1084. package/src/tooling/node-file-fetch.d.ts +1 -0
  1085. package/src/tooling/node-file-fetch.js +38 -0
  1086. package/src/tooling/node-source-runtime.d.ts +19 -0
  1087. package/src/tooling/node-source-runtime.js +469 -0
  1088. package/src/tooling/node-webgpu.d.ts +6 -0
  1089. package/src/tooling/node-webgpu.js +321 -0
  1090. package/src/tooling/opfs-cache.d.ts +11 -0
  1091. package/src/tooling/opfs-cache.js +174 -0
  1092. package/src/tooling/source-runtime-bundle.d.ts +102 -0
  1093. package/src/tooling/source-runtime-bundle.js +484 -0
  1094. package/src/tooling-exports.browser.d.ts +7 -0
  1095. package/src/tooling-exports.browser.js +2 -0
  1096. package/src/tooling-exports.d.ts +22 -0
  1097. package/src/tooling-exports.js +7 -0
  1098. package/src/tooling-exports.shared.d.ts +105 -0
  1099. package/src/tooling-exports.shared.js +92 -0
  1100. package/src/training/README.md +153 -0
  1101. package/src/training/artifacts.d.ts +160 -0
  1102. package/src/training/artifacts.js +896 -0
  1103. package/src/training/attention-backward.d.ts +30 -0
  1104. package/src/training/attention-backward.js +217 -0
  1105. package/src/training/attention-forward.d.ts +22 -0
  1106. package/src/training/attention-forward.js +82 -0
  1107. package/src/training/autograd.d.ts +51 -0
  1108. package/src/training/autograd.js +380 -0
  1109. package/src/training/checkpoint.d.ts +31 -0
  1110. package/src/training/checkpoint.js +238 -0
  1111. package/src/training/clip.d.ts +9 -0
  1112. package/src/training/clip.js +54 -0
  1113. package/src/training/dataloader.d.ts +8 -0
  1114. package/src/training/dataloader.js +44 -0
  1115. package/src/training/datasets/index.d.ts +12 -0
  1116. package/src/training/datasets/index.js +6 -0
  1117. package/src/training/datasets/jsonl.d.ts +11 -0
  1118. package/src/training/datasets/jsonl.js +50 -0
  1119. package/src/training/datasets/reploid.d.ts +3 -0
  1120. package/src/training/datasets/reploid.js +36 -0
  1121. package/src/training/datasets/text-pairs.d.ts +21 -0
  1122. package/src/training/datasets/text-pairs.js +42 -0
  1123. package/src/training/datasets/token-batch.d.ts +21 -0
  1124. package/src/training/datasets/token-batch.js +40 -0
  1125. package/src/training/datasets/translation-pairs.d.ts +34 -0
  1126. package/src/training/datasets/translation-pairs.js +49 -0
  1127. package/src/training/export.d.ts +32 -0
  1128. package/src/training/export.js +112 -0
  1129. package/src/training/index.d.ts +52 -0
  1130. package/src/training/index.js +41 -0
  1131. package/src/training/lora.d.ts +19 -0
  1132. package/src/training/lora.js +57 -0
  1133. package/src/training/loss-scaling.d.ts +21 -0
  1134. package/src/training/loss-scaling.js +80 -0
  1135. package/src/training/loss.d.ts +10 -0
  1136. package/src/training/loss.js +41 -0
  1137. package/src/training/objectives/base.d.ts +58 -0
  1138. package/src/training/objectives/base.js +38 -0
  1139. package/src/training/objectives/cross_entropy.d.ts +18 -0
  1140. package/src/training/objectives/cross_entropy.js +37 -0
  1141. package/src/training/objectives/distill_kd.d.ts +16 -0
  1142. package/src/training/objectives/distill_kd.js +369 -0
  1143. package/src/training/objectives/distill_triplet.d.ts +16 -0
  1144. package/src/training/objectives/distill_triplet.js +412 -0
  1145. package/src/training/objectives/index.d.ts +12 -0
  1146. package/src/training/objectives/index.js +6 -0
  1147. package/src/training/objectives/ul_stage1_joint.d.ts +16 -0
  1148. package/src/training/objectives/ul_stage1_joint.js +188 -0
  1149. package/src/training/objectives/ul_stage2_base.d.ts +16 -0
  1150. package/src/training/objectives/ul_stage2_base.js +222 -0
  1151. package/src/training/optimizer.d.ts +22 -0
  1152. package/src/training/optimizer.js +115 -0
  1153. package/src/training/runner.d.ts +196 -0
  1154. package/src/training/runner.js +1194 -0
  1155. package/src/training/suite.d.ts +187 -0
  1156. package/src/training/suite.js +3156 -0
  1157. package/src/training/trainer.d.ts +89 -0
  1158. package/src/training/trainer.js +301 -0
  1159. package/src/training/ul_dataset.d.ts +47 -0
  1160. package/src/training/ul_dataset.js +153 -0
  1161. package/src/training/ul_schedule.d.ts +6 -0
  1162. package/src/training/ul_schedule.js +29 -0
  1163. package/src/types/chrome.d.ts +36 -0
  1164. package/src/types/chrome.js +1 -0
  1165. package/src/types/gpu.d.ts +185 -0
  1166. package/src/types/gpu.js +5 -0
  1167. package/src/types/index.d.ts +3 -0
  1168. package/src/types/index.js +3 -0
  1169. package/src/types/inference.d.ts +197 -0
  1170. package/src/types/inference.js +5 -0
  1171. package/src/types/model.d.ts +125 -0
  1172. package/src/types/model.js +5 -0
  1173. package/src/utils/index.d.ts +7 -0
  1174. package/src/utils/index.js +7 -0
  1175. package/src/utils/load-json.d.ts +5 -0
  1176. package/src/utils/load-json.js +23 -0
  1177. package/src/utils/plain-object.d.ts +1 -0
  1178. package/src/utils/plain-object.js +3 -0
  1179. package/src/utils/sha256.d.ts +4 -0
  1180. package/src/utils/sha256.js +135 -0
  1181. package/tools/convert-safetensors-node.js +180 -0
  1182. package/tools/doppler-cli.js +1170 -0
package/src/config/presets/kernel-paths/gemma3-f16-fused-f16a-online.json ADDED
@@ -0,0 +1,200 @@
1
+ {
2
+ "id": "gemma3-f16-fused-f16a-online",
3
+ "name": "Gemma 3 F16 (F16 activations, fused online)",
4
+ "description": "Online-tuned Gemma 3 F16 path with decode lm_head retile overrides and explicitly defined Fused FFN kernels.",
5
+ "activationDtype": "f16",
6
+ "decode": {
7
+ "steps": [
8
+ {
9
+ "op": "input_norm",
10
+ "kernel": "rmsnorm_f16.wgsl",
11
+ "entry": "main"
12
+ },
13
+ {
14
+ "op": "q_proj",
15
+ "kernel": "matmul_gemv_subgroup_f16a.wgsl",
16
+ "entry": "main_vec4",
17
+ "weights": "layer.{L}.self_attn.q_proj"
18
+ },
19
+ {
20
+ "op": "k_proj",
21
+ "kernel": "matmul_gemv_subgroup_f16a.wgsl",
22
+ "entry": "main_vec4",
23
+ "weights": "layer.{L}.self_attn.k_proj"
24
+ },
25
+ {
26
+ "op": "v_proj",
27
+ "kernel": "matmul_gemv_subgroup_f16a.wgsl",
28
+ "entry": "main_vec4",
29
+ "weights": "layer.{L}.self_attn.v_proj"
30
+ },
31
+ {
32
+ "op": "rope_q",
33
+ "kernel": "rope_f16.wgsl",
34
+ "entry": "main"
35
+ },
36
+ {
37
+ "op": "rope_k",
38
+ "kernel": "rope_f16.wgsl",
39
+ "entry": "main"
40
+ },
41
+ {
42
+ "op": "attention",
43
+ "kernel": "attention_decode_online_f16.wgsl",
44
+ "entry": "main"
45
+ },
46
+ {
47
+ "op": "o_proj",
48
+ "kernel": "matmul_gemv_subgroup_f16a.wgsl",
49
+ "entry": "main_vec4",
50
+ "weights": "layer.{L}.self_attn.o_proj"
51
+ },
52
+ {
53
+ "op": "attn_residual",
54
+ "kernel": "residual_f16.wgsl",
55
+ "entry": "main"
56
+ },
57
+ {
58
+ "op": "post_attn_norm",
59
+ "kernel": "rmsnorm_f16.wgsl",
60
+ "entry": "main"
61
+ },
62
+ {
63
+ "op": "ffn_gate_up",
64
+ "kernel": "fused_ffn_f16.wgsl",
65
+ "entry": "main",
66
+ "weights": "layer.{L}.mlp",
67
+ "constants": {
68
+ "ACTIVATION": 1
69
+ }
70
+ },
71
+ {
72
+ "op": "down_proj",
73
+ "kernel": "matmul_gemv_subgroup_f16a.wgsl",
74
+ "entry": "main_vec4",
75
+ "weights": "layer.{L}.mlp.down_proj"
76
+ },
77
+ {
78
+ "op": "ffn_residual",
79
+ "kernel": "residual_f16.wgsl",
80
+ "entry": "main"
81
+ }
82
+ ]
83
+ },
84
+ "prefill": {
85
+ "steps": [
86
+ {
87
+ "op": "input_norm",
88
+ "kernel": "rmsnorm_f16.wgsl",
89
+ "entry": "main"
90
+ },
91
+ {
92
+ "op": "q_proj",
93
+ "kernel": "matmul_f16_tiled.wgsl",
94
+ "entry": "main",
95
+ "weights": "layer.{L}.self_attn.q_proj"
96
+ },
97
+ {
98
+ "op": "k_proj",
99
+ "kernel": "matmul_f16_tiled.wgsl",
100
+ "entry": "main",
101
+ "weights": "layer.{L}.self_attn.k_proj"
102
+ },
103
+ {
104
+ "op": "v_proj",
105
+ "kernel": "matmul_f16_tiled.wgsl",
106
+ "entry": "main",
107
+ "weights": "layer.{L}.self_attn.v_proj"
108
+ },
109
+ {
110
+ "op": "rope_q",
111
+ "kernel": "rope_f16.wgsl",
112
+ "entry": "main"
113
+ },
114
+ {
115
+ "op": "rope_k",
116
+ "kernel": "rope_f16.wgsl",
117
+ "entry": "main"
118
+ },
119
+ {
120
+ "op": "attention",
121
+ "kernel": "attention_small_f16.wgsl",
122
+ "entry": "main"
123
+ },
124
+ {
125
+ "op": "o_proj",
126
+ "kernel": "matmul_f16_tiled.wgsl",
127
+ "entry": "main",
128
+ "weights": "layer.{L}.self_attn.o_proj"
129
+ },
130
+ {
131
+ "op": "attn_residual",
132
+ "kernel": "residual_f16.wgsl",
133
+ "entry": "main"
134
+ },
135
+ {
136
+ "op": "post_attn_norm",
137
+ "kernel": "rmsnorm_f16.wgsl",
138
+ "entry": "main"
139
+ },
140
+ {
141
+ "op": "ffn_gate_up",
142
+ "kernel": "fused_ffn_f16.wgsl",
143
+ "entry": "main_batched",
144
+ "weights": "layer.{L}.mlp",
145
+ "constants": {
146
+ "ACTIVATION": 1
147
+ }
148
+ },
149
+ {
150
+ "op": "down_proj",
151
+ "kernel": "matmul_f16_tiled.wgsl",
152
+ "entry": "main",
153
+ "weights": "layer.{L}.mlp.down_proj"
154
+ },
155
+ {
156
+ "op": "ffn_residual",
157
+ "kernel": "residual_f16.wgsl",
158
+ "entry": "main"
159
+ }
160
+ ]
161
+ },
162
+ "preLayer": [
163
+ {
164
+ "op": "embed",
165
+ "kernel": "gather_f16.wgsl",
166
+ "entry": "main",
167
+ "weights": "embed_tokens"
168
+ }
169
+ ],
170
+ "postLayer": [
171
+ {
172
+ "op": "final_norm",
173
+ "kernel": "rmsnorm_f16.wgsl",
174
+ "entry": "main"
175
+ },
176
+ {
177
+ "op": "lm_head",
178
+ "kernel": "matmul_gemv_subgroup_f16a.wgsl",
179
+ "entry": "main_multicol",
180
+ "weights": "lm_head",
181
+ "constants": {
182
+ "MULTICOL_COLS_PER_WG": 64,
183
+ "MULTICOL_THREADS_PER_COL": 4
184
+ }
185
+ },
186
+ {
187
+ "op": "lm_head_prefill",
188
+ "kernel": "matmul_f16_tiled.wgsl",
189
+ "entry": "main",
190
+ "weights": "lm_head"
191
+ }
192
+ ],
193
+ "sampling": [
194
+ {
195
+ "op": "sample",
196
+ "kernel": "sample_f16.wgsl",
197
+ "entry": "sample_single_pass"
198
+ }
199
+ ]
200
+ }
package/src/config/presets/kernel-paths/gemma3-f16-fused-f32a-online.json ADDED
@@ -0,0 +1,223 @@
1
+ {
2
+ "id": "gemma3-f16-fused-f32a-online",
3
+ "name": "Gemma 3 F16 (F32 activations, online)",
4
+ "description": "F16 weights with F32 activations, online decode attention, and explicit gate/up matmul kernel overrides.",
5
+ "activationDtype": "f32",
6
+ "kvDtype": "f16",
7
+ "decode": {
8
+ "steps": [
9
+ {
10
+ "op": "input_norm",
11
+ "kernel": "rmsnorm.wgsl",
12
+ "entry": "main"
13
+ },
14
+ {
15
+ "op": "q_proj",
16
+ "kernel": "matmul_gemv_subgroup.wgsl",
17
+ "entry": "main_vec4",
18
+ "weights": "layer.{L}.self_attn.q_proj"
19
+ },
20
+ {
21
+ "op": "k_proj",
22
+ "kernel": "matmul_gemv_subgroup.wgsl",
23
+ "entry": "main_vec4",
24
+ "weights": "layer.{L}.self_attn.k_proj"
25
+ },
26
+ {
27
+ "op": "v_proj",
28
+ "kernel": "matmul_gemv_subgroup.wgsl",
29
+ "entry": "main_vec4",
30
+ "weights": "layer.{L}.self_attn.v_proj"
31
+ },
32
+ {
33
+ "op": "rope_q",
34
+ "kernel": "rope.wgsl",
35
+ "entry": "main"
36
+ },
37
+ {
38
+ "op": "rope_k",
39
+ "kernel": "rope.wgsl",
40
+ "entry": "main"
41
+ },
42
+ {
43
+ "op": "attention",
44
+ "kernel": "attention_decode_online_f16kv.wgsl",
45
+ "entry": "main"
46
+ },
47
+ {
48
+ "op": "o_proj",
49
+ "kernel": "matmul_gemv_subgroup.wgsl",
50
+ "entry": "main_vec4",
51
+ "weights": "layer.{L}.self_attn.o_proj"
52
+ },
53
+ {
54
+ "op": "attn_residual",
55
+ "kernel": "residual.wgsl",
56
+ "entry": "main"
57
+ },
58
+ {
59
+ "op": "post_attn_norm",
60
+ "kernel": "rmsnorm.wgsl",
61
+ "entry": "main"
62
+ },
63
+ {
64
+ "op": "gate_proj",
65
+ "kernel": "matmul_gemv_subgroup.wgsl",
66
+ "entry": "main_vec4",
67
+ "weights": "layer.{L}.mlp.gate_proj"
68
+ },
69
+ {
70
+ "op": "up_proj",
71
+ "kernel": "matmul_gemv_subgroup.wgsl",
72
+ "entry": "main_vec4",
73
+ "weights": "layer.{L}.mlp.up_proj"
74
+ },
75
+ {
76
+ "op": "activation",
77
+ "kernel": "gelu.wgsl",
78
+ "entry": "main",
79
+ "constants": {
80
+ "HAS_GATE": true
81
+ }
82
+ },
83
+ {
84
+ "op": "down_proj",
85
+ "kernel": "matmul_gemv_subgroup.wgsl",
86
+ "entry": "main_vec4",
87
+ "weights": "layer.{L}.mlp.down_proj"
88
+ },
89
+ {
90
+ "op": "ffn_residual",
91
+ "kernel": "residual.wgsl",
92
+ "entry": "main"
93
+ }
94
+ ]
95
+ },
96
+ "prefill": {
97
+ "steps": [
98
+ {
99
+ "op": "input_norm",
100
+ "kernel": "rmsnorm.wgsl",
101
+ "entry": "main"
102
+ },
103
+ {
104
+ "op": "q_proj",
105
+ "kernel": "matmul_f16w_f32a_tiled.wgsl",
106
+ "entry": "main",
107
+ "weights": "layer.{L}.self_attn.q_proj"
108
+ },
109
+ {
110
+ "op": "k_proj",
111
+ "kernel": "matmul_f16w_f32a_tiled.wgsl",
112
+ "entry": "main",
113
+ "weights": "layer.{L}.self_attn.k_proj"
114
+ },
115
+ {
116
+ "op": "v_proj",
117
+ "kernel": "matmul_f16w_f32a_tiled.wgsl",
118
+ "entry": "main",
119
+ "weights": "layer.{L}.self_attn.v_proj"
120
+ },
121
+ {
122
+ "op": "rope_q",
123
+ "kernel": "rope.wgsl",
124
+ "entry": "main"
125
+ },
126
+ {
127
+ "op": "rope_k",
128
+ "kernel": "rope.wgsl",
129
+ "entry": "main"
130
+ },
131
+ {
132
+ "op": "attention",
133
+ "kernel": "attention_small_f16kv.wgsl",
134
+ "entry": "main"
135
+ },
136
+ {
137
+ "op": "o_proj",
138
+ "kernel": "matmul_f16w_f32a_tiled.wgsl",
139
+ "entry": "main",
140
+ "weights": "layer.{L}.self_attn.o_proj"
141
+ },
142
+ {
143
+ "op": "attn_residual",
144
+ "kernel": "residual.wgsl",
145
+ "entry": "main"
146
+ },
147
+ {
148
+ "op": "post_attn_norm",
149
+ "kernel": "rmsnorm.wgsl",
150
+ "entry": "main"
151
+ },
152
+ {
153
+ "op": "gate_proj",
154
+ "kernel": "matmul_f16w_f32a_tiled.wgsl",
155
+ "entry": "main",
156
+ "weights": "layer.{L}.mlp.gate_proj"
157
+ },
158
+ {
159
+ "op": "up_proj",
160
+ "kernel": "matmul_f16w_f32a_tiled.wgsl",
161
+ "entry": "main",
162
+ "weights": "layer.{L}.mlp.up_proj"
163
+ },
164
+ {
165
+ "op": "activation",
166
+ "kernel": "gelu.wgsl",
167
+ "entry": "main",
168
+ "constants": {
169
+ "HAS_GATE": true
170
+ }
171
+ },
172
+ {
173
+ "op": "down_proj",
174
+ "kernel": "matmul_f16w_f32a_tiled.wgsl",
175
+ "entry": "main",
176
+ "weights": "layer.{L}.mlp.down_proj"
177
+ },
178
+ {
179
+ "op": "ffn_residual",
180
+ "kernel": "residual.wgsl",
181
+ "entry": "main"
182
+ }
183
+ ]
184
+ },
185
+ "preLayer": [
186
+ {
187
+ "op": "embed",
188
+ "kernel": "gather.wgsl",
189
+ "entry": "main",
190
+ "weights": "embed_tokens"
191
+ }
192
+ ],
193
+ "postLayer": [
194
+ {
195
+ "op": "final_norm",
196
+ "kernel": "rmsnorm.wgsl",
197
+ "entry": "main"
198
+ },
199
+ {
200
+ "op": "lm_head",
201
+ "kernel": "matmul_gemv_subgroup.wgsl",
202
+ "entry": "main_multicol",
203
+ "weights": "lm_head",
204
+ "constants": {
205
+ "MULTICOL_COLS_PER_WG": 64,
206
+ "MULTICOL_THREADS_PER_COL": 4
207
+ }
208
+ },
209
+ {
210
+ "op": "lm_head_prefill",
211
+ "kernel": "matmul_f16w_f32a_tiled.wgsl",
212
+ "entry": "main",
213
+ "weights": "lm_head"
214
+ }
215
+ ],
216
+ "sampling": [
217
+ {
218
+ "op": "sample",
219
+ "kernel": "sample.wgsl",
220
+ "entry": "sample_single_pass"
221
+ }
222
+ ]
223
+ }
package/src/config/presets/kernel-paths/gemma3-q4k-dequant-f16a-online.json ADDED
@@ -0,0 +1,60 @@
1
+ {
2
+ "id": "gemma3-q4k-dequant-f16a-online",
3
+ "name": "Gemma 3 Q4K Dequant (F16 activations, online decode)",
4
+ "description": "Q4K dequantized to F16 with online decode attention (subgroup-capable devices).",
5
+ "activationDtype": "f16",
6
+
7
+ "decode": {
8
+ "steps": [
9
+ { "op": "input_norm", "kernel": "rmsnorm_f16.wgsl", "entry": "main" },
10
+ { "op": "q_proj", "kernel": "matmul_gemv_subgroup_f16a.wgsl", "entry": "main_vec4", "weights": "layer.{L}.self_attn.q_proj" },
11
+ { "op": "k_proj", "kernel": "matmul_gemv_subgroup_f16a.wgsl", "entry": "main_vec4", "weights": "layer.{L}.self_attn.k_proj" },
12
+ { "op": "v_proj", "kernel": "matmul_gemv_subgroup_f16a.wgsl", "entry": "main_vec4", "weights": "layer.{L}.self_attn.v_proj" },
13
+ { "op": "rope_q", "kernel": "rope_f16.wgsl", "entry": "main" },
14
+ { "op": "rope_k", "kernel": "rope_f16.wgsl", "entry": "main" },
15
+ { "op": "attention", "kernel": "attention_decode_online_f16.wgsl", "entry": "main" },
16
+ { "op": "o_proj", "kernel": "matmul_gemv_subgroup_f16a.wgsl", "entry": "main_vec4", "weights": "layer.{L}.self_attn.o_proj" },
17
+ { "op": "attn_residual","kernel": "residual_f16.wgsl", "entry": "main" },
18
+ { "op": "post_attn_norm","kernel": "rmsnorm_f16.wgsl", "entry": "main" },
19
+ { "op": "gate_proj", "kernel": "matmul_gemv_subgroup_f16a.wgsl", "entry": "main_vec4", "weights": "layer.{L}.mlp.gate_proj" },
20
+ { "op": "up_proj", "kernel": "matmul_gemv_subgroup_f16a.wgsl", "entry": "main_vec4", "weights": "layer.{L}.mlp.up_proj" },
21
+ { "op": "activation", "kernel": "gelu_f16.wgsl", "entry": "main", "constants": { "HAS_GATE": true } },
22
+ { "op": "down_proj", "kernel": "matmul_gemv_subgroup_f16a.wgsl", "entry": "main_vec4", "weights": "layer.{L}.mlp.down_proj" },
23
+ { "op": "ffn_residual", "kernel": "residual_f16.wgsl", "entry": "main" }
24
+ ]
25
+ },
26
+
27
+ "prefill": {
28
+ "steps": [
29
+ { "op": "input_norm", "kernel": "rmsnorm_f16.wgsl", "entry": "main" },
30
+ { "op": "q_proj", "kernel": "matmul_f16.wgsl", "entry": "main", "weights": "layer.{L}.self_attn.q_proj" },
31
+ { "op": "k_proj", "kernel": "matmul_f16.wgsl", "entry": "main", "weights": "layer.{L}.self_attn.k_proj" },
32
+ { "op": "v_proj", "kernel": "matmul_f16.wgsl", "entry": "main", "weights": "layer.{L}.self_attn.v_proj" },
33
+ { "op": "rope_q", "kernel": "rope_f16.wgsl", "entry": "main" },
34
+ { "op": "rope_k", "kernel": "rope_f16.wgsl", "entry": "main" },
35
+ { "op": "attention", "kernel": "attention_streaming_f16.wgsl", "entry": "main" },
36
+ { "op": "o_proj", "kernel": "matmul_f16.wgsl", "entry": "main", "weights": "layer.{L}.self_attn.o_proj" },
37
+ { "op": "attn_residual","kernel": "residual_f16.wgsl", "entry": "main" },
38
+ { "op": "post_attn_norm","kernel": "rmsnorm_f16.wgsl", "entry": "main" },
39
+ { "op": "gate_proj", "kernel": "matmul_f16.wgsl", "entry": "main", "weights": "layer.{L}.mlp.gate_proj" },
40
+ { "op": "up_proj", "kernel": "matmul_f16.wgsl", "entry": "main", "weights": "layer.{L}.mlp.up_proj" },
41
+ { "op": "activation", "kernel": "gelu_f16.wgsl", "entry": "main", "constants": { "HAS_GATE": true } },
42
+ { "op": "down_proj", "kernel": "matmul_f16.wgsl", "entry": "main", "weights": "layer.{L}.mlp.down_proj" },
43
+ { "op": "ffn_residual", "kernel": "residual_f16.wgsl", "entry": "main" }
44
+ ]
45
+ },
46
+
47
+ "preLayer": [
48
+ { "op": "embed", "kernel": "gather_f16.wgsl", "entry": "main", "weights": "embed_tokens" }
49
+ ],
50
+
51
+ "postLayer": [
52
+ { "op": "final_norm", "kernel": "rmsnorm_f16.wgsl", "entry": "main" },
53
+ { "op": "lm_head", "kernel": "matmul_gemv_subgroup_f16a.wgsl", "entry": "main_multicol", "weights": "lm_head", "constants": { "MULTICOL_COLS_PER_WG": 64, "MULTICOL_THREADS_PER_COL": 4 } },
54
+ { "op": "lm_head_prefill", "kernel": "matmul_f16.wgsl", "entry": "main", "weights": "lm_head" }
55
+ ],
56
+
57
+ "sampling": [
58
+ { "op": "sample", "kernel": "sample_f16.wgsl", "entry": "sample_single_pass" }
59
+ ]
60
+ }
package/src/config/presets/kernel-paths/gemma3-q4k-dequant-f32a-online.json ADDED
@@ -0,0 +1,61 @@
1
+ {
2
+ "id": "gemma3-q4k-dequant-f32a-online",
3
+ "name": "Gemma 3 Q4K Dequant (F32 activations, online decode)",
4
+ "description": "Q4K dequantized to F16 with F32 activations. Decode uses subgroup GEMV and online attention; prefill uses tiled matmul.",
5
+ "activationDtype": "f32",
6
+ "kvDtype": "f16",
7
+
8
+ "decode": {
9
+ "steps": [
10
+ { "op": "input_norm", "kernel": "rmsnorm.wgsl", "entry": "main" },
11
+ { "op": "q_proj", "kernel": "matmul_gemv_subgroup.wgsl", "entry": "main_vec4", "weights": "layer.{L}.self_attn.q_proj" },
12
+ { "op": "k_proj", "kernel": "matmul_gemv_subgroup.wgsl", "entry": "main_vec4", "weights": "layer.{L}.self_attn.k_proj" },
13
+ { "op": "v_proj", "kernel": "matmul_gemv_subgroup.wgsl", "entry": "main_vec4", "weights": "layer.{L}.self_attn.v_proj" },
14
+ { "op": "rope_q", "kernel": "rope.wgsl", "entry": "main" },
15
+ { "op": "rope_k", "kernel": "rope.wgsl", "entry": "main" },
16
+ { "op": "attention", "kernel": "attention_decode_online_f16kv.wgsl", "entry": "main" },
17
+ { "op": "o_proj", "kernel": "matmul_gemv_subgroup.wgsl", "entry": "main_vec4", "weights": "layer.{L}.self_attn.o_proj" },
18
+ { "op": "attn_residual", "kernel": "residual.wgsl", "entry": "main" },
19
+ { "op": "post_attn_norm","kernel": "rmsnorm.wgsl", "entry": "main" },
20
+ { "op": "gate_proj", "kernel": "matmul_gemv_subgroup.wgsl", "entry": "main_vec4", "weights": "layer.{L}.mlp.gate_proj" },
21
+ { "op": "up_proj", "kernel": "matmul_gemv_subgroup.wgsl", "entry": "main_vec4", "weights": "layer.{L}.mlp.up_proj" },
22
+ { "op": "activation", "kernel": "gelu.wgsl", "entry": "main", "constants": { "HAS_GATE": true } },
23
+ { "op": "down_proj", "kernel": "matmul_gemv_subgroup.wgsl", "entry": "main_vec4", "weights": "layer.{L}.mlp.down_proj" },
24
+ { "op": "ffn_residual", "kernel": "residual.wgsl", "entry": "main" }
25
+ ]
26
+ },
27
+
28
+ "prefill": {
29
+ "steps": [
30
+ { "op": "input_norm", "kernel": "rmsnorm.wgsl", "entry": "main" },
31
+ { "op": "q_proj", "kernel": "matmul_f16w_f32a.wgsl", "entry": "main", "weights": "layer.{L}.self_attn.q_proj" },
32
+ { "op": "k_proj", "kernel": "matmul_f16w_f32a.wgsl", "entry": "main", "weights": "layer.{L}.self_attn.k_proj" },
33
+ { "op": "v_proj", "kernel": "matmul_f16w_f32a.wgsl", "entry": "main", "weights": "layer.{L}.self_attn.v_proj" },
34
+ { "op": "rope_q", "kernel": "rope.wgsl", "entry": "main" },
35
+ { "op": "rope_k", "kernel": "rope.wgsl", "entry": "main" },
36
+ { "op": "attention", "kernel": "attention_streaming_f16kv.wgsl", "entry": "main" },
37
+ { "op": "o_proj", "kernel": "matmul_f16w_f32a.wgsl", "entry": "main", "weights": "layer.{L}.self_attn.o_proj" },
38
+ { "op": "attn_residual", "kernel": "residual.wgsl", "entry": "main" },
39
+ { "op": "post_attn_norm","kernel": "rmsnorm.wgsl", "entry": "main" },
40
+ { "op": "gate_proj", "kernel": "matmul_f16w_f32a.wgsl", "entry": "main", "weights": "layer.{L}.mlp.gate_proj" },
41
+ { "op": "up_proj", "kernel": "matmul_f16w_f32a.wgsl", "entry": "main", "weights": "layer.{L}.mlp.up_proj" },
42
+ { "op": "activation", "kernel": "gelu.wgsl", "entry": "main", "constants": { "HAS_GATE": true } },
43
+ { "op": "down_proj", "kernel": "matmul_f16w_f32a.wgsl", "entry": "main", "weights": "layer.{L}.mlp.down_proj" },
44
+ { "op": "ffn_residual", "kernel": "residual.wgsl", "entry": "main" }
45
+ ]
46
+ },
47
+
48
+ "preLayer": [
49
+ { "op": "embed", "kernel": "gather_f16.wgsl", "entry": "main", "weights": "embed_tokens" }
50
+ ],
51
+
52
+ "postLayer": [
53
+ { "op": "final_norm", "kernel": "rmsnorm.wgsl", "entry": "main" },
54
+ { "op": "lm_head", "kernel": "matmul_gemv_subgroup.wgsl", "entry": "main_multicol", "weights": "lm_head", "constants": { "MULTICOL_COLS_PER_WG": 64, "MULTICOL_THREADS_PER_COL": 4 } },
55
+ { "op": "lm_head_prefill", "kernel": "matmul_f16w_f32a.wgsl", "entry": "main", "weights": "lm_head" }
56
+ ],
57
+
58
+ "sampling": [
59
+ { "op": "sample", "kernel": "sample.wgsl", "entry": "sample_single_pass" }
60
+ ]
61
+ }
package/src/config/presets/kernel-paths/gemma3-q4k-dequant-f32a.json ADDED
@@ -0,0 +1,61 @@
1
+ {
2
+ "id": "gemma3-q4k-dequant-f32a",
3
+ "name": "Gemma 3 Q4K Dequant (F32 activations)",
4
+ "description": "Q4K weights dequantized to F16 with F32 activations for improved stability on small Gemma 3 checkpoints.",
5
+ "activationDtype": "f32",
6
+ "kvDtype": "f16",
7
+
8
+ "decode": {
9
+ "steps": [
10
+ { "op": "input_norm", "kernel": "rmsnorm.wgsl", "entry": "main" },
11
+ { "op": "q_proj", "kernel": "matmul_f16w_f32a.wgsl", "entry": "main", "weights": "layer.{L}.self_attn.q_proj" },
12
+ { "op": "k_proj", "kernel": "matmul_f16w_f32a.wgsl", "entry": "main", "weights": "layer.{L}.self_attn.k_proj" },
13
+ { "op": "v_proj", "kernel": "matmul_f16w_f32a.wgsl", "entry": "main", "weights": "layer.{L}.self_attn.v_proj" },
14
+ { "op": "rope_q", "kernel": "rope.wgsl", "entry": "main" },
15
+ { "op": "rope_k", "kernel": "rope.wgsl", "entry": "main" },
16
+ { "op": "attention", "kernel": "attention_decode_chunked_f16kv.wgsl", "entry": "main" },
17
+ { "op": "o_proj", "kernel": "matmul_f16w_f32a.wgsl", "entry": "main", "weights": "layer.{L}.self_attn.o_proj" },
18
+ { "op": "attn_residual","kernel": "residual.wgsl", "entry": "main" },
19
+ { "op": "post_attn_norm","kernel": "rmsnorm.wgsl", "entry": "main" },
20
+ { "op": "gate_proj", "kernel": "matmul_f16w_f32a.wgsl", "entry": "main", "weights": "layer.{L}.mlp.gate_proj" },
21
+ { "op": "up_proj", "kernel": "matmul_f16w_f32a.wgsl", "entry": "main", "weights": "layer.{L}.mlp.up_proj" },
22
+ { "op": "activation", "kernel": "gelu.wgsl", "entry": "main", "constants": { "HAS_GATE": true } },
23
+ { "op": "down_proj", "kernel": "matmul_f16w_f32a.wgsl", "entry": "main", "weights": "layer.{L}.mlp.down_proj" },
24
+ { "op": "ffn_residual", "kernel": "residual.wgsl", "entry": "main" }
25
+ ]
26
+ },
27
+
28
+ "prefill": {
29
+ "steps": [
30
+ { "op": "input_norm", "kernel": "rmsnorm.wgsl", "entry": "main" },
31
+ { "op": "q_proj", "kernel": "matmul_f16w_f32a.wgsl", "entry": "main", "weights": "layer.{L}.self_attn.q_proj" },
32
+ { "op": "k_proj", "kernel": "matmul_f16w_f32a.wgsl", "entry": "main", "weights": "layer.{L}.self_attn.k_proj" },
33
+ { "op": "v_proj", "kernel": "matmul_f16w_f32a.wgsl", "entry": "main", "weights": "layer.{L}.self_attn.v_proj" },
34
+ { "op": "rope_q", "kernel": "rope.wgsl", "entry": "main" },
35
+ { "op": "rope_k", "kernel": "rope.wgsl", "entry": "main" },
36
+ { "op": "attention", "kernel": "attention_streaming_f16kv.wgsl", "entry": "main" },
37
+ { "op": "o_proj", "kernel": "matmul_f16w_f32a.wgsl", "entry": "main", "weights": "layer.{L}.self_attn.o_proj" },
38
+ { "op": "attn_residual","kernel": "residual.wgsl", "entry": "main" },
39
+ { "op": "post_attn_norm","kernel": "rmsnorm.wgsl", "entry": "main" },
40
+ { "op": "gate_proj", "kernel": "matmul_f16w_f32a.wgsl", "entry": "main", "weights": "layer.{L}.mlp.gate_proj" },
41
+ { "op": "up_proj", "kernel": "matmul_f16w_f32a.wgsl", "entry": "main", "weights": "layer.{L}.mlp.up_proj" },
42
+ { "op": "activation", "kernel": "gelu.wgsl", "entry": "main", "constants": { "HAS_GATE": true } },
43
+ { "op": "down_proj", "kernel": "matmul_f16w_f32a.wgsl", "entry": "main", "weights": "layer.{L}.mlp.down_proj" },
44
+ { "op": "ffn_residual", "kernel": "residual.wgsl", "entry": "main" }
45
+ ]
46
+ },
47
+
48
+ "preLayer": [
49
+ { "op": "embed", "kernel": "gather_f16.wgsl", "entry": "main", "weights": "embed_tokens" }
50
+ ],
51
+
52
+ "postLayer": [
53
+ { "op": "final_norm", "kernel": "rmsnorm.wgsl", "entry": "main" },
54
+ { "op": "lm_head", "kernel": "matmul_f16w_f32a.wgsl", "entry": "main", "weights": "lm_head" },
55
+ { "op": "lm_head_prefill", "kernel": "matmul_f16w_f32a.wgsl", "entry": "main", "weights": "lm_head" }
56
+ ],
57
+
58
+ "sampling": [
59
+ { "op": "sample", "kernel": "sample.wgsl", "entry": "sample_single_pass" }
60
+ ]
61
+ }