doppler-gpu 0.4.3 → 0.4.4

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 (433) hide show
  1. package/CHANGELOG.md +28 -0
  2. package/README.md +54 -33
  3. package/assets/doppler.svg +57 -1601
  4. package/package.json +36 -5
  5. package/src/cli/cli-output.js +34 -1
  6. package/src/cli/doppler-cli.d.ts +18 -0
  7. package/src/cli/doppler-cli.js +68 -35
  8. package/src/cli/doppler-serve.d.ts +210 -0
  9. package/src/cli/doppler-serve.js +468 -42
  10. package/src/client/doppler-registry.json +3 -5
  11. package/src/client/provider.d.ts +1 -1
  12. package/src/client/provider.js +1 -1
  13. package/src/client/runtime/index.js +13 -0
  14. package/src/client/runtime/lora.d.ts +11 -1
  15. package/src/client/runtime/lora.js +28 -7
  16. package/src/client/runtime/model-manager.js +35 -11
  17. package/src/client/runtime/model-session.d.ts +2 -1
  18. package/src/client/runtime/model-session.js +2 -2
  19. package/src/client/runtime/model-source.js +20 -1
  20. package/src/client/runtime/types.d.ts +8 -0
  21. package/src/client/wrap-pipeline-handle.d.ts +62 -0
  22. package/src/client/wrap-pipeline-handle.js +188 -13
  23. package/src/config/conversion/README.md +5 -0
  24. package/src/config/conversion/diffusiongemma/diffusiongemma-26b-a4b-it-q4k-ehf16-af16.json +646 -0
  25. package/src/config/conversion/embeddinggemma/google-embeddinggemma-300m-q4k-ehf16-af32.json +3 -3
  26. package/src/config/conversion/gemma3/gemma-3-1b-it-f16-af32.json +4 -4
  27. package/src/config/conversion/gemma3/gemma-3-1b-it-q4k-ehf16-af32.json +4 -4
  28. package/src/config/conversion/gemma3/gemma-3-270m-it-f16-af32.json +393 -0
  29. package/src/config/conversion/gemma3/gemma-3-270m-it-q4k-ehf16-af16.json +425 -0
  30. package/src/config/conversion/gemma3/gemma-3-270m-it-q4k-ehf16-af32.json +17 -9
  31. package/src/config/conversion/gemma3/translategemma-4b-1b-enes-q4k-ehf16-af32.json +4 -4
  32. package/src/config/conversion/gemma3/translategemma-4b-it-q4k-ehf16-af32.json +4 -4
  33. package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-af16.json +737 -0
  34. package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-af32.json +730 -0
  35. package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-hq4k-af16.json +780 -0
  36. package/src/config/conversion/gemma4/gemma-4-12b-it-text-w4a16-ct-ehf16-af16.json +716 -0
  37. package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af16.json +5 -5
  38. package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af32.json +4 -4
  39. package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af16-int4ple.json +14 -12
  40. package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32-int4ple.json +7 -7
  41. package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32.json +7 -7
  42. package/src/config/conversion/gemma4/gemma-4-moe-q4k-ehf16-af32.json +3 -3
  43. package/src/config/conversion/gpt-oss-20b-f16-xmxfp4.json +3 -3
  44. package/src/config/conversion/janus/janus-pro-1b-text-q4k-ehaf16.json +3 -3
  45. package/src/config/conversion/lfm2/lfm2.5-1.2b-instruct-q4k-ehf16-af32.json +3 -3
  46. package/src/config/conversion/qwen3/qwen-3-5-0-8b-q4k-ehaf16.json +8 -4
  47. package/src/config/conversion/qwen3/qwen-3-5-2b-q4k-ehaf16.json +8 -4
  48. package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-eaf16.json +9 -5
  49. package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-ehaf16.json +8 -4
  50. package/src/config/goal-completion-matrix.json +552 -0
  51. package/src/config/kernel-path-loader.js +192 -38
  52. package/src/config/kernels/kernel-ref-digests.js +52 -18
  53. package/src/config/kernels/registry.json +3017 -744
  54. package/src/config/merge-contract-check.js +7 -3
  55. package/src/config/merge-helpers.d.ts +2 -0
  56. package/src/config/merge-helpers.js +1 -0
  57. package/src/config/merge.d.ts +1 -0
  58. package/src/config/merge.js +30 -0
  59. package/src/config/model-lanes/catalog-lane-resolver.d.ts +66 -0
  60. package/src/config/model-lanes/catalog-lane-resolver.js +170 -0
  61. package/src/config/param-validator.js +4 -0
  62. package/src/config/runtime/experiments/verify/lfm2-verify.json +3 -3
  63. package/src/config/runtime/profiles/diffusiongemma-26b-a4b-throughput.json +26 -0
  64. package/src/config/runtime/profiles/gemma3-270m-q4k-throughput-overlapped-probe.json +42 -0
  65. package/src/config/runtime/profiles/gemma3-270m-q4k-throughput-probe.json +39 -0
  66. package/src/config/runtime/profiles/gemma4-12b-rdrr-safe-b4.json +48 -0
  67. package/src/config/runtime/profiles/gemma4-12b-rdrr-safe-single.json +46 -0
  68. package/src/config/schema/buffer-pool.schema.js +3 -3
  69. package/src/config/schema/claim-evidence-contract.schema.json +93 -0
  70. package/src/config/schema/command-surface-contract.schema.json +81 -0
  71. package/src/config/schema/converter.schema.d.ts +11 -0
  72. package/src/config/schema/converter.schema.js +8 -0
  73. package/src/config/schema/doppler.schema.d.ts +1 -0
  74. package/src/config/schema/doppler.schema.js +38 -6
  75. package/src/config/schema/execution-v1.schema.d.ts +27 -0
  76. package/src/config/schema/execution-v1.schema.js +28 -19
  77. package/src/config/schema/goal-completion-matrix.schema.json +223 -0
  78. package/src/config/schema/inference.schema.d.ts +11 -0
  79. package/src/config/schema/manifest.schema.d.ts +107 -1
  80. package/src/config/schema/manifest.schema.js +10 -1
  81. package/src/config/schema/model-artifact-contract.schema.json +119 -0
  82. package/src/config/schema/moe.schema.d.ts +13 -0
  83. package/src/config/schema/moe.schema.js +1 -1
  84. package/src/config/schema/policy-schema-registry.json +26 -0
  85. package/src/config/schema/subsystem-support-contract.schema.json +48 -0
  86. package/src/config/schema/training-metrics.schema.d.ts +1 -1
  87. package/src/config/schema/training-metrics.schema.js +4 -3
  88. package/src/config/source-packages/litert/gemma-4-12b-it.json +390 -0
  89. package/src/config/source-packages/litert/gemma-4-e2b-it.json +7 -7
  90. package/src/config/source-packages/litert/gemma-4-e4b-it.json +370 -0
  91. package/src/config/source-packages/registry.json +25 -0
  92. package/src/config/supported-operations.json +5 -0
  93. package/src/config/transforms/capability-transform-resolver.d.ts +13 -2
  94. package/src/config/transforms/execution-graph-transforms.d.ts +42 -0
  95. package/src/config/transforms/execution-graph-transforms.js +227 -230
  96. package/src/converter/conversion-plan.d.ts +2 -0
  97. package/src/converter/conversion-plan.js +18 -4
  98. package/src/converter/core.d.ts +8 -0
  99. package/src/converter/core.js +574 -12
  100. package/src/converter/quantization-info.js +179 -2
  101. package/src/converter/quantizer.d.ts +5 -4
  102. package/src/converter/quantizer.js +16 -2
  103. package/src/experimental/adapters/lora-loader.js +109 -24
  104. package/src/experimental/diffusion/index.d.ts +24 -0
  105. package/src/experimental/diffusion/index.js +10 -0
  106. package/src/experimental/distribution/descriptor-transport.d.ts +117 -0
  107. package/src/experimental/distribution/descriptor-transport.js +342 -0
  108. package/src/experimental/training/README.md +11 -0
  109. package/src/experimental/training/autograd.js +33 -8
  110. package/src/experimental/training/datasets/gepa-frontier.d.ts +41 -0
  111. package/src/experimental/training/datasets/gepa-frontier.js +220 -0
  112. package/src/experimental/training/datasets/index.d.ts +13 -1
  113. package/src/experimental/training/datasets/index.js +8 -1
  114. package/src/experimental/training/datasets/teacher-traces.d.ts +51 -0
  115. package/src/experimental/training/datasets/teacher-traces.js +279 -0
  116. package/src/experimental/training/datasets/text-pairs.d.ts +45 -0
  117. package/src/experimental/training/datasets/text-pairs.js +82 -5
  118. package/src/experimental/training/distillation/student-fixture.d.ts +5 -0
  119. package/src/experimental/training/distillation/student-fixture.js +290 -32
  120. package/src/experimental/training/export.d.ts +9 -0
  121. package/src/experimental/training/export.js +113 -7
  122. package/src/experimental/training/index.d.ts +13 -1
  123. package/src/experimental/training/index.js +13 -1
  124. package/src/experimental/training/lora-pipeline.d.ts +143 -0
  125. package/src/experimental/training/lora-pipeline.js +1737 -49
  126. package/src/experimental/training/lora.js +23 -1
  127. package/src/experimental/training/loss.js +1 -1
  128. package/src/experimental/training/operator-agent-eval.d.ts +65 -0
  129. package/src/experimental/training/operator-agent-eval.js +326 -0
  130. package/src/experimental/training/operator-command.js +310 -51
  131. package/src/experimental/training/runner.d.ts +1 -0
  132. package/src/experimental/training/runner.js +4 -0
  133. package/src/experimental/training/workload-packs/distill-glm52-doppler-agent-harness-gemma270m-f16-sft-tiny.json +160 -0
  134. package/src/experimental/training/workload-packs/distill-glm52-doppler-code-agent-gemma270m-f16-sft-tiny.json +175 -0
  135. package/src/experimental/training/workload-packs/distill-glm52-doppler-code-agent-sft-tiny.json +175 -0
  136. package/src/experimental/training/workload-packs/distill-glm52-doppler-js-json-gemma270m-f16-sft-tiny.json +160 -0
  137. package/src/experimental/training/workload-packs/distill-glm52-doppler-review-gemma270m-f16-sft-tiny.json +162 -0
  138. package/src/experimental/training/workload-packs/distill-glm52-doppler-wgsl-gemma270m-f16-sft-tiny.json +160 -0
  139. package/src/experimental/training/workload-packs/lora-doppler-agent-harness-gemma270m-f16-tiny.json +137 -0
  140. package/src/experimental/training/workload-packs/lora-doppler-code-agent-gemma270m-f16-tiny.json +152 -0
  141. package/src/experimental/training/workload-packs/lora-doppler-code-agent-tiny.json +152 -0
  142. package/src/experimental/training/workload-packs/lora-doppler-js-json-gemma270m-f16-tiny.json +137 -0
  143. package/src/experimental/training/workload-packs/lora-doppler-review-gemma270m-f16-tiny.json +139 -0
  144. package/src/experimental/training/workload-packs/lora-doppler-wgsl-gemma270m-f16-tiny.json +137 -0
  145. package/src/experimental/training/workload-packs/registry.json +186 -6
  146. package/src/experimental/training/workloads.d.ts +37 -0
  147. package/src/experimental/training/workloads.js +110 -1
  148. package/src/formats/rdrr/functional-descriptor.d.ts +73 -0
  149. package/src/formats/rdrr/functional-descriptor.js +230 -0
  150. package/src/formats/rdrr/index.d.ts +1 -0
  151. package/src/formats/rdrr/index.js +1 -0
  152. package/src/formats/rdrr/parsing.js +22 -6
  153. package/src/formats/rdrr/source-transform-contract.js +10 -1
  154. package/src/formats/rdrr/storage-descriptor.js +9 -2
  155. package/src/formats/rdrr/types.d.ts +5 -3
  156. package/src/formats/rdrr/validation.js +15 -1
  157. package/src/gpu/command-recorder.d.ts +43 -3
  158. package/src/gpu/command-recorder.js +122 -25
  159. package/src/gpu/device.js +111 -11
  160. package/src/gpu/kernel-tuner/cache.js +1 -5
  161. package/src/gpu/kernels/attention.js +17 -9
  162. package/src/gpu/kernels/attention_decode_online_head256_f16kv.wgsl +236 -0
  163. package/src/gpu/kernels/attention_streaming_f16.wgsl +7 -5
  164. package/src/gpu/kernels/backward/cross_entropy_backward.js +72 -3
  165. package/src/gpu/kernels/backward/cross_entropy_backward.wgsl +2 -2
  166. package/src/gpu/kernels/cast.js +74 -67
  167. package/src/gpu/kernels/codegen/patches/attention_streaming_f16.from.attention_streaming.diff +46 -1
  168. package/src/gpu/kernels/codegen/patches/gelu_f16.from.gelu.diff +0 -20
  169. package/src/gpu/kernels/codegen/patches/residual_f16.from.residual.diff +15 -12
  170. package/src/gpu/kernels/codegen/patches/rmsnorm_f16.from.rmsnorm.diff +59 -44
  171. package/src/gpu/kernels/diffusion-gemma-sampling.d.ts +19 -0
  172. package/src/gpu/kernels/diffusion-gemma-sampling.js +104 -0
  173. package/src/gpu/kernels/diffusion_gemma_sampling.wgsl +130 -0
  174. package/src/gpu/kernels/dispatch.js +13 -3
  175. package/src/gpu/kernels/fused_ffn.js +3 -4
  176. package/src/gpu/kernels/fused_matmul_litert_int4_f16a.wgsl +78 -0
  177. package/src/gpu/kernels/fused_matmul_litert_int4_f16a_f32out.wgsl +78 -0
  178. package/src/gpu/kernels/fused_matmul_litert_int4_f32a_f32out.wgsl +76 -0
  179. package/src/gpu/kernels/fused_matmul_rmsnorm.js +101 -80
  180. package/src/gpu/kernels/fused_matmul_w4a16_batched_f16a.wgsl +116 -0
  181. package/src/gpu/kernels/fused_matmul_w4a16_batched_f16a_f32out.wgsl +116 -0
  182. package/src/gpu/kernels/fused_matmul_w4a16_f16a.wgsl +95 -0
  183. package/src/gpu/kernels/fused_matmul_w4a16_f16a_f32out.wgsl +95 -0
  184. package/src/gpu/kernels/gather.d.ts +84 -1
  185. package/src/gpu/kernels/gather.js +353 -5
  186. package/src/gpu/kernels/gather_litert_int4_vec4_f16_out.wgsl +75 -0
  187. package/src/gpu/kernels/gather_litert_int4_vec4_f32_out.wgsl +72 -0
  188. package/src/gpu/kernels/gather_split4_f16_vec4_f16_out.wgsl +103 -0
  189. package/src/gpu/kernels/gather_split8_f16_vec4_f16_out.wgsl +143 -0
  190. package/src/gpu/kernels/gather_split8_f16_vec4_f32_out.wgsl +143 -0
  191. package/src/gpu/kernels/gelu.wgsl +11 -7
  192. package/src/gpu/kernels/gelu_f16.wgsl +11 -7
  193. package/src/gpu/kernels/gemma4-route-expert.d.ts +35 -0
  194. package/src/gpu/kernels/gemma4-route-expert.js +126 -0
  195. package/src/gpu/kernels/gemma4_route_q4_matmul_f16a.wgsl +210 -0
  196. package/src/gpu/kernels/index.d.ts +71 -0
  197. package/src/gpu/kernels/index.js +52 -0
  198. package/src/gpu/kernels/kv-cache-write.d.ts +22 -0
  199. package/src/gpu/kernels/kv-cache-write.js +126 -0
  200. package/src/gpu/kernels/kv_cache_write_f32_to_f16.wgsl +32 -0
  201. package/src/gpu/kernels/lm-head-argmax.d.ts +19 -0
  202. package/src/gpu/kernels/lm-head-argmax.js +218 -0
  203. package/src/gpu/kernels/lm_head_argmax_f16.wgsl +189 -0
  204. package/src/gpu/kernels/matmul-dispatch.d.ts +11 -0
  205. package/src/gpu/kernels/matmul-dispatch.js +37 -1
  206. package/src/gpu/kernels/matmul-selection.d.ts +5 -5
  207. package/src/gpu/kernels/matmul-selection.js +104 -9
  208. package/src/gpu/kernels/matmul.d.ts +1 -1
  209. package/src/gpu/kernels/matmul.js +151 -43
  210. package/src/gpu/kernels/moe.d.ts +1 -0
  211. package/src/gpu/kernels/moe.js +46 -18
  212. package/src/gpu/kernels/pipeline-cache.d.ts +8 -0
  213. package/src/gpu/kernels/pipeline-cache.js +24 -0
  214. package/src/gpu/kernels/rep-penalty.d.ts +21 -0
  215. package/src/gpu/kernels/rep-penalty.js +17 -9
  216. package/src/gpu/kernels/residual.d.ts +1 -0
  217. package/src/gpu/kernels/residual.js +13 -1
  218. package/src/gpu/kernels/residual.wgsl +3 -3
  219. package/src/gpu/kernels/residual_f16.wgsl +5 -3
  220. package/src/gpu/kernels/residual_f16_vec4.wgsl +6 -6
  221. package/src/gpu/kernels/residual_vec4.wgsl +6 -6
  222. package/src/gpu/kernels/rmsnorm-pair.d.ts +39 -0
  223. package/src/gpu/kernels/rmsnorm-pair.js +152 -0
  224. package/src/gpu/kernels/rmsnorm-qk.d.ts +41 -0
  225. package/src/gpu/kernels/rmsnorm-qk.js +128 -0
  226. package/src/gpu/kernels/rmsnorm.d.ts +31 -0
  227. package/src/gpu/kernels/rmsnorm.js +32 -7
  228. package/src/gpu/kernels/rmsnorm.wgsl +10 -6
  229. package/src/gpu/kernels/rmsnorm_f16.wgsl +7 -3
  230. package/src/gpu/kernels/rmsnorm_pair.wgsl +224 -0
  231. package/src/gpu/kernels/rmsnorm_qk.wgsl +118 -0
  232. package/src/gpu/kernels/rmsnorm_qk_f16.wgsl +143 -0
  233. package/src/gpu/kernels/rope-qk.d.ts +41 -0
  234. package/src/gpu/kernels/rope-qk.js +143 -0
  235. package/src/gpu/kernels/rope_qk.wgsl +95 -0
  236. package/src/gpu/kernels/rope_qk_f16.wgsl +97 -0
  237. package/src/gpu/kernels/rule-registry.js +50 -10
  238. package/src/gpu/kernels/sample.js +212 -187
  239. package/src/gpu/kernels/scatter_add_dynamic_expert_scale.wgsl +47 -0
  240. package/src/gpu/kernels/scatter_add_dynamic_f16_expert_scale.wgsl +49 -0
  241. package/src/gpu/kernels/scatter_add_dynamic_f16_weights_expert_scale.wgsl +49 -0
  242. package/src/gpu/kernels/scatter_add_routes_f16_weights_expert_scale.wgsl +45 -0
  243. package/src/gpu/kernels/shader-cache.js +1 -21
  244. package/src/gpu/kernels/soft-embedding.d.ts +28 -0
  245. package/src/gpu/kernels/soft-embedding.js +447 -0
  246. package/src/gpu/kernels/soft_embedding_logits_f16.wgsl +107 -0
  247. package/src/gpu/kernels/soft_embedding_split_f16.wgsl +46 -0
  248. package/src/gpu/kernels/softmax.wgsl +5 -80
  249. package/src/gpu/kernels/split-qkv-rmsnorm-qk.d.ts +44 -0
  250. package/src/gpu/kernels/split-qkv-rmsnorm-qk.js +154 -0
  251. package/src/gpu/kernels/split-qkv-rmsnorm-rope-qk.d.ts +59 -0
  252. package/src/gpu/kernels/split-qkv-rmsnorm-rope-qk.js +245 -0
  253. package/src/gpu/kernels/split_qkv_rmsnorm_qk.wgsl +146 -0
  254. package/src/gpu/kernels/split_qkv_rmsnorm_rope_qk.wgsl +185 -0
  255. package/src/gpu/kernels/split_qkv_rmsnorm_rope_qk_f16kv_cache.wgsl +187 -0
  256. package/src/gpu/kernels/uniform-utils.js +2 -5
  257. package/src/gpu/kernels/utils.d.ts +16 -1
  258. package/src/gpu/kernels/utils.js +42 -13
  259. package/src/gpu/readback-utils.js +10 -2
  260. package/src/gpu/uniform-cache.d.ts +10 -1
  261. package/src/gpu/uniform-cache.js +25 -7
  262. package/src/gpu/weight-buffer.d.ts +76 -7
  263. package/src/gpu/weight-buffer.js +85 -1
  264. package/src/inference/browser-harness-diffusion-energy-suites.js +269 -1
  265. package/src/inference/browser-harness-suite-helpers.d.ts +9 -0
  266. package/src/inference/browser-harness-suite-helpers.js +438 -2
  267. package/src/inference/browser-harness-text-helpers.d.ts +77 -0
  268. package/src/inference/browser-harness-text-helpers.js +236 -11
  269. package/src/inference/browser-harness.d.ts +8 -0
  270. package/src/inference/browser-harness.js +292 -8
  271. package/src/inference/decode-ring.d.ts +1 -0
  272. package/src/inference/decode-ring.js +7 -5
  273. package/src/inference/kv-cache/base.d.ts +25 -0
  274. package/src/inference/kv-cache/base.js +109 -1
  275. package/src/inference/kv-cache/mixed-geometry.d.ts +10 -0
  276. package/src/inference/kv-cache/mixed-geometry.js +109 -2
  277. package/src/inference/kv-cache/sliding-window.d.ts +14 -0
  278. package/src/inference/kv-cache/sliding-window.js +88 -0
  279. package/src/inference/model-lanes/execution-lane-resolver.d.ts +61 -0
  280. package/src/inference/model-lanes/execution-lane-resolver.js +126 -0
  281. package/src/inference/moe-router.d.ts +14 -1
  282. package/src/inference/moe-router.js +24 -1
  283. package/src/inference/pipelines/audio/gemma4.js +52 -1
  284. package/src/inference/pipelines/diffusion-gemma/config.d.ts +27 -0
  285. package/src/inference/pipelines/diffusion-gemma/config.js +149 -0
  286. package/src/inference/pipelines/diffusion-gemma/index.d.ts +31 -0
  287. package/src/inference/pipelines/diffusion-gemma/index.js +12 -0
  288. package/src/inference/pipelines/diffusion-gemma/pipeline.d.ts +69 -0
  289. package/src/inference/pipelines/diffusion-gemma/pipeline.js +380 -0
  290. package/src/inference/pipelines/diffusion-gemma/sampling.d.ts +128 -0
  291. package/src/inference/pipelines/diffusion-gemma/sampling.js +395 -0
  292. package/src/inference/pipelines/text/attention/dispatch-params.d.ts +28 -4
  293. package/src/inference/pipelines/text/attention/dispatch-params.js +176 -2
  294. package/src/inference/pipelines/text/attention/projections.d.ts +71 -2
  295. package/src/inference/pipelines/text/attention/projections.js +329 -49
  296. package/src/inference/pipelines/text/attention/quant-options.d.ts +38 -0
  297. package/src/inference/pipelines/text/attention/record.js +446 -73
  298. package/src/inference/pipelines/text/attention/run.js +251 -49
  299. package/src/inference/pipelines/text/attention/types.d.ts +5 -1
  300. package/src/inference/pipelines/text/config.d.ts +22 -1
  301. package/src/inference/pipelines/text/config.js +199 -22
  302. package/src/inference/pipelines/text/dtype-contract.d.ts +1 -1
  303. package/src/inference/pipelines/text/dtype-contract.js +4 -1
  304. package/src/inference/pipelines/text/embed.d.ts +4 -3
  305. package/src/inference/pipelines/text/embed.js +16 -6
  306. package/src/inference/pipelines/text/execution-plan.d.ts +8 -0
  307. package/src/inference/pipelines/text/execution-plan.js +6 -0
  308. package/src/inference/pipelines/text/execution-runtime-builders.d.ts +4 -0
  309. package/src/inference/pipelines/text/execution-runtime-builders.js +44 -28
  310. package/src/inference/pipelines/text/execution-v1.d.ts +2 -0
  311. package/src/inference/pipelines/text/execution-v1.js +373 -139
  312. package/src/inference/pipelines/text/ffn/dense.js +44 -13
  313. package/src/inference/pipelines/text/ffn/moe.d.ts +4 -1
  314. package/src/inference/pipelines/text/ffn/moe.js +9 -2
  315. package/src/inference/pipelines/text/ffn/sandwich.d.ts +3 -1
  316. package/src/inference/pipelines/text/ffn/sandwich.js +237 -15
  317. package/src/inference/pipelines/text/ffn/standard.d.ts +3 -1
  318. package/src/inference/pipelines/text/ffn/standard.js +44 -4
  319. package/src/inference/pipelines/text/generator-decode-policy.d.ts +81 -0
  320. package/src/inference/pipelines/text/generator-decode-policy.js +2 -12
  321. package/src/inference/pipelines/text/generator-helpers.js +5 -2
  322. package/src/inference/pipelines/text/generator-prefill-helpers.d.ts +84 -0
  323. package/src/inference/pipelines/text/generator-prefill-helpers.js +67 -1
  324. package/src/inference/pipelines/text/generator-runtime.d.ts +1 -0
  325. package/src/inference/pipelines/text/generator-steps.d.ts +38 -3
  326. package/src/inference/pipelines/text/generator-steps.js +268 -77
  327. package/src/inference/pipelines/text/generator.d.ts +39 -0
  328. package/src/inference/pipelines/text/generator.js +1105 -18
  329. package/src/inference/pipelines/text/init.d.ts +8 -1
  330. package/src/inference/pipelines/text/init.js +30 -1
  331. package/src/inference/pipelines/text/layer-plan-gpu.js +62 -14
  332. package/src/inference/pipelines/text/layer.js +272 -25
  333. package/src/inference/pipelines/text/linear-attention.js +6 -4
  334. package/src/inference/pipelines/text/logits/gpu.d.ts +49 -2
  335. package/src/inference/pipelines/text/logits/gpu.js +417 -66
  336. package/src/inference/pipelines/text/logits/index.d.ts +2 -2
  337. package/src/inference/pipelines/text/logits/index.js +96 -42
  338. package/src/inference/pipelines/text/logits/types.d.ts +2 -2
  339. package/src/inference/pipelines/text/lora-apply.js +17 -0
  340. package/src/inference/pipelines/text/moe-gpu.d.ts +13 -0
  341. package/src/inference/pipelines/text/moe-gpu.js +445 -16
  342. package/src/inference/pipelines/text/moe-impl.d.ts +13 -3
  343. package/src/inference/pipelines/text/moe-shape-validator.js +13 -1
  344. package/src/inference/pipelines/text/ops.d.ts +29 -2
  345. package/src/inference/pipelines/text/ops.js +19 -0
  346. package/src/inference/pipelines/text/per-layer-inputs.js +5 -0
  347. package/src/inference/pipelines/text/probes.js +4 -5
  348. package/src/inference/pipelines/text/sampling-config.d.ts +4 -10
  349. package/src/inference/pipelines/text/sampling-config.js +103 -64
  350. package/src/inference/pipelines/text/sampling.js +13 -9
  351. package/src/inference/pipelines/text/state.js +17 -0
  352. package/src/inference/pipelines/text/tsir-fixture-writer.d.ts +57 -0
  353. package/src/inference/pipelines/text/tsir-fixture-writer.js +8 -9
  354. package/src/inference/pipelines/text/types.d.ts +84 -3
  355. package/src/inference/pipelines/text/weights.d.ts +4 -2
  356. package/src/inference/pipelines/text/weights.js +4 -5
  357. package/src/inference/pipelines/text.d.ts +24 -0
  358. package/src/inference/pipelines/text.js +162 -38
  359. package/src/inference/pipelines/vision/gemma4.js +20 -9
  360. package/src/inference/tokenizer.d.ts +30 -0
  361. package/src/inference/tokenizer.js +340 -164
  362. package/src/inference/tokenizers/bundled.js +71 -49
  363. package/src/loader/doppler-loader.d.ts +11 -0
  364. package/src/loader/doppler-loader.js +290 -15
  365. package/src/loader/embedding-limit-preflight.d.ts +31 -0
  366. package/src/loader/embedding-limit-preflight.js +216 -0
  367. package/src/loader/embedding-loader.d.ts +8 -2
  368. package/src/loader/embedding-loader.js +263 -8
  369. package/src/loader/experts/expert-loader.d.ts +4 -1
  370. package/src/loader/experts/expert-loader.js +174 -5
  371. package/src/loader/final-weights-loader.d.ts +5 -0
  372. package/src/loader/final-weights-loader.js +167 -4
  373. package/src/loader/layer-loader.d.ts +2 -0
  374. package/src/loader/layer-loader.js +74 -6
  375. package/src/loader/loader-types.d.ts +59 -2
  376. package/src/loader/manifest-config.d.ts +8 -0
  377. package/src/loader/shard-resolver.js +37 -4
  378. package/src/loader/tensors/functional-descriptor-loader.d.ts +9 -0
  379. package/src/loader/tensors/functional-descriptor-loader.js +377 -0
  380. package/src/loader/tensors/tensor-loader.d.ts +36 -1
  381. package/src/loader/tensors/tensor-loader.js +437 -21
  382. package/src/loader/tensors/tensor-reader.d.ts +4 -1
  383. package/src/loader/tensors/tensor-reader.js +11 -4
  384. package/src/loader/weights.d.ts +3 -1
  385. package/src/memory/buffer-pool.js +12 -9
  386. package/src/models/diffusiongemma.d.ts +7 -0
  387. package/src/models/diffusiongemma.js +25 -0
  388. package/src/models/gemma4.js +24 -0
  389. package/src/models/qwen3.js +2 -2
  390. package/src/rules/execution-rules-contract-check.js +93 -10
  391. package/src/rules/inference/capability-transforms.rules.json +26 -2
  392. package/src/rules/inference/config.rules.json +1 -0
  393. package/src/rules/inference/dtype.rules.json +7 -0
  394. package/src/rules/inference/execution.rules.json +108 -17
  395. package/src/rules/inference/ffn.rules.json +18 -0
  396. package/src/rules/kernels/gather.rules.json +81 -0
  397. package/src/rules/kernels/matmul.rules.json +19 -0
  398. package/src/rules/kernels/moe.rules.json +3 -0
  399. package/src/rules/kernels/rep-penalty.rules.json +6 -0
  400. package/src/rules/kernels/rmsnorm-qk.rules.json +6 -0
  401. package/src/rules/kernels/rmsnorm.rules.json +42 -7
  402. package/src/rules/kernels/rope-qk.rules.json +6 -0
  403. package/src/rules/loader/tensor-loader.rules.json +8 -0
  404. package/src/rules/rule-registry.js +6 -0
  405. package/src/storage/artifact-storage-context.d.ts +1 -0
  406. package/src/storage/artifact-storage-context.js +10 -0
  407. package/src/tooling/browser-command-runner.d.ts +0 -3
  408. package/src/tooling/browser-command-runner.js +2 -14
  409. package/src/tooling/command-api-family-normalizers.js +4 -2
  410. package/src/tooling/command-api-helpers.js +2 -1
  411. package/src/tooling/command-api.d.ts +2 -0
  412. package/src/tooling/command-api.js +1 -0
  413. package/src/tooling/command-runner-shared.d.ts +4 -0
  414. package/src/tooling/command-runner-shared.js +2 -1
  415. package/src/tooling/diagnose-runner.js +2 -0
  416. package/src/tooling/litert-package-runtime.js +553 -16
  417. package/src/tooling/node-browser-command-runner.js +10 -4
  418. package/src/tooling/node-command-runner.js +7 -2
  419. package/src/tooling/node-convert-worker-pool.js +3 -0
  420. package/src/tooling/node-convert-worker.js +1 -0
  421. package/src/tooling/node-converter.js +12 -0
  422. package/src/tooling/node-webgpu.d.ts +1 -0
  423. package/src/tooling/node-webgpu.js +23 -11
  424. package/src/tooling/program-bundle.d.ts +52 -1
  425. package/src/tooling/rdrr-integrity-refresh.d.ts +18 -0
  426. package/src/tooling/source-artifact-adapter.js +7 -2
  427. package/src/tooling/source-package-profiles.d.ts +18 -1
  428. package/src/tooling/source-package-profiles.js +109 -1
  429. package/src/tooling/source-runtime-bundle.d.ts +1 -0
  430. package/src/tooling/source-runtime-bundle.js +27 -0
  431. package/src/tooling/source-runtime-converter-config.js +12 -3
  432. package/src/tooling-experimental-exports.shared.d.ts +10 -0
  433. package/src/tooling-experimental-exports.shared.js +10 -0
@@ -12,14 +12,33 @@
12
12
  // Canonical base rule registry: src/rules/rule-registry.js
13
13
  // ============================================================================
14
14
  import { getRuleSet as getBaseRuleSet, selectRuleValue as selectBaseRuleValue } from '../../rules/rule-registry.js';
15
- import { getKernelCapabilities, getPlatformConfig } from '../device.js';
15
+ import { getDeviceEpoch, getKernelCapabilities, getPlatformConfig } from '../device.js';
16
+
17
+ let cachedKernelRuleContextBase = null;
18
+ let cachedKernelRuleContextEpoch = null;
16
19
 
17
20
  export function getRuleSet(group, name) {
18
21
  return getBaseRuleSet('kernels', group, name);
19
22
  }
20
23
 
21
- function enrichKernelRuleContext(context) {
22
- const next = (context && typeof context === 'object') ? { ...context } : {};
24
+ function getKernelRuleContextBase() {
25
+ let epoch = null;
26
+ try {
27
+ epoch = getDeviceEpoch();
28
+ } catch {
29
+ epoch = null;
30
+ }
31
+
32
+ if (
33
+ epoch !== null
34
+ && cachedKernelRuleContextEpoch === epoch
35
+ && cachedKernelRuleContextBase
36
+ ) {
37
+ return cachedKernelRuleContextBase;
38
+ }
39
+
40
+ const base = {};
41
+ let cacheable = false;
23
42
 
24
43
  let capabilities = null;
25
44
  try {
@@ -28,9 +47,10 @@ function enrichKernelRuleContext(context) {
28
47
  capabilities = null;
29
48
  }
30
49
  if (capabilities) {
31
- if (next.hasSubgroups === undefined) next.hasSubgroups = capabilities.hasSubgroups;
32
- if (next.hasSubgroupsF16 === undefined) next.hasSubgroupsF16 = capabilities.hasSubgroupsF16;
33
- if (next.hasF16 === undefined) next.hasF16 = capabilities.hasF16;
50
+ base.hasSubgroups = capabilities.hasSubgroups;
51
+ base.hasSubgroupsF16 = capabilities.hasSubgroupsF16;
52
+ base.hasF16 = capabilities.hasF16;
53
+ cacheable = true;
34
54
  }
35
55
 
36
56
  let platform = null;
@@ -40,12 +60,32 @@ function enrichKernelRuleContext(context) {
40
60
  platform = null;
41
61
  }
42
62
  if (platform) {
43
- if (next.platformId === undefined) next.platformId = platform.id;
44
- if (next.platformVendor === undefined) next.platformVendor = platform.detection?.vendor ?? null;
45
- if (next.platformArchitecture === undefined) next.platformArchitecture = platform.detection?.architecture ?? null;
46
- if (next.platformDevice === undefined) next.platformDevice = platform.detection?.device ?? null;
63
+ base.platformId = platform.id;
64
+ base.platformVendor = platform.detection?.vendor ?? null;
65
+ base.platformArchitecture = platform.detection?.architecture ?? null;
66
+ base.platformDevice = platform.detection?.device ?? null;
47
67
  }
48
68
 
69
+ if (epoch !== null && cacheable) {
70
+ cachedKernelRuleContextEpoch = epoch;
71
+ cachedKernelRuleContextBase = Object.freeze(base);
72
+ return cachedKernelRuleContextBase;
73
+ }
74
+
75
+ return base;
76
+ }
77
+
78
+ function enrichKernelRuleContext(context) {
79
+ const base = getKernelRuleContextBase();
80
+ const next = { ...base };
81
+ if (!context || typeof context !== 'object') {
82
+ return next;
83
+ }
84
+ for (const [key, value] of Object.entries(context)) {
85
+ if (value !== undefined || !(key in base)) {
86
+ next[key] = value;
87
+ }
88
+ }
49
89
  return next;
50
90
  }
51
91
 
@@ -7,6 +7,8 @@ import { createPipeline, createUniformBufferWithView, getOrCreateBindGroupLayout
7
7
  import { allowReadback } from '../perf-guards.js';
8
8
  import { selectRuleValue as selectKernelRuleValue } from './rule-registry.js';
9
9
  import { selectRuleValue as selectSharedRuleValue } from '../../rules/rule-registry.js';
10
+ import { getKernelThresholds } from '../../config/schema/index.js';
11
+ import { recordDispatch } from './dispatch.js';
10
12
 
11
13
 
12
14
  function getSampleBindGroupLayout(device) {
@@ -93,8 +95,10 @@ function assertSampleOptions(options, recordMode = false) {
93
95
 
94
96
  async function resolveArgmaxPipelines(device, vocabSize, variants) {
95
97
  const argmaxPipeline = await createSamplePipeline(device, variants.argmax);
96
- const numWorkgroups = Math.min(WORKGROUP_SIZES.DEFAULT, Math.ceil(vocabSize / WORKGROUP_SIZES.DEFAULT));
97
- const useSinglePassArgmax = numWorkgroups === 1;
98
+ const reduceWorkgroups = Math.min(WORKGROUP_SIZES.DEFAULT, Math.ceil(vocabSize / WORKGROUP_SIZES.DEFAULT));
99
+ const { argmaxReduceVocabThreshold } = getKernelThresholds().sample;
100
+ const useSinglePassArgmax = vocabSize <= argmaxReduceVocabThreshold;
101
+ const numWorkgroups = useSinglePassArgmax ? 1 : reduceWorkgroups;
98
102
  const reducePipeline = useSinglePassArgmax
99
103
  ? null
100
104
  : await createSamplePipeline(device, variants.argmaxReduce);
@@ -189,55 +193,59 @@ async function executeArgmaxRun(logits, vocabSize, options) {
189
193
  useSinglePassArgmax,
190
194
  } = await resolveArgmaxPipelines(device, vocabSize, variants);
191
195
 
192
- const tempLogits = acquireBuffer(WORKGROUP_SIZES.DEFAULT * 4, undefined, 'argmax_temp_logits');
193
- const tempIndices = acquireBuffer(WORKGROUP_SIZES.DEFAULT * 4, undefined, 'argmax_temp_indices');
194
- const outputIndex = options.outputIndex;
195
- const minOutputBytes = Math.max(4, (outputIndex + 1) * 4);
196
- const outputBuffer = options.outputBuffer ?? acquireBuffer(minOutputBytes, undefined, 'argmax_output');
197
- const ownsOutputBuffer = !options.outputBuffer;
198
- ensureOutputBufferSize(outputBuffer, minOutputBytes, 'argmax outputIndex');
199
-
200
- const uniformBuffer = createArgmaxUniformBuffer(device, null, vocabSize, options);
201
-
202
- const bindGroupLayout = getSampleBindGroupLayout(device);
203
- const entries = [
204
- { binding: 0, resource: { buffer: uniformBuffer } },
205
- { binding: 1, resource: { buffer: logits } },
206
- { binding: 2, resource: { buffer: outputBuffer } },
207
- { binding: 3, resource: { buffer: tempIndices } },
208
- { binding: 4, resource: { buffer: tempLogits } },
209
- ];
210
- const argmaxBindGroup = device.createBindGroup({
211
- label: 'argmax_bind_group',
212
- layout: bindGroupLayout,
213
- entries,
214
- });
215
-
216
- const encoder = device.createCommandEncoder({ label: 'argmax_encoder' });
217
-
218
- const pass1 = encoder.beginComputePass({ label: 'argmax_pass1' });
219
- pass1.setPipeline(singlePassPipeline ?? argmaxPipeline);
220
- pass1.setBindGroup(0, argmaxBindGroup);
221
- pass1.dispatchWorkgroups(useSinglePassArgmax ? 1 : numWorkgroups);
222
- pass1.end();
223
-
224
- if (reducePipeline) {
225
- const reduceBindGroup = device.createBindGroup({
226
- label: 'argmax_reduce_bind_group',
196
+ let tempLogits = null;
197
+ let tempIndices = null;
198
+ let outputBuffer = null;
199
+ let ownsOutputBuffer = false;
200
+ let uniformBuffer = null;
201
+ try {
202
+ tempLogits = acquireBuffer(WORKGROUP_SIZES.DEFAULT * 4, undefined, 'argmax_temp_logits');
203
+ tempIndices = acquireBuffer(WORKGROUP_SIZES.DEFAULT * 4, undefined, 'argmax_temp_indices');
204
+ const outputIndex = options.outputIndex;
205
+ const minOutputBytes = Math.max(4, (outputIndex + 1) * 4);
206
+ outputBuffer = options.outputBuffer ?? acquireBuffer(minOutputBytes, undefined, 'argmax_output');
207
+ ownsOutputBuffer = !options.outputBuffer;
208
+ ensureOutputBufferSize(outputBuffer, minOutputBytes, 'argmax outputIndex');
209
+
210
+ uniformBuffer = createArgmaxUniformBuffer(device, null, vocabSize, options);
211
+
212
+ const bindGroupLayout = getSampleBindGroupLayout(device);
213
+ const entries = [
214
+ { binding: 0, resource: { buffer: uniformBuffer } },
215
+ { binding: 1, resource: { buffer: logits } },
216
+ { binding: 2, resource: { buffer: outputBuffer } },
217
+ { binding: 3, resource: { buffer: tempIndices } },
218
+ { binding: 4, resource: { buffer: tempLogits } },
219
+ ];
220
+ const argmaxBindGroup = device.createBindGroup({
221
+ label: 'argmax_bind_group',
227
222
  layout: bindGroupLayout,
228
223
  entries,
229
224
  });
230
225
 
231
- const pass2 = encoder.beginComputePass({ label: 'argmax_pass2' });
232
- pass2.setPipeline(reducePipeline);
233
- pass2.setBindGroup(0, reduceBindGroup);
234
- pass2.dispatchWorkgroups(1);
235
- pass2.end();
236
- }
237
-
238
- device.queue.submit([encoder.finish()]);
226
+ const encoder = device.createCommandEncoder({ label: 'argmax_encoder' });
227
+
228
+ const pass1 = encoder.beginComputePass({ label: 'argmax_pass1' });
229
+ pass1.setPipeline(singlePassPipeline ?? argmaxPipeline);
230
+ pass1.setBindGroup(0, argmaxBindGroup);
231
+ pass1.dispatchWorkgroups(useSinglePassArgmax ? 1 : numWorkgroups);
232
+ pass1.end();
233
+
234
+ if (reducePipeline) {
235
+ const reduceBindGroup = device.createBindGroup({
236
+ label: 'argmax_reduce_bind_group',
237
+ layout: bindGroupLayout,
238
+ entries,
239
+ });
240
+
241
+ const pass2 = encoder.beginComputePass({ label: 'argmax_pass2' });
242
+ pass2.setPipeline(reducePipeline);
243
+ pass2.setBindGroup(0, reduceBindGroup);
244
+ pass2.dispatchWorkgroups(1);
245
+ pass2.end();
246
+ }
239
247
 
240
- try {
248
+ device.queue.submit([encoder.finish()]);
241
249
  return await readTokenFromOutput(outputBuffer, outputIndex);
242
250
  } finally {
243
251
  cleanupRunResources(
@@ -262,53 +270,65 @@ async function executeArgmaxRecord(recorder, logits, vocabSize, options) {
262
270
  useSinglePassArgmax,
263
271
  } = await resolveArgmaxPipelines(device, vocabSize, variants);
264
272
 
265
- const tempLogits = acquireBuffer(WORKGROUP_SIZES.DEFAULT * 4, undefined, 'argmax_temp_logits');
266
- const tempIndices = acquireBuffer(WORKGROUP_SIZES.DEFAULT * 4, undefined, 'argmax_temp_indices');
267
- const outputIndex = options.outputIndex;
268
- const minOutputBytes = Math.max(4, (outputIndex + 1) * 4);
269
- const outputBuffer = options.outputBuffer ?? acquireBuffer(minOutputBytes, undefined, 'argmax_output');
270
- ensureOutputBufferSize(outputBuffer, minOutputBytes, 'argmax outputIndex');
271
-
272
- const uniformBuffer = createArgmaxUniformBuffer(device, recorder, vocabSize, options);
273
-
274
- const bindGroupLayout = getSampleBindGroupLayout(device);
275
- const entries = [
276
- { binding: 0, resource: { buffer: uniformBuffer } },
277
- { binding: 1, resource: { buffer: logits } },
278
- { binding: 2, resource: { buffer: outputBuffer } },
279
- { binding: 3, resource: { buffer: tempIndices } },
280
- { binding: 4, resource: { buffer: tempLogits } },
281
- ];
282
- const bindGroup = device.createBindGroup({
283
- label: 'argmax_bind_group',
284
- layout: bindGroupLayout,
285
- entries,
286
- });
287
-
288
- const pass1 = recorder.beginComputePass('argmax_phase1');
289
- pass1.setPipeline(singlePassPipeline ?? argmaxPipeline);
290
- pass1.setBindGroup(0, bindGroup);
291
- pass1.dispatchWorkgroups(useSinglePassArgmax ? 1 : numWorkgroups);
292
- pass1.end();
293
-
294
- if (reducePipeline) {
295
- const reduceBindGroup = device.createBindGroup({
296
- label: 'argmax_reduce_bind_group',
273
+ let tempLogits = null;
274
+ let tempIndices = null;
275
+ let outputBuffer = null;
276
+ let ownsOutputBuffer = false;
277
+ let completed = false;
278
+ try {
279
+ tempLogits = acquireBuffer(WORKGROUP_SIZES.DEFAULT * 4, undefined, 'argmax_temp_logits');
280
+ tempIndices = acquireBuffer(WORKGROUP_SIZES.DEFAULT * 4, undefined, 'argmax_temp_indices');
281
+ const outputIndex = options.outputIndex;
282
+ const minOutputBytes = Math.max(4, (outputIndex + 1) * 4);
283
+ outputBuffer = options.outputBuffer ?? acquireBuffer(minOutputBytes, undefined, 'argmax_output');
284
+ ownsOutputBuffer = !options.outputBuffer;
285
+ ensureOutputBufferSize(outputBuffer, minOutputBytes, 'argmax outputIndex');
286
+
287
+ const uniformBuffer = createArgmaxUniformBuffer(device, recorder, vocabSize, options);
288
+
289
+ const bindGroupLayout = getSampleBindGroupLayout(device);
290
+ const entries = [
291
+ { binding: 0, resource: { buffer: uniformBuffer } },
292
+ { binding: 1, resource: { buffer: logits } },
293
+ { binding: 2, resource: { buffer: outputBuffer } },
294
+ { binding: 3, resource: { buffer: tempIndices } },
295
+ { binding: 4, resource: { buffer: tempLogits } },
296
+ ];
297
+ const bindGroup = device.createBindGroup({
298
+ label: 'argmax_bind_group',
297
299
  layout: bindGroupLayout,
298
300
  entries,
299
301
  });
300
302
 
301
- const pass2 = recorder.beginComputePass('argmax_phase2');
302
- pass2.setPipeline(reducePipeline);
303
- pass2.setBindGroup(0, reduceBindGroup);
304
- pass2.dispatchWorkgroups(1);
305
- pass2.end();
306
- }
303
+ recordDispatch(
304
+ recorder,
305
+ singlePassPipeline ?? argmaxPipeline,
306
+ bindGroup,
307
+ useSinglePassArgmax ? 1 : numWorkgroups,
308
+ 'argmax_phase1'
309
+ );
307
310
 
308
- recorder.trackTemporaryBuffer(tempLogits);
309
- recorder.trackTemporaryBuffer(tempIndices);
311
+ if (reducePipeline) {
312
+ const reduceBindGroup = device.createBindGroup({
313
+ label: 'argmax_reduce_bind_group',
314
+ layout: bindGroupLayout,
315
+ entries,
316
+ });
317
+ recordDispatch(recorder, reducePipeline, reduceBindGroup, 1, 'argmax_phase2');
318
+ }
310
319
 
311
- return outputBuffer;
320
+ recorder.trackTemporaryBuffer(tempLogits);
321
+ recorder.trackTemporaryBuffer(tempIndices);
322
+ completed = true;
323
+ return outputBuffer;
324
+ } finally {
325
+ if (!completed) {
326
+ cleanupRunResources(
327
+ null,
328
+ [tempLogits, tempIndices, ownsOutputBuffer ? outputBuffer : null]
329
+ );
330
+ }
331
+ }
312
332
  }
313
333
 
314
334
  export async function runArgmax(
@@ -367,61 +387,65 @@ export async function runGPUSample(
367
387
 
368
388
  const numWorkgroups = Math.min(WORKGROUP_SIZES.DEFAULT, Math.ceil(vocabSize / WORKGROUP_SIZES.DEFAULT));
369
389
 
370
- const topkLogits = acquireBuffer(WORKGROUP_SIZES.DEFAULT * 4, undefined, 'topk_logits');
371
- const topkIndices = acquireBuffer(WORKGROUP_SIZES.DEFAULT * 4, undefined, 'topk_indices');
372
- const minOutputBytes = Math.max(4, (outputIndex + 1) * 4);
373
- const outputBuffer = outputBufferOverride ?? acquireBuffer(minOutputBytes, undefined, 'sample_output');
374
- const ownsOutputBuffer = !outputBufferOverride;
375
- ensureOutputBufferSize(outputBuffer, minOutputBytes, 'sample outputIndex');
376
-
377
- const uniformBuffer = createSampleUniformBuffer(
378
- device,
379
- null,
380
- vocabSize,
381
- topK,
382
- temperature,
383
- randomValue,
384
- padTokenId,
385
- logitSoftcap,
386
- outputIndex
387
- );
388
-
389
- const bindGroupLayout = getSampleBindGroupLayout(device);
390
- const bindGroup = device.createBindGroup({
391
- label: 'sample_bind_group',
392
- layout: bindGroupLayout,
393
- entries: [
394
- { binding: 0, resource: { buffer: uniformBuffer } },
395
- { binding: 1, resource: { buffer: logits } },
396
- { binding: 2, resource: { buffer: outputBuffer } },
397
- { binding: 3, resource: { buffer: topkIndices } },
398
- { binding: 4, resource: { buffer: topkLogits } },
399
- ],
400
- });
390
+ let topkLogits = null;
391
+ let topkIndices = null;
392
+ let outputBuffer = null;
393
+ let ownsOutputBuffer = false;
394
+ let uniformBuffer = null;
395
+ try {
396
+ topkLogits = acquireBuffer(WORKGROUP_SIZES.DEFAULT * 4, undefined, 'topk_logits');
397
+ topkIndices = acquireBuffer(WORKGROUP_SIZES.DEFAULT * 4, undefined, 'topk_indices');
398
+ const minOutputBytes = Math.max(4, (outputIndex + 1) * 4);
399
+ outputBuffer = outputBufferOverride ?? acquireBuffer(minOutputBytes, undefined, 'sample_output');
400
+ ownsOutputBuffer = !outputBufferOverride;
401
+ ensureOutputBufferSize(outputBuffer, minOutputBytes, 'sample outputIndex');
402
+
403
+ uniformBuffer = createSampleUniformBuffer(
404
+ device,
405
+ null,
406
+ vocabSize,
407
+ topK,
408
+ temperature,
409
+ randomValue,
410
+ padTokenId,
411
+ logitSoftcap,
412
+ outputIndex
413
+ );
401
414
 
402
- const encoder = device.createCommandEncoder({ label: 'sample_encoder' });
415
+ const bindGroupLayout = getSampleBindGroupLayout(device);
416
+ const bindGroup = device.createBindGroup({
417
+ label: 'sample_bind_group',
418
+ layout: bindGroupLayout,
419
+ entries: [
420
+ { binding: 0, resource: { buffer: uniformBuffer } },
421
+ { binding: 1, resource: { buffer: logits } },
422
+ { binding: 2, resource: { buffer: outputBuffer } },
423
+ { binding: 3, resource: { buffer: topkIndices } },
424
+ { binding: 4, resource: { buffer: topkLogits } },
425
+ ],
426
+ });
403
427
 
404
- const pass1 = encoder.beginComputePass({ label: 'sample_phase1' });
405
- pass1.setPipeline(phase1Pipeline);
406
- pass1.setBindGroup(0, bindGroup);
407
- pass1.dispatchWorkgroups(numWorkgroups);
408
- pass1.end();
428
+ const encoder = device.createCommandEncoder({ label: 'sample_encoder' });
409
429
 
410
- const pass2 = encoder.beginComputePass({ label: 'sample_phase2' });
411
- pass2.setPipeline(phase2Pipeline);
412
- pass2.setBindGroup(0, bindGroup);
413
- pass2.dispatchWorkgroups(1);
414
- pass2.end();
430
+ const pass1 = encoder.beginComputePass({ label: 'sample_phase1' });
431
+ pass1.setPipeline(phase1Pipeline);
432
+ pass1.setBindGroup(0, bindGroup);
433
+ pass1.dispatchWorkgroups(numWorkgroups);
434
+ pass1.end();
415
435
 
416
- const pass3 = encoder.beginComputePass({ label: 'sample_phase3' });
417
- pass3.setPipeline(phase3Pipeline);
418
- pass3.setBindGroup(0, bindGroup);
419
- pass3.dispatchWorkgroups(1);
420
- pass3.end();
436
+ const pass2 = encoder.beginComputePass({ label: 'sample_phase2' });
437
+ pass2.setPipeline(phase2Pipeline);
438
+ pass2.setBindGroup(0, bindGroup);
439
+ pass2.dispatchWorkgroups(1);
440
+ pass2.end();
421
441
 
422
- device.queue.submit([encoder.finish()]);
442
+ const pass3 = encoder.beginComputePass({ label: 'sample_phase3' });
443
+ pass3.setPipeline(phase3Pipeline);
444
+ pass3.setBindGroup(0, bindGroup);
445
+ pass3.dispatchWorkgroups(1);
446
+ pass3.end();
423
447
 
424
- try {
448
+ device.queue.submit([encoder.finish()]);
425
449
  return await readTokenFromOutput(outputBuffer, outputIndex);
426
450
  } finally {
427
451
  cleanupRunResources(
@@ -485,59 +509,60 @@ export async function recordGPUSample(
485
509
 
486
510
  const numWorkgroups = Math.min(WORKGROUP_SIZES.DEFAULT, Math.ceil(vocabSize / WORKGROUP_SIZES.DEFAULT));
487
511
 
488
- const topkLogits = acquireBuffer(WORKGROUP_SIZES.DEFAULT * 4, undefined, 'topk_logits');
489
- const topkIndices = acquireBuffer(WORKGROUP_SIZES.DEFAULT * 4, undefined, 'topk_indices');
490
- const minOutputBytes = Math.max(4, (outputIndex + 1) * 4);
491
- const outputBuffer = outputBufferOverride ?? acquireBuffer(minOutputBytes, undefined, 'sample_output');
492
- ensureOutputBufferSize(outputBuffer, minOutputBytes, 'sample outputIndex');
512
+ let topkLogits = null;
513
+ let topkIndices = null;
514
+ let outputBuffer = null;
515
+ let ownsOutputBuffer = false;
516
+ let completed = false;
517
+ try {
518
+ topkLogits = acquireBuffer(WORKGROUP_SIZES.DEFAULT * 4, undefined, 'topk_logits');
519
+ topkIndices = acquireBuffer(WORKGROUP_SIZES.DEFAULT * 4, undefined, 'topk_indices');
520
+ const minOutputBytes = Math.max(4, (outputIndex + 1) * 4);
521
+ outputBuffer = outputBufferOverride ?? acquireBuffer(minOutputBytes, undefined, 'sample_output');
522
+ ownsOutputBuffer = !outputBufferOverride;
523
+ ensureOutputBufferSize(outputBuffer, minOutputBytes, 'sample outputIndex');
524
+
525
+ const uniformBuffer = createSampleUniformBuffer(
526
+ device,
527
+ recorder,
528
+ vocabSize,
529
+ topK,
530
+ temperature,
531
+ randomValue,
532
+ padTokenId,
533
+ logitSoftcap,
534
+ outputIndex
535
+ );
493
536
 
494
- const uniformBuffer = createSampleUniformBuffer(
495
- device,
496
- recorder,
497
- vocabSize,
498
- topK,
499
- temperature,
500
- randomValue,
501
- padTokenId,
502
- logitSoftcap,
503
- outputIndex
504
- );
537
+ const bindGroupLayout = getSampleBindGroupLayout(device);
538
+ const bindGroup = device.createBindGroup({
539
+ label: 'sample_bind_group',
540
+ layout: bindGroupLayout,
541
+ entries: [
542
+ { binding: 0, resource: { buffer: uniformBuffer } },
543
+ { binding: 1, resource: { buffer: logits } },
544
+ { binding: 2, resource: { buffer: outputBuffer } },
545
+ { binding: 3, resource: { buffer: topkIndices } },
546
+ { binding: 4, resource: { buffer: topkLogits } },
547
+ ],
548
+ });
505
549
 
506
- const bindGroupLayout = getSampleBindGroupLayout(device);
507
- const bindGroup = device.createBindGroup({
508
- label: 'sample_bind_group',
509
- layout: bindGroupLayout,
510
- entries: [
511
- { binding: 0, resource: { buffer: uniformBuffer } },
512
- { binding: 1, resource: { buffer: logits } },
513
- { binding: 2, resource: { buffer: outputBuffer } },
514
- { binding: 3, resource: { buffer: topkIndices } },
515
- { binding: 4, resource: { buffer: topkLogits } },
516
- ],
517
- });
518
-
519
- const pass1 = recorder.beginComputePass('sample_phase1');
520
- pass1.setPipeline(phase1Pipeline);
521
- pass1.setBindGroup(0, bindGroup);
522
- pass1.dispatchWorkgroups(numWorkgroups);
523
- pass1.end();
524
-
525
- const pass2 = recorder.beginComputePass('sample_phase2');
526
- pass2.setPipeline(phase2Pipeline);
527
- pass2.setBindGroup(0, bindGroup);
528
- pass2.dispatchWorkgroups(1);
529
- pass2.end();
530
-
531
- const pass3 = recorder.beginComputePass('sample_phase3');
532
- pass3.setPipeline(phase3Pipeline);
533
- pass3.setBindGroup(0, bindGroup);
534
- pass3.dispatchWorkgroups(1);
535
- pass3.end();
536
-
537
- recorder.trackTemporaryBuffer(topkLogits);
538
- recorder.trackTemporaryBuffer(topkIndices);
539
-
540
- return outputBuffer;
550
+ recordDispatch(recorder, phase1Pipeline, bindGroup, numWorkgroups, 'sample_phase1');
551
+ recordDispatch(recorder, phase2Pipeline, bindGroup, 1, 'sample_phase2');
552
+ recordDispatch(recorder, phase3Pipeline, bindGroup, 1, 'sample_phase3');
553
+
554
+ recorder.trackTemporaryBuffer(topkLogits);
555
+ recorder.trackTemporaryBuffer(topkIndices);
556
+ completed = true;
557
+ return outputBuffer;
558
+ } finally {
559
+ if (!completed) {
560
+ cleanupRunResources(
561
+ null,
562
+ [topkLogits, topkIndices, ownsOutputBuffer ? outputBuffer : null]
563
+ );
564
+ }
565
+ }
541
566
  }
542
567
 
543
568
 
@@ -0,0 +1,47 @@
1
+ // scatter_add_dynamic_expert_scale.wgsl
2
+
3
+ override WORKGROUP_SIZE: u32 = 256u;
4
+
5
+ struct DynamicUniforms {
6
+ num_tokens: u32,
7
+ hidden_size: u32,
8
+ top_k: u32,
9
+ _pad: u32,
10
+ }
11
+
12
+ @group(0) @binding(0) var<uniform> u_dyn: DynamicUniforms;
13
+ @group(0) @binding(1) var<storage, read> expert_outputs_flat: array<f32>;
14
+ @group(0) @binding(2) var<storage, read> routing_indices: array<u32>;
15
+ @group(0) @binding(3) var<storage, read> routing_weights: array<f32>;
16
+ @group(0) @binding(4) var<storage, read> token_offsets: array<u32>;
17
+ @group(0) @binding(5) var<storage, read> expert_scales: array<f32>;
18
+ @group(0) @binding(6) var<storage, read_write> output_dyn: array<f32>;
19
+
20
+ @compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
21
+ fn scatter_add_dynamic(@builtin(global_invocation_id) gid: vec3<u32>) {
22
+ let tid = gid.x;
23
+ let total_elements = u_dyn.num_tokens * u_dyn.hidden_size;
24
+
25
+ if (tid >= total_elements) {
26
+ return;
27
+ }
28
+
29
+ let token_idx = tid / u_dyn.hidden_size;
30
+ let dim_idx = tid % u_dyn.hidden_size;
31
+ let top_k = u_dyn.top_k;
32
+ let hidden_size = u_dyn.hidden_size;
33
+
34
+ var sum: f32 = 0.0;
35
+ let routing_base = token_idx * top_k;
36
+
37
+ for (var k: u32 = 0u; k < top_k; k = k + 1u) {
38
+ let expert_idx = routing_indices[routing_base + k];
39
+ let weight = routing_weights[routing_base + k] * expert_scales[expert_idx];
40
+ let data_offset = token_offsets[routing_base + k];
41
+ let expert_data_idx = data_offset * hidden_size + dim_idx;
42
+
43
+ sum = sum + weight * expert_outputs_flat[expert_data_idx];
44
+ }
45
+
46
+ output_dyn[tid] = sum;
47
+ }
@@ -0,0 +1,49 @@
1
+ // scatter_add_dynamic_f16_expert_scale.wgsl
2
+
3
+ enable f16;
4
+
5
+ override WORKGROUP_SIZE: u32 = 256u;
6
+
7
+ struct DynamicUniforms {
8
+ num_tokens: u32,
9
+ hidden_size: u32,
10
+ top_k: u32,
11
+ _pad: u32,
12
+ }
13
+
14
+ @group(0) @binding(0) var<uniform> u_dyn: DynamicUniforms;
15
+ @group(0) @binding(1) var<storage, read> expert_outputs_flat: array<f16>;
16
+ @group(0) @binding(2) var<storage, read> routing_indices: array<u32>;
17
+ @group(0) @binding(3) var<storage, read> routing_weights: array<f32>;
18
+ @group(0) @binding(4) var<storage, read> token_offsets: array<u32>;
19
+ @group(0) @binding(5) var<storage, read> expert_scales: array<f32>;
20
+ @group(0) @binding(6) var<storage, read_write> output_dyn: array<f16>;
21
+
22
+ @compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
23
+ fn scatter_add_dynamic(@builtin(global_invocation_id) gid: vec3<u32>) {
24
+ let tid = gid.x;
25
+ let total_elements = u_dyn.num_tokens * u_dyn.hidden_size;
26
+
27
+ if (tid >= total_elements) {
28
+ return;
29
+ }
30
+
31
+ let token_idx = tid / u_dyn.hidden_size;
32
+ let dim_idx = tid % u_dyn.hidden_size;
33
+ let top_k = u_dyn.top_k;
34
+ let hidden_size = u_dyn.hidden_size;
35
+
36
+ var sum: f32 = 0.0;
37
+ let routing_base = token_idx * top_k;
38
+
39
+ for (var k: u32 = 0u; k < top_k; k = k + 1u) {
40
+ let expert_idx = routing_indices[routing_base + k];
41
+ let weight = routing_weights[routing_base + k] * expert_scales[expert_idx];
42
+ let data_offset = token_offsets[routing_base + k];
43
+ let expert_data_idx = data_offset * hidden_size + dim_idx;
44
+
45
+ sum = sum + weight * f32(expert_outputs_flat[expert_data_idx]);
46
+ }
47
+
48
+ output_dyn[tid] = f16(sum);
49
+ }