doppler-gpu 0.4.3 → 0.4.5

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 (435) hide show
  1. package/CHANGELOG.md +28 -0
  2. package/README.md +67 -34
  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/model-load.js +5 -1
  341. package/src/inference/pipelines/text/moe-gpu.d.ts +13 -0
  342. package/src/inference/pipelines/text/moe-gpu.js +445 -16
  343. package/src/inference/pipelines/text/moe-impl.d.ts +13 -3
  344. package/src/inference/pipelines/text/moe-shape-validator.js +13 -1
  345. package/src/inference/pipelines/text/ops.d.ts +29 -2
  346. package/src/inference/pipelines/text/ops.js +19 -0
  347. package/src/inference/pipelines/text/per-layer-inputs.js +5 -0
  348. package/src/inference/pipelines/text/probes.js +4 -5
  349. package/src/inference/pipelines/text/sampling-config.d.ts +4 -10
  350. package/src/inference/pipelines/text/sampling-config.js +103 -64
  351. package/src/inference/pipelines/text/sampling.js +13 -9
  352. package/src/inference/pipelines/text/state.js +17 -0
  353. package/src/inference/pipelines/text/tsir-fixture-writer.d.ts +57 -0
  354. package/src/inference/pipelines/text/tsir-fixture-writer.js +8 -9
  355. package/src/inference/pipelines/text/types.d.ts +84 -3
  356. package/src/inference/pipelines/text/weights.d.ts +4 -2
  357. package/src/inference/pipelines/text/weights.js +4 -5
  358. package/src/inference/pipelines/text.d.ts +24 -0
  359. package/src/inference/pipelines/text.js +162 -38
  360. package/src/inference/pipelines/vision/gemma4.js +20 -9
  361. package/src/inference/tokenizer.d.ts +30 -0
  362. package/src/inference/tokenizer.js +340 -164
  363. package/src/inference/tokenizers/bundled.js +71 -49
  364. package/src/loader/doppler-loader.d.ts +11 -0
  365. package/src/loader/doppler-loader.js +290 -15
  366. package/src/loader/embedding-limit-preflight.d.ts +31 -0
  367. package/src/loader/embedding-limit-preflight.js +216 -0
  368. package/src/loader/embedding-loader.d.ts +8 -2
  369. package/src/loader/embedding-loader.js +263 -8
  370. package/src/loader/experts/expert-loader.d.ts +4 -1
  371. package/src/loader/experts/expert-loader.js +174 -5
  372. package/src/loader/final-weights-loader.d.ts +5 -0
  373. package/src/loader/final-weights-loader.js +167 -4
  374. package/src/loader/layer-loader.d.ts +2 -0
  375. package/src/loader/layer-loader.js +74 -6
  376. package/src/loader/loader-types.d.ts +59 -2
  377. package/src/loader/manifest-config.d.ts +8 -0
  378. package/src/loader/shard-resolver.js +37 -4
  379. package/src/loader/tensors/functional-descriptor-loader.d.ts +9 -0
  380. package/src/loader/tensors/functional-descriptor-loader.js +377 -0
  381. package/src/loader/tensors/tensor-loader.d.ts +36 -1
  382. package/src/loader/tensors/tensor-loader.js +437 -21
  383. package/src/loader/tensors/tensor-reader.d.ts +4 -1
  384. package/src/loader/tensors/tensor-reader.js +11 -4
  385. package/src/loader/weights.d.ts +3 -1
  386. package/src/memory/buffer-pool.js +12 -9
  387. package/src/models/diffusiongemma.d.ts +7 -0
  388. package/src/models/diffusiongemma.js +25 -0
  389. package/src/models/gemma4.js +24 -0
  390. package/src/models/qwen3.js +2 -2
  391. package/src/rules/execution-rules-contract-check.js +93 -10
  392. package/src/rules/inference/capability-transforms.rules.json +26 -2
  393. package/src/rules/inference/config.rules.json +1 -0
  394. package/src/rules/inference/dtype.rules.json +7 -0
  395. package/src/rules/inference/execution.rules.json +108 -17
  396. package/src/rules/inference/ffn.rules.json +18 -0
  397. package/src/rules/kernels/gather.rules.json +81 -0
  398. package/src/rules/kernels/matmul.rules.json +19 -0
  399. package/src/rules/kernels/moe.rules.json +3 -0
  400. package/src/rules/kernels/rep-penalty.rules.json +6 -0
  401. package/src/rules/kernels/rmsnorm-qk.rules.json +6 -0
  402. package/src/rules/kernels/rmsnorm.rules.json +42 -7
  403. package/src/rules/kernels/rope-qk.rules.json +6 -0
  404. package/src/rules/loader/tensor-loader.rules.json +8 -0
  405. package/src/rules/rule-registry.js +6 -0
  406. package/src/storage/artifact-storage-context.d.ts +1 -0
  407. package/src/storage/artifact-storage-context.js +10 -0
  408. package/src/tooling/browser-command-runner.d.ts +0 -3
  409. package/src/tooling/browser-command-runner.js +2 -14
  410. package/src/tooling/command-api-family-normalizers.js +4 -2
  411. package/src/tooling/command-api-helpers.js +2 -1
  412. package/src/tooling/command-api.d.ts +2 -0
  413. package/src/tooling/command-api.js +1 -0
  414. package/src/tooling/command-runner-shared.d.ts +4 -0
  415. package/src/tooling/command-runner-shared.js +2 -1
  416. package/src/tooling/diagnose-runner.js +2 -0
  417. package/src/tooling/litert-package-runtime.js +553 -16
  418. package/src/tooling/node-browser-command-runner.js +10 -4
  419. package/src/tooling/node-command-runner.js +7 -2
  420. package/src/tooling/node-convert-worker-pool.js +3 -0
  421. package/src/tooling/node-convert-worker.js +1 -0
  422. package/src/tooling/node-converter.js +12 -0
  423. package/src/tooling/node-webgpu.d.ts +1 -0
  424. package/src/tooling/node-webgpu.js +23 -11
  425. package/src/tooling/program-bundle.d.ts +52 -1
  426. package/src/tooling/rdrr-integrity-refresh.d.ts +18 -0
  427. package/src/tooling/source-artifact-adapter.js +7 -2
  428. package/src/tooling/source-package-profiles.d.ts +18 -1
  429. package/src/tooling/source-package-profiles.js +109 -1
  430. package/src/tooling/source-runtime-bundle.d.ts +1 -0
  431. package/src/tooling/source-runtime-bundle.js +27 -0
  432. package/src/tooling/source-runtime-converter-config.js +12 -3
  433. package/src/tooling-experimental-exports.shared.d.ts +10 -0
  434. package/src/tooling-experimental-exports.shared.js +10 -0
  435. package/src/version.js +1 -1
@@ -0,0 +1,218 @@
1
+ import { acquireBuffer, releaseBuffer } from '../../memory/buffer-pool.js';
2
+ import {
3
+ getBuffer,
4
+ getLayout,
5
+ getWeightDtype,
6
+ isWeightBuffer,
7
+ resolveWeightBufferMaterialization,
8
+ } from '../weight-buffer.js';
9
+ import {
10
+ createPipeline,
11
+ createUniformBufferWithView,
12
+ getOrCreateBindGroupLayout,
13
+ } from './utils.js';
14
+ import { recordDispatch } from './dispatch.js';
15
+
16
+ const WORKGROUP_SIZE = 256;
17
+ const COLS_PER_WG = 64;
18
+ const THREADS_PER_COL = 4;
19
+
20
+ function getLmHeadArgmaxBindGroupLayout(device) {
21
+ return getOrCreateBindGroupLayout(
22
+ 'lm_head_argmax_bind_group_layout',
23
+ [
24
+ { binding: 0, visibility: GPUShaderStage.COMPUTE, buffer: { type: 'uniform' } },
25
+ { binding: 1, visibility: GPUShaderStage.COMPUTE, buffer: { type: 'read-only-storage' } },
26
+ { binding: 2, visibility: GPUShaderStage.COMPUTE, buffer: { type: 'read-only-storage' } },
27
+ { binding: 3, visibility: GPUShaderStage.COMPUTE, buffer: { type: 'storage' } },
28
+ { binding: 4, visibility: GPUShaderStage.COMPUTE, buffer: { type: 'storage' } },
29
+ { binding: 5, visibility: GPUShaderStage.COMPUTE, buffer: { type: 'storage' } },
30
+ ],
31
+ device
32
+ );
33
+ }
34
+
35
+ async function createLmHeadArgmaxPipeline(device, variant) {
36
+ return createPipeline(
37
+ 'lm_head_argmax',
38
+ variant,
39
+ getLmHeadArgmaxBindGroupLayout(device),
40
+ {
41
+ WORKGROUP_SIZE,
42
+ COLS_PER_WG,
43
+ THREADS_PER_COL,
44
+ }
45
+ );
46
+ }
47
+
48
+ function assertOutputBufferSize(outputBuffer, outputIndex) {
49
+ const minBytes = Math.max(4, (outputIndex + 1) * 4);
50
+ if (outputBuffer.size < minBytes) {
51
+ throw new Error(
52
+ `[LmHeadArgmax] outputBuffer too small for outputIndex=${outputIndex}: ` +
53
+ `${outputBuffer.size} < ${minBytes}.`
54
+ );
55
+ }
56
+ }
57
+
58
+ function resolveLmHeadWeight(lmHead, vocabSize, hiddenSize) {
59
+ if (!isWeightBuffer(lmHead)) {
60
+ throw new Error('[LmHeadArgmax] LM head fusion requires a GPU-resident WeightBuffer.');
61
+ }
62
+ const resolved = resolveWeightBufferMaterialization(lmHead, 'f16');
63
+ if (!isWeightBuffer(resolved)) {
64
+ throw new Error('[LmHeadArgmax] LM head fusion requires a single GPU-resident weight buffer.');
65
+ }
66
+ const dtype = getWeightDtype(resolved);
67
+ if (dtype !== 'f16') {
68
+ throw new Error(`[LmHeadArgmax] LM head fusion requires f16 weights, got "${dtype ?? 'unknown'}".`);
69
+ }
70
+ const layout = getLayout(resolved);
71
+ if (layout !== 'row' && layout !== 'column') {
72
+ throw new Error(`[LmHeadArgmax] LM head fusion requires row or column layout, got "${layout ?? 'unknown'}".`);
73
+ }
74
+ const shape = resolved.shape;
75
+ if (!Array.isArray(shape) || shape.length !== 2) {
76
+ throw new Error(`[LmHeadArgmax] LM head fusion requires 2D weights, got [${shape?.join?.(', ') ?? ''}].`);
77
+ }
78
+ const shapeVocab = layout === 'row' ? shape[0] : shape[1];
79
+ const shapeHidden = layout === 'row' ? shape[1] : shape[0];
80
+ if (shapeVocab < vocabSize || shapeHidden !== hiddenSize) {
81
+ throw new Error(
82
+ `[LmHeadArgmax] LM head shape mismatch: layout=${layout}, ` +
83
+ `shape=[${shape.join(', ')}], vocab=${vocabSize}, hidden=${hiddenSize}.`
84
+ );
85
+ }
86
+ return {
87
+ buffer: getBuffer(resolved),
88
+ transposeB: layout === 'row',
89
+ };
90
+ }
91
+
92
+ function createUniformBuffer(device, recorder, options) {
93
+ const padTokenId = options.padTokenId == null ? 0xFFFFFFFF : options.padTokenId;
94
+ return createUniformBufferWithView(
95
+ 'lm_head_argmax_uniforms',
96
+ 32,
97
+ (view) => {
98
+ view.setUint32(0, options.vocabSize, true);
99
+ view.setUint32(4, options.hiddenSize, true);
100
+ view.setUint32(8, options.transposeB ? 1 : 0, true);
101
+ view.setUint32(12, options.workgroupsX, true);
102
+ view.setUint32(16, padTokenId, true);
103
+ view.setFloat32(20, options.logitSoftcap, true);
104
+ view.setUint32(24, options.outputIndex, true);
105
+ view.setUint32(28, options.numGroups, true);
106
+ },
107
+ recorder,
108
+ device
109
+ );
110
+ }
111
+
112
+ function planDispatch(device, vocabSize) {
113
+ const numGroups = Math.ceil(vocabSize / COLS_PER_WG);
114
+ const maxWorkgroups = Number.isFinite(device.limits?.maxComputeWorkgroupsPerDimension)
115
+ ? device.limits.maxComputeWorkgroupsPerDimension
116
+ : 65535;
117
+ const workgroupsX = Math.min(numGroups, maxWorkgroups);
118
+ const workgroupsY = Math.ceil(numGroups / workgroupsX);
119
+ if (workgroupsY > maxWorkgroups) {
120
+ throw new Error(
121
+ `[LmHeadArgmax] dispatch exceeds WebGPU workgroup limits: ` +
122
+ `workgroupsX=${workgroupsX}, workgroupsY=${workgroupsY}, max=${maxWorkgroups}.`
123
+ );
124
+ }
125
+ return {
126
+ numGroups,
127
+ workgroups: [workgroupsX, workgroupsY, 1],
128
+ workgroupsX,
129
+ };
130
+ }
131
+
132
+ export async function recordLmHeadArgmaxF16(recorder, inputTensor, lmHead, options = {}) {
133
+ if (!recorder?.device) {
134
+ throw new Error('[LmHeadArgmax] CommandRecorder is required.');
135
+ }
136
+ if (inputTensor?.dtype !== 'f32') {
137
+ throw new Error(`[LmHeadArgmax] input tensor must be f32, got "${inputTensor?.dtype ?? 'unknown'}".`);
138
+ }
139
+ const vocabSize = options.vocabSize;
140
+ const hiddenSize = options.hiddenSize;
141
+ if (!Number.isInteger(vocabSize) || vocabSize <= 0) {
142
+ throw new Error(`[LmHeadArgmax] vocabSize must be a positive integer, got "${vocabSize}".`);
143
+ }
144
+ if (!Number.isInteger(hiddenSize) || hiddenSize <= 0) {
145
+ throw new Error(`[LmHeadArgmax] hiddenSize must be a positive integer, got "${hiddenSize}".`);
146
+ }
147
+ if (options.outputIndex == null) {
148
+ throw new Error('[LmHeadArgmax] outputIndex is required.');
149
+ }
150
+ if (options.logitSoftcap === undefined) {
151
+ throw new Error('[LmHeadArgmax] logitSoftcap is required.');
152
+ }
153
+ if (options.padTokenId === undefined) {
154
+ throw new Error('[LmHeadArgmax] padTokenId is required.');
155
+ }
156
+
157
+ const device = recorder.device;
158
+ const weight = resolveLmHeadWeight(lmHead, vocabSize, hiddenSize);
159
+ const dispatchPlan = planDispatch(device, vocabSize);
160
+ const minOutputBytes = Math.max(4, (options.outputIndex + 1) * 4);
161
+ const outputBuffer = options.outputBuffer ?? acquireBuffer(minOutputBytes, undefined, 'lm_head_argmax_output');
162
+ const ownsOutput = !options.outputBuffer;
163
+ assertOutputBufferSize(outputBuffer, options.outputIndex);
164
+
165
+ let tempIndices = null;
166
+ let tempLogits = null;
167
+ let completed = false;
168
+ try {
169
+ tempIndices = acquireBuffer(dispatchPlan.numGroups * 4, undefined, 'lm_head_argmax_temp_indices');
170
+ tempLogits = acquireBuffer(dispatchPlan.numGroups * 4, undefined, 'lm_head_argmax_temp_logits');
171
+ const uniformBuffer = createUniformBuffer(device, recorder, {
172
+ vocabSize,
173
+ hiddenSize,
174
+ transposeB: weight.transposeB,
175
+ workgroupsX: dispatchPlan.workgroupsX,
176
+ padTokenId: options.padTokenId,
177
+ logitSoftcap: options.logitSoftcap,
178
+ outputIndex: options.outputIndex,
179
+ numGroups: dispatchPlan.numGroups,
180
+ });
181
+ const layout = getLmHeadArgmaxBindGroupLayout(device);
182
+ const entries = [
183
+ { binding: 0, resource: { buffer: uniformBuffer } },
184
+ { binding: 1, resource: { buffer: inputTensor.buffer } },
185
+ { binding: 2, resource: { buffer: weight.buffer } },
186
+ { binding: 3, resource: { buffer: outputBuffer } },
187
+ { binding: 4, resource: { buffer: tempIndices } },
188
+ { binding: 5, resource: { buffer: tempLogits } },
189
+ ];
190
+ const bindGroup = device.createBindGroup({
191
+ label: 'lm_head_argmax_bind_group',
192
+ layout,
193
+ entries,
194
+ });
195
+ const phase1Pipeline = await createLmHeadArgmaxPipeline(device, 'phase1_f16w_f32a');
196
+ recordDispatch(
197
+ recorder,
198
+ phase1Pipeline,
199
+ bindGroup,
200
+ dispatchPlan.workgroups,
201
+ 'lm_head_argmax_phase1'
202
+ );
203
+
204
+ const phase2Pipeline = await createLmHeadArgmaxPipeline(device, 'phase2');
205
+ recordDispatch(recorder, phase2Pipeline, bindGroup, [1, 1, 1], 'lm_head_argmax_phase2');
206
+
207
+ recorder.trackTemporaryBuffer(tempIndices);
208
+ recorder.trackTemporaryBuffer(tempLogits);
209
+ completed = true;
210
+ return outputBuffer;
211
+ } finally {
212
+ if (!completed) {
213
+ if (tempIndices) releaseBuffer(tempIndices);
214
+ if (tempLogits) releaseBuffer(tempLogits);
215
+ if (ownsOutput) releaseBuffer(outputBuffer);
216
+ }
217
+ }
218
+ }
@@ -0,0 +1,189 @@
1
+ // lm-head-argmax-f16.wgsl
2
+
3
+ enable f16;
4
+
5
+ override WORKGROUP_SIZE: u32 = 256u;
6
+ override COLS_PER_WG: u32 = 64u;
7
+ override THREADS_PER_COL: u32 = 4u;
8
+
9
+ const MAX_WORKGROUP_SIZE: u32 = 256u;
10
+ const MAX_COLS_PER_WG: u32 = 64u;
11
+ const NEG_INF: f32 = -3.402823e+38;
12
+
13
+ struct Uniforms {
14
+ vocab_size: u32,
15
+ hidden_size: u32,
16
+ transpose_b: u32,
17
+ workgroups_x: u32,
18
+ pad_token_id: u32,
19
+ logit_softcap: f32,
20
+ output_index: u32,
21
+ num_groups: u32,
22
+ }
23
+
24
+ @group(0) @binding(0) var<uniform> u: Uniforms;
25
+ @group(0) @binding(1) var<storage, read> hidden: array<f32>;
26
+ @group(0) @binding(2) var<storage, read> weights: array<f16>;
27
+ @group(0) @binding(3) var<storage, read_write> output: array<u32>;
28
+ @group(0) @binding(4) var<storage, read_write> temp_indices: array<u32>;
29
+ @group(0) @binding(5) var<storage, read_write> temp_logits: array<f32>;
30
+
31
+ var<workgroup> partial_sums: array<f32, MAX_WORKGROUP_SIZE>;
32
+ var<workgroup> candidate_values: array<f32, MAX_WORKGROUP_SIZE>;
33
+ var<workgroup> candidate_indices: array<u32, MAX_WORKGROUP_SIZE>;
34
+
35
+ fn apply_softcap(x: f32, softcap: f32) -> f32 {
36
+ if (softcap <= 0.0) {
37
+ return x;
38
+ }
39
+ return softcap * tanh(x / softcap);
40
+ }
41
+
42
+ fn candidate_beats(candidate_value: f32, candidate_index: u32, best_value: f32, best_index: u32) -> bool {
43
+ if (candidate_value > best_value) {
44
+ return true;
45
+ }
46
+ if (candidate_value < best_value) {
47
+ return false;
48
+ }
49
+ return candidate_index < best_index;
50
+ }
51
+
52
+ @compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
53
+ fn phase1(
54
+ @builtin(local_invocation_id) lid: vec3<u32>,
55
+ @builtin(workgroup_id) wg_id: vec3<u32>
56
+ ) {
57
+ let local_id = lid.x;
58
+ let col_in_wg = local_id / THREADS_PER_COL;
59
+ let thread_in_col = local_id % THREADS_PER_COL;
60
+ let wg_linear = wg_id.y * u.workgroups_x + wg_id.x;
61
+ let base_col = wg_linear * COLS_PER_WG;
62
+ let col = base_col + col_in_wg;
63
+ let is_valid = col < u.vocab_size && col != u.pad_token_id;
64
+
65
+ var partial_sum: f32 = 0.0;
66
+ if (is_valid) {
67
+ let k_per_thread = (u.hidden_size + THREADS_PER_COL - 1u) / THREADS_PER_COL;
68
+ let k_start = thread_in_col * k_per_thread;
69
+ let k_end = min(k_start + k_per_thread, u.hidden_size);
70
+ var k = k_start;
71
+ let k_aligned_end = k_start + ((k_end - k_start) / 4u) * 4u;
72
+
73
+ if (u.transpose_b == 1u) {
74
+ let row_offset = col * u.hidden_size;
75
+ for (; k < k_aligned_end; k = k + 4u) {
76
+ let a = vec4<f32>(hidden[k], hidden[k + 1u], hidden[k + 2u], hidden[k + 3u]);
77
+ let b = vec4<f32>(
78
+ f32(weights[row_offset + k]),
79
+ f32(weights[row_offset + k + 1u]),
80
+ f32(weights[row_offset + k + 2u]),
81
+ f32(weights[row_offset + k + 3u])
82
+ );
83
+ partial_sum = partial_sum + dot(a, b);
84
+ }
85
+ for (; k < k_end; k = k + 1u) {
86
+ partial_sum = partial_sum + hidden[k] * f32(weights[row_offset + k]);
87
+ }
88
+ } else {
89
+ for (; k < k_aligned_end; k = k + 4u) {
90
+ let a = vec4<f32>(hidden[k], hidden[k + 1u], hidden[k + 2u], hidden[k + 3u]);
91
+ let b = vec4<f32>(
92
+ f32(weights[k * u.vocab_size + col]),
93
+ f32(weights[(k + 1u) * u.vocab_size + col]),
94
+ f32(weights[(k + 2u) * u.vocab_size + col]),
95
+ f32(weights[(k + 3u) * u.vocab_size + col])
96
+ );
97
+ partial_sum = partial_sum + dot(a, b);
98
+ }
99
+ for (; k < k_end; k = k + 1u) {
100
+ partial_sum = partial_sum + hidden[k] * f32(weights[k * u.vocab_size + col]);
101
+ }
102
+ }
103
+ }
104
+
105
+ partial_sums[local_id] = partial_sum;
106
+ workgroupBarrier();
107
+
108
+ if (thread_in_col == 0u && col_in_wg < MAX_COLS_PER_WG) {
109
+ let base = col_in_wg * THREADS_PER_COL;
110
+ var sum = partial_sums[base];
111
+ for (var i: u32 = 1u; i < THREADS_PER_COL; i = i + 1u) {
112
+ sum = sum + partial_sums[base + i];
113
+ }
114
+ candidate_values[col_in_wg] = select(NEG_INF, apply_softcap(sum, u.logit_softcap), is_valid);
115
+ candidate_indices[col_in_wg] = col;
116
+ }
117
+ workgroupBarrier();
118
+
119
+ if (local_id < COLS_PER_WG) {
120
+ candidate_values[local_id] = select(NEG_INF, candidate_values[local_id], local_id < MAX_COLS_PER_WG);
121
+ candidate_indices[local_id] = select(0u, candidate_indices[local_id], local_id < MAX_COLS_PER_WG);
122
+ }
123
+ workgroupBarrier();
124
+
125
+ var stride = COLS_PER_WG / 2u;
126
+ while (stride > 0u) {
127
+ if (local_id < stride) {
128
+ if (candidate_beats(
129
+ candidate_values[local_id + stride],
130
+ candidate_indices[local_id + stride],
131
+ candidate_values[local_id],
132
+ candidate_indices[local_id]
133
+ )) {
134
+ candidate_values[local_id] = candidate_values[local_id + stride];
135
+ candidate_indices[local_id] = candidate_indices[local_id + stride];
136
+ }
137
+ }
138
+ workgroupBarrier();
139
+ stride = stride / 2u;
140
+ }
141
+
142
+ if (local_id == 0u) {
143
+ temp_logits[wg_linear] = candidate_values[0];
144
+ temp_indices[wg_linear] = candidate_indices[0];
145
+ }
146
+ }
147
+
148
+ @compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
149
+ fn phase2(@builtin(local_invocation_id) lid: vec3<u32>) {
150
+ let local_id = lid.x;
151
+ var local_best_value: f32 = NEG_INF;
152
+ var local_best_index: u32 = 0u;
153
+
154
+ var group = local_id;
155
+ while (group < u.num_groups) {
156
+ let candidate_value = temp_logits[group];
157
+ let candidate_index = temp_indices[group];
158
+ if (candidate_beats(candidate_value, candidate_index, local_best_value, local_best_index)) {
159
+ local_best_value = candidate_value;
160
+ local_best_index = candidate_index;
161
+ }
162
+ group = group + WORKGROUP_SIZE;
163
+ }
164
+
165
+ candidate_values[local_id] = local_best_value;
166
+ candidate_indices[local_id] = local_best_index;
167
+ workgroupBarrier();
168
+
169
+ var stride = WORKGROUP_SIZE / 2u;
170
+ while (stride > 0u) {
171
+ if (local_id < stride) {
172
+ if (candidate_beats(
173
+ candidate_values[local_id + stride],
174
+ candidate_indices[local_id + stride],
175
+ candidate_values[local_id],
176
+ candidate_indices[local_id]
177
+ )) {
178
+ candidate_values[local_id] = candidate_values[local_id + stride];
179
+ candidate_indices[local_id] = candidate_indices[local_id + stride];
180
+ }
181
+ }
182
+ workgroupBarrier();
183
+ stride = stride / 2u;
184
+ }
185
+
186
+ if (local_id == 0u) {
187
+ output[u.output_index] = candidate_indices[0];
188
+ }
189
+ }
@@ -17,6 +17,17 @@ export declare function calculateMatmulDispatch(
17
17
  config: KernelConfig
18
18
  ): { workgroups: GPUExtent3D; uniformWorkgroupsX?: number };
19
19
 
20
+ export declare function calculateMatmulDispatch(
21
+ variant: string,
22
+ useQ4KFused: boolean,
23
+ useGemv: boolean,
24
+ useLiteRTInt4Fused: boolean,
25
+ M: number,
26
+ N: number,
27
+ config: KernelConfig,
28
+ useW4A16Fused?: boolean
29
+ ): { workgroups: GPUExtent3D; uniformWorkgroupsX?: number };
30
+
20
31
  export declare function createMatmulUniformBuffer(
21
32
  label: string,
22
33
  M: number,
@@ -21,7 +21,13 @@ export class MatmulKernel extends KernelBase {
21
21
  }
22
22
 
23
23
 
24
- export function calculateMatmulDispatch(variant, useQ4KFused, useGemv, M, N, config) {
24
+ export function calculateMatmulDispatch(variant, useQ4KFused, useGemv, useLiteRTInt4Fused, M, N, config, useW4A16Fused = false) {
25
+ if (typeof useLiteRTInt4Fused === 'number') {
26
+ config = N;
27
+ N = M;
28
+ M = useLiteRTInt4Fused;
29
+ useLiteRTInt4Fused = false;
30
+ }
25
31
  const maxWorkgroups = GPU_LIMITS.MAX_WORKGROUPS;
26
32
  const [wgX, wgY] = config.workgroupSize;
27
33
  let workgroupsX = 1;
@@ -40,6 +46,36 @@ export function calculateMatmulDispatch(variant, useQ4KFused, useGemv, M, N, con
40
46
  if (useQ4KFused && variant.includes('batched') && tileM == null) {
41
47
  throw new Error(`Matmul kernel "${variant}" is missing variantMetadata.tileM.`);
42
48
  }
49
+ if ((useLiteRTInt4Fused || useW4A16Fused) && colsPerWg == null) {
50
+ throw new Error(`Matmul kernel "${variant}" is missing variantMetadata.colsPerWg.`);
51
+ }
52
+ if (useW4A16Fused && variant.includes('batched') && tileM == null) {
53
+ throw new Error(`Matmul kernel "${variant}" is missing variantMetadata.tileM.`);
54
+ }
55
+
56
+ if (useW4A16Fused && variant.includes('batched')) {
57
+ workgroupsX = Math.ceil(N / colsPerWg);
58
+ workgroupsY = Math.ceil(M / tileM);
59
+ if (workgroupsX > maxWorkgroups || workgroupsY > maxWorkgroups) {
60
+ throw new Error(
61
+ `Matmul kernel "${variant}" dispatch exceeds WebGPU workgroup limits: ` +
62
+ `workgroupsX=${workgroupsX}, workgroupsY=${workgroupsY}, max=${maxWorkgroups}.`
63
+ );
64
+ }
65
+ return { workgroups: [workgroupsX, workgroupsY, 1], uniformWorkgroupsX };
66
+ }
67
+
68
+ if (useLiteRTInt4Fused || useW4A16Fused) {
69
+ workgroupsX = Math.ceil(N / colsPerWg);
70
+ workgroupsY = M;
71
+ if (workgroupsX > maxWorkgroups || workgroupsY > maxWorkgroups) {
72
+ throw new Error(
73
+ `Matmul kernel "${variant}" dispatch exceeds WebGPU workgroup limits: ` +
74
+ `workgroupsX=${workgroupsX}, workgroupsY=${workgroupsY}, max=${maxWorkgroups}.`
75
+ );
76
+ }
77
+ return { workgroups: [workgroupsX, workgroupsY, 1], uniformWorkgroupsX };
78
+ }
43
79
 
44
80
  if (useGemv && variant.startsWith('gemv_subgroup')) {
45
81
  if (colsPerWg == null) {
@@ -24,14 +24,14 @@ export declare function isFusedQ4KDisabled(options?: {
24
24
  kernelPath?: KernelPathSchema | null;
25
25
  }): boolean;
26
26
 
27
- export declare function toMatmulDtype(dtype: string | null | undefined): 'f16' | 'f32' | 'q4k';
27
+ export declare function toMatmulDtype(dtype: string | null | undefined): 'f16' | 'f32' | 'q4k' | 'litert_int4' | 'w4a16';
28
28
 
29
29
  export declare function selectMatmulKernel(options?: {
30
30
  preferF16?: boolean;
31
31
  useVec4?: boolean;
32
32
  outputDtype?: TensorDtype | 'f16' | 'f32';
33
33
  aDtype?: 'f16' | 'f32' | null;
34
- bDtype?: 'f16' | 'f32' | 'q4k' | null;
34
+ bDtype?: 'f16' | 'f32' | 'q4k' | 'litert_int4' | 'w4a16' | null;
35
35
  isPrefill?: boolean;
36
36
  prefillRows?: number;
37
37
  transposeB?: boolean;
@@ -54,7 +54,7 @@ export declare function getMatmulBindingSizes(
54
54
  N: number,
55
55
  K: number,
56
56
  aDtype: 'f16' | 'f32',
57
- bDtype: 'f16' | 'f32' | 'q4k',
57
+ bDtype: 'f16' | 'f32' | 'q4k' | 'litert_int4' | 'w4a16',
58
58
  transposeB: boolean,
59
59
  aOffset: number,
60
60
  bOffset: number
@@ -68,11 +68,11 @@ export declare function selectMatmulVariantAndFlags(
68
68
  N: number,
69
69
  K: number,
70
70
  aDtype: 'f16' | 'f32',
71
- bDtype: 'f16' | 'f32' | 'q4k',
71
+ bDtype: 'f16' | 'f32' | 'q4k' | 'litert_int4' | 'w4a16',
72
72
  transposeB: boolean,
73
73
  requestedOutputDtype: TensorDtype | 'f16' | 'f32',
74
74
  options: { role?: string; layerIdx?: number; kernelPath?: KernelPathSchema | null; [key: string]: unknown }
75
- ): { variant: string; useQ4KFused: boolean; useGemv: boolean };
75
+ ): { variant: string; useQ4KFused: boolean; useGemv: boolean; useLiteRTInt4Fused?: boolean; useW4A16Fused?: boolean };
76
76
 
77
77
  export declare function resolveMatmulOutput(
78
78
  variant: string,