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
@@ -0,0 +1,46 @@
1
+ enable f16;
2
+
3
+ struct Uniforms {
4
+ canvas_length: u32,
5
+ hidden_size: u32,
6
+ vocab_size: u32,
7
+ row_start: u32,
8
+ row_count: u32,
9
+ output_mode: u32,
10
+ _pad0: u32,
11
+ _pad1: u32,
12
+ }
13
+
14
+ @group(0) @binding(0) var<uniform> u: Uniforms;
15
+ @group(0) @binding(1) var<storage, read> softmax_rows: array<f32>;
16
+ @group(0) @binding(2) var<storage, read> embedding_section: array<f16>;
17
+ @group(0) @binding(3) var<storage, read_write> output: array<f32>;
18
+
19
+ @compute @workgroup_size(8, 8, 1)
20
+ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
21
+ let hidden_idx = gid.x;
22
+ let token_idx = gid.y;
23
+ if (hidden_idx >= u.hidden_size || token_idx >= u.canvas_length) {
24
+ return;
25
+ }
26
+
27
+ var sum = 0.0;
28
+ var row = 0u;
29
+ loop {
30
+ if (row >= u.row_count) {
31
+ break;
32
+ }
33
+ let vocab_idx = u.row_start + row;
34
+ let probability = softmax_rows[token_idx * u.vocab_size + vocab_idx];
35
+ let embedding = f32(embedding_section[row * u.hidden_size + hidden_idx]);
36
+ sum += probability * embedding;
37
+ row += 1u;
38
+ }
39
+
40
+ let out_idx = token_idx * u.hidden_size + hidden_idx;
41
+ if (u.output_mode == 0u) {
42
+ output[out_idx] = sum;
43
+ } else {
44
+ output[out_idx] = output[out_idx] + sum;
45
+ }
46
+ }
@@ -1,9 +1,9 @@
1
- // Online Softmax Kernel
1
+ // Softmax Kernels
2
2
  //
3
- // Numerically stable softmax using online algorithm:
4
- // 1. Track running max while iterating
5
- // 2. Compute exp(x - max) and sum in same pass
6
- // 3. Normalize by sum
3
+ // Numerically stable softmax:
4
+ // 1. Reduce max for each row.
5
+ // 2. Compute exp(x - max) and reduce sum.
6
+ // 3. Normalize by the reduced sum.
7
7
  //
8
8
  // Supports softmax along last dimension (axis=-1).
9
9
  // Subgroup variants use subgroupMax/subgroupAdd for faster reductions.
@@ -171,78 +171,3 @@ fn softmax_small(
171
171
  output[base_offset + thread_idx] = exp_val * inv_sum;
172
172
  }
173
173
  }
174
-
175
- // Online softmax - single pass algorithm
176
- // More memory efficient but requires careful implementation
177
- @compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
178
- fn softmax_online(
179
- @builtin(local_invocation_id) local_id: vec3<u32>,
180
- @builtin(workgroup_id) wg_id: vec3<u32>
181
- ) {
182
- let row_idx = wg_id.x;
183
- let thread_idx = local_id.x;
184
- let inner_size = u.inner_size;
185
- let temperature = u.temperature;
186
-
187
- if (row_idx >= u.outer_size) {
188
- return;
189
- }
190
-
191
- let base_offset = row_idx * inner_size;
192
- let elements_per_thread = (inner_size + WORKGROUP_SIZE - 1u) / WORKGROUP_SIZE;
193
-
194
- // Online algorithm: track max and sum simultaneously
195
- var m: f32 = NEG_INF; // Running max
196
- var d: f32 = 0.0; // Running sum of exp(x - m)
197
-
198
- for (var i: u32 = 0u; i < elements_per_thread; i = i + 1u) {
199
- let idx = thread_idx * elements_per_thread + i;
200
- if (idx < inner_size) {
201
- let x = input[base_offset + idx] / temperature;
202
-
203
- // Update running max and rescale sum
204
- let m_new = max(m, x);
205
- d = d * exp(m - m_new) + exp(x - m_new);
206
- m = m_new;
207
- }
208
- }
209
-
210
- // Store for reduction
211
- shared_max[thread_idx] = m;
212
- shared_sum[thread_idx] = d;
213
- workgroupBarrier();
214
-
215
- // Pairwise reduction combining (max, sum) pairs
216
- for (var stride: u32 = WORKGROUP_SIZE / 2u; stride > 0u; stride = stride >> 1u) {
217
- if (thread_idx < stride) {
218
- let m1 = shared_max[thread_idx];
219
- let d1 = shared_sum[thread_idx];
220
- let m2 = shared_max[thread_idx + stride];
221
- let d2 = shared_sum[thread_idx + stride];
222
-
223
- let m_new = max(m1, m2);
224
- let d_new = d1 * exp(m1 - m_new) + d2 * exp(m2 - m_new);
225
-
226
- shared_max[thread_idx] = m_new;
227
- shared_sum[thread_idx] = d_new;
228
- }
229
- workgroupBarrier();
230
- }
231
-
232
- let global_max = shared_max[0];
233
- let global_sum = shared_sum[0];
234
- // Guard against division by zero when all exp values underflow
235
- let inv_sum = select(0.0, 1.0 / global_sum, global_sum > 0.0);
236
-
237
- workgroupBarrier();
238
-
239
- // Write normalized output
240
- for (var i: u32 = 0u; i < elements_per_thread; i = i + 1u) {
241
- let idx = thread_idx * elements_per_thread + i;
242
- if (idx < inner_size) {
243
- let x = input[base_offset + idx] / temperature;
244
- output[base_offset + idx] = exp(x - global_max) * inv_sum;
245
- }
246
- }
247
- }
248
-
@@ -0,0 +1,44 @@
1
+ import type { Tensor } from '../tensor.js';
2
+ import type { CommandRecorder } from '../command-recorder.js';
3
+ import type { TensorLike, WeightBuffer } from '../weight-buffer.js';
4
+
5
+ export interface SplitQKVRMSNormQKOptions {
6
+ numTokens: number;
7
+ numHeads: number;
8
+ numKVHeads: number;
9
+ headDim: number;
10
+ qSize: number;
11
+ kSize: number;
12
+ vSize: number;
13
+ rmsNormWeightOffset?: boolean;
14
+ skipKNorm?: boolean;
15
+ allowUnitQKNorm?: boolean;
16
+ }
17
+
18
+ export interface SplitQKVRMSNormQKResult {
19
+ Q: Tensor;
20
+ K: Tensor;
21
+ V: Tensor;
22
+ }
23
+
24
+ export function canUseSplitQKVRMSNormQK(
25
+ qkvTensor: Tensor | null | undefined,
26
+ options?: Pick<SplitQKVRMSNormQKOptions, 'skipKNorm' | 'allowUnitQKNorm'>
27
+ ): boolean;
28
+
29
+ export function runSplitQKVRMSNormQK(
30
+ qkvTensor: Tensor,
31
+ qWeight: GPUBuffer | WeightBuffer | TensorLike,
32
+ kWeight: GPUBuffer | WeightBuffer | TensorLike,
33
+ eps: number,
34
+ options: SplitQKVRMSNormQKOptions
35
+ ): Promise<SplitQKVRMSNormQKResult>;
36
+
37
+ export function recordSplitQKVRMSNormQK(
38
+ recorder: CommandRecorder,
39
+ qkvTensor: Tensor,
40
+ qWeight: GPUBuffer | WeightBuffer | TensorLike,
41
+ kWeight: GPUBuffer | WeightBuffer | TensorLike,
42
+ eps: number,
43
+ options: SplitQKVRMSNormQKOptions
44
+ ): Promise<SplitQKVRMSNormQKResult>;
@@ -0,0 +1,154 @@
1
+ import { acquireBuffer, releaseBuffer } from '../../memory/buffer-pool.js';
2
+ import { padToQ4KBlock } from '../../config/schema/index.js';
3
+ import { createTensor, dtypeBytes } from '../tensor.js';
4
+ import { getBuffer } from '../weight-buffer.js';
5
+ import { unifiedKernelWrapper } from './utils.js';
6
+ import {
7
+ assertRMSNormWeightBuffer,
8
+ planRMSNormDispatch,
9
+ resolveNormWeightDtype,
10
+ } from './rmsnorm.js';
11
+ import { WORKGROUP_SIZES } from './constants.js';
12
+
13
+ function assertPositiveInteger(value, label) {
14
+ if (!Number.isInteger(value) || value <= 0) {
15
+ throw new Error(`split_qkv_rmsnorm_qk requires ${label} to be a positive integer.`);
16
+ }
17
+ }
18
+
19
+ function assertQKVSizes(options) {
20
+ const {
21
+ qSize,
22
+ kSize,
23
+ vSize,
24
+ numHeads,
25
+ numKVHeads,
26
+ headDim,
27
+ } = options;
28
+ if (qSize !== numHeads * headDim) {
29
+ throw new Error(`split_qkv_rmsnorm_qk qSize mismatch: got ${qSize}, expected ${numHeads * headDim}.`);
30
+ }
31
+ if (kSize !== numKVHeads * headDim) {
32
+ throw new Error(`split_qkv_rmsnorm_qk kSize mismatch: got ${kSize}, expected ${numKVHeads * headDim}.`);
33
+ }
34
+ if (vSize !== numKVHeads * headDim) {
35
+ throw new Error(`split_qkv_rmsnorm_qk vSize mismatch: got ${vSize}, expected ${numKVHeads * headDim}.`);
36
+ }
37
+ }
38
+
39
+ export function canUseSplitQKVRMSNormQK(qkvTensor, options = {}) {
40
+ if (!qkvTensor || qkvTensor.dtype !== 'f32') {
41
+ return false;
42
+ }
43
+ if (options.skipKNorm === true || options.allowUnitQKNorm === true) {
44
+ return false;
45
+ }
46
+ return true;
47
+ }
48
+
49
+ async function splitQKVRMSNormQK(target, qkvTensor, qWeight, kWeight, eps, options = {}) {
50
+ const {
51
+ numTokens,
52
+ numHeads,
53
+ numKVHeads,
54
+ headDim,
55
+ qSize,
56
+ kSize,
57
+ vSize,
58
+ rmsNormWeightOffset = false,
59
+ } = options;
60
+
61
+ assertPositiveInteger(numTokens, 'numTokens');
62
+ assertPositiveInteger(numHeads, 'numHeads');
63
+ assertPositiveInteger(numKVHeads, 'numKVHeads');
64
+ assertPositiveInteger(headDim, 'headDim');
65
+ assertPositiveInteger(qSize, 'qSize');
66
+ assertPositiveInteger(kSize, 'kSize');
67
+ assertPositiveInteger(vSize, 'vSize');
68
+ assertQKVSizes(options);
69
+ if (qkvTensor.dtype !== 'f32') {
70
+ throw new Error(`split_qkv_rmsnorm_qk requires f32 QKV input, got ${qkvTensor.dtype}.`);
71
+ }
72
+
73
+ const qWeightBuffer = getBuffer(qWeight);
74
+ const kWeightBuffer = getBuffer(kWeight);
75
+ assertRMSNormWeightBuffer(qWeight, qWeightBuffer, headDim);
76
+ assertRMSNormWeightBuffer(kWeight, kWeightBuffer, headDim);
77
+ const qWeightDtype = resolveNormWeightDtype(qWeight, headDim);
78
+ const kWeightDtype = resolveNormWeightDtype(kWeight, headDim);
79
+
80
+ const qRows = numTokens * numHeads;
81
+ const kRows = numTokens * numKVHeads;
82
+ const qkRows = qRows + kRows;
83
+ const totalV = numTokens * vSize;
84
+ const vWorkgroups = Math.ceil(totalV / WORKGROUP_SIZES.DEFAULT);
85
+ const totalWorkgroups = qkRows + vWorkgroups;
86
+ const dispatchPlan = planRMSNormDispatch(target, totalWorkgroups);
87
+ const bytesPerElement = dtypeBytes(qkvTensor.dtype);
88
+ const paddedHeadDim = padToQ4KBlock(headDim);
89
+ const qOutputSize = qRows * paddedHeadDim * bytesPerElement;
90
+ const kOutputSize = kRows * paddedHeadDim * bytesPerElement;
91
+ const vOutputSize = totalV * bytesPerElement;
92
+ let qOutputBuffer = null;
93
+ let kOutputBuffer = null;
94
+ let vOutputBuffer = null;
95
+
96
+ try {
97
+ qOutputBuffer = acquireBuffer(qOutputSize, undefined, 'split_qkv_rmsnorm_qk_q_output');
98
+ kOutputBuffer = acquireBuffer(kOutputSize, undefined, 'split_qkv_rmsnorm_qk_k_output');
99
+ vOutputBuffer = acquireBuffer(vOutputSize, undefined, 'split_qkv_rmsnorm_qk_v_output');
100
+ await unifiedKernelWrapper(
101
+ 'split_qkv_rmsnorm_qk',
102
+ target,
103
+ 'default',
104
+ [qkvTensor, qWeightBuffer, kWeightBuffer, qOutputBuffer, kOutputBuffer, vOutputBuffer],
105
+ {
106
+ num_tokens: numTokens,
107
+ q_size: qSize,
108
+ k_size: kSize,
109
+ v_size: vSize,
110
+ num_heads: numHeads,
111
+ num_kv_heads: numKVHeads,
112
+ head_dim: headDim,
113
+ workgroup_stride: dispatchPlan.tokenStride,
114
+ qk_rows: qkRows,
115
+ total_v: totalV,
116
+ eps,
117
+ _pad0: 0,
118
+ },
119
+ dispatchPlan.workgroups,
120
+ {
121
+ RMS_NORM_OFFSET: rmsNormWeightOffset,
122
+ Q_WEIGHT_IS_F16: qWeightDtype === 'f16',
123
+ K_WEIGHT_IS_F16: kWeightDtype === 'f16',
124
+ },
125
+ null,
126
+ 'split_qkv_rmsnorm_qk'
127
+ );
128
+
129
+ return {
130
+ Q: createTensor(qOutputBuffer, 'f32', [qRows, headDim], 'split_qkv_rmsnorm_qk_q_output'),
131
+ K: createTensor(kOutputBuffer, 'f32', [kRows, headDim], 'split_qkv_rmsnorm_qk_k_output'),
132
+ V: createTensor(vOutputBuffer, 'f32', [numTokens, vSize], 'split_qkv_rmsnorm_qk_v_output'),
133
+ };
134
+ } catch (error) {
135
+ if (qOutputBuffer) {
136
+ releaseBuffer(qOutputBuffer);
137
+ }
138
+ if (kOutputBuffer) {
139
+ releaseBuffer(kOutputBuffer);
140
+ }
141
+ if (vOutputBuffer) {
142
+ releaseBuffer(vOutputBuffer);
143
+ }
144
+ throw error;
145
+ }
146
+ }
147
+
148
+ export async function runSplitQKVRMSNormQK(qkvTensor, qWeight, kWeight, eps, options = {}) {
149
+ return splitQKVRMSNormQK(null, qkvTensor, qWeight, kWeight, eps, options);
150
+ }
151
+
152
+ export async function recordSplitQKVRMSNormQK(recorder, qkvTensor, qWeight, kWeight, eps, options = {}) {
153
+ return splitQKVRMSNormQK(recorder, qkvTensor, qWeight, kWeight, eps, options);
154
+ }
@@ -0,0 +1,59 @@
1
+ import type { Tensor } from '../tensor.js';
2
+ import type { CommandRecorder } from '../command-recorder.js';
3
+ import type { TensorLike, WeightBuffer } from '../weight-buffer.js';
4
+
5
+ export interface SplitQKVRMSNormRoPEQKOptions {
6
+ numTokens: number;
7
+ numHeads: number;
8
+ numKVHeads: number;
9
+ headDim: number;
10
+ qSize: number;
11
+ kSize: number;
12
+ vSize: number;
13
+ startPos?: number;
14
+ rotaryDim?: number;
15
+ pairSpanDim?: number;
16
+ interleaved?: boolean;
17
+ rmsNormWeightOffset?: boolean;
18
+ reusesSharedKV?: boolean;
19
+ skipKNorm?: boolean;
20
+ allowUnitQKNorm?: boolean;
21
+ f16KVCacheWrite?: {
22
+ keysBuffer: GPUBuffer;
23
+ valuesBuffer: GPUBuffer;
24
+ dstOffset: number;
25
+ } | null;
26
+ }
27
+
28
+ export interface SplitQKVRMSNormRoPEQKResult {
29
+ Q: Tensor;
30
+ K: Tensor | null;
31
+ V: Tensor | null;
32
+ wroteF16KVCache?: boolean;
33
+ }
34
+
35
+ export function canUseSplitQKVRMSNormRoPEQK(
36
+ qkvTensor: Tensor | null | undefined,
37
+ options?: Partial<SplitQKVRMSNormRoPEQKOptions>
38
+ ): boolean;
39
+
40
+ export function runSplitQKVRMSNormRoPEQK(
41
+ qkvTensor: Tensor,
42
+ qWeight: GPUBuffer | WeightBuffer | TensorLike,
43
+ kWeight: GPUBuffer | WeightBuffer | TensorLike,
44
+ freqsCos: GPUBuffer | Tensor,
45
+ freqsSin: GPUBuffer | Tensor,
46
+ eps: number,
47
+ options: SplitQKVRMSNormRoPEQKOptions
48
+ ): Promise<SplitQKVRMSNormRoPEQKResult>;
49
+
50
+ export function recordSplitQKVRMSNormRoPEQK(
51
+ recorder: CommandRecorder,
52
+ qkvTensor: Tensor,
53
+ qWeight: GPUBuffer | WeightBuffer | TensorLike,
54
+ kWeight: GPUBuffer | WeightBuffer | TensorLike,
55
+ freqsCos: GPUBuffer | Tensor,
56
+ freqsSin: GPUBuffer | Tensor,
57
+ eps: number,
58
+ options: SplitQKVRMSNormRoPEQKOptions
59
+ ): Promise<SplitQKVRMSNormRoPEQKResult>;
@@ -0,0 +1,245 @@
1
+ import { acquireBuffer, releaseBuffer } from '../../memory/buffer-pool.js';
2
+ import { padToQ4KBlock } from '../../config/schema/index.js';
3
+ import { createTensor, dtypeBytes } from '../tensor.js';
4
+ import { getBuffer } from '../weight-buffer.js';
5
+ import { unifiedKernelWrapper } from './utils.js';
6
+ import {
7
+ assertRMSNormWeightBuffer,
8
+ planRMSNormDispatch,
9
+ resolveNormWeightDtype,
10
+ } from './rmsnorm.js';
11
+ import { WORKGROUP_SIZES } from './constants.js';
12
+
13
+ function assertPositiveInteger(value, label) {
14
+ if (!Number.isInteger(value) || value <= 0) {
15
+ throw new Error(`split_qkv_rmsnorm_rope_qk requires ${label} to be a positive integer.`);
16
+ }
17
+ }
18
+
19
+ function assertNonNegativeInteger(value, label) {
20
+ if (!Number.isInteger(value) || value < 0) {
21
+ throw new Error(`split_qkv_rmsnorm_rope_qk requires ${label} to be a non-negative integer.`);
22
+ }
23
+ }
24
+
25
+ function assertQKVSizes(options) {
26
+ const {
27
+ qSize,
28
+ kSize,
29
+ vSize,
30
+ numHeads,
31
+ numKVHeads,
32
+ headDim,
33
+ } = options;
34
+ if (qSize !== numHeads * headDim) {
35
+ throw new Error(`split_qkv_rmsnorm_rope_qk qSize mismatch: got ${qSize}, expected ${numHeads * headDim}.`);
36
+ }
37
+ if (kSize !== numKVHeads * headDim) {
38
+ throw new Error(`split_qkv_rmsnorm_rope_qk kSize mismatch: got ${kSize}, expected ${numKVHeads * headDim}.`);
39
+ }
40
+ if (vSize !== numKVHeads * headDim) {
41
+ throw new Error(`split_qkv_rmsnorm_rope_qk vSize mismatch: got ${vSize}, expected ${numKVHeads * headDim}.`);
42
+ }
43
+ }
44
+
45
+ function bufferOf(bufferOrTensor) {
46
+ return bufferOrTensor?.buffer || bufferOrTensor;
47
+ }
48
+
49
+ function assertBufferCapacity(buffer, requiredBytes, label) {
50
+ if (Number.isFinite(buffer?.size) && requiredBytes > buffer.size) {
51
+ throw new Error(
52
+ `split_qkv_rmsnorm_rope_qk ${label} buffer is smaller than requested range ` +
53
+ `(${requiredBytes} > ${buffer.size} bytes).`
54
+ );
55
+ }
56
+ }
57
+
58
+ function supportsFullHeadNonInterleavedRoPE(options) {
59
+ const {
60
+ headDim,
61
+ rotaryDim = headDim,
62
+ pairSpanDim = rotaryDim,
63
+ interleaved = false,
64
+ } = options;
65
+ return Number.isInteger(headDim)
66
+ && headDim > 0
67
+ && (headDim % 2) === 0
68
+ && rotaryDim === headDim
69
+ && pairSpanDim === headDim
70
+ && interleaved === false;
71
+ }
72
+
73
+ function assertSupportedRoPE(options) {
74
+ const {
75
+ headDim,
76
+ rotaryDim = headDim,
77
+ pairSpanDim = rotaryDim,
78
+ interleaved = false,
79
+ } = options;
80
+ if (!supportsFullHeadNonInterleavedRoPE({ headDim, rotaryDim, pairSpanDim, interleaved })) {
81
+ throw new Error(
82
+ 'split_qkv_rmsnorm_rope_qk supports only full-head non-interleaved RoPE ' +
83
+ `(headDim=${headDim}, rotaryDim=${rotaryDim}, pairSpanDim=${pairSpanDim}, interleaved=${interleaved}).`
84
+ );
85
+ }
86
+ }
87
+
88
+ export function canUseSplitQKVRMSNormRoPEQK(qkvTensor, options = {}) {
89
+ if (!qkvTensor || qkvTensor.dtype !== 'f32') {
90
+ return false;
91
+ }
92
+ if (options.reusesSharedKV === true || options.skipKNorm === true || options.allowUnitQKNorm === true) {
93
+ return false;
94
+ }
95
+ return supportsFullHeadNonInterleavedRoPE(options);
96
+ }
97
+
98
+ async function splitQKVRMSNormRoPEQK(target, qkvTensor, qWeight, kWeight, freqsCos, freqsSin, eps, options = {}) {
99
+ const {
100
+ numTokens,
101
+ numHeads,
102
+ numKVHeads,
103
+ headDim,
104
+ qSize,
105
+ kSize,
106
+ vSize,
107
+ startPos = 0,
108
+ rmsNormWeightOffset = false,
109
+ f16KVCacheWrite = null,
110
+ } = options;
111
+
112
+ assertPositiveInteger(numTokens, 'numTokens');
113
+ assertPositiveInteger(numHeads, 'numHeads');
114
+ assertPositiveInteger(numKVHeads, 'numKVHeads');
115
+ assertPositiveInteger(headDim, 'headDim');
116
+ assertPositiveInteger(qSize, 'qSize');
117
+ assertPositiveInteger(kSize, 'kSize');
118
+ assertPositiveInteger(vSize, 'vSize');
119
+ assertNonNegativeInteger(startPos, 'startPos');
120
+ assertQKVSizes(options);
121
+ assertSupportedRoPE(options);
122
+ if (qkvTensor.dtype !== 'f32') {
123
+ throw new Error(`split_qkv_rmsnorm_rope_qk requires f32 QKV input, got ${qkvTensor.dtype}.`);
124
+ }
125
+
126
+ const qWeightBuffer = getBuffer(qWeight);
127
+ const kWeightBuffer = getBuffer(kWeight);
128
+ const cosBuffer = bufferOf(freqsCos);
129
+ const sinBuffer = bufferOf(freqsSin);
130
+ assertRMSNormWeightBuffer(qWeight, qWeightBuffer, headDim);
131
+ assertRMSNormWeightBuffer(kWeight, kWeightBuffer, headDim);
132
+ const qWeightDtype = resolveNormWeightDtype(qWeight, headDim);
133
+ const kWeightDtype = resolveNormWeightDtype(kWeight, headDim);
134
+
135
+ const qRows = numTokens * numHeads;
136
+ const kRows = numTokens * numKVHeads;
137
+ const qkRows = qRows + kRows;
138
+ const totalV = numTokens * vSize;
139
+ const vWorkgroups = Math.ceil(totalV / WORKGROUP_SIZES.DEFAULT);
140
+ const totalWorkgroups = qkRows + vWorkgroups;
141
+ const dispatchPlan = planRMSNormDispatch(target, totalWorkgroups);
142
+ const bytesPerElement = dtypeBytes(qkvTensor.dtype);
143
+ const paddedHeadDim = padToQ4KBlock(headDim);
144
+ const qOutputSize = qRows * paddedHeadDim * bytesPerElement;
145
+ const kOutputSize = kRows * paddedHeadDim * bytesPerElement;
146
+ const vOutputSize = totalV * bytesPerElement;
147
+ const halfDim = headDim / 2;
148
+ const freqElementCount = (startPos + numTokens) * halfDim;
149
+ const writesF16KVCache = f16KVCacheWrite != null;
150
+ assertBufferCapacity(qkvTensor.buffer, numTokens * (qSize + kSize + vSize) * bytesPerElement, 'QKV input');
151
+ assertBufferCapacity(cosBuffer, freqElementCount * 4, 'cos frequencies');
152
+ assertBufferCapacity(sinBuffer, freqElementCount * 4, 'sin frequencies');
153
+ if (writesF16KVCache) {
154
+ assertNonNegativeInteger(f16KVCacheWrite.dstOffset, 'f16KVCacheWrite.dstOffset');
155
+ const requiredCacheElements = f16KVCacheWrite.dstOffset + (numTokens * numKVHeads * headDim);
156
+ assertBufferCapacity(f16KVCacheWrite.keysBuffer, requiredCacheElements * 2, 'f16 KV keys cache');
157
+ assertBufferCapacity(f16KVCacheWrite.valuesBuffer, requiredCacheElements * 2, 'f16 KV values cache');
158
+ }
159
+ let qOutputBuffer = null;
160
+ let kOutputBuffer = null;
161
+ let vOutputBuffer = null;
162
+
163
+ try {
164
+ qOutputBuffer = acquireBuffer(qOutputSize, undefined, 'split_qkv_rmsnorm_rope_qk_q_output');
165
+ if (!writesF16KVCache) {
166
+ kOutputBuffer = acquireBuffer(kOutputSize, undefined, 'split_qkv_rmsnorm_rope_qk_k_output');
167
+ vOutputBuffer = acquireBuffer(vOutputSize, undefined, 'split_qkv_rmsnorm_rope_qk_v_output');
168
+ }
169
+ const opName = writesF16KVCache
170
+ ? 'split_qkv_rmsnorm_rope_qk_f16kv_cache'
171
+ : 'split_qkv_rmsnorm_rope_qk';
172
+ const bindings = writesF16KVCache
173
+ ? [
174
+ qkvTensor,
175
+ qWeightBuffer,
176
+ kWeightBuffer,
177
+ cosBuffer,
178
+ sinBuffer,
179
+ qOutputBuffer,
180
+ f16KVCacheWrite.keysBuffer,
181
+ f16KVCacheWrite.valuesBuffer,
182
+ ]
183
+ : [qkvTensor, qWeightBuffer, kWeightBuffer, cosBuffer, sinBuffer, qOutputBuffer, kOutputBuffer, vOutputBuffer];
184
+ await unifiedKernelWrapper(
185
+ opName,
186
+ target,
187
+ 'default',
188
+ bindings,
189
+ {
190
+ num_tokens: numTokens,
191
+ q_size: qSize,
192
+ k_size: kSize,
193
+ v_size: vSize,
194
+ num_heads: numHeads,
195
+ num_kv_heads: numKVHeads,
196
+ head_dim: headDim,
197
+ workgroup_stride: dispatchPlan.tokenStride,
198
+ qk_rows: qkRows,
199
+ total_v: totalV,
200
+ start_pos: startPos,
201
+ half_dim: halfDim,
202
+ eps,
203
+ kv_dst_offset: writesF16KVCache ? f16KVCacheWrite.dstOffset : 0,
204
+ _pad1: 0,
205
+ _pad2: 0,
206
+ },
207
+ dispatchPlan.workgroups,
208
+ {
209
+ RMS_NORM_OFFSET: rmsNormWeightOffset,
210
+ Q_WEIGHT_IS_F16: qWeightDtype === 'f16',
211
+ K_WEIGHT_IS_F16: kWeightDtype === 'f16',
212
+ },
213
+ null,
214
+ writesF16KVCache
215
+ ? 'split_qkv_rmsnorm_rope_qk:f16kv_cache'
216
+ : 'split_qkv_rmsnorm_rope_qk'
217
+ );
218
+
219
+ return {
220
+ Q: createTensor(qOutputBuffer, 'f32', [qRows, headDim], 'split_qkv_rmsnorm_rope_qk_q_output'),
221
+ K: writesF16KVCache ? null : createTensor(kOutputBuffer, 'f32', [kRows, headDim], 'split_qkv_rmsnorm_rope_qk_k_output'),
222
+ V: writesF16KVCache ? null : createTensor(vOutputBuffer, 'f32', [numTokens, vSize], 'split_qkv_rmsnorm_rope_qk_v_output'),
223
+ wroteF16KVCache: writesF16KVCache,
224
+ };
225
+ } catch (error) {
226
+ if (qOutputBuffer) {
227
+ releaseBuffer(qOutputBuffer);
228
+ }
229
+ if (kOutputBuffer) {
230
+ releaseBuffer(kOutputBuffer);
231
+ }
232
+ if (vOutputBuffer) {
233
+ releaseBuffer(vOutputBuffer);
234
+ }
235
+ throw error;
236
+ }
237
+ }
238
+
239
+ export async function runSplitQKVRMSNormRoPEQK(qkvTensor, qWeight, kWeight, freqsCos, freqsSin, eps, options = {}) {
240
+ return splitQKVRMSNormRoPEQK(null, qkvTensor, qWeight, kWeight, freqsCos, freqsSin, eps, options);
241
+ }
242
+
243
+ export async function recordSplitQKVRMSNormRoPEQK(recorder, qkvTensor, qWeight, kWeight, freqsCos, freqsSin, eps, options = {}) {
244
+ return splitQKVRMSNormRoPEQK(recorder, qkvTensor, qWeight, kWeight, freqsCos, freqsSin, eps, options);
245
+ }