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 +1,118 @@
1
+ override WORKGROUP_SIZE: u32 = 256u;
2
+ const MAX_WORKGROUP_SIZE: u32 = 256u;
3
+
4
+ override RMS_NORM_OFFSET: bool = false;
5
+ override Q_WEIGHT_IS_F16: bool = false;
6
+ override K_WEIGHT_IS_F16: bool = false;
7
+
8
+ struct Uniforms {
9
+ q_rows: u32,
10
+ k_rows: u32,
11
+ head_dim: u32,
12
+ row_stride: u32,
13
+ eps: f32,
14
+ _pad0: u32,
15
+ _pad1: u32,
16
+ _pad2: u32,
17
+ }
18
+
19
+ @group(0) @binding(0) var<uniform> u: Uniforms;
20
+ @group(0) @binding(1) var<storage, read> q_input: array<f32>;
21
+ @group(0) @binding(2) var<storage, read> q_weight: array<u32>;
22
+ @group(0) @binding(3) var<storage, read_write> q_output: array<f32>;
23
+ @group(0) @binding(4) var<storage, read> k_input: array<f32>;
24
+ @group(0) @binding(5) var<storage, read> k_weight: array<u32>;
25
+ @group(0) @binding(6) var<storage, read_write> k_output: array<f32>;
26
+
27
+ var<workgroup> shared_sum: array<f32, MAX_WORKGROUP_SIZE>;
28
+
29
+ fn apply_weight(w: f32) -> f32 {
30
+ if (RMS_NORM_OFFSET) {
31
+ return 1.0 + w;
32
+ }
33
+ return w;
34
+ }
35
+
36
+ fn row_index(wg_id: vec3<u32>) -> u32 {
37
+ return wg_id.y * max(u.row_stride, 1u) + wg_id.x;
38
+ }
39
+
40
+ fn load_value(row: u32, idx: u32) -> f32 {
41
+ if (row < u.q_rows) {
42
+ return q_input[row * u.head_dim + idx];
43
+ }
44
+ let k_row = row - u.q_rows;
45
+ return k_input[k_row * u.head_dim + idx];
46
+ }
47
+
48
+ fn load_weight(row: u32, idx: u32) -> f32 {
49
+ if (row < u.q_rows) {
50
+ if (Q_WEIGHT_IS_F16) {
51
+ let packed = q_weight[idx >> 1u];
52
+ let pair = unpack2x16float(packed);
53
+ return select(pair.x, pair.y, (idx & 1u) == 1u);
54
+ }
55
+ return bitcast<f32>(q_weight[idx]);
56
+ }
57
+ if (K_WEIGHT_IS_F16) {
58
+ let packed = k_weight[idx >> 1u];
59
+ let pair = unpack2x16float(packed);
60
+ return select(pair.x, pair.y, (idx & 1u) == 1u);
61
+ }
62
+ return bitcast<f32>(k_weight[idx]);
63
+ }
64
+
65
+ fn store_value(row: u32, idx: u32, value: f32) {
66
+ if (row < u.q_rows) {
67
+ q_output[row * u.head_dim + idx] = value;
68
+ } else {
69
+ let k_row = row - u.q_rows;
70
+ k_output[k_row * u.head_dim + idx] = value;
71
+ }
72
+ }
73
+
74
+ @compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
75
+ fn main(
76
+ @builtin(local_invocation_id) local_id: vec3<u32>,
77
+ @builtin(workgroup_id) wg_id: vec3<u32>
78
+ ) {
79
+ let row = row_index(wg_id);
80
+ let total_rows = u.q_rows + u.k_rows;
81
+ if (row >= total_rows) {
82
+ return;
83
+ }
84
+
85
+ let thread_idx = local_id.x;
86
+ let size = u.head_dim;
87
+ let elements_per_thread = (size + WORKGROUP_SIZE - 1u) / WORKGROUP_SIZE;
88
+
89
+ var local_sum_sq: f32 = 0.0;
90
+ for (var i: u32 = 0u; i < elements_per_thread; i = i + 1u) {
91
+ let idx = thread_idx * elements_per_thread + i;
92
+ if (idx < size) {
93
+ let x = load_value(row, idx);
94
+ local_sum_sq = local_sum_sq + x * x;
95
+ }
96
+ }
97
+
98
+ shared_sum[thread_idx] = local_sum_sq;
99
+ workgroupBarrier();
100
+
101
+ for (var stride: u32 = WORKGROUP_SIZE / 2u; stride > 0u; stride = stride >> 1u) {
102
+ if (thread_idx < stride) {
103
+ shared_sum[thread_idx] = shared_sum[thread_idx] + shared_sum[thread_idx + stride];
104
+ }
105
+ workgroupBarrier();
106
+ }
107
+
108
+ let inv_rms = 1.0 / sqrt(shared_sum[0] / f32(size) + u.eps);
109
+ workgroupBarrier();
110
+
111
+ for (var i: u32 = 0u; i < elements_per_thread; i = i + 1u) {
112
+ let idx = thread_idx * elements_per_thread + i;
113
+ if (idx < size) {
114
+ let result = load_value(row, idx) * inv_rms * apply_weight(load_weight(row, idx));
115
+ store_value(row, idx, result);
116
+ }
117
+ }
118
+ }
@@ -0,0 +1,143 @@
1
+ enable f16;
2
+
3
+ override WORKGROUP_SIZE: u32 = 256u;
4
+ const MAX_WORKGROUP_SIZE: u32 = 256u;
5
+
6
+ override RMS_NORM_OFFSET: bool = false;
7
+ override Q_WEIGHT_IS_F16: bool = false;
8
+ override K_WEIGHT_IS_F16: bool = false;
9
+
10
+ struct Uniforms {
11
+ q_rows: u32,
12
+ k_rows: u32,
13
+ head_dim: u32,
14
+ row_stride: u32,
15
+ eps: f32,
16
+ _pad0: u32,
17
+ _pad1: u32,
18
+ _pad2: u32,
19
+ }
20
+
21
+ @group(0) @binding(0) var<uniform> u: Uniforms;
22
+ @group(0) @binding(1) var<storage, read> q_input: array<f16>;
23
+ @group(0) @binding(2) var<storage, read> q_weight: array<u32>;
24
+ @group(0) @binding(3) var<storage, read_write> q_output: array<f16>;
25
+ @group(0) @binding(4) var<storage, read> k_input: array<f16>;
26
+ @group(0) @binding(5) var<storage, read> k_weight: array<u32>;
27
+ @group(0) @binding(6) var<storage, read_write> k_output: array<f16>;
28
+
29
+ var<workgroup> shared_sum: array<f16, MAX_WORKGROUP_SIZE>;
30
+
31
+ fn apply_weight(w: f16) -> f16 {
32
+ if (RMS_NORM_OFFSET) {
33
+ return f16(1.0) + w;
34
+ }
35
+ return w;
36
+ }
37
+
38
+ fn row_index(wg_id: vec3<u32>) -> u32 {
39
+ return wg_id.y * max(u.row_stride, 1u) + wg_id.x;
40
+ }
41
+
42
+ fn load_value(row: u32, idx: u32) -> f16 {
43
+ if (row < u.q_rows) {
44
+ return q_input[row * u.head_dim + idx];
45
+ }
46
+ let k_row = row - u.q_rows;
47
+ return k_input[k_row * u.head_dim + idx];
48
+ }
49
+
50
+ fn load_weight(row: u32, idx: u32) -> f16 {
51
+ if (row < u.q_rows) {
52
+ if (Q_WEIGHT_IS_F16) {
53
+ let packed = q_weight[idx >> 1u];
54
+ let pair = unpack2x16float(packed);
55
+ return f16(select(pair.x, pair.y, (idx & 1u) == 1u));
56
+ }
57
+ return f16(bitcast<f32>(q_weight[idx]));
58
+ }
59
+ if (K_WEIGHT_IS_F16) {
60
+ let packed = k_weight[idx >> 1u];
61
+ let pair = unpack2x16float(packed);
62
+ return f16(select(pair.x, pair.y, (idx & 1u) == 1u));
63
+ }
64
+ return f16(bitcast<f32>(k_weight[idx]));
65
+ }
66
+
67
+ fn store_value(row: u32, idx: u32, value: f16) {
68
+ if (row < u.q_rows) {
69
+ q_output[row * u.head_dim + idx] = value;
70
+ } else {
71
+ let k_row = row - u.q_rows;
72
+ k_output[k_row * u.head_dim + idx] = value;
73
+ }
74
+ }
75
+
76
+ @compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
77
+ fn main(
78
+ @builtin(local_invocation_id) local_id: vec3<u32>,
79
+ @builtin(workgroup_id) wg_id: vec3<u32>
80
+ ) {
81
+ let row = row_index(wg_id);
82
+ let total_rows = u.q_rows + u.k_rows;
83
+ if (row >= total_rows) {
84
+ return;
85
+ }
86
+
87
+ let thread_idx = local_id.x;
88
+ let size = u.head_dim;
89
+ let elements_per_thread = (size + WORKGROUP_SIZE - 1u) / WORKGROUP_SIZE;
90
+
91
+ var local_max: f16 = f16(0.0);
92
+ for (var i: u32 = 0u; i < elements_per_thread; i = i + 1u) {
93
+ let idx = thread_idx * elements_per_thread + i;
94
+ if (idx < size) {
95
+ local_max = max(local_max, abs(load_value(row, idx)));
96
+ }
97
+ }
98
+
99
+ shared_sum[thread_idx] = local_max;
100
+ workgroupBarrier();
101
+
102
+ for (var stride: u32 = WORKGROUP_SIZE / 2u; stride > 0u; stride = stride >> 1u) {
103
+ if (thread_idx < stride) {
104
+ shared_sum[thread_idx] = max(shared_sum[thread_idx], shared_sum[thread_idx + stride]);
105
+ }
106
+ workgroupBarrier();
107
+ }
108
+
109
+ let row_scale = select(f16(1.0), shared_sum[0], shared_sum[0] > f16(0.0));
110
+ workgroupBarrier();
111
+
112
+ var local_sum_sq: f16 = f16(0.0);
113
+ let inv_size = f16(1.0) / f16(size);
114
+ for (var i: u32 = 0u; i < elements_per_thread; i = i + 1u) {
115
+ let idx = thread_idx * elements_per_thread + i;
116
+ if (idx < size) {
117
+ let x = load_value(row, idx) / row_scale;
118
+ local_sum_sq = local_sum_sq + (x * x) * inv_size;
119
+ }
120
+ }
121
+
122
+ shared_sum[thread_idx] = local_sum_sq;
123
+ workgroupBarrier();
124
+
125
+ for (var stride: u32 = WORKGROUP_SIZE / 2u; stride > 0u; stride = stride >> 1u) {
126
+ if (thread_idx < stride) {
127
+ shared_sum[thread_idx] = shared_sum[thread_idx] + shared_sum[thread_idx + stride];
128
+ }
129
+ workgroupBarrier();
130
+ }
131
+
132
+ let eps_scaled = f16(u.eps) / (row_scale * row_scale);
133
+ let inv_rms = (f16(1.0) / row_scale) / sqrt(shared_sum[0] + eps_scaled);
134
+ workgroupBarrier();
135
+
136
+ for (var i: u32 = 0u; i < elements_per_thread; i = i + 1u) {
137
+ let idx = thread_idx * elements_per_thread + i;
138
+ if (idx < size) {
139
+ let result = load_value(row, idx) * inv_rms * apply_weight(load_weight(row, idx));
140
+ store_value(row, idx, result);
141
+ }
142
+ }
143
+ }
@@ -0,0 +1,41 @@
1
+ import type { Tensor } from '../tensor.js';
2
+ import type { CommandRecorder } from '../command-recorder.js';
3
+ import type { RoPEOptions } from './rope.js';
4
+
5
+ export interface RoPEQKOptions extends Omit<RoPEOptions, 'numHeads'> {
6
+ numQHeads: number;
7
+ numKVHeads: number;
8
+ headDim: number;
9
+ rotaryDim?: number;
10
+ pairSpanDim?: number;
11
+ }
12
+
13
+ export interface RoPEQKResult {
14
+ q: Tensor;
15
+ k: Tensor;
16
+ }
17
+
18
+ export function canUseRoPEQK(
19
+ qTensor: Tensor | null | undefined,
20
+ kTensor: Tensor | null | undefined,
21
+ options?: { reusesSharedKV?: boolean }
22
+ ): boolean;
23
+
24
+ export declare function runRoPEQK(
25
+ qTensor: Tensor,
26
+ kTensor: Tensor,
27
+ freqsCos: GPUBuffer | Tensor,
28
+ freqsSin: GPUBuffer | Tensor,
29
+ seqLen: number,
30
+ options: RoPEQKOptions
31
+ ): Promise<RoPEQKResult>;
32
+
33
+ export declare function recordRoPEQK(
34
+ recorder: CommandRecorder,
35
+ qTensor: Tensor,
36
+ kTensor: Tensor,
37
+ freqsCos: GPUBuffer | Tensor,
38
+ freqsSin: GPUBuffer | Tensor,
39
+ seqLen: number,
40
+ options: RoPEQKOptions
41
+ ): Promise<RoPEQKResult>;
@@ -0,0 +1,143 @@
1
+ import { createTensor, dtypeBytes } from '../tensor.js';
2
+ import { WORKGROUP_SIZES } from './constants.js';
3
+ import { unifiedKernelWrapper } from './utils.js';
4
+ import { selectRuleValue } from './rule-registry.js';
5
+
6
+ function assertPositiveInteger(value, label) {
7
+ if (!Number.isInteger(value) || value <= 0) {
8
+ throw new Error(`RoPE QK requires ${label} to be a positive integer.`);
9
+ }
10
+ }
11
+
12
+ function assertNonNegativeInteger(value, label) {
13
+ if (!Number.isInteger(value) || value < 0) {
14
+ throw new Error(`RoPE QK requires ${label} to be a non-negative integer.`);
15
+ }
16
+ }
17
+
18
+ function assertRopeGeometry(options) {
19
+ const { headDim, rotaryDim, pairSpanDim } = options;
20
+ assertPositiveInteger(headDim, 'headDim');
21
+ if ((headDim % 2) !== 0) {
22
+ throw new Error(`RoPE QK headDim must be even, got ${headDim}.`);
23
+ }
24
+ assertPositiveInteger(rotaryDim, 'rotaryDim');
25
+ if ((rotaryDim % 2) !== 0) {
26
+ throw new Error(`RoPE QK rotaryDim must be even, got ${rotaryDim}.`);
27
+ }
28
+ if (rotaryDim > headDim) {
29
+ throw new Error(`RoPE QK rotaryDim must be <= headDim; got ${rotaryDim} for headDim ${headDim}.`);
30
+ }
31
+ assertPositiveInteger(pairSpanDim, 'pairSpanDim');
32
+ if ((pairSpanDim % 2) !== 0) {
33
+ throw new Error(`RoPE QK pairSpanDim must be even, got ${pairSpanDim}.`);
34
+ }
35
+ if (pairSpanDim < rotaryDim || pairSpanDim > headDim) {
36
+ throw new Error(
37
+ `RoPE QK pairSpanDim must be in [rotaryDim, headDim]; got ${pairSpanDim} ` +
38
+ `for rotaryDim ${rotaryDim} and headDim ${headDim}.`
39
+ );
40
+ }
41
+ }
42
+
43
+ function bufferOf(bufferOrTensor) {
44
+ return bufferOrTensor?.buffer || bufferOrTensor;
45
+ }
46
+
47
+ function assertBufferCapacity(buffer, requiredBytes, label) {
48
+ if (Number.isFinite(buffer?.size) && requiredBytes > buffer.size) {
49
+ throw new Error(`RoPE QK ${label} buffer is smaller than requested range (${requiredBytes} > ${buffer.size} bytes).`);
50
+ }
51
+ }
52
+
53
+ function selectRoPEQKVariant(dtype) {
54
+ return selectRuleValue('ropeQk', 'variant', { useF16: dtype === 'f16' });
55
+ }
56
+
57
+ export function canUseRoPEQK(qTensor, kTensor, options = {}) {
58
+ if (options.reusesSharedKV === true) {
59
+ return false;
60
+ }
61
+ if (!qTensor || !kTensor || qTensor.buffer === kTensor.buffer) {
62
+ return false;
63
+ }
64
+ return qTensor.dtype === kTensor.dtype && (qTensor.dtype === 'f32' || qTensor.dtype === 'f16');
65
+ }
66
+
67
+ async function _ropeQK(target, qTensor, kTensor, freqsCos, freqsSin, seqLen, options = {}) {
68
+ const {
69
+ numQHeads,
70
+ numKVHeads,
71
+ headDim,
72
+ rotaryDim = headDim,
73
+ pairSpanDim = rotaryDim,
74
+ interleaved = false,
75
+ startPos = 0,
76
+ } = options;
77
+
78
+ assertNonNegativeInteger(seqLen, 'seqLen');
79
+ assertNonNegativeInteger(startPos, 'startPos');
80
+ assertPositiveInteger(numQHeads, 'numQHeads');
81
+ assertPositiveInteger(numKVHeads, 'numKVHeads');
82
+ assertRopeGeometry({ headDim, rotaryDim, pairSpanDim });
83
+ if (qTensor.dtype !== kTensor.dtype) {
84
+ throw new Error(`RoPE QK requires matching Q/K dtypes, got ${qTensor.dtype} and ${kTensor.dtype}.`);
85
+ }
86
+ if (qTensor.dtype !== 'f32' && qTensor.dtype !== 'f16') {
87
+ throw new Error(`RoPE QK requires f32 or f16 tensors, got ${qTensor.dtype}.`);
88
+ }
89
+ if (seqLen === 0) {
90
+ return {
91
+ q: createTensor(qTensor.buffer, qTensor.dtype, [...qTensor.shape], 'rope_q_output'),
92
+ k: createTensor(kTensor.buffer, kTensor.dtype, [...kTensor.shape], 'rope_k_output'),
93
+ };
94
+ }
95
+
96
+ const bytesPerElement = dtypeBytes(qTensor.dtype);
97
+ const qElementCount = seqLen * numQHeads * headDim;
98
+ const kElementCount = seqLen * numKVHeads * headDim;
99
+ const halfDim = rotaryDim / 2;
100
+ const freqElementCount = (startPos + seqLen) * halfDim;
101
+ assertBufferCapacity(qTensor.buffer, qElementCount * bytesPerElement, 'Q input');
102
+ assertBufferCapacity(kTensor.buffer, kElementCount * bytesPerElement, 'K input');
103
+ assertBufferCapacity(bufferOf(freqsCos), freqElementCount * 4, 'cos frequencies');
104
+ assertBufferCapacity(bufferOf(freqsSin), freqElementCount * 4, 'sin frequencies');
105
+
106
+ const variant = selectRoPEQKVariant(qTensor.dtype);
107
+ const totalPairs = seqLen * (numQHeads + numKVHeads) * halfDim;
108
+ const workgroups = Math.ceil(totalPairs / WORKGROUP_SIZES.DEFAULT);
109
+
110
+ await unifiedKernelWrapper(
111
+ 'rope_qk',
112
+ target,
113
+ variant,
114
+ [qTensor, kTensor, freqsCos, freqsSin],
115
+ {
116
+ seq_len: seqLen,
117
+ num_q_heads: numQHeads,
118
+ num_kv_heads: numKVHeads,
119
+ head_dim: headDim,
120
+ start_pos: startPos,
121
+ rotary_dim: rotaryDim,
122
+ interleaved: interleaved ? 1 : 0,
123
+ pair_span_dim: pairSpanDim,
124
+ },
125
+ workgroups,
126
+ null,
127
+ null,
128
+ 'rope_qk'
129
+ );
130
+
131
+ return {
132
+ q: createTensor(qTensor.buffer, qTensor.dtype, [...qTensor.shape], 'rope_q_output'),
133
+ k: createTensor(kTensor.buffer, kTensor.dtype, [...kTensor.shape], 'rope_k_output'),
134
+ };
135
+ }
136
+
137
+ export async function runRoPEQK(qTensor, kTensor, freqsCos, freqsSin, seqLen, options = {}) {
138
+ return _ropeQK(null, qTensor, kTensor, freqsCos, freqsSin, seqLen, options);
139
+ }
140
+
141
+ export async function recordRoPEQK(recorder, qTensor, kTensor, freqsCos, freqsSin, seqLen, options = {}) {
142
+ return _ropeQK(recorder, qTensor, kTensor, freqsCos, freqsSin, seqLen, options);
143
+ }
@@ -0,0 +1,95 @@
1
+ // Rotary Position Embeddings for separate Q/K buffers.
2
+
3
+ override WORKGROUP_SIZE: u32 = 256u;
4
+
5
+ struct Uniforms {
6
+ seq_len: u32,
7
+ num_q_heads: u32,
8
+ num_kv_heads: u32,
9
+ head_dim: u32,
10
+ start_pos: u32,
11
+ rotary_dim: u32,
12
+ interleaved: u32,
13
+ pair_span_dim: u32,
14
+ }
15
+
16
+ @group(0) @binding(0) var<uniform> u: Uniforms;
17
+ @group(0) @binding(1) var<storage, read_write> q_input: array<f32>;
18
+ @group(0) @binding(2) var<storage, read_write> k_input: array<f32>;
19
+ @group(0) @binding(3) var<storage, read> freqs_cos: array<f32>;
20
+ @group(0) @binding(4) var<storage, read> freqs_sin: array<f32>;
21
+
22
+ fn get_first_rotary_idx(pair_idx: u32) -> u32 {
23
+ if (u.interleaved == 1u) {
24
+ return pair_idx * 2u;
25
+ }
26
+ return pair_idx;
27
+ }
28
+
29
+ fn get_second_rotary_idx(pair_idx: u32, pair_span_dim: u32) -> u32 {
30
+ if (u.interleaved == 1u) {
31
+ return pair_idx * 2u + 1u;
32
+ }
33
+ return pair_idx + (pair_span_dim / 2u);
34
+ }
35
+
36
+ fn rotate_pair(x0: f32, x1: f32, cos_val: f32, sin_val: f32) -> vec2<f32> {
37
+ return vec2<f32>(
38
+ x0 * cos_val - x1 * sin_val,
39
+ x0 * sin_val + x1 * cos_val
40
+ );
41
+ }
42
+
43
+ @compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
44
+ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {
45
+ let half_dim = u.rotary_dim / 2u;
46
+ let q_pairs = u.seq_len * u.num_q_heads * half_dim;
47
+ let k_pairs = u.seq_len * u.num_kv_heads * half_dim;
48
+ let total_pairs = q_pairs + k_pairs;
49
+ let idx = global_id.x;
50
+
51
+ if (idx >= total_pairs) {
52
+ return;
53
+ }
54
+
55
+ if (idx < q_pairs) {
56
+ let pos = idx / (u.num_q_heads * half_dim);
57
+ let remainder = idx % (u.num_q_heads * half_dim);
58
+ let head_idx = remainder / half_dim;
59
+ let pair_idx = remainder % half_dim;
60
+ let freq_idx = (u.start_pos + pos) * half_dim + pair_idx;
61
+ let cos_val = freqs_cos[freq_idx];
62
+ let sin_val = freqs_sin[freq_idx];
63
+ let base_idx = pos * u.num_q_heads * u.head_dim + head_idx * u.head_dim;
64
+ let first_idx = get_first_rotary_idx(pair_idx);
65
+ let second_idx = get_second_rotary_idx(pair_idx, u.pair_span_dim);
66
+ let rotated = rotate_pair(
67
+ q_input[base_idx + first_idx],
68
+ q_input[base_idx + second_idx],
69
+ cos_val,
70
+ sin_val
71
+ );
72
+ q_input[base_idx + first_idx] = rotated.x;
73
+ q_input[base_idx + second_idx] = rotated.y;
74
+ } else {
75
+ let k_idx = idx - q_pairs;
76
+ let pos = k_idx / (u.num_kv_heads * half_dim);
77
+ let remainder = k_idx % (u.num_kv_heads * half_dim);
78
+ let head_idx = remainder / half_dim;
79
+ let pair_idx = remainder % half_dim;
80
+ let freq_idx = (u.start_pos + pos) * half_dim + pair_idx;
81
+ let cos_val = freqs_cos[freq_idx];
82
+ let sin_val = freqs_sin[freq_idx];
83
+ let base_idx = pos * u.num_kv_heads * u.head_dim + head_idx * u.head_dim;
84
+ let first_idx = get_first_rotary_idx(pair_idx);
85
+ let second_idx = get_second_rotary_idx(pair_idx, u.pair_span_dim);
86
+ let rotated = rotate_pair(
87
+ k_input[base_idx + first_idx],
88
+ k_input[base_idx + second_idx],
89
+ cos_val,
90
+ sin_val
91
+ );
92
+ k_input[base_idx + first_idx] = rotated.x;
93
+ k_input[base_idx + second_idx] = rotated.y;
94
+ }
95
+ }
@@ -0,0 +1,97 @@
1
+ // Rotary Position Embeddings for separate f16 Q/K buffers.
2
+
3
+ enable f16;
4
+
5
+ override WORKGROUP_SIZE: u32 = 256u;
6
+
7
+ struct Uniforms {
8
+ seq_len: u32,
9
+ num_q_heads: u32,
10
+ num_kv_heads: u32,
11
+ head_dim: u32,
12
+ start_pos: u32,
13
+ rotary_dim: u32,
14
+ interleaved: u32,
15
+ pair_span_dim: u32,
16
+ }
17
+
18
+ @group(0) @binding(0) var<uniform> u: Uniforms;
19
+ @group(0) @binding(1) var<storage, read_write> q_input: array<f16>;
20
+ @group(0) @binding(2) var<storage, read_write> k_input: array<f16>;
21
+ @group(0) @binding(3) var<storage, read> freqs_cos: array<f32>;
22
+ @group(0) @binding(4) var<storage, read> freqs_sin: array<f32>;
23
+
24
+ fn get_first_rotary_idx(pair_idx: u32) -> u32 {
25
+ if (u.interleaved == 1u) {
26
+ return pair_idx * 2u;
27
+ }
28
+ return pair_idx;
29
+ }
30
+
31
+ fn get_second_rotary_idx(pair_idx: u32, pair_span_dim: u32) -> u32 {
32
+ if (u.interleaved == 1u) {
33
+ return pair_idx * 2u + 1u;
34
+ }
35
+ return pair_idx + (pair_span_dim / 2u);
36
+ }
37
+
38
+ fn rotate_pair(x0: f16, x1: f16, cos_val: f16, sin_val: f16) -> vec2<f16> {
39
+ return vec2<f16>(
40
+ x0 * cos_val - x1 * sin_val,
41
+ x0 * sin_val + x1 * cos_val
42
+ );
43
+ }
44
+
45
+ @compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
46
+ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {
47
+ let half_dim = u.rotary_dim / 2u;
48
+ let q_pairs = u.seq_len * u.num_q_heads * half_dim;
49
+ let k_pairs = u.seq_len * u.num_kv_heads * half_dim;
50
+ let total_pairs = q_pairs + k_pairs;
51
+ let idx = global_id.x;
52
+
53
+ if (idx >= total_pairs) {
54
+ return;
55
+ }
56
+
57
+ if (idx < q_pairs) {
58
+ let pos = idx / (u.num_q_heads * half_dim);
59
+ let remainder = idx % (u.num_q_heads * half_dim);
60
+ let head_idx = remainder / half_dim;
61
+ let pair_idx = remainder % half_dim;
62
+ let freq_idx = (u.start_pos + pos) * half_dim + pair_idx;
63
+ let cos_val = f16(freqs_cos[freq_idx]);
64
+ let sin_val = f16(freqs_sin[freq_idx]);
65
+ let base_idx = pos * u.num_q_heads * u.head_dim + head_idx * u.head_dim;
66
+ let first_idx = get_first_rotary_idx(pair_idx);
67
+ let second_idx = get_second_rotary_idx(pair_idx, u.pair_span_dim);
68
+ let rotated = rotate_pair(
69
+ q_input[base_idx + first_idx],
70
+ q_input[base_idx + second_idx],
71
+ cos_val,
72
+ sin_val
73
+ );
74
+ q_input[base_idx + first_idx] = rotated.x;
75
+ q_input[base_idx + second_idx] = rotated.y;
76
+ } else {
77
+ let k_idx = idx - q_pairs;
78
+ let pos = k_idx / (u.num_kv_heads * half_dim);
79
+ let remainder = k_idx % (u.num_kv_heads * half_dim);
80
+ let head_idx = remainder / half_dim;
81
+ let pair_idx = remainder % half_dim;
82
+ let freq_idx = (u.start_pos + pos) * half_dim + pair_idx;
83
+ let cos_val = f16(freqs_cos[freq_idx]);
84
+ let sin_val = f16(freqs_sin[freq_idx]);
85
+ let base_idx = pos * u.num_kv_heads * u.head_dim + head_idx * u.head_dim;
86
+ let first_idx = get_first_rotary_idx(pair_idx);
87
+ let second_idx = get_second_rotary_idx(pair_idx, u.pair_span_dim);
88
+ let rotated = rotate_pair(
89
+ k_input[base_idx + first_idx],
90
+ k_input[base_idx + second_idx],
91
+ cos_val,
92
+ sin_val
93
+ );
94
+ k_input[base_idx + first_idx] = rotated.x;
95
+ k_input[base_idx + second_idx] = rotated.y;
96
+ }
97
+ }