doppler-gpu 0.4.6 → 0.4.8

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 (705) hide show
  1. package/CHANGELOG.md +64 -0
  2. package/README.md +98 -137
  3. package/assets/doppler.svg +1 -3
  4. package/models/catalog.json +1575 -0
  5. package/package.json +59 -26
  6. package/src/cli/doppler-cli.js +28 -7
  7. package/src/cli/doppler-quickstart.js +1 -1
  8. package/src/cli/doppler-serve.js +2 -2
  9. package/src/client/doppler-registry.js +12 -0
  10. package/src/client/doppler-registry.json +52 -3
  11. package/src/client/runtime/index.js +25 -2
  12. package/src/client/runtime/model-session.d.ts +31 -1
  13. package/src/client/runtime/model-session.js +31 -0
  14. package/src/client/runtime/model-source.d.ts +19 -2
  15. package/src/client/runtime/model-source.js +45 -5
  16. package/src/config/conversion/diffusiongemma/diffusiongemma-26b-a4b-it-q4k-ehf16-af16.json +28 -4
  17. package/src/config/conversion/embeddinggemma/google-embeddinggemma-300m-q4k-ehf16-af32.json +29 -4
  18. package/src/config/conversion/gemma3/gemma-3-1b-it-f16-af32.json +29 -4
  19. package/src/config/conversion/gemma3/gemma-3-1b-it-q4k-ehf16-af32.json +29 -4
  20. package/src/config/conversion/gemma3/gemma-3-270m-it-f16-af32.json +29 -4
  21. package/src/config/conversion/gemma3/gemma-3-270m-it-q4k-ehf16-af16.json +29 -4
  22. package/src/config/conversion/gemma3/gemma-3-270m-it-q4k-ehf16-af32.json +29 -4
  23. package/src/config/conversion/gemma3/translategemma-4b-1b-enes-q4k-ehf16-af32.json +50 -9
  24. package/src/config/conversion/gemma3/translategemma-4b-it-q4k-ehf16-af32.json +35 -5
  25. package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-af16.json +31 -5
  26. package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-af32.json +30 -4
  27. package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-hq4k-af16.json +30 -4
  28. package/src/config/conversion/gemma4/gemma-4-12b-it-text-w4a16-ct-ehf16-af16.json +30 -4
  29. package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af16.json +30 -5
  30. package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af32.json +29 -4
  31. package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af16-int4ple.json +30 -5
  32. package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32-int4ple.json +29 -4
  33. package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32.json +29 -4
  34. package/src/config/conversion/gemma4/gemma-4-moe-q4k-ehf16-af32.json +29 -4
  35. package/src/config/conversion/gpt-oss-20b-f16-xmxfp4.json +29 -4
  36. package/src/config/conversion/janus/janus-pro-1b-text-q4k-ehaf16.json +29 -4
  37. package/src/config/conversion/lfm2/lfm2.5-1.2b-instruct-q4k-ehf16-af32.json +29 -4
  38. package/src/config/conversion/minicpm/minicpm4-0-5b-f16-af32.json +481 -0
  39. package/src/config/conversion/qwen3/qwen-3-5-0-8b-q4k-ehaf16.json +29 -4
  40. package/src/config/conversion/qwen3/qwen-3-5-2b-q4k-ehaf16.json +29 -4
  41. package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-eaf16.json +30 -5
  42. package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-ehaf16.json +29 -4
  43. package/src/config/conversion/qwen3/qwen-3-embedding-0-6b-q4k-ehf16-af32.json +395 -0
  44. package/src/config/conversion/qwen3/qwen-3-reranker-0-6b-f16-af32.json +429 -0
  45. package/src/config/conversion/qwen3/qwen-3-reranker-0-6b-q4k-ehf16-af32.json +548 -0
  46. package/src/config/execution-contract-check.js +14 -10
  47. package/src/config/kernel-path-loader.js +1 -0
  48. package/src/config/kernels/kernel-ref-digests.js +26 -17
  49. package/src/config/kernels/registry.json +448 -27
  50. package/src/config/merge.js +41 -0
  51. package/src/config/param-validator.js +67 -8
  52. package/src/config/required-inference-fields-contract-check.js +14 -0
  53. package/src/config/runtime/default.json +3 -1
  54. package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-attn128-candidate.json +20 -0
  55. package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-maxbatch64-candidate.json +52 -0
  56. package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-normed-ffn-candidate.json +18 -0
  57. package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-overlap-ringtokens1-candidate.json +52 -0
  58. package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-parity.json +85 -0
  59. package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-sequential-candidate.json +52 -0
  60. package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-throughput.json +85 -0
  61. package/src/config/runtime/profiles/qwen-3-5-2b-f16-primary-throughput.json +23 -0
  62. package/src/config/runtime/profiles/qwen-3-5-2b-metal-parity.json +38 -0
  63. package/src/config/runtime/profiles/qwen-3-5-2b-metal-throughput.json +135 -0
  64. package/src/config/runtime/profiles/qwen-3-embedding-0-6b-metal-stability.json +37 -0
  65. package/src/config/runtime/profiles/qwen-3-embedding-0-6b-metal-throughput.json +21 -0
  66. package/src/config/runtime/profiles/qwen-3-reranker-0-6b-metal-stability.json +41 -0
  67. package/src/config/runtime/profiles/qwen-3-reranker-0-6b-metal-throughput.json +41 -0
  68. package/src/config/runtime/profiles/rerank-stability.json +43 -0
  69. package/src/config/runtime/profiles/rerank-throughput.json +40 -0
  70. package/src/config/runtime/profiles/translategemma-1b-nativekd2-q4k-rdna3-throughput-probe.json +22 -0
  71. package/src/config/schema/debug.schema.d.ts +2 -0
  72. package/src/config/schema/debug.schema.js +1 -0
  73. package/src/config/schema/doppler.schema.d.ts +5 -1
  74. package/src/config/schema/doppler.schema.js +117 -292
  75. package/src/config/schema/execution-v1.schema.d.ts +41 -0
  76. package/src/config/schema/execution-v1.schema.js +10 -0
  77. package/src/config/schema/index.d.ts +3 -1
  78. package/src/config/schema/inference.schema.d.ts +17 -3
  79. package/src/config/schema/kvcache.schema.d.ts +19 -3
  80. package/src/config/schema/kvcache.schema.js +2 -1
  81. package/src/config/schema/manifest.schema.d.ts +38 -1
  82. package/src/config/schema/manifest.schema.js +18 -0
  83. package/src/config/schema/speculation-self.schema.d.ts +16 -0
  84. package/src/config/source-packages/litert/gemma-4-12b-it.json +20 -1
  85. package/src/config/source-packages/litert/gemma-4-e2b-it.json +20 -1
  86. package/src/config/source-packages/litert/gemma-4-e4b-it.json +21 -1
  87. package/src/config/transforms/capability-transform-resolver.js +7 -1
  88. package/src/config/transforms/execution-graph-transforms.js +10 -2
  89. package/src/converter/conversion-plan.js +2 -0
  90. package/src/converter/parsers/transformer.js +27 -1
  91. package/src/converter/quantization-info.js +16 -4
  92. package/src/converter/rope-config.js +55 -3
  93. package/src/debug/config.d.ts +4 -0
  94. package/src/experimental/orchestration/index.d.ts +10 -1
  95. package/src/experimental/training/distillation/student-fixture.js +1 -1
  96. package/src/experimental/training/lora-pipeline.js +1 -1
  97. package/src/experimental/training/lora.js +2 -2
  98. package/src/experimental/training/suite.js +2 -2
  99. package/src/formats/rdrr/classification.js +3 -0
  100. package/src/formats/rdrr/integrity.d.ts +0 -2
  101. package/src/formats/rdrr/integrity.js +0 -2
  102. package/src/formats/rdrr/types.d.ts +1 -1
  103. package/src/gpu/kernels/attention.d.ts +2 -0
  104. package/src/gpu/kernels/attention.js +75 -16
  105. package/src/gpu/kernels/attention_decode_online_head256_f16kv_output_gate.wgsl +268 -0
  106. package/src/gpu/kernels/attention_head128_f16kv.wgsl +269 -0
  107. package/src/gpu/kernels/backward/attention_backward.js +8 -0
  108. package/src/gpu/kernels/dequant.d.ts +6 -6
  109. package/src/gpu/kernels/dequant.js +17 -6
  110. package/src/gpu/kernels/fused_ffn.d.ts +32 -0
  111. package/src/gpu/kernels/fused_ffn.js +540 -19
  112. package/src/gpu/kernels/fused_ffn.wgsl +2 -1
  113. package/src/gpu/kernels/fused_ffn_f16.wgsl +2 -1
  114. package/src/gpu/kernels/fused_ffn_f16w_f32a_batched.wgsl +135 -0
  115. package/src/gpu/kernels/fused_ffn_q4k.wgsl +5 -3
  116. package/src/gpu/kernels/fused_ffn_q4k_f16.wgsl +2 -1
  117. package/src/gpu/kernels/fused_ffn_q4k_metal_simd16.wgsl +183 -0
  118. package/src/gpu/kernels/fused_gate_up_gelu_gated_f16.wgsl +2 -1
  119. package/src/gpu/kernels/fused_matmul_q4.wgsl +15 -6
  120. package/src/gpu/kernels/fused_matmul_q4_metal_simd16.wgsl +144 -0
  121. package/src/gpu/kernels/index.d.ts +21 -0
  122. package/src/gpu/kernels/index.js +13 -0
  123. package/src/gpu/kernels/linear-attention-core.js +414 -21
  124. package/src/gpu/kernels/lm-head-argmax.d.ts +10 -1
  125. package/src/gpu/kernels/lm-head-argmax.js +153 -57
  126. package/src/gpu/kernels/lm-head-select-logits.d.ts +23 -0
  127. package/src/gpu/kernels/lm-head-select-logits.js +186 -0
  128. package/src/gpu/kernels/lm_head_argmax_q4k.wgsl +289 -0
  129. package/src/gpu/kernels/lm_head_select_logits.wgsl +77 -0
  130. package/src/gpu/kernels/matmul-selection.js +9 -2
  131. package/src/gpu/kernels/matmul.d.ts +1 -1
  132. package/src/gpu/kernels/matmul.js +24 -6
  133. package/src/gpu/kernels/moe.d.ts +5 -1
  134. package/src/gpu/kernels/moe.js +5 -2
  135. package/src/gpu/kernels/residual_rmsnorm_pair.wgsl +127 -0
  136. package/src/gpu/kernels/rmsnorm-pair.d.ts +33 -0
  137. package/src/gpu/kernels/rmsnorm-pair.js +113 -0
  138. package/src/gpu/kernels/rmsnorm-stats.d.ts +29 -0
  139. package/src/gpu/kernels/rmsnorm-stats.js +250 -0
  140. package/src/gpu/kernels/silu.wgsl +7 -1
  141. package/src/gpu/kernels/silu_f16.wgsl +7 -1
  142. package/src/gpu/kernels/softmax.d.ts +3 -3
  143. package/src/gpu/kernels/softmax.js +9 -2
  144. package/src/gpu/kernels/split-qkv-rmsnorm-rope-qk.js +15 -10
  145. package/src/gpu/kernels/split_qkv_rmsnorm_rope_qk.wgsl +37 -7
  146. package/src/gpu/kernels/split_qkv_rmsnorm_rope_qk_f16kv_cache.wgsl +37 -7
  147. package/src/gpu/kernels/turboquant-codebook.js +1 -1
  148. package/src/gpu/kernels/types.d.ts +1 -1
  149. package/src/gpu/profiler.js +11 -1
  150. package/src/gpu/tensor.d.ts +1 -0
  151. package/src/index.d.ts +1 -1
  152. package/src/index.js +1 -1
  153. package/src/inference/browser-harness-model-helpers.js +4 -1
  154. package/src/inference/browser-harness-suite-helpers.js +1 -1
  155. package/src/inference/browser-harness-text-helpers.js +589 -117
  156. package/src/inference/browser-harness.d.ts +3 -3
  157. package/src/inference/browser-harness.js +486 -13
  158. package/src/inference/fixtures/rerank-semantic-fixtures.json +38 -0
  159. package/src/inference/kv-cache/base.js +42 -1
  160. package/src/inference/kv-cache/quantized.js +16 -2
  161. package/src/inference/kv-cache/tiered.js +50 -15
  162. package/src/inference/moe-router.js +38 -7
  163. package/src/inference/pipelines/diffusion/text-encoder-gpu.js +8 -6
  164. package/src/inference/pipelines/structured/json-grammar-mask.d.ts +18 -0
  165. package/src/inference/pipelines/text/attention/dispatch-params.js +8 -3
  166. package/src/inference/pipelines/text/attention/output-gate-fusion.js +37 -0
  167. package/src/inference/pipelines/text/attention/projections.js +174 -4
  168. package/src/inference/pipelines/text/attention/record.js +44 -21
  169. package/src/inference/pipelines/text/attention/run.js +18 -3
  170. package/src/inference/pipelines/text/config.d.ts +23 -1
  171. package/src/inference/pipelines/text/config.js +235 -112
  172. package/src/inference/pipelines/text/embed.js +39 -12
  173. package/src/inference/pipelines/text/execution-plan.d.ts +2 -0
  174. package/src/inference/pipelines/text/execution-plan.js +45 -2
  175. package/src/inference/pipelines/text/execution-v1.d.ts +6 -0
  176. package/src/inference/pipelines/text/execution-v1.js +122 -8
  177. package/src/inference/pipelines/text/ffn/dense.js +254 -25
  178. package/src/inference/pipelines/text/ffn/moe.js +4 -3
  179. package/src/inference/pipelines/text/ffn/standard.js +145 -39
  180. package/src/inference/pipelines/text/generator-decode-policy.js +1 -0
  181. package/src/inference/pipelines/text/generator-helpers.js +4 -1
  182. package/src/inference/pipelines/text/generator-runtime.js +54 -20
  183. package/src/inference/pipelines/text/generator-steps.js +5 -2
  184. package/src/inference/pipelines/text/generator.d.ts +18 -1
  185. package/src/inference/pipelines/text/generator.js +303 -20
  186. package/src/inference/pipelines/text/init-qkv-fusion.js +155 -62
  187. package/src/inference/pipelines/text/init-rope.js +54 -6
  188. package/src/inference/pipelines/text/init.d.ts +22 -3
  189. package/src/inference/pipelines/text/init.js +65 -17
  190. package/src/inference/pipelines/text/layer.js +111 -5
  191. package/src/inference/pipelines/text/linear-attention-ab-fusion.js +208 -0
  192. package/src/inference/pipelines/text/linear-attention.d.ts +1 -0
  193. package/src/inference/pipelines/text/linear-attention.js +185 -24
  194. package/src/inference/pipelines/text/logits/gpu.d.ts +1 -1
  195. package/src/inference/pipelines/text/logits/gpu.js +78 -115
  196. package/src/inference/pipelines/text/logits/index.js +105 -87
  197. package/src/inference/pipelines/text/logits/precision-policy.js +103 -0
  198. package/src/inference/pipelines/text/logits/types.d.ts +1 -0
  199. package/src/inference/pipelines/text/logits/utils.d.ts +2 -0
  200. package/src/inference/pipelines/text/logits/utils.js +11 -0
  201. package/src/inference/pipelines/text/lora-apply.js +6 -4
  202. package/src/inference/pipelines/text/model-load.js +6 -0
  203. package/src/inference/pipelines/text/moe-cpu.js +25 -0
  204. package/src/inference/pipelines/text/moe-gpu.js +148 -97
  205. package/src/inference/pipelines/text/moe-shape-validator.js +149 -31
  206. package/src/inference/pipelines/text/ops.js +48 -1
  207. package/src/inference/pipelines/text/per-layer-inputs.js +105 -21
  208. package/src/inference/pipelines/text/probes.js +9 -4
  209. package/src/inference/pipelines/text/state.d.ts +4 -4
  210. package/src/inference/pipelines/text/state.js +4 -0
  211. package/src/inference/pipelines/text/types.d.ts +27 -2
  212. package/src/inference/pipelines/text/weights.js +7 -2
  213. package/src/inference/pipelines/text.d.ts +21 -1
  214. package/src/inference/pipelines/text.js +60 -14
  215. package/src/inference/tokenizers/behavior-flags.js +23 -2
  216. package/src/inference/tokenizers/bundled.js +45 -23
  217. package/src/inference/tokenizers/types.d.ts +7 -4
  218. package/src/loader/doppler-loader.js +25 -2
  219. package/src/loader/final-weights-loader.js +12 -1
  220. package/src/loader/index.d.ts +3 -1
  221. package/src/loader/index.js +2 -0
  222. package/src/loader/loader-types.d.ts +3 -1
  223. package/src/loader/manifest-config.d.ts +18 -0
  224. package/src/loader/manifest-config.js +11 -3
  225. package/src/loader/shard-cache.d.ts +2 -0
  226. package/src/loader/shard-cache.js +29 -1
  227. package/src/models/gemma4.js +8 -0
  228. package/src/rules/execution-rules-contract-check.js +12 -3
  229. package/src/rules/inference/capability-transforms.rules.json +159 -22
  230. package/src/rules/inference/config.rules.json +26 -2
  231. package/src/rules/inference/dtype.rules.json +12 -5
  232. package/src/rules/inference/execution.rules.json +5 -19
  233. package/src/rules/inference/ffn.rules.json +33 -0
  234. package/src/rules/inference/moe.rules.json +97 -0
  235. package/src/rules/kernels/fused-ffn.rules.json +1 -1
  236. package/src/rules/kernels/lm-head-argmax.rules.json +10 -0
  237. package/src/rules/kernels/matmul.rules.json +1 -0
  238. package/src/rules/kernels/silu.rules.json +1 -0
  239. package/src/rules/rule-registry.js +2 -0
  240. package/src/storage/artifact-storage-context.js +36 -0
  241. package/src/tooling/command-api-constants.js +2 -2
  242. package/src/tooling/command-api-family-normalizers.js +2 -1
  243. package/src/tooling/command-api.d.ts +1 -1
  244. package/src/tooling/command-runner-shared.js +3 -1
  245. package/src/tooling/node-browser-command-runner.js +42 -7
  246. package/src/tooling/node-converter.js +20 -0
  247. package/src/tooling/node-source-runtime.js +5 -1
  248. package/src/tooling/opfs-cache.d.ts +21 -1
  249. package/src/tooling/opfs-cache.js +168 -32
  250. package/src/tooling-exports/storage.d.ts +1 -1
  251. package/src/tooling-exports/storage.js +1 -1
  252. package/src/types/gpu.d.ts +25 -0
  253. package/src/version.js +1 -1
  254. package/tests/kernels/browser/kernel-suite.js +1593 -0
  255. package/tests/kernels/browser/test-page.js +2629 -0
  256. package/tests/kernels/harness/benchmark.js +225 -0
  257. package/tests/kernels/harness/buffer-utils.js +88 -0
  258. package/tests/kernels/harness/tolerance.js +174 -0
  259. package/tests/kernels/reference/adam.js +15 -0
  260. package/tests/kernels/reference/attention.js +158 -0
  261. package/tests/kernels/reference/backward.js +96 -0
  262. package/tests/kernels/reference/bias-add.js +20 -0
  263. package/tests/kernels/reference/conv2d.js +58 -0
  264. package/tests/kernels/reference/cross-entropy.js +26 -0
  265. package/tests/kernels/reference/dequant.js +312 -0
  266. package/tests/kernels/reference/embed.js +15 -0
  267. package/tests/kernels/reference/gather.js +61 -0
  268. package/tests/kernels/reference/gelu.js +37 -0
  269. package/tests/kernels/reference/groupnorm.js +36 -0
  270. package/tests/kernels/reference/index.js +89 -0
  271. package/tests/kernels/reference/layernorm.js +22 -0
  272. package/tests/kernels/reference/matmul.js +57 -0
  273. package/tests/kernels/reference/modulate.js +24 -0
  274. package/tests/kernels/reference/moe-gather.js +83 -0
  275. package/tests/kernels/reference/pixel-shuffle.js +20 -0
  276. package/tests/kernels/reference/residual.js +55 -0
  277. package/tests/kernels/reference/rmsnorm.js +52 -0
  278. package/tests/kernels/reference/rope.js +79 -0
  279. package/tests/kernels/reference/sample.js +97 -0
  280. package/tests/kernels/reference/scatter-add.js +58 -0
  281. package/tests/kernels/reference/silu.js +52 -0
  282. package/tests/kernels/reference/softmax.js +86 -0
  283. package/tests/kernels/reference/split-qkv.js +58 -0
  284. package/tests/kernels/reference/topk.js +88 -0
  285. package/tests/kernels/reference/upsample2d.js +37 -0
  286. package/BRANDING.md +0 -14
  287. package/SECURITY.md +0 -19
  288. package/assets/doppler-readme-hero.svg +0 -84
  289. package/assets/doppler-runtime-map.svg +0 -84
  290. package/src/bootstrap.js +0 -30
  291. package/src/cli/doppler-cli.d.ts +0 -46
  292. package/src/cli/doppler-quickstart.d.ts +0 -45
  293. package/src/cli/doppler-serve.d.ts +0 -210
  294. package/src/cli/runtime-profiles.d.ts +0 -40
  295. package/src/client/doppler-api.browser.d.ts +0 -1
  296. package/src/client/doppler-api.browser.js +0 -25
  297. package/src/client/doppler-registry.d.ts +0 -35
  298. package/src/client/runtime/node-quickstart-cache.d.ts +0 -22
  299. package/src/client/runtime/source-runtime.d.ts +0 -26
  300. package/src/client/runtime/storage.d.ts +0 -5
  301. package/src/config/README.md +0 -61
  302. package/src/config/conversion/README.md +0 -58
  303. package/src/config/goal-completion-matrix.json +0 -552
  304. package/src/config/kernels/kernel-ref-digests.d.ts +0 -1
  305. package/src/config/kernels/moe/gpt-oss.paths.json +0 -49
  306. package/src/config/kernels/moe/mixtral.paths.json +0 -46
  307. package/src/config/kernels/registry.d.ts +0 -51
  308. package/src/config/merge-contract-check.d.ts +0 -16
  309. package/src/config/merge-contract-check.js +0 -318
  310. package/src/config/merge-helpers.d.ts +0 -60
  311. package/src/config/model-lanes/catalog-lane-resolver.d.ts +0 -66
  312. package/src/config/model-lanes/catalog-lane-resolver.js +0 -170
  313. package/src/config/param-categories.d.ts +0 -8
  314. package/src/config/param-validator.d.ts +0 -28
  315. package/src/config/platforms/loader.d.ts +0 -38
  316. package/src/config/quantization-contract-check.d.ts +0 -12
  317. package/src/config/quantization-contract-check.js +0 -91
  318. package/src/config/runtime/README.md +0 -28
  319. package/src/config/runtime-merge.d.ts +0 -5
  320. package/src/config/schema/claim-evidence-contract.schema.json +0 -93
  321. package/src/config/schema/command-surface-contract.schema.json +0 -81
  322. package/src/config/schema/goal-completion-matrix.schema.json +0 -223
  323. package/src/config/schema/model-artifact-contract.schema.json +0 -119
  324. package/src/config/schema/platform.schema.js +0 -1
  325. package/src/config/schema/policy-schema-registry.json +0 -26
  326. package/src/config/schema/subsystem-support-contract.schema.json +0 -48
  327. package/src/config/support-tiers/subsystems.json +0 -502
  328. package/src/config/supported-operations.d.ts +0 -14
  329. package/src/config/transforms/capability-transform-resolver.d.ts +0 -56
  330. package/src/config/transforms/execution-graph-transforms.d.ts +0 -300
  331. package/src/converter/conversion-plan.d.ts +0 -56
  332. package/src/converter/index.d.ts +0 -98
  333. package/src/converter/index.js +0 -59
  334. package/src/converter/parsers/diffusion.d.ts +0 -50
  335. package/src/converter/parsers/gguf.d.ts +0 -31
  336. package/src/converter/parsers/index.d.ts +0 -21
  337. package/src/converter/parsers/index.js +0 -12
  338. package/src/converter/parsers/transformer.d.ts +0 -31
  339. package/src/converter/quantization-info.d.ts +0 -38
  340. package/src/converter/rope-config.d.ts +0 -16
  341. package/src/converter/tokenizer-utils.d.ts +0 -12
  342. package/src/debug/reference/README.md +0 -114
  343. package/src/debug/reference/hf_attn_debug.py +0 -114
  344. package/src/debug/reference/hf_embed_check.py +0 -89
  345. package/src/debug/reference/hf_layer_out.py +0 -100
  346. package/src/debug/reference/hf_qwen35_linear_attn_debug.py +0 -268
  347. package/src/debug/reference/hf_rope_check.py +0 -116
  348. package/src/debug/reference/hf_weights.py +0 -75
  349. package/src/errors/doppler-error.d.ts +0 -20
  350. package/src/experimental/bridge/extension/background.d.ts +0 -14
  351. package/src/experimental/bridge/extension/background.js +0 -168
  352. package/src/experimental/bridge/extension/manifest.json +0 -34
  353. package/src/experimental/bridge/extension-client.d.ts +0 -114
  354. package/src/experimental/bridge/index.d.ts +0 -69
  355. package/src/experimental/bridge/node-bridge.js +0 -24
  356. package/src/experimental/bridge/protocol.d.ts +0 -96
  357. package/src/experimental/browser/gguf-importer.d.ts +0 -136
  358. package/src/experimental/browser/gguf-importer.js +0 -532
  359. package/src/experimental/browser/gguf-parser-browser.d.ts +0 -14
  360. package/src/experimental/browser/quantization.d.ts +0 -74
  361. package/src/experimental/browser/safetensors-parser-browser.d.ts +0 -193
  362. package/src/experimental/browser/shard-io-browser.d.ts +0 -57
  363. package/src/experimental/browser/shard-io-browser.js +0 -89
  364. package/src/experimental/browser/tensor-source-download.d.ts +0 -27
  365. package/src/experimental/diffusion/config/expansion-lanes.json +0 -29
  366. package/src/experimental/distribution/p2p-transport-contract.d.ts +0 -57
  367. package/src/experimental/distribution/p2p-transport.js +0 -75
  368. package/src/experimental/hotswap/intent-bundle.d.ts +0 -37
  369. package/src/experimental/hotswap/manifest.d.ts +0 -42
  370. package/src/experimental/hotswap/model-swap.js +0 -42
  371. package/src/experimental/hotswap/runtime.d.ts +0 -31
  372. package/src/experimental/training/README.md +0 -164
  373. package/src/experimental/training/artifacts.d.ts +0 -160
  374. package/src/experimental/training/attention-backward.d.ts +0 -30
  375. package/src/experimental/training/attention-forward.d.ts +0 -22
  376. package/src/experimental/training/attention-forward.js +0 -82
  377. package/src/experimental/training/autograd.d.ts +0 -51
  378. package/src/experimental/training/backward-kernels.js +0 -39
  379. package/src/experimental/training/checkpoint-watch.d.ts +0 -8
  380. package/src/experimental/training/checkpoint.d.ts +0 -36
  381. package/src/experimental/training/clip.d.ts +0 -9
  382. package/src/experimental/training/dataloader.d.ts +0 -8
  383. package/src/experimental/training/datasets/gepa-frontier.d.ts +0 -41
  384. package/src/experimental/training/datasets/gepa-frontier.js +0 -220
  385. package/src/experimental/training/datasets/index.d.ts +0 -24
  386. package/src/experimental/training/datasets/index.js +0 -13
  387. package/src/experimental/training/datasets/jsonl.d.ts +0 -11
  388. package/src/experimental/training/datasets/reploid.d.ts +0 -3
  389. package/src/experimental/training/datasets/reploid.js +0 -36
  390. package/src/experimental/training/datasets/teacher-traces.d.ts +0 -51
  391. package/src/experimental/training/datasets/teacher-traces.js +0 -279
  392. package/src/experimental/training/datasets/text-pairs.d.ts +0 -66
  393. package/src/experimental/training/datasets/token-batch.d.ts +0 -21
  394. package/src/experimental/training/datasets/token-batch.js +0 -52
  395. package/src/experimental/training/datasets/translation-pairs.d.ts +0 -34
  396. package/src/experimental/training/datasets/translation-pairs.js +0 -49
  397. package/src/experimental/training/distillation/artifacts.d.ts +0 -71
  398. package/src/experimental/training/distillation/checkpoint-watch.d.ts +0 -10
  399. package/src/experimental/training/distillation/dataset.d.ts +0 -59
  400. package/src/experimental/training/distillation/eval.d.ts +0 -34
  401. package/src/experimental/training/distillation/index.d.ts +0 -29
  402. package/src/experimental/training/distillation/runtime.d.ts +0 -20
  403. package/src/experimental/training/distillation/scoreboard.d.ts +0 -6
  404. package/src/experimental/training/distillation/stage-a.d.ts +0 -45
  405. package/src/experimental/training/distillation/stage-b.d.ts +0 -24
  406. package/src/experimental/training/distillation/student-fixture.d.ts +0 -27
  407. package/src/experimental/training/distillation/suite-data.d.ts +0 -45
  408. package/src/experimental/training/export.d.ts +0 -41
  409. package/src/experimental/training/index.d.ts +0 -74
  410. package/src/experimental/training/index.js +0 -63
  411. package/src/experimental/training/lora-pipeline.d.ts +0 -183
  412. package/src/experimental/training/lora.d.ts +0 -19
  413. package/src/experimental/training/loss-scaling.d.ts +0 -21
  414. package/src/experimental/training/loss.d.ts +0 -10
  415. package/src/experimental/training/objectives/base.d.ts +0 -58
  416. package/src/experimental/training/objectives/cross_entropy.d.ts +0 -18
  417. package/src/experimental/training/objectives/distill_kd.d.ts +0 -16
  418. package/src/experimental/training/objectives/distill_triplet.d.ts +0 -16
  419. package/src/experimental/training/objectives/index.d.ts +0 -12
  420. package/src/experimental/training/objectives/index.js +0 -6
  421. package/src/experimental/training/objectives/ul_stage1_joint.d.ts +0 -16
  422. package/src/experimental/training/objectives/ul_stage2_base.d.ts +0 -16
  423. package/src/experimental/training/operator-agent-eval.d.ts +0 -65
  424. package/src/experimental/training/operator-artifacts.d.ts +0 -62
  425. package/src/experimental/training/operator-command.d.ts +0 -5
  426. package/src/experimental/training/operator-eval.d.ts +0 -48
  427. package/src/experimental/training/operator-scoreboard.d.ts +0 -5
  428. package/src/experimental/training/optimizer.d.ts +0 -22
  429. package/src/experimental/training/runner.d.ts +0 -249
  430. package/src/experimental/training/suite.d.ts +0 -299
  431. package/src/experimental/training/tensor-factory.d.ts +0 -9
  432. package/src/experimental/training/trainer.d.ts +0 -89
  433. package/src/experimental/training/ul_dataset.d.ts +0 -47
  434. package/src/experimental/training/ul_schedule.d.ts +0 -6
  435. package/src/experimental/training/workload-packs/distill-glm52-doppler-agent-harness-gemma270m-f16-sft-tiny.json +0 -160
  436. package/src/experimental/training/workload-packs/distill-glm52-doppler-code-agent-gemma270m-f16-sft-tiny.json +0 -175
  437. package/src/experimental/training/workload-packs/distill-glm52-doppler-code-agent-sft-tiny.json +0 -175
  438. package/src/experimental/training/workload-packs/distill-glm52-doppler-js-json-gemma270m-f16-sft-tiny.json +0 -160
  439. package/src/experimental/training/workload-packs/distill-glm52-doppler-review-gemma270m-f16-sft-tiny.json +0 -162
  440. package/src/experimental/training/workload-packs/distill-glm52-doppler-wgsl-gemma270m-f16-sft-tiny.json +0 -160
  441. package/src/experimental/training/workload-packs/distill-translategemma-medium.json +0 -103
  442. package/src/experimental/training/workload-packs/distill-translategemma-tiny.json +0 -103
  443. package/src/experimental/training/workload-packs/lora-doppler-agent-harness-gemma270m-f16-tiny.json +0 -137
  444. package/src/experimental/training/workload-packs/lora-doppler-code-agent-gemma270m-f16-tiny.json +0 -152
  445. package/src/experimental/training/workload-packs/lora-doppler-code-agent-tiny.json +0 -152
  446. package/src/experimental/training/workload-packs/lora-doppler-js-json-gemma270m-f16-tiny.json +0 -137
  447. package/src/experimental/training/workload-packs/lora-doppler-review-gemma270m-f16-tiny.json +0 -139
  448. package/src/experimental/training/workload-packs/lora-doppler-wgsl-gemma270m-f16-tiny.json +0 -137
  449. package/src/experimental/training/workload-packs/lora-toy-medium.json +0 -91
  450. package/src/experimental/training/workload-packs/lora-toy-tiny.json +0 -91
  451. package/src/experimental/training/workload-packs/registry.json +0 -278
  452. package/src/experimental/training/workload-packs/ul-training-medium.json +0 -55
  453. package/src/experimental/training/workload-packs/ul-training-tiny.json +0 -55
  454. package/src/experimental/training/workloads.d.ts +0 -201
  455. package/src/formats/gguf/types.d.ts +0 -74
  456. package/src/formats/index.d.ts +0 -75
  457. package/src/formats/index.js +0 -19
  458. package/src/formats/litert/types.d.ts +0 -74
  459. package/src/formats/rdrr/distributed/parsing.d.ts +0 -6
  460. package/src/formats/rdrr/distributed/validation.d.ts +0 -23
  461. package/src/formats/rdrr/storage-descriptor.d.ts +0 -36
  462. package/src/formats/rdrr/tensor-config-validator.d.ts +0 -33
  463. package/src/formats/safetensors/types.d.ts +0 -67
  464. package/src/formats/tflite/types.d.ts +0 -61
  465. package/src/formats/tokenizer/index.d.ts +0 -5
  466. package/src/formats/tokenizer/index.js +0 -3
  467. package/src/formats/tokenizer/types.d.ts +0 -9
  468. package/src/gpu/kernel-runtime.d.ts +0 -13
  469. package/src/gpu/kernel-selection-cache.d.ts +0 -13
  470. package/src/gpu/kernel-selection-log.d.ts +0 -4
  471. package/src/gpu/kernel-tuner/benchmarks.d.ts +0 -144
  472. package/src/gpu/kernel-tuner/cache.d.ts +0 -55
  473. package/src/gpu/kernel-tuner/index.d.ts +0 -59
  474. package/src/gpu/kernel-tuner/tuner.d.ts +0 -82
  475. package/src/gpu/kernel-tuner/types.d.ts +0 -101
  476. package/src/gpu/kernel-tuner/types.js +0 -4
  477. package/src/gpu/kernel-tuner.d.ts +0 -33
  478. package/src/gpu/kernels/README.md +0 -159
  479. package/src/gpu/kernels/check-finiteness.d.ts +0 -15
  480. package/src/gpu/kernels/check-hot-vocab-stop.d.ts +0 -18
  481. package/src/gpu/kernels/check-stop.d.ts +0 -31
  482. package/src/gpu/kernels/codegen/patches/attention_decode_f16.from.attention_decode.diff +0 -213
  483. package/src/gpu/kernels/codegen/patches/attention_decode_f16kv.from.attention_decode.diff +0 -89
  484. package/src/gpu/kernels/codegen/patches/attention_f16.from.attention.diff +0 -62
  485. package/src/gpu/kernels/codegen/patches/attention_f16kv.from.attention.diff +0 -40
  486. package/src/gpu/kernels/codegen/patches/attention_small_f16.from.attention_small.diff +0 -226
  487. package/src/gpu/kernels/codegen/patches/attention_small_f16kv.from.attention_small.diff +0 -264
  488. package/src/gpu/kernels/codegen/patches/attention_streaming_f16.from.attention_streaming.diff +0 -71
  489. package/src/gpu/kernels/codegen/patches/attention_streaming_f16kv.from.attention_streaming.diff +0 -53
  490. package/src/gpu/kernels/codegen/patches/bias_add_f16.from.bias_add.diff +0 -54
  491. package/src/gpu/kernels/codegen/patches/dequant_mxfp4_expert_f16.from.dequant_mxfp4_expert.diff +0 -125
  492. package/src/gpu/kernels/codegen/patches/fused_ffn_f16.from.fused_ffn.diff +0 -428
  493. package/src/gpu/kernels/codegen/patches/fused_matmul_q4_batched_f16.from.fused_matmul_q4_batched.diff +0 -114
  494. package/src/gpu/kernels/codegen/patches/fused_matmul_rmsnorm_f16.from.fused_matmul_rmsnorm.diff +0 -477
  495. package/src/gpu/kernels/codegen/patches/fused_swiglu_f16.from.fused_swiglu.diff +0 -60
  496. package/src/gpu/kernels/codegen/patches/gelu_f16.from.gelu.diff +0 -41
  497. package/src/gpu/kernels/codegen/patches/layernorm_f16.from.layernorm.diff +0 -82
  498. package/src/gpu/kernels/codegen/patches/matmul_gemv_residual_f16.from.matmul_gemv_residual.diff +0 -87
  499. package/src/gpu/kernels/codegen/patches/moe_gather_f16.from.moe_gather.diff +0 -181
  500. package/src/gpu/kernels/codegen/patches/residual_f16.from.residual.diff +0 -65
  501. package/src/gpu/kernels/codegen/patches/rmsnorm_f16.from.rmsnorm.diff +0 -515
  502. package/src/gpu/kernels/codegen/patches/rope_f16.from.rope.diff +0 -292
  503. package/src/gpu/kernels/codegen/patches/sample_f16.from.sample.diff +0 -336
  504. package/src/gpu/kernels/codegen/patches/scatter_add_dynamic_f16.from.scatter_add_dynamic.diff +0 -65
  505. package/src/gpu/kernels/codegen/patches/silu_f16.from.silu.diff +0 -81
  506. package/src/gpu/kernels/codegen/patches/topk_f16.from.topk.diff +0 -231
  507. package/src/gpu/kernels/codegen/wgsl-patch-variants.js +0 -152
  508. package/src/gpu/kernels/codegen/wgsl-variants.js +0 -885
  509. package/src/gpu/kernels/constants.d.ts +0 -129
  510. package/src/gpu/kernels/dispatch.d.ts +0 -72
  511. package/src/gpu/kernels/gated-short-conv.d.ts +0 -63
  512. package/src/gpu/kernels/kernel-base.d.ts +0 -33
  513. package/src/gpu/kernels/linear-attention-core.d.ts +0 -42
  514. package/src/gpu/kernels/matmul-dispatch.d.ts +0 -49
  515. package/src/gpu/kernels/matmul-selection.d.ts +0 -87
  516. package/src/gpu/kernels/rep-penalty.d.ts +0 -21
  517. package/src/gpu/kernels/rule-matcher.d.ts +0 -30
  518. package/src/gpu/kernels/rule-registry.d.ts +0 -7
  519. package/src/gpu/kernels/types.js +0 -4
  520. package/src/gpu/perf-guards.d.ts +0 -13
  521. package/src/gpu/profiler.d.ts +0 -93
  522. package/src/gpu/readback-utils.d.ts +0 -16
  523. package/src/gpu/readback-utils.js +0 -49
  524. package/src/gpu/submit-probe.d.ts +0 -12
  525. package/src/index-browser.d.ts +0 -2
  526. package/src/index-browser.js +0 -2
  527. package/src/inference/README.md +0 -593
  528. package/src/inference/browser-harness-contract-helpers.d.ts +0 -5
  529. package/src/inference/browser-harness-diffusion-energy-suites.d.ts +0 -2
  530. package/src/inference/browser-harness-model-helpers.d.ts +0 -10
  531. package/src/inference/browser-harness-report-helpers.d.ts +0 -7
  532. package/src/inference/browser-harness-suite-helpers.d.ts +0 -40
  533. package/src/inference/browser-harness-text-helpers.d.ts +0 -124
  534. package/src/inference/model-lanes/execution-lane-resolver.d.ts +0 -61
  535. package/src/inference/model-lanes/execution-lane-resolver.js +0 -126
  536. package/src/inference/network-evolution.js +0 -79
  537. package/src/inference/pipelines/context.d.ts +0 -21
  538. package/src/inference/pipelines/diffusion/helpers.d.ts +0 -29
  539. package/src/inference/pipelines/diffusion/index.d.ts +0 -3
  540. package/src/inference/pipelines/diffusion/index.js +0 -3
  541. package/src/inference/pipelines/diffusion/scheduler.d.ts +0 -35
  542. package/src/inference/pipelines/diffusion/sd3-transformer.d.ts +0 -20
  543. package/src/inference/pipelines/diffusion/sd3-weights.d.ts +0 -21
  544. package/src/inference/pipelines/diffusion/text-encoder-gpu.d.ts +0 -82
  545. package/src/inference/pipelines/diffusion/text-encoder.d.ts +0 -29
  546. package/src/inference/pipelines/diffusion/types.js +0 -1
  547. package/src/inference/pipelines/diffusion/vae.d.ts +0 -20
  548. package/src/inference/pipelines/dream/energy-head-pipeline.d.ts +0 -29
  549. package/src/inference/pipelines/dream/energy-head-pipeline.js +0 -6
  550. package/src/inference/pipelines/dream/pipeline.d.ts +0 -17
  551. package/src/inference/pipelines/dream/pipeline.js +0 -8
  552. package/src/inference/pipelines/energy/index.d.ts +0 -1
  553. package/src/inference/pipelines/energy/index.js +0 -1
  554. package/src/inference/pipelines/energy/types.js +0 -1
  555. package/src/inference/pipelines/energy-head/index.d.ts +0 -6
  556. package/src/inference/pipelines/energy-head/index.js +0 -6
  557. package/src/inference/pipelines/factory.d.ts +0 -10
  558. package/src/inference/pipelines/index.d.ts +0 -22
  559. package/src/inference/pipelines/index.js +0 -19
  560. package/src/inference/pipelines/registry.d.ts +0 -15
  561. package/src/inference/pipelines/rng.d.ts +0 -2
  562. package/src/inference/pipelines/structured/tsv-grammar-mask.js +0 -166
  563. package/src/inference/pipelines/text/attention/attn-config.d.ts +0 -23
  564. package/src/inference/pipelines/text/attention/dispatch-params.d.ts +0 -173
  565. package/src/inference/pipelines/text/attention/index.d.ts +0 -24
  566. package/src/inference/pipelines/text/attention/output-projection.d.ts +0 -12
  567. package/src/inference/pipelines/text/attention/precision-contract.d.ts +0 -36
  568. package/src/inference/pipelines/text/attention/projections.d.ts +0 -208
  569. package/src/inference/pipelines/text/attention/quant-options.d.ts +0 -38
  570. package/src/inference/pipelines/text/attention/record.d.ts +0 -36
  571. package/src/inference/pipelines/text/attention/rmsnorm-fusion-gate.d.ts +0 -8
  572. package/src/inference/pipelines/text/attention/run.d.ts +0 -38
  573. package/src/inference/pipelines/text/attention/types.d.ts +0 -125
  574. package/src/inference/pipelines/text/attention.d.ts +0 -23
  575. package/src/inference/pipelines/text/attention.js +0 -12
  576. package/src/inference/pipelines/text/bdpa-steamroller.d.ts +0 -22
  577. package/src/inference/pipelines/text/buffer-types.js +0 -4
  578. package/src/inference/pipelines/text/debug-utils/config.d.ts +0 -144
  579. package/src/inference/pipelines/text/debug-utils/index.d.ts +0 -53
  580. package/src/inference/pipelines/text/debug-utils/logging.d.ts +0 -106
  581. package/src/inference/pipelines/text/debug-utils/tensor.d.ts +0 -119
  582. package/src/inference/pipelines/text/debug-utils/utils.d.ts +0 -77
  583. package/src/inference/pipelines/text/debug-utils.d.ts +0 -42
  584. package/src/inference/pipelines/text/drift-policy.d.ts +0 -57
  585. package/src/inference/pipelines/text/dtype-contract.d.ts +0 -53
  586. package/src/inference/pipelines/text/embed.d.ts +0 -68
  587. package/src/inference/pipelines/text/execution-runtime-builders.d.ts +0 -49
  588. package/src/inference/pipelines/text/failure-reduction.d.ts +0 -84
  589. package/src/inference/pipelines/text/failure-reduction.js +0 -180
  590. package/src/inference/pipelines/text/ffn/dense.d.ts +0 -99
  591. package/src/inference/pipelines/text/ffn/index.d.ts +0 -23
  592. package/src/inference/pipelines/text/ffn/moe.d.ts +0 -24
  593. package/src/inference/pipelines/text/ffn/sandwich.d.ts +0 -27
  594. package/src/inference/pipelines/text/ffn/standard.d.ts +0 -25
  595. package/src/inference/pipelines/text/ffn/types.d.ts +0 -30
  596. package/src/inference/pipelines/text/ffn.d.ts +0 -31
  597. package/src/inference/pipelines/text/ffn.js +0 -18
  598. package/src/inference/pipelines/text/finiteness-guard-status.d.ts +0 -11
  599. package/src/inference/pipelines/text/finiteness-policy.d.ts +0 -39
  600. package/src/inference/pipelines/text/generator-decode-policy.d.ts +0 -81
  601. package/src/inference/pipelines/text/generator-helpers.d.ts +0 -45
  602. package/src/inference/pipelines/text/generator-prefill-helpers.d.ts +0 -84
  603. package/src/inference/pipelines/text/generator-runtime.d.ts +0 -89
  604. package/src/inference/pipelines/text/generator-steps.d.ts +0 -186
  605. package/src/inference/pipelines/text/index.d.ts +0 -5
  606. package/src/inference/pipelines/text/index.js +0 -6
  607. package/src/inference/pipelines/text/kernel-trace.d.ts +0 -119
  608. package/src/inference/pipelines/text/layer.d.ts +0 -118
  609. package/src/inference/pipelines/text/load-errors.d.ts +0 -21
  610. package/src/inference/pipelines/text/logits/types.js +0 -4
  611. package/src/inference/pipelines/text/logits.d.ts +0 -27
  612. package/src/inference/pipelines/text/logits.js +0 -16
  613. package/src/inference/pipelines/text/lora-apply.d.ts +0 -28
  614. package/src/inference/pipelines/text/moe-cache.d.ts +0 -32
  615. package/src/inference/pipelines/text/moe-cpu-gptoss.d.ts +0 -9
  616. package/src/inference/pipelines/text/moe-cpu.d.ts +0 -13
  617. package/src/inference/pipelines/text/moe-gpu.d.ts +0 -26
  618. package/src/inference/pipelines/text/moe-helpers.d.ts +0 -12
  619. package/src/inference/pipelines/text/moe-shape-validator.d.ts +0 -40
  620. package/src/inference/pipelines/text/operator-events.d.ts +0 -125
  621. package/src/inference/pipelines/text/operator-identity.d.ts +0 -60
  622. package/src/inference/pipelines/text/ops.d.ts +0 -219
  623. package/src/inference/pipelines/text/per-layer-inputs.d.ts +0 -107
  624. package/src/inference/pipelines/text/probes.d.ts +0 -28
  625. package/src/inference/pipelines/text/sampling-config.d.ts +0 -48
  626. package/src/inference/pipelines/text/stage-names.d.ts +0 -88
  627. package/src/inference/pipelines/text/tsir-fixture-writer.d.ts +0 -57
  628. package/src/inference/pipelines/text/types.js +0 -4
  629. package/src/inference/tokenizers/base.d.ts +0 -44
  630. package/src/inference/tokenizers/behavior-flags.d.ts +0 -9
  631. package/src/inference/tokenizers/bpe.d.ts +0 -27
  632. package/src/inference/tokenizers/bundled.d.ts +0 -67
  633. package/src/inference/tokenizers/sentencepiece.d.ts +0 -30
  634. package/src/inference/tokenizers/types.js +0 -7
  635. package/src/loader/embedding-limit-preflight.d.ts +0 -31
  636. package/src/loader/per-layer-input-loader.d.ts +0 -30
  637. package/src/loader/tensors/functional-descriptor-loader.d.ts +0 -9
  638. package/src/loader/tensors/source-transform.d.ts +0 -65
  639. package/src/memory/address-table.d.ts +0 -96
  640. package/src/rules/execution-rules-contract-check.d.ts +0 -21
  641. package/src/rules/layer-pattern-contract-check.d.ts +0 -17
  642. package/src/rules/rule-registry.d.ts +0 -77
  643. package/src/storage/artifact-storage-context.d.ts +0 -63
  644. package/src/storage/backends/idb-store.d.ts +0 -52
  645. package/src/storage/backends/memory-store.d.ts +0 -36
  646. package/src/storage/backends/opfs-store.d.ts +0 -41
  647. package/src/storage/blake3.d.ts +0 -17
  648. package/src/storage/download-types.d.ts +0 -157
  649. package/src/storage/downloader.d.ts +0 -103
  650. package/src/storage/index.d.ts +0 -274
  651. package/src/storage/index.js +0 -201
  652. package/src/storage/preflight.d.ts +0 -144
  653. package/src/storage/preflight.js +0 -360
  654. package/src/storage/quickstart-downloader.d.ts +0 -157
  655. package/src/storage/quickstart-downloader.js +0 -289
  656. package/src/storage/source-artifact-store.d.ts +0 -59
  657. package/src/sw.js +0 -184
  658. package/src/tooling/command-api-constants.d.ts +0 -9
  659. package/src/tooling/command-api-family-normalizers.d.ts +0 -10
  660. package/src/tooling/command-api-helpers.d.ts +0 -23
  661. package/src/tooling/command-runner-shared.d.ts +0 -89
  662. package/src/tooling/conversion-config-materializer.d.ts +0 -14
  663. package/src/tooling/conversion-config-materializer.js +0 -96
  664. package/src/tooling/diagnose-runner.d.ts +0 -28
  665. package/src/tooling/hf-registry-utils.d.ts +0 -55
  666. package/src/tooling/kernel-path-builder/custom-runtime-facts.d.ts +0 -18
  667. package/src/tooling/kernel-path-builder/custom-runtime-facts.js +0 -72
  668. package/src/tooling/kernel-path-builder/index.d.ts +0 -65
  669. package/src/tooling/kernel-path-builder/index.js +0 -1361
  670. package/src/tooling/kernel-path-builder/runtime-overlay.d.ts +0 -9
  671. package/src/tooling/kernel-path-builder/runtime-overlay.js +0 -147
  672. package/src/tooling/lean-execution-contract-runner.d.ts +0 -33
  673. package/src/tooling/lean-execution-contract-runner.js +0 -158
  674. package/src/tooling/lean-execution-contract.d.ts +0 -16
  675. package/src/tooling/lean-execution-contract.js +0 -81
  676. package/src/tooling/litert-package-runtime.d.ts +0 -46
  677. package/src/tooling/node-convert-worker-pool.d.ts +0 -16
  678. package/src/tooling/node-convert-worker.d.ts +0 -1
  679. package/src/tooling/node-converter.d.ts +0 -1
  680. package/src/tooling/node-file-fetch.d.ts +0 -1
  681. package/src/tooling/node-source-runtime.d.ts +0 -24
  682. package/src/tooling/node-webgpu.d.ts +0 -17
  683. package/src/tooling/precision-replay-math.js +0 -174
  684. package/src/tooling/program-bundle-host.d.ts +0 -16
  685. package/src/tooling/program-bundle-host.js +0 -20
  686. package/src/tooling/runtime-input-composition.d.ts +0 -37
  687. package/src/tooling/source-package-profiles.d.ts +0 -50
  688. package/src/tooling/source-runtime-converter-config.d.ts +0 -19
  689. package/src/tooling/source-runtime-materializer.d.ts +0 -6
  690. package/src/tooling/source-runtime-materializer.js +0 -89
  691. package/src/tooling-exports.browser.d.ts +0 -7
  692. package/src/types/chrome.d.ts +0 -36
  693. package/src/types/chrome.js +0 -1
  694. package/src/types/index.d.ts +0 -1
  695. package/src/types/index.js +0 -1
  696. package/src/utils/array-buffer.d.ts +0 -4
  697. package/src/utils/canonical-hash.d.ts +0 -7
  698. package/src/utils/clone-json.d.ts +0 -1
  699. package/src/utils/hf-resolve-url.d.ts +0 -16
  700. package/src/utils/index.d.ts +0 -10
  701. package/src/utils/index.js +0 -10
  702. package/src/utils/load-json.d.ts +0 -5
  703. package/src/utils/plain-object.d.ts +0 -1
  704. package/src/utils/sha256.d.ts +0 -5
  705. package/src/utils/stable-sort-object.d.ts +0 -1
@@ -1,477 +0,0 @@
1
- --- src/gpu/kernels/fused_matmul_rmsnorm.wgsl
2
- +++ src/gpu/kernels/fused_matmul_rmsnorm_f16.wgsl
3
- @@ -1,16 +1,21 @@
4
- -// Fused GEMV + RMSNorm Kernel
5
- +// Fused GEMV + RMSNorm Kernel (F16 variant)
6
- //
7
- // For decode (M=1), combines the down projection matmul with RMSNorm in a single kernel:
8
- // 1. Compute GEMV: C[1, N] = A[1, K] × B[K, N] (down projection)
9
- // 2. Compute RMSNorm on C: output = C / sqrt(mean(C^2) + eps) * weight
10
- // 3. Optional residual: output = output + residual
11
- //
12
- +// F16 variant: All buffers are f16, accumulation in f32 for stability.
13
- +//
14
- // Benefits:
15
- // - Single GPU dispatch instead of 2
16
- // - No intermediate buffer for matmul output
17
- // - Better cache locality
18
- +// - 2x memory bandwidth vs F32 variant
19
- //
20
- -// Expected speedup: 1.2-1.5x for post-FFN normalization path
21
- +// Expected speedup: 1.3-1.8x for post-FFN normalization path
22
- +
23
- +enable f16;
24
-
25
- override WORKGROUP_SIZE: u32 = 256u;
26
- const MAX_WORKGROUP_SIZE: u32 = 256u;
27
- @@ -25,19 +30,19 @@
28
- eps: f32, // RMSNorm epsilon
29
- has_residual: u32, // 1 if residual provided, 0 otherwise
30
- transpose_b: u32, // 1 if weight is [N,K] (row-major), 0 if [K,N] (column-major)
31
- - _pad0: u32, // Padding to 24 bytes for alignment
32
- + _pad0: u32,
33
- _pad1: u32,
34
- _pad2: u32,
35
- }
36
-
37
- @group(0) @binding(0) var<uniform> u: Uniforms;
38
- -@group(0) @binding(1) var<storage, read> input: array<f32>; // [1, K] - activation from FFN
39
- -@group(0) @binding(2) var<storage, read> weight: array<f32>; // [K, N] - down projection weight
40
- +@group(0) @binding(1) var<storage, read> input: array<f16>; // [1, K] - activation from FFN
41
- +@group(0) @binding(2) var<storage, read> weight: array<f16>; // [K, N] - down projection weight
42
- @group(0) @binding(3) var<storage, read> norm_weight: array<u32>; // [N] - RMSNorm weight (F16 or F32)
43
- -@group(0) @binding(4) var<storage, read_write> output: array<f32>;// [1, N] - final output
44
- -@group(0) @binding(5) var<storage, read> residual: array<f32>; // [1, N] - optional residual
45
- -
46
- -// Shared memory for reduction
47
- +@group(0) @binding(4) var<storage, read_write> output: array<f16>;// [1, N] - final output
48
- +@group(0) @binding(5) var<storage, read> residual: array<f16>; // [1, N] - optional residual
49
- +
50
- +// Shared memory for reduction (f32 for numerical stability)
51
- var<workgroup> shared_partial: array<f32, MAX_WORKGROUP_SIZE>; // Partial dot products
52
- var<workgroup> shared_output: array<f32, MAX_WORKGROUP_SIZE>; // Output values for RMSNorm
53
- var<workgroup> shared_sum_sq: array<f32, MAX_WORKGROUP_SIZE>; // Sum of squares for reduction
54
- @@ -58,18 +63,166 @@
55
- return bitcast<f32>(norm_weight[idx]);
56
- }
57
-
58
- -// Fused GEMV + RMSNorm: Main entry point
59
- -// Each workgroup computes COLS_PER_WG output columns, then all workgroups
60
- -// cooperate to compute the RMSNorm across the full output row.
61
- -//
62
- -// Phase 1: GEMV - compute dot products for each output column
63
- -// Phase 2: Store outputs and compute partial sum of squares
64
- -// Phase 3: Global reduction of sum of squares (via atomic or multi-pass)
65
- -// Phase 4: Normalize outputs and optionally add residual
66
- -//
67
- -// Note: This kernel requires N <= WG_SIZE * COLS_PER_WG for single-pass RMSNorm.
68
- -// For larger N, use the multi-workgroup variant.
69
- -
70
- +// Optimized single-workgroup variant for small N (hidden_size <= WORKGROUP_SIZE)
71
- +@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
72
- +fn gemv_rmsnorm_small(
73
- + @builtin(local_invocation_id) local_id: vec3<u32>
74
- +) {
75
- + let tid = local_id.x;
76
- + let N = u.N;
77
- + let K = u.K;
78
- +
79
- + // Each thread computes one output column (for N <= WORKGROUP_SIZE)
80
- + var dot_sum: f32 = 0.0;
81
- + if (tid < N) {
82
- + // Unroll by 4 for better ILP
83
- + var k: u32 = 0u;
84
- + let k_aligned = (K / 4u) * 4u;
85
- +
86
- + for (; k < k_aligned; k = k + 4u) {
87
- + let a0 = f32(input[k]);
88
- + let a1 = f32(input[k + 1u]);
89
- + let a2 = f32(input[k + 2u]);
90
- + let a3 = f32(input[k + 3u]);
91
- +
92
- + if (u.transpose_b == 1u) {
93
- + // Row-major: weight[tid, k]
94
- + let base = tid * K + k;
95
- + dot_sum = dot_sum + a0 * f32(weight[base])
96
- + + a1 * f32(weight[base + 1u])
97
- + + a2 * f32(weight[base + 2u])
98
- + + a3 * f32(weight[base + 3u]);
99
- + } else {
100
- + // Column-major: weight[k, tid]
101
- + dot_sum = dot_sum + a0 * f32(weight[k * N + tid])
102
- + + a1 * f32(weight[(k + 1u) * N + tid])
103
- + + a2 * f32(weight[(k + 2u) * N + tid])
104
- + + a3 * f32(weight[(k + 3u) * N + tid]);
105
- + }
106
- + }
107
- +
108
- + // Handle remainder
109
- + for (; k < K; k = k + 1u) {
110
- + let w_idx = select(k * N + tid, tid * K + k, u.transpose_b == 1u);
111
- + dot_sum = dot_sum + f32(input[k]) * f32(weight[w_idx]);
112
- + }
113
- + }
114
- +
115
- + // Store output value and compute sum of squares
116
- + shared_output[tid] = dot_sum;
117
- + shared_sum_sq[tid] = select(0.0, dot_sum * dot_sum, tid < N);
118
- + workgroupBarrier();
119
- +
120
- + // Parallel reduction for sum of squares
121
- + for (var stride: u32 = WORKGROUP_SIZE / 2u; stride > 0u; stride = stride >> 1u) {
122
- + if (tid < stride) {
123
- + shared_sum_sq[tid] = shared_sum_sq[tid] + shared_sum_sq[tid + stride];
124
- + }
125
- + workgroupBarrier();
126
- + }
127
- +
128
- + // Compute inverse RMS
129
- + let mean_sq = shared_sum_sq[0] / f32(N);
130
- + let inv_rms = 1.0 / sqrt(mean_sq + u.eps);
131
- +
132
- + // Write normalized output
133
- + if (tid < N) {
134
- + var result = shared_output[tid] * inv_rms * apply_weight(load_norm_weight(tid));
135
- + if (u.has_residual == 1u) {
136
- + result = result + f32(residual[tid]);
137
- + }
138
- + output[tid] = f16(result);
139
- + }
140
- +}
141
- +
142
- +// Medium variant for N up to ~4096 (single workgroup, multiple elements per thread)
143
- +// Handles Gemma's hidden_size=2304 and similar models
144
- +@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
145
- +fn gemv_rmsnorm_medium(
146
- + @builtin(local_invocation_id) local_id: vec3<u32>
147
- +) {
148
- + let tid = local_id.x;
149
- + let N = u.N;
150
- + let K = u.K;
151
- +
152
- + // Each thread handles multiple output columns
153
- + let cols_per_thread = (N + WORKGROUP_SIZE - 1u) / WORKGROUP_SIZE;
154
- +
155
- + // Phase 1: Compute dot products for all columns assigned to this thread
156
- + var local_sum_sq: f32 = 0.0;
157
- +
158
- + for (var c: u32 = 0u; c < cols_per_thread; c = c + 1u) {
159
- + let col = tid + c * WORKGROUP_SIZE;
160
- + if (col < N) {
161
- + // Compute dot product for this column
162
- + var dot_sum: f32 = 0.0;
163
- +
164
- + // Unroll by 4 for better ILP
165
- + var k: u32 = 0u;
166
- + let k_aligned = (K / 4u) * 4u;
167
- +
168
- + for (; k < k_aligned; k = k + 4u) {
169
- + let a0 = f32(input[k]);
170
- + let a1 = f32(input[k + 1u]);
171
- + let a2 = f32(input[k + 2u]);
172
- + let a3 = f32(input[k + 3u]);
173
- +
174
- + if (u.transpose_b == 1u) {
175
- + let base = col * K + k;
176
- + dot_sum = dot_sum + a0 * f32(weight[base])
177
- + + a1 * f32(weight[base + 1u])
178
- + + a2 * f32(weight[base + 2u])
179
- + + a3 * f32(weight[base + 3u]);
180
- + } else {
181
- + dot_sum = dot_sum + a0 * f32(weight[k * N + col])
182
- + + a1 * f32(weight[(k + 1u) * N + col])
183
- + + a2 * f32(weight[(k + 2u) * N + col])
184
- + + a3 * f32(weight[(k + 3u) * N + col]);
185
- + }
186
- + }
187
- +
188
- + // Handle remainder
189
- + for (; k < K; k = k + 1u) {
190
- + let w_idx = select(k * N + col, col * K + k, u.transpose_b == 1u);
191
- + dot_sum = dot_sum + f32(input[k]) * f32(weight[w_idx]);
192
- + }
193
- +
194
- + // Store to output buffer (will normalize in-place later)
195
- + output[col] = f16(dot_sum);
196
- + local_sum_sq = local_sum_sq + dot_sum * dot_sum;
197
- + }
198
- + }
199
- +
200
- + // Phase 2: Parallel reduction for sum of squares
201
- + shared_sum_sq[tid] = local_sum_sq;
202
- + workgroupBarrier();
203
- +
204
- + for (var stride: u32 = WORKGROUP_SIZE / 2u; stride > 0u; stride = stride >> 1u) {
205
- + if (tid < stride) {
206
- + shared_sum_sq[tid] = shared_sum_sq[tid] + shared_sum_sq[tid + stride];
207
- + }
208
- + workgroupBarrier();
209
- + }
210
- +
211
- + // Compute inverse RMS (all threads read the same value)
212
- + let mean_sq = shared_sum_sq[0] / f32(N);
213
- + let inv_rms = 1.0 / sqrt(mean_sq + u.eps);
214
- +
215
- + // Phase 3: Normalize outputs in-place
216
- + for (var c: u32 = 0u; c < cols_per_thread; c = c + 1u) {
217
- + let col = tid + c * WORKGROUP_SIZE;
218
- + if (col < N) {
219
- + var result = f32(output[col]) * inv_rms * apply_weight(load_norm_weight(col));
220
- + if (u.has_residual == 1u) {
221
- + result = result + f32(residual[col]);
222
- + }
223
- + output[col] = f16(result);
224
- + }
225
- + }
226
- +}
227
- +
228
- +// Main entry point - dispatches based on N size
229
- +// Uses multiple workgroups for large N
230
- @compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
231
- fn main(
232
- @builtin(global_invocation_id) global_id: vec3<u32>,
233
- @@ -82,23 +235,20 @@
234
- let K = u.K;
235
-
236
- // Each workgroup handles COLS_PER_WG columns of output
237
- - let col_in_wg = tid / THREADS_PER_COL; // Which column within this workgroup (0-3)
238
- - let thread_in_col = tid % THREADS_PER_COL; // Thread index within column (0-63)
239
- - let global_col = wg_id.x * COLS_PER_WG + col_in_wg; // Global column index
240
- + let col_in_wg = tid / THREADS_PER_COL;
241
- + let thread_in_col = tid % THREADS_PER_COL;
242
- + let global_col = wg_id.x * COLS_PER_WG + col_in_wg;
243
-
244
- // Phase 1: Compute dot product for this column
245
- var dot_sum: f32 = 0.0;
246
- if (global_col < N) {
247
- - // Each thread in a column processes a stripe of K
248
- let k_per_thread = (K + THREADS_PER_COL - 1u) / THREADS_PER_COL;
249
- let k_start = thread_in_col * k_per_thread;
250
- let k_end = min(k_start + k_per_thread, K);
251
-
252
- for (var k: u32 = k_start; k < k_end; k = k + 1u) {
253
- - // transpose_b=1: weight is [N,K], access weight[global_col * K + k]
254
- - // transpose_b=0: weight is [K,N], access weight[k * N + global_col]
255
- let w_idx = select(k * N + global_col, global_col * K + k, u.transpose_b == 1u);
256
- - dot_sum = dot_sum + input[k] * weight[w_idx];
257
- + dot_sum = dot_sum + f32(input[k]) * f32(weight[w_idx]);
258
- }
259
- }
260
-
261
- @@ -106,8 +256,7 @@
262
- shared_partial[tid] = dot_sum;
263
- workgroupBarrier();
264
-
265
- - // Reduce within each column (THREADS_PER_COL threads -> 1 value)
266
- - // Tree reduction: 64 -> 32 -> 16 -> 8 -> 4 -> 2 -> 1
267
- + // Reduce within each column
268
- for (var stride: u32 = THREADS_PER_COL / 2u; stride > 0u; stride = stride >> 1u) {
269
- if (thread_in_col < stride) {
270
- let idx = col_in_wg * THREADS_PER_COL + thread_in_col;
271
- @@ -125,14 +274,7 @@
272
- }
273
- workgroupBarrier();
274
-
275
- - // Phase 2: For single-workgroup case (N <= 1024), we can do RMSNorm directly
276
- - // For multi-workgroup, we'd need a second pass or atomic accumulation
277
- -
278
- - // This version handles N <= WORKGROUP_SIZE * COLS_PER_WG
279
- - // For larger N, spawn multiple workgroups and do global reduction
280
- -
281
- // Reduce sum of squares across columns in this workgroup
282
- - // All threads must participate in barriers, so move barrier outside condition
283
- for (var stride: u32 = COLS_PER_WG / 2u; stride > 0u; stride = stride >> 1u) {
284
- if (tid < stride && tid + stride < COLS_PER_WG) {
285
- shared_sum_sq[tid] = shared_sum_sq[tid] + shared_sum_sq[tid + stride];
286
- @@ -140,185 +282,20 @@
287
- workgroupBarrier();
288
- }
289
-
290
- - // For single-workgroup: compute RMS and normalize
291
- - // Thread 0 has partial sum; need global sum across all workgroups
292
- - // For now, assume single workgroup (N <= COLS_PER_WG = 4)
293
- - // Single-workgroup kernel; selection code enforces N within limits.
294
- -
295
- let local_sum_sq = shared_sum_sq[0];
296
-
297
- // Phase 3: Write normalized output
298
- if (thread_in_col == 0u && global_col < N) {
299
- let val = shared_output[col_in_wg];
300
- -
301
- - // For multi-WG: would need global sum here
302
- - // For now: just use local (works for small N)
303
- let mean_sq = local_sum_sq / f32(min(COLS_PER_WG, N));
304
- let inv_rms = 1.0 / sqrt(mean_sq + u.eps);
305
-
306
- var result = val * inv_rms * apply_weight(load_norm_weight(global_col));
307
-
308
- if (u.has_residual == 1u) {
309
- - result = result + residual[global_col];
310
- - }
311
- -
312
- - output[global_col] = result;
313
- - }
314
- -}
315
- -
316
- -// Optimized single-workgroup variant for small N (hidden_size <= WORKGROUP_SIZE)
317
- -// All output columns computed by one workgroup, RMSNorm in single pass
318
- -@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
319
- -fn gemv_rmsnorm_small(
320
- - @builtin(local_invocation_id) local_id: vec3<u32>
321
- -) {
322
- - let tid = local_id.x;
323
- - let N = u.N;
324
- - let K = u.K;
325
- -
326
- - // Each thread computes one output column (for N <= WORKGROUP_SIZE)
327
- - var dot_sum: f32 = 0.0;
328
- - if (tid < N) {
329
- - for (var k: u32 = 0u; k < K; k = k + 1u) {
330
- - let w_idx = select(k * N + tid, tid * K + k, u.transpose_b == 1u);
331
- - dot_sum = dot_sum + input[k] * weight[w_idx];
332
- - }
333
- - }
334
- -
335
- - // Store output value and compute sum of squares
336
- - shared_output[tid] = dot_sum;
337
- - shared_sum_sq[tid] = select(0.0, dot_sum * dot_sum, tid < N);
338
- - workgroupBarrier();
339
- -
340
- - // Parallel reduction for sum of squares
341
- - for (var stride: u32 = WORKGROUP_SIZE / 2u; stride > 0u; stride = stride >> 1u) {
342
- - if (tid < stride) {
343
- - shared_sum_sq[tid] = shared_sum_sq[tid] + shared_sum_sq[tid + stride];
344
- - }
345
- - workgroupBarrier();
346
- - }
347
- -
348
- - // Compute inverse RMS
349
- - let mean_sq = shared_sum_sq[0] / f32(N);
350
- - let inv_rms = 1.0 / sqrt(mean_sq + u.eps);
351
- -
352
- - // Write normalized output
353
- - if (tid < N) {
354
- - var result = shared_output[tid] * inv_rms * apply_weight(load_norm_weight(tid));
355
- - if (u.has_residual == 1u) {
356
- - result = result + residual[tid];
357
- - }
358
- - output[tid] = result;
359
- - }
360
- -}
361
- -
362
- -// Medium variant for N up to ~4096 (single workgroup, multiple elements per thread)
363
- -// Handles Gemma 3's hidden_size=1152 and similar models
364
- -// Each thread computes ceil(N/WORKGROUP_SIZE) output columns sequentially
365
- -@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
366
- -fn gemv_rmsnorm_medium(
367
- - @builtin(local_invocation_id) local_id: vec3<u32>
368
- -) {
369
- - let tid = local_id.x;
370
- - let N = u.N;
371
- - let K = u.K;
372
- -
373
- - // Each thread handles multiple output columns
374
- - let cols_per_thread = (N + WORKGROUP_SIZE - 1u) / WORKGROUP_SIZE;
375
- -
376
- - // Phase 1: Compute dot products for all columns assigned to this thread
377
- - // Store partial sum of squares as we go
378
- - var local_sum_sq: f32 = 0.0;
379
- -
380
- - for (var c: u32 = 0u; c < cols_per_thread; c = c + 1u) {
381
- - let col = tid + c * WORKGROUP_SIZE;
382
- - if (col < N) {
383
- - // Compute dot product for this column
384
- - var dot_sum: f32 = 0.0;
385
- - for (var k: u32 = 0u; k < K; k = k + 1u) {
386
- - let w_idx = select(k * N + col, col * K + k, u.transpose_b == 1u);
387
- - dot_sum = dot_sum + input[k] * weight[w_idx];
388
- - }
389
- -
390
- - // Store to output buffer (will normalize in-place later)
391
- - output[col] = dot_sum;
392
- - local_sum_sq = local_sum_sq + dot_sum * dot_sum;
393
- - }
394
- - }
395
- -
396
- - // Phase 2: Parallel reduction for sum of squares
397
- - shared_sum_sq[tid] = local_sum_sq;
398
- - workgroupBarrier();
399
- -
400
- - for (var stride: u32 = WORKGROUP_SIZE / 2u; stride > 0u; stride = stride >> 1u) {
401
- - if (tid < stride) {
402
- - shared_sum_sq[tid] = shared_sum_sq[tid] + shared_sum_sq[tid + stride];
403
- - }
404
- - workgroupBarrier();
405
- - }
406
- -
407
- - // Compute inverse RMS (all threads read the same value)
408
- - let mean_sq = shared_sum_sq[0] / f32(N);
409
- - let inv_rms = 1.0 / sqrt(mean_sq + u.eps);
410
- -
411
- - // Phase 3: Normalize outputs in-place
412
- - for (var c: u32 = 0u; c < cols_per_thread; c = c + 1u) {
413
- - let col = tid + c * WORKGROUP_SIZE;
414
- - if (col < N) {
415
- - var result = output[col] * inv_rms * apply_weight(load_norm_weight(col));
416
- - if (u.has_residual == 1u) {
417
- - result = result + residual[col];
418
- - }
419
- - output[col] = result;
420
- - }
421
- - }
422
- -}
423
- -
424
- -// Multi-workgroup variant using two-phase approach
425
- -// Phase 1: Each workgroup computes output columns and partial sum of squares
426
- -// Phase 2: Separate kernel reduces partials and normalizes
427
- -// This entry point is Phase 1 only
428
- -@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
429
- -fn gemv_rmsnorm_phase1(
430
- - @builtin(global_invocation_id) global_id: vec3<u32>,
431
- - @builtin(local_invocation_id) local_id: vec3<u32>,
432
- - @builtin(workgroup_id) wg_id: vec3<u32>
433
- -) {
434
- - let tid = local_id.x;
435
- - let N = u.N;
436
- - let K = u.K;
437
- -
438
- - let col_in_wg = tid / THREADS_PER_COL;
439
- - let thread_in_col = tid % THREADS_PER_COL;
440
- - let global_col = wg_id.x * COLS_PER_WG + col_in_wg;
441
- -
442
- - // Compute dot product
443
- - var dot_sum: f32 = 0.0;
444
- - if (global_col < N) {
445
- - let k_per_thread = (K + THREADS_PER_COL - 1u) / THREADS_PER_COL;
446
- - let k_start = thread_in_col * k_per_thread;
447
- - let k_end = min(k_start + k_per_thread, K);
448
- -
449
- - for (var k: u32 = k_start; k < k_end; k = k + 1u) {
450
- - let w_idx = select(k * N + global_col, global_col * K + k, u.transpose_b == 1u);
451
- - dot_sum = dot_sum + input[k] * weight[w_idx];
452
- - }
453
- - }
454
- -
455
- - shared_partial[tid] = dot_sum;
456
- - workgroupBarrier();
457
- -
458
- - // Reduce within column
459
- - for (var stride: u32 = THREADS_PER_COL / 2u; stride > 0u; stride = stride >> 1u) {
460
- - if (thread_in_col < stride) {
461
- - let idx = col_in_wg * THREADS_PER_COL + thread_in_col;
462
- - shared_partial[idx] = shared_partial[idx] + shared_partial[idx + stride];
463
- - }
464
- - workgroupBarrier();
465
- - }
466
- -
467
- - // Write matmul output (RMSNorm will be done in phase 2)
468
- - if (thread_in_col == 0u && global_col < N) {
469
- - output[global_col] = shared_partial[col_in_wg * THREADS_PER_COL];
470
- - }
471
- -}
472
- + result = result + f32(residual[global_col]);
473
- + }
474
- +
475
- + output[global_col] = f16(result);
476
- + }
477
- +}
@@ -1,60 +0,0 @@
1
- --- src/gpu/kernels/fused_swiglu.wgsl
2
- +++ src/gpu/kernels/fused_swiglu_f16.wgsl
3
- @@ -1,30 +1,24 @@
4
- -// fused_swiglu.wgsl
5
- +// fused_swiglu_f16.wgsl
6
-
7
- /**
8
- - * SwiGLU Activation Kernel (Row-Split Fused Gate/Up)
9
- - *
10
- - * Computes: output = SiLU(gate + gate_bias) * (up + up_bias)
11
- - *
12
- - * input layout: [num_tokens, 2 * dim] flattened, row-major
13
- - * row = [gate[0..dim), up[0..dim)]
14
- - * bias layout: [2 * dim]
15
- - * bias = [gate_bias[0..dim), up_bias[0..dim)]
16
- - * output layout: [num_tokens, dim]
17
- + * SwiGLU Activation Kernel (Row-Split Fused Gate/Up, f16 IO)
18
- */
19
- +
20
- +enable f16;
21
-
22
- override WORKGROUP_SIZE: u32 = 256u;
23
-
24
- struct Uniforms {
25
- num_tokens: u32,
26
- dim: u32,
27
- - bias_offset: u32, // byte offset into bias buffer (divide by 4 for F32 index)
28
- - clamp_max: f32, // SwiGLU clamp (0 = disabled)
29
- + bias_offset: u32,
30
- + clamp_max: f32,
31
- }
32
-
33
- @group(0) @binding(0) var<uniform> u: Uniforms;
34
- -@group(0) @binding(1) var<storage, read> input: array<f32>;
35
- -@group(0) @binding(2) var<storage, read> bias: array<f32>;
36
- -@group(0) @binding(3) var<storage, read_write> output: array<f32>;
37
- +@group(0) @binding(1) var<storage, read> input: array<f16>;
38
- +@group(0) @binding(2) var<storage, read> bias: array<f16>;
39
- +@group(0) @binding(3) var<storage, read_write> output: array<f16>;
40
-
41
- fn sigmoid(x: f32) -> f32 {
42
- return 1.0 / (1.0 + exp(-x));
43
- @@ -51,13 +45,11 @@
44
-
45
- let token_idx = idx / u.dim;
46
- let dim_idx = idx % u.dim;
47
- -
48
- - // bias_offset is in bytes, convert to F32 index
49
- - let bias_base = u.bias_offset / 4u;
50
- + let bias_base = u.bias_offset / 2u;
51
-
52
- let row_base = token_idx * u.dim * 2u;
53
- - let gate = input[row_base + dim_idx] + bias[bias_base + dim_idx];
54
- - let up = input[row_base + u.dim + dim_idx] + bias[bias_base + u.dim + dim_idx];
55
- + let gate = f32(input[row_base + dim_idx]) + f32(bias[bias_base + dim_idx]);
56
- + let up = f32(input[row_base + u.dim + dim_idx]) + f32(bias[bias_base + u.dim + dim_idx]);
57
-
58
- - output[idx] = clamp_swiglu(silu(gate) * up);
59
- + output[idx] = f16(clamp_swiglu(silu(gate) * up));
60
- }
@@ -1,41 +0,0 @@
1
- --- src/gpu/kernels/gelu.wgsl
2
- +++ src/gpu/kernels/gelu_f16.wgsl
3
- @@ -1,9 +1,9 @@
4
- -// GeLU Activation Kernel
5
- +// GeLU Activation Kernel with F16 Input/Output
6
- //
7
- -// GeLU(x) ≈ 0.5 * x * (1 + tanh(sqrt(2/π) * (x + 0.044715 * x^3)))
8
- -// Variants:
9
- -// - geglu: GeLU(gate) * up (separate gate buffer)
10
- -// - geglu_rowsplit: input = [numTokens, 2*dim] row-split
11
- +// F16 variant for reduced memory bandwidth when using F16 activations.
12
- +
13
- +enable f16;
14
-
15
- override WORKGROUP_SIZE: u32 = 256u;
16
- override HAS_GATE: bool = false;
17
- @@ -17,9 +17,9 @@
18
- }
19
-
20
- @group(0) @binding(0) var<uniform> u: Uniforms;
21
- -@group(0) @binding(1) var<storage, read> input: array<f32>;
22
- -@group(0) @binding(2) var<storage, read_write> output: array<f32>;
23
- -@group(0) @binding(3) var<storage, read> gate: array<f32>;
24
- +@group(0) @binding(1) var<storage, read> input: array<f16>;
25
- +@group(0) @binding(2) var<storage, read_write> output: array<f16>;
26
- +@group(0) @binding(3) var<storage, read> gate: array<f16>;
27
-
28
- -fn gelu(x: f32) -> f32 {
29
- - let sqrt_2_over_pi: f32 = 0.7978845608;
30
- - let c: f32 = 0.044715;
31
- +fn gelu(x: f16) -> f16 {
32
- + let sqrt_2_over_pi = f16(0.7978845608);
33
- + let c = f16(0.044715);
34
- let inner = sqrt_2_over_pi * (x + c * x * x * x);
35
- - let inner_clamped = clamp(inner, -15.0, 15.0);
36
- - return 0.5 * x * (1.0 + tanh(inner_clamped));
37
- + let inner_clamped = clamp(inner, f16(-15.0), f16(15.0));
38
- + return f16(0.5) * x * (f16(1.0) + tanh(inner_clamped));
39
- }
40
-
41
- @compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
@@ -1,82 +0,0 @@
1
- --- src/gpu/kernels/layernorm.wgsl
2
- +++ src/gpu/kernels/layernorm_f16.wgsl
3
- @@ -1,20 +1,12 @@
4
- -// LayerNorm Kernel
5
- +// LayerNorm Kernel (F16 Input/Output)
6
- //
7
- // LayerNorm(x) = (x - mean) / sqrt(var + eps) * weight + bias
8
- -//
9
- -// Uses workgroup reduction for mean and variance per token.
10
-
11
- -// =============================================================================
12
- -// Override Constants
13
- -// =============================================================================
14
- +enable f16;
15
-
16
- override WORKGROUP_SIZE: u32 = 256u;
17
- const MAX_WORKGROUP_SIZE: u32 = 256u;
18
- override PARAMS_IS_F16: bool = false; // Weight/bias packed as f16 pairs
19
- -
20
- -// =============================================================================
21
- -// Uniforms
22
- -// =============================================================================
23
-
24
- struct Uniforms {
25
- size: u32,
26
- @@ -24,21 +16,13 @@
27
- }
28
-
29
- @group(0) @binding(0) var<uniform> u: Uniforms;
30
- -@group(0) @binding(1) var<storage, read> input: array<f32>;
31
- +@group(0) @binding(1) var<storage, read> input: array<f16>;
32
- @group(0) @binding(2) var<storage, read> weight: array<u32>;
33
- @group(0) @binding(3) var<storage, read> bias: array<u32>;
34
- -@group(0) @binding(4) var<storage, read_write> output: array<f32>;
35
- -
36
- -// =============================================================================
37
- -// Shared Memory
38
- -// =============================================================================
39
- +@group(0) @binding(4) var<storage, read_write> output: array<f16>;
40
-
41
- var<workgroup> shared_sum: array<f32, MAX_WORKGROUP_SIZE>;
42
- var<workgroup> shared_sum_sq: array<f32, MAX_WORKGROUP_SIZE>;
43
- -
44
- -// =============================================================================
45
- -// Helpers
46
- -// =============================================================================
47
-
48
- fn load_weight(idx: u32) -> f32 {
49
- if (PARAMS_IS_F16) {
50
- @@ -57,10 +41,6 @@
51
- }
52
- return bitcast<f32>(bias[idx]);
53
- }
54
- -
55
- -// =============================================================================
56
- -// Entry Point
57
- -// =============================================================================
58
-
59
- @compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
60
- fn main(
61
- @@ -84,7 +64,7 @@
62
- for (var i: u32 = 0u; i < elements_per_thread; i = i + 1u) {
63
- let idx = thread_idx * elements_per_thread + i;
64
- if (idx < size) {
65
- - let x = input[base_offset + idx];
66
- + let x = f32(input[base_offset + idx]);
67
- local_sum = local_sum + x;
68
- local_sum_sq = local_sum_sq + x * x;
69
- }
70
- @@ -112,10 +92,10 @@
71
- for (var i: u32 = 0u; i < elements_per_thread; i = i + 1u) {
72
- let idx = thread_idx * elements_per_thread + i;
73
- if (idx < size) {
74
- - let x = input[base_offset + idx];
75
- + let x = f32(input[base_offset + idx]);
76
- let norm = (x - mean) * inv_std;
77
- let y = norm * load_weight(idx) + load_bias(idx);
78
- - output[base_offset + idx] = y;
79
- + output[base_offset + idx] = f16(y);
80
- }
81
- }
82
- }