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
@@ -9,13 +9,14 @@
9
9
  import type { CpuWeightBuffer, WeightBuffer } from '../gpu/weight-buffer.js';
10
10
  import type { TensorRole } from '../config/schema/index.js';
11
11
  import type { TensorSourceTransform } from '../formats/rdrr/index.js';
12
+ import type { FunctionalDescriptorManifest } from '../formats/rdrr/functional-descriptor.js';
12
13
 
13
14
  /**
14
15
  * Tensor location in loaded model
15
16
  */
16
17
  export interface TensorLocation {
17
- shardIndex: number;
18
- offset: number;
18
+ shardIndex?: number;
19
+ offset?: number;
19
20
  size: number;
20
21
  shape: number[];
21
22
  dtype: string;
@@ -28,6 +29,8 @@ export interface TensorLocation {
28
29
  originalShape?: number[];
29
30
  /** Optional direct-source transform applied before dtype-specific loading */
30
31
  sourceTransform?: TensorSourceTransform;
32
+ /** Parsed manifoldgguf descriptor for FUNCTIONAL_DESCRIPTOR tensors */
33
+ descriptorManifest?: FunctionalDescriptorManifest;
31
34
  }
32
35
 
33
36
  /**
@@ -53,7 +56,10 @@ export interface LayerWeights {
53
56
  kNorm: GPUBuffer | Float32Array | null;
54
57
  postAttentionNorm: GPUBuffer | Float32Array | null;
55
58
  preFeedforwardNorm: GPUBuffer | Float32Array | null;
59
+ preFeedforwardNorm2?: GPUBuffer | Float32Array | null;
56
60
  postFeedforwardNorm: GPUBuffer | Float32Array | null;
61
+ postFeedforwardNorm1?: GPUBuffer | Float32Array | null;
62
+ postFeedforwardNorm2?: GPUBuffer | Float32Array | null;
57
63
  postNorm: GPUBuffer | Float32Array | null;
58
64
  postAttnNorm: GPUBuffer | Float32Array | null;
59
65
  convInProj?: GPUBuffer | WeightBuffer | Float32Array | null;
@@ -72,6 +78,8 @@ export interface LayerWeights {
72
78
  gateUp?: GPUBuffer | WeightBuffer | Float32Array | null;
73
79
  routerWeight?: GPUBuffer | import('../gpu/weight-buffer.js').WeightBuffer | Float32Array | null;
74
80
  routerBias?: GPUBuffer | Float32Array | null;
81
+ routerScale?: GPUBuffer | Float32Array | null;
82
+ routerPerExpertScale?: GPUBuffer | Float32Array | null;
75
83
  attentionSinks?: GPUBuffer | Float32Array | null;
76
84
  perLayerInputGate?: GPUBuffer | WeightBuffer | Float32Array | null;
77
85
  perLayerProjection?: GPUBuffer | WeightBuffer | Float32Array | null;
@@ -79,6 +87,14 @@ export interface LayerWeights {
79
87
  layerScalar?: GPUBuffer | Float32Array | null;
80
88
  }
81
89
 
90
+ export interface DiffusionGemmaSelfConditioningWeights {
91
+ preNorm: GPUBuffer | Float32Array;
92
+ postNorm: GPUBuffer | Float32Array | null;
93
+ gateProj: GPUBuffer | WeightBuffer | Float32Array;
94
+ upProj: GPUBuffer | WeightBuffer | Float32Array;
95
+ downProj: GPUBuffer | WeightBuffer | Float32Array;
96
+ }
97
+
82
98
  export interface PerLayerInputWeights {
83
99
  embedTokensPerLayer: GPUBuffer | WeightBuffer | CpuWeightBuffer | Float32Array | null;
84
100
  embedTokensPerLayerSplit?: (GPUBuffer | WeightBuffer | CpuWeightBuffer | Float32Array | null)[] | null;
@@ -120,6 +136,41 @@ export interface LoadOptions {
120
136
  verifyHashes: boolean;
121
137
  }
122
138
 
139
+ export type LoaderLoadTimingPhase =
140
+ | 'preflight'
141
+ | 'tensorLocations'
142
+ | 'embeddings'
143
+ | 'layers'
144
+ | 'finalWeights'
145
+ | 'cleanup';
146
+
147
+ export interface LoaderLayerLoadTiming {
148
+ count: number | null;
149
+ totalMs: number | null;
150
+ meanMs: number | null;
151
+ maxMs: number | null;
152
+ maxLayer: number | null;
153
+ }
154
+
155
+ export interface LoaderLoadTiming {
156
+ schemaVersion: 1;
157
+ source: 'doppler-loader';
158
+ modelId: string | null;
159
+ status: 'running' | 'complete' | 'failed';
160
+ customShardLoader: boolean;
161
+ byteAccountingMode: 'full-shard-progress' | 'custom-loader-progress-unavailable';
162
+ totalBytes: number | null;
163
+ totalShards: number | null;
164
+ bytesLoaded: number;
165
+ shardsLoaded: number;
166
+ bytesPerSecond: number | null;
167
+ phasesMs: Record<LoaderLoadTimingPhase, number | null>;
168
+ layers: LoaderLayerLoadTiming;
169
+ totalMs: number | null;
170
+ failedPhase: LoaderLoadTimingPhase | null;
171
+ error: string | null;
172
+ }
173
+
123
174
  /**
124
175
  * Shard load priority.
125
176
  */
@@ -139,6 +190,7 @@ export interface CustomShardLoaderOptions {
139
190
  verify?: boolean;
140
191
  loadShardRange?: CustomShardRangeLoader;
141
192
  streamShardRange?: CustomShardStreamLoader;
193
+ loadAuxiliaryFile?: CustomAuxiliaryFileLoader | null;
142
194
  }
143
195
 
144
196
  /**
@@ -174,6 +226,10 @@ export type CustomShardStreamLoader = (
174
226
  options?: CustomShardStreamOptions
175
227
  ) => AsyncIterable<Uint8Array>;
176
228
 
229
+ export type CustomAuxiliaryFileLoader = (
230
+ path: string
231
+ ) => Promise<ArrayBuffer | Uint8Array | null | undefined>;
232
+
177
233
  /**
178
234
  * Loader statistics
179
235
  */
@@ -185,6 +241,7 @@ export interface LoaderStats {
185
241
  layersLoaded: number;
186
242
  expertsLoaded: number;
187
243
  gpuBuffers: number;
244
+ loadTiming?: LoaderLoadTiming | null;
188
245
  }
189
246
 
190
247
  /**
@@ -37,6 +37,14 @@ export declare function getLargeWeightConfig(): LargeWeightConfigSchema;
37
37
  */
38
38
  export declare function getLargeWeightMaxBytes(): number | null;
39
39
 
40
+ /**
41
+ * Runtime large-weight overrides replace manifest overrides when present.
42
+ */
43
+ export declare function resolveLargeWeightOverrides(
44
+ manifestOverrides: string[] | null | undefined,
45
+ runtimeOverrides: string[] | null | undefined
46
+ ): string[] | null;
47
+
40
48
  /**
41
49
  * Estimate GPU memory required for a matmul weight after dequantization.
42
50
  *
@@ -1,5 +1,10 @@
1
1
  import { loadTensorsFromStore } from '../storage/shard-manager.js';
2
2
  import { parseTensorMap } from '../formats/rdrr/index.js';
3
+ import {
4
+ assertFunctionalDescriptorManifest,
5
+ getFunctionalDescriptorManifest,
6
+ isFunctionalDescriptorDtype,
7
+ } from '../formats/rdrr/functional-descriptor.js';
3
8
  import { log, trace as debugTrace } from '../debug/index.js';
4
9
 
5
10
  function normalizeLocationSpans(spans, name, sourceLabel) {
@@ -24,6 +29,22 @@ function normalizeLocationSpans(spans, name, sourceLabel) {
24
29
  });
25
30
  }
26
31
 
32
+ function normalizeDescriptorManifest(info, name, sourceLabel) {
33
+ const descriptorManifest = getFunctionalDescriptorManifest(info);
34
+ if (descriptorManifest == null) {
35
+ return null;
36
+ }
37
+ return assertFunctionalDescriptorManifest(
38
+ descriptorManifest,
39
+ `Tensor "${name}" ${sourceLabel} descriptorManifest`
40
+ );
41
+ }
42
+
43
+ function normalizeShardIndex(info, fallback = undefined) {
44
+ const shardIndex = info?.shardIndex ?? info?.shard;
45
+ return typeof shardIndex === 'number' ? shardIndex : fallback;
46
+ }
47
+
27
48
  /**
28
49
  * Builds the tensor name-to-location map from the manifest.
29
50
  *
@@ -101,10 +122,14 @@ export async function buildTensorLocations(manifest, options = {}) {
101
122
  if (!info.role) {
102
123
  throw new Error(`Tensor "${name}" missing role in tensors.json`);
103
124
  }
125
+ const descriptorManifest = normalizeDescriptorManifest(info, name, 'tensors.json');
126
+ if (isFunctionalDescriptorDtype(info.dtype) && descriptorManifest == null) {
127
+ throw new Error(`Tensor "${name}" missing descriptorManifest in tensors.json`);
128
+ }
104
129
  locations.set(name, {
105
- shardIndex: info.shardIndex ?? info.shard,
130
+ shardIndex: normalizeShardIndex(info),
106
131
  offset: info.offset,
107
- size: info.size,
132
+ size: info.size ?? 0,
108
133
  shape: info.shape,
109
134
  dtype: info.dtype,
110
135
  role: info.role,
@@ -112,7 +137,9 @@ export async function buildTensorLocations(manifest, options = {}) {
112
137
  spans: normalizeLocationSpans(info.spans, name, 'tensors.json'),
113
138
  layout: info.layout,
114
139
  originalShape: info.originalShape,
140
+ storage: info.storage,
115
141
  sourceTransform: info.sourceTransform,
142
+ descriptorManifest,
116
143
  });
117
144
  }
118
145
  const resolvedSource = trace.find((entry) => entry.outcome === 'resolved')?.source ?? 'unknown';
@@ -141,10 +168,14 @@ export async function buildTensorLocations(manifest, options = {}) {
141
168
  if (!tensorInfo.role) {
142
169
  throw new Error(`Tensor "${name}" missing role in manifest.tensors`);
143
170
  }
171
+ const descriptorManifest = normalizeDescriptorManifest(tensorInfo, name, 'manifest.tensors');
172
+ if (isFunctionalDescriptorDtype(tensorInfo.dtype) && descriptorManifest == null) {
173
+ throw new Error(`Tensor "${name}" missing descriptorManifest in manifest.tensors`);
174
+ }
144
175
  locations.set(name, {
145
- shardIndex: tensorInfo.shardIndex ?? tensorInfo.shard ?? 0,
176
+ shardIndex: normalizeShardIndex(tensorInfo, isFunctionalDescriptorDtype(tensorInfo.dtype) ? undefined : 0),
146
177
  offset: tensorInfo.offset,
147
- size: tensorInfo.size,
178
+ size: tensorInfo.size ?? 0,
148
179
  shape: tensorInfo.shape,
149
180
  dtype: tensorInfo.dtype,
150
181
  role: tensorInfo.role,
@@ -152,7 +183,9 @@ export async function buildTensorLocations(manifest, options = {}) {
152
183
  spans: normalizeLocationSpans(tensorInfo.spans, name, 'manifest.tensors'),
153
184
  layout: tensorInfo.layout,
154
185
  originalShape: tensorInfo.originalShape,
186
+ storage: tensorInfo.storage,
155
187
  sourceTransform: tensorInfo.sourceTransform,
188
+ descriptorManifest,
156
189
  });
157
190
  }
158
191
  debugTrace.loader(`Tensor map: ${locations.size} tensors (inline)`);
@@ -0,0 +1,9 @@
1
+ import type { TensorLocation } from '../loader-types.js';
2
+ import type { TensorLoadConfig, TensorLoadResult } from './tensor-loader.js';
3
+
4
+ export declare function loadFunctionalDescriptor(
5
+ shardData: Uint8Array & { descriptorShards?: Map<string, Uint8Array> },
6
+ location: TensorLocation,
7
+ name: string,
8
+ config?: TensorLoadConfig
9
+ ): Promise<TensorLoadResult>;
@@ -0,0 +1,377 @@
1
+ import { acquireBuffer, uploadData } from '../../memory/buffer-pool.js';
2
+ import { createWeightBuffer } from '../../gpu/weight-buffer.js';
3
+ import { assertFunctionalDescriptorManifest } from '../../formats/rdrr/functional-descriptor.js';
4
+ import { log } from '../../debug/index.js';
5
+
6
+ // Reconstruct a functional_descriptor tensor from its component shards and
7
+ // upload the result to a GPU weight buffer. All reconstruction runs on CPU;
8
+ // the result is a row-major F32 (or F16 when hardware supports it) buffer.
9
+ //
10
+ // shardData must carry descriptorShards: Map<string, Uint8Array> keyed by the
11
+ // shard_file names declared in the descriptor manifest (kron, siren, sparse).
12
+ // location must carry descriptorManifest: the parsed manifoldgguf.v0.1 object.
13
+
14
+ // ============================================================================
15
+ // Coord-deterministic PRNG — must match tools/manifoldgguf-harness.js exactly.
16
+ // Algorithm identifier: coord_hash_normal_v1
17
+ // ============================================================================
18
+
19
+ function splitmix64(x) {
20
+ x = BigInt.asUintN(64, x + 0x9E3779B97F4A7C15n);
21
+ x = BigInt.asUintN(64, (x ^ (x >> 30n)) * 0xBF58476D1CE4E5B9n);
22
+ x = BigInt.asUintN(64, (x ^ (x >> 27n)) * 0x94D049BB133111EBn);
23
+ return x ^ (x >> 31n);
24
+ }
25
+
26
+ function coordUniform(seed, row, col) {
27
+ const s = BigInt(seed);
28
+ const r = BigInt(row);
29
+ const c = BigInt(col);
30
+ const z = splitmix64(
31
+ BigInt.asUintN(64, s ^ BigInt.asUintN(64, r * 0x9E3779B97F4A7C15n) ^ BigInt.asUintN(64, c * 0x6C62272E07BB0142n))
32
+ );
33
+ return Number(BigInt.asUintN(53, z)) / Number(1n << 53n);
34
+ }
35
+
36
+ function coordHashNormal(seed, row, col) {
37
+ const u1 = Math.max(coordUniform(seed, row, col), 1e-10);
38
+ const u2 = coordUniform(seed + 1, row, col);
39
+ return Math.sqrt(-2 * Math.log(u1)) * Math.cos(2 * Math.PI * u2);
40
+ }
41
+
42
+ // ============================================================================
43
+ // PRNG component
44
+ // ============================================================================
45
+
46
+ function buildPRNG(prngSpec, rows, cols) {
47
+ if (prngSpec.algorithm !== 'coord_hash_normal_v1') {
48
+ throw new Error(
49
+ `[FunctionalDescriptor] Unknown PRNG algorithm "${prngSpec.algorithm}". ` +
50
+ 'Only coord_hash_normal_v1 is supported in this runtime version.'
51
+ );
52
+ }
53
+ const { seed, learned_scale: scale } = prngSpec;
54
+ const out = new Float32Array(rows * cols);
55
+ for (let r = 0; r < rows; r++) {
56
+ for (let c = 0; c < cols; c++) {
57
+ out[r * cols + c] = scale * coordHashNormal(seed, r, c);
58
+ }
59
+ }
60
+ return out;
61
+ }
62
+
63
+ // ============================================================================
64
+ // Kronecker component
65
+ // Binary layout: [rank u32][a,b,c,d u32x4][A f32...][B f32...] per term
66
+ // ============================================================================
67
+
68
+ function blockUnpermute(P, a, b, c, d) {
69
+ const rows = a * b;
70
+ const cols = c * d;
71
+ const W = new Float32Array(rows * cols);
72
+ for (let ra = 0; ra < a; ra++) {
73
+ for (let rb = 0; rb < b; rb++) {
74
+ const i = ra * b + rb;
75
+ for (let cc = 0; cc < c; cc++) {
76
+ for (let cd = 0; cd < d; cd++) {
77
+ const j = cc * d + cd;
78
+ W[i * cols + j] = P[(ra * c + cc) * (b * d) + (rb * d + cd)];
79
+ }
80
+ }
81
+ }
82
+ }
83
+ return W;
84
+ }
85
+
86
+ function deserializeKronecker(buf) {
87
+ const view = new DataView(buf.buffer, buf.byteOffset, buf.byteLength);
88
+ const rank = view.getUint32(0, true);
89
+ let off = 4;
90
+ const factors = [];
91
+ for (let r = 0; r < rank; r++) {
92
+ const a = view.getUint32(off, true); off += 4;
93
+ const b = view.getUint32(off, true); off += 4;
94
+ const c = view.getUint32(off, true); off += 4;
95
+ const d = view.getUint32(off, true); off += 4;
96
+ const aLen = a * c;
97
+ const bLen = b * d;
98
+ const A = new Float32Array(buf.buffer, buf.byteOffset + off, aLen); off += aLen * 4;
99
+ const B = new Float32Array(buf.buffer, buf.byteOffset + off, bLen); off += bLen * 4;
100
+ factors.push({ a, b, c, d, A, B });
101
+ }
102
+ return factors;
103
+ }
104
+
105
+ function reconstructKronecker(factors, rows, cols) {
106
+ if (factors.length === 0) return new Float32Array(rows * cols);
107
+ const { a, b, c, d } = factors[0];
108
+ const p = a * c;
109
+ const q = b * d;
110
+ const P = new Float32Array(p * q);
111
+ for (const { A, B } of factors) {
112
+ for (let i = 0; i < p; i++) {
113
+ for (let j = 0; j < q; j++) {
114
+ P[i * q + j] += A[i] * B[j];
115
+ }
116
+ }
117
+ }
118
+ return blockUnpermute(P, a, b, c, d);
119
+ }
120
+
121
+ // ============================================================================
122
+ // SIREN component
123
+ // Binary layout: [depth u32][inDim,outDim u32x2][W f32...][b f32...] per layer
124
+ // ============================================================================
125
+
126
+ function deserializeSIREN(buf) {
127
+ const view = new DataView(buf.buffer, buf.byteOffset, buf.byteLength);
128
+ const depth = view.getUint32(0, true);
129
+ let off = 4;
130
+ const layers = [];
131
+ for (let l = 0; l < depth; l++) {
132
+ const inDim = view.getUint32(off, true); off += 4;
133
+ const outDim = view.getUint32(off, true); off += 4;
134
+ const W = new Float32Array(buf.buffer, buf.byteOffset + off, inDim * outDim); off += inDim * outDim * 4;
135
+ const b = new Float32Array(buf.buffer, buf.byteOffset + off, outDim); off += outDim * 4;
136
+ const isLast = l === depth - 1;
137
+ layers.push({ inDim, outDim, W, b, omega: l === 0 ? 30.0 : 1.0, isLast });
138
+ }
139
+ return layers;
140
+ }
141
+
142
+ function sirenForward(layers, rn, cn) {
143
+ let x = [rn, cn];
144
+ for (const { W, b, inDim, outDim, omega, isLast } of layers) {
145
+ const next = new Array(outDim);
146
+ for (let o = 0; o < outDim; o++) {
147
+ let s = b[o];
148
+ for (let i = 0; i < inDim; i++) s += W[o * inDim + i] * x[i];
149
+ next[o] = isLast ? s : Math.sin(omega * s);
150
+ }
151
+ x = next;
152
+ }
153
+ return x[0];
154
+ }
155
+
156
+ function reconstructSIREN(layers, rows, cols) {
157
+ const out = new Float32Array(rows * cols);
158
+ for (let r = 0; r < rows; r++) {
159
+ const rn = rows > 1 ? (r / (rows - 1)) * 2 - 1 : 0;
160
+ for (let c = 0; c < cols; c++) {
161
+ const cn = cols > 1 ? (c / (cols - 1)) * 2 - 1 : 0;
162
+ out[r * cols + c] = sirenForward(layers, rn, cn);
163
+ }
164
+ }
165
+ return out;
166
+ }
167
+
168
+ function cropMatrix(matrix, rows, cols, cropRows, cropCols) {
169
+ if (rows === cropRows && cols === cropCols) {
170
+ return matrix;
171
+ }
172
+ const out = new Float32Array(cropRows * cropCols);
173
+ for (let r = 0; r < cropRows; r += 1) {
174
+ out.set(
175
+ matrix.subarray(r * cols, (r * cols) + cropCols),
176
+ r * cropCols
177
+ );
178
+ }
179
+ return out;
180
+ }
181
+
182
+ function normalizeShape2(value, label) {
183
+ if (!Array.isArray(value) || value.length !== 2) {
184
+ throw new Error(`[FunctionalDescriptor] ${label} must be [rows, cols].`);
185
+ }
186
+ const rows = Number(value[0]);
187
+ const cols = Number(value[1]);
188
+ if (!Number.isInteger(rows) || rows <= 0 || !Number.isInteger(cols) || cols <= 0) {
189
+ throw new Error(`[FunctionalDescriptor] ${label} has invalid shape [${value.join(',')}].`);
190
+ }
191
+ return [rows, cols];
192
+ }
193
+
194
+ function getDescriptorShardBytes(shards) {
195
+ let total = 0;
196
+ for (const bytes of shards.values()) {
197
+ total += bytes.byteLength;
198
+ }
199
+ return total;
200
+ }
201
+
202
+ function assertOptionalIntegerEquals(value, expected, label, name) {
203
+ if (value === undefined) {
204
+ return;
205
+ }
206
+ const parsed = Number(value);
207
+ if (!Number.isInteger(parsed) || parsed !== expected) {
208
+ throw new Error(
209
+ `[FunctionalDescriptor] ${label} for "${name}" must equal ${expected}, got ${value}.`
210
+ );
211
+ }
212
+ }
213
+
214
+ function assertRuntimeProofGates(manifest, name, descriptorBytes, denseF16Bytes) {
215
+ assertOptionalIntegerEquals(manifest.descriptor_bytes, descriptorBytes, 'descriptor_bytes', name);
216
+ assertOptionalIntegerEquals(manifest.dense_f16_bytes, denseF16Bytes, 'dense_f16_bytes', name);
217
+
218
+ const gate = manifest.proof_status_gate;
219
+ const proofStatus = typeof manifest.proof_status === 'string'
220
+ ? manifest.proof_status.trim().toLowerCase()
221
+ : null;
222
+ const sensitivityGate = typeof gate?.sensitivity === 'string'
223
+ ? gate.sensitivity.trim().toLowerCase()
224
+ : null;
225
+ const compressionGate = typeof gate?.compression === 'string'
226
+ ? gate.compression.trim().toLowerCase()
227
+ : null;
228
+
229
+ if (proofStatus === 'passed' && sensitivityGate !== 'passed') {
230
+ throw new Error(
231
+ `[FunctionalDescriptor] proof_status is "passed" for "${name}" but proof_status_gate.sensitivity is not "passed".`
232
+ );
233
+ }
234
+ if (proofStatus === 'passed' && compressionGate !== 'passed') {
235
+ throw new Error(
236
+ `[FunctionalDescriptor] proof_status is "passed" for "${name}" but proof_status_gate.compression is not "passed".`
237
+ );
238
+ }
239
+ if (compressionGate === 'passed' && descriptorBytes >= denseF16Bytes) {
240
+ throw new Error(
241
+ `[FunctionalDescriptor] compression proof gate failed for "${name}": ` +
242
+ `descriptor_bytes=${descriptorBytes} must be lower than dense_f16_bytes=${denseF16Bytes}.`
243
+ );
244
+ }
245
+ }
246
+
247
+ // ============================================================================
248
+ // Sparse component (coo_v1)
249
+ // Binary layout: [nnz u32][rowIdx i32...][colIdx i32...][values f32...]
250
+ // ============================================================================
251
+
252
+ function deserializeSparse(buf) {
253
+ const view = new DataView(buf.buffer, buf.byteOffset, buf.byteLength);
254
+ const nnz = view.getUint32(0, true);
255
+ let off = 4;
256
+ const rowIdx = new Int32Array(buf.buffer, buf.byteOffset + off, nnz); off += nnz * 4;
257
+ const colIdx = new Int32Array(buf.buffer, buf.byteOffset + off, nnz); off += nnz * 4;
258
+ const values = new Float32Array(buf.buffer, buf.byteOffset + off, nnz);
259
+ return { rowIdx, colIdx, values };
260
+ }
261
+
262
+ function applySparse(out, sparse, cols) {
263
+ const { rowIdx, colIdx, values } = sparse;
264
+ for (let k = 0; k < values.length; k++) {
265
+ out[rowIdx[k] * cols + colIdx[k]] += values[k];
266
+ }
267
+ }
268
+
269
+ // ============================================================================
270
+ // Public entry point
271
+ // ============================================================================
272
+
273
+ export async function loadFunctionalDescriptor(shardData, location, name) {
274
+ if (!location.descriptorManifest) {
275
+ throw new Error(
276
+ `[FunctionalDescriptor] location.descriptorManifest is required for tensor "${name}". ` +
277
+ 'The layer loader must parse the manifoldgguf manifest and attach it before calling loadTensorToGPU.'
278
+ );
279
+ }
280
+ const manifest = assertFunctionalDescriptorManifest(
281
+ location.descriptorManifest,
282
+ `FUNCTIONAL_DESCRIPTOR tensor "${name}" descriptorManifest`
283
+ );
284
+
285
+ const shards = shardData?.descriptorShards;
286
+ if (!shards || typeof shards.get !== 'function') {
287
+ throw new Error(
288
+ `[FunctionalDescriptor] shardData.descriptorShards (Map) is required for tensor "${name}". ` +
289
+ 'The shard manager must load all component shard files before dispatch.'
290
+ );
291
+ }
292
+
293
+ const { components } = manifest;
294
+ const [rows, cols] = normalizeShape2(
295
+ manifest.padded_shape ?? manifest.slice_shape,
296
+ 'descriptor reconstruction shape'
297
+ );
298
+ const [cropRows, cropCols] = normalizeShape2(
299
+ manifest.crop_shape ?? location.shape ?? manifest.slice_shape,
300
+ 'descriptor crop shape'
301
+ );
302
+ if (cropRows > rows || cropCols > cols) {
303
+ throw new Error(
304
+ `[FunctionalDescriptor] crop shape [${cropRows},${cropCols}] exceeds descriptor shape [${rows},${cols}] for "${name}".`
305
+ );
306
+ }
307
+ const descriptorBytes = getDescriptorShardBytes(shards);
308
+ const denseF16Bytes = cropRows * cropCols * 2;
309
+ assertRuntimeProofGates(manifest, name, descriptorBytes, denseF16Bytes);
310
+
311
+ log.debug('FunctionalDescriptor', `Reconstructing "${name}" [${rows}x${cols}]`);
312
+
313
+ const reconstructed = new Float32Array(rows * cols);
314
+
315
+ // PRNG substrate
316
+ const prngSpec = components.prng_substrate;
317
+ const prng = buildPRNG(prngSpec, rows, cols);
318
+ for (let i = 0; i < reconstructed.length; i++) reconstructed[i] += prng[i];
319
+
320
+ // Kronecker sum
321
+ const kronSpec = components.kronecker_sum;
322
+ const kronShard = shards.get(kronSpec.shard_file);
323
+ if (!kronShard) {
324
+ throw new Error(`[FunctionalDescriptor] Kron shard "${kronSpec.shard_file}" not found in descriptorShards for "${name}".`);
325
+ }
326
+ const kronFactors = deserializeKronecker(
327
+ kronShard instanceof Uint8Array ? kronShard : new Uint8Array(kronShard)
328
+ );
329
+ const kron = reconstructKronecker(kronFactors, rows, cols);
330
+ for (let i = 0; i < reconstructed.length; i++) reconstructed[i] += kron[i];
331
+
332
+ // SIREN INR
333
+ const sirenSpec = components.coordinate_inr;
334
+ const sirenShard = shards.get(sirenSpec.shard_file);
335
+ if (!sirenShard) {
336
+ throw new Error(`[FunctionalDescriptor] SIREN shard "${sirenSpec.shard_file}" not found in descriptorShards for "${name}".`);
337
+ }
338
+ const sirenLayers = deserializeSIREN(
339
+ sirenShard instanceof Uint8Array ? sirenShard : new Uint8Array(sirenShard)
340
+ );
341
+ const siren = reconstructSIREN(sirenLayers, rows, cols);
342
+ for (let i = 0; i < reconstructed.length; i++) reconstructed[i] += siren[i];
343
+
344
+ // Sparse residuals
345
+ const sparseSpec = components.sparse_outliers;
346
+ const sparseShard = shards.get(sparseSpec.shard_file);
347
+ if (!sparseShard) {
348
+ throw new Error(`[FunctionalDescriptor] Sparse shard "${sparseSpec.shard_file}" not found in descriptorShards for "${name}".`);
349
+ }
350
+ const sparse = deserializeSparse(
351
+ sparseShard instanceof Uint8Array ? sparseShard : new Uint8Array(sparseShard)
352
+ );
353
+ applySparse(reconstructed, sparse, cols);
354
+
355
+ log.debug('FunctionalDescriptor', `"${name}" reconstructed, uploading to GPU`);
356
+ const out = cropMatrix(reconstructed, rows, cols, cropRows, cropCols);
357
+
358
+ // Upload as dense F32 weight buffer
359
+ const gpuBuf = acquireBuffer(out.byteLength, GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST);
360
+ uploadData(gpuBuf, out);
361
+
362
+ const data = createWeightBuffer(gpuBuf, 'f32', 'row', [cropRows, cropCols], name, null, {
363
+ storageType: 'functional_descriptor',
364
+ descriptorHash: manifest.descriptor_hash,
365
+ descriptorBytes,
366
+ denseF16Bytes,
367
+ compressionRatio: descriptorBytes > 0 ? denseF16Bytes / descriptorBytes : null,
368
+ proofStatus: manifest.proof_status ?? null,
369
+ proofStatusGate: manifest.proof_status_gate ?? null,
370
+ descriptorShape: [rows, cols],
371
+ cropShape: [cropRows, cropCols],
372
+ });
373
+ return {
374
+ data,
375
+ allocatedBuffers: [gpuBuf],
376
+ };
377
+ }
@@ -72,6 +72,16 @@ export declare function getWeightLayout(
72
72
  config: TensorLoadConfig
73
73
  ): WeightLayout;
74
74
 
75
+ export declare function isLiteRTAffineInt4FusedEligible(
76
+ location: TensorLocation,
77
+ config: Partial<TensorLoadConfig> | null
78
+ ): boolean;
79
+
80
+ export declare function isW4A16FusedEligible(
81
+ location: TensorLocation,
82
+ config: Partial<TensorLoadConfig> | null
83
+ ): boolean;
84
+
75
85
  /**
76
86
  * Convert BF16 data to F32 on CPU.
77
87
  */
@@ -91,6 +101,20 @@ export declare function loadQ4KFused(
91
101
  name: string
92
102
  ): Promise<TensorLoadResult>;
93
103
 
104
+ export declare function loadLiteRTInt4Fused(
105
+ shardData: Uint8Array,
106
+ location: TensorLocation,
107
+ name: string,
108
+ config?: TensorLoadConfig | null
109
+ ): Promise<TensorLoadResult>;
110
+
111
+ export declare function loadW4A16Fused(
112
+ shardData: Uint8Array,
113
+ location: TensorLocation,
114
+ name: string,
115
+ config?: TensorLoadConfig | null
116
+ ): Promise<TensorLoadResult>;
117
+
94
118
  /**
95
119
  * Load Q4K tensor to GPU with dequantization.
96
120
  */
@@ -131,6 +155,16 @@ export declare function loadFloat(
131
155
  config: TensorLoadConfig
132
156
  ): Promise<TensorLoadResult>;
133
157
 
158
+ /**
159
+ * Load W4A16 packed tensor through the reference dequantization path.
160
+ */
161
+ export declare function loadW4A16Dequant(
162
+ shardData: Uint8Array,
163
+ location: TensorLocation,
164
+ name: string,
165
+ config: TensorLoadConfig
166
+ ): Promise<TensorLoadResult>;
167
+
134
168
  /**
135
169
  * Load tensor data to GPU based on dtype.
136
170
  *
@@ -158,5 +192,6 @@ export declare function loadTensorToGPU(
158
192
  */
159
193
  export declare function loadTensorToCPU(
160
194
  shardData: Uint8Array,
161
- location: TensorLocation
195
+ location: TensorLocation,
196
+ name?: string | null
162
197
  ): Float32Array | Uint8Array;