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
@@ -707,15 +707,26 @@ export async function encodeGemma4Image(params) {
707
707
  );
708
708
  uploadData(pooledTensor.buffer, pooled.output, 0);
709
709
 
710
- const projected = await runMatmul(
711
- pooledTensor,
712
- weights.projector,
713
- pooled.outputLength,
714
- weights.textHiddenSize,
715
- hiddenSize,
716
- { outputDtype: 'f32', transposeB: 'auto' }
717
- );
718
- releaseBuffer(pooledTensor.buffer);
710
+ let projected = null;
711
+ if (weights.projector) {
712
+ projected = await runMatmul(
713
+ pooledTensor,
714
+ weights.projector,
715
+ pooled.outputLength,
716
+ weights.textHiddenSize,
717
+ hiddenSize,
718
+ { outputDtype: 'f32', transposeB: 'auto' }
719
+ );
720
+ releaseBuffer(pooledTensor.buffer);
721
+ } else {
722
+ if (hiddenSize !== weights.textHiddenSize) {
723
+ throw new Error(
724
+ `[Vision] Gemma 4 vision encoder-free mode has no projector, but vision hiddenSize (${hiddenSize}) ` +
725
+ `does not match text hiddenSize (${weights.textHiddenSize}).`
726
+ );
727
+ }
728
+ projected = pooledTensor;
729
+ }
719
730
 
720
731
  return {
721
732
  features: projected.buffer,
@@ -32,6 +32,30 @@ export interface TokenizerInitOptions {
32
32
  loadTokenizerModel?: ((path?: string) => Promise<ArrayBuffer | Uint8Array | null | undefined>) | null;
33
33
  }
34
34
 
35
+ export type TokenizerLoadTimingPhase =
36
+ | 'configResolution'
37
+ | 'cacheLookup'
38
+ | 'backendCreate'
39
+ | 'assetLoad'
40
+ | 'assetParse'
41
+ | 'backendLoad'
42
+ | 'cacheStore';
43
+
44
+ export interface TokenizerLoadTiming {
45
+ schemaVersion: 1;
46
+ source: 'doppler-tokenizer';
47
+ modelId: string | null;
48
+ status: 'running' | 'complete' | 'failed';
49
+ tokenizerType: string | null;
50
+ tokenizerFile: string | null;
51
+ backend: string | null;
52
+ assetSource: string | null;
53
+ cacheHit: boolean;
54
+ phasesMs: Record<TokenizerLoadTimingPhase, number | null>;
55
+ totalMs: number | null;
56
+ error: string | null;
57
+ }
58
+
35
59
  /**
36
60
  * Tokenizer wrapper that auto-detects backend from model manifest
37
61
  * This is a thin wrapper over the backend implementations
@@ -39,6 +63,7 @@ export interface TokenizerInitOptions {
39
63
  export declare class Tokenizer {
40
64
  private backend;
41
65
  private config;
66
+ private loadTiming;
42
67
 
43
68
  /**
44
69
  * Initialize from model manifest.
@@ -72,6 +97,11 @@ export declare class Tokenizer {
72
97
  * Get high-priority token IDs for bounded runtime hot caches, or null when unavailable
73
98
  */
74
99
  getHotTokenIds(limit: number): number[] | null;
100
+
101
+ /**
102
+ * Get report-only tokenizer initialization timing.
103
+ */
104
+ getLoadTiming(): TokenizerLoadTiming | null;
75
105
  }
76
106
 
77
107
  export default Tokenizer;
@@ -8,6 +8,15 @@ import { BPETokenizer } from './tokenizers/bpe.js';
8
8
  import { toArrayBuffer } from '../utils/array-buffer.js';
9
9
 
10
10
  const BUNDLED_TOKENIZER_CACHE = new Map();
11
+ const TOKENIZER_LOAD_PHASES = Object.freeze([
12
+ 'configResolution',
13
+ 'cacheLookup',
14
+ 'backendCreate',
15
+ 'assetLoad',
16
+ 'assetParse',
17
+ 'backendLoad',
18
+ 'cacheStore',
19
+ ]);
11
20
 
12
21
  function hasScheme(value) {
13
22
  return /^[a-zA-Z][a-zA-Z0-9+.-]*:/.test(value);
@@ -26,6 +35,69 @@ function parseTokenizerJsonPayload(value) {
26
35
  throw new Error('Tokenizer JSON loader must return an object, JSON string, or null.');
27
36
  }
28
37
 
38
+ function roundTokenizerTimingMs(value) {
39
+ return Number.isFinite(value) ? Number(value.toFixed(3)) : null;
40
+ }
41
+
42
+ function createTokenizerLoadTiming(modelId) {
43
+ return {
44
+ schemaVersion: 1,
45
+ source: 'doppler-tokenizer',
46
+ modelId: typeof modelId === 'string' ? modelId : null,
47
+ status: 'running',
48
+ tokenizerType: null,
49
+ tokenizerFile: null,
50
+ backend: null,
51
+ assetSource: null,
52
+ cacheHit: false,
53
+ phasesMs: Object.fromEntries(TOKENIZER_LOAD_PHASES.map((phase) => [phase, null])),
54
+ totalMs: null,
55
+ error: null,
56
+ };
57
+ }
58
+
59
+ function cloneTokenizerLoadTiming(loadTiming) {
60
+ if (!loadTiming || typeof loadTiming !== 'object') {
61
+ return null;
62
+ }
63
+ return JSON.parse(JSON.stringify(loadTiming));
64
+ }
65
+
66
+ function finishTokenizerLoadPhase(loadTiming, phase, startMs) {
67
+ if (!loadTiming?.phasesMs || !phase) return;
68
+ const elapsedMs = roundTokenizerTimingMs(performance.now() - startMs);
69
+ const currentMs = loadTiming.phasesMs[phase];
70
+ loadTiming.phasesMs[phase] = Number.isFinite(currentMs)
71
+ ? roundTokenizerTimingMs(currentMs + elapsedMs)
72
+ : elapsedMs;
73
+ }
74
+
75
+ async function timedTokenizerLoadPhase(loadTiming, phase, run) {
76
+ const startMs = performance.now();
77
+ try {
78
+ return await run();
79
+ } finally {
80
+ finishTokenizerLoadPhase(loadTiming, phase, startMs);
81
+ }
82
+ }
83
+
84
+ function timedTokenizerLoadPhaseSync(loadTiming, phase, run) {
85
+ const startMs = performance.now();
86
+ try {
87
+ return run();
88
+ } finally {
89
+ finishTokenizerLoadPhase(loadTiming, phase, startMs);
90
+ }
91
+ }
92
+
93
+ function finishTokenizerLoadTiming(loadTiming, status, startMs, error = null) {
94
+ loadTiming.status = status;
95
+ loadTiming.totalMs = roundTokenizerTimingMs(performance.now() - startMs);
96
+ loadTiming.error = error == null
97
+ ? null
98
+ : (error instanceof Error ? error.message : String(error));
99
+ }
100
+
29
101
  function buildBundledTokenizerCacheKey(modelId, tokenizerConfig) {
30
102
  const tokenizerType = typeof tokenizerConfig?.type === 'string'
31
103
  ? tokenizerConfig.type.trim().toLowerCase()
@@ -55,199 +127,299 @@ export class Tokenizer {
55
127
 
56
128
  config = null;
57
129
 
130
+ loadTiming = null;
131
+
58
132
 
59
133
  async initialize(manifest, options = {}) {
134
+ const loadStart = performance.now();
60
135
  const modelId = manifest?.modelId || 'unknown';
61
- // Merge tokenizer hints from the caller, but let the manifest stay authoritative.
62
- const tokenizerHints = options.tokenizerHints || {};
63
- const tokenizerConfig = { ...tokenizerHints, ...(manifest.tokenizer || {}) };
64
- const eosTokenId = Array.isArray(manifest.eos_token_id)
65
- ? manifest.eos_token_id[0]
66
- : manifest.eos_token_id;
67
-
68
- if (tokenizerConfig.eosToken == null && Array.isArray(tokenizerConfig.eosTokens) && tokenizerConfig.eosTokens.length > 0) {
69
- tokenizerConfig.eosToken = tokenizerConfig.eosTokens[0];
70
- }
71
- if (tokenizerConfig.eosToken == null && typeof eosTokenId === 'number') {
72
- tokenizerConfig.eosToken = eosTokenId;
73
- }
74
- if (tokenizerConfig.bosToken == null && tokenizerConfig.bosTokenId != null) {
75
- tokenizerConfig.bosToken = tokenizerConfig.bosTokenId;
76
- }
77
- if (tokenizerConfig.padToken == null && tokenizerConfig.padTokenId != null) {
78
- tokenizerConfig.padToken = tokenizerConfig.padTokenId;
79
- }
80
- if (tokenizerConfig.unkToken == null && tokenizerConfig.unkTokenId != null) {
81
- tokenizerConfig.unkToken = tokenizerConfig.unkTokenId;
82
- }
136
+ const loadTiming = createTokenizerLoadTiming(modelId);
137
+ this.loadTiming = loadTiming;
138
+ try {
139
+ const tokenizerConfig = timedTokenizerLoadPhaseSync(loadTiming, 'configResolution', () => {
140
+ const tokenizerHints = options.tokenizerHints || {};
141
+ const resolvedTokenizerConfig = { ...tokenizerHints, ...(manifest.tokenizer || {}) };
142
+ const eosTokenId = Array.isArray(manifest.eos_token_id)
143
+ ? manifest.eos_token_id[0]
144
+ : manifest.eos_token_id;
83
145
 
84
- // Check for bundled or HuggingFace tokenizer first (eliminates transformers.js dependency)
85
- const isBundled = tokenizerConfig.type === 'bundled' || tokenizerConfig.type === 'huggingface';
86
- if (isBundled && tokenizerConfig.file) {
87
- tokenizerConfig.deferSpecialTokens = true;
88
- log.info('Tokenizer', `Loading ${tokenizerConfig.type} tokenizer from ${tokenizerConfig.file}`);
89
- const cacheKey = buildBundledTokenizerCacheKey(modelId, tokenizerConfig);
90
- const cachedBackend = cacheKey ? BUNDLED_TOKENIZER_CACHE.get(cacheKey) : null;
91
- if (cachedBackend) {
92
- log.info('Tokenizer', `Bundled tokenizer cache hit: ${tokenizerConfig.file}`);
93
- this.backend = cachedBackend;
94
- this.config = tokenizerConfig;
95
- return;
96
- }
146
+ if (resolvedTokenizerConfig.eosToken == null && Array.isArray(resolvedTokenizerConfig.eosTokens) && resolvedTokenizerConfig.eosTokens.length > 0) {
147
+ resolvedTokenizerConfig.eosToken = resolvedTokenizerConfig.eosTokens[0];
148
+ }
149
+ if (resolvedTokenizerConfig.eosToken == null && typeof eosTokenId === 'number') {
150
+ resolvedTokenizerConfig.eosToken = eosTokenId;
151
+ }
152
+ if (resolvedTokenizerConfig.bosToken == null && resolvedTokenizerConfig.bosTokenId != null) {
153
+ resolvedTokenizerConfig.bosToken = resolvedTokenizerConfig.bosTokenId;
154
+ }
155
+ if (resolvedTokenizerConfig.padToken == null && resolvedTokenizerConfig.padTokenId != null) {
156
+ resolvedTokenizerConfig.padToken = resolvedTokenizerConfig.padTokenId;
157
+ }
158
+ if (resolvedTokenizerConfig.unkToken == null && resolvedTokenizerConfig.unkTokenId != null) {
159
+ resolvedTokenizerConfig.unkToken = resolvedTokenizerConfig.unkTokenId;
160
+ }
161
+ return resolvedTokenizerConfig;
162
+ });
163
+ loadTiming.tokenizerType = typeof tokenizerConfig.type === 'string' ? tokenizerConfig.type : null;
164
+ loadTiming.tokenizerFile = typeof tokenizerConfig.file === 'string' ? tokenizerConfig.file : null;
165
+
166
+ // Check for bundled or HuggingFace tokenizer first (eliminates transformers.js dependency)
167
+ const isBundled = tokenizerConfig.type === 'bundled' || tokenizerConfig.type === 'huggingface';
168
+ if (isBundled && tokenizerConfig.file) {
169
+ loadTiming.backend = 'bundled';
170
+ tokenizerConfig.deferSpecialTokens = true;
171
+ log.info('Tokenizer', `Loading ${tokenizerConfig.type} tokenizer from ${tokenizerConfig.file}`);
172
+ const { cacheKey, cachedBackend } = timedTokenizerLoadPhaseSync(loadTiming, 'cacheLookup', () => {
173
+ const resolvedCacheKey = buildBundledTokenizerCacheKey(modelId, tokenizerConfig);
174
+ return {
175
+ cacheKey: resolvedCacheKey,
176
+ cachedBackend: resolvedCacheKey ? BUNDLED_TOKENIZER_CACHE.get(resolvedCacheKey) : null,
177
+ };
178
+ });
179
+ loadTiming.cacheHit = Boolean(cachedBackend);
180
+ if (cachedBackend) {
181
+ log.info('Tokenizer', `Bundled tokenizer cache hit: ${tokenizerConfig.file}`);
182
+ this.backend = cachedBackend;
183
+ this.config = tokenizerConfig;
184
+ loadTiming.assetSource = 'cache';
185
+ finishTokenizerLoadTiming(loadTiming, 'complete', loadStart);
186
+ return;
187
+ }
97
188
 
98
- this.backend = new BundledTokenizer(tokenizerConfig);
189
+ this.backend = timedTokenizerLoadPhaseSync(
190
+ loadTiming,
191
+ 'backendCreate',
192
+ () => new BundledTokenizer(tokenizerConfig)
193
+ );
99
194
 
100
- const baseUrl = options.baseUrl;
195
+ const baseUrl = options.baseUrl;
101
196
 
102
- let tokenizerJson = null;
103
- const attemptedPaths = [];
197
+ let tokenizerJson = null;
198
+ const attemptedPaths = [];
104
199
 
105
- // Try to load tokenizer.json
106
- if (typeof options.loadTokenizerJson === 'function') {
107
- attemptedPaths.push('custom-loader');
108
- try {
109
- const loaded = await options.loadTokenizerJson();
110
- tokenizerJson = parseTokenizerJsonPayload(loaded);
111
- } catch (err) {
112
- const message = err instanceof Error ? err.message : String(err);
113
- log.warn('Tokenizer', `Failed to load bundled tokenizer from custom loader: ${message}`);
200
+ // Try to load tokenizer.json
201
+ if (typeof options.loadTokenizerJson === 'function') {
202
+ attemptedPaths.push('custom-loader');
203
+ try {
204
+ const loaded = await timedTokenizerLoadPhase(
205
+ loadTiming,
206
+ 'assetLoad',
207
+ () => options.loadTokenizerJson()
208
+ );
209
+ tokenizerJson = timedTokenizerLoadPhaseSync(
210
+ loadTiming,
211
+ 'assetParse',
212
+ () => parseTokenizerJsonPayload(loaded)
213
+ );
214
+ if (tokenizerJson) {
215
+ loadTiming.assetSource = 'custom-loader';
216
+ }
217
+ } catch (err) {
218
+ const message = err instanceof Error ? err.message : String(err);
219
+ log.warn('Tokenizer', `Failed to load bundled tokenizer from custom loader: ${message}`);
220
+ }
114
221
  }
115
- }
116
222
 
117
- if (!tokenizerJson && baseUrl) {
118
- // Load from remote URL
119
- const tokenizerUrl = `${baseUrl}/${tokenizerConfig.file}`;
120
- attemptedPaths.push(tokenizerUrl);
121
- try {
122
- const response = await fetch(tokenizerUrl);
123
- if (!response.ok) {
124
- throw new Error(`Failed to fetch tokenizer: ${response.status}`);
223
+ if (!tokenizerJson && baseUrl) {
224
+ // Load from remote URL
225
+ const tokenizerUrl = `${baseUrl}/${tokenizerConfig.file}`;
226
+ attemptedPaths.push(tokenizerUrl);
227
+ try {
228
+ const tokenizerText = await timedTokenizerLoadPhase(loadTiming, 'assetLoad', async () => {
229
+ const response = await fetch(tokenizerUrl);
230
+ if (!response.ok) {
231
+ throw new Error(`Failed to fetch tokenizer: ${response.status}`);
232
+ }
233
+ return response.text();
234
+ });
235
+ tokenizerJson = timedTokenizerLoadPhaseSync(
236
+ loadTiming,
237
+ 'assetParse',
238
+ () => JSON.parse(tokenizerText)
239
+ );
240
+ loadTiming.assetSource = tokenizerUrl;
241
+ } catch (err) {
242
+ const message = err instanceof Error ? err.message : String(err);
243
+ log.warn('Tokenizer', `Failed to fetch bundled tokenizer from URL: ${message}`);
244
+ }
245
+ } else if (!tokenizerJson) {
246
+ // Try to load from OPFS (for cached models)
247
+ attemptedPaths.push('OPFS:tokenizer.json');
248
+ try {
249
+ const tokenizerStr = await timedTokenizerLoadPhase(loadTiming, 'assetLoad', async () => {
250
+ const { loadTokenizerFromStore } = await import('../storage/shard-manager.js');
251
+ return loadTokenizerFromStore();
252
+ });
253
+ if (tokenizerStr) {
254
+ tokenizerJson = timedTokenizerLoadPhaseSync(
255
+ loadTiming,
256
+ 'assetParse',
257
+ () => JSON.parse(tokenizerStr)
258
+ );
259
+ loadTiming.assetSource = 'OPFS:tokenizer.json';
260
+ }
261
+ } catch (err) {
262
+ const message = err instanceof Error ? err.message : String(err);
263
+ log.warn('Tokenizer', `Failed to load bundled tokenizer from OPFS: ${message}`);
125
264
  }
126
- tokenizerJson = await response.json();
127
- } catch (err) {
128
- const message = err instanceof Error ? err.message : String(err);
129
- log.warn('Tokenizer', `Failed to fetch bundled tokenizer from URL: ${message}`);
130
265
  }
131
- } else if (!tokenizerJson) {
132
- // Try to load from OPFS (for cached models)
133
- attemptedPaths.push('OPFS:tokenizer.json');
134
- try {
135
- const { loadTokenizerFromStore } = await import('../storage/shard-manager.js');
136
- const tokenizerStr = await loadTokenizerFromStore();
137
- if (tokenizerStr) {
138
- tokenizerJson = JSON.parse(tokenizerStr);
266
+
267
+ if (tokenizerJson) {
268
+ timedTokenizerLoadPhaseSync(
269
+ loadTiming,
270
+ 'backendLoad',
271
+ () => (this.backend).load(tokenizerJson)
272
+ );
273
+ if (cacheKey) {
274
+ timedTokenizerLoadPhaseSync(
275
+ loadTiming,
276
+ 'cacheStore',
277
+ () => BUNDLED_TOKENIZER_CACHE.set(cacheKey, this.backend)
278
+ );
139
279
  }
140
- } catch (err) {
141
- const message = err instanceof Error ? err.message : String(err);
142
- log.warn('Tokenizer', `Failed to load bundled tokenizer from OPFS: ${message}`);
280
+ this.config = tokenizerConfig;
281
+ finishTokenizerLoadTiming(loadTiming, 'complete', loadStart);
282
+ return;
143
283
  }
284
+
285
+ // No external fallback - bundled tokenizer is required
286
+ throw new Error(
287
+ `[Tokenizer] Bundled tokenizer not found for model "${modelId}". ` +
288
+ `Expected tokenizer file: "${tokenizerConfig.file}". ` +
289
+ `Attempted paths: [${attemptedPaths.join(', ')}]. ` +
290
+ 'Ensure tokenizer.json is in OPFS or model directory. ' +
291
+ 'Clear browser storage and re-download the model.'
292
+ );
144
293
  }
145
294
 
146
- if (tokenizerJson) {
147
- (this.backend).load(tokenizerJson);
148
- if (cacheKey) {
149
- BUNDLED_TOKENIZER_CACHE.set(cacheKey, this.backend);
150
- }
151
- this.config = tokenizerConfig;
152
- return;
295
+ let hfModel = tokenizerConfig.hfModel ?? tokenizerConfig.modelId ?? null;
296
+ const allowArchFallback = tokenizerConfig.allowArchFallback === true;
297
+ if (allowArchFallback && !hfModel) {
298
+ throw new Error(
299
+ `[Tokenizer] tokenizer.allowArchFallback requires explicit tokenizer.hfModel or tokenizer.modelId for model "${modelId}".`
300
+ );
153
301
  }
154
302
 
155
- // No external fallback - bundled tokenizer is required
156
- throw new Error(
157
- `[Tokenizer] Bundled tokenizer not found for model "${modelId}". ` +
158
- `Expected tokenizer file: "${tokenizerConfig.file}". ` +
159
- `Attempted paths: [${attemptedPaths.join(', ')}]. ` +
160
- 'Ensure tokenizer.json is in OPFS or model directory. ' +
161
- 'Clear browser storage and re-download the model.'
162
- );
163
- }
303
+ if (hfModel) {
304
+ // Use Transformers.js for HuggingFace models (fallback)
305
+ loadTiming.backend = 'transformersjs';
306
+ loadTiming.assetSource = hfModel;
307
+ log.info('Tokenizer', `Loading from HuggingFace: ${hfModel}`);
308
+ this.backend = timedTokenizerLoadPhaseSync(loadTiming, 'backendCreate', () => new TransformersTokenizer({
309
+ modelId: hfModel,
310
+ ...tokenizerConfig
311
+ }));
312
+ await timedTokenizerLoadPhase(loadTiming, 'backendLoad', () => (this.backend).load(hfModel));
313
+ } else if (tokenizerConfig.sentencepieceModel) {
314
+ // Load SentencePiece model
315
+ loadTiming.backend = 'sentencepiece';
316
+ loadTiming.tokenizerFile = typeof tokenizerConfig.sentencepieceModel === 'string'
317
+ ? tokenizerConfig.sentencepieceModel
318
+ : loadTiming.tokenizerFile;
319
+ this.backend = timedTokenizerLoadPhaseSync(
320
+ loadTiming,
321
+ 'backendCreate',
322
+ () => new SentencePieceTokenizer(tokenizerConfig)
323
+ );
164
324
 
165
- let hfModel = tokenizerConfig.hfModel ?? tokenizerConfig.modelId ?? null;
166
- const allowArchFallback = tokenizerConfig.allowArchFallback === true;
167
- if (allowArchFallback && !hfModel) {
168
- throw new Error(
169
- `[Tokenizer] tokenizer.allowArchFallback requires explicit tokenizer.hfModel or tokenizer.modelId for model "${modelId}".`
170
- );
171
- }
325
+ // Load the model data from the provided source
172
326
 
173
- if (hfModel) {
174
- // Use Transformers.js for HuggingFace models (fallback)
175
- log.info('Tokenizer', `Loading from HuggingFace: ${hfModel}`);
176
- this.backend = new TransformersTokenizer({
177
- modelId: hfModel,
178
- ...tokenizerConfig
179
- });
180
- await (this.backend).load(hfModel);
181
- } else if (tokenizerConfig.sentencepieceModel) {
182
- // Load SentencePiece model
183
- this.backend = new SentencePieceTokenizer(tokenizerConfig);
184
-
185
- // Load the model data from the provided source
186
-
187
- let modelData;
188
- const spAttemptedPaths = [];
189
- if (typeof options.loadTokenizerModel === 'function') {
190
- spAttemptedPaths.push('custom-loader');
191
- try {
192
- const loaded = await options.loadTokenizerModel(
193
- typeof tokenizerConfig.sentencepieceModel === 'string'
194
- ? tokenizerConfig.sentencepieceModel
195
- : undefined
196
- );
197
- if (loaded != null) {
198
- modelData = toArrayBuffer(loaded, 'options.loadTokenizerModel');
199
- }
200
- } catch (err) {
201
- const message = err instanceof Error ? err.message : String(err);
202
- log.warn('Tokenizer', `Failed to load sentencepiece model from custom loader: ${message}`);
203
- }
204
- }
205
- if (!modelData && tokenizerConfig.sentencepieceModel instanceof ArrayBuffer) {
206
- spAttemptedPaths.push('inline-ArrayBuffer');
207
- modelData = tokenizerConfig.sentencepieceModel;
208
- } else if (!modelData && tokenizerConfig.loadShard) {
209
- // Use provided shard loader
210
- spAttemptedPaths.push('shard-loader');
211
- modelData = await tokenizerConfig.loadShard(tokenizerConfig.sentencepieceModel);
212
- } else if (!modelData && typeof tokenizerConfig.sentencepieceModel === 'string') {
213
- if (options.baseUrl && !hasScheme(tokenizerConfig.sentencepieceModel)) {
214
- const url = `${options.baseUrl}/${tokenizerConfig.sentencepieceModel}`;
215
- spAttemptedPaths.push(url);
216
- const response = await fetch(url);
217
- modelData = await response.arrayBuffer();
218
- } else if (hasScheme(tokenizerConfig.sentencepieceModel)) {
219
- spAttemptedPaths.push(tokenizerConfig.sentencepieceModel);
220
- const response = await fetch(tokenizerConfig.sentencepieceModel);
221
- modelData = await response.arrayBuffer();
222
- } else {
223
- spAttemptedPaths.push('OPFS:tokenizer.model');
327
+ let modelData;
328
+ const spAttemptedPaths = [];
329
+ if (typeof options.loadTokenizerModel === 'function') {
330
+ spAttemptedPaths.push('custom-loader');
224
331
  try {
225
- const { loadTokenizerModelFromStore } = await import('../storage/shard-manager.js');
226
- modelData = await loadTokenizerModelFromStore();
332
+ modelData = await timedTokenizerLoadPhase(loadTiming, 'assetLoad', async () => {
333
+ const loaded = await options.loadTokenizerModel(
334
+ typeof tokenizerConfig.sentencepieceModel === 'string'
335
+ ? tokenizerConfig.sentencepieceModel
336
+ : undefined
337
+ );
338
+ return loaded == null
339
+ ? null
340
+ : toArrayBuffer(loaded, 'options.loadTokenizerModel');
341
+ });
342
+ if (modelData) {
343
+ loadTiming.assetSource = 'custom-loader';
344
+ }
227
345
  } catch (err) {
228
346
  const message = err instanceof Error ? err.message : String(err);
229
- log.warn('Tokenizer', `Failed to load tokenizer.model from OPFS: ${message}`);
347
+ log.warn('Tokenizer', `Failed to load sentencepiece model from custom loader: ${message}`);
348
+ }
349
+ }
350
+ if (!modelData && tokenizerConfig.sentencepieceModel instanceof ArrayBuffer) {
351
+ spAttemptedPaths.push('inline-ArrayBuffer');
352
+ loadTiming.assetSource = 'inline-ArrayBuffer';
353
+ modelData = tokenizerConfig.sentencepieceModel;
354
+ } else if (!modelData && tokenizerConfig.loadShard) {
355
+ // Use provided shard loader
356
+ spAttemptedPaths.push('shard-loader');
357
+ modelData = await timedTokenizerLoadPhase(
358
+ loadTiming,
359
+ 'assetLoad',
360
+ () => tokenizerConfig.loadShard(tokenizerConfig.sentencepieceModel)
361
+ );
362
+ loadTiming.assetSource = 'shard-loader';
363
+ } else if (!modelData && typeof tokenizerConfig.sentencepieceModel === 'string') {
364
+ if (options.baseUrl && !hasScheme(tokenizerConfig.sentencepieceModel)) {
365
+ const url = `${options.baseUrl}/${tokenizerConfig.sentencepieceModel}`;
366
+ spAttemptedPaths.push(url);
367
+ modelData = await timedTokenizerLoadPhase(loadTiming, 'assetLoad', async () => {
368
+ const response = await fetch(url);
369
+ return response.arrayBuffer();
370
+ });
371
+ loadTiming.assetSource = url;
372
+ } else if (hasScheme(tokenizerConfig.sentencepieceModel)) {
373
+ spAttemptedPaths.push(tokenizerConfig.sentencepieceModel);
374
+ modelData = await timedTokenizerLoadPhase(loadTiming, 'assetLoad', async () => {
375
+ const response = await fetch(tokenizerConfig.sentencepieceModel);
376
+ return response.arrayBuffer();
377
+ });
378
+ loadTiming.assetSource = tokenizerConfig.sentencepieceModel;
379
+ } else {
380
+ spAttemptedPaths.push('OPFS:tokenizer.model');
381
+ try {
382
+ modelData = await timedTokenizerLoadPhase(loadTiming, 'assetLoad', async () => {
383
+ const { loadTokenizerModelFromStore } = await import('../storage/shard-manager.js');
384
+ return loadTokenizerModelFromStore();
385
+ });
386
+ if (modelData) {
387
+ loadTiming.assetSource = 'OPFS:tokenizer.model';
388
+ }
389
+ } catch (err) {
390
+ const message = err instanceof Error ? err.message : String(err);
391
+ log.warn('Tokenizer', `Failed to load tokenizer.model from OPFS: ${message}`);
392
+ }
230
393
  }
231
394
  }
232
- }
233
395
 
234
- if (modelData) {
235
- await (this.backend).load(modelData);
396
+ if (modelData) {
397
+ await timedTokenizerLoadPhase(loadTiming, 'backendLoad', () => (this.backend).load(modelData));
398
+ } else {
399
+ throw new Error(`Could not load SentencePiece model data. Attempted paths: [${spAttemptedPaths.join(', ')}]`);
400
+ }
401
+ } else if (tokenizerConfig.vocab && tokenizerConfig.merges) {
402
+ // BPE with vocab + merges
403
+ loadTiming.backend = 'bpe';
404
+ this.backend = timedTokenizerLoadPhaseSync(loadTiming, 'backendCreate', () => new BPETokenizer(tokenizerConfig));
405
+ timedTokenizerLoadPhaseSync(
406
+ loadTiming,
407
+ 'backendLoad',
408
+ () => (this.backend).load(tokenizerConfig.vocab, tokenizerConfig.merges)
409
+ );
236
410
  } else {
237
- throw new Error(`Could not load SentencePiece model data. Attempted paths: [${spAttemptedPaths.join(', ')}]`);
411
+ throw new Error(
412
+ `[Tokenizer] No valid tokenizer configuration in manifest for model "${modelId}". ` +
413
+ 'Provide tokenizer.hfModel or bundle tokenizer.json (tokenizer.type="bundled", tokenizer.file="tokenizer.json").'
414
+ );
238
415
  }
239
- } else if (tokenizerConfig.vocab && tokenizerConfig.merges) {
240
- // BPE with vocab + merges
241
- this.backend = new BPETokenizer(tokenizerConfig);
242
- (this.backend).load(tokenizerConfig.vocab, tokenizerConfig.merges);
243
- } else {
244
- throw new Error(
245
- `[Tokenizer] No valid tokenizer configuration in manifest for model "${modelId}". ` +
246
- 'Provide tokenizer.hfModel or bundle tokenizer.json (tokenizer.type="bundled", tokenizer.file="tokenizer.json").'
247
- );
248
- }
249
416
 
250
- this.config = tokenizerConfig;
417
+ this.config = tokenizerConfig;
418
+ finishTokenizerLoadTiming(loadTiming, 'complete', loadStart);
419
+ } catch (error) {
420
+ finishTokenizerLoadTiming(loadTiming, 'failed', loadStart, error);
421
+ throw error;
422
+ }
251
423
  }
252
424
  encode(text) {
253
425
  if (!this.backend) {
@@ -280,6 +452,10 @@ export class Tokenizer {
280
452
  }
281
453
  return this.backend.getHotTokenIds(limit);
282
454
  }
455
+
456
+ getLoadTiming() {
457
+ return cloneTokenizerLoadTiming(this.loadTiming);
458
+ }
283
459
  }
284
460
 
285
461
  export default Tokenizer;