doppler-gpu 0.5.1 → 0.6.0

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 (927) hide show
  1. package/CHANGELOG.md +33 -577
  2. package/LICENSE +21 -201
  3. package/README.md +165 -222
  4. package/models/catalog.json +241 -33
  5. package/package.json +99 -86
  6. package/src/capsule-runtime.d.ts +23 -0
  7. package/src/capsule-runtime.js +23 -0
  8. package/src/capsule.d.ts +8 -0
  9. package/src/capsule.js +4 -0
  10. package/src/cli/argument-contract.js +200 -0
  11. package/src/cli/browser-report-output.js +24 -0
  12. package/src/cli/cli-model-resolution.js +2 -2
  13. package/src/cli/cli-output.js +10 -0
  14. package/src/cli/commands/bundle.js +828 -0
  15. package/src/cli/doppler-cli.js +50 -1457
  16. package/src/cli/doppler-serve.js +18 -244
  17. package/src/cli/output/manifest.js +368 -0
  18. package/src/cli/runtime-profiles.js +1 -1
  19. package/src/cli/serve/capsule-handler.d.ts +33 -0
  20. package/src/cli/serve/capsule-handler.js +105 -0
  21. package/src/cli/serve/capsule-http.js +102 -0
  22. package/src/cli/serve/receipt.js +269 -0
  23. package/src/client/capsule-host.browser.js +1 -0
  24. package/src/client/capsule-host.d.ts +3 -0
  25. package/src/client/capsule-host.js +1 -0
  26. package/src/client/doppler-api.browser.js +42 -0
  27. package/src/client/doppler-api.d.ts +36 -1
  28. package/src/client/doppler-api.js +40 -1
  29. package/src/client/doppler-registry.js +15 -7
  30. package/src/client/electron/index.d.ts +4 -0
  31. package/src/client/electron/index.js +13 -0
  32. package/src/client/electron/ipc-contract.d.ts +9 -0
  33. package/src/client/electron/ipc-contract.js +56 -0
  34. package/src/client/electron/release-state.d.ts +53 -0
  35. package/src/client/electron/release-state.js +425 -0
  36. package/src/client/electron/renderer-runtime.d.ts +19 -0
  37. package/src/client/electron/renderer-runtime.js +104 -0
  38. package/src/client/failure-taxonomy.js +1 -28
  39. package/src/client/fault-injection.js +0 -8
  40. package/src/client/inspection.js +1 -1
  41. package/src/client/model-host/index.d.ts +140 -0
  42. package/src/client/model-host/index.js +548 -0
  43. package/src/client/model-host/model-evidence.js +255 -0
  44. package/src/client/model-host/model-session.d.ts +322 -0
  45. package/src/client/{runtime → model-host}/model-session.js +190 -148
  46. package/src/client/provider.d.ts +18 -2
  47. package/src/client/provider.js +26 -3
  48. package/src/client/receipt.js +30 -72
  49. package/src/client/runtime/capsule-acquisition.d.ts +18 -0
  50. package/src/client/runtime/capsule-acquisition.js +103 -0
  51. package/src/client/runtime/capsule-adapter-execution.d.ts +17 -0
  52. package/src/client/runtime/capsule-adapter-execution.js +63 -0
  53. package/src/client/runtime/capsule-artifact-source.js +62 -0
  54. package/src/client/runtime/capsule-embedding.d.ts +29 -0
  55. package/src/client/runtime/capsule-embedding.js +76 -0
  56. package/src/client/runtime/capsule-forecast-program.d.ts +2 -0
  57. package/src/client/runtime/capsule-forecast-program.js +85 -0
  58. package/src/client/runtime/capsule-forecast.d.ts +17 -0
  59. package/src/client/runtime/capsule-forecast.js +31 -0
  60. package/src/client/runtime/capsule-operation-adapters.d.ts +12 -0
  61. package/src/client/runtime/capsule-operation-adapters.js +68 -0
  62. package/src/client/runtime/capsule-operation-executor.d.ts +25 -0
  63. package/src/client/runtime/capsule-operation-executor.js +100 -0
  64. package/src/client/runtime/capsule-program-adapter.js +151 -0
  65. package/src/client/runtime/capsule-release-authorization.js +25 -0
  66. package/src/client/runtime/capsule-rerank.d.ts +50 -0
  67. package/src/client/runtime/capsule-rerank.js +105 -0
  68. package/src/client/runtime/capsule-session-execution.js +85 -0
  69. package/src/client/runtime/command-executor.d.ts +15 -0
  70. package/src/client/runtime/command-executor.js +128 -0
  71. package/src/client/runtime/composition-root.d.ts +74 -0
  72. package/src/client/runtime/composition-root.js +344 -0
  73. package/src/client/runtime/fetch-capsule-artifact-store.d.ts +10 -0
  74. package/src/client/runtime/fetch-capsule-artifact-store.js +25 -0
  75. package/src/client/runtime/index.d.ts +1 -117
  76. package/src/client/runtime/index.js +1 -433
  77. package/src/client/runtime/lora.d.ts +14 -0
  78. package/src/client/runtime/lora.js +19 -1
  79. package/src/client/runtime/model-manager.js +12 -0
  80. package/src/client/runtime/model-rerank.js +30 -0
  81. package/src/client/runtime/model-session.d.ts +1 -221
  82. package/src/client/runtime/model-source.d.ts +9 -0
  83. package/src/client/runtime/model-source.js +46 -17
  84. package/src/client/runtime/node-quickstart-cache.js +1 -1
  85. package/src/client/runtime/resolution-policy.d.ts +20 -0
  86. package/src/client/runtime/resolution-policy.js +84 -0
  87. package/src/client/runtime/resource-binder.d.ts +24 -0
  88. package/src/client/runtime/resource-binder.js +135 -0
  89. package/src/client/runtime/scoped-session.d.ts +10 -1
  90. package/src/client/runtime/scoped-session.js +20 -1
  91. package/src/client/runtime/session-controller.d.ts +30 -0
  92. package/src/client/runtime/session-controller.js +170 -0
  93. package/src/client/runtime/source-runtime.js +1 -1
  94. package/src/client/runtime/target-selector.d.ts +9 -0
  95. package/src/client/runtime/target-selector.js +2 -0
  96. package/src/client/runtime/types.d.ts +1 -0
  97. package/src/client/runtime/verified-capsule-artifact-store.js +72 -0
  98. package/src/client/wrap-pipeline-handle.js +3 -9
  99. package/src/config/backward-registry-loader.js +1 -1
  100. package/src/config/capsule-adapter-policy.js +30 -0
  101. package/src/config/capsule-adapters.d.ts +21 -0
  102. package/src/config/capsule-adapters.js +40 -0
  103. package/src/config/capsule-adapters.json +9 -0
  104. package/src/config/capsule-loading.d.ts +2 -0
  105. package/src/config/capsule-loading.js +14 -0
  106. package/src/config/capsule-loading.schema.json +11 -0
  107. package/src/config/capsule-operation.d.ts +19 -0
  108. package/src/config/capsule-operation.js +58 -0
  109. package/src/config/capsule-operations.json +18 -0
  110. package/src/config/capsule-release-contract.d.ts +66 -0
  111. package/src/config/capsule-release-contract.js +197 -0
  112. package/src/config/capsule-release-events.d.ts +53 -0
  113. package/src/config/capsule-release-events.js +172 -0
  114. package/src/config/capsule-sequence-receipt.js +19 -0
  115. package/src/config/capsule-serve.d.ts +9 -0
  116. package/src/config/capsule-serve.js +28 -0
  117. package/src/config/capsule-serve.schema.json +14 -0
  118. package/src/config/capsule-signature.d.ts +6 -0
  119. package/src/config/capsule-signature.js +50 -0
  120. package/src/config/capsule-v2.d.ts +70 -0
  121. package/src/config/capsule-v2.js +431 -0
  122. package/src/config/capsule-v3.d.ts +19 -0
  123. package/src/config/capsule-v3.js +72 -0
  124. package/src/config/capsule.d.ts +22 -0
  125. package/src/config/capsule.js +52 -0
  126. package/src/config/conversion/amplify/amplify-120m-f16-af32.json +1 -1
  127. package/src/config/conversion/embeddinggemma/google-embeddinggemma-300m-q4k-ehf16-af32.json +1 -1
  128. package/src/config/conversion/esm/esm2-t12-35m-ur50d-f32-af32.json +41 -1
  129. package/src/config/conversion/esmc/esmc-300m-f32-af32.json +1 -1
  130. package/src/config/conversion/gemma3/gemma-3-1b-it-q4k-ehf16-af32.json +1 -1
  131. package/src/config/conversion/gemma3/gemma-3-270m-it-f16-af32.json +1 -1
  132. package/src/config/conversion/gemma3/gemma-3-270m-it-q4k-ehf16-af16.json +1 -1
  133. package/src/config/conversion/gemma3/gemma-3-270m-it-q4k-ehf16-af32.json +1 -1
  134. package/src/config/conversion/gemma3/translategemma-4b-1b-enes-q4k-ehf16-af32.json +1 -1
  135. package/src/config/conversion/gemma3/translategemma-4b-it-q4k-ehf16-af32.json +1 -1
  136. package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-af16.json +2 -2
  137. package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-af32.json +1 -1
  138. package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-hq4k-af16.json +1 -1
  139. package/src/config/conversion/gemma4/gemma-4-12b-it-text-w4a16-ct-ehf16-af16.json +1 -1
  140. package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af16.json +2 -2
  141. package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af32.json +1 -1
  142. package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af16-int4ple.json +3 -3
  143. package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32-int4ple.json +1 -1
  144. package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32.json +1 -1
  145. package/src/config/conversion/glmocr/glm-ocr-f16-af32.json +442 -0
  146. package/src/config/conversion/minicpm/minicpm4-0-5b-f16-af32.json +1 -1
  147. package/src/config/conversion/muse-glimmer/muse-glimmer-30b-text-bf16-af32.json +529 -0
  148. package/src/config/conversion/muse-glimmer/muse-glimmer-30b-text-f16-af32.json +529 -0
  149. package/src/config/conversion/nucleotide-transformer/nucleotide-transformer-v2-50m-f32-af32.json +1 -1
  150. package/src/config/conversion/qwen3/qwen-3-5-0-8b-q4k-ehaf16.json +1 -1
  151. package/src/config/conversion/qwen3/qwen-3-5-2b-q4k-ehaf16.json +1 -1
  152. package/src/config/conversion/qwen3/qwen-3-5-9b-f16-af32.json +1 -1
  153. package/src/config/conversion/qwen3/qwen-3-5-9b-q4k-ehaf16.json +1 -1
  154. package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-eaf16.json +2 -2
  155. package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-ehaf16.json +1 -1
  156. package/src/config/conversion/qwen3/qwen-3-embedding-0-6b-f16-af32.json +463 -0
  157. package/src/config/conversion/qwen3/qwen-3-embedding-0-6b-q4k-ehf16-af32.json +1 -1
  158. package/src/config/conversion/qwen3/qwen-3-reranker-0-6b-f16-af32.json +1 -1
  159. package/src/config/conversion/qwen3/qwen-3-reranker-0-6b-f16-true-logit-af32.json +481 -0
  160. package/src/config/conversion/qwen3/qwen-3-reranker-0-6b-q4k-ehf16-af32.json +59 -7
  161. package/src/config/conversion/qwen3/qwen3.8-27b-text-f16-af32.json +561 -0
  162. package/src/config/conversion/qwen3/qwen3.8-27b-text-q4k-ehaf16.json +558 -0
  163. package/src/config/dtype-transition-contract.js +40 -0
  164. package/src/config/embedding-contract.js +26 -0
  165. package/src/config/embedding-reference.js +85 -0
  166. package/src/config/evidence/source-intake-policy.json +1 -1
  167. package/src/config/execution-contract-check.d.ts +16 -73
  168. package/src/config/execution-contract-check.js +7 -429
  169. package/src/config/forecast-manifest.js +56 -0
  170. package/src/config/heterogeneous-attention-contract.js +46 -0
  171. package/src/config/initial-execution-identity.d.ts +58 -0
  172. package/src/config/initial-execution-identity.js +331 -0
  173. package/src/config/kernel-path-loader.d.ts +10 -0
  174. package/src/config/kernel-path-loader.js +23 -8
  175. package/src/config/kernel-registry-contract.d.ts +25 -0
  176. package/src/config/kernel-registry-contract.js +46 -0
  177. package/src/config/kernels/kernel-ref-digests.js +50 -2
  178. package/src/config/kernels/registry.js +1 -1
  179. package/src/config/kernels/registry.json +13592 -11377
  180. package/src/config/lora-layouts.d.ts +10 -0
  181. package/src/config/lora-layouts.js +17 -0
  182. package/src/config/lora-layouts.json +9 -0
  183. package/src/config/manifest-inference-validation.js +575 -0
  184. package/src/config/merge/execution-patches.js +12 -0
  185. package/src/config/{merge-helpers.js → merge/kernel-path-policy.js} +0 -52
  186. package/src/config/merge/value-selection.js +38 -0
  187. package/src/config/merge.js +43 -11
  188. package/src/config/model-ir-v2.d.ts +57 -0
  189. package/src/config/model-ir-v2.js +258 -0
  190. package/src/config/model-ir.d.ts +49 -0
  191. package/src/config/model-ir.js +180 -0
  192. package/src/config/param-validator.js +1 -1
  193. package/src/config/platforms/amd-strix-halo-gfx1151.json +17 -0
  194. package/src/config/platforms/loader.js +2 -1
  195. package/src/config/production-release-evidence.d.ts +23 -0
  196. package/src/config/production-release-evidence.js +423 -0
  197. package/src/config/production-release.js +321 -0
  198. package/src/{client/doppler-registry.json → config/quickstart-registry.json} +7 -7
  199. package/src/config/required-inference-fields-contract-check.js +2 -2
  200. package/src/config/rerank-reference.d.ts +28 -0
  201. package/src/config/rerank-reference.js +89 -0
  202. package/src/config/revocation-policy.d.ts +65 -0
  203. package/src/config/revocation-policy.js +290 -0
  204. package/src/config/revocation-registry.json +11 -0
  205. package/src/config/revocation-updates.d.ts +68 -0
  206. package/src/config/revocation-updates.js +440 -0
  207. package/src/config/runtime/profiles/gemma4-31b-f16-activations-probe.json +1 -1
  208. package/src/config/runtime/profiles/muse-glimmer-30b-source-parity-f32kv.json +30 -0
  209. package/src/config/runtime/profiles/muse-glimmer-30b-source-parity.json +110 -0
  210. package/src/config/runtime/profiles/qwen3-6-27b-f16-activations-probe.json +1 -1
  211. package/src/config/runtime/profiles/qwen3-8-27b-source-parity.json +87 -0
  212. package/src/config/runtime-merge.js +1 -1
  213. package/src/config/runtime.js +1 -1
  214. package/src/config/schema/browser-suite-metrics.schema.d.ts +1 -0
  215. package/src/config/schema/browser-suite-metrics.schema.js +2 -0
  216. package/src/config/schema/conversion-report.schema.d.ts +3 -0
  217. package/src/config/schema/conversion-report.schema.js +29 -0
  218. package/src/config/schema/converter.schema.d.ts +1 -0
  219. package/src/config/schema/debug.schema.d.ts +1 -0
  220. package/src/config/schema/doppler.schema.js +5 -5
  221. package/src/config/schema/emulation.schema.js +1 -1
  222. package/src/config/schema/execution-v1.schema.d.ts +3 -0
  223. package/src/config/schema/execution-v1.schema.js +19 -0
  224. package/src/config/schema/harness.schema.d.ts +13 -0
  225. package/src/config/schema/harness.schema.js +6 -0
  226. package/src/config/schema/inference.schema.d.ts +21 -2
  227. package/src/config/schema/kernel-thresholds.schema.js +1 -1
  228. package/src/config/schema/manifest.schema.d.ts +26 -7
  229. package/src/config/schema/manifest.schema.js +10 -1
  230. package/src/config/schema/program-bundle.schema.d.ts +52 -6
  231. package/src/config/schema/program-bundle.schema.js +145 -10
  232. package/src/config/sequence-reference.d.ts +26 -0
  233. package/src/config/sequence-reference.js +100 -0
  234. package/src/config/target-plan.d.ts +108 -0
  235. package/src/config/target-plan.js +411 -0
  236. package/src/config/training-defaults.js +1 -1
  237. package/src/config/transforms/capability-transform-resolver.js +4 -105
  238. package/src/config/transforms/execution-graph/dtype.js +591 -0
  239. package/src/config/transforms/execution-graph/fusion.js +436 -0
  240. package/src/config/transforms/execution-graph/validation.js +414 -0
  241. package/src/config/transforms/execution-graph-transforms.js +13 -1419
  242. package/src/converter/artifact-identity.js +475 -0
  243. package/src/converter/conversion-plan.js +1 -1
  244. package/src/converter/core.js +94 -2096
  245. package/src/converter/forge-candidate-evaluation.js +271 -0
  246. package/src/converter/forge-qualification.js +162 -0
  247. package/src/converter/forge-stages.js +798 -0
  248. package/src/converter/manifest-builder.js +683 -0
  249. package/src/converter/quantization-info.js +4 -1
  250. package/src/converter/quantizer.js +58 -9
  251. package/src/converter/shard-packer.d.ts +1 -1
  252. package/src/converter/shard-packer.js +2 -2
  253. package/src/converter/tensor-transform.js +640 -0
  254. package/src/converter/tokenizer-contract.js +243 -0
  255. package/src/converter/tokenizer-manifest.js +243 -0
  256. package/src/debug/history.js +0 -8
  257. package/src/debug/index.js +0 -10
  258. package/src/debug/tensor.js +1 -1
  259. package/src/debug/trace.js +0 -7
  260. package/src/experimental/adapters/adapter-manager.js +2 -0
  261. package/src/experimental/adapters/adapter-manifest.d.ts +6 -0
  262. package/src/experimental/adapters/adapter-manifest.js +11 -0
  263. package/src/experimental/adapters/artifact-contract.d.ts +2 -2
  264. package/src/experimental/adapters/artifact-contract.js +5 -2
  265. package/src/experimental/adapters/lora-loader.d.ts +10 -1
  266. package/src/experimental/adapters/lora-loader.js +110 -48
  267. package/src/{inference/pipelines → experimental}/audio/gemma4.js +9 -31
  268. package/src/experimental/bdpa/basis-decomposed-paged.d.ts +50 -0
  269. package/src/experimental/bdpa/basis-decomposed-paged.js +276 -0
  270. package/src/experimental/bridge/trainer-artifact-bridge.js +3 -3
  271. package/src/experimental/browser/browser-converter.js +11 -11
  272. package/src/experimental/diffusion/index.d.ts +1 -1
  273. package/src/experimental/diffusion/index.js +1 -1
  274. package/src/{inference/pipelines → experimental}/diffusion/pipeline.d.ts +2 -2
  275. package/src/{inference/pipelines → experimental}/diffusion/pipeline.js +17 -17
  276. package/src/{inference/pipelines → experimental}/diffusion/text-encoder-gpu.js +21 -19
  277. package/src/experimental/diffusion/vae.js +596 -0
  278. package/src/experimental/distribution/shard-delivery/receipt.js +588 -0
  279. package/src/experimental/distribution/shard-delivery/retry.js +392 -0
  280. package/src/experimental/distribution/shard-delivery/transport.js +831 -0
  281. package/src/experimental/distribution/shard-delivery.js +4 -1709
  282. package/src/experimental/energy/index.d.ts +2 -2
  283. package/src/experimental/energy/index.js +2 -2
  284. package/src/{inference/pipelines → experimental}/energy/pipeline.js +13 -13
  285. package/src/{inference/pipelines → experimental}/energy/types.d.ts +2 -2
  286. package/src/experimental/logits/cpu-output.d.ts +27 -0
  287. package/src/{inference/pipelines/text/logits/utils.js → experimental/logits/cpu-output.js} +19 -27
  288. package/src/experimental/logits/cpu.d.ts +87 -0
  289. package/src/experimental/logits/cpu.js +95 -0
  290. package/src/experimental/orchestration/energy-row-head-pipeline.d.ts +103 -0
  291. package/src/{inference/pipelines/energy-head/row-head-pipeline.js → experimental/orchestration/energy-row-head-pipeline.js} +13 -13
  292. package/src/experimental/orchestration/index.d.ts +1 -1
  293. package/src/experimental/orchestration/index.js +1 -1
  294. package/src/experimental/training/artifacts.js +2 -2
  295. package/src/experimental/training/checkpoint.js +2 -2
  296. package/src/experimental/training/distillation/artifacts.js +1 -1
  297. package/src/experimental/training/distillation/dataset.js +2 -2
  298. package/src/experimental/training/distillation/eval.js +1 -1
  299. package/src/experimental/training/distillation/runtime.js +2 -2
  300. package/src/experimental/training/distillation/student-fixture/execution.js +166 -0
  301. package/src/experimental/training/distillation/student-fixture/model.js +353 -0
  302. package/src/experimental/training/distillation/student-fixture.js +2 -469
  303. package/src/experimental/training/lora/checkpoint.js +746 -0
  304. package/src/experimental/training/lora/execution.js +550 -0
  305. package/src/experimental/training/lora/plan.js +387 -0
  306. package/src/experimental/training/lora/recovery.js +418 -0
  307. package/src/experimental/training/lora-pipeline.js +7 -1949
  308. package/src/experimental/training/operator-artifacts.js +2 -2
  309. package/src/experimental/training/operator-command.js +4 -4
  310. package/src/experimental/training/runner/evaluation.js +62 -0
  311. package/src/experimental/training/runner/lifecycle.js +257 -0
  312. package/src/experimental/training/runner.js +5 -268
  313. package/src/experimental/training/suite/execution.js +879 -0
  314. package/src/experimental/training/suite/plan.js +855 -0
  315. package/src/experimental/training/suite.js +5 -1659
  316. package/src/experimental/training/workloads.d.ts +5 -5
  317. package/src/experimental/training/workloads.js +6 -6
  318. package/src/experimental/turboquant/codebook.js +470 -0
  319. package/src/formats/quantization-constants.d.ts +11 -0
  320. package/src/formats/quantization-constants.js +11 -0
  321. package/src/formats/rdrr/integrity.js +1 -1
  322. package/src/formats/rdrr/lowerings.js +0 -29
  323. package/src/formats/rdrr/merkle.js +1 -1
  324. package/src/formats/rdrr/tensor-config-validator.js +25 -5
  325. package/src/formats/rdrr/validation.js +3 -3
  326. package/src/gpu/command-recorder.d.ts +10 -1
  327. package/src/gpu/command-recorder.js +58 -14
  328. package/src/gpu/constant-buffer.js +29 -0
  329. package/src/gpu/device-state.js +36 -0
  330. package/src/gpu/device.js +6 -41
  331. package/src/gpu/kernel-runtime.js +2 -1
  332. package/src/gpu/kernels/activation-static-qdq.js +1 -1
  333. package/src/gpu/kernels/append_embedding.wgsl +14 -0
  334. package/src/gpu/kernels/attention/executors.js +960 -0
  335. package/src/gpu/kernels/attention/plan.js +472 -0
  336. package/src/gpu/kernels/attention.d.ts +4 -0
  337. package/src/gpu/kernels/attention.js +145 -1545
  338. package/src/gpu/kernels/attention_relative.wgsl +59 -0
  339. package/src/gpu/kernels/backward/adam.js +2 -1
  340. package/src/gpu/kernels/backward/attention_backward.js +1 -1
  341. package/src/gpu/kernels/backward/bias_add_backward.d.ts +1 -1
  342. package/src/gpu/kernels/backward/bias_add_backward.js +1 -1
  343. package/src/gpu/kernels/backward/causal_conv1d_silu_backward.js +2 -1
  344. package/src/gpu/kernels/backward/conv2d_backward.d.ts +1 -1
  345. package/src/gpu/kernels/backward/conv2d_backward.js +2 -1
  346. package/src/gpu/kernels/backward/cross_entropy_backward.d.ts +1 -1
  347. package/src/gpu/kernels/backward/cross_entropy_backward.js +1 -1
  348. package/src/gpu/kernels/backward/embed_backward.js +2 -1
  349. package/src/gpu/kernels/backward/executor.d.ts +34 -0
  350. package/src/gpu/kernels/backward/executor.js +158 -0
  351. package/src/gpu/kernels/backward/gated_delta_recurrent_backward.js +2 -1
  352. package/src/gpu/kernels/backward/gated_rmsnorm_backward.js +2 -1
  353. package/src/gpu/kernels/backward/gelu_backward.d.ts +1 -1
  354. package/src/gpu/kernels/backward/gelu_backward.js +1 -1
  355. package/src/gpu/kernels/backward/groupnorm_backward.d.ts +1 -1
  356. package/src/gpu/kernels/backward/groupnorm_backward.js +1 -1
  357. package/src/gpu/kernels/backward/layernorm_backward.d.ts +1 -1
  358. package/src/gpu/kernels/backward/layernorm_backward.js +2 -1
  359. package/src/gpu/kernels/backward/matmul.js +133 -0
  360. package/src/gpu/kernels/backward/matmul_backward.d.ts +0 -1
  361. package/src/gpu/kernels/backward/matmul_backward.js +1 -1
  362. package/src/gpu/kernels/backward/pixel_shuffle_backward.d.ts +1 -1
  363. package/src/gpu/kernels/backward/pixel_shuffle_backward.js +1 -1
  364. package/src/gpu/kernels/backward/rmsnorm_backward.d.ts +1 -1
  365. package/src/gpu/kernels/backward/rmsnorm_backward.js +2 -1
  366. package/src/gpu/kernels/backward/rope_backward.d.ts +1 -1
  367. package/src/gpu/kernels/backward/rope_backward.js +2 -1
  368. package/src/gpu/kernels/backward/scale_backward.d.ts +1 -1
  369. package/src/gpu/kernels/backward/scale_backward.js +2 -1
  370. package/src/gpu/kernels/backward/silu_backward.d.ts +1 -1
  371. package/src/gpu/kernels/backward/silu_backward.js +1 -1
  372. package/src/gpu/kernels/backward/softmax_backward.d.ts +1 -1
  373. package/src/gpu/kernels/backward/softmax_backward.js +1 -1
  374. package/src/gpu/kernels/backward/upsample2d_backward.d.ts +1 -1
  375. package/src/gpu/kernels/backward/upsample2d_backward.js +1 -1
  376. package/src/gpu/kernels/backward/utils.js +2 -392
  377. package/src/gpu/kernels/bind-group-validation.js +16 -0
  378. package/src/gpu/kernels/cast.js +2 -1
  379. package/src/gpu/kernels/check-finiteness.js +13 -49
  380. package/src/gpu/kernels/check-hot-vocab-stop.js +13 -39
  381. package/src/gpu/kernels/check-stop.js +14 -33
  382. package/src/gpu/kernels/check_finiteness.wgsl +33 -0
  383. package/src/gpu/kernels/check_hot_vocab_stop.wgsl +29 -0
  384. package/src/gpu/kernels/check_stop.wgsl +20 -0
  385. package/src/gpu/kernels/clamp.js +1 -1
  386. package/src/gpu/kernels/conv2d.js +1 -1
  387. package/src/gpu/kernels/cross_entropy_loss.js +1 -1
  388. package/src/gpu/kernels/depthwise_conv1d.js +1 -1
  389. package/src/gpu/kernels/depthwise_conv2d.js +1 -1
  390. package/src/gpu/kernels/dequant.js +7 -2
  391. package/src/gpu/kernels/diffusion-gemma-sampling.js +3 -5
  392. package/src/gpu/kernels/dispatch.js +12 -1
  393. package/src/gpu/kernels/embedding-pool.js +72 -0
  394. package/src/gpu/kernels/embedding_pool_f16.wgsl +38 -0
  395. package/src/gpu/kernels/embedding_pool_f32.wgsl +36 -0
  396. package/src/gpu/kernels/energy.js +2 -1
  397. package/src/gpu/kernels/forecast_output.wgsl +17 -0
  398. package/src/gpu/kernels/fused-gate-up-gelu.js +1 -1
  399. package/src/gpu/kernels/fused_ffn.js +10 -220
  400. package/src/gpu/kernels/fused_matmul_residual.js +2 -1
  401. package/src/gpu/kernels/fused_matmul_rmsnorm.js +2 -1
  402. package/src/gpu/kernels/fused_normed_ffn_q4k.wgsl +208 -0
  403. package/src/gpu/kernels/gated-short-conv.js +11 -69
  404. package/src/gpu/kernels/gated_delta_conv.wgsl +66 -0
  405. package/src/gpu/kernels/gated_delta_conv_f16.wgsl +68 -0
  406. package/src/gpu/kernels/gated_delta_fused_decode.wgsl +173 -0
  407. package/src/gpu/kernels/gated_delta_fused_decode_f16.wgsl +175 -0
  408. package/src/gpu/kernels/gated_delta_recurrent.wgsl +165 -0
  409. package/src/gpu/kernels/gated_delta_recurrent_f16.wgsl +167 -0
  410. package/src/gpu/kernels/gated_short_conv.wgsl +47 -0
  411. package/src/gpu/kernels/gather.d.ts +1 -1
  412. package/src/gpu/kernels/gather.js +30 -6
  413. package/src/gpu/kernels/gather_bf16.wgsl +69 -0
  414. package/src/gpu/kernels/gelu.js +1 -1
  415. package/src/gpu/kernels/gemma4-route-expert.js +1 -1
  416. package/src/gpu/kernels/grouped_pointwise_conv2d.js +1 -1
  417. package/src/gpu/kernels/groupnorm.js +1 -1
  418. package/src/gpu/kernels/index.d.ts +38 -5
  419. package/src/gpu/kernels/index.js +18 -4
  420. package/src/gpu/kernels/instance_stats.wgsl +26 -0
  421. package/src/gpu/kernels/kernel-base.js +1 -1
  422. package/src/gpu/kernels/kernel-cache-lifecycle.d.ts +9 -0
  423. package/src/gpu/kernels/kernel-cache-lifecycle.js +26 -0
  424. package/src/gpu/kernels/kernel-configs.d.ts +10 -61
  425. package/src/gpu/kernels/kernel-configs.js +26 -54
  426. package/src/gpu/kernels/kernel-execution.js +141 -0
  427. package/src/gpu/kernels/kv-cache-write.js +4 -4
  428. package/src/gpu/kernels/kv-quantize.js +4 -1
  429. package/src/gpu/kernels/l2-normalize.js +36 -0
  430. package/src/gpu/kernels/l2_normalize.wgsl +53 -0
  431. package/src/gpu/kernels/layernorm.d.ts +4 -4
  432. package/src/gpu/kernels/layernorm.js +44 -7
  433. package/src/gpu/kernels/linear-attention-core.js +29 -449
  434. package/src/gpu/kernels/linear_attention.js +1 -1
  435. package/src/gpu/kernels/lm-head-argmax.js +2 -2
  436. package/src/gpu/kernels/lm-head-select-logits.js +2 -2
  437. package/src/gpu/kernels/logit-finalize.js +103 -0
  438. package/src/gpu/kernels/logit-merge.js +13 -125
  439. package/src/gpu/kernels/logit-scatter.js +40 -0
  440. package/src/gpu/kernels/logit_finalize_f16.wgsl +43 -0
  441. package/src/gpu/kernels/logit_finalize_f32.wgsl +41 -0
  442. package/src/gpu/kernels/logit_merge_geometric.wgsl +23 -0
  443. package/src/gpu/kernels/logit_merge_max.wgsl +23 -0
  444. package/src/gpu/kernels/logit_merge_weighted.wgsl +23 -0
  445. package/src/gpu/kernels/logit_scatter_f16.wgsl +26 -0
  446. package/src/gpu/kernels/logit_scatter_f32.wgsl +24 -0
  447. package/src/gpu/kernels/matmul-dispatch.js +6 -1
  448. package/src/gpu/kernels/matmul-selection.js +3 -2
  449. package/src/gpu/kernels/matmul.js +4 -15
  450. package/src/gpu/kernels/matmul_bf16w_f32a.wgsl +93 -0
  451. package/src/gpu/kernels/matmul_gemv_subgroup_bf16w.wgsl +433 -0
  452. package/src/gpu/kernels/modulate.js +1 -1
  453. package/src/gpu/kernels/moe.js +3 -1
  454. package/src/gpu/kernels/moe_router_bias_add_f16_f16.wgsl +23 -0
  455. package/src/gpu/kernels/moe_router_bias_add_f16_f32.wgsl +23 -0
  456. package/src/gpu/kernels/moe_router_bias_add_f32_f16.wgsl +23 -0
  457. package/src/gpu/kernels/moe_router_bias_add_f32_f32.wgsl +21 -0
  458. package/src/gpu/kernels/patch_context.wgsl +29 -0
  459. package/src/gpu/kernels/pipeline-cache.js +3 -1
  460. package/src/gpu/kernels/pixel_shuffle.js +1 -1
  461. package/src/gpu/kernels/probe_shader_f16.wgsl +7 -0
  462. package/src/gpu/kernels/relu.js +1 -1
  463. package/src/gpu/kernels/repeat_channels.js +1 -1
  464. package/src/gpu/kernels/residual.js +2 -2
  465. package/src/gpu/kernels/rmsnorm-pair.js +1 -1
  466. package/src/gpu/kernels/rmsnorm-qk.js +1 -1
  467. package/src/gpu/kernels/rmsnorm-stats.js +14 -113
  468. package/src/gpu/kernels/rmsnorm.js +1 -1
  469. package/src/gpu/kernels/rmsnorm_serial.wgsl +19 -0
  470. package/src/gpu/kernels/rmsnorm_stats.wgsl +59 -0
  471. package/src/gpu/kernels/rmsnorm_stats_subgroups.wgsl +76 -0
  472. package/src/gpu/kernels/rope-precompute.js +160 -0
  473. package/src/gpu/kernels/rope-qk.js +1 -1
  474. package/src/gpu/kernels/rope.js +2 -2
  475. package/src/gpu/kernels/rope_precompute.wgsl +74 -0
  476. package/src/gpu/kernels/sample.js +5 -1
  477. package/src/gpu/kernels/scale.js +1 -1
  478. package/src/gpu/kernels/shader-cache.js +8 -4
  479. package/src/gpu/kernels/shader-source-scope.js +68 -0
  480. package/src/gpu/kernels/silu.js +2 -1
  481. package/src/gpu/kernels/soft-embedding.js +3 -1
  482. package/src/gpu/kernels/softmax.js +4 -2
  483. package/src/gpu/kernels/split-qkv-rmsnorm-qk.js +1 -1
  484. package/src/gpu/kernels/split-qkv-rmsnorm-rope-qk.js +1 -1
  485. package/src/gpu/kernels/split_qg.js +1 -1
  486. package/src/gpu/kernels/split_qkv.d.ts +2 -0
  487. package/src/gpu/kernels/split_qkv.js +18 -3
  488. package/src/gpu/kernels/split_qkv.wgsl +5 -2
  489. package/src/gpu/kernels/split_qkv_f16.wgsl +5 -2
  490. package/src/gpu/kernels/submit_probe.wgsl +6 -0
  491. package/src/gpu/kernels/transpose.js +1 -1
  492. package/src/gpu/kernels/turboquant-codebook.js +4 -535
  493. package/src/gpu/kernels/uniform-utils.js +7 -2
  494. package/src/gpu/kernels/upsample2d.js +1 -1
  495. package/src/gpu/kernels/utils.js +9 -330
  496. package/src/gpu/kernels/vision-average-pool.d.ts +13 -0
  497. package/src/gpu/kernels/vision-average-pool.js +68 -0
  498. package/src/gpu/kernels/vision-patch-embed.d.ts +17 -0
  499. package/src/gpu/kernels/vision-patch-embed.js +117 -0
  500. package/src/gpu/kernels/vision-position-embedding.d.ts +13 -0
  501. package/src/gpu/kernels/vision-position-embedding.js +79 -0
  502. package/src/gpu/kernels/vision-rope-2d.d.ts +16 -0
  503. package/src/gpu/kernels/vision-rope-2d.js +73 -0
  504. package/src/gpu/kernels/vision-spatial-merge.d.ts +19 -0
  505. package/src/gpu/kernels/vision-spatial-merge.js +100 -0
  506. package/src/gpu/kernels/vision_average_pool.wgsl +42 -0
  507. package/src/gpu/kernels/vision_patch_embed.wgsl +59 -0
  508. package/src/gpu/kernels/vision_patch_embed_f16w.wgsl +61 -0
  509. package/src/gpu/kernels/vision_position_embedding.wgsl +33 -0
  510. package/src/gpu/kernels/vision_position_embedding_f16.wgsl +35 -0
  511. package/src/gpu/kernels/vision_rope_2d.wgsl +60 -0
  512. package/src/gpu/kernels/vision_spatial_merge.wgsl +48 -0
  513. package/src/gpu/readback-buffer.js +20 -0
  514. package/src/gpu/submit-probe.js +2 -26
  515. package/src/gpu/weight-buffer.d.ts +1 -0
  516. package/src/gpu/weight-buffer.js +7 -0
  517. package/src/index.d.ts +10 -0
  518. package/src/index.js +1 -0
  519. package/src/inference/browser-harness/benchmark-execution.js +872 -0
  520. package/src/inference/browser-harness/benchmark-rerank.js +284 -0
  521. package/src/inference/browser-harness/decode-diagnostics.js +241 -0
  522. package/src/inference/browser-harness/diffusion-performance.js +102 -0
  523. package/src/inference/browser-harness/inference-execution.js +543 -0
  524. package/src/inference/{browser-harness-model-helpers.js → browser-harness/model-initialization.js} +14 -145
  525. package/src/inference/browser-harness/model-resolution.js +137 -0
  526. package/src/inference/browser-harness/report.js +338 -0
  527. package/src/inference/browser-harness/request.js +263 -0
  528. package/src/inference/browser-harness/runtime-config.d.ts +25 -0
  529. package/src/inference/browser-harness/runtime-config.js +255 -0
  530. package/src/inference/browser-harness/runtime-isolation.d.ts +13 -0
  531. package/src/inference/browser-harness/runtime-isolation.js +45 -0
  532. package/src/inference/browser-harness/runtime-report.d.ts +26 -0
  533. package/src/inference/browser-harness/runtime-report.js +117 -0
  534. package/src/inference/browser-harness/suite-summary.js +39 -0
  535. package/src/inference/browser-harness/text-evidence.js +694 -0
  536. package/src/inference/browser-harness/text-execution.js +757 -0
  537. package/src/inference/browser-harness/text-input.js +725 -0
  538. package/src/inference/{browser-harness-suite-helpers.js → browser-harness/timing-diagnostics.js} +1 -379
  539. package/src/inference/browser-harness-contract-helpers.js +1 -1
  540. package/src/inference/browser-harness-diffusion-energy-suites.js +8 -3
  541. package/src/inference/browser-harness-runtime-helpers.d.ts +3 -56
  542. package/src/inference/browser-harness-runtime-helpers.js +3 -423
  543. package/src/inference/browser-harness-text-helpers.js +31 -2199
  544. package/src/inference/browser-harness.d.ts +15 -2
  545. package/src/inference/browser-harness.js +38 -1952
  546. package/src/inference/kv-cache/base.js +67 -381
  547. package/src/inference/kv-cache/basis-decomposed-paged.d.ts +2 -50
  548. package/src/inference/kv-cache/basis-decomposed-paged.js +1 -276
  549. package/src/inference/kv-cache/quantized.d.ts +2 -2
  550. package/src/inference/kv-cache/quantized.js +4 -8
  551. package/src/inference/kv-cache/update.js +309 -0
  552. package/src/inference/model-checkpoint-evidence.d.ts +75 -0
  553. package/src/inference/model-checkpoint-evidence.js +264 -0
  554. package/src/inference/moe-router.js +36 -216
  555. package/src/inference/multi-model-network.js +5 -341
  556. package/src/inference/pipelines/audio/index.js +1 -20
  557. package/src/inference/pipelines/audio/mel.js +0 -38
  558. package/src/inference/pipelines/diffusion/profile-timing.js +10 -0
  559. package/src/inference/pipelines/diffusion/runtime-resources.js +54 -0
  560. package/src/inference/pipelines/diffusion/sd3/execution.js +805 -0
  561. package/src/inference/pipelines/diffusion/sd3/plan.js +125 -0
  562. package/src/inference/pipelines/diffusion/sd3-transformer.js +44 -895
  563. package/src/inference/pipelines/diffusion/{vae.js → vae/execution.js} +46 -618
  564. package/src/inference/pipelines/diffusion/{helpers.js → weight-contract.js} +1 -68
  565. package/src/inference/pipelines/energy-head/row-head-pipeline.d.ts +1 -103
  566. package/src/inference/pipelines/factory.js +14 -3
  567. package/src/inference/pipelines/shader-scoped-pipeline.js +29 -0
  568. package/src/inference/pipelines/structured/json-grammar-mask.js +0 -6
  569. package/src/inference/pipelines/text/abort-contract.js +22 -0
  570. package/src/inference/pipelines/text/attention/attn-config.js +1 -16
  571. package/src/inference/pipelines/text/attention/gate-projection.js +30 -0
  572. package/src/inference/pipelines/text/attention/heterogeneous-contract.js +8 -0
  573. package/src/inference/pipelines/text/attention/interpreter/immediate.js +211 -0
  574. package/src/inference/pipelines/text/attention/interpreter/recorded.js +451 -0
  575. package/src/inference/pipelines/text/attention/interpreter.js +89 -439
  576. package/src/inference/pipelines/text/attention/plan.d.ts +1 -0
  577. package/src/inference/pipelines/text/attention/plan.js +17 -0
  578. package/src/inference/pipelines/text/attention/projection-executor.js +331 -0
  579. package/src/inference/pipelines/text/attention/projection-plan.js +259 -0
  580. package/src/inference/pipelines/text/attention/projections.js +20 -537
  581. package/src/inference/pipelines/text/attention/query-transform.js +11 -0
  582. package/src/inference/pipelines/text/attention/record.js +0 -6
  583. package/src/inference/pipelines/text/attention/recorded-health.js +21 -0
  584. package/src/inference/pipelines/text/attention/rope-observation.js +40 -0
  585. package/src/inference/pipelines/text/attention/run.js +0 -6
  586. package/src/inference/pipelines/text/attention/types.js +3 -31
  587. package/src/inference/pipelines/text/chat-format.d.ts +3 -0
  588. package/src/inference/pipelines/text/chat-format.js +42 -0
  589. package/src/inference/pipelines/text/config/normalization.js +729 -0
  590. package/src/inference/pipelines/text/config/validation.js +9 -0
  591. package/src/inference/pipelines/text/config.d.ts +12 -1
  592. package/src/inference/pipelines/text/config.js +14 -1228
  593. package/src/inference/pipelines/text/debug-utils/health-metrics.js +57 -0
  594. package/src/inference/pipelines/text/debug-utils/index.js +3 -8
  595. package/src/inference/pipelines/text/debug-utils/profiles.js +8 -0
  596. package/src/inference/pipelines/text/debug-utils/readback-decoding.js +24 -0
  597. package/src/inference/pipelines/text/debug-utils/tensor.js +1 -1
  598. package/src/inference/pipelines/text/dtype-contract.js +1 -68
  599. package/src/inference/pipelines/text/embed.js +178 -241
  600. package/src/inference/pipelines/text/embedding-contract.d.ts +7 -0
  601. package/src/inference/pipelines/text/embedding-contract.js +21 -0
  602. package/src/inference/pipelines/text/embedding-extraction.js +225 -0
  603. package/src/inference/pipelines/text/embedding-normalization.js +39 -0
  604. package/src/inference/pipelines/text/execution-runtime-builders.js +0 -32
  605. package/src/inference/pipelines/text/execution-v1/plan.js +69 -0
  606. package/src/inference/pipelines/text/execution-v1/validation.js +233 -0
  607. package/src/inference/pipelines/text/execution-v1.js +7 -283
  608. package/src/inference/pipelines/text/execution.js +517 -0
  609. package/src/inference/pipelines/text/ffn/dense-executor.js +718 -0
  610. package/src/inference/pipelines/text/ffn/dense-plan.js +327 -0
  611. package/src/inference/pipelines/text/ffn/dense.js +4 -968
  612. package/src/inference/pipelines/text/ffn/sandwich.js +10 -9
  613. package/src/inference/pipelines/text/ffn/standard.js +2 -1
  614. package/src/inference/pipelines/text/generator/attention-lifecycle.js +24 -0
  615. package/src/inference/pipelines/text/generator/decode-runtime.js +991 -0
  616. package/src/inference/pipelines/text/generator/decode.js +952 -0
  617. package/src/inference/pipelines/text/generator/diffusion.js +607 -0
  618. package/src/inference/pipelines/text/generator/logits-config.js +63 -0
  619. package/src/inference/pipelines/text/generator/prefill-observation.js +29 -0
  620. package/src/inference/pipelines/text/generator/prefill-policy.js +56 -0
  621. package/src/inference/pipelines/text/generator/prefill-runtime.js +896 -0
  622. package/src/inference/pipelines/text/{generator-prefill-helpers.js → generator/prefix-embedding.js} +16 -121
  623. package/src/inference/pipelines/text/generator/prompt-input.js +47 -0
  624. package/src/inference/pipelines/text/generator/recovery.js +310 -0
  625. package/src/inference/pipelines/text/{generator-helpers.js → generator/session-context.js} +42 -110
  626. package/src/inference/pipelines/text/generator/text.js +448 -0
  627. package/src/inference/pipelines/text/generator/training.js +752 -0
  628. package/src/inference/pipelines/text/generator-decode-policy.js +1 -1
  629. package/src/inference/pipelines/text/generator-logits-observation.js +45 -0
  630. package/src/inference/pipelines/text/generator-runtime.js +5 -135
  631. package/src/inference/pipelines/text/generator-steps.js +30 -972
  632. package/src/inference/pipelines/text/generator.js +165 -3544
  633. package/src/inference/pipelines/text/image-transcription.js +304 -0
  634. package/src/inference/pipelines/text/init-qkv-fusion.js +18 -9
  635. package/src/inference/pipelines/text/init-rope.js +37 -189
  636. package/src/inference/pipelines/text/init.d.ts +8 -3
  637. package/src/inference/pipelines/text/init.js +6 -216
  638. package/src/inference/pipelines/text/layer-execution-contract.js +97 -0
  639. package/src/inference/pipelines/text/layer-input-execution.js +435 -0
  640. package/src/inference/pipelines/text/layer-plan-gpu.js +13 -6
  641. package/src/inference/pipelines/text/layer-plan.d.ts +21 -5
  642. package/src/inference/pipelines/text/layer-plan.js +45 -3
  643. package/src/inference/pipelines/text/layer.js +52 -522
  644. package/src/inference/pipelines/text/lifecycle.js +839 -0
  645. package/src/inference/pipelines/text/linear-attention/plan.js +554 -0
  646. package/src/inference/pipelines/text/linear-attention-ab-fusion.js +2 -2
  647. package/src/inference/pipelines/text/linear-attention.d.ts +2 -2
  648. package/src/inference/pipelines/text/linear-attention.js +7 -537
  649. package/src/inference/pipelines/text/logits/cpu-output.d.ts +1 -0
  650. package/src/inference/pipelines/text/logits/cpu-output.js +5 -0
  651. package/src/inference/pipelines/text/logits/cpu.d.ts +1 -87
  652. package/src/inference/pipelines/text/logits/cpu.js +8 -95
  653. package/src/inference/pipelines/text/logits/gpu-executor.js +368 -0
  654. package/src/inference/pipelines/text/logits/gpu.d.ts +27 -0
  655. package/src/inference/pipelines/text/logits/gpu.js +116 -579
  656. package/src/inference/pipelines/text/logits/index.d.ts +2 -1
  657. package/src/inference/pipelines/text/logits/index.js +62 -102
  658. package/src/inference/pipelines/text/logits/output-transform.js +47 -0
  659. package/src/inference/pipelines/text/logits/plan.js +184 -0
  660. package/src/inference/pipelines/text/logits/readback.d.ts +6 -0
  661. package/src/inference/pipelines/text/logits/readback.js +9 -0
  662. package/src/inference/pipelines/text/logits/scale-policy.js +20 -0
  663. package/src/inference/pipelines/text/logits/types.d.ts +1 -0
  664. package/src/inference/pipelines/text/lora-apply.js +12 -4
  665. package/src/inference/pipelines/text/lora-types.d.ts +12 -0
  666. package/src/inference/pipelines/text/modality-token-contract.js +115 -0
  667. package/src/inference/pipelines/text/model-load.js +0 -20
  668. package/src/inference/pipelines/text/moe/expert-loading.js +9 -0
  669. package/src/inference/pipelines/text/moe/gpu-executor.js +513 -0
  670. package/src/inference/pipelines/text/moe/plan.js +47 -0
  671. package/src/inference/pipelines/text/moe-cpu.js +9 -135
  672. package/src/inference/pipelines/text/moe-gpu.js +21 -545
  673. package/src/inference/pipelines/text/normalization-contract.js +17 -0
  674. package/src/inference/pipelines/text/per-layer/materialize.js +736 -0
  675. package/src/inference/pipelines/text/per-layer/plan.js +362 -0
  676. package/src/inference/pipelines/text/per-layer-inputs.js +4 -1037
  677. package/src/inference/pipelines/text/pipeline-load-timing.js +77 -0
  678. package/src/inference/pipelines/text/probes.js +20 -0
  679. package/src/inference/pipelines/text/state.d.ts +8 -0
  680. package/src/inference/pipelines/text/state.js +1 -0
  681. package/src/inference/pipelines/text/tsir-fixture-writer.js +88 -16
  682. package/src/inference/pipelines/text/types.d.ts +4 -0
  683. package/src/inference/pipelines/text/weight-loading-config.js +194 -0
  684. package/src/inference/pipelines/text.d.ts +2 -2
  685. package/src/inference/pipelines/text.js +46 -1633
  686. package/src/inference/pipelines/video/frame-sampler.js +0 -13
  687. package/src/inference/pipelines/video/index.js +0 -21
  688. package/src/inference/pipelines/vision/attention.js +59 -0
  689. package/src/inference/pipelines/vision/encoder.js +254 -369
  690. package/src/inference/pipelines/vision/gemma4.js +204 -318
  691. package/src/inference/pipelines/vision/glmocr-preprocess.js +273 -0
  692. package/src/inference/pipelines/vision/glmocr-rope.js +106 -0
  693. package/src/inference/pipelines/vision/glmocr.js +557 -0
  694. package/src/inference/pipelines/vision/image-preprocess.js +0 -27
  695. package/src/inference/pipelines/vision/index.js +8 -27
  696. package/src/inference/pipelines/vision/ops.js +11 -33
  697. package/src/inference/pipelines/vision/patch-embed.js +51 -145
  698. package/src/inference/pipelines/vision/spatial-merge.js +78 -0
  699. package/src/inference/pipelines/vision/weight-loading.js +79 -0
  700. package/src/inference/refactor-receipt.js +2 -2
  701. package/src/inference/rerank.js +182 -0
  702. package/src/inference/tokenizer.js +1 -1
  703. package/src/inference/tokenizers/behavior-flags.js +6 -3
  704. package/src/inference/tokenizers/bundled/decoders.js +133 -0
  705. package/src/inference/tokenizers/bundled/registry.js +216 -0
  706. package/src/inference/tokenizers/bundled.js +23 -353
  707. package/src/loader/doppler-loader.d.ts +4 -0
  708. package/src/loader/doppler-loader.js +119 -930
  709. package/src/loader/embedding-limit-preflight.js +11 -4
  710. package/src/loader/embedding-loader.js +1 -3
  711. package/src/loader/gpu-tensor-materialization.js +20 -0
  712. package/src/loader/layer-loader.js +15 -119
  713. package/src/loader/layer-tensor-aliases.js +99 -0
  714. package/src/loader/layer-weight-shape.js +29 -0
  715. package/src/loader/load-timing.js +64 -0
  716. package/src/loader/loader-types.d.ts +3 -0
  717. package/src/loader/model-executor.js +819 -0
  718. package/src/loader/model-load-validation.js +15 -0
  719. package/src/loader/quantization-constants.d.ts +11 -29
  720. package/src/loader/quantization-constants.js +1 -2
  721. package/src/loader/shard-resolver.js +0 -14
  722. package/src/loader/tensor-load-error.js +14 -0
  723. package/src/loader/tensor-shard-indices.js +8 -0
  724. package/src/loader/tensors/bf16-loader.js +125 -0
  725. package/src/loader/tensors/functional-descriptor-loader.js +3 -375
  726. package/src/loader/tensors/load-executor.js +577 -0
  727. package/src/loader/tensors/source-transform/executor.js +123 -0
  728. package/src/loader/tensors/source-transform/plan.js +107 -0
  729. package/src/loader/tensors/source-transform.js +15 -264
  730. package/src/loader/tensors/tensor-loader.d.ts +4 -4
  731. package/src/loader/tensors/tensor-loader.js +7 -653
  732. package/src/memory/buffer-pool.js +7 -29
  733. package/src/rules/capability-transform-resolver.js +80 -0
  734. package/src/rules/execution-contract.d.ts +73 -0
  735. package/src/rules/execution-contract.js +429 -0
  736. package/src/rules/execution-rules-contract-check.js +1 -1
  737. package/src/rules/generated/rule-bundle.json +4305 -0
  738. package/src/rules/inference/capability-transforms.rules.json +3 -3
  739. package/src/rules/layer-pattern-contract-check.js +1 -1
  740. package/src/rules/rule-registry.js +70 -57
  741. package/src/storage/artifact-storage-context.js +2 -2
  742. package/src/storage/backend-selection.js +43 -0
  743. package/src/storage/distribution-transport.js +5 -0
  744. package/src/storage/download/integrity.d.ts +33 -0
  745. package/src/storage/download/integrity.js +130 -0
  746. package/src/storage/download/resume-inspection.d.ts +22 -0
  747. package/src/storage/download/resume-inspection.js +134 -0
  748. package/src/storage/download/retry.js +57 -0
  749. package/src/storage/download/state.js +131 -0
  750. package/src/storage/download/transport.js +99 -0
  751. package/src/storage/downloader.js +74 -453
  752. package/src/storage/model-cache.d.ts +1 -0
  753. package/src/storage/model-cache.js +51 -11
  754. package/src/storage/shard-manager.d.ts +5 -0
  755. package/src/storage/shard-manager.js +50 -369
  756. package/src/storage/shards/index.js +60 -0
  757. package/src/storage/shards/integrity.js +225 -0
  758. package/src/storage/shards/lifecycle.js +22 -0
  759. package/src/storage/source-artifact-store.js +14 -4
  760. package/src/tooling/boundary-evidence.d.ts +4 -4
  761. package/src/tooling/boundary-evidence.js +20 -20
  762. package/src/tooling/browser-command-runner.js +13 -6
  763. package/src/tooling/capsule-v2.d.ts +11 -0
  764. package/src/tooling/capsule-v2.js +58 -0
  765. package/src/tooling/capsule.js +20 -0
  766. package/src/tooling/command-api-constants.js +1 -1
  767. package/src/tooling/command-api-family-normalizers.js +163 -5
  768. package/src/tooling/command-api.d.ts +42 -3
  769. package/src/tooling/command-api.js +7 -2
  770. package/src/tooling/command-envelope.js +1 -1
  771. package/src/tooling/{command-api-helpers.js → command-request-normalization.js} +1 -1
  772. package/src/tooling/command-runner.html +49 -1
  773. package/src/tooling/command-validation.js +1 -24
  774. package/src/tooling/conversion-run-timing.js +24 -0
  775. package/src/tooling/diagnose-runner.js +11 -9
  776. package/src/tooling/distribution-shard-transport.js +11 -0
  777. package/src/tooling/execution-cost-ledger.d.ts +76 -2
  778. package/src/tooling/execution-cost-ledger.js +197 -15
  779. package/src/tooling/hf-registry/catalog-io.js +63 -0
  780. package/src/tooling/litert/package.js +863 -0
  781. package/src/tooling/litert-package-runtime.js +3 -849
  782. package/src/tooling/model-capsule-forge.js +478 -0
  783. package/src/tooling/node-browser/execution.d.ts +40 -0
  784. package/src/tooling/node-browser/execution.js +308 -0
  785. package/src/tooling/node-browser/static-server.js +188 -0
  786. package/src/tooling/node-browser/transport.d.ts +80 -0
  787. package/src/tooling/node-browser/transport.js +622 -0
  788. package/src/tooling/node-browser-command-runner.d.ts +18 -78
  789. package/src/tooling/node-browser-command-runner.js +10 -1112
  790. package/src/tooling/node-capsule-artifact-store.js +44 -0
  791. package/src/tooling/node-command-execution.js +237 -0
  792. package/src/tooling/node-command-runner.js +26 -205
  793. package/src/tooling/node-converter/input.js +521 -0
  794. package/src/tooling/node-converter/output.js +244 -0
  795. package/src/tooling/node-converter-chunk-batches.js +21 -0
  796. package/src/tooling/node-converter.js +7 -720
  797. package/src/tooling/node-source-runtime.js +16 -276
  798. package/src/tooling/node-webgpu-provider.v1.json +33 -0
  799. package/src/tooling/node-webgpu.js +178 -453
  800. package/src/tooling/opfs-cache.d.ts +5 -0
  801. package/src/tooling/opfs-cache.js +4 -0
  802. package/src/tooling/production-release.d.ts +19 -0
  803. package/src/tooling/production-release.js +647 -0
  804. package/src/tooling/program-bundle/embedding-reference.js +19 -0
  805. package/src/tooling/program-bundle/materialize.js +324 -0
  806. package/src/tooling/program-bundle/rerank-reference.js +22 -0
  807. package/src/tooling/program-bundle/sequence-reference.js +42 -0
  808. package/src/tooling/program-bundle/validation.js +31 -0
  809. package/src/tooling/program-bundle/wgsl-closure.js +254 -0
  810. package/src/tooling/program-bundle-parity.d.ts +11 -4
  811. package/src/tooling/program-bundle-parity.js +262 -109
  812. package/src/tooling/program-bundle.d.ts +7 -0
  813. package/src/tooling/program-bundle.js +186 -591
  814. package/src/tooling/provider-v1-contract.d.ts +61 -0
  815. package/src/tooling/provider-v1-contract.js +192 -0
  816. package/src/tooling/registered-variant-calibration-job.d.ts +98 -0
  817. package/src/tooling/registered-variant-calibration-job.js +210 -0
  818. package/src/tooling/registered-variant-calibration.d.ts +167 -7
  819. package/src/tooling/registered-variant-calibration.js +57 -7
  820. package/src/tooling/runtime-optimization/candidates.js +664 -0
  821. package/src/tooling/runtime-optimization-campaign.js +166 -0
  822. package/src/tooling/runtime-optimization-index.js +16 -1
  823. package/src/tooling/runtime-optimization.d.ts +44 -0
  824. package/src/tooling/runtime-optimization.js +13 -656
  825. package/src/tooling/source-intake.d.ts +1 -1
  826. package/src/tooling/source-intake.js +34 -17
  827. package/src/tooling/source-package-profiles.js +1 -1
  828. package/src/tooling/source-runtime/bundle-materialize.js +298 -0
  829. package/src/tooling/source-runtime/bundle-plan.js +121 -0
  830. package/src/tooling/source-runtime/node-loader.js +258 -0
  831. package/src/tooling/source-runtime-bundle.js +26 -411
  832. package/src/tooling/source-runtime-converter-config.js +1 -1
  833. package/src/tooling-experimental-exports.js +0 -5
  834. package/src/tooling-exports/evidence.d.ts +14 -3
  835. package/src/tooling-exports/evidence.js +14 -3
  836. package/src/tooling-exports/storage.d.ts +5 -0
  837. package/src/tooling-exports/storage.js +4 -0
  838. package/src/tooling-exports.d.ts +3 -0
  839. package/src/tooling-exports.js +2 -0
  840. package/src/tooling-exports.shared.d.ts +17 -3
  841. package/src/tooling-exports.shared.js +9 -3
  842. package/src/training.d.ts +6 -6
  843. package/src/training.js +4 -4
  844. package/src/version.js +1 -1
  845. package/tests/kernels/browser/kernel-suite.js +512 -0
  846. package/tests/kernels/browser/test-page.js +169 -14
  847. package/tests/kernels/harness/tolerance.js +1 -0
  848. package/tests/kernels/reference/index.js +13 -1
  849. package/tests/kernels/reference/vision.js +160 -0
  850. package/src/converter/quantizer.d.ts +0 -121
  851. package/src/gpu/kernels/backward/utils.d.ts +0 -90
  852. package/src/gpu/kernels/uniform-utils.d.ts +0 -66
  853. package/src/gpu/kernels/utils.d.ts +0 -147
  854. package/src/inference/pipelines/text/debug-utils/utils.js +0 -118
  855. package/src/inference/pipelines/text/logits/utils.d.ts +0 -58
  856. package/src/inference/pipelines/text/moe-cpu-gptoss.js +0 -115
  857. package/src/inference/pipelines/text/moe-helpers.js +0 -21
  858. package/src/rules/converter/execution.rules.json +0 -20
  859. package/src/rules/converter/tensor-roles.rules.json +0 -13
  860. package/src/rules/converter/tokenizer.rules.json +0 -12
  861. package/src/rules/inference/attention.rules.json +0 -69
  862. package/src/rules/inference/config.rules.json +0 -82
  863. package/src/rules/inference/dtype.rules.json +0 -113
  864. package/src/rules/inference/execution.rules.json +0 -195
  865. package/src/rules/inference/ffn.rules.json +0 -123
  866. package/src/rules/inference/layer-pattern.rules.json +0 -16
  867. package/src/rules/inference/layer.rules.json +0 -7
  868. package/src/rules/inference/moe.rules.json +0 -145
  869. package/src/rules/kernels/attention.rules.json +0 -68
  870. package/src/rules/kernels/conv2d.rules.json +0 -6
  871. package/src/rules/kernels/depthwise-conv2d.rules.json +0 -6
  872. package/src/rules/kernels/dequant.rules.json +0 -58
  873. package/src/rules/kernels/energy.rules.json +0 -22
  874. package/src/rules/kernels/fused-ffn.rules.json +0 -15
  875. package/src/rules/kernels/fused-matmul-residual.rules.json +0 -6
  876. package/src/rules/kernels/fused-matmul-rmsnorm.rules.json +0 -8
  877. package/src/rules/kernels/gather.rules.json +0 -93
  878. package/src/rules/kernels/gelu.rules.json +0 -11
  879. package/src/rules/kernels/grouped-pointwise-conv2d.rules.json +0 -6
  880. package/src/rules/kernels/groupnorm.rules.json +0 -10
  881. package/src/rules/kernels/kv_quantize.rules.json +0 -9
  882. package/src/rules/kernels/layernorm.rules.json +0 -6
  883. package/src/rules/kernels/linear-attention.rules.json +0 -6
  884. package/src/rules/kernels/lm-head-argmax.rules.json +0 -10
  885. package/src/rules/kernels/matmul.rules.json +0 -95
  886. package/src/rules/kernels/modulate.rules.json +0 -6
  887. package/src/rules/kernels/moe.rules.gptoss.json +0 -105
  888. package/src/rules/kernels/moe.rules.json +0 -14
  889. package/src/rules/kernels/moe.rules.mixtral.json +0 -75
  890. package/src/rules/kernels/pixel_shuffle.rules.json +0 -6
  891. package/src/rules/kernels/relu.rules.json +0 -6
  892. package/src/rules/kernels/rep-penalty.rules.json +0 -6
  893. package/src/rules/kernels/repeat-channels.rules.json +0 -6
  894. package/src/rules/kernels/residual.rules.json +0 -12
  895. package/src/rules/kernels/rmsnorm-qk.rules.json +0 -6
  896. package/src/rules/kernels/rmsnorm.rules.json +0 -46
  897. package/src/rules/kernels/rope-qk.rules.json +0 -6
  898. package/src/rules/kernels/rope.rules.json +0 -6
  899. package/src/rules/kernels/sample.rules.json +0 -6
  900. package/src/rules/kernels/scale.rules.json +0 -8
  901. package/src/rules/kernels/silu.rules.json +0 -22
  902. package/src/rules/kernels/softmax.rules.json +0 -25
  903. package/src/rules/kernels/split-qg.rules.json +0 -6
  904. package/src/rules/kernels/split-qkv.rules.json +0 -6
  905. package/src/rules/kernels/upsample2d.rules.json +0 -6
  906. package/src/rules/loader/tensor-loader.rules.json +0 -42
  907. package/src/rules/loader/weights.rules.json +0 -42
  908. package/src/rules/tooling/command-runtime.rules.json +0 -65
  909. package/src/tooling/hf-registry-utils.js +0 -395
  910. /package/src/converter/{tokenizer-utils.js → tokenizer-special-tokens.js} +0 -0
  911. /package/src/{inference/pipelines/text/bdpa-steamroller.js → experimental/bdpa/steamroller.js} +0 -0
  912. /package/src/{inference/pipelines → experimental}/diffusion/scheduler.js +0 -0
  913. /package/src/{inference/pipelines → experimental}/energy/pipeline.d.ts +0 -0
  914. /package/src/{inference/pipelines → experimental}/energy/quintel.d.ts +0 -0
  915. /package/src/{inference/pipelines → experimental}/energy/quintel.js +0 -0
  916. /package/src/{utils → formats}/array-buffer.js +0 -0
  917. /package/src/{utils → formats}/canonical-hash.js +0 -0
  918. /package/src/{utils → formats}/clone-json.js +0 -0
  919. /package/src/{utils → formats}/encode-utf8.js +0 -0
  920. /package/src/{utils → formats}/load-json.js +0 -0
  921. /package/src/{utils → formats}/plain-object.js +0 -0
  922. /package/src/{utils → formats}/sha256.js +0 -0
  923. /package/src/{utils → formats}/stable-sort-object.js +0 -0
  924. /package/src/{gpu/kernels → rules}/rule-matcher.js +0 -0
  925. /package/src/{utils → storage}/hf-resolve-url.js +0 -0
  926. /package/src/{utils → storage}/runtime-env.js +0 -0
  927. /package/src/tooling/{command-runner-shared.js → command-runtime-execution.js} +0 -0
package/README.md CHANGED
@@ -6,278 +6,221 @@
6
6
 
7
7
  [![Build](https://img.shields.io/github/actions/workflow/status/clocksmith/doppler/check-green.yml?branch=main&label=build)](https://github.com/clocksmith/doppler/actions/workflows/check-green.yml)
8
8
  [![npm version](https://img.shields.io/npm/v/doppler-gpu.svg?label=version)](https://www.npmjs.com/package/doppler-gpu)
9
- [![License: Apache-2.0](https://img.shields.io/badge/license-Apache--2.0-blue.svg)](https://github.com/clocksmith/doppler/blob/main/LICENSE)
10
- [![PRs welcome](https://img.shields.io/badge/PRs-welcome-brightgreen.svg)](https://github.com/clocksmith/doppler/pulls)
9
+ [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://github.com/clocksmith/doppler/blob/main/LICENSE)
11
10
 
12
- Doppler turns inference engineering into a verifiable search problem. Optimizers
13
- edit JSON manifests, execution plans, and WGSL kernels; parity and benchmark gates
14
- accept or reject each candidate against correctness and speed.
11
+ Doppler is a model compiler, release foundry, and WebGPU runtime for JavaScript.
12
+ Forge turns supported model sources into signed immutable Capsules with
13
+ ModelIR-derived, qualified TargetPlans. Runtime verifies, selects, binds, and
14
+ executes the declared JavaScript/WGSL program. Browser and Node paths are
15
+ qualified separately; Bun remains experimental.
15
16
 
16
- These contracts and gates are used by operators to tune a configurable and pure JavaScript and WGSL sourced WebGPU
17
- runtime for supported
18
- [RDRR artifacts](https://github.com/clocksmith/doppler/blob/main/docs/rdrr-format.md)
19
- in browser, Node, and Bun. Doppler runs text generation, embeddings, and reranking
20
- locally, with CLI and OpenAI-compatible server entry points.
17
+ ## Mission, goal, and value
21
18
 
22
- **Read this first:** [getting started](https://github.com/clocksmith/doppler/blob/main/docs/getting-started.md) → [Root API](https://github.com/clocksmith/doppler/blob/main/docs/api/root.md) → [open source evidence](https://github.com/clocksmith/doppler/blob/main/docs/model-competition-scoreboard.md)
19
+ Make useful AI capabilities installable, inspectable, and maintainable as
20
+ application dependencies. Success means an independent application voluntarily
21
+ retains Doppler for a measured improvement and ships a second revision with less
22
+ release effort. Free adoption counts; benchmarks alone do not prove it.
23
23
 
24
- **[Try the live demo](https://d4da.com/doppler)** | **[npm](https://www.npmjs.com/package/doppler-gpu)** | **[docs](https://github.com/clocksmith/doppler/blob/main/docs/INDEX.md)**
24
+ Applications control trust and updates. This checkout introduces the breaking
25
+ [Capsule naming migration](docs/capsule-naming-migration.md); former Pack APIs
26
+ and schemas are not accepted. Capsule v2 and v3 are supported;
27
+ [Capsule v3](docs/capsule-identity-migration.md) separates executable identity from
28
+ release events. Doe, Poolday, and Reploid are optional, not prerequisites.
25
29
 
26
- ```bash
27
- npx doppler-gpu
28
- ```
30
+ Build local generation, embeddings, or reranking, or contribute improvements to
31
+ source interpretation, kernels, loading, and release reliability. See the
32
+ [goals](docs/goals.md) and [contribution guide](docs/contributing.md).
29
33
 
30
- <!-- model-type-clusters:start -->
34
+ ## How to use Doppler
31
35
 
32
- ## Supported RDRR model types
33
-
34
- Doppler classifies artifacts by what they consume and produce. This is
35
- separate from lineage (`family`), runtime implementation (`modelType`), and
36
- artifact-size tier.
37
-
38
- | Type | Verified / Cataloged | What it does |
39
- | --- | --- | --- |
40
- | Text generators | 12 / 14 | `text → text` |
41
- | Multimodal generators | 3 / 3 | `audio + image + text → text` |
42
- | Diffusion language models | 0 / 1 | `text → text` |
43
- | Translation specialists | 2 / 2 | `text → text` |
44
- | Language embedders | 2 / 2 | `text → embedding` |
45
- | Rerankers | 2 / 2 | `text-pair → relevance-score` |
46
- | Protein encoders | 3 / 3 | `sequence → embedding + token outputs` |
47
- | Nucleotide encoders | 1 / 1 | `dna-seq → embedding` |
48
-
49
- The [full model-support matrix](https://github.com/clocksmith/doppler/blob/main/docs/model-support-matrix.md)
50
- lists every lane and its lifecycle evidence. Classification says what an
51
- artifact is shaped to do; only lifecycle receipts establish what is
52
- verified, and a runtime pass does not by itself qualify every declared input
53
- modality.
54
-
55
- <!-- model-type-clusters:end -->
56
-
57
- ## LoRA: runtime loading and training
58
-
59
- Doppler supports:
60
-
61
- - SafeTensors LoRA loading and hot-swap at runtime.
62
- - Experimental SFT/LoRA training in Node, Bun, and browser via `doppler-gpu/training`.
63
- - Native packed-Q4K LoRA for `qwen-3-5-0-8b-q4k-ehaf16` in `webgpu_native` (frozen base, LoRA A/B updates only).
64
- - External backends for other packed-Q4K training targets.
36
+ Run the CLI without installing a global package:
65
37
 
66
38
  ```bash
67
- npx doppler-gpu lora --config ./workload.json --surface node
39
+ npx doppler-gpu "Summarize WebGPU in one sentence"
40
+ npx doppler-gpu --model qwen3-0.8b --prompt "Write a haiku about GPUs"
41
+ npx doppler-gpu --list-models
68
42
  ```
69
43
 
70
- Active adapters are registered in [models/adapters/catalog.json](models/adapters/catalog.json).
71
- The training API, workload schema, and adapter families live in the
72
- [LoRA Format specification](docs/lora-format.md),
73
- [Training handbook](docs/training-handbook.md), and [Training API](docs/api/training.md).
74
-
75
- ## Evidence
76
-
77
- Doppler has higher steady-state inference throughput than Transformers.js in
78
- accepted, throughput-comparable browser WebGPU comparisons indexed below where
79
- the throughput gate passes.
80
-
81
- ![Metal and Vulkan browser WebGPU throughput distributions](https://raw.githubusercontent.com/clocksmith/doppler/main/assets/doppler-webgpu-evidence.svg)
44
+ The live browser demo is at [d4da.com/doppler](https://d4da.com/doppler).
45
+ The first documentation path is [getting started](https://github.com/clocksmith/doppler/blob/main/docs/getting-started.md),
46
+ followed by the [Capsule Runtime API](https://github.com/clocksmith/doppler/blob/main/docs/api/root.md).
82
47
 
83
- Text uses tok/s, embeddings use embedding/s, and reranking uses rerank/s;
84
- higher is faster. Every comparison passes its declared workload
85
- correctness gate. Model loading is separate: Transformers.js loads the Vulkan
86
- embedding and reranker artifacts faster. The
87
- [competition scoreboard](https://github.com/clocksmith/doppler/blob/main/docs/model-competition-scoreboard.md) links every
88
- receipt, and the [methodology](https://github.com/clocksmith/doppler/blob/main/docs/benchmark-methodology.md) defines the gates.
48
+ For the release-foundry path, the installed command is `doppler release`; from
49
+ npm use `npx --package doppler-gpu doppler release`. It consumes a pinned
50
+ `production-release/v1` manifest and signed exact-device receipts, then emits an
51
+ `eligible` or `blocked` decision and retained evidence. It never activates or
52
+ deploys the customer application. See the [release platform contract](docs/model-release-platform.md)
53
+ and [CLI reference](docs/cli.md).
89
54
 
90
- ## Why these lanes are faster
55
+ ### Capsule Runtime API
91
56
 
92
- Doppler speedups come from specific, explicit runtime changes:
93
-
94
- - **Fusion:** projection + FFN path reductions reduce kernel count.
95
- - **Attention/layout tuning:** lowers prefill overhead.
96
- - **Batched decode:** amortizes readback overhead.
97
- - **Packed kernels:** Q4_K + INT4-PLE reduce projection traffic.
57
+ ```js
58
+ import { openCapsule } from 'doppler-gpu/host';
59
+ import { reviewedRelease } from './reviewed-release.js';
98
60
 
99
- Evidence receipts:
61
+ const session = await openCapsule(reviewedRelease.capsuleUrl, {
62
+ trustedSigners: reviewedRelease.trustedSigners,
63
+ acceptedTargetPlanDigests: reviewedRelease.acceptedTargetPlanDigests,
64
+ });
65
+ try {
66
+ const result = await session.rerank({
67
+ application: reviewedRelease.application,
68
+ query: 'Which API provides browser GPU compute?',
69
+ documents: ['WebGPU exposes GPU compute.', 'WebSocket transports messages.'],
70
+ options: {},
71
+ });
72
+ console.log(result, session.selectedTargetPlanDigest);
73
+ } finally {
74
+ await session.close();
75
+ }
76
+ ```
100
77
 
101
- - [Qwen 3.5 0.8B (Metal)](https://github.com/clocksmith/doppler/blob/main/benchmarks/vendors/results/compare_20260709T154633.json)
102
- - [Qwen embedding + rerank](https://github.com/clocksmith/doppler/blob/main/benchmarks/vendors/results/embedding_compare_qwen-3-embedding-0-6b-q4k-ehf16-af32_20260709T180853.json), [Qwen rerank](https://github.com/clocksmith/doppler/blob/main/benchmarks/vendors/results/rerank_compare_qwen-3-reranker-0-6b-q4k-ehf16-af32_20260709T192830.json)
103
- - [Qwen 3.5 2B (Vulkan)](https://github.com/clocksmith/doppler/blob/main/benchmarks/vendors/results/compare_20260707T161623.json)
104
- - [Gemma 4 (Vulkan)](https://github.com/clocksmith/doppler/blob/main/benchmarks/vendors/results/compare_20260707T170557.json)
78
+ `reviewedRelease` is application-owned configuration, not metadata trusted merely
79
+ because it was downloaded. It identifies a reranker Capsule, accepted plans, trusted
80
+ publishers, and the application/workload/oracle contract. The host entry composes
81
+ the existing device, artifact-store, and program ports; trust and upgrades remain
82
+ explicit. See the [Electron integration](examples/electron-document-search/README.md)
83
+ and [retained evaluation](docs/integration/reranker-evaluation.md).
105
84
 
106
- Retention is gated: a lane moves forward only when both parity and benchmark
107
- receipts pass in the
108
- [challenger framework](https://github.com/clocksmith/doppler/blob/main/docs/local-gpu-challenger-framework.md).
85
+ Advanced applications can still inject every port through `doppler-gpu` or
86
+ `doppler-gpu/runtime`. Both routes verify the same Capsule and initial execution
87
+ identity. The host facade does not add model operations or broaden qualification.
109
88
 
110
- ### Long-term direction
89
+ Discover repository commands with `npm run` or `npm pkg get scripts` (JSON).
90
+ Use `doppler --help` for the installed CLI.
111
91
 
112
- Humans run this loop today. WGSL-distillation is experimental.
113
- Automation in kernel generation/autotuning is a roadmap item, not shipped yet.
114
- New checkpoints and GPUs still go through the same parity and benchmark gate flow.
92
+ The former manifest-loading facade remains available only as an explicit
93
+ compatibility import:
115
94
 
116
- ## How it works
95
+ ```js
96
+ import { dr } from 'doppler-gpu/compat';
117
97
 
118
- ```mermaid
119
- flowchart TB
120
- subgraph Propose["1 · Propose"]
121
- direction LR
122
- HUMAN["Engineer or optimizer"]
123
- CANDIDATE["Candidate edit<br/>manifest · plan · WGSL"]
124
- HUMAN --> CANDIDATE
125
- end
126
-
127
- subgraph Run["2 · Resolve and run"]
128
- direction LR
129
- ARTIFACT["RDRR artifact<br/>manifest · weights · tokenizer"]
130
- REQUEST["Request + runtime profile"]
131
- VALIDATE["Validate contracts"]
132
- RESOLVE["Resolve execution graph<br/>kernel refs · session"]
133
- LOAD["Load + bind<br/>GPU buffers · cache"]
134
- DISPATCH["Dispatch WGSL<br/>prefill · decode · retrieval"]
135
- OUTPUT["Tokens · embeddings · scores"]
136
-
137
- ARTIFACT --> VALIDATE
138
- REQUEST --> VALIDATE
139
- CANDIDATE --> VALIDATE
140
- VALIDATE --> RESOLVE --> LOAD --> DISPATCH --> OUTPUT
141
- end
142
-
143
- subgraph Verify["3 · Evidence disposes"]
144
- direction LR
145
- PARITY{"Parity gate<br/>correct vs reference"}
146
- BENCH{"Benchmark gate<br/>comparable speed vs prior"}
147
- RETAIN["Retain<br/>versioned lane + receipt"]
148
- REJECT["Reject<br/>finding + negative result"]
149
-
150
- OUTPUT --> PARITY
151
- PARITY -- pass --> BENCH
152
- PARITY -- mismatch --> REJECT
153
- BENCH -- pass --> RETAIN
154
- BENCH -- fail --> REJECT
155
- end
156
-
157
- RETAIN -. reusable asset .-> HUMAN
158
- REJECT -. search memory .-> HUMAN
159
-
160
- classDef proposal fill:#fff1f3,stroke:#ff6b7a,color:#08090c,stroke-width:2px;
161
- classDef runtime fill:#eef4ff,stroke:#5d8dff,color:#08090c,stroke-width:2px;
162
- classDef gate fill:#f7f1ff,stroke:#b98cff,color:#08090c,stroke-width:2px;
163
- classDef accepted fill:#08090c,stroke:#5d8dff,color:#ffffff,stroke-width:3px;
164
- classDef rejected fill:#fff1f3,stroke:#ff6b7a,color:#08090c,stroke-width:2px;
165
-
166
- class HUMAN,CANDIDATE proposal;
167
- class ARTIFACT,REQUEST,VALIDATE,RESOLVE,LOAD,DISPATCH,OUTPUT runtime;
168
- class PARITY,BENCH gate;
169
- class RETAIN accepted;
170
- class REJECT rejected;
171
-
172
- style Propose fill:#fffafa,stroke:#ff6b7a,stroke-width:1px
173
- style Run fill:#f8faff,stroke:#5d8dff,stroke-width:1px
174
- style Verify fill:#fcfaff,stroke:#b98cff,stroke-width:1px
175
- linkStyle default stroke:#5d8dff,stroke-width:2px;
98
+ const session = await dr.open('qwen3-0.8b');
99
+ const result = await session.generate('Describe WebGPU briefly');
100
+ await session.close();
176
101
  ```
177
102
 
178
- The full resolve, load, bind, dispatch, and readback flow lives in the
179
- [architecture](https://github.com/clocksmith/doppler/blob/main/docs/architecture.md)
180
- document. Unsupported paths fail closed. Doppler owns artifact and execution
181
- contracts; applications own policy.
182
-
183
- New model families need RDRR conversion and may need tokenizer, graph, or kernel
184
- support.
185
-
186
- [Ouroboros/Reploid](https://github.com/clocksmith/doppler/blob/main/docs/architecture.md#optional-ouroborosreploid-integration)
187
- keeps orchestration above this boundary.
188
- [Program Bundles](https://github.com/clocksmith/doppler/blob/main/docs/integration/program-bundle.md)
189
- preserve program identity for downstream backends.
190
-
191
- ## Quick start
103
+ ### OpenAI-compatible server
192
104
 
193
- ### Browser
105
+ ```bash
106
+ npx doppler-serve --model qwen3-0.8b --port 8080
107
+ ```
194
108
 
195
- The live demo runs locally and works offline after its first model download.
196
- Its public API boundary, observation tiers, generated PWA shell, and independent
197
- contract/hardware smoke receipts are documented in
198
- [demo/README.md](demo/README.md).
109
+ The server accepts requests at `http://localhost:8080/v1`. Registry IDs resolve
110
+ to hosted RDRR artifacts from `clocksmith/rdrr` by default.
199
111
 
200
- ### CLI
112
+ ### LoRA loading and training
201
113
 
202
114
  ```bash
203
- npx doppler-gpu "Summarize WebGPU in one sentence"
204
- npx doppler-gpu --model qwen3-0.8b --prompt "Write a haiku about GPUs"
205
- npx doppler-gpu --list-models
115
+ npx doppler-gpu lora --config ./workload.json --surface node
206
116
  ```
207
117
 
208
- ### Root API
118
+ Doppler supports SafeTensors LoRA loading and hot swap at runtime. SFT/LoRA
119
+ training is available through the experimental Node, Bun, and browser training
120
+ surface. Cataloged adapter identities and lifecycle states are listed in
121
+ [`models/adapters/catalog.json`](models/adapters/catalog.json). See the
122
+ [LoRA format](docs/lora-format.md), [training handbook](docs/training-handbook.md),
123
+ and [Training API](docs/api/training.md).
209
124
 
210
- The `dr` facade is the primary app API. Advanced APIs use package subpaths.
125
+ <!-- model-type-clusters:start -->
211
126
 
212
- ```js
213
- import { dr } from 'doppler-gpu';
127
+ ## Supported RDRR model types
214
128
 
215
- // Explicit scoped API with a stable result envelope
216
- const session = await dr.open('qwen3-0.8b');
217
- const result = await session.generate('Describe WebGPU briefly');
218
- console.log(result.outputText, result.fingerprint);
219
- await session.close();
129
+ Doppler classifies artifacts by what they consume and produce. This is
130
+ separate from lineage (`family`), runtime implementation (`modelType`), and
131
+ artifact-size tier.
220
132
 
221
- // Stream tokens
222
- const model = await dr.load('qwen3-0.8b');
223
- for await (const token of model.generate('Describe WebGPU briefly')) {
224
- process.stdout.write(token);
225
- }
133
+ | Type | Input → output | Runtime-verified / cataloged | Representative lanes |
134
+ | --- | --- | --- | --- |
135
+ | Text generators | text → text | 13 / 17 | gemma-3-1b-it-q4k-ehf16-af32<br>gemma-3-270m-it-f16-af32<br>gemma-3-270m-it-q4k-ehf16-af32<br>+14 more |
136
+ | Multimodal generators | audio + image + text → text | 3 / 3 | gemma-4-e2b-it-q4k-ehf16-af16-int4ple<br>gemma-4-e2b-it-q4k-ehf16-af32<br>gemma-4-e2b-it-q4k-ehf16-af32-int4ple |
137
+ | Diffusion language models | text → text | 0 / 1 | diffusiongemma-26b-a4b-it-q4k-ehf16-af16 |
138
+ | Translation specialists | text → text | 2 / 2 | translategemma-4b-1b-enes-q4k-ehf16-af32<br>translategemma-4b-it-q4k-ehf16-af32 |
139
+ | Language embedders | text → pooled-embedding | 2 / 2 | google-embeddinggemma-300m-q4k-ehf16-af32<br>qwen-3-embedding-0-6b-q4k-ehf16-af32 |
140
+ | Rerankers | text-pair → relevance-score | 2 / 2 | qwen-3-reranker-0-6b-f16-af32<br>qwen-3-reranker-0-6b-q4k-ehf16-af32 |
141
+ | Protein encoders | protein-sequence → pooled-embedding + token-embedding + token-logits | 3 / 3 | amplify-120m-f16-af32<br>esm2-t12-35m-ur50d-f32-af32<br>esmc-300m-f32-af32 |
142
+ | Nucleotide encoders | dna-sequence → pooled-embedding + token-embedding | 1 / 1 | nucleotide-transformer-v2-50m-f32-af32 |
226
143
 
227
- // One-shot
228
- const text = await model.generateText('Explain WebGPU in one sentence');
144
+ The [full model-support matrix](https://github.com/clocksmith/doppler/blob/main/docs/model-support-matrix.md)
145
+ lists every lane and its lifecycle evidence. Classification says what an
146
+ artifact is shaped to do; only lifecycle receipts establish what is
147
+ verified, and a runtime pass does not by itself qualify every declared input
148
+ modality.
229
149
 
230
- // One-shot with a hash-bound execution receipt
231
- const evidence = await model.generateWithEvidence('Explain WebGPU in one sentence', {
232
- maxTokens: 64,
233
- temperature: 0,
234
- });
235
- console.log(evidence.outputText, evidence.transcriptHash);
150
+ <!-- model-type-clusters:end -->
236
151
 
237
- await model.unload();
238
- ```
152
+ ## Evidence
239
153
 
240
- `generateWithEvidence()` returns the text and token IDs together with canonical
241
- SHA-256 hashes for the transcript, resolved generation config, runtime profile,
242
- WebGPU backend, and execution-plan identity. The receipt records what ran; it
243
- does not by itself establish output correctness.
154
+ Doppler has accepted browser WebGPU comparisons with higher steady-state inference throughput
155
+ (higher is faster) than Transformers.js where the declared workload correctness and
156
+ throughput gates pass. Loading is a separate measurement; the referenced
157
+ Vulkan embedding and reranker artifacts load faster in Transformers.js. The
158
+ [scoreboard](https://github.com/clocksmith/doppler/blob/main/docs/model-competition-scoreboard.md)
159
+ links the receipts and the [benchmark methodology](https://github.com/clocksmith/doppler/blob/main/docs/benchmark-methodology.md)
160
+ defines the gates.
244
161
 
245
- ### OpenAI-compatible server
162
+ ![Metal and Vulkan browser WebGPU throughput distributions](https://raw.githubusercontent.com/clocksmith/doppler/main/assets/doppler-webgpu-evidence.svg)
246
163
 
247
- ```bash
248
- npx doppler-serve --model qwen3-0.8b --port 8080
164
+ ## Release and execution flow
165
+
166
+ ```mermaid
167
+ flowchart TB
168
+ S[Pinned source checkpoint] --> F[Forge: inspect, normalize, lower]
169
+ F --> V[Verify and qualify TargetPlans]
170
+ V --> P[Package and sign immutable Capsule]
171
+ P --> A{Application acceptance}
172
+ A -- fail --> N[Reject with finding]
173
+ A -- pass --> K[Promote supported release]
174
+ K --> R[Runtime: validate, select, bind, execute, observe]
175
+ R --> Q[Requalify, roll back, or revoke]
249
176
  ```
250
177
 
251
- Point an OpenAI client at `http://localhost:8080/v1`:
178
+ Candidates begin as pinned source truth. Forge owns graph-changing work and
179
+ emits an immutable Capsule only after verification and qualification. Application
180
+ acceptance authorizes promotion. Runtime never repairs or specializes the Capsule;
181
+ it selects a qualified TargetPlan, binds resources, executes declared commands,
182
+ and emits evidence for continuing qualification and recovery decisions.
252
183
 
253
- ```js
254
- import OpenAI from 'openai';
255
- const client = new OpenAI({ baseURL: 'http://localhost:8080/v1', apiKey: 'unused' });
256
- const response = await client.chat.completions.create({
257
- model: 'qwen3-0.8b',
258
- messages: [{ role: 'user', content: 'Hello' }],
259
- });
260
- ```
184
+ ## Long-term vision
185
+
186
+ Doppler is intended to support a growing set of local model families and
187
+ runtime variants without hiding the model contract or execution path. Registered
188
+ variant calibration, paired performance gates, and WGSL experiments remain
189
+ human-reviewed. Ouroboros and Reploid sit above Doppler as orchestration or
190
+ product layers; Doppler owns the artifact and execution boundary.
261
191
 
262
- Registry IDs resolve to hosted RDRR artifacts from `clocksmith/rdrr` by default. See the [Root API guide](https://github.com/clocksmith/doppler/blob/main/docs/api/root.md).
192
+ New model families require RDRR conversion and may require tokenizer, graph, or
193
+ kernel support. Native packed-Q4K LoRA support is available for the declared
194
+ Qwen target; other packed-Q4K training targets use external backends.
263
195
 
264
- ## Start here
196
+ ## Limits and current status
265
197
 
266
- ### Pick a path
198
+ WebGPU is required. Use a current Chromium browser; Node installs the WebGPU
199
+ provider as an optional dependency. A runtime pass does not verify every input
200
+ modality, and a receipt records what ran without establishing output quality.
201
+ Throughput comparisons are valid only when the workload, timing scope, and
202
+ correctness path are comparable. Unsupported paths fail closed.
267
203
 
268
- - **Application builders:** start with [getting started](https://github.com/clocksmith/doppler/blob/main/docs/getting-started.md), [Root API](https://github.com/clocksmith/doppler/blob/main/docs/api/root.md), and the [OpenAI-compatible server](#openai-compatible-server).
269
- - **Integrators:** use the [RDRR format](https://github.com/clocksmith/doppler/blob/main/docs/rdrr-format.md), [support matrix](https://github.com/clocksmith/doppler/blob/main/docs/model-support-matrix.md), and [Program Bundles](https://github.com/clocksmith/doppler/blob/main/docs/integration/program-bundle.md).
270
- - **Kernel/inference engineers:** follow [architecture](https://github.com/clocksmith/doppler/blob/main/docs/architecture.md), [performance optimization](https://github.com/clocksmith/doppler/blob/main/docs/developer-guides/16-kernel-performance-optimization.md), and the [challenger framework](https://github.com/clocksmith/doppler/blob/main/docs/local-gpu-challenger-framework.md).
271
- - **Evidence reviewers:** review the [scoreboard](https://github.com/clocksmith/doppler/blob/main/docs/model-competition-scoreboard.md), [methodology](https://github.com/clocksmith/doppler/blob/main/docs/benchmark-methodology.md), and [release matrix](https://github.com/clocksmith/doppler/blob/main/docs/release-matrix.md).
204
+ ## Repository map
272
205
 
273
- The [docs index](https://github.com/clocksmith/doppler/blob/main/docs/INDEX.md)
274
- owns the complete model, subsystem, API, architecture, and operator inventory.
206
+ - [Component charters](https://github.com/clocksmith/doppler/blob/main/CATSCAN.md) — recursive repository and subsystem intent
207
+ - [Component index](https://github.com/clocksmith/doppler/blob/main/docs/component-index.md) — generated authority and parent map
208
+ - [`src/`](src/) — runtime, model loading, inference, and execution contracts
209
+ - [`demo/`](demo/) — browser demo and its public API boundary
210
+ - [`models/adapters/`](models/adapters/) — adapter catalog, lifecycle, identity, and evidence metadata
211
+ - [`benchmarks/`](benchmarks/) — vendor comparisons and retained results
212
+ - [`docs/`](docs/) — APIs, architecture, formats, methodology, and release matrices
213
+ - [`tests/`](tests/) — runtime, contract, browser, and benchmark tests
214
+ - [`tools/`](tools/) — conversion, qualification, and operator tools
275
215
 
276
- ## Environment requirements
216
+ ## Read next
277
217
 
278
- WebGPU is required. Use a current Chromium browser; Node installs the `webgpu`
279
- provider as an optional dependency.
218
+ - [Documentation index](https://github.com/clocksmith/doppler/blob/main/docs/INDEX.md)
219
+ - [Architecture](https://github.com/clocksmith/doppler/blob/main/docs/architecture.md)
220
+ - [RDRR format](https://github.com/clocksmith/doppler/blob/main/docs/rdrr-format.md)
221
+ - [Local GPU challenger framework](https://github.com/clocksmith/doppler/blob/main/docs/local-gpu-challenger-framework.md)
222
+ - [Program Bundles](https://github.com/clocksmith/doppler/blob/main/docs/integration/program-bundle.md)
280
223
 
281
224
  ## License
282
225
 
283
- Apache License 2.0 (`Apache-2.0`). See [LICENSE](LICENSE) and [NOTICE](NOTICE).
226
+ [MIT License](LICENSE). See [NOTICE](NOTICE) for attribution.