@simulatte/doppler 0.1.5 → 0.1.7

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 (392) hide show
  1. package/CHANGELOG.md +126 -0
  2. package/README.md +25 -17
  3. package/package.json +20 -4
  4. package/src/adapters/adapter-registry.js +12 -1
  5. package/src/adapters/lora-loader.js +23 -6
  6. package/src/bridge/extension-client.d.ts +5 -0
  7. package/src/bridge/extension-client.js +40 -0
  8. package/src/bridge/index.d.ts +2 -1
  9. package/src/bridge/index.js +6 -4
  10. package/src/browser/browser-converter.js +26 -1
  11. package/src/browser/file-picker.js +6 -0
  12. package/src/browser/safetensors-parser-browser.js +84 -1
  13. package/src/browser/shard-io-browser.js +2 -2
  14. package/src/browser/tensor-source-download.js +8 -2
  15. package/src/browser/tensor-source-http.d.ts +1 -0
  16. package/src/browser/tensor-source-http.js +5 -1
  17. package/src/client/doppler-api.browser.js +20 -4
  18. package/src/client/doppler-api.js +19 -3
  19. package/src/client/doppler-provider/generation.js +12 -0
  20. package/src/client/doppler-provider/model-manager.d.ts +10 -0
  21. package/src/client/doppler-provider/model-manager.js +91 -19
  22. package/src/client/doppler-provider/source-runtime.d.ts +2 -1
  23. package/src/client/doppler-provider/source-runtime.js +132 -13
  24. package/src/client/doppler-registry.json +8 -7
  25. package/src/config/backward-registry-loader.js +17 -2
  26. package/src/config/execution-v0-contract-check.js +113 -15
  27. package/src/config/kernel-path-contract-check.js +57 -29
  28. package/src/config/kernel-path-loader.js +5 -36
  29. package/src/config/kernels/kernel-ref-digests.js +39 -39
  30. package/src/config/kernels/registry.js +14 -1
  31. package/src/config/kernels/registry.json +49 -7
  32. package/src/config/loader.d.ts +1 -1
  33. package/src/config/loader.js +43 -4
  34. package/src/config/merge-contract-check.js +59 -4
  35. package/src/config/merge-helpers.js +128 -7
  36. package/src/config/merge.d.ts +1 -0
  37. package/src/config/merge.js +28 -0
  38. package/src/config/param-validator.js +47 -2
  39. package/src/config/presets/kernel-paths/{gemma2-q4k-dequant-f32a.json → gemma2-q4k-dequant-f32a-nosubgroups.json} +3 -3
  40. package/src/config/presets/kernel-paths/gemma3-f16-fused-f32a-online-streamingprefill.json +223 -0
  41. package/src/config/presets/kernel-paths/{gemma3-q4k-dequant-f32a.json → gemma3-q4k-dequant-f32a-nosubgroups.json} +3 -3
  42. package/src/config/presets/kernel-paths/registry.json +29 -8
  43. package/src/config/presets/models/gemma2.json +2 -2
  44. package/src/config/presets/models/qwen3.json +9 -2
  45. package/src/config/presets/models/transformer.json +5 -0
  46. package/src/config/presets/runtime/experiments/bench/gemma3-bench-q4k.json +1 -1
  47. package/src/config/presets/runtime/experiments/debug/gemma3-debug-q4k.json +1 -1
  48. package/src/config/presets/runtime/experiments/verify/gemma3-verify.json +1 -1
  49. package/src/config/presets/runtime/kernels/dequant-f16-q4k.json +6 -13
  50. package/src/config/presets/runtime/kernels/dequant-f32-q4k.json +6 -13
  51. package/src/config/presets/runtime/kernels/embeddinggemma-q4k-dequant-f32a.json +37 -0
  52. package/src/config/presets/runtime/kernels/fused-q4k.json +6 -13
  53. package/src/config/presets/runtime/kernels/gemma2-q4k-dequant-f16a.json +33 -0
  54. package/src/config/presets/runtime/kernels/gemma2-q4k-dequant-f32a-nosubgroups.json +33 -0
  55. package/src/config/presets/runtime/kernels/gemma2-q4k-fused-f32a.json +33 -0
  56. package/src/config/presets/runtime/kernels/safe-q4k.json +6 -13
  57. package/src/config/presets/runtime/platform/metal-apple-q4k.json +1 -1
  58. package/src/config/required-inference-fields-contract-check.js +6 -0
  59. package/src/config/runtime.js +6 -1
  60. package/src/config/schema/debug.schema.d.ts +5 -0
  61. package/src/config/schema/doppler.schema.js +16 -21
  62. package/src/config/schema/inference-defaults.schema.js +6 -3
  63. package/src/config/schema/inference.schema.d.ts +9 -0
  64. package/src/config/schema/kernel-path.schema.d.ts +11 -1
  65. package/src/config/schema/kernel-thresholds.schema.js +12 -4
  66. package/src/config/schema/manifest.schema.d.ts +8 -1
  67. package/src/config/schema/manifest.schema.js +19 -3
  68. package/src/config/training-defaults.js +30 -22
  69. package/src/converter/conversion-plan.js +94 -9
  70. package/src/converter/core.d.ts +7 -0
  71. package/src/converter/core.js +14 -9
  72. package/src/converter/execution-v0-manifest.js +4 -1
  73. package/src/converter/index.d.ts +1 -0
  74. package/src/converter/index.js +1 -0
  75. package/src/converter/manifest-inference.js +43 -12
  76. package/src/converter/parsers/diffusion.js +0 -3
  77. package/src/converter/quantization-info.js +35 -15
  78. package/src/converter/rope-config.js +42 -0
  79. package/src/converter/shard-packer.d.ts +1 -1
  80. package/src/converter/shard-packer.js +4 -1
  81. package/src/debug/config.js +123 -11
  82. package/src/debug/signals.js +7 -1
  83. package/src/debug/tensor.d.ts +2 -0
  84. package/src/debug/tensor.js +13 -2
  85. package/src/distribution/p2p-control-plane.js +52 -12
  86. package/src/distribution/p2p-observability.js +43 -7
  87. package/src/distribution/p2p-webrtc-browser.js +20 -0
  88. package/src/distribution/shard-delivery.js +77 -26
  89. package/src/formats/gguf/types.js +33 -16
  90. package/src/formats/rdrr/groups.d.ts +12 -4
  91. package/src/formats/rdrr/groups.js +3 -6
  92. package/src/formats/rdrr/parsing.js +39 -2
  93. package/src/formats/rdrr/types.d.ts +2 -1
  94. package/src/gpu/command-recorder.js +86 -61
  95. package/src/gpu/device.d.ts +1 -0
  96. package/src/gpu/device.js +131 -19
  97. package/src/gpu/kernel-tuner/benchmarks.js +326 -316
  98. package/src/gpu/kernel-tuner/cache.js +71 -4
  99. package/src/gpu/kernel-tuner/tuner.js +22 -4
  100. package/src/gpu/kernels/attention.js +113 -34
  101. package/src/gpu/kernels/backward/adam.js +62 -58
  102. package/src/gpu/kernels/backward/attention_backward.js +257 -169
  103. package/src/gpu/kernels/backward/conv2d_backward.js +14 -1
  104. package/src/gpu/kernels/bias_add.wgsl +8 -6
  105. package/src/gpu/kernels/bias_add_f16.wgsl +8 -5
  106. package/src/gpu/kernels/cast.js +191 -149
  107. package/src/gpu/kernels/check-stop.js +33 -44
  108. package/src/gpu/kernels/conv2d.js +27 -17
  109. package/src/gpu/kernels/conv2d.wgsl +7 -8
  110. package/src/gpu/kernels/conv2d_f16.wgsl +7 -8
  111. package/src/gpu/kernels/cross_entropy_loss.js +21 -15
  112. package/src/gpu/kernels/depthwise_conv2d.js +37 -26
  113. package/src/gpu/kernels/depthwise_conv2d.wgsl +6 -9
  114. package/src/gpu/kernels/depthwise_conv2d_f16.wgsl +6 -9
  115. package/src/gpu/kernels/dequant.js +178 -126
  116. package/src/gpu/kernels/energy.d.ts +3 -21
  117. package/src/gpu/kernels/energy.js +111 -88
  118. package/src/gpu/kernels/feature-check.js +1 -1
  119. package/src/gpu/kernels/fused_ffn.js +84 -65
  120. package/src/gpu/kernels/fused_matmul_residual.js +56 -33
  121. package/src/gpu/kernels/fused_matmul_rmsnorm.js +62 -45
  122. package/src/gpu/kernels/gather.js +33 -15
  123. package/src/gpu/kernels/gelu.js +19 -11
  124. package/src/gpu/kernels/grouped_pointwise_conv2d.js +34 -23
  125. package/src/gpu/kernels/grouped_pointwise_conv2d.wgsl +6 -9
  126. package/src/gpu/kernels/grouped_pointwise_conv2d_f16.wgsl +6 -9
  127. package/src/gpu/kernels/groupnorm.js +34 -23
  128. package/src/gpu/kernels/kv-quantize.js +5 -2
  129. package/src/gpu/kernels/layernorm.js +35 -19
  130. package/src/gpu/kernels/logit-merge.js +5 -3
  131. package/src/gpu/kernels/matmul.js +83 -39
  132. package/src/gpu/kernels/modulate.js +23 -15
  133. package/src/gpu/kernels/moe.js +221 -175
  134. package/src/gpu/kernels/pixel_shuffle.js +22 -14
  135. package/src/gpu/kernels/pixel_shuffle.wgsl +4 -5
  136. package/src/gpu/kernels/pixel_shuffle_f16.wgsl +4 -5
  137. package/src/gpu/kernels/relu.js +31 -10
  138. package/src/gpu/kernels/relu.wgsl +2 -1
  139. package/src/gpu/kernels/relu_f16.wgsl +2 -1
  140. package/src/gpu/kernels/repeat_channels.js +25 -17
  141. package/src/gpu/kernels/repeat_channels.wgsl +4 -5
  142. package/src/gpu/kernels/repeat_channels_f16.wgsl +4 -5
  143. package/src/gpu/kernels/residual.js +69 -23
  144. package/src/gpu/kernels/residual.wgsl +6 -3
  145. package/src/gpu/kernels/residual_f16.wgsl +2 -1
  146. package/src/gpu/kernels/residual_f16_vec4.wgsl +2 -1
  147. package/src/gpu/kernels/residual_vec4.wgsl +2 -1
  148. package/src/gpu/kernels/rmsnorm.js +96 -28
  149. package/src/gpu/kernels/rmsnorm.wgsl +14 -6
  150. package/src/gpu/kernels/rmsnorm_f16.wgsl +10 -2
  151. package/src/gpu/kernels/rope.d.ts +2 -0
  152. package/src/gpu/kernels/rope.js +14 -1
  153. package/src/gpu/kernels/rope.wgsl +56 -40
  154. package/src/gpu/kernels/sample.js +27 -38
  155. package/src/gpu/kernels/sana_linear_attention.js +19 -12
  156. package/src/gpu/kernels/sana_linear_attention_apply.wgsl +4 -5
  157. package/src/gpu/kernels/sana_linear_attention_apply_f16.wgsl +4 -5
  158. package/src/gpu/kernels/sana_linear_attention_summary.wgsl +4 -0
  159. package/src/gpu/kernels/sana_linear_attention_summary_f16.wgsl +4 -0
  160. package/src/gpu/kernels/scale.js +18 -11
  161. package/src/gpu/kernels/shader-cache.js +4 -2
  162. package/src/gpu/kernels/silu.d.ts +1 -0
  163. package/src/gpu/kernels/silu.js +148 -82
  164. package/src/gpu/kernels/silu.wgsl +19 -9
  165. package/src/gpu/kernels/silu_f16.wgsl +19 -9
  166. package/src/gpu/kernels/softmax.js +44 -25
  167. package/src/gpu/kernels/split_qkv.js +23 -13
  168. package/src/gpu/kernels/transpose.js +31 -10
  169. package/src/gpu/kernels/transpose.wgsl +6 -5
  170. package/src/gpu/kernels/upsample2d.js +22 -13
  171. package/src/gpu/kernels/upsample2d.wgsl +6 -9
  172. package/src/gpu/kernels/upsample2d_f16.wgsl +6 -9
  173. package/src/gpu/kernels/utils.js +35 -13
  174. package/src/gpu/partitioned-buffer-pool.js +10 -2
  175. package/src/gpu/perf-guards.js +2 -9
  176. package/src/gpu/profiler.js +27 -22
  177. package/src/gpu/readback-utils.d.ts +16 -0
  178. package/src/gpu/readback-utils.js +41 -0
  179. package/src/gpu/submit-tracker.js +13 -0
  180. package/src/gpu/uniform-cache.d.ts +1 -0
  181. package/src/gpu/uniform-cache.js +30 -9
  182. package/src/hotswap/intent-bundle.js +6 -0
  183. package/src/hotswap/manifest.d.ts +10 -1
  184. package/src/hotswap/manifest.js +12 -2
  185. package/src/hotswap/runtime.js +30 -8
  186. package/src/index-browser.d.ts +44 -0
  187. package/src/index-browser.js +14 -0
  188. package/src/inference/browser-harness-contract-helpers.d.ts +5 -0
  189. package/src/inference/browser-harness-contract-helpers.js +28 -0
  190. package/src/inference/browser-harness-diffusion-energy-suites.d.ts +2 -0
  191. package/src/inference/browser-harness-diffusion-energy-suites.js +269 -0
  192. package/src/inference/browser-harness-model-helpers.d.ts +16 -0
  193. package/src/inference/browser-harness-model-helpers.js +217 -0
  194. package/src/inference/browser-harness-report-helpers.d.ts +7 -0
  195. package/src/inference/browser-harness-report-helpers.js +42 -0
  196. package/src/inference/browser-harness-runtime-helpers.d.ts +61 -0
  197. package/src/inference/browser-harness-runtime-helpers.js +415 -0
  198. package/src/inference/browser-harness-suite-helpers.d.ts +28 -0
  199. package/src/inference/browser-harness-suite-helpers.js +268 -0
  200. package/src/inference/browser-harness-text-helpers.d.ts +27 -0
  201. package/src/inference/browser-harness-text-helpers.js +788 -0
  202. package/src/inference/browser-harness.d.ts +6 -0
  203. package/src/inference/browser-harness.js +130 -1950
  204. package/src/inference/kv-cache/base.js +140 -94
  205. package/src/inference/kv-cache/tiered.js +5 -3
  206. package/src/inference/moe-router.js +88 -56
  207. package/src/inference/multi-model-network.js +5 -3
  208. package/src/inference/network-evolution.d.ts +11 -2
  209. package/src/inference/network-evolution.js +20 -21
  210. package/src/inference/pipelines/context.d.ts +3 -0
  211. package/src/inference/pipelines/context.js +142 -2
  212. package/src/inference/pipelines/diffusion/helpers.js +7 -2
  213. package/src/inference/pipelines/diffusion/pipeline.js +17 -7
  214. package/src/inference/pipelines/diffusion/sd3-transformer.js +10 -10
  215. package/src/inference/pipelines/diffusion/text-encoder-gpu.d.ts +5 -0
  216. package/src/inference/pipelines/diffusion/text-encoder-gpu.js +27 -15
  217. package/src/inference/pipelines/diffusion/vae.js +3 -7
  218. package/src/inference/pipelines/energy/pipeline.js +27 -21
  219. package/src/inference/pipelines/energy/quintel.d.ts +5 -0
  220. package/src/inference/pipelines/energy/quintel.js +11 -0
  221. package/src/inference/pipelines/energy-head/row-head-pipeline.js +17 -13
  222. package/src/inference/pipelines/structured/json-head-pipeline.js +26 -11
  223. package/src/inference/pipelines/text/attention/projections.js +151 -101
  224. package/src/inference/pipelines/text/attention/record.js +73 -10
  225. package/src/inference/pipelines/text/attention/run.js +73 -10
  226. package/src/inference/pipelines/text/chat-format.js +25 -1
  227. package/src/inference/pipelines/text/config.d.ts +4 -0
  228. package/src/inference/pipelines/text/config.js +71 -5
  229. package/src/inference/pipelines/text/embed.js +2 -8
  230. package/src/inference/pipelines/text/execution-plan.js +64 -50
  231. package/src/inference/pipelines/text/execution-v0-contract-helpers.d.ts +59 -0
  232. package/src/inference/pipelines/text/execution-v0-contract-helpers.js +937 -0
  233. package/src/inference/pipelines/text/execution-v0-runtime-builders.d.ts +15 -0
  234. package/src/inference/pipelines/text/execution-v0-runtime-builders.js +279 -0
  235. package/src/inference/pipelines/text/execution-v0.js +78 -1002
  236. package/src/inference/pipelines/text/ffn/standard.js +3 -0
  237. package/src/inference/pipelines/text/generator-steps.d.ts +46 -0
  238. package/src/inference/pipelines/text/generator-steps.js +298 -207
  239. package/src/inference/pipelines/text/generator.js +6 -23
  240. package/src/inference/pipelines/text/init.d.ts +4 -0
  241. package/src/inference/pipelines/text/init.js +134 -29
  242. package/src/inference/pipelines/text/kernel-path-auto-select.js +2 -0
  243. package/src/inference/pipelines/text/kernel-trace.d.ts +2 -0
  244. package/src/inference/pipelines/text/kernel-trace.js +6 -0
  245. package/src/inference/pipelines/text/layer.js +14 -9
  246. package/src/inference/pipelines/text/linear-attention.d.ts +10 -0
  247. package/src/inference/pipelines/text/linear-attention.js +80 -6
  248. package/src/inference/pipelines/text/logits/gpu.js +10 -5
  249. package/src/inference/pipelines/text/logits/index.js +10 -11
  250. package/src/inference/pipelines/text/logits/utils.d.ts +7 -0
  251. package/src/inference/pipelines/text/logits/utils.js +9 -0
  252. package/src/inference/pipelines/text/lora-apply.js +50 -32
  253. package/src/inference/pipelines/text/model-load.js +279 -104
  254. package/src/inference/pipelines/text/moe-cache.js +5 -4
  255. package/src/inference/pipelines/text/moe-cpu-gptoss.js +74 -69
  256. package/src/inference/pipelines/text/moe-cpu.js +42 -38
  257. package/src/inference/pipelines/text/moe-gpu.js +110 -86
  258. package/src/inference/pipelines/text/ops.js +90 -90
  259. package/src/inference/pipelines/text/probes.js +9 -9
  260. package/src/inference/pipelines/text/weights.js +17 -7
  261. package/src/inference/pipelines/text.js +17 -1
  262. package/src/inference/speculative.d.ts +2 -2
  263. package/src/inference/speculative.js +4 -18
  264. package/src/inference/test-harness.d.ts +1 -1
  265. package/src/inference/test-harness.js +15 -5
  266. package/src/inference/tokenizer.d.ts +0 -5
  267. package/src/inference/tokenizer.js +4 -23
  268. package/src/inference/tokenizers/bpe.js +9 -0
  269. package/src/inference/tokenizers/bundled.js +176 -33
  270. package/src/inference/tokenizers/sentencepiece.js +12 -0
  271. package/src/loader/doppler-loader.js +38 -22
  272. package/src/loader/dtype-utils.js +3 -44
  273. package/src/loader/embedding-loader.js +7 -3
  274. package/src/loader/experts/expert-cache.js +13 -6
  275. package/src/loader/experts/expert-loader.js +10 -6
  276. package/src/loader/final-weights-loader.js +8 -4
  277. package/src/loader/layer-loader.js +2 -1
  278. package/src/loader/loader-state.js +2 -2
  279. package/src/loader/memory-monitor.js +8 -0
  280. package/src/loader/multi-model-loader.d.ts +14 -0
  281. package/src/loader/multi-model-loader.js +70 -24
  282. package/src/loader/shard-cache.js +81 -12
  283. package/src/loader/shard-resolver.js +25 -3
  284. package/src/loader/tensors/tensor-loader.js +209 -144
  285. package/src/loader/tensors/tensor-reader.js +76 -19
  286. package/src/loader/weight-downcast.js +1 -1
  287. package/src/memory/buffer-pool.d.ts +9 -1
  288. package/src/memory/buffer-pool.js +109 -44
  289. package/src/memory/unified-detect.js +1 -1
  290. package/src/rules/inference/kernel-path.rules.json +24 -8
  291. package/src/rules/rule-registry.js +25 -1
  292. package/src/rules/tooling/command-runtime.rules.json +18 -0
  293. package/src/storage/backends/opfs-store.js +68 -24
  294. package/src/storage/downloader.js +364 -83
  295. package/src/storage/index.d.ts +3 -0
  296. package/src/storage/index.js +3 -0
  297. package/src/storage/preflight.d.ts +2 -2
  298. package/src/storage/preflight.js +24 -2
  299. package/src/storage/quickstart-downloader.js +11 -5
  300. package/src/storage/registry.js +10 -4
  301. package/src/storage/reports.js +1 -1
  302. package/src/storage/shard-manager.d.ts +15 -1
  303. package/src/storage/shard-manager.js +51 -3
  304. package/src/storage/source-artifact-store.d.ts +52 -0
  305. package/src/storage/source-artifact-store.js +234 -0
  306. package/src/tooling/command-api-constants.d.ts +9 -0
  307. package/src/tooling/command-api-constants.js +9 -0
  308. package/src/tooling/command-api-family-normalizers.d.ts +9 -0
  309. package/src/tooling/command-api-family-normalizers.js +343 -0
  310. package/src/tooling/command-api-helpers.d.ts +25 -0
  311. package/src/tooling/command-api-helpers.js +262 -0
  312. package/src/tooling/command-api.d.ts +27 -1
  313. package/src/tooling/command-api.js +26 -473
  314. package/src/tooling/command-envelope.js +4 -1
  315. package/src/tooling/command-runner-shared.js +52 -18
  316. package/src/tooling/lean-execution-contract.js +150 -3
  317. package/src/tooling/node-browser-command-runner.d.ts +4 -0
  318. package/src/tooling/node-browser-command-runner.js +218 -273
  319. package/src/tooling/node-command-runner.js +44 -3
  320. package/src/tooling/node-converter.js +27 -1
  321. package/src/tooling/node-source-runtime.d.ts +1 -1
  322. package/src/tooling/node-source-runtime.js +84 -3
  323. package/src/tooling/node-webgpu.js +30 -105
  324. package/src/tooling/opfs-cache.js +21 -4
  325. package/src/tooling/runtime-input-composition.d.ts +38 -0
  326. package/src/tooling/runtime-input-composition.js +86 -0
  327. package/src/tooling/source-runtime-bundle.d.ts +40 -5
  328. package/src/tooling/source-runtime-bundle.js +261 -34
  329. package/src/tooling/source-runtime-materializer.d.ts +6 -0
  330. package/src/tooling/source-runtime-materializer.js +93 -0
  331. package/src/training/attention-backward.js +32 -17
  332. package/src/training/autograd.js +80 -52
  333. package/src/training/checkpoint-watch.d.ts +8 -0
  334. package/src/training/checkpoint-watch.js +139 -0
  335. package/src/training/checkpoint.d.ts +6 -1
  336. package/src/training/checkpoint.js +46 -7
  337. package/src/training/clip.js +2 -1
  338. package/src/training/datasets/token-batch.js +20 -8
  339. package/src/training/distillation/artifacts.d.ts +71 -0
  340. package/src/training/distillation/artifacts.js +132 -0
  341. package/src/training/distillation/checkpoint-watch.d.ts +10 -0
  342. package/src/training/distillation/checkpoint-watch.js +58 -0
  343. package/src/training/distillation/dataset.d.ts +59 -0
  344. package/src/training/distillation/dataset.js +337 -0
  345. package/src/training/distillation/eval.d.ts +34 -0
  346. package/src/training/distillation/eval.js +310 -0
  347. package/src/training/distillation/index.d.ts +29 -0
  348. package/src/training/distillation/index.js +29 -0
  349. package/src/training/distillation/runtime.d.ts +20 -0
  350. package/src/training/distillation/runtime.js +121 -0
  351. package/src/training/distillation/scoreboard.d.ts +6 -0
  352. package/src/training/distillation/scoreboard.js +8 -0
  353. package/src/training/distillation/stage-a.d.ts +45 -0
  354. package/src/training/distillation/stage-a.js +338 -0
  355. package/src/training/distillation/stage-b.d.ts +24 -0
  356. package/src/training/distillation/stage-b.js +20 -0
  357. package/src/training/distillation/student-fixture.d.ts +22 -0
  358. package/src/training/distillation/student-fixture.js +846 -0
  359. package/src/training/distillation/suite-data.d.ts +45 -0
  360. package/src/training/distillation/suite-data.js +189 -0
  361. package/src/training/index.d.ts +10 -0
  362. package/src/training/index.js +10 -0
  363. package/src/training/lora-pipeline.d.ts +40 -0
  364. package/src/training/lora-pipeline.js +793 -0
  365. package/src/training/lora.js +26 -12
  366. package/src/training/loss.js +5 -6
  367. package/src/training/objectives/cross_entropy.js +2 -5
  368. package/src/training/objectives/distill_kd.js +4 -8
  369. package/src/training/objectives/distill_triplet.js +4 -8
  370. package/src/training/objectives/ul_stage2_base.js +4 -8
  371. package/src/training/operator-artifacts.d.ts +62 -0
  372. package/src/training/operator-artifacts.js +140 -0
  373. package/src/training/operator-command.d.ts +5 -0
  374. package/src/training/operator-command.js +455 -0
  375. package/src/training/operator-eval.d.ts +48 -0
  376. package/src/training/operator-eval.js +230 -0
  377. package/src/training/operator-scoreboard.d.ts +5 -0
  378. package/src/training/operator-scoreboard.js +44 -0
  379. package/src/training/optimizer.js +19 -7
  380. package/src/training/runner.d.ts +52 -0
  381. package/src/training/runner.js +31 -5
  382. package/src/training/suite.d.ts +112 -0
  383. package/src/training/suite.js +24 -984
  384. package/src/training/tensor-factory.d.ts +9 -0
  385. package/src/training/tensor-factory.js +13 -0
  386. package/src/training/trainer.js +3 -5
  387. package/src/training/ul_dataset.js +3 -5
  388. package/src/training/workloads.d.ts +164 -0
  389. package/src/training/workloads.js +530 -0
  390. package/src/version.js +1 -1
  391. package/tools/convert-safetensors-node.js +22 -16
  392. package/tools/doppler-cli.js +179 -63
@@ -0,0 +1,455 @@
1
+ import { readFile, readdir } from 'node:fs/promises';
2
+ import { join, resolve } from 'node:path';
3
+
4
+ import { loadTrainingWorkloadPack } from './workloads.js';
5
+ import {
6
+ buildFrozenSubset,
7
+ createDistillationRunArtifacts,
8
+ evaluateDistillationCheckpoint,
9
+ runDistillationStageA,
10
+ runDistillationStageB,
11
+ watchDistillationCheckpoints,
12
+ } from './distillation/index.js';
13
+ import {
14
+ compareLoraRun,
15
+ evaluateLoraCheckpoint,
16
+ exportLoraCheckpoint,
17
+ qualityGateLoraRun,
18
+ runLoraPipeline,
19
+ watchLoraCheckpoints,
20
+ } from './lora-pipeline.js';
21
+ import { writeDistillCompareReport, writeDistillQualityGateReport } from './distillation/artifacts.js';
22
+
23
+ async function listJsonFiles(rootDir) {
24
+ const results = [];
25
+ const entries = await readdir(rootDir, { withFileTypes: true });
26
+ for (const entry of entries) {
27
+ const absolutePath = join(rootDir, entry.name);
28
+ if (entry.isDirectory()) {
29
+ results.push(...await listJsonFiles(absolutePath));
30
+ continue;
31
+ }
32
+ if (entry.isFile() && entry.name.endsWith('.json')) {
33
+ results.push(absolutePath);
34
+ }
35
+ }
36
+ return results.sort((left, right) => left.localeCompare(right));
37
+ }
38
+
39
+ async function readJson(filePath) {
40
+ const raw = await readFile(filePath, 'utf8');
41
+ return JSON.parse(raw);
42
+ }
43
+
44
+ async function loadWorkloadFromRunRoot(runRoot) {
45
+ const lockPath = join(resolve(String(runRoot)), 'workload.lock.json');
46
+ const payload = await readJson(lockPath);
47
+ return {
48
+ absolutePath: payload.workloadPath,
49
+ path: payload.workloadPath,
50
+ raw: JSON.stringify(payload.workload),
51
+ workloadSha256: payload.workloadSha256,
52
+ workload: payload.workload,
53
+ };
54
+ }
55
+
56
+ function resolveStageEntry(workload, action) {
57
+ const stagePlan = Array.isArray(workload.pipeline?.stagePlan) ? workload.pipeline.stagePlan : [];
58
+ if (action === 'stage-a') {
59
+ return stagePlan.find((entry) => String(entry.trainingStage || entry.id || '').includes('stage_a')) || stagePlan[0];
60
+ }
61
+ if (action === 'stage-b') {
62
+ return stagePlan.find((entry) => String(entry.trainingStage || entry.id || '').includes('stage_b'))
63
+ || stagePlan[stagePlan.length - 1];
64
+ }
65
+ return null;
66
+ }
67
+
68
+ async function buildDistillSubsetIfNeeded(loadedWorkload, layout, subsetManifestPath = null) {
69
+ if (subsetManifestPath) {
70
+ const manifest = await readJson(subsetManifestPath);
71
+ return {
72
+ subsetManifestPath: resolve(String(subsetManifestPath)),
73
+ subsetJsonlPath: manifest.output?.subsetJsonlPath || loadedWorkload.workload.datasetPath,
74
+ manifest,
75
+ };
76
+ }
77
+ if (!loadedWorkload.workload.pipeline.subsetSpec) {
78
+ return null;
79
+ }
80
+ const subset = await buildFrozenSubset({
81
+ datasetPath: loadedWorkload.workload.datasetPath,
82
+ outputDir: join(layout.exports, 'subset'),
83
+ strictPairContract: loadedWorkload.workload.pipeline.strictPairContract === true,
84
+ sourceLangs: loadedWorkload.workload.pipeline.sourceLangs,
85
+ targetLangs: loadedWorkload.workload.pipeline.targetLangs,
86
+ pairAllowlist: loadedWorkload.workload.pipeline.pairAllowlist,
87
+ subsetSpec: loadedWorkload.workload.pipeline.subsetSpec,
88
+ });
89
+ return {
90
+ subsetManifestPath: subset.manifestPath,
91
+ subsetJsonlPath: subset.subsetJsonlPath,
92
+ manifest: subset.manifest,
93
+ };
94
+ }
95
+
96
+ async function compareDistillRun(runRoot) {
97
+ const evalDir = join(runRoot, 'eval');
98
+ const files = await listJsonFiles(evalDir);
99
+ const reports = [];
100
+ for (const filePath of files) {
101
+ const report = await readJson(filePath);
102
+ if (report?.artifactType === 'training_eval_report') {
103
+ reports.push(report);
104
+ }
105
+ }
106
+ const sorted = reports
107
+ .slice()
108
+ .sort((left, right) => Number(right?.primaryScore ?? Number.NEGATIVE_INFINITY) - Number(left?.primaryScore ?? Number.NEGATIVE_INFINITY));
109
+ const payload = {
110
+ artifactType: 'training_compare_report',
111
+ schemaVersion: 1,
112
+ generatedAt: new Date().toISOString(),
113
+ runRoot,
114
+ count: sorted.length,
115
+ best: sorted[0] || null,
116
+ reports: sorted.map((report) => ({
117
+ stage: report.stage || null,
118
+ checkpointId: report.checkpointId || null,
119
+ evalDatasetId: report.evalDatasetId || null,
120
+ primaryMetric: report.primaryMetric || null,
121
+ primaryScore: report.primaryScore ?? null,
122
+ bleu: report.bleu ?? null,
123
+ chrf: report.chrf ?? null,
124
+ reportPath: report.reportPath || null,
125
+ })),
126
+ };
127
+ const artifact = await writeDistillCompareReport({
128
+ compare: join(runRoot, 'compare'),
129
+ }, payload);
130
+ return {
131
+ ...payload,
132
+ comparePath: artifact.path,
133
+ };
134
+ }
135
+
136
+ async function qualityGateDistillRun(runRoot, loadedWorkload) {
137
+ const checks = [];
138
+ const required = [
139
+ join(runRoot, 'run_contract.json'),
140
+ join(runRoot, 'workload.lock.json'),
141
+ ];
142
+ for (const filePath of required) {
143
+ try {
144
+ await readFile(filePath, 'utf8');
145
+ checks.push({ path: filePath, ok: true });
146
+ } catch (error) {
147
+ checks.push({ path: filePath, ok: false, error: error?.message || String(error) });
148
+ }
149
+ }
150
+ const stageManifests = await listJsonFiles(join(runRoot, 'checkpoints'));
151
+ const expectedStageCount = Array.isArray(loadedWorkload.workload.pipeline?.stagePlan)
152
+ ? loadedWorkload.workload.pipeline.stagePlan.length
153
+ : 0;
154
+ const actualStageCount = stageManifests.filter((filePath) => filePath.endsWith('distill_stage_manifest.json')).length;
155
+ checks.push({
156
+ path: join(runRoot, 'checkpoints'),
157
+ ok: actualStageCount >= expectedStageCount,
158
+ expectedStageCount,
159
+ actualStageCount,
160
+ });
161
+ const passed = checks.every((entry) => entry.ok === true);
162
+ const payload = {
163
+ artifactType: 'training_quality_gate',
164
+ schemaVersion: 1,
165
+ generatedAt: new Date().toISOString(),
166
+ runRoot,
167
+ passed,
168
+ checks,
169
+ };
170
+ const artifact = await writeDistillQualityGateReport({
171
+ qualityGate: join(runRoot, 'quality-gate'),
172
+ }, payload);
173
+ return {
174
+ ...payload,
175
+ reportPath: artifact.path,
176
+ };
177
+ }
178
+
179
+ async function runDistillCommand(request) {
180
+ const action = String(request.action || '').trim();
181
+ const loadedWorkload = request.workloadPath
182
+ ? await loadTrainingWorkloadPack(request.workloadPath)
183
+ : await loadWorkloadFromRunRoot(request.runRoot);
184
+ const runArtifacts = await createDistillationRunArtifacts({
185
+ loadedWorkload,
186
+ runRoot: request.runRoot || null,
187
+ timestamp: request.timestamp || null,
188
+ });
189
+ if (action === 'subsets') {
190
+ const subset = await buildDistillSubsetIfNeeded(loadedWorkload, runArtifacts.layout, request.subsetManifest || null);
191
+ if (!subset) {
192
+ throw new Error(`Distill workload "${loadedWorkload.workload.id}" has no subsetSpec and no subsetManifest was provided.`);
193
+ }
194
+ return {
195
+ ok: true,
196
+ kind: 'distill',
197
+ action,
198
+ workloadId: loadedWorkload.workload.id,
199
+ runRoot: runArtifacts.layout.runRoot,
200
+ subset,
201
+ };
202
+ }
203
+ if (action === 'run' || action === 'stage-a' || action === 'stage-b') {
204
+ const subset = await buildDistillSubsetIfNeeded(loadedWorkload, runArtifacts.layout, request.subsetManifest || null);
205
+ const datasetPath = subset?.subsetJsonlPath || loadedWorkload.workload.datasetPath;
206
+ const stageResults = [];
207
+ if (action === 'run') {
208
+ for (const stageEntry of loadedWorkload.workload.pipeline.stagePlan) {
209
+ const stageResult = String(stageEntry.trainingStage || stageEntry.id || '').includes('stage_b')
210
+ ? await runDistillationStageB({
211
+ loadedWorkload,
212
+ stageEntry,
213
+ layout: runArtifacts.layout,
214
+ datasetPath,
215
+ priorStageResult: stageResults[stageResults.length - 1] || null,
216
+ legacyArtifactDir: join(runArtifacts.layout.runRoot, 'legacy-stage-artifacts'),
217
+ timestamp: request.timestamp || null,
218
+ })
219
+ : await runDistillationStageA({
220
+ loadedWorkload,
221
+ stageEntry,
222
+ layout: runArtifacts.layout,
223
+ datasetPath,
224
+ legacyArtifactDir: join(runArtifacts.layout.runRoot, 'legacy-stage-artifacts'),
225
+ timestamp: request.timestamp || null,
226
+ });
227
+ stageResults.push(stageResult);
228
+ }
229
+ } else {
230
+ const stageEntry = resolveStageEntry(loadedWorkload.workload, action);
231
+ if (!stageEntry) {
232
+ throw new Error(`Unable to resolve stage entry for action "${action}".`);
233
+ }
234
+ const stageResult = action === 'stage-b'
235
+ ? await runDistillationStageB({
236
+ loadedWorkload,
237
+ stageEntry,
238
+ layout: runArtifacts.layout,
239
+ datasetPath,
240
+ stageAArtifact: request.stageArtifact || null,
241
+ legacyArtifactDir: join(runArtifacts.layout.runRoot, 'legacy-stage-artifacts'),
242
+ timestamp: request.timestamp || null,
243
+ })
244
+ : await runDistillationStageA({
245
+ loadedWorkload,
246
+ stageEntry,
247
+ layout: runArtifacts.layout,
248
+ datasetPath,
249
+ legacyArtifactDir: join(runArtifacts.layout.runRoot, 'legacy-stage-artifacts'),
250
+ timestamp: request.timestamp || null,
251
+ });
252
+ stageResults.push(stageResult);
253
+ }
254
+ const compare = await compareDistillRun(runArtifacts.layout.runRoot);
255
+ const qualityGate = await qualityGateDistillRun(runArtifacts.layout.runRoot, loadedWorkload);
256
+ return {
257
+ ok: true,
258
+ kind: 'distill',
259
+ action,
260
+ workloadId: loadedWorkload.workload.id,
261
+ runRoot: runArtifacts.layout.runRoot,
262
+ subsetManifest: subset?.subsetManifestPath || null,
263
+ stageResults,
264
+ compare,
265
+ qualityGate,
266
+ };
267
+ }
268
+ if (action === 'eval') {
269
+ if (request.checkpointPath) {
270
+ return {
271
+ ok: true,
272
+ kind: 'distill',
273
+ action,
274
+ reports: await evaluateDistillationCheckpoint({
275
+ loadedWorkload,
276
+ checkpointPath: request.checkpointPath,
277
+ checkpointId: request.checkpointId || null,
278
+ checkpointStep: request.checkpointStep ?? null,
279
+ stageId: request.stageId || null,
280
+ evalDatasetId: request.evalDatasetId || null,
281
+ layout: runArtifacts.layout,
282
+ }),
283
+ };
284
+ }
285
+ const markerFiles = (await listJsonFiles(join(runArtifacts.layout.runRoot, 'checkpoints')))
286
+ .filter((filePath) => filePath.endsWith('checkpoint.complete.json'));
287
+ const reports = [];
288
+ for (const markerPath of markerFiles) {
289
+ const marker = await readJson(markerPath);
290
+ reports.push(...await evaluateDistillationCheckpoint({
291
+ loadedWorkload,
292
+ checkpointPath: marker.checkpointPath,
293
+ checkpointId: marker.checkpointId || null,
294
+ checkpointStep: marker.checkpointStep ?? null,
295
+ stageId: marker.stage || null,
296
+ evalDatasetId: request.evalDatasetId || null,
297
+ layout: runArtifacts.layout,
298
+ stageAArtifact: marker.stageArtifact || null,
299
+ stageAArtifactHash: marker.stageArtifactHash || null,
300
+ }));
301
+ }
302
+ return {
303
+ ok: true,
304
+ kind: 'distill',
305
+ action,
306
+ reports,
307
+ };
308
+ }
309
+ if (action === 'watch') {
310
+ return {
311
+ ok: true,
312
+ kind: 'distill',
313
+ action,
314
+ ...(await watchDistillationCheckpoints({
315
+ loadedWorkload,
316
+ layout: runArtifacts.layout,
317
+ pollIntervalMs: request.pollIntervalMs || null,
318
+ stopWhenIdle: request.stopWhenIdle === true,
319
+ signal: request.signal ?? null,
320
+ })),
321
+ };
322
+ }
323
+ if (action === 'compare') {
324
+ return {
325
+ ok: true,
326
+ kind: 'distill',
327
+ action,
328
+ ...(await compareDistillRun(runArtifacts.layout.runRoot)),
329
+ };
330
+ }
331
+ if (action === 'quality-gate') {
332
+ return {
333
+ ok: true,
334
+ kind: 'distill',
335
+ action,
336
+ ...(await qualityGateDistillRun(runArtifacts.layout.runRoot, loadedWorkload)),
337
+ };
338
+ }
339
+ throw new Error(`Unsupported distill action "${action}".`);
340
+ }
341
+
342
+ async function runLoraCommand(request) {
343
+ const action = String(request.action || '').trim();
344
+ const loadedWorkload = request.workloadPath
345
+ ? await loadTrainingWorkloadPack(request.workloadPath)
346
+ : await loadWorkloadFromRunRoot(request.runRoot);
347
+ if (action === 'run') {
348
+ return runLoraPipeline({
349
+ loadedWorkload,
350
+ runRoot: request.runRoot || null,
351
+ timestamp: request.timestamp || null,
352
+ });
353
+ }
354
+ if (action === 'eval') {
355
+ const checkpointPath = request.checkpointPath
356
+ || (await selectLoraCheckpointPath(request.runRoot)).checkpointPath;
357
+ return {
358
+ ok: true,
359
+ kind: 'lora',
360
+ action,
361
+ reports: await evaluateLoraCheckpoint({
362
+ loadedWorkload,
363
+ checkpointPath,
364
+ checkpointId: request.checkpointId || null,
365
+ checkpointStep: request.checkpointStep ?? null,
366
+ layout: request.runRoot
367
+ ? { eval: join(resolve(String(request.runRoot)), 'eval') }
368
+ : null,
369
+ }),
370
+ };
371
+ }
372
+ if (action === 'watch') {
373
+ return {
374
+ ok: true,
375
+ kind: 'lora',
376
+ action,
377
+ ...(await watchLoraCheckpoints({
378
+ loadedWorkload,
379
+ runRoot: resolve(String(request.runRoot)),
380
+ pollIntervalMs: request.pollIntervalMs || null,
381
+ stopWhenIdle: request.stopWhenIdle === true,
382
+ signal: request.signal ?? null,
383
+ })),
384
+ };
385
+ }
386
+ if (action === 'export') {
387
+ const checkpointSelection = request.checkpointPath
388
+ ? { checkpointPath: request.checkpointPath, checkpointId: request.checkpointId || null }
389
+ : await selectLoraCheckpointPath(request.runRoot);
390
+ return {
391
+ ok: true,
392
+ kind: 'lora',
393
+ action,
394
+ ...(await exportLoraCheckpoint({
395
+ loadedWorkload,
396
+ checkpointPath: checkpointSelection.checkpointPath,
397
+ checkpointId: checkpointSelection.checkpointId || null,
398
+ layout: request.runRoot
399
+ ? { exports: join(resolve(String(request.runRoot)), 'exports') }
400
+ : null,
401
+ })),
402
+ };
403
+ }
404
+ if (action === 'compare') {
405
+ return {
406
+ ok: true,
407
+ kind: 'lora',
408
+ action,
409
+ ...(await compareLoraRun({
410
+ runRoot: resolve(String(request.runRoot)),
411
+ })),
412
+ };
413
+ }
414
+ if (action === 'quality-gate') {
415
+ return {
416
+ ok: true,
417
+ kind: 'lora',
418
+ action,
419
+ ...(await qualityGateLoraRun({
420
+ runRoot: resolve(String(request.runRoot)),
421
+ })),
422
+ };
423
+ }
424
+ if (action === 'activate') {
425
+ throw new Error(
426
+ 'lora activate is not supported in the Node operator runner. The active-model adapter surface currently lives in the browser provider.'
427
+ );
428
+ }
429
+ throw new Error(`Unsupported lora action "${action}".`);
430
+ }
431
+
432
+ async function selectLoraCheckpointPath(runRoot) {
433
+ const checkpointsDir = join(resolve(String(runRoot)), 'checkpoints');
434
+ const markers = (await listJsonFiles(checkpointsDir))
435
+ .filter((filePath) => filePath.endsWith('checkpoint.complete.json'));
436
+ const latest = markers[markers.length - 1];
437
+ if (!latest) {
438
+ throw new Error(`No finalized LoRA checkpoints found in ${checkpointsDir}.`);
439
+ }
440
+ const marker = await readJson(latest);
441
+ return {
442
+ checkpointPath: marker.checkpointPath,
443
+ checkpointId: marker.checkpointId || null,
444
+ };
445
+ }
446
+
447
+ export async function runTrainingOperatorCommand(request) {
448
+ if (request.command === 'distill') {
449
+ return runDistillCommand(request);
450
+ }
451
+ if (request.command === 'lora') {
452
+ return runLoraCommand(request);
453
+ }
454
+ throw new Error(`Unsupported training operator command "${request.command}".`);
455
+ }
@@ -0,0 +1,48 @@
1
+ export interface BleuResult {
2
+ score: number;
3
+ brevityPenalty: number;
4
+ precisions: number[];
5
+ hypothesisLength: number;
6
+ referenceLength: number;
7
+ }
8
+
9
+ export interface ChrfResult {
10
+ score: number;
11
+ precision: number;
12
+ recall: number;
13
+ }
14
+
15
+ export declare function computeBleuScore(
16
+ hypotheses: string[],
17
+ references: string[],
18
+ options?: { maxOrder?: number }
19
+ ): BleuResult;
20
+
21
+ export declare function computeChrfScore(
22
+ hypotheses: string[],
23
+ references: string[],
24
+ options?: { maxOrder?: number; beta?: number }
25
+ ): ChrfResult;
26
+
27
+ export declare function computeExactMatch(
28
+ hypotheses: string[],
29
+ references: string[]
30
+ ): { score: number; matches: number; total: number };
31
+
32
+ export declare function computeAccuracy(
33
+ labels: string[],
34
+ predictions: string[]
35
+ ): { score: number; matches: number; total: number };
36
+
37
+ export declare function computeEvalMetrics(
38
+ evalKind: string,
39
+ hypotheses: string[],
40
+ references: string[],
41
+ options?: Record<string, unknown>
42
+ ): Record<string, unknown>;
43
+
44
+ export declare function loadEvalDataset(datasetPath: string): Promise<{
45
+ absolutePath: string;
46
+ rows: unknown[];
47
+ raw: string;
48
+ }>;