@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
@@ -1,12 +1,159 @@
1
1
  import {
2
- extractExecutionContractFacts,
3
2
  sanitizeLeanModuleName,
4
3
  } from '../config/execution-contract-check.js';
4
+ import { DEFAULT_BATCHING_DEFAULTS, DEFAULT_GENERATION_CONFIG } from '../config/schema/inference-defaults.schema.js';
5
+ import { DEFAULT_KVCACHE_CONFIG } from '../config/schema/kvcache.schema.js';
5
6
 
6
7
  export {
7
- extractExecutionContractFacts,
8
8
  sanitizeLeanModuleName,
9
- } from '../config/execution-contract-check.js';
9
+ };
10
+
11
+ const KV_LAYOUTS = new Set(['contiguous', 'paged', 'tiered', 'bdpa', 'bdpa_paged']);
12
+ const PHASES = new Set(['prefill', 'decode', 'both']);
13
+ const COLD_QUANT_MODES = new Set(['none', 'int8', 'int4']);
14
+ const ATTENTION_OPS = new Set(['attention']);
15
+ const EMBED_OPS = new Set(['embed', 'gather']);
16
+ const SAMPLE_OPS = new Set(['sample']);
17
+
18
+ function isPlainObject(value) {
19
+ return value != null && typeof value === 'object' && !Array.isArray(value);
20
+ }
21
+
22
+ function assertManifestObject(manifest) {
23
+ if (!isPlainObject(manifest)) {
24
+ throw new Error('lean execution contract: manifest must be an object.');
25
+ }
26
+ return manifest;
27
+ }
28
+
29
+ function assertPositiveInteger(value, label) {
30
+ if (!Number.isInteger(value) || value < 0) {
31
+ throw new Error(`lean execution contract: ${label} must be a non-negative integer.`);
32
+ }
33
+ return value;
34
+ }
35
+
36
+ function assertPositiveIntegerOrDefault(value, fallback, label) {
37
+ if (value == null) {
38
+ return assertPositiveInteger(fallback, `${label} fallback`);
39
+ }
40
+ return assertPositiveInteger(value, label);
41
+ }
42
+
43
+ function normalizeKVLayout(value) {
44
+ const normalized = String(value ?? DEFAULT_KVCACHE_CONFIG.layout).trim().toLowerCase();
45
+ if (!KV_LAYOUTS.has(normalized)) {
46
+ throw new Error(
47
+ `lean execution contract: unsupported KV layout "${value}". ` +
48
+ `Expected one of ${[...KV_LAYOUTS].join(', ')}.`
49
+ );
50
+ }
51
+ return normalized;
52
+ }
53
+
54
+ function normalizePhase(value, label) {
55
+ const normalized = String(value ?? '').trim().toLowerCase();
56
+ if (!PHASES.has(normalized)) {
57
+ throw new Error(
58
+ `lean execution contract: ${label} must be one of ${[...PHASES].join(', ')}.`
59
+ );
60
+ }
61
+ return normalized;
62
+ }
63
+
64
+ function normalizeColdQuantMode(kvcache) {
65
+ const tieringMode = String(
66
+ kvcache?.tiering?.mode
67
+ ?? DEFAULT_KVCACHE_CONFIG.tiering.mode
68
+ ).trim().toLowerCase();
69
+ if (tieringMode === 'off' || tieringMode === 'fp16') {
70
+ return 'none';
71
+ }
72
+ if (!COLD_QUANT_MODES.has(tieringMode)) {
73
+ throw new Error(
74
+ `lean execution contract: unsupported tiered cold quant mode "${tieringMode}".`
75
+ );
76
+ }
77
+ return tieringMode;
78
+ }
79
+
80
+ function classifyOp(op) {
81
+ const normalized = String(op ?? '').trim().toLowerCase();
82
+ if (!normalized) {
83
+ throw new Error('lean execution contract: execution step op is required.');
84
+ }
85
+ if (ATTENTION_OPS.has(normalized)) return 'attention';
86
+ if (EMBED_OPS.has(normalized)) return 'embed';
87
+ if (SAMPLE_OPS.has(normalized)) return 'sample';
88
+ if (normalized.includes('norm')) return 'norm';
89
+ if (normalized.includes('residual')) return 'residual';
90
+ if (normalized.endsWith('_proj') || normalized.startsWith('rope_') || normalized === 'activation') {
91
+ return 'projection';
92
+ }
93
+ return 'other';
94
+ }
95
+
96
+ export function extractExecutionContractFacts(manifest) {
97
+ const normalizedManifest = assertManifestObject(manifest);
98
+ const modelId = String(normalizedManifest.modelId ?? 'model').trim() || 'model';
99
+ const architecture = isPlainObject(normalizedManifest.architecture)
100
+ ? normalizedManifest.architecture
101
+ : {};
102
+ const inference = isPlainObject(normalizedManifest.inference)
103
+ ? normalizedManifest.inference
104
+ : {};
105
+ const sessionDefaults = isPlainObject(inference.sessionDefaults)
106
+ ? inference.sessionDefaults
107
+ : {};
108
+ const execution = isPlainObject(inference.execution)
109
+ ? inference.execution
110
+ : {};
111
+ const kvcache = isPlainObject(sessionDefaults.kvcache)
112
+ ? sessionDefaults.kvcache
113
+ : {};
114
+ const decodeLoop = isPlainObject(sessionDefaults.decodeLoop)
115
+ ? sessionDefaults.decodeLoop
116
+ : {};
117
+
118
+ const steps = Array.isArray(execution.steps)
119
+ ? execution.steps.map((step, index) => {
120
+ if (!isPlainObject(step)) {
121
+ throw new Error(`lean execution contract: execution.steps[${index}] must be an object.`);
122
+ }
123
+ const id = String(step.id ?? '').trim();
124
+ if (!id) {
125
+ throw new Error(`lean execution contract: execution.steps[${index}].id is required.`);
126
+ }
127
+ return {
128
+ id,
129
+ phase: normalizePhase(step.phase, `execution.steps[${index}].phase`),
130
+ opClass: classifyOp(step.op),
131
+ };
132
+ })
133
+ : [];
134
+
135
+ return {
136
+ modelId,
137
+ session: {
138
+ layout: normalizeKVLayout(kvcache.layout),
139
+ disableCommandBatching: decodeLoop.disableCommandBatching
140
+ ?? DEFAULT_GENERATION_CONFIG.disableCommandBatching,
141
+ decodeBatchSize: assertPositiveIntegerOrDefault(
142
+ decodeLoop.batchSize,
143
+ DEFAULT_BATCHING_DEFAULTS.batchSize,
144
+ 'sessionDefaults.decodeLoop.batchSize'
145
+ ),
146
+ headDim: assertPositiveInteger(architecture.headDim, 'architecture.headDim'),
147
+ kvLen: assertPositiveIntegerOrDefault(
148
+ architecture.maxSeqLen ?? kvcache.maxSeqLen,
149
+ DEFAULT_KVCACHE_CONFIG.maxSeqLen,
150
+ 'architecture.maxSeqLen'
151
+ ),
152
+ coldQuantMode: normalizeColdQuantMode(kvcache),
153
+ },
154
+ steps,
155
+ };
156
+ }
10
157
 
11
158
  function asLeanString(value) {
12
159
  return JSON.stringify(String(value));
@@ -7,6 +7,10 @@ import type { BrowserCommandRunResult } from './browser-command-runner.js';
7
7
 
8
8
  export interface NodeBrowserCommandRunOptions {
9
9
  staticRootDir?: string;
10
+ staticMounts?: Array<{
11
+ urlPrefix: string;
12
+ rootDir: string;
13
+ }>;
10
14
  baseUrl?: string;
11
15
  host?: string;
12
16
  port?: number;