mlx-bun 0.0.13 → 0.3.0

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 (427) hide show
  1. package/README.md +71 -578
  2. package/bin/mlx-bun.mjs +1 -1
  3. package/docs/README.md +91 -0
  4. package/docs/{design → archive}/hlg-sampling.md +3 -3
  5. package/docs/{design → archive/investigations}/audio-input-plan.md +4 -4
  6. package/docs/{design → archive/investigations}/colibri-glm52-port.md +1 -1
  7. package/docs/{design → archive/investigations}/diffusion-gemma-port.md +3 -3
  8. package/docs/{design → archive/investigations}/faithful-l1-consolidation.md +4 -4
  9. package/docs/{investigations → archive/investigations}/hlg-sampling-investigation.md +1 -1
  10. package/docs/{design → archive/investigations}/minicpm5-decode-megakernel.md +11 -11
  11. package/docs/{design → archive/investigations}/optimization_plan.md +1 -1
  12. package/docs/archive/investigations/pareto-specialized-runtime-findings.md +92 -0
  13. package/docs/{design → archive/investigations}/pre-colibri-stabilization.md +5 -5
  14. package/docs/{design → archive/investigations}/the-dreaming-handoff.md +2 -2
  15. package/docs/{planning → archive/planning}/chat-ui-vision.md +1 -1
  16. package/docs/{planning → archive/planning}/chunk-task-roadmap.md +1 -1
  17. package/docs/{planning → archive/planning}/curve-sampler-research-plan.md +4 -4
  18. package/docs/{planning → archive/planning}/memory-docs-and-dag-plan.md +1 -1
  19. package/docs/{planning → archive/planning}/release-notes-v0.0.9.md +1 -1
  20. package/docs/archive/planning/release-notes-v0.2.0.md +49 -0
  21. package/docs/{planning → archive/planning}/website-readme-pass-plan.md +1 -1
  22. package/docs/design/batching.md +569 -0
  23. package/docs/design/decode-speed-program.md +3178 -169
  24. package/docs/design/dreaming-nightly-pipeline.md +675 -282
  25. package/docs/design/generic-model-support.md +579 -334
  26. package/docs/design/kv-cache.md +680 -0
  27. package/docs/design/orpo-training.md +673 -1239
  28. package/docs/design/speculative-decoding.md +1627 -0
  29. package/docs/design/turboquant.md +2102 -0
  30. package/docs/design/unified-engine-frontier-plan.md +1831 -727
  31. package/docs/design/web-chat-redesign.md +666 -1083
  32. package/docs/planning/PRODUCT_ROADMAP.md +34 -1
  33. package/docs/planning/release-notes-v0.3.0.md +35 -0
  34. package/docs/planning/video-series.md +421 -0
  35. package/docs/reference/benchmarks.md +1162 -0
  36. package/docs/reference/cli.md +395 -149
  37. package/docs/reference/distribution.md +321 -181
  38. package/docs/reference/environment.md +348 -0
  39. package/docs/reference/glossary.md +39 -0
  40. package/docs/reference/library-api.md +580 -142
  41. package/docs/reference/memory.md +227 -64
  42. package/docs/reference/models.md +137 -6
  43. package/docs/reference/server-api.md +938 -422
  44. package/docs/reference/server-config.md +652 -320
  45. package/docs/reference/training.md +393 -254
  46. package/docs/reference/troubleshooting.md +166 -0
  47. package/package.json +10 -5
  48. package/src/admission-plan.ts +18 -0
  49. package/src/anthropic.ts +92 -3
  50. package/src/audio/decode.ts +2 -2
  51. package/src/audio/features.ts +2 -2
  52. package/src/backends/mlx/autoregressive.ts +108 -0
  53. package/src/backends/mlx/batch-group.ts +1646 -0
  54. package/src/backends/mlx/diffusion.ts +39 -0
  55. package/src/backends/mlx/draft-target.ts +47 -0
  56. package/src/backends/mlx/evaluation.ts +105 -0
  57. package/src/backends/mlx/fill-append.ts +45 -0
  58. package/src/backends/mlx/gateway-binding.ts +71 -0
  59. package/src/backends/mlx/graph.ts +38 -0
  60. package/src/backends/mlx/memory-client.ts +189 -0
  61. package/src/backends/mlx/memory-guard.ts +40 -0
  62. package/src/backends/mlx/methods.ts +90 -0
  63. package/src/backends/mlx/model-prompt.ts +164 -0
  64. package/src/backends/mlx/model-serving.ts +79 -0
  65. package/src/backends/mlx/prefill.ts +33 -0
  66. package/src/backends/mlx/request-policy.ts +20 -0
  67. package/src/backends/mlx/rollback.ts +51 -0
  68. package/src/backends/mlx/serial-executor.ts +273 -0
  69. package/src/backends/mlx/speculative.ts +114 -0
  70. package/src/backends/mlx/state-views.ts +26 -0
  71. package/src/backends/mlx/text-engine.ts +73 -0
  72. package/src/cli.ts +189 -61
  73. package/src/client.ts +66 -0
  74. package/src/config.ts +21 -4
  75. package/src/contracts/completion.ts +16 -0
  76. package/src/contracts/execution.ts +49 -0
  77. package/src/contracts/generation.ts +103 -0
  78. package/src/contracts/host.ts +7 -0
  79. package/src/contracts/jobs.ts +66 -0
  80. package/src/contracts/pi-provider.ts +30 -0
  81. package/src/contracts/pi.ts +259 -0
  82. package/src/contracts/resources.ts +18 -0
  83. package/src/contracts/scheduling.ts +27 -0
  84. package/src/contracts/tsconfig.json +14 -0
  85. package/src/diffusion/diffusion-generate.ts +218 -153
  86. package/src/embed.ts +1 -1
  87. package/src/engine/admission.ts +59 -0
  88. package/src/engine/cancellation.ts +38 -0
  89. package/src/engine/engine.ts +56 -0
  90. package/src/engine/execution-coordinator.ts +70 -0
  91. package/src/engine/execution-plan.ts +51 -0
  92. package/src/engine/resources.ts +42 -0
  93. package/src/engine/scheduler.ts +67 -0
  94. package/src/engine/session.ts +247 -0
  95. package/src/eval/README.md +44 -1
  96. package/src/eval/kl.ts +263 -23
  97. package/src/eval/reference-logits.ts +207 -0
  98. package/src/eval/runner.ts +10 -79
  99. package/src/eval/tasks/bfcl.ts +2 -1
  100. package/src/eval/tasks/gsm8k.ts +2 -1
  101. package/src/eval/tasks/hashhop.ts +2 -1
  102. package/src/eval/tasks/humaneval.ts +5 -4
  103. package/src/eval/tasks/ifeval.ts +2 -1
  104. package/src/eval/tasks/mmlu.ts +2 -1
  105. package/src/expert-io.ts +3 -1
  106. package/src/expert-offload.ts +3 -1
  107. package/src/fill/echo-index.ts +188 -0
  108. package/src/fill/fill-session.ts +452 -0
  109. package/src/fill/proposal.ts +68 -0
  110. package/src/fill/schema-rows.ts +506 -0
  111. package/src/fill/tool-boundary.ts +103 -0
  112. package/src/fit.ts +24 -43
  113. package/src/generate.ts +740 -337
  114. package/src/grammar.ts +14 -8
  115. package/src/harness-pi.ts +15 -34
  116. package/src/hub-rest.ts +4 -4
  117. package/src/index.ts +27 -0
  118. package/src/inference/denoising.ts +17 -0
  119. package/src/inference/graph.ts +33 -0
  120. package/src/inference/prefill.ts +39 -0
  121. package/src/inference/rollback.ts +7 -0
  122. package/src/jobs/db.ts +4 -0
  123. package/src/jobs/index.ts +2 -0
  124. package/src/jobs/job-entry.ts +15 -14
  125. package/src/jobs/runner.ts +98 -13
  126. package/src/jobs/task-client.ts +19 -0
  127. package/src/jobs/types.ts +2 -58
  128. package/src/kv-scheme.ts +199 -0
  129. package/src/kv-store.ts +459 -234
  130. package/src/lab/README.md +18 -0
  131. package/src/{assets → lab/curve}/curve-designer.html +2 -2
  132. package/src/{curve-sampler.ts → lab/curve/curve-sampler.ts} +2 -2
  133. package/src/{expert-trace.ts → lab/expert-trace/expert-trace.ts} +7 -5
  134. package/src/{model → lab/paged-kv}/paged-kv.ts +9 -6
  135. package/src/library.ts +40 -0
  136. package/src/lora.ts +18 -0
  137. package/src/media-fetch.ts +20 -4
  138. package/src/memory/article.ts +1 -1
  139. package/src/memory/chunk.ts +1 -1
  140. package/src/memory/model.ts +32 -170
  141. package/src/memory/prompts.ts +1 -1
  142. package/src/memory/rest.ts +1 -1
  143. package/src/memory/schedule.ts +1 -1
  144. package/src/memory/stages.ts +4 -1
  145. package/src/memory/vault.ts +3 -4
  146. package/src/mlx/array.ts +9 -10
  147. package/src/mlx/compile.ts +12 -2
  148. package/src/mlx/ffi.ts +66 -22
  149. package/src/mlx/materialize.ts +23 -0
  150. package/src/mlx/metal-kernel.ts +10 -11
  151. package/src/mlx/ops.ts +64 -19
  152. package/src/mlx/quantized-rows.ts +36 -0
  153. package/src/mlx/token-bitmask.ts +52 -0
  154. package/src/mlx/turboquant-kv-decode.ts +82 -0
  155. package/src/mlx/turboquant-kv-inverse.ts +122 -0
  156. package/src/mlx/turboquant-ops.ts +33 -3
  157. package/src/mlx/turboquant-tables.ts +2 -2
  158. package/src/model/batched-mask.ts +9 -3
  159. package/src/model/batched-quant.ts +2 -2
  160. package/src/model/batched-rotating-quant.ts +45 -99
  161. package/src/model/batched-rotating-state.ts +73 -0
  162. package/src/model/batched-rotating.ts +57 -106
  163. package/src/model/batched-row-storage.ts +127 -0
  164. package/src/model/compiled-decode.ts +4 -3
  165. package/src/model/diffusion-gemma.ts +8 -8
  166. package/src/model/factory.ts +87 -49
  167. package/src/model/fingerprint.ts +1 -1
  168. package/src/model/flash-attention.ts +3 -3
  169. package/src/model/gemma4-base.ts +187 -17
  170. package/src/model/gemma4.ts +6 -6
  171. package/src/model/generated/gemma4-12b.ts +52 -148
  172. package/src/model/generated/gemma4-26b.ts +34 -94
  173. package/src/model/generated/gemma4-e4b.ts +28 -76
  174. package/src/model/glm52-cache.ts +19 -9
  175. package/src/model/glm52-dsa.ts +13 -9
  176. package/src/model/glm52-memory.ts +28 -1
  177. package/src/model/implementation.ts +69 -0
  178. package/src/model/minicpm5.ts +7 -8
  179. package/src/model/profile.ts +443 -0
  180. package/src/model/qwen-append.ts +12 -0
  181. package/src/model/qwen-conv.ts +52 -0
  182. package/src/model/qwen3-delta.ts +98 -2
  183. package/src/model/qwen3-mrope.ts +252 -0
  184. package/src/model/qwen3_5-checkpoint.ts +164 -0
  185. package/src/model/qwen3_5.ts +336 -60
  186. package/src/model/support.ts +1 -1
  187. package/src/model/trellis-balanced-scatter.ts +51 -0
  188. package/src/model/trellis-linear.ts +672 -0
  189. package/src/model/trellis-shared-m.ts +151 -0
  190. package/src/model/trellis-shared-scatter.ts +144 -0
  191. package/src/model/trellis-splitk-prefill.ts +158 -0
  192. package/src/model/trellis-tiled-prefill.ts +136 -0
  193. package/src/model/trellis-vector-expand.ts +62 -0
  194. package/src/model/trellis-wide-prefill.ts +94 -0
  195. package/src/model/universal/dense.ts +1 -1
  196. package/src/model/universal/modules.ts +8 -1
  197. package/src/native/frame_extract.swift +140 -0
  198. package/src/native-pack.ts +15 -7
  199. package/src/pi-provider.ts +6 -64
  200. package/src/pi-web.ts +14 -320
  201. package/src/prompt-cache.ts +97 -54
  202. package/src/quantize/allocator.ts +1 -1
  203. package/src/quantize/atomic-output.ts +38 -0
  204. package/src/quantize/calibration.ts +2 -1
  205. package/src/quantize/config-writer.ts +16 -3
  206. package/src/quantize/index.ts +28 -1
  207. package/src/quantize/job.ts +11 -0
  208. package/src/quantize/quantizer.ts +265 -100
  209. package/src/quantize/rotate.ts +603 -0
  210. package/src/quantize/safetensors-writer.ts +79 -1
  211. package/src/quantize/sensitivity.ts +41 -59
  212. package/src/quantize/trellis.ts +586 -0
  213. package/src/quantize/weight-transform.ts +292 -0
  214. package/src/responses.ts +130 -3
  215. package/src/runtime-config.ts +81 -0
  216. package/src/sampler.ts +276 -4
  217. package/src/serve/admin-routes.ts +184 -0
  218. package/src/serve/aux-routes.ts +132 -0
  219. package/src/serve/batch-scheduler.ts +8 -1558
  220. package/src/serve/chat-request.ts +391 -0
  221. package/src/serve/chat-stage.ts +195 -0
  222. package/src/serve/checkpoint-identity.ts +64 -0
  223. package/src/serve/completion-executor.ts +354 -0
  224. package/src/serve/completion-sink.ts +157 -0
  225. package/src/serve/discovery-routes.ts +204 -0
  226. package/src/serve/generation-gateway.ts +211 -243
  227. package/src/serve/http.ts +159 -0
  228. package/src/serve/inference-request.ts +79 -0
  229. package/src/serve/isolate.ts +297 -105
  230. package/src/serve/kv-budget.ts +14 -0
  231. package/src/serve/lab-routes.ts +341 -0
  232. package/src/serve/lane-registry.ts +2 -1
  233. package/src/serve/model-admin-routes.ts +155 -0
  234. package/src/serve/model-binding.ts +43 -0
  235. package/src/serve/model-host.ts +512 -0
  236. package/src/serve/openai-wire.ts +103 -0
  237. package/src/serve/parent-application.ts +56 -0
  238. package/src/serve/pipeline.ts +33 -0
  239. package/src/serve/preparation.ts +33 -0
  240. package/src/serve/prompt-response-trace.ts +147 -0
  241. package/src/serve/request-plan.ts +162 -0
  242. package/src/serve/request-prep.ts +346 -0
  243. package/src/serve/response-lease.ts +24 -0
  244. package/src/serve/responses-client.ts +69 -0
  245. package/src/serve/session-completion-engine.ts +110 -0
  246. package/src/serve/session-search.ts +1 -1
  247. package/src/serve/static-routes.ts +101 -0
  248. package/src/serve/text-completion-stage.ts +104 -0
  249. package/src/serve/token-streams.ts +519 -0
  250. package/src/server.ts +523 -3551
  251. package/src/spec/assistant-source.ts +10 -44
  252. package/src/spec/deepspec-source.ts +6 -6
  253. package/src/spec/dflash-source.ts +9 -8
  254. package/src/spec/dspark/ab-stats.ts +3 -3
  255. package/src/spec/dspark/calibration.ts +2 -2
  256. package/src/spec/dspark/deepspec-module.ts +2 -2
  257. package/src/spec/dspark/generate-dflash.ts +2 -2
  258. package/src/spec/dspark/generate.ts +2 -2
  259. package/src/spec/dspark/loader.ts +1 -1
  260. package/src/spec/dspark/module-dflash.ts +5 -4
  261. package/src/spec/dspark/module.ts +3 -3
  262. package/src/spec/dspark/quantize-drafter.ts +1 -1
  263. package/src/spec/generate.ts +1 -1
  264. package/src/spec/glm52-mtp-source.ts +9 -1
  265. package/src/spec/ngram-source.ts +1 -1
  266. package/src/spec/prefix-state.ts +45 -0
  267. package/src/spec/qwen-mtp-source.ts +528 -0
  268. package/src/spec/serve-loop.ts +179 -213
  269. package/src/spec/source.ts +42 -10
  270. package/src/spec/two-model.ts +1 -1
  271. package/src/ssd-cache.ts +95 -24
  272. package/src/ssd-durability.ts +207 -0
  273. package/src/tool-call.ts +51 -14
  274. package/src/train/flash-cce.ts +14 -13
  275. package/src/train/forward.ts +38 -1
  276. package/src/train/loss.ts +3 -2
  277. package/src/train/prefix-shared.ts +5 -1
  278. package/src/train/segmented.ts +9 -6
  279. package/src/train/trainer.ts +9 -8
  280. package/src/vision/preprocess.ts +1 -1
  281. package/src/vision/prompt.ts +48 -2
  282. package/src/vision/qwen3vl-preprocess.ts +409 -0
  283. package/src/vision/qwen3vl-prompt.ts +196 -0
  284. package/src/vision/qwen3vl-tower.ts +555 -0
  285. package/src/vision/video-frames.ts +148 -0
  286. package/src/web/app.html +1 -1
  287. package/src/web/app.js +15 -6
  288. package/src/web/src/adapters-panel.ts +3 -3
  289. package/src/web/src/assistant.ts +2 -36
  290. package/src/web/src/chat.ts +16 -9
  291. package/src/web/src/composer.ts +4 -4
  292. package/src/web/src/hub.ts +1 -1
  293. package/src/web/src/markdown.ts +0 -0
  294. package/src/web/src/memory-panel.ts +1 -1
  295. package/src/web/src/model-picker.ts +1 -1
  296. package/src/web/src/palette.ts +1 -1
  297. package/src/web/src/protocol.ts +3 -39
  298. package/src/web/src/sessions.ts +1 -1
  299. package/src/web/src/ui-catalog.ts +4 -18
  300. package/src/web-assets.ts +47 -0
  301. package/src/weights.ts +134 -33
  302. package/docs/archive/mlx-bun-lab-report.html +0 -275
  303. package/docs/archive/wwdc-mlx-bun.html +0 -848
  304. package/docs/design/adapters-end-to-end.md +0 -205
  305. package/docs/design/batching-perf-path.md +0 -255
  306. package/docs/design/batching-v2-plan.md +0 -529
  307. package/docs/design/bucketing-stage.md +0 -197
  308. package/docs/design/compat-cli-surface-design.md +0 -230
  309. package/docs/design/docs-reorg-plan.md +0 -165
  310. package/docs/design/dspark-serving-program.md +0 -371
  311. package/docs/design/dspark-speculative-decoding.md +0 -268
  312. package/docs/design/grammar-spec-batching-integration.md +0 -301
  313. package/docs/design/memory-inference-path.md +0 -422
  314. package/docs/design/memory-synthesis.md +0 -287
  315. package/docs/design/memory-system.md +0 -237
  316. package/docs/design/mlx-lm-tool-parity-plan.md +0 -545
  317. package/docs/design/omlx-adoption-map.md +0 -98
  318. package/docs/design/orpo-dynamic-lambda.md +0 -220
  319. package/docs/design/orpo-future-enhancements.md +0 -340
  320. package/docs/design/paged-kv-cache.md +0 -214
  321. package/docs/design/parallel-slots.md +0 -567
  322. package/docs/design/parity-tier-dag.md +0 -161
  323. package/docs/design/repo-cleanup-plan.md +0 -189
  324. package/docs/design/runtime-isolation.md +0 -263
  325. package/docs/design/segmented-backward-training.md +0 -543
  326. package/docs/design/spec-decode-larger-targets.md +0 -229
  327. package/docs/design/ssd-kv-cold-tier.md +0 -288
  328. package/docs/design/structured-output.md +0 -484
  329. package/docs/design/tauri-desktop-app.md +0 -127
  330. package/docs/design/the-dreaming-master-plan.md +0 -445
  331. package/docs/design/turboquant-kv.md +0 -235
  332. package/docs/design/w4a16-compute-precision-spike.md +0 -232
  333. package/docs/design/web-chat-beat-matrix.md +0 -315
  334. package/docs/design/write-pipeline-entity.md +0 -468
  335. package/docs/investigations/curve-runs/bisector-route.json +0 -98
  336. package/docs/investigations/curve-runs/bisector-trace.json +0 -1393
  337. package/docs/investigations/curve-runs/curve-terrain.json +0 -1179
  338. package/docs/investigations/curve-runs/curve-terrain.svg +0 -42
  339. package/docs/investigations/hlg-ab.html +0 -34
  340. package/docs/investigations/hlg-explorer.html +0 -238
  341. package/docs/investigations/hlg-figs/cliff.svg +0 -1
  342. package/docs/investigations/hlg-figs/cross.svg +0 -1
  343. package/docs/investigations/hlg-figs/edges.svg +0 -1
  344. package/docs/investigations/hlg-figs/k20-grids.svg +0 -1
  345. package/docs/investigations/hlg-figs/k20-peaks.svg +0 -1
  346. package/docs/investigations/hlg-figs/lw-curve.svg +0 -1
  347. package/docs/investigations/hlg-figs/terrain.svg +0 -1
  348. package/docs/investigations/hlg-report.html +0 -262
  349. package/docs/investigations/hlg-runs/hlg-ab.json +0 -149
  350. package/docs/investigations/hlg-runs/hlg-blind-prompts.md +0 -87
  351. package/docs/investigations/hlg-runs/hlg-blind.json +0 -107
  352. package/docs/investigations/hlg-runs/hlg-cliff.json +0 -197
  353. package/docs/investigations/hlg-runs/hlg-cross.json +0 -505
  354. package/docs/investigations/hlg-runs/hlg-edge.json +0 -388
  355. package/docs/investigations/hlg-runs/hlg-eval1.json +0 -25
  356. package/docs/investigations/hlg-runs/hlg-frontier.json +0 -86
  357. package/docs/investigations/hlg-runs/hlg-grid.json +0 -242
  358. package/docs/investigations/hlg-runs/hlg-grid2.json +0 -343
  359. package/docs/investigations/hlg-runs/hlg-lw-coarse.json +0 -110
  360. package/docs/investigations/hlg-runs/hlg-map.json +0 -383
  361. package/docs/investigations/hlg-runs/run0-original.log +0 -159
  362. package/docs/investigations/hlg-runs/runA-recommended.log +0 -131
  363. package/docs/investigations/hlg-runs/runAA-cliff-cpm5.log +0 -26
  364. package/docs/investigations/hlg-runs/runAB-grid2-e4b.log +0 -18
  365. package/docs/investigations/hlg-runs/runAC-peak-e4b-partial.log +0 -25
  366. package/docs/investigations/hlg-runs/runAD-ab-samples.log +0 -8
  367. package/docs/investigations/hlg-runs/runAE-edge-e4b.log +0 -51
  368. package/docs/investigations/hlg-runs/runAF-logit-calibration.log +0 -10
  369. package/docs/investigations/hlg-runs/runAG-eval-config2.log +0 -16
  370. package/docs/investigations/hlg-runs/runB-gain-variants.log +0 -192
  371. package/docs/investigations/hlg-runs/runC-pivots.log +0 -162
  372. package/docs/investigations/hlg-runs/runD-oetf.log +0 -132
  373. package/docs/investigations/hlg-runs/runE-eotf.log +0 -162
  374. package/docs/investigations/hlg-runs/runF-shaper-working.log +0 -86
  375. package/docs/investigations/hlg-runs/runG-shaper-sweep.log +0 -101
  376. package/docs/investigations/hlg-runs/runH-wsweep.log +0 -176
  377. package/docs/investigations/hlg-runs/runI-asweep.log +0 -176
  378. package/docs/investigations/hlg-runs/runJ-knobmap.log +0 -49
  379. package/docs/investigations/hlg-runs/runK-frontier.log +0 -19
  380. package/docs/investigations/hlg-runs/runL-widemap.log +0 -65
  381. package/docs/investigations/hlg-runs/runM-frontier-emb.log +0 -15
  382. package/docs/investigations/hlg-runs/runN-verify.log +0 -116
  383. package/docs/investigations/hlg-runs/runO-frontier-seed5000.log +0 -15
  384. package/docs/investigations/hlg-runs/runP-grid-AxSm.log +0 -15
  385. package/docs/investigations/hlg-runs/runQ-grid-tgW.log +0 -15
  386. package/docs/investigations/hlg-runs/runR-blind-transcript.log +0 -79
  387. package/docs/investigations/hlg-runs/runS-cross-coherence.log +0 -33
  388. package/docs/investigations/hlg-runs/runT-cross-cpm5.log +0 -33
  389. package/docs/investigations/hlg-runs/runU-cross-12b.log +0 -33
  390. package/docs/investigations/hlg-runs/runV-lw-e4b.log +0 -22
  391. package/docs/investigations/hlg-runs/runW-lw-12b.log +0 -22
  392. package/docs/investigations/hlg-runs/runX-lw-cpm5.log +0 -22
  393. package/docs/investigations/hlg-runs/runY-cliff-e4b.log +0 -26
  394. package/docs/investigations/hlg-runs/runZ-cliff-12b.log +0 -26
  395. package/docs/planning/IDEAS.md +0 -69
  396. package/docs/reference/embedding.md +0 -197
  397. package/docs/reference/features-matrix.md +0 -100
  398. package/docs/reference/orpo-quickstart.md +0 -135
  399. package/src/flags.ts +0 -13
  400. /package/docs/{investigations → archive/investigations}/chunk-finetune-experiment.md +0 -0
  401. /package/docs/{investigations → archive/investigations}/colibri-oracle-pin.md +0 -0
  402. /package/docs/{investigations → archive/investigations}/curve-bisector-routing.md +0 -0
  403. /package/docs/{investigations → archive/investigations}/decode-roofline-lookagain.md +0 -0
  404. /package/docs/{investigations → archive/investigations}/dspark-handoff.md +0 -0
  405. /package/docs/{investigations → archive/investigations}/expert-offload-single-user-moe.md +0 -0
  406. /package/docs/{investigations → archive/investigations}/indexshare-performance-spike.md +0 -0
  407. /package/docs/{investigations → archive/investigations}/kernel-perf-review-2026-07.md +0 -0
  408. /package/docs/{investigations → archive/investigations}/lab-build-journal.md +0 -0
  409. /package/docs/{investigations → archive/investigations}/openwiki-evaluation.md +0 -0
  410. /package/docs/{investigations → archive/investigations}/orpo-base-uf-experiment-and-directions.md +0 -0
  411. /package/docs/{investigations → archive/investigations}/orpo-flash-cce-pin-leak.md +0 -0
  412. /package/docs/{investigations → archive/investigations}/orpo-uf-testing-handoff.md +0 -0
  413. /package/docs/{investigations → archive/investigations}/pi-builtin-investigation.md +0 -0
  414. /package/docs/{investigations → archive/investigations}/segmented-backward-handoff.md +0 -0
  415. /package/docs/{investigations → archive/investigations}/starter-model-port-handoff.md +0 -0
  416. /package/docs/{investigations → archive/investigations}/steel-flash-cce-handoff.md +0 -0
  417. /package/docs/{investigations → archive/investigations}/trainer-validation-experiment.md +0 -0
  418. /package/docs/{planning → archive/planning}/chat-ui-vision-memory-first.md +0 -0
  419. /package/docs/{planning → archive/planning}/journal.md +0 -0
  420. /package/docs/{planning → archive/planning}/pi-integration-review.md +0 -0
  421. /package/docs/{planning → archive/planning}/release-notes-v0.0.10.md +0 -0
  422. /package/docs/{planning → archive/planning}/release-notes-v0.0.11.md +0 -0
  423. /package/docs/{planning → archive/planning}/release-notes-v0.0.12.md +0 -0
  424. /package/docs/{planning → archive/planning}/release-notes-v0.0.13.md +0 -0
  425. /package/docs/{planning → archive/planning}/web-ui-pass-plan.md +0 -0
  426. /package/{docs/investigations → src/serve/assets}/curve-terrain.html +0 -0
  427. /package/{docs/dag → src/serve/assets}/training-inference-map.html +0 -0
package/docs/investigations/hlg-runs/hlg-edge.json DELETED
@@ -1,388 +0,0 @@
1
- {
2
- "model": "gemma-4-e4b-it-OptiQ-4bit",
3
- "k": 20,
4
- "base": {
5
- "window": 6,
6
- "shoulderA": 0.35,
7
- "sM": 0.7,
8
- "targetGap": 15,
9
- "refGap": 4,
10
- "lw": 1200
11
- },
12
- "zooms": [
13
- {
14
- "name": "target_gap ramp (base) — sub-unit",
15
- "key": "targetGap",
16
- "over": {},
17
- "values": [
18
- 7,
19
- 7.5,
20
- 8,
21
- 8.5,
22
- 9,
23
- 9.5,
24
- 10,
25
- 10.5,
26
- 11
27
- ]
28
- },
29
- {
30
- "name": "W channel-close @ A=100 — sub-unit",
31
- "key": "window",
32
- "over": {
33
- "shoulderA": 100
34
- },
35
- "values": [
36
- 6,
37
- 6.5,
38
- 7,
39
- 7.25,
40
- 7.5,
41
- 7.75,
42
- 8,
43
- 8.5,
44
- 9
45
- ]
46
- },
47
- {
48
- "name": "L_W low wall (γ→0 inversion)",
49
- "key": "lw",
50
- "over": {},
51
- "values": [
52
- 1.4,
53
- 1.6,
54
- 1.8,
55
- 2,
56
- 2.3,
57
- 2.7,
58
- 3.2,
59
- 4,
60
- 6,
61
- 10
62
- ],
63
- "gamma": true
64
- },
65
- {
66
- "name": "L_W high wall (γ→2.4 over-sharpen)",
67
- "key": "lw",
68
- "over": {},
69
- "values": [
70
- 150000,
71
- 200000,
72
- 300000,
73
- 500000,
74
- 700000,
75
- 1000000
76
- ],
77
- "gamma": true
78
- }
79
- ],
80
- "grid": [
81
- {
82
- "zoom": "target_gap ramp (base) — sub-unit",
83
- "key": "targetGap",
84
- "value": 7,
85
- "gamma": null,
86
- "meanJunk": 0.24456282327979945,
87
- "maxJunk": 0.4318181818181818,
88
- "breakRate": 1
89
- },
90
- {
91
- "zoom": "target_gap ramp (base) — sub-unit",
92
- "key": "targetGap",
93
- "value": 7.5,
94
- "gamma": null,
95
- "meanJunk": 0.22452164954945486,
96
- "maxJunk": 0.4090909090909091,
97
- "breakRate": 1
98
- },
99
- {
100
- "zoom": "target_gap ramp (base) — sub-unit",
101
- "key": "targetGap",
102
- "value": 8,
103
- "gamma": null,
104
- "meanJunk": 0.19681853451251913,
105
- "maxJunk": 0.5535714285714286,
106
- "breakRate": 1
107
- },
108
- {
109
- "zoom": "target_gap ramp (base) — sub-unit",
110
- "key": "targetGap",
111
- "value": 8.5,
112
- "gamma": null,
113
- "meanJunk": 0.13733405619367972,
114
- "maxJunk": 0.8372093023255814,
115
- "breakRate": 0.85
116
- },
117
- {
118
- "zoom": "target_gap ramp (base) — sub-unit",
119
- "key": "targetGap",
120
- "value": 9,
121
- "gamma": null,
122
- "meanJunk": 0.0688070885356468,
123
- "maxJunk": 0.6744186046511628,
124
- "breakRate": 0.7
125
- },
126
- {
127
- "zoom": "target_gap ramp (base) — sub-unit",
128
- "key": "targetGap",
129
- "value": 9.5,
130
- "gamma": null,
131
- "meanJunk": 0.05854233213812482,
132
- "maxJunk": 0.7727272727272727,
133
- "breakRate": 0.5
134
- },
135
- {
136
- "zoom": "target_gap ramp (base) — sub-unit",
137
- "key": "targetGap",
138
- "value": 10,
139
- "gamma": null,
140
- "meanJunk": 0.0024470899470899468,
141
- "maxJunk": 0.037037037037037035,
142
- "breakRate": 0.1
143
- },
144
- {
145
- "zoom": "target_gap ramp (base) — sub-unit",
146
- "key": "targetGap",
147
- "value": 10.5,
148
- "gamma": null,
149
- "meanJunk": 0.002339424141749723,
150
- "maxJunk": 0.03488372093023256,
151
- "breakRate": 0.1
152
- },
153
- {
154
- "zoom": "target_gap ramp (base) — sub-unit",
155
- "key": "targetGap",
156
- "value": 11,
157
- "gamma": null,
158
- "meanJunk": 0.003655213615433933,
159
- "maxJunk": 0.03488372093023256,
160
- "breakRate": 0.15
161
- },
162
- {
163
- "zoom": "W channel-close @ A=100 — sub-unit",
164
- "key": "window",
165
- "value": 6,
166
- "gamma": null,
167
- "meanJunk": 0.16097243758841046,
168
- "maxJunk": 0.6190476190476191,
169
- "breakRate": 0.9
170
- },
171
- {
172
- "zoom": "W channel-close @ A=100 — sub-unit",
173
- "key": "window",
174
- "value": 6.5,
175
- "gamma": null,
176
- "meanJunk": 0.08816660696226027,
177
- "maxJunk": 0.42857142857142855,
178
- "breakRate": 0.9
179
- },
180
- {
181
- "zoom": "W channel-close @ A=100 — sub-unit",
182
- "key": "window",
183
- "value": 7,
184
- "gamma": null,
185
- "meanJunk": 0.05842702422655176,
186
- "maxJunk": 0.7678571428571429,
187
- "breakRate": 0.5
188
- },
189
- {
190
- "zoom": "W channel-close @ A=100 — sub-unit",
191
- "key": "window",
192
- "value": 7.25,
193
- "gamma": null,
194
- "meanJunk": 0.050461198587243336,
195
- "maxJunk": 0.7678571428571429,
196
- "breakRate": 0.35
197
- },
198
- {
199
- "zoom": "W channel-close @ A=100 — sub-unit",
200
- "key": "window",
201
- "value": 7.5,
202
- "gamma": null,
203
- "meanJunk": 0.048574950540700275,
204
- "maxJunk": 0.88,
205
- "breakRate": 0.2
206
- },
207
- {
208
- "zoom": "W channel-close @ A=100 — sub-unit",
209
- "key": "window",
210
- "value": 7.75,
211
- "gamma": null,
212
- "meanJunk": 0.0073722988083298555,
213
- "maxJunk": 0.07042253521126761,
214
- "breakRate": 0.2
215
- },
216
- {
217
- "zoom": "W channel-close @ A=100 — sub-unit",
218
- "key": "window",
219
- "value": 8,
220
- "gamma": null,
221
- "meanJunk": 0.002995346586677856,
222
- "maxJunk": 0.03529411764705882,
223
- "breakRate": 0.1
224
- },
225
- {
226
- "zoom": "W channel-close @ A=100 — sub-unit",
227
- "key": "window",
228
- "value": 8.5,
229
- "gamma": null,
230
- "meanJunk": 0,
231
- "maxJunk": 0,
232
- "breakRate": 0
233
- },
234
- {
235
- "zoom": "W channel-close @ A=100 — sub-unit",
236
- "key": "window",
237
- "value": 9,
238
- "gamma": null,
239
- "meanJunk": 0,
240
- "maxJunk": 0,
241
- "breakRate": 0
242
- },
243
- {
244
- "zoom": "L_W low wall (γ→0 inversion)",
245
- "key": "lw",
246
- "value": 1.4,
247
- "gamma": 0.0013737749848601322,
248
- "meanJunk": 0,
249
- "maxJunk": 0,
250
- "breakRate": 0
251
- },
252
- {
253
- "zoom": "L_W low wall (γ→0 inversion)",
254
- "key": "lw",
255
- "value": 1.6,
256
- "gamma": 0.025730392715488604,
257
- "meanJunk": 0,
258
- "maxJunk": 0,
259
- "breakRate": 0
260
- },
261
- {
262
- "zoom": "L_W low wall (γ→0 inversion)",
263
- "key": "lw",
264
- "value": 1.8,
265
- "gamma": 0.0472144521433886,
266
- "meanJunk": 0,
267
- "maxJunk": 0,
268
- "breakRate": 0
269
- },
270
- {
271
- "zoom": "L_W low wall (γ→0 inversion)",
272
- "key": "lw",
273
- "value": 2,
274
- "gamma": 0.06643259817887204,
275
- "meanJunk": 0,
276
- "maxJunk": 0,
277
- "breakRate": 0
278
- },
279
- {
280
- "zoom": "L_W low wall (γ→0 inversion)",
281
- "key": "lw",
282
- "value": 2.3,
283
- "gamma": 0.09192569112738913,
284
- "meanJunk": 0,
285
- "maxJunk": 0,
286
- "breakRate": 0
287
- },
288
- {
289
- "zoom": "L_W low wall (γ→0 inversion)",
290
- "key": "lw",
291
- "value": 2.7,
292
- "gamma": 0.12117278094677464,
293
- "meanJunk": 0,
294
- "maxJunk": 0,
295
- "breakRate": 0
296
- },
297
- {
298
- "zoom": "L_W low wall (γ→0 inversion)",
299
- "key": "lw",
300
- "value": 3.2,
301
- "gamma": 0.15216299089436047,
302
- "meanJunk": 0,
303
- "maxJunk": 0,
304
- "breakRate": 0
305
- },
306
- {
307
- "zoom": "L_W low wall (γ→0 inversion)",
308
- "key": "lw",
309
- "value": 4,
310
- "gamma": 0.19286519635774435,
311
- "meanJunk": 0,
312
- "maxJunk": 0,
313
- "breakRate": 0
314
- },
315
- {
316
- "zoom": "L_W low wall (γ→0 inversion)",
317
- "key": "lw",
318
- "value": 6,
319
- "gamma": 0.2668235251611304,
320
- "meanJunk": 0,
321
- "maxJunk": 0,
322
- "breakRate": 0
323
- },
324
- {
325
- "zoom": "L_W low wall (γ→0 inversion)",
326
- "key": "lw",
327
- "value": 10,
328
- "gamma": 0.36,
329
- "meanJunk": 0,
330
- "maxJunk": 0,
331
- "breakRate": 0
332
- },
333
- {
334
- "zoom": "L_W high wall (γ→2.4 over-sharpen)",
335
- "key": "lw",
336
- "value": 150000,
337
- "gamma": 2.113958328803386,
338
- "meanJunk": 0,
339
- "maxJunk": 0,
340
- "breakRate": 0
341
- },
342
- {
343
- "zoom": "L_W high wall (γ→2.4 over-sharpen)",
344
- "key": "lw",
345
- "value": 200000,
346
- "gamma": 2.166432598178872,
347
- "meanJunk": 0.0013157894736842105,
348
- "maxJunk": 0.02631578947368421,
349
- "breakRate": 0.05
350
- },
351
- {
352
- "zoom": "L_W high wall (γ→2.4 over-sharpen)",
353
- "key": "lw",
354
- "value": 300000,
355
- "gamma": 2.240390926982258,
356
- "meanJunk": 0.0013157894736842105,
357
- "maxJunk": 0.02631578947368421,
358
- "breakRate": 0.05
359
- },
360
- {
361
- "zoom": "L_W high wall (γ→2.4 over-sharpen)",
362
- "key": "lw",
363
- "value": 500000,
364
- "gamma": 2.333567401821128,
365
- "meanJunk": 0.003059975520195838,
366
- "maxJunk": 0.03488372093023256,
367
- "breakRate": 0.1
368
- },
369
- {
370
- "zoom": "L_W high wall (γ→2.4 over-sharpen)",
371
- "key": "lw",
372
- "value": 700000,
373
- "gamma": 2.394941176805988,
374
- "meanJunk": 0.003059975520195838,
375
- "maxJunk": 0.03488372093023256,
376
- "breakRate": 0.1
377
- },
378
- {
379
- "zoom": "L_W high wall (γ→2.4 over-sharpen)",
380
- "key": "lw",
381
- "value": 1000000,
382
- "gamma": 2.46,
383
- "meanJunk": 0.003655213615433933,
384
- "maxJunk": 0.03488372093023256,
385
- "breakRate": 0.15
386
- }
387
- ]
388
- }
package/docs/investigations/hlg-runs/hlg-eval1.json DELETED
@@ -1,25 +0,0 @@
1
- {
2
- "model": "gemma-4-e4b-it-OptiQ-4bit",
3
- "k": 20,
4
- "config": {
5
- "window": 8,
6
- "shoulderA": 0.02,
7
- "sM": 0.15,
8
- "targetGap": 15,
9
- "refGap": 4,
10
- "lw": 1000000,
11
- "xM": 0.77,
12
- "yM": 0.7,
13
- "xFloor": 0.45,
14
- "yFloor": 0.05,
15
- "p": 4
16
- },
17
- "gamma": 2.46,
18
- "refDiv": 0.6333071413347174,
19
- "refJunk": 0,
20
- "canMean": 0.0012987012987012987,
21
- "canMax": 0.025974025974025976,
22
- "broke": 1,
23
- "div": 0.619160697663867,
24
- "openJunk": 0
25
- }
package/docs/investigations/hlg-runs/hlg-frontier.json DELETED
@@ -1,86 +0,0 @@
1
- {
2
- "model": "gemma-4-e4b-it-OptiQ-4bit",
3
- "N": 10,
4
- "reference": {
5
- "sbleu": 0.7546809695639246,
6
- "emb": 0.12136846955801817,
7
- "d2": 0.7729703451628585,
8
- "junk": 0
9
- },
10
- "dominance": true,
11
- "results": [
12
- {
13
- "name": "base (W6 tg15 A.35)",
14
- "over": {},
15
- "canaryJunk": 0,
16
- "sbleu": 0.5195901989780582,
17
- "emb": 0.09224795558066878,
18
- "d2": 0.6371700348032079,
19
- "junk": 0,
20
- "verdict": "acceptable"
21
- },
22
- {
23
- "name": "A=0.01 (loose)",
24
- "over": {
25
- "shoulderA": 0.01
26
- },
27
- "canaryJunk": 0,
28
- "sbleu": 0.7217647475356881,
29
- "emb": 0.1381691980997049,
30
- "d2": 0.7599092778348939,
31
- "junk": 0,
32
- "verdict": "★ beats emb"
33
- },
34
- {
35
- "name": "sM=0.05 (loose)",
36
- "over": {
37
- "sM": 0.05
38
- },
39
- "canaryJunk": 0,
40
- "sbleu": 0.7702710786696176,
41
- "emb": 0.13287383091062793,
42
- "d2": 0.7575513538748833,
43
- "junk": 0,
44
- "verdict": "★★ BEATS both"
45
- },
46
- {
47
- "name": "Lw=10 (loose)",
48
- "over": {
49
- "lw": 10
50
- },
51
- "canaryJunk": 0,
52
- "sbleu": 0.5555123068361736,
53
- "emb": 0.1095044115287212,
54
- "d2": 0.6640863402202801,
55
- "junk": 0,
56
- "verdict": "acceptable"
57
- },
58
- {
59
- "name": "A.01 × sM.05",
60
- "over": {
61
- "shoulderA": 0.01,
62
- "sM": 0.05
63
- },
64
- "canaryJunk": 0,
65
- "sbleu": 0.7465954477438622,
66
- "emb": 0.1410434919251719,
67
- "d2": 0.7616955465075907,
68
- "junk": 0,
69
- "verdict": "★ beats emb"
70
- },
71
- {
72
- "name": "A.01 × sM.05 × Lw10",
73
- "over": {
74
- "shoulderA": 0.01,
75
- "sM": 0.05,
76
- "lw": 10
77
- },
78
- "canaryJunk": 0,
79
- "sbleu": 0.7414095926628356,
80
- "emb": 0.13942094473615174,
81
- "d2": 0.7537338673787272,
82
- "junk": 0,
83
- "verdict": "★ beats emb"
84
- }
85
- ]
86
- }