mlx-bun 0.4.0 → 1.0.0

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 (601) hide show
  1. package/README.md +1453 -88
  2. package/THIRD_PARTY_NOTICES.md +882 -0
  3. package/bin/mlx-bun.convert.mjs +4 -0
  4. package/bin/mlx-bun.fuse.mjs +4 -0
  5. package/bin/mlx-bun.generate.mjs +4 -0
  6. package/bin/mlx-bun.lora.mjs +4 -0
  7. package/bin/mlx-bun.mjs +13 -21
  8. package/bin/mlx-bun.server.mjs +4 -0
  9. package/bin/mlx-bun.upload.mjs +4 -0
  10. package/dist/web/app.js +9584 -0
  11. package/package.json +50 -53
  12. package/src/cli/ane-install.ts +76 -0
  13. package/src/cli/args.ts +179 -0
  14. package/src/cli/engine-entry.ts +68 -0
  15. package/src/cli/inference.ts +214 -0
  16. package/src/cli/job-entry.ts +97 -0
  17. package/src/cli/library-host.ts +117 -0
  18. package/src/cli/main.ts +57 -0
  19. package/src/cli/memory-engine.ts +163 -0
  20. package/src/cli/memory.ts +36 -0
  21. package/src/cli/mlx-lm-aliases.ts +309 -0
  22. package/src/cli/model-choice.ts +33 -0
  23. package/src/cli/model-selection.ts +84 -0
  24. package/src/cli/module-host.ts +11 -0
  25. package/src/cli/module-verbs.ts +50 -0
  26. package/src/cli/numerical-policy.ts +12 -0
  27. package/src/cli/perplexity.ts +89 -0
  28. package/src/cli/serve-host.ts +423 -0
  29. package/src/cli/serve-isolated.ts +217 -0
  30. package/src/cli/serve-options.ts +106 -0
  31. package/src/cli/serve-state.ts +205 -0
  32. package/src/cli/serve.ts +500 -0
  33. package/src/cli/served-model-host.ts +102 -0
  34. package/src/cli/server-entry.ts +148 -0
  35. package/src/cli/serving-unit.ts +177 -0
  36. package/src/{tui.ts → cli/terminal.ts} +13 -8
  37. package/src/cli/worker-adapters.ts +34 -0
  38. package/src/cli/worker-entry.ts +272 -0
  39. package/src/cli/worker-unit.ts +179 -0
  40. package/src/engine/adapter-operation.ts +26 -0
  41. package/src/engine/cache-services.ts +114 -0
  42. package/src/engine/completion.ts +39 -0
  43. package/src/engine/generation-gateway.ts +358 -0
  44. package/src/engine/index.ts +59 -0
  45. package/src/engine/kv-budget.ts +120 -0
  46. package/src/engine/media-preparation.ts +232 -0
  47. package/src/engine/model-binding.ts +24 -0
  48. package/src/engine/model-host.ts +513 -0
  49. package/src/engine/model-serving.ts +89 -0
  50. package/src/{serve → engine}/preparation.ts +10 -8
  51. package/src/{serve → engine}/session-completion-engine.ts +6 -6
  52. package/src/engine/telemetry.ts +109 -0
  53. package/src/hub/downloads.ts +116 -0
  54. package/src/jobs/db.ts +16 -7
  55. package/src/jobs/events.ts +29 -0
  56. package/src/jobs/executable.ts +5 -0
  57. package/src/jobs/host.ts +51 -0
  58. package/src/jobs/in-process.ts +55 -0
  59. package/src/{contracts/jobs.ts → jobs/protocol.ts} +7 -9
  60. package/src/jobs/runner.ts +150 -231
  61. package/src/jobs/service.ts +75 -0
  62. package/src/jobs/sse.ts +59 -23
  63. package/src/jobs/worker-memory.ts +46 -0
  64. package/src/jobs/worker-process.ts +163 -0
  65. package/src/jobs/worker-supervisor.ts +258 -0
  66. package/src/modules.ts +58 -0
  67. package/src/publishing/catalog-hub.ts +18 -0
  68. package/src/publishing/credentials.ts +31 -0
  69. package/src/publishing/upload.ts +37 -0
  70. package/src/residency/leased-adapters.ts +21 -0
  71. package/src/residency/local-records.ts +16 -0
  72. package/src/residency/model-residency.ts +405 -0
  73. package/src/residency/record-index.ts +28 -0
  74. package/src/residency/resident-estimate.ts +50 -0
  75. package/src/{anthropic.ts → server/anthropic.ts} +3 -81
  76. package/src/server/cache-routes.ts +36 -0
  77. package/src/{serve → server}/chat-request.ts +12 -10
  78. package/src/{serve → server}/chat-stage.ts +27 -22
  79. package/src/{client.ts → server/client.ts} +28 -2
  80. package/src/{serve → server}/completion-executor.ts +42 -42
  81. package/src/server/decision-body.ts +55 -0
  82. package/src/server/decision-routes.ts +134 -0
  83. package/src/server/decision-wire.ts +228 -0
  84. package/src/{serve → server}/discovery-routes.ts +42 -83
  85. package/src/server/generated-token-history.ts +57 -0
  86. package/src/{serve → server}/http.ts +25 -13
  87. package/src/{serve → server}/inference-request.ts +5 -5
  88. package/src/server/job-routes.ts +20 -0
  89. package/src/server/media-prompt.ts +63 -0
  90. package/src/server/memory-completion-client.ts +46 -0
  91. package/src/server/model-routes.ts +192 -0
  92. package/src/{serve → server}/openai-wire.ts +3 -2
  93. package/src/{serve → server}/pipeline.ts +1 -1
  94. package/src/server/prompt-builder.ts +13 -0
  95. package/src/server/prompt-contracts.ts +17 -0
  96. package/src/server/proxy-routes.ts +187 -0
  97. package/src/server/publishing-routes.ts +41 -0
  98. package/src/{serve → server}/request-plan.ts +8 -8
  99. package/src/{serve → server}/request-prep.ts +82 -68
  100. package/src/{serve → server}/responses-client.ts +30 -13
  101. package/src/{responses.ts → server/responses.ts} +12 -93
  102. package/src/server/routes.ts +205 -0
  103. package/src/server/start.ts +72 -0
  104. package/src/server/status-routes.ts +108 -0
  105. package/src/{serve → server}/text-completion-stage.ts +25 -4
  106. package/src/server/text-prompt.ts +12 -0
  107. package/src/{serve → server}/token-streams.ts +36 -38
  108. package/src/server/worker-routes.ts +311 -0
  109. package/src/storage/paths.ts +61 -0
  110. package/src/web/assets.ts +37 -0
  111. package/src/web/{src → browser}/api.ts +5 -23
  112. package/src/web/browser/chat-host.ts +17 -0
  113. package/src/web/browser/installed-panels.ts +6 -0
  114. package/src/web/browser/main.ts +55 -0
  115. package/src/web/browser/markdown.ts +279 -0
  116. package/src/web/browser/memory-panel.ts +9 -0
  117. package/src/web/browser/palette.ts +64 -0
  118. package/src/web/browser/protocol.ts +13 -0
  119. package/src/web/browser/shell.ts +467 -0
  120. package/src/web/{src → browser}/status.ts +32 -36
  121. package/src/web/build.ts +51 -0
  122. package/src/web/public/app.html +1054 -0
  123. package/src/web/{sw.js → public/sw.js} +4 -3
  124. package/src/web/{vendor → public/vendor}/hljs-theme.css +2 -2
  125. package/src/web/{vendor → public/vendor}/hljs.js +2 -2
  126. package/THIRD_PARTY_LICENSES.md +0 -51
  127. package/docs/README.md +0 -94
  128. package/docs/archive/hlg-sampling.md +0 -350
  129. package/docs/archive/investigations/audio-input-plan.md +0 -356
  130. package/docs/archive/investigations/chunk-finetune-experiment.md +0 -84
  131. package/docs/archive/investigations/colibri-glm52-port.md +0 -1486
  132. package/docs/archive/investigations/colibri-oracle-pin.md +0 -1284
  133. package/docs/archive/investigations/curve-bisector-routing.md +0 -123
  134. package/docs/archive/investigations/decode-roofline-lookagain.md +0 -348
  135. package/docs/archive/investigations/diffusion-gemma-port.md +0 -602
  136. package/docs/archive/investigations/dspark-handoff.md +0 -224
  137. package/docs/archive/investigations/expert-offload-single-user-moe.md +0 -292
  138. package/docs/archive/investigations/faithful-l1-consolidation.md +0 -310
  139. package/docs/archive/investigations/hlg-sampling-investigation.md +0 -970
  140. package/docs/archive/investigations/indexshare-performance-spike.md +0 -357
  141. package/docs/archive/investigations/kernel-perf-review-2026-07.md +0 -171
  142. package/docs/archive/investigations/lab-build-journal.md +0 -157
  143. package/docs/archive/investigations/minicpm5-decode-megakernel.md +0 -489
  144. package/docs/archive/investigations/openjarvis-fit.md +0 -198
  145. package/docs/archive/investigations/openwiki-evaluation.md +0 -249
  146. package/docs/archive/investigations/optimization_plan.md +0 -346
  147. package/docs/archive/investigations/orpo-base-uf-experiment-and-directions.md +0 -120
  148. package/docs/archive/investigations/orpo-flash-cce-pin-leak.md +0 -167
  149. package/docs/archive/investigations/orpo-uf-testing-handoff.md +0 -126
  150. package/docs/archive/investigations/pareto-specialized-runtime-findings.md +0 -92
  151. package/docs/archive/investigations/pi-builtin-investigation.md +0 -308
  152. package/docs/archive/investigations/pre-colibri-stabilization.md +0 -329
  153. package/docs/archive/investigations/qwen38-27b-campaign-log.md +0 -1092
  154. package/docs/archive/investigations/segmented-backward-handoff.md +0 -148
  155. package/docs/archive/investigations/starter-model-port-handoff.md +0 -342
  156. package/docs/archive/investigations/steel-flash-cce-handoff.md +0 -489
  157. package/docs/archive/investigations/the-dreaming-handoff.md +0 -141
  158. package/docs/archive/investigations/trainer-validation-experiment.md +0 -173
  159. package/docs/archive/planning/chat-ui-vision-memory-first.md +0 -199
  160. package/docs/archive/planning/chat-ui-vision.md +0 -50
  161. package/docs/archive/planning/chunk-task-roadmap.md +0 -77
  162. package/docs/archive/planning/curve-sampler-research-plan.md +0 -331
  163. package/docs/archive/planning/journal.md +0 -316
  164. package/docs/archive/planning/memory-docs-and-dag-plan.md +0 -194
  165. package/docs/archive/planning/pi-integration-review.md +0 -178
  166. package/docs/archive/planning/release-notes-v0.0.10.md +0 -57
  167. package/docs/archive/planning/release-notes-v0.0.11.md +0 -239
  168. package/docs/archive/planning/release-notes-v0.0.12.md +0 -80
  169. package/docs/archive/planning/release-notes-v0.0.13.md +0 -77
  170. package/docs/archive/planning/release-notes-v0.0.9.md +0 -122
  171. package/docs/archive/planning/release-notes-v0.2.0.md +0 -49
  172. package/docs/archive/planning/release-notes-v0.3.0.md +0 -35
  173. package/docs/archive/planning/web-ui-pass-plan.md +0 -213
  174. package/docs/archive/planning/website-readme-pass-plan.md +0 -224
  175. package/docs/design/batching.md +0 -1841
  176. package/docs/design/decode-speed-program.md +0 -2349
  177. package/docs/design/dreaming-nightly-pipeline.md +0 -684
  178. package/docs/design/generic-model-support.md +0 -705
  179. package/docs/design/kv-cache.md +0 -824
  180. package/docs/design/orpo-training.md +0 -680
  181. package/docs/design/speculative-decoding.md +0 -1990
  182. package/docs/design/turboquant.md +0 -2212
  183. package/docs/design/unified-engine-frontier-plan.md +0 -1880
  184. package/docs/design/web-chat-redesign.md +0 -695
  185. package/docs/planning/PRODUCT_ROADMAP.md +0 -594
  186. package/docs/planning/ResearchTopics.md +0 -119
  187. package/docs/planning/release-notes-v0.4.0.md +0 -45
  188. package/docs/planning/video-series.md +0 -421
  189. package/docs/reference/benchmarks.md +0 -3613
  190. package/docs/reference/cli.md +0 -585
  191. package/docs/reference/distribution.md +0 -355
  192. package/docs/reference/environment.md +0 -398
  193. package/docs/reference/glossary.md +0 -39
  194. package/docs/reference/library-api.md +0 -674
  195. package/docs/reference/memory.md +0 -260
  196. package/docs/reference/models.md +0 -334
  197. package/docs/reference/server-api.md +0 -1775
  198. package/docs/reference/server-config.md +0 -846
  199. package/docs/reference/training.md +0 -462
  200. package/docs/reference/troubleshooting.md +0 -158
  201. package/src/admission-plan.ts +0 -18
  202. package/src/audio/conformer.ts +0 -647
  203. package/src/audio/decode.ts +0 -133
  204. package/src/audio/features.ts +0 -202
  205. package/src/audio/transcode.ts +0 -67
  206. package/src/backends/mlx/assistant-target.ts +0 -46
  207. package/src/backends/mlx/autoregressive-draft-rows.ts +0 -120
  208. package/src/backends/mlx/autoregressive.ts +0 -108
  209. package/src/backends/mlx/batch-group.ts +0 -1790
  210. package/src/backends/mlx/cache-attachment.ts +0 -67
  211. package/src/backends/mlx/cache-layout.ts +0 -31
  212. package/src/backends/mlx/checkpoint-state.ts +0 -53
  213. package/src/backends/mlx/continuation-persistence.ts +0 -106
  214. package/src/backends/mlx/continuation-request.ts +0 -43
  215. package/src/backends/mlx/continuation.ts +0 -77
  216. package/src/backends/mlx/diffusion.ts +0 -39
  217. package/src/backends/mlx/draft-target.ts +0 -32
  218. package/src/backends/mlx/evaluation.ts +0 -105
  219. package/src/backends/mlx/fill-append.ts +0 -45
  220. package/src/backends/mlx/full-kv-row-append.ts +0 -32
  221. package/src/backends/mlx/full-kv-row-donor.ts +0 -54
  222. package/src/backends/mlx/full-transitioning-kv-rows.ts +0 -65
  223. package/src/backends/mlx/gateway-binding.ts +0 -130
  224. package/src/backends/mlx/graph.ts +0 -38
  225. package/src/backends/mlx/kv-codec.ts +0 -9
  226. package/src/backends/mlx/kv-maintenance.ts +0 -133
  227. package/src/backends/mlx/kv-tensor-rows.ts +0 -212
  228. package/src/backends/mlx/memory-client.ts +0 -189
  229. package/src/backends/mlx/methods.ts +0 -90
  230. package/src/backends/mlx/model-prompt.ts +0 -164
  231. package/src/backends/mlx/model-serving.ts +0 -79
  232. package/src/backends/mlx/prefill-cohort.ts +0 -89
  233. package/src/backends/mlx/prefill-rows.ts +0 -200
  234. package/src/backends/mlx/prefill.ts +0 -33
  235. package/src/backends/mlx/projected-context-rows.ts +0 -72
  236. package/src/backends/mlx/request-policy.ts +0 -20
  237. package/src/backends/mlx/request-state-policy.ts +0 -41
  238. package/src/backends/mlx/rollback.ts +0 -89
  239. package/src/backends/mlx/row-sampling.ts +0 -49
  240. package/src/backends/mlx/serial-executor.ts +0 -242
  241. package/src/backends/mlx/speculative-group.ts +0 -359
  242. package/src/backends/mlx/speculative-round.ts +0 -161
  243. package/src/backends/mlx/speculative.ts +0 -114
  244. package/src/backends/mlx/state-rows.ts +0 -104
  245. package/src/backends/mlx/state-views.ts +0 -46
  246. package/src/backends/mlx/target-layout-capability.ts +0 -35
  247. package/src/backends/mlx/text-engine.ts +0 -73
  248. package/src/backends/mlx/transitioning-kv-rows.ts +0 -165
  249. package/src/chat-template.ts +0 -382
  250. package/src/cli.ts +0 -3168
  251. package/src/config.ts +0 -399
  252. package/src/contracts/completion.ts +0 -16
  253. package/src/contracts/execution.ts +0 -60
  254. package/src/contracts/generation.ts +0 -103
  255. package/src/contracts/host.ts +0 -7
  256. package/src/contracts/pi-provider.ts +0 -30
  257. package/src/contracts/pi.ts +0 -259
  258. package/src/contracts/prefix-cache.ts +0 -23
  259. package/src/contracts/resources.ts +0 -25
  260. package/src/contracts/sampling.ts +0 -23
  261. package/src/contracts/scheduling.ts +0 -45
  262. package/src/contracts/tsconfig.json +0 -14
  263. package/src/dataset/generators.ts +0 -1087
  264. package/src/dataset/index.ts +0 -12
  265. package/src/dataset/job.ts +0 -81
  266. package/src/dataset/llm.ts +0 -71
  267. package/src/dataset/registry.ts +0 -528
  268. package/src/diffusion/diffusion-generate.ts +0 -507
  269. package/src/download.ts +0 -505
  270. package/src/embed.ts +0 -77
  271. package/src/engine/admission.ts +0 -59
  272. package/src/engine/cancellation.ts +0 -38
  273. package/src/engine/engine.ts +0 -56
  274. package/src/engine/execution-coordinator.ts +0 -70
  275. package/src/engine/execution-plan.ts +0 -53
  276. package/src/engine/resources.ts +0 -53
  277. package/src/engine/scheduler.ts +0 -72
  278. package/src/engine/session.ts +0 -247
  279. package/src/eval/README.md +0 -130
  280. package/src/eval/ifeval.ts +0 -81
  281. package/src/eval/kl-prompts.ts +0 -29
  282. package/src/eval/kl.ts +0 -590
  283. package/src/eval/perplexity.ts +0 -164
  284. package/src/eval/quality-db.ts +0 -92
  285. package/src/eval/reference-logits.ts +0 -207
  286. package/src/eval/runner.ts +0 -124
  287. package/src/eval/score.ts +0 -49
  288. package/src/eval/tasks/bfcl.ts +0 -535
  289. package/src/eval/tasks/gsm8k.ts +0 -98
  290. package/src/eval/tasks/hashhop.ts +0 -139
  291. package/src/eval/tasks/humaneval.ts +0 -275
  292. package/src/eval/tasks/ifeval.ts +0 -525
  293. package/src/eval/tasks/mmlu.ts +0 -157
  294. package/src/eval/ultrafeedback.ts +0 -60
  295. package/src/evaldb.ts +0 -199
  296. package/src/expert-io.ts +0 -393
  297. package/src/expert-offload-build.ts +0 -77
  298. package/src/expert-offload.ts +0 -76
  299. package/src/expert-residency.ts +0 -1134
  300. package/src/expert-usage.ts +0 -415
  301. package/src/fill/echo-index.ts +0 -188
  302. package/src/fill/fill-session.ts +0 -452
  303. package/src/fill/proposal.ts +0 -68
  304. package/src/fill/schema-rows.ts +0 -506
  305. package/src/fill/tool-boundary.ts +0 -103
  306. package/src/fit.ts +0 -249
  307. package/src/generate.ts +0 -1346
  308. package/src/grammar.ts +0 -591
  309. package/src/harness-pi.ts +0 -184
  310. package/src/hf-push.ts +0 -628
  311. package/src/hub-rest.ts +0 -242
  312. package/src/index.ts +0 -63
  313. package/src/inference/denoising.ts +0 -17
  314. package/src/inference/draft-acceptance.ts +0 -41
  315. package/src/inference/draft-output.ts +0 -49
  316. package/src/inference/graph.ts +0 -33
  317. package/src/inference/prefill.ts +0 -39
  318. package/src/inference/rollback.ts +0 -7
  319. package/src/jobs/index.ts +0 -31
  320. package/src/jobs/job-entry.ts +0 -107
  321. package/src/jobs/task-client.ts +0 -19
  322. package/src/jobs/types.ts +0 -2
  323. package/src/kv-scheme.ts +0 -257
  324. package/src/kv-store.ts +0 -1099
  325. package/src/lab/README.md +0 -18
  326. package/src/lab/curve/curve-designer.html +0 -303
  327. package/src/lab/curve/curve-sampler.ts +0 -122
  328. package/src/lab/expert-trace/expert-trace.ts +0 -69
  329. package/src/lab/paged-kv/paged-kv-rows.ts +0 -83
  330. package/src/lab/paged-kv/paged-kv.ts +0 -270
  331. package/src/library.ts +0 -40
  332. package/src/lora.ts +0 -408
  333. package/src/md-modules.d.ts +0 -49
  334. package/src/media-fetch.ts +0 -442
  335. package/src/memory/article.ts +0 -558
  336. package/src/memory/chunk-validation.ts +0 -150
  337. package/src/memory/chunk.ts +0 -318
  338. package/src/memory/cluster.ts +0 -351
  339. package/src/memory/coverage.ts +0 -134
  340. package/src/memory/crosslink.ts +0 -0
  341. package/src/memory/db.ts +0 -253
  342. package/src/memory/entity.ts +0 -203
  343. package/src/memory/gate.ts +0 -187
  344. package/src/memory/history.ts +0 -152
  345. package/src/memory/ingest.ts +0 -335
  346. package/src/memory/model.ts +0 -183
  347. package/src/memory/normalize.ts +0 -671
  348. package/src/memory/parse.ts +0 -22
  349. package/src/memory/pipeline.ts +0 -275
  350. package/src/memory/prompts.ts +0 -65
  351. package/src/memory/query.ts +0 -334
  352. package/src/memory/reconcile.ts +0 -734
  353. package/src/memory/redirect.ts +0 -108
  354. package/src/memory/reindex.ts +0 -325
  355. package/src/memory/resolve.ts +0 -307
  356. package/src/memory/rest.ts +0 -245
  357. package/src/memory/route.ts +0 -272
  358. package/src/memory/schedule.ts +0 -165
  359. package/src/memory/stages.ts +0 -777
  360. package/src/memory/synthesize.ts +0 -1638
  361. package/src/memory/tools.ts +0 -749
  362. package/src/memory/vault.ts +0 -1194
  363. package/src/memory/wikify.ts +0 -1064
  364. package/src/mlx/array.ts +0 -388
  365. package/src/mlx/autograd.ts +0 -304
  366. package/src/mlx/checkpoint.ts +0 -149
  367. package/src/mlx/compile.ts +0 -168
  368. package/src/mlx/custom-vjp.ts +0 -188
  369. package/src/mlx/ffi.ts +0 -512
  370. package/src/mlx/materialize.ts +0 -23
  371. package/src/mlx/metal-kernel.ts +0 -159
  372. package/src/mlx/ops.ts +0 -1021
  373. package/src/mlx/quantized-rows.ts +0 -36
  374. package/src/mlx/token-bitmask.ts +0 -52
  375. package/src/mlx/turboquant-kv-decode.ts +0 -82
  376. package/src/mlx/turboquant-kv-inverse.ts +0 -122
  377. package/src/mlx/turboquant-ops.ts +0 -861
  378. package/src/mlx/turboquant-tables.ts +0 -237
  379. package/src/mmap.ts +0 -90
  380. package/src/model/artifact-identity.ts +0 -73
  381. package/src/model/batched-kv.ts +0 -235
  382. package/src/model/batched-mask.ts +0 -279
  383. package/src/model/batched-quant.ts +0 -212
  384. package/src/model/batched-quantized-kv.ts +0 -144
  385. package/src/model/batched-rotating-quant.ts +0 -353
  386. package/src/model/batched-rotating-state.ts +0 -149
  387. package/src/model/batched-rotating.ts +0 -338
  388. package/src/model/batched-row-storage.ts +0 -179
  389. package/src/model/batched-ssm.ts +0 -52
  390. package/src/model/batched-turboquant-kv.ts +0 -93
  391. package/src/model/compiled-decode.ts +0 -712
  392. package/src/model/decoded-kv-donor.ts +0 -12
  393. package/src/model/delayed-quantized-kv.ts +0 -72
  394. package/src/model/delayed-rotating-quantized-kv.ts +0 -93
  395. package/src/model/delayed-turboquant-kv.ts +0 -55
  396. package/src/model/diffusion-gemma.ts +0 -1026
  397. package/src/model/factory.ts +0 -189
  398. package/src/model/fingerprint.ts +0 -33
  399. package/src/model/flash-attention.ts +0 -694
  400. package/src/model/full-prefill-padding.ts +0 -56
  401. package/src/model/full-prefill-row.ts +0 -105
  402. package/src/model/gemma4-base.ts +0 -2351
  403. package/src/model/gemma4.ts +0 -1315
  404. package/src/model/generated/gemma4-12b.ts +0 -429
  405. package/src/model/generated/gemma4-26b.ts +0 -448
  406. package/src/model/generated/gemma4-e4b.ts +0 -717
  407. package/src/model/generated/index.ts +0 -18
  408. package/src/model/glm52-atlas.ts +0 -511
  409. package/src/model/glm52-cache.ts +0 -792
  410. package/src/model/glm52-config.ts +0 -262
  411. package/src/model/glm52-container.ts +0 -305
  412. package/src/model/glm52-coupling.ts +0 -388
  413. package/src/model/glm52-dsa.ts +0 -606
  414. package/src/model/glm52-expert-layout.ts +0 -253
  415. package/src/model/glm52-layout.ts +0 -152
  416. package/src/model/glm52-memory.ts +0 -500
  417. package/src/model/glm52-mla.ts +0 -792
  418. package/src/model/glm52-moe.ts +0 -223
  419. package/src/model/glm52-pilot.ts +0 -537
  420. package/src/model/glm52-reference.ts +0 -445
  421. package/src/model/glm52-residency.ts +0 -557
  422. package/src/model/glm52-resident-weights.ts +0 -135
  423. package/src/model/glm52-streamed-experts.ts +0 -453
  424. package/src/model/glm52-streamed-metal.ts +0 -610
  425. package/src/model/glm52-weights.ts +0 -503
  426. package/src/model/glm52.ts +0 -1132
  427. package/src/model/implementation.ts +0 -69
  428. package/src/model/kv-attention-view.ts +0 -67
  429. package/src/model/minicpm5.ts +0 -376
  430. package/src/model/profile.ts +0 -443
  431. package/src/model/qwen-append.ts +0 -12
  432. package/src/model/qwen-conv.ts +0 -52
  433. package/src/model/qwen3-delta-state.ts +0 -74
  434. package/src/model/qwen3-delta.ts +0 -451
  435. package/src/model/qwen3-moe.ts +0 -404
  436. package/src/model/qwen3-mrope.ts +0 -252
  437. package/src/model/qwen3.ts +0 -277
  438. package/src/model/qwen3_5-checkpoint.ts +0 -164
  439. package/src/model/qwen3_5.ts +0 -819
  440. package/src/model/rotating-kv-donor.ts +0 -35
  441. package/src/model/rotating-kv-layout.ts +0 -262
  442. package/src/model/rotating-row-transaction.ts +0 -35
  443. package/src/model/speculative-rotating-kv.ts +0 -101
  444. package/src/model/ssm-prefill-padding.ts +0 -39
  445. package/src/model/support.ts +0 -102
  446. package/src/model/trellis-balanced-scatter.ts +0 -51
  447. package/src/model/trellis-linear.ts +0 -672
  448. package/src/model/trellis-shared-m.ts +0 -151
  449. package/src/model/trellis-shared-scatter.ts +0 -144
  450. package/src/model/trellis-splitk-prefill.ts +0 -158
  451. package/src/model/trellis-tiled-prefill.ts +0 -136
  452. package/src/model/trellis-vector-expand.ts +0 -62
  453. package/src/model/trellis-wide-prefill.ts +0 -94
  454. package/src/model/turboquant-codec.ts +0 -83
  455. package/src/model/universal/archs.ts +0 -332
  456. package/src/model/universal/dense.ts +0 -612
  457. package/src/model/universal/modules.ts +0 -170
  458. package/src/model/universal/rope.ts +0 -381
  459. package/src/native/expert_io.c +0 -523
  460. package/src/native/frame_extract.swift +0 -140
  461. package/src/native-pack.ts +0 -162
  462. package/src/pi-provider.ts +0 -53
  463. package/src/pi-session.ts +0 -81
  464. package/src/pi-terminal.ts +0 -274
  465. package/src/pi-web.ts +0 -1932
  466. package/src/preflight.ts +0 -158
  467. package/src/prompt-cache.ts +0 -460
  468. package/src/quantize/allocator.ts +0 -627
  469. package/src/quantize/atomic-output.ts +0 -38
  470. package/src/quantize/calibration.ts +0 -181
  471. package/src/quantize/config-writer.ts +0 -145
  472. package/src/quantize/index.ts +0 -84
  473. package/src/quantize/job.ts +0 -162
  474. package/src/quantize/quantizer.ts +0 -666
  475. package/src/quantize/rotate.ts +0 -603
  476. package/src/quantize/safetensors-writer.ts +0 -249
  477. package/src/quantize/sensitivity.ts +0 -216
  478. package/src/quantize/trellis.ts +0 -586
  479. package/src/quantize/weight-transform.ts +0 -292
  480. package/src/registry.ts +0 -602
  481. package/src/runtime-config.ts +0 -81
  482. package/src/safetensors.ts +0 -158
  483. package/src/sampler.ts +0 -1011
  484. package/src/serve/admin-routes.ts +0 -184
  485. package/src/serve/assets/curve-terrain.html +0 -105
  486. package/src/serve/assets/training-inference-map.html +0 -575
  487. package/src/serve/aux-routes.ts +0 -132
  488. package/src/serve/batch-scheduler.ts +0 -8
  489. package/src/serve/checkpoint-identity.ts +0 -64
  490. package/src/serve/generation-gateway.ts +0 -496
  491. package/src/serve/isolate.ts +0 -609
  492. package/src/serve/kv-budget.ts +0 -14
  493. package/src/serve/lab-routes.ts +0 -341
  494. package/src/serve/lane-registry.ts +0 -45
  495. package/src/serve/model-admin-routes.ts +0 -155
  496. package/src/serve/model-binding.ts +0 -43
  497. package/src/serve/model-host.ts +0 -511
  498. package/src/serve/parent-application.ts +0 -56
  499. package/src/serve/prompt-response-trace.ts +0 -147
  500. package/src/serve/response-lease.ts +0 -24
  501. package/src/serve/session-search.ts +0 -326
  502. package/src/serve/static-routes.ts +0 -101
  503. package/src/server.ts +0 -1314
  504. package/src/spec/assistant-rows.ts +0 -92
  505. package/src/spec/assistant-source.ts +0 -126
  506. package/src/spec/cache-identity.ts +0 -22
  507. package/src/spec/deepspec-rows.ts +0 -29
  508. package/src/spec/deepspec-source.ts +0 -145
  509. package/src/spec/dflash-source.ts +0 -149
  510. package/src/spec/drafter.ts +0 -304
  511. package/src/spec/dspark/ab-stats.ts +0 -136
  512. package/src/spec/dspark/calibration.ts +0 -112
  513. package/src/spec/dspark/data-dflash.ts +0 -156
  514. package/src/spec/dspark/data.ts +0 -201
  515. package/src/spec/dspark/deepspec-module.ts +0 -776
  516. package/src/spec/dspark/generate-dflash.ts +0 -245
  517. package/src/spec/dspark/generate.ts +0 -290
  518. package/src/spec/dspark/loader.ts +0 -23
  519. package/src/spec/dspark/loss.ts +0 -213
  520. package/src/spec/dspark/module-dflash.ts +0 -560
  521. package/src/spec/dspark/module.ts +0 -540
  522. package/src/spec/dspark/quantize-drafter.ts +0 -91
  523. package/src/spec/dspark/sample.ts +0 -112
  524. package/src/spec/generate.ts +0 -276
  525. package/src/spec/glm52-mtp-graph.ts +0 -29
  526. package/src/spec/glm52-mtp-rows.ts +0 -132
  527. package/src/spec/glm52-mtp-source.ts +0 -317
  528. package/src/spec/glm52-mtp-state.ts +0 -33
  529. package/src/spec/ngram-source.ts +0 -204
  530. package/src/spec/projected-draft-rows.ts +0 -73
  531. package/src/spec/qwen-mtp-module.ts +0 -171
  532. package/src/spec/qwen-mtp-rows.ts +0 -219
  533. package/src/spec/qwen-mtp-source.ts +0 -383
  534. package/src/spec/qwen-mtp-state.ts +0 -30
  535. package/src/spec/serve-loop.ts +0 -585
  536. package/src/spec/source.ts +0 -215
  537. package/src/spec/two-model-rows.ts +0 -61
  538. package/src/spec/two-model.ts +0 -113
  539. package/src/ssd-cache.ts +0 -372
  540. package/src/ssd-durability.ts +0 -211
  541. package/src/tokenizer.ts +0 -335
  542. package/src/tool-approvals.ts +0 -131
  543. package/src/tool-call.ts +0 -695
  544. package/src/train/dataset.ts +0 -484
  545. package/src/train/diffusion-lora.ts +0 -216
  546. package/src/train/export.ts +0 -32
  547. package/src/train/flash-cce.ts +0 -1234
  548. package/src/train/forward.ts +0 -266
  549. package/src/train/fuse.ts +0 -243
  550. package/src/train/index.ts +0 -46
  551. package/src/train/job.ts +0 -205
  552. package/src/train/lora-params.ts +0 -254
  553. package/src/train/loss.ts +0 -1591
  554. package/src/train/merge.ts +0 -195
  555. package/src/train/optimizer.ts +0 -183
  556. package/src/train/prefix-shared.ts +0 -490
  557. package/src/train/rank.ts +0 -137
  558. package/src/train/segmented.ts +0 -1828
  559. package/src/train/steel-qmm-header.ts +0 -1461
  560. package/src/train/trainer.ts +0 -1221
  561. package/src/train/watch.ts +0 -542
  562. package/src/vision/diffusion-vision.ts +0 -620
  563. package/src/vision/embedder.ts +0 -168
  564. package/src/vision/preprocess.ts +0 -247
  565. package/src/vision/prompt.ts +0 -373
  566. package/src/vision/qwen3vl-preprocess.ts +0 -409
  567. package/src/vision/qwen3vl-prompt.ts +0 -196
  568. package/src/vision/qwen3vl-tower.ts +0 -555
  569. package/src/vision/siglip.ts +0 -599
  570. package/src/vision/video-frames.ts +0 -148
  571. package/src/web/app.html +0 -2143
  572. package/src/web/app.js +0 -5703
  573. package/src/web/skills/memory/SKILL.md +0 -120
  574. package/src/web/skills/web-research/SKILL.md +0 -46
  575. package/src/web/skills.ts +0 -97
  576. package/src/web/src/adapters-panel.ts +0 -288
  577. package/src/web/src/assistant.ts +0 -451
  578. package/src/web/src/chat.ts +0 -1057
  579. package/src/web/src/composer.ts +0 -959
  580. package/src/web/src/dataset.ts +0 -132
  581. package/src/web/src/finetune.ts +0 -267
  582. package/src/web/src/hub.ts +0 -428
  583. package/src/web/src/main.ts +0 -98
  584. package/src/web/src/markdown.ts +0 -0
  585. package/src/web/src/memory-panel.ts +0 -572
  586. package/src/web/src/model-picker.ts +0 -159
  587. package/src/web/src/palette.ts +0 -346
  588. package/src/web/src/protocol.ts +0 -26
  589. package/src/web/src/quantize.ts +0 -167
  590. package/src/web/src/rag.ts +0 -274
  591. package/src/web/src/sessions.ts +0 -469
  592. package/src/web/src/shell.ts +0 -804
  593. package/src/web/src/ui-catalog.ts +0 -78
  594. package/src/web/vendor/README.md +0 -61
  595. package/src/web-assets.ts +0 -47
  596. package/src/web-tools.ts +0 -478
  597. package/src/weights.ts +0 -255
  598. /package/src/{serve → server}/completion-sink.ts +0 -0
  599. /package/src/web/{icon.svg → public/icon.svg} +0 -0
  600. /package/src/web/{manifest.webmanifest → public/manifest.webmanifest} +0 -0
  601. /package/src/web/{vendor → public/vendor}/hljs-LICENSE +0 -0
@@ -1,575 +0,0 @@
1
- <!doctype html>
2
- <html lang="en">
3
- <head>
4
- <meta charset="utf-8" />
5
- <meta name="viewport" content="width=device-width, initial-scale=1" />
6
- <title>mlx-bun · control-flow DAG (MiniCPM5-1B)</title>
7
- <script src="https://cdnjs.cloudflare.com/ajax/libs/cytoscape/3.28.1/cytoscape.min.js"></script>
8
- <script src="https://cdnjs.cloudflare.com/ajax/libs/dagre/0.8.5/dagre.min.js"></script>
9
- <script src="https://cdn.jsdelivr.net/npm/cytoscape-dagre@2.5.0/cytoscape-dagre.min.js"></script>
10
- <style>
11
- :root{--bg:#0b0f14;--panel:#11161d;--line:#243040;--txt:#cdd9e5;--dim:#7d8aa0;--cond:#39c5cf;}
12
- *{box-sizing:border-box;}
13
- html,body{margin:0;height:100%;background:var(--bg);color:var(--txt);font:13px/1.5 ui-monospace,SFMono-Regular,Menlo,monospace;}
14
- #top{display:flex;align-items:center;gap:12px;padding:8px 14px;flex-wrap:wrap;border-bottom:1px solid var(--line);background:var(--panel);}
15
- #top b{color:#fff;font-size:14px;}
16
- .seg{display:flex;gap:2px;background:#0b0f14;border:1px solid var(--line);border-radius:6px;}
17
- .seg button,.btn{background:transparent;color:var(--dim);border:0;padding:5px 11px;cursor:pointer;font:inherit;border-radius:5px;}
18
- .seg button.on{background:#388bfd;color:#fff;}
19
- .btn{border:1px solid var(--line);}
20
- .legend{margin-left:auto;display:flex;gap:11px;color:var(--dim);align-items:center;flex-wrap:wrap;font-size:11px;}
21
- .legend span{display:flex;align-items:center;gap:5px;}
22
- .dot{width:9px;height:9px;border-radius:50%;display:inline-block;}
23
- .dash{width:15px;border-top:2px dashed var(--cond);display:inline-block;}
24
- #wrap{display:flex;height:calc(100% - 46px);}
25
- #cy{flex:1;height:100%;}
26
- #side{width:340px;border-left:1px solid var(--line);background:var(--panel);padding:14px;overflow:auto;}
27
- #side h2{margin:0 0 4px;font-size:14px;color:#fff;}
28
- #side .file{color:#388bfd;font-size:12px;word-break:break-all;margin-bottom:8px;}
29
- #side .k{color:var(--dim);text-transform:uppercase;font-size:10.5px;letter-spacing:.06em;margin-top:10px;}
30
- #side .v{white-space:pre-wrap;}
31
- .hint{color:var(--dim);font-size:12px;margin-top:6px;}
32
- code{color:#e3b341;}
33
- </style>
34
- </head>
35
- <body>
36
- <div id="top">
37
- <b>mlx-bun control-flow DAG</b><span style="color:var(--dim);font-size:11px">MiniCPM5-1B</span>
38
- <div class="seg" id="graph">
39
- <button data-g="infer" class="on">inference</button>
40
- <button data-g="train">training</button>
41
- </div>
42
- <button class="btn" id="fit">⤢ fit</button>
43
- <div class="legend" id="legend"></div>
44
- </div>
45
- <div id="wrap">
46
- <div id="cy"></div>
47
- <div id="side">
48
- <h2 id="s-title">Click a step</h2>
49
- <div class="file" id="s-file"></div>
50
- <div class="hint" id="hint">Every node = one actual step the code runs. Solid arrows = the next step in execution order. <b style="color:var(--cond)">Dashed labelled arrows</b> = a branch: the code can go several ways, and the label is the exact condition (flag / variable / check) that takes that edge. Loops are back-edges. Scroll to zoom, drag to pan. Colour = subsystem (legend). Traced from the actual source — file:line on each node.</div>
51
- <div id="s-body" style="display:none">
52
- <div class="k">file</div><div class="v" id="s-fileb"></div>
53
- <div class="k">subsystem</div><div class="v" id="s-grp"></div>
54
- <div class="k">outgoing</div><div class="v" id="s-out"></div>
55
- </div>
56
- </div>
57
- </div>
58
- <script>
59
- // node = [id, label, group, file]. edge = [from, to, condition('' = unconditional)].
60
- // A node with several edges of different conditions IS a branch point.
61
- const GRP={ // subsystem → colour
62
- drive:'#388bfd', fwd:'#2ea043', layer:'#3fb950', attn:'#d29922', sdpa:'#a371f7',
63
- unfused:'#6e7b91', tiled:'#39c5cf', mlp:'#56d364', kv:'#f0883e', samp:'#db61a2',
64
- setup:'#6e7b91', loss:'#d29922', seg:'#a371f7', accum:'#388bfd', bwd:'#f85149',
65
- norm:'#39c5cf', opt:'#f0883e', metrics:'#db61a2' };
66
-
67
- const INFER={ n:[
68
- ['n1','generate() entry','drive','generate.ts:168'],
69
- ['n2','generateInner() entry','drive','generate.ts:208'],
70
- ['n3','? wired-scope: FORCE_WIRE ∨ weights>thresh','drive','generate.ts:175'],
71
- ['n4','enter wired scope','drive','generate.ts:200'],
72
- ['n5','? adapters?.length','drive','generate.ts:174'],
73
- ['n6','set loraState.active','drive','generate.ts:187'],
74
- ['n7','makeCache/Sampler/LogitsProcessors','drive','generate.ts:218'],
75
- ['n8','maybeQuantizeKv (prefill)','drive','generate.ts:281'],
76
- ['n9','? prompt-pos > prefillChunkSize','drive','generate.ts:298'],
77
- ['n10','chunked prefill: loop chunk','drive','generate.ts:299'],
78
- ['n11','forwardHidden(chunk, cache)','drive','generate.ts:301'],
79
- ['n12','evalAll(cache) after chunk','drive','generate.ts:304'],
80
- ['n13','clearCache() after chunk','drive','generate.ts:312'],
81
- ['n14','? options.promptEmbeddings','drive','generate.ts:284'],
82
- ['n15','forwardEmbeddings(...)','drive','generate.ts:292'],
83
- ['n16','forwardHidden(lastChunk, cache)','drive','generate.ts:320'],
84
- ['n17','? needsTokenHistory init','drive','generate.ts:315'],
85
- ['n18','slice last position h0[:,L-1]','drive','generate.ts:324'],
86
- ['n19','logitsFromHidden(hLast)','drive','generate.ts:326'],
87
- ['n20','sampleStep(logits0, 0)','drive','generate.ts:328'],
88
- ['n21','asyncEvalAll([pending])','drive','generate.ts:338'],
89
- ['n22','? CompiledDecode.for (env·adapters·gemma4)','drive','generate.ts:352'],
90
- ['n23','decode loop: while !stop','drive','generate.ts:360'],
91
- ['n24','? generated+1 < maxTokens','drive','generate.ts:364'],
92
- ['n25','maybeQuantizeKv (decode)','drive','generate.ts:365'],
93
- ['n26','pushHistory(cur)','drive','generate.ts:366'],
94
- ['n27','? compiled && supports(cache)','drive','generate.ts:369'],
95
- ['n28','compiled.step(cur, cache)','drive','generate.ts:371'],
96
- ['n29','catch → compiled=null (fallback)','drive','generate.ts:374'],
97
- ['n30','forwardHidden(ids, cache)','drive','generate.ts:383'],
98
- ['n31','logitsFromHidden(h)','drive','generate.ts:385'],
99
- ['n32','sampleStep(logits, n+1)','drive','generate.ts:388'],
100
- ['n33','asyncEvalAll([nextPending,…])','drive','generate.ts:390'],
101
- ['n34','itemUint32(cur) sync-read token','drive','generate.ts:394'],
102
- ['n35','? first token: prefillMs/tDecode','drive','generate.ts:399'],
103
- ['n36','cur.dispose; generated++','drive','generate.ts:402'],
104
- ['n37','? nextPending !== null','drive','generate.ts:406'],
105
- ['n38','? eosTokenIds.includes(token)','drive','generate.ts:408'],
106
- ['n39','yield {token, index}','drive','generate.ts:413'],
107
- ['n40','? clearCache every 256','drive','generate.ts:416'],
108
- ['n42','stop = true (no next)','drive','generate.ts:418'],
109
- ['n43','pending = nextPending','drive','generate.ts:420'],
110
- ['n44','stop = true (EOS / done)','drive','generate.ts:411'],
111
- ['n45','decodeMs','drive','generate.ts:425'],
112
- ['n46','return makeStats()','drive','generate.ts:427'],
113
- ['n47','finally: dispose pending/cache/history','drive','generate.ts:432'],
114
- ['n49','loraState.active=[] (scope exit)','drive','generate.ts:191'],
115
- ['n50','exitWiredScope()','drive','generate.ts:204'],
116
- // model forward
117
- ['n51','embed.encode(ids)','fwd','minicpm5.ts:250'],
118
- ['n52','forwardLayers(h, cache)','fwd','minicpm5.ts:258'],
119
- ['n53','runLayerRange(h,0,N,cache)','fwd','minicpm5.ts:259'],
120
- ['n54','? makeMask(L) (N==1? offset==0?)','fwd','minicpm5.ts:275'],
121
- ['n55','for i in [0,N): layers','fwd','minicpm5.ts:277'],
122
- ['n56','layers[i].forward(cur,mask,cache[i])','fwd','minicpm5.ts:278'],
123
- ['n57','? dispose(cur) if i>0','fwd','minicpm5.ts:279'],
124
- ['n58','mask.arr?.dispose()','fwd','minicpm5.ts:282'],
125
- ['n59','finalNorm.forward(h)','fwd','minicpm5.ts:261'],
126
- ['n60','lmHead.forward(h) → logits','fwd','minicpm5.ts:287'],
127
- // layer
128
- ['n61','inputNorm RMSNorm','layer','minicpm5.ts:189'],
129
- ['n62','attn.forward(xn,mask,cache)','layer','minicpm5.ts:190'],
130
- ['n63','add(x, attnOut) residual','layer','minicpm5.ts:192'],
131
- ['n64','postAttnNorm RMSNorm','layer','minicpm5.ts:194'],
132
- ['n65','mlp.forward(hn)','layer','minicpm5.ts:195'],
133
- ['n66','add(h, mlpOut) residual','layer','minicpm5.ts:197'],
134
- // attention
135
- ['n67','q_proj quantized_matmul','attn','minicpm5.ts:89'],
136
- ['n68','k_proj quantized_matmul','attn','minicpm5.ts:90'],
137
- ['n69','v_proj quantized_matmul','attn','minicpm5.ts:91'],
138
- ['n70','reshape q [B,L,H,D]','attn','minicpm5.ts:93'],
139
- ['n71','reshape k [B,L,Kv,D]','attn','minicpm5.ts:94'],
140
- ['n72','reshape v [B,L,Kv,D]','attn','minicpm5.ts:95'],
141
- ['n73','transpose q →[B,H,L,D]','attn','minicpm5.ts:97'],
142
- ['n74','transpose k','attn','minicpm5.ts:98'],
143
- ['n75','transpose v','attn','minicpm5.ts:99'],
144
- ['n76','? rope: prefixPlan / offsetArr / static','attn','minicpm5.ts:106'],
145
- ['n77','ropeBlocks(q) (prefix)','attn','minicpm5.ts:108'],
146
- ['n78','ropeDynamic(q) (offsetArr)','attn','minicpm5.ts:110'],
147
- ['n79','rope(q, cache.offset)','attn','minicpm5.ts:111'],
148
- ['n80','rope k (same offset)','attn','minicpm5.ts:113'],
149
- ['n81','? cache instanceof QuantizedKVCache','attn','minicpm5.ts:115'],
150
- ['n82','updateAndFetchQuantized(k,v)','attn','minicpm5.ts:116'],
151
- ['n83','quantizedSdpa(q,kq,vq,…)','attn','minicpm5.ts:119'],
152
- ['n84','updateAndFetch(k,v) bf16','attn','minicpm5.ts:123'],
153
- ['n85','? getTrainingAttn()=="flash"','attn','minicpm5.ts:128'],
154
- ['n86','flashAttention(...)','attn','minicpm5.ts:130'],
155
- ['n87','ops.sdpa(q,keys,values,…)','attn','minicpm5.ts:132'],
156
- ['n88','transpose attn →[B,L,H,D]','attn','minicpm5.ts:138'],
157
- ['n89','reshape merged [B,L,H*D]','attn','minicpm5.ts:140'],
158
- ['n90','o_proj quantized_matmul','attn','minicpm5.ts:142'],
159
- // sdpa dispatch
160
- ['n91','? mask=="" && scale==1.0','sdpa','gemma4-base.ts:1514'],
161
- ['n92','? perfKernelEnabled()','sdpa','fused-decode-kernel.ts:36'],
162
- ['n93','? !isCompiledTrace()','sdpa','gemma4-base.ts:1515'],
163
- ['n94','? fusedDecodeKernelSupported','sdpa','gemma4-base.ts:1515'],
164
- ['n95','⚙ fusedDecodeSdpa (L=1)','sdpa','gemma4-base.ts:1517'],
165
- ['n96','? tile: L>1 ∨ FUSED_DECODE','sdpa','gemma4-base.ts:1518'],
166
- ['n97','? fusedSdpaSupported','sdpa','gemma4-base.ts:1519'],
167
- ['n98','⚙ quantizedSdpaTiled','sdpa','gemma4-base.ts:1520'],
168
- ['n99','quantizedSdpaUnfused','sdpa','gemma4-base.ts:1521'],
169
- // unfused internals
170
- ['n100','dims [B,H,L,D,KV,nRep,N]','unfused','gemma4-base.ts:1221'],
171
- ['n101','? scale !== 1.0','unfused','gemma4-base.ts:1229'],
172
- ['n102','mulScalar(q, scale)','unfused','gemma4-base.ts:1229'],
173
- ['n103','? nRep > 1 (GQA)','unfused','gemma4-base.ts:1233'],
174
- ['n104','reshape q [B,KV,nRep,L,D]','unfused','gemma4-base.ts:1234'],
175
- ['n105','expandDims(k.packed,-3)','unfused','gemma4-base.ts:1245'],
176
- ['n106','quantizedMatmulQT(q,kᵀ) → scores','unfused','gemma4-base.ts:1251'],
177
- ['n107','? mask.mode=="causal"','unfused','gemma4-base.ts:1256'],
178
- ['n108','greaterEqual → causal mask','unfused','gemma4-base.ts:1261'],
179
- ['n109','? mask.mode=="array"','unfused','gemma4-base.ts:1264'],
180
- ['n110','where(mask, scores, -inf)','unfused','gemma4-base.ts:1269'],
181
- ['n111','softmaxAxis(scores,-1)','unfused','gemma4-base.ts:1276'],
182
- ['n112','quantizedMatmulQT(probs,vᵀ) → out','unfused','gemma4-base.ts:1278'],
183
- ['n113','? nRep > 1','unfused','gemma4-base.ts:1279'],
184
- ['n114','reshape out [B,H,L,D]','unfused','gemma4-base.ts:1280'],
185
- // tiled internals
186
- ['n115','dims [B,H,L,D,KV,nRep,N]','tiled','gemma4-base.ts:1304'],
187
- ['n116','? scale !== 1.0','tiled','gemma4-base.ts:1312'],
188
- ['n117','mulScalar(q, scale)','tiled','gemma4-base.ts:1312'],
189
- ['n118','? nRep > 1 (GQA)','tiled','gemma4-base.ts:1317'],
190
- ['n119','reshape q [B,KV,nRep,L,D]','tiled','gemma4-base.ts:1318'],
191
- ['n120','expandDims(k.packed,-3)','tiled','gemma4-base.ts:1324'],
192
- ['n121','? mask.mode=="causal"','tiled','gemma4-base.ts:1340'],
193
- ['n122','greaterEqual → causal mask','tiled','gemma4-base.ts:1345'],
194
- ['n123','? mask.mode=="array"','tiled','gemma4-base.ts:1348'],
195
- ['n124','for n0=0..N step 512 (KV chunks)','tiled','gemma4-base.ts:1378'],
196
- ['n125','slice kᵀ.packed chunk','tiled','gemma4-base.ts:1381'],
197
- ['n126','slice kᵀ.scales chunk','tiled','gemma4-base.ts:1382'],
198
- ['n127','slice kᵀ.biases chunk','tiled','gemma4-base.ts:1383'],
199
- ['n128','slice vᵀ.packed chunk','tiled','gemma4-base.ts:1386'],
200
- ['n129','quantizedMatmulQT(q,kChunk) → scores','tiled','gemma4-base.ts:1391'],
201
- ['n130','? maskChunk dtype bool','tiled','gemma4-base.ts:1397'],
202
- ['n131','where(mask, scores, -inf)','tiled','gemma4-base.ts:1399'],
203
- ['n132','add(scores, maskChunk) float','tiled','gemma4-base.ts:1402'],
204
- ['n133','maxAxis(scores) → chunkMax','tiled','gemma4-base.ts:1409'],
205
- ['n134','? oAcc===null (first tile)','tiled','gemma4-base.ts:1410'],
206
- ['n135','init rowMax/rowSum/oAcc','tiled','gemma4-base.ts:1411'],
207
- ['n136','maximum(rowMax,chunkMax)=newMax','tiled','gemma4-base.ts:1419'],
208
- ['n137','exp(rowMax-newMax)=factor','tiled','gemma4-base.ts:1422'],
209
- ['n138','exp(scores-newMax)=exps','tiled','gemma4-base.ts:1425'],
210
- ['n139','newSum = factor·rowSum + Σexps','tiled','gemma4-base.ts:1428'],
211
- ['n140','nextAcc = oAcc·factor + exps·vChunk','tiled','gemma4-base.ts:1436'],
212
- ['n141','rowMax/rowSum = new','tiled','gemma4-base.ts:1443'],
213
- ['n142','div(oAcc, rowSum) → out','tiled','gemma4-base.ts:1452'],
214
- ['n143','? nRep > 1','tiled','gemma4-base.ts:1459'],
215
- ['n144','reshape out [B,H,L,D]','tiled','gemma4-base.ts:1460'],
216
- // mlp
217
- ['n145','gate_proj quantized_matmul','mlp','minicpm5.ts:160'],
218
- ['n146','up_proj quantized_matmul','mlp','minicpm5.ts:161'],
219
- ['n147','sigmoid(gate)','mlp','minicpm5.ts:162'],
220
- ['n148','mul → SiLU (gate·σ(gate))','mlp','minicpm5.ts:163'],
221
- ['n149','mul(silu, up)','mlp','minicpm5.ts:166'],
222
- ['n150','down_proj quantized_matmul','mlp','minicpm5.ts:169'],
223
- // KV bf16
224
- ['n151','offset=prev, L=k.shape[2]','kv','gemma4-base.ts:252'],
225
- ['n152','? capacity: !keys ∨ prev+L>cap','kv','gemma4-base.ts:254'],
226
- ['n153','alloc newK/newV zeros','kv','gemma4-base.ts:257'],
227
- ['n154','? cache exists','kv','gemma4-base.ts:260'],
228
- ['n155','? prev%STEP!=0 (trim)','kv','gemma4-base.ts:263'],
229
- ['n156','slice old, dispose','kv','gemma4-base.ts:264'],
230
- ['n157','concat([old,new]) grow','kv','gemma4-base.ts:271'],
231
- ['n158','sliceUpdate keys←k','kv','gemma4-base.ts:283'],
232
- ['n159','sliceUpdate values←v','kv','gemma4-base.ts:284'],
233
- ['n160','slice fetch keys','kv','gemma4-base.ts:290'],
234
- ['n161','slice fetch values','kv','gemma4-base.ts:291'],
235
- // KV quantized
236
- ['n162','prev/L/elPerInt','kv','gemma4-base.ts:443'],
237
- ['n163','? capacity (packed)','kv','gemma4-base.ts:446'],
238
- ['n164','grow cache (6 comps)','kv','gemma4-base.ts:447'],
239
- ['n165','? prev%STEP!=0 (trim)','kv','gemma4-base.ts:448'],
240
- ['n166','trimTo (slice 6)','kv','gemma4-base.ts:449'],
241
- ['n167','quantize(k) → kq','kv','gemma4-base.ts:467'],
242
- ['n168','quantize(v) → vq','kv','gemma4-base.ts:468'],
243
- ['n169','write k packed/scales/biases','kv','gemma4-base.ts:479'],
244
- ['n175','fetch keys (3 comps)','kv','gemma4-base.ts:488'],
245
- ['n176','fetch values (3 comps)','kv','gemma4-base.ts:488'],
246
- // sampler
247
- ['n177','? curve sampler active','samp','sampler.ts:418'],
248
- ['n178','applyCurve(lp, curve)','samp','sampler.ts:420'],
249
- ['n179','randomCategorical (curve)','samp','sampler.ts:421'],
250
- ['n180','? temperature === 0','samp','sampler.ts:429'],
251
- ['n181','argmaxAxis(lp) greedy','samp','sampler.ts:430'],
252
- ['n182','? hlg?.enabled','samp','sampler.ts:437'],
253
- ['n183','applyHlgShaper','samp','sampler.ts:453'],
254
- ['n184','applyHlgPipeline','samp','sampler.ts:454'],
255
- ['n185','applyHlgEotf','samp','sampler.ts:454'],
256
- ['n186','applyHlgOetf','samp','sampler.ts:455'],
257
- ['n187','applyHlg (parametric)','samp','sampler.ts:455'],
258
- ['n188','randomCategorical (HLG)','samp','sampler.ts:457'],
259
- ['n189','? topP in (0,1)','samp','sampler.ts:467'],
260
- ['n190','applyTopP(lp, topP)','samp','sampler.ts:467'],
261
- ['n191','? topK > 0','samp','sampler.ts:468'],
262
- ['n192','applyTopK(lp, topK)','samp','sampler.ts:468'],
263
- ['n193','mulScalar(lp, 1/temp)','samp','sampler.ts:469'],
264
- ['n194','randomCategorical (temp)','samp','sampler.ts:471'],
265
- ],e:[
266
- ['n1','n2',''],['n2','n3',''],
267
- ['n3','n4','FORCE_WIRE ∨ weights>thresh'],['n3','n5','else'],
268
- ['n4','n5',''],
269
- ['n5','n6','adapters?.length'],['n5','n7','else'],
270
- ['n6','n7',''],['n7','n8',''],['n8','n9',''],
271
- ['n9','n10','pos > prefillChunkSize'],['n9','n14','else (one-shot)'],
272
- ['n10','n11',''],['n11','n51','call forward'],['n11','n12',''],['n12','n13',''],
273
- ['n13','n10','more chunks'],['n13','n14','done'],
274
- ['n14','n15','promptEmbeddings'],['n14','n16','else'],
275
- ['n15','n17',''],['n16','n51','call forward'],['n16','n17',''],
276
- ['n17','n18',''],['n18','n19',''],['n19','n60','call lmHead'],['n19','n20',''],['n20','n21',''],['n21','n22',''],['n22','n23',''],
277
- ['n23','n24',''],
278
- ['n24','n25','generated+1<maxTokens'],['n24','n34','else'],
279
- ['n25','n26',''],['n26','n27',''],
280
- ['n27','n28','compiled && supports'],['n27','n30','else (uncompiled)'],
281
- ['n28','n29','throws'],['n28','n31','ok'],['n29','n30',''],
282
- ['n30','n51','call forward'],['n30','n31',''],['n31','n60','call lmHead'],['n31','n32',''],['n32','n33',''],['n33','n34',''],
283
- ['n34','n35','first token'],['n34','n36','else'],['n35','n36',''],['n36','n37',''],
284
- ['n37','n38','nextPending!=null'],['n37','n42','else'],
285
- ['n38','n39','!EOS'],['n38','n44','EOS'],
286
- ['n39','n40',''],['n40','n43','nextPending'],['n40','n42','none'],
287
- ['n42','n44',''],['n43','n23','loop'],
288
- ['n44','n45',''],['n45','n46',''],['n46','n47',''],['n47','n49',''],['n49','n50',''],
289
- // forward
290
- ['n51','n52',''],['n52','n53',''],['n53','n54',''],['n54','n55',''],
291
- ['n55','n56','i<N'],['n56','n57',''],['n57','n55','i++'],['n57','n58','done'],
292
- ['n58','n59',''],['n59','n12','return (prefill)'],['n59','n17','return (prefill)'],['n59','n31','return (decode)'],
293
- // layer (called by n56)
294
- ['n56','n61','call layer'],
295
- ['n61','n62',''],['n62','n63',''],['n63','n64',''],['n64','n65',''],['n65','n66',''],
296
- // attention (called by n62)
297
- ['n62','n67','call attn'],
298
- ['n67','n68',''],['n68','n69',''],['n69','n70',''],['n70','n71',''],['n71','n72',''],['n72','n73',''],['n73','n74',''],['n74','n75',''],['n75','n76',''],
299
- ['n76','n77','_prefixPlan'],['n76','n78','offsetArr'],['n76','n79','else (static)'],
300
- ['n77','n80',''],['n78','n80',''],['n79','n80',''],['n80','n81',''],
301
- ['n81','n82','QuantizedKVCache'],['n81','n84','else (bf16)'],
302
- ['n82','n83',''],['n83','n88',''],
303
- ['n84','n85',''],
304
- ['n85','n86','flash & supported'],['n85','n87','else'],
305
- ['n86','n88',''],['n87','n88',''],['n88','n89',''],['n89','n90',''],
306
- // sdpa dispatch (called by n83)
307
- ['n83','n91','call quantizedSdpa'],
308
- ['n91','n92','mask=="" & scale==1'],['n91','n96','else'],
309
- ['n92','n93','perfKernel'],['n92','n96','else'],
310
- ['n93','n94','!compiledTrace'],['n93','n96','else'],
311
- ['n94','n95','supported'],['n94','n96','else'],
312
- ['n96','n97','tile (L>1 ∨ FUSED_DECODE)'],['n96','n99','else'],
313
- ['n97','n98','fusedSdpaSupported'],['n97','n99','else'],
314
- // unfused (n99)
315
- ['n99','n100','call'],['n100','n101',''],
316
- ['n101','n102','scale!=1'],['n101','n103','else'],['n102','n103',''],
317
- ['n103','n104','nRep>1'],['n103','n106','else'],['n104','n105',''],['n105','n106',''],
318
- ['n106','n107',''],['n107','n108','causal'],['n107','n109','else'],
319
- ['n108','n110',''],['n109','n110','array'],['n109','n111','none'],['n110','n111',''],
320
- ['n111','n112',''],['n112','n113',''],['n113','n114','nRep>1'],
321
- // tiled (n98)
322
- ['n98','n115','call'],['n115','n116',''],
323
- ['n116','n117','scale!=1'],['n116','n118','else'],['n117','n118',''],
324
- ['n118','n119','nRep>1'],['n118','n121','else'],['n119','n120',''],['n120','n121',''],
325
- ['n121','n122','causal'],['n121','n123','else'],['n122','n124',''],['n123','n124',''],
326
- ['n124','n125','n0<N'],['n125','n126',''],['n126','n127',''],['n127','n128',''],['n128','n129',''],
327
- ['n129','n130',''],['n130','n131','bool'],['n130','n132','float'],['n131','n133',''],['n132','n133',''],
328
- ['n133','n134',''],['n134','n135','first'],['n134','n136','else'],
329
- ['n135','n124','next chunk'],
330
- ['n136','n137',''],['n137','n138',''],['n138','n139',''],['n139','n140',''],['n140','n141',''],['n141','n124','next chunk'],
331
- ['n124','n142','n0>=N'],['n142','n143',''],['n143','n144','nRep>1'],
332
- // mlp (called by n65)
333
- ['n65','n145','call mlp'],
334
- ['n145','n146',''],['n146','n147',''],['n147','n148',''],['n148','n149',''],['n149','n150',''],
335
- // KV bf16 (called by n84)
336
- ['n84','n151','call updateAndFetch'],
337
- ['n151','n152',''],['n152','n153','grow'],['n152','n158','fits'],
338
- ['n153','n154',''],['n154','n155','exists'],['n154','n157','first'],
339
- ['n155','n156','prev%STEP!=0'],['n155','n157','aligned'],['n156','n157',''],['n157','n158',''],
340
- ['n158','n159',''],['n159','n160',''],['n160','n161',''],
341
- // KV quant (called by n82)
342
- ['n82','n162','call updateAndFetchQuantized'],
343
- ['n162','n163',''],['n163','n164','grow'],['n163','n167','fits'],
344
- ['n164','n165',''],['n165','n166','prev%STEP!=0'],['n165','n167','aligned'],['n166','n167',''],
345
- ['n167','n168',''],['n168','n169',''],['n169','n175',''],['n175','n176',''],
346
- // sampler (called by n20, n32)
347
- ['n20','n177','call sampleStep'],['n32','n177','call sampleStep'],
348
- ['n177','n178','curve'],['n177','n180','else'],['n178','n179',''],
349
- ['n180','n181','temp==0'],['n180','n182','else'],
350
- ['n182','n183','shaper'],['n182','n184','pipeline'],['n182','n185','eotf'],['n182','n186','oetf'],['n182','n187','parametric'],['n182','n189','hlg off'],
351
- ['n183','n188',''],['n184','n188',''],['n185','n188',''],['n186','n188',''],['n187','n188',''],
352
- ['n189','n190','topP'],['n189','n191','else'],['n190','n191',''],
353
- ['n191','n192','topK'],['n191','n193','else'],['n192','n193',''],['n193','n194',''],
354
- ]};
355
-
356
- const TRAIN={ n:[
357
- ['t1','step loop begins','setup','trainer.ts:479'],
358
- ['t2','per-step dropout seed','setup','trainer.ts:876'],
359
- ['t3','enter accumulation loop','accum','trainer.ts:487'],
360
- ['t4','fetch micro-batch','accum','trainer.ts:490'],
361
- ['t5','? segmentSize > 0','setup','trainer.ts:409'],
362
- ['t6','segmented forward start','seg','segmented.ts:196'],
363
- ['t7','embed.encode(ids)','seg','segmented.ts:196'],
364
- ['t8','detach embedding boundary','seg','segmented.ts:196'],
365
- ['t9','for segment k=0..nSeg','seg','segmented.ts:197'],
366
- ['t10','runLayerRange[lo:hi] (swap LoRA)','seg','segmented.ts:199'],
367
- ['t11','detach output boundary k','seg','segmented.ts:199'],
368
- ['t12','? ORPO prefix-shared & segmented','seg','trainer.ts:880'],
369
- ['t13','splitPrefixBatch','seg','trainer.ts:881'],
370
- ['t14','single forward [P;C;R] block-mask','seg','prefix-shared.ts:159'],
371
- ['t15','prefix fallback (two-forward)','seg','trainer.ts:884'],
372
- ['t16','non-segmented forward start','accum','trainer.ts:492'],
373
- ['t17','value_and_grad: swap primals','accum','trainer.ts:432'],
374
- ['t18','? B==1 response-only vs full','loss','loss.ts:123'],
375
- ['t19','forwardHidden → [B,T,hid]','fwd','forward.ts:72'],
376
- ['t20','embedding forward','fwd','forward.ts:92'],
377
- ['t21','layer-stack forward','fwd','forward.ts:92'],
378
- ['t22','? MLX_BUN_TRAIN_ATTN=="flash"','fwd','trainer.ts:469'],
379
- ['t23','sdpa attention (O(L²) bwd)','fwd','loss.ts:131'],
380
- ['t24','flash attention (O(L) bwd)','fwd','trainer.ts:469'],
381
- ['t25','final norm','fwd','forward.ts:113'],
382
- ['t26','lm-head → logits','fwd','loss.ts:131'],
383
- ['t27','? B==1 responseCe vs B>1 maskedCe','loss','loss.ts:123'],
384
- ['t28','slice hidden to response span','loss','loss.ts:160'],
385
- ['t29','head at response positions','loss','loss.ts:161'],
386
- ['t30','gather target logits','loss','loss.ts:170'],
387
- ['t31','logsumexp vocab','loss','loss.ts:169'],
388
- ['t32','CE = lse - target','loss','loss.ts:172'],
389
- ['t33','sum/normalize response CE','loss','loss.ts:174'],
390
- ['t34','build response mask [B,T]','loss','loss.ts:233'],
391
- ['t35','reshape logits [B*T,V]','loss','loss.ts:251'],
392
- ['t36','gather targets','loss','loss.ts:252'],
393
- ['t37','logsumexp','loss','loss.ts:256'],
394
- ['t38','CE * mask','loss','loss.ts:262'],
395
- ['t39','sum masked CE / norm','loss','loss.ts:264'],
396
- ['t40','DPO ref logps (scale=0, detach)','loss','loss.ts:361'],
397
- ['t41','DPO policy logps','loss','loss.ts:389'],
398
- ['t42','logr_c = piC - refC','loss','loss.ts:392'],
399
- ['t43','logr_r = piR - refR','loss','loss.ts:393'],
400
- ['t44','adv = β(logr_c - logr_r)','loss','loss.ts:394'],
401
- ['t45','-logσ(adv) = softplus; mean','loss','loss.ts:399'],
402
- ['t46','ORPO chosen logp','loss','loss.ts:1242'],
403
- ['t47','? orpoChunkSize > 0','loss','loss.ts:481'],
404
- ['t48','? orpoFusedCe ∨ orpoFlashCe','loss','loss.ts:482'],
405
- ['t49','chunked head (Checkpoint)','loss','loss.ts:567'],
406
- ['t50','fused linear-CE head (CustomVjp)','loss','loss.ts:726'],
407
- ['t51','vocab-blocked head','loss','loss.ts:822'],
408
- ['t52','flash-CCE Metal head','loss','loss.ts:1009'],
409
- ['t53','whole-vocab head (oracle)','loss','loss.ts:1087'],
410
- ['t54','ORPO rejected logp','loss','loss.ts:1243'],
411
- ['t55','NLL = mean(-ℓw)','loss','loss.ts:1225'],
412
- ['t56','log_odds calc','loss','loss.ts:1207'],
413
- ['t57','OR = mean softplus(-log_odds)','loss','loss.ts:1228'],
414
- ['t58','L = NLL + λ·OR','loss','loss.ts:1231'],
415
- ['t59','seg loss-head VJP (cot=1)','seg','segmented.ts:141'],
416
- ['t60','seg loss scalar (detached)','seg','segmented.ts:204'],
417
- ['t61','seg dh_out = dLoss/dhidden','seg','segmented.ts:205'],
418
- ['t62','evalAll([loss, …grads])','accum','trainer.ts:1027'],
419
- ['t63','loss scalar toFloat32','accum','trainer.ts:1028'],
420
- ['t64','scale grad by 1/accum','accum','trainer.ts:1042'],
421
- ['t65','accumulate grad','accum','trainer.ts:1048'],
422
- ['t66','dispose checkpoint closures','accum','trainer.ts:496'],
423
- ['t67','sum micro losses','accum','trainer.ts:1061'],
424
- ['t69','value_and_grad backward','bwd','trainer.ts:492'],
425
- ['t70','MLX reverse-mode AD','bwd','trainer.ts:492'],
426
- ['t71','restore LoRA leaves','bwd','trainer.ts:436'],
427
- ['t72','seg backward: k=nSeg-1..0','bwd','segmented.ts:210'],
428
- ['t73','vjp segment k (cot=dh_out)','bwd','segmented.ts:215'],
429
- ['t74','scatter per-seg LoRA grads','bwd','segmented.ts:220'],
430
- ['t75','dh_in → cotangent for k-1','bwd','segmented.ts:225'],
431
- ['t76','dispose boundary k','bwd','segmented.ts:226'],
432
- ['t80','globalNorm(grads)','norm','trainer.ts:1076'],
433
- ['t81','square each grad','norm','trainer.ts:1078'],
434
- ['t82','sum squares, sqrt','norm','trainer.ts:1086'],
435
- ['t83','? ‖g‖ > maxNorm','norm','trainer.ts:1099'],
436
- ['t84','scale g *= max/‖g‖','norm','trainer.ts:1103'],
437
- ['t85','AdamW.step(grads)','opt','trainer.ts:504'],
438
- ['t86','t++ step counter','opt','optimizer.ts:84'],
439
- ['t87','bias-correct β1^t','opt','optimizer.ts:85'],
440
- ['t88','bias-correct β2^t','opt','optimizer.ts:86'],
441
- ['t89','for each param i (A,B)','opt','optimizer.ts:88'],
442
- ['t90','grad → float32','opt','optimizer.ts:92'],
443
- ['t91','m ← β1·m + (1-β1)g','opt','optimizer.ts:95'],
444
- ['t92','v ← β2·v + (1-β2)g²','opt','optimizer.ts:105'],
445
- ['t93','? weight decay','opt','optimizer.ts:119'],
446
- ['t94','m̂ = m/(1-β1^t)','opt','optimizer.ts:125'],
447
- ['t95','v̂ = v/(1-β2^t)','opt','optimizer.ts:126'],
448
- ['t96','√v̂ + ε','opt','optimizer.ts:127'],
449
- ['t97','m̂/(√v̂+ε)','opt','optimizer.ts:130'],
450
- ['t98','scale by lr (LoRA+ per-param)','opt','optimizer.ts:131'],
451
- ['t99','p = p - lr·update','opt','optimizer.ts:132'],
452
- ['t100','writeParam → swap into LoRA','opt','trainer.ts:402'],
453
- ['t101','opt.evalState()','opt','trainer.ts:505'],
454
- ['t102','clearCache()','metrics','trainer.ts:511'],
455
- ['t103','emit step metrics','metrics','trainer.ts:516'],
456
- ['t104','? step%report==0 ∨ step==1','metrics','trainer.ts:514'],
457
- ['t105','recompute dpo/orpo metrics','metrics','trainer.ts:900'],
458
- ['t106','? step%eval==0 & valid>0','metrics','trainer.ts:529'],
459
- ['t107','disable grad-ckpt for eval','metrics','trainer.ts:532'],
460
- ['t108','loop validation examples','metrics','trainer.ts:554'],
461
- ['t109','eval loss per example','metrics','trainer.ts:564'],
462
- ['t110','reduce eval (mean/acc/margin)','metrics','trainer.ts:684'],
463
- ['t111','re-enable grad-ckpt','metrics','trainer.ts:534'],
464
- ['t112','step iteration completes','setup','trainer.ts:538'],
465
- ],e:[
466
- ['t1','t2',''],['t2','t3',''],['t3','t4',''],['t4','t5',''],
467
- ['t5','t6','segmentSize>0'],['t5','t16','else'],
468
- ['t6','t7',''],['t7','t8',''],['t8','t9',''],
469
- ['t9','t10','k<nSeg'],['t10','t11',''],['t11','t9','k++'],
470
- ['t11','t12','done (ORPO)'],['t11','t59','done (SFT seg)'],
471
- ['t12','t13','prefixShared & segmented'],['t12','t59','else'],
472
- ['t13','t14','split non-null'],['t13','t15','null (mismatch)'],
473
- ['t14','t59',''],['t15','t72',''],
474
- ['t16','t17',''],['t17','t18',''],
475
- ['t18','t19','B==1'],['t18','t20','full-logits'],
476
- ['t19','t20',''],['t20','t21',''],['t21','t22',''],
477
- ['t22','t24','flash'],['t22','t23','else'],
478
- ['t23','t25',''],['t24','t25',''],['t25','t26',''],
479
- ['t19','t27','SFT B==1'],['t26','t27','SFT B>1'],
480
- ['t27','t28','B==1'],['t27','t34','B>1'],
481
- ['t28','t29',''],['t29','t30',''],['t30','t31',''],['t31','t32',''],['t32','t33',''],
482
- ['t34','t35',''],['t35','t36',''],['t36','t37',''],['t37','t38',''],['t38','t39',''],
483
- // DPO
484
- ['t4','t40','method==dpo'],['t40','t41',''],['t41','t42',''],['t42','t43',''],['t43','t44',''],['t44','t45',''],
485
- // ORPO
486
- ['t4','t12','method==orpo & prefixShared'],['t12','t46','else'],['t13','t46','null → two-forward'],
487
- ['t46','t47',''],
488
- ['t47','t48','chunkSize>0'],['t47','t54','batched'],
489
- ['t48','t49','neither'],['t48','t50','fused'],['t48','t52','flash'],
490
- ['t50','t51','vocabBlock>0'],['t50','t53','else'],
491
- ['t49','t54',''],['t51','t54',''],['t52','t54',''],['t53','t54',''],
492
- ['t14','t55','prefix-shared done'],['t54','t55',''],
493
- ['t55','t56',''],['t56','t57',''],['t57','t58',''],
494
- // seg head
495
- ['t59','t60',''],['t60','t61',''],
496
- // accumulation
497
- ['t33','t62','SFT B==1'],['t39','t62','SFT B>1'],['t45','t62','DPO'],['t58','t62','ORPO'],['t60','t62','seg'],
498
- ['t62','t63',''],
499
- ['t63','t64','accum>1'],['t63','t66','else'],
500
- ['t64','t65',''],['t65','t66',''],
501
- ['t66','t64','next micro'],['t66','t67','last micro'],
502
- // backward
503
- ['t62','t69','non-seg'],['t69','t70',''],['t70','t71',''],
504
- ['t61','t72','seg'],['t72','t73','k>=0'],['t73','t74',''],['t74','t75',''],['t75','t76',''],['t76','t72','k--'],
505
- // norm/clip
506
- ['t71','t80',''],['t72','t80','seg done'],['t67','t80','accum done'],
507
- ['t80','t81',''],['t81','t82',''],['t82','t83',''],
508
- ['t83','t84','‖g‖>max'],['t83','t85','else'],['t84','t85',''],
509
- // optimizer
510
- ['t85','t86',''],['t86','t87',''],['t87','t88',''],['t88','t89',''],
511
- ['t89','t90','i<len'],['t90','t91',''],['t91','t92',''],['t92','t93',''],
512
- ['t93','t94','wd>0 / else'],['t94','t95',''],['t95','t96',''],['t96','t97',''],['t97','t98',''],['t98','t99',''],['t99','t100',''],['t100','t89','i++'],
513
- ['t89','t101','done'],
514
- // metrics
515
- ['t101','t102',''],['t102','t103',''],['t103','t104',''],
516
- ['t104','t105','report'],['t104','t106','skip'],['t105','t106',''],
517
- ['t106','t107','eval'],['t106','t112','skip'],
518
- ['t107','t108',''],['t108','t109','ex'],['t109','t108','next'],['t108','t110','done'],['t110','t111',''],['t111','t112',''],
519
- ['t112','t1','step++'],
520
- ]};
521
-
522
- let MODE_G='infer';
523
- const cy=cytoscape({container:document.getElementById('cy'),minZoom:0.06,maxZoom:3,wheelSensitivity:0.25,
524
- autoungrabify:true, boxSelectionEnabled:false, selectionType:'single',
525
- style:[
526
- {selector:'node',style:{'background-color':d=>GRP[d.data('grp')]||'#6e7b91','label':'data(label)',
527
- 'color':'#0b0f14','font-family':'ui-monospace, Menlo, monospace','font-size':'11px','font-weight':'600',
528
- 'text-valign':'center','text-halign':'center','shape':'round-rectangle','width':'label','height':22,
529
- 'padding':'7px','text-wrap':'wrap','text-max-width':'210px','border-width':1,'border-color':'#0b0f14'}},
530
- {selector:'node[branch]',style:{'border-width':2.5,'border-color':'#39c5cf'}},
531
- {selector:'node:selected',style:{'border-width':3,'border-color':'#fff'}},
532
- {selector:'edge',style:{'width':1.6,'line-color':'#3a4760','target-arrow-color':'#3a4760',
533
- 'target-arrow-shape':'triangle','curve-style':'bezier','arrow-scale':0.8}},
534
- {selector:'edge[cond]',style:{'line-color':'#39c5cf','target-arrow-color':'#39c5cf','line-style':'dashed',
535
- 'label':'data(cond)','font-size':'8px','font-family':'ui-monospace, Menlo, monospace','color':'#39c5cf',
536
- 'text-background-color':'#0b0f14','text-background-opacity':0.85,'text-background-padding':'1px','text-rotation':'autorotate'}},
537
- ]});
538
-
539
- function build(g){
540
- // count outgoing per node to mark branch points (>1 distinct outgoing edge)
541
- const out={}; g.e.forEach(([a])=>out[a]=(out[a]||0)+1);
542
- const nodes=g.n.map(([id,label,grp,file])=>({data:{id,label,grp,file,branch:out[id]>1?1:undefined}}));
543
- const edges=g.e.map(([a,b,c])=>({data:{id:a+'>'+b+(c||''),source:a,target:b,...(c?{cond:c}:{})}}));
544
- return {nodes,edges};
545
- }
546
- function load(g){const el=build(g);cy.elements().remove();cy.add(el.nodes);cy.add(el.edges);
547
- cy.layout({name:'dagre',rankDir:'TB',nodeSep:18,rankSep:30,padding:30}).run();}
548
-
549
- function show(d){
550
- document.getElementById('s-title').textContent=d.label;
551
- document.getElementById('s-file').textContent=d.file||'';
552
- document.getElementById('hint').style.display='none';
553
- document.getElementById('s-body').style.display='block';
554
- document.getElementById('s-fileb').textContent=d.file||'';
555
- document.getElementById('s-grp').textContent=d.grp||'';
556
- const outs=cy.getElementById(d.id).outgoers('edge').map(e=>{
557
- const t=e.target().data('label'); const c=e.data('cond'); return c?`• [${c}] → ${t}`:`• → ${t}`;}).join('\n');
558
- document.getElementById('s-out').textContent=outs||'(terminal)';
559
- }
560
- cy.on('tap','node',e=>show(e.target.data()));
561
- document.getElementById('fit').onclick=()=>cy.fit(undefined,40);
562
- const LABELS={drive:'generate',fwd:'forward',layer:'layer',attn:'attention',sdpa:'sdpa dispatch',unfused:'unfused sdpa',tiled:'tiled sdpa',mlp:'mlp',kv:'kv cache',samp:'sampler',
563
- setup:'setup/loop',loss:'loss head',seg:'segmented',accum:'accum',bwd:'backward',norm:'grad norm/clip',opt:'optimizer',metrics:'metrics/eval'};
564
- function legend(){
565
- const groups = MODE_G==='infer' ? ['drive','fwd','attn','sdpa','tiled','unfused','mlp','kv','samp'] : ['setup','fwd','loss','seg','accum','bwd','norm','opt','metrics'];
566
- document.getElementById('legend').innerHTML='<span><i class="dash"></i>branch (condition)</span>'+
567
- groups.map(g=>`<span><i class="dot" style="background:${GRP[g]}"></i>${LABELS[g]}</span>`).join('');
568
- }
569
- document.querySelectorAll('#graph button').forEach(b=>b.onclick=()=>{
570
- document.querySelectorAll('#graph button').forEach(x=>x.classList.remove('on'));b.classList.add('on');
571
- MODE_G=b.dataset.g;load(MODE_G==='infer'?INFER:TRAIN);legend();});
572
- legend();load(INFER);
573
- </script>
574
- </body>
575
- </html>