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
@@ -1,18 +0,0 @@
1
- // Registry of generated per-architecture specializations (hand-kept;
2
- // the files themselves come from scripts/gen-model.ts). Dispatch is by
3
- // config fingerprint; anything unregistered runs the monolith.
4
-
5
- import type { Weights } from "../../weights";
6
- import type { ModelConfig } from "../../config";
7
- import { Gemma4Model } from "../gemma4";
8
- import * as g12b from "./gemma4-12b";
9
- import * as ge4b from "./gemma4-e4b";
10
- import * as g26b from "./gemma4-26b";
11
-
12
- type ModelCtor = new (weights: Weights, config: ModelConfig) => Gemma4Model;
13
-
14
- export const GENERATED = new Map<string, ModelCtor>([
15
- [g12b.FINGERPRINT, g12b.GeneratedGemma4],
16
- [ge4b.FINGERPRINT, ge4b.GeneratedGemma4],
17
- [g26b.FINGERPRINT, g26b.GeneratedGemma4],
18
- ]);
@@ -1,511 +0,0 @@
1
- // Offline GLM-5.2 Expert Atlas analysis.
2
- //
3
- // Each prompt is one independent observation. Selection counts are normalized
4
- // within a run, averaged within a category, and only promoted to an affinity
5
- // label when the top topic fires in multiple independent prompts. Nothing in
6
- // this module feeds residency, routing, or prefetch policy.
7
-
8
- import type { Glm52RouteTraceRecord } from "./glm52-coupling";
9
-
10
- export interface Glm52AtlasProbeSet {
11
- readonly schemaVersion: 1;
12
- readonly provenance: {
13
- readonly source: string;
14
- readonly sourceCommit: string;
15
- readonly method: string;
16
- };
17
- readonly categories: Readonly<Record<string, readonly string[]>>;
18
- }
19
-
20
- export interface Glm52AtlasRun {
21
- readonly category: string;
22
- readonly promptIndex: number;
23
- readonly totalSelections: number;
24
- readonly counts: ReadonlyMap<string, number>;
25
- }
26
-
27
- export interface Glm52AtlasExpert {
28
- readonly layer: number;
29
- readonly expert: number;
30
- readonly total: number;
31
- readonly specialization: number;
32
- readonly topTopic: string;
33
- readonly topLift: number;
34
- readonly reliability: {
35
- readonly firedRuns: number;
36
- readonly totalRuns: number;
37
- };
38
- readonly affinity: Readonly<Record<string, number>>;
39
- }
40
-
41
- export interface Glm52AtlasAnalysis {
42
- readonly schemaVersion: 1;
43
- readonly categories: readonly string[];
44
- readonly runCount: number;
45
- readonly method: {
46
- readonly minCount: number;
47
- readonly minRuns: number;
48
- readonly strongThreshold: number;
49
- readonly normalization: "mean per-run selection share";
50
- };
51
- readonly summary: {
52
- readonly expertsSeen: number;
53
- readonly expertsKept: number;
54
- readonly droppedSparse: number;
55
- readonly droppedUnreplicated: number;
56
- readonly strongSpecialists: number;
57
- readonly strongSpecialistRate: number;
58
- };
59
- readonly experts: readonly Glm52AtlasExpert[];
60
- }
61
-
62
- export interface Glm52AtlasValidationTrial {
63
- readonly category: string;
64
- readonly promptIndex: number;
65
- readonly predicted: string;
66
- readonly correct: boolean;
67
- readonly ownShare: number;
68
- readonly bestOtherShare: number;
69
- readonly scores: Readonly<Record<string, number>>;
70
- }
71
-
72
- export interface Glm52AtlasValidation {
73
- readonly topK: number;
74
- readonly hits: number;
75
- readonly trials: number;
76
- readonly accuracy: number;
77
- readonly chance: number;
78
- readonly protocol: "global leave-one-prompt-out";
79
- readonly trialsByPrompt: readonly Glm52AtlasValidationTrial[];
80
- }
81
-
82
- export interface Glm52AtlasBuildOptions {
83
- readonly minCount?: number;
84
- readonly minRuns?: number;
85
- readonly strongThreshold?: number;
86
- }
87
-
88
- const CATEGORY = /^[a-z0-9][a-z0-9_]*$/;
89
-
90
- function finiteUnit(value: number, label: string): void {
91
- if (!Number.isFinite(value) || value < 0 || value > 1)
92
- throw new RangeError(`${label} must be in [0, 1]`);
93
- }
94
-
95
- function positiveInteger(value: number, label: string): void {
96
- if (!Number.isSafeInteger(value) || value <= 0)
97
- throw new RangeError(`${label} must be a positive safe integer`);
98
- }
99
-
100
- function round(value: number, digits = 6): number {
101
- const scale = 10 ** digits;
102
- return Math.round(value * scale) / scale;
103
- }
104
-
105
- function expertKey(layer: number, expert: number): string {
106
- return `${layer}:${expert}`;
107
- }
108
-
109
- function parseExpertKey(key: string): [number, number] {
110
- const fields = key.split(":");
111
- const layer = Number(fields[0]);
112
- const expert = Number(fields[1]);
113
- if (fields.length !== 2 || !Number.isSafeInteger(layer) || layer < 0 ||
114
- !Number.isSafeInteger(expert) || expert < 0) {
115
- throw new Error(`invalid GLM Atlas expert key ${key}`);
116
- }
117
- return [layer, expert];
118
- }
119
-
120
- export function glm52AtlasRunId(category: string, promptIndex: number): string {
121
- if (!CATEGORY.test(category))
122
- throw new Error(`invalid GLM Atlas category ${category}`);
123
- if (!Number.isSafeInteger(promptIndex) || promptIndex < 0)
124
- throw new RangeError(`invalid GLM Atlas prompt index ${promptIndex}`);
125
- return `${category}:${promptIndex}`;
126
- }
127
-
128
- export function validateGlm52AtlasProbeSet(
129
- value: unknown,
130
- ): Glm52AtlasProbeSet {
131
- if (!value || typeof value !== "object")
132
- throw new Error("GLM Atlas probe set must be an object");
133
- const input = value as Record<string, unknown>;
134
- if (input.schemaVersion !== 1)
135
- throw new Error("unsupported GLM Atlas probe schema");
136
- const provenance = input.provenance as Record<string, unknown> | undefined;
137
- if (!provenance || ["source", "sourceCommit", "method"].some(
138
- (key) => typeof provenance[key] !== "string" ||
139
- !(provenance[key] as string).trim(),
140
- )) {
141
- throw new Error("GLM Atlas probe provenance is incomplete");
142
- }
143
- const raw = input.categories;
144
- if (!raw || typeof raw !== "object" || Array.isArray(raw))
145
- throw new Error("GLM Atlas categories must be an object");
146
- const entries = Object.entries(raw as Record<string, unknown>);
147
- if (entries.length < 2)
148
- throw new Error("GLM Atlas requires at least two categories");
149
- const categories: Record<string, readonly string[]> = {};
150
- for (const [category, prompts] of entries) {
151
- if (!CATEGORY.test(category))
152
- throw new Error(`invalid GLM Atlas category ${category}`);
153
- if (!Array.isArray(prompts) || prompts.length < 3 ||
154
- prompts.some((prompt) => typeof prompt !== "string" || !prompt.trim())) {
155
- throw new Error(
156
- `GLM Atlas category ${category} requires at least three prompts`,
157
- );
158
- }
159
- categories[category] = Object.freeze(prompts.slice());
160
- }
161
- return Object.freeze({
162
- schemaVersion: 1,
163
- provenance: Object.freeze({
164
- source: provenance.source as string,
165
- sourceCommit: provenance.sourceCommit as string,
166
- method: provenance.method as string,
167
- }),
168
- categories: Object.freeze(categories),
169
- });
170
- }
171
-
172
- export function buildGlm52AtlasRun(
173
- category: string,
174
- promptIndex: number,
175
- records: readonly Glm52RouteTraceRecord[],
176
- ): Glm52AtlasRun {
177
- const segment = glm52AtlasRunId(category, promptIndex);
178
- const counts = new Map<string, number>();
179
- let totalSelections = 0;
180
- if (records.length === 0)
181
- throw new Error(`GLM Atlas run ${segment} has no route records`);
182
- for (const record of records) {
183
- if (record.segment !== segment)
184
- throw new Error(
185
- `GLM Atlas run ${segment} contains segment ${record.segment}`,
186
- );
187
- if (!Number.isSafeInteger(record.forward) || record.forward < 0 ||
188
- !Number.isSafeInteger(record.row) || record.row < 0 ||
189
- !Number.isSafeInteger(record.layer) || record.layer < 0 ||
190
- record.indices.length === 0 ||
191
- new Set(record.indices).size !== record.indices.length) {
192
- throw new Error(`GLM Atlas run ${segment} has an invalid route row`);
193
- }
194
- for (const expert of record.indices) {
195
- if (!Number.isSafeInteger(expert) || expert < 0)
196
- throw new Error(`GLM Atlas run ${segment} has invalid expert ${expert}`);
197
- const key = expertKey(record.layer, expert);
198
- counts.set(key, (counts.get(key) ?? 0) + 1);
199
- totalSelections++;
200
- }
201
- }
202
- return Object.freeze({
203
- category,
204
- promptIndex,
205
- totalSelections,
206
- counts,
207
- });
208
- }
209
-
210
- function checkedRuns(
211
- probes: Glm52AtlasProbeSet,
212
- runs: readonly Glm52AtlasRun[],
213
- ): {
214
- categories: string[];
215
- ordered: Glm52AtlasRun[];
216
- } {
217
- const categories = Object.keys(probes.categories).sort();
218
- const expected = new Set<string>();
219
- for (const category of categories) {
220
- const prompts = probes.categories[category]!;
221
- for (let index = 0; index < prompts.length; index++)
222
- expected.add(glm52AtlasRunId(category, index));
223
- }
224
- const seen = new Set<string>();
225
- for (const run of runs) {
226
- const id = glm52AtlasRunId(run.category, run.promptIndex);
227
- if (!expected.has(id)) throw new Error(`unexpected GLM Atlas run ${id}`);
228
- if (seen.has(id)) throw new Error(`duplicate GLM Atlas run ${id}`);
229
- if (!Number.isSafeInteger(run.totalSelections) || run.totalSelections <= 0)
230
- throw new Error(`GLM Atlas run ${id} has no selections`);
231
- let counted = 0;
232
- for (const [key, count] of run.counts) {
233
- parseExpertKey(key);
234
- positiveInteger(count, `GLM Atlas count ${id}/${key}`);
235
- counted += count;
236
- }
237
- if (counted !== run.totalSelections)
238
- throw new Error(
239
- `GLM Atlas run ${id} count total ${counted} != ${run.totalSelections}`,
240
- );
241
- seen.add(id);
242
- }
243
- const missing = [...expected].filter((id) => !seen.has(id));
244
- if (missing.length)
245
- throw new Error(`missing GLM Atlas runs: ${missing.join(", ")}`);
246
- return {
247
- categories,
248
- ordered: runs.slice().sort((left, right) =>
249
- left.category.localeCompare(right.category) ||
250
- left.promptIndex - right.promptIndex),
251
- };
252
- }
253
-
254
- function categoryShares(
255
- categories: readonly string[],
256
- runs: readonly Glm52AtlasRun[],
257
- ): Map<string, Float64Array> {
258
- const index = new Map(categories.map((category, i) => [category, i]));
259
- const byCategory = new Map<string, Glm52AtlasRun[]>();
260
- for (const category of categories) byCategory.set(category, []);
261
- for (const run of runs) byCategory.get(run.category)?.push(run);
262
- const shares = new Map<string, Float64Array>();
263
- for (const category of categories) {
264
- const categoryRuns = byCategory.get(category)!;
265
- if (categoryRuns.length === 0) continue;
266
- const c = index.get(category)!;
267
- for (const run of categoryRuns) {
268
- for (const [key, count] of run.counts) {
269
- let values = shares.get(key);
270
- if (!values) {
271
- values = new Float64Array(categories.length);
272
- shares.set(key, values);
273
- }
274
- values[c] = values[c]! +
275
- count / run.totalSelections / categoryRuns.length;
276
- }
277
- }
278
- }
279
- return shares;
280
- }
281
-
282
- function deterministicTop(
283
- categories: readonly string[],
284
- values: ArrayLike<number>,
285
- ): number {
286
- let best = 0;
287
- for (let index = 1; index < categories.length; index++)
288
- if (values[index]! > values[best]!) best = index;
289
- return best;
290
- }
291
-
292
- export function buildGlm52Atlas(
293
- probes: Glm52AtlasProbeSet,
294
- runs: readonly Glm52AtlasRun[],
295
- options: Glm52AtlasBuildOptions = {},
296
- ): Glm52AtlasAnalysis {
297
- const minCount = options.minCount ?? 30;
298
- const minRuns = options.minRuns ?? 2;
299
- const strongThreshold = options.strongThreshold ?? 0.5;
300
- positiveInteger(minCount, "GLM Atlas minCount");
301
- positiveInteger(minRuns, "GLM Atlas minRuns");
302
- finiteUnit(strongThreshold, "GLM Atlas strongThreshold");
303
- const checked = checkedRuns(probes, runs);
304
- const { categories, ordered } = checked;
305
- if (categories.some((category) =>
306
- probes.categories[category]!.length < minRuns)) {
307
- throw new Error("GLM Atlas minRuns exceeds a category's prompt count");
308
- }
309
- const shares = categoryShares(categories, ordered);
310
- const runsByCategory = new Map<string, Glm52AtlasRun[]>();
311
- for (const category of categories) runsByCategory.set(category, []);
312
- for (const run of ordered) runsByCategory.get(run.category)!.push(run);
313
-
314
- let droppedSparse = 0;
315
- let droppedUnreplicated = 0;
316
- const experts: Glm52AtlasExpert[] = [];
317
- for (const [key, meanShares] of shares) {
318
- const total = ordered.reduce(
319
- (sum, run) => sum + (run.counts.get(key) ?? 0),
320
- 0,
321
- );
322
- if (total < minCount) {
323
- droppedSparse++;
324
- continue;
325
- }
326
- const sum = meanShares.reduce((value, next) => value + next, 0);
327
- if (sum <= 0) continue;
328
- const probabilities = Array.from(meanShares, (value) => value / sum);
329
- const topIndex = deterministicTop(categories, probabilities);
330
- const topTopic = categories[topIndex]!;
331
- const topRuns = runsByCategory.get(topTopic)!;
332
- const firedRuns = topRuns.filter((run) =>
333
- (run.counts.get(key) ?? 0) > 0).length;
334
- if (firedRuns < minRuns) {
335
- droppedUnreplicated++;
336
- continue;
337
- }
338
- const entropy = -probabilities.reduce(
339
- (value, probability) =>
340
- value + (probability > 0 ? probability * Math.log(probability) : 0),
341
- 0,
342
- );
343
- const specialization = Math.max(
344
- 0,
345
- Math.min(1, 1 - entropy / Math.log(categories.length)),
346
- );
347
- const affinity: Record<string, number> = {};
348
- for (let index = 0; index < categories.length; index++)
349
- affinity[categories[index]!] = round(probabilities[index]!);
350
- const [layer, expert] = parseExpertKey(key);
351
- experts.push(Object.freeze({
352
- layer,
353
- expert,
354
- total,
355
- specialization: round(specialization),
356
- topTopic,
357
- topLift: round(probabilities[topIndex]! * categories.length),
358
- reliability: Object.freeze({
359
- firedRuns,
360
- totalRuns: topRuns.length,
361
- }),
362
- affinity: Object.freeze(affinity),
363
- }));
364
- }
365
- experts.sort((left, right) =>
366
- right.specialization - left.specialization ||
367
- right.total - left.total ||
368
- left.layer - right.layer ||
369
- left.expert - right.expert);
370
- const strongSpecialists = experts.filter(
371
- (expert) => expert.specialization >= strongThreshold,
372
- ).length;
373
- return Object.freeze({
374
- schemaVersion: 1,
375
- categories: Object.freeze(categories),
376
- runCount: ordered.length,
377
- method: Object.freeze({
378
- minCount,
379
- minRuns,
380
- strongThreshold,
381
- normalization: "mean per-run selection share",
382
- }),
383
- summary: Object.freeze({
384
- expertsSeen: shares.size,
385
- expertsKept: experts.length,
386
- droppedSparse,
387
- droppedUnreplicated,
388
- strongSpecialists,
389
- strongSpecialistRate: round(
390
- strongSpecialists / Math.max(1, experts.length),
391
- ),
392
- }),
393
- experts: Object.freeze(experts),
394
- });
395
- }
396
-
397
- function specialistSets(
398
- categories: readonly string[],
399
- training: readonly Glm52AtlasRun[],
400
- topK: number,
401
- ): Map<string, Set<string>> {
402
- const shares = categoryShares(categories, training);
403
- const sets = new Map<string, Set<string>>();
404
- for (let categoryIndex = 0;
405
- categoryIndex < categories.length;
406
- categoryIndex++
407
- ) {
408
- const scored: Array<{ key: string; score: number; layer: number; expert: number }> = [];
409
- for (const [key, values] of shares) {
410
- if (values[categoryIndex]! <= 0) continue;
411
- const sum = values.reduce((value, next) => value + next, 0);
412
- const [layer, expert] = parseExpertKey(key);
413
- scored.push({
414
- key,
415
- score: values[categoryIndex]! / sum,
416
- layer,
417
- expert,
418
- });
419
- }
420
- scored.sort((left, right) =>
421
- right.score - left.score ||
422
- left.layer - right.layer ||
423
- left.expert - right.expert);
424
- sets.set(
425
- categories[categoryIndex]!,
426
- new Set(scored.slice(0, topK).map((entry) => entry.key)),
427
- );
428
- }
429
- return sets;
430
- }
431
-
432
- export function validateGlm52Atlas(
433
- probes: Glm52AtlasProbeSet,
434
- runs: readonly Glm52AtlasRun[],
435
- topK = 200,
436
- ): Glm52AtlasValidation {
437
- positiveInteger(topK, "GLM Atlas validation topK");
438
- const { categories, ordered } = checkedRuns(probes, runs);
439
- const trialsByPrompt: Glm52AtlasValidationTrial[] = [];
440
- let hits = 0;
441
- for (const held of ordered) {
442
- // The held-out prompt is absent from every topic model, including the
443
- // background distribution used to score other categories.
444
- const training = ordered.filter((run) => run !== held);
445
- const sets = specialistSets(categories, training, topK);
446
- const scores: Record<string, number> = {};
447
- for (const category of categories) {
448
- const selected = sets.get(category)!;
449
- let count = 0;
450
- for (const [key, value] of held.counts)
451
- if (selected.has(key)) count += value;
452
- scores[category] = count / held.totalSelections;
453
- }
454
- const predictedIndex = deterministicTop(
455
- categories,
456
- categories.map((category) => scores[category]!),
457
- );
458
- const predicted = categories[predictedIndex]!;
459
- const correct = predicted === held.category;
460
- if (correct) hits++;
461
- const bestOtherShare = Math.max(
462
- ...categories.filter((category) => category !== held.category)
463
- .map((category) => scores[category]!),
464
- );
465
- trialsByPrompt.push(Object.freeze({
466
- category: held.category,
467
- promptIndex: held.promptIndex,
468
- predicted,
469
- correct,
470
- ownShare: round(scores[held.category]!),
471
- bestOtherShare: round(bestOtherShare),
472
- scores: Object.freeze(Object.fromEntries(
473
- categories.map((category) => [category, round(scores[category]!)]),
474
- )),
475
- }));
476
- }
477
- return Object.freeze({
478
- topK,
479
- hits,
480
- trials: trialsByPrompt.length,
481
- accuracy: round(hits / trialsByPrompt.length),
482
- chance: round(1 / categories.length),
483
- protocol: "global leave-one-prompt-out",
484
- trialsByPrompt: Object.freeze(trialsByPrompt),
485
- });
486
- }
487
-
488
- export function glm52AtlasWebExperts(
489
- analysis: Glm52AtlasAnalysis,
490
- ): Readonly<Record<string, unknown>> {
491
- const experts: Record<string, unknown> = {};
492
- for (const expert of analysis.experts) {
493
- const entropy = -Object.values(expert.affinity).reduce(
494
- (value, probability) => value +
495
- (probability > 0 ? probability * Math.log2(probability) : 0),
496
- 0,
497
- );
498
- experts[expertKey(expert.layer, expert.expert)] = Object.freeze({
499
- affinity: expert.affinity,
500
- entropy: round(entropy, 3),
501
- specialization: expert.specialization,
502
- top: expert.topTopic,
503
- label: expert.specialization >= analysis.method.strongThreshold
504
- ? `specialist: ${expert.topTopic}`
505
- : "generalist",
506
- total: expert.total,
507
- reliability: expert.reliability,
508
- });
509
- }
510
- return Object.freeze(experts);
511
- }