@mlx-bun/inference 0.0.0-stage → 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 (467) hide show
  1. package/LICENSE +21 -0
  2. package/README.md +1475 -2
  3. package/THIRD_PARTY_NOTICES.md +3404 -0
  4. package/dist/native/libmlx_bun_ane.dylib +0 -0
  5. package/dist/native/libmlx_bun_expert_io.dylib +0 -0
  6. package/dist/native/mlx-bun-frame-extract +0 -0
  7. package/examples/fit-model.ts +25 -0
  8. package/examples/qwen3-forward.ts +42 -0
  9. package/examples/qwen3-generate.ts +28 -0
  10. package/examples/trellis-expand.ts +23 -0
  11. package/package.json +129 -4
  12. package/src/adapters/available.ts +110 -0
  13. package/src/adapters/index.ts +3 -0
  14. package/src/adapters/manager.ts +316 -0
  15. package/src/artifacts/auxiliary-files.ts +45 -0
  16. package/src/artifacts/config-dialects.ts +94 -0
  17. package/src/artifacts/config.ts +341 -0
  18. package/src/artifacts/draft-vocabulary.ts +30 -0
  19. package/src/artifacts/expert-offload-build.ts +75 -0
  20. package/src/artifacts/expert-offload.ts +89 -0
  21. package/src/artifacts/experts/index.ts +3 -0
  22. package/src/artifacts/experts/io.ts +386 -0
  23. package/src/artifacts/experts/residency.ts +1134 -0
  24. package/src/artifacts/experts/usage.ts +415 -0
  25. package/src/artifacts/fingerprint.ts +44 -0
  26. package/src/artifacts/glm52/container.ts +305 -0
  27. package/src/artifacts/glm52/expert-layout.ts +249 -0
  28. package/src/artifacts/glm52/index.ts +7 -0
  29. package/src/artifacts/glm52/kernel-layout.ts +124 -0
  30. package/src/artifacts/glm52/layout.ts +152 -0
  31. package/src/artifacts/glm52/memory.ts +500 -0
  32. package/src/artifacts/glm52/resident-weights.ts +135 -0
  33. package/src/artifacts/glm52/weights.ts +503 -0
  34. package/src/artifacts/glm52-config.ts +262 -0
  35. package/src/artifacts/identity.ts +94 -0
  36. package/src/artifacts/index.ts +15 -0
  37. package/src/artifacts/mmap.ts +90 -0
  38. package/src/artifacts/safetensors-writer.ts +249 -0
  39. package/src/artifacts/safetensors.ts +164 -0
  40. package/src/artifacts/turboquant-fold.ts +42 -0
  41. package/src/artifacts/weight-audit.ts +43 -0
  42. package/src/artifacts/weights.ts +255 -0
  43. package/src/contracts/index.ts +2 -0
  44. package/src/contracts/mlx/attention.ts +8 -0
  45. package/src/contracts/mlx/cache.ts +510 -0
  46. package/src/contracts/mlx/checkpoint.ts +10 -0
  47. package/src/contracts/mlx/draft-checkpoint.ts +14 -0
  48. package/src/contracts/mlx/draft-projection.ts +27 -0
  49. package/src/contracts/mlx/draft-target.ts +84 -0
  50. package/src/contracts/mlx/drafter.ts +176 -0
  51. package/src/contracts/mlx/forward-work.ts +23 -0
  52. package/src/contracts/mlx/glm52-weights.ts +25 -0
  53. package/src/contracts/mlx/graph.ts +309 -0
  54. package/src/contracts/mlx/head.ts +6 -0
  55. package/src/contracts/mlx/index.ts +15 -0
  56. package/src/contracts/mlx/kv-codec.ts +9 -0
  57. package/src/contracts/mlx/media.ts +113 -0
  58. package/src/contracts/mlx/positions.ts +16 -0
  59. package/src/contracts/mlx/row-table.ts +2 -0
  60. package/src/contracts/mlx/trainable.ts +103 -0
  61. package/src/contracts/mlx/whisper.ts +96 -0
  62. package/src/contracts/portable/composition.ts +106 -0
  63. package/src/contracts/portable/denoising.ts +17 -0
  64. package/src/contracts/portable/execution.ts +79 -0
  65. package/src/contracts/portable/generation.ts +103 -0
  66. package/src/contracts/portable/graph.ts +80 -0
  67. package/src/contracts/portable/index.ts +15 -0
  68. package/src/contracts/portable/memory-plan.ts +26 -0
  69. package/src/contracts/portable/object-cache.ts +8 -0
  70. package/src/contracts/portable/prefill.ts +63 -0
  71. package/src/contracts/portable/prefix-cache.ts +33 -0
  72. package/src/contracts/portable/prompt.ts +2 -0
  73. package/src/contracts/portable/resources.ts +25 -0
  74. package/src/contracts/portable/rollback.ts +7 -0
  75. package/src/contracts/portable/sampling.ts +44 -0
  76. package/src/contracts/portable/scheduling.ts +54 -0
  77. package/src/contracts/portable/token-delivery.ts +4 -0
  78. package/src/contracts/portable/tsconfig.json +14 -0
  79. package/src/embeddings/index.ts +1 -0
  80. package/src/embeddings/text.ts +85 -0
  81. package/src/execution/admission.ts +59 -0
  82. package/src/execution/batch-group.ts +1240 -0
  83. package/src/execution/batch-types.ts +225 -0
  84. package/src/execution/cancellation.ts +38 -0
  85. package/src/execution/continuation-request.ts +45 -0
  86. package/src/execution/continuation-types.ts +10 -0
  87. package/src/execution/coordinator.ts +70 -0
  88. package/src/execution/denoising-group.ts +180 -0
  89. package/src/execution/engine.ts +56 -0
  90. package/src/execution/fill-group.ts +505 -0
  91. package/src/execution/fit.ts +184 -0
  92. package/src/execution/gateway-binding.ts +178 -0
  93. package/src/execution/grammar-group.ts +224 -0
  94. package/src/execution/index.ts +18 -0
  95. package/src/execution/methods.ts +75 -0
  96. package/src/execution/mixed-iteration.ts +96 -0
  97. package/src/execution/plan.ts +65 -0
  98. package/src/execution/prefill-cohort.ts +100 -0
  99. package/src/execution/prefill-rows.ts +263 -0
  100. package/src/execution/request-policy.ts +19 -0
  101. package/src/execution/row-sampling.ts +49 -0
  102. package/src/execution/scheduler.ts +92 -0
  103. package/src/execution/session.ts +247 -0
  104. package/src/execution/speculative-group.ts +797 -0
  105. package/src/execution/tasks.ts +44 -0
  106. package/src/generation/autoregressive.ts +840 -0
  107. package/src/generation/bindings/autoregressive.ts +104 -0
  108. package/src/generation/bindings/denoising.ts +60 -0
  109. package/src/generation/bindings/prefill-policy.ts +36 -0
  110. package/src/generation/checkpoint-identity.ts +60 -0
  111. package/src/generation/denoising.ts +85 -0
  112. package/src/generation/diffusion.ts +583 -0
  113. package/src/generation/fill/echo-index.ts +188 -0
  114. package/src/generation/fill/index.ts +3 -0
  115. package/src/generation/fill/proposal.ts +69 -0
  116. package/src/generation/fill/schema-rows.ts +506 -0
  117. package/src/generation/fill/session.ts +464 -0
  118. package/src/generation/fill/tool-boundary.ts +103 -0
  119. package/src/generation/fill-append.ts +58 -0
  120. package/src/generation/generate.ts +44 -0
  121. package/src/generation/grammar-step.ts +39 -0
  122. package/src/generation/index.ts +14 -0
  123. package/src/generation/phase-dispatch.ts +115 -0
  124. package/src/generation/prefill.ts +33 -0
  125. package/src/generation/result.ts +49 -0
  126. package/src/generation/scopes.ts +157 -0
  127. package/src/generation/speculative/acceptance.ts +41 -0
  128. package/src/generation/speculative/adaptive-draft-gate.ts +166 -0
  129. package/src/generation/speculative/bindings/assistant-rows.ts +102 -0
  130. package/src/generation/speculative/bindings/autoregressive-rows.ts +124 -0
  131. package/src/generation/speculative/bindings/binding.ts +106 -0
  132. package/src/generation/speculative/bindings/deepspec-rows.ts +33 -0
  133. package/src/generation/speculative/bindings/draft-target.ts +10 -0
  134. package/src/generation/speculative/bindings/glm52-mtp-rows.ts +128 -0
  135. package/src/generation/speculative/bindings/projected-context-rows.ts +78 -0
  136. package/src/generation/speculative/bindings/projected-draft-rows.ts +95 -0
  137. package/src/generation/speculative/bindings/qwen-mtp-rows.ts +336 -0
  138. package/src/generation/speculative/bindings/two-model-rows.ts +63 -0
  139. package/src/generation/speculative/cache-identity.ts +16 -0
  140. package/src/generation/speculative/draft-kind.ts +149 -0
  141. package/src/generation/speculative/draft-policy.ts +41 -0
  142. package/src/generation/speculative/draft-registry.ts +118 -0
  143. package/src/generation/speculative/draft-vocab.ts +110 -0
  144. package/src/generation/speculative/index.ts +17 -0
  145. package/src/generation/speculative/output.ts +49 -0
  146. package/src/generation/speculative/round.ts +307 -0
  147. package/src/generation/speculative/source.ts +144 -0
  148. package/src/generation/speculative/sources/assistant-source.ts +64 -0
  149. package/src/generation/speculative/sources/deepspec-source.ts +70 -0
  150. package/src/generation/speculative/sources/dflash-source.ts +84 -0
  151. package/src/generation/speculative/sources/dflash2-source.ts +86 -0
  152. package/src/generation/speculative/sources/glm52-mtp-source.ts +66 -0
  153. package/src/generation/speculative/sources/ngram-source.ts +165 -0
  154. package/src/generation/speculative/sources/qwen-mtp-source.ts +154 -0
  155. package/src/generation/speculative/sources/two-model.ts +75 -0
  156. package/src/generation/types.ts +194 -0
  157. package/src/index.ts +20 -0
  158. package/src/input/audio/audiotoolbox.ts +112 -0
  159. package/src/input/audio/decode.ts +133 -0
  160. package/src/input/audio/features.ts +202 -0
  161. package/src/input/audio/index.ts +6 -0
  162. package/src/input/audio/transcode.ts +67 -0
  163. package/src/input/audio/whisper-mel.ts +185 -0
  164. package/src/input/audio/whisper-tokenizer.ts +238 -0
  165. package/src/input/chat-template.ts +389 -0
  166. package/src/input/index.ts +6 -0
  167. package/src/input/media-fetch.ts +439 -0
  168. package/src/input/packed-mixed.ts +43 -0
  169. package/src/input/special-tokens.ts +75 -0
  170. package/src/input/token-groups.ts +49 -0
  171. package/src/input/tokenizer.ts +336 -0
  172. package/src/input/tool-call.ts +689 -0
  173. package/src/input/vision/index.ts +3 -0
  174. package/src/input/vision/preprocess.ts +247 -0
  175. package/src/input/vision/prompt.ts +398 -0
  176. package/src/input/vision/qwen3vl-preprocess.ts +409 -0
  177. package/src/input/vision/qwen3vl-prompt.ts +237 -0
  178. package/src/input/vision/video-frames.ts +100 -0
  179. package/src/kernels/ane/fill.ts +101 -0
  180. package/src/kernels/ane/linear.ts +416 -0
  181. package/src/kernels/attention/flash.ts +694 -0
  182. package/src/kernels/attention/kvq4-decode.ts +155 -0
  183. package/src/kernels/attention/masks.ts +40 -0
  184. package/src/kernels/attention/multi-query.ts +193 -0
  185. package/src/kernels/attention/paged.ts +130 -0
  186. package/src/kernels/delta/gated.ts +306 -0
  187. package/src/kernels/delta/index.ts +3 -0
  188. package/src/kernels/delta/silu-rms-heads.ts +69 -0
  189. package/src/kernels/delta/state.ts +74 -0
  190. package/src/kernels/gated-rms-norm.ts +64 -0
  191. package/src/kernels/glm52/layout.ts +37 -0
  192. package/src/kernels/glm52/streamed-experts.ts +360 -0
  193. package/src/kernels/logits.ts +38 -0
  194. package/src/kernels/quantization/affine-row.ts +154 -0
  195. package/src/kernels/quantization/affine-rows.ts +36 -0
  196. package/src/kernels/quantization/affine-split-mma.ts +144 -0
  197. package/src/kernels/quantization/affine3-mma.ts +117 -0
  198. package/src/kernels/quantization/affine3-rows.ts +103 -0
  199. package/src/kernels/quantization/affine4-mma.ts +113 -0
  200. package/src/kernels/quantization/affine4-rows.ts +102 -0
  201. package/src/kernels/residual-rms-norm.ts +65 -0
  202. package/src/kernels/sampling/index.ts +2 -0
  203. package/src/kernels/sampling/normalized-argmax.ts +116 -0
  204. package/src/kernels/sampling/token-bitmask.ts +52 -0
  205. package/src/kernels/softcap.ts +35 -0
  206. package/src/kernels/trellis/codebook.ts +123 -0
  207. package/src/kernels/trellis/down-codebook.ts +184 -0
  208. package/src/kernels/trellis/down-factored.ts +180 -0
  209. package/src/kernels/trellis/down-k3-interleaved-factored.ts +146 -0
  210. package/src/kernels/trellis/down-mma-wide.ts +195 -0
  211. package/src/kernels/trellis/down-mma.ts +145 -0
  212. package/src/kernels/trellis/expand.ts +42 -0
  213. package/src/kernels/trellis/gate-up-codebook-mma.ts +141 -0
  214. package/src/kernels/trellis/gate-up-codebook.ts +137 -0
  215. package/src/kernels/trellis/gate-up-factored.ts +197 -0
  216. package/src/kernels/trellis/gate-up-mma-wide.ts +147 -0
  217. package/src/kernels/trellis/gate-up-mma.ts +148 -0
  218. package/src/kernels/trellis/gate-up.ts +248 -0
  219. package/src/kernels/trellis/geometry.ts +22 -0
  220. package/src/kernels/trellis/index.ts +103 -0
  221. package/src/kernels/trellis/launch.ts +12 -0
  222. package/src/kernels/trellis/mixed-gate-up-factored.ts +119 -0
  223. package/src/kernels/trellis/mixed-gate-up.ts +146 -0
  224. package/src/kernels/trellis/qualified.ts +34 -0
  225. package/src/kernels/trellis/reduce.ts +150 -0
  226. package/src/kernels/trellis/scatter.ts +407 -0
  227. package/src/kernels/trellis/splitk-prefill.ts +152 -0
  228. package/src/kernels/trellis/tiled-prefill.ts +131 -0
  229. package/src/kernels/trellis/vector-expand.ts +58 -0
  230. package/src/kernels/trellis/wide-prefill.ts +89 -0
  231. package/src/kernels/turboquant/decode.ts +82 -0
  232. package/src/kernels/turboquant/index.ts +5 -0
  233. package/src/kernels/turboquant/inverse.ts +122 -0
  234. package/src/kernels/turboquant/ops.ts +860 -0
  235. package/src/kernels/turboquant/signs.ts +130 -0
  236. package/src/kernels/turboquant/tables.ts +124 -0
  237. package/src/layers/affine-verify-linear.ts +75 -0
  238. package/src/layers/ane-prefill-split.ts +283 -0
  239. package/src/layers/compiled-normalization.ts +41 -0
  240. package/src/layers/delta-heads.ts +23 -0
  241. package/src/layers/dense-embedding.ts +29 -0
  242. package/src/layers/dense-linear.ts +32 -0
  243. package/src/layers/dequantized-linear.ts +72 -0
  244. package/src/layers/folded-attention.ts +42 -0
  245. package/src/layers/geglu.ts +130 -0
  246. package/src/layers/helpers.ts +7 -0
  247. package/src/layers/index.ts +36 -0
  248. package/src/layers/kv4-decode-attention.ts +15 -0
  249. package/src/layers/layer-norm.ts +24 -0
  250. package/src/layers/loading.ts +92 -0
  251. package/src/layers/lora.ts +150 -0
  252. package/src/layers/normalization.ts +9 -0
  253. package/src/layers/prefix-layout.ts +33 -0
  254. package/src/layers/quantized-append-attention.ts +32 -0
  255. package/src/layers/quantized-attention.ts +353 -0
  256. package/src/layers/quantized-embedding.ts +39 -0
  257. package/src/layers/quantized-linear.ts +137 -0
  258. package/src/layers/quantized-switch-linear.ts +37 -0
  259. package/src/layers/qwen-append.ts +12 -0
  260. package/src/layers/qwen-mrope.ts +238 -0
  261. package/src/layers/rope-2d.ts +110 -0
  262. package/src/layers/rope.ts +389 -0
  263. package/src/layers/swiglu.ts +23 -0
  264. package/src/layers/trellis-codebook-down.ts +52 -0
  265. package/src/layers/trellis-codebook-gate-up.ts +42 -0
  266. package/src/layers/trellis-down.ts +117 -0
  267. package/src/layers/trellis-gate-up.ts +71 -0
  268. package/src/layers/trellis-linear.ts +211 -0
  269. package/src/layers/trellis-prompt.ts +102 -0
  270. package/src/layers/trellis-reduce-gate-up.ts +73 -0
  271. package/src/layers/trellis-scatter-down.ts +98 -0
  272. package/src/models/ane-install.ts +74 -0
  273. package/src/models/audio/conformer.ts +656 -0
  274. package/src/models/cache-composition.ts +44 -0
  275. package/src/models/capabilities.ts +89 -0
  276. package/src/models/chat-template.ts +10 -0
  277. package/src/models/composition.ts +107 -0
  278. package/src/models/diffusion-gemma/experts.ts +59 -0
  279. package/src/models/diffusion-gemma/family.ts +21 -0
  280. package/src/models/diffusion-gemma/model.ts +900 -0
  281. package/src/models/diffusion-gemma/vision.ts +549 -0
  282. package/src/models/drafter-loaders.ts +49 -0
  283. package/src/models/drafters.ts +39 -0
  284. package/src/models/factory.ts +351 -0
  285. package/src/models/families.ts +70 -0
  286. package/src/models/family.ts +125 -0
  287. package/src/models/gemma4/assistant-target.ts +46 -0
  288. package/src/models/gemma4/assistant.ts +327 -0
  289. package/src/models/gemma4/blocks.ts +755 -0
  290. package/src/models/gemma4/compiled-decode.ts +452 -0
  291. package/src/models/gemma4/family.ts +39 -0
  292. package/src/models/gemma4/gemma4-12b.ts +677 -0
  293. package/src/models/gemma4/gemma4-26b.ts +544 -0
  294. package/src/models/gemma4/gemma4-e4b.ts +599 -0
  295. package/src/models/gemma4/graphs.ts +147 -0
  296. package/src/models/gemma4/media.ts +88 -0
  297. package/src/models/gemma4/model.ts +860 -0
  298. package/src/models/gemma4/serving-blocks.ts +72 -0
  299. package/src/models/gemma4/trainable.ts +186 -0
  300. package/src/models/gemma4/written-blocks.ts +232 -0
  301. package/src/models/gemma4/written-forward.ts +159 -0
  302. package/src/models/gemma4/written-input.ts +29 -0
  303. package/src/models/glm52/atlas.ts +507 -0
  304. package/src/models/glm52/coupling.ts +388 -0
  305. package/src/models/glm52/dsa.ts +601 -0
  306. package/src/models/glm52/family.ts +20 -0
  307. package/src/models/glm52/mla.ts +812 -0
  308. package/src/models/glm52/model.ts +1127 -0
  309. package/src/models/glm52/moe.ts +223 -0
  310. package/src/models/glm52/mtp.ts +60 -0
  311. package/src/models/glm52/pilot.ts +537 -0
  312. package/src/models/glm52/reference.ts +445 -0
  313. package/src/models/glm52/residency.ts +557 -0
  314. package/src/models/glm52/streamed-experts.ts +447 -0
  315. package/src/models/graph.ts +59 -0
  316. package/src/models/implementation.ts +230 -0
  317. package/src/models/index.ts +15 -0
  318. package/src/models/media-input.ts +15 -0
  319. package/src/models/memory-plan.ts +39 -0
  320. package/src/models/minicpm5/family.ts +29 -0
  321. package/src/models/minicpm5/model.ts +394 -0
  322. package/src/models/minicpm5/trainable.ts +90 -0
  323. package/src/models/profile.ts +337 -0
  324. package/src/models/qwen/blocks.ts +658 -0
  325. package/src/models/qwen/checkpoint.ts +190 -0
  326. package/src/models/qwen/families.ts +57 -0
  327. package/src/models/qwen/media-input.ts +161 -0
  328. package/src/models/qwen/mtp.ts +101 -0
  329. package/src/models/qwen/qwen3-moe.ts +355 -0
  330. package/src/models/qwen/qwen3.ts +323 -0
  331. package/src/models/qwen/qwen38-27b-trellis-m1max.ts +1532 -0
  332. package/src/models/qwen/qwen38-27b-trellis-m4pro.ts +1092 -0
  333. package/src/models/qwen/qwen3_5.ts +433 -0
  334. package/src/models/qwen/written-layers.ts +285 -0
  335. package/src/models/row-table.ts +1 -0
  336. package/src/models/runtime.ts +73 -0
  337. package/src/models/speculative/deepspec-artifact.ts +10 -0
  338. package/src/models/speculative/deepspec.ts +744 -0
  339. package/src/models/speculative/dflash.ts +553 -0
  340. package/src/models/speculative/dflash2-artifact.ts +9 -0
  341. package/src/models/speculative/dflash2-operations.ts +1 -0
  342. package/src/models/speculative/dflash2.ts +576 -0
  343. package/src/models/speculative/loader.ts +29 -0
  344. package/src/models/support.ts +79 -0
  345. package/src/models/universal/archs.ts +365 -0
  346. package/src/models/universal/blocks.ts +479 -0
  347. package/src/models/universal/dense.ts +255 -0
  348. package/src/models/universal/family.ts +17 -0
  349. package/src/models/vision/qwen3vl.ts +566 -0
  350. package/src/models/vision/siglip.ts +520 -0
  351. package/src/models/vision/unified.ts +176 -0
  352. package/src/models/whisper/family.ts +19 -0
  353. package/src/models/whisper/fast.ts +734 -0
  354. package/src/models/whisper/model.ts +430 -0
  355. package/src/runtime/compiled-trace.ts +16 -0
  356. package/src/runtime/config.ts +77 -0
  357. package/src/runtime/expert-trace.ts +71 -0
  358. package/src/runtime/graph-forward.ts +65 -0
  359. package/src/runtime/integers.ts +6 -0
  360. package/src/runtime/native.ts +16 -0
  361. package/src/runtime/resources.ts +53 -0
  362. package/src/runtime/row-operations.ts +42 -0
  363. package/src/runtime/trace.ts +161 -0
  364. package/src/sampling/curve.ts +122 -0
  365. package/src/sampling/draft.ts +55 -0
  366. package/src/sampling/extras.ts +45 -0
  367. package/src/sampling/filters.ts +134 -0
  368. package/src/sampling/grammar.ts +630 -0
  369. package/src/sampling/hlg.ts +290 -0
  370. package/src/sampling/index.ts +9 -0
  371. package/src/sampling/processors.ts +142 -0
  372. package/src/sampling/sampler.ts +120 -0
  373. package/src/sampling/step.ts +343 -0
  374. package/src/sampling/types.ts +161 -0
  375. package/src/sampling/window-filters.ts +46 -0
  376. package/src/scoring/decisions.ts +96 -0
  377. package/src/scoring/full-sequence.ts +103 -0
  378. package/src/scoring/index.ts +9 -0
  379. package/src/scoring/kl.ts +29 -0
  380. package/src/scoring/perplexity.ts +165 -0
  381. package/src/scoring/response-logps.ts +65 -0
  382. package/src/state/affine-attention.ts +180 -0
  383. package/src/state/attachment.ts +68 -0
  384. package/src/state/attention-read.ts +215 -0
  385. package/src/state/batched-kv.ts +294 -0
  386. package/src/state/batched-mask.ts +445 -0
  387. package/src/state/batched-quant.ts +389 -0
  388. package/src/state/batched-quantized-kv.ts +209 -0
  389. package/src/state/batched-rotating-quant.ts +457 -0
  390. package/src/state/batched-rotating-state.ts +149 -0
  391. package/src/state/batched-rotating.ts +430 -0
  392. package/src/state/batched-row-storage.ts +179 -0
  393. package/src/state/batched-ssm.ts +54 -0
  394. package/src/state/batched-turboquant-kv.ts +144 -0
  395. package/src/state/bf16-first-kv.ts +663 -0
  396. package/src/state/cache-service.ts +592 -0
  397. package/src/state/capabilities.ts +65 -0
  398. package/src/state/checkpoint.ts +46 -0
  399. package/src/state/continuation-persistence.ts +126 -0
  400. package/src/state/decode-step-inputs.ts +31 -0
  401. package/src/state/decoded-kv-donor.ts +12 -0
  402. package/src/state/delayed-quantized-kv.ts +141 -0
  403. package/src/state/delayed-rotating-quantized-kv.ts +178 -0
  404. package/src/state/delayed-turboquant-kv.ts +123 -0
  405. package/src/state/durability.ts +214 -0
  406. package/src/state/encoder-cache.ts +26 -0
  407. package/src/state/full-kv-row-append.ts +33 -0
  408. package/src/state/full-kv-row-donor.ts +54 -0
  409. package/src/state/full-prefill-padding.ts +56 -0
  410. package/src/state/full-prefill-row.ts +112 -0
  411. package/src/state/full-transitioning-kv-rows.ts +70 -0
  412. package/src/state/glm52-cache.ts +783 -0
  413. package/src/state/grouped-heads-kv4-cache.ts +26 -0
  414. package/src/state/index.ts +38 -0
  415. package/src/state/kv-attention-view.ts +77 -0
  416. package/src/state/kv-budget.ts +14 -0
  417. package/src/state/kv-scheme.ts +315 -0
  418. package/src/state/kv-storage.ts +57 -0
  419. package/src/state/kv-tensor-rows.ts +212 -0
  420. package/src/state/kv.ts +324 -0
  421. package/src/state/kv4-head256-cache.ts +46 -0
  422. package/src/state/latent-legos.ts +895 -0
  423. package/src/state/layout.ts +277 -0
  424. package/src/state/leases.ts +26 -0
  425. package/src/state/paged/cache.ts +689 -0
  426. package/src/state/paged/codec.ts +36 -0
  427. package/src/state/paged/rows.ts +223 -0
  428. package/src/state/persistence-types.ts +173 -0
  429. package/src/state/persistence-worker.ts +83 -0
  430. package/src/state/persistence.ts +1036 -0
  431. package/src/state/persistence.worker.d.ts +3 -0
  432. package/src/state/persistence.worker.js +264 -0
  433. package/src/state/prefix-cache.ts +685 -0
  434. package/src/state/quantized-kv.ts +369 -0
  435. package/src/state/quantized-tensor.ts +26 -0
  436. package/src/state/request-policy.ts +48 -0
  437. package/src/state/retention-policy.ts +37 -0
  438. package/src/state/rollback.ts +131 -0
  439. package/src/state/rotating-kv-donor.ts +35 -0
  440. package/src/state/rotating-kv-layout.ts +341 -0
  441. package/src/state/rotating-kv.ts +529 -0
  442. package/src/state/rotating-quantized-kv.ts +616 -0
  443. package/src/state/rotating-row-transaction.ts +35 -0
  444. package/src/state/rows.ts +104 -0
  445. package/src/state/softcap-attention.ts +118 -0
  446. package/src/state/speculative/glm52-mtp-state.ts +38 -0
  447. package/src/state/speculative/qwen-mtp-state.ts +34 -0
  448. package/src/state/speculative-rotating-kv.ts +113 -0
  449. package/src/state/ssd-cache.ts +434 -0
  450. package/src/state/ssm-prefill-padding.ts +39 -0
  451. package/src/state/ssm.ts +422 -0
  452. package/src/state/target-layout.ts +7 -0
  453. package/src/state/tiered-prefix-cache.ts +45 -0
  454. package/src/state/training-cache.ts +300 -0
  455. package/src/state/transitioning-kv-rows.ts +267 -0
  456. package/src/state/turboquant-attention.ts +97 -0
  457. package/src/state/turboquant-codec.ts +90 -0
  458. package/src/state/turboquant-kv.ts +442 -0
  459. package/src/state/views.ts +26 -0
  460. package/src/transcription/index.ts +3 -0
  461. package/src/transcription/silero-vad.ts +379 -0
  462. package/src/transcription/whisper/decode.ts +773 -0
  463. package/src/transcription/whisper/format.ts +74 -0
  464. package/src/transcription/whisper/index.ts +4 -0
  465. package/src/transcription/whisper/timing.ts +274 -0
  466. package/src/transcription/whisper/transcribe.ts +394 -0
  467. package/src/transcription/whisper/types.ts +22 -0
@@ -0,0 +1,689 @@
1
+ // Gemma 4 tool-call parsing — port of mlx-lm tool_parsers/gemma4.py.
2
+ //
3
+ // The model emits <|tool_call>call:name{key:value,...}<tool_call|>
4
+ // where strings are delimited by the <|"|> special token, keys are bare,
5
+ // and numbers/bools/objects/arrays are JSON-ish. After its tool_call
6
+ // blocks the model emits <|tool_response> (token 50, in the EOS set) —
7
+ // the request to the runtime for results.
8
+ //
9
+ // The marker tokens (<|tool_call>, <tool_call|>, <|channel>, <channel|>) are
10
+ // declared by the model profile as token text and resolved to ids against the
11
+ // tokenizer in use (input/special-tokens.ts); nothing here names vocabulary ids.
12
+ // With thinking on, the model wraps its chain-of-thought as
13
+ // <|channel>thought\n…<channel|> before the final answer; those special tokens
14
+ // are stripped at decode, so reasoning is split at the TOKEN level
15
+ // (ToolAwareStream), not from decoded text.
16
+
17
+ const QUOTE = '<|"|>';
18
+
19
+ export interface ParsedToolCall {
20
+ name: string;
21
+ /** Parsed argument object (callers serialize for OpenAI's string field). */
22
+ arguments: Record<string, unknown>;
23
+ /** Set only when strict parsing failed and a repair pass recovered the
24
+ * call — never silent (`02d723a:docs/design/web-chat-redesign.md` Appendix A,
25
+ * beat-matrix Axis 7, self-healing row). `repairs` lists
26
+ * every malformation class that had to be corrected, in application
27
+ * order, so failures are observable (logs/telemetry) instead of masked. */
28
+ repaired?: true;
29
+ repairs?: RepairKind[];
30
+ }
31
+
32
+ interface ToolSpec {
33
+ function?: { name?: string; parameters?: Record<string, unknown> };
34
+ name?: string;
35
+ }
36
+
37
+ /** Convert gemma4 argument syntax to JSON: extract <|"|>-strings to
38
+ * placeholders, quote bare keys, restore strings JSON-escaped. */
39
+ export function gemmaArgsToJson(text: string): string {
40
+ const strings: string[] = [];
41
+ let out = "";
42
+ let i = 0;
43
+ while (i < text.length) {
44
+ const start = text.indexOf(QUOTE, i);
45
+ if (start === -1) {
46
+ out += text.slice(i);
47
+ break;
48
+ }
49
+ const end = text.indexOf(QUOTE, start + QUOTE.length);
50
+ if (end === -1) throw new Error("unterminated <|\"|> string");
51
+ out += text.slice(i, start) + `\x00${strings.length}\x00`;
52
+ strings.push(text.slice(start + QUOTE.length, end));
53
+ i = end + QUOTE.length;
54
+ }
55
+ // quote bare keys (after { or ,)
56
+ out = out.replace(/([{,])\s*([\w-]+)\s*:/g, (_, pre, key) => `${pre}${JSON.stringify(key)}:`);
57
+ // restore strings as JSON literals
58
+ out = out.replace(/\x00(\d+)\x00/g, (_, n) => JSON.stringify(strings[Number(n)]!));
59
+ return out;
60
+ }
61
+
62
+ /** Parse every `call:name{...}` block in a decoded tool-call segment.
63
+ *
64
+ * Strict scan first (unchanged behavior on well-formed input). On the two
65
+ * ways a small model mangles THIS family's syntax specifically —
66
+ * (a) generation truncated mid-call (an unterminated `<|"|>` string or a
67
+ * brace that never closes, both hit_max_tokens artifacts) and (b) a
68
+ * trailing comma surviving `gemmaArgsToJson`'s conversion to JSON — the
69
+ * repair pass recovers by treating "ran off the end of the text" as an
70
+ * implicit close (mirrors closeUnbalanced's brace-stack unwind, adapted
71
+ * for the `<|"|>`-quoted syntax) and re-running stripTrailingCommas on the
72
+ * converted JSON. Recovered calls are tagged repaired/repairs; a segment
73
+ * that still can't be recovered throws exactly as before. */
74
+ export function parseToolCalls(text: string): ParsedToolCall[] {
75
+ const calls: ParsedToolCall[] = [];
76
+ let i = 0;
77
+ while (i < text.length) {
78
+ const m = /call:([\w-]+)\{/.exec(text.slice(i));
79
+ if (!m) break;
80
+ const nameEnd = i + m.index + m[0].length - 1; // position of '{'
81
+ // balanced-brace scan, skipping <|"|>-delimited strings
82
+ let depth = 0;
83
+ let j = nameEnd;
84
+ let unterminatedString = false;
85
+ while (j < text.length) {
86
+ if (text.startsWith(QUOTE, j)) {
87
+ const close = text.indexOf(QUOTE, j + QUOTE.length);
88
+ if (close === -1) { unterminatedString = true; break; }
89
+ j = close + QUOTE.length;
90
+ continue;
91
+ }
92
+ const c = text[j];
93
+ if (c === "{") depth++;
94
+ else if (c === "}") {
95
+ depth--;
96
+ if (depth === 0) break;
97
+ }
98
+ j++;
99
+ }
100
+ const truncated = unterminatedString || depth !== 0;
101
+ let argsBlock: string;
102
+ const repairs: RepairKind[] = [];
103
+ if (truncated) {
104
+ // Ran off the end of the text without closing — truncated generation.
105
+ // Close the dangling <|"|> string (if any) then append enough `}` to
106
+ // balance whatever braces are still open, using the SAME depth
107
+ // counter's remaining value (skip <|"|>-quoted spans again so a
108
+ // brace character inside a string doesn't miscount).
109
+ let closeDepth = 0;
110
+ let k = nameEnd;
111
+ let inQuote = false;
112
+ while (k < text.length) {
113
+ if (text.startsWith(QUOTE, k) && !inQuote) { inQuote = true; k += QUOTE.length; continue; }
114
+ if (inQuote) {
115
+ if (text.startsWith(QUOTE, k)) { inQuote = false; k += QUOTE.length; continue; }
116
+ k++;
117
+ continue;
118
+ }
119
+ const c = text[k];
120
+ if (c === "{") closeDepth++;
121
+ else if (c === "}") closeDepth--;
122
+ k++;
123
+ }
124
+ argsBlock = text.slice(nameEnd) + (inQuote ? QUOTE : "") + "}".repeat(Math.max(closeDepth, 1));
125
+ repairs.push("unbalanced_braces");
126
+ j = text.length - 1;
127
+ } else {
128
+ argsBlock = text.slice(nameEnd, j + 1);
129
+ }
130
+ let converted = gemmaArgsToJson(argsBlock);
131
+ let parsed: unknown;
132
+ try {
133
+ parsed = JSON.parse(converted);
134
+ } catch {
135
+ const commas = stripTrailingCommas(converted);
136
+ if (!commas.changed) throw new Error("unbalanced braces in tool call");
137
+ try {
138
+ parsed = JSON.parse(commas.out);
139
+ } catch {
140
+ throw new Error("unbalanced braces in tool call");
141
+ }
142
+ converted = commas.out;
143
+ repairs.push("trailing_commas");
144
+ }
145
+ calls.push({
146
+ name: m[1]!,
147
+ arguments: parsed as Record<string, unknown>,
148
+ ...(repairs.length ? { repaired: true, repairs } : {}),
149
+ });
150
+ i = j + 1;
151
+ }
152
+ return calls;
153
+ }
154
+
155
+ const TOOL_CALL_BLOCK_RE = /<tool_call>\s*([\s\S]*?)\s*<\/tool_call>/gi;
156
+ const XML_FUNCTION_EQUALS_RE = /^\s*<function=([^>\s]+)>\s*([\s\S]*?)\s*<\/function>\s*$/i;
157
+ const XML_PARAMETER_EQUALS_RE = /<parameter=([^>\s]+)>([\s\S]*?)<\/parameter>/gi;
158
+ const GLM52_ARG_KEY_RE = /<arg_key>/i;
159
+ const GLM52_ARG_PAIR_RE = /<arg_key>\s*([\s\S]*?)\s*<\/arg_key>\s*<arg_value>\s*([\s\S]*?)\s*<\/arg_value>/gi;
160
+ const XML_FUNCTION_ATTR_RE = /<function\s+name=["']([^"']+)["']\s*>\s*([\s\S]*?)\s*<\/function>/gi;
161
+ // CDATA alternative first so a `</param>` inside a CDATA block never
162
+ // terminates the value early.
163
+ const XML_PARAM_ATTR_RE = /<param\s+name=["']([^"']+)["']\s*>((?:<!\[CDATA\[[\s\S]*?\]\]>|[\s\S])*?)<\/param>/gi;
164
+
165
+ function toolSpecName(tool: ToolSpec): string | null {
166
+ const name = tool.function?.name ?? tool.name;
167
+ return typeof name === "string" && name.trim() ? name.trim() : null;
168
+ }
169
+
170
+ function decodeXml(text: string): string {
171
+ return text
172
+ .replace(/<!\[CDATA\[([\s\S]*?)\]\]>/g, "$1")
173
+ .replace(/&quot;/g, '"')
174
+ .replace(/&apos;/g, "'")
175
+ .replace(/&lt;/g, "<")
176
+ .replace(/&gt;/g, ">")
177
+ .replace(/&amp;/g, "&")
178
+ .trim();
179
+ }
180
+
181
+ /** JSON-schema type names for a parameter (handles `type: [..]` unions). */
182
+ function schemaTypeNames(schema: unknown): Set<string> {
183
+ if (!schema || typeof schema !== "object") return new Set();
184
+ const raw = (schema as Record<string, unknown>).type;
185
+ if (typeof raw === "string") return new Set([raw]);
186
+ if (Array.isArray(raw)) return new Set(raw.filter((t): t is string => typeof t === "string"));
187
+ return new Set();
188
+ }
189
+
190
+ function toolParameterSchema(
191
+ tools: ToolSpec[], toolName: string, parameterName: string,
192
+ ): Record<string, unknown> | null {
193
+ for (const tool of tools) {
194
+ if (toolSpecName(tool) !== toolName) continue;
195
+ const params = tool.function?.parameters ?? (tool as Record<string, any>).parameters;
196
+ const props = params && typeof params === "object"
197
+ ? (params as Record<string, any>).properties : null;
198
+ const schema = props && typeof props === "object" ? props[parameterName] : null;
199
+ return schema && typeof schema === "object" ? schema : null;
200
+ }
201
+ return null;
202
+ }
203
+
204
+ /** Decode a value after its caller removes protocol-specific delimiters.
205
+ * String-typed params stay raw; everything else is JSON-decoded when possible. */
206
+ function decodeToolValue(text: string, schema: Record<string, unknown> | null): unknown {
207
+ if (!text) return "";
208
+ const types = schemaTypeNames(schema);
209
+ if (types.size && [...types].every((t) => t === "string" || t === "null")) return text;
210
+ try {
211
+ return JSON.parse(text);
212
+ } catch {
213
+ return text;
214
+ }
215
+ }
216
+
217
+ // ---- Self-healing repair layer ---------------------------------------
218
+ // (`02d723a:docs/design/web-chat-redesign.md` Appendix A, beat-matrix Axis 7)
219
+ //
220
+ // pi-ai (node_modules/@earendil-works/pi-ai/dist/utils/json-parse.js)
221
+ // already ships a generic JSON repair pass for cloud-model streaming
222
+ // tool calls: `repairJson`/`parseJsonWithRepair` fix invalid backslash
223
+ // escapes and raw control characters inside strings, and
224
+ // `parseStreamingJson` falls back to the `partial-json` package for
225
+ // truncated mid-stream JSON. Its `validateToolArguments`
226
+ // (utils/validation.js) additionally coerces JSON-Schema primitive
227
+ // TYPES (string "3" -> number 3, etc.) once a call is already
228
+ // structurally valid JSON. None of that touches the STRUCTURAL
229
+ // malformations small local models actually produce here: single
230
+ // quotes instead of double, trailing commas, unbalanced/truncated
231
+ // braces, a whole call wrapped in a markdown ```json fence, stray
232
+ // prose ("Sure, here's the call:") glued in front of the JSON, or the
233
+ // function name landing inside `arguments` instead of the envelope
234
+ // (toolCallFromPayload above already handles that last one for
235
+ // already-valid JSON — the repair pass below reuses it once the text
236
+ // is coerced to valid JSON). This layer is therefore purely additive
237
+ // to pi-ai's, applied ONLY when strict parsing has already failed, and
238
+ // every repair is tagged on the returned ParsedToolCall (`repaired`,
239
+ // `repairs`) rather than silently swallowed — see ParsedToolCall above.
240
+ export type RepairKind =
241
+ | "fenced_wrapper" // ```json ... ``` or ``` ... ``` wrapper stripped
242
+ | "prose_prefix" // leading non-JSON prose trimmed to the first { or [
243
+ | "single_quotes" // 'like this' -> "like this" (keys and string values)
244
+ | "trailing_commas" // {"a":1,} / [1,2,] -> valid JSON
245
+ | "unbalanced_braces" // truncated generation: missing closing }/]/" appended
246
+ | "python_literals" // True/False/None -> true/false/null
247
+ | "name_in_args"; // function name found inside `arguments` instead of the envelope
248
+
249
+ /** Strip a ```json fenced block (or bare ```) wrapping the whole payload. */
250
+ function stripFence(text: string): { out: string; changed: boolean } {
251
+ const m = /^\s*```(?:json)?\s*\n?([\s\S]*?)\n?```\s*$/i.exec(text.trim());
252
+ if (!m) return { out: text, changed: false };
253
+ return { out: m[1] ?? "", changed: true };
254
+ }
255
+
256
+ /** Trim stray prose before the first `{`/`[` and after the matching
257
+ * document's last `}`/`]` (a small model narrating "Sure, here's the
258
+ * call: {...}" or trailing chatter after the JSON). Leaves the text
259
+ * alone if it doesn't start with prose (no-op safe to always run). */
260
+ function stripProsePrefix(text: string): { out: string; changed: boolean } {
261
+ const trimmed = text.trim();
262
+ const firstBrace = trimmed.search(/[{[]/);
263
+ if (firstBrace <= 0) return { out: trimmed, changed: false };
264
+ // Only strip when what precedes genuinely looks like prose (no braces
265
+ // of its own) — never chew into legitimate leading whitespace-only JSON.
266
+ const prefix = trimmed.slice(0, firstBrace);
267
+ if (/[{}[\]]/.test(prefix)) return { out: trimmed, changed: false };
268
+ const lastClose = Math.max(trimmed.lastIndexOf("}"), trimmed.lastIndexOf("]"));
269
+ const out = lastClose >= firstBrace ? trimmed.slice(firstBrace, lastClose + 1) : trimmed.slice(firstBrace);
270
+ return { out, changed: true };
271
+ }
272
+
273
+ /** Convert Python-ish literals (True/False/None) used outside string
274
+ * literals into JSON ones. Scans char-by-char tracking string state so
275
+ * a value like "Nonetheless" or a string containing "True" is untouched. */
276
+ function pythonLiteralsToJson(text: string): { out: string; changed: boolean } {
277
+ let out = "";
278
+ let changed = false;
279
+ let inString: '"' | "'" | null = null;
280
+ for (let i = 0; i < text.length; i++) {
281
+ const ch = text[i]!;
282
+ if (inString) {
283
+ out += ch;
284
+ if (ch === "\\") { out += text[++i] ?? ""; continue; }
285
+ if (ch === inString) inString = null;
286
+ continue;
287
+ }
288
+ if (ch === '"' || ch === "'") { inString = ch; out += ch; continue; }
289
+ const rest = text.slice(i);
290
+ const wordMatch = /^(True|False|None)\b/.exec(rest);
291
+ if (wordMatch) {
292
+ const replacement = wordMatch[1] === "True" ? "true" : wordMatch[1] === "False" ? "false" : "null";
293
+ out += replacement;
294
+ i += wordMatch[1]!.length - 1;
295
+ changed = true;
296
+ continue;
297
+ }
298
+ out += ch;
299
+ }
300
+ return { out, changed };
301
+ }
302
+
303
+ /** Convert single-quoted JSON (keys and string values) to double-quoted,
304
+ * tracking string state so an apostrophe INSIDE a double-quoted string
305
+ * (e.g. "don't") is left alone. Only fires when the text has no double
306
+ * quotes at all in "quote position" — i.e. looks like it was written
307
+ * with single quotes throughout — to avoid mangling a string that
308
+ * legitimately contains a single quote inside double-quoted JSON. */
309
+ function singleQuotesToDouble(text: string): { out: string; changed: boolean } {
310
+ if (!/'/.test(text)) return { out: text, changed: false };
311
+ // Heuristic gate: real JSON never opens a key/value with a bare `'`.
312
+ // If the text already parses with plain JSON.parse it wouldn't have
313
+ // reached here, so any '"'-delimited string is assumed intentional and
314
+ // left byte-for-byte; only '...'-delimited runs are retargeted.
315
+ let out = "";
316
+ let changed = false;
317
+ let inDouble = false;
318
+ let inSingle = false;
319
+ for (let i = 0; i < text.length; i++) {
320
+ const ch = text[i]!;
321
+ if (inDouble) {
322
+ out += ch;
323
+ if (ch === "\\") { out += text[++i] ?? ""; continue; }
324
+ if (ch === '"') inDouble = false;
325
+ continue;
326
+ }
327
+ if (inSingle) {
328
+ if (ch === "\\") { out += ch + (text[++i] ?? ""); continue; }
329
+ if (ch === "'") { out += '"'; inSingle = false; continue; }
330
+ if (ch === '"') { out += '\\"'; continue; } // embedded " must be escaped now
331
+ out += ch;
332
+ continue;
333
+ }
334
+ if (ch === '"') { inDouble = true; out += ch; continue; }
335
+ if (ch === "'") { inSingle = true; changed = true; out += '"'; continue; }
336
+ out += ch;
337
+ }
338
+ return { out, changed };
339
+ }
340
+
341
+ /** Remove trailing commas before a closing `}`/`]` (whitespace/newlines
342
+ * between the comma and the close are preserved otherwise). */
343
+ function stripTrailingCommas(text: string): { out: string; changed: boolean } {
344
+ let changed = false;
345
+ const out = text.replace(/,(\s*[}\]])/g, (_m, close) => { changed = true; return close; });
346
+ return { out, changed };
347
+ }
348
+
349
+ /** Append the minimal closers to balance braces/brackets/an unterminated
350
+ * string — a small model that got truncated (or just stopped early)
351
+ * mid-object. Walks once tracking a stack of open delimiters and string
352
+ * state; on EOF, closes the open string (if any) then unwinds the stack
353
+ * in reverse. No-op (unchanged) when already balanced. */
354
+ function closeUnbalanced(text: string): { out: string; changed: boolean } {
355
+ const stack: ("}" | "]")[] = [];
356
+ let inString = false;
357
+ for (let i = 0; i < text.length; i++) {
358
+ const ch = text[i]!;
359
+ if (inString) {
360
+ if (ch === "\\") { i++; continue; }
361
+ if (ch === '"') inString = false;
362
+ continue;
363
+ }
364
+ if (ch === '"') { inString = true; continue; }
365
+ if (ch === "{") stack.push("}");
366
+ else if (ch === "[") stack.push("]");
367
+ else if (ch === "}" || ch === "]") stack.pop();
368
+ }
369
+ if (!inString && stack.length === 0) return { out: text, changed: false };
370
+ let out = text;
371
+ if (inString) out += '"';
372
+ out += stack.reverse().join("");
373
+ return { out, changed: true };
374
+ }
375
+
376
+ /** Try strict JSON.parse first (pi-ai's own escape-level repair already
377
+ * ran upstream where applicable — see the comment above); on failure,
378
+ * apply the structural repair passes in order, re-testing after each so
379
+ * only the ones that were actually needed get tagged. Returns null (no
380
+ * throw) when nothing recovers a parseable object — callers decide how
381
+ * to surface that. */
382
+ function repairJsonToolText(text: string): { value: unknown; repairs: RepairKind[] } | null {
383
+ const repairs: RepairKind[] = [];
384
+ let candidate = text;
385
+
386
+ const fence = stripFence(candidate);
387
+ if (fence.changed) { candidate = fence.out; repairs.push("fenced_wrapper"); }
388
+
389
+ const prose = stripProsePrefix(candidate);
390
+ if (prose.changed) { candidate = prose.out; repairs.push("prose_prefix"); }
391
+
392
+ const tryParse = (s: string): unknown | undefined => {
393
+ try { return JSON.parse(s); } catch { return undefined; }
394
+ };
395
+
396
+ let parsed = tryParse(candidate);
397
+ if (parsed !== undefined) return { value: parsed, repairs };
398
+
399
+ const literals = pythonLiteralsToJson(candidate);
400
+ if (literals.changed) {
401
+ parsed = tryParse(literals.out);
402
+ if (parsed !== undefined) return { value: parsed, repairs: [...repairs, "python_literals"] };
403
+ candidate = literals.out;
404
+ }
405
+
406
+ const commas = stripTrailingCommas(candidate);
407
+ if (commas.changed) {
408
+ parsed = tryParse(commas.out);
409
+ if (parsed !== undefined) return { value: parsed, repairs: [...repairs, "trailing_commas"] };
410
+ candidate = commas.out;
411
+ }
412
+
413
+ const closed = closeUnbalanced(candidate);
414
+ if (closed.changed) {
415
+ parsed = tryParse(closed.out);
416
+ if (parsed !== undefined) return { value: parsed, repairs: [...repairs, "unbalanced_braces"] };
417
+ candidate = closed.out;
418
+ }
419
+
420
+ const quotes = singleQuotesToDouble(candidate);
421
+ if (quotes.changed) {
422
+ parsed = tryParse(quotes.out);
423
+ if (parsed !== undefined) return { value: parsed, repairs: [...repairs, "single_quotes"] };
424
+ candidate = quotes.out;
425
+ // Single-quote conversion can itself leave a trailing comma or an
426
+ // unbalanced document (e.g. `{'a': 1,}` truncated) — retry those two
427
+ // passes once more on the requoted text before giving up.
428
+ const commas2 = stripTrailingCommas(candidate);
429
+ if (commas2.changed) {
430
+ parsed = tryParse(commas2.out);
431
+ if (parsed !== undefined) return { value: parsed, repairs: [...repairs, "single_quotes", "trailing_commas"] };
432
+ candidate = commas2.out;
433
+ }
434
+ const closed2 = closeUnbalanced(candidate);
435
+ if (closed2.changed) {
436
+ parsed = tryParse(closed2.out);
437
+ if (parsed !== undefined) return { value: parsed, repairs: [...repairs, "single_quotes", "unbalanced_braces"] };
438
+ }
439
+ }
440
+
441
+ return null;
442
+ }
443
+
444
+ /** Lenient JSON.parse used for argument STRINGS nested inside an already
445
+ * JSON-valid envelope (e.g. `"arguments": "{'a': 1}"`, a small model
446
+ * double-encoding with the wrong quote style). Falls back through the
447
+ * same repair passes; throws (not silently {}) when nothing recovers,
448
+ * same contract as JSON.parse for its callers. */
449
+ function parseJsonLoose(text: string): unknown {
450
+ try {
451
+ return JSON.parse(text);
452
+ } catch (err) {
453
+ const repaired = repairJsonToolText(text);
454
+ if (repaired) return repaired.value;
455
+ throw err;
456
+ }
457
+ }
458
+
459
+ /** Build a ParsedToolCall from an already-parsed JSON payload (object or
460
+ * {function:{name,arguments}} envelope). Shared by the strict and
461
+ * repaired paths so name/arguments extraction logic lives once. Handles
462
+ * the "name-in-args confusion" malformation: a small model sometimes
463
+ * emits the function name INSIDE the arguments object instead of (or in
464
+ * addition to) the envelope's own `name` field — e.g.
465
+ * {"arguments":{"name":"read","path":"x"}} with no top-level name at
466
+ * all. When the envelope has no usable name, an args-embedded `name` (or
467
+ * `function`) string is promoted and stripped out of the arguments. */
468
+ function toolCallFromPayload(
469
+ payload: unknown,
470
+ ): { name: string; arguments: Record<string, unknown>; nameInArgs: boolean } | null {
471
+ if (!payload || typeof payload !== "object") return null;
472
+ const obj = payload as Record<string, any>;
473
+ const fn = obj.function && typeof obj.function === "object" ? obj.function : obj;
474
+ let name = typeof fn.name === "string" ? fn.name.trim() : "";
475
+ let args = fn.arguments ?? obj.arguments ?? {};
476
+ if (typeof args === "string") {
477
+ const parsed = args.trim() ? parseJsonLoose(args) : {};
478
+ args = parsed && typeof parsed === "object" && !Array.isArray(parsed) ? parsed : {};
479
+ }
480
+ if (!args || typeof args !== "object" || Array.isArray(args)) return null;
481
+ const argsObj = { ...(args as Record<string, unknown>) };
482
+ let nameInArgs = false;
483
+ if (!name) {
484
+ const embedded = argsObj.name ?? argsObj.function;
485
+ if (typeof embedded === "string" && embedded.trim()) {
486
+ name = embedded.trim();
487
+ delete argsObj.name;
488
+ delete argsObj.function;
489
+ nameInArgs = true;
490
+ }
491
+ }
492
+ if (!name) return null;
493
+ return { name, arguments: argsObj, nameInArgs };
494
+ }
495
+
496
+ /** Shape guard for the wrapper-less last-resort repair path only (see its
497
+ * call site in parseGeneratedToolCalls). A `<tool_call>`/`<function...>`
498
+ * tag already proves the model intended a tool call, so toolCallFromPayload
499
+ * can be lenient there; a bare JSON object with NO such tag is ambiguous —
500
+ * it could just as easily be an ordinary conversational reply that happens
501
+ * to be JSON-shaped (e.g. `{"name": "read", "age": 42}` answering "give me
502
+ * an example record"). Require an actual envelope shape: a `function`
503
+ * wrapper object, or a top-level/nested `arguments` object (including the
504
+ * name-in-args case, where the name lives inside that same `arguments`
505
+ * object) — never just a bare `name`/`function` string sitting next to
506
+ * unrelated fields. */
507
+ function looksLikeToolEnvelope(payload: unknown): boolean {
508
+ if (!payload || typeof payload !== "object" || Array.isArray(payload)) return false;
509
+ const obj = payload as Record<string, unknown>;
510
+ if (obj.function && typeof obj.function === "object" && !Array.isArray(obj.function)) return true;
511
+ const args = obj.arguments;
512
+ if (args && typeof args === "object" && !Array.isArray(args)) return true;
513
+ if (typeof args === "string") return true; // double-encoded arguments string
514
+ return false;
515
+ }
516
+
517
+ function parseJsonToolCall(block: string): ParsedToolCall | null {
518
+ let payload: unknown;
519
+ let repairs: RepairKind[] = [];
520
+ try {
521
+ payload = JSON.parse(block);
522
+ } catch {
523
+ // Strict JSON.parse failed — this is the ONLY point the syntax-repair
524
+ // pass engages (never attempted when the model emitted valid JSON). A
525
+ // repair failure here returns null (not throw): parseGeneratedToolCalls
526
+ // falls through to the XML parsers next, which is the right behavior
527
+ // for a `<tool_call>` block that isn't actually JSON-shaped at all.
528
+ const repaired = repairJsonToolText(block);
529
+ if (!repaired) return null;
530
+ payload = repaired.value;
531
+ repairs = repaired.repairs;
532
+ }
533
+ if (!payload || typeof payload !== "object") throw new Error("JSON tool_call payload must be an object");
534
+ const result = toolCallFromPayload(payload);
535
+ if (!result) throw new Error("tool_call arguments must be a JSON object");
536
+ if (result.nameInArgs) repairs = [...repairs, "name_in_args"];
537
+ const { nameInArgs: _nameInArgs, ...call } = result;
538
+ return repairs.length ? { ...call, repaired: true, repairs } : call;
539
+ }
540
+
541
+ function parseXmlEqualsToolCall(block: string, tools: ToolSpec[], strict = false): ParsedToolCall | null {
542
+ const match = XML_FUNCTION_EQUALS_RE.exec(block);
543
+ XML_FUNCTION_EQUALS_RE.lastIndex = 0;
544
+ if (!match) return null;
545
+ if (strict && (match.index !== 0 || match[0].length !== block.length)) return null;
546
+ const name = match[1]!.trim();
547
+ const body = match[2]!;
548
+ const args: Record<string, unknown> = {};
549
+ let cursor = 0;
550
+ for (const param of body.matchAll(XML_PARAMETER_EQUALS_RE)) {
551
+ if (strict && body.slice(cursor, param.index).trim()) return null;
552
+ const key = param[1]!.trim();
553
+ if (!key) throw new Error(`tool '${name}' contains an empty parameter name`);
554
+ if (strict && Object.hasOwn(args, key)) return null;
555
+ // Qwen's format carries raw text, not XML entity-encoded text. Match
556
+ // mlx_lm.tool_parsers.qwen3_coder: remove one framing newline at each
557
+ // end, preserving indentation, literal entities and JSON string escapes.
558
+ const value = param[2]!.replace(/^\n|\n$/g, "");
559
+ args[key] = decodeToolValue(value, toolParameterSchema(tools, name, key));
560
+ cursor = param.index + param[0].length;
561
+ }
562
+ XML_PARAMETER_EQUALS_RE.lastIndex = 0;
563
+ if (strict && body.slice(cursor).trim()) return null;
564
+ return { name, arguments: args };
565
+ }
566
+
567
+ /** One complete envelope, with full consumption and no syntax repair. Used
568
+ * to establish structural fill boundaries, never to verify model predictions. */
569
+ export function parseStrictToolCall(text: string, tools: ToolSpec[]): ParsedToolCall | null {
570
+ const open = "<tool_call>", close = "</tool_call>";
571
+ if (!text.startsWith(open) || !text.endsWith(close)) return null;
572
+ const body = text.slice(open.length, -close.length).trim();
573
+ if (body.includes(open) || body.includes(close)) return null;
574
+ try {
575
+ let call: ParsedToolCall | null;
576
+ if (body.startsWith("{")) {
577
+ // JSON.parse must succeed before the existing payload interpretation.
578
+ JSON.parse(body);
579
+ call = parseJsonToolCall(body);
580
+ } else {
581
+ call = parseXmlEqualsToolCall(body, tools, true) ?? parseGlm52ToolCall(body, tools, true);
582
+ }
583
+ if (!call || call.repaired || !tools.some(tool => toolSpecName(tool) === call.name)) return null;
584
+ return call;
585
+ } catch { return null; }
586
+ }
587
+
588
+ /** A textual argument delimiter is complete only after its closing bracket. */
589
+ export function hasCompleteToolValueDelimiter(text: string): boolean {
590
+ return /^(?:<\/parameter>|<\/arg_value>)/.test(text.trimStart());
591
+ }
592
+
593
+ function parseGlm52ToolCall(block: string, tools: ToolSpec[], strict = false): ParsedToolCall | null {
594
+ const firstArg = block.search(GLM52_ARG_KEY_RE);
595
+ const name = (firstArg < 0 ? block : block.slice(0, firstArg)).trim();
596
+ if (!name || !/^[^\s<>]+$/.test(name)) return null;
597
+ if (firstArg < 0) return { name, arguments: {} };
598
+
599
+ const body = block.slice(firstArg);
600
+ const args: Record<string, unknown> = {};
601
+ let cursor = 0;
602
+ let matched = false;
603
+ for (const pair of body.matchAll(GLM52_ARG_PAIR_RE)) {
604
+ if (body.slice(cursor, pair.index).trim()) return null;
605
+ const key = pair[1]!.trim();
606
+ if (!key) throw new Error(`tool '${name}' contains an empty argument name`);
607
+ if (strict && Object.hasOwn(args, key)) return null;
608
+ args[key] = decodeToolValue(decodeXml(pair[2]!), toolParameterSchema(tools, name, key));
609
+ cursor = pair.index + pair[0].length;
610
+ matched = true;
611
+ }
612
+ GLM52_ARG_PAIR_RE.lastIndex = 0;
613
+ if (!matched || body.slice(cursor).trim()) return null;
614
+ return { name, arguments: args };
615
+ }
616
+
617
+ function parseXmlAttrToolCalls(text: string, tools: ToolSpec[]): ParsedToolCall[] {
618
+ const calls: ParsedToolCall[] = [];
619
+ for (const fn of text.matchAll(XML_FUNCTION_ATTR_RE)) {
620
+ const name = fn[1]!.trim();
621
+ const body = fn[2]!;
622
+ const args: Record<string, unknown> = {};
623
+ for (const param of body.matchAll(XML_PARAM_ATTR_RE)) {
624
+ const key = param[1]!.trim();
625
+ if (!key) throw new Error(`tool '${name}' contains an empty parameter name`);
626
+ args[key] = decodeToolValue(decodeXml(param[2]!), toolParameterSchema(tools, name, key));
627
+ }
628
+ XML_PARAM_ATTR_RE.lastIndex = 0;
629
+ calls.push({ name, arguments: args });
630
+ }
631
+ XML_FUNCTION_ATTR_RE.lastIndex = 0;
632
+ return calls;
633
+ }
634
+
635
+ /** OptiQ-style decoded-text tool parsing. Tool markup is parsed only
636
+ * after generation, and only when tools are active; tokenizer ids are
637
+ * model-family-specific and must not be used globally. Supports the
638
+ * OpenAI JSON `<tool_call>...</tool_call>` contract, Qwen-style
639
+ * `<function=name><parameter=...>`, GLM-5.2's
640
+ * `name<arg_key>...<arg_value>...`, and MiniCPM5's native
641
+ * `<function name="..."><param name="...">...` template shape. */
642
+ export function parseGeneratedToolCalls(text: string, tools: ToolSpec[]): ParsedToolCall[] {
643
+ if (!tools.length) return [];
644
+ const known = new Set(tools.map(toolSpecName).filter((n): n is string => !!n));
645
+ const calls: ParsedToolCall[] = [];
646
+ for (const block of text.matchAll(TOOL_CALL_BLOCK_RE)) {
647
+ const body = block[1]!.trim();
648
+ // A valid Qwen parameter may itself contain JSON. Dispatch by the outer
649
+ // format before JSON repair can mistake that value for the tool envelope.
650
+ const parsed = /^<function=/i.test(body)
651
+ ? parseXmlEqualsToolCall(body, tools)
652
+ : parseJsonToolCall(body) ?? parseGlm52ToolCall(body, tools);
653
+ if (!parsed) throw new Error("unsupported tool_call payload format");
654
+ calls.push(parsed);
655
+ }
656
+ TOOL_CALL_BLOCK_RE.lastIndex = 0;
657
+ if (calls.length === 0) calls.push(...parseXmlAttrToolCalls(text, tools));
658
+ // Last-resort repair: no `<tool_call>` tags AND no `<function name=...>`
659
+ // XML matched at all — a small model sometimes drops the wrapper tags
660
+ // entirely and emits just (possibly fenced/prose-prefixed/malformed)
661
+ // JSON. Only attempted when everything else found nothing, so a
662
+ // genuinely tool-free assistant reply never risks being misread as a
663
+ // call. Tagged repaired like every other path here.
664
+ //
665
+ // Unlike the `<tool_call>`-wrapped path, there is no surrounding markup
666
+ // here to prove the model actually intended a tool call — an ordinary
667
+ // conversational reply can itself BE a bare JSON object (e.g. "give me
668
+ // an example record" -> `{"name": "read", "age": 42}`), which happens to
669
+ // collide with a real tool name. Require the recovered payload to
670
+ // actually look like a tool-call envelope (an `arguments`/
671
+ // `function.arguments` object, or a `function` wrapper, or the
672
+ // name-in-args shape) before accepting it — a flat object whose only
673
+ // tool-shaped feature is a `name` key that matches a tool is NOT enough
674
+ // signal on its own.
675
+ if (calls.length === 0) {
676
+ const repaired = repairJsonToolText(text.trim());
677
+ if (repaired && looksLikeToolEnvelope(repaired.value)) {
678
+ const result = toolCallFromPayload(repaired.value);
679
+ if (result) {
680
+ const repairs = result.nameInArgs ? [...repaired.repairs, "name_in_args" as const] : repaired.repairs;
681
+ calls.push({ name: result.name, arguments: result.arguments, repaired: true, repairs });
682
+ }
683
+ }
684
+ }
685
+ for (const c of calls) {
686
+ if (!known.has(c.name)) throw new Error(`unknown tool '${c.name}'`);
687
+ }
688
+ return calls;
689
+ }