doppler-gpu 0.2.2 → 0.4.3

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 (658) hide show
  1. package/CHANGELOG.md +182 -0
  2. package/README.md +20 -8
  3. package/assets/doppler.svg +1601 -0
  4. package/package.json +82 -33
  5. package/src/cli/cli-model-resolution.js +275 -0
  6. package/src/cli/cli-output.js +453 -0
  7. package/src/cli/doppler-cli.js +817 -709
  8. package/src/cli/doppler-quickstart.d.ts +7 -23
  9. package/src/cli/doppler-quickstart.js +20 -4
  10. package/src/cli/doppler-serve.js +1 -1
  11. package/src/cli/runtime-profiles.d.ts +40 -0
  12. package/src/cli/runtime-profiles.js +225 -0
  13. package/src/client/doppler-api.js +1 -6
  14. package/src/client/doppler-registry.d.ts +12 -0
  15. package/src/client/doppler-registry.js +27 -0
  16. package/src/client/doppler-registry.json +79 -5
  17. package/src/client/failure-taxonomy.js +139 -0
  18. package/src/client/fault-injection.js +30 -0
  19. package/src/client/provider.d.ts +151 -0
  20. package/src/client/provider.js +374 -0
  21. package/src/client/receipt.js +156 -0
  22. package/src/client/runtime/index.js +56 -4
  23. package/src/client/runtime/model-manager.d.ts +0 -14
  24. package/src/client/runtime/model-manager.js +2 -42
  25. package/src/client/runtime/model-session.js +33 -0
  26. package/src/client/runtime/model-source.d.ts +13 -1
  27. package/src/client/runtime/model-source.js +128 -2
  28. package/src/client/runtime/node-quickstart-cache.d.ts +22 -0
  29. package/src/client/runtime/node-quickstart-cache.js +353 -0
  30. package/src/client/runtime/source-runtime.js +1 -10
  31. package/src/client/runtime/types.d.ts +0 -60
  32. package/src/client/wrap-pipeline-handle.d.ts +20 -0
  33. package/src/client/wrap-pipeline-handle.js +84 -0
  34. package/src/config/README.md +3 -1
  35. package/src/config/conversion/README.md +23 -157
  36. package/src/config/conversion/embeddinggemma/google-embeddinggemma-300m-q4k-ehf16-af32.json +19 -4
  37. package/src/config/conversion/gemma3/gemma-3-1b-it-f16-af32.json +16 -9
  38. package/src/config/conversion/gemma3/gemma-3-1b-it-q4k-ehf16-af32.json +24 -9
  39. package/src/config/conversion/gemma3/gemma-3-270m-it-q4k-ehf16-af32.json +24 -9
  40. package/src/config/conversion/gemma3/translategemma-4b-1b-enes-q4k-ehf16-af32.json +16 -9
  41. package/src/config/conversion/gemma3/translategemma-4b-it-q4k-ehf16-af32.json +24 -9
  42. package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af16.json +719 -0
  43. package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af32.json +715 -0
  44. package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af16-int4ple.json +589 -0
  45. package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32-int4ple.json +566 -0
  46. package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32.json +368 -62
  47. package/src/config/conversion/gemma4/gemma-4-moe-q4k-ehf16-af32.json +212 -55
  48. package/src/config/conversion/gpt-oss-20b-f16-xmxfp4.json +216 -57
  49. package/src/config/conversion/janus/janus-pro-1b-text-q4k-ehaf16.json +229 -57
  50. package/src/config/conversion/lfm2/lfm2.5-1.2b-instruct-q4k-ehf16-af32.json +17 -10
  51. package/src/config/conversion/qwen3/qwen-3-5-0-8b-q4k-ehaf16.json +75 -33
  52. package/src/config/conversion/qwen3/qwen-3-5-2b-q4k-ehaf16.json +91 -59
  53. package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-eaf16.json +517 -0
  54. package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-ehaf16.json +511 -0
  55. package/src/config/execution-contract-check.js +6 -5
  56. package/src/config/kernel-path-loader.d.ts +8 -2
  57. package/src/config/kernels/kernel-ref-digests.js +125 -122
  58. package/src/config/kernels/registry.d.ts +0 -35
  59. package/src/config/kernels/registry.js +5 -9
  60. package/src/config/kernels/registry.json +615 -368
  61. package/src/config/merge-helpers.d.ts +1 -1
  62. package/src/config/merge-helpers.js +1 -1
  63. package/src/config/merge.js +35 -7
  64. package/src/config/param-categories.d.ts +0 -9
  65. package/src/config/param-categories.js +1 -5
  66. package/src/config/param-validator.d.ts +2 -0
  67. package/src/config/param-validator.js +16 -0
  68. package/src/config/platforms/loader.d.ts +0 -27
  69. package/src/config/platforms/loader.js +0 -30
  70. package/src/config/required-inference-fields-contract-check.js +1 -7
  71. package/src/config/runtime/README.md +28 -0
  72. package/src/config/runtime/experiments/bench/gemma4-bench-q4k.json +32 -0
  73. package/src/config/runtime/experiments/bench/qwen3-bench-q4k-batched.json +2 -2
  74. package/src/config/runtime/experiments/debug/gemma4-e2b-browser-diagnose-q4k.json +85 -0
  75. package/src/config/runtime/platform/metal-apple-q4k.json +19 -3
  76. package/src/config/runtime/profiles/gemma4-31b-f16-activations-probe.json +65 -0
  77. package/src/config/runtime/profiles/gemma4-e2b-baseline-probe.json +41 -0
  78. package/src/config/runtime/profiles/gemma4-e2b-fused-rmsnorm-probe.json +44 -0
  79. package/src/config/runtime/profiles/gemma4-e2b-fused-rmsnorm-test.json +19 -0
  80. package/src/config/runtime/profiles/gemma4-e2b-int4ple-baseline-probe.json +31 -0
  81. package/src/config/runtime/profiles/gemma4-e2b-int4ple-f16a-probe.json +34 -0
  82. package/src/config/runtime/profiles/gemma4-e2b-int4ple-reference-export.json +64 -0
  83. package/src/config/runtime/profiles/gemma4-e2b-int4ple-retain-q4k-probe.json +47 -0
  84. package/src/config/runtime/profiles/gemma4-e2b-prefill-profile.json +35 -0
  85. package/src/config/runtime/profiles/gemma4-e2b-throughput-probe.json +16 -0
  86. package/src/config/runtime/profiles/qwen-3-5-0-8b-prefill-profile.json +35 -0
  87. package/src/config/runtime/profiles/qwen-3-5-0-8b-throughput-probe.json +16 -0
  88. package/src/config/runtime/profiles/qwen3-6-27b-f16-activations-probe.json +62 -0
  89. package/src/config/schema/benchmark.schema.d.ts +2 -0
  90. package/src/config/schema/benchmark.schema.js +1 -0
  91. package/src/config/schema/bridge.schema.d.ts +16 -0
  92. package/src/config/schema/browser-suite-metrics.schema.d.ts +1 -0
  93. package/src/config/schema/browser-suite-metrics.schema.js +2 -0
  94. package/src/config/schema/converter.schema.d.ts +5 -0
  95. package/src/config/schema/converter.schema.js +3 -0
  96. package/src/config/schema/debug.schema.d.ts +22 -1
  97. package/src/config/schema/debug.schema.js +1 -0
  98. package/src/config/schema/distill-training.schema.d.ts +0 -1
  99. package/src/config/schema/distill-training.schema.js +1 -1
  100. package/src/config/schema/doppler.schema.d.ts +21 -3
  101. package/src/config/schema/doppler.schema.js +116 -9
  102. package/src/config/schema/emulation.schema.d.ts +3 -0
  103. package/src/config/schema/emulation.schema.js +16 -16
  104. package/src/config/schema/execution-v1.schema.d.ts +12 -1
  105. package/src/config/schema/execution-v1.schema.js +7 -1
  106. package/src/config/schema/index.d.ts +54 -1
  107. package/src/config/schema/index.js +15 -0
  108. package/src/config/schema/inference.schema.d.ts +3 -3
  109. package/src/config/schema/kernel-path.schema.d.ts +6 -27
  110. package/src/config/schema/kernel-thresholds.schema.d.ts +0 -2
  111. package/src/config/schema/kernel-thresholds.schema.js +3 -7
  112. package/src/config/schema/kvcache.schema.js +0 -9
  113. package/src/config/schema/loading.schema.d.ts +12 -0
  114. package/src/config/schema/loading.schema.js +4 -0
  115. package/src/config/schema/manifest.schema.d.ts +73 -1
  116. package/src/config/schema/manifest.schema.js +48 -2
  117. package/src/config/schema/moe.schema.js +0 -8
  118. package/src/config/schema/program-bundle.schema.d.ts +233 -0
  119. package/src/config/schema/program-bundle.schema.js +393 -0
  120. package/src/config/schema/ul-training.schema.d.ts +0 -1
  121. package/src/config/schema/ul-training.schema.js +1 -1
  122. package/src/config/source-packages/litert/gemma-4-e2b-it.json +36 -15
  123. package/src/config/supported-operations.d.ts +14 -0
  124. package/src/config/supported-operations.js +34 -0
  125. package/src/config/supported-operations.json +178 -0
  126. package/src/config/training-defaults.js +1 -7
  127. package/src/config/transforms/capability-transform-resolver.js +3 -1
  128. package/src/config/transforms/execution-graph-transforms.d.ts +55 -0
  129. package/src/config/transforms/execution-graph-transforms.js +693 -29
  130. package/src/converter/conversion-plan.js +13 -10
  131. package/src/converter/core.d.ts +57 -0
  132. package/src/converter/core.js +582 -13
  133. package/src/converter/parsers/diffusion.js +1 -33
  134. package/src/converter/quantization-info.d.ts +5 -4
  135. package/src/converter/quantization-info.js +39 -3
  136. package/src/converter/quantizer.d.ts +19 -0
  137. package/src/converter/quantizer.js +81 -4
  138. package/src/debug/config.d.ts +0 -4
  139. package/src/debug/config.js +0 -17
  140. package/src/debug/index.d.ts +28 -30
  141. package/src/debug/index.js +0 -18
  142. package/src/debug/stats.d.ts +0 -8
  143. package/src/debug/stats.js +3 -3
  144. package/src/debug/tensor.d.ts +9 -2
  145. package/src/debug/tensor.js +11 -4
  146. package/src/debug/trace.js +2 -15
  147. package/src/errors/doppler-error.js +20 -0
  148. package/src/experimental/browser/quantization.d.ts +5 -0
  149. package/src/experimental/distribution/shard-delivery.js +179 -23
  150. package/src/experimental/training/artifacts.js +1 -14
  151. package/src/experimental/training/checkpoint.js +1 -14
  152. package/src/experimental/training/datasets/token-batch.d.ts +1 -1
  153. package/src/experimental/training/datasets/token-batch.js +2 -2
  154. package/src/experimental/training/distillation/dataset.js +2 -15
  155. package/src/experimental/training/distillation/eval.js +3 -3
  156. package/src/experimental/training/distillation/runtime.js +3 -16
  157. package/src/experimental/training/distillation/student-fixture.js +6 -6
  158. package/src/experimental/training/lora-pipeline.js +1 -14
  159. package/src/experimental/training/objectives/base.d.ts +2 -2
  160. package/src/experimental/training/objectives/cross_entropy.d.ts +2 -2
  161. package/src/experimental/training/objectives/distill_kd.d.ts +2 -2
  162. package/src/experimental/training/objectives/distill_kd.js +2 -2
  163. package/src/experimental/training/objectives/distill_triplet.d.ts +2 -2
  164. package/src/experimental/training/objectives/distill_triplet.js +2 -2
  165. package/src/experimental/training/objectives/ul_stage1_joint.d.ts +2 -2
  166. package/src/experimental/training/objectives/ul_stage2_base.d.ts +2 -2
  167. package/src/experimental/training/objectives/ul_stage2_base.js +1 -1
  168. package/src/experimental/training/operator-artifacts.js +1 -14
  169. package/src/experimental/training/runner.js +1 -14
  170. package/src/experimental/training/workload-packs/registry.json +6 -6
  171. package/src/experimental/training/workloads.js +1 -1
  172. package/src/formats/gguf/types.d.ts +0 -63
  173. package/src/formats/gguf/types.js +58 -38
  174. package/src/formats/litert/types.d.ts +0 -13
  175. package/src/formats/litert/types.js +2 -2
  176. package/src/formats/rdrr/distributed/index.d.ts +34 -0
  177. package/src/formats/rdrr/distributed/index.js +3 -0
  178. package/src/formats/rdrr/distributed/parsing.d.ts +6 -0
  179. package/src/formats/rdrr/distributed/parsing.js +23 -0
  180. package/src/formats/rdrr/distributed/validation.d.ts +23 -0
  181. package/src/formats/rdrr/distributed/validation.js +131 -0
  182. package/src/formats/rdrr/groups.d.ts +1 -13
  183. package/src/formats/rdrr/groups.js +0 -34
  184. package/src/formats/rdrr/index.d.ts +9 -0
  185. package/src/formats/rdrr/index.js +9 -0
  186. package/src/formats/rdrr/integrity.d.ts +51 -0
  187. package/src/formats/rdrr/integrity.js +200 -0
  188. package/src/formats/rdrr/lowerings.d.ts +28 -0
  189. package/src/formats/rdrr/lowerings.js +112 -0
  190. package/src/formats/rdrr/manifest.d.ts +0 -14
  191. package/src/formats/rdrr/manifest.js +5 -58
  192. package/src/formats/rdrr/merkle.d.ts +24 -0
  193. package/src/formats/rdrr/merkle.js +118 -0
  194. package/src/formats/rdrr/parsing.js +4 -0
  195. package/src/formats/rdrr/slice-compiler.d.ts +29 -0
  196. package/src/formats/rdrr/slice-compiler.js +188 -0
  197. package/src/formats/rdrr/source-transform-contract.js +390 -0
  198. package/src/formats/rdrr/storage-descriptor.d.ts +36 -0
  199. package/src/formats/rdrr/storage-descriptor.js +186 -0
  200. package/src/formats/rdrr/tensor-config-validator.d.ts +0 -9
  201. package/src/formats/rdrr/tensor-config-validator.js +0 -30
  202. package/src/formats/rdrr/types.d.ts +115 -1
  203. package/src/formats/rdrr/validation.js +245 -0
  204. package/src/formats/tflite/types.d.ts +8 -31
  205. package/src/formats/tflite/types.js +50 -5
  206. package/src/gpu/command-recorder.d.ts +9 -21
  207. package/src/gpu/command-recorder.js +41 -2
  208. package/src/gpu/device.d.ts +17 -3
  209. package/src/gpu/device.js +20 -4
  210. package/src/gpu/kernel-runtime.d.ts +8 -15
  211. package/src/gpu/kernel-selection-log.d.ts +0 -8
  212. package/src/gpu/kernel-selection-log.js +0 -15
  213. package/src/gpu/kernel-tuner/cache.js +0 -3
  214. package/src/gpu/kernels/README.md +50 -18
  215. package/src/gpu/kernels/activation-static-qdq.d.ts +22 -0
  216. package/src/gpu/kernels/activation-static-qdq.js +69 -0
  217. package/src/gpu/kernels/activation_static_qdq.wgsl +25 -0
  218. package/src/gpu/kernels/attention.d.ts +28 -19
  219. package/src/gpu/kernels/attention.js +443 -9
  220. package/src/gpu/kernels/attention_decode_contiguous_turboquant_f16kv.wgsl +0 -1
  221. package/src/gpu/kernels/attention_decode_contiguous_turboquant_prod_f16kv.wgsl +0 -1
  222. package/src/gpu/kernels/attention_decode_f16.wgsl +1 -1
  223. package/src/gpu/kernels/attention_decode_f16kv.wgsl +1 -1
  224. package/src/gpu/kernels/attention_decode_online_f16.wgsl +32 -29
  225. package/src/gpu/kernels/attention_decode_online_f16kv.wgsl +1 -1
  226. package/src/gpu/kernels/attention_decode_paged_f16kv.wgsl +1 -1
  227. package/src/gpu/kernels/attention_decode_tiered_f16kv.wgsl +1 -1
  228. package/src/gpu/kernels/attention_decode_tiered_turboquant_f16kv.wgsl +0 -1
  229. package/src/gpu/kernels/attention_decode_tiered_turboquant_prod_f16kv.wgsl +0 -1
  230. package/src/gpu/kernels/attention_f16.wgsl +88 -30
  231. package/src/gpu/kernels/attention_f16kv.wgsl +45 -15
  232. package/src/gpu/kernels/attention_head512_f16.wgsl +270 -0
  233. package/src/gpu/kernels/attention_head512_f16kv.wgsl +271 -0
  234. package/src/gpu/kernels/attention_prefill_flash_head256_f16kv.wgsl +306 -0
  235. package/src/gpu/kernels/attention_prefill_flash_reduce.wgsl +101 -0
  236. package/src/gpu/kernels/attention_small_f16.wgsl +22 -20
  237. package/src/gpu/kernels/attention_small_f16kv.wgsl +1 -1
  238. package/src/gpu/kernels/attention_streaming_f16.wgsl +39 -16
  239. package/src/gpu/kernels/attention_streaming_f16kv.wgsl +22 -3
  240. package/src/gpu/kernels/backward/index.d.ts +6 -0
  241. package/src/gpu/kernels/backward/layernorm_backward.wgsl +0 -1
  242. package/src/gpu/kernels/backward/utils.d.ts +35 -0
  243. package/src/gpu/kernels/bf16_to_f32.wgsl +0 -16
  244. package/src/gpu/kernels/bias_add_f16.wgsl +1 -1
  245. package/src/gpu/kernels/cast_f32_to_f16.wgsl +0 -1
  246. package/src/gpu/kernels/codegen/patches/attention_f16.from.attention.diff +19 -215
  247. package/src/gpu/kernels/codegen/patches/attention_f16kv.from.attention.diff +8 -82
  248. package/src/gpu/kernels/codegen/patches/attention_small_f16.from.attention_small.diff +92 -59
  249. package/src/gpu/kernels/codegen/patches/attention_streaming_f16.from.attention_streaming.diff +2 -85
  250. package/src/gpu/kernels/codegen/patches/gelu_f16.from.gelu.diff +27 -24
  251. package/src/gpu/kernels/codegen/patches/residual_f16.from.residual.diff +1 -1
  252. package/src/gpu/kernels/codegen/patches/rmsnorm_f16.from.rmsnorm.diff +249 -179
  253. package/src/gpu/kernels/codegen/patches/rope_f16.from.rope.diff +133 -235
  254. package/src/gpu/kernels/codegen/patches/sample_f16.from.sample.diff +119 -90
  255. package/src/gpu/kernels/codegen/wgsl-variants.js +142 -8
  256. package/src/gpu/kernels/constants.d.ts +9 -48
  257. package/src/gpu/kernels/constants.js +0 -23
  258. package/src/gpu/kernels/conv2d.d.ts +0 -4
  259. package/src/gpu/kernels/conv2d_f16.wgsl +1 -1
  260. package/src/gpu/kernels/dequant.d.ts +0 -16
  261. package/src/gpu/kernels/dequant.js +1 -71
  262. package/src/gpu/kernels/dequant_mxfp4_expert_f16.wgsl +1 -1
  263. package/src/gpu/kernels/dequant_shared.wgsl +1 -55
  264. package/src/gpu/kernels/dispatch.d.ts +12 -97
  265. package/src/gpu/kernels/dispatch.js +0 -117
  266. package/src/gpu/kernels/energy_eval_f16.wgsl +1 -1
  267. package/src/gpu/kernels/energy_quintel_grad_f16.wgsl +1 -1
  268. package/src/gpu/kernels/energy_quintel_reduce_f16.wgsl +1 -1
  269. package/src/gpu/kernels/energy_quintel_update_f16.wgsl +1 -1
  270. package/src/gpu/kernels/energy_update_f16.wgsl +1 -1
  271. package/src/gpu/kernels/fused-gate-up-gelu.js +121 -0
  272. package/src/gpu/kernels/fused_ffn_f16.wgsl +1 -1
  273. package/src/gpu/kernels/fused_ffn_q4k.wgsl +0 -19
  274. package/src/gpu/kernels/fused_ffn_q4k_f16.wgsl +0 -18
  275. package/src/gpu/kernels/fused_gate_up_gelu_gated_f16.wgsl +124 -0
  276. package/src/gpu/kernels/fused_matmul_q4.wgsl +23 -439
  277. package/src/gpu/kernels/fused_matmul_q4_batched_f16.wgsl +1 -1
  278. package/src/gpu/kernels/fused_matmul_q4_batched_f16a.wgsl +0 -1
  279. package/src/gpu/kernels/fused_matmul_q4_batched_f16acc_f16a.wgsl +158 -0
  280. package/src/gpu/kernels/fused_matmul_q4_f16a.wgsl +0 -1
  281. package/src/gpu/kernels/fused_matmul_q4_multicol_f16.wgsl +0 -1
  282. package/src/gpu/kernels/fused_matmul_q4_multicol_f16a.wgsl +32 -33
  283. package/src/gpu/kernels/fused_matmul_q4_multicol_f16a_f32acc.wgsl +209 -0
  284. package/src/gpu/kernels/fused_matmul_q4_prefill_tiled_f16.wgsl +261 -0
  285. package/src/gpu/kernels/fused_matmul_q4_widetile.wgsl +164 -0
  286. package/src/gpu/kernels/fused_matmul_q4_widetile_f16.wgsl +198 -0
  287. package/src/gpu/kernels/fused_matmul_q4_widetile_f16a.wgsl +161 -0
  288. package/src/gpu/kernels/fused_matmul_q4_widetile_residual.wgsl +173 -0
  289. package/src/gpu/kernels/fused_matmul_rmsnorm_f16.wgsl +1 -1
  290. package/src/gpu/kernels/fused_rmsnorm_q4_widetile.wgsl +271 -0
  291. package/src/gpu/kernels/fused_swiglu_f16.wgsl +1 -1
  292. package/src/gpu/kernels/gather.d.ts +0 -7
  293. package/src/gpu/kernels/gather_f16.wgsl +1 -1
  294. package/src/gpu/kernels/gelu_f16.wgsl +14 -15
  295. package/src/gpu/kernels/groupnorm.d.ts +0 -4
  296. package/src/gpu/kernels/groupnorm_apply_f16.wgsl +1 -1
  297. package/src/gpu/kernels/groupnorm_stats_f16.wgsl +1 -1
  298. package/src/gpu/kernels/index.d.ts +13 -5
  299. package/src/gpu/kernels/index.js +9 -4
  300. package/src/gpu/kernels/kv_quantize_turboquant.wgsl +0 -2
  301. package/src/gpu/kernels/kv_quantize_turboquant_prod.wgsl +0 -2
  302. package/src/gpu/kernels/layernorm.d.ts +0 -4
  303. package/src/gpu/kernels/layernorm_f16.wgsl +1 -1
  304. package/src/gpu/kernels/{sana_linear_attention.d.ts → linear_attention.d.ts} +5 -5
  305. package/src/gpu/kernels/{sana_linear_attention.js → linear_attention.js} +16 -16
  306. package/src/gpu/kernels/logit-merge.d.ts +0 -24
  307. package/src/gpu/kernels/logit-merge.js +1 -17
  308. package/src/gpu/kernels/matmul-dispatch.js +7 -1
  309. package/src/gpu/kernels/matmul-selection.js +68 -1
  310. package/src/gpu/kernels/matmul.d.ts +0 -28
  311. package/src/gpu/kernels/matmul.js +101 -7
  312. package/src/gpu/kernels/matmul_f16_tiled.wgsl +33 -34
  313. package/src/gpu/kernels/matmul_gemv_residual_f16.wgsl +1 -1
  314. package/src/gpu/kernels/matmul_gemv_subgroup.wgsl +0 -1
  315. package/src/gpu/kernels/matmul_gemv_subgroup_f16a.wgsl +24 -25
  316. package/src/gpu/kernels/modulate_f16.wgsl +1 -1
  317. package/src/gpu/kernels/moe.d.ts +0 -56
  318. package/src/gpu/kernels/moe.js +0 -15
  319. package/src/gpu/kernels/moe_gather_f16.wgsl +1 -1
  320. package/src/gpu/kernels/pixel_shuffle_f16.wgsl +1 -1
  321. package/src/gpu/kernels/residual.d.ts +0 -12
  322. package/src/gpu/kernels/residual_f16.wgsl +2 -2
  323. package/src/gpu/kernels/residual_f16_vec4.wgsl +5 -5
  324. package/src/gpu/kernels/rmsnorm.d.ts +0 -8
  325. package/src/gpu/kernels/rmsnorm.js +17 -1
  326. package/src/gpu/kernels/rmsnorm.wgsl +5 -6
  327. package/src/gpu/kernels/rmsnorm_f16.wgsl +65 -39
  328. package/src/gpu/kernels/rope.d.ts +1 -5
  329. package/src/gpu/kernels/rope.js +12 -2
  330. package/src/gpu/kernels/rope.wgsl +33 -19
  331. package/src/gpu/kernels/rope_f16.wgsl +39 -25
  332. package/src/gpu/kernels/sample.d.ts +0 -5
  333. package/src/gpu/kernels/sample.wgsl +7 -6
  334. package/src/gpu/kernels/sample_f16.wgsl +28 -30
  335. package/src/gpu/kernels/scatter_add_dynamic_f16.wgsl +1 -1
  336. package/src/gpu/kernels/shader-cache.d.ts +13 -0
  337. package/src/gpu/kernels/shader-cache.js +27 -0
  338. package/src/gpu/kernels/silu_f16.wgsl +1 -1
  339. package/src/gpu/kernels/softmax.wgsl +4 -144
  340. package/src/gpu/kernels/split_qg_f16.wgsl +1 -1
  341. package/src/gpu/kernels/split_qkv_f16.wgsl +1 -1
  342. package/src/gpu/kernels/topk_f16.wgsl +1 -1
  343. package/src/gpu/kernels/turboquant-codebook.js +11 -40
  344. package/src/gpu/kernels/uniform-utils.d.ts +18 -0
  345. package/src/gpu/kernels/upsample2d.d.ts +0 -4
  346. package/src/gpu/kernels/upsample2d_f16.wgsl +1 -1
  347. package/src/gpu/kernels/utils.d.ts +28 -0
  348. package/src/gpu/kernels/utils.js +41 -0
  349. package/src/gpu/perf-guards.d.ts +0 -12
  350. package/src/gpu/perf-guards.js +0 -62
  351. package/src/gpu/profiler.d.ts +0 -21
  352. package/src/gpu/profiler.js +0 -34
  353. package/src/gpu/submit-tracker.d.ts +0 -57
  354. package/src/gpu/submit-tracker.js +3 -62
  355. package/src/gpu/tensor.d.ts +0 -23
  356. package/src/gpu/tensor.js +0 -41
  357. package/src/gpu/uniform-cache.d.ts +0 -11
  358. package/src/gpu/uniform-cache.js +0 -23
  359. package/src/index.d.ts +1 -0
  360. package/src/index.js +1 -0
  361. package/src/inference/browser-harness-model-helpers.d.ts +0 -9
  362. package/src/inference/browser-harness-model-helpers.js +22 -13
  363. package/src/inference/browser-harness-runtime-helpers.d.ts +0 -5
  364. package/src/inference/browser-harness-runtime-helpers.js +2 -2
  365. package/src/inference/browser-harness-suite-helpers.d.ts +4 -1
  366. package/src/inference/browser-harness-suite-helpers.js +87 -1
  367. package/src/inference/browser-harness-text-helpers.d.ts +13 -0
  368. package/src/inference/browser-harness-text-helpers.js +267 -18
  369. package/src/inference/browser-harness.d.ts +0 -29
  370. package/src/inference/browser-harness.js +271 -71
  371. package/src/inference/kv-cache/base.d.ts +6 -0
  372. package/src/inference/kv-cache/base.js +16 -1
  373. package/src/inference/kv-cache/mixed-geometry.d.ts +1 -0
  374. package/src/inference/kv-cache/mixed-geometry.js +12 -2
  375. package/src/inference/kv-cache/quantized.js +2 -0
  376. package/src/inference/kv-cache/tiered.js +7 -0
  377. package/src/inference/kv-cache/types.d.ts +2 -0
  378. package/src/inference/multi-model-network.d.ts +1 -1
  379. package/src/inference/multi-model-network.js +1 -2
  380. package/src/inference/network-evolution.d.ts +0 -12
  381. package/src/inference/network-evolution.js +2 -2
  382. package/src/inference/pipelines/context.js +21 -2
  383. package/src/inference/pipelines/diffusion/init.js +2 -2
  384. package/src/inference/pipelines/diffusion/pipeline.js +41 -96
  385. package/src/inference/pipelines/diffusion/text-encoder-gpu.d.ts +0 -5
  386. package/src/inference/pipelines/diffusion/text-encoder-gpu.js +0 -283
  387. package/src/inference/pipelines/diffusion/vae.js +4 -4
  388. package/src/inference/pipelines/energy/quintel.d.ts +15 -0
  389. package/src/inference/pipelines/structured/index.d.ts +2 -0
  390. package/src/inference/pipelines/structured/index.js +2 -0
  391. package/src/inference/pipelines/structured/json-grammar-mask.js +143 -0
  392. package/src/inference/pipelines/structured/tsv-grammar-mask.js +166 -0
  393. package/src/inference/pipelines/text/attention/attn-config.d.ts +0 -5
  394. package/src/inference/pipelines/text/attention/attn-config.js +1 -1
  395. package/src/inference/pipelines/text/attention/projections.d.ts +2 -0
  396. package/src/inference/pipelines/text/attention/projections.js +70 -23
  397. package/src/inference/pipelines/text/attention/record.js +118 -5
  398. package/src/inference/pipelines/text/attention/rmsnorm-fusion-gate.d.ts +8 -0
  399. package/src/inference/pipelines/text/attention/rmsnorm-fusion-gate.js +19 -0
  400. package/src/inference/pipelines/text/attention/run.js +97 -6
  401. package/src/inference/pipelines/text/attention/types.d.ts +2 -0
  402. package/src/inference/pipelines/text/attention/types.js +9 -4
  403. package/src/inference/pipelines/text/chat-format.d.ts +0 -10
  404. package/src/inference/pipelines/text/chat-format.js +1 -8
  405. package/src/inference/pipelines/text/config.d.ts +3 -25
  406. package/src/inference/pipelines/text/config.js +105 -3
  407. package/src/inference/pipelines/text/debug-utils/utils.js +2 -23
  408. package/src/inference/pipelines/text/drift-policy.js +1 -1
  409. package/src/inference/pipelines/text/embed.d.ts +23 -27
  410. package/src/inference/pipelines/text/embed.js +10 -43
  411. package/src/inference/pipelines/text/execution-plan.d.ts +5 -3
  412. package/src/inference/pipelines/text/execution-plan.js +15 -2
  413. package/src/inference/pipelines/text/execution-runtime-builders.d.ts +0 -4
  414. package/src/inference/pipelines/text/execution-runtime-builders.js +0 -11
  415. package/src/inference/pipelines/text/execution-v1.d.ts +18 -0
  416. package/src/inference/pipelines/text/execution-v1.js +94 -25
  417. package/src/inference/pipelines/text/ffn/dense.js +180 -3
  418. package/src/inference/pipelines/text/ffn/standard.js +40 -13
  419. package/src/inference/pipelines/text/generator-decode-policy.js +243 -0
  420. package/src/inference/pipelines/text/generator-helpers.d.ts +6 -2
  421. package/src/inference/pipelines/text/generator-helpers.js +20 -2
  422. package/src/inference/pipelines/text/generator-prefill-helpers.js +293 -0
  423. package/src/inference/pipelines/text/generator-runtime.d.ts +0 -7
  424. package/src/inference/pipelines/text/generator-runtime.js +8 -6
  425. package/src/inference/pipelines/text/generator-steps.d.ts +22 -7
  426. package/src/inference/pipelines/text/generator-steps.js +23 -9
  427. package/src/inference/pipelines/text/generator.d.ts +0 -6
  428. package/src/inference/pipelines/text/generator.js +254 -530
  429. package/src/inference/pipelines/text/init-chat-templates.js +86 -0
  430. package/src/inference/pipelines/text/init-qkv-fusion.js +201 -0
  431. package/src/inference/pipelines/text/init-rope.js +408 -0
  432. package/src/inference/pipelines/text/init.d.ts +8 -2
  433. package/src/inference/pipelines/text/init.js +190 -710
  434. package/src/inference/pipelines/text/kernel-trace.d.ts +0 -33
  435. package/src/inference/pipelines/text/kernel-trace.js +0 -33
  436. package/src/inference/pipelines/text/layer-plan-gpu.js +542 -0
  437. package/src/inference/pipelines/text/layer-plan.d.ts +0 -5
  438. package/src/inference/pipelines/text/layer-plan.js +1 -1
  439. package/src/inference/pipelines/text/layer.d.ts +70 -8
  440. package/src/inference/pipelines/text/layer.js +105 -537
  441. package/src/inference/pipelines/text/linear-attention.d.ts +2 -1
  442. package/src/inference/pipelines/text/load-errors.d.ts +21 -0
  443. package/src/inference/pipelines/text/load-errors.js +59 -0
  444. package/src/inference/pipelines/text/logits/cpu.js +2 -20
  445. package/src/inference/pipelines/text/logits/gpu.d.ts +5 -4
  446. package/src/inference/pipelines/text/logits/gpu.js +148 -11
  447. package/src/inference/pipelines/text/logits/index.d.ts +1 -1
  448. package/src/inference/pipelines/text/model-load.d.ts +5 -4
  449. package/src/inference/pipelines/text/model-load.js +74 -17
  450. package/src/inference/pipelines/text/ops.d.ts +13 -2
  451. package/src/inference/pipelines/text/ops.js +1 -1
  452. package/src/inference/pipelines/text/per-layer-inputs.js +347 -59
  453. package/src/inference/pipelines/text/probes.d.ts +3 -6
  454. package/src/inference/pipelines/text/probes.js +198 -31
  455. package/src/inference/pipelines/text/sampling.d.ts +0 -5
  456. package/src/inference/pipelines/text/sampling.js +140 -0
  457. package/src/inference/pipelines/text/stage-names.js +12 -0
  458. package/src/inference/pipelines/text/state.js +2 -0
  459. package/src/inference/pipelines/text/tsir-fixture-writer.js +409 -0
  460. package/src/inference/pipelines/text/types.d.ts +8 -78
  461. package/src/inference/pipelines/text/weights.d.ts +0 -50
  462. package/src/inference/pipelines/text/weights.js +2 -51
  463. package/src/inference/pipelines/text.d.ts +9 -1
  464. package/src/inference/pipelines/text.js +233 -29
  465. package/src/inference/pipelines/vision/index.js +0 -107
  466. package/src/inference/runtime-model.d.ts +0 -7
  467. package/src/inference/runtime-model.js +0 -23
  468. package/src/inference/test-harness.d.ts +6 -22
  469. package/src/inference/test-harness.js +97 -23
  470. package/src/inference/tokenizer.js +1 -10
  471. package/src/inference/tokenizers/sentencepiece.js +65 -15
  472. package/src/inference/tokenizers/types.d.ts +0 -21
  473. package/src/loader/doppler-loader.d.ts +3 -0
  474. package/src/loader/doppler-loader.js +90 -12
  475. package/src/loader/embedding-loader.js +30 -5
  476. package/src/loader/loader-types.d.ts +4 -0
  477. package/src/loader/manifest-config.d.ts +2 -7
  478. package/src/loader/manifest-config.js +25 -1
  479. package/src/loader/memory-monitor.d.ts +17 -5
  480. package/src/loader/per-layer-input-loader.js +114 -1
  481. package/src/loader/quantization-constants.d.ts +6 -0
  482. package/src/loader/shard-resolver.d.ts +1 -0
  483. package/src/loader/shard-resolver.js +15 -0
  484. package/src/loader/tensors/source-transform.d.ts +15 -0
  485. package/src/loader/tensors/source-transform.js +317 -34
  486. package/src/loader/tensors/tensor-loader.d.ts +2 -9
  487. package/src/loader/tensors/tensor-loader.js +84 -1
  488. package/src/loader/tensors/tensor-reader.js +28 -2
  489. package/src/memory/address-table.d.ts +1 -9
  490. package/src/memory/buffer-pool.d.ts +0 -16
  491. package/src/memory/buffer-pool.js +1 -33
  492. package/src/memory/heap-manager.d.ts +18 -43
  493. package/src/memory/unified-detect.d.ts +19 -34
  494. package/src/models/embeddinggemma.d.ts +7 -0
  495. package/src/models/embeddinggemma.js +24 -0
  496. package/src/models/family.d.ts +24 -0
  497. package/src/models/family.js +25 -0
  498. package/src/models/gemma3.d.ts +7 -0
  499. package/src/models/gemma3.js +33 -0
  500. package/src/models/gemma4.d.ts +7 -0
  501. package/src/models/gemma4.js +33 -0
  502. package/src/models/qwen3.d.ts +7 -0
  503. package/src/models/qwen3.js +36 -0
  504. package/src/rules/converter/tokenizer.rules.json +5 -0
  505. package/src/rules/execution-rules-contract-check.js +72 -3
  506. package/src/rules/inference/capability-transforms.rules.json +87 -0
  507. package/src/rules/inference/execution.rules.json +11 -1
  508. package/src/rules/inference/ffn.rules.json +6 -2
  509. package/src/rules/loader/tensor-loader.rules.json +9 -0
  510. package/src/rules/rule-registry.js +53 -75
  511. package/src/storage/artifact-storage-context.d.ts +12 -3
  512. package/src/storage/artifact-storage-context.js +389 -98
  513. package/src/storage/backends/opfs-store.js +138 -88
  514. package/src/storage/blake3.js +9 -1
  515. package/src/storage/export.js +1 -5
  516. package/src/storage/normalize-model-id.js +3 -0
  517. package/src/storage/preflight.js +22 -0
  518. package/src/storage/quickstart-downloader.js +44 -56
  519. package/src/storage/reports.js +35 -11
  520. package/src/storage/shard-manager.d.ts +8 -2
  521. package/src/storage/shard-manager.js +176 -15
  522. package/src/storage/source-artifact-store.d.ts +0 -2
  523. package/src/storage/source-artifact-store.js +3 -9
  524. package/src/sw.js +0 -3
  525. package/src/tooling/browser-command-runner.js +39 -0
  526. package/src/tooling/command-api-constants.js +1 -1
  527. package/src/tooling/command-api-family-normalizers.d.ts +1 -0
  528. package/src/tooling/command-api-family-normalizers.js +49 -7
  529. package/src/tooling/command-api-helpers.d.ts +0 -2
  530. package/src/tooling/command-api-helpers.js +6 -2
  531. package/src/tooling/command-api.d.ts +20 -1
  532. package/src/tooling/command-api.js +10 -1
  533. package/src/tooling/command-envelope.d.ts +0 -1
  534. package/src/tooling/command-envelope.js +1 -1
  535. package/src/tooling/command-runner-shared.d.ts +10 -0
  536. package/src/tooling/conversion-config-materializer.d.ts +0 -10
  537. package/src/tooling/conversion-config-materializer.js +1 -1
  538. package/src/tooling/diagnose-runner.js +1 -4
  539. package/src/tooling/hf-registry-utils.d.ts +2 -6
  540. package/src/tooling/hf-registry-utils.js +90 -19
  541. package/src/tooling/kernel-path-builder/index.js +1 -4
  542. package/src/tooling/kernel-path-builder/runtime-overlay.js +1 -3
  543. package/src/tooling/lean-execution-contract-runner.d.ts +0 -10
  544. package/src/tooling/lean-execution-contract-runner.js +2 -2
  545. package/src/tooling/litert-package-runtime.d.ts +1 -1
  546. package/src/tooling/litert-package-runtime.js +373 -21
  547. package/src/tooling/node-browser-command-runner.d.ts +15 -0
  548. package/src/tooling/node-browser-command-runner.js +97 -13
  549. package/src/tooling/node-command-runner.js +36 -0
  550. package/src/tooling/node-convert-worker-pool.js +12 -0
  551. package/src/tooling/node-convert-worker.js +10 -1
  552. package/src/tooling/node-converter.js +90 -1
  553. package/src/tooling/node-source-runtime.js +257 -150
  554. package/src/tooling/opfs-cache.js +1 -4
  555. package/src/tooling/precision-replay-math.js +174 -0
  556. package/src/tooling/program-bundle-host.d.ts +16 -0
  557. package/src/tooling/program-bundle-host.js +20 -0
  558. package/src/tooling/program-bundle-parity.d.ts +28 -0
  559. package/src/tooling/program-bundle-parity.js +210 -0
  560. package/src/tooling/program-bundle.d.ts +61 -0
  561. package/src/tooling/program-bundle.js +1119 -0
  562. package/src/tooling/rdrr-integrity-refresh.d.ts +32 -0
  563. package/src/tooling/rdrr-integrity-refresh.js +271 -0
  564. package/src/tooling/reference-verify.js +191 -0
  565. package/src/tooling/source-artifact-adapter.d.ts +0 -20
  566. package/src/tooling/source-artifact-adapter.js +4 -4
  567. package/src/tooling/source-package-profiles.js +1 -7
  568. package/src/tooling/source-runtime-bundle.d.ts +27 -0
  569. package/src/tooling/source-runtime-bundle.js +56 -23
  570. package/src/tooling/source-runtime-converter-config.d.ts +0 -5
  571. package/src/tooling/source-runtime-converter-config.js +13 -9
  572. package/src/tooling/source-runtime-materializer.js +4 -8
  573. package/src/tooling-exports/device.d.ts +11 -0
  574. package/src/tooling-exports/device.js +14 -0
  575. package/src/tooling-exports/manifest.d.ts +8 -0
  576. package/src/tooling-exports/manifest.js +11 -0
  577. package/src/tooling-exports/storage.d.ts +25 -0
  578. package/src/tooling-exports/storage.js +29 -0
  579. package/src/tooling-exports/structured.d.ts +1 -0
  580. package/src/tooling-exports/structured.js +5 -0
  581. package/src/tooling-exports.d.ts +23 -0
  582. package/src/tooling-exports.js +17 -0
  583. package/src/tooling-exports.shared.d.ts +27 -30
  584. package/src/tooling-exports.shared.js +28 -29
  585. package/src/types/index.d.ts +0 -3
  586. package/src/types/index.js +0 -3
  587. package/src/utils/array-buffer.d.ts +4 -0
  588. package/src/utils/array-buffer.js +9 -0
  589. package/src/utils/canonical-hash.d.ts +7 -0
  590. package/src/utils/canonical-hash.js +74 -0
  591. package/src/utils/clone-json.d.ts +1 -0
  592. package/src/utils/clone-json.js +7 -0
  593. package/src/utils/encode-utf8.js +5 -0
  594. package/src/utils/index.d.ts +4 -1
  595. package/src/utils/index.js +4 -1
  596. package/src/utils/runtime-env.js +7 -0
  597. package/src/utils/sha256.d.ts +1 -0
  598. package/src/utils/sha256.js +18 -4
  599. package/src/utils/stable-sort-object.d.ts +1 -0
  600. package/src/utils/stable-sort-object.js +13 -0
  601. package/src/version.js +1 -1
  602. package/tools/convert-safetensors-node.js +1 -2
  603. package/src/bootstrap.d.ts +0 -1
  604. package/src/config/conversion/sana/sana-sprint-0.6b-f16.json +0 -247
  605. package/src/config/kernels/kernel-ref.d.ts +0 -17
  606. package/src/config/kernels/kernel-ref.js +0 -75
  607. package/src/config/runtime/kernels/dequant-f16-q4k.json +0 -13
  608. package/src/config/runtime/kernels/dequant-f32-q4k.json +0 -13
  609. package/src/config/runtime/kernels/embeddinggemma-q4k-dequant-f32a.json +0 -37
  610. package/src/config/runtime/kernels/fused-q4k.json +0 -13
  611. package/src/config/runtime/kernels/gemma2-q4k-dequant-f16a.json +0 -33
  612. package/src/config/runtime/kernels/gemma2-q4k-dequant-f32a-nosubgroups.json +0 -33
  613. package/src/config/runtime/kernels/gemma2-q4k-fused-f32a.json +0 -33
  614. package/src/config/runtime/kernels/safe-q4k.json +0 -13
  615. package/src/config/runtime/overlays/debug-f32a.json +0 -3
  616. package/src/config/runtime/overlays/trace-config.json +0 -28
  617. package/src/config/runtime/profiles/gemma4-e2b-throughput.json +0 -39
  618. package/src/debug/perf.d.ts +0 -33
  619. package/src/debug/perf.js +0 -51
  620. package/src/errors/index.d.ts +0 -1
  621. package/src/errors/index.js +0 -1
  622. package/src/formats/gguf/index.d.ts +0 -8
  623. package/src/formats/gguf/index.js +0 -4
  624. package/src/formats/litert/index.d.ts +0 -1
  625. package/src/formats/litert/index.js +0 -1
  626. package/src/formats/safetensors/index.d.ts +0 -8
  627. package/src/formats/safetensors/index.js +0 -4
  628. package/src/formats/tflite/index.d.ts +0 -1
  629. package/src/formats/tflite/index.js +0 -1
  630. package/src/gpu/kernels/attention_decode_optimized.wgsl +0 -445
  631. package/src/gpu/kernels/dequant_mxfp4_vec4.wgsl +0 -116
  632. package/src/gpu/kernels/dequant_q8_0.wgsl +0 -98
  633. package/src/gpu/kernels/dequant_shared_vec4.wgsl +0 -160
  634. package/src/gpu/kernels/moe_gather_vec4.wgsl +0 -74
  635. package/src/inference/functiongemma.d.ts +0 -15
  636. package/src/inference/functiongemma.js +0 -1
  637. package/src/inference/pipelines/diffusion/sana-transformer.d.ts +0 -53
  638. package/src/inference/pipelines/diffusion/sana-transformer.js +0 -738
  639. package/src/rules/kernels/kernel-validator.d.ts +0 -24
  640. package/src/rules/kernels/kernel-validator.js +0 -160
  641. package/src/sw.d.ts +0 -1
  642. package/src/tooling/bench-runner.js +0 -64
  643. package/src/tooling/distill-runner.js +0 -56
  644. package/src/tooling/lora-runner.js +0 -37
  645. package/src/tooling/verify-runner.js +0 -63
  646. package/src/types/gpu.d.ts +0 -185
  647. package/src/types/gpu.js +0 -5
  648. package/src/types/inference.d.ts +0 -197
  649. package/src/types/inference.js +0 -5
  650. package/src/types/model.d.ts +0 -130
  651. package/src/types/model.js +0 -5
  652. package/src/utils/format-bytes.js +0 -20
  653. /package/src/config/{runtime/diagnostics → diagnostics}/drift-policies.json +0 -0
  654. /package/src/gpu/kernels/{sana_linear_attention_apply.wgsl → linear_attention_apply.wgsl} +0 -0
  655. /package/src/gpu/kernels/{sana_linear_attention_apply_f16.wgsl → linear_attention_apply_f16.wgsl} +0 -0
  656. /package/src/gpu/kernels/{sana_linear_attention_summary.wgsl → linear_attention_summary.wgsl} +0 -0
  657. /package/src/gpu/kernels/{sana_linear_attention_summary_f16.wgsl → linear_attention_summary_f16.wgsl} +0 -0
  658. /package/src/rules/kernels/{sana-linear-attention.rules.json → linear-attention.rules.json} +0 -0
@@ -38,7 +38,9 @@ import {
38
38
  quantizeToQ4KM,
39
39
  quantizeToQ4KMRowWise,
40
40
  quantizeToQ4KMColumnWise,
41
+ quantizeToInt4PerRowSymmetric,
41
42
  } from './quantizer.js';
43
+ import { cloneJsonValue } from '../utils/clone-json.js';
42
44
 
43
45
  // ============================================================================
44
46
  // Re-exports for Backward Compatibility
@@ -255,7 +257,16 @@ function resolveConversionTensors(model, converterConfig) {
255
257
  if (hasLanguageModelNamespace) {
256
258
  return source.filter((tensor) => {
257
259
  const lower = normalizeTensorName(tensor).toLowerCase();
258
- return lower.startsWith('language_model.') || lower.startsWith('model.language_model.');
260
+ // Keep top-level lm_head/output tensors alongside the language_model.*
261
+ // namespace. Multimodal HF models (e.g. Qwen 3.6-27B) place the
262
+ // language model body under model.language_model.* but expose the
263
+ // language modeling head at the bare top level (`lm_head.weight`).
264
+ // Dropping it leaves text-only conversion without an LM head and
265
+ // pipeline init fails at loadWeights.
266
+ return lower.startsWith('language_model.')
267
+ || lower.startsWith('model.language_model.')
268
+ || lower === 'lm_head.weight'
269
+ || lower === 'model.lm_head.weight';
259
270
  });
260
271
  }
261
272
 
@@ -264,6 +275,77 @@ function resolveConversionTensors(model, converterConfig) {
264
275
  ));
265
276
  }
266
277
 
278
+ function shouldMaterializeTiedLmHead(tensors, options) {
279
+ if (options?.inference?.output?.tieWordEmbeddings !== true) {
280
+ return false;
281
+ }
282
+ if (normalizeStorageQuant(options?.quantizationInfo?.lmHead ?? null) !== 'q4k') {
283
+ return false;
284
+ }
285
+ const embeddingQuant = normalizeStorageQuant(options?.quantizationInfo?.embeddings ?? null);
286
+ if (embeddingQuant === 'q4k') {
287
+ return false;
288
+ }
289
+ return !tensors.some((tensor) => resolveTensorRole(tensor) === 'lm_head');
290
+ }
291
+
292
+ function resolveTiedLmHeadName(embeddingName) {
293
+ const name = typeof embeddingName === 'string' ? embeddingName.trim() : '';
294
+ if (name === 'model.language_model.model.embed_tokens.weight') {
295
+ return 'model.language_model.lm_head.weight';
296
+ }
297
+ if (name === 'language_model.model.embed_tokens.weight') {
298
+ return 'language_model.lm_head.weight';
299
+ }
300
+ if (name === 'model.language_model.embed_tokens.weight') {
301
+ return 'model.language_model.lm_head.weight';
302
+ }
303
+ if (name === 'language_model.embed_tokens.weight') {
304
+ return 'language_model.lm_head.weight';
305
+ }
306
+ if (name === 'model.embed_tokens.weight' || name === 'embed_tokens.weight') {
307
+ return 'lm_head.weight';
308
+ }
309
+ return 'lm_head.weight';
310
+ }
311
+
312
+ function resolveTiedEmbeddingTensor(tensors, modelType) {
313
+ const candidates = tensors.filter((tensor) => (
314
+ resolveTensorRole(tensor) === 'embedding'
315
+ && resolveTensorGroup(tensor, modelType) === 'embed'
316
+ && Array.isArray(tensor?.shape)
317
+ && tensor.shape.length === 2
318
+ ));
319
+ return candidates[0] ?? null;
320
+ }
321
+
322
+ function materializeTiedLmHeadTensor(tensors, options) {
323
+ if (!shouldMaterializeTiedLmHead(tensors, options)) {
324
+ return tensors;
325
+ }
326
+ const embedding = resolveTiedEmbeddingTensor(tensors, options?.modelType ?? 'transformer');
327
+ if (!embedding) {
328
+ throw new Error(
329
+ 'Cannot materialize tied Q4K LM head: no 2D token embedding tensor was selected. '
330
+ + 'Check inference.output.tieWordEmbeddings and the conversion tensor filter.'
331
+ );
332
+ }
333
+ const name = resolveTiedLmHeadName(embedding.name);
334
+ if (tensors.some((tensor) => tensor?.name === name)) {
335
+ throw new Error(`Cannot materialize tied Q4K LM head: synthetic tensor name "${name}" already exists.`);
336
+ }
337
+ return [
338
+ ...tensors,
339
+ {
340
+ ...embedding,
341
+ name,
342
+ role: 'lm_head',
343
+ group: 'head',
344
+ sourceTensorName: embedding.name,
345
+ },
346
+ ];
347
+ }
348
+
267
349
  function toFloat32ForQ4K(tensorData, sourceDtype, tensorName) {
268
350
  const dtype = String(sourceDtype || '').toUpperCase();
269
351
  if (dtype === 'F32') {
@@ -722,6 +804,143 @@ export function shouldQuantize(tensorName, shape, options = {}) {
722
804
  return true;
723
805
  }
724
806
 
807
+ const GEMMA4_PLE_TENSOR_SUFFIXES = [
808
+ 'embed_tokens_per_layer.weight',
809
+ 'per_layer_embeddings.weight',
810
+ // GGUF (unsloth / ggml-org) naming for Gemma 4's PLE table.
811
+ 'per_layer_token_embd.weight',
812
+ ];
813
+
814
+ export function isGemma4PerLayerEmbedTensor(tensorName) {
815
+ if (typeof tensorName !== 'string') return false;
816
+ const normalized = tensorName.trim().toLowerCase();
817
+ return GEMMA4_PLE_TENSOR_SUFFIXES.some((suffix) => normalized.endsWith(suffix));
818
+ }
819
+
820
+ function validateInt4PleMaterializationContract(tensorLocations, inference, modelId) {
821
+ const materialization = inference?.session?.perLayerInputs?.materialization;
822
+ if (materialization !== 'gpu_split_tables') {
823
+ return;
824
+ }
825
+ for (const [name, location] of Object.entries(tensorLocations ?? {})) {
826
+ const sourceTransform = location?.sourceTransform ?? null;
827
+ if (
828
+ isGemma4PerLayerEmbedTensor(name)
829
+ && sourceTransform?.kind === 'litert_axis_dequant'
830
+ && String(sourceTransform?.sourceDtype ?? '').toUpperCase() === 'INT4'
831
+ ) {
832
+ throw new Error(
833
+ `Manifest "${modelId}" Gemma 4 INT4 PLE tensor "${name}" cannot use ` +
834
+ 'inference.session.perLayerInputs.materialization="gpu_split_tables". ' +
835
+ 'Use materialization="range_backed" or disable INT4 PLE quantization.'
836
+ );
837
+ }
838
+ }
839
+ }
840
+
841
+ function resolveOriginalTensorShape(options) {
842
+ const shape = options?.originalTensorShape;
843
+ if (!Array.isArray(shape) || shape.length !== 2) {
844
+ return null;
845
+ }
846
+ const rows = Number(shape[0]);
847
+ const cols = Number(shape[1]);
848
+ if (!Number.isInteger(rows) || rows <= 0 || !Number.isInteger(cols) || cols <= 0) {
849
+ return null;
850
+ }
851
+ return [rows, cols];
852
+ }
853
+
854
+ function resolvePerLayerEmbeddingQuant(options) {
855
+ const value = (
856
+ options?.perLayerEmbeddings
857
+ ?? options?.quantizationInfo?.perLayerEmbeddings
858
+ ?? null
859
+ );
860
+ if (value == null) return null;
861
+ return String(value).trim().toLowerCase().replace(/\s+/g, '_') || null;
862
+ }
863
+
864
+ function canInt4QuantizePerRow(tensor, options) {
865
+ if (options?.skipInt4PlePerRow === true) return false;
866
+ if (resolvePerLayerEmbeddingQuant(options) !== 'int4_per_row') return false;
867
+ if (!Array.isArray(tensor.shape) || tensor.shape.length !== 2) return false;
868
+ const [rows, cols] = tensor.shape;
869
+ if (!Number.isInteger(rows) || rows <= 0) return false;
870
+ if (!Number.isInteger(cols) || cols <= 0) return false;
871
+ if ((cols & 1) !== 0) return false;
872
+ const originalShape = resolveOriginalTensorShape(options);
873
+ if (originalShape && originalShape[1] !== cols) return false;
874
+ // Per-row symmetric INT4 produces one scale per row. PLE semantics require
875
+ // one scale per VOCAB token (typically 262144 for Gemma 4). HF stores PLE
876
+ // as [vocab, hidden] (rows = vocab ≫ cols). GGUF stores it transposed as
877
+ // [hidden, vocab] (rows = hidden ≪ cols). Quantizing the GGUF layout with
878
+ // per-row would wrongly share a single scale across all vocab tokens.
879
+ // Require the tensor to already be in [vocab, hidden] layout. Callers that
880
+ // have a GGUF PLE must transpose before reaching this function.
881
+ if (rows <= cols && (!originalShape || originalShape[0] <= originalShape[1])) return false;
882
+ const srcDtype = String(tensor.dtype || '').toUpperCase();
883
+ return srcDtype === 'F32' || srcDtype === 'F16' || srcDtype === 'BF16';
884
+ }
885
+
886
+ function toFloat32FromTensor(bytes, sourceDtype, tensorName) {
887
+ const src = String(sourceDtype || '').toUpperCase();
888
+ if (src === 'F32') {
889
+ if (bytes.byteLength % 4 !== 0) {
890
+ throw new Error(`Invalid F32 tensor byte length for ${tensorName}: ${bytes.byteLength}`);
891
+ }
892
+ return new Float32Array(bytes.buffer, bytes.byteOffset, bytes.byteLength / 4);
893
+ }
894
+ if (src === 'F16') {
895
+ if (bytes.byteLength % 2 !== 0) {
896
+ throw new Error(`Invalid F16 tensor byte length for ${tensorName}: ${bytes.byteLength}`);
897
+ }
898
+ const src16 = new Uint16Array(bytes.buffer, bytes.byteOffset, bytes.byteLength / 2);
899
+ const out = new Float32Array(src16.length);
900
+ for (let i = 0; i < src16.length; i++) out[i] = float16ToFloat32(src16[i]);
901
+ return out;
902
+ }
903
+ if (src === 'BF16') {
904
+ if (bytes.byteLength % 2 !== 0) {
905
+ throw new Error(`Invalid BF16 tensor byte length for ${tensorName}: ${bytes.byteLength}`);
906
+ }
907
+ const src16 = new Uint16Array(bytes.buffer, bytes.byteOffset, bytes.byteLength / 2);
908
+ const out = new Float32Array(src16.length);
909
+ for (let i = 0; i < src16.length; i++) out[i] = bf16ToFloat32(src16[i]);
910
+ return out;
911
+ }
912
+ throw new Error(`Unsupported source dtype "${sourceDtype}" for PLE INT4 quantization of ${tensorName}`);
913
+ }
914
+
915
+ function buildInt4PerRowPleTransform(tensor, bytes, sourceDtype, options = {}) {
916
+ const f32 = toFloat32FromTensor(bytes, sourceDtype, tensor.name);
917
+ const { quantized, scales } = quantizeToInt4PerRowSymmetric(f32, tensor.shape);
918
+ const scalesBytes = new Uint8Array(scales.buffer, scales.byteOffset, scales.byteLength);
919
+ const [rows, cols] = resolveOriginalTensorShape(options) ?? tensor.shape;
920
+ return {
921
+ tensorData: quantized,
922
+ companionData: scalesBytes,
923
+ // outDtype is the LOGICAL dtype the tensor resolves to after dequant at
924
+ // load time. Storage dtype (INT4) is carried in sourceTransform.sourceDtype.
925
+ outDtype: 'F16',
926
+ outLayout: 'row',
927
+ sourceDtype: String(sourceDtype || '').toUpperCase(),
928
+ tensorTargetQuant: 'int4_per_row_ple',
929
+ sourceTransform: {
930
+ kind: 'litert_axis_dequant',
931
+ scheme: 'per_axis_affine',
932
+ sourceDtype: 'INT4',
933
+ targetDtype: 'F16',
934
+ storageEncoding: 'offset_binary',
935
+ scaleSemantics: 'step',
936
+ storageShape: [rows, cols],
937
+ quantAxis: 1,
938
+ // scaleSource gets filled in by the writer with the scales companion's
939
+ // shard/offset/size after appendTensorBytes().
940
+ },
941
+ };
942
+ }
943
+
725
944
  export function transformTensorBytes(tensor, rawData, options = {}) {
726
945
  const tensorDataInput = rawData instanceof Uint8Array ? rawData : new Uint8Array(rawData);
727
946
  let tensorData = tensorDataInput;
@@ -750,6 +969,15 @@ export function transformTensorBytes(tensor, rawData, options = {}) {
750
969
  : null
751
970
  );
752
971
 
972
+ // Gemma 4 per-layer embeddings use MediaPipe's INT4 per-row symmetric
973
+ // quantization (verified against gemma-4-E2B-it.litertlm composites:
974
+ // quantizedDimension=0, one F32 scale per vocab row, zero_point=0). Saves
975
+ // ~3.5 GB per model vs the default F16 path. Runtime reads via
976
+ // sourceTransform.kind=litert_axis_dequant with scaleSemantics=step.
977
+ if (isGemma4PerLayerEmbedTensor(tensor.name) && canInt4QuantizePerRow(tensor, options)) {
978
+ return buildInt4PerRowPleTransform(tensor, tensorDataInput, sourceDtype, options);
979
+ }
980
+
753
981
  if (tensorTargetQuant === 'q4k') {
754
982
  const sourceQuant = normalizeStorageQuant(sourceDtype);
755
983
  if (sourceQuant === 'q4k') {
@@ -963,6 +1191,7 @@ export function extractArchitecture(config, ggufConfig) {
963
1191
  'num_attention_heads'
964
1192
  );
965
1193
  const numKVHeads = fromConfig('num_key_value_heads', 'num_kv_heads') ?? numHeads;
1194
+ const numGlobalKVHeads = fromConfig('num_global_key_value_heads', 'num_global_kv_heads');
966
1195
  const headDimFromConfig = fromConfig('head_dim') ?? Math.floor(hiddenSize / numHeads);
967
1196
  const vocabSize = requireNumber(
968
1197
  fromConfig('vocab_size', 'n_vocab'),
@@ -1006,6 +1235,7 @@ export function extractArchitecture(config, ggufConfig) {
1006
1235
  intermediateSize,
1007
1236
  numAttentionHeads: numHeads,
1008
1237
  numKeyValueHeads: numKVHeads,
1238
+ numGlobalKeyValueHeads: numGlobalKVHeads ?? undefined,
1009
1239
  headDim: headDimFromConfig,
1010
1240
  vocabSize,
1011
1241
  maxSeqLen,
@@ -1052,15 +1282,35 @@ export function extractArchitecture(config, ggufConfig) {
1052
1282
  'contextLength'
1053
1283
  );
1054
1284
 
1285
+ // Gemma 4-specific fields (optional — undefined on non-Gemma-4 GGUFs).
1286
+ // key_length is per-head; pick the larger of (key_length, key_length_swa)
1287
+ // for globalHeadDim and the smaller for headDim. Matches the mixed-geometry
1288
+ // KV cache expected by src/inference/pipelines/text/layer.js.
1289
+ const keyLen = firstNumber(c.attentionKeyLength);
1290
+ const keyLenSwa = firstNumber(c.attentionKeyLengthSwa);
1291
+ const headDimFromGguf = keyLenSwa != null && keyLen != null
1292
+ ? Math.min(keyLen, keyLenSwa)
1293
+ : (keyLenSwa ?? Math.floor(hiddenSize / numHeads));
1294
+ const globalHeadDim = (keyLen != null && keyLenSwa != null && keyLen !== keyLenSwa)
1295
+ ? Math.max(keyLen, keyLenSwa)
1296
+ : undefined;
1297
+ const numKvSharedLayers = firstNumber(c.numKvSharedLayers);
1298
+ const hiddenSizePerLayerInput = firstNumber(c.hiddenSizePerLayerInput);
1299
+ const vocabSizePerLayerInput = hiddenSizePerLayerInput != null ? vocabSize : undefined;
1300
+
1055
1301
  return {
1056
1302
  numLayers,
1057
1303
  hiddenSize,
1058
1304
  intermediateSize,
1059
1305
  numAttentionHeads: numHeads,
1060
1306
  numKeyValueHeads: numKVHeads,
1061
- headDim: Math.floor(hiddenSize / numHeads),
1307
+ headDim: headDimFromGguf,
1062
1308
  vocabSize,
1063
1309
  maxSeqLen,
1310
+ ...(globalHeadDim != null ? { globalHeadDim } : {}),
1311
+ ...(numKvSharedLayers != null ? { numKvSharedLayers } : {}),
1312
+ ...(hiddenSizePerLayerInput != null ? { hiddenSizePerLayerInput } : {}),
1313
+ ...(vocabSizePerLayerInput != null ? { vocabSizePerLayerInput } : {}),
1064
1314
  };
1065
1315
  }
1066
1316
 
@@ -1229,14 +1479,6 @@ export function resolveConvertedAt(value) {
1229
1479
  return new Date(parsed).toISOString();
1230
1480
  }
1231
1481
 
1232
- function cloneJsonValue(value) {
1233
- if (value == null) return value;
1234
- if (typeof structuredClone === 'function') {
1235
- return structuredClone(value);
1236
- }
1237
- return JSON.parse(JSON.stringify(value));
1238
- }
1239
-
1240
1482
  export function resolveManifestMultimodalConfig(rawConfig, manifestConfig = null) {
1241
1483
  const explicitVisionConfig = manifestConfig?.visionConfig;
1242
1484
  const explicitAudioConfig = manifestConfig?.audioConfig;
@@ -1248,6 +1490,215 @@ export function resolveManifestMultimodalConfig(rawConfig, manifestConfig = null
1248
1490
  };
1249
1491
  }
1250
1492
 
1493
+ function isPlainRecord(value) {
1494
+ return value != null && typeof value === 'object' && !Array.isArray(value);
1495
+ }
1496
+
1497
+ function stripUndefined(value) {
1498
+ if (Array.isArray(value)) {
1499
+ return value.map(stripUndefined);
1500
+ }
1501
+ if (!isPlainRecord(value)) {
1502
+ return value;
1503
+ }
1504
+ const out = {};
1505
+ for (const key of Object.keys(value).sort()) {
1506
+ const item = value[key];
1507
+ if (item !== undefined) {
1508
+ out[key] = stripUndefined(item);
1509
+ }
1510
+ }
1511
+ return out;
1512
+ }
1513
+
1514
+ function canonicalJson(value) {
1515
+ return JSON.stringify(stripUndefined(value));
1516
+ }
1517
+
1518
+ function normalizeDigest(value, label) {
1519
+ const digest = typeof value === 'string' ? value.trim() : '';
1520
+ if (!digest) {
1521
+ throw new Error(`Missing ${label} digest`);
1522
+ }
1523
+ return digest.startsWith('sha256:') ? digest : `sha256:${digest}`;
1524
+ }
1525
+
1526
+ function digestSuffix(value, length = 12) {
1527
+ return String(value || '').replace(/^sha256:/, '').slice(0, length);
1528
+ }
1529
+
1530
+ async function hashArtifactValue(hashString, value, label) {
1531
+ if (typeof hashString !== 'function') {
1532
+ return null;
1533
+ }
1534
+ const digest = await hashString(canonicalJson(value));
1535
+ return normalizeDigest(digest, label);
1536
+ }
1537
+
1538
+ function resolveArtifactSourceFormat(options) {
1539
+ const explicit = typeof options?.sourceFormat === 'string' ? options.sourceFormat.trim() : '';
1540
+ if (explicit) return explicit;
1541
+ const sourcePath = typeof options?.sourcePath === 'string' ? options.sourcePath.trim().toLowerCase() : '';
1542
+ if (sourcePath.endsWith('.gguf')) return 'gguf';
1543
+ if (sourcePath.endsWith('.tflite')) return 'tflite';
1544
+ if (sourcePath.endsWith('.task')) return 'task';
1545
+ if (sourcePath.endsWith('.litertlm')) return 'litertlm';
1546
+ return 'safetensors';
1547
+ }
1548
+
1549
+ function resolveSourceCheckpointIdentity(explicit, options) {
1550
+ const sourceRepo = typeof explicit?.sourceRepo === 'string' && explicit.sourceRepo.trim()
1551
+ ? explicit.sourceRepo.trim()
1552
+ : null;
1553
+ const sourceRevision = typeof explicit?.sourceRevision === 'string' && explicit.sourceRevision.trim()
1554
+ ? explicit.sourceRevision.trim()
1555
+ : null;
1556
+ const sourcePath = typeof options?.sourcePath === 'string' && options.sourcePath.trim()
1557
+ ? options.sourcePath.trim()
1558
+ : null;
1559
+ const source = typeof options?.source === 'string' && options.source.trim()
1560
+ ? options.source.trim()
1561
+ : null;
1562
+ const sourceCheckpointId = typeof explicit?.sourceCheckpointId === 'string' && explicit.sourceCheckpointId.trim()
1563
+ ? explicit.sourceCheckpointId.trim()
1564
+ : (
1565
+ sourceRepo && sourceRevision
1566
+ ? `${sourceRepo}@${sourceRevision}`
1567
+ : (sourceRepo ?? source ?? sourcePath ?? null)
1568
+ );
1569
+ return {
1570
+ sourceCheckpointId,
1571
+ sourceRepo,
1572
+ sourceRevision,
1573
+ };
1574
+ }
1575
+
1576
+ function inferArtifactModalitySet(modelType, tensorLocations, converterConfig) {
1577
+ if (modelType === 'diffusion') {
1578
+ return ['image'];
1579
+ }
1580
+ const names = Object.keys(tensorLocations ?? {}).map((name) => name.toLowerCase());
1581
+ const modalities = new Set();
1582
+ if (modelType === 'embedding') {
1583
+ modalities.add('embedding');
1584
+ } else {
1585
+ modalities.add('text');
1586
+ }
1587
+ if (converterConfig?.output?.textOnly === true) {
1588
+ return [...modalities].sort();
1589
+ }
1590
+ if (names.some((name) => name.includes('vision') || name.includes('visual') || name.includes('image'))) {
1591
+ modalities.add('vision');
1592
+ }
1593
+ if (names.some((name) => name.includes('audio'))) {
1594
+ modalities.add('audio');
1595
+ }
1596
+ if (names.some((name) => name.includes('projector') || name.includes('mm_projector'))) {
1597
+ modalities.add('projector');
1598
+ }
1599
+ return [...modalities].sort();
1600
+ }
1601
+
1602
+ function resolveMaterializationProfile(quantizationInfo, inference) {
1603
+ const materialization = typeof inference?.session?.perLayerInputs?.materialization === 'string'
1604
+ ? inference.session.perLayerInputs.materialization
1605
+ : 'standard';
1606
+ const perLayerEmbeddings = typeof quantizationInfo?.perLayerEmbeddings === 'string'
1607
+ ? quantizationInfo.perLayerEmbeddings
1608
+ : null;
1609
+ return perLayerEmbeddings ? `${materialization}-${perLayerEmbeddings}` : materialization;
1610
+ }
1611
+
1612
+ async function buildArtifactIdentity(options) {
1613
+ const explicit = isPlainRecord(options?.explicitArtifactIdentity)
1614
+ ? options.explicitArtifactIdentity
1615
+ : {};
1616
+ const hashString = options?.hashString;
1617
+ if (typeof hashString !== 'function' && Object.keys(explicit).length > 0) {
1618
+ return stripUndefined(explicit);
1619
+ }
1620
+ if (typeof hashString !== 'function') {
1621
+ return null;
1622
+ }
1623
+
1624
+ const sourceIdentity = resolveSourceCheckpointIdentity(explicit, options);
1625
+ const sourceFormat = typeof explicit.sourceFormat === 'string' && explicit.sourceFormat.trim()
1626
+ ? explicit.sourceFormat.trim()
1627
+ : resolveArtifactSourceFormat(options);
1628
+ const conversionConfigDigest = explicit.conversionConfigDigest
1629
+ ?? (options.conversionConfig
1630
+ ? await hashArtifactValue(hashString, options.conversionConfig, 'conversionConfig')
1631
+ : null);
1632
+ const shardSetHash = explicit.shardSetHash ?? explicit.weightPackHash
1633
+ ?? await hashArtifactValue(
1634
+ hashString,
1635
+ {
1636
+ hashAlgorithm: options.hashAlgorithm,
1637
+ shards: (options.shards ?? []).map((shard) => ({
1638
+ index: shard.index,
1639
+ filename: shard.filename,
1640
+ size: shard.size,
1641
+ hash: shard.hash,
1642
+ offset: shard.offset,
1643
+ })),
1644
+ },
1645
+ 'shardSet'
1646
+ );
1647
+ const modalitySet = Array.isArray(explicit.modalitySet) && explicit.modalitySet.length > 0
1648
+ ? [...explicit.modalitySet]
1649
+ : inferArtifactModalitySet(options.modelType, options.tensorLocations, options.converterConfig);
1650
+ const materializationProfile = explicit.materializationProfile
1651
+ ?? resolveMaterializationProfile(options.quantizationInfo, options.inference);
1652
+ const weightPackInput = {
1653
+ sourceCheckpointId: sourceIdentity.sourceCheckpointId,
1654
+ sourceFormat,
1655
+ modelType: options.modelType,
1656
+ modalitySet,
1657
+ quantizationInfo: options.quantizationInfo,
1658
+ materializationProfile,
1659
+ shardSetHash,
1660
+ sharding: {
1661
+ shardSizeBytes: options.converterConfig?.sharding?.shardSizeBytes ?? null,
1662
+ },
1663
+ output: {
1664
+ textOnly: options.converterConfig?.output?.textOnly === true,
1665
+ },
1666
+ };
1667
+ const weightPackHash = explicit.weightPackHash
1668
+ ?? await hashArtifactValue(hashString, weightPackInput, 'weightPack');
1669
+ const weightPackId = explicit.weightPackId
1670
+ ?? `${sanitizeModelId(options.modelId) ?? 'model'}-wp-${digestSuffix(weightPackHash)}`;
1671
+ const manifestVariantHash = await hashArtifactValue(
1672
+ hashString,
1673
+ {
1674
+ weightPackId,
1675
+ modelType: options.modelType,
1676
+ inference: options.inference,
1677
+ config: options.manifestConfig ?? null,
1678
+ },
1679
+ 'manifestVariant'
1680
+ );
1681
+ const manifestVariantId = explicit.manifestVariantId
1682
+ ?? `${sanitizeModelId(options.modelId) ?? 'model'}-mv-${digestSuffix(manifestVariantHash)}`;
1683
+
1684
+ return stripUndefined({
1685
+ ...explicit,
1686
+ sourceCheckpointId: sourceIdentity.sourceCheckpointId,
1687
+ sourceRepo: sourceIdentity.sourceRepo ?? undefined,
1688
+ sourceRevision: sourceIdentity.sourceRevision ?? undefined,
1689
+ sourceFormat,
1690
+ conversionConfigPath: explicit.conversionConfigPath ?? options.conversionConfigPath ?? undefined,
1691
+ conversionConfigDigest: conversionConfigDigest ?? undefined,
1692
+ weightPackId,
1693
+ weightPackHash,
1694
+ shardSetHash,
1695
+ manifestVariantId,
1696
+ modalitySet,
1697
+ materializationProfile,
1698
+ artifactCompleteness: explicit.artifactCompleteness ?? 'complete',
1699
+ });
1700
+ }
1701
+
1251
1702
 
1252
1703
  export function createManifest(
1253
1704
  modelId,
@@ -1344,6 +1795,8 @@ export function createManifest(
1344
1795
  modelType: resolvedModelType,
1345
1796
  quantization: resolvedQuantization,
1346
1797
  quantizationInfo: options.quantizationInfo,
1798
+ ...(options.artifactIdentity ? { artifactIdentity: options.artifactIdentity } : {}),
1799
+ ...(options.weightsRef ? { weightsRef: options.weightsRef } : {}),
1347
1800
  architecture: resolvedArchitecture,
1348
1801
  moeConfig,
1349
1802
  inference,
@@ -1394,7 +1847,7 @@ export function createManifest(
1394
1847
  // Main Converter (uses I/O adapter)
1395
1848
  // ============================================================================
1396
1849
 
1397
- const MAX_TENSOR_TYPED_ARRAY_BYTES = 0xffff_ffff;
1850
+ const MAX_TENSOR_TYPED_ARRAY_BYTES = 0x7fff_ffff;
1398
1851
 
1399
1852
  export async function convertModel(model, io, options = {}) {
1400
1853
  const { onProgress, signal } = options;
@@ -1413,7 +1866,14 @@ export async function convertModel(model, io, options = {}) {
1413
1866
  if (!modelId) {
1414
1867
  throw new Error('Missing modelId for conversion');
1415
1868
  }
1416
- const tensors = resolveConversionTensors(model, converterConfig);
1869
+ const tensors = materializeTiedLmHeadTensor(
1870
+ resolveConversionTensors(model, converterConfig),
1871
+ {
1872
+ inference: options.inference ?? converterConfig?.inference ?? null,
1873
+ quantizationInfo: options.quantizationInfo ?? null,
1874
+ modelType: options.modelType ?? model.modelType ?? 'transformer',
1875
+ }
1876
+ );
1417
1877
  if (!Array.isArray(tensors) || tensors.length === 0) {
1418
1878
  const textOnly = converterConfig?.output?.textOnly === true;
1419
1879
  if (textOnly) {
@@ -1550,6 +2010,12 @@ export async function convertModel(model, io, options = {}) {
1550
2010
  }
1551
2011
 
1552
2012
  let emittedChunk = false;
2013
+ // For PLE INT4 per-row quantization, each row-chunk returns its own
2014
+ // per-row F32 scale slice via companionData. Accumulate them across
2015
+ // chunks; after the stream completes, write the concatenated scales
2016
+ // blob and attach sourceTransform.scaleSource pointing at it.
2017
+ const companionChunks = [];
2018
+ let accumulatedSourceTransform = null;
1553
2019
  await options.largeTensorTransformer({
1554
2020
  tensor,
1555
2021
  transformContext,
@@ -1575,12 +2041,57 @@ export async function convertModel(model, io, options = {}) {
1575
2041
  }
1576
2042
  tensorSize += tensorData.byteLength;
1577
2043
  await appendTensorBytes(tensorData, tensorSpans);
2044
+
2045
+ if (result?.companionData instanceof Uint8Array && result.companionData.byteLength > 0) {
2046
+ if (!result.sourceTransform) {
2047
+ throw new Error(
2048
+ `Large tensor chunk returned companionData without sourceTransform for ${tensor.name}.`
2049
+ );
2050
+ }
2051
+ companionChunks.push(result.companionData);
2052
+ if (!accumulatedSourceTransform) {
2053
+ accumulatedSourceTransform = { ...result.sourceTransform };
2054
+ }
2055
+ } else if (result?.sourceTransform && !accumulatedSourceTransform) {
2056
+ accumulatedSourceTransform = { ...result.sourceTransform };
2057
+ }
1578
2058
  },
1579
2059
  });
1580
2060
 
1581
2061
  if (!emittedChunk) {
1582
2062
  throw new Error(`Large tensor transformer did not emit any bytes for ${tensor.name}.`);
1583
2063
  }
2064
+
2065
+ if (accumulatedSourceTransform) {
2066
+ let companionBytes = null;
2067
+ if (companionChunks.length > 0) {
2068
+ let totalCompanionBytes = 0;
2069
+ for (const chunk of companionChunks) totalCompanionBytes += chunk.byteLength;
2070
+ companionBytes = new Uint8Array(totalCompanionBytes);
2071
+ let offset = 0;
2072
+ for (const chunk of companionChunks) {
2073
+ companionBytes.set(chunk, offset);
2074
+ offset += chunk.byteLength;
2075
+ }
2076
+ const companionSpans = [];
2077
+ await appendTensorBytes(companionBytes, companionSpans);
2078
+ if (companionSpans.length !== 1) {
2079
+ throw new Error(
2080
+ `Companion scales for ${tensor.name} must land in a single shard (got ${companionSpans.length}).`
2081
+ );
2082
+ }
2083
+ transformContext._pleSourceTransform = {
2084
+ ...accumulatedSourceTransform,
2085
+ scaleSource: {
2086
+ shard: companionSpans[0].shardIndex,
2087
+ offset: companionSpans[0].offset,
2088
+ size: companionBytes.byteLength,
2089
+ },
2090
+ };
2091
+ } else {
2092
+ transformContext._pleSourceTransform = accumulatedSourceTransform;
2093
+ }
2094
+ }
1584
2095
  } else {
1585
2096
  const data = await io.readTensorData(tensor);
1586
2097
  const tensorDataInput = new Uint8Array(data);
@@ -1603,11 +2114,44 @@ export async function convertModel(model, io, options = {}) {
1603
2114
  outLayout = transformResult?.outLayout ?? null;
1604
2115
  tensorSize = tensorData.byteLength;
1605
2116
  await appendTensorBytes(tensorData, tensorSpans);
2117
+
2118
+ // Companion data (e.g., INT4 per-row scales for PLE) writes to its own
2119
+ // spans and is referenced from sourceTransform.scaleSource on the
2120
+ // primary tensor location.
2121
+ const companionData = transformResult?.companionData;
2122
+ if (companionData instanceof Uint8Array && companionData.byteLength > 0) {
2123
+ if (!transformResult?.sourceTransform) {
2124
+ throw new Error(
2125
+ `Tensor transformer returned companionData without sourceTransform for ${tensor.name}.`
2126
+ );
2127
+ }
2128
+ const companionSpans = [];
2129
+ await appendTensorBytes(companionData, companionSpans);
2130
+ if (companionSpans.length !== 1) {
2131
+ throw new Error(
2132
+ `Companion scales for ${tensor.name} must land in a single shard (got ${companionSpans.length}).`
2133
+ );
2134
+ }
2135
+ transformContext._pleSourceTransform = {
2136
+ ...transformResult.sourceTransform,
2137
+ scaleSource: {
2138
+ shard: companionSpans[0].shardIndex,
2139
+ offset: companionSpans[0].offset,
2140
+ size: companionData.byteLength,
2141
+ },
2142
+ };
2143
+ } else if (transformResult?.sourceTransform) {
2144
+ transformContext._pleSourceTransform = transformResult.sourceTransform;
2145
+ } else {
2146
+ transformContext._pleSourceTransform = null;
2147
+ }
1606
2148
  }
1607
2149
 
1608
2150
  // Record tensor location
1609
2151
  const role = resolveTensorRole(tensor);
1610
2152
  const group = resolveTensorGroup(tensor, tensorGroupModelType);
2153
+ const pleSourceTransform = transformContext._pleSourceTransform ?? null;
2154
+ transformContext._pleSourceTransform = null;
1611
2155
 
1612
2156
  if (tensorSpans.length === 1) {
1613
2157
  tensorLocations[tensor.name] = {
@@ -1619,6 +2163,7 @@ export async function convertModel(model, io, options = {}) {
1619
2163
  role,
1620
2164
  group,
1621
2165
  ...(outLayout ? { layout: outLayout } : {}),
2166
+ ...(pleSourceTransform ? { sourceTransform: pleSourceTransform } : {}),
1622
2167
  };
1623
2168
  } else {
1624
2169
  tensorLocations[tensor.name] = {
@@ -1629,6 +2174,7 @@ export async function convertModel(model, io, options = {}) {
1629
2174
  role,
1630
2175
  group,
1631
2176
  ...(outLayout ? { layout: outLayout } : {}),
2177
+ ...(pleSourceTransform ? { sourceTransform: pleSourceTransform } : {}),
1632
2178
  };
1633
2179
  }
1634
2180
 
@@ -1662,11 +2208,34 @@ export async function convertModel(model, io, options = {}) {
1662
2208
  options.quantization ?? model.quantization
1663
2209
  );
1664
2210
 
2211
+ validateInt4PleMaterializationContract(tensorLocations, options.inference, modelId);
2212
+
2213
+ const artifactIdentity = await buildArtifactIdentity({
2214
+ modelId,
2215
+ modelType: options.modelType,
2216
+ source: options.source,
2217
+ sourcePath: options.sourcePath,
2218
+ sourceFormat: options.sourceFormat,
2219
+ conversionConfigPath: options.conversionConfigPath,
2220
+ conversionConfig: options.conversionConfig,
2221
+ explicitArtifactIdentity: converterConfig?.manifest?.artifactIdentity ?? options.artifactIdentity ?? null,
2222
+ hashString: options.hashString,
2223
+ hashAlgorithm: converterConfig.manifest.hashAlgorithm,
2224
+ shards,
2225
+ tensorLocations,
2226
+ quantizationInfo: effectiveQuantizationInfo,
2227
+ inference: options.inference,
2228
+ manifestConfig: converterConfig?.manifest ?? null,
2229
+ converterConfig,
2230
+ });
2231
+
1665
2232
  const manifest = createManifest(modelId, model, shards, tensorLocations, {
1666
- source: 'convert-core',
2233
+ source: options.source ?? 'convert-core',
1667
2234
  modelType: options.modelType,
1668
2235
  quantization: effectiveManifestQuantization,
1669
2236
  quantizationInfo: effectiveQuantizationInfo,
2237
+ artifactIdentity,
2238
+ weightsRef: converterConfig?.manifest?.weightsRef ?? options.weightsRef ?? null,
1670
2239
  hashAlgorithm: converterConfig.manifest.hashAlgorithm,
1671
2240
  architecture: options.architecture,
1672
2241
  inference: options.inference,