doppler-gpu 0.4.3 → 0.4.4

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 (433) hide show
  1. package/CHANGELOG.md +28 -0
  2. package/README.md +54 -33
  3. package/assets/doppler.svg +57 -1601
  4. package/package.json +36 -5
  5. package/src/cli/cli-output.js +34 -1
  6. package/src/cli/doppler-cli.d.ts +18 -0
  7. package/src/cli/doppler-cli.js +68 -35
  8. package/src/cli/doppler-serve.d.ts +210 -0
  9. package/src/cli/doppler-serve.js +468 -42
  10. package/src/client/doppler-registry.json +3 -5
  11. package/src/client/provider.d.ts +1 -1
  12. package/src/client/provider.js +1 -1
  13. package/src/client/runtime/index.js +13 -0
  14. package/src/client/runtime/lora.d.ts +11 -1
  15. package/src/client/runtime/lora.js +28 -7
  16. package/src/client/runtime/model-manager.js +35 -11
  17. package/src/client/runtime/model-session.d.ts +2 -1
  18. package/src/client/runtime/model-session.js +2 -2
  19. package/src/client/runtime/model-source.js +20 -1
  20. package/src/client/runtime/types.d.ts +8 -0
  21. package/src/client/wrap-pipeline-handle.d.ts +62 -0
  22. package/src/client/wrap-pipeline-handle.js +188 -13
  23. package/src/config/conversion/README.md +5 -0
  24. package/src/config/conversion/diffusiongemma/diffusiongemma-26b-a4b-it-q4k-ehf16-af16.json +646 -0
  25. package/src/config/conversion/embeddinggemma/google-embeddinggemma-300m-q4k-ehf16-af32.json +3 -3
  26. package/src/config/conversion/gemma3/gemma-3-1b-it-f16-af32.json +4 -4
  27. package/src/config/conversion/gemma3/gemma-3-1b-it-q4k-ehf16-af32.json +4 -4
  28. package/src/config/conversion/gemma3/gemma-3-270m-it-f16-af32.json +393 -0
  29. package/src/config/conversion/gemma3/gemma-3-270m-it-q4k-ehf16-af16.json +425 -0
  30. package/src/config/conversion/gemma3/gemma-3-270m-it-q4k-ehf16-af32.json +17 -9
  31. package/src/config/conversion/gemma3/translategemma-4b-1b-enes-q4k-ehf16-af32.json +4 -4
  32. package/src/config/conversion/gemma3/translategemma-4b-it-q4k-ehf16-af32.json +4 -4
  33. package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-af16.json +737 -0
  34. package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-af32.json +730 -0
  35. package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-hq4k-af16.json +780 -0
  36. package/src/config/conversion/gemma4/gemma-4-12b-it-text-w4a16-ct-ehf16-af16.json +716 -0
  37. package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af16.json +5 -5
  38. package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af32.json +4 -4
  39. package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af16-int4ple.json +14 -12
  40. package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32-int4ple.json +7 -7
  41. package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32.json +7 -7
  42. package/src/config/conversion/gemma4/gemma-4-moe-q4k-ehf16-af32.json +3 -3
  43. package/src/config/conversion/gpt-oss-20b-f16-xmxfp4.json +3 -3
  44. package/src/config/conversion/janus/janus-pro-1b-text-q4k-ehaf16.json +3 -3
  45. package/src/config/conversion/lfm2/lfm2.5-1.2b-instruct-q4k-ehf16-af32.json +3 -3
  46. package/src/config/conversion/qwen3/qwen-3-5-0-8b-q4k-ehaf16.json +8 -4
  47. package/src/config/conversion/qwen3/qwen-3-5-2b-q4k-ehaf16.json +8 -4
  48. package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-eaf16.json +9 -5
  49. package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-ehaf16.json +8 -4
  50. package/src/config/goal-completion-matrix.json +552 -0
  51. package/src/config/kernel-path-loader.js +192 -38
  52. package/src/config/kernels/kernel-ref-digests.js +52 -18
  53. package/src/config/kernels/registry.json +3017 -744
  54. package/src/config/merge-contract-check.js +7 -3
  55. package/src/config/merge-helpers.d.ts +2 -0
  56. package/src/config/merge-helpers.js +1 -0
  57. package/src/config/merge.d.ts +1 -0
  58. package/src/config/merge.js +30 -0
  59. package/src/config/model-lanes/catalog-lane-resolver.d.ts +66 -0
  60. package/src/config/model-lanes/catalog-lane-resolver.js +170 -0
  61. package/src/config/param-validator.js +4 -0
  62. package/src/config/runtime/experiments/verify/lfm2-verify.json +3 -3
  63. package/src/config/runtime/profiles/diffusiongemma-26b-a4b-throughput.json +26 -0
  64. package/src/config/runtime/profiles/gemma3-270m-q4k-throughput-overlapped-probe.json +42 -0
  65. package/src/config/runtime/profiles/gemma3-270m-q4k-throughput-probe.json +39 -0
  66. package/src/config/runtime/profiles/gemma4-12b-rdrr-safe-b4.json +48 -0
  67. package/src/config/runtime/profiles/gemma4-12b-rdrr-safe-single.json +46 -0
  68. package/src/config/schema/buffer-pool.schema.js +3 -3
  69. package/src/config/schema/claim-evidence-contract.schema.json +93 -0
  70. package/src/config/schema/command-surface-contract.schema.json +81 -0
  71. package/src/config/schema/converter.schema.d.ts +11 -0
  72. package/src/config/schema/converter.schema.js +8 -0
  73. package/src/config/schema/doppler.schema.d.ts +1 -0
  74. package/src/config/schema/doppler.schema.js +38 -6
  75. package/src/config/schema/execution-v1.schema.d.ts +27 -0
  76. package/src/config/schema/execution-v1.schema.js +28 -19
  77. package/src/config/schema/goal-completion-matrix.schema.json +223 -0
  78. package/src/config/schema/inference.schema.d.ts +11 -0
  79. package/src/config/schema/manifest.schema.d.ts +107 -1
  80. package/src/config/schema/manifest.schema.js +10 -1
  81. package/src/config/schema/model-artifact-contract.schema.json +119 -0
  82. package/src/config/schema/moe.schema.d.ts +13 -0
  83. package/src/config/schema/moe.schema.js +1 -1
  84. package/src/config/schema/policy-schema-registry.json +26 -0
  85. package/src/config/schema/subsystem-support-contract.schema.json +48 -0
  86. package/src/config/schema/training-metrics.schema.d.ts +1 -1
  87. package/src/config/schema/training-metrics.schema.js +4 -3
  88. package/src/config/source-packages/litert/gemma-4-12b-it.json +390 -0
  89. package/src/config/source-packages/litert/gemma-4-e2b-it.json +7 -7
  90. package/src/config/source-packages/litert/gemma-4-e4b-it.json +370 -0
  91. package/src/config/source-packages/registry.json +25 -0
  92. package/src/config/supported-operations.json +5 -0
  93. package/src/config/transforms/capability-transform-resolver.d.ts +13 -2
  94. package/src/config/transforms/execution-graph-transforms.d.ts +42 -0
  95. package/src/config/transforms/execution-graph-transforms.js +227 -230
  96. package/src/converter/conversion-plan.d.ts +2 -0
  97. package/src/converter/conversion-plan.js +18 -4
  98. package/src/converter/core.d.ts +8 -0
  99. package/src/converter/core.js +574 -12
  100. package/src/converter/quantization-info.js +179 -2
  101. package/src/converter/quantizer.d.ts +5 -4
  102. package/src/converter/quantizer.js +16 -2
  103. package/src/experimental/adapters/lora-loader.js +109 -24
  104. package/src/experimental/diffusion/index.d.ts +24 -0
  105. package/src/experimental/diffusion/index.js +10 -0
  106. package/src/experimental/distribution/descriptor-transport.d.ts +117 -0
  107. package/src/experimental/distribution/descriptor-transport.js +342 -0
  108. package/src/experimental/training/README.md +11 -0
  109. package/src/experimental/training/autograd.js +33 -8
  110. package/src/experimental/training/datasets/gepa-frontier.d.ts +41 -0
  111. package/src/experimental/training/datasets/gepa-frontier.js +220 -0
  112. package/src/experimental/training/datasets/index.d.ts +13 -1
  113. package/src/experimental/training/datasets/index.js +8 -1
  114. package/src/experimental/training/datasets/teacher-traces.d.ts +51 -0
  115. package/src/experimental/training/datasets/teacher-traces.js +279 -0
  116. package/src/experimental/training/datasets/text-pairs.d.ts +45 -0
  117. package/src/experimental/training/datasets/text-pairs.js +82 -5
  118. package/src/experimental/training/distillation/student-fixture.d.ts +5 -0
  119. package/src/experimental/training/distillation/student-fixture.js +290 -32
  120. package/src/experimental/training/export.d.ts +9 -0
  121. package/src/experimental/training/export.js +113 -7
  122. package/src/experimental/training/index.d.ts +13 -1
  123. package/src/experimental/training/index.js +13 -1
  124. package/src/experimental/training/lora-pipeline.d.ts +143 -0
  125. package/src/experimental/training/lora-pipeline.js +1737 -49
  126. package/src/experimental/training/lora.js +23 -1
  127. package/src/experimental/training/loss.js +1 -1
  128. package/src/experimental/training/operator-agent-eval.d.ts +65 -0
  129. package/src/experimental/training/operator-agent-eval.js +326 -0
  130. package/src/experimental/training/operator-command.js +310 -51
  131. package/src/experimental/training/runner.d.ts +1 -0
  132. package/src/experimental/training/runner.js +4 -0
  133. package/src/experimental/training/workload-packs/distill-glm52-doppler-agent-harness-gemma270m-f16-sft-tiny.json +160 -0
  134. package/src/experimental/training/workload-packs/distill-glm52-doppler-code-agent-gemma270m-f16-sft-tiny.json +175 -0
  135. package/src/experimental/training/workload-packs/distill-glm52-doppler-code-agent-sft-tiny.json +175 -0
  136. package/src/experimental/training/workload-packs/distill-glm52-doppler-js-json-gemma270m-f16-sft-tiny.json +160 -0
  137. package/src/experimental/training/workload-packs/distill-glm52-doppler-review-gemma270m-f16-sft-tiny.json +162 -0
  138. package/src/experimental/training/workload-packs/distill-glm52-doppler-wgsl-gemma270m-f16-sft-tiny.json +160 -0
  139. package/src/experimental/training/workload-packs/lora-doppler-agent-harness-gemma270m-f16-tiny.json +137 -0
  140. package/src/experimental/training/workload-packs/lora-doppler-code-agent-gemma270m-f16-tiny.json +152 -0
  141. package/src/experimental/training/workload-packs/lora-doppler-code-agent-tiny.json +152 -0
  142. package/src/experimental/training/workload-packs/lora-doppler-js-json-gemma270m-f16-tiny.json +137 -0
  143. package/src/experimental/training/workload-packs/lora-doppler-review-gemma270m-f16-tiny.json +139 -0
  144. package/src/experimental/training/workload-packs/lora-doppler-wgsl-gemma270m-f16-tiny.json +137 -0
  145. package/src/experimental/training/workload-packs/registry.json +186 -6
  146. package/src/experimental/training/workloads.d.ts +37 -0
  147. package/src/experimental/training/workloads.js +110 -1
  148. package/src/formats/rdrr/functional-descriptor.d.ts +73 -0
  149. package/src/formats/rdrr/functional-descriptor.js +230 -0
  150. package/src/formats/rdrr/index.d.ts +1 -0
  151. package/src/formats/rdrr/index.js +1 -0
  152. package/src/formats/rdrr/parsing.js +22 -6
  153. package/src/formats/rdrr/source-transform-contract.js +10 -1
  154. package/src/formats/rdrr/storage-descriptor.js +9 -2
  155. package/src/formats/rdrr/types.d.ts +5 -3
  156. package/src/formats/rdrr/validation.js +15 -1
  157. package/src/gpu/command-recorder.d.ts +43 -3
  158. package/src/gpu/command-recorder.js +122 -25
  159. package/src/gpu/device.js +111 -11
  160. package/src/gpu/kernel-tuner/cache.js +1 -5
  161. package/src/gpu/kernels/attention.js +17 -9
  162. package/src/gpu/kernels/attention_decode_online_head256_f16kv.wgsl +236 -0
  163. package/src/gpu/kernels/attention_streaming_f16.wgsl +7 -5
  164. package/src/gpu/kernels/backward/cross_entropy_backward.js +72 -3
  165. package/src/gpu/kernels/backward/cross_entropy_backward.wgsl +2 -2
  166. package/src/gpu/kernels/cast.js +74 -67
  167. package/src/gpu/kernels/codegen/patches/attention_streaming_f16.from.attention_streaming.diff +46 -1
  168. package/src/gpu/kernels/codegen/patches/gelu_f16.from.gelu.diff +0 -20
  169. package/src/gpu/kernels/codegen/patches/residual_f16.from.residual.diff +15 -12
  170. package/src/gpu/kernels/codegen/patches/rmsnorm_f16.from.rmsnorm.diff +59 -44
  171. package/src/gpu/kernels/diffusion-gemma-sampling.d.ts +19 -0
  172. package/src/gpu/kernels/diffusion-gemma-sampling.js +104 -0
  173. package/src/gpu/kernels/diffusion_gemma_sampling.wgsl +130 -0
  174. package/src/gpu/kernels/dispatch.js +13 -3
  175. package/src/gpu/kernels/fused_ffn.js +3 -4
  176. package/src/gpu/kernels/fused_matmul_litert_int4_f16a.wgsl +78 -0
  177. package/src/gpu/kernels/fused_matmul_litert_int4_f16a_f32out.wgsl +78 -0
  178. package/src/gpu/kernels/fused_matmul_litert_int4_f32a_f32out.wgsl +76 -0
  179. package/src/gpu/kernels/fused_matmul_rmsnorm.js +101 -80
  180. package/src/gpu/kernels/fused_matmul_w4a16_batched_f16a.wgsl +116 -0
  181. package/src/gpu/kernels/fused_matmul_w4a16_batched_f16a_f32out.wgsl +116 -0
  182. package/src/gpu/kernels/fused_matmul_w4a16_f16a.wgsl +95 -0
  183. package/src/gpu/kernels/fused_matmul_w4a16_f16a_f32out.wgsl +95 -0
  184. package/src/gpu/kernels/gather.d.ts +84 -1
  185. package/src/gpu/kernels/gather.js +353 -5
  186. package/src/gpu/kernels/gather_litert_int4_vec4_f16_out.wgsl +75 -0
  187. package/src/gpu/kernels/gather_litert_int4_vec4_f32_out.wgsl +72 -0
  188. package/src/gpu/kernels/gather_split4_f16_vec4_f16_out.wgsl +103 -0
  189. package/src/gpu/kernels/gather_split8_f16_vec4_f16_out.wgsl +143 -0
  190. package/src/gpu/kernels/gather_split8_f16_vec4_f32_out.wgsl +143 -0
  191. package/src/gpu/kernels/gelu.wgsl +11 -7
  192. package/src/gpu/kernels/gelu_f16.wgsl +11 -7
  193. package/src/gpu/kernels/gemma4-route-expert.d.ts +35 -0
  194. package/src/gpu/kernels/gemma4-route-expert.js +126 -0
  195. package/src/gpu/kernels/gemma4_route_q4_matmul_f16a.wgsl +210 -0
  196. package/src/gpu/kernels/index.d.ts +71 -0
  197. package/src/gpu/kernels/index.js +52 -0
  198. package/src/gpu/kernels/kv-cache-write.d.ts +22 -0
  199. package/src/gpu/kernels/kv-cache-write.js +126 -0
  200. package/src/gpu/kernels/kv_cache_write_f32_to_f16.wgsl +32 -0
  201. package/src/gpu/kernels/lm-head-argmax.d.ts +19 -0
  202. package/src/gpu/kernels/lm-head-argmax.js +218 -0
  203. package/src/gpu/kernels/lm_head_argmax_f16.wgsl +189 -0
  204. package/src/gpu/kernels/matmul-dispatch.d.ts +11 -0
  205. package/src/gpu/kernels/matmul-dispatch.js +37 -1
  206. package/src/gpu/kernels/matmul-selection.d.ts +5 -5
  207. package/src/gpu/kernels/matmul-selection.js +104 -9
  208. package/src/gpu/kernels/matmul.d.ts +1 -1
  209. package/src/gpu/kernels/matmul.js +151 -43
  210. package/src/gpu/kernels/moe.d.ts +1 -0
  211. package/src/gpu/kernels/moe.js +46 -18
  212. package/src/gpu/kernels/pipeline-cache.d.ts +8 -0
  213. package/src/gpu/kernels/pipeline-cache.js +24 -0
  214. package/src/gpu/kernels/rep-penalty.d.ts +21 -0
  215. package/src/gpu/kernels/rep-penalty.js +17 -9
  216. package/src/gpu/kernels/residual.d.ts +1 -0
  217. package/src/gpu/kernels/residual.js +13 -1
  218. package/src/gpu/kernels/residual.wgsl +3 -3
  219. package/src/gpu/kernels/residual_f16.wgsl +5 -3
  220. package/src/gpu/kernels/residual_f16_vec4.wgsl +6 -6
  221. package/src/gpu/kernels/residual_vec4.wgsl +6 -6
  222. package/src/gpu/kernels/rmsnorm-pair.d.ts +39 -0
  223. package/src/gpu/kernels/rmsnorm-pair.js +152 -0
  224. package/src/gpu/kernels/rmsnorm-qk.d.ts +41 -0
  225. package/src/gpu/kernels/rmsnorm-qk.js +128 -0
  226. package/src/gpu/kernels/rmsnorm.d.ts +31 -0
  227. package/src/gpu/kernels/rmsnorm.js +32 -7
  228. package/src/gpu/kernels/rmsnorm.wgsl +10 -6
  229. package/src/gpu/kernels/rmsnorm_f16.wgsl +7 -3
  230. package/src/gpu/kernels/rmsnorm_pair.wgsl +224 -0
  231. package/src/gpu/kernels/rmsnorm_qk.wgsl +118 -0
  232. package/src/gpu/kernels/rmsnorm_qk_f16.wgsl +143 -0
  233. package/src/gpu/kernels/rope-qk.d.ts +41 -0
  234. package/src/gpu/kernels/rope-qk.js +143 -0
  235. package/src/gpu/kernels/rope_qk.wgsl +95 -0
  236. package/src/gpu/kernels/rope_qk_f16.wgsl +97 -0
  237. package/src/gpu/kernels/rule-registry.js +50 -10
  238. package/src/gpu/kernels/sample.js +212 -187
  239. package/src/gpu/kernels/scatter_add_dynamic_expert_scale.wgsl +47 -0
  240. package/src/gpu/kernels/scatter_add_dynamic_f16_expert_scale.wgsl +49 -0
  241. package/src/gpu/kernels/scatter_add_dynamic_f16_weights_expert_scale.wgsl +49 -0
  242. package/src/gpu/kernels/scatter_add_routes_f16_weights_expert_scale.wgsl +45 -0
  243. package/src/gpu/kernels/shader-cache.js +1 -21
  244. package/src/gpu/kernels/soft-embedding.d.ts +28 -0
  245. package/src/gpu/kernels/soft-embedding.js +447 -0
  246. package/src/gpu/kernels/soft_embedding_logits_f16.wgsl +107 -0
  247. package/src/gpu/kernels/soft_embedding_split_f16.wgsl +46 -0
  248. package/src/gpu/kernels/softmax.wgsl +5 -80
  249. package/src/gpu/kernels/split-qkv-rmsnorm-qk.d.ts +44 -0
  250. package/src/gpu/kernels/split-qkv-rmsnorm-qk.js +154 -0
  251. package/src/gpu/kernels/split-qkv-rmsnorm-rope-qk.d.ts +59 -0
  252. package/src/gpu/kernels/split-qkv-rmsnorm-rope-qk.js +245 -0
  253. package/src/gpu/kernels/split_qkv_rmsnorm_qk.wgsl +146 -0
  254. package/src/gpu/kernels/split_qkv_rmsnorm_rope_qk.wgsl +185 -0
  255. package/src/gpu/kernels/split_qkv_rmsnorm_rope_qk_f16kv_cache.wgsl +187 -0
  256. package/src/gpu/kernels/uniform-utils.js +2 -5
  257. package/src/gpu/kernels/utils.d.ts +16 -1
  258. package/src/gpu/kernels/utils.js +42 -13
  259. package/src/gpu/readback-utils.js +10 -2
  260. package/src/gpu/uniform-cache.d.ts +10 -1
  261. package/src/gpu/uniform-cache.js +25 -7
  262. package/src/gpu/weight-buffer.d.ts +76 -7
  263. package/src/gpu/weight-buffer.js +85 -1
  264. package/src/inference/browser-harness-diffusion-energy-suites.js +269 -1
  265. package/src/inference/browser-harness-suite-helpers.d.ts +9 -0
  266. package/src/inference/browser-harness-suite-helpers.js +438 -2
  267. package/src/inference/browser-harness-text-helpers.d.ts +77 -0
  268. package/src/inference/browser-harness-text-helpers.js +236 -11
  269. package/src/inference/browser-harness.d.ts +8 -0
  270. package/src/inference/browser-harness.js +292 -8
  271. package/src/inference/decode-ring.d.ts +1 -0
  272. package/src/inference/decode-ring.js +7 -5
  273. package/src/inference/kv-cache/base.d.ts +25 -0
  274. package/src/inference/kv-cache/base.js +109 -1
  275. package/src/inference/kv-cache/mixed-geometry.d.ts +10 -0
  276. package/src/inference/kv-cache/mixed-geometry.js +109 -2
  277. package/src/inference/kv-cache/sliding-window.d.ts +14 -0
  278. package/src/inference/kv-cache/sliding-window.js +88 -0
  279. package/src/inference/model-lanes/execution-lane-resolver.d.ts +61 -0
  280. package/src/inference/model-lanes/execution-lane-resolver.js +126 -0
  281. package/src/inference/moe-router.d.ts +14 -1
  282. package/src/inference/moe-router.js +24 -1
  283. package/src/inference/pipelines/audio/gemma4.js +52 -1
  284. package/src/inference/pipelines/diffusion-gemma/config.d.ts +27 -0
  285. package/src/inference/pipelines/diffusion-gemma/config.js +149 -0
  286. package/src/inference/pipelines/diffusion-gemma/index.d.ts +31 -0
  287. package/src/inference/pipelines/diffusion-gemma/index.js +12 -0
  288. package/src/inference/pipelines/diffusion-gemma/pipeline.d.ts +69 -0
  289. package/src/inference/pipelines/diffusion-gemma/pipeline.js +380 -0
  290. package/src/inference/pipelines/diffusion-gemma/sampling.d.ts +128 -0
  291. package/src/inference/pipelines/diffusion-gemma/sampling.js +395 -0
  292. package/src/inference/pipelines/text/attention/dispatch-params.d.ts +28 -4
  293. package/src/inference/pipelines/text/attention/dispatch-params.js +176 -2
  294. package/src/inference/pipelines/text/attention/projections.d.ts +71 -2
  295. package/src/inference/pipelines/text/attention/projections.js +329 -49
  296. package/src/inference/pipelines/text/attention/quant-options.d.ts +38 -0
  297. package/src/inference/pipelines/text/attention/record.js +446 -73
  298. package/src/inference/pipelines/text/attention/run.js +251 -49
  299. package/src/inference/pipelines/text/attention/types.d.ts +5 -1
  300. package/src/inference/pipelines/text/config.d.ts +22 -1
  301. package/src/inference/pipelines/text/config.js +199 -22
  302. package/src/inference/pipelines/text/dtype-contract.d.ts +1 -1
  303. package/src/inference/pipelines/text/dtype-contract.js +4 -1
  304. package/src/inference/pipelines/text/embed.d.ts +4 -3
  305. package/src/inference/pipelines/text/embed.js +16 -6
  306. package/src/inference/pipelines/text/execution-plan.d.ts +8 -0
  307. package/src/inference/pipelines/text/execution-plan.js +6 -0
  308. package/src/inference/pipelines/text/execution-runtime-builders.d.ts +4 -0
  309. package/src/inference/pipelines/text/execution-runtime-builders.js +44 -28
  310. package/src/inference/pipelines/text/execution-v1.d.ts +2 -0
  311. package/src/inference/pipelines/text/execution-v1.js +373 -139
  312. package/src/inference/pipelines/text/ffn/dense.js +44 -13
  313. package/src/inference/pipelines/text/ffn/moe.d.ts +4 -1
  314. package/src/inference/pipelines/text/ffn/moe.js +9 -2
  315. package/src/inference/pipelines/text/ffn/sandwich.d.ts +3 -1
  316. package/src/inference/pipelines/text/ffn/sandwich.js +237 -15
  317. package/src/inference/pipelines/text/ffn/standard.d.ts +3 -1
  318. package/src/inference/pipelines/text/ffn/standard.js +44 -4
  319. package/src/inference/pipelines/text/generator-decode-policy.d.ts +81 -0
  320. package/src/inference/pipelines/text/generator-decode-policy.js +2 -12
  321. package/src/inference/pipelines/text/generator-helpers.js +5 -2
  322. package/src/inference/pipelines/text/generator-prefill-helpers.d.ts +84 -0
  323. package/src/inference/pipelines/text/generator-prefill-helpers.js +67 -1
  324. package/src/inference/pipelines/text/generator-runtime.d.ts +1 -0
  325. package/src/inference/pipelines/text/generator-steps.d.ts +38 -3
  326. package/src/inference/pipelines/text/generator-steps.js +268 -77
  327. package/src/inference/pipelines/text/generator.d.ts +39 -0
  328. package/src/inference/pipelines/text/generator.js +1105 -18
  329. package/src/inference/pipelines/text/init.d.ts +8 -1
  330. package/src/inference/pipelines/text/init.js +30 -1
  331. package/src/inference/pipelines/text/layer-plan-gpu.js +62 -14
  332. package/src/inference/pipelines/text/layer.js +272 -25
  333. package/src/inference/pipelines/text/linear-attention.js +6 -4
  334. package/src/inference/pipelines/text/logits/gpu.d.ts +49 -2
  335. package/src/inference/pipelines/text/logits/gpu.js +417 -66
  336. package/src/inference/pipelines/text/logits/index.d.ts +2 -2
  337. package/src/inference/pipelines/text/logits/index.js +96 -42
  338. package/src/inference/pipelines/text/logits/types.d.ts +2 -2
  339. package/src/inference/pipelines/text/lora-apply.js +17 -0
  340. package/src/inference/pipelines/text/moe-gpu.d.ts +13 -0
  341. package/src/inference/pipelines/text/moe-gpu.js +445 -16
  342. package/src/inference/pipelines/text/moe-impl.d.ts +13 -3
  343. package/src/inference/pipelines/text/moe-shape-validator.js +13 -1
  344. package/src/inference/pipelines/text/ops.d.ts +29 -2
  345. package/src/inference/pipelines/text/ops.js +19 -0
  346. package/src/inference/pipelines/text/per-layer-inputs.js +5 -0
  347. package/src/inference/pipelines/text/probes.js +4 -5
  348. package/src/inference/pipelines/text/sampling-config.d.ts +4 -10
  349. package/src/inference/pipelines/text/sampling-config.js +103 -64
  350. package/src/inference/pipelines/text/sampling.js +13 -9
  351. package/src/inference/pipelines/text/state.js +17 -0
  352. package/src/inference/pipelines/text/tsir-fixture-writer.d.ts +57 -0
  353. package/src/inference/pipelines/text/tsir-fixture-writer.js +8 -9
  354. package/src/inference/pipelines/text/types.d.ts +84 -3
  355. package/src/inference/pipelines/text/weights.d.ts +4 -2
  356. package/src/inference/pipelines/text/weights.js +4 -5
  357. package/src/inference/pipelines/text.d.ts +24 -0
  358. package/src/inference/pipelines/text.js +162 -38
  359. package/src/inference/pipelines/vision/gemma4.js +20 -9
  360. package/src/inference/tokenizer.d.ts +30 -0
  361. package/src/inference/tokenizer.js +340 -164
  362. package/src/inference/tokenizers/bundled.js +71 -49
  363. package/src/loader/doppler-loader.d.ts +11 -0
  364. package/src/loader/doppler-loader.js +290 -15
  365. package/src/loader/embedding-limit-preflight.d.ts +31 -0
  366. package/src/loader/embedding-limit-preflight.js +216 -0
  367. package/src/loader/embedding-loader.d.ts +8 -2
  368. package/src/loader/embedding-loader.js +263 -8
  369. package/src/loader/experts/expert-loader.d.ts +4 -1
  370. package/src/loader/experts/expert-loader.js +174 -5
  371. package/src/loader/final-weights-loader.d.ts +5 -0
  372. package/src/loader/final-weights-loader.js +167 -4
  373. package/src/loader/layer-loader.d.ts +2 -0
  374. package/src/loader/layer-loader.js +74 -6
  375. package/src/loader/loader-types.d.ts +59 -2
  376. package/src/loader/manifest-config.d.ts +8 -0
  377. package/src/loader/shard-resolver.js +37 -4
  378. package/src/loader/tensors/functional-descriptor-loader.d.ts +9 -0
  379. package/src/loader/tensors/functional-descriptor-loader.js +377 -0
  380. package/src/loader/tensors/tensor-loader.d.ts +36 -1
  381. package/src/loader/tensors/tensor-loader.js +437 -21
  382. package/src/loader/tensors/tensor-reader.d.ts +4 -1
  383. package/src/loader/tensors/tensor-reader.js +11 -4
  384. package/src/loader/weights.d.ts +3 -1
  385. package/src/memory/buffer-pool.js +12 -9
  386. package/src/models/diffusiongemma.d.ts +7 -0
  387. package/src/models/diffusiongemma.js +25 -0
  388. package/src/models/gemma4.js +24 -0
  389. package/src/models/qwen3.js +2 -2
  390. package/src/rules/execution-rules-contract-check.js +93 -10
  391. package/src/rules/inference/capability-transforms.rules.json +26 -2
  392. package/src/rules/inference/config.rules.json +1 -0
  393. package/src/rules/inference/dtype.rules.json +7 -0
  394. package/src/rules/inference/execution.rules.json +108 -17
  395. package/src/rules/inference/ffn.rules.json +18 -0
  396. package/src/rules/kernels/gather.rules.json +81 -0
  397. package/src/rules/kernels/matmul.rules.json +19 -0
  398. package/src/rules/kernels/moe.rules.json +3 -0
  399. package/src/rules/kernels/rep-penalty.rules.json +6 -0
  400. package/src/rules/kernels/rmsnorm-qk.rules.json +6 -0
  401. package/src/rules/kernels/rmsnorm.rules.json +42 -7
  402. package/src/rules/kernels/rope-qk.rules.json +6 -0
  403. package/src/rules/loader/tensor-loader.rules.json +8 -0
  404. package/src/rules/rule-registry.js +6 -0
  405. package/src/storage/artifact-storage-context.d.ts +1 -0
  406. package/src/storage/artifact-storage-context.js +10 -0
  407. package/src/tooling/browser-command-runner.d.ts +0 -3
  408. package/src/tooling/browser-command-runner.js +2 -14
  409. package/src/tooling/command-api-family-normalizers.js +4 -2
  410. package/src/tooling/command-api-helpers.js +2 -1
  411. package/src/tooling/command-api.d.ts +2 -0
  412. package/src/tooling/command-api.js +1 -0
  413. package/src/tooling/command-runner-shared.d.ts +4 -0
  414. package/src/tooling/command-runner-shared.js +2 -1
  415. package/src/tooling/diagnose-runner.js +2 -0
  416. package/src/tooling/litert-package-runtime.js +553 -16
  417. package/src/tooling/node-browser-command-runner.js +10 -4
  418. package/src/tooling/node-command-runner.js +7 -2
  419. package/src/tooling/node-convert-worker-pool.js +3 -0
  420. package/src/tooling/node-convert-worker.js +1 -0
  421. package/src/tooling/node-converter.js +12 -0
  422. package/src/tooling/node-webgpu.d.ts +1 -0
  423. package/src/tooling/node-webgpu.js +23 -11
  424. package/src/tooling/program-bundle.d.ts +52 -1
  425. package/src/tooling/rdrr-integrity-refresh.d.ts +18 -0
  426. package/src/tooling/source-artifact-adapter.js +7 -2
  427. package/src/tooling/source-package-profiles.d.ts +18 -1
  428. package/src/tooling/source-package-profiles.js +109 -1
  429. package/src/tooling/source-runtime-bundle.d.ts +1 -0
  430. package/src/tooling/source-runtime-bundle.js +27 -0
  431. package/src/tooling/source-runtime-converter-config.js +12 -3
  432. package/src/tooling-experimental-exports.shared.d.ts +10 -0
  433. package/src/tooling-experimental-exports.shared.js +10 -0
@@ -170,9 +170,77 @@ export function normalizeStorageQuant(value) {
170
170
  if (lower === 'fp32' || lower === 'float32') return 'f32';
171
171
  if (lower === 'bfloat16') return 'bf16';
172
172
  if (lower === 'q4_k_m' || lower === 'q4km') return 'q4k';
173
+ if (lower === 'q4_0' || lower === 'q4-0') return 'q4_0';
174
+ if (
175
+ lower === 'w4a16-ct'
176
+ || lower === 'w4a16_ct'
177
+ || lower === 'compressed-tensors-w4a16'
178
+ || lower === 'compressed_tensors_w4a16'
179
+ ) return 'w4a16';
180
+ if (lower === 'wna8-o8' || lower === 'wna8_o8') return 'wna8o8';
173
181
  return lower;
174
182
  }
175
183
 
184
+ const SOURCE_PACKED_QUANT_DTYPES = new Set(['q4_0', 'w4a16', 'wna8o8']);
185
+
186
+ const SOURCE_PACKED_MANIFEST_DTYPES = {
187
+ q4_0: 'Q4_0',
188
+ w4a16: 'W4A16',
189
+ wna8o8: 'WNA8O8',
190
+ };
191
+
192
+ const SOURCE_PACKED_STORAGE_DESCRIPTORS = {
193
+ q4_0: {
194
+ packing: 'q4_0',
195
+ blockShape: [32],
196
+ blockBytes: 18,
197
+ },
198
+ w4a16: {
199
+ packing: 'w4a16',
200
+ blockShape: [32],
201
+ blockBytes: 16,
202
+ },
203
+ };
204
+
205
+ const COMPRESSED_TENSORS_W4A16_SUFFIXES = {
206
+ packed: '.weight_packed',
207
+ scale: '.weight_scale',
208
+ shape: '.weight_shape',
209
+ };
210
+
211
+ function cloneSourcePackedStorageDescriptor(targetQuant) {
212
+ const descriptor = SOURCE_PACKED_STORAGE_DESCRIPTORS[targetQuant];
213
+ if (!descriptor) return null;
214
+ return {
215
+ ...descriptor,
216
+ blockShape: [...descriptor.blockShape],
217
+ ...(Array.isArray(descriptor.companions)
218
+ ? { companions: descriptor.companions.map((companion) => ({ ...companion })) }
219
+ : {}),
220
+ };
221
+ }
222
+
223
+ function resolveExplicitRoleQuant(tensor, quantizationInfo) {
224
+ if (!quantizationInfo || typeof quantizationInfo !== 'object') {
225
+ return null;
226
+ }
227
+ const role = resolveTensorRole(tensor);
228
+ if (role === 'embedding') {
229
+ return normalizeStorageQuant(quantizationInfo.embeddings ?? null);
230
+ }
231
+ if (role === 'lm_head') {
232
+ return normalizeStorageQuant(
233
+ quantizationInfo.lmHead
234
+ ?? quantizationInfo.embeddings
235
+ ?? null
236
+ );
237
+ }
238
+ if (role === 'matmul' || role === 'expert' || role === 'router') {
239
+ return normalizeStorageQuant(quantizationInfo.weights ?? null);
240
+ }
241
+ return null;
242
+ }
243
+
176
244
  export function resolveTensorTargetQuant(tensorOrName, fallbackQuant, quantizationInfo) {
177
245
  const fallback = normalizeStorageQuant(fallbackQuant);
178
246
  if (!quantizationInfo || typeof quantizationInfo !== 'object') {
@@ -214,30 +282,45 @@ function shouldExcludeTextOnlyTensor(name) {
214
282
  const lower = name.toLowerCase();
215
283
  return lower.startsWith('vision_tower.')
216
284
  || lower.startsWith('model.vision_tower.')
285
+ || lower.startsWith('model.encoder.vision_tower.')
217
286
  || lower.startsWith('vision_model.')
218
287
  || lower.startsWith('model.vision_model.')
288
+ || lower.startsWith('model.encoder.vision_model.')
219
289
  || lower.startsWith('visual.')
220
290
  || lower.startsWith('model.visual.')
291
+ || lower.startsWith('model.encoder.visual.')
221
292
  || lower.startsWith('embed_vision.')
222
293
  || lower.startsWith('model.embed_vision.')
294
+ || lower.startsWith('model.encoder.embed_vision.')
223
295
  || lower.startsWith('vision.')
224
296
  || lower.startsWith('model.vision.')
297
+ || lower.startsWith('model.encoder.vision.')
225
298
  || lower.startsWith('vision_encoder.')
299
+ || lower.startsWith('model.encoder.vision_encoder.')
226
300
  || lower.startsWith('image_encoder.')
301
+ || lower.startsWith('model.encoder.image_encoder.')
227
302
  || lower.startsWith('image_tower.')
303
+ || lower.startsWith('model.encoder.image_tower.')
228
304
  || lower.startsWith('image.')
229
305
  || lower.startsWith('model.image.')
306
+ || lower.startsWith('model.encoder.image.')
230
307
  || lower.startsWith('audio_tower.')
231
308
  || lower.startsWith('model.audio_tower.')
309
+ || lower.startsWith('model.encoder.audio_tower.')
232
310
  || lower.startsWith('audio_model.')
233
311
  || lower.startsWith('model.audio_model.')
312
+ || lower.startsWith('model.encoder.audio_model.')
234
313
  || lower.startsWith('audio.')
235
314
  || lower.startsWith('model.audio.')
315
+ || lower.startsWith('model.encoder.audio.')
236
316
  || lower.startsWith('audio_encoder.')
317
+ || lower.startsWith('model.encoder.audio_encoder.')
237
318
  || lower.startsWith('multi_modal_projector.')
238
319
  || lower.startsWith('model.multi_modal_projector.')
320
+ || lower.startsWith('model.encoder.multi_modal_projector.')
239
321
  || lower.startsWith('mm_projector.')
240
- || lower.startsWith('model.mm_projector.');
322
+ || lower.startsWith('model.mm_projector.')
323
+ || lower.startsWith('model.encoder.mm_projector.');
241
324
  }
242
325
 
243
326
  function resolveConversionTensors(model, converterConfig) {
@@ -275,15 +358,212 @@ function resolveConversionTensors(model, converterConfig) {
275
358
  ));
276
359
  }
277
360
 
361
+ function normalizePositiveIntegerShape(value, tensorName) {
362
+ if (!Array.isArray(value) || value.length === 0) {
363
+ throw new Error(`Compressed-tensors W4A16 tensor "${tensorName}" is missing a logical shape.`);
364
+ }
365
+ return value.map((entry, index) => {
366
+ const number = Number(entry);
367
+ if (!Number.isInteger(number) || number <= 0) {
368
+ throw new Error(
369
+ `Compressed-tensors W4A16 tensor "${tensorName}" has invalid shape[${index}]=${JSON.stringify(entry)}.`
370
+ );
371
+ }
372
+ return number;
373
+ });
374
+ }
375
+
376
+ function resolveCompressedTensorBaseName(name, suffix) {
377
+ return name.endsWith(suffix)
378
+ ? name.slice(0, -suffix.length)
379
+ : null;
380
+ }
381
+
382
+ function collectCompressedTensorsW4A16(tensors) {
383
+ const groups = new Map();
384
+ const groupFor = (baseName) => {
385
+ const existing = groups.get(baseName);
386
+ if (existing) return existing;
387
+ const created = {
388
+ baseName,
389
+ packed: null,
390
+ scale: null,
391
+ shape: null,
392
+ };
393
+ groups.set(baseName, created);
394
+ return created;
395
+ };
396
+
397
+ for (const tensor of tensors) {
398
+ const name = normalizeTensorName(tensor);
399
+ const packedBase = resolveCompressedTensorBaseName(name, COMPRESSED_TENSORS_W4A16_SUFFIXES.packed);
400
+ if (packedBase) {
401
+ groupFor(packedBase).packed = tensor;
402
+ continue;
403
+ }
404
+ const scaleBase = resolveCompressedTensorBaseName(name, COMPRESSED_TENSORS_W4A16_SUFFIXES.scale);
405
+ if (scaleBase) {
406
+ groupFor(scaleBase).scale = tensor;
407
+ continue;
408
+ }
409
+ const shapeBase = resolveCompressedTensorBaseName(name, COMPRESSED_TENSORS_W4A16_SUFFIXES.shape);
410
+ if (shapeBase) {
411
+ groupFor(shapeBase).shape = tensor;
412
+ }
413
+ }
414
+
415
+ return groups;
416
+ }
417
+
418
+ function assertCompressedTensorsW4A16Group(group) {
419
+ const missing = [];
420
+ if (!group.packed) missing.push('weight_packed');
421
+ if (!group.scale) missing.push('weight_scale');
422
+ if (!group.shape) missing.push('weight_shape');
423
+ if (missing.length > 0) {
424
+ throw new Error(
425
+ `Compressed-tensors W4A16 tensor "${group.baseName}.weight" is missing companion tensors: ${missing.join(', ')}.`
426
+ );
427
+ }
428
+ const packedShape = normalizePositiveIntegerShape(group.packed.shape, group.packed.name);
429
+ if (packedShape.length !== 2) {
430
+ throw new Error(
431
+ `Compressed-tensors W4A16 tensor "${group.packed.name}" must be 2D; got shape ${JSON.stringify(packedShape)}.`
432
+ );
433
+ }
434
+ const scaleShape = normalizePositiveIntegerShape(group.scale.shape, group.scale.name);
435
+ if (scaleShape.length !== 2) {
436
+ throw new Error(
437
+ `Compressed-tensors W4A16 scale tensor "${group.scale.name}" must be 2D; got shape ${JSON.stringify(scaleShape)}.`
438
+ );
439
+ }
440
+ normalizePositiveIntegerShape(group.shape.shape, group.shape.name);
441
+ const shapeDtype = String(group.shape.dtype || '').toUpperCase();
442
+ if (shapeDtype !== 'I64' && shapeDtype !== 'I32' && shapeDtype !== 'U32') {
443
+ throw new Error(
444
+ `Compressed-tensors W4A16 shape tensor "${group.shape.name}" must use I64, I32, or U32; got "${group.shape.dtype}".`
445
+ );
446
+ }
447
+ }
448
+
449
+ const COMPRESSED_TENSORS_W4A16_PACKED_VALUES_PER_ELEMENT = {
450
+ U8: 2,
451
+ I8: 2,
452
+ U16: 4,
453
+ I16: 4,
454
+ U32: 8,
455
+ I32: 8,
456
+ };
457
+
458
+ function inferCompressedTensorsW4A16LogicalShape(group) {
459
+ const packedShape = normalizePositiveIntegerShape(group.packed.shape, group.packed.name);
460
+ const packedDtype = String(group.packed.dtype || '').toUpperCase();
461
+ const valuesPerElement = COMPRESSED_TENSORS_W4A16_PACKED_VALUES_PER_ELEMENT[packedDtype];
462
+ if (!valuesPerElement) {
463
+ throw new Error(
464
+ `Compressed-tensors W4A16 tensor "${group.packed.name}" has unsupported packed dtype "${group.packed.dtype}".`
465
+ );
466
+ }
467
+ return [packedShape[0], packedShape[1] * valuesPerElement];
468
+ }
469
+
470
+ function shouldNormalizeCompressedTensorsW4A16(converterConfig) {
471
+ return normalizeStorageQuant(converterConfig?.quantization?.weights) === 'w4a16'
472
+ || normalizeStorageQuant(converterConfig?.quantization?.sourceQuantizationTarget) === 'w4a16';
473
+ }
474
+
475
+ function sortTensorsForConversion(tensors) {
476
+ return [...tensors].sort((left, right) => normalizeTensorName(left).localeCompare(normalizeTensorName(right)));
477
+ }
478
+
479
+ function normalizeCompressedTensorsW4A16(tensors, converterConfig) {
480
+ const groups = collectCompressedTensorsW4A16(tensors);
481
+ if (groups.size === 0) {
482
+ return tensors;
483
+ }
484
+ if (!shouldNormalizeCompressedTensorsW4A16(converterConfig)) {
485
+ throw new Error(
486
+ 'Compressed-tensors W4A16 tensors were detected, but converter.quantization.weights or ' +
487
+ 'converter.quantization.sourceQuantizationTarget is not "w4a16".'
488
+ );
489
+ }
490
+
491
+ const byName = new Map(tensors.map((tensor) => [normalizeTensorName(tensor), tensor]));
492
+ const consumed = new Set();
493
+ const companionByName = new Map();
494
+ const synthetic = [];
495
+ const sortedGroups = [...groups.values()].sort((left, right) => left.baseName.localeCompare(right.baseName));
496
+ for (const group of sortedGroups) {
497
+ assertCompressedTensorsW4A16Group(group);
498
+ const logicalName = `${group.baseName}.weight`;
499
+ if (byName.has(logicalName)) {
500
+ throw new Error(
501
+ `Compressed-tensors W4A16 logical tensor "${logicalName}" conflicts with an existing source tensor.`
502
+ );
503
+ }
504
+ consumed.add(group.packed.name);
505
+ companionByName.set(group.scale.name, {
506
+ ...group.scale,
507
+ compressedTensorsW4A16Companion: {
508
+ role: 'scales',
509
+ primary: logicalName,
510
+ },
511
+ });
512
+ companionByName.set(group.shape.name, {
513
+ ...group.shape,
514
+ compressedTensorsW4A16Companion: {
515
+ role: 'shape',
516
+ primary: logicalName,
517
+ },
518
+ });
519
+ synthetic.push({
520
+ ...group.packed,
521
+ name: logicalName,
522
+ dtype: 'W4A16',
523
+ shape: inferCompressedTensorsW4A16LogicalShape(group),
524
+ packedSourceName: group.packed.name,
525
+ compressedTensorsW4A16: {
526
+ packed: group.packed.name,
527
+ scales: group.scale.name,
528
+ shape: group.shape.name,
529
+ },
530
+ storage: {
531
+ packing: 'w4a16',
532
+ blockShape: [32],
533
+ blockBytes: 16,
534
+ companions: [
535
+ { role: 'scales', tensorId: group.scale.name },
536
+ { role: 'shape', tensorId: group.shape.name },
537
+ ],
538
+ },
539
+ });
540
+ }
541
+
542
+ return sortTensorsForConversion([
543
+ ...tensors
544
+ .filter((tensor) => !consumed.has(normalizeTensorName(tensor)))
545
+ .map((tensor) => companionByName.get(normalizeTensorName(tensor)) ?? tensor),
546
+ ...synthetic,
547
+ ]);
548
+ }
549
+
550
+ function isCompressedTensorsW4A16CompanionTensor(tensor) {
551
+ return Boolean(
552
+ tensor?.compressedTensorsW4A16Companion
553
+ && typeof tensor.compressedTensorsW4A16Companion === 'object'
554
+ );
555
+ }
556
+
278
557
  function shouldMaterializeTiedLmHead(tensors, options) {
279
558
  if (options?.inference?.output?.tieWordEmbeddings !== true) {
280
559
  return false;
281
560
  }
282
- if (normalizeStorageQuant(options?.quantizationInfo?.lmHead ?? null) !== 'q4k') {
561
+ const lmHeadQuant = normalizeStorageQuant(options?.quantizationInfo?.lmHead ?? null);
562
+ if (!SOURCE_PACKED_QUANT_DTYPES.has(lmHeadQuant) && lmHeadQuant !== 'q4k') {
283
563
  return false;
284
564
  }
285
565
  const embeddingQuant = normalizeStorageQuant(options?.quantizationInfo?.embeddings ?? null);
286
- if (embeddingQuant === 'q4k') {
566
+ if (embeddingQuant === lmHeadQuant) {
287
567
  return false;
288
568
  }
289
569
  return !tensors.some((tensor) => resolveTensorRole(tensor) === 'lm_head');
@@ -300,6 +580,15 @@ function resolveTiedLmHeadName(embeddingName) {
300
580
  if (name === 'model.language_model.embed_tokens.weight') {
301
581
  return 'model.language_model.lm_head.weight';
302
582
  }
583
+ if (name === 'model.decoder.embed_tokens.weight') {
584
+ return 'lm_head.weight';
585
+ }
586
+ if (name === 'decoder.embed_tokens.weight') {
587
+ return 'lm_head.weight';
588
+ }
589
+ if (name === 'model.encoder.language_model.embed_tokens.weight') {
590
+ return 'lm_head.weight';
591
+ }
303
592
  if (name === 'language_model.embed_tokens.weight') {
304
593
  return 'language_model.lm_head.weight';
305
594
  }
@@ -536,6 +825,9 @@ function resolveMoEExpertFormat(rawConfig, resolvedModelType, quantizationInfo,
536
825
  if (modelType.includes('gpt_oss') || modelType.includes('gpt-oss') || modelType.includes('gptoss')) {
537
826
  return 'gpt-oss';
538
827
  }
828
+ if (modelType === 'diffusion_gemma' || modelType === 'diffusion_gemma_text') {
829
+ return 'gemma4';
830
+ }
539
831
  return 'mixtral';
540
832
  }
541
833
 
@@ -544,7 +836,10 @@ function normalizeMoEConfig(config, contextLabel) {
544
836
  const numExperts = Number(config.numExperts);
545
837
  const numExpertsPerToken = Number(config.numExpertsPerToken);
546
838
  const expertFormat = String(config.expertFormat || '').trim();
547
- const allowedExpertFormat = expertFormat === 'gpt-oss' || expertFormat === 'mixtral';
839
+ const expertIntermediateSize = config.expertIntermediateSize == null
840
+ ? null
841
+ : Number(config.expertIntermediateSize);
842
+ const allowedExpertFormat = expertFormat === 'gpt-oss' || expertFormat === 'mixtral' || expertFormat === 'gemma4';
548
843
  if (!Number.isFinite(numExperts) || numExperts <= 0) {
549
844
  throw new Error(`Invalid moeConfig.numExperts for ${contextLabel}`);
550
845
  }
@@ -557,10 +852,20 @@ function normalizeMoEConfig(config, contextLabel) {
557
852
  if (!allowedExpertFormat) {
558
853
  throw new Error(`Invalid moeConfig.expertFormat for ${contextLabel}: "${expertFormat}"`);
559
854
  }
855
+ if (
856
+ expertIntermediateSize != null
857
+ && (!Number.isFinite(expertIntermediateSize) || expertIntermediateSize <= 0)
858
+ ) {
859
+ throw new Error(`Invalid moeConfig.expertIntermediateSize for ${contextLabel}`);
860
+ }
861
+ if (expertFormat === 'gemma4' && expertIntermediateSize == null) {
862
+ throw new Error(`Invalid moeConfig for ${contextLabel}: gemma4 experts require expertIntermediateSize`);
863
+ }
560
864
  return {
561
865
  numExperts,
562
866
  numExpertsPerToken,
563
867
  expertFormat,
868
+ ...(expertIntermediateSize == null ? {} : { expertIntermediateSize }),
564
869
  };
565
870
  }
566
871
 
@@ -605,9 +910,17 @@ export function resolveManifestMoEConfig(model, options, rawConfig, resolvedMode
605
910
  options?.quantizationInfo ?? null,
606
911
  null
607
912
  );
913
+ const expertIntermediateSize = expertFormat === 'gemma4'
914
+ ? resolveMoEConfigNumber(rawConfig, 'moe_intermediate_size', 'expert_intermediate_size')
915
+ : null;
608
916
 
609
917
  return normalizeMoEConfig(
610
- { numExperts, numExpertsPerToken, expertFormat },
918
+ {
919
+ numExperts,
920
+ numExpertsPerToken,
921
+ expertFormat,
922
+ ...(expertIntermediateSize == null ? {} : { expertIntermediateSize }),
923
+ },
611
924
  options?.modelId ?? 'model'
612
925
  );
613
926
  }
@@ -978,11 +1291,74 @@ export function transformTensorBytes(tensor, rawData, options = {}) {
978
1291
  return buildInt4PerRowPleTransform(tensor, tensorDataInput, sourceDtype, options);
979
1292
  }
980
1293
 
1294
+ if (isCompressedTensorsW4A16CompanionTensor(tensor)) {
1295
+ return {
1296
+ tensorData,
1297
+ outDtype: sourceDtype,
1298
+ outLayout: null,
1299
+ sourceDtype,
1300
+ tensorTargetQuant: null,
1301
+ };
1302
+ }
1303
+
1304
+ if (SOURCE_PACKED_QUANT_DTYPES.has(tensorTargetQuant)) {
1305
+ const sourceQuant = normalizeStorageQuant(sourceDtype);
1306
+ if (sourceQuant === tensorTargetQuant) {
1307
+ const descriptor = cloneSourcePackedStorageDescriptor(tensorTargetQuant);
1308
+ const sourceCompanions = Array.isArray(tensor?.storage?.companions)
1309
+ ? tensor.storage.companions.map((companion) => ({ ...companion }))
1310
+ : null;
1311
+ return {
1312
+ tensorData,
1313
+ outDtype: SOURCE_PACKED_MANIFEST_DTYPES[tensorTargetQuant],
1314
+ outLayout: null,
1315
+ sourceDtype,
1316
+ tensorTargetQuant,
1317
+ storage: {
1318
+ ...descriptor,
1319
+ ...(sourceCompanions ? { companions: sourceCompanions } : {}),
1320
+ },
1321
+ };
1322
+ }
1323
+ const roleQuant = resolveExplicitRoleQuant(tensor, quantizationInfo);
1324
+ const requiresPackedSource = (
1325
+ forceQuantizeDecision
1326
+ ?? (
1327
+ roleQuant === tensorTargetQuant
1328
+ || shouldQuantize(tensor.name, tensor.shape, {
1329
+ quantizeEmbeddings,
1330
+ modulesToNotConvert,
1331
+ role: tensor.role ?? null,
1332
+ })
1333
+ )
1334
+ );
1335
+ if (!requiresPackedSource) {
1336
+ return {
1337
+ tensorData,
1338
+ outDtype: sourceDtype,
1339
+ outLayout: null,
1340
+ sourceDtype,
1341
+ tensorTargetQuant,
1342
+ };
1343
+ }
1344
+ throw new Error(
1345
+ `Cannot materialize ${tensorTargetQuant} for ${tensor.name}: ` +
1346
+ `native import requires source dtype ${SOURCE_PACKED_MANIFEST_DTYPES[tensorTargetQuant]}; ` +
1347
+ 'the converter does not re-quantize tensors into this packed format.'
1348
+ );
1349
+ }
1350
+
981
1351
  if (tensorTargetQuant === 'q4k') {
982
1352
  const sourceQuant = normalizeStorageQuant(sourceDtype);
1353
+ const tensorRole = resolveTensorRole(tensor);
1354
+ const isMatrixLikeShape = Array.isArray(tensor.shape) && tensor.shape.length >= 2;
1355
+ const is2DMatrixShape = Array.isArray(tensor.shape) && tensor.shape.length === 2;
1356
+ const useQ4KRowWise = isMatrixLikeShape
1357
+ && q4kLayout === 'row'
1358
+ && (is2DMatrixShape || tensorRole === 'expert');
983
1359
  if (sourceQuant === 'q4k') {
984
1360
  outDtype = 'Q4_K_M';
985
- if (Array.isArray(tensor.shape) && tensor.shape.length === 2) {
1361
+ if (is2DMatrixShape) {
986
1362
  outLayout = q4kLayout;
987
1363
  }
988
1364
  return {
@@ -1004,15 +1380,17 @@ export function transformTensorBytes(tensor, rawData, options = {}) {
1004
1380
  if (shouldQuantizeTensor) {
1005
1381
  const f32Data = toFloat32ForQ4K(tensorData, sourceDtype, tensor.name);
1006
1382
  const quantized = (
1007
- Array.isArray(tensor.shape) && tensor.shape.length === 2
1383
+ is2DMatrixShape
1008
1384
  ? (q4kLayout === 'col'
1009
1385
  ? quantizeToQ4KMColumnWise(f32Data, tensor.shape)
1010
1386
  : quantizeToQ4KMRowWise(f32Data, tensor.shape))
1387
+ : useQ4KRowWise
1388
+ ? quantizeToQ4KMRowWise(f32Data, tensor.shape)
1011
1389
  : quantizeToQ4KM(f32Data, tensor.shape)
1012
1390
  );
1013
1391
  tensorData = quantized.quantized;
1014
1392
  outDtype = 'Q4_K_M';
1015
- if (Array.isArray(tensor.shape) && tensor.shape.length === 2) {
1393
+ if (is2DMatrixShape || useQ4KRowWise) {
1016
1394
  outLayout = q4kLayout;
1017
1395
  }
1018
1396
  } else if (sourceDtype === 'BF16') {
@@ -1333,12 +1711,37 @@ function getNestedTextConfig(config) {
1333
1711
  function resolveGemma4TextConfig(rawConfig) {
1334
1712
  const textConfig = getNestedTextConfig(rawConfig);
1335
1713
  const modelType = String(textConfig?.model_type ?? rawConfig?.model_type ?? '').trim().toLowerCase();
1336
- if (modelType !== 'gemma4' && modelType !== 'gemma4_text') {
1714
+ if (
1715
+ modelType !== 'gemma4'
1716
+ && modelType !== 'gemma4_text'
1717
+ && modelType !== 'gemma4_unified'
1718
+ && modelType !== 'gemma4_unified_text'
1719
+ ) {
1337
1720
  return null;
1338
1721
  }
1339
1722
  return textConfig ?? rawConfig ?? null;
1340
1723
  }
1341
1724
 
1725
+ function resolveDiffusionGemmaConfig(rawConfig, resolvedModelType = null) {
1726
+ const textConfig = getNestedTextConfig(rawConfig);
1727
+ const modelType = String(
1728
+ resolvedModelType
1729
+ ?? rawConfig?.model_type
1730
+ ?? textConfig?.model_type
1731
+ ?? ''
1732
+ ).trim().toLowerCase();
1733
+ const textModelType = String(textConfig?.model_type ?? '').trim().toLowerCase();
1734
+ if (
1735
+ modelType !== 'diffusion_gemma'
1736
+ && modelType !== 'diffusion_gemma_text'
1737
+ && textModelType !== 'diffusion_gemma'
1738
+ && textModelType !== 'diffusion_gemma_text'
1739
+ ) {
1740
+ return null;
1741
+ }
1742
+ return textConfig ?? rawConfig ?? {};
1743
+ }
1744
+
1342
1745
  function collectGemma4UnsupportedTensorFlags(tensors) {
1343
1746
  const names = Array.isArray(tensors) ? tensors.map((tensor) => normalizeTensorName(tensor).toLowerCase()) : [];
1344
1747
  const flags = [];
@@ -1410,6 +1813,143 @@ function assertSupportedGemma4Conversion(model, tensors, modelId) {
1410
1813
  );
1411
1814
  }
1412
1815
 
1816
+ function readDiffusionGemmaPositiveInteger(value, label, modelId) {
1817
+ if (!Number.isInteger(value) || value <= 0) {
1818
+ throw new Error(`DiffusionGemma model "${modelId}" requires ${label} to be a positive integer.`);
1819
+ }
1820
+ return value;
1821
+ }
1822
+
1823
+ function readDiffusionGemmaNonNegativeNumber(value, label, modelId) {
1824
+ if (typeof value !== 'number' || !Number.isFinite(value) || value < 0) {
1825
+ throw new Error(`DiffusionGemma model "${modelId}" requires ${label} to be a non-negative finite number.`);
1826
+ }
1827
+ return value;
1828
+ }
1829
+
1830
+ function readDiffusionGemmaEntropyBound(generationConfig, modelId) {
1831
+ if (generationConfig && Object.hasOwn(generationConfig, 'entropy_bound')) {
1832
+ return readDiffusionGemmaNonNegativeNumber(
1833
+ generationConfig.entropy_bound,
1834
+ 'generation_config.entropy_bound',
1835
+ modelId
1836
+ );
1837
+ }
1838
+ return readDiffusionGemmaNonNegativeNumber(
1839
+ generationConfig?.sampler_config?.entropy_bound,
1840
+ 'generation_config.sampler_config.entropy_bound',
1841
+ modelId
1842
+ );
1843
+ }
1844
+
1845
+ function readDiffusionGemmaNullableTokenId(value, label, modelId) {
1846
+ if (value == null) return null;
1847
+ if (!Number.isInteger(value) || value < 0) {
1848
+ throw new Error(`DiffusionGemma model "${modelId}" requires ${label} to be null or a non-negative integer.`);
1849
+ }
1850
+ return value;
1851
+ }
1852
+
1853
+ function readDiffusionGemmaTokenId(value, label, modelId) {
1854
+ if (!Number.isInteger(value) || value < 0) {
1855
+ throw new Error(`DiffusionGemma model "${modelId}" requires ${label} to be a non-negative integer.`);
1856
+ }
1857
+ return value;
1858
+ }
1859
+
1860
+ function readDiffusionGemmaEosTokenIds(rawConfig, generationConfig, modelId) {
1861
+ const eos = resolveTokenizerIds(
1862
+ generationConfig?.eos_token_id
1863
+ ?? generationConfig?.eos_token_ids
1864
+ ?? rawConfig?.eos_token_id
1865
+ ?? rawConfig?.eos_token_ids
1866
+ );
1867
+ if (!Array.isArray(eos) || eos.length === 0) {
1868
+ throw new Error(`DiffusionGemma model "${modelId}" requires generation_config.eos_token_id.`);
1869
+ }
1870
+ return eos;
1871
+ }
1872
+
1873
+ function resolveDiffusionGemmaInferenceContract(rawConfig, generationConfig, modelId) {
1874
+ const canvasLength = readDiffusionGemmaPositiveInteger(
1875
+ rawConfig?.canvas_length,
1876
+ 'config.canvas_length',
1877
+ modelId
1878
+ );
1879
+ const maxDenoisingSteps = readDiffusionGemmaPositiveInteger(
1880
+ generationConfig?.max_denoising_steps,
1881
+ 'generation_config.max_denoising_steps',
1882
+ modelId
1883
+ );
1884
+ const maxNewTokens = readDiffusionGemmaPositiveInteger(
1885
+ generationConfig?.max_new_tokens,
1886
+ 'generation_config.max_new_tokens',
1887
+ modelId
1888
+ );
1889
+ const stabilityThreshold = readDiffusionGemmaPositiveInteger(
1890
+ generationConfig?.stability_threshold,
1891
+ 'generation_config.stability_threshold',
1892
+ modelId
1893
+ );
1894
+ const padTokenId = readDiffusionGemmaTokenId(
1895
+ generationConfig?.pad_token_id ?? rawConfig?.pad_token_id,
1896
+ 'generation_config.pad_token_id',
1897
+ modelId
1898
+ );
1899
+ const tMin = readDiffusionGemmaNonNegativeNumber(
1900
+ generationConfig?.t_min,
1901
+ 'generation_config.t_min',
1902
+ modelId
1903
+ );
1904
+ const tMax = readDiffusionGemmaNonNegativeNumber(
1905
+ generationConfig?.t_max,
1906
+ 'generation_config.t_max',
1907
+ modelId
1908
+ );
1909
+ if (tMax < tMin) {
1910
+ throw new Error(`DiffusionGemma model "${modelId}" requires generation_config.t_max >= t_min.`);
1911
+ }
1912
+ return {
1913
+ canvasLength,
1914
+ maxDenoisingSteps,
1915
+ maxNewTokens,
1916
+ tMin,
1917
+ tMax,
1918
+ entropyBound: readDiffusionGemmaEntropyBound(generationConfig, modelId),
1919
+ confidenceThreshold: readDiffusionGemmaNonNegativeNumber(
1920
+ generationConfig?.confidence_threshold,
1921
+ 'generation_config.confidence_threshold',
1922
+ modelId
1923
+ ),
1924
+ stabilityThreshold,
1925
+ padTokenId,
1926
+ eosTokenIds: readDiffusionGemmaEosTokenIds(rawConfig, generationConfig, modelId),
1927
+ boiTokenId: readDiffusionGemmaNullableTokenId(rawConfig?.boi_token_id, 'config.boi_token_id', modelId),
1928
+ eoiTokenId: readDiffusionGemmaNullableTokenId(rawConfig?.eoi_token_id, 'config.eoi_token_id', modelId),
1929
+ imageTokenId: readDiffusionGemmaNullableTokenId(rawConfig?.image_token_id, 'config.image_token_id', modelId),
1930
+ selfConditioning: true,
1931
+ decoderCacheMode: 'encoder_kv_readonly_canvas_concat',
1932
+ router: {
1933
+ scaleHiddenStates: true,
1934
+ normalizeTopK: true,
1935
+ perExpertScale: true,
1936
+ },
1937
+ };
1938
+ }
1939
+
1940
+ function applyDiffusionGemmaInferenceContract(inference, rawConfig, generationConfig, modelId, resolvedModelType) {
1941
+ if (!resolveDiffusionGemmaConfig(rawConfig, resolvedModelType)) {
1942
+ return inference;
1943
+ }
1944
+ if (inference?.diffusionGemma && typeof inference.diffusionGemma === 'object') {
1945
+ return inference;
1946
+ }
1947
+ return {
1948
+ ...inference,
1949
+ diffusionGemma: resolveDiffusionGemmaInferenceContract(rawConfig, generationConfig, modelId),
1950
+ };
1951
+ }
1952
+
1413
1953
 
1414
1954
  export function buildTensorMap(tensors, shardSize) {
1415
1955
  if (!shardSize || shardSize <= 0) {
@@ -1740,6 +2280,13 @@ export function createManifest(
1740
2280
  if (!inference) {
1741
2281
  throw new Error('inference config is required — use a v1 conversion config');
1742
2282
  }
2283
+ inference = applyDiffusionGemmaInferenceContract(
2284
+ inference,
2285
+ rawConfig,
2286
+ generationConfig,
2287
+ modelId,
2288
+ resolvedModelType
2289
+ );
1743
2290
 
1744
2291
  const embeddingOutput = inferEmbeddingOutputConfig(tensorLocations);
1745
2292
  const hasExplicitEmbeddingPostprocessor = Object.prototype.hasOwnProperty.call(
@@ -1779,7 +2326,8 @@ export function createManifest(
1779
2326
  throw new Error('Missing hashAlgorithm for manifest');
1780
2327
  }
1781
2328
 
1782
- const multimodalConfig = isDiffusion
2329
+ const isTextOnlyArtifact = options.textOnly === true;
2330
+ const multimodalConfig = isDiffusion || isTextOnlyArtifact
1783
2331
  ? { vision_config: null, audio_config: null }
1784
2332
  : resolveManifestMultimodalConfig(rawConfig, manifestPolicy);
1785
2333
  const manifestConfig = isDiffusion
@@ -1867,7 +2415,10 @@ export async function convertModel(model, io, options = {}) {
1867
2415
  throw new Error('Missing modelId for conversion');
1868
2416
  }
1869
2417
  const tensors = materializeTiedLmHeadTensor(
1870
- resolveConversionTensors(model, converterConfig),
2418
+ normalizeCompressedTensorsW4A16(
2419
+ resolveConversionTensors(model, converterConfig),
2420
+ converterConfig
2421
+ ),
1871
2422
  {
1872
2423
  inference: options.inference ?? converterConfig?.inference ?? null,
1873
2424
  quantizationInfo: options.quantizationInfo ?? null,
@@ -1996,6 +2547,7 @@ export async function convertModel(model, io, options = {}) {
1996
2547
  const sourceTensorSize = Number.isFinite(tensor?.size) ? Number(tensor.size) : null;
1997
2548
  let outDtype = tensor.dtype;
1998
2549
  let outLayout = null;
2550
+ let tensorStorage = null;
1999
2551
  let tensorSize = 0;
2000
2552
 
2001
2553
  if (
@@ -2027,9 +2579,11 @@ export async function convertModel(model, io, options = {}) {
2027
2579
  }
2028
2580
  const chunkOutDtype = result?.outDtype ?? tensor.dtype;
2029
2581
  const chunkOutLayout = result?.outLayout ?? null;
2582
+ const chunkStorage = result?.storage ?? null;
2030
2583
  if (!emittedChunk) {
2031
2584
  outDtype = chunkOutDtype;
2032
2585
  outLayout = chunkOutLayout;
2586
+ tensorStorage = chunkStorage;
2033
2587
  emittedChunk = true;
2034
2588
  } else {
2035
2589
  if (chunkOutDtype !== outDtype) {
@@ -2038,6 +2592,9 @@ export async function convertModel(model, io, options = {}) {
2038
2592
  if (chunkOutLayout !== outLayout) {
2039
2593
  throw new Error(`Large tensor transformer returned inconsistent layout for ${tensor.name}.`);
2040
2594
  }
2595
+ if (JSON.stringify(chunkStorage) !== JSON.stringify(tensorStorage)) {
2596
+ throw new Error(`Large tensor transformer returned inconsistent storage descriptor for ${tensor.name}.`);
2597
+ }
2041
2598
  }
2042
2599
  tensorSize += tensorData.byteLength;
2043
2600
  await appendTensorBytes(tensorData, tensorSpans);
@@ -2112,6 +2669,7 @@ export async function convertModel(model, io, options = {}) {
2112
2669
  }
2113
2670
  outDtype = transformResult?.outDtype ?? tensor.dtype;
2114
2671
  outLayout = transformResult?.outLayout ?? null;
2672
+ tensorStorage = transformResult?.storage ?? null;
2115
2673
  tensorSize = tensorData.byteLength;
2116
2674
  await appendTensorBytes(tensorData, tensorSpans);
2117
2675
 
@@ -2163,6 +2721,7 @@ export async function convertModel(model, io, options = {}) {
2163
2721
  role,
2164
2722
  group,
2165
2723
  ...(outLayout ? { layout: outLayout } : {}),
2724
+ ...(tensorStorage ? { storage: tensorStorage } : {}),
2166
2725
  ...(pleSourceTransform ? { sourceTransform: pleSourceTransform } : {}),
2167
2726
  };
2168
2727
  } else {
@@ -2174,6 +2733,7 @@ export async function convertModel(model, io, options = {}) {
2174
2733
  role,
2175
2734
  group,
2176
2735
  ...(outLayout ? { layout: outLayout } : {}),
2736
+ ...(tensorStorage ? { storage: tensorStorage } : {}),
2177
2737
  ...(pleSourceTransform ? { sourceTransform: pleSourceTransform } : {}),
2178
2738
  };
2179
2739
  }
@@ -2234,15 +2794,17 @@ export async function convertModel(model, io, options = {}) {
2234
2794
  modelType: options.modelType,
2235
2795
  quantization: effectiveManifestQuantization,
2236
2796
  quantizationInfo: effectiveQuantizationInfo,
2797
+ moeConfig: converterConfig?.moeConfig ?? options.moeConfig ?? null,
2237
2798
  artifactIdentity,
2238
2799
  weightsRef: converterConfig?.manifest?.weightsRef ?? options.weightsRef ?? null,
2239
2800
  hashAlgorithm: converterConfig.manifest.hashAlgorithm,
2240
2801
  architecture: options.architecture,
2241
2802
  inference: options.inference,
2242
- eosTokenId: options.eosTokenId,
2803
+ eosTokenId: converterConfig?.manifest?.eosTokenId ?? options.eosTokenId,
2243
2804
  convertedAt: converterConfig?.manifest?.conversion?.convertedAt ?? null,
2244
2805
  conversionInfo: converterConfig?.manifest?.conversion ?? null,
2245
2806
  manifestConfig: converterConfig?.manifest ?? null,
2807
+ textOnly: converterConfig?.output?.textOnly === true,
2246
2808
  });
2247
2809
 
2248
2810
  // Write manifest