doppler-gpu 0.4.7 → 0.4.8

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 (610) hide show
  1. package/CHANGELOG.md +64 -0
  2. package/README.md +93 -144
  3. package/models/catalog.json +1575 -0
  4. package/package.json +37 -12
  5. package/src/cli/doppler-cli.js +28 -7
  6. package/src/client/doppler-registry.json +2 -2
  7. package/src/client/runtime/index.js +25 -2
  8. package/src/client/runtime/model-session.d.ts +31 -1
  9. package/src/client/runtime/model-session.js +31 -0
  10. package/src/client/runtime/model-source.d.ts +19 -2
  11. package/src/client/runtime/model-source.js +45 -5
  12. package/src/config/conversion/gemma3/translategemma-4b-1b-enes-q4k-ehf16-af32.json +21 -5
  13. package/src/config/conversion/gemma3/translategemma-4b-it-q4k-ehf16-af32.json +6 -1
  14. package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af16-int4ple.json +1 -1
  15. package/src/config/conversion/qwen3/qwen-3-reranker-0-6b-q4k-ehf16-af32.json +9 -1
  16. package/src/config/kernel-path-loader.js +1 -0
  17. package/src/config/kernels/kernel-ref-digests.js +26 -17
  18. package/src/config/kernels/registry.json +448 -27
  19. package/src/config/merge.js +11 -0
  20. package/src/config/param-validator.js +25 -0
  21. package/src/config/runtime/default.json +3 -1
  22. package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-attn128-candidate.json +20 -0
  23. package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-maxbatch64-candidate.json +52 -0
  24. package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-normed-ffn-candidate.json +18 -0
  25. package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-overlap-ringtokens1-candidate.json +52 -0
  26. package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-parity.json +85 -0
  27. package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-sequential-candidate.json +52 -0
  28. package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-throughput.json +85 -0
  29. package/src/config/runtime/profiles/qwen-3-5-2b-f16-primary-throughput.json +23 -0
  30. package/src/config/runtime/profiles/qwen-3-5-2b-metal-parity.json +38 -0
  31. package/src/config/runtime/profiles/qwen-3-5-2b-metal-throughput.json +135 -0
  32. package/src/config/runtime/profiles/qwen-3-embedding-0-6b-metal-stability.json +37 -0
  33. package/src/config/runtime/profiles/qwen-3-embedding-0-6b-metal-throughput.json +21 -0
  34. package/src/config/runtime/profiles/qwen-3-reranker-0-6b-metal-stability.json +41 -0
  35. package/src/config/runtime/profiles/qwen-3-reranker-0-6b-metal-throughput.json +41 -0
  36. package/src/config/runtime/profiles/rerank-stability.json +43 -0
  37. package/src/config/runtime/profiles/rerank-throughput.json +40 -0
  38. package/src/config/runtime/profiles/translategemma-1b-nativekd2-q4k-rdna3-throughput-probe.json +22 -0
  39. package/src/config/schema/doppler.schema.js +40 -5
  40. package/src/config/schema/execution-v1.schema.d.ts +41 -0
  41. package/src/config/schema/execution-v1.schema.js +10 -0
  42. package/src/config/schema/index.d.ts +3 -1
  43. package/src/config/schema/speculation-self.schema.d.ts +16 -0
  44. package/src/config/transforms/capability-transform-resolver.js +7 -1
  45. package/src/config/transforms/execution-graph-transforms.js +2 -2
  46. package/src/debug/config.d.ts +4 -0
  47. package/src/experimental/orchestration/index.d.ts +10 -1
  48. package/src/experimental/training/suite.js +1 -1
  49. package/src/formats/rdrr/classification.js +3 -0
  50. package/src/formats/rdrr/integrity.d.ts +0 -2
  51. package/src/formats/rdrr/integrity.js +0 -2
  52. package/src/formats/rdrr/types.d.ts +1 -1
  53. package/src/gpu/kernels/attention.d.ts +2 -0
  54. package/src/gpu/kernels/attention.js +75 -16
  55. package/src/gpu/kernels/attention_decode_online_head256_f16kv_output_gate.wgsl +268 -0
  56. package/src/gpu/kernels/attention_head128_f16kv.wgsl +269 -0
  57. package/src/gpu/kernels/fused_ffn.d.ts +32 -0
  58. package/src/gpu/kernels/fused_ffn.js +540 -19
  59. package/src/gpu/kernels/fused_ffn.wgsl +2 -1
  60. package/src/gpu/kernels/fused_ffn_f16.wgsl +2 -1
  61. package/src/gpu/kernels/fused_ffn_f16w_f32a_batched.wgsl +135 -0
  62. package/src/gpu/kernels/fused_ffn_q4k.wgsl +5 -3
  63. package/src/gpu/kernels/fused_ffn_q4k_f16.wgsl +2 -1
  64. package/src/gpu/kernels/fused_ffn_q4k_metal_simd16.wgsl +183 -0
  65. package/src/gpu/kernels/fused_gate_up_gelu_gated_f16.wgsl +2 -1
  66. package/src/gpu/kernels/fused_matmul_q4.wgsl +15 -6
  67. package/src/gpu/kernels/fused_matmul_q4_metal_simd16.wgsl +144 -0
  68. package/src/gpu/kernels/index.d.ts +21 -0
  69. package/src/gpu/kernels/index.js +13 -0
  70. package/src/gpu/kernels/linear-attention-core.js +413 -15
  71. package/src/gpu/kernels/lm-head-argmax.d.ts +10 -1
  72. package/src/gpu/kernels/lm-head-argmax.js +153 -57
  73. package/src/gpu/kernels/lm-head-select-logits.d.ts +23 -0
  74. package/src/gpu/kernels/lm-head-select-logits.js +186 -0
  75. package/src/gpu/kernels/lm_head_argmax_q4k.wgsl +289 -0
  76. package/src/gpu/kernels/lm_head_select_logits.wgsl +77 -0
  77. package/src/gpu/kernels/matmul-selection.js +5 -1
  78. package/src/gpu/kernels/residual_rmsnorm_pair.wgsl +127 -0
  79. package/src/gpu/kernels/rmsnorm-pair.d.ts +33 -0
  80. package/src/gpu/kernels/rmsnorm-pair.js +113 -0
  81. package/src/gpu/kernels/rmsnorm-stats.d.ts +29 -0
  82. package/src/gpu/kernels/rmsnorm-stats.js +250 -0
  83. package/src/gpu/kernels/silu.wgsl +7 -1
  84. package/src/gpu/kernels/silu_f16.wgsl +7 -1
  85. package/src/gpu/kernels/split-qkv-rmsnorm-rope-qk.js +15 -10
  86. package/src/gpu/kernels/split_qkv_rmsnorm_rope_qk.wgsl +37 -7
  87. package/src/gpu/kernels/split_qkv_rmsnorm_rope_qk_f16kv_cache.wgsl +37 -7
  88. package/src/gpu/tensor.d.ts +1 -0
  89. package/src/index.d.ts +1 -1
  90. package/src/index.js +1 -1
  91. package/src/inference/browser-harness-suite-helpers.js +1 -1
  92. package/src/inference/browser-harness-text-helpers.js +251 -20
  93. package/src/inference/browser-harness.js +221 -3
  94. package/src/inference/kv-cache/base.js +42 -1
  95. package/src/inference/pipelines/structured/json-grammar-mask.d.ts +18 -0
  96. package/src/inference/pipelines/text/attention/dispatch-params.js +8 -3
  97. package/src/inference/pipelines/text/attention/output-gate-fusion.js +37 -0
  98. package/src/inference/pipelines/text/attention/projections.js +174 -4
  99. package/src/inference/pipelines/text/attention/record.js +44 -21
  100. package/src/inference/pipelines/text/attention/run.js +18 -3
  101. package/src/inference/pipelines/text/config.d.ts +18 -0
  102. package/src/inference/pipelines/text/config.js +100 -0
  103. package/src/inference/pipelines/text/execution-plan.d.ts +2 -0
  104. package/src/inference/pipelines/text/execution-plan.js +37 -0
  105. package/src/inference/pipelines/text/execution-v1.d.ts +6 -0
  106. package/src/inference/pipelines/text/execution-v1.js +122 -8
  107. package/src/inference/pipelines/text/ffn/dense.js +239 -21
  108. package/src/inference/pipelines/text/ffn/standard.js +121 -38
  109. package/src/inference/pipelines/text/generator-decode-policy.js +1 -0
  110. package/src/inference/pipelines/text/generator-helpers.js +2 -0
  111. package/src/inference/pipelines/text/generator-runtime.js +1 -0
  112. package/src/inference/pipelines/text/generator-steps.js +2 -2
  113. package/src/inference/pipelines/text/generator.d.ts +18 -1
  114. package/src/inference/pipelines/text/generator.js +297 -19
  115. package/src/inference/pipelines/text/init-qkv-fusion.js +155 -62
  116. package/src/inference/pipelines/text/init.d.ts +22 -3
  117. package/src/inference/pipelines/text/init.js +2 -2
  118. package/src/inference/pipelines/text/layer.js +87 -8
  119. package/src/inference/pipelines/text/linear-attention-ab-fusion.js +208 -0
  120. package/src/inference/pipelines/text/linear-attention.d.ts +1 -0
  121. package/src/inference/pipelines/text/linear-attention.js +173 -16
  122. package/src/inference/pipelines/text/logits/gpu.js +2 -2
  123. package/src/inference/pipelines/text/logits/index.js +62 -1
  124. package/src/inference/pipelines/text/model-load.js +6 -0
  125. package/src/inference/pipelines/text/ops.js +22 -0
  126. package/src/inference/pipelines/text/per-layer-inputs.js +80 -11
  127. package/src/inference/pipelines/text/probes.js +9 -4
  128. package/src/inference/pipelines/text/state.d.ts +4 -4
  129. package/src/inference/pipelines/text/state.js +4 -0
  130. package/src/inference/pipelines/text/types.d.ts +27 -2
  131. package/src/inference/pipelines/text.d.ts +21 -1
  132. package/src/inference/pipelines/text.js +40 -14
  133. package/src/inference/tokenizers/bundled.js +45 -23
  134. package/src/inference/tokenizers/types.d.ts +7 -4
  135. package/src/loader/doppler-loader.js +25 -2
  136. package/src/loader/index.d.ts +3 -1
  137. package/src/loader/index.js +2 -0
  138. package/src/loader/loader-types.d.ts +3 -1
  139. package/src/loader/manifest-config.d.ts +18 -0
  140. package/src/loader/manifest-config.js +11 -3
  141. package/src/loader/shard-cache.d.ts +2 -0
  142. package/src/loader/shard-cache.js +29 -1
  143. package/src/models/gemma4.js +8 -0
  144. package/src/rules/inference/capability-transforms.rules.json +145 -20
  145. package/src/rules/inference/ffn.rules.json +33 -0
  146. package/src/rules/kernels/fused-ffn.rules.json +1 -1
  147. package/src/rules/kernels/lm-head-argmax.rules.json +10 -0
  148. package/src/rules/kernels/matmul.rules.json +1 -0
  149. package/src/rules/kernels/silu.rules.json +1 -0
  150. package/src/rules/rule-registry.js +2 -0
  151. package/src/storage/artifact-storage-context.js +36 -0
  152. package/src/tooling/node-browser-command-runner.js +42 -7
  153. package/src/tooling/opfs-cache.d.ts +21 -1
  154. package/src/tooling/opfs-cache.js +168 -32
  155. package/src/tooling-exports/storage.d.ts +1 -1
  156. package/src/tooling-exports/storage.js +1 -1
  157. package/src/types/gpu.d.ts +25 -0
  158. package/src/version.js +1 -1
  159. package/tests/kernels/browser/kernel-suite.js +1593 -0
  160. package/tests/kernels/browser/test-page.js +2629 -0
  161. package/tests/kernels/harness/benchmark.js +225 -0
  162. package/tests/kernels/harness/buffer-utils.js +88 -0
  163. package/tests/kernels/harness/tolerance.js +174 -0
  164. package/tests/kernels/reference/adam.js +15 -0
  165. package/tests/kernels/reference/attention.js +158 -0
  166. package/tests/kernels/reference/backward.js +96 -0
  167. package/tests/kernels/reference/bias-add.js +20 -0
  168. package/tests/kernels/reference/conv2d.js +58 -0
  169. package/tests/kernels/reference/cross-entropy.js +26 -0
  170. package/tests/kernels/reference/dequant.js +312 -0
  171. package/tests/kernels/reference/embed.js +15 -0
  172. package/tests/kernels/reference/gather.js +61 -0
  173. package/tests/kernels/reference/gelu.js +37 -0
  174. package/tests/kernels/reference/groupnorm.js +36 -0
  175. package/tests/kernels/reference/index.js +89 -0
  176. package/tests/kernels/reference/layernorm.js +22 -0
  177. package/tests/kernels/reference/matmul.js +57 -0
  178. package/tests/kernels/reference/modulate.js +24 -0
  179. package/tests/kernels/reference/moe-gather.js +83 -0
  180. package/tests/kernels/reference/pixel-shuffle.js +20 -0
  181. package/tests/kernels/reference/residual.js +55 -0
  182. package/tests/kernels/reference/rmsnorm.js +52 -0
  183. package/tests/kernels/reference/rope.js +79 -0
  184. package/tests/kernels/reference/sample.js +97 -0
  185. package/tests/kernels/reference/scatter-add.js +58 -0
  186. package/tests/kernels/reference/silu.js +52 -0
  187. package/tests/kernels/reference/softmax.js +86 -0
  188. package/tests/kernels/reference/split-qkv.js +58 -0
  189. package/tests/kernels/reference/topk.js +88 -0
  190. package/tests/kernels/reference/upsample2d.js +37 -0
  191. package/BRANDING.md +0 -14
  192. package/SECURITY.md +0 -19
  193. package/assets/doppler-runtime-map.svg +0 -82
  194. package/src/bootstrap.js +0 -30
  195. package/src/cli/doppler-cli.d.ts +0 -46
  196. package/src/cli/doppler-quickstart.d.ts +0 -45
  197. package/src/cli/doppler-serve.d.ts +0 -210
  198. package/src/cli/runtime-profiles.d.ts +0 -40
  199. package/src/client/doppler-api.browser.d.ts +0 -1
  200. package/src/client/doppler-api.browser.js +0 -25
  201. package/src/client/doppler-registry.d.ts +0 -54
  202. package/src/client/runtime/node-quickstart-cache.d.ts +0 -22
  203. package/src/client/runtime/source-runtime.d.ts +0 -26
  204. package/src/client/runtime/storage.d.ts +0 -5
  205. package/src/config/README.md +0 -61
  206. package/src/config/conversion/README.md +0 -62
  207. package/src/config/goal-completion-matrix.json +0 -552
  208. package/src/config/kernels/kernel-ref-digests.d.ts +0 -1
  209. package/src/config/kernels/moe/gpt-oss.paths.json +0 -49
  210. package/src/config/kernels/moe/mixtral.paths.json +0 -46
  211. package/src/config/kernels/registry.d.ts +0 -51
  212. package/src/config/merge-contract-check.d.ts +0 -16
  213. package/src/config/merge-contract-check.js +0 -319
  214. package/src/config/merge-helpers.d.ts +0 -60
  215. package/src/config/model-lanes/catalog-lane-resolver.d.ts +0 -66
  216. package/src/config/model-lanes/catalog-lane-resolver.js +0 -170
  217. package/src/config/param-categories.d.ts +0 -8
  218. package/src/config/param-validator.d.ts +0 -33
  219. package/src/config/platforms/loader.d.ts +0 -38
  220. package/src/config/quantization-contract-check.d.ts +0 -12
  221. package/src/config/quantization-contract-check.js +0 -91
  222. package/src/config/runtime/README.md +0 -28
  223. package/src/config/runtime-merge.d.ts +0 -5
  224. package/src/config/schema/claim-evidence-contract.schema.json +0 -93
  225. package/src/config/schema/command-surface-contract.schema.json +0 -81
  226. package/src/config/schema/goal-completion-matrix.schema.json +0 -223
  227. package/src/config/schema/model-artifact-contract.schema.json +0 -119
  228. package/src/config/schema/platform.schema.js +0 -1
  229. package/src/config/schema/policy-schema-registry.json +0 -31
  230. package/src/config/schema/subsystem-support-contract.schema.json +0 -48
  231. package/src/config/support-tiers/subsystems.json +0 -502
  232. package/src/config/supported-operations.d.ts +0 -14
  233. package/src/config/transforms/capability-transform-resolver.d.ts +0 -56
  234. package/src/config/transforms/execution-graph-transforms.d.ts +0 -306
  235. package/src/converter/conversion-plan.d.ts +0 -56
  236. package/src/converter/index.d.ts +0 -98
  237. package/src/converter/index.js +0 -59
  238. package/src/converter/parsers/diffusion.d.ts +0 -50
  239. package/src/converter/parsers/gguf.d.ts +0 -31
  240. package/src/converter/parsers/index.d.ts +0 -21
  241. package/src/converter/parsers/index.js +0 -12
  242. package/src/converter/parsers/transformer.d.ts +0 -31
  243. package/src/converter/quantization-info.d.ts +0 -38
  244. package/src/converter/rope-config.d.ts +0 -16
  245. package/src/converter/tokenizer-utils.d.ts +0 -12
  246. package/src/debug/reference/README.md +0 -114
  247. package/src/debug/reference/hf_attn_debug.py +0 -114
  248. package/src/debug/reference/hf_embed_check.py +0 -89
  249. package/src/debug/reference/hf_layer_out.py +0 -100
  250. package/src/debug/reference/hf_qwen35_linear_attn_debug.py +0 -268
  251. package/src/debug/reference/hf_rope_check.py +0 -116
  252. package/src/debug/reference/hf_weights.py +0 -75
  253. package/src/errors/doppler-error.d.ts +0 -20
  254. package/src/experimental/bridge/extension/background.d.ts +0 -14
  255. package/src/experimental/bridge/extension/background.js +0 -168
  256. package/src/experimental/bridge/extension/manifest.json +0 -34
  257. package/src/experimental/bridge/extension-client.d.ts +0 -114
  258. package/src/experimental/bridge/index.d.ts +0 -69
  259. package/src/experimental/bridge/node-bridge.js +0 -24
  260. package/src/experimental/bridge/protocol.d.ts +0 -96
  261. package/src/experimental/browser/gguf-importer.d.ts +0 -136
  262. package/src/experimental/browser/gguf-importer.js +0 -532
  263. package/src/experimental/browser/gguf-parser-browser.d.ts +0 -14
  264. package/src/experimental/browser/quantization.d.ts +0 -74
  265. package/src/experimental/browser/safetensors-parser-browser.d.ts +0 -193
  266. package/src/experimental/browser/shard-io-browser.d.ts +0 -57
  267. package/src/experimental/browser/shard-io-browser.js +0 -89
  268. package/src/experimental/browser/tensor-source-download.d.ts +0 -27
  269. package/src/experimental/diffusion/config/expansion-lanes.json +0 -29
  270. package/src/experimental/distribution/p2p-transport-contract.d.ts +0 -57
  271. package/src/experimental/distribution/p2p-transport.js +0 -75
  272. package/src/experimental/hotswap/intent-bundle.d.ts +0 -37
  273. package/src/experimental/hotswap/manifest.d.ts +0 -42
  274. package/src/experimental/hotswap/model-swap.js +0 -42
  275. package/src/experimental/hotswap/runtime.d.ts +0 -31
  276. package/src/experimental/training/README.md +0 -164
  277. package/src/experimental/training/artifacts.d.ts +0 -160
  278. package/src/experimental/training/attention-backward.d.ts +0 -30
  279. package/src/experimental/training/attention-forward.d.ts +0 -22
  280. package/src/experimental/training/attention-forward.js +0 -82
  281. package/src/experimental/training/autograd.d.ts +0 -51
  282. package/src/experimental/training/backward-kernels.js +0 -39
  283. package/src/experimental/training/checkpoint-watch.d.ts +0 -8
  284. package/src/experimental/training/checkpoint.d.ts +0 -36
  285. package/src/experimental/training/clip.d.ts +0 -9
  286. package/src/experimental/training/dataloader.d.ts +0 -8
  287. package/src/experimental/training/datasets/gepa-frontier.d.ts +0 -41
  288. package/src/experimental/training/datasets/gepa-frontier.js +0 -220
  289. package/src/experimental/training/datasets/index.d.ts +0 -24
  290. package/src/experimental/training/datasets/index.js +0 -13
  291. package/src/experimental/training/datasets/jsonl.d.ts +0 -11
  292. package/src/experimental/training/datasets/reploid.d.ts +0 -3
  293. package/src/experimental/training/datasets/reploid.js +0 -36
  294. package/src/experimental/training/datasets/teacher-traces.d.ts +0 -51
  295. package/src/experimental/training/datasets/teacher-traces.js +0 -279
  296. package/src/experimental/training/datasets/text-pairs.d.ts +0 -66
  297. package/src/experimental/training/datasets/token-batch.d.ts +0 -21
  298. package/src/experimental/training/datasets/token-batch.js +0 -52
  299. package/src/experimental/training/datasets/translation-pairs.d.ts +0 -34
  300. package/src/experimental/training/datasets/translation-pairs.js +0 -49
  301. package/src/experimental/training/distillation/artifacts.d.ts +0 -71
  302. package/src/experimental/training/distillation/checkpoint-watch.d.ts +0 -10
  303. package/src/experimental/training/distillation/dataset.d.ts +0 -59
  304. package/src/experimental/training/distillation/eval.d.ts +0 -34
  305. package/src/experimental/training/distillation/index.d.ts +0 -29
  306. package/src/experimental/training/distillation/runtime.d.ts +0 -20
  307. package/src/experimental/training/distillation/scoreboard.d.ts +0 -6
  308. package/src/experimental/training/distillation/stage-a.d.ts +0 -45
  309. package/src/experimental/training/distillation/stage-b.d.ts +0 -24
  310. package/src/experimental/training/distillation/student-fixture.d.ts +0 -27
  311. package/src/experimental/training/distillation/suite-data.d.ts +0 -45
  312. package/src/experimental/training/export.d.ts +0 -41
  313. package/src/experimental/training/index.d.ts +0 -74
  314. package/src/experimental/training/index.js +0 -63
  315. package/src/experimental/training/lora-pipeline.d.ts +0 -183
  316. package/src/experimental/training/lora.d.ts +0 -19
  317. package/src/experimental/training/loss-scaling.d.ts +0 -21
  318. package/src/experimental/training/loss.d.ts +0 -10
  319. package/src/experimental/training/objectives/base.d.ts +0 -58
  320. package/src/experimental/training/objectives/cross_entropy.d.ts +0 -18
  321. package/src/experimental/training/objectives/distill_kd.d.ts +0 -16
  322. package/src/experimental/training/objectives/distill_triplet.d.ts +0 -16
  323. package/src/experimental/training/objectives/index.d.ts +0 -12
  324. package/src/experimental/training/objectives/index.js +0 -6
  325. package/src/experimental/training/objectives/ul_stage1_joint.d.ts +0 -16
  326. package/src/experimental/training/objectives/ul_stage2_base.d.ts +0 -16
  327. package/src/experimental/training/operator-agent-eval.d.ts +0 -65
  328. package/src/experimental/training/operator-artifacts.d.ts +0 -62
  329. package/src/experimental/training/operator-command.d.ts +0 -5
  330. package/src/experimental/training/operator-eval.d.ts +0 -48
  331. package/src/experimental/training/operator-scoreboard.d.ts +0 -5
  332. package/src/experimental/training/optimizer.d.ts +0 -22
  333. package/src/experimental/training/runner.d.ts +0 -249
  334. package/src/experimental/training/suite.d.ts +0 -299
  335. package/src/experimental/training/tensor-factory.d.ts +0 -9
  336. package/src/experimental/training/trainer.d.ts +0 -89
  337. package/src/experimental/training/ul_dataset.d.ts +0 -47
  338. package/src/experimental/training/ul_schedule.d.ts +0 -6
  339. package/src/experimental/training/workload-packs/distill-glm52-doppler-agent-harness-gemma270m-f16-sft-tiny.json +0 -160
  340. package/src/experimental/training/workload-packs/distill-glm52-doppler-code-agent-gemma270m-f16-sft-tiny.json +0 -175
  341. package/src/experimental/training/workload-packs/distill-glm52-doppler-code-agent-sft-tiny.json +0 -175
  342. package/src/experimental/training/workload-packs/distill-glm52-doppler-js-json-gemma270m-f16-sft-tiny.json +0 -160
  343. package/src/experimental/training/workload-packs/distill-glm52-doppler-review-gemma270m-f16-sft-tiny.json +0 -162
  344. package/src/experimental/training/workload-packs/distill-glm52-doppler-wgsl-gemma270m-f16-sft-tiny.json +0 -160
  345. package/src/experimental/training/workload-packs/distill-translategemma-medium.json +0 -103
  346. package/src/experimental/training/workload-packs/distill-translategemma-tiny.json +0 -103
  347. package/src/experimental/training/workload-packs/lora-doppler-agent-harness-gemma270m-f16-tiny.json +0 -137
  348. package/src/experimental/training/workload-packs/lora-doppler-code-agent-gemma270m-f16-tiny.json +0 -152
  349. package/src/experimental/training/workload-packs/lora-doppler-code-agent-tiny.json +0 -152
  350. package/src/experimental/training/workload-packs/lora-doppler-js-json-gemma270m-f16-tiny.json +0 -137
  351. package/src/experimental/training/workload-packs/lora-doppler-review-gemma270m-f16-tiny.json +0 -139
  352. package/src/experimental/training/workload-packs/lora-doppler-wgsl-gemma270m-f16-tiny.json +0 -137
  353. package/src/experimental/training/workload-packs/lora-toy-medium.json +0 -91
  354. package/src/experimental/training/workload-packs/lora-toy-tiny.json +0 -91
  355. package/src/experimental/training/workload-packs/registry.json +0 -278
  356. package/src/experimental/training/workload-packs/ul-training-medium.json +0 -55
  357. package/src/experimental/training/workload-packs/ul-training-tiny.json +0 -55
  358. package/src/experimental/training/workloads.d.ts +0 -201
  359. package/src/formats/gguf/types.d.ts +0 -74
  360. package/src/formats/index.d.ts +0 -75
  361. package/src/formats/index.js +0 -19
  362. package/src/formats/litert/types.d.ts +0 -74
  363. package/src/formats/rdrr/distributed/parsing.d.ts +0 -6
  364. package/src/formats/rdrr/distributed/validation.d.ts +0 -23
  365. package/src/formats/rdrr/storage-descriptor.d.ts +0 -36
  366. package/src/formats/rdrr/tensor-config-validator.d.ts +0 -33
  367. package/src/formats/safetensors/types.d.ts +0 -67
  368. package/src/formats/tflite/types.d.ts +0 -61
  369. package/src/formats/tokenizer/index.d.ts +0 -5
  370. package/src/formats/tokenizer/index.js +0 -3
  371. package/src/formats/tokenizer/types.d.ts +0 -9
  372. package/src/gpu/kernel-runtime.d.ts +0 -13
  373. package/src/gpu/kernel-selection-cache.d.ts +0 -13
  374. package/src/gpu/kernel-selection-log.d.ts +0 -4
  375. package/src/gpu/kernel-tuner/benchmarks.d.ts +0 -144
  376. package/src/gpu/kernel-tuner/cache.d.ts +0 -55
  377. package/src/gpu/kernel-tuner/index.d.ts +0 -59
  378. package/src/gpu/kernel-tuner/tuner.d.ts +0 -82
  379. package/src/gpu/kernel-tuner/types.d.ts +0 -101
  380. package/src/gpu/kernel-tuner/types.js +0 -4
  381. package/src/gpu/kernel-tuner.d.ts +0 -33
  382. package/src/gpu/kernels/README.md +0 -166
  383. package/src/gpu/kernels/check-finiteness.d.ts +0 -15
  384. package/src/gpu/kernels/check-hot-vocab-stop.d.ts +0 -18
  385. package/src/gpu/kernels/check-stop.d.ts +0 -31
  386. package/src/gpu/kernels/codegen/patches/attention_decode_f16.from.attention_decode.diff +0 -213
  387. package/src/gpu/kernels/codegen/patches/attention_decode_f16kv.from.attention_decode.diff +0 -89
  388. package/src/gpu/kernels/codegen/patches/attention_f16.from.attention.diff +0 -62
  389. package/src/gpu/kernels/codegen/patches/attention_f16kv.from.attention.diff +0 -40
  390. package/src/gpu/kernels/codegen/patches/attention_small_f16.from.attention_small.diff +0 -226
  391. package/src/gpu/kernels/codegen/patches/attention_small_f16kv.from.attention_small.diff +0 -264
  392. package/src/gpu/kernels/codegen/patches/attention_streaming_f16.from.attention_streaming.diff +0 -71
  393. package/src/gpu/kernels/codegen/patches/attention_streaming_f16kv.from.attention_streaming.diff +0 -53
  394. package/src/gpu/kernels/codegen/patches/bias_add_f16.from.bias_add.diff +0 -54
  395. package/src/gpu/kernels/codegen/patches/dequant_mxfp4_expert_f16.from.dequant_mxfp4_expert.diff +0 -125
  396. package/src/gpu/kernels/codegen/patches/fused_ffn_f16.from.fused_ffn.diff +0 -428
  397. package/src/gpu/kernels/codegen/patches/fused_matmul_q4_batched_f16.from.fused_matmul_q4_batched.diff +0 -114
  398. package/src/gpu/kernels/codegen/patches/fused_matmul_rmsnorm_f16.from.fused_matmul_rmsnorm.diff +0 -477
  399. package/src/gpu/kernels/codegen/patches/fused_swiglu_f16.from.fused_swiglu.diff +0 -60
  400. package/src/gpu/kernels/codegen/patches/gelu_f16.from.gelu.diff +0 -41
  401. package/src/gpu/kernels/codegen/patches/layernorm_f16.from.layernorm.diff +0 -82
  402. package/src/gpu/kernels/codegen/patches/matmul_gemv_residual_f16.from.matmul_gemv_residual.diff +0 -87
  403. package/src/gpu/kernels/codegen/patches/moe_gather_f16.from.moe_gather.diff +0 -181
  404. package/src/gpu/kernels/codegen/patches/residual_f16.from.residual.diff +0 -65
  405. package/src/gpu/kernels/codegen/patches/rmsnorm_f16.from.rmsnorm.diff +0 -515
  406. package/src/gpu/kernels/codegen/patches/rope_f16.from.rope.diff +0 -292
  407. package/src/gpu/kernels/codegen/patches/sample_f16.from.sample.diff +0 -336
  408. package/src/gpu/kernels/codegen/patches/scatter_add_dynamic_f16.from.scatter_add_dynamic.diff +0 -65
  409. package/src/gpu/kernels/codegen/patches/silu_f16.from.silu.diff +0 -81
  410. package/src/gpu/kernels/codegen/patches/topk_f16.from.topk.diff +0 -231
  411. package/src/gpu/kernels/codegen/wgsl-patch-variants.js +0 -152
  412. package/src/gpu/kernels/codegen/wgsl-variants.js +0 -885
  413. package/src/gpu/kernels/constants.d.ts +0 -129
  414. package/src/gpu/kernels/dispatch.d.ts +0 -72
  415. package/src/gpu/kernels/gated-short-conv.d.ts +0 -63
  416. package/src/gpu/kernels/kernel-base.d.ts +0 -33
  417. package/src/gpu/kernels/linear-attention-core.d.ts +0 -42
  418. package/src/gpu/kernels/matmul-dispatch.d.ts +0 -49
  419. package/src/gpu/kernels/matmul-selection.d.ts +0 -87
  420. package/src/gpu/kernels/rep-penalty.d.ts +0 -21
  421. package/src/gpu/kernels/rule-matcher.d.ts +0 -30
  422. package/src/gpu/kernels/rule-registry.d.ts +0 -7
  423. package/src/gpu/kernels/types.js +0 -4
  424. package/src/gpu/perf-guards.d.ts +0 -13
  425. package/src/gpu/profiler.d.ts +0 -93
  426. package/src/gpu/readback-utils.d.ts +0 -16
  427. package/src/gpu/readback-utils.js +0 -49
  428. package/src/gpu/submit-probe.d.ts +0 -12
  429. package/src/index-browser.d.ts +0 -2
  430. package/src/index-browser.js +0 -2
  431. package/src/inference/README.md +0 -593
  432. package/src/inference/browser-harness-contract-helpers.d.ts +0 -5
  433. package/src/inference/browser-harness-diffusion-energy-suites.d.ts +0 -2
  434. package/src/inference/browser-harness-model-helpers.d.ts +0 -10
  435. package/src/inference/browser-harness-report-helpers.d.ts +0 -7
  436. package/src/inference/browser-harness-suite-helpers.d.ts +0 -40
  437. package/src/inference/browser-harness-text-helpers.d.ts +0 -152
  438. package/src/inference/model-lanes/execution-lane-resolver.d.ts +0 -61
  439. package/src/inference/model-lanes/execution-lane-resolver.js +0 -126
  440. package/src/inference/network-evolution.js +0 -79
  441. package/src/inference/pipelines/context.d.ts +0 -21
  442. package/src/inference/pipelines/diffusion/helpers.d.ts +0 -29
  443. package/src/inference/pipelines/diffusion/index.d.ts +0 -3
  444. package/src/inference/pipelines/diffusion/index.js +0 -3
  445. package/src/inference/pipelines/diffusion/scheduler.d.ts +0 -35
  446. package/src/inference/pipelines/diffusion/sd3-transformer.d.ts +0 -20
  447. package/src/inference/pipelines/diffusion/sd3-weights.d.ts +0 -21
  448. package/src/inference/pipelines/diffusion/text-encoder-gpu.d.ts +0 -82
  449. package/src/inference/pipelines/diffusion/text-encoder.d.ts +0 -29
  450. package/src/inference/pipelines/diffusion/types.js +0 -1
  451. package/src/inference/pipelines/diffusion/vae.d.ts +0 -20
  452. package/src/inference/pipelines/dream/energy-head-pipeline.d.ts +0 -29
  453. package/src/inference/pipelines/dream/energy-head-pipeline.js +0 -6
  454. package/src/inference/pipelines/dream/pipeline.d.ts +0 -17
  455. package/src/inference/pipelines/dream/pipeline.js +0 -8
  456. package/src/inference/pipelines/energy/index.d.ts +0 -1
  457. package/src/inference/pipelines/energy/index.js +0 -1
  458. package/src/inference/pipelines/energy/types.js +0 -1
  459. package/src/inference/pipelines/energy-head/index.d.ts +0 -6
  460. package/src/inference/pipelines/energy-head/index.js +0 -6
  461. package/src/inference/pipelines/factory.d.ts +0 -10
  462. package/src/inference/pipelines/index.d.ts +0 -22
  463. package/src/inference/pipelines/index.js +0 -19
  464. package/src/inference/pipelines/registry.d.ts +0 -15
  465. package/src/inference/pipelines/rng.d.ts +0 -2
  466. package/src/inference/pipelines/structured/tsv-grammar-mask.js +0 -166
  467. package/src/inference/pipelines/text/attention/attn-config.d.ts +0 -23
  468. package/src/inference/pipelines/text/attention/dispatch-params.d.ts +0 -173
  469. package/src/inference/pipelines/text/attention/index.d.ts +0 -24
  470. package/src/inference/pipelines/text/attention/output-projection.d.ts +0 -12
  471. package/src/inference/pipelines/text/attention/precision-contract.d.ts +0 -36
  472. package/src/inference/pipelines/text/attention/projections.d.ts +0 -208
  473. package/src/inference/pipelines/text/attention/quant-options.d.ts +0 -38
  474. package/src/inference/pipelines/text/attention/record.d.ts +0 -36
  475. package/src/inference/pipelines/text/attention/rmsnorm-fusion-gate.d.ts +0 -8
  476. package/src/inference/pipelines/text/attention/run.d.ts +0 -38
  477. package/src/inference/pipelines/text/attention/types.d.ts +0 -125
  478. package/src/inference/pipelines/text/attention.d.ts +0 -23
  479. package/src/inference/pipelines/text/attention.js +0 -12
  480. package/src/inference/pipelines/text/bdpa-steamroller.d.ts +0 -22
  481. package/src/inference/pipelines/text/buffer-types.js +0 -4
  482. package/src/inference/pipelines/text/debug-utils/config.d.ts +0 -144
  483. package/src/inference/pipelines/text/debug-utils/index.d.ts +0 -53
  484. package/src/inference/pipelines/text/debug-utils/logging.d.ts +0 -106
  485. package/src/inference/pipelines/text/debug-utils/tensor.d.ts +0 -119
  486. package/src/inference/pipelines/text/debug-utils/utils.d.ts +0 -77
  487. package/src/inference/pipelines/text/debug-utils.d.ts +0 -42
  488. package/src/inference/pipelines/text/drift-policy.d.ts +0 -57
  489. package/src/inference/pipelines/text/dtype-contract.d.ts +0 -53
  490. package/src/inference/pipelines/text/embed.d.ts +0 -69
  491. package/src/inference/pipelines/text/execution-runtime-builders.d.ts +0 -49
  492. package/src/inference/pipelines/text/failure-reduction.d.ts +0 -84
  493. package/src/inference/pipelines/text/failure-reduction.js +0 -180
  494. package/src/inference/pipelines/text/ffn/dense.d.ts +0 -99
  495. package/src/inference/pipelines/text/ffn/index.d.ts +0 -23
  496. package/src/inference/pipelines/text/ffn/moe.d.ts +0 -24
  497. package/src/inference/pipelines/text/ffn/sandwich.d.ts +0 -27
  498. package/src/inference/pipelines/text/ffn/standard.d.ts +0 -26
  499. package/src/inference/pipelines/text/ffn/types.d.ts +0 -30
  500. package/src/inference/pipelines/text/ffn.d.ts +0 -31
  501. package/src/inference/pipelines/text/ffn.js +0 -18
  502. package/src/inference/pipelines/text/finiteness-guard-status.d.ts +0 -11
  503. package/src/inference/pipelines/text/finiteness-policy.d.ts +0 -39
  504. package/src/inference/pipelines/text/generator-decode-policy.d.ts +0 -81
  505. package/src/inference/pipelines/text/generator-helpers.d.ts +0 -45
  506. package/src/inference/pipelines/text/generator-prefill-helpers.d.ts +0 -84
  507. package/src/inference/pipelines/text/generator-runtime.d.ts +0 -93
  508. package/src/inference/pipelines/text/generator-steps.d.ts +0 -186
  509. package/src/inference/pipelines/text/index.d.ts +0 -5
  510. package/src/inference/pipelines/text/index.js +0 -6
  511. package/src/inference/pipelines/text/kernel-trace.d.ts +0 -119
  512. package/src/inference/pipelines/text/layer.d.ts +0 -118
  513. package/src/inference/pipelines/text/load-errors.d.ts +0 -21
  514. package/src/inference/pipelines/text/logits/precision-policy.d.ts +0 -11
  515. package/src/inference/pipelines/text/logits/types.js +0 -4
  516. package/src/inference/pipelines/text/logits.d.ts +0 -27
  517. package/src/inference/pipelines/text/logits.js +0 -16
  518. package/src/inference/pipelines/text/lora-apply.d.ts +0 -28
  519. package/src/inference/pipelines/text/moe-cache.d.ts +0 -32
  520. package/src/inference/pipelines/text/moe-cpu-gptoss.d.ts +0 -9
  521. package/src/inference/pipelines/text/moe-cpu.d.ts +0 -13
  522. package/src/inference/pipelines/text/moe-gpu.d.ts +0 -26
  523. package/src/inference/pipelines/text/moe-helpers.d.ts +0 -12
  524. package/src/inference/pipelines/text/moe-shape-validator.d.ts +0 -70
  525. package/src/inference/pipelines/text/operator-events.d.ts +0 -125
  526. package/src/inference/pipelines/text/operator-identity.d.ts +0 -60
  527. package/src/inference/pipelines/text/ops.d.ts +0 -222
  528. package/src/inference/pipelines/text/per-layer-inputs.d.ts +0 -107
  529. package/src/inference/pipelines/text/probes.d.ts +0 -28
  530. package/src/inference/pipelines/text/sampling-config.d.ts +0 -48
  531. package/src/inference/pipelines/text/stage-names.d.ts +0 -88
  532. package/src/inference/pipelines/text/tsir-fixture-writer.d.ts +0 -57
  533. package/src/inference/pipelines/text/types.js +0 -4
  534. package/src/inference/tokenizers/base.d.ts +0 -44
  535. package/src/inference/tokenizers/behavior-flags.d.ts +0 -9
  536. package/src/inference/tokenizers/bpe.d.ts +0 -27
  537. package/src/inference/tokenizers/bundled.d.ts +0 -67
  538. package/src/inference/tokenizers/sentencepiece.d.ts +0 -30
  539. package/src/inference/tokenizers/types.js +0 -7
  540. package/src/loader/embedding-limit-preflight.d.ts +0 -31
  541. package/src/loader/per-layer-input-loader.d.ts +0 -30
  542. package/src/loader/tensors/functional-descriptor-loader.d.ts +0 -9
  543. package/src/loader/tensors/source-transform.d.ts +0 -65
  544. package/src/memory/address-table.d.ts +0 -96
  545. package/src/rules/execution-rules-contract-check.d.ts +0 -21
  546. package/src/rules/layer-pattern-contract-check.d.ts +0 -17
  547. package/src/rules/rule-registry.d.ts +0 -77
  548. package/src/storage/artifact-storage-context.d.ts +0 -63
  549. package/src/storage/backends/idb-store.d.ts +0 -52
  550. package/src/storage/backends/memory-store.d.ts +0 -36
  551. package/src/storage/backends/opfs-store.d.ts +0 -41
  552. package/src/storage/blake3.d.ts +0 -17
  553. package/src/storage/download-types.d.ts +0 -157
  554. package/src/storage/downloader.d.ts +0 -103
  555. package/src/storage/index.d.ts +0 -274
  556. package/src/storage/index.js +0 -201
  557. package/src/storage/preflight.d.ts +0 -144
  558. package/src/storage/preflight.js +0 -360
  559. package/src/storage/quickstart-downloader.d.ts +0 -157
  560. package/src/storage/quickstart-downloader.js +0 -289
  561. package/src/storage/source-artifact-store.d.ts +0 -59
  562. package/src/sw.js +0 -184
  563. package/src/tooling/command-api-constants.d.ts +0 -9
  564. package/src/tooling/command-api-family-normalizers.d.ts +0 -10
  565. package/src/tooling/command-api-helpers.d.ts +0 -23
  566. package/src/tooling/command-runner-shared.d.ts +0 -89
  567. package/src/tooling/conversion-config-materializer.d.ts +0 -14
  568. package/src/tooling/conversion-config-materializer.js +0 -96
  569. package/src/tooling/diagnose-runner.d.ts +0 -28
  570. package/src/tooling/hf-registry-utils.d.ts +0 -55
  571. package/src/tooling/kernel-path-builder/custom-runtime-facts.d.ts +0 -18
  572. package/src/tooling/kernel-path-builder/custom-runtime-facts.js +0 -72
  573. package/src/tooling/kernel-path-builder/index.d.ts +0 -65
  574. package/src/tooling/kernel-path-builder/index.js +0 -1361
  575. package/src/tooling/kernel-path-builder/runtime-overlay.d.ts +0 -9
  576. package/src/tooling/kernel-path-builder/runtime-overlay.js +0 -147
  577. package/src/tooling/lean-execution-contract-runner.d.ts +0 -33
  578. package/src/tooling/lean-execution-contract-runner.js +0 -158
  579. package/src/tooling/lean-execution-contract.d.ts +0 -16
  580. package/src/tooling/lean-execution-contract.js +0 -81
  581. package/src/tooling/litert-package-runtime.d.ts +0 -46
  582. package/src/tooling/node-convert-worker-pool.d.ts +0 -16
  583. package/src/tooling/node-convert-worker.d.ts +0 -1
  584. package/src/tooling/node-converter.d.ts +0 -1
  585. package/src/tooling/node-file-fetch.d.ts +0 -1
  586. package/src/tooling/node-source-runtime.d.ts +0 -24
  587. package/src/tooling/node-webgpu.d.ts +0 -17
  588. package/src/tooling/precision-replay-math.js +0 -174
  589. package/src/tooling/program-bundle-host.d.ts +0 -16
  590. package/src/tooling/program-bundle-host.js +0 -20
  591. package/src/tooling/runtime-input-composition.d.ts +0 -37
  592. package/src/tooling/source-package-profiles.d.ts +0 -50
  593. package/src/tooling/source-runtime-converter-config.d.ts +0 -19
  594. package/src/tooling/source-runtime-materializer.d.ts +0 -6
  595. package/src/tooling/source-runtime-materializer.js +0 -89
  596. package/src/tooling-exports.browser.d.ts +0 -7
  597. package/src/types/chrome.d.ts +0 -36
  598. package/src/types/chrome.js +0 -1
  599. package/src/types/index.d.ts +0 -1
  600. package/src/types/index.js +0 -1
  601. package/src/utils/array-buffer.d.ts +0 -4
  602. package/src/utils/canonical-hash.d.ts +0 -7
  603. package/src/utils/clone-json.d.ts +0 -1
  604. package/src/utils/hf-resolve-url.d.ts +0 -16
  605. package/src/utils/index.d.ts +0 -10
  606. package/src/utils/index.js +0 -10
  607. package/src/utils/load-json.d.ts +0 -5
  608. package/src/utils/plain-object.d.ts +0 -1
  609. package/src/utils/sha256.d.ts +0 -5
  610. package/src/utils/stable-sort-object.d.ts +0 -1
package/CHANGELOG.md CHANGED
@@ -8,6 +8,70 @@ docs so the `0.1.x` line has one conventional npm-visible history surface.
8
8
 
9
9
  ## [Unreleased]
10
10
 
11
+ ## [0.4.8] - 2026-07-11
12
+
13
+ ### Added
14
+
15
+ - Advanced model-handle APIs for deterministic downstream scoring and prefix
16
+ reuse: `tokenizeText()`, `prefillWithTokenLogits()`,
17
+ `prefillWithTokenLogitsFromKV()`, and stateful `resetToSeqLen()` truncation.
18
+ - NativeKD2 English/Spanish translation-student evidence, including the
19
+ converted Q4K artifact, WMT13 quality receipts, translation benchmark
20
+ tooling, and an experimental Radeon/RDNA3 throughput profile.
21
+ - Experimental-tier catalog, hosted-artifact, and demo compare promotion for
22
+ the 1B EN/ES translation student, with frozen Q4K metrics and explicit
23
+ language, platform, and cross-runtime evidence boundaries, plus a pinned
24
+ hosted browser/WebGPU benchmark receipt.
25
+ - Machine-qualified host teacher selection for Doppler JavaScript and WGSL,
26
+ including disjoint qualification/label/student-holdout tasks, disposable
27
+ pinned source snapshots, exact repair replay, command and path audits, and
28
+ receipt-linked training-row export for accepted labels.
29
+ - Claim-grade, release-claimable Apple M3 Metal evidence for the current Tier 1
30
+ Qwen text-generation, embedding, and reranking lanes.
31
+ - Claim-grade, release-claimable Radeon 8060S Vulkan evidence for the current
32
+ Tier 1 Qwen lanes, including fresh 15-run embedding and reranking receipts.
33
+ - Local-comparable Radeon 8060S Vulkan evidence for the Tier 2 Qwen 3.5 2B and
34
+ Gemma 4 E2B INT4-PLE generation lanes.
35
+ - A receipt-linked README comparison chart for the promoted Metal and Vulkan
36
+ results.
37
+
38
+ ### Changed
39
+
40
+ - Consolidated GitHub Actions on the deterministic CPU contract; browser,
41
+ model-download, and GPU validation remain explicit hardware workflows.
42
+ - Kept the receipt-linked README evidence chart in GitHub without duplicating
43
+ that documentation asset in the npm tarball.
44
+ - Renamed nested repository guides so npm no longer force-includes them,
45
+ preserving the 1,349-file package boundary without removing runtime code.
46
+ - Refreshed the competition scoreboard, model support inventory, local GPU
47
+ challenger matrix, and model roadmap from the promoted receipts.
48
+ - Added Metal runtime profiles and optimized WebGPU paths for Qwen 3.5 text,
49
+ embedding, and reranking, including fixed `head_dim=128` f16-KV prefill
50
+ attention for the reranker lane.
51
+ - Reduced the npm package from 1,733 files to the enforced 1,349-file runtime,
52
+ declaration, and resource closure.
53
+ - Reframed the README around the verifiable optimization loop, measured
54
+ performance mechanisms, and one runnable entry point.
55
+
56
+ ### Fixed
57
+
58
+ - Preserved loaded-manifest provenance through `manifestHash` and reset
59
+ embedding state after both successful and failed calls so repeated model
60
+ sessions remain deterministic.
61
+ - Regenerated kernel reachability metadata after the new runtime profiles and
62
+ kernel paths changed the selectable registry surface.
63
+ - Replaced the stale Vulkan reranker loss with the current 15-run,
64
+ release-claimable Doppler throughput win while preserving the TJS model-load
65
+ advantage in the evidence ledger.
66
+ - Published the browser runtime capability rules, model-source resolver,
67
+ execution-v1 runtime, and schema updates required by downstream static
68
+ consumers such as Simulatte without patching installed package files.
69
+ - Added an installed-tarball smoke for every public export, all three CLIs,
70
+ browser-runner assets, optional-dependency omission, and strict TypeScript
71
+ declaration resolution; repaired the declaration defects it exposed.
72
+ - Restamped the Linux/Vulkan Qwen 3.5 2B fixture after a Darwin-only profile
73
+ addition and updated the harness contract test for the renamed testing guide.
74
+
11
75
  ## [0.4.4] - 2026-06-30
12
76
 
13
77
  ### Added
package/README.md CHANGED
@@ -1,74 +1,104 @@
1
1
  # doppler-gpu
2
2
 
3
- WebGPU inference runtime for browser, Node, Bun, CLI, and local server use.
4
- Doppler runs JavaScript orchestration over WGSL kernels and loads models from
5
- RDRR manifests.
3
+ [![Build](https://img.shields.io/github/actions/workflow/status/clocksmith/doppler/check-green.yml?branch=main&label=build)](https://github.com/clocksmith/doppler/actions/workflows/check-green.yml)
4
+ [![npm version](https://img.shields.io/npm/v/doppler-gpu.svg?label=version)](https://www.npmjs.com/package/doppler-gpu)
5
+ [![License: Apache-2.0](https://img.shields.io/badge/license-Apache--2.0-blue.svg)](https://github.com/clocksmith/doppler/blob/main/LICENSE)
6
+ [![PRs welcome](https://img.shields.io/badge/PRs-welcome-brightgreen.svg)](https://github.com/clocksmith/doppler/pulls)
7
+
8
+ Doppler turns inference engineering into a verifiable search problem. Optimizers
9
+ edit RDRR manifests, execution plans, and kernels; parity and benchmark gates
10
+ accept or reject each candidate against correctness and speed.
11
+
12
+ Today engineers use these contracts and gates to tune a JavaScript/WGSL WebGPU
13
+ runtime for supported
14
+ [RDRR artifacts](https://github.com/clocksmith/doppler/blob/main/docs/rdrr-format.md)
15
+ in browser and Node. Doppler runs text generation, embeddings, and reranking
16
+ locally, with CLI and OpenAI-compatible server entry points. Bun WebGPU support
17
+ is experimental.
6
18
 
7
19
  **[Try the live demo](https://d4da.com/doppler)** | **[npm](https://www.npmjs.com/package/doppler-gpu)** | **[docs](https://github.com/clocksmith/doppler/blob/main/docs/INDEX.md)**
8
20
 
9
- Broader model status and compare evidence live in the support and release
10
- matrices. See the
11
- [benchmark methodology](https://github.com/clocksmith/doppler/blob/main/docs/benchmark-methodology.md)
12
- for the receipt contract and disclosure rules.
21
+ ```bash
22
+ npx doppler-gpu
23
+ ```
24
+
25
+ ## Evidence
26
+
27
+ Doppler has lower steady-state inference latency than Transformers.js in each
28
+ of the comparable browser WebGPU results indexed below.
13
29
 
14
- ## What it is
30
+ ![Metal and Vulkan browser WebGPU latency distributions](https://raw.githubusercontent.com/clocksmith/doppler/main/assets/doppler-webgpu-evidence.svg)
15
31
 
16
- - A WebGPU inference engine written in JavaScript and WGSL.
17
- - A shared browser, Node, Bun, CLI, and OpenAI-compatible server surface.
18
- - An RDRR model loader for sharded weights, manifest-owned config, and tokenizer
19
- metadata.
20
- - A runtime that keeps kernel paths, dtype policy, and benchmark contracts
21
- visible in JSON, JavaScript, and WGSL.
32
+ For text, faster means lower decode ms/token; for retrieval it means lower
33
+ ms/embedding or ms/rerank. Every comparison passes its declared workload
34
+ correctness gate. Model loading is separate: Transformers.js loads the Vulkan
35
+ embedding and reranker artifacts faster. The
36
+ [competition scoreboard](https://github.com/clocksmith/doppler/blob/main/docs/model-competition-scoreboard.md) links every
37
+ receipt, and the [methodology](https://github.com/clocksmith/doppler/blob/main/docs/benchmark-methodology.md) defines the gates.
22
38
 
23
- ## Goals
39
+ ## Why these lanes are faster
24
40
 
25
- Doppler's mainline work is organized around three concrete surfaces:
41
+ Doppler authors the WGSL path and pins it in RDRR. Runtime profiles pin session
42
+ cadence. The measured wins come from different mechanisms in different phases:
26
43
 
27
- 1. Make local WebGPU inference a real product surface across the hosted browser
28
- demo, `npx doppler-gpu`, root API, CLI, Node/Bun, and OpenAI-compatible
29
- localhost server.
30
- 2. Own the model artifact and runtime contract through RDRR manifests, hosted
31
- registry IDs, explicit runtime config, tokenizer/shard identity, and support
32
- matrices.
33
- 3. Make correctness and performance evidence-backed through release receipts,
34
- benchmark artifacts, command parity, explicit kernel paths, and fail-closed
35
- unsupported paths.
44
+ | Lever | Affected phase | Measured receipt |
45
+ | --- | --- | --- |
46
+ | Fused Q4_K projection and FFN work removes separate dispatches | Text decode | [Qwen 3.5 0.8B Metal](https://github.com/clocksmith/doppler/blob/main/benchmarks/vendors/results/compare_20260709T154633.json) |
47
+ | Fixed `head_dim=128` attention avoids the generic prefill path | Retrieval attention | [embedding](https://github.com/clocksmith/doppler/blob/main/benchmarks/vendors/results/embedding_compare_qwen-3-embedding-0-6b-q4k-ehf16-af32_20260709T180853.json), [reranking](https://github.com/clocksmith/doppler/blob/main/benchmarks/vendors/results/rerank_compare_qwen-3-reranker-0-6b-q4k-ehf16-af32_20260709T192830.json) |
48
+ | Batch-four decode amortizes one readback across four tokens | Submit and map waits | [Qwen 3.5 2B Vulkan](https://github.com/clocksmith/doppler/blob/main/benchmarks/vendors/results/compare_20260707T161623.json) |
49
+ | INT4-PLE and Q4_K layouts lower projection traffic | Text decode | [Gemma 4 Vulkan](https://github.com/clocksmith/doppler/blob/main/benchmarks/vendors/results/compare_20260707T170557.json) |
36
50
 
37
- See
38
- [docs/goals.md](https://github.com/clocksmith/doppler/blob/main/docs/goals.md)
39
- for the product and technical contract behind those goals. The current
40
- completion gate is `src/config/goal-completion-matrix.json`, checked by
41
- `npm run goals:check`.
51
+ The Gemma 4 comparison uses its declared product-format output policy. It does
52
+ not claim exact greedy token parity.
53
+
54
+ Parity and comparable benchmark gates decide what is retained. The
55
+ [challenger framework](https://github.com/clocksmith/doppler/blob/main/docs/local-gpu-challenger-framework.md)
56
+ records accepted and rejected trials, and each verdict's receipt binds it to a
57
+ model, profile, workload, and device; the
58
+ [negative-results findings](https://github.com/clocksmith/doppler/blob/main/docs/developer-guides/16-kernel-performance-optimization.md#what-we-ruled-out)
59
+ keep the dead ends.
60
+
61
+ ### Long-term direction
62
+
63
+ Humans run this loop today. The WGSL-distillation workload is experimental.
64
+ Automated kernel generation and autotuning are directions, not shipped product
65
+ paths. Doppler aims to reduce the manual steps between a new checkpoint or GPU
66
+ and a verified runtime lane. Automated proposals will face the same parity and
67
+ benchmark gates.
42
68
 
43
69
  ## How it works
44
70
 
45
- ![Doppler runtime contract map](./assets/doppler-runtime-map.svg)
71
+ ```text
72
+ +----------+ +------------------+ +--------------+ +-----------------+ +------------------+
73
+ | Contract | --> | Candidate | --> | Parity gate | --> | Benchmark gate | --> | Retain or reject |
74
+ | RDRR / | | model / plan / | | correctness | | comparable | | receipts / |
75
+ | request | | kernel edit | | vs reference | | speed vs prior | | findings |
76
+ +----------+ +------------------+ +--------------+ +-----------------+ +--------+---------+
77
+ |
78
+ next search <------+
79
+ ```
80
+
81
+ The full resolve, load, bind, dispatch, and readback flow lives in the
82
+ [architecture](https://github.com/clocksmith/doppler/blob/main/docs/architecture.md)
83
+ document. Unsupported paths fail closed. Doppler owns artifact and execution
84
+ contracts; applications own policy.
85
+
86
+ New model families need RDRR conversion and may need tokenizer, graph, or kernel
87
+ support.
46
88
 
47
- 1. A registry ID or model URL resolves to an RDRR manifest and weight shards.
48
- 2. The manifest owns model parameters, tokenizer metadata, session policy, and
49
- execution graph.
50
- 3. The loader caches shards in OPFS or disk and uploads weights to WebGPU
51
- buffers.
52
- 4. JavaScript orchestrates prefill, decode, KV cache, and streaming.
53
- 5. WGSL kernels run the tensor work selected by the manifest and runtime config.
89
+ [Ouroboros/Reploid](https://github.com/clocksmith/doppler/blob/main/docs/architecture.md#optional-ouroborosreploid-integration)
90
+ keeps orchestration above this boundary.
91
+ [Program Bundles](https://github.com/clocksmith/doppler/blob/main/docs/integration/program-bundle.md)
92
+ preserve program identity for downstream backends.
54
93
 
55
94
  ## Quick start
56
95
 
57
96
  ### Browser
58
97
 
59
- Use the live demo link above — it runs entirely in the browser with no server required. Models load into the browser cache and work offline after first download.
98
+ The live demo runs locally and works offline after its first model download.
60
99
 
61
100
  ### CLI
62
101
 
63
- ```bash
64
- npx doppler-gpu
65
- ```
66
-
67
- Downloads the default quickstart model, runs a local prompt, and prints the answer.
68
- Node quickstart artifacts are cached in `~/.cache/doppler-gpu/models` after the
69
- first run; set `DOPPLER_QUICKSTART_CACHE_DIR` to move the cache or
70
- `DOPPLER_QUICKSTART_CACHE=0` to disable it.
71
-
72
102
  ```bash
73
103
  npx doppler-gpu "Summarize WebGPU in one sentence"
74
104
  npx doppler-gpu --model qwen3-0.8b --prompt "Write a haiku about GPUs"
@@ -77,18 +107,13 @@ npx doppler-gpu --list-models
77
107
 
78
108
  ### Root API
79
109
 
80
- The `doppler` facade is the primary app-facing API.
81
- The root package intentionally stays small: it exports `doppler` and `DOPPLER_VERSION`.
82
- Advanced surfaces now live on explicit subpaths such as `doppler-gpu/loaders`,
83
- `doppler-gpu/generation`, `doppler-gpu/tooling`, and `doppler-gpu/orchestration`.
84
- Support tiers for those subpaths are tracked in the subsystem support matrix rather
85
- than assumed from export shape alone.
110
+ The `dr` facade is the primary app API. Advanced APIs use package subpaths.
86
111
 
87
112
  ```js
88
- import { doppler } from 'doppler-gpu';
113
+ import { dr } from 'doppler-gpu';
89
114
 
90
115
  // Stream tokens
91
- const model = await doppler.load('qwen3-0.8b');
116
+ const model = await dr.load('qwen3-0.8b');
92
117
  for await (const token of model.generate('Describe WebGPU briefly')) {
93
118
  process.stdout.write(token);
94
119
  }
@@ -99,13 +124,11 @@ const text = await model.generateText('Explain WebGPU in one sentence');
99
124
 
100
125
  ### OpenAI-compatible server
101
126
 
102
- For existing apps, SDKs, and eval stacks that speak the OpenAI protocol:
103
-
104
127
  ```bash
105
128
  npx doppler-serve --model qwen3-0.8b --port 8080
106
129
  ```
107
130
 
108
- Then point any OpenAI client at `http://localhost:8080/v1`:
131
+ Point an OpenAI client at `http://localhost:8080/v1`:
109
132
 
110
133
  ```js
111
134
  import OpenAI from 'openai';
@@ -116,98 +139,24 @@ const response = await client.chat.completions.create({
116
139
  });
117
140
  ```
118
141
 
119
- This is a compatibility bridge — the core engine runs identically in the browser or Node.
120
-
121
142
  Registry IDs resolve to hosted RDRR artifacts from `Clocksmith/rdrr` by default. See the [Root API guide](https://github.com/clocksmith/doppler/blob/main/docs/api/root.md).
122
143
 
123
- ## Support contract
124
-
125
- Doppler keeps model support and subsystem support separate:
126
-
127
- - [model support matrix](https://github.com/clocksmith/doppler/blob/main/docs/model-support-matrix.md): which models are verified right now
128
- - [subsystem support matrix](https://github.com/clocksmith/doppler/blob/main/docs/subsystem-support-matrix.md): which runtime and API surfaces are `tier1`, `experimental`, or `internal-only`
129
-
130
- The tier1 proof surface is the hosted browser demo, the root `doppler` API,
131
- the quickstart CLI, the OpenAI-compatible localhost server, and the verified
132
- Qwen text, embedding, and rerank paths behind them.
133
-
134
- ## Tier 1 Qwen registry lanes
135
-
136
- The hosted `Clocksmith/rdrr` registry carries the current Tier 1 Qwen lanes:
137
-
138
- - `qwen3-0.8b` / `qwen-3-5-0-8b-q4k-ehaf16`: text generation.
139
- - `qwen3-embedding-0.6b` / `qwen-3-embedding-0-6b-q4k-ehf16-af32`: sentence embeddings.
140
- - `qwen3-reranker-0.6b-q4k` / `qwen-3-reranker-0-6b-q4k-ehf16-af32`: document reranking.
141
-
142
- Each lane has a catalog-owned hosted artifact pointer and explicit verify
143
- contract. The model support matrix is the source of truth for current receipts,
144
- runtime surfaces, and benchmark evidence.
145
-
146
- Defaults and policy choices must be represented in schema, manifest, config,
147
- profile, or rule assets. Runtime code must not invent hidden fallbacks or
148
- surface-specific behavior.
149
-
150
- ## Benchmark evidence
151
-
152
- The npm package includes the runtime, config contracts, schemas, and quickstart
153
- surface. It does not bundle the full benchmark result tree, local hardware
154
- artifacts, browser run outputs, or release evidence reports.
155
-
156
- The GitHub release matrix lists checked benchmark fixtures with hardware and
157
- backend, including Apple Metal and AMD Vulkan rows. Each row states backend,
158
- surface, comparator, metric direction, result, claim state, and evidence path.
159
- Metal and Vulkan rows are separate evidence lanes and should only be compared
160
- within a row.
161
-
162
- - Benchmark methodology:
163
- [docs/benchmark-methodology.md](https://github.com/clocksmith/doppler/blob/main/docs/benchmark-methodology.md)
164
- - Backend evidence summary:
165
- [benchmarks/vendors/results/doppler-backend-evidence-summary.svg](https://github.com/clocksmith/doppler/blob/main/benchmarks/vendors/results/doppler-backend-evidence-summary.svg)
166
-
167
- ## Model roadmap
168
-
169
- The README tracks model priority by source model, not by implementation lane.
170
- Doppler selects the best RDRR/runtime implementation from committed verification
171
- and benchmark evidence.
172
-
173
- The current roadmap is maintained in
174
- [docs/model-roadmap.md](https://github.com/clocksmith/doppler/blob/main/docs/model-roadmap.md).
175
- It replaces the old README quickstart-model table as the product-facing model
176
- plan.
177
-
178
- Current tiers:
179
-
180
- - Tier 1: Qwen 3.5 0.8B, Qwen 3 Embedding 0.6B, Qwen 3 Reranker 0.6B.
181
- - Tier 2: Qwen 3.5 2B, Gemma 4 E2B.
182
- - Tier 3A: Qwen 3.6 27B, Gemma 4 12B.
183
- - Tier 3B: DiffusionGemma 26B A4B, Gemma 4 MoE after a concrete catalog target exists.
184
- - Stretch: Gemma 4 31B and larger Qwen 3.6/3.7-class dense targets after a real catalog/HF target exists.
185
-
186
- For exact runtime evidence, supported registry IDs, and benchmark claims, use
187
- the [model support matrix](https://github.com/clocksmith/doppler/blob/main/docs/model-support-matrix.md),
188
- [model support inventory](https://github.com/clocksmith/doppler/blob/main/docs/model-support-inventory.md),
189
- and [release matrix](https://github.com/clocksmith/doppler/blob/main/docs/release-matrix.md).
190
- Subsystem support tiers for direct-source inputs, advanced subpaths, diffusion,
191
- energy, and training live in the
192
- [subsystem support matrix](https://github.com/clocksmith/doppler/blob/main/docs/subsystem-support-matrix.md).
144
+ ## Start here
193
145
 
194
- ## Documentation
146
+ | Reader | Entry points |
147
+ | --- | --- |
148
+ | Application developers | [Getting started](https://github.com/clocksmith/doppler/blob/main/docs/getting-started.md), [Root API](https://github.com/clocksmith/doppler/blob/main/docs/api/root.md), and the [OpenAI-compatible server](#openai-compatible-server) |
149
+ | Model integrators | [RDRR format](https://github.com/clocksmith/doppler/blob/main/docs/rdrr-format.md), [model support](https://github.com/clocksmith/doppler/blob/main/docs/model-support-matrix.md), and [Program Bundles](https://github.com/clocksmith/doppler/blob/main/docs/integration/program-bundle.md) |
150
+ | Runtime and kernel engineers | [Architecture](https://github.com/clocksmith/doppler/blob/main/docs/architecture.md), [kernel optimization](https://github.com/clocksmith/doppler/blob/main/docs/developer-guides/16-kernel-performance-optimization.md), and the [challenger framework](https://github.com/clocksmith/doppler/blob/main/docs/local-gpu-challenger-framework.md) |
151
+ | Evidence reviewers | [Competition scoreboard](https://github.com/clocksmith/doppler/blob/main/docs/model-competition-scoreboard.md), [benchmark methodology](https://github.com/clocksmith/doppler/blob/main/docs/benchmark-methodology.md), and [release matrix](https://github.com/clocksmith/doppler/blob/main/docs/release-matrix.md) |
195
152
 
196
- - npm quickstart: run `npx doppler-gpu --help`
197
- - Docs index (canonical navigation): [docs/INDEX.md](https://github.com/clocksmith/doppler/blob/main/docs/INDEX.md)
198
- - First-run workflow: [docs/getting-started.md](https://github.com/clocksmith/doppler/blob/main/docs/getting-started.md)
199
- - CLI reference: [docs/cli.md](https://github.com/clocksmith/doppler/blob/main/docs/cli.md)
200
- - Runtime config contract: [docs/config.md](https://github.com/clocksmith/doppler/blob/main/docs/config.md)
201
- - Architecture: [docs/architecture.md](https://github.com/clocksmith/doppler/blob/main/docs/architecture.md)
202
- - Model roadmap: [docs/model-roadmap.md](https://github.com/clocksmith/doppler/blob/main/docs/model-roadmap.md)
203
- - Model support matrix: [docs/model-support-matrix.md](https://github.com/clocksmith/doppler/blob/main/docs/model-support-matrix.md)
153
+ The [docs index](https://github.com/clocksmith/doppler/blob/main/docs/INDEX.md)
154
+ owns the complete model, subsystem, API, architecture, and operator inventory.
204
155
 
205
156
  ## Environment requirements
206
157
 
207
- - WebGPU is required.
208
- - **Browser**: Current Chromium browsers with WebGPU enabled, including Chrome and Edge.
209
- WebGPU shipped in Chrome/Edge 113+. Firefox and Safari support varies.
210
- - **Node**: Requires a WebGPU provider (`webgpu` npm package). Installed automatically as an optional dependency.
158
+ WebGPU is required. Use a current Chromium browser; Node installs the `webgpu`
159
+ provider as an optional dependency.
211
160
 
212
161
  ## License
213
162