doppler-gpu 0.4.3 → 0.4.5
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.
- package/CHANGELOG.md +28 -0
- package/README.md +67 -34
- package/assets/doppler.svg +57 -1601
- package/package.json +36 -5
- package/src/cli/cli-output.js +34 -1
- package/src/cli/doppler-cli.d.ts +18 -0
- package/src/cli/doppler-cli.js +68 -35
- package/src/cli/doppler-serve.d.ts +210 -0
- package/src/cli/doppler-serve.js +468 -42
- package/src/client/doppler-registry.json +3 -5
- package/src/client/provider.d.ts +1 -1
- package/src/client/provider.js +1 -1
- package/src/client/runtime/index.js +13 -0
- package/src/client/runtime/lora.d.ts +11 -1
- package/src/client/runtime/lora.js +28 -7
- package/src/client/runtime/model-manager.js +35 -11
- package/src/client/runtime/model-session.d.ts +2 -1
- package/src/client/runtime/model-session.js +2 -2
- package/src/client/runtime/model-source.js +20 -1
- package/src/client/runtime/types.d.ts +8 -0
- package/src/client/wrap-pipeline-handle.d.ts +62 -0
- package/src/client/wrap-pipeline-handle.js +188 -13
- package/src/config/conversion/README.md +5 -0
- package/src/config/conversion/diffusiongemma/diffusiongemma-26b-a4b-it-q4k-ehf16-af16.json +646 -0
- package/src/config/conversion/embeddinggemma/google-embeddinggemma-300m-q4k-ehf16-af32.json +3 -3
- package/src/config/conversion/gemma3/gemma-3-1b-it-f16-af32.json +4 -4
- package/src/config/conversion/gemma3/gemma-3-1b-it-q4k-ehf16-af32.json +4 -4
- package/src/config/conversion/gemma3/gemma-3-270m-it-f16-af32.json +393 -0
- package/src/config/conversion/gemma3/gemma-3-270m-it-q4k-ehf16-af16.json +425 -0
- package/src/config/conversion/gemma3/gemma-3-270m-it-q4k-ehf16-af32.json +17 -9
- package/src/config/conversion/gemma3/translategemma-4b-1b-enes-q4k-ehf16-af32.json +4 -4
- package/src/config/conversion/gemma3/translategemma-4b-it-q4k-ehf16-af32.json +4 -4
- package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-af16.json +737 -0
- package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-af32.json +730 -0
- package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-hq4k-af16.json +780 -0
- package/src/config/conversion/gemma4/gemma-4-12b-it-text-w4a16-ct-ehf16-af16.json +716 -0
- package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af16.json +5 -5
- package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af32.json +4 -4
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af16-int4ple.json +14 -12
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32-int4ple.json +7 -7
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32.json +7 -7
- package/src/config/conversion/gemma4/gemma-4-moe-q4k-ehf16-af32.json +3 -3
- package/src/config/conversion/gpt-oss-20b-f16-xmxfp4.json +3 -3
- package/src/config/conversion/janus/janus-pro-1b-text-q4k-ehaf16.json +3 -3
- package/src/config/conversion/lfm2/lfm2.5-1.2b-instruct-q4k-ehf16-af32.json +3 -3
- package/src/config/conversion/qwen3/qwen-3-5-0-8b-q4k-ehaf16.json +8 -4
- package/src/config/conversion/qwen3/qwen-3-5-2b-q4k-ehaf16.json +8 -4
- package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-eaf16.json +9 -5
- package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-ehaf16.json +8 -4
- package/src/config/goal-completion-matrix.json +552 -0
- package/src/config/kernel-path-loader.js +192 -38
- package/src/config/kernels/kernel-ref-digests.js +52 -18
- package/src/config/kernels/registry.json +3017 -744
- package/src/config/merge-contract-check.js +7 -3
- package/src/config/merge-helpers.d.ts +2 -0
- package/src/config/merge-helpers.js +1 -0
- package/src/config/merge.d.ts +1 -0
- package/src/config/merge.js +30 -0
- package/src/config/model-lanes/catalog-lane-resolver.d.ts +66 -0
- package/src/config/model-lanes/catalog-lane-resolver.js +170 -0
- package/src/config/param-validator.js +4 -0
- package/src/config/runtime/experiments/verify/lfm2-verify.json +3 -3
- package/src/config/runtime/profiles/diffusiongemma-26b-a4b-throughput.json +26 -0
- package/src/config/runtime/profiles/gemma3-270m-q4k-throughput-overlapped-probe.json +42 -0
- package/src/config/runtime/profiles/gemma3-270m-q4k-throughput-probe.json +39 -0
- package/src/config/runtime/profiles/gemma4-12b-rdrr-safe-b4.json +48 -0
- package/src/config/runtime/profiles/gemma4-12b-rdrr-safe-single.json +46 -0
- package/src/config/schema/buffer-pool.schema.js +3 -3
- package/src/config/schema/claim-evidence-contract.schema.json +93 -0
- package/src/config/schema/command-surface-contract.schema.json +81 -0
- package/src/config/schema/converter.schema.d.ts +11 -0
- package/src/config/schema/converter.schema.js +8 -0
- package/src/config/schema/doppler.schema.d.ts +1 -0
- package/src/config/schema/doppler.schema.js +38 -6
- package/src/config/schema/execution-v1.schema.d.ts +27 -0
- package/src/config/schema/execution-v1.schema.js +28 -19
- package/src/config/schema/goal-completion-matrix.schema.json +223 -0
- package/src/config/schema/inference.schema.d.ts +11 -0
- package/src/config/schema/manifest.schema.d.ts +107 -1
- package/src/config/schema/manifest.schema.js +10 -1
- package/src/config/schema/model-artifact-contract.schema.json +119 -0
- package/src/config/schema/moe.schema.d.ts +13 -0
- package/src/config/schema/moe.schema.js +1 -1
- package/src/config/schema/policy-schema-registry.json +26 -0
- package/src/config/schema/subsystem-support-contract.schema.json +48 -0
- package/src/config/schema/training-metrics.schema.d.ts +1 -1
- package/src/config/schema/training-metrics.schema.js +4 -3
- package/src/config/source-packages/litert/gemma-4-12b-it.json +390 -0
- package/src/config/source-packages/litert/gemma-4-e2b-it.json +7 -7
- package/src/config/source-packages/litert/gemma-4-e4b-it.json +370 -0
- package/src/config/source-packages/registry.json +25 -0
- package/src/config/supported-operations.json +5 -0
- package/src/config/transforms/capability-transform-resolver.d.ts +13 -2
- package/src/config/transforms/execution-graph-transforms.d.ts +42 -0
- package/src/config/transforms/execution-graph-transforms.js +227 -230
- package/src/converter/conversion-plan.d.ts +2 -0
- package/src/converter/conversion-plan.js +18 -4
- package/src/converter/core.d.ts +8 -0
- package/src/converter/core.js +574 -12
- package/src/converter/quantization-info.js +179 -2
- package/src/converter/quantizer.d.ts +5 -4
- package/src/converter/quantizer.js +16 -2
- package/src/experimental/adapters/lora-loader.js +109 -24
- package/src/experimental/diffusion/index.d.ts +24 -0
- package/src/experimental/diffusion/index.js +10 -0
- package/src/experimental/distribution/descriptor-transport.d.ts +117 -0
- package/src/experimental/distribution/descriptor-transport.js +342 -0
- package/src/experimental/training/README.md +11 -0
- package/src/experimental/training/autograd.js +33 -8
- package/src/experimental/training/datasets/gepa-frontier.d.ts +41 -0
- package/src/experimental/training/datasets/gepa-frontier.js +220 -0
- package/src/experimental/training/datasets/index.d.ts +13 -1
- package/src/experimental/training/datasets/index.js +8 -1
- package/src/experimental/training/datasets/teacher-traces.d.ts +51 -0
- package/src/experimental/training/datasets/teacher-traces.js +279 -0
- package/src/experimental/training/datasets/text-pairs.d.ts +45 -0
- package/src/experimental/training/datasets/text-pairs.js +82 -5
- package/src/experimental/training/distillation/student-fixture.d.ts +5 -0
- package/src/experimental/training/distillation/student-fixture.js +290 -32
- package/src/experimental/training/export.d.ts +9 -0
- package/src/experimental/training/export.js +113 -7
- package/src/experimental/training/index.d.ts +13 -1
- package/src/experimental/training/index.js +13 -1
- package/src/experimental/training/lora-pipeline.d.ts +143 -0
- package/src/experimental/training/lora-pipeline.js +1737 -49
- package/src/experimental/training/lora.js +23 -1
- package/src/experimental/training/loss.js +1 -1
- package/src/experimental/training/operator-agent-eval.d.ts +65 -0
- package/src/experimental/training/operator-agent-eval.js +326 -0
- package/src/experimental/training/operator-command.js +310 -51
- package/src/experimental/training/runner.d.ts +1 -0
- package/src/experimental/training/runner.js +4 -0
- package/src/experimental/training/workload-packs/distill-glm52-doppler-agent-harness-gemma270m-f16-sft-tiny.json +160 -0
- package/src/experimental/training/workload-packs/distill-glm52-doppler-code-agent-gemma270m-f16-sft-tiny.json +175 -0
- package/src/experimental/training/workload-packs/distill-glm52-doppler-code-agent-sft-tiny.json +175 -0
- package/src/experimental/training/workload-packs/distill-glm52-doppler-js-json-gemma270m-f16-sft-tiny.json +160 -0
- package/src/experimental/training/workload-packs/distill-glm52-doppler-review-gemma270m-f16-sft-tiny.json +162 -0
- package/src/experimental/training/workload-packs/distill-glm52-doppler-wgsl-gemma270m-f16-sft-tiny.json +160 -0
- package/src/experimental/training/workload-packs/lora-doppler-agent-harness-gemma270m-f16-tiny.json +137 -0
- package/src/experimental/training/workload-packs/lora-doppler-code-agent-gemma270m-f16-tiny.json +152 -0
- package/src/experimental/training/workload-packs/lora-doppler-code-agent-tiny.json +152 -0
- package/src/experimental/training/workload-packs/lora-doppler-js-json-gemma270m-f16-tiny.json +137 -0
- package/src/experimental/training/workload-packs/lora-doppler-review-gemma270m-f16-tiny.json +139 -0
- package/src/experimental/training/workload-packs/lora-doppler-wgsl-gemma270m-f16-tiny.json +137 -0
- package/src/experimental/training/workload-packs/registry.json +186 -6
- package/src/experimental/training/workloads.d.ts +37 -0
- package/src/experimental/training/workloads.js +110 -1
- package/src/formats/rdrr/functional-descriptor.d.ts +73 -0
- package/src/formats/rdrr/functional-descriptor.js +230 -0
- package/src/formats/rdrr/index.d.ts +1 -0
- package/src/formats/rdrr/index.js +1 -0
- package/src/formats/rdrr/parsing.js +22 -6
- package/src/formats/rdrr/source-transform-contract.js +10 -1
- package/src/formats/rdrr/storage-descriptor.js +9 -2
- package/src/formats/rdrr/types.d.ts +5 -3
- package/src/formats/rdrr/validation.js +15 -1
- package/src/gpu/command-recorder.d.ts +43 -3
- package/src/gpu/command-recorder.js +122 -25
- package/src/gpu/device.js +111 -11
- package/src/gpu/kernel-tuner/cache.js +1 -5
- package/src/gpu/kernels/attention.js +17 -9
- package/src/gpu/kernels/attention_decode_online_head256_f16kv.wgsl +236 -0
- package/src/gpu/kernels/attention_streaming_f16.wgsl +7 -5
- package/src/gpu/kernels/backward/cross_entropy_backward.js +72 -3
- package/src/gpu/kernels/backward/cross_entropy_backward.wgsl +2 -2
- package/src/gpu/kernels/cast.js +74 -67
- package/src/gpu/kernels/codegen/patches/attention_streaming_f16.from.attention_streaming.diff +46 -1
- package/src/gpu/kernels/codegen/patches/gelu_f16.from.gelu.diff +0 -20
- package/src/gpu/kernels/codegen/patches/residual_f16.from.residual.diff +15 -12
- package/src/gpu/kernels/codegen/patches/rmsnorm_f16.from.rmsnorm.diff +59 -44
- package/src/gpu/kernels/diffusion-gemma-sampling.d.ts +19 -0
- package/src/gpu/kernels/diffusion-gemma-sampling.js +104 -0
- package/src/gpu/kernels/diffusion_gemma_sampling.wgsl +130 -0
- package/src/gpu/kernels/dispatch.js +13 -3
- package/src/gpu/kernels/fused_ffn.js +3 -4
- package/src/gpu/kernels/fused_matmul_litert_int4_f16a.wgsl +78 -0
- package/src/gpu/kernels/fused_matmul_litert_int4_f16a_f32out.wgsl +78 -0
- package/src/gpu/kernels/fused_matmul_litert_int4_f32a_f32out.wgsl +76 -0
- package/src/gpu/kernels/fused_matmul_rmsnorm.js +101 -80
- package/src/gpu/kernels/fused_matmul_w4a16_batched_f16a.wgsl +116 -0
- package/src/gpu/kernels/fused_matmul_w4a16_batched_f16a_f32out.wgsl +116 -0
- package/src/gpu/kernels/fused_matmul_w4a16_f16a.wgsl +95 -0
- package/src/gpu/kernels/fused_matmul_w4a16_f16a_f32out.wgsl +95 -0
- package/src/gpu/kernels/gather.d.ts +84 -1
- package/src/gpu/kernels/gather.js +353 -5
- package/src/gpu/kernels/gather_litert_int4_vec4_f16_out.wgsl +75 -0
- package/src/gpu/kernels/gather_litert_int4_vec4_f32_out.wgsl +72 -0
- package/src/gpu/kernels/gather_split4_f16_vec4_f16_out.wgsl +103 -0
- package/src/gpu/kernels/gather_split8_f16_vec4_f16_out.wgsl +143 -0
- package/src/gpu/kernels/gather_split8_f16_vec4_f32_out.wgsl +143 -0
- package/src/gpu/kernels/gelu.wgsl +11 -7
- package/src/gpu/kernels/gelu_f16.wgsl +11 -7
- package/src/gpu/kernels/gemma4-route-expert.d.ts +35 -0
- package/src/gpu/kernels/gemma4-route-expert.js +126 -0
- package/src/gpu/kernels/gemma4_route_q4_matmul_f16a.wgsl +210 -0
- package/src/gpu/kernels/index.d.ts +71 -0
- package/src/gpu/kernels/index.js +52 -0
- package/src/gpu/kernels/kv-cache-write.d.ts +22 -0
- package/src/gpu/kernels/kv-cache-write.js +126 -0
- package/src/gpu/kernels/kv_cache_write_f32_to_f16.wgsl +32 -0
- package/src/gpu/kernels/lm-head-argmax.d.ts +19 -0
- package/src/gpu/kernels/lm-head-argmax.js +218 -0
- package/src/gpu/kernels/lm_head_argmax_f16.wgsl +189 -0
- package/src/gpu/kernels/matmul-dispatch.d.ts +11 -0
- package/src/gpu/kernels/matmul-dispatch.js +37 -1
- package/src/gpu/kernels/matmul-selection.d.ts +5 -5
- package/src/gpu/kernels/matmul-selection.js +104 -9
- package/src/gpu/kernels/matmul.d.ts +1 -1
- package/src/gpu/kernels/matmul.js +151 -43
- package/src/gpu/kernels/moe.d.ts +1 -0
- package/src/gpu/kernels/moe.js +46 -18
- package/src/gpu/kernels/pipeline-cache.d.ts +8 -0
- package/src/gpu/kernels/pipeline-cache.js +24 -0
- package/src/gpu/kernels/rep-penalty.d.ts +21 -0
- package/src/gpu/kernels/rep-penalty.js +17 -9
- package/src/gpu/kernels/residual.d.ts +1 -0
- package/src/gpu/kernels/residual.js +13 -1
- package/src/gpu/kernels/residual.wgsl +3 -3
- package/src/gpu/kernels/residual_f16.wgsl +5 -3
- package/src/gpu/kernels/residual_f16_vec4.wgsl +6 -6
- package/src/gpu/kernels/residual_vec4.wgsl +6 -6
- package/src/gpu/kernels/rmsnorm-pair.d.ts +39 -0
- package/src/gpu/kernels/rmsnorm-pair.js +152 -0
- package/src/gpu/kernels/rmsnorm-qk.d.ts +41 -0
- package/src/gpu/kernels/rmsnorm-qk.js +128 -0
- package/src/gpu/kernels/rmsnorm.d.ts +31 -0
- package/src/gpu/kernels/rmsnorm.js +32 -7
- package/src/gpu/kernels/rmsnorm.wgsl +10 -6
- package/src/gpu/kernels/rmsnorm_f16.wgsl +7 -3
- package/src/gpu/kernels/rmsnorm_pair.wgsl +224 -0
- package/src/gpu/kernels/rmsnorm_qk.wgsl +118 -0
- package/src/gpu/kernels/rmsnorm_qk_f16.wgsl +143 -0
- package/src/gpu/kernels/rope-qk.d.ts +41 -0
- package/src/gpu/kernels/rope-qk.js +143 -0
- package/src/gpu/kernels/rope_qk.wgsl +95 -0
- package/src/gpu/kernels/rope_qk_f16.wgsl +97 -0
- package/src/gpu/kernels/rule-registry.js +50 -10
- package/src/gpu/kernels/sample.js +212 -187
- package/src/gpu/kernels/scatter_add_dynamic_expert_scale.wgsl +47 -0
- package/src/gpu/kernels/scatter_add_dynamic_f16_expert_scale.wgsl +49 -0
- package/src/gpu/kernels/scatter_add_dynamic_f16_weights_expert_scale.wgsl +49 -0
- package/src/gpu/kernels/scatter_add_routes_f16_weights_expert_scale.wgsl +45 -0
- package/src/gpu/kernels/shader-cache.js +1 -21
- package/src/gpu/kernels/soft-embedding.d.ts +28 -0
- package/src/gpu/kernels/soft-embedding.js +447 -0
- package/src/gpu/kernels/soft_embedding_logits_f16.wgsl +107 -0
- package/src/gpu/kernels/soft_embedding_split_f16.wgsl +46 -0
- package/src/gpu/kernels/softmax.wgsl +5 -80
- package/src/gpu/kernels/split-qkv-rmsnorm-qk.d.ts +44 -0
- package/src/gpu/kernels/split-qkv-rmsnorm-qk.js +154 -0
- package/src/gpu/kernels/split-qkv-rmsnorm-rope-qk.d.ts +59 -0
- package/src/gpu/kernels/split-qkv-rmsnorm-rope-qk.js +245 -0
- package/src/gpu/kernels/split_qkv_rmsnorm_qk.wgsl +146 -0
- package/src/gpu/kernels/split_qkv_rmsnorm_rope_qk.wgsl +185 -0
- package/src/gpu/kernels/split_qkv_rmsnorm_rope_qk_f16kv_cache.wgsl +187 -0
- package/src/gpu/kernels/uniform-utils.js +2 -5
- package/src/gpu/kernels/utils.d.ts +16 -1
- package/src/gpu/kernels/utils.js +42 -13
- package/src/gpu/readback-utils.js +10 -2
- package/src/gpu/uniform-cache.d.ts +10 -1
- package/src/gpu/uniform-cache.js +25 -7
- package/src/gpu/weight-buffer.d.ts +76 -7
- package/src/gpu/weight-buffer.js +85 -1
- package/src/inference/browser-harness-diffusion-energy-suites.js +269 -1
- package/src/inference/browser-harness-suite-helpers.d.ts +9 -0
- package/src/inference/browser-harness-suite-helpers.js +438 -2
- package/src/inference/browser-harness-text-helpers.d.ts +77 -0
- package/src/inference/browser-harness-text-helpers.js +236 -11
- package/src/inference/browser-harness.d.ts +8 -0
- package/src/inference/browser-harness.js +292 -8
- package/src/inference/decode-ring.d.ts +1 -0
- package/src/inference/decode-ring.js +7 -5
- package/src/inference/kv-cache/base.d.ts +25 -0
- package/src/inference/kv-cache/base.js +109 -1
- package/src/inference/kv-cache/mixed-geometry.d.ts +10 -0
- package/src/inference/kv-cache/mixed-geometry.js +109 -2
- package/src/inference/kv-cache/sliding-window.d.ts +14 -0
- package/src/inference/kv-cache/sliding-window.js +88 -0
- package/src/inference/model-lanes/execution-lane-resolver.d.ts +61 -0
- package/src/inference/model-lanes/execution-lane-resolver.js +126 -0
- package/src/inference/moe-router.d.ts +14 -1
- package/src/inference/moe-router.js +24 -1
- package/src/inference/pipelines/audio/gemma4.js +52 -1
- package/src/inference/pipelines/diffusion-gemma/config.d.ts +27 -0
- package/src/inference/pipelines/diffusion-gemma/config.js +149 -0
- package/src/inference/pipelines/diffusion-gemma/index.d.ts +31 -0
- package/src/inference/pipelines/diffusion-gemma/index.js +12 -0
- package/src/inference/pipelines/diffusion-gemma/pipeline.d.ts +69 -0
- package/src/inference/pipelines/diffusion-gemma/pipeline.js +380 -0
- package/src/inference/pipelines/diffusion-gemma/sampling.d.ts +128 -0
- package/src/inference/pipelines/diffusion-gemma/sampling.js +395 -0
- package/src/inference/pipelines/text/attention/dispatch-params.d.ts +28 -4
- package/src/inference/pipelines/text/attention/dispatch-params.js +176 -2
- package/src/inference/pipelines/text/attention/projections.d.ts +71 -2
- package/src/inference/pipelines/text/attention/projections.js +329 -49
- package/src/inference/pipelines/text/attention/quant-options.d.ts +38 -0
- package/src/inference/pipelines/text/attention/record.js +446 -73
- package/src/inference/pipelines/text/attention/run.js +251 -49
- package/src/inference/pipelines/text/attention/types.d.ts +5 -1
- package/src/inference/pipelines/text/config.d.ts +22 -1
- package/src/inference/pipelines/text/config.js +199 -22
- package/src/inference/pipelines/text/dtype-contract.d.ts +1 -1
- package/src/inference/pipelines/text/dtype-contract.js +4 -1
- package/src/inference/pipelines/text/embed.d.ts +4 -3
- package/src/inference/pipelines/text/embed.js +16 -6
- package/src/inference/pipelines/text/execution-plan.d.ts +8 -0
- package/src/inference/pipelines/text/execution-plan.js +6 -0
- package/src/inference/pipelines/text/execution-runtime-builders.d.ts +4 -0
- package/src/inference/pipelines/text/execution-runtime-builders.js +44 -28
- package/src/inference/pipelines/text/execution-v1.d.ts +2 -0
- package/src/inference/pipelines/text/execution-v1.js +373 -139
- package/src/inference/pipelines/text/ffn/dense.js +44 -13
- package/src/inference/pipelines/text/ffn/moe.d.ts +4 -1
- package/src/inference/pipelines/text/ffn/moe.js +9 -2
- package/src/inference/pipelines/text/ffn/sandwich.d.ts +3 -1
- package/src/inference/pipelines/text/ffn/sandwich.js +237 -15
- package/src/inference/pipelines/text/ffn/standard.d.ts +3 -1
- package/src/inference/pipelines/text/ffn/standard.js +44 -4
- package/src/inference/pipelines/text/generator-decode-policy.d.ts +81 -0
- package/src/inference/pipelines/text/generator-decode-policy.js +2 -12
- package/src/inference/pipelines/text/generator-helpers.js +5 -2
- package/src/inference/pipelines/text/generator-prefill-helpers.d.ts +84 -0
- package/src/inference/pipelines/text/generator-prefill-helpers.js +67 -1
- package/src/inference/pipelines/text/generator-runtime.d.ts +1 -0
- package/src/inference/pipelines/text/generator-steps.d.ts +38 -3
- package/src/inference/pipelines/text/generator-steps.js +268 -77
- package/src/inference/pipelines/text/generator.d.ts +39 -0
- package/src/inference/pipelines/text/generator.js +1105 -18
- package/src/inference/pipelines/text/init.d.ts +8 -1
- package/src/inference/pipelines/text/init.js +30 -1
- package/src/inference/pipelines/text/layer-plan-gpu.js +62 -14
- package/src/inference/pipelines/text/layer.js +272 -25
- package/src/inference/pipelines/text/linear-attention.js +6 -4
- package/src/inference/pipelines/text/logits/gpu.d.ts +49 -2
- package/src/inference/pipelines/text/logits/gpu.js +417 -66
- package/src/inference/pipelines/text/logits/index.d.ts +2 -2
- package/src/inference/pipelines/text/logits/index.js +96 -42
- package/src/inference/pipelines/text/logits/types.d.ts +2 -2
- package/src/inference/pipelines/text/lora-apply.js +17 -0
- package/src/inference/pipelines/text/model-load.js +5 -1
- package/src/inference/pipelines/text/moe-gpu.d.ts +13 -0
- package/src/inference/pipelines/text/moe-gpu.js +445 -16
- package/src/inference/pipelines/text/moe-impl.d.ts +13 -3
- package/src/inference/pipelines/text/moe-shape-validator.js +13 -1
- package/src/inference/pipelines/text/ops.d.ts +29 -2
- package/src/inference/pipelines/text/ops.js +19 -0
- package/src/inference/pipelines/text/per-layer-inputs.js +5 -0
- package/src/inference/pipelines/text/probes.js +4 -5
- package/src/inference/pipelines/text/sampling-config.d.ts +4 -10
- package/src/inference/pipelines/text/sampling-config.js +103 -64
- package/src/inference/pipelines/text/sampling.js +13 -9
- package/src/inference/pipelines/text/state.js +17 -0
- package/src/inference/pipelines/text/tsir-fixture-writer.d.ts +57 -0
- package/src/inference/pipelines/text/tsir-fixture-writer.js +8 -9
- package/src/inference/pipelines/text/types.d.ts +84 -3
- package/src/inference/pipelines/text/weights.d.ts +4 -2
- package/src/inference/pipelines/text/weights.js +4 -5
- package/src/inference/pipelines/text.d.ts +24 -0
- package/src/inference/pipelines/text.js +162 -38
- package/src/inference/pipelines/vision/gemma4.js +20 -9
- package/src/inference/tokenizer.d.ts +30 -0
- package/src/inference/tokenizer.js +340 -164
- package/src/inference/tokenizers/bundled.js +71 -49
- package/src/loader/doppler-loader.d.ts +11 -0
- package/src/loader/doppler-loader.js +290 -15
- package/src/loader/embedding-limit-preflight.d.ts +31 -0
- package/src/loader/embedding-limit-preflight.js +216 -0
- package/src/loader/embedding-loader.d.ts +8 -2
- package/src/loader/embedding-loader.js +263 -8
- package/src/loader/experts/expert-loader.d.ts +4 -1
- package/src/loader/experts/expert-loader.js +174 -5
- package/src/loader/final-weights-loader.d.ts +5 -0
- package/src/loader/final-weights-loader.js +167 -4
- package/src/loader/layer-loader.d.ts +2 -0
- package/src/loader/layer-loader.js +74 -6
- package/src/loader/loader-types.d.ts +59 -2
- package/src/loader/manifest-config.d.ts +8 -0
- package/src/loader/shard-resolver.js +37 -4
- package/src/loader/tensors/functional-descriptor-loader.d.ts +9 -0
- package/src/loader/tensors/functional-descriptor-loader.js +377 -0
- package/src/loader/tensors/tensor-loader.d.ts +36 -1
- package/src/loader/tensors/tensor-loader.js +437 -21
- package/src/loader/tensors/tensor-reader.d.ts +4 -1
- package/src/loader/tensors/tensor-reader.js +11 -4
- package/src/loader/weights.d.ts +3 -1
- package/src/memory/buffer-pool.js +12 -9
- package/src/models/diffusiongemma.d.ts +7 -0
- package/src/models/diffusiongemma.js +25 -0
- package/src/models/gemma4.js +24 -0
- package/src/models/qwen3.js +2 -2
- package/src/rules/execution-rules-contract-check.js +93 -10
- package/src/rules/inference/capability-transforms.rules.json +26 -2
- package/src/rules/inference/config.rules.json +1 -0
- package/src/rules/inference/dtype.rules.json +7 -0
- package/src/rules/inference/execution.rules.json +108 -17
- package/src/rules/inference/ffn.rules.json +18 -0
- package/src/rules/kernels/gather.rules.json +81 -0
- package/src/rules/kernels/matmul.rules.json +19 -0
- package/src/rules/kernels/moe.rules.json +3 -0
- package/src/rules/kernels/rep-penalty.rules.json +6 -0
- package/src/rules/kernels/rmsnorm-qk.rules.json +6 -0
- package/src/rules/kernels/rmsnorm.rules.json +42 -7
- package/src/rules/kernels/rope-qk.rules.json +6 -0
- package/src/rules/loader/tensor-loader.rules.json +8 -0
- package/src/rules/rule-registry.js +6 -0
- package/src/storage/artifact-storage-context.d.ts +1 -0
- package/src/storage/artifact-storage-context.js +10 -0
- package/src/tooling/browser-command-runner.d.ts +0 -3
- package/src/tooling/browser-command-runner.js +2 -14
- package/src/tooling/command-api-family-normalizers.js +4 -2
- package/src/tooling/command-api-helpers.js +2 -1
- package/src/tooling/command-api.d.ts +2 -0
- package/src/tooling/command-api.js +1 -0
- package/src/tooling/command-runner-shared.d.ts +4 -0
- package/src/tooling/command-runner-shared.js +2 -1
- package/src/tooling/diagnose-runner.js +2 -0
- package/src/tooling/litert-package-runtime.js +553 -16
- package/src/tooling/node-browser-command-runner.js +10 -4
- package/src/tooling/node-command-runner.js +7 -2
- package/src/tooling/node-convert-worker-pool.js +3 -0
- package/src/tooling/node-convert-worker.js +1 -0
- package/src/tooling/node-converter.js +12 -0
- package/src/tooling/node-webgpu.d.ts +1 -0
- package/src/tooling/node-webgpu.js +23 -11
- package/src/tooling/program-bundle.d.ts +52 -1
- package/src/tooling/rdrr-integrity-refresh.d.ts +18 -0
- package/src/tooling/source-artifact-adapter.js +7 -2
- package/src/tooling/source-package-profiles.d.ts +18 -1
- package/src/tooling/source-package-profiles.js +109 -1
- package/src/tooling/source-runtime-bundle.d.ts +1 -0
- package/src/tooling/source-runtime-bundle.js +27 -0
- package/src/tooling/source-runtime-converter-config.js +12 -3
- package/src/tooling-experimental-exports.shared.d.ts +10 -0
- package/src/tooling-experimental-exports.shared.js +10 -0
- package/src/version.js +1 -1
|
@@ -0,0 +1,137 @@
|
|
|
1
|
+
{
|
|
2
|
+
"schemaVersion": 1,
|
|
3
|
+
"kind": "lora",
|
|
4
|
+
"id": "lora-doppler-js-json-gemma270m-f16-tiny",
|
|
5
|
+
"description": "Tiny native local LoRA workload for Doppler JavaScript and JSON writing behavior.",
|
|
6
|
+
"claimBoundary": "Executable local LoRA adapter training receipt for JS/JSON traces on a Doppler f16 base; not a model-quality claim.",
|
|
7
|
+
"seed": 1337,
|
|
8
|
+
"baseModelId": "gemma-3-270m-it-f16-af32",
|
|
9
|
+
"studentModelId": null,
|
|
10
|
+
"teacherModelId": "zai-org/GLM-5.2",
|
|
11
|
+
"datasetId": "doppler-js-json-sft-tiny",
|
|
12
|
+
"datasetPath": "tests/fixtures/training/doppler-js-json-sft-tiny.jsonl",
|
|
13
|
+
"evalDatasets": [
|
|
14
|
+
{
|
|
15
|
+
"id": "doppler-js-json-heldout-tiny-eval",
|
|
16
|
+
"datasetPath": "tests/fixtures/training/doppler-js-json-heldout-tiny.jsonl",
|
|
17
|
+
"evalKind": "text_generation",
|
|
18
|
+
"metrics": [
|
|
19
|
+
"loss",
|
|
20
|
+
"agent_heldout_gate"
|
|
21
|
+
],
|
|
22
|
+
"decodePolicy": {
|
|
23
|
+
"maxTokens": 160,
|
|
24
|
+
"stopOnEos": true
|
|
25
|
+
},
|
|
26
|
+
"scoreboardColumns": [
|
|
27
|
+
"loss",
|
|
28
|
+
"baseline.loss",
|
|
29
|
+
"qualityClaim.absoluteImprovement",
|
|
30
|
+
"agentEval.passRate"
|
|
31
|
+
],
|
|
32
|
+
"quality": {
|
|
33
|
+
"baseline": "base_model",
|
|
34
|
+
"requireImprovement": true,
|
|
35
|
+
"minAbsoluteImprovement": 0,
|
|
36
|
+
"minRelativeImprovement": 0
|
|
37
|
+
},
|
|
38
|
+
"sourceLangs": [],
|
|
39
|
+
"targetLangs": [],
|
|
40
|
+
"pairAllowlist": [],
|
|
41
|
+
"agentEval": {
|
|
42
|
+
"suiteId": "doppler-js-json-heldout-v1",
|
|
43
|
+
"categories": [
|
|
44
|
+
"js_patching",
|
|
45
|
+
"manifest_config_review",
|
|
46
|
+
"no_hallucinated_files_tools"
|
|
47
|
+
],
|
|
48
|
+
"minPassRate": 1,
|
|
49
|
+
"requirePatchApplies": false,
|
|
50
|
+
"requireNoHallucinatedFiles": true,
|
|
51
|
+
"requireNoHallucinatedTools": true,
|
|
52
|
+
"allowedFiles": [
|
|
53
|
+
"src/experimental/training/workloads.js",
|
|
54
|
+
"src/config/merge.js"
|
|
55
|
+
],
|
|
56
|
+
"allowedTools": [
|
|
57
|
+
"rg",
|
|
58
|
+
"node",
|
|
59
|
+
"npm",
|
|
60
|
+
"git",
|
|
61
|
+
"exec_command",
|
|
62
|
+
"apply_patch"
|
|
63
|
+
]
|
|
64
|
+
}
|
|
65
|
+
}
|
|
66
|
+
],
|
|
67
|
+
"trainingSchemaVersion": 1,
|
|
68
|
+
"checkpointEvery": 1,
|
|
69
|
+
"selectionMetric": "loss",
|
|
70
|
+
"selectionGoal": "min",
|
|
71
|
+
"surfaceSupport": "node",
|
|
72
|
+
"training": {
|
|
73
|
+
"optimizer": {
|
|
74
|
+
"type": "adamw",
|
|
75
|
+
"lr": 0.0001,
|
|
76
|
+
"beta1": 0.9,
|
|
77
|
+
"beta2": 0.999,
|
|
78
|
+
"eps": 1e-8,
|
|
79
|
+
"weightDecay": 0,
|
|
80
|
+
"scheduler": {
|
|
81
|
+
"enabled": false,
|
|
82
|
+
"type": "constant",
|
|
83
|
+
"warmupSteps": 0,
|
|
84
|
+
"stepSize": 1,
|
|
85
|
+
"gamma": 1,
|
|
86
|
+
"totalSteps": 1,
|
|
87
|
+
"minLr": 0
|
|
88
|
+
}
|
|
89
|
+
},
|
|
90
|
+
"batchSize": 1,
|
|
91
|
+
"accumSteps": 1,
|
|
92
|
+
"steps": 1,
|
|
93
|
+
"precision": {
|
|
94
|
+
"activations": "f16",
|
|
95
|
+
"gradients": "f16",
|
|
96
|
+
"loraParams": "f32"
|
|
97
|
+
},
|
|
98
|
+
"gradientClipping": {
|
|
99
|
+
"maxNorm": 1
|
|
100
|
+
}
|
|
101
|
+
},
|
|
102
|
+
"lora": {
|
|
103
|
+
"datasetFormat": "text-pairs",
|
|
104
|
+
"taskType": "text_generation",
|
|
105
|
+
"maxLength": 128,
|
|
106
|
+
"joinWith": "\n",
|
|
107
|
+
"adapter": {
|
|
108
|
+
"rank": 4,
|
|
109
|
+
"alpha": 8,
|
|
110
|
+
"dropout": 0,
|
|
111
|
+
"targetModules": [
|
|
112
|
+
"q_proj",
|
|
113
|
+
"v_proj"
|
|
114
|
+
]
|
|
115
|
+
},
|
|
116
|
+
"freeze": {
|
|
117
|
+
"encoder": false,
|
|
118
|
+
"prior": false,
|
|
119
|
+
"decoder": false,
|
|
120
|
+
"base": true,
|
|
121
|
+
"lora": false
|
|
122
|
+
},
|
|
123
|
+
"export": {
|
|
124
|
+
"enabled": true,
|
|
125
|
+
"atCheckpoints": false,
|
|
126
|
+
"select": "final",
|
|
127
|
+
"id": "doppler-js-json-gemma270m-f16-lora",
|
|
128
|
+
"name": "Doppler JS/JSON Gemma 270M F16 LoRA",
|
|
129
|
+
"format": "manifest_json"
|
|
130
|
+
},
|
|
131
|
+
"activation": {
|
|
132
|
+
"enabled": false,
|
|
133
|
+
"autoActivate": false,
|
|
134
|
+
"smokePrompt": "Patch this runtime JSON config."
|
|
135
|
+
}
|
|
136
|
+
}
|
|
137
|
+
}
|
|
@@ -0,0 +1,139 @@
|
|
|
1
|
+
{
|
|
2
|
+
"schemaVersion": 1,
|
|
3
|
+
"kind": "lora",
|
|
4
|
+
"id": "lora-doppler-review-gemma270m-f16-tiny",
|
|
5
|
+
"description": "Tiny native local LoRA workload for Doppler JavaScript, JSON, and WGSL review behavior.",
|
|
6
|
+
"claimBoundary": "Executable local LoRA adapter training receipt for review traces on a Doppler f16 base; not a model-quality claim.",
|
|
7
|
+
"seed": 1337,
|
|
8
|
+
"baseModelId": "gemma-3-270m-it-f16-af32",
|
|
9
|
+
"studentModelId": null,
|
|
10
|
+
"teacherModelId": "zai-org/GLM-5.2",
|
|
11
|
+
"datasetId": "doppler-review-sft-tiny",
|
|
12
|
+
"datasetPath": "tests/fixtures/training/doppler-review-sft-tiny.jsonl",
|
|
13
|
+
"evalDatasets": [
|
|
14
|
+
{
|
|
15
|
+
"id": "doppler-review-heldout-tiny-eval",
|
|
16
|
+
"datasetPath": "tests/fixtures/training/doppler-review-heldout-tiny.jsonl",
|
|
17
|
+
"evalKind": "text_generation",
|
|
18
|
+
"metrics": [
|
|
19
|
+
"loss",
|
|
20
|
+
"agent_heldout_gate"
|
|
21
|
+
],
|
|
22
|
+
"decodePolicy": {
|
|
23
|
+
"maxTokens": 160,
|
|
24
|
+
"stopOnEos": true
|
|
25
|
+
},
|
|
26
|
+
"scoreboardColumns": [
|
|
27
|
+
"loss",
|
|
28
|
+
"baseline.loss",
|
|
29
|
+
"qualityClaim.absoluteImprovement",
|
|
30
|
+
"agentEval.passRate"
|
|
31
|
+
],
|
|
32
|
+
"quality": {
|
|
33
|
+
"baseline": "base_model",
|
|
34
|
+
"requireImprovement": true,
|
|
35
|
+
"minAbsoluteImprovement": 0,
|
|
36
|
+
"minRelativeImprovement": 0
|
|
37
|
+
},
|
|
38
|
+
"sourceLangs": [],
|
|
39
|
+
"targetLangs": [],
|
|
40
|
+
"pairAllowlist": [],
|
|
41
|
+
"agentEval": {
|
|
42
|
+
"suiteId": "doppler-review-heldout-v1",
|
|
43
|
+
"categories": [
|
|
44
|
+
"manifest_config_review",
|
|
45
|
+
"wgsl_review",
|
|
46
|
+
"no_hallucinated_files_tools"
|
|
47
|
+
],
|
|
48
|
+
"minPassRate": 1,
|
|
49
|
+
"requirePatchApplies": false,
|
|
50
|
+
"requireNoHallucinatedFiles": true,
|
|
51
|
+
"requireNoHallucinatedTools": true,
|
|
52
|
+
"allowedFiles": [
|
|
53
|
+
"src/config/schema/manifest.schema.js",
|
|
54
|
+
"src/converter/manifest-inference.js",
|
|
55
|
+
"src/config/merge.js",
|
|
56
|
+
"src/gpu/kernels/matmul.wgsl"
|
|
57
|
+
],
|
|
58
|
+
"allowedTools": [
|
|
59
|
+
"rg",
|
|
60
|
+
"node",
|
|
61
|
+
"npm",
|
|
62
|
+
"git",
|
|
63
|
+
"exec_command",
|
|
64
|
+
"apply_patch"
|
|
65
|
+
]
|
|
66
|
+
}
|
|
67
|
+
}
|
|
68
|
+
],
|
|
69
|
+
"trainingSchemaVersion": 1,
|
|
70
|
+
"checkpointEvery": 1,
|
|
71
|
+
"selectionMetric": "loss",
|
|
72
|
+
"selectionGoal": "min",
|
|
73
|
+
"surfaceSupport": "node",
|
|
74
|
+
"training": {
|
|
75
|
+
"optimizer": {
|
|
76
|
+
"type": "adamw",
|
|
77
|
+
"lr": 0.0001,
|
|
78
|
+
"beta1": 0.9,
|
|
79
|
+
"beta2": 0.999,
|
|
80
|
+
"eps": 1e-8,
|
|
81
|
+
"weightDecay": 0,
|
|
82
|
+
"scheduler": {
|
|
83
|
+
"enabled": false,
|
|
84
|
+
"type": "constant",
|
|
85
|
+
"warmupSteps": 0,
|
|
86
|
+
"stepSize": 1,
|
|
87
|
+
"gamma": 1,
|
|
88
|
+
"totalSteps": 1,
|
|
89
|
+
"minLr": 0
|
|
90
|
+
}
|
|
91
|
+
},
|
|
92
|
+
"batchSize": 1,
|
|
93
|
+
"accumSteps": 1,
|
|
94
|
+
"steps": 1,
|
|
95
|
+
"precision": {
|
|
96
|
+
"activations": "f16",
|
|
97
|
+
"gradients": "f16",
|
|
98
|
+
"loraParams": "f32"
|
|
99
|
+
},
|
|
100
|
+
"gradientClipping": {
|
|
101
|
+
"maxNorm": 1
|
|
102
|
+
}
|
|
103
|
+
},
|
|
104
|
+
"lora": {
|
|
105
|
+
"datasetFormat": "text-pairs",
|
|
106
|
+
"taskType": "text_generation",
|
|
107
|
+
"maxLength": 128,
|
|
108
|
+
"joinWith": "\n",
|
|
109
|
+
"adapter": {
|
|
110
|
+
"rank": 4,
|
|
111
|
+
"alpha": 8,
|
|
112
|
+
"dropout": 0,
|
|
113
|
+
"targetModules": [
|
|
114
|
+
"q_proj",
|
|
115
|
+
"v_proj"
|
|
116
|
+
]
|
|
117
|
+
},
|
|
118
|
+
"freeze": {
|
|
119
|
+
"encoder": false,
|
|
120
|
+
"prior": false,
|
|
121
|
+
"decoder": false,
|
|
122
|
+
"base": true,
|
|
123
|
+
"lora": false
|
|
124
|
+
},
|
|
125
|
+
"export": {
|
|
126
|
+
"enabled": true,
|
|
127
|
+
"atCheckpoints": false,
|
|
128
|
+
"select": "final",
|
|
129
|
+
"id": "doppler-review-gemma270m-f16-lora",
|
|
130
|
+
"name": "Doppler Review Gemma 270M F16 LoRA",
|
|
131
|
+
"format": "manifest_json"
|
|
132
|
+
},
|
|
133
|
+
"activation": {
|
|
134
|
+
"enabled": false,
|
|
135
|
+
"autoActivate": false,
|
|
136
|
+
"smokePrompt": "Review this JS and WGSL patch."
|
|
137
|
+
}
|
|
138
|
+
}
|
|
139
|
+
}
|
|
@@ -0,0 +1,137 @@
|
|
|
1
|
+
{
|
|
2
|
+
"schemaVersion": 1,
|
|
3
|
+
"kind": "lora",
|
|
4
|
+
"id": "lora-doppler-wgsl-gemma270m-f16-tiny",
|
|
5
|
+
"description": "Tiny native local LoRA workload for Doppler WGSL writing behavior.",
|
|
6
|
+
"claimBoundary": "Executable local LoRA adapter training receipt for WGSL traces on a Doppler f16 base; not a model-quality claim.",
|
|
7
|
+
"seed": 1337,
|
|
8
|
+
"baseModelId": "gemma-3-270m-it-f16-af32",
|
|
9
|
+
"studentModelId": null,
|
|
10
|
+
"teacherModelId": "zai-org/GLM-5.2",
|
|
11
|
+
"datasetId": "doppler-wgsl-sft-tiny",
|
|
12
|
+
"datasetPath": "tests/fixtures/training/doppler-wgsl-sft-tiny.jsonl",
|
|
13
|
+
"evalDatasets": [
|
|
14
|
+
{
|
|
15
|
+
"id": "doppler-wgsl-heldout-tiny-eval",
|
|
16
|
+
"datasetPath": "tests/fixtures/training/doppler-wgsl-heldout-tiny.jsonl",
|
|
17
|
+
"evalKind": "text_generation",
|
|
18
|
+
"metrics": [
|
|
19
|
+
"loss",
|
|
20
|
+
"agent_heldout_gate"
|
|
21
|
+
],
|
|
22
|
+
"decodePolicy": {
|
|
23
|
+
"maxTokens": 160,
|
|
24
|
+
"stopOnEos": true
|
|
25
|
+
},
|
|
26
|
+
"scoreboardColumns": [
|
|
27
|
+
"loss",
|
|
28
|
+
"baseline.loss",
|
|
29
|
+
"qualityClaim.absoluteImprovement",
|
|
30
|
+
"agentEval.passRate"
|
|
31
|
+
],
|
|
32
|
+
"quality": {
|
|
33
|
+
"baseline": "base_model",
|
|
34
|
+
"requireImprovement": true,
|
|
35
|
+
"minAbsoluteImprovement": 0,
|
|
36
|
+
"minRelativeImprovement": 0
|
|
37
|
+
},
|
|
38
|
+
"sourceLangs": [],
|
|
39
|
+
"targetLangs": [],
|
|
40
|
+
"pairAllowlist": [],
|
|
41
|
+
"agentEval": {
|
|
42
|
+
"suiteId": "doppler-wgsl-heldout-v1",
|
|
43
|
+
"categories": [
|
|
44
|
+
"wgsl_review",
|
|
45
|
+
"no_hallucinated_files_tools"
|
|
46
|
+
],
|
|
47
|
+
"minPassRate": 1,
|
|
48
|
+
"requirePatchApplies": false,
|
|
49
|
+
"requireNoHallucinatedFiles": true,
|
|
50
|
+
"requireNoHallucinatedTools": true,
|
|
51
|
+
"allowedFiles": [
|
|
52
|
+
"src/gpu/kernels/matmul.wgsl",
|
|
53
|
+
"src/gpu/kernels/matmul.js",
|
|
54
|
+
"tests/kernels/matmul.test.js"
|
|
55
|
+
],
|
|
56
|
+
"allowedTools": [
|
|
57
|
+
"rg",
|
|
58
|
+
"node",
|
|
59
|
+
"npm",
|
|
60
|
+
"git",
|
|
61
|
+
"exec_command",
|
|
62
|
+
"apply_patch"
|
|
63
|
+
]
|
|
64
|
+
}
|
|
65
|
+
}
|
|
66
|
+
],
|
|
67
|
+
"trainingSchemaVersion": 1,
|
|
68
|
+
"checkpointEvery": 1,
|
|
69
|
+
"selectionMetric": "loss",
|
|
70
|
+
"selectionGoal": "min",
|
|
71
|
+
"surfaceSupport": "node",
|
|
72
|
+
"training": {
|
|
73
|
+
"optimizer": {
|
|
74
|
+
"type": "adamw",
|
|
75
|
+
"lr": 0.0001,
|
|
76
|
+
"beta1": 0.9,
|
|
77
|
+
"beta2": 0.999,
|
|
78
|
+
"eps": 1e-8,
|
|
79
|
+
"weightDecay": 0,
|
|
80
|
+
"scheduler": {
|
|
81
|
+
"enabled": false,
|
|
82
|
+
"type": "constant",
|
|
83
|
+
"warmupSteps": 0,
|
|
84
|
+
"stepSize": 1,
|
|
85
|
+
"gamma": 1,
|
|
86
|
+
"totalSteps": 1,
|
|
87
|
+
"minLr": 0
|
|
88
|
+
}
|
|
89
|
+
},
|
|
90
|
+
"batchSize": 1,
|
|
91
|
+
"accumSteps": 1,
|
|
92
|
+
"steps": 1,
|
|
93
|
+
"precision": {
|
|
94
|
+
"activations": "f16",
|
|
95
|
+
"gradients": "f16",
|
|
96
|
+
"loraParams": "f32"
|
|
97
|
+
},
|
|
98
|
+
"gradientClipping": {
|
|
99
|
+
"maxNorm": 1
|
|
100
|
+
}
|
|
101
|
+
},
|
|
102
|
+
"lora": {
|
|
103
|
+
"datasetFormat": "text-pairs",
|
|
104
|
+
"taskType": "text_generation",
|
|
105
|
+
"maxLength": 128,
|
|
106
|
+
"joinWith": "\n",
|
|
107
|
+
"adapter": {
|
|
108
|
+
"rank": 4,
|
|
109
|
+
"alpha": 8,
|
|
110
|
+
"dropout": 0,
|
|
111
|
+
"targetModules": [
|
|
112
|
+
"q_proj",
|
|
113
|
+
"v_proj"
|
|
114
|
+
]
|
|
115
|
+
},
|
|
116
|
+
"freeze": {
|
|
117
|
+
"encoder": false,
|
|
118
|
+
"prior": false,
|
|
119
|
+
"decoder": false,
|
|
120
|
+
"base": true,
|
|
121
|
+
"lora": false
|
|
122
|
+
},
|
|
123
|
+
"export": {
|
|
124
|
+
"enabled": true,
|
|
125
|
+
"atCheckpoints": false,
|
|
126
|
+
"select": "final",
|
|
127
|
+
"id": "doppler-wgsl-gemma270m-f16-lora",
|
|
128
|
+
"name": "Doppler WGSL Gemma 270M F16 LoRA",
|
|
129
|
+
"format": "manifest_json"
|
|
130
|
+
},
|
|
131
|
+
"activation": {
|
|
132
|
+
"enabled": false,
|
|
133
|
+
"autoActivate": false,
|
|
134
|
+
"smokePrompt": "Patch this WGSL bounds check."
|
|
135
|
+
}
|
|
136
|
+
}
|
|
137
|
+
}
|
|
@@ -1,9 +1,99 @@
|
|
|
1
1
|
{
|
|
2
2
|
"schemaVersion": 1,
|
|
3
3
|
"artifactType": "training_workload_registry",
|
|
4
|
-
"generatedAt": "2026-
|
|
5
|
-
"registryHash": "
|
|
4
|
+
"generatedAt": "2026-06-28T04:21:34.833Z",
|
|
5
|
+
"registryHash": "a8335d4b9d79d96d92109c3c4a46ffb01aea9fdb57588bca1ab5c2b7ac67ff52",
|
|
6
6
|
"workloads": [
|
|
7
|
+
{
|
|
8
|
+
"id": "distill-glm52-doppler-agent-harness-gemma270m-f16-sft-tiny",
|
|
9
|
+
"path": "src/experimental/training/workload-packs/distill-glm52-doppler-agent-harness-gemma270m-f16-sft-tiny.json",
|
|
10
|
+
"sha256": "922b33b7ecffe0b7ce958c17014c8c07bf78e36a61bc047004d91a958701469f",
|
|
11
|
+
"workloadKind": "distill",
|
|
12
|
+
"trainingSchemaVersion": 1,
|
|
13
|
+
"seed": 1337,
|
|
14
|
+
"checkpointEvery": 1,
|
|
15
|
+
"selectionMetric": "loss",
|
|
16
|
+
"selectionGoal": "min",
|
|
17
|
+
"surfaceSupport": "node",
|
|
18
|
+
"baselineReportId": "trn_distill-glm52-doppler-agent-harness-gemma270m-f16-sft-tiny_922b33b7ecff",
|
|
19
|
+
"claimBoundary": "Executable SFT distillation receipt for agent harness teacher traces into a Doppler f16 LoRA adapter; not a model-quality claim.",
|
|
20
|
+
"configHash": "c4e8522a40e2b90d21dc60ab938516283e51ba0a5154f55fbb2c61fac2fec33c"
|
|
21
|
+
},
|
|
22
|
+
{
|
|
23
|
+
"id": "distill-glm52-doppler-code-agent-gemma270m-f16-sft-tiny",
|
|
24
|
+
"path": "src/experimental/training/workload-packs/distill-glm52-doppler-code-agent-gemma270m-f16-sft-tiny.json",
|
|
25
|
+
"sha256": "53b048d76c320e8060427f854cbb17ce6731e18e90aba0f0afe46b9f8af4f66b",
|
|
26
|
+
"workloadKind": "distill",
|
|
27
|
+
"trainingSchemaVersion": 1,
|
|
28
|
+
"seed": 1337,
|
|
29
|
+
"checkpointEvery": 1,
|
|
30
|
+
"selectionMetric": "loss",
|
|
31
|
+
"selectionGoal": "min",
|
|
32
|
+
"surfaceSupport": "node",
|
|
33
|
+
"baselineReportId": "trn_distill-glm52-doppler-code-agent-gemma270m-f16-sft-tiny_53b048d76c32",
|
|
34
|
+
"claimBoundary": "Executable local SFT sequence distillation receipt from teacher-generated text-pair traces into a Doppler f16 LoRA adapter; not a model-quality claim.",
|
|
35
|
+
"configHash": "9478cae4cc89bb80443ad796657c3c9fe96f15808efe1e112cc48d16ac6d25ff"
|
|
36
|
+
},
|
|
37
|
+
{
|
|
38
|
+
"id": "distill-glm52-doppler-code-agent-sft-tiny",
|
|
39
|
+
"path": "src/experimental/training/workload-packs/distill-glm52-doppler-code-agent-sft-tiny.json",
|
|
40
|
+
"sha256": "97ae3d3db9fbf19bc700013bc9ac5c79a47305ef3aa658be90ccd84a9dd6cfcf",
|
|
41
|
+
"workloadKind": "distill",
|
|
42
|
+
"trainingSchemaVersion": 1,
|
|
43
|
+
"seed": 1337,
|
|
44
|
+
"checkpointEvery": 1,
|
|
45
|
+
"selectionMetric": "loss",
|
|
46
|
+
"selectionGoal": "min",
|
|
47
|
+
"surfaceSupport": "node",
|
|
48
|
+
"baselineReportId": "trn_distill-glm52-doppler-code-agent-sft-tiny_97ae3d3db9fb",
|
|
49
|
+
"claimBoundary": "Operator workload shape for SFT sequence distillation from teacher-generated text-pair traces into a Doppler-compatible LoRA adapter; not a model-quality claim.",
|
|
50
|
+
"configHash": "ecb07e0a97bce9762ad7d139db4a75e78d8ba771d1cc3b364607a5ed3107a2b0"
|
|
51
|
+
},
|
|
52
|
+
{
|
|
53
|
+
"id": "distill-glm52-doppler-js-json-gemma270m-f16-sft-tiny",
|
|
54
|
+
"path": "src/experimental/training/workload-packs/distill-glm52-doppler-js-json-gemma270m-f16-sft-tiny.json",
|
|
55
|
+
"sha256": "a011740d9a218d94fa587a1f0e2efb825daddd83e7eaab11f533c504bb6190c1",
|
|
56
|
+
"workloadKind": "distill",
|
|
57
|
+
"trainingSchemaVersion": 1,
|
|
58
|
+
"seed": 1337,
|
|
59
|
+
"checkpointEvery": 1,
|
|
60
|
+
"selectionMetric": "loss",
|
|
61
|
+
"selectionGoal": "min",
|
|
62
|
+
"surfaceSupport": "node",
|
|
63
|
+
"baselineReportId": "trn_distill-glm52-doppler-js-json-gemma270m-f16-sft-tiny_a011740d9a21",
|
|
64
|
+
"claimBoundary": "Executable SFT distillation receipt for JS/JSON teacher traces into a Doppler f16 LoRA adapter; not a model-quality claim.",
|
|
65
|
+
"configHash": "fb796350325110a6d506d06464085b28ddcff9aad30b495af1f1d0a07abfd517"
|
|
66
|
+
},
|
|
67
|
+
{
|
|
68
|
+
"id": "distill-glm52-doppler-review-gemma270m-f16-sft-tiny",
|
|
69
|
+
"path": "src/experimental/training/workload-packs/distill-glm52-doppler-review-gemma270m-f16-sft-tiny.json",
|
|
70
|
+
"sha256": "dbf198ed223b7390011a274629e8143ca801af08ce384a2f2d70cc838fa93b83",
|
|
71
|
+
"workloadKind": "distill",
|
|
72
|
+
"trainingSchemaVersion": 1,
|
|
73
|
+
"seed": 1337,
|
|
74
|
+
"checkpointEvery": 1,
|
|
75
|
+
"selectionMetric": "loss",
|
|
76
|
+
"selectionGoal": "min",
|
|
77
|
+
"surfaceSupport": "node",
|
|
78
|
+
"baselineReportId": "trn_distill-glm52-doppler-review-gemma270m-f16-sft-tiny_dbf198ed223b",
|
|
79
|
+
"claimBoundary": "Executable SFT distillation receipt for review teacher traces into a Doppler f16 LoRA adapter; not a model-quality claim.",
|
|
80
|
+
"configHash": "ecea2b77bffbd309207b61c4ea3f790d01a2df5d5ec067e4be37fbef0e929b61"
|
|
81
|
+
},
|
|
82
|
+
{
|
|
83
|
+
"id": "distill-glm52-doppler-wgsl-gemma270m-f16-sft-tiny",
|
|
84
|
+
"path": "src/experimental/training/workload-packs/distill-glm52-doppler-wgsl-gemma270m-f16-sft-tiny.json",
|
|
85
|
+
"sha256": "2a398efeec87073f25540c018f2b3c2904d699783151d0deeb1d00ab554ab5db",
|
|
86
|
+
"workloadKind": "distill",
|
|
87
|
+
"trainingSchemaVersion": 1,
|
|
88
|
+
"seed": 1337,
|
|
89
|
+
"checkpointEvery": 1,
|
|
90
|
+
"selectionMetric": "loss",
|
|
91
|
+
"selectionGoal": "min",
|
|
92
|
+
"surfaceSupport": "node",
|
|
93
|
+
"baselineReportId": "trn_distill-glm52-doppler-wgsl-gemma270m-f16-sft-tiny_2a398efeec87",
|
|
94
|
+
"claimBoundary": "Executable SFT distillation receipt for WGSL teacher traces into a Doppler f16 LoRA adapter; not a model-quality claim.",
|
|
95
|
+
"configHash": "4e3550c146a43d3fc3dd3a1ceca81b12569018bf88bc71e162a31db72a1e5c00"
|
|
96
|
+
},
|
|
7
97
|
{
|
|
8
98
|
"id": "distill-translategemma-medium",
|
|
9
99
|
"path": "src/experimental/training/workload-packs/distill-translategemma-medium.json",
|
|
@@ -17,7 +107,7 @@
|
|
|
17
107
|
"surfaceSupport": "node",
|
|
18
108
|
"baselineReportId": "trn_distill-translategemma-medium_af12934081cf",
|
|
19
109
|
"claimBoundary": "Practical deterministic distillation traceability for translation checkpoints and eval artifacts.",
|
|
20
|
-
"configHash": "
|
|
110
|
+
"configHash": "bf24974207f62fb2acc40708d3c5d6915582b84effe93bbcd30e7b4850035b79"
|
|
21
111
|
},
|
|
22
112
|
{
|
|
23
113
|
"id": "distill-translategemma-tiny",
|
|
@@ -32,7 +122,97 @@
|
|
|
32
122
|
"surfaceSupport": "node",
|
|
33
123
|
"baselineReportId": "trn_distill-translategemma-tiny_7a50fb0fd7c0",
|
|
34
124
|
"claimBoundary": "Practical deterministic distillation traceability for translation checkpoints and eval artifacts.",
|
|
35
|
-
"configHash": "
|
|
125
|
+
"configHash": "4585409faea75057aa286627c3f7e77244b814e710e9a0d9f69d9f9124453bdf"
|
|
126
|
+
},
|
|
127
|
+
{
|
|
128
|
+
"id": "lora-doppler-agent-harness-gemma270m-f16-tiny",
|
|
129
|
+
"path": "src/experimental/training/workload-packs/lora-doppler-agent-harness-gemma270m-f16-tiny.json",
|
|
130
|
+
"sha256": "ca334d7038fe2e4914cc347e78daa456685c27a8c5f0663576b455679323dac4",
|
|
131
|
+
"workloadKind": "lora",
|
|
132
|
+
"trainingSchemaVersion": 1,
|
|
133
|
+
"seed": 1337,
|
|
134
|
+
"checkpointEvery": 1,
|
|
135
|
+
"selectionMetric": "loss",
|
|
136
|
+
"selectionGoal": "min",
|
|
137
|
+
"surfaceSupport": "node",
|
|
138
|
+
"baselineReportId": "trn_lora-doppler-agent-harness-gemma270m-f16-tiny_ca334d7038fe",
|
|
139
|
+
"claimBoundary": "Executable local LoRA adapter training receipt for agent harness traces on a Doppler f16 base; not a model-quality claim.",
|
|
140
|
+
"configHash": "6297bb50f00283ca7c341560b6c07bec5b2219f4fbe662a20cb9ef5ab8beea01"
|
|
141
|
+
},
|
|
142
|
+
{
|
|
143
|
+
"id": "lora-doppler-code-agent-gemma270m-f16-tiny",
|
|
144
|
+
"path": "src/experimental/training/workload-packs/lora-doppler-code-agent-gemma270m-f16-tiny.json",
|
|
145
|
+
"sha256": "6e2546b248810d3206466eea2ddedcb2cd6f583552936e44dfc81200a74fbbae",
|
|
146
|
+
"workloadKind": "lora",
|
|
147
|
+
"trainingSchemaVersion": 1,
|
|
148
|
+
"seed": 1337,
|
|
149
|
+
"checkpointEvery": 1,
|
|
150
|
+
"selectionMetric": "loss",
|
|
151
|
+
"selectionGoal": "min",
|
|
152
|
+
"surfaceSupport": "node",
|
|
153
|
+
"baselineReportId": "trn_lora-doppler-code-agent-gemma270m-f16-tiny_6e2546b24881",
|
|
154
|
+
"claimBoundary": "Executable local causal-LM LoRA adapter training receipt on a Doppler f16 base; not a model-quality claim.",
|
|
155
|
+
"configHash": "9e9718b798d231d72c4a4709ec8568d85b74ac9831f039ed8902f8539a88e516"
|
|
156
|
+
},
|
|
157
|
+
{
|
|
158
|
+
"id": "lora-doppler-code-agent-tiny",
|
|
159
|
+
"path": "src/experimental/training/workload-packs/lora-doppler-code-agent-tiny.json",
|
|
160
|
+
"sha256": "e5127b6749bafd9d0ab70cad4caf32e87bebca8ac63616a19d30351133f522f2",
|
|
161
|
+
"workloadKind": "lora",
|
|
162
|
+
"trainingSchemaVersion": 1,
|
|
163
|
+
"seed": 1337,
|
|
164
|
+
"checkpointEvery": 1,
|
|
165
|
+
"selectionMetric": "loss",
|
|
166
|
+
"selectionGoal": "min",
|
|
167
|
+
"surfaceSupport": "node",
|
|
168
|
+
"baselineReportId": "trn_lora-doppler-code-agent-tiny_e5127b6749ba",
|
|
169
|
+
"claimBoundary": "Operator workload shape for causal-LM LoRA adapter training on code-agent text-pair traces; not a model-quality claim.",
|
|
170
|
+
"configHash": "4379b823776d8c30ca4168f389bc1fe84d10544536af2ae38277b9dc2a561f62"
|
|
171
|
+
},
|
|
172
|
+
{
|
|
173
|
+
"id": "lora-doppler-js-json-gemma270m-f16-tiny",
|
|
174
|
+
"path": "src/experimental/training/workload-packs/lora-doppler-js-json-gemma270m-f16-tiny.json",
|
|
175
|
+
"sha256": "87c159cf5936224b735ee61beeede2fb5ea2053cbb9eb72deb982ecf4cbbe46e",
|
|
176
|
+
"workloadKind": "lora",
|
|
177
|
+
"trainingSchemaVersion": 1,
|
|
178
|
+
"seed": 1337,
|
|
179
|
+
"checkpointEvery": 1,
|
|
180
|
+
"selectionMetric": "loss",
|
|
181
|
+
"selectionGoal": "min",
|
|
182
|
+
"surfaceSupport": "node",
|
|
183
|
+
"baselineReportId": "trn_lora-doppler-js-json-gemma270m-f16-tiny_87c159cf5936",
|
|
184
|
+
"claimBoundary": "Executable local LoRA adapter training receipt for JS/JSON traces on a Doppler f16 base; not a model-quality claim.",
|
|
185
|
+
"configHash": "ac2d418059fabdf1ede18af696c65e1a2955ef2eef7cb0ef5023a9bd3e718786"
|
|
186
|
+
},
|
|
187
|
+
{
|
|
188
|
+
"id": "lora-doppler-review-gemma270m-f16-tiny",
|
|
189
|
+
"path": "src/experimental/training/workload-packs/lora-doppler-review-gemma270m-f16-tiny.json",
|
|
190
|
+
"sha256": "75c514d464a89be5c3e84b1848c6096d1304f7383e9611c5d030a584b084e71f",
|
|
191
|
+
"workloadKind": "lora",
|
|
192
|
+
"trainingSchemaVersion": 1,
|
|
193
|
+
"seed": 1337,
|
|
194
|
+
"checkpointEvery": 1,
|
|
195
|
+
"selectionMetric": "loss",
|
|
196
|
+
"selectionGoal": "min",
|
|
197
|
+
"surfaceSupport": "node",
|
|
198
|
+
"baselineReportId": "trn_lora-doppler-review-gemma270m-f16-tiny_75c514d464a8",
|
|
199
|
+
"claimBoundary": "Executable local LoRA adapter training receipt for review traces on a Doppler f16 base; not a model-quality claim.",
|
|
200
|
+
"configHash": "36576b62d00057a30325fb4eb44e8332cd6cb493f17266bb86a3f62215432aa6"
|
|
201
|
+
},
|
|
202
|
+
{
|
|
203
|
+
"id": "lora-doppler-wgsl-gemma270m-f16-tiny",
|
|
204
|
+
"path": "src/experimental/training/workload-packs/lora-doppler-wgsl-gemma270m-f16-tiny.json",
|
|
205
|
+
"sha256": "40b6ef063ff13f42d4a380f007e6983619a2b649cc7696583b3b584285afc709",
|
|
206
|
+
"workloadKind": "lora",
|
|
207
|
+
"trainingSchemaVersion": 1,
|
|
208
|
+
"seed": 1337,
|
|
209
|
+
"checkpointEvery": 1,
|
|
210
|
+
"selectionMetric": "loss",
|
|
211
|
+
"selectionGoal": "min",
|
|
212
|
+
"surfaceSupport": "node",
|
|
213
|
+
"baselineReportId": "trn_lora-doppler-wgsl-gemma270m-f16-tiny_40b6ef063ff1",
|
|
214
|
+
"claimBoundary": "Executable local LoRA adapter training receipt for WGSL traces on a Doppler f16 base; not a model-quality claim.",
|
|
215
|
+
"configHash": "a8af4a7f4262cda72d5e1e4be0d31871e1ec16cdde93ec128d5d0ce874f286e1"
|
|
36
216
|
},
|
|
37
217
|
{
|
|
38
218
|
"id": "lora-toy-medium",
|
|
@@ -47,7 +227,7 @@
|
|
|
47
227
|
"surfaceSupport": "node",
|
|
48
228
|
"baselineReportId": "trn_lora-toy-medium_622511a55294",
|
|
49
229
|
"claimBoundary": "Practical deterministic LoRA traceability for toy adapter training, export, and eval artifacts.",
|
|
50
|
-
"configHash": "
|
|
230
|
+
"configHash": "8695a8c7b7393b0d826c1cb117be3145815e70361d243bfd7ebba17429b7842d"
|
|
51
231
|
},
|
|
52
232
|
{
|
|
53
233
|
"id": "lora-toy-tiny",
|
|
@@ -62,7 +242,7 @@
|
|
|
62
242
|
"surfaceSupport": "node",
|
|
63
243
|
"baselineReportId": "trn_lora-toy-tiny_99a3663403a3",
|
|
64
244
|
"claimBoundary": "Practical deterministic LoRA traceability for toy adapter training, export, and eval artifacts.",
|
|
65
|
-
"configHash": "
|
|
245
|
+
"configHash": "456e78833cac8bd2c376079e8336c50bffc193be18212a8052e2f8fed99f4341"
|
|
66
246
|
},
|
|
67
247
|
{
|
|
68
248
|
"id": "ul-training-medium",
|