doppler-gpu 0.5.1 → 0.6.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +33 -577
- package/LICENSE +21 -201
- package/README.md +165 -222
- package/models/catalog.json +241 -33
- package/package.json +99 -86
- package/src/capsule-runtime.d.ts +23 -0
- package/src/capsule-runtime.js +23 -0
- package/src/capsule.d.ts +8 -0
- package/src/capsule.js +4 -0
- package/src/cli/argument-contract.js +200 -0
- package/src/cli/browser-report-output.js +24 -0
- package/src/cli/cli-model-resolution.js +2 -2
- package/src/cli/cli-output.js +10 -0
- package/src/cli/commands/bundle.js +828 -0
- package/src/cli/doppler-cli.js +50 -1457
- package/src/cli/doppler-serve.js +18 -244
- package/src/cli/output/manifest.js +368 -0
- package/src/cli/runtime-profiles.js +1 -1
- package/src/cli/serve/capsule-handler.d.ts +33 -0
- package/src/cli/serve/capsule-handler.js +105 -0
- package/src/cli/serve/capsule-http.js +102 -0
- package/src/cli/serve/receipt.js +269 -0
- package/src/client/capsule-host.browser.js +1 -0
- package/src/client/capsule-host.d.ts +3 -0
- package/src/client/capsule-host.js +1 -0
- package/src/client/doppler-api.browser.js +42 -0
- package/src/client/doppler-api.d.ts +36 -1
- package/src/client/doppler-api.js +40 -1
- package/src/client/doppler-registry.js +15 -7
- package/src/client/electron/index.d.ts +4 -0
- package/src/client/electron/index.js +13 -0
- package/src/client/electron/ipc-contract.d.ts +9 -0
- package/src/client/electron/ipc-contract.js +56 -0
- package/src/client/electron/release-state.d.ts +53 -0
- package/src/client/electron/release-state.js +425 -0
- package/src/client/electron/renderer-runtime.d.ts +19 -0
- package/src/client/electron/renderer-runtime.js +104 -0
- package/src/client/failure-taxonomy.js +1 -28
- package/src/client/fault-injection.js +0 -8
- package/src/client/inspection.js +1 -1
- package/src/client/model-host/index.d.ts +140 -0
- package/src/client/model-host/index.js +548 -0
- package/src/client/model-host/model-evidence.js +255 -0
- package/src/client/model-host/model-session.d.ts +322 -0
- package/src/client/{runtime → model-host}/model-session.js +190 -148
- package/src/client/provider.d.ts +18 -2
- package/src/client/provider.js +26 -3
- package/src/client/receipt.js +30 -72
- package/src/client/runtime/capsule-acquisition.d.ts +18 -0
- package/src/client/runtime/capsule-acquisition.js +103 -0
- package/src/client/runtime/capsule-adapter-execution.d.ts +17 -0
- package/src/client/runtime/capsule-adapter-execution.js +63 -0
- package/src/client/runtime/capsule-artifact-source.js +62 -0
- package/src/client/runtime/capsule-embedding.d.ts +29 -0
- package/src/client/runtime/capsule-embedding.js +76 -0
- package/src/client/runtime/capsule-forecast-program.d.ts +2 -0
- package/src/client/runtime/capsule-forecast-program.js +85 -0
- package/src/client/runtime/capsule-forecast.d.ts +17 -0
- package/src/client/runtime/capsule-forecast.js +31 -0
- package/src/client/runtime/capsule-operation-adapters.d.ts +12 -0
- package/src/client/runtime/capsule-operation-adapters.js +68 -0
- package/src/client/runtime/capsule-operation-executor.d.ts +25 -0
- package/src/client/runtime/capsule-operation-executor.js +100 -0
- package/src/client/runtime/capsule-program-adapter.js +151 -0
- package/src/client/runtime/capsule-release-authorization.js +25 -0
- package/src/client/runtime/capsule-rerank.d.ts +50 -0
- package/src/client/runtime/capsule-rerank.js +105 -0
- package/src/client/runtime/capsule-session-execution.js +85 -0
- package/src/client/runtime/command-executor.d.ts +15 -0
- package/src/client/runtime/command-executor.js +128 -0
- package/src/client/runtime/composition-root.d.ts +74 -0
- package/src/client/runtime/composition-root.js +344 -0
- package/src/client/runtime/fetch-capsule-artifact-store.d.ts +10 -0
- package/src/client/runtime/fetch-capsule-artifact-store.js +25 -0
- package/src/client/runtime/index.d.ts +1 -117
- package/src/client/runtime/index.js +1 -433
- package/src/client/runtime/lora.d.ts +14 -0
- package/src/client/runtime/lora.js +19 -1
- package/src/client/runtime/model-manager.js +12 -0
- package/src/client/runtime/model-rerank.js +30 -0
- package/src/client/runtime/model-session.d.ts +1 -221
- package/src/client/runtime/model-source.d.ts +9 -0
- package/src/client/runtime/model-source.js +46 -17
- package/src/client/runtime/node-quickstart-cache.js +1 -1
- package/src/client/runtime/resolution-policy.d.ts +20 -0
- package/src/client/runtime/resolution-policy.js +84 -0
- package/src/client/runtime/resource-binder.d.ts +24 -0
- package/src/client/runtime/resource-binder.js +135 -0
- package/src/client/runtime/scoped-session.d.ts +10 -1
- package/src/client/runtime/scoped-session.js +20 -1
- package/src/client/runtime/session-controller.d.ts +30 -0
- package/src/client/runtime/session-controller.js +170 -0
- package/src/client/runtime/source-runtime.js +1 -1
- package/src/client/runtime/target-selector.d.ts +9 -0
- package/src/client/runtime/target-selector.js +2 -0
- package/src/client/runtime/types.d.ts +1 -0
- package/src/client/runtime/verified-capsule-artifact-store.js +72 -0
- package/src/client/wrap-pipeline-handle.js +3 -9
- package/src/config/backward-registry-loader.js +1 -1
- package/src/config/capsule-adapter-policy.js +30 -0
- package/src/config/capsule-adapters.d.ts +21 -0
- package/src/config/capsule-adapters.js +40 -0
- package/src/config/capsule-adapters.json +9 -0
- package/src/config/capsule-loading.d.ts +2 -0
- package/src/config/capsule-loading.js +14 -0
- package/src/config/capsule-loading.schema.json +11 -0
- package/src/config/capsule-operation.d.ts +19 -0
- package/src/config/capsule-operation.js +58 -0
- package/src/config/capsule-operations.json +18 -0
- package/src/config/capsule-release-contract.d.ts +66 -0
- package/src/config/capsule-release-contract.js +197 -0
- package/src/config/capsule-release-events.d.ts +53 -0
- package/src/config/capsule-release-events.js +172 -0
- package/src/config/capsule-sequence-receipt.js +19 -0
- package/src/config/capsule-serve.d.ts +9 -0
- package/src/config/capsule-serve.js +28 -0
- package/src/config/capsule-serve.schema.json +14 -0
- package/src/config/capsule-signature.d.ts +6 -0
- package/src/config/capsule-signature.js +50 -0
- package/src/config/capsule-v2.d.ts +70 -0
- package/src/config/capsule-v2.js +431 -0
- package/src/config/capsule-v3.d.ts +19 -0
- package/src/config/capsule-v3.js +72 -0
- package/src/config/capsule.d.ts +22 -0
- package/src/config/capsule.js +52 -0
- package/src/config/conversion/amplify/amplify-120m-f16-af32.json +1 -1
- package/src/config/conversion/embeddinggemma/google-embeddinggemma-300m-q4k-ehf16-af32.json +1 -1
- package/src/config/conversion/esm/esm2-t12-35m-ur50d-f32-af32.json +41 -1
- package/src/config/conversion/esmc/esmc-300m-f32-af32.json +1 -1
- package/src/config/conversion/gemma3/gemma-3-1b-it-q4k-ehf16-af32.json +1 -1
- package/src/config/conversion/gemma3/gemma-3-270m-it-f16-af32.json +1 -1
- package/src/config/conversion/gemma3/gemma-3-270m-it-q4k-ehf16-af16.json +1 -1
- package/src/config/conversion/gemma3/gemma-3-270m-it-q4k-ehf16-af32.json +1 -1
- package/src/config/conversion/gemma3/translategemma-4b-1b-enes-q4k-ehf16-af32.json +1 -1
- package/src/config/conversion/gemma3/translategemma-4b-it-q4k-ehf16-af32.json +1 -1
- package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-af16.json +2 -2
- package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-af32.json +1 -1
- package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-hq4k-af16.json +1 -1
- package/src/config/conversion/gemma4/gemma-4-12b-it-text-w4a16-ct-ehf16-af16.json +1 -1
- package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af16.json +2 -2
- package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af32.json +1 -1
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af16-int4ple.json +3 -3
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32-int4ple.json +1 -1
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32.json +1 -1
- package/src/config/conversion/glmocr/glm-ocr-f16-af32.json +442 -0
- package/src/config/conversion/minicpm/minicpm4-0-5b-f16-af32.json +1 -1
- package/src/config/conversion/muse-glimmer/muse-glimmer-30b-text-bf16-af32.json +529 -0
- package/src/config/conversion/muse-glimmer/muse-glimmer-30b-text-f16-af32.json +529 -0
- package/src/config/conversion/nucleotide-transformer/nucleotide-transformer-v2-50m-f32-af32.json +1 -1
- package/src/config/conversion/qwen3/qwen-3-5-0-8b-q4k-ehaf16.json +1 -1
- package/src/config/conversion/qwen3/qwen-3-5-2b-q4k-ehaf16.json +1 -1
- package/src/config/conversion/qwen3/qwen-3-5-9b-f16-af32.json +1 -1
- package/src/config/conversion/qwen3/qwen-3-5-9b-q4k-ehaf16.json +1 -1
- package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-eaf16.json +2 -2
- package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-ehaf16.json +1 -1
- package/src/config/conversion/qwen3/qwen-3-embedding-0-6b-f16-af32.json +463 -0
- package/src/config/conversion/qwen3/qwen-3-embedding-0-6b-q4k-ehf16-af32.json +1 -1
- package/src/config/conversion/qwen3/qwen-3-reranker-0-6b-f16-af32.json +1 -1
- package/src/config/conversion/qwen3/qwen-3-reranker-0-6b-f16-true-logit-af32.json +481 -0
- package/src/config/conversion/qwen3/qwen-3-reranker-0-6b-q4k-ehf16-af32.json +59 -7
- package/src/config/conversion/qwen3/qwen3.8-27b-text-f16-af32.json +561 -0
- package/src/config/conversion/qwen3/qwen3.8-27b-text-q4k-ehaf16.json +558 -0
- package/src/config/dtype-transition-contract.js +40 -0
- package/src/config/embedding-contract.js +26 -0
- package/src/config/embedding-reference.js +85 -0
- package/src/config/evidence/source-intake-policy.json +1 -1
- package/src/config/execution-contract-check.d.ts +16 -73
- package/src/config/execution-contract-check.js +7 -429
- package/src/config/forecast-manifest.js +56 -0
- package/src/config/heterogeneous-attention-contract.js +46 -0
- package/src/config/initial-execution-identity.d.ts +58 -0
- package/src/config/initial-execution-identity.js +331 -0
- package/src/config/kernel-path-loader.d.ts +10 -0
- package/src/config/kernel-path-loader.js +23 -8
- package/src/config/kernel-registry-contract.d.ts +25 -0
- package/src/config/kernel-registry-contract.js +46 -0
- package/src/config/kernels/kernel-ref-digests.js +50 -2
- package/src/config/kernels/registry.js +1 -1
- package/src/config/kernels/registry.json +13592 -11377
- package/src/config/lora-layouts.d.ts +10 -0
- package/src/config/lora-layouts.js +17 -0
- package/src/config/lora-layouts.json +9 -0
- package/src/config/manifest-inference-validation.js +575 -0
- package/src/config/merge/execution-patches.js +12 -0
- package/src/config/{merge-helpers.js → merge/kernel-path-policy.js} +0 -52
- package/src/config/merge/value-selection.js +38 -0
- package/src/config/merge.js +43 -11
- package/src/config/model-ir-v2.d.ts +57 -0
- package/src/config/model-ir-v2.js +258 -0
- package/src/config/model-ir.d.ts +49 -0
- package/src/config/model-ir.js +180 -0
- package/src/config/param-validator.js +1 -1
- package/src/config/platforms/amd-strix-halo-gfx1151.json +17 -0
- package/src/config/platforms/loader.js +2 -1
- package/src/config/production-release-evidence.d.ts +23 -0
- package/src/config/production-release-evidence.js +423 -0
- package/src/config/production-release.js +321 -0
- package/src/{client/doppler-registry.json → config/quickstart-registry.json} +7 -7
- package/src/config/required-inference-fields-contract-check.js +2 -2
- package/src/config/rerank-reference.d.ts +28 -0
- package/src/config/rerank-reference.js +89 -0
- package/src/config/revocation-policy.d.ts +65 -0
- package/src/config/revocation-policy.js +290 -0
- package/src/config/revocation-registry.json +11 -0
- package/src/config/revocation-updates.d.ts +68 -0
- package/src/config/revocation-updates.js +440 -0
- package/src/config/runtime/profiles/gemma4-31b-f16-activations-probe.json +1 -1
- package/src/config/runtime/profiles/muse-glimmer-30b-source-parity-f32kv.json +30 -0
- package/src/config/runtime/profiles/muse-glimmer-30b-source-parity.json +110 -0
- package/src/config/runtime/profiles/qwen3-6-27b-f16-activations-probe.json +1 -1
- package/src/config/runtime/profiles/qwen3-8-27b-source-parity.json +87 -0
- package/src/config/runtime-merge.js +1 -1
- package/src/config/runtime.js +1 -1
- package/src/config/schema/browser-suite-metrics.schema.d.ts +1 -0
- package/src/config/schema/browser-suite-metrics.schema.js +2 -0
- package/src/config/schema/conversion-report.schema.d.ts +3 -0
- package/src/config/schema/conversion-report.schema.js +29 -0
- package/src/config/schema/converter.schema.d.ts +1 -0
- package/src/config/schema/debug.schema.d.ts +1 -0
- package/src/config/schema/doppler.schema.js +5 -5
- package/src/config/schema/emulation.schema.js +1 -1
- package/src/config/schema/execution-v1.schema.d.ts +3 -0
- package/src/config/schema/execution-v1.schema.js +19 -0
- package/src/config/schema/harness.schema.d.ts +13 -0
- package/src/config/schema/harness.schema.js +6 -0
- package/src/config/schema/inference.schema.d.ts +21 -2
- package/src/config/schema/kernel-thresholds.schema.js +1 -1
- package/src/config/schema/manifest.schema.d.ts +26 -7
- package/src/config/schema/manifest.schema.js +10 -1
- package/src/config/schema/program-bundle.schema.d.ts +52 -6
- package/src/config/schema/program-bundle.schema.js +145 -10
- package/src/config/sequence-reference.d.ts +26 -0
- package/src/config/sequence-reference.js +100 -0
- package/src/config/target-plan.d.ts +108 -0
- package/src/config/target-plan.js +411 -0
- package/src/config/training-defaults.js +1 -1
- package/src/config/transforms/capability-transform-resolver.js +4 -105
- package/src/config/transforms/execution-graph/dtype.js +591 -0
- package/src/config/transforms/execution-graph/fusion.js +436 -0
- package/src/config/transforms/execution-graph/validation.js +414 -0
- package/src/config/transforms/execution-graph-transforms.js +13 -1419
- package/src/converter/artifact-identity.js +475 -0
- package/src/converter/conversion-plan.js +1 -1
- package/src/converter/core.js +94 -2096
- package/src/converter/forge-candidate-evaluation.js +271 -0
- package/src/converter/forge-qualification.js +162 -0
- package/src/converter/forge-stages.js +798 -0
- package/src/converter/manifest-builder.js +683 -0
- package/src/converter/quantization-info.js +4 -1
- package/src/converter/quantizer.js +58 -9
- package/src/converter/shard-packer.d.ts +1 -1
- package/src/converter/shard-packer.js +2 -2
- package/src/converter/tensor-transform.js +640 -0
- package/src/converter/tokenizer-contract.js +243 -0
- package/src/converter/tokenizer-manifest.js +243 -0
- package/src/debug/history.js +0 -8
- package/src/debug/index.js +0 -10
- package/src/debug/tensor.js +1 -1
- package/src/debug/trace.js +0 -7
- package/src/experimental/adapters/adapter-manager.js +2 -0
- package/src/experimental/adapters/adapter-manifest.d.ts +6 -0
- package/src/experimental/adapters/adapter-manifest.js +11 -0
- package/src/experimental/adapters/artifact-contract.d.ts +2 -2
- package/src/experimental/adapters/artifact-contract.js +5 -2
- package/src/experimental/adapters/lora-loader.d.ts +10 -1
- package/src/experimental/adapters/lora-loader.js +110 -48
- package/src/{inference/pipelines → experimental}/audio/gemma4.js +9 -31
- package/src/experimental/bdpa/basis-decomposed-paged.d.ts +50 -0
- package/src/experimental/bdpa/basis-decomposed-paged.js +276 -0
- package/src/experimental/bridge/trainer-artifact-bridge.js +3 -3
- package/src/experimental/browser/browser-converter.js +11 -11
- package/src/experimental/diffusion/index.d.ts +1 -1
- package/src/experimental/diffusion/index.js +1 -1
- package/src/{inference/pipelines → experimental}/diffusion/pipeline.d.ts +2 -2
- package/src/{inference/pipelines → experimental}/diffusion/pipeline.js +17 -17
- package/src/{inference/pipelines → experimental}/diffusion/text-encoder-gpu.js +21 -19
- package/src/experimental/diffusion/vae.js +596 -0
- package/src/experimental/distribution/shard-delivery/receipt.js +588 -0
- package/src/experimental/distribution/shard-delivery/retry.js +392 -0
- package/src/experimental/distribution/shard-delivery/transport.js +831 -0
- package/src/experimental/distribution/shard-delivery.js +4 -1709
- package/src/experimental/energy/index.d.ts +2 -2
- package/src/experimental/energy/index.js +2 -2
- package/src/{inference/pipelines → experimental}/energy/pipeline.js +13 -13
- package/src/{inference/pipelines → experimental}/energy/types.d.ts +2 -2
- package/src/experimental/logits/cpu-output.d.ts +27 -0
- package/src/{inference/pipelines/text/logits/utils.js → experimental/logits/cpu-output.js} +19 -27
- package/src/experimental/logits/cpu.d.ts +87 -0
- package/src/experimental/logits/cpu.js +95 -0
- package/src/experimental/orchestration/energy-row-head-pipeline.d.ts +103 -0
- package/src/{inference/pipelines/energy-head/row-head-pipeline.js → experimental/orchestration/energy-row-head-pipeline.js} +13 -13
- package/src/experimental/orchestration/index.d.ts +1 -1
- package/src/experimental/orchestration/index.js +1 -1
- package/src/experimental/training/artifacts.js +2 -2
- package/src/experimental/training/checkpoint.js +2 -2
- package/src/experimental/training/distillation/artifacts.js +1 -1
- package/src/experimental/training/distillation/dataset.js +2 -2
- package/src/experimental/training/distillation/eval.js +1 -1
- package/src/experimental/training/distillation/runtime.js +2 -2
- package/src/experimental/training/distillation/student-fixture/execution.js +166 -0
- package/src/experimental/training/distillation/student-fixture/model.js +353 -0
- package/src/experimental/training/distillation/student-fixture.js +2 -469
- package/src/experimental/training/lora/checkpoint.js +746 -0
- package/src/experimental/training/lora/execution.js +550 -0
- package/src/experimental/training/lora/plan.js +387 -0
- package/src/experimental/training/lora/recovery.js +418 -0
- package/src/experimental/training/lora-pipeline.js +7 -1949
- package/src/experimental/training/operator-artifacts.js +2 -2
- package/src/experimental/training/operator-command.js +4 -4
- package/src/experimental/training/runner/evaluation.js +62 -0
- package/src/experimental/training/runner/lifecycle.js +257 -0
- package/src/experimental/training/runner.js +5 -268
- package/src/experimental/training/suite/execution.js +879 -0
- package/src/experimental/training/suite/plan.js +855 -0
- package/src/experimental/training/suite.js +5 -1659
- package/src/experimental/training/workloads.d.ts +5 -5
- package/src/experimental/training/workloads.js +6 -6
- package/src/experimental/turboquant/codebook.js +470 -0
- package/src/formats/quantization-constants.d.ts +11 -0
- package/src/formats/quantization-constants.js +11 -0
- package/src/formats/rdrr/integrity.js +1 -1
- package/src/formats/rdrr/lowerings.js +0 -29
- package/src/formats/rdrr/merkle.js +1 -1
- package/src/formats/rdrr/tensor-config-validator.js +25 -5
- package/src/formats/rdrr/validation.js +3 -3
- package/src/gpu/command-recorder.d.ts +10 -1
- package/src/gpu/command-recorder.js +58 -14
- package/src/gpu/constant-buffer.js +29 -0
- package/src/gpu/device-state.js +36 -0
- package/src/gpu/device.js +6 -41
- package/src/gpu/kernel-runtime.js +2 -1
- package/src/gpu/kernels/activation-static-qdq.js +1 -1
- package/src/gpu/kernels/append_embedding.wgsl +14 -0
- package/src/gpu/kernels/attention/executors.js +960 -0
- package/src/gpu/kernels/attention/plan.js +472 -0
- package/src/gpu/kernels/attention.d.ts +4 -0
- package/src/gpu/kernels/attention.js +145 -1545
- package/src/gpu/kernels/attention_relative.wgsl +59 -0
- package/src/gpu/kernels/backward/adam.js +2 -1
- package/src/gpu/kernels/backward/attention_backward.js +1 -1
- package/src/gpu/kernels/backward/bias_add_backward.d.ts +1 -1
- package/src/gpu/kernels/backward/bias_add_backward.js +1 -1
- package/src/gpu/kernels/backward/causal_conv1d_silu_backward.js +2 -1
- package/src/gpu/kernels/backward/conv2d_backward.d.ts +1 -1
- package/src/gpu/kernels/backward/conv2d_backward.js +2 -1
- package/src/gpu/kernels/backward/cross_entropy_backward.d.ts +1 -1
- package/src/gpu/kernels/backward/cross_entropy_backward.js +1 -1
- package/src/gpu/kernels/backward/embed_backward.js +2 -1
- package/src/gpu/kernels/backward/executor.d.ts +34 -0
- package/src/gpu/kernels/backward/executor.js +158 -0
- package/src/gpu/kernels/backward/gated_delta_recurrent_backward.js +2 -1
- package/src/gpu/kernels/backward/gated_rmsnorm_backward.js +2 -1
- package/src/gpu/kernels/backward/gelu_backward.d.ts +1 -1
- package/src/gpu/kernels/backward/gelu_backward.js +1 -1
- package/src/gpu/kernels/backward/groupnorm_backward.d.ts +1 -1
- package/src/gpu/kernels/backward/groupnorm_backward.js +1 -1
- package/src/gpu/kernels/backward/layernorm_backward.d.ts +1 -1
- package/src/gpu/kernels/backward/layernorm_backward.js +2 -1
- package/src/gpu/kernels/backward/matmul.js +133 -0
- package/src/gpu/kernels/backward/matmul_backward.d.ts +0 -1
- package/src/gpu/kernels/backward/matmul_backward.js +1 -1
- package/src/gpu/kernels/backward/pixel_shuffle_backward.d.ts +1 -1
- package/src/gpu/kernels/backward/pixel_shuffle_backward.js +1 -1
- package/src/gpu/kernels/backward/rmsnorm_backward.d.ts +1 -1
- package/src/gpu/kernels/backward/rmsnorm_backward.js +2 -1
- package/src/gpu/kernels/backward/rope_backward.d.ts +1 -1
- package/src/gpu/kernels/backward/rope_backward.js +2 -1
- package/src/gpu/kernels/backward/scale_backward.d.ts +1 -1
- package/src/gpu/kernels/backward/scale_backward.js +2 -1
- package/src/gpu/kernels/backward/silu_backward.d.ts +1 -1
- package/src/gpu/kernels/backward/silu_backward.js +1 -1
- package/src/gpu/kernels/backward/softmax_backward.d.ts +1 -1
- package/src/gpu/kernels/backward/softmax_backward.js +1 -1
- package/src/gpu/kernels/backward/upsample2d_backward.d.ts +1 -1
- package/src/gpu/kernels/backward/upsample2d_backward.js +1 -1
- package/src/gpu/kernels/backward/utils.js +2 -392
- package/src/gpu/kernels/bind-group-validation.js +16 -0
- package/src/gpu/kernels/cast.js +2 -1
- package/src/gpu/kernels/check-finiteness.js +13 -49
- package/src/gpu/kernels/check-hot-vocab-stop.js +13 -39
- package/src/gpu/kernels/check-stop.js +14 -33
- package/src/gpu/kernels/check_finiteness.wgsl +33 -0
- package/src/gpu/kernels/check_hot_vocab_stop.wgsl +29 -0
- package/src/gpu/kernels/check_stop.wgsl +20 -0
- package/src/gpu/kernels/clamp.js +1 -1
- package/src/gpu/kernels/conv2d.js +1 -1
- package/src/gpu/kernels/cross_entropy_loss.js +1 -1
- package/src/gpu/kernels/depthwise_conv1d.js +1 -1
- package/src/gpu/kernels/depthwise_conv2d.js +1 -1
- package/src/gpu/kernels/dequant.js +7 -2
- package/src/gpu/kernels/diffusion-gemma-sampling.js +3 -5
- package/src/gpu/kernels/dispatch.js +12 -1
- package/src/gpu/kernels/embedding-pool.js +72 -0
- package/src/gpu/kernels/embedding_pool_f16.wgsl +38 -0
- package/src/gpu/kernels/embedding_pool_f32.wgsl +36 -0
- package/src/gpu/kernels/energy.js +2 -1
- package/src/gpu/kernels/forecast_output.wgsl +17 -0
- package/src/gpu/kernels/fused-gate-up-gelu.js +1 -1
- package/src/gpu/kernels/fused_ffn.js +10 -220
- package/src/gpu/kernels/fused_matmul_residual.js +2 -1
- package/src/gpu/kernels/fused_matmul_rmsnorm.js +2 -1
- package/src/gpu/kernels/fused_normed_ffn_q4k.wgsl +208 -0
- package/src/gpu/kernels/gated-short-conv.js +11 -69
- package/src/gpu/kernels/gated_delta_conv.wgsl +66 -0
- package/src/gpu/kernels/gated_delta_conv_f16.wgsl +68 -0
- package/src/gpu/kernels/gated_delta_fused_decode.wgsl +173 -0
- package/src/gpu/kernels/gated_delta_fused_decode_f16.wgsl +175 -0
- package/src/gpu/kernels/gated_delta_recurrent.wgsl +165 -0
- package/src/gpu/kernels/gated_delta_recurrent_f16.wgsl +167 -0
- package/src/gpu/kernels/gated_short_conv.wgsl +47 -0
- package/src/gpu/kernels/gather.d.ts +1 -1
- package/src/gpu/kernels/gather.js +30 -6
- package/src/gpu/kernels/gather_bf16.wgsl +69 -0
- package/src/gpu/kernels/gelu.js +1 -1
- package/src/gpu/kernels/gemma4-route-expert.js +1 -1
- package/src/gpu/kernels/grouped_pointwise_conv2d.js +1 -1
- package/src/gpu/kernels/groupnorm.js +1 -1
- package/src/gpu/kernels/index.d.ts +38 -5
- package/src/gpu/kernels/index.js +18 -4
- package/src/gpu/kernels/instance_stats.wgsl +26 -0
- package/src/gpu/kernels/kernel-base.js +1 -1
- package/src/gpu/kernels/kernel-cache-lifecycle.d.ts +9 -0
- package/src/gpu/kernels/kernel-cache-lifecycle.js +26 -0
- package/src/gpu/kernels/kernel-configs.d.ts +10 -61
- package/src/gpu/kernels/kernel-configs.js +26 -54
- package/src/gpu/kernels/kernel-execution.js +141 -0
- package/src/gpu/kernels/kv-cache-write.js +4 -4
- package/src/gpu/kernels/kv-quantize.js +4 -1
- package/src/gpu/kernels/l2-normalize.js +36 -0
- package/src/gpu/kernels/l2_normalize.wgsl +53 -0
- package/src/gpu/kernels/layernorm.d.ts +4 -4
- package/src/gpu/kernels/layernorm.js +44 -7
- package/src/gpu/kernels/linear-attention-core.js +29 -449
- package/src/gpu/kernels/linear_attention.js +1 -1
- package/src/gpu/kernels/lm-head-argmax.js +2 -2
- package/src/gpu/kernels/lm-head-select-logits.js +2 -2
- package/src/gpu/kernels/logit-finalize.js +103 -0
- package/src/gpu/kernels/logit-merge.js +13 -125
- package/src/gpu/kernels/logit-scatter.js +40 -0
- package/src/gpu/kernels/logit_finalize_f16.wgsl +43 -0
- package/src/gpu/kernels/logit_finalize_f32.wgsl +41 -0
- package/src/gpu/kernels/logit_merge_geometric.wgsl +23 -0
- package/src/gpu/kernels/logit_merge_max.wgsl +23 -0
- package/src/gpu/kernels/logit_merge_weighted.wgsl +23 -0
- package/src/gpu/kernels/logit_scatter_f16.wgsl +26 -0
- package/src/gpu/kernels/logit_scatter_f32.wgsl +24 -0
- package/src/gpu/kernels/matmul-dispatch.js +6 -1
- package/src/gpu/kernels/matmul-selection.js +3 -2
- package/src/gpu/kernels/matmul.js +4 -15
- package/src/gpu/kernels/matmul_bf16w_f32a.wgsl +93 -0
- package/src/gpu/kernels/matmul_gemv_subgroup_bf16w.wgsl +433 -0
- package/src/gpu/kernels/modulate.js +1 -1
- package/src/gpu/kernels/moe.js +3 -1
- package/src/gpu/kernels/moe_router_bias_add_f16_f16.wgsl +23 -0
- package/src/gpu/kernels/moe_router_bias_add_f16_f32.wgsl +23 -0
- package/src/gpu/kernels/moe_router_bias_add_f32_f16.wgsl +23 -0
- package/src/gpu/kernels/moe_router_bias_add_f32_f32.wgsl +21 -0
- package/src/gpu/kernels/patch_context.wgsl +29 -0
- package/src/gpu/kernels/pipeline-cache.js +3 -1
- package/src/gpu/kernels/pixel_shuffle.js +1 -1
- package/src/gpu/kernels/probe_shader_f16.wgsl +7 -0
- package/src/gpu/kernels/relu.js +1 -1
- package/src/gpu/kernels/repeat_channels.js +1 -1
- package/src/gpu/kernels/residual.js +2 -2
- package/src/gpu/kernels/rmsnorm-pair.js +1 -1
- package/src/gpu/kernels/rmsnorm-qk.js +1 -1
- package/src/gpu/kernels/rmsnorm-stats.js +14 -113
- package/src/gpu/kernels/rmsnorm.js +1 -1
- package/src/gpu/kernels/rmsnorm_serial.wgsl +19 -0
- package/src/gpu/kernels/rmsnorm_stats.wgsl +59 -0
- package/src/gpu/kernels/rmsnorm_stats_subgroups.wgsl +76 -0
- package/src/gpu/kernels/rope-precompute.js +160 -0
- package/src/gpu/kernels/rope-qk.js +1 -1
- package/src/gpu/kernels/rope.js +2 -2
- package/src/gpu/kernels/rope_precompute.wgsl +74 -0
- package/src/gpu/kernels/sample.js +5 -1
- package/src/gpu/kernels/scale.js +1 -1
- package/src/gpu/kernels/shader-cache.js +8 -4
- package/src/gpu/kernels/shader-source-scope.js +68 -0
- package/src/gpu/kernels/silu.js +2 -1
- package/src/gpu/kernels/soft-embedding.js +3 -1
- package/src/gpu/kernels/softmax.js +4 -2
- package/src/gpu/kernels/split-qkv-rmsnorm-qk.js +1 -1
- package/src/gpu/kernels/split-qkv-rmsnorm-rope-qk.js +1 -1
- package/src/gpu/kernels/split_qg.js +1 -1
- package/src/gpu/kernels/split_qkv.d.ts +2 -0
- package/src/gpu/kernels/split_qkv.js +18 -3
- package/src/gpu/kernels/split_qkv.wgsl +5 -2
- package/src/gpu/kernels/split_qkv_f16.wgsl +5 -2
- package/src/gpu/kernels/submit_probe.wgsl +6 -0
- package/src/gpu/kernels/transpose.js +1 -1
- package/src/gpu/kernels/turboquant-codebook.js +4 -535
- package/src/gpu/kernels/uniform-utils.js +7 -2
- package/src/gpu/kernels/upsample2d.js +1 -1
- package/src/gpu/kernels/utils.js +9 -330
- package/src/gpu/kernels/vision-average-pool.d.ts +13 -0
- package/src/gpu/kernels/vision-average-pool.js +68 -0
- package/src/gpu/kernels/vision-patch-embed.d.ts +17 -0
- package/src/gpu/kernels/vision-patch-embed.js +117 -0
- package/src/gpu/kernels/vision-position-embedding.d.ts +13 -0
- package/src/gpu/kernels/vision-position-embedding.js +79 -0
- package/src/gpu/kernels/vision-rope-2d.d.ts +16 -0
- package/src/gpu/kernels/vision-rope-2d.js +73 -0
- package/src/gpu/kernels/vision-spatial-merge.d.ts +19 -0
- package/src/gpu/kernels/vision-spatial-merge.js +100 -0
- package/src/gpu/kernels/vision_average_pool.wgsl +42 -0
- package/src/gpu/kernels/vision_patch_embed.wgsl +59 -0
- package/src/gpu/kernels/vision_patch_embed_f16w.wgsl +61 -0
- package/src/gpu/kernels/vision_position_embedding.wgsl +33 -0
- package/src/gpu/kernels/vision_position_embedding_f16.wgsl +35 -0
- package/src/gpu/kernels/vision_rope_2d.wgsl +60 -0
- package/src/gpu/kernels/vision_spatial_merge.wgsl +48 -0
- package/src/gpu/readback-buffer.js +20 -0
- package/src/gpu/submit-probe.js +2 -26
- package/src/gpu/weight-buffer.d.ts +1 -0
- package/src/gpu/weight-buffer.js +7 -0
- package/src/index.d.ts +10 -0
- package/src/index.js +1 -0
- package/src/inference/browser-harness/benchmark-execution.js +872 -0
- package/src/inference/browser-harness/benchmark-rerank.js +284 -0
- package/src/inference/browser-harness/decode-diagnostics.js +241 -0
- package/src/inference/browser-harness/diffusion-performance.js +102 -0
- package/src/inference/browser-harness/inference-execution.js +543 -0
- package/src/inference/{browser-harness-model-helpers.js → browser-harness/model-initialization.js} +14 -145
- package/src/inference/browser-harness/model-resolution.js +137 -0
- package/src/inference/browser-harness/report.js +338 -0
- package/src/inference/browser-harness/request.js +263 -0
- package/src/inference/browser-harness/runtime-config.d.ts +25 -0
- package/src/inference/browser-harness/runtime-config.js +255 -0
- package/src/inference/browser-harness/runtime-isolation.d.ts +13 -0
- package/src/inference/browser-harness/runtime-isolation.js +45 -0
- package/src/inference/browser-harness/runtime-report.d.ts +26 -0
- package/src/inference/browser-harness/runtime-report.js +117 -0
- package/src/inference/browser-harness/suite-summary.js +39 -0
- package/src/inference/browser-harness/text-evidence.js +694 -0
- package/src/inference/browser-harness/text-execution.js +757 -0
- package/src/inference/browser-harness/text-input.js +725 -0
- package/src/inference/{browser-harness-suite-helpers.js → browser-harness/timing-diagnostics.js} +1 -379
- package/src/inference/browser-harness-contract-helpers.js +1 -1
- package/src/inference/browser-harness-diffusion-energy-suites.js +8 -3
- package/src/inference/browser-harness-runtime-helpers.d.ts +3 -56
- package/src/inference/browser-harness-runtime-helpers.js +3 -423
- package/src/inference/browser-harness-text-helpers.js +31 -2199
- package/src/inference/browser-harness.d.ts +15 -2
- package/src/inference/browser-harness.js +38 -1952
- package/src/inference/kv-cache/base.js +67 -381
- package/src/inference/kv-cache/basis-decomposed-paged.d.ts +2 -50
- package/src/inference/kv-cache/basis-decomposed-paged.js +1 -276
- package/src/inference/kv-cache/quantized.d.ts +2 -2
- package/src/inference/kv-cache/quantized.js +4 -8
- package/src/inference/kv-cache/update.js +309 -0
- package/src/inference/model-checkpoint-evidence.d.ts +75 -0
- package/src/inference/model-checkpoint-evidence.js +264 -0
- package/src/inference/moe-router.js +36 -216
- package/src/inference/multi-model-network.js +5 -341
- package/src/inference/pipelines/audio/index.js +1 -20
- package/src/inference/pipelines/audio/mel.js +0 -38
- package/src/inference/pipelines/diffusion/profile-timing.js +10 -0
- package/src/inference/pipelines/diffusion/runtime-resources.js +54 -0
- package/src/inference/pipelines/diffusion/sd3/execution.js +805 -0
- package/src/inference/pipelines/diffusion/sd3/plan.js +125 -0
- package/src/inference/pipelines/diffusion/sd3-transformer.js +44 -895
- package/src/inference/pipelines/diffusion/{vae.js → vae/execution.js} +46 -618
- package/src/inference/pipelines/diffusion/{helpers.js → weight-contract.js} +1 -68
- package/src/inference/pipelines/energy-head/row-head-pipeline.d.ts +1 -103
- package/src/inference/pipelines/factory.js +14 -3
- package/src/inference/pipelines/shader-scoped-pipeline.js +29 -0
- package/src/inference/pipelines/structured/json-grammar-mask.js +0 -6
- package/src/inference/pipelines/text/abort-contract.js +22 -0
- package/src/inference/pipelines/text/attention/attn-config.js +1 -16
- package/src/inference/pipelines/text/attention/gate-projection.js +30 -0
- package/src/inference/pipelines/text/attention/heterogeneous-contract.js +8 -0
- package/src/inference/pipelines/text/attention/interpreter/immediate.js +211 -0
- package/src/inference/pipelines/text/attention/interpreter/recorded.js +451 -0
- package/src/inference/pipelines/text/attention/interpreter.js +89 -439
- package/src/inference/pipelines/text/attention/plan.d.ts +1 -0
- package/src/inference/pipelines/text/attention/plan.js +17 -0
- package/src/inference/pipelines/text/attention/projection-executor.js +331 -0
- package/src/inference/pipelines/text/attention/projection-plan.js +259 -0
- package/src/inference/pipelines/text/attention/projections.js +20 -537
- package/src/inference/pipelines/text/attention/query-transform.js +11 -0
- package/src/inference/pipelines/text/attention/record.js +0 -6
- package/src/inference/pipelines/text/attention/recorded-health.js +21 -0
- package/src/inference/pipelines/text/attention/rope-observation.js +40 -0
- package/src/inference/pipelines/text/attention/run.js +0 -6
- package/src/inference/pipelines/text/attention/types.js +3 -31
- package/src/inference/pipelines/text/chat-format.d.ts +3 -0
- package/src/inference/pipelines/text/chat-format.js +42 -0
- package/src/inference/pipelines/text/config/normalization.js +729 -0
- package/src/inference/pipelines/text/config/validation.js +9 -0
- package/src/inference/pipelines/text/config.d.ts +12 -1
- package/src/inference/pipelines/text/config.js +14 -1228
- package/src/inference/pipelines/text/debug-utils/health-metrics.js +57 -0
- package/src/inference/pipelines/text/debug-utils/index.js +3 -8
- package/src/inference/pipelines/text/debug-utils/profiles.js +8 -0
- package/src/inference/pipelines/text/debug-utils/readback-decoding.js +24 -0
- package/src/inference/pipelines/text/debug-utils/tensor.js +1 -1
- package/src/inference/pipelines/text/dtype-contract.js +1 -68
- package/src/inference/pipelines/text/embed.js +178 -241
- package/src/inference/pipelines/text/embedding-contract.d.ts +7 -0
- package/src/inference/pipelines/text/embedding-contract.js +21 -0
- package/src/inference/pipelines/text/embedding-extraction.js +225 -0
- package/src/inference/pipelines/text/embedding-normalization.js +39 -0
- package/src/inference/pipelines/text/execution-runtime-builders.js +0 -32
- package/src/inference/pipelines/text/execution-v1/plan.js +69 -0
- package/src/inference/pipelines/text/execution-v1/validation.js +233 -0
- package/src/inference/pipelines/text/execution-v1.js +7 -283
- package/src/inference/pipelines/text/execution.js +517 -0
- package/src/inference/pipelines/text/ffn/dense-executor.js +718 -0
- package/src/inference/pipelines/text/ffn/dense-plan.js +327 -0
- package/src/inference/pipelines/text/ffn/dense.js +4 -968
- package/src/inference/pipelines/text/ffn/sandwich.js +10 -9
- package/src/inference/pipelines/text/ffn/standard.js +2 -1
- package/src/inference/pipelines/text/generator/attention-lifecycle.js +24 -0
- package/src/inference/pipelines/text/generator/decode-runtime.js +991 -0
- package/src/inference/pipelines/text/generator/decode.js +952 -0
- package/src/inference/pipelines/text/generator/diffusion.js +607 -0
- package/src/inference/pipelines/text/generator/logits-config.js +63 -0
- package/src/inference/pipelines/text/generator/prefill-observation.js +29 -0
- package/src/inference/pipelines/text/generator/prefill-policy.js +56 -0
- package/src/inference/pipelines/text/generator/prefill-runtime.js +896 -0
- package/src/inference/pipelines/text/{generator-prefill-helpers.js → generator/prefix-embedding.js} +16 -121
- package/src/inference/pipelines/text/generator/prompt-input.js +47 -0
- package/src/inference/pipelines/text/generator/recovery.js +310 -0
- package/src/inference/pipelines/text/{generator-helpers.js → generator/session-context.js} +42 -110
- package/src/inference/pipelines/text/generator/text.js +448 -0
- package/src/inference/pipelines/text/generator/training.js +752 -0
- package/src/inference/pipelines/text/generator-decode-policy.js +1 -1
- package/src/inference/pipelines/text/generator-logits-observation.js +45 -0
- package/src/inference/pipelines/text/generator-runtime.js +5 -135
- package/src/inference/pipelines/text/generator-steps.js +30 -972
- package/src/inference/pipelines/text/generator.js +165 -3544
- package/src/inference/pipelines/text/image-transcription.js +304 -0
- package/src/inference/pipelines/text/init-qkv-fusion.js +18 -9
- package/src/inference/pipelines/text/init-rope.js +37 -189
- package/src/inference/pipelines/text/init.d.ts +8 -3
- package/src/inference/pipelines/text/init.js +6 -216
- package/src/inference/pipelines/text/layer-execution-contract.js +97 -0
- package/src/inference/pipelines/text/layer-input-execution.js +435 -0
- package/src/inference/pipelines/text/layer-plan-gpu.js +13 -6
- package/src/inference/pipelines/text/layer-plan.d.ts +21 -5
- package/src/inference/pipelines/text/layer-plan.js +45 -3
- package/src/inference/pipelines/text/layer.js +52 -522
- package/src/inference/pipelines/text/lifecycle.js +839 -0
- package/src/inference/pipelines/text/linear-attention/plan.js +554 -0
- package/src/inference/pipelines/text/linear-attention-ab-fusion.js +2 -2
- package/src/inference/pipelines/text/linear-attention.d.ts +2 -2
- package/src/inference/pipelines/text/linear-attention.js +7 -537
- package/src/inference/pipelines/text/logits/cpu-output.d.ts +1 -0
- package/src/inference/pipelines/text/logits/cpu-output.js +5 -0
- package/src/inference/pipelines/text/logits/cpu.d.ts +1 -87
- package/src/inference/pipelines/text/logits/cpu.js +8 -95
- package/src/inference/pipelines/text/logits/gpu-executor.js +368 -0
- package/src/inference/pipelines/text/logits/gpu.d.ts +27 -0
- package/src/inference/pipelines/text/logits/gpu.js +116 -579
- package/src/inference/pipelines/text/logits/index.d.ts +2 -1
- package/src/inference/pipelines/text/logits/index.js +62 -102
- package/src/inference/pipelines/text/logits/output-transform.js +47 -0
- package/src/inference/pipelines/text/logits/plan.js +184 -0
- package/src/inference/pipelines/text/logits/readback.d.ts +6 -0
- package/src/inference/pipelines/text/logits/readback.js +9 -0
- package/src/inference/pipelines/text/logits/scale-policy.js +20 -0
- package/src/inference/pipelines/text/logits/types.d.ts +1 -0
- package/src/inference/pipelines/text/lora-apply.js +12 -4
- package/src/inference/pipelines/text/lora-types.d.ts +12 -0
- package/src/inference/pipelines/text/modality-token-contract.js +115 -0
- package/src/inference/pipelines/text/model-load.js +0 -20
- package/src/inference/pipelines/text/moe/expert-loading.js +9 -0
- package/src/inference/pipelines/text/moe/gpu-executor.js +513 -0
- package/src/inference/pipelines/text/moe/plan.js +47 -0
- package/src/inference/pipelines/text/moe-cpu.js +9 -135
- package/src/inference/pipelines/text/moe-gpu.js +21 -545
- package/src/inference/pipelines/text/normalization-contract.js +17 -0
- package/src/inference/pipelines/text/per-layer/materialize.js +736 -0
- package/src/inference/pipelines/text/per-layer/plan.js +362 -0
- package/src/inference/pipelines/text/per-layer-inputs.js +4 -1037
- package/src/inference/pipelines/text/pipeline-load-timing.js +77 -0
- package/src/inference/pipelines/text/probes.js +20 -0
- package/src/inference/pipelines/text/state.d.ts +8 -0
- package/src/inference/pipelines/text/state.js +1 -0
- package/src/inference/pipelines/text/tsir-fixture-writer.js +88 -16
- package/src/inference/pipelines/text/types.d.ts +4 -0
- package/src/inference/pipelines/text/weight-loading-config.js +194 -0
- package/src/inference/pipelines/text.d.ts +2 -2
- package/src/inference/pipelines/text.js +46 -1633
- package/src/inference/pipelines/video/frame-sampler.js +0 -13
- package/src/inference/pipelines/video/index.js +0 -21
- package/src/inference/pipelines/vision/attention.js +59 -0
- package/src/inference/pipelines/vision/encoder.js +254 -369
- package/src/inference/pipelines/vision/gemma4.js +204 -318
- package/src/inference/pipelines/vision/glmocr-preprocess.js +273 -0
- package/src/inference/pipelines/vision/glmocr-rope.js +106 -0
- package/src/inference/pipelines/vision/glmocr.js +557 -0
- package/src/inference/pipelines/vision/image-preprocess.js +0 -27
- package/src/inference/pipelines/vision/index.js +8 -27
- package/src/inference/pipelines/vision/ops.js +11 -33
- package/src/inference/pipelines/vision/patch-embed.js +51 -145
- package/src/inference/pipelines/vision/spatial-merge.js +78 -0
- package/src/inference/pipelines/vision/weight-loading.js +79 -0
- package/src/inference/refactor-receipt.js +2 -2
- package/src/inference/rerank.js +182 -0
- package/src/inference/tokenizer.js +1 -1
- package/src/inference/tokenizers/behavior-flags.js +6 -3
- package/src/inference/tokenizers/bundled/decoders.js +133 -0
- package/src/inference/tokenizers/bundled/registry.js +216 -0
- package/src/inference/tokenizers/bundled.js +23 -353
- package/src/loader/doppler-loader.d.ts +4 -0
- package/src/loader/doppler-loader.js +119 -930
- package/src/loader/embedding-limit-preflight.js +11 -4
- package/src/loader/embedding-loader.js +1 -3
- package/src/loader/gpu-tensor-materialization.js +20 -0
- package/src/loader/layer-loader.js +15 -119
- package/src/loader/layer-tensor-aliases.js +99 -0
- package/src/loader/layer-weight-shape.js +29 -0
- package/src/loader/load-timing.js +64 -0
- package/src/loader/loader-types.d.ts +3 -0
- package/src/loader/model-executor.js +819 -0
- package/src/loader/model-load-validation.js +15 -0
- package/src/loader/quantization-constants.d.ts +11 -29
- package/src/loader/quantization-constants.js +1 -2
- package/src/loader/shard-resolver.js +0 -14
- package/src/loader/tensor-load-error.js +14 -0
- package/src/loader/tensor-shard-indices.js +8 -0
- package/src/loader/tensors/bf16-loader.js +125 -0
- package/src/loader/tensors/functional-descriptor-loader.js +3 -375
- package/src/loader/tensors/load-executor.js +577 -0
- package/src/loader/tensors/source-transform/executor.js +123 -0
- package/src/loader/tensors/source-transform/plan.js +107 -0
- package/src/loader/tensors/source-transform.js +15 -264
- package/src/loader/tensors/tensor-loader.d.ts +4 -4
- package/src/loader/tensors/tensor-loader.js +7 -653
- package/src/memory/buffer-pool.js +7 -29
- package/src/rules/capability-transform-resolver.js +80 -0
- package/src/rules/execution-contract.d.ts +73 -0
- package/src/rules/execution-contract.js +429 -0
- package/src/rules/execution-rules-contract-check.js +1 -1
- package/src/rules/generated/rule-bundle.json +4305 -0
- package/src/rules/inference/capability-transforms.rules.json +3 -3
- package/src/rules/layer-pattern-contract-check.js +1 -1
- package/src/rules/rule-registry.js +70 -57
- package/src/storage/artifact-storage-context.js +2 -2
- package/src/storage/backend-selection.js +43 -0
- package/src/storage/distribution-transport.js +5 -0
- package/src/storage/download/integrity.d.ts +33 -0
- package/src/storage/download/integrity.js +130 -0
- package/src/storage/download/resume-inspection.d.ts +22 -0
- package/src/storage/download/resume-inspection.js +134 -0
- package/src/storage/download/retry.js +57 -0
- package/src/storage/download/state.js +131 -0
- package/src/storage/download/transport.js +99 -0
- package/src/storage/downloader.js +74 -453
- package/src/storage/model-cache.d.ts +1 -0
- package/src/storage/model-cache.js +51 -11
- package/src/storage/shard-manager.d.ts +5 -0
- package/src/storage/shard-manager.js +50 -369
- package/src/storage/shards/index.js +60 -0
- package/src/storage/shards/integrity.js +225 -0
- package/src/storage/shards/lifecycle.js +22 -0
- package/src/storage/source-artifact-store.js +14 -4
- package/src/tooling/boundary-evidence.d.ts +4 -4
- package/src/tooling/boundary-evidence.js +20 -20
- package/src/tooling/browser-command-runner.js +13 -6
- package/src/tooling/capsule-v2.d.ts +11 -0
- package/src/tooling/capsule-v2.js +58 -0
- package/src/tooling/capsule.js +20 -0
- package/src/tooling/command-api-constants.js +1 -1
- package/src/tooling/command-api-family-normalizers.js +163 -5
- package/src/tooling/command-api.d.ts +42 -3
- package/src/tooling/command-api.js +7 -2
- package/src/tooling/command-envelope.js +1 -1
- package/src/tooling/{command-api-helpers.js → command-request-normalization.js} +1 -1
- package/src/tooling/command-runner.html +49 -1
- package/src/tooling/command-validation.js +1 -24
- package/src/tooling/conversion-run-timing.js +24 -0
- package/src/tooling/diagnose-runner.js +11 -9
- package/src/tooling/distribution-shard-transport.js +11 -0
- package/src/tooling/execution-cost-ledger.d.ts +76 -2
- package/src/tooling/execution-cost-ledger.js +197 -15
- package/src/tooling/hf-registry/catalog-io.js +63 -0
- package/src/tooling/litert/package.js +863 -0
- package/src/tooling/litert-package-runtime.js +3 -849
- package/src/tooling/model-capsule-forge.js +478 -0
- package/src/tooling/node-browser/execution.d.ts +40 -0
- package/src/tooling/node-browser/execution.js +308 -0
- package/src/tooling/node-browser/static-server.js +188 -0
- package/src/tooling/node-browser/transport.d.ts +80 -0
- package/src/tooling/node-browser/transport.js +622 -0
- package/src/tooling/node-browser-command-runner.d.ts +18 -78
- package/src/tooling/node-browser-command-runner.js +10 -1112
- package/src/tooling/node-capsule-artifact-store.js +44 -0
- package/src/tooling/node-command-execution.js +237 -0
- package/src/tooling/node-command-runner.js +26 -205
- package/src/tooling/node-converter/input.js +521 -0
- package/src/tooling/node-converter/output.js +244 -0
- package/src/tooling/node-converter-chunk-batches.js +21 -0
- package/src/tooling/node-converter.js +7 -720
- package/src/tooling/node-source-runtime.js +16 -276
- package/src/tooling/node-webgpu-provider.v1.json +33 -0
- package/src/tooling/node-webgpu.js +178 -453
- package/src/tooling/opfs-cache.d.ts +5 -0
- package/src/tooling/opfs-cache.js +4 -0
- package/src/tooling/production-release.d.ts +19 -0
- package/src/tooling/production-release.js +647 -0
- package/src/tooling/program-bundle/embedding-reference.js +19 -0
- package/src/tooling/program-bundle/materialize.js +324 -0
- package/src/tooling/program-bundle/rerank-reference.js +22 -0
- package/src/tooling/program-bundle/sequence-reference.js +42 -0
- package/src/tooling/program-bundle/validation.js +31 -0
- package/src/tooling/program-bundle/wgsl-closure.js +254 -0
- package/src/tooling/program-bundle-parity.d.ts +11 -4
- package/src/tooling/program-bundle-parity.js +262 -109
- package/src/tooling/program-bundle.d.ts +7 -0
- package/src/tooling/program-bundle.js +186 -591
- package/src/tooling/provider-v1-contract.d.ts +61 -0
- package/src/tooling/provider-v1-contract.js +192 -0
- package/src/tooling/registered-variant-calibration-job.d.ts +98 -0
- package/src/tooling/registered-variant-calibration-job.js +210 -0
- package/src/tooling/registered-variant-calibration.d.ts +167 -7
- package/src/tooling/registered-variant-calibration.js +57 -7
- package/src/tooling/runtime-optimization/candidates.js +664 -0
- package/src/tooling/runtime-optimization-campaign.js +166 -0
- package/src/tooling/runtime-optimization-index.js +16 -1
- package/src/tooling/runtime-optimization.d.ts +44 -0
- package/src/tooling/runtime-optimization.js +13 -656
- package/src/tooling/source-intake.d.ts +1 -1
- package/src/tooling/source-intake.js +34 -17
- package/src/tooling/source-package-profiles.js +1 -1
- package/src/tooling/source-runtime/bundle-materialize.js +298 -0
- package/src/tooling/source-runtime/bundle-plan.js +121 -0
- package/src/tooling/source-runtime/node-loader.js +258 -0
- package/src/tooling/source-runtime-bundle.js +26 -411
- package/src/tooling/source-runtime-converter-config.js +1 -1
- package/src/tooling-experimental-exports.js +0 -5
- package/src/tooling-exports/evidence.d.ts +14 -3
- package/src/tooling-exports/evidence.js +14 -3
- package/src/tooling-exports/storage.d.ts +5 -0
- package/src/tooling-exports/storage.js +4 -0
- package/src/tooling-exports.d.ts +3 -0
- package/src/tooling-exports.js +2 -0
- package/src/tooling-exports.shared.d.ts +17 -3
- package/src/tooling-exports.shared.js +9 -3
- package/src/training.d.ts +6 -6
- package/src/training.js +4 -4
- package/src/version.js +1 -1
- package/tests/kernels/browser/kernel-suite.js +512 -0
- package/tests/kernels/browser/test-page.js +169 -14
- package/tests/kernels/harness/tolerance.js +1 -0
- package/tests/kernels/reference/index.js +13 -1
- package/tests/kernels/reference/vision.js +160 -0
- package/src/converter/quantizer.d.ts +0 -121
- package/src/gpu/kernels/backward/utils.d.ts +0 -90
- package/src/gpu/kernels/uniform-utils.d.ts +0 -66
- package/src/gpu/kernels/utils.d.ts +0 -147
- package/src/inference/pipelines/text/debug-utils/utils.js +0 -118
- package/src/inference/pipelines/text/logits/utils.d.ts +0 -58
- package/src/inference/pipelines/text/moe-cpu-gptoss.js +0 -115
- package/src/inference/pipelines/text/moe-helpers.js +0 -21
- package/src/rules/converter/execution.rules.json +0 -20
- package/src/rules/converter/tensor-roles.rules.json +0 -13
- package/src/rules/converter/tokenizer.rules.json +0 -12
- package/src/rules/inference/attention.rules.json +0 -69
- package/src/rules/inference/config.rules.json +0 -82
- package/src/rules/inference/dtype.rules.json +0 -113
- package/src/rules/inference/execution.rules.json +0 -195
- package/src/rules/inference/ffn.rules.json +0 -123
- package/src/rules/inference/layer-pattern.rules.json +0 -16
- package/src/rules/inference/layer.rules.json +0 -7
- package/src/rules/inference/moe.rules.json +0 -145
- package/src/rules/kernels/attention.rules.json +0 -68
- package/src/rules/kernels/conv2d.rules.json +0 -6
- package/src/rules/kernels/depthwise-conv2d.rules.json +0 -6
- package/src/rules/kernels/dequant.rules.json +0 -58
- package/src/rules/kernels/energy.rules.json +0 -22
- package/src/rules/kernels/fused-ffn.rules.json +0 -15
- package/src/rules/kernels/fused-matmul-residual.rules.json +0 -6
- package/src/rules/kernels/fused-matmul-rmsnorm.rules.json +0 -8
- package/src/rules/kernels/gather.rules.json +0 -93
- package/src/rules/kernels/gelu.rules.json +0 -11
- package/src/rules/kernels/grouped-pointwise-conv2d.rules.json +0 -6
- package/src/rules/kernels/groupnorm.rules.json +0 -10
- package/src/rules/kernels/kv_quantize.rules.json +0 -9
- package/src/rules/kernels/layernorm.rules.json +0 -6
- package/src/rules/kernels/linear-attention.rules.json +0 -6
- package/src/rules/kernels/lm-head-argmax.rules.json +0 -10
- package/src/rules/kernels/matmul.rules.json +0 -95
- package/src/rules/kernels/modulate.rules.json +0 -6
- package/src/rules/kernels/moe.rules.gptoss.json +0 -105
- package/src/rules/kernels/moe.rules.json +0 -14
- package/src/rules/kernels/moe.rules.mixtral.json +0 -75
- package/src/rules/kernels/pixel_shuffle.rules.json +0 -6
- package/src/rules/kernels/relu.rules.json +0 -6
- package/src/rules/kernels/rep-penalty.rules.json +0 -6
- package/src/rules/kernels/repeat-channels.rules.json +0 -6
- package/src/rules/kernels/residual.rules.json +0 -12
- package/src/rules/kernels/rmsnorm-qk.rules.json +0 -6
- package/src/rules/kernels/rmsnorm.rules.json +0 -46
- package/src/rules/kernels/rope-qk.rules.json +0 -6
- package/src/rules/kernels/rope.rules.json +0 -6
- package/src/rules/kernels/sample.rules.json +0 -6
- package/src/rules/kernels/scale.rules.json +0 -8
- package/src/rules/kernels/silu.rules.json +0 -22
- package/src/rules/kernels/softmax.rules.json +0 -25
- package/src/rules/kernels/split-qg.rules.json +0 -6
- package/src/rules/kernels/split-qkv.rules.json +0 -6
- package/src/rules/kernels/upsample2d.rules.json +0 -6
- package/src/rules/loader/tensor-loader.rules.json +0 -42
- package/src/rules/loader/weights.rules.json +0 -42
- package/src/rules/tooling/command-runtime.rules.json +0 -65
- package/src/tooling/hf-registry-utils.js +0 -395
- /package/src/converter/{tokenizer-utils.js → tokenizer-special-tokens.js} +0 -0
- /package/src/{inference/pipelines/text/bdpa-steamroller.js → experimental/bdpa/steamroller.js} +0 -0
- /package/src/{inference/pipelines → experimental}/diffusion/scheduler.js +0 -0
- /package/src/{inference/pipelines → experimental}/energy/pipeline.d.ts +0 -0
- /package/src/{inference/pipelines → experimental}/energy/quintel.d.ts +0 -0
- /package/src/{inference/pipelines → experimental}/energy/quintel.js +0 -0
- /package/src/{utils → formats}/array-buffer.js +0 -0
- /package/src/{utils → formats}/canonical-hash.js +0 -0
- /package/src/{utils → formats}/clone-json.js +0 -0
- /package/src/{utils → formats}/encode-utf8.js +0 -0
- /package/src/{utils → formats}/load-json.js +0 -0
- /package/src/{utils → formats}/plain-object.js +0 -0
- /package/src/{utils → formats}/sha256.js +0 -0
- /package/src/{utils → formats}/stable-sort-object.js +0 -0
- /package/src/{gpu/kernels → rules}/rule-matcher.js +0 -0
- /package/src/{utils → storage}/hf-resolve-url.js +0 -0
- /package/src/{utils → storage}/runtime-env.js +0 -0
- /package/src/tooling/{command-runner-shared.js → command-runtime-execution.js} +0 -0
|
@@ -15,19 +15,20 @@ import {
|
|
|
15
15
|
isDecodeRecorderEnabled,
|
|
16
16
|
isProfileDecodeRecorderEnabled,
|
|
17
17
|
} from './execution-plan.js';
|
|
18
|
-
|
|
19
18
|
import { sample, applyRepetitionPenalty, logitsSanity, getTopK } from './sampling.js';
|
|
20
19
|
import { isStopToken } from './init.js';
|
|
21
20
|
import { embed } from './embed.js';
|
|
22
|
-
import { resolvePerLayerInputsSession } from './generator-
|
|
21
|
+
import { resolvePerLayerInputsSession } from './generator/session-context.js';
|
|
23
22
|
import { processLayer } from './layer.js';
|
|
24
|
-
import { computeLogits, computeLogitsGPU, recordLogitsGPU, recordGreedyLmHeadArgmaxGPU, extractLastPositionLogits
|
|
23
|
+
import { computeLogits, computeLogitsGPU, recordLogitsGPU, recordGreedyLmHeadArgmaxGPU, extractLastPositionLogits } from './logits/index.js';
|
|
25
24
|
import { isWeightBuffer, isCpuWeightBuffer, isGpuBufferInstance, isSplitWeightBuffer, getWeightDtype, getWeightMetadata } from '../../../gpu/weight-buffer.js';
|
|
26
25
|
import { decodeReadback } from './debug-utils/index.js';
|
|
27
|
-
import {
|
|
26
|
+
import { captureObservedFusedDecodeLogits, emitObservedLogits } from './generator-logits-observation.js';
|
|
27
|
+
import { extractEmbeddingFromHiddenGPU } from './embedding-extraction.js';
|
|
28
28
|
import { parseFinitenessStatusWords } from './finiteness-guard-status.js';
|
|
29
29
|
import { hasLinearAttentionLayers } from './linear-attention.js';
|
|
30
30
|
import { hasConvLayers } from './layer.js';
|
|
31
|
+
import { advanceDecodeStepCount } from './tsir-fixture-writer.js';
|
|
31
32
|
import {
|
|
32
33
|
preparePerLayerInputs,
|
|
33
34
|
createPleBufferCache,
|
|
@@ -36,22 +37,11 @@ import {
|
|
|
36
37
|
hasGpuSplitPerLayerInputEmbeddings,
|
|
37
38
|
getPleHotVocabularyRuntime,
|
|
38
39
|
} from './per-layer-inputs.js';
|
|
40
|
+
import { FINITENESS_RESET_WORDS, FinitenessError, decodeStep, getEffectiveActivationDtype, isOwnedDecodeBuffer, recordDecodeProfileStep, releasePerLayerInputBuffer, runDecodeLayers, shouldLogProfileStep, sumProfileTimings } from './generator/decode.js';
|
|
41
|
+
export { FinitenessError, decodeStep, decodeStepLogits, readMappedBufferCopy, readSampledTokenFromStagingBuffer, shouldUseFusedDecodeSampling, sumProfileTimings } from './generator/decode.js';
|
|
39
42
|
|
|
40
|
-
const UNKNOWN_TOKEN_TEXT = '<unknown>';
|
|
41
|
-
const FINITENESS_RESET_WORDS = new Uint32Array(4);
|
|
42
43
|
const FINITENESS_STATUS_BYTES = FINITENESS_RESET_WORDS.byteLength;
|
|
43
44
|
|
|
44
|
-
export function sumProfileTimings(timings) {
|
|
45
|
-
if (!timings || Object.keys(timings).length === 0) return null;
|
|
46
|
-
let total = 0;
|
|
47
|
-
for (const value of Object.values(timings)) {
|
|
48
|
-
if (Number.isFinite(value)) {
|
|
49
|
-
total += value;
|
|
50
|
-
}
|
|
51
|
-
}
|
|
52
|
-
return total;
|
|
53
|
-
}
|
|
54
|
-
|
|
55
45
|
function mergeRecordOpLabelCounts(target, source) {
|
|
56
46
|
const merged = target && typeof target === 'object' && !Array.isArray(target)
|
|
57
47
|
? target
|
|
@@ -69,134 +59,13 @@ function mergeRecordOpLabelCounts(target, source) {
|
|
|
69
59
|
return merged;
|
|
70
60
|
}
|
|
71
61
|
|
|
72
|
-
function getEffectiveActivationDtype(state, opts) {
|
|
73
|
-
const executionPlanDtype = opts?.executionPlan?.activationDtype;
|
|
74
|
-
if (executionPlanDtype !== undefined && executionPlanDtype !== null) {
|
|
75
|
-
return executionPlanDtype;
|
|
76
|
-
}
|
|
77
|
-
if (executionPlanDtype === null) {
|
|
78
|
-
throw new Error('[Pipeline] executionPlan.activationDtype is required when provided and cannot be null.');
|
|
79
|
-
}
|
|
80
|
-
return state.runtimeConfig.inference.compute.activationDtype;
|
|
81
|
-
}
|
|
82
|
-
|
|
83
|
-
function getTokenTextOrUnknown(tokenizer, tokenId) {
|
|
84
|
-
if (!tokenizer || typeof tokenizer.decode !== 'function') {
|
|
85
|
-
return UNKNOWN_TOKEN_TEXT;
|
|
86
|
-
}
|
|
87
|
-
|
|
88
|
-
const tokenText = tokenizer.decode([tokenId], false, false);
|
|
89
|
-
if (typeof tokenText !== 'string' || tokenText.length === 0) {
|
|
90
|
-
return UNKNOWN_TOKEN_TEXT;
|
|
91
|
-
}
|
|
92
|
-
|
|
93
|
-
return tokenText;
|
|
94
|
-
}
|
|
95
|
-
|
|
96
|
-
function isOwnedDecodeBuffer(candidate, decodeHiddenBuffer, decodeAltBuffer) {
|
|
97
|
-
if (candidate === decodeHiddenBuffer) {
|
|
98
|
-
return true;
|
|
99
|
-
}
|
|
100
|
-
return candidate === decodeAltBuffer;
|
|
101
|
-
}
|
|
102
|
-
|
|
103
|
-
function releasePerLayerInputBuffer(buffer, recorder, decodeBuffers, pleCache = null) {
|
|
104
|
-
if (!buffer) {
|
|
105
|
-
return;
|
|
106
|
-
}
|
|
107
|
-
const ownsBuffer = decodeBuffers?.ownsBuffer(buffer) ?? false;
|
|
108
|
-
if (ownsBuffer) {
|
|
109
|
-
return;
|
|
110
|
-
}
|
|
111
|
-
const cachedPleBuffer = pleCache?.ownedBuffers instanceof Set && pleCache.ownedBuffers.has(buffer);
|
|
112
|
-
if (cachedPleBuffer) {
|
|
113
|
-
return;
|
|
114
|
-
}
|
|
115
|
-
if (recorder) {
|
|
116
|
-
recorder.trackTemporaryBuffer(buffer);
|
|
117
|
-
return;
|
|
118
|
-
}
|
|
119
|
-
releaseBuffer(buffer);
|
|
120
|
-
}
|
|
121
|
-
|
|
122
|
-
function schedulePlePrefetchForToken(state, tokenId) {
|
|
123
|
-
if (state?.prefetchPleNextToken !== true) {
|
|
124
|
-
return;
|
|
125
|
-
}
|
|
126
|
-
const config = state.modelConfig;
|
|
127
|
-
const pleHiddenSize = Number(config?.hiddenSizePerLayerInput ?? 0);
|
|
128
|
-
if (!Number.isFinite(pleHiddenSize) || pleHiddenSize <= 0) {
|
|
129
|
-
return;
|
|
130
|
-
}
|
|
131
|
-
const pleWeights = state.weights.get('per_layer_inputs');
|
|
132
|
-
if (!pleWeights?.embedTokensPerLayer) {
|
|
133
|
-
return;
|
|
134
|
-
}
|
|
135
|
-
const resolvedPerLayerInputsSession = resolvePerLayerInputsSession(
|
|
136
|
-
config.perLayerInputsSession ?? null,
|
|
137
|
-
state.runtimeConfig?.inference?.session?.perLayerInputs ?? null
|
|
138
|
-
);
|
|
139
|
-
state.plePrefetchPending = prefetchPerLayerRow(
|
|
140
|
-
tokenId,
|
|
141
|
-
pleWeights.embedTokensPerLayer,
|
|
142
|
-
config.numLayers * pleHiddenSize,
|
|
143
|
-
resolvedPerLayerInputsSession
|
|
144
|
-
);
|
|
145
|
-
}
|
|
146
|
-
|
|
147
|
-
function getReusableSampleReadbackBuffer(state, device, size) {
|
|
148
|
-
const existing = state.sampleReadbackBuffer;
|
|
149
|
-
if (existing && existing.size >= size) {
|
|
150
|
-
return existing;
|
|
151
|
-
}
|
|
152
|
-
if (existing) {
|
|
153
|
-
existing.destroy();
|
|
154
|
-
}
|
|
155
|
-
const buffer = device.createBuffer({
|
|
156
|
-
label: 'sample_staging_reuse',
|
|
157
|
-
size,
|
|
158
|
-
usage: GPUBufferUsage.MAP_READ | GPUBufferUsage.COPY_DST,
|
|
159
|
-
});
|
|
160
|
-
state.sampleReadbackBuffer = buffer;
|
|
161
|
-
return buffer;
|
|
162
|
-
}
|
|
163
|
-
|
|
164
|
-
export class FinitenessError extends Error {
|
|
165
|
-
constructor(message) {
|
|
166
|
-
super(message);
|
|
167
|
-
this.name = 'FinitenessError';
|
|
168
|
-
}
|
|
169
|
-
}
|
|
170
|
-
|
|
171
|
-
function shouldLogProfileStep(state, step) {
|
|
172
|
-
const profilerConfig = state.runtimeConfig?.shared?.debug?.profiler;
|
|
173
|
-
const every = profilerConfig?.logEveryDecodeSteps ?? 1;
|
|
174
|
-
if (!Number.isFinite(every) || every <= 1) return true;
|
|
175
|
-
return step === 1 || step % every === 0;
|
|
176
|
-
}
|
|
177
|
-
|
|
178
|
-
function recordDecodeProfileStep(state, entry) {
|
|
179
|
-
if (!entry || (!entry.timings && !entry.recorderStats)) return;
|
|
180
|
-
if (!state.stats.decodeProfileSteps) {
|
|
181
|
-
state.stats.decodeProfileSteps = [];
|
|
182
|
-
}
|
|
183
|
-
state.stats.decodeProfileSteps.push(entry);
|
|
184
|
-
}
|
|
185
|
-
|
|
186
62
|
export function shouldUseBatchDecode(config) {
|
|
187
63
|
return isBatchDecodeEnabled(config);
|
|
188
64
|
}
|
|
189
65
|
|
|
190
|
-
export function shouldUseFusedDecodeSampling(config) {
|
|
191
|
-
return config.recorderEnabled === true
|
|
192
|
-
&& config.gpuSamplingEnabled === true
|
|
193
|
-
&& config.fusedDecodeDisabled !== true
|
|
194
|
-
&& !hasConvLayers(config.layerTypes ?? []);
|
|
195
|
-
}
|
|
196
|
-
|
|
197
66
|
function shouldUseGreedyLmHeadArgmaxFusion(state, opts, samplingDefaults, repetitionPenalty) {
|
|
198
67
|
const probes = state.runtimeConfig?.shared?.debug?.probes;
|
|
199
|
-
return state.runtimeConfig?.inference?.session?.useGreedyLmHeadArgmaxFusion === true
|
|
68
|
+
return state.runtimeConfig?.inference?.session?.useGreedyLmHeadArgmaxFusion === true && (state.modelConfig?.logitOutputScale ?? 1) === 1
|
|
200
69
|
&& opts.temperature < samplingDefaults.greedyThreshold
|
|
201
70
|
&& repetitionPenalty === 1.0
|
|
202
71
|
&& state.operatorDiagnostics == null
|
|
@@ -276,72 +145,6 @@ export function findInvalidGeneratedToken(tokens, vocabSize, padTokenId = null)
|
|
|
276
145
|
return null;
|
|
277
146
|
}
|
|
278
147
|
|
|
279
|
-
export async function readSampledTokenFromStagingBuffer(stagingBuffer, options = {}) {
|
|
280
|
-
const ownsStagingBuffer = options.ownsStagingBuffer === true;
|
|
281
|
-
const hasFinitenessBuffer = options.hasFinitenessBuffer === true;
|
|
282
|
-
const ring = options.ring ?? null;
|
|
283
|
-
const cleanupRecorder = options.cleanupRecorder ?? null;
|
|
284
|
-
const timing = {
|
|
285
|
-
mapWaitMs: 0,
|
|
286
|
-
cleanupMs: 0,
|
|
287
|
-
copyMs: 0,
|
|
288
|
-
};
|
|
289
|
-
let mapped = false;
|
|
290
|
-
let cleanupCompleted = false;
|
|
291
|
-
|
|
292
|
-
try {
|
|
293
|
-
const mapStart = performance.now();
|
|
294
|
-
await stagingBuffer.mapAsync(GPUMapMode.READ);
|
|
295
|
-
timing.mapWaitMs = performance.now() - mapStart;
|
|
296
|
-
mapped = true;
|
|
297
|
-
const cleanupStart = performance.now();
|
|
298
|
-
await cleanupRecorder?.completeDeferredCleanup();
|
|
299
|
-
timing.cleanupMs = performance.now() - cleanupStart;
|
|
300
|
-
cleanupCompleted = true;
|
|
301
|
-
const copyStart = performance.now();
|
|
302
|
-
const mappedWords = new Uint32Array(stagingBuffer.getMappedRange());
|
|
303
|
-
const result = {
|
|
304
|
-
nextToken: mappedWords[0],
|
|
305
|
-
finitenessStatus: hasFinitenessBuffer
|
|
306
|
-
? parseFinitenessStatusWords(mappedWords, 1)
|
|
307
|
-
: parseFinitenessStatusWords(mappedWords, 0),
|
|
308
|
-
timing,
|
|
309
|
-
};
|
|
310
|
-
timing.copyMs = performance.now() - copyStart;
|
|
311
|
-
return result;
|
|
312
|
-
} finally {
|
|
313
|
-
if (mapped) {
|
|
314
|
-
stagingBuffer.unmap();
|
|
315
|
-
}
|
|
316
|
-
if (!cleanupCompleted) {
|
|
317
|
-
await cleanupRecorder?.completeDeferredCleanup({ discardPooled: true });
|
|
318
|
-
}
|
|
319
|
-
if (ownsStagingBuffer) {
|
|
320
|
-
stagingBuffer.destroy();
|
|
321
|
-
}
|
|
322
|
-
ring?.advance();
|
|
323
|
-
}
|
|
324
|
-
}
|
|
325
|
-
|
|
326
|
-
|
|
327
|
-
export async function readMappedBufferCopy(stagingBuffer, options = {}) {
|
|
328
|
-
const ownsStagingBuffer = options.ownsStagingBuffer !== false;
|
|
329
|
-
let mapped = false;
|
|
330
|
-
|
|
331
|
-
try {
|
|
332
|
-
await stagingBuffer.mapAsync(GPUMapMode.READ);
|
|
333
|
-
mapped = true;
|
|
334
|
-
return stagingBuffer.getMappedRange().slice(0);
|
|
335
|
-
} finally {
|
|
336
|
-
if (mapped) {
|
|
337
|
-
stagingBuffer.unmap();
|
|
338
|
-
}
|
|
339
|
-
if (ownsStagingBuffer) {
|
|
340
|
-
stagingBuffer.destroy();
|
|
341
|
-
}
|
|
342
|
-
}
|
|
343
|
-
}
|
|
344
|
-
|
|
345
148
|
export async function readBatchTokensFromStagingBuffers(options) {
|
|
346
149
|
const {
|
|
347
150
|
tokensStagingBuffer,
|
|
@@ -440,730 +243,8 @@ export async function readBatchTokensFromStagingBuffers(options) {
|
|
|
440
243
|
}
|
|
441
244
|
}
|
|
442
245
|
|
|
443
|
-
async function runDecodeLayers(state, tokenId, opts, helpers) {
|
|
444
|
-
const config = state.modelConfig;
|
|
445
|
-
const debugCheckBuffer = state.debug ? helpers.debugCheckBuffer : undefined;
|
|
446
|
-
|
|
447
|
-
const context = helpers.buildLayerContext(undefined, true, opts.debugLayers, opts.executionPlan);
|
|
448
|
-
context.currentTokenIds = [tokenId];
|
|
449
|
-
|
|
450
|
-
state.decodeBuffers.resetPingPong();
|
|
451
|
-
|
|
452
|
-
const decodeHiddenBuffer = state.decodeBuffers.getHiddenBuffer();
|
|
453
|
-
const decodeAltBuffer = state.decodeBuffers.getOutputHiddenBuffer();
|
|
454
|
-
|
|
455
|
-
const embedBufferRaw = state.weights.get('embed');
|
|
456
|
-
if (!isGpuBufferInstance(embedBufferRaw) && !isWeightBuffer(embedBufferRaw) && !isCpuWeightBuffer(embedBufferRaw) && !isSplitWeightBuffer(embedBufferRaw) && !(embedBufferRaw instanceof Float32Array)) {
|
|
457
|
-
throw new Error('Embed buffer not found or not a supported buffer type');
|
|
458
|
-
}
|
|
459
|
-
const embedBuffer = isWeightBuffer(embedBufferRaw) ? embedBufferRaw.buffer : embedBufferRaw;
|
|
460
|
-
const embedDtype = isCpuWeightBuffer(embedBufferRaw)
|
|
461
|
-
? embedBufferRaw.dtype
|
|
462
|
-
: getWeightDtype(embedBufferRaw);
|
|
463
|
-
const embedMetadata = getWeightMetadata(embedBufferRaw);
|
|
464
|
-
const activationDtype = getEffectiveActivationDtype(state, opts);
|
|
465
|
-
|
|
466
|
-
const embedTensor = await embed([tokenId], embedBuffer, {
|
|
467
|
-
hiddenSize: config.hiddenSize,
|
|
468
|
-
vocabSize: config.vocabSize,
|
|
469
|
-
scaleEmbeddings: config.scaleEmbeddings,
|
|
470
|
-
embeddingScale: config.embeddingScale,
|
|
471
|
-
outputBuffer: decodeHiddenBuffer ?? undefined,
|
|
472
|
-
transpose: state.embeddingTranspose,
|
|
473
|
-
debugProbes: state.runtimeConfig.shared.debug.probes,
|
|
474
|
-
operatorDiagnostics: state.operatorDiagnostics,
|
|
475
|
-
activationDtype,
|
|
476
|
-
embeddingDtype: selectRuleValue('inference', 'dtype', 'embeddingDtype', { dtype: embedDtype }),
|
|
477
|
-
embeddingStorageEncoding: embedMetadata?.storageEncoding ?? null,
|
|
478
|
-
executionPolicies: state.executionV1State?.policies ?? null,
|
|
479
|
-
});
|
|
480
|
-
|
|
481
|
-
let hiddenStates = embedTensor.buffer;
|
|
482
|
-
|
|
483
|
-
// Resolve pending PLE prefetch from previous decode step
|
|
484
|
-
let plePrefetchResult = null;
|
|
485
|
-
if (state.plePrefetchPending) {
|
|
486
|
-
plePrefetchResult = await state.plePrefetchPending;
|
|
487
|
-
state.plePrefetchPending = null;
|
|
488
|
-
}
|
|
489
|
-
|
|
490
|
-
const perLayerInputs = await preparePerLayerInputs([tokenId], embedTensor, context, {
|
|
491
|
-
numTokens: 1,
|
|
492
|
-
pleCache: state.pleCache ?? null,
|
|
493
|
-
prefetchedRow: plePrefetchResult,
|
|
494
|
-
});
|
|
495
|
-
|
|
496
|
-
try {
|
|
497
|
-
for (let l = 0; l < config.numLayers; l++) {
|
|
498
|
-
context.perLayerInputBuffer = perLayerInputs?.[l] ?? null;
|
|
499
|
-
const prevStates = hiddenStates;
|
|
500
|
-
hiddenStates = (await processLayer(l, hiddenStates, 1, false, context));
|
|
501
|
-
state.decodeBuffers.swapPingPong();
|
|
502
|
-
releasePerLayerInputBuffer(context.perLayerInputBuffer, null, context.decodeBuffers, state.pleCache ?? null);
|
|
503
|
-
if (perLayerInputs) {
|
|
504
|
-
perLayerInputs[l] = null;
|
|
505
|
-
}
|
|
506
|
-
context.perLayerInputBuffer = null;
|
|
507
|
-
if (isGpuBufferInstance(prevStates) && prevStates !== hiddenStates) {
|
|
508
|
-
const isPreAllocated = isOwnedDecodeBuffer(prevStates, decodeHiddenBuffer, decodeAltBuffer);
|
|
509
|
-
if (!isPreAllocated) {
|
|
510
|
-
releaseBuffer(prevStates);
|
|
511
|
-
}
|
|
512
|
-
}
|
|
513
|
-
}
|
|
514
|
-
} finally {
|
|
515
|
-
context.perLayerInputBuffer = null;
|
|
516
|
-
if (perLayerInputs) {
|
|
517
|
-
for (const buffer of perLayerInputs) {
|
|
518
|
-
releasePerLayerInputBuffer(buffer, null, context.decodeBuffers, state.pleCache ?? null);
|
|
519
|
-
}
|
|
520
|
-
}
|
|
521
|
-
helpers.releaseSharedAttentionState?.(context.sharedAttentionState, null);
|
|
522
|
-
}
|
|
523
|
-
|
|
524
|
-
return { hiddenStates, decodeHiddenBuffer, decodeAltBuffer, debugCheckBuffer, context };
|
|
525
|
-
}
|
|
526
|
-
|
|
527
|
-
function createDecodeRecorder(state, opts) {
|
|
528
|
-
const device = getDevice();
|
|
529
|
-
const executionPlan = opts.executionPlan;
|
|
530
|
-
const recorderConfig = {
|
|
531
|
-
hasDevice: Boolean(device),
|
|
532
|
-
debug: opts.debug,
|
|
533
|
-
disableCommandBatching: executionPlan?.disableCommandBatching ?? opts.disableCommandBatching,
|
|
534
|
-
kvLayout: state.kvCache?.layout ?? null,
|
|
535
|
-
};
|
|
536
|
-
const recorderEnabled = opts.profile
|
|
537
|
-
? isProfileDecodeRecorderEnabled(recorderConfig)
|
|
538
|
-
: isDecodeRecorderEnabled(recorderConfig);
|
|
539
|
-
let recorder;
|
|
540
|
-
if (recorderEnabled) {
|
|
541
|
-
recorder = opts.profile
|
|
542
|
-
? createProfilingRecorder('decode', device)
|
|
543
|
-
: createCommandRecorder('decode', { recordLabels: opts.debug === true || opts.benchmark === true }, device);
|
|
544
|
-
}
|
|
545
|
-
if (state.decodeStepCount === 1) {
|
|
546
|
-
const path = selectRuleValue('inference', 'config', 'tracePath', { useRecorder: Boolean(recorder) });
|
|
547
|
-
log.debug('Decode', `Using ${path} path (recorder=${!!recorder}, debug=${opts.debug})`);
|
|
548
|
-
}
|
|
549
|
-
return recorder;
|
|
550
|
-
}
|
|
551
|
-
|
|
552
|
-
async function submitDecodeRecorderProfile(state, opts, recorder, profileLabel) {
|
|
553
|
-
if (!recorder) {
|
|
554
|
-
return;
|
|
555
|
-
}
|
|
556
|
-
await recorder.submitAndWait();
|
|
557
|
-
|
|
558
|
-
if (!opts.profile) {
|
|
559
|
-
return;
|
|
560
|
-
}
|
|
561
|
-
|
|
562
|
-
const timings = recorder.isProfilingEnabled()
|
|
563
|
-
? await recorder.resolveProfileTimings()
|
|
564
|
-
: null;
|
|
565
|
-
const total = sumProfileTimings(timings);
|
|
566
|
-
if (total !== null) {
|
|
567
|
-
state.stats.gpuTimeDecodeMs = (state.stats.gpuTimeDecodeMs ?? 0) + total;
|
|
568
|
-
}
|
|
569
|
-
if (timings || opts.executionObserver) {
|
|
570
|
-
recordDecodeProfileStep(state, {
|
|
571
|
-
step: state.decodeStepCount,
|
|
572
|
-
timings,
|
|
573
|
-
totalMs: total ?? undefined,
|
|
574
|
-
recorderStats: recorder.getStats(),
|
|
575
|
-
});
|
|
576
|
-
if (!opts.executionObserver && shouldLogProfileStep(state, state.decodeStepCount)) {
|
|
577
|
-
log.warn('Profile', `Decode step ${state.decodeStepCount}${profileLabel}:`);
|
|
578
|
-
log.warn('Profile', CommandRecorder.formatProfileReport(timings));
|
|
579
|
-
}
|
|
580
|
-
}
|
|
581
|
-
}
|
|
582
|
-
|
|
583
|
-
export async function decodeStep(state, currentIds, opts, helpers) {
|
|
584
|
-
const stepWallStart = performance.now();
|
|
585
|
-
const lastToken = currentIds[currentIds.length - 1];
|
|
586
|
-
const numTokens = 1;
|
|
587
|
-
const config = state.modelConfig;
|
|
588
|
-
const samplingDefaults = state.runtimeConfig.inference.sampling;
|
|
589
|
-
const executionPlan = opts.executionPlan;
|
|
590
|
-
const debugCheckBuffer = state.debug ? helpers.debugCheckBuffer : undefined;
|
|
591
|
-
|
|
592
|
-
state.decodeStepCount++;
|
|
593
|
-
const isDebugStep = opts.debug && state.decodeStepCount <= 5;
|
|
594
|
-
if (isDebugStep) {
|
|
595
|
-
const tokenText = getTokenTextOrUnknown(state.tokenizer, lastToken);
|
|
596
|
-
log.debug('Decode', `[${state.decodeStepCount}] token="${tokenText}" pos=${state.currentSeqLen}`);
|
|
597
|
-
}
|
|
598
|
-
|
|
599
|
-
const device = getDevice();
|
|
600
|
-
const recorder = createDecodeRecorder(state, opts);
|
|
601
|
-
|
|
602
|
-
if (state.finitenessBuffer && device) {
|
|
603
|
-
device.queue.writeBuffer(state.finitenessBuffer, 0, FINITENESS_RESET_WORDS);
|
|
604
|
-
}
|
|
605
|
-
|
|
606
|
-
const context = helpers.buildLayerContext(recorder, true, opts.debugLayers, executionPlan);
|
|
607
|
-
context.currentTokenIds = [lastToken];
|
|
608
|
-
|
|
609
|
-
state.decodeBuffers.resetPingPong();
|
|
610
|
-
|
|
611
|
-
const decodeHiddenBuffer = state.decodeBuffers.getHiddenBuffer();
|
|
612
|
-
const decodeAltBuffer = state.decodeBuffers.getOutputHiddenBuffer();
|
|
613
|
-
|
|
614
|
-
const embedBufferRaw = state.weights.get('embed');
|
|
615
|
-
if (!isGpuBufferInstance(embedBufferRaw) && !isWeightBuffer(embedBufferRaw) && !isCpuWeightBuffer(embedBufferRaw) && !isSplitWeightBuffer(embedBufferRaw) && !(embedBufferRaw instanceof Float32Array)) {
|
|
616
|
-
throw new Error('Embed buffer not found or not a supported buffer type');
|
|
617
|
-
}
|
|
618
|
-
const embedBuffer = isWeightBuffer(embedBufferRaw) ? embedBufferRaw.buffer : embedBufferRaw;
|
|
619
|
-
const embedDtype = isCpuWeightBuffer(embedBufferRaw)
|
|
620
|
-
? embedBufferRaw.dtype
|
|
621
|
-
: getWeightDtype(embedBufferRaw);
|
|
622
|
-
const embedMetadata = getWeightMetadata(embedBufferRaw);
|
|
623
|
-
const activationDtype = getEffectiveActivationDtype(state, opts);
|
|
624
|
-
const activationBytes = selectRuleValue('shared', 'dtype', 'bytesFromDtype', { dtype: activationDtype });
|
|
625
|
-
|
|
626
|
-
const embedTensor = await embed([lastToken], embedBuffer, {
|
|
627
|
-
hiddenSize: config.hiddenSize,
|
|
628
|
-
vocabSize: config.vocabSize,
|
|
629
|
-
scaleEmbeddings: config.scaleEmbeddings,
|
|
630
|
-
embeddingScale: config.embeddingScale,
|
|
631
|
-
recorder,
|
|
632
|
-
outputBuffer: decodeHiddenBuffer ?? undefined,
|
|
633
|
-
transpose: state.embeddingTranspose,
|
|
634
|
-
debugProbes: state.runtimeConfig.shared.debug.probes,
|
|
635
|
-
operatorDiagnostics: state.operatorDiagnostics,
|
|
636
|
-
activationDtype,
|
|
637
|
-
embeddingDtype: selectRuleValue('inference', 'dtype', 'embeddingDtype', { dtype: embedDtype }),
|
|
638
|
-
embeddingStorageEncoding: embedMetadata?.storageEncoding ?? null,
|
|
639
|
-
executionPolicies: state.executionV1State?.policies ?? null,
|
|
640
|
-
});
|
|
641
|
-
|
|
642
|
-
let hiddenStates = embedTensor.buffer;
|
|
643
|
-
|
|
644
|
-
// Resolve pending PLE prefetch from previous decode step
|
|
645
|
-
let plePrefetchResult = null;
|
|
646
|
-
if (state.plePrefetchPending) {
|
|
647
|
-
plePrefetchResult = await state.plePrefetchPending;
|
|
648
|
-
state.plePrefetchPending = null;
|
|
649
|
-
}
|
|
650
|
-
|
|
651
|
-
const perLayerInputs = await preparePerLayerInputs([lastToken], embedTensor, context, {
|
|
652
|
-
numTokens: 1,
|
|
653
|
-
pleCache: state.pleCache ?? null,
|
|
654
|
-
prefetchedRow: plePrefetchResult,
|
|
655
|
-
});
|
|
656
|
-
|
|
657
|
-
if (opts.debug && state.decodeStepCount === 1) {
|
|
658
|
-
const validSize = config.hiddenSize * activationBytes;
|
|
659
|
-
const embedData = await readBuffer(hiddenStates, validSize);
|
|
660
|
-
const embedArr = decodeReadback(embedData, activationDtype);
|
|
661
|
-
const sample = embedArr.slice(0, 5);
|
|
662
|
-
const maxAbs = Math.max(...embedArr.map(Math.abs));
|
|
663
|
-
const nonZero = embedArr.filter(x => Math.abs(x) > 1e-10).length;
|
|
664
|
-
log.debug('Decode', `[1] Embed check: maxAbs=${maxAbs.toFixed(2)}, nonZero=${nonZero}/${embedArr.length}, sample=[${Array.from(sample).map(v => v.toFixed(3)).join(', ')}]`);
|
|
665
|
-
}
|
|
666
|
-
|
|
667
|
-
const benchmarkSubmits = state.decodeStepCount <= 3 && opts.debug;
|
|
668
|
-
if (benchmarkSubmits) {
|
|
669
|
-
setTrackSubmits(true);
|
|
670
|
-
resetSubmitStats();
|
|
671
|
-
}
|
|
672
|
-
|
|
673
|
-
const hasGPUCache = context.kvCache?.hasGPUCache?.() ?? false;
|
|
674
|
-
if (opts.debug && state.decodeStepCount === 1) {
|
|
675
|
-
log.debug('Decode', `KV cache check: hasGPUCache=${hasGPUCache}, currentSeqLen=${context.currentSeqLen}`);
|
|
676
|
-
}
|
|
677
|
-
|
|
678
|
-
try {
|
|
679
|
-
for (let l = 0; l < config.numLayers; l++) {
|
|
680
|
-
context.perLayerInputBuffer = perLayerInputs?.[l] ?? null;
|
|
681
|
-
const prevStates = hiddenStates;
|
|
682
|
-
hiddenStates = (await processLayer(l, hiddenStates, numTokens, false, context));
|
|
683
|
-
|
|
684
|
-
state.decodeBuffers.swapPingPong();
|
|
685
|
-
releasePerLayerInputBuffer(
|
|
686
|
-
context.perLayerInputBuffer,
|
|
687
|
-
recorder,
|
|
688
|
-
context.decodeBuffers,
|
|
689
|
-
state.pleCache ?? null
|
|
690
|
-
);
|
|
691
|
-
if (perLayerInputs) {
|
|
692
|
-
perLayerInputs[l] = null;
|
|
693
|
-
}
|
|
694
|
-
context.perLayerInputBuffer = null;
|
|
695
|
-
|
|
696
|
-
if (isGpuBufferInstance(prevStates) && prevStates !== hiddenStates) {
|
|
697
|
-
const isPreAllocated = isOwnedDecodeBuffer(prevStates, decodeHiddenBuffer, decodeAltBuffer);
|
|
698
|
-
if (!isPreAllocated) {
|
|
699
|
-
if (recorder) {
|
|
700
|
-
recorder.trackTemporaryBuffer(prevStates);
|
|
701
|
-
} else {
|
|
702
|
-
releaseBuffer(prevStates);
|
|
703
|
-
}
|
|
704
|
-
}
|
|
705
|
-
}
|
|
706
|
-
}
|
|
707
|
-
} finally {
|
|
708
|
-
context.perLayerInputBuffer = null;
|
|
709
|
-
if (perLayerInputs) {
|
|
710
|
-
for (const buffer of perLayerInputs) {
|
|
711
|
-
releasePerLayerInputBuffer(
|
|
712
|
-
buffer,
|
|
713
|
-
recorder,
|
|
714
|
-
context.decodeBuffers,
|
|
715
|
-
state.pleCache ?? null
|
|
716
|
-
);
|
|
717
|
-
}
|
|
718
|
-
}
|
|
719
|
-
helpers.releaseSharedAttentionState?.(context.sharedAttentionState, recorder);
|
|
720
|
-
}
|
|
721
|
-
|
|
722
|
-
const logitSoftcap = config.finalLogitSoftcapping === null
|
|
723
|
-
? 0
|
|
724
|
-
: config.finalLogitSoftcapping;
|
|
725
|
-
const padTokenId = state.tokenizer?.getSpecialTokens?.()?.pad ?? null;
|
|
726
|
-
const lmHeadIsCpu = isCpuWeightBuffer(state.weights.get('lm_head'));
|
|
727
|
-
const useGPUSampling = state.useGPU && isGPUSamplingAvailable() && !lmHeadIsCpu;
|
|
728
|
-
const useFusedDecode = shouldUseFusedDecodeSampling({
|
|
729
|
-
recorderEnabled: Boolean(recorder),
|
|
730
|
-
gpuSamplingEnabled: useGPUSampling,
|
|
731
|
-
fusedDecodeDisabled: state.disableFusedDecode,
|
|
732
|
-
layerTypes: config.layerTypes,
|
|
733
|
-
});
|
|
734
|
-
|
|
735
|
-
if (useFusedDecode) {
|
|
736
|
-
const ring = state.decodeRing;
|
|
737
|
-
let ringSlot = null;
|
|
738
|
-
if (ring) {
|
|
739
|
-
ring.ensure({
|
|
740
|
-
batchSize: 1,
|
|
741
|
-
tokensPerInterval: 1,
|
|
742
|
-
stopCheckMode: executionPlan?.stopCheckMode ?? opts.stopCheckMode,
|
|
743
|
-
ringTokens: executionPlan?.ringTokens ?? state.runtimeConfig.inference.batching.ringTokens,
|
|
744
|
-
ringStop: executionPlan?.ringStop ?? state.runtimeConfig.inference.batching.ringStop,
|
|
745
|
-
ringStaging: executionPlan?.ringStaging ?? state.runtimeConfig.inference.batching.ringStaging,
|
|
746
|
-
});
|
|
747
|
-
ringSlot = ring.acquire();
|
|
748
|
-
}
|
|
749
|
-
|
|
750
|
-
const { logitsBuffer, vocabSize, logitsDtype } = await recordLogitsGPU(
|
|
751
|
-
recorder,
|
|
752
|
-
hiddenStates,
|
|
753
|
-
numTokens,
|
|
754
|
-
helpers.getLogitsWeights(),
|
|
755
|
-
helpers.getLogitsConfig(),
|
|
756
|
-
state.operatorDiagnostics,
|
|
757
|
-
);
|
|
758
|
-
|
|
759
|
-
const ringTokensBuffer = ringSlot?.tokens ?? null;
|
|
760
|
-
const sampleOutputBuffer = opts.temperature < samplingDefaults.greedyThreshold
|
|
761
|
-
? await recordArgmax(recorder, logitsBuffer, vocabSize, {
|
|
762
|
-
padTokenId,
|
|
763
|
-
logitSoftcap,
|
|
764
|
-
logitsDtype,
|
|
765
|
-
outputBuffer: ringTokensBuffer ?? undefined,
|
|
766
|
-
outputIndex: 0,
|
|
767
|
-
})
|
|
768
|
-
: await recordGPUSample(recorder, logitsBuffer, vocabSize, {
|
|
769
|
-
temperature: opts.temperature,
|
|
770
|
-
topK: opts.topK,
|
|
771
|
-
padTokenId,
|
|
772
|
-
logitSoftcap,
|
|
773
|
-
logitsDtype,
|
|
774
|
-
outputBuffer: ringTokensBuffer ?? undefined,
|
|
775
|
-
outputIndex: 0,
|
|
776
|
-
greedyThreshold: samplingDefaults.greedyThreshold,
|
|
777
|
-
randomSeed: opts.seed,
|
|
778
|
-
});
|
|
779
|
-
|
|
780
|
-
const ringStagingBuffer = ringSlot?.stagingTokens ?? null;
|
|
781
|
-
const stagingSize = state.finitenessBuffer ? 20 : 4;
|
|
782
|
-
const stagingBuffer = ringStagingBuffer && ringStagingBuffer.size >= stagingSize
|
|
783
|
-
? ringStagingBuffer
|
|
784
|
-
: getReusableSampleReadbackBuffer(state, device, stagingSize);
|
|
785
|
-
const ownsStagingBuffer = false;
|
|
786
|
-
const ownsSampleOutputBuffer = !ringTokensBuffer || sampleOutputBuffer !== ringTokensBuffer;
|
|
787
|
-
|
|
788
|
-
const isPreAllocated = isOwnedDecodeBuffer(hiddenStates, decodeHiddenBuffer, decodeAltBuffer);
|
|
789
|
-
const encoder = recorder.getEncoder();
|
|
790
|
-
encoder.copyBufferToBuffer(sampleOutputBuffer, 0, stagingBuffer, 0, 4);
|
|
791
|
-
if (state.finitenessBuffer) {
|
|
792
|
-
encoder.copyBufferToBuffer(state.finitenessBuffer, 0, stagingBuffer, 4, 16);
|
|
793
|
-
}
|
|
794
|
-
|
|
795
|
-
const readbackMode = executionPlan?.readbackMode;
|
|
796
|
-
const isOverlapped = readbackMode === 'overlapped';
|
|
797
|
-
|
|
798
|
-
// In overlapped mode, advance ring BEFORE submit so the GPU's next copy
|
|
799
|
-
// target is a fresh slot while we read the current one.
|
|
800
|
-
if (isOverlapped) {
|
|
801
|
-
ring?.advance();
|
|
802
|
-
}
|
|
803
|
-
|
|
804
|
-
if (!allowReadback('pipeline.decode.sample')) {
|
|
805
|
-
throw new Error('[Pipeline] GPU readback disabled for sampling');
|
|
806
|
-
}
|
|
807
|
-
|
|
808
|
-
const submitStart = performance.now();
|
|
809
|
-
recorder.submit({ cleanup: 'deferred' });
|
|
810
|
-
const submitWaitMs = performance.now() - submitStart;
|
|
811
|
-
|
|
812
|
-
const readbackStart = performance.now();
|
|
813
|
-
const readbackResult = await readSampledTokenFromStagingBuffer(stagingBuffer, {
|
|
814
|
-
ownsStagingBuffer,
|
|
815
|
-
hasFinitenessBuffer: Boolean(state.finitenessBuffer),
|
|
816
|
-
ring: isOverlapped ? null : ring,
|
|
817
|
-
cleanupRecorder: recorder,
|
|
818
|
-
});
|
|
819
|
-
const readbackWaitMs = performance.now() - readbackStart;
|
|
820
|
-
|
|
821
|
-
state.stats.singleTokenSubmitWaitMs = (state.stats.singleTokenSubmitWaitMs ?? 0) + submitWaitMs;
|
|
822
|
-
state.stats.singleTokenReadbackWaitMs = (state.stats.singleTokenReadbackWaitMs ?? 0) + readbackWaitMs;
|
|
823
|
-
if (readbackResult.timing) {
|
|
824
|
-
state.stats.singleTokenReadbackMapWaitMs = (state.stats.singleTokenReadbackMapWaitMs ?? 0)
|
|
825
|
-
+ readbackResult.timing.mapWaitMs;
|
|
826
|
-
state.stats.singleTokenReadbackCleanupMs = (state.stats.singleTokenReadbackCleanupMs ?? 0)
|
|
827
|
-
+ readbackResult.timing.cleanupMs;
|
|
828
|
-
state.stats.singleTokenReadbackCopyMs = (state.stats.singleTokenReadbackCopyMs ?? 0)
|
|
829
|
-
+ readbackResult.timing.copyMs;
|
|
830
|
-
}
|
|
831
|
-
|
|
832
|
-
const { nextToken: fusedNextToken, finitenessStatus } = readbackResult;
|
|
833
|
-
|
|
834
|
-
if (finitenessStatus.triggered) {
|
|
835
|
-
releaseBuffer(logitsBuffer);
|
|
836
|
-
if (ownsSampleOutputBuffer) releaseBuffer(sampleOutputBuffer);
|
|
837
|
-
if (!isPreAllocated) releaseBuffer(hiddenStates);
|
|
838
|
-
throw new FinitenessError(`F16 bounds exceeded during generation${finitenessStatus.metadata}`);
|
|
839
|
-
}
|
|
840
|
-
|
|
841
|
-
log.debug('Decode', `Step ${state.decodeStepCount}: token=${fusedNextToken} (vocabSize=${config.vocabSize})`);
|
|
842
|
-
|
|
843
|
-
const invalidToken = fusedNextToken >= config.vocabSize
|
|
844
|
-
|| (padTokenId != null && fusedNextToken === padTokenId)
|
|
845
|
-
|| (padTokenId == null && fusedNextToken === 0);
|
|
846
|
-
if (!invalidToken) {
|
|
847
|
-
schedulePlePrefetchForToken(state, fusedNextToken);
|
|
848
|
-
}
|
|
849
|
-
if (invalidToken) {
|
|
850
|
-
log.warn('Decode', `Suspicious token ${fusedNextToken} (vocabSize=${config.vocabSize}, step=${state.decodeStepCount})`);
|
|
851
|
-
if (allowReadback('pipeline.decode.debug-logits')) {
|
|
852
|
-
const logitsBytes = selectRuleValue('shared', 'dtype', 'bytesFromDtype', { dtype: logitsDtype });
|
|
853
|
-
const logitSample = await readBuffer(logitsBuffer, Math.min(config.vocabSize * logitsBytes, 4096));
|
|
854
|
-
const logitArr = decodeReadback(logitSample, logitsDtype);
|
|
855
|
-
const maxLogit = Math.max(...logitArr);
|
|
856
|
-
const minLogit = Math.min(...logitArr);
|
|
857
|
-
const hasNaN = logitArr.some((v) => Number.isNaN(v));
|
|
858
|
-
const hasInf = logitArr.some((v) => !Number.isFinite(v));
|
|
859
|
-
let argmaxIdx = 0;
|
|
860
|
-
let argmaxVal = logitArr[0];
|
|
861
|
-
for (let i = 1; i < logitArr.length; i++) {
|
|
862
|
-
if (logitArr[i] > argmaxVal) {
|
|
863
|
-
argmaxVal = logitArr[i];
|
|
864
|
-
argmaxIdx = i;
|
|
865
|
-
}
|
|
866
|
-
}
|
|
867
|
-
log.warn('Decode', `Logits: max=${maxLogit.toFixed(4)} at [${argmaxIdx}], min=${minLogit.toFixed(4)}, hasNaN=${hasNaN}, hasInf=${hasInf}`);
|
|
868
|
-
log.warn('Decode', `First 10 logits: ${Array.from(logitSample.slice(0, 10)).map((v) => v.toFixed(4)).join(', ')}`);
|
|
869
|
-
log.warn('Decode', `Logit[0] (pad): ${logitArr[0].toFixed(4)}, Logit[${argmaxIdx}]: ${argmaxVal.toFixed(4)}`);
|
|
870
|
-
}
|
|
871
|
-
}
|
|
872
|
-
|
|
873
|
-
releaseBuffer(logitsBuffer);
|
|
874
|
-
if (ownsSampleOutputBuffer) {
|
|
875
|
-
releaseBuffer(sampleOutputBuffer);
|
|
876
|
-
}
|
|
877
|
-
|
|
878
|
-
if (benchmarkSubmits) {
|
|
879
|
-
logSubmitStats(`Decode step ${state.decodeStepCount} (${config.numLayers} layers, fused)`);
|
|
880
|
-
setTrackSubmits(false);
|
|
881
|
-
}
|
|
882
|
-
|
|
883
|
-
if (opts.profile) {
|
|
884
|
-
const timings = recorder.isProfilingEnabled()
|
|
885
|
-
? await recorder.resolveProfileTimings()
|
|
886
|
-
: null;
|
|
887
|
-
const total = sumProfileTimings(timings);
|
|
888
|
-
if (total !== null) {
|
|
889
|
-
state.stats.gpuTimeDecodeMs = (state.stats.gpuTimeDecodeMs ?? 0) + total;
|
|
890
|
-
}
|
|
891
|
-
if (timings || opts.executionObserver) {
|
|
892
|
-
recordDecodeProfileStep(state, {
|
|
893
|
-
step: state.decodeStepCount,
|
|
894
|
-
timings,
|
|
895
|
-
totalMs: total ?? undefined,
|
|
896
|
-
recorderStats: recorder.getStats(),
|
|
897
|
-
});
|
|
898
|
-
if (!opts.executionObserver && shouldLogProfileStep(state, state.decodeStepCount)) {
|
|
899
|
-
log.warn('Profile', `Decode step ${state.decodeStepCount}:`);
|
|
900
|
-
log.warn('Profile', CommandRecorder.formatProfileReport(timings));
|
|
901
|
-
}
|
|
902
|
-
}
|
|
903
|
-
}
|
|
904
|
-
|
|
905
|
-
if (invalidToken) {
|
|
906
|
-
state.disableFusedDecode = true;
|
|
907
|
-
log.warn('Decode', 'Fused sampling produced invalid token; falling back to CPU sampling.');
|
|
908
|
-
const fallbackLogits = await computeLogits(
|
|
909
|
-
hiddenStates,
|
|
910
|
-
numTokens,
|
|
911
|
-
helpers.getLogitsWeights(),
|
|
912
|
-
helpers.getLogitsConfig(),
|
|
913
|
-
state.useGPU,
|
|
914
|
-
state.debugFlags,
|
|
915
|
-
undefined,
|
|
916
|
-
debugCheckBuffer,
|
|
917
|
-
state.runtimeConfig.shared.debug.probes,
|
|
918
|
-
null,
|
|
919
|
-
state.operatorDiagnostics
|
|
920
|
-
);
|
|
921
|
-
applyRepetitionPenalty(fallbackLogits, currentIds, opts.repetitionPenalty);
|
|
922
|
-
const fallbackToken = sample(fallbackLogits, {
|
|
923
|
-
temperature: opts.temperature,
|
|
924
|
-
topP: opts.topP,
|
|
925
|
-
topK: opts.topK,
|
|
926
|
-
padTokenId,
|
|
927
|
-
seed: opts.seed,
|
|
928
|
-
});
|
|
929
|
-
schedulePlePrefetchForToken(state, fallbackToken);
|
|
930
|
-
if (!isPreAllocated) {
|
|
931
|
-
releaseBuffer(hiddenStates);
|
|
932
|
-
}
|
|
933
|
-
state.currentSeqLen++;
|
|
934
|
-
return fallbackToken;
|
|
935
|
-
}
|
|
936
|
-
|
|
937
|
-
if (!isPreAllocated) {
|
|
938
|
-
releaseBuffer(hiddenStates);
|
|
939
|
-
}
|
|
940
|
-
|
|
941
|
-
state.currentSeqLen++;
|
|
942
|
-
const stepWallMs = performance.now() - stepWallStart;
|
|
943
|
-
state.stats.singleTokenOrchestrationMs = (state.stats.singleTokenOrchestrationMs ?? 0)
|
|
944
|
-
+ Math.max(0, stepWallMs - submitWaitMs - readbackWaitMs);
|
|
945
|
-
return fusedNextToken;
|
|
946
|
-
}
|
|
947
|
-
|
|
948
|
-
await submitDecodeRecorderProfile(state, opts, recorder, ' (layers only)');
|
|
949
|
-
|
|
950
|
-
if (benchmarkSubmits) {
|
|
951
|
-
logSubmitStats(`Decode step ${state.decodeStepCount} (${config.numLayers} layers)`);
|
|
952
|
-
setTrackSubmits(false);
|
|
953
|
-
}
|
|
954
|
-
|
|
955
|
-
if (opts.debug && state.decodeStepCount === 1 && isGpuBufferInstance(hiddenStates)) {
|
|
956
|
-
const debugDevice = getDevice();
|
|
957
|
-
if (debugDevice) {
|
|
958
|
-
if (allowReadback('pipeline.decode.debug-hidden')) {
|
|
959
|
-
const debugReadbackSize = state.runtimeConfig.shared.debug.pipeline.readbackSampleSize;
|
|
960
|
-
const sampleSize = Math.min(debugReadbackSize, hiddenStates.size);
|
|
961
|
-
const staging = debugDevice.createBuffer({
|
|
962
|
-
size: sampleSize,
|
|
963
|
-
usage: GPUBufferUsage.COPY_DST | GPUBufferUsage.MAP_READ,
|
|
964
|
-
});
|
|
965
|
-
const enc = debugDevice.createCommandEncoder();
|
|
966
|
-
enc.copyBufferToBuffer(hiddenStates, 0, staging, 0, sampleSize);
|
|
967
|
-
debugDevice.queue.submit([enc.finish()]);
|
|
968
|
-
const data = new Float32Array(await readMappedBufferCopy(staging));
|
|
969
|
-
const nanCount = Array.from(data).filter(x => !Number.isFinite(x)).length;
|
|
970
|
-
const nonZero = Array.from(data).filter(x => Number.isFinite(x) && x !== 0).slice(0, 5);
|
|
971
|
-
log.debug('Decode', `[1] HIDDEN_AFTER_LAYERS: nan=${nanCount}/${data.length}, nonZero=${nonZero.length}, sample=[${nonZero.map(x => x.toFixed(4)).join(', ')}]`);
|
|
972
|
-
}
|
|
973
|
-
}
|
|
974
|
-
}
|
|
975
|
-
|
|
976
|
-
if (useGPUSampling) {
|
|
977
|
-
const logitsResult = await computeLogitsGPU(
|
|
978
|
-
hiddenStates,
|
|
979
|
-
numTokens,
|
|
980
|
-
helpers.getLogitsWeights(),
|
|
981
|
-
helpers.getLogitsConfig(),
|
|
982
|
-
state.debugFlags,
|
|
983
|
-
state.operatorDiagnostics
|
|
984
|
-
);
|
|
985
|
-
if (logitsResult) {
|
|
986
|
-
const { logitsBuffer, vocabSize, logitsDtype } = logitsResult;
|
|
987
|
-
const logitsBytes = selectRuleValue('shared', 'dtype', 'bytesFromDtype', { dtype: logitsDtype });
|
|
988
|
-
const nfReadbackStart = performance.now();
|
|
989
|
-
const logitsData = await readBuffer(logitsBuffer, numTokens * vocabSize * logitsBytes);
|
|
990
|
-
const nfReadbackMs = performance.now() - nfReadbackStart;
|
|
991
|
-
state.stats.singleTokenReadbackWaitMs = (state.stats.singleTokenReadbackWaitMs ?? 0) + nfReadbackMs;
|
|
992
|
-
releaseBuffer(logitsBuffer);
|
|
993
|
-
|
|
994
|
-
const rawLogits = decodeReadback(logitsData, logitsDtype);
|
|
995
|
-
const finalizedLogits = await finalizeLogits(
|
|
996
|
-
rawLogits,
|
|
997
|
-
numTokens,
|
|
998
|
-
vocabSize,
|
|
999
|
-
config.vocabSize,
|
|
1000
|
-
config,
|
|
1001
|
-
state.runtimeConfig.shared.debug.probes,
|
|
1002
|
-
state.operatorDiagnostics
|
|
1003
|
-
);
|
|
1004
|
-
const sampledLogits = extractLastPositionLogits(finalizedLogits, numTokens, config.vocabSize);
|
|
1005
|
-
|
|
1006
|
-
applyRepetitionPenalty(sampledLogits, currentIds, opts.repetitionPenalty);
|
|
1007
|
-
const nextToken = sample(sampledLogits, {
|
|
1008
|
-
temperature: opts.temperature,
|
|
1009
|
-
topP: opts.topP,
|
|
1010
|
-
topK: opts.topK,
|
|
1011
|
-
padTokenId,
|
|
1012
|
-
seed: opts.seed,
|
|
1013
|
-
});
|
|
1014
|
-
|
|
1015
|
-
if (!context.decodeBuffers?.ownsBuffer(hiddenStates)) {
|
|
1016
|
-
releaseBuffer(hiddenStates);
|
|
1017
|
-
}
|
|
1018
|
-
state.currentSeqLen++;
|
|
1019
|
-
const nfStepWallMs = performance.now() - stepWallStart;
|
|
1020
|
-
state.stats.singleTokenOrchestrationMs = (state.stats.singleTokenOrchestrationMs ?? 0)
|
|
1021
|
-
+ Math.max(0, nfStepWallMs - nfReadbackMs);
|
|
1022
|
-
return nextToken;
|
|
1023
|
-
}
|
|
1024
|
-
}
|
|
1025
|
-
|
|
1026
|
-
if (state.finitenessBuffer) {
|
|
1027
|
-
const isInfiniteData = await readBuffer(state.finitenessBuffer, 16);
|
|
1028
|
-
const u32 = new Uint32Array(isInfiniteData.buffer, isInfiniteData.byteOffset, 4);
|
|
1029
|
-
const finitenessStatus = parseFinitenessStatusWords(u32, 0);
|
|
1030
|
-
if (finitenessStatus.triggered) {
|
|
1031
|
-
if (!context.decodeBuffers?.ownsBuffer(hiddenStates)) {
|
|
1032
|
-
releaseBuffer(hiddenStates);
|
|
1033
|
-
}
|
|
1034
|
-
throw new FinitenessError(`F16 bounds exceeded during generation${finitenessStatus.metadata}`);
|
|
1035
|
-
}
|
|
1036
|
-
}
|
|
1037
|
-
|
|
1038
|
-
const logits = await computeLogits(
|
|
1039
|
-
hiddenStates,
|
|
1040
|
-
numTokens,
|
|
1041
|
-
helpers.getLogitsWeights(),
|
|
1042
|
-
helpers.getLogitsConfig(),
|
|
1043
|
-
state.useGPU,
|
|
1044
|
-
state.debugFlags,
|
|
1045
|
-
undefined,
|
|
1046
|
-
debugCheckBuffer,
|
|
1047
|
-
state.runtimeConfig.shared.debug.probes,
|
|
1048
|
-
null,
|
|
1049
|
-
state.operatorDiagnostics
|
|
1050
|
-
);
|
|
1051
|
-
|
|
1052
|
-
if (!context.decodeBuffers?.ownsBuffer(hiddenStates)) {
|
|
1053
|
-
releaseBuffer(hiddenStates);
|
|
1054
|
-
}
|
|
1055
|
-
|
|
1056
|
-
if (isDebugStep) {
|
|
1057
|
-
logitsSanity(logits, `Decode[${state.decodeStepCount}]`, opts.decode);
|
|
1058
|
-
}
|
|
1059
|
-
|
|
1060
|
-
applyRepetitionPenalty(logits, currentIds, opts.repetitionPenalty);
|
|
1061
|
-
const nextToken = sample(logits, {
|
|
1062
|
-
temperature: opts.temperature,
|
|
1063
|
-
topP: opts.topP,
|
|
1064
|
-
topK: opts.topK,
|
|
1065
|
-
padTokenId,
|
|
1066
|
-
seed: opts.seed,
|
|
1067
|
-
});
|
|
1068
|
-
|
|
1069
|
-
state.currentSeqLen++;
|
|
1070
|
-
const cpuStepWallMs = performance.now() - stepWallStart;
|
|
1071
|
-
state.stats.singleTokenOrchestrationMs = (state.stats.singleTokenOrchestrationMs ?? 0) + cpuStepWallMs;
|
|
1072
|
-
return nextToken;
|
|
1073
|
-
}
|
|
1074
|
-
|
|
1075
|
-
export async function decodeStepLogits(state, currentIds, opts, helpers) {
|
|
1076
|
-
const lastToken = currentIds[currentIds.length - 1];
|
|
1077
|
-
const numTokens = 1;
|
|
1078
|
-
const config = state.modelConfig;
|
|
1079
|
-
|
|
1080
|
-
state.decodeStepCount++;
|
|
1081
|
-
const recorder = createDecodeRecorder(state, opts);
|
|
1082
|
-
|
|
1083
|
-
const { hiddenStates, decodeHiddenBuffer, decodeAltBuffer, debugCheckBuffer } = await runDecodeLayers(
|
|
1084
|
-
state,
|
|
1085
|
-
lastToken,
|
|
1086
|
-
opts,
|
|
1087
|
-
{
|
|
1088
|
-
...helpers,
|
|
1089
|
-
buildLayerContext: (ignoredRecorder, isDecode, debugLayers, executionPlan) =>
|
|
1090
|
-
helpers.buildLayerContext(recorder, isDecode, debugLayers, executionPlan),
|
|
1091
|
-
}
|
|
1092
|
-
);
|
|
1093
|
-
|
|
1094
|
-
await submitDecodeRecorderProfile(state, opts, recorder, ' (layers only)');
|
|
1095
|
-
|
|
1096
|
-
let logitsBuffer = null;
|
|
1097
|
-
let logitsDtype = null;
|
|
1098
|
-
let rawVocabSize = config.vocabSize;
|
|
1099
|
-
let logits = null;
|
|
1100
|
-
|
|
1101
|
-
if (state.useGPU && !isCpuWeightBuffer(state.weights.get('lm_head'))) {
|
|
1102
|
-
const logitsResult = await computeLogitsGPU(
|
|
1103
|
-
hiddenStates,
|
|
1104
|
-
numTokens,
|
|
1105
|
-
helpers.getLogitsWeights(),
|
|
1106
|
-
helpers.getLogitsConfig(),
|
|
1107
|
-
state.debugFlags,
|
|
1108
|
-
state.operatorDiagnostics
|
|
1109
|
-
);
|
|
1110
|
-
|
|
1111
|
-
if (logitsResult) {
|
|
1112
|
-
logitsBuffer = logitsResult.logitsBuffer;
|
|
1113
|
-
logitsDtype = logitsResult.logitsDtype;
|
|
1114
|
-
rawVocabSize = logitsResult.vocabSize;
|
|
1115
|
-
|
|
1116
|
-
const logitsBytes = selectRuleValue('shared', 'dtype', 'bytesFromDtype', { dtype: logitsDtype });
|
|
1117
|
-
const logitsData = await readBuffer(logitsBuffer, numTokens * rawVocabSize * logitsBytes);
|
|
1118
|
-
const rawLogits = decodeReadback(logitsData, logitsDtype);
|
|
1119
|
-
const finalized = await finalizeLogits(
|
|
1120
|
-
rawLogits,
|
|
1121
|
-
numTokens,
|
|
1122
|
-
rawVocabSize,
|
|
1123
|
-
config.vocabSize,
|
|
1124
|
-
config,
|
|
1125
|
-
state.runtimeConfig.shared.debug.probes,
|
|
1126
|
-
state.operatorDiagnostics
|
|
1127
|
-
);
|
|
1128
|
-
logits = extractLastPositionLogits(finalized, numTokens, config.vocabSize);
|
|
1129
|
-
}
|
|
1130
|
-
}
|
|
1131
|
-
|
|
1132
|
-
if (!logits) {
|
|
1133
|
-
const rawLogits = await computeLogits(
|
|
1134
|
-
hiddenStates,
|
|
1135
|
-
numTokens,
|
|
1136
|
-
helpers.getLogitsWeights(),
|
|
1137
|
-
helpers.getLogitsConfig(),
|
|
1138
|
-
state.useGPU,
|
|
1139
|
-
state.debugFlags,
|
|
1140
|
-
undefined,
|
|
1141
|
-
debugCheckBuffer,
|
|
1142
|
-
state.runtimeConfig.shared.debug.probes,
|
|
1143
|
-
null,
|
|
1144
|
-
state.operatorDiagnostics
|
|
1145
|
-
);
|
|
1146
|
-
logits = extractLastPositionLogits(rawLogits, numTokens, config.vocabSize);
|
|
1147
|
-
}
|
|
1148
|
-
|
|
1149
|
-
const isPreAllocated = isOwnedDecodeBuffer(hiddenStates, decodeHiddenBuffer, decodeAltBuffer);
|
|
1150
|
-
if (!isPreAllocated) {
|
|
1151
|
-
releaseBuffer(hiddenStates);
|
|
1152
|
-
}
|
|
1153
|
-
|
|
1154
|
-
state.currentSeqLen++;
|
|
1155
|
-
|
|
1156
|
-
return {
|
|
1157
|
-
logits,
|
|
1158
|
-
logitsBuffer,
|
|
1159
|
-
logitsDtype,
|
|
1160
|
-
rawVocabSize,
|
|
1161
|
-
vocabSize: config.vocabSize,
|
|
1162
|
-
};
|
|
1163
|
-
}
|
|
1164
|
-
|
|
1165
246
|
export async function advanceWithToken(state, tokenId, opts, helpers) {
|
|
1166
|
-
state
|
|
247
|
+
advanceDecodeStepCount(state);
|
|
1167
248
|
|
|
1168
249
|
const { hiddenStates, decodeHiddenBuffer, decodeAltBuffer } = await runDecodeLayers(
|
|
1169
250
|
state,
|
|
@@ -1182,7 +263,7 @@ export async function advanceWithToken(state, tokenId, opts, helpers) {
|
|
|
1182
263
|
|
|
1183
264
|
export async function advanceWithTokenAndEmbedding(state, tokenId, opts, helpers, embeddingMode) {
|
|
1184
265
|
|
|
1185
|
-
state
|
|
266
|
+
advanceDecodeStepCount(state);
|
|
1186
267
|
|
|
1187
268
|
const { hiddenStates, decodeHiddenBuffer, decodeAltBuffer } = await runDecodeLayers(
|
|
1188
269
|
state,
|
|
@@ -1195,51 +276,23 @@ export async function advanceWithTokenAndEmbedding(state, tokenId, opts, helpers
|
|
|
1195
276
|
throw new Error('GPU readback disabled; cannot return embedding');
|
|
1196
277
|
}
|
|
1197
278
|
|
|
1198
|
-
const device = getDevice();
|
|
1199
|
-
if (!device) {
|
|
1200
|
-
throw new Error('GPU device not available');
|
|
1201
|
-
}
|
|
1202
|
-
|
|
1203
279
|
const config = state.modelConfig;
|
|
1204
280
|
const activationDtype = getEffectiveActivationDtype(state, opts);
|
|
1205
|
-
const activationBytes = selectRuleValue('shared', 'dtype', 'bytesFromDtype', { dtype: activationDtype });
|
|
1206
281
|
|
|
1207
282
|
let embedding;
|
|
1208
283
|
try {
|
|
1209
|
-
const
|
|
1210
|
-
|
|
1211
|
-
|
|
1212
|
-
|
|
1213
|
-
|
|
1214
|
-
|
|
1215
|
-
let decodedHidden;
|
|
1216
|
-
let stagingMapped = false;
|
|
1217
|
-
try {
|
|
1218
|
-
const enc = device.createCommandEncoder({ label: 'advance_with_embedding_copy' });
|
|
1219
|
-
enc.copyBufferToBuffer(hiddenStates, 0, staging, 0, sampleSize);
|
|
1220
|
-
device.queue.submit([enc.finish()]);
|
|
1221
|
-
|
|
1222
|
-
await staging.mapAsync(GPUMapMode.READ);
|
|
1223
|
-
stagingMapped = true;
|
|
1224
|
-
decodedHidden = decodeReadback(staging.getMappedRange().slice(0), activationDtype);
|
|
1225
|
-
} finally {
|
|
1226
|
-
if (stagingMapped) {
|
|
1227
|
-
staging.unmap();
|
|
1228
|
-
}
|
|
1229
|
-
staging.destroy();
|
|
1230
|
-
}
|
|
1231
|
-
const finalNormWeights = await getFinalNormWeights(state);
|
|
1232
|
-
embedding = extractEmbeddingFromHidden(
|
|
1233
|
-
decodedHidden,
|
|
1234
|
-
1,
|
|
1235
|
-
config.hiddenSize,
|
|
284
|
+
const extracted = await extractEmbeddingFromHiddenGPU({
|
|
285
|
+
hiddenBuffer: hiddenStates,
|
|
286
|
+
activationDtype,
|
|
287
|
+
numTokens: 1,
|
|
288
|
+
hiddenSize: config.hiddenSize,
|
|
1236
289
|
embeddingMode,
|
|
1237
|
-
|
|
290
|
+
finalNorm: state.weights.get('final_norm'),
|
|
291
|
+
finalNormBias: state.weights.get('final_norm_bias') ?? null,
|
|
1238
292
|
config,
|
|
1239
|
-
state.embeddingPostprocessor,
|
|
1240
|
-
|
|
1241
|
-
|
|
1242
|
-
);
|
|
293
|
+
embeddingPostprocessor: state.embeddingPostprocessor,
|
|
294
|
+
});
|
|
295
|
+
embedding = extracted.embedding;
|
|
1243
296
|
} finally {
|
|
1244
297
|
const isPreAllocated = isOwnedDecodeBuffer(hiddenStates, decodeHiddenBuffer, decodeAltBuffer);
|
|
1245
298
|
if (!isPreAllocated) {
|
|
@@ -1337,9 +390,14 @@ export async function generateNTokensGPU(state, startToken, N, currentIds, opts,
|
|
|
1337
390
|
|
|
1338
391
|
state.batchingStats.batchedForwardCalls += 1;
|
|
1339
392
|
const tokensPerInterval = batchSize * readbackInterval;
|
|
393
|
+
const recorderOptions = {
|
|
394
|
+
recordLabels: opts.debug === true || opts.benchmark === true || opts.executionObserver === true,
|
|
395
|
+
recordDispatches: opts.debug === true || opts.executionObserver === true,
|
|
396
|
+
aggregateDispatches: opts.executionObserver === true,
|
|
397
|
+
};
|
|
1340
398
|
const recorder = opts.profile
|
|
1341
|
-
? createProfilingRecorder('batch_decode', device)
|
|
1342
|
-
: createCommandRecorder('batch_decode',
|
|
399
|
+
? createProfilingRecorder('batch_decode', device, recorderOptions)
|
|
400
|
+
: createCommandRecorder('batch_decode', recorderOptions, device);
|
|
1343
401
|
const lmHead = state.weights.get('lm_head');
|
|
1344
402
|
if (lmHead && isCpuWeightBuffer(lmHead)) {
|
|
1345
403
|
throw new Error('[Pipeline] GPU-only decode not supported with CPU-resident LM head.');
|
|
@@ -1545,7 +603,7 @@ export async function generateNTokensGPU(state, startToken, N, currentIds, opts,
|
|
|
1545
603
|
hiddenSize: config.hiddenSize,
|
|
1546
604
|
vocabSize: config.vocabSize,
|
|
1547
605
|
scaleEmbeddings: config.scaleEmbeddings,
|
|
1548
|
-
embeddingScale: config.embeddingScale,
|
|
606
|
+
embeddingScale: config.embeddingScale, embeddingNormalization: config.embeddingNormalization,
|
|
1549
607
|
recorder,
|
|
1550
608
|
outputBuffer: context.decodeBuffers?.getHiddenBuffer() ?? undefined,
|
|
1551
609
|
transpose: state.embeddingTranspose,
|
|
@@ -1670,7 +728,7 @@ export async function generateNTokensGPU(state, startToken, N, currentIds, opts,
|
|
|
1670
728
|
}
|
|
1671
729
|
|
|
1672
730
|
const stopCheck = canUseHotVocabularyBatchDecode
|
|
1673
|
-
? recordCheckHotVocabStop(recorder, {
|
|
731
|
+
? await recordCheckHotVocabStop(recorder, {
|
|
1674
732
|
sampledTokenBuffer: tokensBuffer,
|
|
1675
733
|
nextInputTokenBuffer: pleInputTokensBuffer,
|
|
1676
734
|
hotTokenIndexMapBuffer: pleHotVocabularyRuntime.hotTokenIndexMapBuffer,
|
|
@@ -1682,7 +740,7 @@ export async function generateNTokensGPU(state, startToken, N, currentIds, opts,
|
|
|
1682
740
|
currentPos,
|
|
1683
741
|
})
|
|
1684
742
|
: useGpuStopFlags
|
|
1685
|
-
? recordCheckStop(recorder, {
|
|
743
|
+
? await recordCheckStop(recorder, {
|
|
1686
744
|
sampledTokenBuffer: tokensBuffer,
|
|
1687
745
|
shouldStopBuffer: stopBuffer,
|
|
1688
746
|
tokenIndex: outputIndex,
|