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
|
@@ -6,12 +6,10 @@ import { createTensor } from '../tensor.js';
|
|
|
6
6
|
import { KernelBase } from './kernel-base.js';
|
|
7
7
|
import { TILE_SIZES } from './constants.js';
|
|
8
8
|
import { getKernelThresholds, padToQ4KBlock } from '../../config/schema/index.js';
|
|
9
|
-
import {
|
|
10
|
-
|
|
11
|
-
|
|
12
|
-
|
|
13
|
-
hasRequiredFeatures,
|
|
14
|
-
} from './utils.js';
|
|
9
|
+
import { createUniformBufferWithView } from './uniform-utils.js';
|
|
10
|
+
import { getKernelConfig } from './kernel-configs.js';
|
|
11
|
+
import { getPipelineBindGroupLayout } from './pipeline-cache.js';
|
|
12
|
+
import { hasRequiredFeatures } from './feature-check.js';
|
|
15
13
|
import { dispatchIndirect, recordDispatchIndirect } from './dispatch.js';
|
|
16
14
|
import { releaseUniformBuffer } from '../uniform-cache.js';
|
|
17
15
|
import { log, trace } from '../../debug/index.js';
|
|
@@ -20,22 +18,10 @@ import { selectRuleValue as selectKernelRuleValue } from './rule-registry.js';
|
|
|
20
18
|
import { selectRuleValue as selectSharedRuleValue } from '../../rules/rule-registry.js';
|
|
21
19
|
import { logKernelSelectionOnce } from '../kernel-selection-log.js';
|
|
22
20
|
import { getRuntimeConfig } from '../../config/runtime.js';
|
|
21
|
+
import { getRequiredVariantMaxKVLen, resolveAttentionPlan } from './attention/plan.js';
|
|
22
|
+
import { assertAttentionBindGroupBuffer, dispatchAttentionKernel, executeAttention, executeAttentionBDPA, getPageTableFallbackBuffer, releaseAttentionUniform, resolveAttentionExecution } from './attention/executors.js';
|
|
23
23
|
|
|
24
24
|
// Track if we've logged the attention tier selection (avoid spam)
|
|
25
|
-
let loggedAttentionTier = false;
|
|
26
|
-
|
|
27
|
-
function getRequiredVariantMaxKVLen(operation, variant, errorLabel) {
|
|
28
|
-
const config = getKernelConfig(operation, variant);
|
|
29
|
-
const maxKVLen = config.variantMetadata?.maxKVLen;
|
|
30
|
-
if (!Number.isFinite(maxKVLen)) {
|
|
31
|
-
throw new Error(`Kernel config missing ${errorLabel} maxKVLen`);
|
|
32
|
-
}
|
|
33
|
-
return maxKVLen;
|
|
34
|
-
}
|
|
35
|
-
|
|
36
|
-
function getChunkedMaxKVLen() {
|
|
37
|
-
return getRequiredVariantMaxKVLen('attention', 'decode_chunked_f16kv', 'attention.decode_chunked_f16kv');
|
|
38
|
-
}
|
|
39
25
|
|
|
40
26
|
function getTieredMaxKVLen() {
|
|
41
27
|
return getRequiredVariantMaxKVLen('attention_tiered', 'decode_tiered_f16', 'attention_tiered.decode_tiered_f16');
|
|
@@ -57,107 +43,12 @@ function getContiguousQuantMaxKVLen() {
|
|
|
57
43
|
);
|
|
58
44
|
}
|
|
59
45
|
|
|
60
|
-
|
|
61
|
-
let kvLenFallbackBuffer = null;
|
|
62
|
-
let kvLenFallbackBufferEpoch = -1;
|
|
63
|
-
const U32_BYTES = Uint32Array.BYTES_PER_ELEMENT;
|
|
64
|
-
const F32_BYTES = Float32Array.BYTES_PER_ELEMENT;
|
|
65
|
-
|
|
66
|
-
|
|
67
|
-
function getKvLenFallbackBuffer(device) {
|
|
68
|
-
const epoch = getDeviceEpoch();
|
|
69
|
-
if (!kvLenFallbackBuffer || kvLenFallbackBufferEpoch !== epoch) {
|
|
70
|
-
kvLenFallbackBuffer = device.createBuffer({
|
|
71
|
-
label: 'attention_kv_len_fallback',
|
|
72
|
-
size: U32_BYTES,
|
|
73
|
-
usage: GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST,
|
|
74
|
-
});
|
|
75
|
-
device.queue.writeBuffer(kvLenFallbackBuffer, 0, new Uint32Array([0]));
|
|
76
|
-
kvLenFallbackBufferEpoch = epoch;
|
|
77
|
-
}
|
|
78
|
-
return kvLenFallbackBuffer;
|
|
79
|
-
}
|
|
80
|
-
|
|
81
|
-
let pageTableFallbackBuffer = null;
|
|
82
|
-
let pageTableFallbackBufferEpoch = -1;
|
|
83
|
-
|
|
84
|
-
function getPageTableFallbackBuffer(device) {
|
|
85
|
-
const epoch = getDeviceEpoch();
|
|
86
|
-
if (!pageTableFallbackBuffer || pageTableFallbackBufferEpoch !== epoch) {
|
|
87
|
-
pageTableFallbackBuffer = device.createBuffer({
|
|
88
|
-
label: 'attention_page_table_fallback',
|
|
89
|
-
size: U32_BYTES,
|
|
90
|
-
usage: GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST,
|
|
91
|
-
});
|
|
92
|
-
device.queue.writeBuffer(pageTableFallbackBuffer, 0, new Uint32Array([0]));
|
|
93
|
-
pageTableFallbackBufferEpoch = epoch;
|
|
94
|
-
}
|
|
95
|
-
return pageTableFallbackBuffer;
|
|
96
|
-
}
|
|
97
|
-
|
|
98
|
-
class AttentionKernel extends KernelBase {
|
|
99
|
-
|
|
100
|
-
async getPipeline(variant, constants = null) {
|
|
101
|
-
return this.getPipelineFor('attention', variant, null, constants);
|
|
102
|
-
}
|
|
103
|
-
|
|
104
|
-
|
|
105
|
-
dispatch(
|
|
106
|
-
pipeline,
|
|
107
|
-
bindGroup,
|
|
108
|
-
workgroups
|
|
109
|
-
) {
|
|
110
|
-
this.dispatchKernel(pipeline, bindGroup, workgroups, 'attention');
|
|
111
|
-
}
|
|
112
|
-
|
|
113
|
-
|
|
114
|
-
record(
|
|
115
|
-
recorder,
|
|
116
|
-
pipeline,
|
|
117
|
-
bindGroup,
|
|
118
|
-
workgroups
|
|
119
|
-
) {
|
|
120
|
-
this.recordKernel(recorder, pipeline, bindGroup, workgroups, 'attention');
|
|
121
|
-
}
|
|
122
|
-
}
|
|
123
|
-
|
|
124
|
-
function resolveOnlineDecodePipelineConstants(variant, headDim, options = {}) {
|
|
125
|
-
if (!variant?.startsWith('decode_online')) {
|
|
126
|
-
return null;
|
|
127
|
-
}
|
|
128
|
-
const policy = getRuntimeConfig()?.inference?.session?.attentionDecodeOnline;
|
|
129
|
-
const workgroupSize = policy?.workgroupSize;
|
|
130
|
-
const constants = {};
|
|
131
|
-
if (workgroupSize != null) {
|
|
132
|
-
if (workgroupSize !== 128 && workgroupSize !== 256) {
|
|
133
|
-
throw new Error(`attentionDecodeOnline.workgroupSize must be 128 or 256, got ${workgroupSize}.`);
|
|
134
|
-
}
|
|
135
|
-
if (headDim > workgroupSize * 2) {
|
|
136
|
-
throw new Error(
|
|
137
|
-
`attentionDecodeOnline.workgroupSize=${workgroupSize} cannot cover headDim=${headDim}; ` +
|
|
138
|
-
`requires headDim <= ${workgroupSize * 2}.`
|
|
139
|
-
);
|
|
140
|
-
}
|
|
141
|
-
constants.WORKGROUP_SIZE = workgroupSize;
|
|
142
|
-
}
|
|
143
|
-
if (
|
|
144
|
-
variant === 'decode_online_head256_f16kv_output_gate'
|
|
145
|
-
&& policy?.useDirectContiguousKVLayout === true
|
|
146
|
-
&& options.kvLayout === 'contiguous'
|
|
147
|
-
&& options.slidingWindow === 0
|
|
148
|
-
) {
|
|
149
|
-
constants.USE_DIRECT_KV_LAYOUT = true;
|
|
150
|
-
}
|
|
151
|
-
return Object.keys(constants).length > 0 ? constants : null;
|
|
152
|
-
}
|
|
153
|
-
|
|
154
46
|
class AttentionTieredKernel extends KernelBase {
|
|
155
47
|
|
|
156
48
|
async getPipeline(variant) {
|
|
157
49
|
return this.getPipelineFor('attention_tiered', variant);
|
|
158
50
|
}
|
|
159
51
|
|
|
160
|
-
|
|
161
52
|
dispatch(
|
|
162
53
|
pipeline,
|
|
163
54
|
bindGroup,
|
|
@@ -166,7 +57,6 @@ class AttentionTieredKernel extends KernelBase {
|
|
|
166
57
|
this.dispatchKernel(pipeline, bindGroup, workgroups, 'attention_tiered');
|
|
167
58
|
}
|
|
168
59
|
|
|
169
|
-
|
|
170
60
|
record(
|
|
171
61
|
recorder,
|
|
172
62
|
pipeline,
|
|
@@ -183,1478 +73,189 @@ class AttentionTieredQuantKernel extends KernelBase {
|
|
|
183
73
|
return this.getPipelineFor('attention_tiered_quant', variant);
|
|
184
74
|
}
|
|
185
75
|
|
|
186
|
-
|
|
187
|
-
dispatch(
|
|
188
|
-
pipeline,
|
|
189
|
-
bindGroup,
|
|
190
|
-
workgroups
|
|
191
|
-
) {
|
|
192
|
-
this.dispatchKernel(pipeline, bindGroup, workgroups, 'attention_tiered_quant');
|
|
193
|
-
}
|
|
194
|
-
|
|
195
|
-
|
|
196
|
-
record(
|
|
197
|
-
recorder,
|
|
198
|
-
pipeline,
|
|
199
|
-
bindGroup,
|
|
200
|
-
workgroups
|
|
201
|
-
) {
|
|
202
|
-
this.recordKernel(recorder, pipeline, bindGroup, workgroups, 'attention_tiered_quant');
|
|
203
|
-
}
|
|
204
|
-
}
|
|
205
|
-
|
|
206
|
-
class AttentionContiguousQuantKernel extends KernelBase {
|
|
207
|
-
|
|
208
|
-
async getPipeline(variant) {
|
|
209
|
-
return this.getPipelineFor('attention_contiguous_quant', variant);
|
|
210
|
-
}
|
|
211
|
-
|
|
212
|
-
dispatch(
|
|
213
|
-
pipeline,
|
|
214
|
-
bindGroup,
|
|
215
|
-
workgroups
|
|
216
|
-
) {
|
|
217
|
-
this.dispatchKernel(pipeline, bindGroup, workgroups, 'attention_contiguous_quant');
|
|
218
|
-
}
|
|
219
|
-
|
|
220
|
-
record(
|
|
221
|
-
recorder,
|
|
222
|
-
pipeline,
|
|
223
|
-
bindGroup,
|
|
224
|
-
workgroups
|
|
225
|
-
) {
|
|
226
|
-
this.recordKernel(recorder, pipeline, bindGroup, workgroups, 'attention_contiguous_quant');
|
|
227
|
-
}
|
|
228
|
-
}
|
|
229
|
-
|
|
230
|
-
class AttentionBDPAKernel extends KernelBase {
|
|
231
|
-
async getPipeline(variant) {
|
|
232
|
-
return this.getPipelineFor('attention_bdpa', variant);
|
|
233
|
-
}
|
|
234
|
-
|
|
235
76
|
dispatch(
|
|
236
77
|
pipeline,
|
|
237
78
|
bindGroup,
|
|
238
|
-
workgroups
|
|
239
|
-
) {
|
|
240
|
-
this.dispatchKernel(pipeline, bindGroup, workgroups, '
|
|
241
|
-
}
|
|
242
|
-
|
|
243
|
-
record(
|
|
244
|
-
recorder,
|
|
245
|
-
pipeline,
|
|
246
|
-
bindGroup,
|
|
247
|
-
workgroups
|
|
248
|
-
) {
|
|
249
|
-
this.recordKernel(recorder, pipeline, bindGroup, workgroups, 'attention_bdpa');
|
|
250
|
-
}
|
|
251
|
-
}
|
|
252
|
-
|
|
253
|
-
function selectAttentionTier(
|
|
254
|
-
headDim,
|
|
255
|
-
seqLen,
|
|
256
|
-
useF16KV,
|
|
257
|
-
forcedTier,
|
|
258
|
-
sharedLimit,
|
|
259
|
-
caps
|
|
260
|
-
) {
|
|
261
|
-
const isDecode = seqLen === 1;
|
|
262
|
-
const thresholds = getKernelThresholds().attention;
|
|
263
|
-
const largeRequired = useF16KV
|
|
264
|
-
? thresholds.largeSharedF16
|
|
265
|
-
: thresholds.largeSharedF32;
|
|
266
|
-
const canLarge =
|
|
267
|
-
headDim <= thresholds.largeMaxHeadDim &&
|
|
268
|
-
sharedLimit >= largeRequired;
|
|
269
|
-
const smallRequired = useF16KV
|
|
270
|
-
? thresholds.smallSharedF16
|
|
271
|
-
: thresholds.smallSharedF32;
|
|
272
|
-
const canSmall =
|
|
273
|
-
headDim <= thresholds.smallMaxHeadDim &&
|
|
274
|
-
sharedLimit >= smallRequired;
|
|
275
|
-
const canSubgroup =
|
|
276
|
-
caps.hasSubgroups &&
|
|
277
|
-
headDim <= thresholds.subgroupMaxHeadDim &&
|
|
278
|
-
sharedLimit >= thresholds.subgroupShared &&
|
|
279
|
-
isDecode;
|
|
280
|
-
|
|
281
|
-
|
|
282
|
-
let tier = forcedTier;
|
|
283
|
-
let reason = forcedTier ? `forced:${forcedTier}` : '';
|
|
284
|
-
|
|
285
|
-
if (tier === 'tiled_large' && !canLarge) {
|
|
286
|
-
throw new Error(`Requested tiled_large but device doesn't support it (headDim=${headDim}, shared=${sharedLimit}).`);
|
|
287
|
-
}
|
|
288
|
-
if (tier === 'tiled_small' && !canSmall) {
|
|
289
|
-
throw new Error(`Requested tiled_small but device doesn't support it (headDim=${headDim}, shared=${sharedLimit}).`);
|
|
290
|
-
}
|
|
291
|
-
if (tier === 'subgroup' && !canSubgroup) {
|
|
292
|
-
throw new Error(`Requested subgroup attention but device doesn't support it (headDim=${headDim}, shared=${sharedLimit}, subgroups=${caps.hasSubgroups}).`);
|
|
293
|
-
}
|
|
294
|
-
|
|
295
|
-
if (!tier) {
|
|
296
|
-
tier = selectKernelRuleValue('attention', 'tier', { canSubgroup, canLarge, canSmall, isDecode });
|
|
297
|
-
if (!reason) {
|
|
298
|
-
if (canSubgroup) {
|
|
299
|
-
reason = 'subgroup_capable';
|
|
300
|
-
} else if (canLarge) {
|
|
301
|
-
reason = 'tiled_large_capable';
|
|
302
|
-
} else if (canSmall) {
|
|
303
|
-
reason = 'tiled_small_capable';
|
|
304
|
-
} else if (isDecode) {
|
|
305
|
-
reason = 'decode_streaming_fallback';
|
|
306
|
-
} else {
|
|
307
|
-
reason = 'streaming_fallback';
|
|
308
|
-
}
|
|
309
|
-
}
|
|
310
|
-
if (tier === 'subgroup' && !loggedAttentionTier) {
|
|
311
|
-
trace.attn(0, `Using subgroup decode kernel (headDim=${headDim}, hasSubgroups=true)`);
|
|
312
|
-
loggedAttentionTier = true;
|
|
313
|
-
}
|
|
314
|
-
}
|
|
315
|
-
|
|
316
|
-
return { tier, reason };
|
|
317
|
-
}
|
|
318
|
-
|
|
319
|
-
// Track if we've logged chunked kernel selection
|
|
320
|
-
let loggedChunkedKernel = false;
|
|
321
|
-
|
|
322
|
-
|
|
323
|
-
function resolveAttentionVariant(
|
|
324
|
-
tier,
|
|
325
|
-
isDecode,
|
|
326
|
-
useF16KV,
|
|
327
|
-
useF16Q,
|
|
328
|
-
numHeads,
|
|
329
|
-
headDim,
|
|
330
|
-
kvLen,
|
|
331
|
-
isPaged,
|
|
332
|
-
caps,
|
|
333
|
-
sharedLimit
|
|
334
|
-
) {
|
|
335
|
-
const base = selectKernelRuleValue('attention', 'phase', { isDecode });
|
|
336
|
-
const useF16 = useF16KV && useF16Q;
|
|
337
|
-
const suffix = selectKernelRuleValue('attention', 'suffix', { useF16, useF16KV });
|
|
338
|
-
|
|
339
|
-
// Check if chunked kernel is viable:
|
|
340
|
-
// - Decode only (seqLen=1)
|
|
341
|
-
// - F16 KV cache
|
|
342
|
-
// - Large headDim (parallelizes across dimensions)
|
|
343
|
-
// - KV length within shared memory limit (from kernel config)
|
|
344
|
-
const chunkedMaxKVLen = getChunkedMaxKVLen();
|
|
345
|
-
const minHeadDimForChunked = getKernelThresholds().attention.minHeadDimForChunked;
|
|
346
|
-
const canUseChunked = isDecode && useF16KV && headDim >= minHeadDimForChunked && kvLen <= chunkedMaxKVLen;
|
|
347
|
-
const decodeSubgroupMaxKVLen = chunkedMaxKVLen;
|
|
348
|
-
const decodeSubgroupMaxHeadDim = getKernelThresholds().attention.subgroupMaxHeadDim;
|
|
349
|
-
const canUseDecodeSubgroup = isDecode && !useF16KV && !useF16Q && headDim <= decodeSubgroupMaxHeadDim && kvLen <= decodeSubgroupMaxKVLen;
|
|
350
|
-
const canUseDecodeOptimized = isDecode
|
|
351
|
-
&& useF16KV
|
|
352
|
-
&& caps.hasF16
|
|
353
|
-
&& caps.hasSubgroups
|
|
354
|
-
&& headDim <= decodeSubgroupMaxHeadDim
|
|
355
|
-
&& sharedLimit >= getKernelThresholds().attention.subgroupShared;
|
|
356
|
-
const chunkedVariant = selectKernelRuleValue('attention', 'chunkedVariant', { useF16 });
|
|
357
|
-
const pagedVariant = selectKernelRuleValue('attention', 'pagedVariant', { useF16 });
|
|
358
|
-
const optimizedVariant = selectKernelRuleValue('attention', 'optimizedVariant', { useF16 });
|
|
359
|
-
const variant = selectKernelRuleValue(
|
|
360
|
-
'attention',
|
|
361
|
-
'variant',
|
|
362
|
-
{
|
|
363
|
-
tier,
|
|
364
|
-
useF16KV,
|
|
365
|
-
canUseChunked,
|
|
366
|
-
canUseDecodeSubgroup,
|
|
367
|
-
canUseDecodeOptimized,
|
|
368
|
-
base,
|
|
369
|
-
suffix,
|
|
370
|
-
chunkedVariant,
|
|
371
|
-
pagedVariant,
|
|
372
|
-
optimizedVariant,
|
|
373
|
-
isPaged,
|
|
374
|
-
isDecode,
|
|
375
|
-
}
|
|
376
|
-
);
|
|
377
|
-
|
|
378
|
-
if (variant === chunkedVariant && !loggedChunkedKernel) {
|
|
379
|
-
trace.attn(0, `Using chunked decode kernel (headDim=${headDim}, numHeads=${numHeads}, f16kv=${!useF16Q})`);
|
|
380
|
-
loggedChunkedKernel = true;
|
|
381
|
-
}
|
|
382
|
-
|
|
383
|
-
return variant;
|
|
384
|
-
}
|
|
385
|
-
|
|
386
|
-
|
|
387
|
-
function resolveAttentionQueryBlockSize(tier, variant = null) {
|
|
388
|
-
if (variant) {
|
|
389
|
-
const metadataBlockSize = getKernelConfig('attention', variant).variantMetadata?.queryBlockSize;
|
|
390
|
-
if (metadataBlockSize != null) {
|
|
391
|
-
if (!Number.isInteger(metadataBlockSize) || metadataBlockSize <= 0) {
|
|
392
|
-
throw new Error(`Attention kernel "${variant}" has invalid variantMetadata.queryBlockSize=${metadataBlockSize}.`);
|
|
393
|
-
}
|
|
394
|
-
return metadataBlockSize;
|
|
395
|
-
}
|
|
396
|
-
}
|
|
397
|
-
if (tier === 'tiled_large') {
|
|
398
|
-
return TILE_SIZES.ATTENTION_LARGE_BLOCK_SIZE;
|
|
399
|
-
}
|
|
400
|
-
return TILE_SIZES.ATTENTION_SMALL_BLOCK_SIZE;
|
|
401
|
-
}
|
|
402
|
-
|
|
403
|
-
function calculateAttentionWorkgroups(tier, seqLen, numHeads, variant = null) {
|
|
404
|
-
if (tier === 'subgroup') {
|
|
405
|
-
return numHeads;
|
|
406
|
-
}
|
|
407
|
-
if (tier === 'streaming') {
|
|
408
|
-
return seqLen * numHeads;
|
|
409
|
-
}
|
|
410
|
-
const queryBlockSize = resolveAttentionQueryBlockSize(tier, variant);
|
|
411
|
-
return Math.ceil(seqLen / queryBlockSize) * numHeads;
|
|
412
|
-
}
|
|
413
|
-
|
|
414
|
-
|
|
415
|
-
function inferAttentionTierFromVariant(variant) {
|
|
416
|
-
const config = getKernelConfig('attention', variant);
|
|
417
|
-
const tier = config.variantMetadata?.tier;
|
|
418
|
-
if (!tier) {
|
|
419
|
-
throw new Error(`Attention kernel "${variant}" missing variantMetadata.tier in registry.`);
|
|
420
|
-
}
|
|
421
|
-
return tier;
|
|
422
|
-
}
|
|
423
|
-
|
|
424
|
-
|
|
425
|
-
function validateAttentionVariant(
|
|
426
|
-
variant,
|
|
427
|
-
isDecode,
|
|
428
|
-
useF16KV,
|
|
429
|
-
useF16Q,
|
|
430
|
-
caps,
|
|
431
|
-
headDim,
|
|
432
|
-
kvLen,
|
|
433
|
-
sharedLimit
|
|
434
|
-
) {
|
|
435
|
-
const normalized = variant.trim();
|
|
436
|
-
|
|
437
|
-
let config;
|
|
438
|
-
try {
|
|
439
|
-
config = getKernelConfig('attention', normalized);
|
|
440
|
-
} catch {
|
|
441
|
-
throw new Error(`Unknown attention kernel variant "${variant}".`);
|
|
442
|
-
}
|
|
443
|
-
|
|
444
|
-
if (!hasRequiredFeatures(config.requires, caps)) {
|
|
445
|
-
throw new Error(`Attention kernel "${variant}" requires unsupported GPU features.`);
|
|
446
|
-
}
|
|
447
|
-
|
|
448
|
-
const expectsF16KV = normalized.includes('_f16kv');
|
|
449
|
-
const expectsF16 = normalized.includes('_f16') && !expectsF16KV;
|
|
450
|
-
if (expectsF16) {
|
|
451
|
-
if (!(useF16KV && useF16Q)) {
|
|
452
|
-
const kvLabel = selectSharedRuleValue('shared', 'dtype', 'f16OrF32', { useF16: useF16KV });
|
|
453
|
-
const qLabel = selectSharedRuleValue('shared', 'dtype', 'f16OrF32', { useF16: useF16Q });
|
|
454
|
-
throw new Error(`Attention kernel "${variant}" requires f16 Q/K/V but got Q=${qLabel}, KV=${kvLabel}.`);
|
|
455
|
-
}
|
|
456
|
-
} else if (expectsF16KV) {
|
|
457
|
-
if (!useF16KV || useF16Q) {
|
|
458
|
-
const kvLabel = selectSharedRuleValue('shared', 'dtype', 'f16OrF32', { useF16: useF16KV });
|
|
459
|
-
const qLabel = selectSharedRuleValue('shared', 'dtype', 'f16OrF32', { useF16: useF16Q });
|
|
460
|
-
throw new Error(`Attention kernel "${variant}" requires f32 Q with f16 KV but got Q=${qLabel}, KV=${kvLabel}.`);
|
|
461
|
-
}
|
|
462
|
-
} else {
|
|
463
|
-
if (useF16KV || useF16Q) {
|
|
464
|
-
const kvLabel = selectSharedRuleValue('shared', 'dtype', 'f16OrF32', { useF16: useF16KV });
|
|
465
|
-
const qLabel = selectSharedRuleValue('shared', 'dtype', 'f16OrF32', { useF16: useF16Q });
|
|
466
|
-
throw new Error(`Attention kernel "${variant}" requires f32 Q/K/V but got Q=${qLabel}, KV=${kvLabel}.`);
|
|
467
|
-
}
|
|
468
|
-
}
|
|
469
|
-
|
|
470
|
-
const isDecodeVariant = normalized.startsWith('decode');
|
|
471
|
-
const isPrefillVariant = normalized.startsWith('prefill');
|
|
472
|
-
if (isDecode && isPrefillVariant) {
|
|
473
|
-
throw new Error(`Attention kernel "${variant}" is prefill-only but decode requested.`);
|
|
474
|
-
}
|
|
475
|
-
if (!isDecode && isDecodeVariant) {
|
|
476
|
-
throw new Error(`Attention kernel "${variant}" is decode-only but prefill requested.`);
|
|
477
|
-
}
|
|
478
|
-
|
|
479
|
-
const thresholds = getKernelThresholds().attention;
|
|
480
|
-
const chunkedMaxKVLen = getChunkedMaxKVLen();
|
|
481
|
-
const isChunked = normalized.startsWith('decode_chunked');
|
|
482
|
-
if (isChunked) {
|
|
483
|
-
const minHeadDimForChunked = thresholds.minHeadDimForChunked;
|
|
484
|
-
if (headDim < minHeadDimForChunked) {
|
|
485
|
-
throw new Error(`Attention kernel "${variant}" requires headDim >= ${minHeadDimForChunked} but got ${headDim}.`);
|
|
486
|
-
}
|
|
487
|
-
if (kvLen > chunkedMaxKVLen) {
|
|
488
|
-
throw new Error(`Attention kernel "${variant}" requires kvLen <= ${chunkedMaxKVLen} but got ${kvLen}.`);
|
|
489
|
-
}
|
|
490
|
-
}
|
|
491
|
-
|
|
492
|
-
if (normalized === 'decode_subgroup') {
|
|
493
|
-
if (!caps.hasSubgroups) {
|
|
494
|
-
throw new Error(`Attention kernel "${variant}" requires subgroup support.`);
|
|
495
|
-
}
|
|
496
|
-
if (headDim > thresholds.subgroupMaxHeadDim) {
|
|
497
|
-
throw new Error(`Attention kernel "${variant}" requires headDim <= ${thresholds.subgroupMaxHeadDim} but got ${headDim}.`);
|
|
498
|
-
}
|
|
499
|
-
if (kvLen > chunkedMaxKVLen) {
|
|
500
|
-
throw new Error(`Attention kernel "${variant}" requires kvLen <= ${chunkedMaxKVLen} but got ${kvLen}.`);
|
|
501
|
-
}
|
|
502
|
-
if (sharedLimit < thresholds.subgroupShared) {
|
|
503
|
-
throw new Error(`Attention kernel "${variant}" requires shared >= ${thresholds.subgroupShared} but got ${sharedLimit}.`);
|
|
504
|
-
}
|
|
505
|
-
}
|
|
506
|
-
|
|
507
|
-
if (normalized.startsWith('decode_online')) {
|
|
508
|
-
const exactHeadDim = config.variantMetadata?.exactHeadDim;
|
|
509
|
-
const maxHeadDim = config.variantMetadata?.maxHeadDim ?? thresholds.subgroupMaxHeadDim;
|
|
510
|
-
if (!caps.hasSubgroups) {
|
|
511
|
-
throw new Error(`Attention kernel "${variant}" requires subgroup support.`);
|
|
512
|
-
}
|
|
513
|
-
if (Number.isFinite(exactHeadDim) && headDim !== exactHeadDim) {
|
|
514
|
-
throw new Error(`Attention kernel "${variant}" requires headDim == ${exactHeadDim} but got ${headDim}.`);
|
|
515
|
-
}
|
|
516
|
-
if (headDim > maxHeadDim) {
|
|
517
|
-
throw new Error(`Attention kernel "${variant}" requires headDim <= ${maxHeadDim} but got ${headDim}.`);
|
|
518
|
-
}
|
|
519
|
-
if (sharedLimit < thresholds.subgroupShared) {
|
|
520
|
-
throw new Error(`Attention kernel "${variant}" requires shared >= ${thresholds.subgroupShared} but got ${sharedLimit}.`);
|
|
521
|
-
}
|
|
522
|
-
}
|
|
523
|
-
|
|
524
|
-
if (normalized.startsWith('prefill') || normalized.startsWith('decode')) {
|
|
525
|
-
const isSmall = normalized.includes('_small');
|
|
526
|
-
const isStreaming = normalized.includes('_streaming');
|
|
527
|
-
const isTiled = !isStreaming
|
|
528
|
-
&& !normalized.startsWith('decode_subgroup')
|
|
529
|
-
&& !normalized.startsWith('decode_online')
|
|
530
|
-
&& !isChunked;
|
|
531
|
-
if (isTiled) {
|
|
532
|
-
const metadata = config.variantMetadata ?? {};
|
|
533
|
-
const requiredShared = metadata.requiredShared ?? (
|
|
534
|
-
isSmall
|
|
535
|
-
? (useF16KV ? thresholds.smallSharedF16 : thresholds.smallSharedF32)
|
|
536
|
-
: (useF16KV ? thresholds.largeSharedF16 : thresholds.largeSharedF32)
|
|
537
|
-
);
|
|
538
|
-
const maxHeadDim = metadata.maxHeadDim ?? (isSmall ? thresholds.smallMaxHeadDim : thresholds.largeMaxHeadDim);
|
|
539
|
-
const minHeadDim = metadata.minHeadDim ?? 0;
|
|
540
|
-
const exactHeadDim = metadata.exactHeadDim;
|
|
541
|
-
if (Number.isFinite(exactHeadDim) && headDim !== exactHeadDim) {
|
|
542
|
-
throw new Error(`Attention kernel "${variant}" requires headDim == ${exactHeadDim} but got ${headDim}.`);
|
|
543
|
-
}
|
|
544
|
-
if (headDim < minHeadDim) {
|
|
545
|
-
throw new Error(`Attention kernel "${variant}" requires headDim >= ${minHeadDim} but got ${headDim}.`);
|
|
546
|
-
}
|
|
547
|
-
if (headDim > maxHeadDim) {
|
|
548
|
-
throw new Error(`Attention kernel "${variant}" requires headDim <= ${maxHeadDim} but got ${headDim}.`);
|
|
549
|
-
}
|
|
550
|
-
if (sharedLimit < requiredShared) {
|
|
551
|
-
throw new Error(`Attention kernel "${variant}" requires shared >= ${requiredShared} but got ${sharedLimit}.`);
|
|
552
|
-
}
|
|
553
|
-
}
|
|
554
|
-
}
|
|
555
|
-
|
|
556
|
-
return normalized;
|
|
557
|
-
}
|
|
558
|
-
|
|
559
|
-
|
|
560
|
-
function resolveAttentionPlan(
|
|
561
|
-
seqLen,
|
|
562
|
-
kvLen,
|
|
563
|
-
headDim,
|
|
564
|
-
numHeads,
|
|
565
|
-
kvDtype,
|
|
566
|
-
qDtype,
|
|
567
|
-
sharedLimit,
|
|
568
|
-
caps,
|
|
569
|
-
layerIdx,
|
|
570
|
-
isPaged,
|
|
571
|
-
kernelPath
|
|
572
|
-
) {
|
|
573
|
-
const useF16KV = kvDtype === 'f16';
|
|
574
|
-
const useF16Q = qDtype === 'f16';
|
|
575
|
-
const isDecode = seqLen === 1;
|
|
576
|
-
const phase = selectKernelRuleValue('attention', 'phase', { isDecode });
|
|
577
|
-
const pathVariant = getKernelPathAttentionVariant(phase, layerIdx, kernelPath);
|
|
578
|
-
const strictPath = getKernelPathStrict();
|
|
579
|
-
|
|
580
|
-
if (pathVariant) {
|
|
581
|
-
let variantOverride;
|
|
582
|
-
try {
|
|
583
|
-
variantOverride = validateAttentionVariant(
|
|
584
|
-
pathVariant,
|
|
585
|
-
isDecode,
|
|
586
|
-
useF16KV,
|
|
587
|
-
useF16Q,
|
|
588
|
-
caps,
|
|
589
|
-
headDim,
|
|
590
|
-
kvLen,
|
|
591
|
-
sharedLimit
|
|
592
|
-
);
|
|
593
|
-
} catch (error) {
|
|
594
|
-
if (strictPath) {
|
|
595
|
-
throw error;
|
|
596
|
-
}
|
|
597
|
-
const reason = error instanceof Error ? error.message : String(error);
|
|
598
|
-
log.warn(
|
|
599
|
-
'Attention',
|
|
600
|
-
`Kernel path override "${pathVariant}" rejected; falling back to capability selection: ${reason}`
|
|
601
|
-
);
|
|
602
|
-
const adaptiveSelection = selectAttentionTier(headDim, seqLen, useF16KV, null, sharedLimit, caps);
|
|
603
|
-
const adaptiveVariant = resolveAttentionVariant(
|
|
604
|
-
adaptiveSelection.tier,
|
|
605
|
-
isDecode,
|
|
606
|
-
useF16KV,
|
|
607
|
-
useF16Q,
|
|
608
|
-
numHeads,
|
|
609
|
-
headDim,
|
|
610
|
-
kvLen,
|
|
611
|
-
isPaged,
|
|
612
|
-
caps,
|
|
613
|
-
sharedLimit
|
|
614
|
-
);
|
|
615
|
-
const workgroups = calculateAttentionWorkgroups(adaptiveSelection.tier, seqLen, numHeads, adaptiveVariant);
|
|
616
|
-
logKernelSelectionOnce('attention', {
|
|
617
|
-
variant: adaptiveVariant,
|
|
618
|
-
reason: `path_override_fallback:${adaptiveSelection.tier}`,
|
|
619
|
-
});
|
|
620
|
-
return {
|
|
621
|
-
tier: adaptiveSelection.tier,
|
|
622
|
-
variant: adaptiveVariant,
|
|
623
|
-
workgroups,
|
|
624
|
-
useF16KV,
|
|
625
|
-
isDecode,
|
|
626
|
-
};
|
|
627
|
-
}
|
|
628
|
-
let selectionReason = 'path_override';
|
|
629
|
-
|
|
630
|
-
if (!isDecode && variantOverride.startsWith('prefill_streaming') && seqLen <= 64) {
|
|
631
|
-
const adaptivePrefillVariant = variantOverride.endsWith('_f16kv')
|
|
632
|
-
? 'prefill_f16kv'
|
|
633
|
-
: variantOverride.endsWith('_f16')
|
|
634
|
-
? 'prefill_f16'
|
|
635
|
-
: 'prefill';
|
|
636
|
-
try {
|
|
637
|
-
const validatedAdaptive = validateAttentionVariant(
|
|
638
|
-
adaptivePrefillVariant,
|
|
639
|
-
isDecode,
|
|
640
|
-
useF16KV,
|
|
641
|
-
useF16Q,
|
|
642
|
-
caps,
|
|
643
|
-
headDim,
|
|
644
|
-
kvLen,
|
|
645
|
-
sharedLimit
|
|
646
|
-
);
|
|
647
|
-
if (validatedAdaptive !== variantOverride) {
|
|
648
|
-
variantOverride = validatedAdaptive;
|
|
649
|
-
selectionReason = 'path_override_adaptive_prefill';
|
|
650
|
-
}
|
|
651
|
-
} catch {
|
|
652
|
-
// Keep original strict-path variant when adaptive fallback is not valid.
|
|
653
|
-
}
|
|
654
|
-
}
|
|
655
|
-
|
|
656
|
-
const tier = inferAttentionTierFromVariant(variantOverride);
|
|
657
|
-
const workgroups = calculateAttentionWorkgroups(tier, seqLen, numHeads, variantOverride);
|
|
658
|
-
logKernelSelectionOnce('attention', {
|
|
659
|
-
variant: variantOverride,
|
|
660
|
-
reason: `${selectionReason}:${tier}`,
|
|
661
|
-
});
|
|
662
|
-
return { tier, variant: variantOverride, workgroups, useF16KV, isDecode };
|
|
663
|
-
}
|
|
664
|
-
|
|
665
|
-
const selection = selectAttentionTier(headDim, seqLen, useF16KV, null, sharedLimit, caps);
|
|
666
|
-
const tier = selection.tier;
|
|
667
|
-
const variant = resolveAttentionVariant(
|
|
668
|
-
tier,
|
|
669
|
-
isDecode,
|
|
670
|
-
useF16KV,
|
|
671
|
-
useF16Q,
|
|
672
|
-
numHeads,
|
|
673
|
-
headDim,
|
|
674
|
-
kvLen,
|
|
675
|
-
isPaged,
|
|
676
|
-
caps,
|
|
677
|
-
sharedLimit
|
|
678
|
-
);
|
|
679
|
-
const validatedVariant = validateAttentionVariant(
|
|
680
|
-
variant,
|
|
681
|
-
isDecode,
|
|
682
|
-
useF16KV,
|
|
683
|
-
useF16Q,
|
|
684
|
-
caps,
|
|
685
|
-
headDim,
|
|
686
|
-
kvLen,
|
|
687
|
-
sharedLimit
|
|
688
|
-
);
|
|
689
|
-
const workgroups = calculateAttentionWorkgroups(tier, seqLen, numHeads, variant);
|
|
690
|
-
|
|
691
|
-
logKernelSelectionOnce('attention', {
|
|
692
|
-
variant: validatedVariant,
|
|
693
|
-
reason: selection.reason,
|
|
694
|
-
});
|
|
695
|
-
|
|
696
|
-
return { tier, variant: validatedVariant, workgroups, useF16KV, isDecode };
|
|
697
|
-
}
|
|
698
|
-
|
|
699
|
-
export function resolveAttentionPlanForTest(
|
|
700
|
-
seqLen,
|
|
701
|
-
kvLen,
|
|
702
|
-
headDim,
|
|
703
|
-
numHeads,
|
|
704
|
-
kvDtype,
|
|
705
|
-
qDtype,
|
|
706
|
-
sharedLimit,
|
|
707
|
-
caps,
|
|
708
|
-
layerIdx,
|
|
709
|
-
isPaged = false,
|
|
710
|
-
kernelPath = null
|
|
711
|
-
) {
|
|
712
|
-
return resolveAttentionPlan(
|
|
713
|
-
seqLen,
|
|
714
|
-
kvLen,
|
|
715
|
-
headDim,
|
|
716
|
-
numHeads,
|
|
717
|
-
kvDtype,
|
|
718
|
-
qDtype,
|
|
719
|
-
sharedLimit,
|
|
720
|
-
caps,
|
|
721
|
-
layerIdx,
|
|
722
|
-
isPaged,
|
|
723
|
-
kernelPath
|
|
724
|
-
);
|
|
725
|
-
}
|
|
726
|
-
|
|
727
|
-
|
|
728
|
-
function createAttentionUniformBuffer(
|
|
729
|
-
device,
|
|
730
|
-
recorder,
|
|
731
|
-
params
|
|
732
|
-
) {
|
|
733
|
-
return createUniformBufferWithView(
|
|
734
|
-
'attention_uniforms',
|
|
735
|
-
80,
|
|
736
|
-
(view) => {
|
|
737
|
-
view.setUint32(0, params.numHeads, true);
|
|
738
|
-
view.setUint32(4, params.numKVHeads, true);
|
|
739
|
-
view.setUint32(8, params.headDim, true);
|
|
740
|
-
view.setUint32(12, params.kvLen, true);
|
|
741
|
-
view.setUint32(16, params.seqLen, true);
|
|
742
|
-
view.setFloat32(20, params.scale, true);
|
|
743
|
-
view.setUint32(24, params.causal ? 1 : 0, true);
|
|
744
|
-
view.setUint32(28, params.startPos, true);
|
|
745
|
-
view.setFloat32(32, params.attnSoftcap, true); // Gemma 2: 50.0, 0 = disabled
|
|
746
|
-
view.setUint32(36, params.slidingWindow, true); // Sliding window size, 0 = disabled
|
|
747
|
-
view.setUint32(40, params.kvLenSource, true); // 0 = uniform kvLen, 1 = buffer
|
|
748
|
-
view.setUint32(44, params.kvStart ?? 0, true);
|
|
749
|
-
view.setUint32(48, params.pageSize ?? 0, true);
|
|
750
|
-
view.setUint32(52, params.kvLayout ?? 0, true);
|
|
751
|
-
view.setUint32(56, params.bidirectionalSpanStart ?? 0, true);
|
|
752
|
-
view.setUint32(60, params.bidirectionalSpanLength ?? 0, true);
|
|
753
|
-
view.setUint32(64, 0, true);
|
|
754
|
-
view.setUint32(68, 0, true);
|
|
755
|
-
view.setUint32(72, 0, true);
|
|
756
|
-
},
|
|
757
|
-
recorder,
|
|
758
|
-
device
|
|
759
|
-
);
|
|
760
|
-
}
|
|
761
|
-
|
|
762
|
-
function createTieredAttentionUniformBuffer(
|
|
763
|
-
device,
|
|
764
|
-
recorder,
|
|
765
|
-
params
|
|
766
|
-
) {
|
|
767
|
-
return createUniformBufferWithView(
|
|
768
|
-
'attention_tiered_uniforms',
|
|
769
|
-
80,
|
|
770
|
-
(view) => {
|
|
771
|
-
view.setUint32(0, params.numHeads, true);
|
|
772
|
-
view.setUint32(4, params.numKVHeads, true);
|
|
773
|
-
view.setUint32(8, params.headDim, true);
|
|
774
|
-
view.setUint32(12, params.coldLen, true);
|
|
775
|
-
view.setUint32(16, params.hotLen, true);
|
|
776
|
-
view.setUint32(20, params.seqLen, true);
|
|
777
|
-
view.setFloat32(24, params.scale, true);
|
|
778
|
-
view.setUint32(28, params.causal ? 1 : 0, true);
|
|
779
|
-
view.setUint32(32, params.startPos, true);
|
|
780
|
-
view.setFloat32(36, params.attnSoftcap, true);
|
|
781
|
-
view.setUint32(40, params.slidingWindow, true);
|
|
782
|
-
view.setUint32(44, params.hotWindow, true);
|
|
783
|
-
view.setUint32(48, params.hotStart, true);
|
|
784
|
-
view.setUint32(52, params.coldPageSize, true);
|
|
785
|
-
view.setUint32(56, params.coldLayout ?? 0, true);
|
|
786
|
-
view.setUint32(60, params.hotLayout ?? 1, true);
|
|
787
|
-
view.setUint32(64, 0, true);
|
|
788
|
-
},
|
|
789
|
-
recorder,
|
|
790
|
-
device
|
|
791
|
-
);
|
|
792
|
-
}
|
|
793
|
-
|
|
794
|
-
function createTieredQuantAttentionUniformBuffer(
|
|
795
|
-
device,
|
|
796
|
-
recorder,
|
|
797
|
-
params
|
|
798
|
-
) {
|
|
799
|
-
return createUniformBufferWithView(
|
|
800
|
-
'attention_tiered_quant_uniforms',
|
|
801
|
-
64,
|
|
802
|
-
(view) => {
|
|
803
|
-
view.setUint32(0, params.numHeads, true);
|
|
804
|
-
view.setUint32(4, params.numKVHeads, true);
|
|
805
|
-
view.setUint32(8, params.headDim, true);
|
|
806
|
-
view.setUint32(12, params.coldLen, true);
|
|
807
|
-
view.setUint32(16, params.hotLen, true);
|
|
808
|
-
view.setUint32(20, params.seqLen, true);
|
|
809
|
-
view.setFloat32(24, params.scale, true);
|
|
810
|
-
view.setUint32(28, params.causal ? 1 : 0, true);
|
|
811
|
-
view.setUint32(32, params.startPos, true);
|
|
812
|
-
view.setFloat32(36, params.attnSoftcap, true);
|
|
813
|
-
view.setUint32(40, params.slidingWindow, true);
|
|
814
|
-
view.setUint32(44, params.hotWindow, true);
|
|
815
|
-
view.setUint32(48, params.hotStart, true);
|
|
816
|
-
view.setUint32(52, params.packedStride, true);
|
|
817
|
-
view.setUint32(56, 0, true);
|
|
818
|
-
},
|
|
819
|
-
recorder,
|
|
820
|
-
device
|
|
821
|
-
);
|
|
822
|
-
}
|
|
823
|
-
|
|
824
|
-
function createContiguousQuantAttentionUniformBuffer(
|
|
825
|
-
device,
|
|
826
|
-
recorder,
|
|
827
|
-
params
|
|
828
|
-
) {
|
|
829
|
-
const hasProdFields = params.packedStrideMSE != null;
|
|
830
|
-
const size = hasProdFields ? 64 : 48;
|
|
831
|
-
return createUniformBufferWithView(
|
|
832
|
-
'attention_contiguous_quant_uniforms',
|
|
833
|
-
size,
|
|
834
|
-
(view) => {
|
|
835
|
-
view.setUint32(0, params.numHeads, true);
|
|
836
|
-
view.setUint32(4, params.numKVHeads, true);
|
|
837
|
-
view.setUint32(8, params.headDim, true);
|
|
838
|
-
view.setUint32(12, params.kvLen, true);
|
|
839
|
-
view.setUint32(16, params.seqLen, true);
|
|
840
|
-
view.setFloat32(20, params.scale, true);
|
|
841
|
-
view.setUint32(24, params.causal ? 1 : 0, true);
|
|
842
|
-
view.setUint32(28, params.startPos, true);
|
|
843
|
-
view.setFloat32(32, params.attnSoftcap, true);
|
|
844
|
-
view.setUint32(36, params.slidingWindow, true);
|
|
845
|
-
if (hasProdFields) {
|
|
846
|
-
view.setUint32(40, params.packedStrideMSE, true);
|
|
847
|
-
view.setUint32(44, params.packedStrideResidual, true);
|
|
848
|
-
view.setUint32(48, 0, true);
|
|
849
|
-
view.setUint32(52, 0, true);
|
|
850
|
-
view.setUint32(56, 0, true);
|
|
851
|
-
view.setUint32(60, 0, true);
|
|
852
|
-
} else {
|
|
853
|
-
view.setUint32(40, params.packedStride, true);
|
|
854
|
-
view.setUint32(44, 0, true);
|
|
855
|
-
}
|
|
856
|
-
},
|
|
857
|
-
recorder,
|
|
858
|
-
device
|
|
859
|
-
);
|
|
860
|
-
}
|
|
861
|
-
|
|
862
|
-
function createBDPAAttentionUniformBuffer(
|
|
863
|
-
device,
|
|
864
|
-
recorder,
|
|
865
|
-
params
|
|
866
|
-
) {
|
|
867
|
-
return createUniformBufferWithView(
|
|
868
|
-
'attention_bdpa_uniforms',
|
|
869
|
-
64,
|
|
870
|
-
(view) => {
|
|
871
|
-
view.setUint32(0, params.numHeads, true);
|
|
872
|
-
view.setUint32(4, params.numKVHeads, true);
|
|
873
|
-
view.setUint32(8, params.headDim, true);
|
|
874
|
-
view.setUint32(12, params.kvLen, true);
|
|
875
|
-
view.setUint32(16, params.seqLen, true);
|
|
876
|
-
view.setFloat32(20, params.scale, true);
|
|
877
|
-
view.setUint32(24, params.causal ? 1 : 0, true);
|
|
878
|
-
view.setUint32(28, params.startPos, true);
|
|
879
|
-
view.setFloat32(32, params.attnSoftcap, true);
|
|
880
|
-
view.setUint32(36, params.slidingWindow, true);
|
|
881
|
-
view.setUint32(40, 0, true); // padding
|
|
882
|
-
view.setUint32(44, 0, true); // padding
|
|
883
|
-
view.setUint32(48, 0, true); // padding
|
|
884
|
-
view.setUint32(52, 0, true); // padding
|
|
885
|
-
view.setUint32(56, 0, true); // padding
|
|
886
|
-
view.setUint32(60, 0, true); // padding
|
|
887
|
-
},
|
|
888
|
-
recorder,
|
|
889
|
-
device
|
|
890
|
-
);
|
|
891
|
-
}
|
|
892
|
-
|
|
893
|
-
function resolveAttentionExecution(recorder) {
|
|
894
|
-
return {
|
|
895
|
-
recorder: recorder || null,
|
|
896
|
-
device: recorder?.device || getDevice(),
|
|
897
|
-
};
|
|
898
|
-
}
|
|
899
|
-
|
|
900
|
-
function assertAttentionBindGroupBuffer(kernelName, variant, bindingIndex, bindingLabel, buffer, details = []) {
|
|
901
|
-
const isGpuBuffer = buffer && (
|
|
902
|
-
typeof GPUBuffer === 'undefined'
|
|
903
|
-
? true
|
|
904
|
-
: buffer instanceof GPUBuffer
|
|
905
|
-
);
|
|
906
|
-
if (isGpuBuffer) {
|
|
907
|
-
return;
|
|
908
|
-
}
|
|
909
|
-
const detailText = details.filter(Boolean).join(', ');
|
|
910
|
-
throw new Error(
|
|
911
|
-
`[${kernelName}] variant="${variant}" binding ${bindingIndex} "${bindingLabel}" requires a GPUBuffer` +
|
|
912
|
-
(detailText ? ` (${detailText})` : '') +
|
|
913
|
-
'.'
|
|
914
|
-
);
|
|
915
|
-
}
|
|
916
|
-
|
|
917
|
-
function releaseAttentionUniform(execution, uniformBuffer) {
|
|
918
|
-
if (!execution.recorder) {
|
|
919
|
-
releaseUniformBuffer(uniformBuffer);
|
|
920
|
-
}
|
|
921
|
-
}
|
|
922
|
-
|
|
923
|
-
function dispatchAttentionKernel(execution, kernel, pipeline, bindGroup, workgroups) {
|
|
924
|
-
if (execution.recorder) {
|
|
925
|
-
kernel.record(execution.recorder, pipeline, bindGroup, workgroups);
|
|
926
|
-
return;
|
|
927
|
-
}
|
|
928
|
-
kernel.dispatch(pipeline, bindGroup, workgroups);
|
|
929
|
-
}
|
|
930
|
-
|
|
931
|
-
async function executeAttentionBDPA(
|
|
932
|
-
recorder,
|
|
933
|
-
Q,
|
|
934
|
-
basisK,
|
|
935
|
-
basisV,
|
|
936
|
-
pagedK,
|
|
937
|
-
pagedV,
|
|
938
|
-
index,
|
|
939
|
-
numHeads,
|
|
940
|
-
headDim,
|
|
941
|
-
options = {}
|
|
942
|
-
) {
|
|
943
|
-
const execution = resolveAttentionExecution(recorder);
|
|
944
|
-
const {
|
|
945
|
-
seqLen = 1,
|
|
946
|
-
kvLen = seqLen,
|
|
947
|
-
numKVHeads = numHeads,
|
|
948
|
-
scale = 1.0 / Math.sqrt(headDim),
|
|
949
|
-
causal = true,
|
|
950
|
-
startPos = 0,
|
|
951
|
-
outputBuffer = null,
|
|
952
|
-
attnSoftcap = 0,
|
|
953
|
-
slidingWindow = 0,
|
|
954
|
-
ropeCos = null,
|
|
955
|
-
ropeSin = null,
|
|
956
|
-
} = options;
|
|
957
|
-
|
|
958
|
-
if (seqLen !== 1) {
|
|
959
|
-
throw new Error(`BDPA attention currently supports decode only (seqLen=1), got seqLen=${seqLen}.`);
|
|
960
|
-
}
|
|
961
|
-
if (Q.dtype !== 'f16' || basisK.dtype !== 'f16' || basisV.dtype !== 'f16') {
|
|
962
|
-
throw new Error(`BDPA attention requires f16 Q/basis tensors; got Q=${Q.dtype}, basisK=${basisK.dtype}, basisV=${basisV.dtype}.`);
|
|
963
|
-
}
|
|
964
|
-
if (!(ropeCos instanceof GPUBuffer) || !(ropeSin instanceof GPUBuffer)) {
|
|
965
|
-
throw new Error('BDPA attention requires GPU ropeCos/ropeSin buffers.');
|
|
966
|
-
}
|
|
967
|
-
|
|
968
|
-
const variant = 'decode_bdpa_f16';
|
|
969
|
-
const caps = getKernelCapabilities();
|
|
970
|
-
const config = getKernelConfig('attention_bdpa', variant);
|
|
971
|
-
if (!hasRequiredFeatures(config.requires, caps)) {
|
|
972
|
-
throw new Error(`BDPA attention kernel "${variant}" requires unsupported GPU features.`);
|
|
973
|
-
}
|
|
974
|
-
const maxKVLen = config.variantMetadata?.maxKVLen;
|
|
975
|
-
if (Number.isFinite(maxKVLen) && kvLen > maxKVLen) {
|
|
976
|
-
throw new Error(`BDPA attention requires kvLen <= ${maxKVLen} but got ${kvLen}.`);
|
|
977
|
-
}
|
|
978
|
-
|
|
979
|
-
const kernel = new AttentionBDPAKernel(execution.device);
|
|
980
|
-
const pipeline = await kernel.getPipeline(variant);
|
|
981
|
-
|
|
982
|
-
const outputDtype = config.outputDtype;
|
|
983
|
-
if (!outputDtype) {
|
|
984
|
-
throw new Error(`Kernel config missing outputDtype for attention_bdpa variant "${variant}".`);
|
|
985
|
-
}
|
|
986
|
-
const bytesPerElement = outputDtype === 'f16' ? 2 : 4;
|
|
987
|
-
const paddedHiddenSize = padToQ4KBlock(numHeads * headDim);
|
|
988
|
-
const outputSize = seqLen * paddedHiddenSize * bytesPerElement;
|
|
989
|
-
const outputBuf = outputBuffer || acquireBuffer(outputSize, undefined, 'attention_bdpa_output');
|
|
990
|
-
|
|
991
|
-
const uniformBuffer = createBDPAAttentionUniformBuffer(execution.device, execution.recorder, {
|
|
992
|
-
numHeads,
|
|
993
|
-
numKVHeads,
|
|
994
|
-
headDim,
|
|
995
|
-
kvLen,
|
|
996
|
-
seqLen,
|
|
997
|
-
scale,
|
|
998
|
-
causal,
|
|
999
|
-
startPos,
|
|
1000
|
-
attnSoftcap,
|
|
1001
|
-
slidingWindow,
|
|
1002
|
-
});
|
|
1003
|
-
|
|
1004
|
-
assertAttentionBindGroupBuffer('attention_bdpa', variant, 0, 'uniforms', uniformBuffer);
|
|
1005
|
-
assertAttentionBindGroupBuffer('attention_bdpa', variant, 1, 'Q', Q?.buffer, [
|
|
1006
|
-
`QLabel=${Q?.label ?? 'unknown'}`,
|
|
1007
|
-
`QDtype=${Q?.dtype ?? 'unknown'}`,
|
|
1008
|
-
]);
|
|
1009
|
-
assertAttentionBindGroupBuffer('attention_bdpa', variant, 2, 'basisK', basisK?.buffer, [
|
|
1010
|
-
`basisKLabel=${basisK?.label ?? 'unknown'}`,
|
|
1011
|
-
`basisKDtype=${basisK?.dtype ?? 'unknown'}`,
|
|
1012
|
-
]);
|
|
1013
|
-
assertAttentionBindGroupBuffer('attention_bdpa', variant, 3, 'basisV', basisV?.buffer, [
|
|
1014
|
-
`basisVLabel=${basisV?.label ?? 'unknown'}`,
|
|
1015
|
-
`basisVDtype=${basisV?.dtype ?? 'unknown'}`,
|
|
1016
|
-
]);
|
|
1017
|
-
assertAttentionBindGroupBuffer('attention_bdpa', variant, 4, 'pagedK', pagedK);
|
|
1018
|
-
assertAttentionBindGroupBuffer('attention_bdpa', variant, 5, 'pagedV', pagedV);
|
|
1019
|
-
assertAttentionBindGroupBuffer('attention_bdpa', variant, 6, 'index', index);
|
|
1020
|
-
assertAttentionBindGroupBuffer('attention_bdpa', variant, 7, 'ropeCos', ropeCos);
|
|
1021
|
-
assertAttentionBindGroupBuffer('attention_bdpa', variant, 8, 'ropeSin', ropeSin);
|
|
1022
|
-
assertAttentionBindGroupBuffer('attention_bdpa', variant, 9, 'output', outputBuf);
|
|
1023
|
-
|
|
1024
|
-
const bindGroup = execution.device.createBindGroup({
|
|
1025
|
-
label: 'attention_bdpa_bind_group',
|
|
1026
|
-
layout: getPipelineBindGroupLayout(pipeline, 0),
|
|
1027
|
-
entries: [
|
|
1028
|
-
{ binding: 0, resource: { buffer: uniformBuffer } },
|
|
1029
|
-
{ binding: 1, resource: { buffer: Q.buffer } },
|
|
1030
|
-
{ binding: 2, resource: { buffer: basisK.buffer } },
|
|
1031
|
-
{ binding: 3, resource: { buffer: basisV.buffer } },
|
|
1032
|
-
{ binding: 4, resource: { buffer: pagedK } },
|
|
1033
|
-
{ binding: 5, resource: { buffer: pagedV } },
|
|
1034
|
-
{ binding: 6, resource: { buffer: index } },
|
|
1035
|
-
{ binding: 7, resource: { buffer: ropeCos } },
|
|
1036
|
-
{ binding: 8, resource: { buffer: ropeSin } },
|
|
1037
|
-
{ binding: 9, resource: { buffer: outputBuf } },
|
|
1038
|
-
],
|
|
1039
|
-
});
|
|
1040
|
-
|
|
1041
|
-
dispatchAttentionKernel(execution, kernel, pipeline, bindGroup, numHeads);
|
|
1042
|
-
releaseAttentionUniform(execution, uniformBuffer);
|
|
1043
|
-
|
|
1044
|
-
return createTensor(outputBuf, outputDtype, [seqLen, numHeads, headDim], 'attention_bdpa_output');
|
|
1045
|
-
}
|
|
1046
|
-
|
|
1047
|
-
async function executeAttention(
|
|
1048
|
-
recorder,
|
|
1049
|
-
Q,
|
|
1050
|
-
K,
|
|
1051
|
-
V,
|
|
1052
|
-
mask,
|
|
1053
|
-
numHeads,
|
|
1054
|
-
headDim,
|
|
1055
|
-
options = {}
|
|
1056
|
-
) {
|
|
1057
|
-
const execution = resolveAttentionExecution(recorder);
|
|
1058
|
-
const {
|
|
1059
|
-
seqLen = 1,
|
|
1060
|
-
kvLen = seqLen,
|
|
1061
|
-
numKVHeads = numHeads,
|
|
1062
|
-
scale = 1.0 / Math.sqrt(headDim),
|
|
1063
|
-
causal = true,
|
|
1064
|
-
bidirectionalSpanStart = 0,
|
|
1065
|
-
bidirectionalSpanLength = 0,
|
|
1066
|
-
startPos = 0,
|
|
1067
|
-
layerIdx,
|
|
1068
|
-
outputBuffer = null,
|
|
1069
|
-
attnSoftcap = 0,
|
|
1070
|
-
slidingWindow = 0,
|
|
1071
|
-
kvLenBuffer = null,
|
|
1072
|
-
indirectBuffer = null,
|
|
1073
|
-
indirectOffset = 0,
|
|
1074
|
-
kvStart = 0,
|
|
1075
|
-
kvLayout = 'contiguous',
|
|
1076
|
-
kvPageTable = null,
|
|
1077
|
-
kvPageSize = 0,
|
|
1078
|
-
kernelPath = null,
|
|
1079
|
-
outputGate = null,
|
|
1080
|
-
} = options;
|
|
1081
|
-
if (!Number.isFinite(bidirectionalSpanStart) || Math.floor(bidirectionalSpanStart) !== bidirectionalSpanStart || bidirectionalSpanStart < 0) {
|
|
1082
|
-
throw new Error(`Attention bidirectionalSpanStart must be a non-negative integer, got ${bidirectionalSpanStart}.`);
|
|
1083
|
-
}
|
|
1084
|
-
if (!Number.isFinite(bidirectionalSpanLength) || Math.floor(bidirectionalSpanLength) !== bidirectionalSpanLength || bidirectionalSpanLength < 0) {
|
|
1085
|
-
throw new Error(`Attention bidirectionalSpanLength must be a non-negative integer, got ${bidirectionalSpanLength}.`);
|
|
1086
|
-
}
|
|
1087
|
-
if (bidirectionalSpanLength > 0 && (bidirectionalSpanStart + bidirectionalSpanLength) > (kvStart + kvLen)) {
|
|
1088
|
-
throw new Error(
|
|
1089
|
-
`Attention bidirectional span [${bidirectionalSpanStart}, ${bidirectionalSpanStart + bidirectionalSpanLength}) ` +
|
|
1090
|
-
`exceeds KV extent [${kvStart}, ${kvStart + kvLen}).`
|
|
1091
|
-
);
|
|
1092
|
-
}
|
|
1093
|
-
|
|
1094
|
-
// ORT-style single-pass flash attention (adapted from microsoft/onnxruntime
|
|
1095
|
-
// flash_attention.wgsl.template). Single kernel, no reduce pass — each WG
|
|
1096
|
-
// handles one (head, query-tile) and processes all K online. Gated by
|
|
1097
|
-
// useOrtFlashPrefill; takes precedence over the split+reduce flash path.
|
|
1098
|
-
if (
|
|
1099
|
-
options.useOrtFlashPrefill === true
|
|
1100
|
-
&& headDim === FLASH_HEAD_DIM
|
|
1101
|
-
&& seqLen > 1
|
|
1102
|
-
&& kvLayout === 'contiguous'
|
|
1103
|
-
&& bidirectionalSpanLength === 0
|
|
1104
|
-
&& indirectBuffer == null
|
|
1105
|
-
&& K?.dtype === 'f16'
|
|
1106
|
-
&& V?.dtype === 'f16'
|
|
1107
|
-
) {
|
|
1108
|
-
return executeOrtFlashAttentionPrefill(recorder, Q, K, V, numHeads, headDim, {
|
|
1109
|
-
seqLen,
|
|
1110
|
-
kvLen,
|
|
1111
|
-
numKVHeads,
|
|
1112
|
-
scale,
|
|
1113
|
-
causal,
|
|
1114
|
-
startPos,
|
|
1115
|
-
outputBuffer,
|
|
1116
|
-
attnSoftcap,
|
|
1117
|
-
slidingWindow,
|
|
1118
|
-
kvLenBuffer,
|
|
1119
|
-
kvStart,
|
|
1120
|
-
kvLayout,
|
|
1121
|
-
kvPageTable,
|
|
1122
|
-
kvPageSize,
|
|
1123
|
-
});
|
|
1124
|
-
}
|
|
1125
|
-
|
|
1126
|
-
// Flash-attention prefill path: raises RDNA3 occupancy via KV-axis workgroup
|
|
1127
|
-
// splitting + online-softmax reduction. Gated by options.useFlashPrefill so
|
|
1128
|
-
// callers opt in deliberately (runtime config flag). Conservative conditions:
|
|
1129
|
-
// head_dim=256, prefill (seqLen>1), contiguous KV, no bidirectional span.
|
|
1130
|
-
if (!globalThis.__DOPPLER_FLASH_TRACE2__ && options.useFlashPrefill === true) {
|
|
1131
|
-
globalThis.__DOPPLER_FLASH_TRACE2__ = true;
|
|
1132
|
-
if (typeof process !== 'undefined' && process?.stderr?.write) process.stderr.write('[FLASH2] useFlash=' + options.useFlashPrefill
|
|
1133
|
-
+ ' hd=' + headDim + '/' + FLASH_HEAD_DIM
|
|
1134
|
-
+ ' seq=' + seqLen + ' kvL=' + kvLayout
|
|
1135
|
-
+ ' biDir=' + bidirectionalSpanLength
|
|
1136
|
-
+ ' indirect=' + (indirectBuffer == null ? 'n' : 'y')
|
|
1137
|
-
+ ' K=' + K?.dtype + ' V=' + V?.dtype + '\n');
|
|
1138
|
-
}
|
|
1139
|
-
if (
|
|
1140
|
-
options.useFlashPrefill === true
|
|
1141
|
-
&& headDim === FLASH_HEAD_DIM
|
|
1142
|
-
&& seqLen > 1
|
|
1143
|
-
&& kvLayout === 'contiguous'
|
|
1144
|
-
&& bidirectionalSpanLength === 0
|
|
1145
|
-
&& indirectBuffer == null
|
|
1146
|
-
&& K?.dtype === 'f16'
|
|
1147
|
-
&& V?.dtype === 'f16'
|
|
1148
|
-
) {
|
|
1149
|
-
return executeFlashAttentionPrefill(recorder, Q, K, V, numHeads, headDim, {
|
|
1150
|
-
seqLen,
|
|
1151
|
-
kvLen,
|
|
1152
|
-
numKVHeads,
|
|
1153
|
-
scale,
|
|
1154
|
-
causal,
|
|
1155
|
-
startPos,
|
|
1156
|
-
outputBuffer,
|
|
1157
|
-
attnSoftcap,
|
|
1158
|
-
slidingWindow,
|
|
1159
|
-
kvLenBuffer,
|
|
1160
|
-
kvStart,
|
|
1161
|
-
kvLayout,
|
|
1162
|
-
kvPageTable,
|
|
1163
|
-
kvPageSize,
|
|
1164
|
-
});
|
|
1165
|
-
}
|
|
1166
|
-
|
|
1167
|
-
const limits = getDeviceLimits();
|
|
1168
|
-
const sharedLimit = limits?.maxComputeWorkgroupStorageSize ?? Infinity;
|
|
1169
|
-
const caps = getKernelCapabilities();
|
|
1170
|
-
|
|
1171
|
-
const kvDtype = K.dtype;
|
|
1172
|
-
const qDtype = Q.dtype;
|
|
1173
|
-
const isPaged = kvLayout === 'paged';
|
|
1174
|
-
const plan = resolveAttentionPlan(
|
|
1175
|
-
seqLen,
|
|
1176
|
-
kvLen,
|
|
1177
|
-
headDim,
|
|
1178
|
-
numHeads,
|
|
1179
|
-
kvDtype,
|
|
1180
|
-
qDtype,
|
|
1181
|
-
sharedLimit,
|
|
1182
|
-
caps,
|
|
1183
|
-
layerIdx,
|
|
1184
|
-
isPaged,
|
|
1185
|
-
kernelPath
|
|
1186
|
-
);
|
|
1187
|
-
|
|
1188
|
-
if (execution.recorder) {
|
|
1189
|
-
trace.attn(0, `recordAttention: isDecode=${plan.isDecode}, tier=${plan.tier}, variant=${plan.variant}, seqLen=${seqLen}, kvLen=${kvLen}, numHeads=${numHeads}, headDim=${headDim}, useF16KV=${plan.useF16KV}`);
|
|
1190
|
-
}
|
|
1191
|
-
|
|
1192
|
-
const kernel = new AttentionKernel(execution.device);
|
|
1193
|
-
const pipelineConstants = resolveOnlineDecodePipelineConstants(plan.variant, headDim, {
|
|
1194
|
-
kvLayout,
|
|
1195
|
-
slidingWindow,
|
|
1196
|
-
});
|
|
1197
|
-
const usesOutputGateFusion = plan.variant === 'decode_online_head256_f16kv_output_gate';
|
|
1198
|
-
const pipeline = await kernel.getPipeline(plan.variant, pipelineConstants);
|
|
1199
|
-
|
|
1200
|
-
const outputConfig = getKernelConfig('attention', plan.variant);
|
|
1201
|
-
const outputDtype = outputConfig.outputDtype;
|
|
1202
|
-
if (!outputDtype) {
|
|
1203
|
-
if (execution.recorder) {
|
|
1204
|
-
throw new Error(`Kernel config missing outputDtype for attention variant "${plan.variant}".`);
|
|
1205
|
-
}
|
|
1206
|
-
throw new Error(`[Attention] outputDtype is required for variant "${plan.variant}".`);
|
|
1207
|
-
}
|
|
1208
|
-
const bytesPerElement = outputDtype === 'f16' ? 2 : 4;
|
|
1209
|
-
const paddedHiddenSize = padToQ4KBlock(numHeads * headDim);
|
|
1210
|
-
const outputSize = seqLen * paddedHiddenSize * bytesPerElement;
|
|
1211
|
-
const outputBuf = outputBuffer || acquireBuffer(outputSize, undefined, 'attention_output');
|
|
1212
|
-
|
|
1213
|
-
const uniformBuffer = createAttentionUniformBuffer(execution.device, execution.recorder, {
|
|
1214
|
-
numHeads,
|
|
1215
|
-
numKVHeads,
|
|
1216
|
-
headDim,
|
|
1217
|
-
kvLen,
|
|
1218
|
-
seqLen,
|
|
1219
|
-
scale,
|
|
1220
|
-
causal,
|
|
1221
|
-
startPos,
|
|
1222
|
-
attnSoftcap,
|
|
1223
|
-
slidingWindow,
|
|
1224
|
-
kvLenSource: kvLenBuffer ? 1 : 0,
|
|
1225
|
-
kvStart,
|
|
1226
|
-
pageSize: kvPageSize,
|
|
1227
|
-
kvLayout: kvLayout === 'paged' ? 2 : (kvLayout === 'ring' ? 1 : 0),
|
|
1228
|
-
bidirectionalSpanStart,
|
|
1229
|
-
bidirectionalSpanLength,
|
|
1230
|
-
});
|
|
1231
|
-
|
|
1232
|
-
const kvLenBinding = kvLenBuffer || getKvLenFallbackBuffer(execution.device);
|
|
1233
|
-
const pageTableBinding = kvPageTable || getPageTableFallbackBuffer(execution.device);
|
|
1234
|
-
const outputGateBinding = usesOutputGateFusion ? outputGate?.buffer : null;
|
|
1235
|
-
if (usesOutputGateFusion) {
|
|
1236
|
-
if (outputGate?.dtype !== 'f32') {
|
|
1237
|
-
throw new Error(`[Attention] outputGate fusion requires f32 gate tensor; got ${String(outputGate?.dtype)}.`);
|
|
1238
|
-
}
|
|
1239
|
-
const requiredGateBytes = numHeads * headDim * F32_BYTES;
|
|
1240
|
-
const actualGateBytes = outputGate?.buffer?.size;
|
|
1241
|
-
if (Number.isFinite(actualGateBytes) && actualGateBytes < requiredGateBytes) {
|
|
1242
|
-
throw new Error(
|
|
1243
|
-
`[Attention] outputGate fusion requires at least ${requiredGateBytes} gate bytes; got ${actualGateBytes}.`
|
|
1244
|
-
);
|
|
1245
|
-
}
|
|
1246
|
-
}
|
|
1247
|
-
assertAttentionBindGroupBuffer('attention', plan.variant, 0, 'uniforms', uniformBuffer);
|
|
1248
|
-
assertAttentionBindGroupBuffer('attention', plan.variant, 1, 'Q', Q?.buffer, [
|
|
1249
|
-
`QLabel=${Q?.label ?? 'unknown'}`,
|
|
1250
|
-
`QDtype=${Q?.dtype ?? 'unknown'}`,
|
|
1251
|
-
]);
|
|
1252
|
-
assertAttentionBindGroupBuffer('attention', plan.variant, 2, 'K', K?.buffer, [
|
|
1253
|
-
`KLabel=${K?.label ?? 'unknown'}`,
|
|
1254
|
-
`KDtype=${K?.dtype ?? 'unknown'}`,
|
|
1255
|
-
]);
|
|
1256
|
-
assertAttentionBindGroupBuffer('attention', plan.variant, 3, 'V', V?.buffer, [
|
|
1257
|
-
`VLabel=${V?.label ?? 'unknown'}`,
|
|
1258
|
-
`VDtype=${V?.dtype ?? 'unknown'}`,
|
|
1259
|
-
]);
|
|
1260
|
-
assertAttentionBindGroupBuffer('attention', plan.variant, 4, 'output', outputBuf);
|
|
1261
|
-
assertAttentionBindGroupBuffer('attention', plan.variant, 5, 'kvLen', kvLenBinding);
|
|
1262
|
-
assertAttentionBindGroupBuffer('attention', plan.variant, 6, 'pageTable', pageTableBinding, [
|
|
1263
|
-
`kvLayout=${kvLayout}`,
|
|
1264
|
-
]);
|
|
1265
|
-
if (usesOutputGateFusion) {
|
|
1266
|
-
assertAttentionBindGroupBuffer('attention', plan.variant, 7, 'outputGate', outputGateBinding, [
|
|
1267
|
-
`useOutputGateFusion=${usesOutputGateFusion}`,
|
|
1268
|
-
]);
|
|
1269
|
-
}
|
|
1270
|
-
const bindGroupEntries = [
|
|
1271
|
-
{ binding: 0, resource: { buffer: uniformBuffer } },
|
|
1272
|
-
{ binding: 1, resource: { buffer: Q.buffer } },
|
|
1273
|
-
{ binding: 2, resource: { buffer: K.buffer } },
|
|
1274
|
-
{ binding: 3, resource: { buffer: V.buffer } },
|
|
1275
|
-
{ binding: 4, resource: { buffer: outputBuf } },
|
|
1276
|
-
{ binding: 5, resource: { buffer: kvLenBinding } },
|
|
1277
|
-
{ binding: 6, resource: { buffer: pageTableBinding } },
|
|
1278
|
-
];
|
|
1279
|
-
if (usesOutputGateFusion) {
|
|
1280
|
-
bindGroupEntries.push({ binding: 7, resource: { buffer: outputGateBinding } });
|
|
1281
|
-
}
|
|
1282
|
-
const bindGroup = execution.device.createBindGroup({
|
|
1283
|
-
label: 'attention_bind_group',
|
|
1284
|
-
layout: getPipelineBindGroupLayout(pipeline, 0),
|
|
1285
|
-
entries: bindGroupEntries,
|
|
1286
|
-
});
|
|
1287
|
-
|
|
1288
|
-
if (!indirectBuffer && limits && plan.workgroups > limits.maxComputeWorkgroupsPerDimension) {
|
|
1289
|
-
throw new Error(
|
|
1290
|
-
`Attention dispatch requires ${plan.workgroups} workgroups but device limit is ` +
|
|
1291
|
-
`${limits.maxComputeWorkgroupsPerDimension}. Reduce prompt length or use streaming attention.`
|
|
1292
|
-
);
|
|
79
|
+
workgroups
|
|
80
|
+
) {
|
|
81
|
+
this.dispatchKernel(pipeline, bindGroup, workgroups, 'attention_tiered_quant');
|
|
1293
82
|
}
|
|
1294
83
|
|
|
1295
|
-
|
|
1296
|
-
|
|
1297
|
-
|
|
1298
|
-
|
|
1299
|
-
|
|
1300
|
-
|
|
1301
|
-
|
|
1302
|
-
dispatchAttentionKernel(execution, kernel, pipeline, bindGroup, plan.workgroups);
|
|
84
|
+
record(
|
|
85
|
+
recorder,
|
|
86
|
+
pipeline,
|
|
87
|
+
bindGroup,
|
|
88
|
+
workgroups
|
|
89
|
+
) {
|
|
90
|
+
this.recordKernel(recorder, pipeline, bindGroup, workgroups, 'attention_tiered_quant');
|
|
1303
91
|
}
|
|
1304
|
-
|
|
1305
|
-
releaseAttentionUniform(execution, uniformBuffer);
|
|
1306
|
-
|
|
1307
|
-
const outputTensor = createTensor(outputBuf, outputDtype, [seqLen, numHeads, headDim], 'attention_output');
|
|
1308
|
-
outputTensor.outputGateFused = usesOutputGateFusion;
|
|
1309
|
-
return outputTensor;
|
|
1310
92
|
}
|
|
1311
93
|
|
|
1312
|
-
|
|
1313
|
-
// Flash-attention prefill path (head_dim = 256, f16 KV)
|
|
1314
|
-
// -----------------------------------------------------------------------------
|
|
1315
|
-
// Two-pass kernel to raise RDNA3 occupancy: pass 1 processes one KV slice per
|
|
1316
|
-
// workgroup and writes per-split (acc, m, l) partials; pass 2 merges across
|
|
1317
|
-
// splits with online softmax. Single recorder, two dispatches — queue order
|
|
1318
|
-
// handles the read-after-write between passes.
|
|
1319
|
-
|
|
1320
|
-
const FLASH_BLOCK_SIZE = 32;
|
|
1321
|
-
const FLASH_HEAD_DIM = 256;
|
|
1322
|
-
const FLASH_HEAD_DIM_VECS = 64;
|
|
1323
|
-
const FLASH_REDUCE_WG = 64;
|
|
1324
|
-
|
|
1325
|
-
let flashPrefillKernel = null;
|
|
1326
|
-
let flashReduceKernel = null;
|
|
94
|
+
class AttentionContiguousQuantKernel extends KernelBase {
|
|
1327
95
|
|
|
1328
|
-
class FlashAttentionPrefillKernel extends KernelBase {
|
|
1329
96
|
async getPipeline(variant) {
|
|
1330
|
-
return this.getPipelineFor('
|
|
1331
|
-
}
|
|
1332
|
-
record(recorder, pipeline, bindGroup, workgroups) {
|
|
1333
|
-
this.recordKernel(recorder, pipeline, bindGroup, workgroups, 'attention');
|
|
97
|
+
return this.getPipelineFor('attention_contiguous_quant', variant);
|
|
1334
98
|
}
|
|
1335
|
-
|
|
1336
|
-
|
|
99
|
+
|
|
100
|
+
dispatch(
|
|
101
|
+
pipeline,
|
|
102
|
+
bindGroup,
|
|
103
|
+
workgroups
|
|
104
|
+
) {
|
|
105
|
+
this.dispatchKernel(pipeline, bindGroup, workgroups, 'attention_contiguous_quant');
|
|
1337
106
|
}
|
|
1338
|
-
}
|
|
1339
107
|
|
|
1340
|
-
|
|
1341
|
-
|
|
1342
|
-
|
|
108
|
+
record(
|
|
109
|
+
recorder,
|
|
110
|
+
pipeline,
|
|
111
|
+
bindGroup,
|
|
112
|
+
workgroups
|
|
113
|
+
) {
|
|
114
|
+
this.recordKernel(recorder, pipeline, bindGroup, workgroups, 'attention_contiguous_quant');
|
|
1343
115
|
}
|
|
1344
|
-
return flashPrefillKernel;
|
|
1345
116
|
}
|
|
1346
117
|
|
|
1347
|
-
|
|
1348
|
-
|
|
1349
|
-
|
|
1350
|
-
|
|
1351
|
-
|
|
118
|
+
// Track if we've logged chunked kernel selection
|
|
119
|
+
|
|
120
|
+
export function resolveAttentionPlanForTest(
|
|
121
|
+
seqLen,
|
|
122
|
+
kvLen,
|
|
123
|
+
headDim,
|
|
124
|
+
numHeads,
|
|
125
|
+
kvDtype,
|
|
126
|
+
qDtype,
|
|
127
|
+
sharedLimit,
|
|
128
|
+
caps,
|
|
129
|
+
layerIdx,
|
|
130
|
+
isPaged = false,
|
|
131
|
+
kernelPath = null
|
|
132
|
+
) {
|
|
133
|
+
return resolveAttentionPlan(
|
|
134
|
+
seqLen,
|
|
135
|
+
kvLen,
|
|
136
|
+
headDim,
|
|
137
|
+
numHeads,
|
|
138
|
+
kvDtype,
|
|
139
|
+
qDtype,
|
|
140
|
+
sharedLimit,
|
|
141
|
+
caps,
|
|
142
|
+
layerIdx,
|
|
143
|
+
isPaged,
|
|
144
|
+
kernelPath
|
|
145
|
+
);
|
|
1352
146
|
}
|
|
1353
147
|
|
|
1354
|
-
function
|
|
1355
|
-
|
|
1356
|
-
|
|
148
|
+
function createTieredAttentionUniformBuffer(
|
|
149
|
+
device,
|
|
150
|
+
recorder,
|
|
151
|
+
params
|
|
152
|
+
) {
|
|
1357
153
|
return createUniformBufferWithView(
|
|
1358
|
-
'
|
|
1359
|
-
|
|
154
|
+
'attention_tiered_uniforms',
|
|
155
|
+
80,
|
|
1360
156
|
(view) => {
|
|
1361
157
|
view.setUint32(0, params.numHeads, true);
|
|
1362
158
|
view.setUint32(4, params.numKVHeads, true);
|
|
1363
159
|
view.setUint32(8, params.headDim, true);
|
|
1364
|
-
view.setUint32(12, params.
|
|
1365
|
-
view.setUint32(16, params.
|
|
1366
|
-
view.
|
|
1367
|
-
view.
|
|
1368
|
-
view.setUint32(28, params.
|
|
1369
|
-
view.
|
|
1370
|
-
view.
|
|
1371
|
-
view.setUint32(40, params.
|
|
1372
|
-
view.setUint32(44, params.
|
|
1373
|
-
view.setUint32(48, params.
|
|
1374
|
-
view.setUint32(52, params.
|
|
1375
|
-
view.setUint32(56, params.
|
|
1376
|
-
view.setUint32(60,
|
|
160
|
+
view.setUint32(12, params.coldLen, true);
|
|
161
|
+
view.setUint32(16, params.hotLen, true);
|
|
162
|
+
view.setUint32(20, params.seqLen, true);
|
|
163
|
+
view.setFloat32(24, params.scale, true);
|
|
164
|
+
view.setUint32(28, params.causal ? 1 : 0, true);
|
|
165
|
+
view.setUint32(32, params.startPos, true);
|
|
166
|
+
view.setFloat32(36, params.attnSoftcap, true);
|
|
167
|
+
view.setUint32(40, params.slidingWindow, true);
|
|
168
|
+
view.setUint32(44, params.hotWindow, true);
|
|
169
|
+
view.setUint32(48, params.hotStart, true);
|
|
170
|
+
view.setUint32(52, params.coldPageSize, true);
|
|
171
|
+
view.setUint32(56, params.coldLayout ?? 0, true);
|
|
172
|
+
view.setUint32(60, params.hotLayout ?? 1, true);
|
|
173
|
+
view.setUint32(64, 0, true);
|
|
1377
174
|
},
|
|
1378
175
|
recorder,
|
|
1379
176
|
device
|
|
1380
177
|
);
|
|
1381
178
|
}
|
|
1382
179
|
|
|
1383
|
-
function
|
|
180
|
+
function createTieredQuantAttentionUniformBuffer(
|
|
181
|
+
device,
|
|
182
|
+
recorder,
|
|
183
|
+
params
|
|
184
|
+
) {
|
|
1384
185
|
return createUniformBufferWithView(
|
|
1385
|
-
'
|
|
1386
|
-
|
|
186
|
+
'attention_tiered_quant_uniforms',
|
|
187
|
+
64,
|
|
1387
188
|
(view) => {
|
|
1388
189
|
view.setUint32(0, params.numHeads, true);
|
|
1389
|
-
view.setUint32(4, params.
|
|
1390
|
-
view.setUint32(8, params.
|
|
1391
|
-
view.setUint32(12,
|
|
190
|
+
view.setUint32(4, params.numKVHeads, true);
|
|
191
|
+
view.setUint32(8, params.headDim, true);
|
|
192
|
+
view.setUint32(12, params.coldLen, true);
|
|
193
|
+
view.setUint32(16, params.hotLen, true);
|
|
194
|
+
view.setUint32(20, params.seqLen, true);
|
|
195
|
+
view.setFloat32(24, params.scale, true);
|
|
196
|
+
view.setUint32(28, params.causal ? 1 : 0, true);
|
|
197
|
+
view.setUint32(32, params.startPos, true);
|
|
198
|
+
view.setFloat32(36, params.attnSoftcap, true);
|
|
199
|
+
view.setUint32(40, params.slidingWindow, true);
|
|
200
|
+
view.setUint32(44, params.hotWindow, true);
|
|
201
|
+
view.setUint32(48, params.hotStart, true);
|
|
202
|
+
view.setUint32(52, params.packedStride, true);
|
|
203
|
+
view.setUint32(56, 0, true);
|
|
1392
204
|
},
|
|
1393
205
|
recorder,
|
|
1394
206
|
device
|
|
1395
207
|
);
|
|
1396
208
|
}
|
|
1397
209
|
|
|
1398
|
-
function
|
|
1399
|
-
|
|
1400
|
-
|
|
1401
|
-
|
|
1402
|
-
|
|
1403
|
-
|
|
1404
|
-
|
|
1405
|
-
|
|
1406
|
-
|
|
1407
|
-
|
|
1408
|
-
|
|
1409
|
-
|
|
1410
|
-
|
|
1411
|
-
|
|
1412
|
-
|
|
1413
|
-
|
|
1414
|
-
|
|
1415
|
-
|
|
1416
|
-
|
|
1417
|
-
|
|
1418
|
-
|
|
1419
|
-
|
|
1420
|
-
|
|
1421
|
-
|
|
1422
|
-
|
|
1423
|
-
|
|
1424
|
-
|
|
1425
|
-
|
|
1426
|
-
|
|
1427
|
-
|
|
1428
|
-
|
|
1429
|
-
} = options;
|
|
1430
|
-
|
|
1431
|
-
if (kvLayout !== 'contiguous') {
|
|
1432
|
-
throw new Error(`[FlashAttention] kvLayout must be "contiguous", got "${kvLayout}".`);
|
|
1433
|
-
}
|
|
1434
|
-
|
|
1435
|
-
const device = execution.device;
|
|
1436
|
-
const numQueryBlocks = Math.max(1, Math.ceil(seqLen / FLASH_BLOCK_SIZE));
|
|
1437
|
-
const numKvSplits = chooseFlashNumKvSplits(kvLen);
|
|
1438
|
-
const singleSplit = numKvSplits === 1;
|
|
1439
|
-
|
|
1440
|
-
// Final output buffer (always allocated; bound to slot 4 on the single-split
|
|
1441
|
-
// fast path where the kernel writes normalised output directly).
|
|
1442
|
-
const paddedHiddenSize = padToQ4KBlock(numHeads * headDim);
|
|
1443
|
-
const outputSize = seqLen * paddedHiddenSize * 4;
|
|
1444
|
-
const outputBuf = outputBuffer || acquireBuffer(outputSize, undefined, 'attention_flash_output');
|
|
1445
|
-
|
|
1446
|
-
// Intermediate buffers for the multi-split path. On the single-split fast
|
|
1447
|
-
// path we bypass the reduce pass entirely and bind the output buffer in
|
|
1448
|
-
// slot 4 (partial_acc slot) while m/l bindings get tiny stub buffers — the
|
|
1449
|
-
// kernel skips writes to them.
|
|
1450
|
-
const partialAccBytes = singleSplit
|
|
1451
|
-
? 4
|
|
1452
|
-
: numQueryBlocks * numHeads * numKvSplits * FLASH_BLOCK_SIZE * FLASH_HEAD_DIM * 4;
|
|
1453
|
-
const partialStatsBytes = singleSplit
|
|
1454
|
-
? 4
|
|
1455
|
-
: numQueryBlocks * numHeads * numKvSplits * FLASH_BLOCK_SIZE * 4;
|
|
1456
|
-
const partialAcc = singleSplit
|
|
1457
|
-
? outputBuf
|
|
1458
|
-
: acquireBuffer(partialAccBytes, GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST, 'flash_partial_acc');
|
|
1459
|
-
const partialM = acquireBuffer(
|
|
1460
|
-
partialStatsBytes,
|
|
1461
|
-
GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST,
|
|
1462
|
-
'flash_partial_m'
|
|
1463
|
-
);
|
|
1464
|
-
const partialL = acquireBuffer(
|
|
1465
|
-
partialStatsBytes,
|
|
1466
|
-
GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST,
|
|
1467
|
-
'flash_partial_l'
|
|
1468
|
-
);
|
|
1469
|
-
|
|
1470
|
-
// Pass 1 uniforms + dispatch.
|
|
1471
|
-
const flashUniform = createFlashAttentionUniformBuffer(device, execution.recorder, {
|
|
1472
|
-
numHeads,
|
|
1473
|
-
numKVHeads,
|
|
1474
|
-
headDim,
|
|
1475
|
-
kvLen,
|
|
1476
|
-
seqLen,
|
|
1477
|
-
scale,
|
|
1478
|
-
causal,
|
|
1479
|
-
startPos,
|
|
1480
|
-
attnSoftcap,
|
|
1481
|
-
slidingWindow,
|
|
1482
|
-
kvLenSource: kvLenBuffer ? 1 : 0,
|
|
1483
|
-
kvStart,
|
|
1484
|
-
pageSize: kvPageSize,
|
|
1485
|
-
kvLayout: 0, // contiguous only for now
|
|
1486
|
-
numKvSplits,
|
|
1487
|
-
});
|
|
1488
|
-
|
|
1489
|
-
const flashKernel = getFlashPrefillKernel(device);
|
|
1490
|
-
const flashPipeline = await flashKernel.getPipeline('prefill_flash_head256_f16kv');
|
|
1491
|
-
const kvLenBinding = kvLenBuffer || getKvLenFallbackBuffer(device);
|
|
1492
|
-
const pageTableBinding = kvPageTable || getPageTableFallbackBuffer(device);
|
|
1493
|
-
|
|
1494
|
-
const flashBindGroup = device.createBindGroup({
|
|
1495
|
-
label: 'attention_flash_prefill_bg',
|
|
1496
|
-
layout: flashPipeline.getBindGroupLayout(0),
|
|
1497
|
-
entries: [
|
|
1498
|
-
{ binding: 0, resource: { buffer: flashUniform } },
|
|
1499
|
-
{ binding: 1, resource: { buffer: Q.buffer } },
|
|
1500
|
-
{ binding: 2, resource: { buffer: K.buffer } },
|
|
1501
|
-
{ binding: 3, resource: { buffer: V.buffer } },
|
|
1502
|
-
{ binding: 4, resource: { buffer: partialAcc } },
|
|
1503
|
-
{ binding: 5, resource: { buffer: partialM } },
|
|
1504
|
-
{ binding: 6, resource: { buffer: partialL } },
|
|
1505
|
-
{ binding: 7, resource: { buffer: kvLenBinding } },
|
|
1506
|
-
{ binding: 8, resource: { buffer: pageTableBinding } },
|
|
1507
|
-
],
|
|
1508
|
-
});
|
|
1509
|
-
|
|
1510
|
-
dispatchAttentionKernel(
|
|
1511
|
-
execution,
|
|
1512
|
-
flashKernel,
|
|
1513
|
-
flashPipeline,
|
|
1514
|
-
flashBindGroup,
|
|
1515
|
-
numQueryBlocks * numHeads * numKvSplits
|
|
1516
|
-
);
|
|
1517
|
-
releaseAttentionUniform(execution, flashUniform);
|
|
1518
|
-
|
|
1519
|
-
// Pass 2 — reduce. Skipped on the single-split fast path where pass 1
|
|
1520
|
-
// already wrote the final normalised output directly to outputBuf.
|
|
1521
|
-
if (!singleSplit) {
|
|
1522
|
-
const reduceUniform = createFlashReduceUniformBuffer(device, execution.recorder, {
|
|
1523
|
-
numHeads,
|
|
1524
|
-
queryLen: seqLen,
|
|
1525
|
-
numKvSplits,
|
|
1526
|
-
});
|
|
1527
|
-
|
|
1528
|
-
const reduceKernel = getFlashReduceKernel(device);
|
|
1529
|
-
const reducePipeline = await reduceKernel.getPipeline('prefill_flash_reduce');
|
|
1530
|
-
|
|
1531
|
-
const reduceBindGroup = device.createBindGroup({
|
|
1532
|
-
label: 'attention_flash_reduce_bg',
|
|
1533
|
-
layout: reducePipeline.getBindGroupLayout(0),
|
|
1534
|
-
entries: [
|
|
1535
|
-
{ binding: 0, resource: { buffer: reduceUniform } },
|
|
1536
|
-
{ binding: 1, resource: { buffer: partialAcc } },
|
|
1537
|
-
{ binding: 2, resource: { buffer: partialM } },
|
|
1538
|
-
{ binding: 3, resource: { buffer: partialL } },
|
|
1539
|
-
{ binding: 4, resource: { buffer: outputBuf } },
|
|
1540
|
-
],
|
|
1541
|
-
});
|
|
1542
|
-
|
|
1543
|
-
const totalQh = seqLen * numHeads;
|
|
1544
|
-
const reduceWgX = Math.ceil(totalQh / FLASH_REDUCE_WG);
|
|
1545
|
-
dispatchAttentionKernel(
|
|
1546
|
-
execution,
|
|
1547
|
-
reduceKernel,
|
|
1548
|
-
reducePipeline,
|
|
1549
|
-
reduceBindGroup,
|
|
1550
|
-
[reduceWgX, FLASH_HEAD_DIM_VECS, 1]
|
|
1551
|
-
);
|
|
1552
|
-
releaseAttentionUniform(execution, reduceUniform);
|
|
1553
|
-
}
|
|
1554
|
-
|
|
1555
|
-
// Release intermediate buffers via the recorder's deferred cleanup so GPU
|
|
1556
|
-
// work completes before they re-enter the pool. On the single-split path
|
|
1557
|
-
// partialAcc IS the output buffer, so we skip it.
|
|
1558
|
-
const intermediates = singleSplit ? [partialM, partialL] : [partialAcc, partialM, partialL];
|
|
1559
|
-
if (execution.recorder) {
|
|
1560
|
-
for (const buf of intermediates) {
|
|
1561
|
-
execution.recorder.trackTemporaryBuffer(buf);
|
|
1562
|
-
}
|
|
1563
|
-
} else {
|
|
1564
|
-
device.queue.onSubmittedWorkDone().then(() => {
|
|
1565
|
-
for (const buf of intermediates) {
|
|
1566
|
-
releaseBuffer(buf);
|
|
210
|
+
function createContiguousQuantAttentionUniformBuffer(
|
|
211
|
+
device,
|
|
212
|
+
recorder,
|
|
213
|
+
params
|
|
214
|
+
) {
|
|
215
|
+
const hasProdFields = params.packedStrideMSE != null;
|
|
216
|
+
const size = hasProdFields ? 64 : 48;
|
|
217
|
+
return createUniformBufferWithView(
|
|
218
|
+
'attention_contiguous_quant_uniforms',
|
|
219
|
+
size,
|
|
220
|
+
(view) => {
|
|
221
|
+
view.setUint32(0, params.numHeads, true);
|
|
222
|
+
view.setUint32(4, params.numKVHeads, true);
|
|
223
|
+
view.setUint32(8, params.headDim, true);
|
|
224
|
+
view.setUint32(12, params.kvLen, true);
|
|
225
|
+
view.setUint32(16, params.seqLen, true);
|
|
226
|
+
view.setFloat32(20, params.scale, true);
|
|
227
|
+
view.setUint32(24, params.causal ? 1 : 0, true);
|
|
228
|
+
view.setUint32(28, params.startPos, true);
|
|
229
|
+
view.setFloat32(32, params.attnSoftcap, true);
|
|
230
|
+
view.setUint32(36, params.slidingWindow, true);
|
|
231
|
+
if (hasProdFields) {
|
|
232
|
+
view.setUint32(40, params.packedStrideMSE, true);
|
|
233
|
+
view.setUint32(44, params.packedStrideResidual, true);
|
|
234
|
+
view.setUint32(48, 0, true);
|
|
235
|
+
view.setUint32(52, 0, true);
|
|
236
|
+
view.setUint32(56, 0, true);
|
|
237
|
+
view.setUint32(60, 0, true);
|
|
238
|
+
} else {
|
|
239
|
+
view.setUint32(40, params.packedStride, true);
|
|
240
|
+
view.setUint32(44, 0, true);
|
|
1567
241
|
}
|
|
1568
|
-
}
|
|
1569
|
-
|
|
1570
|
-
|
|
1571
|
-
|
|
242
|
+
},
|
|
243
|
+
recorder,
|
|
244
|
+
device
|
|
245
|
+
);
|
|
1572
246
|
}
|
|
1573
247
|
|
|
248
|
+
// -----------------------------------------------------------------------------
|
|
249
|
+
// Flash-attention prefill path (head_dim = 256, f16 KV)
|
|
250
|
+
// -----------------------------------------------------------------------------
|
|
251
|
+
// Two-pass kernel to raise RDNA3 occupancy: pass 1 processes one KV slice per
|
|
252
|
+
// workgroup and writes per-split (acc, m, l) partials; pass 2 merges across
|
|
253
|
+
// splits with online softmax. Single recorder, two dispatches — queue order
|
|
254
|
+
// handles the read-after-write between passes.
|
|
255
|
+
|
|
1574
256
|
// Single-pass flash-attention dispatcher (ORT-style). 7-binding contract —
|
|
1575
257
|
// same as attention_head256_f16kv. One kernel launch, no reduce pass.
|
|
1576
258
|
// Workgroups: (num_heads, ceil(seqLen / ORT_FLASH_WG), 1) with ORT_FLASH_WG=64.
|
|
1577
|
-
const ORT_FLASH_WG = 64;
|
|
1578
|
-
|
|
1579
|
-
async function executeOrtFlashAttentionPrefill(recorder, Q, K, V, numHeads, headDim, options = {}) {
|
|
1580
|
-
if (headDim !== FLASH_HEAD_DIM) {
|
|
1581
|
-
throw new Error(`[OrtFlashAttention] headDim must be ${FLASH_HEAD_DIM}, got ${headDim}.`);
|
|
1582
|
-
}
|
|
1583
|
-
const execution = resolveAttentionExecution(recorder);
|
|
1584
|
-
const {
|
|
1585
|
-
seqLen = 1,
|
|
1586
|
-
kvLen = seqLen,
|
|
1587
|
-
numKVHeads = numHeads,
|
|
1588
|
-
scale = 1.0 / Math.sqrt(headDim),
|
|
1589
|
-
causal = true,
|
|
1590
|
-
startPos = 0,
|
|
1591
|
-
outputBuffer = null,
|
|
1592
|
-
attnSoftcap = 0,
|
|
1593
|
-
slidingWindow = 0,
|
|
1594
|
-
kvLenBuffer = null,
|
|
1595
|
-
kvStart = 0,
|
|
1596
|
-
kvLayout = 'contiguous',
|
|
1597
|
-
kvPageTable = null,
|
|
1598
|
-
kvPageSize = 0,
|
|
1599
|
-
} = options;
|
|
1600
|
-
|
|
1601
|
-
if (kvLayout !== 'contiguous') {
|
|
1602
|
-
throw new Error(`[OrtFlashAttention] kvLayout must be "contiguous", got "${kvLayout}".`);
|
|
1603
|
-
}
|
|
1604
|
-
|
|
1605
|
-
const device = execution.device;
|
|
1606
|
-
const paddedHiddenSize = padToQ4KBlock(numHeads * headDim);
|
|
1607
|
-
const outputSize = seqLen * paddedHiddenSize * 4;
|
|
1608
|
-
const outputBuf = outputBuffer || acquireBuffer(outputSize, undefined, 'attention_ort_flash_output');
|
|
1609
|
-
|
|
1610
|
-
const uniform = createAttentionUniformBuffer(device, execution.recorder, {
|
|
1611
|
-
numHeads,
|
|
1612
|
-
numKVHeads,
|
|
1613
|
-
headDim,
|
|
1614
|
-
kvLen,
|
|
1615
|
-
seqLen,
|
|
1616
|
-
scale,
|
|
1617
|
-
causal,
|
|
1618
|
-
startPos,
|
|
1619
|
-
attnSoftcap,
|
|
1620
|
-
slidingWindow,
|
|
1621
|
-
kvLenSource: kvLenBuffer ? 1 : 0,
|
|
1622
|
-
kvStart,
|
|
1623
|
-
pageSize: kvPageSize,
|
|
1624
|
-
kvLayout: 0, // contiguous only
|
|
1625
|
-
});
|
|
1626
|
-
|
|
1627
|
-
const kernel = new AttentionKernel(device);
|
|
1628
|
-
const pipeline = await kernel.getPipeline('prefill_flash_ort_head256_f16kv');
|
|
1629
|
-
const kvLenBinding = kvLenBuffer || getKvLenFallbackBuffer(device);
|
|
1630
|
-
const pageTableBinding = kvPageTable || getPageTableFallbackBuffer(device);
|
|
1631
|
-
|
|
1632
|
-
const bindGroup = execution.device.createBindGroup({
|
|
1633
|
-
label: 'attention_ort_flash_prefill_bg',
|
|
1634
|
-
layout: getPipelineBindGroupLayout(pipeline, 0),
|
|
1635
|
-
entries: [
|
|
1636
|
-
{ binding: 0, resource: { buffer: uniform } },
|
|
1637
|
-
{ binding: 1, resource: { buffer: Q.buffer } },
|
|
1638
|
-
{ binding: 2, resource: { buffer: K.buffer } },
|
|
1639
|
-
{ binding: 3, resource: { buffer: V.buffer } },
|
|
1640
|
-
{ binding: 4, resource: { buffer: outputBuf } },
|
|
1641
|
-
{ binding: 5, resource: { buffer: kvLenBinding } },
|
|
1642
|
-
{ binding: 6, resource: { buffer: pageTableBinding } },
|
|
1643
|
-
],
|
|
1644
|
-
});
|
|
1645
|
-
|
|
1646
|
-
const numSeqTiles = Math.max(1, Math.ceil(seqLen / ORT_FLASH_WG));
|
|
1647
|
-
const workgroups = [numHeads, numSeqTiles, 1];
|
|
1648
|
-
|
|
1649
|
-
if (execution.recorder) {
|
|
1650
|
-
kernel.record(execution.recorder, pipeline, bindGroup, workgroups);
|
|
1651
|
-
} else {
|
|
1652
|
-
kernel.dispatch(pipeline, bindGroup, workgroups);
|
|
1653
|
-
}
|
|
1654
|
-
|
|
1655
|
-
if (uniform) releaseUniformBuffer(uniform);
|
|
1656
|
-
return createTensor(outputBuf, 'f32', [seqLen, numHeads, headDim], 'attention_ort_flash_output');
|
|
1657
|
-
}
|
|
1658
259
|
|
|
1659
260
|
async function executeAttentionTiered(
|
|
1660
261
|
recorder,
|
|
@@ -2111,7 +712,6 @@ export async function recordAttentionTieredQuant(
|
|
|
2111
712
|
);
|
|
2112
713
|
}
|
|
2113
714
|
|
|
2114
|
-
|
|
2115
715
|
// =============================================================================
|
|
2116
716
|
// Contiguous Quantized Attention (TurboQuant for full-attention models)
|
|
2117
717
|
// =============================================================================
|