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
|
@@ -32,7 +32,7 @@ import {
|
|
|
32
32
|
initTokenizerFromManifest,
|
|
33
33
|
assertManifestComputeLaneBinding,
|
|
34
34
|
} from './text/model-load.js';
|
|
35
|
-
import { resolvePerLayerInputsSession } from './text/generator-
|
|
35
|
+
import { resolvePerLayerInputsSession } from './text/generator/session-context.js';
|
|
36
36
|
import { getKernelPathActivationDtype } from '../../config/kernel-path-loader.js';
|
|
37
37
|
import { applyPipelineDebugConfig } from './text/debug-utils.js';
|
|
38
38
|
import { resolveLayerPipeline } from './text/layer-plan.js';
|
|
@@ -51,259 +51,34 @@ import { registerPipeline, getPipelineFactory } from './registry.js';
|
|
|
51
51
|
import { selectRuleValue } from '../../rules/rule-registry.js';
|
|
52
52
|
import { createObservationContext } from '../observation-context.js';
|
|
53
53
|
import { createResolvedRuntimeSession } from './text/resolved-runtime-session.js';
|
|
54
|
-
|
|
55
|
-
|
|
56
|
-
|
|
57
|
-
|
|
58
|
-
|
|
59
|
-
// waiting for the model to finish. Internal checkpoints fire between
|
|
60
|
-
// dispatches; mid-dispatch cancellation is not provided because WebGPU
|
|
61
|
-
// command-buffer cancellation is not exposed by the spec.
|
|
62
|
-
export class AbortError extends Error {
|
|
63
|
-
constructor(message = 'Doppler: aborted') {
|
|
64
|
-
super(message);
|
|
65
|
-
this.name = 'AbortError';
|
|
66
|
-
this.code = 'ABORT_ERR';
|
|
67
|
-
}
|
|
68
|
-
}
|
|
69
|
-
|
|
70
|
-
export function isAbortError(err) {
|
|
71
|
-
return !!err && (err.name === 'AbortError' || err.code === 'ABORT_ERR' || err.code === 20);
|
|
72
|
-
}
|
|
73
|
-
|
|
74
|
-
function assertNotAborted(signal) {
|
|
75
|
-
if (signal && signal.aborted) {
|
|
76
|
-
throw new AbortError(typeof signal.reason === 'string' ? signal.reason : 'Doppler: aborted');
|
|
77
|
-
}
|
|
78
|
-
}
|
|
79
|
-
|
|
80
|
-
function poolSequenceTokenEmbeddings(tokenEmbeddings, tokenIds, hiddenSize, pooling) {
|
|
81
|
-
if (!(tokenEmbeddings instanceof Float32Array)) {
|
|
82
|
-
throw new Error('Sequence token embeddings are unavailable.');
|
|
83
|
-
}
|
|
84
|
-
if (tokenEmbeddings.length !== tokenIds.length * hiddenSize) {
|
|
85
|
-
throw new Error(
|
|
86
|
-
`Sequence embedding shape mismatch: expected ${tokenIds.length * hiddenSize}, got ${tokenEmbeddings.length}.`
|
|
87
|
-
);
|
|
88
|
-
}
|
|
89
|
-
const excluded = new Set(pooling.excludeTokenIds);
|
|
90
|
-
const includedIndices = [];
|
|
91
|
-
for (let index = 0; index < tokenIds.length; index += 1) {
|
|
92
|
-
if (!excluded.has(tokenIds[index])) includedIndices.push(index);
|
|
93
|
-
}
|
|
94
|
-
if (includedIndices.length === 0) {
|
|
95
|
-
throw new Error('Sequence pooling excluded every input token.');
|
|
96
|
-
}
|
|
97
|
-
const pooled = new Float32Array(hiddenSize);
|
|
98
|
-
if (pooling.mode === 'last') {
|
|
99
|
-
const tokenIndex = includedIndices[includedIndices.length - 1];
|
|
100
|
-
const offset = tokenIndex * hiddenSize;
|
|
101
|
-
pooled.set(tokenEmbeddings.subarray(offset, offset + hiddenSize));
|
|
102
|
-
} else {
|
|
103
|
-
for (const tokenIndex of includedIndices) {
|
|
104
|
-
const offset = tokenIndex * hiddenSize;
|
|
105
|
-
for (let column = 0; column < hiddenSize; column += 1) {
|
|
106
|
-
pooled[column] += tokenEmbeddings[offset + column];
|
|
107
|
-
}
|
|
108
|
-
}
|
|
109
|
-
const scale = 1 / includedIndices.length;
|
|
110
|
-
for (let column = 0; column < hiddenSize; column += 1) {
|
|
111
|
-
pooled[column] *= scale;
|
|
112
|
-
}
|
|
113
|
-
}
|
|
114
|
-
return {
|
|
115
|
-
pooled,
|
|
116
|
-
tokenMask: Uint8Array.from(tokenIds, (tokenId) => excluded.has(tokenId) ? 0 : 1),
|
|
117
|
-
includedTokenCount: includedIndices.length,
|
|
118
|
-
};
|
|
119
|
-
}
|
|
54
|
+
import { assertBundledAdapterAuthorized } from '../../config/revocation-policy.js';
|
|
55
|
+
import { assertNotAborted } from './text/abort-contract.js';
|
|
56
|
+
export { AbortError, isAbortError } from './text/abort-contract.js';
|
|
57
|
+
import { destroyMoERouter } from './text/pipeline-load-timing.js';
|
|
58
|
+
export { buildConservativeMultimodalGenerationOptions } from './text/modality-token-contract.js';
|
|
120
59
|
import { initConvLayerState } from './text/ops.js';
|
|
121
60
|
import { destroyPleBufferCache, destroyPleRuntimeCache } from './text/per-layer-inputs.js';
|
|
122
|
-
|
|
123
|
-
|
|
124
|
-
|
|
125
|
-
|
|
126
|
-
|
|
127
|
-
|
|
128
|
-
|
|
129
|
-
|
|
130
|
-
|
|
131
|
-
|
|
132
|
-
|
|
133
|
-
|
|
134
|
-
|
|
135
|
-
|
|
136
|
-
|
|
137
|
-
|
|
138
|
-
|
|
139
|
-
|
|
140
|
-
|
|
141
|
-
|
|
142
|
-
|
|
143
|
-
};
|
|
144
|
-
return wrapped;
|
|
145
|
-
}
|
|
146
|
-
|
|
147
|
-
async function withPipelineLoadPhase(phase, context, run) {
|
|
148
|
-
try {
|
|
149
|
-
return await run();
|
|
150
|
-
} catch (error) {
|
|
151
|
-
if (error?.details?.pipelineLoadPhase) {
|
|
152
|
-
throw error;
|
|
153
|
-
}
|
|
154
|
-
throw createPipelineLoadPhaseError(error, phase, context);
|
|
155
|
-
}
|
|
156
|
-
}
|
|
157
|
-
|
|
158
|
-
function roundPipelineTimingMs(value) {
|
|
159
|
-
return Number.isFinite(value) ? Number(value.toFixed(3)) : null;
|
|
160
|
-
}
|
|
161
|
-
|
|
162
|
-
function createPipelineLoadTiming(modelId) {
|
|
163
|
-
return {
|
|
164
|
-
schemaVersion: 1,
|
|
165
|
-
source: 'doppler-pipeline',
|
|
166
|
-
modelId: typeof modelId === 'string' ? modelId : null,
|
|
167
|
-
status: 'running',
|
|
168
|
-
phasesMs: {
|
|
169
|
-
reset: null,
|
|
170
|
-
configResolution: null,
|
|
171
|
-
kernelWarmup: null,
|
|
172
|
-
tokenizer: null,
|
|
173
|
-
executionSetup: null,
|
|
174
|
-
loadWeights: null,
|
|
175
|
-
rope: null,
|
|
176
|
-
convStates: null,
|
|
177
|
-
},
|
|
178
|
-
details: {
|
|
179
|
-
tokenizer: null,
|
|
180
|
-
},
|
|
181
|
-
totalMs: null,
|
|
182
|
-
};
|
|
183
|
-
}
|
|
184
|
-
|
|
185
|
-
function finishPipelineLoadTimingPhase(loadTiming, phase, startMs) {
|
|
186
|
-
if (!loadTiming?.phasesMs || !phase) return;
|
|
187
|
-
const elapsedMs = roundPipelineTimingMs(performance.now() - startMs);
|
|
188
|
-
const currentMs = loadTiming.phasesMs[phase];
|
|
189
|
-
loadTiming.phasesMs[phase] = Number.isFinite(currentMs)
|
|
190
|
-
? roundPipelineTimingMs(currentMs + elapsedMs)
|
|
191
|
-
: elapsedMs;
|
|
192
|
-
}
|
|
193
|
-
|
|
194
|
-
async function timedPipelineLoadPhase(loadTiming, phase, context, run) {
|
|
195
|
-
const startMs = performance.now();
|
|
196
|
-
try {
|
|
197
|
-
return await withPipelineLoadPhase(phase, context, run);
|
|
198
|
-
} finally {
|
|
199
|
-
finishPipelineLoadTimingPhase(loadTiming, phase, startMs);
|
|
200
|
-
}
|
|
201
|
-
}
|
|
202
|
-
|
|
203
|
-
function resolveSingleSpecialTokenId(tokenizer, tokenText, label) {
|
|
204
|
-
const rawTokenIds = tokenizer?.encode?.(tokenText);
|
|
205
|
-
const tokenIds = Array.isArray(rawTokenIds)
|
|
206
|
-
? rawTokenIds
|
|
207
|
-
: (ArrayBuffer.isView(rawTokenIds) ? Array.from(rawTokenIds) : null);
|
|
208
|
-
if (!Array.isArray(tokenIds) || tokenIds.length !== 1) {
|
|
209
|
-
throw new Error(
|
|
210
|
-
`[Pipeline] transcribeImage: tokenizer must encode ${label} "${tokenText}" as exactly one token.`
|
|
211
|
-
);
|
|
212
|
-
}
|
|
213
|
-
const tokenId = Number(tokenIds[0]);
|
|
214
|
-
if (!Number.isFinite(tokenId) || Math.floor(tokenId) !== tokenId || tokenId < 0) {
|
|
215
|
-
throw new Error(
|
|
216
|
-
`[Pipeline] transcribeImage: tokenizer returned invalid ${label} token id "${tokenIds[0]}".`
|
|
217
|
-
);
|
|
218
|
-
}
|
|
219
|
-
return tokenId;
|
|
220
|
-
}
|
|
221
|
-
|
|
222
|
-
function expandImagePlaceholderTokenIds(tokenIds, imageTokenId, numImageTokens, options = {}) {
|
|
223
|
-
const normalizedTokenIds = Array.isArray(tokenIds)
|
|
224
|
-
? Int32Array.from(tokenIds)
|
|
225
|
-
: (ArrayBuffer.isView(tokenIds) ? Int32Array.from(tokenIds) : null);
|
|
226
|
-
if (!(normalizedTokenIds instanceof Int32Array)) {
|
|
227
|
-
throw new Error(
|
|
228
|
-
'[Pipeline] transcribeImage: tokenizer.encode() must return an array or typed array of token IDs.'
|
|
229
|
-
);
|
|
230
|
-
}
|
|
231
|
-
if (!Number.isFinite(numImageTokens) || Math.floor(numImageTokens) !== numImageTokens || numImageTokens < 1) {
|
|
232
|
-
throw new Error(
|
|
233
|
-
`[Pipeline] transcribeImage: image token span must be a positive integer, got ${numImageTokens}.`
|
|
234
|
-
);
|
|
235
|
-
}
|
|
236
|
-
|
|
237
|
-
let placeholderIndex = -1;
|
|
238
|
-
let placeholderCount = 0;
|
|
239
|
-
for (let i = 0; i < normalizedTokenIds.length; i++) {
|
|
240
|
-
if (normalizedTokenIds[i] !== imageTokenId) continue;
|
|
241
|
-
if (placeholderIndex < 0) {
|
|
242
|
-
placeholderIndex = i;
|
|
243
|
-
}
|
|
244
|
-
placeholderCount++;
|
|
245
|
-
}
|
|
246
|
-
|
|
247
|
-
if (placeholderCount !== 1) {
|
|
248
|
-
throw new Error(
|
|
249
|
-
`[Pipeline] transcribeImage: expected exactly one image_token_id (${imageTokenId}) placeholder ` +
|
|
250
|
-
`from the chat template, got ${placeholderCount}.`
|
|
251
|
-
);
|
|
252
|
-
}
|
|
253
|
-
|
|
254
|
-
const boiTokenId = Number.isInteger(options.boiTokenId) ? options.boiTokenId : null;
|
|
255
|
-
const eoiTokenId = Number.isInteger(options.eoiTokenId) ? options.eoiTokenId : null;
|
|
256
|
-
const prefixExtra = boiTokenId == null ? 0 : 1;
|
|
257
|
-
const suffixExtra = eoiTokenId == null ? 0 : 1;
|
|
258
|
-
const expandedLength = normalizedTokenIds.length - 1 + prefixExtra + numImageTokens + suffixExtra;
|
|
259
|
-
const expanded = new Int32Array(expandedLength);
|
|
260
|
-
expanded.set(normalizedTokenIds.subarray(0, placeholderIndex), 0);
|
|
261
|
-
let writeOffset = placeholderIndex;
|
|
262
|
-
if (boiTokenId != null) {
|
|
263
|
-
expanded[writeOffset++] = boiTokenId;
|
|
264
|
-
}
|
|
265
|
-
expanded.fill(imageTokenId, writeOffset, writeOffset + numImageTokens);
|
|
266
|
-
const imageStartOffset = writeOffset;
|
|
267
|
-
writeOffset += numImageTokens;
|
|
268
|
-
if (eoiTokenId != null) {
|
|
269
|
-
expanded[writeOffset++] = eoiTokenId;
|
|
270
|
-
}
|
|
271
|
-
expanded.set(
|
|
272
|
-
normalizedTokenIds.subarray(placeholderIndex + 1),
|
|
273
|
-
writeOffset
|
|
274
|
-
);
|
|
275
|
-
|
|
276
|
-
return {
|
|
277
|
-
inputIds: expanded,
|
|
278
|
-
imageStartOffset,
|
|
279
|
-
};
|
|
280
|
-
}
|
|
281
|
-
|
|
282
|
-
export function buildConservativeMultimodalGenerationOptions(options = {}) {
|
|
283
|
-
return {
|
|
284
|
-
...options,
|
|
285
|
-
disableCommandBatching: true,
|
|
286
|
-
disableMultiTokenDecode: true,
|
|
287
|
-
stopCheckMode: 'per-token',
|
|
288
|
-
};
|
|
289
|
-
}
|
|
290
|
-
|
|
291
|
-
function resolveMultimodalMaxTokens(runtimeConfig, requestedMaxTokens) {
|
|
292
|
-
if (requestedMaxTokens !== undefined) {
|
|
293
|
-
return requirePositiveInteger(requestedMaxTokens, 'maxTokens');
|
|
294
|
-
}
|
|
295
|
-
return requirePositiveInteger(
|
|
296
|
-
runtimeConfig?.inference?.generation?.multimodalMaxTokens,
|
|
297
|
-
'runtime.inference.generation.multimodalMaxTokens'
|
|
298
|
-
);
|
|
299
|
-
}
|
|
300
|
-
|
|
301
|
-
function requirePositiveInteger(value, label) {
|
|
302
|
-
if (!Number.isInteger(value) || value <= 0) {
|
|
303
|
-
throw new Error(`${label} must be a positive integer.`);
|
|
304
|
-
}
|
|
305
|
-
return value;
|
|
306
|
-
}
|
|
61
|
+
import {
|
|
62
|
+
initialize as initializeImpl,
|
|
63
|
+
loadModel as loadModelImpl,
|
|
64
|
+
_loadWeights as _loadWeightsImpl,
|
|
65
|
+
_initRoPE as _initRoPEImpl,
|
|
66
|
+
_initConvLayerStates as _initConvLayerStatesImpl,
|
|
67
|
+
_loadVisionWeights as _loadVisionWeightsImpl,
|
|
68
|
+
_ensureVisionWeightsLoaded as _ensureVisionWeightsLoadedImpl,
|
|
69
|
+
_loadAudioWeights as _loadAudioWeightsImpl,
|
|
70
|
+
_ensureAudioWeightsLoaded as _ensureAudioWeightsLoadedImpl,
|
|
71
|
+
} from './text/lifecycle.js';
|
|
72
|
+
import {
|
|
73
|
+
transcribeVideo as transcribeVideoImpl,
|
|
74
|
+
transcribeAudio as transcribeAudioImpl,
|
|
75
|
+
embed as embedImpl,
|
|
76
|
+
embedBatch as embedBatchImpl,
|
|
77
|
+
encodeSequence as encodeSequenceImpl,
|
|
78
|
+
embedImage as embedImageImpl,
|
|
79
|
+
embedAudio as embedAudioImpl,
|
|
80
|
+
} from './text/execution.js';
|
|
81
|
+
import { transcribeImage as transcribeImageImpl } from './text/image-transcription.js';
|
|
307
82
|
|
|
308
83
|
// ============================================================================
|
|
309
84
|
// Main Inference Pipeline Class
|
|
@@ -317,7 +92,6 @@ export class InferencePipeline extends PipelineState {
|
|
|
317
92
|
|
|
318
93
|
_onProgress = null;
|
|
319
94
|
|
|
320
|
-
|
|
321
95
|
_preloadedWeights = null;
|
|
322
96
|
runtimeOverrides = null;
|
|
323
97
|
|
|
@@ -333,1213 +107,70 @@ export class InferencePipeline extends PipelineState {
|
|
|
333
107
|
// Initialization
|
|
334
108
|
// ==========================================================================
|
|
335
109
|
|
|
336
|
-
|
|
337
110
|
async initialize(contexts = {}) {
|
|
338
|
-
|
|
339
|
-
assignGpuContext: true,
|
|
340
|
-
assignUseGPU: true,
|
|
341
|
-
assignMemoryContext: true,
|
|
342
|
-
assignStorageContext: true,
|
|
343
|
-
});
|
|
344
|
-
this.runtimeConfig = runtimeConfig;
|
|
345
|
-
this.commandContext = contexts.commandContext ?? null;
|
|
346
|
-
this.observationContext = createObservationContext({
|
|
347
|
-
runtimeConfig,
|
|
348
|
-
commandContext: this.commandContext,
|
|
349
|
-
});
|
|
350
|
-
this.dopplerLoader = contexts.loader || null;
|
|
351
|
-
this.ownsDopplerLoader = contexts.ownsLoader === true;
|
|
352
|
-
this.runtimeOverrides = contexts.runtimeConfig == null
|
|
353
|
-
? null
|
|
354
|
-
: (typeof structuredClone === 'function'
|
|
355
|
-
? structuredClone(contexts.runtimeConfig)
|
|
356
|
-
: JSON.parse(JSON.stringify(contexts.runtimeConfig)));
|
|
357
|
-
applyPipelineDebugConfig(sharedDebug?.pipeline);
|
|
358
|
-
configurePerfGuards(sharedDebug?.perfGuards);
|
|
359
|
-
|
|
360
|
-
if (!this.gpuContext?.device && typeof globalThis.navigator !== 'undefined' && globalThis.navigator?.gpu) {
|
|
361
|
-
const device = await initDevice();
|
|
362
|
-
if (!device || typeof device !== 'object' || typeof device.createBuffer !== 'function' || !device.queue) {
|
|
363
|
-
throw new Error(
|
|
364
|
-
'GPU device initialization returned an invalid device object. ' +
|
|
365
|
-
'Expected an object with queue and createBuffer. Check WebGPU adapter availability.'
|
|
366
|
-
);
|
|
367
|
-
}
|
|
368
|
-
this.gpuContext = { device };
|
|
369
|
-
this.useGPU = true;
|
|
370
|
-
}
|
|
371
|
-
|
|
372
|
-
this.emulation = await initEmulation(this.runtimeConfig);
|
|
373
|
-
|
|
374
|
-
this.debug = sharedDebug?.pipeline?.enabled === true;
|
|
375
|
-
log.debug('Pipeline', 'Initialized', { useGPU: this.useGPU, debug: this.debug });
|
|
111
|
+
return initializeImpl.call(this, contexts);
|
|
376
112
|
}
|
|
377
113
|
|
|
378
|
-
|
|
379
114
|
async loadModel(manifest) {
|
|
380
|
-
|
|
381
|
-
const pipelineLoadTiming = createPipelineLoadTiming(manifest?.modelId ?? null);
|
|
382
|
-
const resetStart = performance.now();
|
|
383
|
-
this.stats.pipelineLoadTiming = pipelineLoadTiming;
|
|
384
|
-
this.manifest = manifest;
|
|
385
|
-
this.decodeRing?.release();
|
|
386
|
-
if (this.sampleReadbackBuffer) {
|
|
387
|
-
this.sampleReadbackBuffer.destroy();
|
|
388
|
-
this.sampleReadbackBuffer = null;
|
|
389
|
-
}
|
|
390
|
-
this.linearAttentionRuntime = resetLinearAttentionRuntime(this.linearAttentionRuntime);
|
|
391
|
-
destroyMoERouter(this.moeRouter);
|
|
392
|
-
this.moeRouter = null;
|
|
393
|
-
finishPipelineLoadTimingPhase(pipelineLoadTiming, 'reset', resetStart);
|
|
394
|
-
|
|
395
|
-
// ========================================================================
|
|
396
|
-
// Config Resolution Passes
|
|
397
|
-
//
|
|
398
|
-
// The following passes mutate this.runtimeConfig in a fixed order.
|
|
399
|
-
// Each pass is allowed to read the full runtimeConfig but must only
|
|
400
|
-
// mutate its own documented subset. Reordering passes may change
|
|
401
|
-
// resolved values.
|
|
402
|
-
//
|
|
403
|
-
// Phase 1 — applyExecutionV1RuntimeConfig
|
|
404
|
-
// Reads: manifest.inference.execution, kernelCapabilities, platform
|
|
405
|
-
// Mutates: runtimeConfig.inference (kernelPath, pipeline, compute,
|
|
406
|
-
// session via runtimeInferencePatch)
|
|
407
|
-
//
|
|
408
|
-
// Phase 2 — parseModelConfig + applyModelBatchingRuntimeDefaults
|
|
409
|
-
// Reads: manifest.architecture, runtimeConfig.inference.modelOverrides
|
|
410
|
-
// Mutates: runtimeConfig.inference.batching,
|
|
411
|
-
// runtimeConfig.inference.generation
|
|
412
|
-
//
|
|
413
|
-
// Phase 3 — resolveKernelPathState
|
|
414
|
-
// Reads: manifest, modelConfig.kernelPath, runtimeConfig.inference.kernelPath
|
|
415
|
-
// Mutates: runtimeConfig.inference.compute.activationDtype,
|
|
416
|
-
// runtimeConfig.inference.session.kvcache.kvDtype,
|
|
417
|
-
// runtimeConfig.inference.session.compute.defaults.outputDtype
|
|
418
|
-
//
|
|
419
|
-
// Phase 4 — _resolveLayerPipeline
|
|
420
|
-
// Reads: runtimeConfig.inference.pipeline, modelConfig.layerPipeline,
|
|
421
|
-
// executionV1State.runtimeInferencePatch.pipeline
|
|
422
|
-
// Mutates: this.layerPipelinePlan (does not mutate runtimeConfig)
|
|
423
|
-
// ========================================================================
|
|
424
|
-
|
|
425
|
-
let configResolutionPhase = 0;
|
|
426
|
-
const configResolutionStart = performance.now();
|
|
427
|
-
|
|
428
|
-
// Phase 1: execution-v1 runtime config
|
|
429
|
-
configResolutionPhase = 1;
|
|
430
|
-
log.debug('Pipeline', `Config resolution phase ${configResolutionPhase}: applyExecutionV1RuntimeConfig`);
|
|
431
|
-
if (hasExecutionV1(manifest.inference)) {
|
|
432
|
-
let capabilities = null;
|
|
433
|
-
let platform = null;
|
|
434
|
-
try {
|
|
435
|
-
capabilities = getKernelCapabilities();
|
|
436
|
-
} catch {
|
|
437
|
-
// Device not yet initialized — transforms will be skipped
|
|
438
|
-
}
|
|
439
|
-
try {
|
|
440
|
-
platform = getPlatform();
|
|
441
|
-
} catch {
|
|
442
|
-
// Platform not yet initialized — use null fallback
|
|
443
|
-
}
|
|
444
|
-
|
|
445
|
-
const executionV1Runtime = applyExecutionV1RuntimeConfig({
|
|
446
|
-
runtimeConfig: this.runtimeConfig,
|
|
447
|
-
runtimeOverrides: this.runtimeOverrides,
|
|
448
|
-
manifest,
|
|
449
|
-
modelId: manifest.modelId ?? 'model',
|
|
450
|
-
numLayers: Number(manifest.architecture?.numLayers ?? 0),
|
|
451
|
-
capabilities,
|
|
452
|
-
platform,
|
|
453
|
-
useGPU: this.useGPU === true,
|
|
454
|
-
});
|
|
455
|
-
if (executionV1Runtime.executionV1State) {
|
|
456
|
-
this.runtimeConfig = executionV1Runtime.runtimeConfig;
|
|
457
|
-
this.executionV1State = executionV1Runtime.executionV1State;
|
|
458
|
-
const transformInfo = this.executionV1State.appliedTransforms?.length > 0
|
|
459
|
-
? `, transforms=[${this.executionV1State.appliedTransforms.join(', ')}]`
|
|
460
|
-
: '';
|
|
461
|
-
const fallbackInfo = this.executionV1State.fallbackKernelPath
|
|
462
|
-
? ', fallbackKernelPath=yes'
|
|
463
|
-
: '';
|
|
464
|
-
const laneIntegrity = this.executionV1State.laneIntegrity;
|
|
465
|
-
const laneInfo = laneIntegrity?.status === 'transformed'
|
|
466
|
-
? `, laneIntegrity=transformed(declared=${laneIntegrity.declared.activationDtype}/${laneIntegrity.declared.kvDtype},` +
|
|
467
|
-
`executed=${laneIntegrity.executed.activationDtype}/${laneIntegrity.executed.kvDtype})`
|
|
468
|
-
: '';
|
|
469
|
-
log.info(
|
|
470
|
-
'Pipeline',
|
|
471
|
-
`Execution v1 enabled (steps=${this.executionV1State.resolvedSteps.all.length}, ` +
|
|
472
|
-
`kernelPathInline=${this.executionV1State.runtimeInferencePatch.kernelPath ? 'yes' : 'no'}, ` +
|
|
473
|
-
`pipelineInline=${this.executionV1State.runtimeInferencePatch.pipeline ? 'yes' : 'no'}` +
|
|
474
|
-
`${transformInfo}${laneInfo}${fallbackInfo})`
|
|
475
|
-
);
|
|
476
|
-
}
|
|
477
|
-
}
|
|
478
|
-
|
|
479
|
-
// Phase 2: model config + batching defaults
|
|
480
|
-
configResolutionPhase = 2;
|
|
481
|
-
log.debug('Pipeline', `Config resolution phase ${configResolutionPhase}: parseModelConfig + applyModelBatchingRuntimeDefaults`);
|
|
482
|
-
const modelOverrides = (this.runtimeConfig.inference.modelOverrides);
|
|
483
|
-
this.modelConfig = parseModelConfig(manifest, modelOverrides);
|
|
484
|
-
this.runtimeConfig = applyModelBatchingRuntimeDefaults(
|
|
485
|
-
this.runtimeConfig,
|
|
486
|
-
manifest,
|
|
487
|
-
this.modelConfig,
|
|
488
|
-
this.runtimeOverrides
|
|
489
|
-
);
|
|
490
|
-
this.useTiedEmbeddings = this.modelConfig.useTiedEmbeddings;
|
|
491
|
-
this.embeddingVocabSize = this.modelConfig.embeddingVocabSize;
|
|
492
|
-
this.embeddingTranspose = this.modelConfig.embeddingTranspose;
|
|
493
|
-
|
|
494
|
-
// Vision capability detection — gated by manifest fields
|
|
495
|
-
const imageTokenId = manifest.image_token_id;
|
|
496
|
-
const hasVisionQuant = manifest.quantizationInfo?.vision != null;
|
|
497
|
-
if (Number.isInteger(imageTokenId) && imageTokenId > 0 && hasVisionQuant) {
|
|
498
|
-
this.visionCapable = true;
|
|
499
|
-
this.imageTokenId = imageTokenId;
|
|
500
|
-
this.visionConfig = this.modelConfig.visionConfig;
|
|
501
|
-
if (!this.visionConfig) {
|
|
502
|
-
throw new Error(
|
|
503
|
-
`Manifest declares image_token_id=${imageTokenId} and quantizationInfo.vision ` +
|
|
504
|
-
'but no vision_config was resolved. Check conversion config.'
|
|
505
|
-
);
|
|
506
|
-
}
|
|
507
|
-
log.info('Pipeline', `Vision capable: imageTokenId=${imageTokenId}`);
|
|
508
|
-
} else {
|
|
509
|
-
this.visionCapable = false;
|
|
510
|
-
}
|
|
511
|
-
|
|
512
|
-
// Audio capability detection — gated by manifest fields
|
|
513
|
-
const audioTokenId = manifest.audio_token_id;
|
|
514
|
-
const hasAudioQuant = manifest.quantizationInfo?.audio != null;
|
|
515
|
-
if (Number.isInteger(audioTokenId) && audioTokenId > 0 && hasAudioQuant) {
|
|
516
|
-
this.audioCapable = true;
|
|
517
|
-
this.audioTokenId = audioTokenId;
|
|
518
|
-
this.audioConfig = this.modelConfig.audioConfig;
|
|
519
|
-
if (!this.audioConfig) {
|
|
520
|
-
throw new Error(
|
|
521
|
-
`Manifest declares audio_token_id=${audioTokenId} and quantizationInfo.audio ` +
|
|
522
|
-
'but no audio_config was resolved. Check conversion config.'
|
|
523
|
-
);
|
|
524
|
-
}
|
|
525
|
-
log.info('Pipeline', `Audio capable: audioTokenId=${audioTokenId}`);
|
|
526
|
-
} else {
|
|
527
|
-
this.audioCapable = false;
|
|
528
|
-
}
|
|
529
|
-
finishPipelineLoadTimingPhase(pipelineLoadTiming, 'configResolution', configResolutionStart);
|
|
530
|
-
|
|
531
|
-
await timedPipelineLoadPhase(
|
|
532
|
-
pipelineLoadTiming,
|
|
533
|
-
'kernelWarmup',
|
|
534
|
-
{ modelId: manifest.modelId ?? null },
|
|
535
|
-
() => runKernelWarmup({
|
|
536
|
-
useGPU: this.useGPU,
|
|
537
|
-
kernelWarmup: this.runtimeConfig.shared?.kernelWarmup,
|
|
538
|
-
})
|
|
539
|
-
);
|
|
540
|
-
|
|
541
|
-
const executionSetupStart = performance.now();
|
|
542
|
-
// Phase 3: kernel path resolution + dtype contract
|
|
543
|
-
configResolutionPhase = 3;
|
|
544
|
-
log.debug('Pipeline', `Config resolution phase ${configResolutionPhase}: resolveKernelPathState`);
|
|
545
|
-
const kernelPathState = resolveKernelPathState({
|
|
546
|
-
manifest,
|
|
547
|
-
runtimeConfig: this.runtimeConfig,
|
|
548
|
-
runtimeOverrides: this.runtimeOverrides,
|
|
549
|
-
modelConfig: this.modelConfig,
|
|
550
|
-
});
|
|
551
|
-
this.resolvedKernelPath = kernelPathState.resolvedKernelPath;
|
|
552
|
-
this.kernelPathSource = kernelPathState.kernelPathSource;
|
|
553
|
-
this.runtimeConfig = kernelPathState.runtimeConfig;
|
|
554
|
-
|
|
555
|
-
// Phase 4: layer pipeline resolution
|
|
556
|
-
configResolutionPhase = 4;
|
|
557
|
-
log.debug('Pipeline', `Config resolution phase ${configResolutionPhase}: _resolveLayerPipeline`);
|
|
558
|
-
this._resolveLayerPipeline();
|
|
559
|
-
log.debug('Pipeline', `Config resolution complete (${configResolutionPhase} phases)`);
|
|
560
|
-
|
|
561
|
-
const cfg = this.modelConfig;
|
|
562
|
-
const moeStr = cfg.useMoE ? `, MoE(${cfg.numExperts}x${cfg.moeTopK})` : '';
|
|
563
|
-
const kernelInfo = this.resolvedKernelPath ? `kernelPath=${this.resolvedKernelPath.id}` : 'kernelPath=none';
|
|
564
|
-
log.info('Pipeline', `${cfg.numLayers}L/${cfg.hiddenSize}H/${cfg.numHeads}heads (${cfg.headDim}dim)${moeStr}, ${kernelInfo}`);
|
|
565
|
-
finishPipelineLoadTimingPhase(pipelineLoadTiming, 'executionSetup', executionSetupStart);
|
|
566
|
-
|
|
567
|
-
this.tokenizer = await timedPipelineLoadPhase(
|
|
568
|
-
pipelineLoadTiming,
|
|
569
|
-
'tokenizer',
|
|
570
|
-
{ modelId: manifest.modelId ?? null },
|
|
571
|
-
() => initTokenizerFromManifest(
|
|
572
|
-
manifest,
|
|
573
|
-
this.baseUrl,
|
|
574
|
-
this.storageContext
|
|
575
|
-
)
|
|
576
|
-
);
|
|
577
|
-
pipelineLoadTiming.details.tokenizer = typeof this.tokenizer?.getLoadTiming === 'function'
|
|
578
|
-
? this.tokenizer.getLoadTiming()
|
|
579
|
-
: null;
|
|
580
|
-
const tokenizerVocabSize = this.tokenizer.getVocabSize();
|
|
581
|
-
if (Number.isFinite(tokenizerVocabSize) && tokenizerVocabSize > 0) {
|
|
582
|
-
if (tokenizerVocabSize !== this.modelConfig.vocabSize) {
|
|
583
|
-
log.info('Pipeline', `Tokenizer vocabSize=${tokenizerVocabSize} differs from model=${this.modelConfig.vocabSize}, using model size`);
|
|
584
|
-
}
|
|
585
|
-
}
|
|
586
|
-
|
|
587
|
-
const postTokenizerExecutionSetupStart = performance.now();
|
|
588
|
-
// Manifest quantizationInfo.compute is the binding lane identity.
|
|
589
|
-
assertManifestComputeLaneBinding({ manifest, runtimeConfig: this.runtimeConfig });
|
|
590
|
-
|
|
591
|
-
// Initialize KV cache
|
|
592
|
-
if (this.modelConfig.decodeStrategy === 'replay_prefill') {
|
|
593
|
-
this.kvCache = null;
|
|
594
|
-
log.warn(
|
|
595
|
-
'Pipeline',
|
|
596
|
-
'Replay-prefill decode enabled for this model. Incremental KV-cache decode is disabled ' +
|
|
597
|
-
'because the model config did not resolve explicit layerTypes for mixed-geometry/shared-KV decode.'
|
|
598
|
-
);
|
|
599
|
-
} else {
|
|
600
|
-
this.kvCache = createKVCache(this.modelConfig, this.useGPU, this.debug, this.runtimeConfig.inference);
|
|
601
|
-
}
|
|
602
|
-
this.executionPlanState = compileExecutionPlanState({
|
|
603
|
-
runtimeConfig: this.runtimeConfig,
|
|
604
|
-
resolvedKernelPath: this.resolvedKernelPath,
|
|
605
|
-
kernelPathSource: this.kernelPathSource,
|
|
606
|
-
fallbackKernelPath: this.executionV1State?.fallbackKernelPath ?? null,
|
|
607
|
-
});
|
|
608
|
-
const activeExecutionPlan = resolveActiveExecutionPlan(this);
|
|
609
|
-
log.info(
|
|
610
|
-
'Pipeline',
|
|
611
|
-
`Execution plan: active=${activeExecutionPlan.id}, dtype=${activeExecutionPlan.activationDtype}, ` +
|
|
612
|
-
`kernelPath=${activeExecutionPlan.kernelPathId ?? 'none'}`
|
|
613
|
-
);
|
|
614
|
-
|
|
615
|
-
// Issue 1: Validate dtype consistency across all three resolution paths
|
|
616
|
-
// (execution plan, runtimeConfig.inference.compute, and layer context).
|
|
617
|
-
// The layer context is not yet built at this point, so pass null for it.
|
|
618
|
-
// This logs a warning if the execution plan and runtimeConfig disagree.
|
|
619
|
-
assertDtypeConsistency(this.executionPlanState, this.runtimeConfig, null);
|
|
620
|
-
|
|
621
|
-
const kpActivation = getKernelPathActivationDtype(this.resolvedKernelPath);
|
|
622
|
-
if (kpActivation && kpActivation !== activeExecutionPlan.activationDtype) {
|
|
623
|
-
throw new Error(
|
|
624
|
-
`Dtype contract violation: execution plan activationDtype="${activeExecutionPlan.activationDtype}" ` +
|
|
625
|
-
`but kernel path "${this.resolvedKernelPath.id}" declares activationDtype="${kpActivation}".`
|
|
626
|
-
);
|
|
627
|
-
}
|
|
628
|
-
this.resolvedRuntimeSession = createResolvedRuntimeSession({
|
|
629
|
-
manifest,
|
|
630
|
-
modelConfig: this.modelConfig,
|
|
631
|
-
runtimeConfig: this.runtimeConfig,
|
|
632
|
-
resolvedKernelPath: this.resolvedKernelPath,
|
|
633
|
-
kernelPathSource: this.kernelPathSource,
|
|
634
|
-
executionV1State: this.executionV1State,
|
|
635
|
-
executionPlanState: this.executionPlanState,
|
|
636
|
-
});
|
|
637
|
-
|
|
638
|
-
// Initialize MoE router if needed
|
|
639
|
-
if (this.modelConfig.useMoE) {
|
|
640
|
-
this.moeRouter = new MoERouter({
|
|
641
|
-
numExperts: this.modelConfig.numExperts,
|
|
642
|
-
topK: this.modelConfig.moeTopK,
|
|
643
|
-
hiddenSize: this.modelConfig.hiddenSize,
|
|
644
|
-
normalizeWeights: this.runtimeConfig.inference.moe.routing.normalizeWeights,
|
|
645
|
-
});
|
|
646
|
-
}
|
|
647
|
-
|
|
648
|
-
// Initialize speculative decoder
|
|
649
|
-
if (manifest.draftModel) {
|
|
650
|
-
this.speculativeDecoder = initSpeculativeDecoder(
|
|
651
|
-
manifest,
|
|
652
|
-
this.runtimeConfig.inference.speculative
|
|
653
|
-
);
|
|
654
|
-
}
|
|
655
|
-
finishPipelineLoadTimingPhase(
|
|
656
|
-
pipelineLoadTiming,
|
|
657
|
-
'executionSetup',
|
|
658
|
-
postTokenizerExecutionSetupStart
|
|
659
|
-
);
|
|
660
|
-
|
|
661
|
-
// Load weights
|
|
662
|
-
await timedPipelineLoadPhase(
|
|
663
|
-
pipelineLoadTiming,
|
|
664
|
-
'loadWeights',
|
|
665
|
-
{ modelId: manifest.modelId ?? null },
|
|
666
|
-
() => this._loadWeights()
|
|
667
|
-
);
|
|
668
|
-
|
|
669
|
-
// Initialize RoPE frequencies
|
|
670
|
-
await timedPipelineLoadPhase(
|
|
671
|
-
pipelineLoadTiming,
|
|
672
|
-
'rope',
|
|
673
|
-
{ modelId: manifest.modelId ?? null },
|
|
674
|
-
() => this._initRoPE()
|
|
675
|
-
);
|
|
676
|
-
|
|
677
|
-
// Initialize conv layer states for gated short conv layers (LFM2)
|
|
678
|
-
await timedPipelineLoadPhase(
|
|
679
|
-
pipelineLoadTiming,
|
|
680
|
-
'convStates',
|
|
681
|
-
{ modelId: manifest.modelId ?? null },
|
|
682
|
-
() => this._initConvLayerStates()
|
|
683
|
-
);
|
|
684
|
-
|
|
685
|
-
this.isLoaded = true;
|
|
686
|
-
const loadMs = performance.now() - loadStart;
|
|
687
|
-
this.stats.modelLoadMs = loadMs;
|
|
688
|
-
pipelineLoadTiming.totalMs = roundPipelineTimingMs(loadMs);
|
|
689
|
-
pipelineLoadTiming.status = 'complete';
|
|
690
|
-
log.info('Pipeline', `Model loaded successfully (${loadMs.toFixed(0)}ms)`);
|
|
115
|
+
return loadModelImpl.call(this, manifest);
|
|
691
116
|
}
|
|
692
117
|
|
|
693
|
-
|
|
694
118
|
async _loadWeights() {
|
|
695
|
-
|
|
696
|
-
(this.manifest),
|
|
697
|
-
(this.modelConfig),
|
|
698
|
-
{
|
|
699
|
-
storageContext: this.storageContext ?? undefined,
|
|
700
|
-
loadingConfig: this.runtimeConfig.loading,
|
|
701
|
-
baseUrl: this.baseUrl ?? undefined,
|
|
702
|
-
resolvedKernelPath: this.resolvedKernelPath,
|
|
703
|
-
kernelPathSource: this.kernelPathSource,
|
|
704
|
-
keepF32Weights: this.runtimeConfig.inference.compute.keepF32Weights === true,
|
|
705
|
-
loaderDebug: this.runtimeConfig?.shared?.debug?.loader ?? null,
|
|
706
|
-
perLayerInputSession: resolvePerLayerInputsSession(
|
|
707
|
-
this.modelConfig.perLayerInputsSession ?? null,
|
|
708
|
-
this.runtimeConfig?.inference?.session?.perLayerInputs ?? null
|
|
709
|
-
),
|
|
710
|
-
loader: this.dopplerLoader ?? undefined,
|
|
711
|
-
onProgress: (info) => {
|
|
712
|
-
if (info.stage !== 'layers' && info.stage !== 'shards') {
|
|
713
|
-
log.verbose('Loader', `${info.stage}: ${Math.round(info.progress * 100)}%${info.message ? ` - ${info.message}` : ''}`);
|
|
714
|
-
}
|
|
715
|
-
if (this._onProgress) {
|
|
716
|
-
this._onProgress({
|
|
717
|
-
percent: info.progress * 100,
|
|
718
|
-
message: info.message,
|
|
719
|
-
stage: info.stage,
|
|
720
|
-
layer: info.layer,
|
|
721
|
-
total: info.total,
|
|
722
|
-
});
|
|
723
|
-
}
|
|
724
|
-
},
|
|
725
|
-
}
|
|
726
|
-
);
|
|
727
|
-
|
|
728
|
-
result.layerWeights.forEach((w, k) => this.weights.set(k, w));
|
|
729
|
-
this.weights.set('embed', result.embeddings);
|
|
730
|
-
this.weights.set('lm_head', result.lmHead);
|
|
731
|
-
this.weights.set('lm_head_bias', result.lmHeadBias);
|
|
732
|
-
if (result.lmHeadBias) {
|
|
733
|
-
// Recorded/fused logits currently keep their result on GPU, while the
|
|
734
|
-
// manifest-owned decoder bias is loaded as an exact CPU tensor. Route
|
|
735
|
-
// biased heads through computeLogits until the execution graph declares
|
|
736
|
-
// and records an explicit GPU bias-add step.
|
|
737
|
-
this.disableRecordedLogits = true;
|
|
738
|
-
this.disableFusedDecode = true;
|
|
739
|
-
}
|
|
740
|
-
this.weights.set('final_norm', result.finalNorm);
|
|
741
|
-
this.weights.set('final_norm_bias', result.finalNormBias);
|
|
742
|
-
if (result.finalNormBias) {
|
|
743
|
-
this.disableRecordedLogits = true;
|
|
744
|
-
this.disableFusedDecode = true;
|
|
745
|
-
}
|
|
746
|
-
this.weights.set('diffusion_gemma_self_conditioning', result.diffusionGemmaSelfConditioning);
|
|
747
|
-
this.weights.set('per_layer_inputs', result.perLayerInputWeights);
|
|
748
|
-
this.embeddingPostprocessor = result.embeddingPostprocessor;
|
|
749
|
-
|
|
750
|
-
this.layerRouterWeights = result.layerRouterWeights;
|
|
751
|
-
|
|
752
|
-
this.dopplerLoader = result.loader ?? getDopplerLoader(this.runtimeConfig.loading);
|
|
753
|
-
this.stats.loadTiming = result.loadTiming ?? this.dopplerLoader?.getLoadTiming?.() ?? null;
|
|
754
|
-
|
|
755
|
-
if ((this.modelConfig).useMoE && this.moeRouter) {
|
|
756
|
-
this.moeRouter = initMoERouter(
|
|
757
|
-
(this.modelConfig),
|
|
758
|
-
this.runtimeConfig.inference.moe.routing,
|
|
759
|
-
result.layerWeights
|
|
760
|
-
);
|
|
761
|
-
}
|
|
762
|
-
|
|
763
|
-
if (this.useGPU && this.modelConfig) {
|
|
764
|
-
const session = this.runtimeConfig?.inference?.session ?? null;
|
|
765
|
-
fuseQKVWeights(result.layerWeights, this.modelConfig, this.resolvedKernelPath, {
|
|
766
|
-
allowQ4K: session?.useFusedQKVSplitQKNorm === true
|
|
767
|
-
|| session?.useFusedQKVSplitQKNormRoPE === true,
|
|
768
|
-
});
|
|
769
|
-
}
|
|
770
|
-
|
|
771
|
-
if (this.useGPU && this.modelConfig) {
|
|
772
|
-
const activeExecutionPlan = resolveActiveExecutionPlan(this);
|
|
773
|
-
try {
|
|
774
|
-
this.decodeBuffers?.ensureBuffers({
|
|
775
|
-
hiddenSize: this.modelConfig.hiddenSize,
|
|
776
|
-
intermediateSize: this.modelConfig.maxIntermediateSize,
|
|
777
|
-
activationDtype: activeExecutionPlan.activationDtype,
|
|
778
|
-
enablePingPong: true,
|
|
779
|
-
});
|
|
780
|
-
|
|
781
|
-
const device = getDevice();
|
|
782
|
-
if (device) {
|
|
783
|
-
this.finitenessBuffer = device.createBuffer({
|
|
784
|
-
label: 'finiteness_status',
|
|
785
|
-
size: 16,
|
|
786
|
-
usage: GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_SRC | GPUBufferUsage.COPY_DST,
|
|
787
|
-
});
|
|
788
|
-
}
|
|
789
|
-
} catch (bufferError) {
|
|
790
|
-
this.decodeBuffers?.release();
|
|
791
|
-
if (this.finitenessBuffer) {
|
|
792
|
-
this.finitenessBuffer.destroy();
|
|
793
|
-
this.finitenessBuffer = null;
|
|
794
|
-
}
|
|
795
|
-
throw bufferError;
|
|
796
|
-
}
|
|
797
|
-
}
|
|
119
|
+
return _loadWeightsImpl.call(this);
|
|
798
120
|
}
|
|
799
121
|
|
|
800
|
-
|
|
801
122
|
setPreloadedWeights(weights) {
|
|
802
123
|
this._preloadedWeights = weights;
|
|
803
124
|
}
|
|
804
125
|
|
|
805
|
-
|
|
806
126
|
async _initRoPE() {
|
|
807
|
-
|
|
808
|
-
const maxSeqLen = config.maxSeqLen;
|
|
809
|
-
const ropeBuffers = await initRoPEFrequencies({
|
|
810
|
-
headDim: config.globalHeadDim ?? config.headDim,
|
|
811
|
-
localHeadDim: config.headDim,
|
|
812
|
-
rotaryDim: config.ropeRotaryDim,
|
|
813
|
-
ropeLocalRotaryDim: config.ropeLocalRotaryDim,
|
|
814
|
-
ropeFrequencyBaseDim: config.ropeFrequencyBaseDim,
|
|
815
|
-
ropeLocalFrequencyBaseDim: config.ropeLocalFrequencyBaseDim,
|
|
816
|
-
maxSeqLen,
|
|
817
|
-
ropeTheta: config.ropeTheta,
|
|
818
|
-
ropeLocalTheta: config.ropeLocalTheta,
|
|
819
|
-
mropeInterleaved: config.mropeInterleaved,
|
|
820
|
-
mropeSection: config.mropeSection,
|
|
821
|
-
partialRotaryFactor: config.partialRotaryFactor,
|
|
822
|
-
ropeLocalPartialRotaryFactor: config.ropeLocalPartialRotaryFactor,
|
|
823
|
-
ropeScale: config.ropeScale,
|
|
824
|
-
ropeLocalScale: config.ropeLocalScale,
|
|
825
|
-
ropeScalingType: config.ropeScalingType,
|
|
826
|
-
ropeLocalScalingType: config.ropeLocalScalingType,
|
|
827
|
-
ropeScaling: config.ropeScaling,
|
|
828
|
-
ropeLocalScaling: config.ropeLocalScaling,
|
|
829
|
-
}, this.useGPU);
|
|
830
|
-
this.ropeFreqsCos = ropeBuffers.cos;
|
|
831
|
-
this.ropeFreqsSin = ropeBuffers.sin;
|
|
832
|
-
this.ropeLocalCos = ropeBuffers.localCos ?? null;
|
|
833
|
-
this.ropeLocalSin = ropeBuffers.localSin ?? null;
|
|
127
|
+
return _initRoPEImpl.call(this);
|
|
834
128
|
}
|
|
835
129
|
|
|
836
|
-
|
|
837
130
|
async _initConvLayerStates() {
|
|
838
|
-
|
|
839
|
-
if (!config?.layerTypes) return;
|
|
840
|
-
const { getDevice } = await import('../../gpu/device.js');
|
|
841
|
-
const device = getDevice();
|
|
842
|
-
if (!device) return;
|
|
843
|
-
|
|
844
|
-
const hiddenSize = config.hiddenSize;
|
|
845
|
-
const convStates = new Map();
|
|
846
|
-
|
|
847
|
-
for (let i = 0; i < config.layerTypes.length; i++) {
|
|
848
|
-
const lt = String(config.layerTypes[i] ?? '').toLowerCase();
|
|
849
|
-
if (lt !== 'conv' && lt !== 'convolution') continue;
|
|
850
|
-
|
|
851
|
-
const layerWeights = this.weights.get(`layer_${i}`);
|
|
852
|
-
if (!layerWeights) continue;
|
|
853
|
-
const convKernel = layerWeights?.convKernel;
|
|
854
|
-
if (!convKernel) continue;
|
|
855
|
-
|
|
856
|
-
const convState = {};
|
|
857
|
-
try {
|
|
858
|
-
await initConvLayerState(
|
|
859
|
-
convState,
|
|
860
|
-
convKernel,
|
|
861
|
-
layerWeights.convInProj ?? null,
|
|
862
|
-
hiddenSize,
|
|
863
|
-
`L${i}.conv`,
|
|
864
|
-
i
|
|
865
|
-
);
|
|
866
|
-
if (!convState.convWeightGPU || !convState.convStateGPU) {
|
|
867
|
-
continue;
|
|
868
|
-
}
|
|
869
|
-
convStates.set(i, convState);
|
|
870
|
-
} catch (e) {
|
|
871
|
-
log.warn('Pipeline', `Conv layer ${i} state init failed: ${e.message}`);
|
|
872
|
-
}
|
|
873
|
-
}
|
|
874
|
-
|
|
875
|
-
if (convStates.size > 0) {
|
|
876
|
-
this.convLayerStates = convStates;
|
|
877
|
-
log.info('Pipeline', `Initialized ${convStates.size} conv layer states (kernelSize=${convStates.values().next().value?.kernelSize})`);
|
|
878
|
-
}
|
|
131
|
+
return _initConvLayerStatesImpl.call(this);
|
|
879
132
|
}
|
|
880
133
|
|
|
881
|
-
|
|
882
134
|
async _loadVisionWeights() {
|
|
883
|
-
|
|
884
|
-
const vc = this.visionConfig;
|
|
885
|
-
const depth = vc.depth;
|
|
886
|
-
|
|
887
|
-
const loadRequiredTensor = async (name, toGPU = true) => {
|
|
888
|
-
const tensor = await loader.loadTensor(name, toGPU, true);
|
|
889
|
-
if (!tensor) {
|
|
890
|
-
throw new Error(`Vision tensor "${name}" is missing from the converted artifact.`);
|
|
891
|
-
}
|
|
892
|
-
return tensor;
|
|
893
|
-
};
|
|
894
|
-
const loadScalar = async (name) => {
|
|
895
|
-
const tensor = await loadRequiredTensor(name, false);
|
|
896
|
-
if (tensor instanceof Float32Array) {
|
|
897
|
-
if (tensor.length !== 1) {
|
|
898
|
-
throw new Error(`Vision scalar "${name}" must be a single-element tensor, got length=${tensor.length}.`);
|
|
899
|
-
}
|
|
900
|
-
return tensor[0];
|
|
901
|
-
}
|
|
902
|
-
if (ArrayBuffer.isView(tensor) && tensor.length === 1) {
|
|
903
|
-
return Number(tensor[0]);
|
|
904
|
-
}
|
|
905
|
-
if (typeof tensor === 'number') {
|
|
906
|
-
return tensor;
|
|
907
|
-
}
|
|
908
|
-
throw new Error(
|
|
909
|
-
`Vision scalar "${name}" must decode to a single numeric value, ` +
|
|
910
|
-
`got ${tensor?.constructor?.name ?? typeof tensor} length=${tensor?.length ?? 'N/A'}.`
|
|
911
|
-
);
|
|
912
|
-
};
|
|
913
|
-
const loadClipRange = async (prefix) => ({
|
|
914
|
-
inputMin: await loadScalar(`${prefix}.input_min`),
|
|
915
|
-
inputMax: await loadScalar(`${prefix}.input_max`),
|
|
916
|
-
outputMin: await loadScalar(`${prefix}.output_min`),
|
|
917
|
-
outputMax: await loadScalar(`${prefix}.output_max`),
|
|
918
|
-
});
|
|
919
|
-
|
|
920
|
-
if (vc.visionArchitecture === 'gemma4') {
|
|
921
|
-
const isEncoderFree = (depth === 0);
|
|
922
|
-
const visionWeights = {
|
|
923
|
-
textHiddenSize: this.modelConfig.hiddenSize,
|
|
924
|
-
patchInputProj: await loadRequiredTensor('model.vision_tower.patch_embedder.input_proj.weight'),
|
|
925
|
-
patchPositionEmbeddingTable: await loadRequiredTensor('model.vision_tower.patch_embedder.position_embedding_table', false),
|
|
926
|
-
projector: isEncoderFree
|
|
927
|
-
? await loader.loadTensor('model.embed_vision.embedding_projection.weight', true, true)
|
|
928
|
-
: await loadRequiredTensor('model.embed_vision.embedding_projection.weight'),
|
|
929
|
-
layers: [],
|
|
930
|
-
};
|
|
931
|
-
|
|
932
|
-
if (!(visionWeights.patchPositionEmbeddingTable instanceof Float32Array)) {
|
|
933
|
-
throw new Error(
|
|
934
|
-
'Gemma 4 vision position_embedding_table must decode to Float32Array on CPU. ' +
|
|
935
|
-
'Re-convert the artifact if this tensor was quantized incorrectly.'
|
|
936
|
-
);
|
|
937
|
-
}
|
|
938
|
-
|
|
939
|
-
for (let i = 0; i < depth; i++) {
|
|
940
|
-
const prefix = `model.vision_tower.encoder.layers.${i}`;
|
|
941
|
-
const attnPrefix = `${prefix}.self_attn`;
|
|
942
|
-
const mlpPrefix = `${prefix}.mlp`;
|
|
943
|
-
visionWeights.layers.push({
|
|
944
|
-
inputLayerNorm: await loadRequiredTensor(`${prefix}.input_layernorm.weight`),
|
|
945
|
-
postAttentionLayerNorm: await loadRequiredTensor(`${prefix}.post_attention_layernorm.weight`),
|
|
946
|
-
preFeedforwardLayerNorm: await loadRequiredTensor(`${prefix}.pre_feedforward_layernorm.weight`),
|
|
947
|
-
postFeedforwardLayerNorm: await loadRequiredTensor(`${prefix}.post_feedforward_layernorm.weight`),
|
|
948
|
-
qNorm: await loadRequiredTensor(`${attnPrefix}.q_norm.weight`),
|
|
949
|
-
kNorm: await loadRequiredTensor(`${attnPrefix}.k_norm.weight`),
|
|
950
|
-
qProj: await loadRequiredTensor(`${attnPrefix}.q_proj.linear.weight`),
|
|
951
|
-
kProj: await loadRequiredTensor(`${attnPrefix}.k_proj.linear.weight`),
|
|
952
|
-
vProj: await loadRequiredTensor(`${attnPrefix}.v_proj.linear.weight`),
|
|
953
|
-
oProj: await loadRequiredTensor(`${attnPrefix}.o_proj.linear.weight`),
|
|
954
|
-
qProjClip: await loadClipRange(`${attnPrefix}.q_proj`),
|
|
955
|
-
kProjClip: await loadClipRange(`${attnPrefix}.k_proj`),
|
|
956
|
-
vProjClip: await loadClipRange(`${attnPrefix}.v_proj`),
|
|
957
|
-
oProjClip: await loadClipRange(`${attnPrefix}.o_proj`),
|
|
958
|
-
gateProj: await loadRequiredTensor(`${mlpPrefix}.gate_proj.linear.weight`),
|
|
959
|
-
upProj: await loadRequiredTensor(`${mlpPrefix}.up_proj.linear.weight`),
|
|
960
|
-
downProj: await loadRequiredTensor(`${mlpPrefix}.down_proj.linear.weight`),
|
|
961
|
-
gateProjClip: await loadClipRange(`${mlpPrefix}.gate_proj`),
|
|
962
|
-
upProjClip: await loadClipRange(`${mlpPrefix}.up_proj`),
|
|
963
|
-
downProjClip: await loadClipRange(`${mlpPrefix}.down_proj`),
|
|
964
|
-
});
|
|
965
|
-
}
|
|
966
|
-
|
|
967
|
-
this.visionWeights = visionWeights;
|
|
968
|
-
log.info('Pipeline', `Vision weights loaded (${depth} Gemma 4 encoder layers)`);
|
|
969
|
-
return;
|
|
970
|
-
}
|
|
971
|
-
|
|
972
|
-
const visionWeights = {};
|
|
973
|
-
|
|
974
|
-
// Patch embedding weights
|
|
975
|
-
const patchProjName = 'visual.patch_embed.proj.weight';
|
|
976
|
-
const patchProjBiasName = 'visual.patch_embed.proj.bias';
|
|
977
|
-
visionWeights.patchProjWeight = await loader.loadTensor(patchProjName, true, true);
|
|
978
|
-
visionWeights.patchProjBias = await loader.loadTensor(patchProjBiasName, true, true);
|
|
979
|
-
|
|
980
|
-
// Vision encoder layer weights
|
|
981
|
-
visionWeights.layers = [];
|
|
982
|
-
for (let i = 0; i < depth; i++) {
|
|
983
|
-
const prefix = `visual.blocks.${i}`;
|
|
984
|
-
const layerW = {
|
|
985
|
-
norm1Weight: await loader.loadTensor(`${prefix}.norm1.weight`, true, true),
|
|
986
|
-
norm2Weight: await loader.loadTensor(`${prefix}.norm2.weight`, true, true),
|
|
987
|
-
qkvWeight: await loader.loadTensor(`${prefix}.attn.qkv.weight`, true, true),
|
|
988
|
-
qkvBias: await loader.loadTensor(`${prefix}.attn.qkv.bias`, true, true),
|
|
989
|
-
projWeight: await loader.loadTensor(`${prefix}.attn.proj.weight`, true, true),
|
|
990
|
-
projBias: await loader.loadTensor(`${prefix}.attn.proj.bias`, true, true),
|
|
991
|
-
fc1Weight: await loader.loadTensor(`${prefix}.mlp.fc1.weight`, true, true),
|
|
992
|
-
fc1Bias: await loader.loadTensor(`${prefix}.mlp.fc1.bias`, true, true),
|
|
993
|
-
fc2Weight: await loader.loadTensor(`${prefix}.mlp.fc2.weight`, true, true),
|
|
994
|
-
fc2Bias: await loader.loadTensor(`${prefix}.mlp.fc2.bias`, true, true),
|
|
995
|
-
};
|
|
996
|
-
visionWeights.layers.push(layerW);
|
|
997
|
-
}
|
|
998
|
-
|
|
999
|
-
// Spatial merge projection
|
|
1000
|
-
visionWeights.mergerLnWeight = await loader.loadTensor('visual.merger.ln_q.weight', true, true);
|
|
1001
|
-
visionWeights.mergerMlp0Weight = await loader.loadTensor('visual.merger.mlp.0.weight', true, true);
|
|
1002
|
-
visionWeights.mergerMlp0Bias = await loader.loadTensor('visual.merger.mlp.0.bias', true, true);
|
|
1003
|
-
visionWeights.mergerMlp2Weight = await loader.loadTensor('visual.merger.mlp.2.weight', true, true);
|
|
1004
|
-
visionWeights.mergerMlp2Bias = await loader.loadTensor('visual.merger.mlp.2.bias', true, true);
|
|
1005
|
-
|
|
1006
|
-
this.visionWeights = visionWeights;
|
|
1007
|
-
log.info('Pipeline', `Vision weights loaded (${depth} encoder layers)`);
|
|
135
|
+
return _loadVisionWeightsImpl.call(this);
|
|
1008
136
|
}
|
|
1009
137
|
|
|
1010
138
|
async _ensureVisionWeightsLoaded() {
|
|
1011
|
-
|
|
1012
|
-
throw new Error(
|
|
1013
|
-
'Pipeline does not support vision weights (no image_token_id in manifest).'
|
|
1014
|
-
);
|
|
1015
|
-
}
|
|
1016
|
-
if (this.visionWeights) {
|
|
1017
|
-
return;
|
|
1018
|
-
}
|
|
1019
|
-
log.info('Pipeline', 'Loading vision weights on demand');
|
|
1020
|
-
await this._loadVisionWeights();
|
|
139
|
+
return _ensureVisionWeightsLoadedImpl.call(this);
|
|
1021
140
|
}
|
|
1022
141
|
|
|
1023
|
-
|
|
1024
142
|
async _loadAudioWeights() {
|
|
1025
|
-
|
|
1026
|
-
const ac = this.audioConfig;
|
|
1027
|
-
const depth = ac.depth;
|
|
1028
|
-
|
|
1029
|
-
const loadRequiredTensor = async (name, toGPU = true) => {
|
|
1030
|
-
const tensor = await loader.loadTensor(name, toGPU, true);
|
|
1031
|
-
if (!tensor) {
|
|
1032
|
-
throw new Error(`Audio tensor "${name}" is missing from the converted artifact.`);
|
|
1033
|
-
}
|
|
1034
|
-
return tensor;
|
|
1035
|
-
};
|
|
1036
|
-
const loadScalar = async (name) => {
|
|
1037
|
-
const tensor = await loadRequiredTensor(name, false);
|
|
1038
|
-
if (tensor instanceof Float32Array) {
|
|
1039
|
-
if (tensor.length !== 1) {
|
|
1040
|
-
throw new Error(`Audio scalar "${name}" must be a single-element tensor, got length=${tensor.length}.`);
|
|
1041
|
-
}
|
|
1042
|
-
return tensor[0];
|
|
1043
|
-
}
|
|
1044
|
-
if (ArrayBuffer.isView(tensor) && tensor.length === 1) {
|
|
1045
|
-
return Number(tensor[0]);
|
|
1046
|
-
}
|
|
1047
|
-
if (typeof tensor === 'number') {
|
|
1048
|
-
return tensor;
|
|
1049
|
-
}
|
|
1050
|
-
throw new Error(
|
|
1051
|
-
`Audio scalar "${name}" must decode to a single numeric value, ` +
|
|
1052
|
-
`got ${tensor?.constructor?.name ?? typeof tensor} length=${tensor?.length ?? 'N/A'}.`
|
|
1053
|
-
);
|
|
1054
|
-
};
|
|
1055
|
-
const loadClipRange = async (prefix) => ({
|
|
1056
|
-
inputMin: await loadScalar(`${prefix}.input_min`),
|
|
1057
|
-
inputMax: await loadScalar(`${prefix}.input_max`),
|
|
1058
|
-
outputMin: await loadScalar(`${prefix}.output_min`),
|
|
1059
|
-
outputMax: await loadScalar(`${prefix}.output_max`),
|
|
1060
|
-
});
|
|
1061
|
-
|
|
1062
|
-
const isEncoderFree = (depth === 0);
|
|
1063
|
-
const audioWeights = {
|
|
1064
|
-
// Subsampling
|
|
1065
|
-
subsampleConv0Weight: isEncoderFree ? null : await loadRequiredTensor('model.audio_tower.subsample_conv_projection.layer0.conv.weight'),
|
|
1066
|
-
subsampleNorm0Weight: isEncoderFree ? null : await loadRequiredTensor('model.audio_tower.subsample_conv_projection.layer0.norm.weight'),
|
|
1067
|
-
subsampleConv1Weight: isEncoderFree ? null : await loadRequiredTensor('model.audio_tower.subsample_conv_projection.layer1.conv.weight'),
|
|
1068
|
-
subsampleNorm1Weight: isEncoderFree ? null : await loadRequiredTensor('model.audio_tower.subsample_conv_projection.layer1.norm.weight'),
|
|
1069
|
-
subsampleInputProjWeight: isEncoderFree ? null : await loadRequiredTensor('model.audio_tower.subsample_conv_projection.input_proj_linear.weight'),
|
|
1070
|
-
// Output
|
|
1071
|
-
outputProjWeight: isEncoderFree ? null : await loadRequiredTensor('model.audio_tower.output_proj.weight'),
|
|
1072
|
-
outputProjBias: isEncoderFree ? null : await loadRequiredTensor('model.audio_tower.output_proj.bias'),
|
|
1073
|
-
audioEmbeddingProjWeight: await loadRequiredTensor('model.embed_audio.embedding_projection.weight'),
|
|
1074
|
-
layers: [],
|
|
1075
|
-
};
|
|
1076
|
-
|
|
1077
|
-
for (let i = 0; i < depth; i++) {
|
|
1078
|
-
const prefix = `model.audio_tower.layers.${i}`;
|
|
1079
|
-
const layer = {
|
|
1080
|
-
// Feed-forward 1 (Macaron half-step)
|
|
1081
|
-
feedForward1: {
|
|
1082
|
-
preLayerNorm: await loadRequiredTensor(`${prefix}.feed_forward1.pre_layer_norm.weight`),
|
|
1083
|
-
ffwLayer1Weight: await loadRequiredTensor(`${prefix}.feed_forward1.ffw_layer_1.linear.weight`),
|
|
1084
|
-
ffwLayer1Clip: await loadClipRange(`${prefix}.feed_forward1.ffw_layer_1`),
|
|
1085
|
-
ffwLayer2Weight: await loadRequiredTensor(`${prefix}.feed_forward1.ffw_layer_2.linear.weight`),
|
|
1086
|
-
ffwLayer2Clip: await loadClipRange(`${prefix}.feed_forward1.ffw_layer_2`),
|
|
1087
|
-
postLayerNorm: await loadRequiredTensor(`${prefix}.feed_forward1.post_layer_norm.weight`),
|
|
1088
|
-
},
|
|
1089
|
-
// Self-attention
|
|
1090
|
-
normPreAttn: await loadRequiredTensor(`${prefix}.norm_pre_attn.weight`),
|
|
1091
|
-
qProj: await loadRequiredTensor(`${prefix}.self_attn.q_proj.linear.weight`),
|
|
1092
|
-
qProjClip: await loadClipRange(`${prefix}.self_attn.q_proj`),
|
|
1093
|
-
kProj: await loadRequiredTensor(`${prefix}.self_attn.k_proj.linear.weight`),
|
|
1094
|
-
kProjClip: await loadClipRange(`${prefix}.self_attn.k_proj`),
|
|
1095
|
-
vProj: await loadRequiredTensor(`${prefix}.self_attn.v_proj.linear.weight`),
|
|
1096
|
-
vProjClip: await loadClipRange(`${prefix}.self_attn.v_proj`),
|
|
1097
|
-
perDimScale: await loadRequiredTensor(`${prefix}.self_attn.per_dim_scale`),
|
|
1098
|
-
relativeKProj: await loadRequiredTensor(`${prefix}.self_attn.relative_k_proj.weight`),
|
|
1099
|
-
postProj: await loadRequiredTensor(`${prefix}.self_attn.post.linear.weight`),
|
|
1100
|
-
postProjClip: await loadClipRange(`${prefix}.self_attn.post`),
|
|
1101
|
-
normPostAttn: await loadRequiredTensor(`${prefix}.norm_post_attn.weight`),
|
|
1102
|
-
// Convolution module (LConv1D)
|
|
1103
|
-
lconvPreLayerNorm: await loadRequiredTensor(`${prefix}.lconv1d.pre_layer_norm.weight`),
|
|
1104
|
-
lconvLinearStartWeight: await loadRequiredTensor(`${prefix}.lconv1d.linear_start.linear.weight`),
|
|
1105
|
-
lconvLinearStartClip: await loadClipRange(`${prefix}.lconv1d.linear_start`),
|
|
1106
|
-
lconvDepthwiseWeight: await loadRequiredTensor(`${prefix}.lconv1d.depthwise_conv1d.weight`),
|
|
1107
|
-
lconvConvNorm: await loadRequiredTensor(`${prefix}.lconv1d.conv_norm.weight`),
|
|
1108
|
-
lconvLinearEndWeight: await loadRequiredTensor(`${prefix}.lconv1d.linear_end.linear.weight`),
|
|
1109
|
-
lconvLinearEndClip: await loadClipRange(`${prefix}.lconv1d.linear_end`),
|
|
1110
|
-
// Feed-forward 2 (Macaron half-step)
|
|
1111
|
-
feedForward2: {
|
|
1112
|
-
preLayerNorm: await loadRequiredTensor(`${prefix}.feed_forward2.pre_layer_norm.weight`),
|
|
1113
|
-
ffwLayer1Weight: await loadRequiredTensor(`${prefix}.feed_forward2.ffw_layer_1.linear.weight`),
|
|
1114
|
-
ffwLayer1Clip: await loadClipRange(`${prefix}.feed_forward2.ffw_layer_1`),
|
|
1115
|
-
ffwLayer2Weight: await loadRequiredTensor(`${prefix}.feed_forward2.ffw_layer_2.linear.weight`),
|
|
1116
|
-
ffwLayer2Clip: await loadClipRange(`${prefix}.feed_forward2.ffw_layer_2`),
|
|
1117
|
-
postLayerNorm: await loadRequiredTensor(`${prefix}.feed_forward2.post_layer_norm.weight`),
|
|
1118
|
-
},
|
|
1119
|
-
// Final layer norm
|
|
1120
|
-
normOut: await loadRequiredTensor(`${prefix}.norm_out.weight`),
|
|
1121
|
-
};
|
|
1122
|
-
audioWeights.layers.push(layer);
|
|
1123
|
-
}
|
|
1124
|
-
|
|
1125
|
-
this.audioWeights = audioWeights;
|
|
1126
|
-
log.info('Pipeline', `Audio weights loaded (${depth} conformer layers)`);
|
|
143
|
+
return _loadAudioWeightsImpl.call(this);
|
|
1127
144
|
}
|
|
1128
145
|
|
|
1129
146
|
async _ensureAudioWeightsLoaded() {
|
|
1130
|
-
|
|
1131
|
-
throw new Error(
|
|
1132
|
-
'Pipeline does not support audio weights (no audio_token_id in manifest).'
|
|
1133
|
-
);
|
|
1134
|
-
}
|
|
1135
|
-
if (this.audioWeights) {
|
|
1136
|
-
return;
|
|
1137
|
-
}
|
|
1138
|
-
log.info('Pipeline', 'Loading audio weights on demand');
|
|
1139
|
-
await this._loadAudioWeights();
|
|
147
|
+
return _ensureAudioWeightsLoadedImpl.call(this);
|
|
1140
148
|
}
|
|
1141
149
|
|
|
1142
|
-
|
|
1143
150
|
// ==========================================================================
|
|
1144
151
|
// Vision: transcribeImage
|
|
1145
152
|
// ==========================================================================
|
|
1146
153
|
|
|
1147
|
-
/**
|
|
1148
|
-
* Transcribe text from an image using the vision encoder and text decoder.
|
|
1149
|
-
*
|
|
1150
|
-
* @param {object} params
|
|
1151
|
-
* @param {Uint8Array|Float32Array} params.imageBytes Raw image pixel data
|
|
1152
|
-
* @param {number} params.width Image width
|
|
1153
|
-
* @param {number} params.height Image height
|
|
1154
|
-
* @param {string} [params.prompt] Custom transcription prompt
|
|
1155
|
-
* @param {number} [params.maxTokens] Max tokens to generate
|
|
1156
|
-
* @param {number} [params.softTokenBudget] Per-request soft token budget (Gemma 4 tiers: 70/140/280/560/1120)
|
|
1157
|
-
* @returns {Promise<{ text: string, tokens: number[] }>}
|
|
1158
|
-
*/
|
|
1159
154
|
async transcribeImage({ imageBytes, width, height, prompt, maxTokens, softTokenBudget, signal }) {
|
|
1160
|
-
|
|
1161
|
-
if (!this.visionCapable) {
|
|
1162
|
-
throw new Error(
|
|
1163
|
-
'Pipeline does not support image transcription (no image_token_id in manifest).'
|
|
1164
|
-
);
|
|
1165
|
-
}
|
|
1166
|
-
await this._ensureVisionWeightsLoaded();
|
|
1167
|
-
assertNotAborted(signal);
|
|
1168
|
-
|
|
1169
|
-
this.reset();
|
|
1170
|
-
|
|
1171
|
-
// Lazy-load vision module (avoids GPU kernel dependency for text-only pipelines)
|
|
1172
|
-
const { encodeImage } = await import('./vision/index.js');
|
|
1173
|
-
|
|
1174
|
-
// Step 1: Encode image through vision pipeline
|
|
1175
|
-
const encodeResult = await encodeImage({
|
|
1176
|
-
pixels: imageBytes,
|
|
1177
|
-
width,
|
|
1178
|
-
height,
|
|
1179
|
-
visionConfig: this.visionConfig,
|
|
1180
|
-
weights: this.visionWeights,
|
|
1181
|
-
softTokenBudget,
|
|
1182
|
-
});
|
|
1183
|
-
|
|
1184
|
-
// Step 2: Build the multimodal prompt from the model's chat template and
|
|
1185
|
-
// expand the single <|image|> placeholder into the exact visual-token span.
|
|
1186
|
-
const requestedPrompt = prompt ?? 'Describe the image in one short sentence.';
|
|
1187
|
-
const imageTokenId = this.visionConfig?.imageTokenId ?? this.modelConfig?.imageTokenId;
|
|
1188
|
-
if (imageTokenId == null) {
|
|
1189
|
-
throw new Error(
|
|
1190
|
-
'Pipeline missing image_token_id. Re-convert the model with image token metadata.'
|
|
1191
|
-
);
|
|
1192
|
-
}
|
|
1193
|
-
if (this.visionConfig?.visionArchitecture !== 'gemma4') {
|
|
1194
|
-
throw new Error(
|
|
1195
|
-
`[Pipeline] transcribeImage: unsupported vision architecture "${this.visionConfig?.visionArchitecture ?? 'unknown'}". ` +
|
|
1196
|
-
'This runtime path currently requires Gemma 4 multimodal prompt expansion.'
|
|
1197
|
-
);
|
|
1198
|
-
}
|
|
1199
|
-
const templateType = this.modelConfig?.chatTemplateType ?? 'gemma4';
|
|
1200
|
-
const chatOptions = this.modelConfig?.chatTemplateThinking === true ? { thinking: true } : undefined;
|
|
1201
|
-
const multimodalPrompt = formatChatMessages([
|
|
1202
|
-
{
|
|
1203
|
-
role: 'user',
|
|
1204
|
-
content: [
|
|
1205
|
-
{ type: 'image' },
|
|
1206
|
-
{ type: 'text', text: requestedPrompt },
|
|
1207
|
-
],
|
|
1208
|
-
},
|
|
1209
|
-
], templateType, chatOptions);
|
|
1210
|
-
const promptTokenIds = this.tokenizer.encode(multimodalPrompt);
|
|
1211
|
-
const imageTokenSpanLength = encodeResult.numTokens;
|
|
1212
|
-
const effectiveBudget = softTokenBudget ?? this.visionConfig?.defaultOutputLength;
|
|
1213
|
-
const maxImageTokenSpanLength = Number(effectiveBudget);
|
|
1214
|
-
if (!Number.isFinite(maxImageTokenSpanLength) || maxImageTokenSpanLength < 1 || Math.floor(maxImageTokenSpanLength) !== maxImageTokenSpanLength) {
|
|
1215
|
-
throw new Error(
|
|
1216
|
-
`[Pipeline] transcribeImage: invalid soft token budget ${effectiveBudget}. ` +
|
|
1217
|
-
'Expected a positive integer from the resolved vision config or softTokenBudget parameter.'
|
|
1218
|
-
);
|
|
1219
|
-
}
|
|
1220
|
-
if (imageTokenSpanLength > maxImageTokenSpanLength) {
|
|
1221
|
-
throw new Error(
|
|
1222
|
-
`[Pipeline] transcribeImage: encoded Gemma 4 image produced ${imageTokenSpanLength} soft tokens, ` +
|
|
1223
|
-
`which exceeds the effective soft token budget=${maxImageTokenSpanLength}.`
|
|
1224
|
-
);
|
|
1225
|
-
}
|
|
1226
|
-
const boiTokenId = resolveSingleSpecialTokenId(this.tokenizer, '<|image>', 'Gemma 4 BOI token');
|
|
1227
|
-
const eoiTokenId = resolveSingleSpecialTokenId(this.tokenizer, '<image|>', 'Gemma 4 EOI token');
|
|
1228
|
-
const { inputIds: fullTokenIds, imageStartOffset } = expandImagePlaceholderTokenIds(
|
|
1229
|
-
promptTokenIds,
|
|
1230
|
-
imageTokenId,
|
|
1231
|
-
imageTokenSpanLength,
|
|
1232
|
-
{ boiTokenId, eoiTokenId }
|
|
1233
|
-
);
|
|
1234
|
-
const padTokenId = this.tokenizer?.getSpecialTokens?.()?.pad;
|
|
1235
|
-
if (!Number.isFinite(padTokenId) || Math.floor(padTokenId) !== padTokenId || padTokenId < 0) {
|
|
1236
|
-
throw new Error(
|
|
1237
|
-
`[Pipeline] transcribeImage: Gemma 4 multimodal prefill requires a tokenizer pad token ID, got ${padTokenId}.`
|
|
1238
|
-
);
|
|
1239
|
-
}
|
|
1240
|
-
|
|
1241
|
-
// Step 3: Generate with embedding override at the image token offset.
|
|
1242
|
-
const tokens = [];
|
|
1243
|
-
const maxGen = resolveMultimodalMaxTokens(this.runtimeConfig, maxTokens);
|
|
1244
|
-
const stopTokenIds = this.modelConfig.stopTokenIds;
|
|
1245
|
-
|
|
1246
|
-
try {
|
|
1247
|
-
const generation = await this.generator.generateTokenIds('', buildConservativeMultimodalGenerationOptions({
|
|
1248
|
-
inputIds: fullTokenIds,
|
|
1249
|
-
embeddingOverrides: {
|
|
1250
|
-
prefixLength: encodeResult.numTokens,
|
|
1251
|
-
offset: imageStartOffset,
|
|
1252
|
-
embeddings: encodeResult.features,
|
|
1253
|
-
},
|
|
1254
|
-
__internalEmbeddingInputSpan: {
|
|
1255
|
-
offset: imageStartOffset,
|
|
1256
|
-
length: encodeResult.numTokens,
|
|
1257
|
-
tokenId: padTokenId,
|
|
1258
|
-
},
|
|
1259
|
-
__internalMultimodalBidirectionalSpan: {
|
|
1260
|
-
offset: imageStartOffset,
|
|
1261
|
-
length: encodeResult.numTokens,
|
|
1262
|
-
},
|
|
1263
|
-
maxTokens: maxGen,
|
|
1264
|
-
temperature: 0,
|
|
1265
|
-
topK: 1,
|
|
1266
|
-
topP: 1,
|
|
1267
|
-
repetitionPenalty: 1,
|
|
1268
|
-
}));
|
|
1269
|
-
for (const token of generation.tokenIds ?? []) {
|
|
1270
|
-
if (Array.isArray(stopTokenIds) && stopTokenIds.includes(token)) break;
|
|
1271
|
-
tokens.push(token);
|
|
1272
|
-
}
|
|
1273
|
-
} finally {
|
|
1274
|
-
if (encodeResult.features) {
|
|
1275
|
-
releaseBuffer(encodeResult.features);
|
|
1276
|
-
}
|
|
1277
|
-
}
|
|
1278
|
-
|
|
1279
|
-
const text = this.tokenizer.decode(tokens);
|
|
1280
|
-
return { text, tokens };
|
|
155
|
+
return transcribeImageImpl.apply(this, arguments);
|
|
1281
156
|
}
|
|
1282
157
|
|
|
1283
|
-
|
|
1284
158
|
// ==========================================================================
|
|
1285
159
|
// Video: transcribeVideo
|
|
1286
160
|
// ==========================================================================
|
|
1287
161
|
|
|
1288
|
-
/**
|
|
1289
|
-
* Transcribe text from video using the vision encoder (per-frame) and text decoder.
|
|
1290
|
-
*
|
|
1291
|
-
* Video is processed as sampled frames through the existing vision encoder.
|
|
1292
|
-
* Each frame is encoded independently and the visual tokens are concatenated.
|
|
1293
|
-
*
|
|
1294
|
-
* @param {object} params
|
|
1295
|
-
* @param {Array<{ pixels: Uint8Array|Float32Array, width: number, height: number }>} params.frames Decoded video frames
|
|
1296
|
-
* @param {string} [params.prompt] Custom transcription prompt
|
|
1297
|
-
* @param {number} [params.maxTokens] Max tokens to generate
|
|
1298
|
-
* @param {number} [params.maxFrames=8] Maximum frames to sample
|
|
1299
|
-
* @param {number} [params.perFrameSoftTokenBudget] Soft token budget per frame
|
|
1300
|
-
* @returns {Promise<{ text: string, tokens: number[] }>}
|
|
1301
|
-
*/
|
|
1302
162
|
async transcribeVideo({ frames, prompt, maxTokens, maxFrames, perFrameSoftTokenBudget, signal }) {
|
|
1303
|
-
|
|
1304
|
-
if (!this.visionCapable) {
|
|
1305
|
-
throw new Error(
|
|
1306
|
-
'Pipeline does not support video transcription (no image_token_id in manifest for vision encoder).'
|
|
1307
|
-
);
|
|
1308
|
-
}
|
|
1309
|
-
await this._ensureVisionWeightsLoaded();
|
|
1310
|
-
|
|
1311
|
-
this.reset();
|
|
1312
|
-
|
|
1313
|
-
// Lazy-load video module
|
|
1314
|
-
const { encodeVideo } = await import('./video/index.js');
|
|
1315
|
-
|
|
1316
|
-
// Step 1: Encode video frames through vision pipeline
|
|
1317
|
-
const encodeResult = await encodeVideo({
|
|
1318
|
-
frames,
|
|
1319
|
-
visionConfig: this.visionConfig,
|
|
1320
|
-
weights: this.visionWeights,
|
|
1321
|
-
maxFrames: maxFrames ?? 8,
|
|
1322
|
-
perFrameSoftTokenBudget,
|
|
1323
|
-
});
|
|
1324
|
-
|
|
1325
|
-
// Step 2: Build the multimodal prompt with <|video|> placeholder
|
|
1326
|
-
const requestedPrompt = prompt ?? 'Describe the video in one short sentence.';
|
|
1327
|
-
const videoTokenId = this.tokenizer?.model?.tokenToId?.('<|video_token|>')
|
|
1328
|
-
?? this.tokenizer?.encode?.('<|video|>')?.find?.((id) => id !== undefined)
|
|
1329
|
-
?? null;
|
|
1330
|
-
// Fall back to image token ID for video placeholder expansion
|
|
1331
|
-
const placeholderTokenId = videoTokenId ?? this.visionConfig?.imageTokenId ?? this.imageTokenId;
|
|
1332
|
-
if (placeholderTokenId == null) {
|
|
1333
|
-
throw new Error(
|
|
1334
|
-
'Pipeline missing video/image token ID for video placeholder expansion.'
|
|
1335
|
-
);
|
|
1336
|
-
}
|
|
1337
|
-
|
|
1338
|
-
const templateType = this.modelConfig?.chatTemplateType ?? 'gemma4';
|
|
1339
|
-
const chatOptions = this.modelConfig?.chatTemplateThinking === true ? { thinking: true } : undefined;
|
|
1340
|
-
const multimodalPrompt = formatChatMessages([
|
|
1341
|
-
{
|
|
1342
|
-
role: 'user',
|
|
1343
|
-
content: [
|
|
1344
|
-
{ type: 'video' },
|
|
1345
|
-
{ type: 'text', text: requestedPrompt },
|
|
1346
|
-
],
|
|
1347
|
-
},
|
|
1348
|
-
], templateType, chatOptions);
|
|
1349
|
-
const promptTokenIds = this.tokenizer.encode(multimodalPrompt);
|
|
1350
|
-
const videoTokenSpanLength = encodeResult.numTokens;
|
|
1351
|
-
|
|
1352
|
-
// Resolve BOV/EOV tokens (reuse image BOI/EOI if video-specific ones don't exist)
|
|
1353
|
-
const bovTokenId = resolveSingleSpecialTokenId(this.tokenizer, '<|video|>', 'Gemma 4 BOV token');
|
|
1354
|
-
const eovTokenId = resolveSingleSpecialTokenId(this.tokenizer, '<video|>', 'Gemma 4 EOV token');
|
|
1355
|
-
|
|
1356
|
-
const { inputIds: fullTokenIds, imageStartOffset: videoStartOffset } = expandImagePlaceholderTokenIds(
|
|
1357
|
-
promptTokenIds,
|
|
1358
|
-
placeholderTokenId,
|
|
1359
|
-
videoTokenSpanLength,
|
|
1360
|
-
{ boiTokenId: bovTokenId, eoiTokenId: eovTokenId }
|
|
1361
|
-
);
|
|
1362
|
-
|
|
1363
|
-
const padTokenId = this.tokenizer?.getSpecialTokens?.()?.pad;
|
|
1364
|
-
if (!Number.isFinite(padTokenId) || Math.floor(padTokenId) !== padTokenId || padTokenId < 0) {
|
|
1365
|
-
throw new Error(
|
|
1366
|
-
`[Pipeline] transcribeVideo: Gemma 4 multimodal prefill requires a tokenizer pad token ID, got ${padTokenId}.`
|
|
1367
|
-
);
|
|
1368
|
-
}
|
|
1369
|
-
|
|
1370
|
-
// Step 3: Generate with embedding override at the video token offset
|
|
1371
|
-
const tokens = [];
|
|
1372
|
-
const maxGen = resolveMultimodalMaxTokens(this.runtimeConfig, maxTokens);
|
|
1373
|
-
const stopTokenIds = this.modelConfig.stopTokenIds;
|
|
1374
|
-
|
|
1375
|
-
try {
|
|
1376
|
-
const generation = await this.generator.generateTokenIds('', buildConservativeMultimodalGenerationOptions({
|
|
1377
|
-
inputIds: fullTokenIds,
|
|
1378
|
-
embeddingOverrides: {
|
|
1379
|
-
prefixLength: encodeResult.numTokens,
|
|
1380
|
-
offset: videoStartOffset,
|
|
1381
|
-
embeddings: encodeResult.features,
|
|
1382
|
-
},
|
|
1383
|
-
__internalEmbeddingInputSpan: {
|
|
1384
|
-
offset: videoStartOffset,
|
|
1385
|
-
length: encodeResult.numTokens,
|
|
1386
|
-
tokenId: padTokenId,
|
|
1387
|
-
},
|
|
1388
|
-
__internalMultimodalBidirectionalSpan: {
|
|
1389
|
-
offset: videoStartOffset,
|
|
1390
|
-
length: encodeResult.numTokens,
|
|
1391
|
-
},
|
|
1392
|
-
maxTokens: maxGen,
|
|
1393
|
-
temperature: 0,
|
|
1394
|
-
topK: 1,
|
|
1395
|
-
topP: 1,
|
|
1396
|
-
repetitionPenalty: 1,
|
|
1397
|
-
}));
|
|
1398
|
-
for (const token of generation.tokenIds ?? []) {
|
|
1399
|
-
if (Array.isArray(stopTokenIds) && stopTokenIds.includes(token)) break;
|
|
1400
|
-
tokens.push(token);
|
|
1401
|
-
}
|
|
1402
|
-
} finally {
|
|
1403
|
-
if (encodeResult.features) {
|
|
1404
|
-
releaseBuffer(encodeResult.features);
|
|
1405
|
-
}
|
|
1406
|
-
}
|
|
1407
|
-
|
|
1408
|
-
const text = this.tokenizer.decode(tokens);
|
|
1409
|
-
return { text, tokens };
|
|
163
|
+
return transcribeVideoImpl.apply(this, arguments);
|
|
1410
164
|
}
|
|
1411
165
|
|
|
1412
|
-
|
|
1413
166
|
// ==========================================================================
|
|
1414
167
|
// Audio: transcribeAudio
|
|
1415
168
|
// ==========================================================================
|
|
1416
169
|
|
|
1417
|
-
/**
|
|
1418
|
-
* Transcribe text from audio using the audio encoder and text decoder.
|
|
1419
|
-
*
|
|
1420
|
-
* @param {object} params
|
|
1421
|
-
* @param {Float32Array} params.audio Raw audio PCM (mono, 16kHz)
|
|
1422
|
-
* @param {string} [params.prompt] Custom transcription prompt
|
|
1423
|
-
* @param {number} [params.maxTokens] Max tokens to generate
|
|
1424
|
-
* @returns {Promise<{ text: string, tokens: number[] }>}
|
|
1425
|
-
*/
|
|
1426
170
|
async transcribeAudio({ audio, prompt, maxTokens, signal }) {
|
|
1427
|
-
|
|
1428
|
-
if (!this.audioCapable) {
|
|
1429
|
-
throw new Error(
|
|
1430
|
-
'Pipeline does not support audio transcription (no audio_token_id in manifest).'
|
|
1431
|
-
);
|
|
1432
|
-
}
|
|
1433
|
-
await this._ensureAudioWeightsLoaded();
|
|
1434
|
-
|
|
1435
|
-
this.reset();
|
|
1436
|
-
|
|
1437
|
-
// Lazy-load audio modules
|
|
1438
|
-
const { encodeAudio } = await import('./audio/index.js');
|
|
1439
|
-
|
|
1440
|
-
let encodeResult;
|
|
1441
|
-
if (this.audioConfig.depth === 0) {
|
|
1442
|
-
encodeResult = await encodeAudio({
|
|
1443
|
-
rawAudio: audio,
|
|
1444
|
-
audioConfig: this.audioConfig,
|
|
1445
|
-
weights: this.audioWeights,
|
|
1446
|
-
});
|
|
1447
|
-
} else {
|
|
1448
|
-
const { extractLogMelSpectrogram } = await import('./audio/mel.js');
|
|
1449
|
-
const { features: melFeatures, numFrames, nMels } = extractLogMelSpectrogram(audio);
|
|
1450
|
-
encodeResult = await encodeAudio({
|
|
1451
|
-
melFeatures,
|
|
1452
|
-
numFrames,
|
|
1453
|
-
nMels,
|
|
1454
|
-
audioConfig: this.audioConfig,
|
|
1455
|
-
weights: this.audioWeights,
|
|
1456
|
-
});
|
|
1457
|
-
}
|
|
1458
|
-
|
|
1459
|
-
// Step 3: Build the multimodal prompt with <|audio|> placeholder
|
|
1460
|
-
const requestedPrompt = prompt ?? 'Transcribe the audio.';
|
|
1461
|
-
const audioTokenId = this.audioConfig?.audioTokenId ?? this.audioTokenId;
|
|
1462
|
-
if (audioTokenId == null) {
|
|
1463
|
-
throw new Error(
|
|
1464
|
-
'Pipeline missing audio_token_id. Re-convert the model with audio token metadata.'
|
|
1465
|
-
);
|
|
1466
|
-
}
|
|
1467
|
-
const templateType = this.modelConfig?.chatTemplateType ?? 'gemma4';
|
|
1468
|
-
const chatOptions = this.modelConfig?.chatTemplateThinking === true ? { thinking: true } : undefined;
|
|
1469
|
-
const multimodalPrompt = formatChatMessages([
|
|
1470
|
-
{
|
|
1471
|
-
role: 'user',
|
|
1472
|
-
content: [
|
|
1473
|
-
{ type: 'audio' },
|
|
1474
|
-
{ type: 'text', text: requestedPrompt },
|
|
1475
|
-
],
|
|
1476
|
-
},
|
|
1477
|
-
], templateType, chatOptions);
|
|
1478
|
-
const promptTokenIds = this.tokenizer.encode(multimodalPrompt);
|
|
1479
|
-
const audioTokenSpanLength = encodeResult.numTokens;
|
|
1480
|
-
|
|
1481
|
-
// Resolve BOA/EOA tokens
|
|
1482
|
-
const boaTokenId = resolveSingleSpecialTokenId(this.tokenizer, '<|audio|>', 'Gemma 4 BOA token');
|
|
1483
|
-
const eoaTokenId = resolveSingleSpecialTokenId(this.tokenizer, '<audio|>', 'Gemma 4 EOA token');
|
|
1484
|
-
|
|
1485
|
-
// Expand single audio placeholder token into the full audio token span
|
|
1486
|
-
const { inputIds: fullTokenIds, imageStartOffset: audioStartOffset } = expandImagePlaceholderTokenIds(
|
|
1487
|
-
promptTokenIds,
|
|
1488
|
-
audioTokenId,
|
|
1489
|
-
audioTokenSpanLength,
|
|
1490
|
-
{ boiTokenId: boaTokenId, eoiTokenId: eoaTokenId }
|
|
1491
|
-
);
|
|
1492
|
-
|
|
1493
|
-
const padTokenId = this.tokenizer?.getSpecialTokens?.()?.pad;
|
|
1494
|
-
if (!Number.isFinite(padTokenId) || Math.floor(padTokenId) !== padTokenId || padTokenId < 0) {
|
|
1495
|
-
throw new Error(
|
|
1496
|
-
`[Pipeline] transcribeAudio: Gemma 4 multimodal prefill requires a tokenizer pad token ID, got ${padTokenId}.`
|
|
1497
|
-
);
|
|
1498
|
-
}
|
|
1499
|
-
|
|
1500
|
-
// Step 4: Generate with embedding override at the audio token offset
|
|
1501
|
-
const tokens = [];
|
|
1502
|
-
const maxGen = resolveMultimodalMaxTokens(this.runtimeConfig, maxTokens);
|
|
1503
|
-
const stopTokenIds = this.modelConfig.stopTokenIds;
|
|
1504
|
-
|
|
1505
|
-
try {
|
|
1506
|
-
const generation = await this.generator.generateTokenIds('', buildConservativeMultimodalGenerationOptions({
|
|
1507
|
-
inputIds: fullTokenIds,
|
|
1508
|
-
embeddingOverrides: {
|
|
1509
|
-
prefixLength: encodeResult.numTokens,
|
|
1510
|
-
offset: audioStartOffset,
|
|
1511
|
-
embeddings: encodeResult.features,
|
|
1512
|
-
},
|
|
1513
|
-
__internalEmbeddingInputSpan: {
|
|
1514
|
-
offset: audioStartOffset,
|
|
1515
|
-
length: encodeResult.numTokens,
|
|
1516
|
-
tokenId: padTokenId,
|
|
1517
|
-
},
|
|
1518
|
-
__internalMultimodalBidirectionalSpan: {
|
|
1519
|
-
offset: audioStartOffset,
|
|
1520
|
-
length: encodeResult.numTokens,
|
|
1521
|
-
},
|
|
1522
|
-
maxTokens: maxGen,
|
|
1523
|
-
temperature: 0,
|
|
1524
|
-
topK: 1,
|
|
1525
|
-
topP: 1,
|
|
1526
|
-
repetitionPenalty: 1,
|
|
1527
|
-
}));
|
|
1528
|
-
for (const token of generation.tokenIds ?? []) {
|
|
1529
|
-
if (Array.isArray(stopTokenIds) && stopTokenIds.includes(token)) break;
|
|
1530
|
-
tokens.push(token);
|
|
1531
|
-
}
|
|
1532
|
-
} finally {
|
|
1533
|
-
if (encodeResult.features) {
|
|
1534
|
-
releaseBuffer(encodeResult.features);
|
|
1535
|
-
}
|
|
1536
|
-
}
|
|
1537
|
-
|
|
1538
|
-
const text = this.tokenizer.decode(tokens);
|
|
1539
|
-
return { text, tokens };
|
|
171
|
+
return transcribeAudioImpl.apply(this, arguments);
|
|
1540
172
|
}
|
|
1541
173
|
|
|
1542
|
-
|
|
1543
174
|
// ==========================================================================
|
|
1544
175
|
// Capability Detection
|
|
1545
176
|
// ==========================================================================
|
|
@@ -1554,7 +185,6 @@ export class InferencePipeline extends PipelineState {
|
|
|
1554
185
|
return Object.freeze(caps);
|
|
1555
186
|
}
|
|
1556
187
|
|
|
1557
|
-
|
|
1558
188
|
// Layer pipeline precedence (lowest to highest):
|
|
1559
189
|
// 1. execution-v1-produced pipeline (via runtimeInferencePatch.pipeline)
|
|
1560
190
|
// 2. model config pipeline (manifest inference.pipeline)
|
|
@@ -1598,7 +228,6 @@ export class InferencePipeline extends PipelineState {
|
|
|
1598
228
|
// Generation Delegates
|
|
1599
229
|
// ==========================================================================
|
|
1600
230
|
|
|
1601
|
-
|
|
1602
231
|
generate(prompt, options = {}) {
|
|
1603
232
|
assertNotAborted(options?.signal);
|
|
1604
233
|
return this.generator.generate(prompt, options);
|
|
@@ -1633,7 +262,6 @@ export class InferencePipeline extends PipelineState {
|
|
|
1633
262
|
return this.generator.advanceWithTokenAndEmbedding(tokenId, options);
|
|
1634
263
|
}
|
|
1635
264
|
|
|
1636
|
-
|
|
1637
265
|
prefillKVOnly(prompt, options = {}) {
|
|
1638
266
|
assertNotAborted(options?.signal);
|
|
1639
267
|
return this.generator.prefillKVOnly(prompt, options);
|
|
@@ -1659,95 +287,15 @@ export class InferencePipeline extends PipelineState {
|
|
|
1659
287
|
}
|
|
1660
288
|
|
|
1661
289
|
async embed(prompt, options = {}) {
|
|
1662
|
-
|
|
1663
|
-
this.resetForBatch();
|
|
1664
|
-
try {
|
|
1665
|
-
const result = await this.prefillWithEmbedding(prompt, {
|
|
1666
|
-
...options,
|
|
1667
|
-
__skipStateSnapshot: true,
|
|
1668
|
-
});
|
|
1669
|
-
assertNotAborted(options?.signal);
|
|
1670
|
-
return {
|
|
1671
|
-
embedding: result.embedding,
|
|
1672
|
-
tokens: result.tokens,
|
|
1673
|
-
seqLen: result.seqLen,
|
|
1674
|
-
embeddingMode: result.embeddingMode,
|
|
1675
|
-
phase: result.phase ?? null,
|
|
1676
|
-
};
|
|
1677
|
-
} finally {
|
|
1678
|
-
this.resetForBatch();
|
|
1679
|
-
}
|
|
290
|
+
return embedImpl.apply(this, arguments);
|
|
1680
291
|
}
|
|
1681
292
|
|
|
1682
293
|
async embedBatch(prompts, options = {}) {
|
|
1683
|
-
|
|
1684
|
-
throw new Error('embedBatch expects an array of prompts');
|
|
1685
|
-
}
|
|
1686
|
-
assertNotAborted(options?.signal);
|
|
1687
|
-
const batchOptions = { ...options, __skipStateSnapshot: true };
|
|
1688
|
-
const outputs = [];
|
|
1689
|
-
for (const prompt of prompts) {
|
|
1690
|
-
// Check between every prompt so a superseded revision drops the rest.
|
|
1691
|
-
assertNotAborted(options?.signal);
|
|
1692
|
-
outputs.push(await this.embed(prompt, batchOptions));
|
|
1693
|
-
}
|
|
1694
|
-
return outputs;
|
|
294
|
+
return embedBatchImpl.apply(this, arguments);
|
|
1695
295
|
}
|
|
1696
296
|
|
|
1697
297
|
async encodeSequence(sequence, options = {}) {
|
|
1698
|
-
|
|
1699
|
-
const contract = this.manifest?.inference?.sequence ?? null;
|
|
1700
|
-
if (this.manifest?.inference?.supportsSequence !== true || !contract) {
|
|
1701
|
-
throw new Error('Model manifest does not declare sequence encoding support.');
|
|
1702
|
-
}
|
|
1703
|
-
if (typeof sequence !== 'string' || sequence.length === 0) {
|
|
1704
|
-
throw new Error('encodeSequence expects a non-empty sequence string.');
|
|
1705
|
-
}
|
|
1706
|
-
const includeTokenEmbeddings = options.includeTokenEmbeddings
|
|
1707
|
-
?? contract.tokenEmbeddings;
|
|
1708
|
-
const includeLogits = options.includeLogits === true;
|
|
1709
|
-
if (includeTokenEmbeddings && contract.tokenEmbeddings !== true) {
|
|
1710
|
-
throw new Error('Model manifest does not permit token-level sequence embeddings.');
|
|
1711
|
-
}
|
|
1712
|
-
if (includeLogits && contract.logits !== true) {
|
|
1713
|
-
throw new Error('Model manifest does not permit sequence logits.');
|
|
1714
|
-
}
|
|
1715
|
-
const needsTokenEmbeddings = includeTokenEmbeddings || contract.pooledEmbedding !== null;
|
|
1716
|
-
|
|
1717
|
-
this.resetForBatch();
|
|
1718
|
-
try {
|
|
1719
|
-
const result = await this.prefillWithEmbedding(sequence, {
|
|
1720
|
-
...options,
|
|
1721
|
-
embeddingMode: contract.pooledEmbedding?.mode ?? 'last',
|
|
1722
|
-
__skipStateSnapshot: true,
|
|
1723
|
-
__returnTokenEmbeddings: needsTokenEmbeddings,
|
|
1724
|
-
__returnSequenceLogits: includeLogits,
|
|
1725
|
-
});
|
|
1726
|
-
assertNotAborted(options?.signal);
|
|
1727
|
-
const hiddenSize = this.modelConfig.hiddenSize;
|
|
1728
|
-
const pooledResult = contract.pooledEmbedding
|
|
1729
|
-
? poolSequenceTokenEmbeddings(
|
|
1730
|
-
result.tokenEmbeddings,
|
|
1731
|
-
result.tokens,
|
|
1732
|
-
hiddenSize,
|
|
1733
|
-
contract.pooledEmbedding
|
|
1734
|
-
)
|
|
1735
|
-
: null;
|
|
1736
|
-
return {
|
|
1737
|
-
alphabet: contract.alphabet,
|
|
1738
|
-
tokens: result.tokens,
|
|
1739
|
-
tokenMask: pooledResult?.tokenMask ?? new Uint8Array(result.tokens.length),
|
|
1740
|
-
includedTokenCount: pooledResult?.includedTokenCount ?? 0,
|
|
1741
|
-
tokenEmbeddings: includeTokenEmbeddings ? result.tokenEmbeddings : null,
|
|
1742
|
-
pooledEmbedding: pooledResult?.pooled ?? null,
|
|
1743
|
-
logits: includeLogits ? result.logits : null,
|
|
1744
|
-
embeddingDim: hiddenSize,
|
|
1745
|
-
vocabSize: this.modelConfig.vocabSize,
|
|
1746
|
-
phase: result.phase ?? null,
|
|
1747
|
-
};
|
|
1748
|
-
} finally {
|
|
1749
|
-
this.resetForBatch();
|
|
1750
|
-
}
|
|
298
|
+
return encodeSequenceImpl.apply(this, arguments);
|
|
1751
299
|
}
|
|
1752
300
|
|
|
1753
301
|
// Run the vision encoder over a single image and return a mean-pooled
|
|
@@ -1755,61 +303,7 @@ export class InferencePipeline extends PipelineState {
|
|
|
1755
303
|
// (jpeg/png -> RGBA pixels) belongs to the caller; this method takes
|
|
1756
304
|
// already-decoded pixel data.
|
|
1757
305
|
async embedImage({ pixels, width, height, softTokenBudget, signal } = {}) {
|
|
1758
|
-
|
|
1759
|
-
if (!this.visionCapable) {
|
|
1760
|
-
throw new Error(
|
|
1761
|
-
'Pipeline does not support image embedding (no image_token_id in manifest).'
|
|
1762
|
-
);
|
|
1763
|
-
}
|
|
1764
|
-
if (pixels == null) {
|
|
1765
|
-
throw new Error('[Pipeline] embedImage: pixels are required.');
|
|
1766
|
-
}
|
|
1767
|
-
if (!Number.isFinite(width) || width <= 0 || !Number.isFinite(height) || height <= 0) {
|
|
1768
|
-
throw new Error('[Pipeline] embedImage: width and height must be positive integers.');
|
|
1769
|
-
}
|
|
1770
|
-
await this._ensureVisionWeightsLoaded();
|
|
1771
|
-
this.reset();
|
|
1772
|
-
|
|
1773
|
-
const { encodeImage } = await import('./vision/index.js');
|
|
1774
|
-
const encodeResult = await encodeImage({
|
|
1775
|
-
pixels,
|
|
1776
|
-
width,
|
|
1777
|
-
height,
|
|
1778
|
-
visionConfig: this.visionConfig,
|
|
1779
|
-
weights: this.visionWeights,
|
|
1780
|
-
softTokenBudget,
|
|
1781
|
-
});
|
|
1782
|
-
|
|
1783
|
-
const hiddenSize = this.modelConfig.hiddenSize;
|
|
1784
|
-
const numTokens = encodeResult.numTokens;
|
|
1785
|
-
if (!Number.isFinite(numTokens) || numTokens < 1) {
|
|
1786
|
-
releaseBuffer(encodeResult.features);
|
|
1787
|
-
throw new Error(`[Pipeline] embedImage: encoder produced ${numTokens} soft tokens; expected >= 1.`);
|
|
1788
|
-
}
|
|
1789
|
-
try {
|
|
1790
|
-
const bytes = await readBuffer(
|
|
1791
|
-
encodeResult.features,
|
|
1792
|
-
numTokens * hiddenSize * Float32Array.BYTES_PER_ELEMENT
|
|
1793
|
-
);
|
|
1794
|
-
const features = new Float32Array(bytes);
|
|
1795
|
-
const pooled = new Float32Array(hiddenSize);
|
|
1796
|
-
for (let t = 0; t < numTokens; t++) {
|
|
1797
|
-
const base = t * hiddenSize;
|
|
1798
|
-
for (let d = 0; d < hiddenSize; d++) {
|
|
1799
|
-
pooled[d] += features[base + d];
|
|
1800
|
-
}
|
|
1801
|
-
}
|
|
1802
|
-
const inv = 1 / numTokens;
|
|
1803
|
-
for (let d = 0; d < hiddenSize; d++) pooled[d] *= inv;
|
|
1804
|
-
return {
|
|
1805
|
-
embedding: pooled,
|
|
1806
|
-
embeddingDim: hiddenSize,
|
|
1807
|
-
numTokens,
|
|
1808
|
-
embeddingMode: 'mean',
|
|
1809
|
-
};
|
|
1810
|
-
} finally {
|
|
1811
|
-
releaseBuffer(encodeResult.features);
|
|
1812
|
-
}
|
|
306
|
+
return embedImageImpl.apply(this, arguments);
|
|
1813
307
|
}
|
|
1814
308
|
|
|
1815
309
|
// Run the audio encoder over a single PCM segment and return a mean-pooled
|
|
@@ -1817,73 +311,7 @@ export class InferencePipeline extends PipelineState {
|
|
|
1817
311
|
// text hidden size in Gemma 4). Decoder responsibility (webm/opus/wav ->
|
|
1818
312
|
// Float32 PCM at the model's expected sample rate) belongs to the caller.
|
|
1819
313
|
async embedAudio({ audio, signal } = {}) {
|
|
1820
|
-
|
|
1821
|
-
if (!this.audioCapable) {
|
|
1822
|
-
throw new Error(
|
|
1823
|
-
'Pipeline does not support audio embedding (no audio_token_id in manifest).'
|
|
1824
|
-
);
|
|
1825
|
-
}
|
|
1826
|
-
if (audio == null) {
|
|
1827
|
-
throw new Error('[Pipeline] embedAudio: audio is required.');
|
|
1828
|
-
}
|
|
1829
|
-
await this._ensureAudioWeightsLoaded();
|
|
1830
|
-
this.reset();
|
|
1831
|
-
|
|
1832
|
-
const { encodeAudio } = await import('./audio/index.js');
|
|
1833
|
-
|
|
1834
|
-
let encodeResult;
|
|
1835
|
-
if (this.audioConfig.depth === 0) {
|
|
1836
|
-
encodeResult = await encodeAudio({
|
|
1837
|
-
rawAudio: audio,
|
|
1838
|
-
audioConfig: this.audioConfig,
|
|
1839
|
-
weights: this.audioWeights,
|
|
1840
|
-
});
|
|
1841
|
-
} else {
|
|
1842
|
-
const { extractLogMelSpectrogram } = await import('./audio/mel.js');
|
|
1843
|
-
const { features: melFeatures, numFrames, nMels } = extractLogMelSpectrogram(audio);
|
|
1844
|
-
encodeResult = await encodeAudio({
|
|
1845
|
-
melFeatures,
|
|
1846
|
-
numFrames,
|
|
1847
|
-
nMels,
|
|
1848
|
-
audioConfig: this.audioConfig,
|
|
1849
|
-
weights: this.audioWeights,
|
|
1850
|
-
});
|
|
1851
|
-
}
|
|
1852
|
-
|
|
1853
|
-
const hiddenSize = Number(this.audioConfig?.outputProjDims ?? this.modelConfig?.hiddenSize);
|
|
1854
|
-
const numTokens = encodeResult.numTokens;
|
|
1855
|
-
if (!Number.isFinite(hiddenSize) || hiddenSize < 1) {
|
|
1856
|
-
releaseBuffer(encodeResult.features);
|
|
1857
|
-
throw new Error('[Pipeline] embedAudio: audioConfig.outputProjDims is missing or invalid.');
|
|
1858
|
-
}
|
|
1859
|
-
if (!Number.isFinite(numTokens) || numTokens < 1) {
|
|
1860
|
-
releaseBuffer(encodeResult.features);
|
|
1861
|
-
throw new Error(`[Pipeline] embedAudio: encoder produced ${numTokens} tokens; expected >= 1.`);
|
|
1862
|
-
}
|
|
1863
|
-
try {
|
|
1864
|
-
const bytes = await readBuffer(
|
|
1865
|
-
encodeResult.features,
|
|
1866
|
-
numTokens * hiddenSize * Float32Array.BYTES_PER_ELEMENT
|
|
1867
|
-
);
|
|
1868
|
-
const features = new Float32Array(bytes);
|
|
1869
|
-
const pooled = new Float32Array(hiddenSize);
|
|
1870
|
-
for (let t = 0; t < numTokens; t++) {
|
|
1871
|
-
const base = t * hiddenSize;
|
|
1872
|
-
for (let d = 0; d < hiddenSize; d++) {
|
|
1873
|
-
pooled[d] += features[base + d];
|
|
1874
|
-
}
|
|
1875
|
-
}
|
|
1876
|
-
const inv = 1 / numTokens;
|
|
1877
|
-
for (let d = 0; d < hiddenSize; d++) pooled[d] *= inv;
|
|
1878
|
-
return {
|
|
1879
|
-
embedding: pooled,
|
|
1880
|
-
embeddingDim: hiddenSize,
|
|
1881
|
-
numTokens,
|
|
1882
|
-
embeddingMode: 'mean',
|
|
1883
|
-
};
|
|
1884
|
-
} finally {
|
|
1885
|
-
releaseBuffer(encodeResult.features);
|
|
1886
|
-
}
|
|
314
|
+
return embedAudioImpl.apply(this, arguments);
|
|
1887
315
|
}
|
|
1888
316
|
|
|
1889
317
|
prefillWithLogits(prompt, options = {}) {
|
|
@@ -1898,7 +326,6 @@ export class InferencePipeline extends PipelineState {
|
|
|
1898
326
|
return this.generator.prefillWithTokenLogitsFromKV(prefix, prompt, tokenIds, options);
|
|
1899
327
|
}
|
|
1900
328
|
|
|
1901
|
-
|
|
1902
329
|
applyKVCacheSnapshot(snapshot) {
|
|
1903
330
|
this.kvCache = snapshot.cache.clone();
|
|
1904
331
|
if (this.useGPU && this.kvCache) {
|
|
@@ -1923,7 +350,6 @@ export class InferencePipeline extends PipelineState {
|
|
|
1923
350
|
this.currentSeqLen = snapshot.seqLen;
|
|
1924
351
|
}
|
|
1925
352
|
|
|
1926
|
-
|
|
1927
353
|
generateWithPrefixKV(prefix, prompt, options = {}) {
|
|
1928
354
|
return this.generator.generateWithPrefixKV(prefix, prompt, options);
|
|
1929
355
|
}
|
|
@@ -1932,7 +358,6 @@ export class InferencePipeline extends PipelineState {
|
|
|
1932
358
|
// Utility Methods
|
|
1933
359
|
// ==========================================================================
|
|
1934
360
|
|
|
1935
|
-
|
|
1936
361
|
getStats() {
|
|
1937
362
|
const stats = { ...this.stats };
|
|
1938
363
|
stats.batching ??= { ...this.batchingStats };
|
|
@@ -1994,12 +419,10 @@ export class InferencePipeline extends PipelineState {
|
|
|
1994
419
|
return stats;
|
|
1995
420
|
}
|
|
1996
421
|
|
|
1997
|
-
|
|
1998
422
|
getBatchingStats() {
|
|
1999
423
|
return { ...this.batchingStats };
|
|
2000
424
|
}
|
|
2001
425
|
|
|
2002
|
-
|
|
2003
426
|
getMemoryStats() {
|
|
2004
427
|
|
|
2005
428
|
const stats = { used: 0 };
|
|
@@ -2025,14 +448,12 @@ export class InferencePipeline extends PipelineState {
|
|
|
2025
448
|
return stats;
|
|
2026
449
|
}
|
|
2027
450
|
|
|
2028
|
-
|
|
2029
451
|
getKVCacheStats() {
|
|
2030
452
|
if (!this.kvCache) return null;
|
|
2031
453
|
const { seqLen, maxSeqLen } = this.kvCache.getMemoryStats();
|
|
2032
454
|
return { seqLen, maxSeqLen };
|
|
2033
455
|
}
|
|
2034
456
|
|
|
2035
|
-
|
|
2036
457
|
getBufferPool() {
|
|
2037
458
|
try {
|
|
2038
459
|
return getGlobalBufferPool();
|
|
@@ -2041,7 +462,6 @@ export class InferencePipeline extends PipelineState {
|
|
|
2041
462
|
}
|
|
2042
463
|
}
|
|
2043
464
|
|
|
2044
|
-
|
|
2045
465
|
async unload() {
|
|
2046
466
|
const storageContext = this.storageContext;
|
|
2047
467
|
this.storageContext = null;
|
|
@@ -2062,6 +482,7 @@ export class InferencePipeline extends PipelineState {
|
|
|
2062
482
|
this.ownsDopplerLoader = false;
|
|
2063
483
|
this.linearAttentionRuntime = resetLinearAttentionRuntime(this.linearAttentionRuntime);
|
|
2064
484
|
this.lora = null;
|
|
485
|
+
this.revocationIdentity = null;
|
|
2065
486
|
destroyMoERouter(this.moeRouter);
|
|
2066
487
|
this.moeRouter = null;
|
|
2067
488
|
if (this.finitenessBuffer) {
|
|
@@ -2081,17 +502,15 @@ export class InferencePipeline extends PipelineState {
|
|
|
2081
502
|
log.info('Pipeline', 'Unloaded');
|
|
2082
503
|
}
|
|
2083
504
|
|
|
2084
|
-
|
|
2085
505
|
setLoRAAdapter(adapter) {
|
|
506
|
+
assertBundledAdapterAuthorized(adapter);
|
|
2086
507
|
this.lora = adapter;
|
|
2087
508
|
}
|
|
2088
509
|
|
|
2089
|
-
|
|
2090
510
|
getActiveLoRA() {
|
|
2091
511
|
return this.lora;
|
|
2092
512
|
}
|
|
2093
513
|
|
|
2094
|
-
|
|
2095
514
|
reset() {
|
|
2096
515
|
this.kvCache?.clear();
|
|
2097
516
|
this.linearAttentionRuntime = resetLinearAttentionRuntime(this.linearAttentionRuntime);
|
|
@@ -2115,10 +534,6 @@ export class InferencePipeline extends PipelineState {
|
|
|
2115
534
|
this.stats.attentionInputs = [];
|
|
2116
535
|
}
|
|
2117
536
|
|
|
2118
|
-
/**
|
|
2119
|
-
* Lightweight reset for batch embedding: clears KV cache and sequence state
|
|
2120
|
-
* without resetting stats or debug flags (they are overwritten on each prefill).
|
|
2121
|
-
*/
|
|
2122
537
|
resetForBatch() {
|
|
2123
538
|
this.kvCache?.clear();
|
|
2124
539
|
this.linearAttentionRuntime = resetLinearAttentionRuntime(this.linearAttentionRuntime);
|
|
@@ -2128,7 +543,6 @@ export class InferencePipeline extends PipelineState {
|
|
|
2128
543
|
this.decodeRing?.reset();
|
|
2129
544
|
}
|
|
2130
545
|
|
|
2131
|
-
|
|
2132
546
|
releaseGPUResources() {
|
|
2133
547
|
this.decodeBuffers?.release();
|
|
2134
548
|
this.decodeRing?.release();
|
|
@@ -2153,7 +567,6 @@ export class InferencePipeline extends PipelineState {
|
|
|
2153
567
|
// Factory Function
|
|
2154
568
|
// ============================================================================
|
|
2155
569
|
|
|
2156
|
-
|
|
2157
570
|
async function createTransformerPipeline(manifest, contexts = {}) {
|
|
2158
571
|
return createInitializedPipeline(InferencePipeline, manifest, contexts);
|
|
2159
572
|
}
|