doppler-gpu 0.4.15 → 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 -213
- package/models/catalog.json +241 -33
- package/package.json +122 -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 +101 -1349
- package/src/cli/doppler-serve.js +18 -244
- package/src/cli/output/manifest.js +368 -0
- package/src/cli/runtime-profiles.js +4 -8
- 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 +64 -1
- package/src/client/doppler-api.js +42 -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.d.ts +122 -0
- package/src/client/inspection.js +343 -0
- 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/model-host/model-session.js +593 -0
- 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 -63
- package/src/client/runtime/index.js +1 -261
- package/src/client/runtime/lora.d.ts +14 -0
- package/src/client/runtime/lora.js +19 -1
- package/src/client/runtime/model-manager.d.ts +0 -4
- package/src/client/runtime/model-manager.js +12 -28
- package/src/client/runtime/model-rerank.js +30 -0
- package/src/client/runtime/model-session.d.ts +1 -174
- package/src/client/runtime/model-source.d.ts +19 -1
- 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 +126 -0
- package/src/client/runtime/scoped-session.js +276 -0
- 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 -1
- package/src/client/runtime/types.js +0 -1
- 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/boundary-evidence-policy.json +54 -0
- package/src/config/evidence/source-intake-policy.json +122 -0
- 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/inspection/observation-policies.json +80 -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 +22 -37
- 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/compute/f16-activations.json +1 -0
- package/src/config/runtime/compute/f16-batched.json +1 -0
- package/src/config/runtime/default.json +1 -0
- package/src/config/runtime/diagnostics/debug-logits.json +1 -0
- package/src/config/runtime/experiments/bench/gemma3-bench-q4k-alt.json +1 -0
- package/src/config/runtime/experiments/bench/gemma3-bench-q4k-batched.json +1 -0
- package/src/config/runtime/experiments/bench/gemma3-bench-q4k.json +1 -0
- package/src/config/runtime/experiments/bench/gemma4-bench-q4k.json +1 -0
- package/src/config/runtime/experiments/bench/qwen3-bench-q4k-alt.json +1 -0
- package/src/config/runtime/experiments/bench/qwen3-bench-q4k-batched.json +1 -0
- package/src/config/runtime/experiments/debug/gemma3-browser-diagnose-q4k.json +1 -1
- package/src/config/runtime/experiments/debug/gemma3-debug-q4k.json +1 -3
- package/src/config/runtime/experiments/debug/gemma4-e2b-browser-diagnose-q4k.json +1 -1
- package/src/config/runtime/experiments/debug/qwen35-browser-diagnose-q4k.json +1 -1
- package/src/config/runtime/experiments/verify/gemma3-verify.json +1 -3
- package/src/config/runtime/experiments/verify/gemma4-e2b-verify.json +1 -3
- package/src/config/runtime/experiments/verify/lfm2-verify.json +1 -3
- package/src/config/runtime/experiments/verify/translategemma-verify.json +1 -3
- package/src/config/runtime/model/gemma2-debug.json +1 -0
- package/src/config/runtime/model/gemma2-pipeline-debug.json +1 -0
- package/src/config/runtime/model/gemma2-pipeline.json +1 -0
- package/src/config/runtime/model/gemma3-layer-probe.json +1 -0
- package/src/config/runtime/model/qwen3-5-layer-probe.json +1 -0
- package/src/config/runtime/model/qwen3-5-layer3-boundary-probe.json +1 -0
- package/src/config/runtime/model/qwen3-5-linear-attn-debug.json +1 -0
- package/src/config/runtime/optimization-candidates/registry.json +5 -0
- package/src/config/runtime/platform/metal-apple-q4k.json +1 -0
- package/src/config/runtime/profiles/default.json +1 -0
- package/src/config/runtime/profiles/diffusiongemma-26b-a4b-throughput.json +1 -0
- package/src/config/runtime/profiles/gemma3-270m-q4k-throughput-overlapped-probe.json +1 -0
- package/src/config/runtime/profiles/gemma3-270m-q4k-throughput-probe.json +1 -0
- package/src/config/runtime/profiles/gemma4-12b-rdrr-safe-b4.json +1 -3
- package/src/config/runtime/profiles/gemma4-12b-rdrr-safe-single.json +1 -3
- package/src/config/runtime/profiles/gemma4-31b-f16-activations-probe.json +2 -6
- package/src/config/runtime/profiles/gemma4-e2b-baseline-probe.json +1 -3
- package/src/config/runtime/profiles/gemma4-e2b-fused-rmsnorm-probe.json +1 -3
- package/src/config/runtime/profiles/gemma4-e2b-fused-rmsnorm-test.json +1 -0
- package/src/config/runtime/profiles/gemma4-e2b-int4ple-baseline-probe.json +1 -5
- package/src/config/runtime/profiles/gemma4-e2b-int4ple-f16a-probe.json +1 -5
- package/src/config/runtime/profiles/gemma4-e2b-int4ple-reference-export.json +1 -3
- package/src/config/runtime/profiles/gemma4-e2b-int4ple-retain-q4k-probe.json +1 -5
- package/src/config/runtime/profiles/gemma4-e2b-low-memory.json +1 -0
- package/src/config/runtime/profiles/gemma4-e2b-prefill-profile.json +1 -3
- package/src/config/runtime/profiles/gemma4-e2b-throughput-probe.json +1 -0
- package/src/config/runtime/profiles/low-memory.json +1 -0
- 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/production.json +1 -3
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-attn128-candidate.json +1 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-maxbatch64-candidate.json +1 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-normed-ffn-candidate.json +1 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-overlap-ringtokens1-candidate.json +1 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-parity.json +1 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-sequential-candidate.json +1 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-throughput.json +1 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-prefill-profile.json +1 -3
- package/src/config/runtime/profiles/qwen-3-5-0-8b-throughput-probe.json +1 -0
- package/src/config/runtime/profiles/qwen-3-5-2b-f16-primary-throughput.json +1 -5
- package/src/config/runtime/profiles/qwen-3-5-2b-metal-parity.json +1 -0
- package/src/config/runtime/profiles/qwen-3-5-2b-metal-throughput.json +1 -0
- package/src/config/runtime/profiles/qwen-3-5-9b-metal-correctness.json +1 -0
- package/src/config/runtime/profiles/qwen-3-embedding-0-6b-metal-stability.json +1 -3
- package/src/config/runtime/profiles/qwen-3-embedding-0-6b-metal-throughput.json +1 -0
- package/src/config/runtime/profiles/qwen-3-reranker-0-6b-metal-stability.json +1 -0
- package/src/config/runtime/profiles/qwen-3-reranker-0-6b-metal-throughput.json +1 -0
- package/src/config/runtime/profiles/qwen3-6-27b-f16-activations-probe.json +2 -6
- package/src/config/runtime/profiles/qwen3-8-27b-source-parity.json +87 -0
- package/src/config/runtime/profiles/rerank-stability.json +1 -3
- package/src/config/runtime/profiles/rerank-throughput.json +1 -3
- package/src/config/runtime/profiles/simulation.json +1 -3
- package/src/config/runtime/profiles/throughput.json +1 -3
- package/src/config/runtime/profiles/trace-layers.json +1 -3
- package/src/config/runtime/profiles/translategemma-1b-nativekd2-q4k-rdna3-throughput-probe.json +1 -0
- package/src/config/runtime/profiles/turboquant-contiguous-prod.json +1 -0
- package/src/config/runtime/profiles/turboquant-contiguous.json +1 -0
- package/src/config/runtime/profiles/turboquant.json +1 -0
- package/src/config/runtime/profiles/vector-stability.json +1 -3
- package/src/config/runtime/profiles/vector-throughput.json +1 -0
- package/src/config/runtime/profiles/verbose-trace.json +1 -3
- package/src/config/runtime/tiers/gemma4-128gb.json +1 -0
- package/src/config/runtime/tiers/gemma4-16gb.json +1 -0
- package/src/config/runtime/tiers/gemma4-24gb.json +1 -0
- package/src/config/runtime/tiers/gemma4-32gb.json +1 -0
- package/src/config/runtime-merge.js +1 -1
- package/src/config/runtime.js +1 -1
- package/src/config/schema/benchmark.schema.d.ts +7 -0
- package/src/config/schema/benchmark.schema.js +5 -0
- 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/index.d.ts +8 -16
- package/src/config/schema/index.js +5 -8
- package/src/config/schema/inference.schema.d.ts +21 -2
- package/src/config/schema/kernel-thresholds.schema.d.ts +0 -12
- package/src/config/schema/kernel-thresholds.schema.js +1 -13
- package/src/config/schema/kernel-warmup.schema.d.ts +1 -3
- package/src/config/schema/kernel-warmup.schema.js +0 -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/schema/runtime-profile.schema.d.ts +24 -0
- package/src/config/schema/runtime-profile.schema.js +92 -0
- package/src/config/schema/shared-runtime.schema.d.ts +1 -4
- package/src/config/schema/shared-runtime.schema.js +0 -2
- package/src/config/schema/tooling.schema.d.ts +8 -5
- package/src/config/schema/tooling.schema.js +3 -2
- 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.d.ts +2 -0
- package/src/debug/tensor.js +21 -2
- 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 +15 -1
- package/src/gpu/command-recorder.js +63 -2
- 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 +3 -10
- 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 -7
- package/src/gpu/kernels/index.js +18 -6
- 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/kernel-prewarm.d.ts +3 -0
- package/src/gpu/kernels/kernel-prewarm.js +60 -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 -332
- 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 +23 -1
- package/src/index.js +8 -1
- 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} +18 -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 -408
- package/src/inference/browser-harness-text-helpers.js +31 -2196
- package/src/inference/browser-harness.d.ts +19 -2
- package/src/inference/browser-harness.js +46 -1873
- 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/observation-context.d.ts +22 -0
- package/src/inference/observation-context.js +41 -0
- 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/executor-immediate.js +42 -0
- package/src/inference/pipelines/text/attention/executor-recorded.js +29 -0
- 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/index.js +7 -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 +984 -0
- package/src/inference/pipelines/text/attention/plan.d.ts +37 -0
- package/src/inference/pipelines/text/attention/plan.js +345 -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/receipt.js +60 -0
- package/src/inference/pipelines/text/attention/record.js +1 -1277
- 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 +1 -1385
- 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} +44 -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 +5 -2
- 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 +37 -962
- package/src/inference/pipelines/text/generator.js +165 -3539
- 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 +11 -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 +15 -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 +54 -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.d.ts +0 -1
- package/src/inference/pipelines/text/model-load.js +2 -31
- 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 +30 -0
- package/src/inference/pipelines/text/resolved-runtime-session.d.ts +56 -0
- package/src/inference/pipelines/text/resolved-runtime-session.js +175 -0
- package/src/inference/pipelines/text/state.d.ts +14 -0
- package/src/inference/pipelines/text/state.js +4 -0
- package/src/inference/pipelines/text/tsir-fixture-writer.js +88 -16
- package/src/inference/pipelines/text/types.d.ts +10 -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 +48 -1620
- 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.d.ts +49 -0
- package/src/inference/refactor-receipt.js +188 -0
- package/src/inference/rerank.js +182 -0
- package/src/inference/resource-scope.d.ts +33 -0
- package/src/inference/resource-scope.js +168 -0
- package/src/inference/test-harness.d.ts +3 -0
- package/src/inference/test-harness.js +1 -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 +54 -0
- package/src/storage/model-cache.js +590 -0
- 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 +51 -0
- package/src/tooling/boundary-evidence.js +408 -0
- package/src/tooling/browser-command-runner.js +17 -10
- 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-context.d.ts +25 -0
- package/src/tooling/command-context.js +53 -0
- 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-runner-shared.js → command-runtime-execution.js} +22 -5
- package/src/tooling/command-validation.js +1 -24
- package/src/tooling/conversion-run-timing.js +24 -0
- package/src/tooling/demo-receipts.d.ts +56 -0
- package/src/tooling/demo-receipts.js +113 -0
- package/src/tooling/diagnose-runner.js +13 -11
- package/src/tooling/distribution-shard-transport.js +11 -0
- package/src/tooling/execution-cost-ledger.d.ts +95 -0
- package/src/tooling/execution-cost-ledger.js +381 -0
- 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 +15 -48
- package/src/tooling/opfs-cache.js +8 -510
- package/src/tooling/precision-replay-math.d.ts +42 -0
- package/src/tooling/precision-replay-math.js +174 -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 +14 -4
- package/src/tooling/program-bundle-parity.js +267 -106
- package/src/tooling/program-bundle.d.ts +8 -52
- package/src/tooling/program-bundle.js +186 -730
- 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 +184 -0
- package/src/tooling/registered-variant-calibration.js +336 -0
- package/src/tooling/runtime-input-composition.js +105 -53
- package/src/tooling/runtime-intent.js +41 -0
- package/src/tooling/runtime-optimization/candidates.js +664 -0
- package/src/tooling/runtime-optimization-campaign.js +166 -0
- package/src/tooling/runtime-optimization-index.d.ts +6 -0
- package/src/tooling/runtime-optimization-index.js +87 -0
- package/src/tooling/runtime-optimization.d.ts +91 -6
- package/src/tooling/runtime-optimization.js +290 -414
- package/src/tooling/source-intake.d.ts +49 -0
- package/src/tooling/source-intake.js +431 -0
- 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 +45 -0
- package/src/tooling-exports/evidence.js +45 -0
- package/src/tooling-exports/runtime.d.ts +7 -0
- package/src/tooling-exports/runtime.js +7 -0
- package/src/tooling-exports/storage.d.ts +5 -0
- package/src/tooling-exports/storage.js +4 -0
- package/src/tooling-exports.d.ts +14 -0
- package/src/tooling-exports.js +8 -3
- package/src/tooling-exports.shared.d.ts +46 -0
- package/src/tooling-exports.shared.js +37 -0
- 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/client/runtime/model-session.js +0 -432
- package/src/config/schema/tuner.schema.d.ts +0 -39
- package/src/config/schema/tuner.schema.js +0 -13
- package/src/converter/quantizer.d.ts +0 -121
- package/src/gpu/kernel-tuner/benchmarks.js +0 -902
- package/src/gpu/kernel-tuner/cache.js +0 -145
- package/src/gpu/kernel-tuner/index.js +0 -38
- package/src/gpu/kernel-tuner/tuner.js +0 -247
- package/src/gpu/kernel-tuner.js +0 -12
- package/src/gpu/kernels/backward/utils.d.ts +0 -90
- package/src/gpu/kernels/kernel-tuning.d.ts +0 -42
- package/src/gpu/kernels/kernel-tuning.js +0 -158
- package/src/gpu/kernels/uniform-utils.d.ts +0 -66
- package/src/gpu/kernels/utils.d.ts +0 -149
- package/src/gpu/profiler.js +0 -382
- package/src/inference/pipelines/text/attention/output-gate-fusion.js +0 -37
- 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
|
@@ -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';
|
|
@@ -49,259 +49,36 @@ import {
|
|
|
49
49
|
import { getDopplerLoader } from '../../loader/doppler-loader.js';
|
|
50
50
|
import { registerPipeline, getPipelineFactory } from './registry.js';
|
|
51
51
|
import { selectRuleValue } from '../../rules/rule-registry.js';
|
|
52
|
-
|
|
53
|
-
|
|
54
|
-
|
|
55
|
-
|
|
56
|
-
|
|
57
|
-
|
|
58
|
-
|
|
59
|
-
// command-buffer cancellation is not exposed by the spec.
|
|
60
|
-
export class AbortError extends Error {
|
|
61
|
-
constructor(message = 'Doppler: aborted') {
|
|
62
|
-
super(message);
|
|
63
|
-
this.name = 'AbortError';
|
|
64
|
-
this.code = 'ABORT_ERR';
|
|
65
|
-
}
|
|
66
|
-
}
|
|
67
|
-
|
|
68
|
-
export function isAbortError(err) {
|
|
69
|
-
return !!err && (err.name === 'AbortError' || err.code === 'ABORT_ERR' || err.code === 20);
|
|
70
|
-
}
|
|
71
|
-
|
|
72
|
-
function assertNotAborted(signal) {
|
|
73
|
-
if (signal && signal.aborted) {
|
|
74
|
-
throw new AbortError(typeof signal.reason === 'string' ? signal.reason : 'Doppler: aborted');
|
|
75
|
-
}
|
|
76
|
-
}
|
|
77
|
-
|
|
78
|
-
function poolSequenceTokenEmbeddings(tokenEmbeddings, tokenIds, hiddenSize, pooling) {
|
|
79
|
-
if (!(tokenEmbeddings instanceof Float32Array)) {
|
|
80
|
-
throw new Error('Sequence token embeddings are unavailable.');
|
|
81
|
-
}
|
|
82
|
-
if (tokenEmbeddings.length !== tokenIds.length * hiddenSize) {
|
|
83
|
-
throw new Error(
|
|
84
|
-
`Sequence embedding shape mismatch: expected ${tokenIds.length * hiddenSize}, got ${tokenEmbeddings.length}.`
|
|
85
|
-
);
|
|
86
|
-
}
|
|
87
|
-
const excluded = new Set(pooling.excludeTokenIds);
|
|
88
|
-
const includedIndices = [];
|
|
89
|
-
for (let index = 0; index < tokenIds.length; index += 1) {
|
|
90
|
-
if (!excluded.has(tokenIds[index])) includedIndices.push(index);
|
|
91
|
-
}
|
|
92
|
-
if (includedIndices.length === 0) {
|
|
93
|
-
throw new Error('Sequence pooling excluded every input token.');
|
|
94
|
-
}
|
|
95
|
-
const pooled = new Float32Array(hiddenSize);
|
|
96
|
-
if (pooling.mode === 'last') {
|
|
97
|
-
const tokenIndex = includedIndices[includedIndices.length - 1];
|
|
98
|
-
const offset = tokenIndex * hiddenSize;
|
|
99
|
-
pooled.set(tokenEmbeddings.subarray(offset, offset + hiddenSize));
|
|
100
|
-
} else {
|
|
101
|
-
for (const tokenIndex of includedIndices) {
|
|
102
|
-
const offset = tokenIndex * hiddenSize;
|
|
103
|
-
for (let column = 0; column < hiddenSize; column += 1) {
|
|
104
|
-
pooled[column] += tokenEmbeddings[offset + column];
|
|
105
|
-
}
|
|
106
|
-
}
|
|
107
|
-
const scale = 1 / includedIndices.length;
|
|
108
|
-
for (let column = 0; column < hiddenSize; column += 1) {
|
|
109
|
-
pooled[column] *= scale;
|
|
110
|
-
}
|
|
111
|
-
}
|
|
112
|
-
return {
|
|
113
|
-
pooled,
|
|
114
|
-
tokenMask: Uint8Array.from(tokenIds, (tokenId) => excluded.has(tokenId) ? 0 : 1),
|
|
115
|
-
includedTokenCount: includedIndices.length,
|
|
116
|
-
};
|
|
117
|
-
}
|
|
52
|
+
import { createObservationContext } from '../observation-context.js';
|
|
53
|
+
import { createResolvedRuntimeSession } from './text/resolved-runtime-session.js';
|
|
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';
|
|
118
59
|
import { initConvLayerState } from './text/ops.js';
|
|
119
60
|
import { destroyPleBufferCache, destroyPleRuntimeCache } from './text/per-layer-inputs.js';
|
|
120
|
-
|
|
121
|
-
|
|
122
|
-
|
|
123
|
-
|
|
124
|
-
|
|
125
|
-
|
|
126
|
-
|
|
127
|
-
|
|
128
|
-
|
|
129
|
-
|
|
130
|
-
|
|
131
|
-
|
|
132
|
-
|
|
133
|
-
|
|
134
|
-
|
|
135
|
-
|
|
136
|
-
|
|
137
|
-
|
|
138
|
-
|
|
139
|
-
|
|
140
|
-
|
|
141
|
-
};
|
|
142
|
-
return wrapped;
|
|
143
|
-
}
|
|
144
|
-
|
|
145
|
-
async function withPipelineLoadPhase(phase, context, run) {
|
|
146
|
-
try {
|
|
147
|
-
return await run();
|
|
148
|
-
} catch (error) {
|
|
149
|
-
if (error?.details?.pipelineLoadPhase) {
|
|
150
|
-
throw error;
|
|
151
|
-
}
|
|
152
|
-
throw createPipelineLoadPhaseError(error, phase, context);
|
|
153
|
-
}
|
|
154
|
-
}
|
|
155
|
-
|
|
156
|
-
function roundPipelineTimingMs(value) {
|
|
157
|
-
return Number.isFinite(value) ? Number(value.toFixed(3)) : null;
|
|
158
|
-
}
|
|
159
|
-
|
|
160
|
-
function createPipelineLoadTiming(modelId) {
|
|
161
|
-
return {
|
|
162
|
-
schemaVersion: 1,
|
|
163
|
-
source: 'doppler-pipeline',
|
|
164
|
-
modelId: typeof modelId === 'string' ? modelId : null,
|
|
165
|
-
status: 'running',
|
|
166
|
-
phasesMs: {
|
|
167
|
-
reset: null,
|
|
168
|
-
configResolution: null,
|
|
169
|
-
kernelWarmup: null,
|
|
170
|
-
tokenizer: null,
|
|
171
|
-
executionSetup: null,
|
|
172
|
-
loadWeights: null,
|
|
173
|
-
rope: null,
|
|
174
|
-
convStates: null,
|
|
175
|
-
},
|
|
176
|
-
details: {
|
|
177
|
-
tokenizer: null,
|
|
178
|
-
},
|
|
179
|
-
totalMs: null,
|
|
180
|
-
};
|
|
181
|
-
}
|
|
182
|
-
|
|
183
|
-
function finishPipelineLoadTimingPhase(loadTiming, phase, startMs) {
|
|
184
|
-
if (!loadTiming?.phasesMs || !phase) return;
|
|
185
|
-
const elapsedMs = roundPipelineTimingMs(performance.now() - startMs);
|
|
186
|
-
const currentMs = loadTiming.phasesMs[phase];
|
|
187
|
-
loadTiming.phasesMs[phase] = Number.isFinite(currentMs)
|
|
188
|
-
? roundPipelineTimingMs(currentMs + elapsedMs)
|
|
189
|
-
: elapsedMs;
|
|
190
|
-
}
|
|
191
|
-
|
|
192
|
-
async function timedPipelineLoadPhase(loadTiming, phase, context, run) {
|
|
193
|
-
const startMs = performance.now();
|
|
194
|
-
try {
|
|
195
|
-
return await withPipelineLoadPhase(phase, context, run);
|
|
196
|
-
} finally {
|
|
197
|
-
finishPipelineLoadTimingPhase(loadTiming, phase, startMs);
|
|
198
|
-
}
|
|
199
|
-
}
|
|
200
|
-
|
|
201
|
-
function resolveSingleSpecialTokenId(tokenizer, tokenText, label) {
|
|
202
|
-
const rawTokenIds = tokenizer?.encode?.(tokenText);
|
|
203
|
-
const tokenIds = Array.isArray(rawTokenIds)
|
|
204
|
-
? rawTokenIds
|
|
205
|
-
: (ArrayBuffer.isView(rawTokenIds) ? Array.from(rawTokenIds) : null);
|
|
206
|
-
if (!Array.isArray(tokenIds) || tokenIds.length !== 1) {
|
|
207
|
-
throw new Error(
|
|
208
|
-
`[Pipeline] transcribeImage: tokenizer must encode ${label} "${tokenText}" as exactly one token.`
|
|
209
|
-
);
|
|
210
|
-
}
|
|
211
|
-
const tokenId = Number(tokenIds[0]);
|
|
212
|
-
if (!Number.isFinite(tokenId) || Math.floor(tokenId) !== tokenId || tokenId < 0) {
|
|
213
|
-
throw new Error(
|
|
214
|
-
`[Pipeline] transcribeImage: tokenizer returned invalid ${label} token id "${tokenIds[0]}".`
|
|
215
|
-
);
|
|
216
|
-
}
|
|
217
|
-
return tokenId;
|
|
218
|
-
}
|
|
219
|
-
|
|
220
|
-
function expandImagePlaceholderTokenIds(tokenIds, imageTokenId, numImageTokens, options = {}) {
|
|
221
|
-
const normalizedTokenIds = Array.isArray(tokenIds)
|
|
222
|
-
? Int32Array.from(tokenIds)
|
|
223
|
-
: (ArrayBuffer.isView(tokenIds) ? Int32Array.from(tokenIds) : null);
|
|
224
|
-
if (!(normalizedTokenIds instanceof Int32Array)) {
|
|
225
|
-
throw new Error(
|
|
226
|
-
'[Pipeline] transcribeImage: tokenizer.encode() must return an array or typed array of token IDs.'
|
|
227
|
-
);
|
|
228
|
-
}
|
|
229
|
-
if (!Number.isFinite(numImageTokens) || Math.floor(numImageTokens) !== numImageTokens || numImageTokens < 1) {
|
|
230
|
-
throw new Error(
|
|
231
|
-
`[Pipeline] transcribeImage: image token span must be a positive integer, got ${numImageTokens}.`
|
|
232
|
-
);
|
|
233
|
-
}
|
|
234
|
-
|
|
235
|
-
let placeholderIndex = -1;
|
|
236
|
-
let placeholderCount = 0;
|
|
237
|
-
for (let i = 0; i < normalizedTokenIds.length; i++) {
|
|
238
|
-
if (normalizedTokenIds[i] !== imageTokenId) continue;
|
|
239
|
-
if (placeholderIndex < 0) {
|
|
240
|
-
placeholderIndex = i;
|
|
241
|
-
}
|
|
242
|
-
placeholderCount++;
|
|
243
|
-
}
|
|
244
|
-
|
|
245
|
-
if (placeholderCount !== 1) {
|
|
246
|
-
throw new Error(
|
|
247
|
-
`[Pipeline] transcribeImage: expected exactly one image_token_id (${imageTokenId}) placeholder ` +
|
|
248
|
-
`from the chat template, got ${placeholderCount}.`
|
|
249
|
-
);
|
|
250
|
-
}
|
|
251
|
-
|
|
252
|
-
const boiTokenId = Number.isInteger(options.boiTokenId) ? options.boiTokenId : null;
|
|
253
|
-
const eoiTokenId = Number.isInteger(options.eoiTokenId) ? options.eoiTokenId : null;
|
|
254
|
-
const prefixExtra = boiTokenId == null ? 0 : 1;
|
|
255
|
-
const suffixExtra = eoiTokenId == null ? 0 : 1;
|
|
256
|
-
const expandedLength = normalizedTokenIds.length - 1 + prefixExtra + numImageTokens + suffixExtra;
|
|
257
|
-
const expanded = new Int32Array(expandedLength);
|
|
258
|
-
expanded.set(normalizedTokenIds.subarray(0, placeholderIndex), 0);
|
|
259
|
-
let writeOffset = placeholderIndex;
|
|
260
|
-
if (boiTokenId != null) {
|
|
261
|
-
expanded[writeOffset++] = boiTokenId;
|
|
262
|
-
}
|
|
263
|
-
expanded.fill(imageTokenId, writeOffset, writeOffset + numImageTokens);
|
|
264
|
-
const imageStartOffset = writeOffset;
|
|
265
|
-
writeOffset += numImageTokens;
|
|
266
|
-
if (eoiTokenId != null) {
|
|
267
|
-
expanded[writeOffset++] = eoiTokenId;
|
|
268
|
-
}
|
|
269
|
-
expanded.set(
|
|
270
|
-
normalizedTokenIds.subarray(placeholderIndex + 1),
|
|
271
|
-
writeOffset
|
|
272
|
-
);
|
|
273
|
-
|
|
274
|
-
return {
|
|
275
|
-
inputIds: expanded,
|
|
276
|
-
imageStartOffset,
|
|
277
|
-
};
|
|
278
|
-
}
|
|
279
|
-
|
|
280
|
-
export function buildConservativeMultimodalGenerationOptions(options = {}) {
|
|
281
|
-
return {
|
|
282
|
-
...options,
|
|
283
|
-
disableCommandBatching: true,
|
|
284
|
-
disableMultiTokenDecode: true,
|
|
285
|
-
stopCheckMode: 'per-token',
|
|
286
|
-
};
|
|
287
|
-
}
|
|
288
|
-
|
|
289
|
-
function resolveMultimodalMaxTokens(runtimeConfig, requestedMaxTokens) {
|
|
290
|
-
if (requestedMaxTokens !== undefined) {
|
|
291
|
-
return requirePositiveInteger(requestedMaxTokens, 'maxTokens');
|
|
292
|
-
}
|
|
293
|
-
return requirePositiveInteger(
|
|
294
|
-
runtimeConfig?.inference?.generation?.multimodalMaxTokens,
|
|
295
|
-
'runtime.inference.generation.multimodalMaxTokens'
|
|
296
|
-
);
|
|
297
|
-
}
|
|
298
|
-
|
|
299
|
-
function requirePositiveInteger(value, label) {
|
|
300
|
-
if (!Number.isInteger(value) || value <= 0) {
|
|
301
|
-
throw new Error(`${label} must be a positive integer.`);
|
|
302
|
-
}
|
|
303
|
-
return value;
|
|
304
|
-
}
|
|
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';
|
|
305
82
|
|
|
306
83
|
// ============================================================================
|
|
307
84
|
// Main Inference Pipeline Class
|
|
@@ -315,7 +92,6 @@ export class InferencePipeline extends PipelineState {
|
|
|
315
92
|
|
|
316
93
|
_onProgress = null;
|
|
317
94
|
|
|
318
|
-
|
|
319
95
|
_preloadedWeights = null;
|
|
320
96
|
runtimeOverrides = null;
|
|
321
97
|
|
|
@@ -331,1200 +107,70 @@ export class InferencePipeline extends PipelineState {
|
|
|
331
107
|
// Initialization
|
|
332
108
|
// ==========================================================================
|
|
333
109
|
|
|
334
|
-
|
|
335
110
|
async initialize(contexts = {}) {
|
|
336
|
-
|
|
337
|
-
assignGpuContext: true,
|
|
338
|
-
assignUseGPU: true,
|
|
339
|
-
assignMemoryContext: true,
|
|
340
|
-
assignStorageContext: true,
|
|
341
|
-
});
|
|
342
|
-
this.runtimeConfig = runtimeConfig;
|
|
343
|
-
this.dopplerLoader = contexts.loader || null;
|
|
344
|
-
this.ownsDopplerLoader = contexts.ownsLoader === true;
|
|
345
|
-
this.runtimeOverrides = contexts.runtimeConfig == null
|
|
346
|
-
? null
|
|
347
|
-
: (typeof structuredClone === 'function'
|
|
348
|
-
? structuredClone(contexts.runtimeConfig)
|
|
349
|
-
: JSON.parse(JSON.stringify(contexts.runtimeConfig)));
|
|
350
|
-
applyPipelineDebugConfig(sharedDebug?.pipeline);
|
|
351
|
-
configurePerfGuards(sharedDebug?.perfGuards);
|
|
352
|
-
|
|
353
|
-
if (!this.gpuContext?.device && typeof globalThis.navigator !== 'undefined' && globalThis.navigator?.gpu) {
|
|
354
|
-
const device = await initDevice();
|
|
355
|
-
if (!device || typeof device !== 'object' || typeof device.createBuffer !== 'function' || !device.queue) {
|
|
356
|
-
throw new Error(
|
|
357
|
-
'GPU device initialization returned an invalid device object. ' +
|
|
358
|
-
'Expected an object with queue and createBuffer. Check WebGPU adapter availability.'
|
|
359
|
-
);
|
|
360
|
-
}
|
|
361
|
-
this.gpuContext = { device };
|
|
362
|
-
this.useGPU = true;
|
|
363
|
-
}
|
|
364
|
-
|
|
365
|
-
this.emulation = await initEmulation(this.runtimeConfig);
|
|
366
|
-
|
|
367
|
-
this.debug = sharedDebug?.pipeline?.enabled === true;
|
|
368
|
-
log.debug('Pipeline', 'Initialized', { useGPU: this.useGPU, debug: this.debug });
|
|
111
|
+
return initializeImpl.call(this, contexts);
|
|
369
112
|
}
|
|
370
113
|
|
|
371
|
-
|
|
372
114
|
async loadModel(manifest) {
|
|
373
|
-
|
|
374
|
-
const pipelineLoadTiming = createPipelineLoadTiming(manifest?.modelId ?? null);
|
|
375
|
-
const resetStart = performance.now();
|
|
376
|
-
this.stats.pipelineLoadTiming = pipelineLoadTiming;
|
|
377
|
-
this.manifest = manifest;
|
|
378
|
-
this.decodeRing?.release();
|
|
379
|
-
if (this.sampleReadbackBuffer) {
|
|
380
|
-
this.sampleReadbackBuffer.destroy();
|
|
381
|
-
this.sampleReadbackBuffer = null;
|
|
382
|
-
}
|
|
383
|
-
this.linearAttentionRuntime = resetLinearAttentionRuntime(this.linearAttentionRuntime);
|
|
384
|
-
destroyMoERouter(this.moeRouter);
|
|
385
|
-
this.moeRouter = null;
|
|
386
|
-
finishPipelineLoadTimingPhase(pipelineLoadTiming, 'reset', resetStart);
|
|
387
|
-
|
|
388
|
-
// ========================================================================
|
|
389
|
-
// Config Resolution Passes
|
|
390
|
-
//
|
|
391
|
-
// The following passes mutate this.runtimeConfig in a fixed order.
|
|
392
|
-
// Each pass is allowed to read the full runtimeConfig but must only
|
|
393
|
-
// mutate its own documented subset. Reordering passes may change
|
|
394
|
-
// resolved values.
|
|
395
|
-
//
|
|
396
|
-
// Phase 1 — applyExecutionV1RuntimeConfig
|
|
397
|
-
// Reads: manifest.inference.execution, kernelCapabilities, platform
|
|
398
|
-
// Mutates: runtimeConfig.inference (kernelPath, pipeline, compute,
|
|
399
|
-
// session via runtimeInferencePatch)
|
|
400
|
-
//
|
|
401
|
-
// Phase 2 — parseModelConfig + applyModelBatchingRuntimeDefaults
|
|
402
|
-
// Reads: manifest.architecture, runtimeConfig.inference.modelOverrides
|
|
403
|
-
// Mutates: runtimeConfig.inference.batching,
|
|
404
|
-
// runtimeConfig.inference.generation
|
|
405
|
-
//
|
|
406
|
-
// Phase 3 — resolveKernelPathState
|
|
407
|
-
// Reads: manifest, modelConfig.kernelPath, runtimeConfig.inference.kernelPath
|
|
408
|
-
// Mutates: runtimeConfig.inference.compute.activationDtype,
|
|
409
|
-
// runtimeConfig.inference.session.kvcache.kvDtype,
|
|
410
|
-
// runtimeConfig.inference.session.compute.defaults.outputDtype
|
|
411
|
-
//
|
|
412
|
-
// Phase 4 — _resolveLayerPipeline
|
|
413
|
-
// Reads: runtimeConfig.inference.pipeline, modelConfig.layerPipeline,
|
|
414
|
-
// executionV1State.runtimeInferencePatch.pipeline
|
|
415
|
-
// Mutates: this.layerPipelinePlan (does not mutate runtimeConfig)
|
|
416
|
-
// ========================================================================
|
|
417
|
-
|
|
418
|
-
let configResolutionPhase = 0;
|
|
419
|
-
const configResolutionStart = performance.now();
|
|
420
|
-
|
|
421
|
-
// Phase 1: execution-v1 runtime config
|
|
422
|
-
configResolutionPhase = 1;
|
|
423
|
-
log.debug('Pipeline', `Config resolution phase ${configResolutionPhase}: applyExecutionV1RuntimeConfig`);
|
|
424
|
-
if (hasExecutionV1(manifest.inference)) {
|
|
425
|
-
let capabilities = null;
|
|
426
|
-
let platform = null;
|
|
427
|
-
try {
|
|
428
|
-
capabilities = getKernelCapabilities();
|
|
429
|
-
} catch {
|
|
430
|
-
// Device not yet initialized — transforms will be skipped
|
|
431
|
-
}
|
|
432
|
-
try {
|
|
433
|
-
platform = getPlatform();
|
|
434
|
-
} catch {
|
|
435
|
-
// Platform not yet initialized — use null fallback
|
|
436
|
-
}
|
|
437
|
-
|
|
438
|
-
const executionV1Runtime = applyExecutionV1RuntimeConfig({
|
|
439
|
-
runtimeConfig: this.runtimeConfig,
|
|
440
|
-
runtimeOverrides: this.runtimeOverrides,
|
|
441
|
-
manifest,
|
|
442
|
-
modelId: manifest.modelId ?? 'model',
|
|
443
|
-
numLayers: Number(manifest.architecture?.numLayers ?? 0),
|
|
444
|
-
capabilities,
|
|
445
|
-
platform,
|
|
446
|
-
useGPU: this.useGPU === true,
|
|
447
|
-
});
|
|
448
|
-
if (executionV1Runtime.executionV1State) {
|
|
449
|
-
this.runtimeConfig = executionV1Runtime.runtimeConfig;
|
|
450
|
-
this.executionV1State = executionV1Runtime.executionV1State;
|
|
451
|
-
const transformInfo = this.executionV1State.appliedTransforms?.length > 0
|
|
452
|
-
? `, transforms=[${this.executionV1State.appliedTransforms.join(', ')}]`
|
|
453
|
-
: '';
|
|
454
|
-
const fallbackInfo = this.executionV1State.fallbackKernelPath
|
|
455
|
-
? ', fallbackKernelPath=yes'
|
|
456
|
-
: '';
|
|
457
|
-
const laneIntegrity = this.executionV1State.laneIntegrity;
|
|
458
|
-
const laneInfo = laneIntegrity?.status === 'transformed'
|
|
459
|
-
? `, laneIntegrity=transformed(declared=${laneIntegrity.declared.activationDtype}/${laneIntegrity.declared.kvDtype},` +
|
|
460
|
-
`executed=${laneIntegrity.executed.activationDtype}/${laneIntegrity.executed.kvDtype})`
|
|
461
|
-
: '';
|
|
462
|
-
log.info(
|
|
463
|
-
'Pipeline',
|
|
464
|
-
`Execution v1 enabled (steps=${this.executionV1State.resolvedSteps.all.length}, ` +
|
|
465
|
-
`kernelPathInline=${this.executionV1State.runtimeInferencePatch.kernelPath ? 'yes' : 'no'}, ` +
|
|
466
|
-
`pipelineInline=${this.executionV1State.runtimeInferencePatch.pipeline ? 'yes' : 'no'}` +
|
|
467
|
-
`${transformInfo}${laneInfo}${fallbackInfo})`
|
|
468
|
-
);
|
|
469
|
-
}
|
|
470
|
-
}
|
|
471
|
-
|
|
472
|
-
// Phase 2: model config + batching defaults
|
|
473
|
-
configResolutionPhase = 2;
|
|
474
|
-
log.debug('Pipeline', `Config resolution phase ${configResolutionPhase}: parseModelConfig + applyModelBatchingRuntimeDefaults`);
|
|
475
|
-
const modelOverrides = (this.runtimeConfig.inference.modelOverrides);
|
|
476
|
-
this.modelConfig = parseModelConfig(manifest, modelOverrides);
|
|
477
|
-
this.runtimeConfig = applyModelBatchingRuntimeDefaults(
|
|
478
|
-
this.runtimeConfig,
|
|
479
|
-
manifest,
|
|
480
|
-
this.modelConfig,
|
|
481
|
-
this.runtimeOverrides
|
|
482
|
-
);
|
|
483
|
-
this.useTiedEmbeddings = this.modelConfig.useTiedEmbeddings;
|
|
484
|
-
this.embeddingVocabSize = this.modelConfig.embeddingVocabSize;
|
|
485
|
-
this.embeddingTranspose = this.modelConfig.embeddingTranspose;
|
|
486
|
-
|
|
487
|
-
// Vision capability detection — gated by manifest fields
|
|
488
|
-
const imageTokenId = manifest.image_token_id;
|
|
489
|
-
const hasVisionQuant = manifest.quantizationInfo?.vision != null;
|
|
490
|
-
if (Number.isInteger(imageTokenId) && imageTokenId > 0 && hasVisionQuant) {
|
|
491
|
-
this.visionCapable = true;
|
|
492
|
-
this.imageTokenId = imageTokenId;
|
|
493
|
-
this.visionConfig = this.modelConfig.visionConfig;
|
|
494
|
-
if (!this.visionConfig) {
|
|
495
|
-
throw new Error(
|
|
496
|
-
`Manifest declares image_token_id=${imageTokenId} and quantizationInfo.vision ` +
|
|
497
|
-
'but no vision_config was resolved. Check conversion config.'
|
|
498
|
-
);
|
|
499
|
-
}
|
|
500
|
-
log.info('Pipeline', `Vision capable: imageTokenId=${imageTokenId}`);
|
|
501
|
-
} else {
|
|
502
|
-
this.visionCapable = false;
|
|
503
|
-
}
|
|
504
|
-
|
|
505
|
-
// Audio capability detection — gated by manifest fields
|
|
506
|
-
const audioTokenId = manifest.audio_token_id;
|
|
507
|
-
const hasAudioQuant = manifest.quantizationInfo?.audio != null;
|
|
508
|
-
if (Number.isInteger(audioTokenId) && audioTokenId > 0 && hasAudioQuant) {
|
|
509
|
-
this.audioCapable = true;
|
|
510
|
-
this.audioTokenId = audioTokenId;
|
|
511
|
-
this.audioConfig = this.modelConfig.audioConfig;
|
|
512
|
-
if (!this.audioConfig) {
|
|
513
|
-
throw new Error(
|
|
514
|
-
`Manifest declares audio_token_id=${audioTokenId} and quantizationInfo.audio ` +
|
|
515
|
-
'but no audio_config was resolved. Check conversion config.'
|
|
516
|
-
);
|
|
517
|
-
}
|
|
518
|
-
log.info('Pipeline', `Audio capable: audioTokenId=${audioTokenId}`);
|
|
519
|
-
} else {
|
|
520
|
-
this.audioCapable = false;
|
|
521
|
-
}
|
|
522
|
-
finishPipelineLoadTimingPhase(pipelineLoadTiming, 'configResolution', configResolutionStart);
|
|
523
|
-
|
|
524
|
-
await timedPipelineLoadPhase(
|
|
525
|
-
pipelineLoadTiming,
|
|
526
|
-
'kernelWarmup',
|
|
527
|
-
{ modelId: manifest.modelId ?? null },
|
|
528
|
-
() => runKernelWarmup({
|
|
529
|
-
useGPU: this.useGPU,
|
|
530
|
-
kernelWarmup: this.runtimeConfig.shared?.kernelWarmup,
|
|
531
|
-
modelConfig: this.modelConfig,
|
|
532
|
-
})
|
|
533
|
-
);
|
|
534
|
-
|
|
535
|
-
const executionSetupStart = performance.now();
|
|
536
|
-
// Phase 3: kernel path resolution + dtype contract
|
|
537
|
-
configResolutionPhase = 3;
|
|
538
|
-
log.debug('Pipeline', `Config resolution phase ${configResolutionPhase}: resolveKernelPathState`);
|
|
539
|
-
const kernelPathState = resolveKernelPathState({
|
|
540
|
-
manifest,
|
|
541
|
-
runtimeConfig: this.runtimeConfig,
|
|
542
|
-
runtimeOverrides: this.runtimeOverrides,
|
|
543
|
-
modelConfig: this.modelConfig,
|
|
544
|
-
});
|
|
545
|
-
this.resolvedKernelPath = kernelPathState.resolvedKernelPath;
|
|
546
|
-
this.kernelPathSource = kernelPathState.kernelPathSource;
|
|
547
|
-
this.runtimeConfig = kernelPathState.runtimeConfig;
|
|
548
|
-
|
|
549
|
-
// Phase 4: layer pipeline resolution
|
|
550
|
-
configResolutionPhase = 4;
|
|
551
|
-
log.debug('Pipeline', `Config resolution phase ${configResolutionPhase}: _resolveLayerPipeline`);
|
|
552
|
-
this._resolveLayerPipeline();
|
|
553
|
-
log.debug('Pipeline', `Config resolution complete (${configResolutionPhase} phases)`);
|
|
554
|
-
|
|
555
|
-
const cfg = this.modelConfig;
|
|
556
|
-
const moeStr = cfg.useMoE ? `, MoE(${cfg.numExperts}x${cfg.moeTopK})` : '';
|
|
557
|
-
const kernelInfo = this.resolvedKernelPath ? `kernelPath=${this.resolvedKernelPath.id}` : 'kernelPath=none';
|
|
558
|
-
log.info('Pipeline', `${cfg.numLayers}L/${cfg.hiddenSize}H/${cfg.numHeads}heads (${cfg.headDim}dim)${moeStr}, ${kernelInfo}`);
|
|
559
|
-
finishPipelineLoadTimingPhase(pipelineLoadTiming, 'executionSetup', executionSetupStart);
|
|
560
|
-
|
|
561
|
-
this.tokenizer = await timedPipelineLoadPhase(
|
|
562
|
-
pipelineLoadTiming,
|
|
563
|
-
'tokenizer',
|
|
564
|
-
{ modelId: manifest.modelId ?? null },
|
|
565
|
-
() => initTokenizerFromManifest(
|
|
566
|
-
manifest,
|
|
567
|
-
this.baseUrl,
|
|
568
|
-
this.storageContext
|
|
569
|
-
)
|
|
570
|
-
);
|
|
571
|
-
pipelineLoadTiming.details.tokenizer = typeof this.tokenizer?.getLoadTiming === 'function'
|
|
572
|
-
? this.tokenizer.getLoadTiming()
|
|
573
|
-
: null;
|
|
574
|
-
const tokenizerVocabSize = this.tokenizer.getVocabSize();
|
|
575
|
-
if (Number.isFinite(tokenizerVocabSize) && tokenizerVocabSize > 0) {
|
|
576
|
-
if (tokenizerVocabSize !== this.modelConfig.vocabSize) {
|
|
577
|
-
log.info('Pipeline', `Tokenizer vocabSize=${tokenizerVocabSize} differs from model=${this.modelConfig.vocabSize}, using model size`);
|
|
578
|
-
}
|
|
579
|
-
}
|
|
580
|
-
|
|
581
|
-
const postTokenizerExecutionSetupStart = performance.now();
|
|
582
|
-
// Manifest quantizationInfo.compute is the binding lane identity.
|
|
583
|
-
assertManifestComputeLaneBinding({ manifest, runtimeConfig: this.runtimeConfig });
|
|
584
|
-
|
|
585
|
-
// Initialize KV cache
|
|
586
|
-
if (this.modelConfig.decodeStrategy === 'replay_prefill') {
|
|
587
|
-
this.kvCache = null;
|
|
588
|
-
log.warn(
|
|
589
|
-
'Pipeline',
|
|
590
|
-
'Replay-prefill decode enabled for this model. Incremental KV-cache decode is disabled ' +
|
|
591
|
-
'because the model config did not resolve explicit layerTypes for mixed-geometry/shared-KV decode.'
|
|
592
|
-
);
|
|
593
|
-
} else {
|
|
594
|
-
this.kvCache = createKVCache(this.modelConfig, this.useGPU, this.debug, this.runtimeConfig.inference);
|
|
595
|
-
}
|
|
596
|
-
this.executionPlanState = compileExecutionPlanState({
|
|
597
|
-
runtimeConfig: this.runtimeConfig,
|
|
598
|
-
resolvedKernelPath: this.resolvedKernelPath,
|
|
599
|
-
kernelPathSource: this.kernelPathSource,
|
|
600
|
-
fallbackKernelPath: this.executionV1State?.fallbackKernelPath ?? null,
|
|
601
|
-
});
|
|
602
|
-
const activeExecutionPlan = resolveActiveExecutionPlan(this);
|
|
603
|
-
log.info(
|
|
604
|
-
'Pipeline',
|
|
605
|
-
`Execution plan: active=${activeExecutionPlan.id}, dtype=${activeExecutionPlan.activationDtype}, ` +
|
|
606
|
-
`kernelPath=${activeExecutionPlan.kernelPathId ?? 'none'}`
|
|
607
|
-
);
|
|
608
|
-
|
|
609
|
-
// Issue 1: Validate dtype consistency across all three resolution paths
|
|
610
|
-
// (execution plan, runtimeConfig.inference.compute, and layer context).
|
|
611
|
-
// The layer context is not yet built at this point, so pass null for it.
|
|
612
|
-
// This logs a warning if the execution plan and runtimeConfig disagree.
|
|
613
|
-
assertDtypeConsistency(this.executionPlanState, this.runtimeConfig, null);
|
|
614
|
-
|
|
615
|
-
const kpActivation = getKernelPathActivationDtype(this.resolvedKernelPath);
|
|
616
|
-
if (kpActivation && kpActivation !== activeExecutionPlan.activationDtype) {
|
|
617
|
-
throw new Error(
|
|
618
|
-
`Dtype contract violation: execution plan activationDtype="${activeExecutionPlan.activationDtype}" ` +
|
|
619
|
-
`but kernel path "${this.resolvedKernelPath.id}" declares activationDtype="${kpActivation}".`
|
|
620
|
-
);
|
|
621
|
-
}
|
|
622
|
-
|
|
623
|
-
// Initialize MoE router if needed
|
|
624
|
-
if (this.modelConfig.useMoE) {
|
|
625
|
-
this.moeRouter = new MoERouter({
|
|
626
|
-
numExperts: this.modelConfig.numExperts,
|
|
627
|
-
topK: this.modelConfig.moeTopK,
|
|
628
|
-
hiddenSize: this.modelConfig.hiddenSize,
|
|
629
|
-
normalizeWeights: this.runtimeConfig.inference.moe.routing.normalizeWeights,
|
|
630
|
-
});
|
|
631
|
-
}
|
|
632
|
-
|
|
633
|
-
// Initialize speculative decoder
|
|
634
|
-
if (manifest.draftModel) {
|
|
635
|
-
this.speculativeDecoder = initSpeculativeDecoder(
|
|
636
|
-
manifest,
|
|
637
|
-
this.runtimeConfig.inference.speculative
|
|
638
|
-
);
|
|
639
|
-
}
|
|
640
|
-
finishPipelineLoadTimingPhase(
|
|
641
|
-
pipelineLoadTiming,
|
|
642
|
-
'executionSetup',
|
|
643
|
-
postTokenizerExecutionSetupStart
|
|
644
|
-
);
|
|
645
|
-
|
|
646
|
-
// Load weights
|
|
647
|
-
await timedPipelineLoadPhase(
|
|
648
|
-
pipelineLoadTiming,
|
|
649
|
-
'loadWeights',
|
|
650
|
-
{ modelId: manifest.modelId ?? null },
|
|
651
|
-
() => this._loadWeights()
|
|
652
|
-
);
|
|
653
|
-
|
|
654
|
-
// Initialize RoPE frequencies
|
|
655
|
-
await timedPipelineLoadPhase(
|
|
656
|
-
pipelineLoadTiming,
|
|
657
|
-
'rope',
|
|
658
|
-
{ modelId: manifest.modelId ?? null },
|
|
659
|
-
() => this._initRoPE()
|
|
660
|
-
);
|
|
661
|
-
|
|
662
|
-
// Initialize conv layer states for gated short conv layers (LFM2)
|
|
663
|
-
await timedPipelineLoadPhase(
|
|
664
|
-
pipelineLoadTiming,
|
|
665
|
-
'convStates',
|
|
666
|
-
{ modelId: manifest.modelId ?? null },
|
|
667
|
-
() => this._initConvLayerStates()
|
|
668
|
-
);
|
|
669
|
-
|
|
670
|
-
this.isLoaded = true;
|
|
671
|
-
const loadMs = performance.now() - loadStart;
|
|
672
|
-
this.stats.modelLoadMs = loadMs;
|
|
673
|
-
pipelineLoadTiming.totalMs = roundPipelineTimingMs(loadMs);
|
|
674
|
-
pipelineLoadTiming.status = 'complete';
|
|
675
|
-
log.info('Pipeline', `Model loaded successfully (${loadMs.toFixed(0)}ms)`);
|
|
115
|
+
return loadModelImpl.call(this, manifest);
|
|
676
116
|
}
|
|
677
117
|
|
|
678
|
-
|
|
679
118
|
async _loadWeights() {
|
|
680
|
-
|
|
681
|
-
(this.manifest),
|
|
682
|
-
(this.modelConfig),
|
|
683
|
-
{
|
|
684
|
-
storageContext: this.storageContext ?? undefined,
|
|
685
|
-
loadingConfig: this.runtimeConfig.loading,
|
|
686
|
-
baseUrl: this.baseUrl ?? undefined,
|
|
687
|
-
resolvedKernelPath: this.resolvedKernelPath,
|
|
688
|
-
kernelPathSource: this.kernelPathSource,
|
|
689
|
-
keepF32Weights: this.runtimeConfig.inference.compute.keepF32Weights === true,
|
|
690
|
-
loaderDebug: this.runtimeConfig?.shared?.debug?.loader ?? null,
|
|
691
|
-
perLayerInputSession: resolvePerLayerInputsSession(
|
|
692
|
-
this.modelConfig.perLayerInputsSession ?? null,
|
|
693
|
-
this.runtimeConfig?.inference?.session?.perLayerInputs ?? null
|
|
694
|
-
),
|
|
695
|
-
loader: this.dopplerLoader ?? undefined,
|
|
696
|
-
onProgress: (info) => {
|
|
697
|
-
if (info.stage !== 'layers' && info.stage !== 'shards') {
|
|
698
|
-
log.verbose('Loader', `${info.stage}: ${Math.round(info.progress * 100)}%${info.message ? ` - ${info.message}` : ''}`);
|
|
699
|
-
}
|
|
700
|
-
if (this._onProgress) {
|
|
701
|
-
this._onProgress({
|
|
702
|
-
percent: info.progress * 100,
|
|
703
|
-
message: info.message,
|
|
704
|
-
stage: info.stage,
|
|
705
|
-
layer: info.layer,
|
|
706
|
-
total: info.total,
|
|
707
|
-
});
|
|
708
|
-
}
|
|
709
|
-
},
|
|
710
|
-
}
|
|
711
|
-
);
|
|
712
|
-
|
|
713
|
-
result.layerWeights.forEach((w, k) => this.weights.set(k, w));
|
|
714
|
-
this.weights.set('embed', result.embeddings);
|
|
715
|
-
this.weights.set('lm_head', result.lmHead);
|
|
716
|
-
this.weights.set('lm_head_bias', result.lmHeadBias);
|
|
717
|
-
if (result.lmHeadBias) {
|
|
718
|
-
// Recorded/fused logits currently keep their result on GPU, while the
|
|
719
|
-
// manifest-owned decoder bias is loaded as an exact CPU tensor. Route
|
|
720
|
-
// biased heads through computeLogits until the execution graph declares
|
|
721
|
-
// and records an explicit GPU bias-add step.
|
|
722
|
-
this.disableRecordedLogits = true;
|
|
723
|
-
this.disableFusedDecode = true;
|
|
724
|
-
}
|
|
725
|
-
this.weights.set('final_norm', result.finalNorm);
|
|
726
|
-
this.weights.set('final_norm_bias', result.finalNormBias);
|
|
727
|
-
if (result.finalNormBias) {
|
|
728
|
-
this.disableRecordedLogits = true;
|
|
729
|
-
this.disableFusedDecode = true;
|
|
730
|
-
}
|
|
731
|
-
this.weights.set('diffusion_gemma_self_conditioning', result.diffusionGemmaSelfConditioning);
|
|
732
|
-
this.weights.set('per_layer_inputs', result.perLayerInputWeights);
|
|
733
|
-
this.embeddingPostprocessor = result.embeddingPostprocessor;
|
|
734
|
-
|
|
735
|
-
this.layerRouterWeights = result.layerRouterWeights;
|
|
736
|
-
|
|
737
|
-
this.dopplerLoader = result.loader ?? getDopplerLoader(this.runtimeConfig.loading);
|
|
738
|
-
this.stats.loadTiming = result.loadTiming ?? this.dopplerLoader?.getLoadTiming?.() ?? null;
|
|
739
|
-
|
|
740
|
-
if ((this.modelConfig).useMoE && this.moeRouter) {
|
|
741
|
-
this.moeRouter = initMoERouter(
|
|
742
|
-
(this.modelConfig),
|
|
743
|
-
this.runtimeConfig.inference.moe.routing,
|
|
744
|
-
result.layerWeights
|
|
745
|
-
);
|
|
746
|
-
}
|
|
747
|
-
|
|
748
|
-
if (this.useGPU && this.modelConfig) {
|
|
749
|
-
const session = this.runtimeConfig?.inference?.session ?? null;
|
|
750
|
-
fuseQKVWeights(result.layerWeights, this.modelConfig, this.resolvedKernelPath, {
|
|
751
|
-
allowQ4K: session?.useFusedQKVSplitQKNorm === true
|
|
752
|
-
|| session?.useFusedQKVSplitQKNormRoPE === true,
|
|
753
|
-
});
|
|
754
|
-
}
|
|
755
|
-
|
|
756
|
-
if (this.useGPU && this.modelConfig) {
|
|
757
|
-
const activeExecutionPlan = resolveActiveExecutionPlan(this);
|
|
758
|
-
try {
|
|
759
|
-
this.decodeBuffers?.ensureBuffers({
|
|
760
|
-
hiddenSize: this.modelConfig.hiddenSize,
|
|
761
|
-
intermediateSize: this.modelConfig.maxIntermediateSize,
|
|
762
|
-
activationDtype: activeExecutionPlan.activationDtype,
|
|
763
|
-
enablePingPong: true,
|
|
764
|
-
});
|
|
765
|
-
|
|
766
|
-
const device = getDevice();
|
|
767
|
-
if (device) {
|
|
768
|
-
this.finitenessBuffer = device.createBuffer({
|
|
769
|
-
label: 'finiteness_status',
|
|
770
|
-
size: 16,
|
|
771
|
-
usage: GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_SRC | GPUBufferUsage.COPY_DST,
|
|
772
|
-
});
|
|
773
|
-
}
|
|
774
|
-
} catch (bufferError) {
|
|
775
|
-
this.decodeBuffers?.release();
|
|
776
|
-
if (this.finitenessBuffer) {
|
|
777
|
-
this.finitenessBuffer.destroy();
|
|
778
|
-
this.finitenessBuffer = null;
|
|
779
|
-
}
|
|
780
|
-
throw bufferError;
|
|
781
|
-
}
|
|
782
|
-
}
|
|
119
|
+
return _loadWeightsImpl.call(this);
|
|
783
120
|
}
|
|
784
121
|
|
|
785
|
-
|
|
786
122
|
setPreloadedWeights(weights) {
|
|
787
123
|
this._preloadedWeights = weights;
|
|
788
124
|
}
|
|
789
125
|
|
|
790
|
-
|
|
791
126
|
async _initRoPE() {
|
|
792
|
-
|
|
793
|
-
const maxSeqLen = config.maxSeqLen;
|
|
794
|
-
const ropeBuffers = await initRoPEFrequencies({
|
|
795
|
-
headDim: config.globalHeadDim ?? config.headDim,
|
|
796
|
-
localHeadDim: config.headDim,
|
|
797
|
-
rotaryDim: config.ropeRotaryDim,
|
|
798
|
-
ropeLocalRotaryDim: config.ropeLocalRotaryDim,
|
|
799
|
-
ropeFrequencyBaseDim: config.ropeFrequencyBaseDim,
|
|
800
|
-
ropeLocalFrequencyBaseDim: config.ropeLocalFrequencyBaseDim,
|
|
801
|
-
maxSeqLen,
|
|
802
|
-
ropeTheta: config.ropeTheta,
|
|
803
|
-
ropeLocalTheta: config.ropeLocalTheta,
|
|
804
|
-
mropeInterleaved: config.mropeInterleaved,
|
|
805
|
-
mropeSection: config.mropeSection,
|
|
806
|
-
partialRotaryFactor: config.partialRotaryFactor,
|
|
807
|
-
ropeLocalPartialRotaryFactor: config.ropeLocalPartialRotaryFactor,
|
|
808
|
-
ropeScale: config.ropeScale,
|
|
809
|
-
ropeLocalScale: config.ropeLocalScale,
|
|
810
|
-
ropeScalingType: config.ropeScalingType,
|
|
811
|
-
ropeLocalScalingType: config.ropeLocalScalingType,
|
|
812
|
-
ropeScaling: config.ropeScaling,
|
|
813
|
-
ropeLocalScaling: config.ropeLocalScaling,
|
|
814
|
-
}, this.useGPU);
|
|
815
|
-
this.ropeFreqsCos = ropeBuffers.cos;
|
|
816
|
-
this.ropeFreqsSin = ropeBuffers.sin;
|
|
817
|
-
this.ropeLocalCos = ropeBuffers.localCos ?? null;
|
|
818
|
-
this.ropeLocalSin = ropeBuffers.localSin ?? null;
|
|
127
|
+
return _initRoPEImpl.call(this);
|
|
819
128
|
}
|
|
820
129
|
|
|
821
|
-
|
|
822
130
|
async _initConvLayerStates() {
|
|
823
|
-
|
|
824
|
-
if (!config?.layerTypes) return;
|
|
825
|
-
const { getDevice } = await import('../../gpu/device.js');
|
|
826
|
-
const device = getDevice();
|
|
827
|
-
if (!device) return;
|
|
828
|
-
|
|
829
|
-
const hiddenSize = config.hiddenSize;
|
|
830
|
-
const convStates = new Map();
|
|
831
|
-
|
|
832
|
-
for (let i = 0; i < config.layerTypes.length; i++) {
|
|
833
|
-
const lt = String(config.layerTypes[i] ?? '').toLowerCase();
|
|
834
|
-
if (lt !== 'conv' && lt !== 'convolution') continue;
|
|
835
|
-
|
|
836
|
-
const layerWeights = this.weights.get(`layer_${i}`);
|
|
837
|
-
if (!layerWeights) continue;
|
|
838
|
-
const convKernel = layerWeights?.convKernel;
|
|
839
|
-
if (!convKernel) continue;
|
|
840
|
-
|
|
841
|
-
const convState = {};
|
|
842
|
-
try {
|
|
843
|
-
await initConvLayerState(
|
|
844
|
-
convState,
|
|
845
|
-
convKernel,
|
|
846
|
-
layerWeights.convInProj ?? null,
|
|
847
|
-
hiddenSize,
|
|
848
|
-
`L${i}.conv`,
|
|
849
|
-
i
|
|
850
|
-
);
|
|
851
|
-
if (!convState.convWeightGPU || !convState.convStateGPU) {
|
|
852
|
-
continue;
|
|
853
|
-
}
|
|
854
|
-
convStates.set(i, convState);
|
|
855
|
-
} catch (e) {
|
|
856
|
-
log.warn('Pipeline', `Conv layer ${i} state init failed: ${e.message}`);
|
|
857
|
-
}
|
|
858
|
-
}
|
|
859
|
-
|
|
860
|
-
if (convStates.size > 0) {
|
|
861
|
-
this.convLayerStates = convStates;
|
|
862
|
-
log.info('Pipeline', `Initialized ${convStates.size} conv layer states (kernelSize=${convStates.values().next().value?.kernelSize})`);
|
|
863
|
-
}
|
|
131
|
+
return _initConvLayerStatesImpl.call(this);
|
|
864
132
|
}
|
|
865
133
|
|
|
866
|
-
|
|
867
134
|
async _loadVisionWeights() {
|
|
868
|
-
|
|
869
|
-
const vc = this.visionConfig;
|
|
870
|
-
const depth = vc.depth;
|
|
871
|
-
|
|
872
|
-
const loadRequiredTensor = async (name, toGPU = true) => {
|
|
873
|
-
const tensor = await loader.loadTensor(name, toGPU, true);
|
|
874
|
-
if (!tensor) {
|
|
875
|
-
throw new Error(`Vision tensor "${name}" is missing from the converted artifact.`);
|
|
876
|
-
}
|
|
877
|
-
return tensor;
|
|
878
|
-
};
|
|
879
|
-
const loadScalar = async (name) => {
|
|
880
|
-
const tensor = await loadRequiredTensor(name, false);
|
|
881
|
-
if (tensor instanceof Float32Array) {
|
|
882
|
-
if (tensor.length !== 1) {
|
|
883
|
-
throw new Error(`Vision scalar "${name}" must be a single-element tensor, got length=${tensor.length}.`);
|
|
884
|
-
}
|
|
885
|
-
return tensor[0];
|
|
886
|
-
}
|
|
887
|
-
if (ArrayBuffer.isView(tensor) && tensor.length === 1) {
|
|
888
|
-
return Number(tensor[0]);
|
|
889
|
-
}
|
|
890
|
-
if (typeof tensor === 'number') {
|
|
891
|
-
return tensor;
|
|
892
|
-
}
|
|
893
|
-
throw new Error(
|
|
894
|
-
`Vision scalar "${name}" must decode to a single numeric value, ` +
|
|
895
|
-
`got ${tensor?.constructor?.name ?? typeof tensor} length=${tensor?.length ?? 'N/A'}.`
|
|
896
|
-
);
|
|
897
|
-
};
|
|
898
|
-
const loadClipRange = async (prefix) => ({
|
|
899
|
-
inputMin: await loadScalar(`${prefix}.input_min`),
|
|
900
|
-
inputMax: await loadScalar(`${prefix}.input_max`),
|
|
901
|
-
outputMin: await loadScalar(`${prefix}.output_min`),
|
|
902
|
-
outputMax: await loadScalar(`${prefix}.output_max`),
|
|
903
|
-
});
|
|
904
|
-
|
|
905
|
-
if (vc.visionArchitecture === 'gemma4') {
|
|
906
|
-
const isEncoderFree = (depth === 0);
|
|
907
|
-
const visionWeights = {
|
|
908
|
-
textHiddenSize: this.modelConfig.hiddenSize,
|
|
909
|
-
patchInputProj: await loadRequiredTensor('model.vision_tower.patch_embedder.input_proj.weight'),
|
|
910
|
-
patchPositionEmbeddingTable: await loadRequiredTensor('model.vision_tower.patch_embedder.position_embedding_table', false),
|
|
911
|
-
projector: isEncoderFree
|
|
912
|
-
? await loader.loadTensor('model.embed_vision.embedding_projection.weight', true, true)
|
|
913
|
-
: await loadRequiredTensor('model.embed_vision.embedding_projection.weight'),
|
|
914
|
-
layers: [],
|
|
915
|
-
};
|
|
916
|
-
|
|
917
|
-
if (!(visionWeights.patchPositionEmbeddingTable instanceof Float32Array)) {
|
|
918
|
-
throw new Error(
|
|
919
|
-
'Gemma 4 vision position_embedding_table must decode to Float32Array on CPU. ' +
|
|
920
|
-
'Re-convert the artifact if this tensor was quantized incorrectly.'
|
|
921
|
-
);
|
|
922
|
-
}
|
|
923
|
-
|
|
924
|
-
for (let i = 0; i < depth; i++) {
|
|
925
|
-
const prefix = `model.vision_tower.encoder.layers.${i}`;
|
|
926
|
-
const attnPrefix = `${prefix}.self_attn`;
|
|
927
|
-
const mlpPrefix = `${prefix}.mlp`;
|
|
928
|
-
visionWeights.layers.push({
|
|
929
|
-
inputLayerNorm: await loadRequiredTensor(`${prefix}.input_layernorm.weight`),
|
|
930
|
-
postAttentionLayerNorm: await loadRequiredTensor(`${prefix}.post_attention_layernorm.weight`),
|
|
931
|
-
preFeedforwardLayerNorm: await loadRequiredTensor(`${prefix}.pre_feedforward_layernorm.weight`),
|
|
932
|
-
postFeedforwardLayerNorm: await loadRequiredTensor(`${prefix}.post_feedforward_layernorm.weight`),
|
|
933
|
-
qNorm: await loadRequiredTensor(`${attnPrefix}.q_norm.weight`),
|
|
934
|
-
kNorm: await loadRequiredTensor(`${attnPrefix}.k_norm.weight`),
|
|
935
|
-
qProj: await loadRequiredTensor(`${attnPrefix}.q_proj.linear.weight`),
|
|
936
|
-
kProj: await loadRequiredTensor(`${attnPrefix}.k_proj.linear.weight`),
|
|
937
|
-
vProj: await loadRequiredTensor(`${attnPrefix}.v_proj.linear.weight`),
|
|
938
|
-
oProj: await loadRequiredTensor(`${attnPrefix}.o_proj.linear.weight`),
|
|
939
|
-
qProjClip: await loadClipRange(`${attnPrefix}.q_proj`),
|
|
940
|
-
kProjClip: await loadClipRange(`${attnPrefix}.k_proj`),
|
|
941
|
-
vProjClip: await loadClipRange(`${attnPrefix}.v_proj`),
|
|
942
|
-
oProjClip: await loadClipRange(`${attnPrefix}.o_proj`),
|
|
943
|
-
gateProj: await loadRequiredTensor(`${mlpPrefix}.gate_proj.linear.weight`),
|
|
944
|
-
upProj: await loadRequiredTensor(`${mlpPrefix}.up_proj.linear.weight`),
|
|
945
|
-
downProj: await loadRequiredTensor(`${mlpPrefix}.down_proj.linear.weight`),
|
|
946
|
-
gateProjClip: await loadClipRange(`${mlpPrefix}.gate_proj`),
|
|
947
|
-
upProjClip: await loadClipRange(`${mlpPrefix}.up_proj`),
|
|
948
|
-
downProjClip: await loadClipRange(`${mlpPrefix}.down_proj`),
|
|
949
|
-
});
|
|
950
|
-
}
|
|
951
|
-
|
|
952
|
-
this.visionWeights = visionWeights;
|
|
953
|
-
log.info('Pipeline', `Vision weights loaded (${depth} Gemma 4 encoder layers)`);
|
|
954
|
-
return;
|
|
955
|
-
}
|
|
956
|
-
|
|
957
|
-
const visionWeights = {};
|
|
958
|
-
|
|
959
|
-
// Patch embedding weights
|
|
960
|
-
const patchProjName = 'visual.patch_embed.proj.weight';
|
|
961
|
-
const patchProjBiasName = 'visual.patch_embed.proj.bias';
|
|
962
|
-
visionWeights.patchProjWeight = await loader.loadTensor(patchProjName, true, true);
|
|
963
|
-
visionWeights.patchProjBias = await loader.loadTensor(patchProjBiasName, true, true);
|
|
964
|
-
|
|
965
|
-
// Vision encoder layer weights
|
|
966
|
-
visionWeights.layers = [];
|
|
967
|
-
for (let i = 0; i < depth; i++) {
|
|
968
|
-
const prefix = `visual.blocks.${i}`;
|
|
969
|
-
const layerW = {
|
|
970
|
-
norm1Weight: await loader.loadTensor(`${prefix}.norm1.weight`, true, true),
|
|
971
|
-
norm2Weight: await loader.loadTensor(`${prefix}.norm2.weight`, true, true),
|
|
972
|
-
qkvWeight: await loader.loadTensor(`${prefix}.attn.qkv.weight`, true, true),
|
|
973
|
-
qkvBias: await loader.loadTensor(`${prefix}.attn.qkv.bias`, true, true),
|
|
974
|
-
projWeight: await loader.loadTensor(`${prefix}.attn.proj.weight`, true, true),
|
|
975
|
-
projBias: await loader.loadTensor(`${prefix}.attn.proj.bias`, true, true),
|
|
976
|
-
fc1Weight: await loader.loadTensor(`${prefix}.mlp.fc1.weight`, true, true),
|
|
977
|
-
fc1Bias: await loader.loadTensor(`${prefix}.mlp.fc1.bias`, true, true),
|
|
978
|
-
fc2Weight: await loader.loadTensor(`${prefix}.mlp.fc2.weight`, true, true),
|
|
979
|
-
fc2Bias: await loader.loadTensor(`${prefix}.mlp.fc2.bias`, true, true),
|
|
980
|
-
};
|
|
981
|
-
visionWeights.layers.push(layerW);
|
|
982
|
-
}
|
|
983
|
-
|
|
984
|
-
// Spatial merge projection
|
|
985
|
-
visionWeights.mergerLnWeight = await loader.loadTensor('visual.merger.ln_q.weight', true, true);
|
|
986
|
-
visionWeights.mergerMlp0Weight = await loader.loadTensor('visual.merger.mlp.0.weight', true, true);
|
|
987
|
-
visionWeights.mergerMlp0Bias = await loader.loadTensor('visual.merger.mlp.0.bias', true, true);
|
|
988
|
-
visionWeights.mergerMlp2Weight = await loader.loadTensor('visual.merger.mlp.2.weight', true, true);
|
|
989
|
-
visionWeights.mergerMlp2Bias = await loader.loadTensor('visual.merger.mlp.2.bias', true, true);
|
|
990
|
-
|
|
991
|
-
this.visionWeights = visionWeights;
|
|
992
|
-
log.info('Pipeline', `Vision weights loaded (${depth} encoder layers)`);
|
|
135
|
+
return _loadVisionWeightsImpl.call(this);
|
|
993
136
|
}
|
|
994
137
|
|
|
995
138
|
async _ensureVisionWeightsLoaded() {
|
|
996
|
-
|
|
997
|
-
throw new Error(
|
|
998
|
-
'Pipeline does not support vision weights (no image_token_id in manifest).'
|
|
999
|
-
);
|
|
1000
|
-
}
|
|
1001
|
-
if (this.visionWeights) {
|
|
1002
|
-
return;
|
|
1003
|
-
}
|
|
1004
|
-
log.info('Pipeline', 'Loading vision weights on demand');
|
|
1005
|
-
await this._loadVisionWeights();
|
|
139
|
+
return _ensureVisionWeightsLoadedImpl.call(this);
|
|
1006
140
|
}
|
|
1007
141
|
|
|
1008
|
-
|
|
1009
142
|
async _loadAudioWeights() {
|
|
1010
|
-
|
|
1011
|
-
const ac = this.audioConfig;
|
|
1012
|
-
const depth = ac.depth;
|
|
1013
|
-
|
|
1014
|
-
const loadRequiredTensor = async (name, toGPU = true) => {
|
|
1015
|
-
const tensor = await loader.loadTensor(name, toGPU, true);
|
|
1016
|
-
if (!tensor) {
|
|
1017
|
-
throw new Error(`Audio tensor "${name}" is missing from the converted artifact.`);
|
|
1018
|
-
}
|
|
1019
|
-
return tensor;
|
|
1020
|
-
};
|
|
1021
|
-
const loadScalar = async (name) => {
|
|
1022
|
-
const tensor = await loadRequiredTensor(name, false);
|
|
1023
|
-
if (tensor instanceof Float32Array) {
|
|
1024
|
-
if (tensor.length !== 1) {
|
|
1025
|
-
throw new Error(`Audio scalar "${name}" must be a single-element tensor, got length=${tensor.length}.`);
|
|
1026
|
-
}
|
|
1027
|
-
return tensor[0];
|
|
1028
|
-
}
|
|
1029
|
-
if (ArrayBuffer.isView(tensor) && tensor.length === 1) {
|
|
1030
|
-
return Number(tensor[0]);
|
|
1031
|
-
}
|
|
1032
|
-
if (typeof tensor === 'number') {
|
|
1033
|
-
return tensor;
|
|
1034
|
-
}
|
|
1035
|
-
throw new Error(
|
|
1036
|
-
`Audio scalar "${name}" must decode to a single numeric value, ` +
|
|
1037
|
-
`got ${tensor?.constructor?.name ?? typeof tensor} length=${tensor?.length ?? 'N/A'}.`
|
|
1038
|
-
);
|
|
1039
|
-
};
|
|
1040
|
-
const loadClipRange = async (prefix) => ({
|
|
1041
|
-
inputMin: await loadScalar(`${prefix}.input_min`),
|
|
1042
|
-
inputMax: await loadScalar(`${prefix}.input_max`),
|
|
1043
|
-
outputMin: await loadScalar(`${prefix}.output_min`),
|
|
1044
|
-
outputMax: await loadScalar(`${prefix}.output_max`),
|
|
1045
|
-
});
|
|
1046
|
-
|
|
1047
|
-
const isEncoderFree = (depth === 0);
|
|
1048
|
-
const audioWeights = {
|
|
1049
|
-
// Subsampling
|
|
1050
|
-
subsampleConv0Weight: isEncoderFree ? null : await loadRequiredTensor('model.audio_tower.subsample_conv_projection.layer0.conv.weight'),
|
|
1051
|
-
subsampleNorm0Weight: isEncoderFree ? null : await loadRequiredTensor('model.audio_tower.subsample_conv_projection.layer0.norm.weight'),
|
|
1052
|
-
subsampleConv1Weight: isEncoderFree ? null : await loadRequiredTensor('model.audio_tower.subsample_conv_projection.layer1.conv.weight'),
|
|
1053
|
-
subsampleNorm1Weight: isEncoderFree ? null : await loadRequiredTensor('model.audio_tower.subsample_conv_projection.layer1.norm.weight'),
|
|
1054
|
-
subsampleInputProjWeight: isEncoderFree ? null : await loadRequiredTensor('model.audio_tower.subsample_conv_projection.input_proj_linear.weight'),
|
|
1055
|
-
// Output
|
|
1056
|
-
outputProjWeight: isEncoderFree ? null : await loadRequiredTensor('model.audio_tower.output_proj.weight'),
|
|
1057
|
-
outputProjBias: isEncoderFree ? null : await loadRequiredTensor('model.audio_tower.output_proj.bias'),
|
|
1058
|
-
audioEmbeddingProjWeight: await loadRequiredTensor('model.embed_audio.embedding_projection.weight'),
|
|
1059
|
-
layers: [],
|
|
1060
|
-
};
|
|
1061
|
-
|
|
1062
|
-
for (let i = 0; i < depth; i++) {
|
|
1063
|
-
const prefix = `model.audio_tower.layers.${i}`;
|
|
1064
|
-
const layer = {
|
|
1065
|
-
// Feed-forward 1 (Macaron half-step)
|
|
1066
|
-
feedForward1: {
|
|
1067
|
-
preLayerNorm: await loadRequiredTensor(`${prefix}.feed_forward1.pre_layer_norm.weight`),
|
|
1068
|
-
ffwLayer1Weight: await loadRequiredTensor(`${prefix}.feed_forward1.ffw_layer_1.linear.weight`),
|
|
1069
|
-
ffwLayer1Clip: await loadClipRange(`${prefix}.feed_forward1.ffw_layer_1`),
|
|
1070
|
-
ffwLayer2Weight: await loadRequiredTensor(`${prefix}.feed_forward1.ffw_layer_2.linear.weight`),
|
|
1071
|
-
ffwLayer2Clip: await loadClipRange(`${prefix}.feed_forward1.ffw_layer_2`),
|
|
1072
|
-
postLayerNorm: await loadRequiredTensor(`${prefix}.feed_forward1.post_layer_norm.weight`),
|
|
1073
|
-
},
|
|
1074
|
-
// Self-attention
|
|
1075
|
-
normPreAttn: await loadRequiredTensor(`${prefix}.norm_pre_attn.weight`),
|
|
1076
|
-
qProj: await loadRequiredTensor(`${prefix}.self_attn.q_proj.linear.weight`),
|
|
1077
|
-
qProjClip: await loadClipRange(`${prefix}.self_attn.q_proj`),
|
|
1078
|
-
kProj: await loadRequiredTensor(`${prefix}.self_attn.k_proj.linear.weight`),
|
|
1079
|
-
kProjClip: await loadClipRange(`${prefix}.self_attn.k_proj`),
|
|
1080
|
-
vProj: await loadRequiredTensor(`${prefix}.self_attn.v_proj.linear.weight`),
|
|
1081
|
-
vProjClip: await loadClipRange(`${prefix}.self_attn.v_proj`),
|
|
1082
|
-
perDimScale: await loadRequiredTensor(`${prefix}.self_attn.per_dim_scale`),
|
|
1083
|
-
relativeKProj: await loadRequiredTensor(`${prefix}.self_attn.relative_k_proj.weight`),
|
|
1084
|
-
postProj: await loadRequiredTensor(`${prefix}.self_attn.post.linear.weight`),
|
|
1085
|
-
postProjClip: await loadClipRange(`${prefix}.self_attn.post`),
|
|
1086
|
-
normPostAttn: await loadRequiredTensor(`${prefix}.norm_post_attn.weight`),
|
|
1087
|
-
// Convolution module (LConv1D)
|
|
1088
|
-
lconvPreLayerNorm: await loadRequiredTensor(`${prefix}.lconv1d.pre_layer_norm.weight`),
|
|
1089
|
-
lconvLinearStartWeight: await loadRequiredTensor(`${prefix}.lconv1d.linear_start.linear.weight`),
|
|
1090
|
-
lconvLinearStartClip: await loadClipRange(`${prefix}.lconv1d.linear_start`),
|
|
1091
|
-
lconvDepthwiseWeight: await loadRequiredTensor(`${prefix}.lconv1d.depthwise_conv1d.weight`),
|
|
1092
|
-
lconvConvNorm: await loadRequiredTensor(`${prefix}.lconv1d.conv_norm.weight`),
|
|
1093
|
-
lconvLinearEndWeight: await loadRequiredTensor(`${prefix}.lconv1d.linear_end.linear.weight`),
|
|
1094
|
-
lconvLinearEndClip: await loadClipRange(`${prefix}.lconv1d.linear_end`),
|
|
1095
|
-
// Feed-forward 2 (Macaron half-step)
|
|
1096
|
-
feedForward2: {
|
|
1097
|
-
preLayerNorm: await loadRequiredTensor(`${prefix}.feed_forward2.pre_layer_norm.weight`),
|
|
1098
|
-
ffwLayer1Weight: await loadRequiredTensor(`${prefix}.feed_forward2.ffw_layer_1.linear.weight`),
|
|
1099
|
-
ffwLayer1Clip: await loadClipRange(`${prefix}.feed_forward2.ffw_layer_1`),
|
|
1100
|
-
ffwLayer2Weight: await loadRequiredTensor(`${prefix}.feed_forward2.ffw_layer_2.linear.weight`),
|
|
1101
|
-
ffwLayer2Clip: await loadClipRange(`${prefix}.feed_forward2.ffw_layer_2`),
|
|
1102
|
-
postLayerNorm: await loadRequiredTensor(`${prefix}.feed_forward2.post_layer_norm.weight`),
|
|
1103
|
-
},
|
|
1104
|
-
// Final layer norm
|
|
1105
|
-
normOut: await loadRequiredTensor(`${prefix}.norm_out.weight`),
|
|
1106
|
-
};
|
|
1107
|
-
audioWeights.layers.push(layer);
|
|
1108
|
-
}
|
|
1109
|
-
|
|
1110
|
-
this.audioWeights = audioWeights;
|
|
1111
|
-
log.info('Pipeline', `Audio weights loaded (${depth} conformer layers)`);
|
|
143
|
+
return _loadAudioWeightsImpl.call(this);
|
|
1112
144
|
}
|
|
1113
145
|
|
|
1114
146
|
async _ensureAudioWeightsLoaded() {
|
|
1115
|
-
|
|
1116
|
-
throw new Error(
|
|
1117
|
-
'Pipeline does not support audio weights (no audio_token_id in manifest).'
|
|
1118
|
-
);
|
|
1119
|
-
}
|
|
1120
|
-
if (this.audioWeights) {
|
|
1121
|
-
return;
|
|
1122
|
-
}
|
|
1123
|
-
log.info('Pipeline', 'Loading audio weights on demand');
|
|
1124
|
-
await this._loadAudioWeights();
|
|
147
|
+
return _ensureAudioWeightsLoadedImpl.call(this);
|
|
1125
148
|
}
|
|
1126
149
|
|
|
1127
|
-
|
|
1128
150
|
// ==========================================================================
|
|
1129
151
|
// Vision: transcribeImage
|
|
1130
152
|
// ==========================================================================
|
|
1131
153
|
|
|
1132
|
-
/**
|
|
1133
|
-
* Transcribe text from an image using the vision encoder and text decoder.
|
|
1134
|
-
*
|
|
1135
|
-
* @param {object} params
|
|
1136
|
-
* @param {Uint8Array|Float32Array} params.imageBytes Raw image pixel data
|
|
1137
|
-
* @param {number} params.width Image width
|
|
1138
|
-
* @param {number} params.height Image height
|
|
1139
|
-
* @param {string} [params.prompt] Custom transcription prompt
|
|
1140
|
-
* @param {number} [params.maxTokens] Max tokens to generate
|
|
1141
|
-
* @param {number} [params.softTokenBudget] Per-request soft token budget (Gemma 4 tiers: 70/140/280/560/1120)
|
|
1142
|
-
* @returns {Promise<{ text: string, tokens: number[] }>}
|
|
1143
|
-
*/
|
|
1144
154
|
async transcribeImage({ imageBytes, width, height, prompt, maxTokens, softTokenBudget, signal }) {
|
|
1145
|
-
|
|
1146
|
-
if (!this.visionCapable) {
|
|
1147
|
-
throw new Error(
|
|
1148
|
-
'Pipeline does not support image transcription (no image_token_id in manifest).'
|
|
1149
|
-
);
|
|
1150
|
-
}
|
|
1151
|
-
await this._ensureVisionWeightsLoaded();
|
|
1152
|
-
assertNotAborted(signal);
|
|
1153
|
-
|
|
1154
|
-
this.reset();
|
|
1155
|
-
|
|
1156
|
-
// Lazy-load vision module (avoids GPU kernel dependency for text-only pipelines)
|
|
1157
|
-
const { encodeImage } = await import('./vision/index.js');
|
|
1158
|
-
|
|
1159
|
-
// Step 1: Encode image through vision pipeline
|
|
1160
|
-
const encodeResult = await encodeImage({
|
|
1161
|
-
pixels: imageBytes,
|
|
1162
|
-
width,
|
|
1163
|
-
height,
|
|
1164
|
-
visionConfig: this.visionConfig,
|
|
1165
|
-
weights: this.visionWeights,
|
|
1166
|
-
softTokenBudget,
|
|
1167
|
-
});
|
|
1168
|
-
|
|
1169
|
-
// Step 2: Build the multimodal prompt from the model's chat template and
|
|
1170
|
-
// expand the single <|image|> placeholder into the exact visual-token span.
|
|
1171
|
-
const requestedPrompt = prompt ?? 'Describe the image in one short sentence.';
|
|
1172
|
-
const imageTokenId = this.visionConfig?.imageTokenId ?? this.modelConfig?.imageTokenId;
|
|
1173
|
-
if (imageTokenId == null) {
|
|
1174
|
-
throw new Error(
|
|
1175
|
-
'Pipeline missing image_token_id. Re-convert the model with image token metadata.'
|
|
1176
|
-
);
|
|
1177
|
-
}
|
|
1178
|
-
if (this.visionConfig?.visionArchitecture !== 'gemma4') {
|
|
1179
|
-
throw new Error(
|
|
1180
|
-
`[Pipeline] transcribeImage: unsupported vision architecture "${this.visionConfig?.visionArchitecture ?? 'unknown'}". ` +
|
|
1181
|
-
'This runtime path currently requires Gemma 4 multimodal prompt expansion.'
|
|
1182
|
-
);
|
|
1183
|
-
}
|
|
1184
|
-
const templateType = this.modelConfig?.chatTemplateType ?? 'gemma4';
|
|
1185
|
-
const chatOptions = this.modelConfig?.chatTemplateThinking === true ? { thinking: true } : undefined;
|
|
1186
|
-
const multimodalPrompt = formatChatMessages([
|
|
1187
|
-
{
|
|
1188
|
-
role: 'user',
|
|
1189
|
-
content: [
|
|
1190
|
-
{ type: 'image' },
|
|
1191
|
-
{ type: 'text', text: requestedPrompt },
|
|
1192
|
-
],
|
|
1193
|
-
},
|
|
1194
|
-
], templateType, chatOptions);
|
|
1195
|
-
const promptTokenIds = this.tokenizer.encode(multimodalPrompt);
|
|
1196
|
-
const imageTokenSpanLength = encodeResult.numTokens;
|
|
1197
|
-
const effectiveBudget = softTokenBudget ?? this.visionConfig?.defaultOutputLength;
|
|
1198
|
-
const maxImageTokenSpanLength = Number(effectiveBudget);
|
|
1199
|
-
if (!Number.isFinite(maxImageTokenSpanLength) || maxImageTokenSpanLength < 1 || Math.floor(maxImageTokenSpanLength) !== maxImageTokenSpanLength) {
|
|
1200
|
-
throw new Error(
|
|
1201
|
-
`[Pipeline] transcribeImage: invalid soft token budget ${effectiveBudget}. ` +
|
|
1202
|
-
'Expected a positive integer from the resolved vision config or softTokenBudget parameter.'
|
|
1203
|
-
);
|
|
1204
|
-
}
|
|
1205
|
-
if (imageTokenSpanLength > maxImageTokenSpanLength) {
|
|
1206
|
-
throw new Error(
|
|
1207
|
-
`[Pipeline] transcribeImage: encoded Gemma 4 image produced ${imageTokenSpanLength} soft tokens, ` +
|
|
1208
|
-
`which exceeds the effective soft token budget=${maxImageTokenSpanLength}.`
|
|
1209
|
-
);
|
|
1210
|
-
}
|
|
1211
|
-
const boiTokenId = resolveSingleSpecialTokenId(this.tokenizer, '<|image>', 'Gemma 4 BOI token');
|
|
1212
|
-
const eoiTokenId = resolveSingleSpecialTokenId(this.tokenizer, '<image|>', 'Gemma 4 EOI token');
|
|
1213
|
-
const { inputIds: fullTokenIds, imageStartOffset } = expandImagePlaceholderTokenIds(
|
|
1214
|
-
promptTokenIds,
|
|
1215
|
-
imageTokenId,
|
|
1216
|
-
imageTokenSpanLength,
|
|
1217
|
-
{ boiTokenId, eoiTokenId }
|
|
1218
|
-
);
|
|
1219
|
-
const padTokenId = this.tokenizer?.getSpecialTokens?.()?.pad;
|
|
1220
|
-
if (!Number.isFinite(padTokenId) || Math.floor(padTokenId) !== padTokenId || padTokenId < 0) {
|
|
1221
|
-
throw new Error(
|
|
1222
|
-
`[Pipeline] transcribeImage: Gemma 4 multimodal prefill requires a tokenizer pad token ID, got ${padTokenId}.`
|
|
1223
|
-
);
|
|
1224
|
-
}
|
|
1225
|
-
|
|
1226
|
-
// Step 3: Generate with embedding override at the image token offset.
|
|
1227
|
-
const tokens = [];
|
|
1228
|
-
const maxGen = resolveMultimodalMaxTokens(this.runtimeConfig, maxTokens);
|
|
1229
|
-
const stopTokenIds = this.modelConfig.stopTokenIds;
|
|
1230
|
-
|
|
1231
|
-
try {
|
|
1232
|
-
const generation = await this.generator.generateTokenIds('', buildConservativeMultimodalGenerationOptions({
|
|
1233
|
-
inputIds: fullTokenIds,
|
|
1234
|
-
embeddingOverrides: {
|
|
1235
|
-
prefixLength: encodeResult.numTokens,
|
|
1236
|
-
offset: imageStartOffset,
|
|
1237
|
-
embeddings: encodeResult.features,
|
|
1238
|
-
},
|
|
1239
|
-
__internalEmbeddingInputSpan: {
|
|
1240
|
-
offset: imageStartOffset,
|
|
1241
|
-
length: encodeResult.numTokens,
|
|
1242
|
-
tokenId: padTokenId,
|
|
1243
|
-
},
|
|
1244
|
-
__internalMultimodalBidirectionalSpan: {
|
|
1245
|
-
offset: imageStartOffset,
|
|
1246
|
-
length: encodeResult.numTokens,
|
|
1247
|
-
},
|
|
1248
|
-
maxTokens: maxGen,
|
|
1249
|
-
temperature: 0,
|
|
1250
|
-
topK: 1,
|
|
1251
|
-
topP: 1,
|
|
1252
|
-
repetitionPenalty: 1,
|
|
1253
|
-
}));
|
|
1254
|
-
for (const token of generation.tokenIds ?? []) {
|
|
1255
|
-
if (Array.isArray(stopTokenIds) && stopTokenIds.includes(token)) break;
|
|
1256
|
-
tokens.push(token);
|
|
1257
|
-
}
|
|
1258
|
-
} finally {
|
|
1259
|
-
if (encodeResult.features) {
|
|
1260
|
-
releaseBuffer(encodeResult.features);
|
|
1261
|
-
}
|
|
1262
|
-
}
|
|
1263
|
-
|
|
1264
|
-
const text = this.tokenizer.decode(tokens);
|
|
1265
|
-
return { text, tokens };
|
|
155
|
+
return transcribeImageImpl.apply(this, arguments);
|
|
1266
156
|
}
|
|
1267
157
|
|
|
1268
|
-
|
|
1269
158
|
// ==========================================================================
|
|
1270
159
|
// Video: transcribeVideo
|
|
1271
160
|
// ==========================================================================
|
|
1272
161
|
|
|
1273
|
-
/**
|
|
1274
|
-
* Transcribe text from video using the vision encoder (per-frame) and text decoder.
|
|
1275
|
-
*
|
|
1276
|
-
* Video is processed as sampled frames through the existing vision encoder.
|
|
1277
|
-
* Each frame is encoded independently and the visual tokens are concatenated.
|
|
1278
|
-
*
|
|
1279
|
-
* @param {object} params
|
|
1280
|
-
* @param {Array<{ pixels: Uint8Array|Float32Array, width: number, height: number }>} params.frames Decoded video frames
|
|
1281
|
-
* @param {string} [params.prompt] Custom transcription prompt
|
|
1282
|
-
* @param {number} [params.maxTokens] Max tokens to generate
|
|
1283
|
-
* @param {number} [params.maxFrames=8] Maximum frames to sample
|
|
1284
|
-
* @param {number} [params.perFrameSoftTokenBudget] Soft token budget per frame
|
|
1285
|
-
* @returns {Promise<{ text: string, tokens: number[] }>}
|
|
1286
|
-
*/
|
|
1287
162
|
async transcribeVideo({ frames, prompt, maxTokens, maxFrames, perFrameSoftTokenBudget, signal }) {
|
|
1288
|
-
|
|
1289
|
-
if (!this.visionCapable) {
|
|
1290
|
-
throw new Error(
|
|
1291
|
-
'Pipeline does not support video transcription (no image_token_id in manifest for vision encoder).'
|
|
1292
|
-
);
|
|
1293
|
-
}
|
|
1294
|
-
await this._ensureVisionWeightsLoaded();
|
|
1295
|
-
|
|
1296
|
-
this.reset();
|
|
1297
|
-
|
|
1298
|
-
// Lazy-load video module
|
|
1299
|
-
const { encodeVideo } = await import('./video/index.js');
|
|
1300
|
-
|
|
1301
|
-
// Step 1: Encode video frames through vision pipeline
|
|
1302
|
-
const encodeResult = await encodeVideo({
|
|
1303
|
-
frames,
|
|
1304
|
-
visionConfig: this.visionConfig,
|
|
1305
|
-
weights: this.visionWeights,
|
|
1306
|
-
maxFrames: maxFrames ?? 8,
|
|
1307
|
-
perFrameSoftTokenBudget,
|
|
1308
|
-
});
|
|
1309
|
-
|
|
1310
|
-
// Step 2: Build the multimodal prompt with <|video|> placeholder
|
|
1311
|
-
const requestedPrompt = prompt ?? 'Describe the video in one short sentence.';
|
|
1312
|
-
const videoTokenId = this.tokenizer?.model?.tokenToId?.('<|video_token|>')
|
|
1313
|
-
?? this.tokenizer?.encode?.('<|video|>')?.find?.((id) => id !== undefined)
|
|
1314
|
-
?? null;
|
|
1315
|
-
// Fall back to image token ID for video placeholder expansion
|
|
1316
|
-
const placeholderTokenId = videoTokenId ?? this.visionConfig?.imageTokenId ?? this.imageTokenId;
|
|
1317
|
-
if (placeholderTokenId == null) {
|
|
1318
|
-
throw new Error(
|
|
1319
|
-
'Pipeline missing video/image token ID for video placeholder expansion.'
|
|
1320
|
-
);
|
|
1321
|
-
}
|
|
1322
|
-
|
|
1323
|
-
const templateType = this.modelConfig?.chatTemplateType ?? 'gemma4';
|
|
1324
|
-
const chatOptions = this.modelConfig?.chatTemplateThinking === true ? { thinking: true } : undefined;
|
|
1325
|
-
const multimodalPrompt = formatChatMessages([
|
|
1326
|
-
{
|
|
1327
|
-
role: 'user',
|
|
1328
|
-
content: [
|
|
1329
|
-
{ type: 'video' },
|
|
1330
|
-
{ type: 'text', text: requestedPrompt },
|
|
1331
|
-
],
|
|
1332
|
-
},
|
|
1333
|
-
], templateType, chatOptions);
|
|
1334
|
-
const promptTokenIds = this.tokenizer.encode(multimodalPrompt);
|
|
1335
|
-
const videoTokenSpanLength = encodeResult.numTokens;
|
|
1336
|
-
|
|
1337
|
-
// Resolve BOV/EOV tokens (reuse image BOI/EOI if video-specific ones don't exist)
|
|
1338
|
-
const bovTokenId = resolveSingleSpecialTokenId(this.tokenizer, '<|video|>', 'Gemma 4 BOV token');
|
|
1339
|
-
const eovTokenId = resolveSingleSpecialTokenId(this.tokenizer, '<video|>', 'Gemma 4 EOV token');
|
|
1340
|
-
|
|
1341
|
-
const { inputIds: fullTokenIds, imageStartOffset: videoStartOffset } = expandImagePlaceholderTokenIds(
|
|
1342
|
-
promptTokenIds,
|
|
1343
|
-
placeholderTokenId,
|
|
1344
|
-
videoTokenSpanLength,
|
|
1345
|
-
{ boiTokenId: bovTokenId, eoiTokenId: eovTokenId }
|
|
1346
|
-
);
|
|
1347
|
-
|
|
1348
|
-
const padTokenId = this.tokenizer?.getSpecialTokens?.()?.pad;
|
|
1349
|
-
if (!Number.isFinite(padTokenId) || Math.floor(padTokenId) !== padTokenId || padTokenId < 0) {
|
|
1350
|
-
throw new Error(
|
|
1351
|
-
`[Pipeline] transcribeVideo: Gemma 4 multimodal prefill requires a tokenizer pad token ID, got ${padTokenId}.`
|
|
1352
|
-
);
|
|
1353
|
-
}
|
|
1354
|
-
|
|
1355
|
-
// Step 3: Generate with embedding override at the video token offset
|
|
1356
|
-
const tokens = [];
|
|
1357
|
-
const maxGen = resolveMultimodalMaxTokens(this.runtimeConfig, maxTokens);
|
|
1358
|
-
const stopTokenIds = this.modelConfig.stopTokenIds;
|
|
1359
|
-
|
|
1360
|
-
try {
|
|
1361
|
-
const generation = await this.generator.generateTokenIds('', buildConservativeMultimodalGenerationOptions({
|
|
1362
|
-
inputIds: fullTokenIds,
|
|
1363
|
-
embeddingOverrides: {
|
|
1364
|
-
prefixLength: encodeResult.numTokens,
|
|
1365
|
-
offset: videoStartOffset,
|
|
1366
|
-
embeddings: encodeResult.features,
|
|
1367
|
-
},
|
|
1368
|
-
__internalEmbeddingInputSpan: {
|
|
1369
|
-
offset: videoStartOffset,
|
|
1370
|
-
length: encodeResult.numTokens,
|
|
1371
|
-
tokenId: padTokenId,
|
|
1372
|
-
},
|
|
1373
|
-
__internalMultimodalBidirectionalSpan: {
|
|
1374
|
-
offset: videoStartOffset,
|
|
1375
|
-
length: encodeResult.numTokens,
|
|
1376
|
-
},
|
|
1377
|
-
maxTokens: maxGen,
|
|
1378
|
-
temperature: 0,
|
|
1379
|
-
topK: 1,
|
|
1380
|
-
topP: 1,
|
|
1381
|
-
repetitionPenalty: 1,
|
|
1382
|
-
}));
|
|
1383
|
-
for (const token of generation.tokenIds ?? []) {
|
|
1384
|
-
if (Array.isArray(stopTokenIds) && stopTokenIds.includes(token)) break;
|
|
1385
|
-
tokens.push(token);
|
|
1386
|
-
}
|
|
1387
|
-
} finally {
|
|
1388
|
-
if (encodeResult.features) {
|
|
1389
|
-
releaseBuffer(encodeResult.features);
|
|
1390
|
-
}
|
|
1391
|
-
}
|
|
1392
|
-
|
|
1393
|
-
const text = this.tokenizer.decode(tokens);
|
|
1394
|
-
return { text, tokens };
|
|
163
|
+
return transcribeVideoImpl.apply(this, arguments);
|
|
1395
164
|
}
|
|
1396
165
|
|
|
1397
|
-
|
|
1398
166
|
// ==========================================================================
|
|
1399
167
|
// Audio: transcribeAudio
|
|
1400
168
|
// ==========================================================================
|
|
1401
169
|
|
|
1402
|
-
/**
|
|
1403
|
-
* Transcribe text from audio using the audio encoder and text decoder.
|
|
1404
|
-
*
|
|
1405
|
-
* @param {object} params
|
|
1406
|
-
* @param {Float32Array} params.audio Raw audio PCM (mono, 16kHz)
|
|
1407
|
-
* @param {string} [params.prompt] Custom transcription prompt
|
|
1408
|
-
* @param {number} [params.maxTokens] Max tokens to generate
|
|
1409
|
-
* @returns {Promise<{ text: string, tokens: number[] }>}
|
|
1410
|
-
*/
|
|
1411
170
|
async transcribeAudio({ audio, prompt, maxTokens, signal }) {
|
|
1412
|
-
|
|
1413
|
-
if (!this.audioCapable) {
|
|
1414
|
-
throw new Error(
|
|
1415
|
-
'Pipeline does not support audio transcription (no audio_token_id in manifest).'
|
|
1416
|
-
);
|
|
1417
|
-
}
|
|
1418
|
-
await this._ensureAudioWeightsLoaded();
|
|
1419
|
-
|
|
1420
|
-
this.reset();
|
|
1421
|
-
|
|
1422
|
-
// Lazy-load audio modules
|
|
1423
|
-
const { encodeAudio } = await import('./audio/index.js');
|
|
1424
|
-
|
|
1425
|
-
let encodeResult;
|
|
1426
|
-
if (this.audioConfig.depth === 0) {
|
|
1427
|
-
encodeResult = await encodeAudio({
|
|
1428
|
-
rawAudio: audio,
|
|
1429
|
-
audioConfig: this.audioConfig,
|
|
1430
|
-
weights: this.audioWeights,
|
|
1431
|
-
});
|
|
1432
|
-
} else {
|
|
1433
|
-
const { extractLogMelSpectrogram } = await import('./audio/mel.js');
|
|
1434
|
-
const { features: melFeatures, numFrames, nMels } = extractLogMelSpectrogram(audio);
|
|
1435
|
-
encodeResult = await encodeAudio({
|
|
1436
|
-
melFeatures,
|
|
1437
|
-
numFrames,
|
|
1438
|
-
nMels,
|
|
1439
|
-
audioConfig: this.audioConfig,
|
|
1440
|
-
weights: this.audioWeights,
|
|
1441
|
-
});
|
|
1442
|
-
}
|
|
1443
|
-
|
|
1444
|
-
// Step 3: Build the multimodal prompt with <|audio|> placeholder
|
|
1445
|
-
const requestedPrompt = prompt ?? 'Transcribe the audio.';
|
|
1446
|
-
const audioTokenId = this.audioConfig?.audioTokenId ?? this.audioTokenId;
|
|
1447
|
-
if (audioTokenId == null) {
|
|
1448
|
-
throw new Error(
|
|
1449
|
-
'Pipeline missing audio_token_id. Re-convert the model with audio token metadata.'
|
|
1450
|
-
);
|
|
1451
|
-
}
|
|
1452
|
-
const templateType = this.modelConfig?.chatTemplateType ?? 'gemma4';
|
|
1453
|
-
const chatOptions = this.modelConfig?.chatTemplateThinking === true ? { thinking: true } : undefined;
|
|
1454
|
-
const multimodalPrompt = formatChatMessages([
|
|
1455
|
-
{
|
|
1456
|
-
role: 'user',
|
|
1457
|
-
content: [
|
|
1458
|
-
{ type: 'audio' },
|
|
1459
|
-
{ type: 'text', text: requestedPrompt },
|
|
1460
|
-
],
|
|
1461
|
-
},
|
|
1462
|
-
], templateType, chatOptions);
|
|
1463
|
-
const promptTokenIds = this.tokenizer.encode(multimodalPrompt);
|
|
1464
|
-
const audioTokenSpanLength = encodeResult.numTokens;
|
|
1465
|
-
|
|
1466
|
-
// Resolve BOA/EOA tokens
|
|
1467
|
-
const boaTokenId = resolveSingleSpecialTokenId(this.tokenizer, '<|audio|>', 'Gemma 4 BOA token');
|
|
1468
|
-
const eoaTokenId = resolveSingleSpecialTokenId(this.tokenizer, '<audio|>', 'Gemma 4 EOA token');
|
|
1469
|
-
|
|
1470
|
-
// Expand single audio placeholder token into the full audio token span
|
|
1471
|
-
const { inputIds: fullTokenIds, imageStartOffset: audioStartOffset } = expandImagePlaceholderTokenIds(
|
|
1472
|
-
promptTokenIds,
|
|
1473
|
-
audioTokenId,
|
|
1474
|
-
audioTokenSpanLength,
|
|
1475
|
-
{ boiTokenId: boaTokenId, eoiTokenId: eoaTokenId }
|
|
1476
|
-
);
|
|
1477
|
-
|
|
1478
|
-
const padTokenId = this.tokenizer?.getSpecialTokens?.()?.pad;
|
|
1479
|
-
if (!Number.isFinite(padTokenId) || Math.floor(padTokenId) !== padTokenId || padTokenId < 0) {
|
|
1480
|
-
throw new Error(
|
|
1481
|
-
`[Pipeline] transcribeAudio: Gemma 4 multimodal prefill requires a tokenizer pad token ID, got ${padTokenId}.`
|
|
1482
|
-
);
|
|
1483
|
-
}
|
|
1484
|
-
|
|
1485
|
-
// Step 4: Generate with embedding override at the audio token offset
|
|
1486
|
-
const tokens = [];
|
|
1487
|
-
const maxGen = resolveMultimodalMaxTokens(this.runtimeConfig, maxTokens);
|
|
1488
|
-
const stopTokenIds = this.modelConfig.stopTokenIds;
|
|
1489
|
-
|
|
1490
|
-
try {
|
|
1491
|
-
const generation = await this.generator.generateTokenIds('', buildConservativeMultimodalGenerationOptions({
|
|
1492
|
-
inputIds: fullTokenIds,
|
|
1493
|
-
embeddingOverrides: {
|
|
1494
|
-
prefixLength: encodeResult.numTokens,
|
|
1495
|
-
offset: audioStartOffset,
|
|
1496
|
-
embeddings: encodeResult.features,
|
|
1497
|
-
},
|
|
1498
|
-
__internalEmbeddingInputSpan: {
|
|
1499
|
-
offset: audioStartOffset,
|
|
1500
|
-
length: encodeResult.numTokens,
|
|
1501
|
-
tokenId: padTokenId,
|
|
1502
|
-
},
|
|
1503
|
-
__internalMultimodalBidirectionalSpan: {
|
|
1504
|
-
offset: audioStartOffset,
|
|
1505
|
-
length: encodeResult.numTokens,
|
|
1506
|
-
},
|
|
1507
|
-
maxTokens: maxGen,
|
|
1508
|
-
temperature: 0,
|
|
1509
|
-
topK: 1,
|
|
1510
|
-
topP: 1,
|
|
1511
|
-
repetitionPenalty: 1,
|
|
1512
|
-
}));
|
|
1513
|
-
for (const token of generation.tokenIds ?? []) {
|
|
1514
|
-
if (Array.isArray(stopTokenIds) && stopTokenIds.includes(token)) break;
|
|
1515
|
-
tokens.push(token);
|
|
1516
|
-
}
|
|
1517
|
-
} finally {
|
|
1518
|
-
if (encodeResult.features) {
|
|
1519
|
-
releaseBuffer(encodeResult.features);
|
|
1520
|
-
}
|
|
1521
|
-
}
|
|
1522
|
-
|
|
1523
|
-
const text = this.tokenizer.decode(tokens);
|
|
1524
|
-
return { text, tokens };
|
|
171
|
+
return transcribeAudioImpl.apply(this, arguments);
|
|
1525
172
|
}
|
|
1526
173
|
|
|
1527
|
-
|
|
1528
174
|
// ==========================================================================
|
|
1529
175
|
// Capability Detection
|
|
1530
176
|
// ==========================================================================
|
|
@@ -1539,7 +185,6 @@ export class InferencePipeline extends PipelineState {
|
|
|
1539
185
|
return Object.freeze(caps);
|
|
1540
186
|
}
|
|
1541
187
|
|
|
1542
|
-
|
|
1543
188
|
// Layer pipeline precedence (lowest to highest):
|
|
1544
189
|
// 1. execution-v1-produced pipeline (via runtimeInferencePatch.pipeline)
|
|
1545
190
|
// 2. model config pipeline (manifest inference.pipeline)
|
|
@@ -1583,7 +228,6 @@ export class InferencePipeline extends PipelineState {
|
|
|
1583
228
|
// Generation Delegates
|
|
1584
229
|
// ==========================================================================
|
|
1585
230
|
|
|
1586
|
-
|
|
1587
231
|
generate(prompt, options = {}) {
|
|
1588
232
|
assertNotAborted(options?.signal);
|
|
1589
233
|
return this.generator.generate(prompt, options);
|
|
@@ -1618,7 +262,6 @@ export class InferencePipeline extends PipelineState {
|
|
|
1618
262
|
return this.generator.advanceWithTokenAndEmbedding(tokenId, options);
|
|
1619
263
|
}
|
|
1620
264
|
|
|
1621
|
-
|
|
1622
265
|
prefillKVOnly(prompt, options = {}) {
|
|
1623
266
|
assertNotAborted(options?.signal);
|
|
1624
267
|
return this.generator.prefillKVOnly(prompt, options);
|
|
@@ -1644,95 +287,15 @@ export class InferencePipeline extends PipelineState {
|
|
|
1644
287
|
}
|
|
1645
288
|
|
|
1646
289
|
async embed(prompt, options = {}) {
|
|
1647
|
-
|
|
1648
|
-
this.resetForBatch();
|
|
1649
|
-
try {
|
|
1650
|
-
const result = await this.prefillWithEmbedding(prompt, {
|
|
1651
|
-
...options,
|
|
1652
|
-
__skipStateSnapshot: true,
|
|
1653
|
-
});
|
|
1654
|
-
assertNotAborted(options?.signal);
|
|
1655
|
-
return {
|
|
1656
|
-
embedding: result.embedding,
|
|
1657
|
-
tokens: result.tokens,
|
|
1658
|
-
seqLen: result.seqLen,
|
|
1659
|
-
embeddingMode: result.embeddingMode,
|
|
1660
|
-
phase: result.phase ?? null,
|
|
1661
|
-
};
|
|
1662
|
-
} finally {
|
|
1663
|
-
this.resetForBatch();
|
|
1664
|
-
}
|
|
290
|
+
return embedImpl.apply(this, arguments);
|
|
1665
291
|
}
|
|
1666
292
|
|
|
1667
293
|
async embedBatch(prompts, options = {}) {
|
|
1668
|
-
|
|
1669
|
-
throw new Error('embedBatch expects an array of prompts');
|
|
1670
|
-
}
|
|
1671
|
-
assertNotAborted(options?.signal);
|
|
1672
|
-
const batchOptions = { ...options, __skipStateSnapshot: true };
|
|
1673
|
-
const outputs = [];
|
|
1674
|
-
for (const prompt of prompts) {
|
|
1675
|
-
// Check between every prompt so a superseded revision drops the rest.
|
|
1676
|
-
assertNotAborted(options?.signal);
|
|
1677
|
-
outputs.push(await this.embed(prompt, batchOptions));
|
|
1678
|
-
}
|
|
1679
|
-
return outputs;
|
|
294
|
+
return embedBatchImpl.apply(this, arguments);
|
|
1680
295
|
}
|
|
1681
296
|
|
|
1682
297
|
async encodeSequence(sequence, options = {}) {
|
|
1683
|
-
|
|
1684
|
-
const contract = this.manifest?.inference?.sequence ?? null;
|
|
1685
|
-
if (this.manifest?.inference?.supportsSequence !== true || !contract) {
|
|
1686
|
-
throw new Error('Model manifest does not declare sequence encoding support.');
|
|
1687
|
-
}
|
|
1688
|
-
if (typeof sequence !== 'string' || sequence.length === 0) {
|
|
1689
|
-
throw new Error('encodeSequence expects a non-empty sequence string.');
|
|
1690
|
-
}
|
|
1691
|
-
const includeTokenEmbeddings = options.includeTokenEmbeddings
|
|
1692
|
-
?? contract.tokenEmbeddings;
|
|
1693
|
-
const includeLogits = options.includeLogits === true;
|
|
1694
|
-
if (includeTokenEmbeddings && contract.tokenEmbeddings !== true) {
|
|
1695
|
-
throw new Error('Model manifest does not permit token-level sequence embeddings.');
|
|
1696
|
-
}
|
|
1697
|
-
if (includeLogits && contract.logits !== true) {
|
|
1698
|
-
throw new Error('Model manifest does not permit sequence logits.');
|
|
1699
|
-
}
|
|
1700
|
-
const needsTokenEmbeddings = includeTokenEmbeddings || contract.pooledEmbedding !== null;
|
|
1701
|
-
|
|
1702
|
-
this.resetForBatch();
|
|
1703
|
-
try {
|
|
1704
|
-
const result = await this.prefillWithEmbedding(sequence, {
|
|
1705
|
-
...options,
|
|
1706
|
-
embeddingMode: contract.pooledEmbedding?.mode ?? 'last',
|
|
1707
|
-
__skipStateSnapshot: true,
|
|
1708
|
-
__returnTokenEmbeddings: needsTokenEmbeddings,
|
|
1709
|
-
__returnSequenceLogits: includeLogits,
|
|
1710
|
-
});
|
|
1711
|
-
assertNotAborted(options?.signal);
|
|
1712
|
-
const hiddenSize = this.modelConfig.hiddenSize;
|
|
1713
|
-
const pooledResult = contract.pooledEmbedding
|
|
1714
|
-
? poolSequenceTokenEmbeddings(
|
|
1715
|
-
result.tokenEmbeddings,
|
|
1716
|
-
result.tokens,
|
|
1717
|
-
hiddenSize,
|
|
1718
|
-
contract.pooledEmbedding
|
|
1719
|
-
)
|
|
1720
|
-
: null;
|
|
1721
|
-
return {
|
|
1722
|
-
alphabet: contract.alphabet,
|
|
1723
|
-
tokens: result.tokens,
|
|
1724
|
-
tokenMask: pooledResult?.tokenMask ?? new Uint8Array(result.tokens.length),
|
|
1725
|
-
includedTokenCount: pooledResult?.includedTokenCount ?? 0,
|
|
1726
|
-
tokenEmbeddings: includeTokenEmbeddings ? result.tokenEmbeddings : null,
|
|
1727
|
-
pooledEmbedding: pooledResult?.pooled ?? null,
|
|
1728
|
-
logits: includeLogits ? result.logits : null,
|
|
1729
|
-
embeddingDim: hiddenSize,
|
|
1730
|
-
vocabSize: this.modelConfig.vocabSize,
|
|
1731
|
-
phase: result.phase ?? null,
|
|
1732
|
-
};
|
|
1733
|
-
} finally {
|
|
1734
|
-
this.resetForBatch();
|
|
1735
|
-
}
|
|
298
|
+
return encodeSequenceImpl.apply(this, arguments);
|
|
1736
299
|
}
|
|
1737
300
|
|
|
1738
301
|
// Run the vision encoder over a single image and return a mean-pooled
|
|
@@ -1740,61 +303,7 @@ export class InferencePipeline extends PipelineState {
|
|
|
1740
303
|
// (jpeg/png -> RGBA pixels) belongs to the caller; this method takes
|
|
1741
304
|
// already-decoded pixel data.
|
|
1742
305
|
async embedImage({ pixels, width, height, softTokenBudget, signal } = {}) {
|
|
1743
|
-
|
|
1744
|
-
if (!this.visionCapable) {
|
|
1745
|
-
throw new Error(
|
|
1746
|
-
'Pipeline does not support image embedding (no image_token_id in manifest).'
|
|
1747
|
-
);
|
|
1748
|
-
}
|
|
1749
|
-
if (pixels == null) {
|
|
1750
|
-
throw new Error('[Pipeline] embedImage: pixels are required.');
|
|
1751
|
-
}
|
|
1752
|
-
if (!Number.isFinite(width) || width <= 0 || !Number.isFinite(height) || height <= 0) {
|
|
1753
|
-
throw new Error('[Pipeline] embedImage: width and height must be positive integers.');
|
|
1754
|
-
}
|
|
1755
|
-
await this._ensureVisionWeightsLoaded();
|
|
1756
|
-
this.reset();
|
|
1757
|
-
|
|
1758
|
-
const { encodeImage } = await import('./vision/index.js');
|
|
1759
|
-
const encodeResult = await encodeImage({
|
|
1760
|
-
pixels,
|
|
1761
|
-
width,
|
|
1762
|
-
height,
|
|
1763
|
-
visionConfig: this.visionConfig,
|
|
1764
|
-
weights: this.visionWeights,
|
|
1765
|
-
softTokenBudget,
|
|
1766
|
-
});
|
|
1767
|
-
|
|
1768
|
-
const hiddenSize = this.modelConfig.hiddenSize;
|
|
1769
|
-
const numTokens = encodeResult.numTokens;
|
|
1770
|
-
if (!Number.isFinite(numTokens) || numTokens < 1) {
|
|
1771
|
-
releaseBuffer(encodeResult.features);
|
|
1772
|
-
throw new Error(`[Pipeline] embedImage: encoder produced ${numTokens} soft tokens; expected >= 1.`);
|
|
1773
|
-
}
|
|
1774
|
-
try {
|
|
1775
|
-
const bytes = await readBuffer(
|
|
1776
|
-
encodeResult.features,
|
|
1777
|
-
numTokens * hiddenSize * Float32Array.BYTES_PER_ELEMENT
|
|
1778
|
-
);
|
|
1779
|
-
const features = new Float32Array(bytes);
|
|
1780
|
-
const pooled = new Float32Array(hiddenSize);
|
|
1781
|
-
for (let t = 0; t < numTokens; t++) {
|
|
1782
|
-
const base = t * hiddenSize;
|
|
1783
|
-
for (let d = 0; d < hiddenSize; d++) {
|
|
1784
|
-
pooled[d] += features[base + d];
|
|
1785
|
-
}
|
|
1786
|
-
}
|
|
1787
|
-
const inv = 1 / numTokens;
|
|
1788
|
-
for (let d = 0; d < hiddenSize; d++) pooled[d] *= inv;
|
|
1789
|
-
return {
|
|
1790
|
-
embedding: pooled,
|
|
1791
|
-
embeddingDim: hiddenSize,
|
|
1792
|
-
numTokens,
|
|
1793
|
-
embeddingMode: 'mean',
|
|
1794
|
-
};
|
|
1795
|
-
} finally {
|
|
1796
|
-
releaseBuffer(encodeResult.features);
|
|
1797
|
-
}
|
|
306
|
+
return embedImageImpl.apply(this, arguments);
|
|
1798
307
|
}
|
|
1799
308
|
|
|
1800
309
|
// Run the audio encoder over a single PCM segment and return a mean-pooled
|
|
@@ -1802,73 +311,7 @@ export class InferencePipeline extends PipelineState {
|
|
|
1802
311
|
// text hidden size in Gemma 4). Decoder responsibility (webm/opus/wav ->
|
|
1803
312
|
// Float32 PCM at the model's expected sample rate) belongs to the caller.
|
|
1804
313
|
async embedAudio({ audio, signal } = {}) {
|
|
1805
|
-
|
|
1806
|
-
if (!this.audioCapable) {
|
|
1807
|
-
throw new Error(
|
|
1808
|
-
'Pipeline does not support audio embedding (no audio_token_id in manifest).'
|
|
1809
|
-
);
|
|
1810
|
-
}
|
|
1811
|
-
if (audio == null) {
|
|
1812
|
-
throw new Error('[Pipeline] embedAudio: audio is required.');
|
|
1813
|
-
}
|
|
1814
|
-
await this._ensureAudioWeightsLoaded();
|
|
1815
|
-
this.reset();
|
|
1816
|
-
|
|
1817
|
-
const { encodeAudio } = await import('./audio/index.js');
|
|
1818
|
-
|
|
1819
|
-
let encodeResult;
|
|
1820
|
-
if (this.audioConfig.depth === 0) {
|
|
1821
|
-
encodeResult = await encodeAudio({
|
|
1822
|
-
rawAudio: audio,
|
|
1823
|
-
audioConfig: this.audioConfig,
|
|
1824
|
-
weights: this.audioWeights,
|
|
1825
|
-
});
|
|
1826
|
-
} else {
|
|
1827
|
-
const { extractLogMelSpectrogram } = await import('./audio/mel.js');
|
|
1828
|
-
const { features: melFeatures, numFrames, nMels } = extractLogMelSpectrogram(audio);
|
|
1829
|
-
encodeResult = await encodeAudio({
|
|
1830
|
-
melFeatures,
|
|
1831
|
-
numFrames,
|
|
1832
|
-
nMels,
|
|
1833
|
-
audioConfig: this.audioConfig,
|
|
1834
|
-
weights: this.audioWeights,
|
|
1835
|
-
});
|
|
1836
|
-
}
|
|
1837
|
-
|
|
1838
|
-
const hiddenSize = Number(this.audioConfig?.outputProjDims ?? this.modelConfig?.hiddenSize);
|
|
1839
|
-
const numTokens = encodeResult.numTokens;
|
|
1840
|
-
if (!Number.isFinite(hiddenSize) || hiddenSize < 1) {
|
|
1841
|
-
releaseBuffer(encodeResult.features);
|
|
1842
|
-
throw new Error('[Pipeline] embedAudio: audioConfig.outputProjDims is missing or invalid.');
|
|
1843
|
-
}
|
|
1844
|
-
if (!Number.isFinite(numTokens) || numTokens < 1) {
|
|
1845
|
-
releaseBuffer(encodeResult.features);
|
|
1846
|
-
throw new Error(`[Pipeline] embedAudio: encoder produced ${numTokens} tokens; expected >= 1.`);
|
|
1847
|
-
}
|
|
1848
|
-
try {
|
|
1849
|
-
const bytes = await readBuffer(
|
|
1850
|
-
encodeResult.features,
|
|
1851
|
-
numTokens * hiddenSize * Float32Array.BYTES_PER_ELEMENT
|
|
1852
|
-
);
|
|
1853
|
-
const features = new Float32Array(bytes);
|
|
1854
|
-
const pooled = new Float32Array(hiddenSize);
|
|
1855
|
-
for (let t = 0; t < numTokens; t++) {
|
|
1856
|
-
const base = t * hiddenSize;
|
|
1857
|
-
for (let d = 0; d < hiddenSize; d++) {
|
|
1858
|
-
pooled[d] += features[base + d];
|
|
1859
|
-
}
|
|
1860
|
-
}
|
|
1861
|
-
const inv = 1 / numTokens;
|
|
1862
|
-
for (let d = 0; d < hiddenSize; d++) pooled[d] *= inv;
|
|
1863
|
-
return {
|
|
1864
|
-
embedding: pooled,
|
|
1865
|
-
embeddingDim: hiddenSize,
|
|
1866
|
-
numTokens,
|
|
1867
|
-
embeddingMode: 'mean',
|
|
1868
|
-
};
|
|
1869
|
-
} finally {
|
|
1870
|
-
releaseBuffer(encodeResult.features);
|
|
1871
|
-
}
|
|
314
|
+
return embedAudioImpl.apply(this, arguments);
|
|
1872
315
|
}
|
|
1873
316
|
|
|
1874
317
|
prefillWithLogits(prompt, options = {}) {
|
|
@@ -1883,7 +326,6 @@ export class InferencePipeline extends PipelineState {
|
|
|
1883
326
|
return this.generator.prefillWithTokenLogitsFromKV(prefix, prompt, tokenIds, options);
|
|
1884
327
|
}
|
|
1885
328
|
|
|
1886
|
-
|
|
1887
329
|
applyKVCacheSnapshot(snapshot) {
|
|
1888
330
|
this.kvCache = snapshot.cache.clone();
|
|
1889
331
|
if (this.useGPU && this.kvCache) {
|
|
@@ -1908,7 +350,6 @@ export class InferencePipeline extends PipelineState {
|
|
|
1908
350
|
this.currentSeqLen = snapshot.seqLen;
|
|
1909
351
|
}
|
|
1910
352
|
|
|
1911
|
-
|
|
1912
353
|
generateWithPrefixKV(prefix, prompt, options = {}) {
|
|
1913
354
|
return this.generator.generateWithPrefixKV(prefix, prompt, options);
|
|
1914
355
|
}
|
|
@@ -1917,7 +358,6 @@ export class InferencePipeline extends PipelineState {
|
|
|
1917
358
|
// Utility Methods
|
|
1918
359
|
// ==========================================================================
|
|
1919
360
|
|
|
1920
|
-
|
|
1921
361
|
getStats() {
|
|
1922
362
|
const stats = { ...this.stats };
|
|
1923
363
|
stats.batching ??= { ...this.batchingStats };
|
|
@@ -1979,12 +419,10 @@ export class InferencePipeline extends PipelineState {
|
|
|
1979
419
|
return stats;
|
|
1980
420
|
}
|
|
1981
421
|
|
|
1982
|
-
|
|
1983
422
|
getBatchingStats() {
|
|
1984
423
|
return { ...this.batchingStats };
|
|
1985
424
|
}
|
|
1986
425
|
|
|
1987
|
-
|
|
1988
426
|
getMemoryStats() {
|
|
1989
427
|
|
|
1990
428
|
const stats = { used: 0 };
|
|
@@ -2010,14 +448,12 @@ export class InferencePipeline extends PipelineState {
|
|
|
2010
448
|
return stats;
|
|
2011
449
|
}
|
|
2012
450
|
|
|
2013
|
-
|
|
2014
451
|
getKVCacheStats() {
|
|
2015
452
|
if (!this.kvCache) return null;
|
|
2016
453
|
const { seqLen, maxSeqLen } = this.kvCache.getMemoryStats();
|
|
2017
454
|
return { seqLen, maxSeqLen };
|
|
2018
455
|
}
|
|
2019
456
|
|
|
2020
|
-
|
|
2021
457
|
getBufferPool() {
|
|
2022
458
|
try {
|
|
2023
459
|
return getGlobalBufferPool();
|
|
@@ -2026,7 +462,6 @@ export class InferencePipeline extends PipelineState {
|
|
|
2026
462
|
}
|
|
2027
463
|
}
|
|
2028
464
|
|
|
2029
|
-
|
|
2030
465
|
async unload() {
|
|
2031
466
|
const storageContext = this.storageContext;
|
|
2032
467
|
this.storageContext = null;
|
|
@@ -2047,6 +482,7 @@ export class InferencePipeline extends PipelineState {
|
|
|
2047
482
|
this.ownsDopplerLoader = false;
|
|
2048
483
|
this.linearAttentionRuntime = resetLinearAttentionRuntime(this.linearAttentionRuntime);
|
|
2049
484
|
this.lora = null;
|
|
485
|
+
this.revocationIdentity = null;
|
|
2050
486
|
destroyMoERouter(this.moeRouter);
|
|
2051
487
|
this.moeRouter = null;
|
|
2052
488
|
if (this.finitenessBuffer) {
|
|
@@ -2066,17 +502,15 @@ export class InferencePipeline extends PipelineState {
|
|
|
2066
502
|
log.info('Pipeline', 'Unloaded');
|
|
2067
503
|
}
|
|
2068
504
|
|
|
2069
|
-
|
|
2070
505
|
setLoRAAdapter(adapter) {
|
|
506
|
+
assertBundledAdapterAuthorized(adapter);
|
|
2071
507
|
this.lora = adapter;
|
|
2072
508
|
}
|
|
2073
509
|
|
|
2074
|
-
|
|
2075
510
|
getActiveLoRA() {
|
|
2076
511
|
return this.lora;
|
|
2077
512
|
}
|
|
2078
513
|
|
|
2079
|
-
|
|
2080
514
|
reset() {
|
|
2081
515
|
this.kvCache?.clear();
|
|
2082
516
|
this.linearAttentionRuntime = resetLinearAttentionRuntime(this.linearAttentionRuntime);
|
|
@@ -2100,10 +534,6 @@ export class InferencePipeline extends PipelineState {
|
|
|
2100
534
|
this.stats.attentionInputs = [];
|
|
2101
535
|
}
|
|
2102
536
|
|
|
2103
|
-
/**
|
|
2104
|
-
* Lightweight reset for batch embedding: clears KV cache and sequence state
|
|
2105
|
-
* without resetting stats or debug flags (they are overwritten on each prefill).
|
|
2106
|
-
*/
|
|
2107
537
|
resetForBatch() {
|
|
2108
538
|
this.kvCache?.clear();
|
|
2109
539
|
this.linearAttentionRuntime = resetLinearAttentionRuntime(this.linearAttentionRuntime);
|
|
@@ -2113,7 +543,6 @@ export class InferencePipeline extends PipelineState {
|
|
|
2113
543
|
this.decodeRing?.reset();
|
|
2114
544
|
}
|
|
2115
545
|
|
|
2116
|
-
|
|
2117
546
|
releaseGPUResources() {
|
|
2118
547
|
this.decodeBuffers?.release();
|
|
2119
548
|
this.decodeRing?.release();
|
|
@@ -2138,7 +567,6 @@ export class InferencePipeline extends PipelineState {
|
|
|
2138
567
|
// Factory Function
|
|
2139
568
|
// ============================================================================
|
|
2140
569
|
|
|
2141
|
-
|
|
2142
570
|
async function createTransformerPipeline(manifest, contexts = {}) {
|
|
2143
571
|
return createInitializedPipeline(InferencePipeline, manifest, contexts);
|
|
2144
572
|
}
|