doppler-gpu 0.7.3 → 0.7.8
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 +60 -0
- package/README.md +9 -17
- package/package.json +5 -48
- package/src/capsule-runtime.d.ts +2 -2
- package/src/client/capsule-host.browser.js +4 -4
- package/src/client/capsule-host.d.ts +8 -8
- package/src/client/capsule-host.js +4 -4
- package/src/client/doppler-api.d.ts +2 -2
- package/src/client/inspection.d.ts +3 -0
- package/src/client/inspection.js +29 -0
- package/src/client/model-host/index.d.ts +5 -4
- package/src/client/model-host/index.js +28 -8
- package/src/client/model-host/model-evidence.js +2 -0
- package/src/client/model-host/model-session.d.ts +7 -5
- package/src/client/model-host/model-session.js +10 -3
- package/src/client/model-host/privacy-receipt.js +22 -0
- package/src/client/resident-partitions.d.ts +3 -3
- package/src/client/resident-partitions.js +2 -2
- package/src/client/runtime/capsule-acquisition.d.ts +3 -3
- package/src/client/runtime/capsule-acquisition.js +1 -1
- package/src/client/runtime/capsule-adapter-execution.d.ts +2 -2
- package/src/client/runtime/capsule-adapter-execution.js +3 -3
- package/src/client/runtime/capsule-embedding.d.ts +2 -2
- package/src/client/runtime/capsule-embedding.js +1 -1
- package/src/client/runtime/capsule-forecast.d.ts +2 -2
- package/src/client/runtime/capsule-forecast.js +1 -1
- package/src/client/runtime/capsule-operation-adapters.d.ts +1 -1
- package/src/client/runtime/capsule-operation-adapters.js +1 -1
- package/src/client/runtime/capsule-operation-deltas.js +1 -1
- package/src/client/runtime/capsule-operation-executor.d.ts +1 -1
- package/src/client/runtime/capsule-operation-executor.js +2 -2
- package/src/client/runtime/capsule-operation-stream.d.ts +1 -1
- package/src/client/runtime/capsule-operation-stream.js +2 -2
- package/src/client/runtime/capsule-piece-reader.js +1 -1
- package/src/client/runtime/capsule-program-adapter.js +1 -1
- package/src/client/runtime/capsule-release-authorization.js +1 -1
- package/src/client/runtime/capsule-rerank.d.ts +2 -2
- package/src/client/runtime/composition-root.d.ts +6 -6
- package/src/client/runtime/composition-root.js +6 -6
- package/src/client/runtime/fetch-capsule-artifact-store.d.ts +1 -1
- package/src/client/runtime/fetch-capsule-artifact-store.js +1 -1
- package/src/client/runtime/model-source.d.ts +4 -2
- package/src/client/runtime/model-source.js +5 -5
- package/src/client/runtime/scoped-session.d.ts +2 -2
- package/src/client/runtime/verified-capsule-artifact-store.d.ts +2 -2
- package/src/client/runtime/verified-capsule-artifact-store.js +1 -1
- package/src/config/backward-registry-loader.js +1 -1
- package/src/config/{capsule-adapter-policy.js → capsule/capsule-adapter-policy.js} +2 -2
- package/src/config/{capsule-adapters.d.ts → capsule/capsule-adapters.d.ts} +1 -1
- package/src/config/{capsule-adapters.js → capsule/capsule-adapters.js} +2 -2
- package/src/config/{capsule-loading.js → capsule/capsule-loading.js} +1 -1
- package/src/config/{capsule-operation.d.ts → capsule/capsule-operation.d.ts} +1 -1
- package/src/config/{capsule-operation.js → capsule/capsule-operation.js} +3 -3
- package/src/config/{capsule-release-contract.js → capsule/capsule-release-contract.js} +2 -2
- package/src/config/{capsule-release-events.d.ts → capsule/capsule-release-events.d.ts} +2 -1
- package/src/config/{capsule-release-events.js → capsule/capsule-release-events.js} +3 -2
- package/src/config/{capsule-sequence-receipt.js → capsule/capsule-sequence-receipt.js} +1 -1
- package/src/config/{capsule-token-selection.d.ts → capsule/capsule-token-selection.d.ts} +1 -1
- package/src/config/{capsule-token-selection.js → capsule/capsule-token-selection.js} +2 -2
- package/src/config/{capsule-v2.d.ts → capsule/capsule-v2.d.ts} +5 -4
- package/src/config/{capsule-v2.js → capsule/capsule-v2.js} +6 -6
- package/src/config/{capsule-v3.js → capsule/capsule-v3.js} +1 -1
- package/src/config/{capsule.js → capsule/capsule.js} +26 -7
- package/src/config/choice-scoring.js +1 -1
- package/src/config/conversion/amplify/amplify-120m-f16-af32.json +1 -1
- package/src/config/conversion/diffusiongemma/diffusiongemma-26b-a4b-it-q4k-ehf16-af16.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 +2 -2
- package/src/config/conversion/esmc/esmc-300m-f32-af32.json +1 -1
- package/src/config/conversion/gemma3/gemma-3-1b-it-f16-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/gemma-3-270m-it-w8a8-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 +8 -8
- package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-af32.json +8 -8
- package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-hq4k-af16.json +8 -8
- package/src/config/conversion/gemma4/gemma-4-12b-it-text-w4a16-ct-ehf16-af16.json +8 -8
- package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af16.json +8 -8
- package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af32.json +8 -8
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af16-int4ple.json +8 -8
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32-int4ple.json +8 -8
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32.json +8 -8
- package/src/config/conversion/gemma4/gemma-4-moe-q4k-ehf16-af32.json +1 -1
- package/src/config/conversion/glmocr/glm-ocr-f16-af32.json +1 -1
- package/src/config/conversion/gpt-oss-20b-f16-xmxfp4.json +1 -1
- package/src/config/conversion/janus/janus-pro-1b-text-q4k-ehaf16.json +1 -1
- package/src/config/conversion/lfm2/lfm2.5-1.2b-instruct-q4k-ehf16-af32.json +1 -1
- 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 +1 -1
- package/src/config/conversion/muse-glimmer/muse-glimmer-30b-text-f16-af32.json +1 -1
- package/src/config/conversion/nucleotide-transformer/nucleotide-transformer-v2-50m-f32-af32.json +1 -1
- package/src/config/conversion/qwen3/qwen-3-4b-instruct-2507-q4k-ehf16-af32.json +2 -2
- package/src/config/conversion/qwen3/qwen-3-5-0-8b-canonical-f32.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 +1 -1
- package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-ehaf16.json +1 -1
- package/src/config/conversion/qwen3/qwen-3-6-35b-a3b-q4_0-passthrough.json +1 -1
- package/src/config/conversion/qwen3/qwen-3-6-35b-a3b-text-q4k-ehf16-af32.json +1 -1
- package/src/config/conversion/qwen3/qwen-3-8-flash-next-text-q4k-ehf16-af32.json +25 -22
- package/src/config/conversion/qwen3/qwen-3-embedding-0-6b-f16-af32.json +2 -2
- 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 +2 -2
- package/src/config/conversion/qwen3/qwen-3-reranker-0-6b-q4k-ehf16-af32.json +2 -2
- package/src/config/conversion/qwen3/qwen3.8-27b-q4_0-passthrough.json +1 -1
- package/src/config/conversion/qwen3/qwen3.8-27b-text-f16-af32.json +1 -1
- package/src/config/conversion/qwen3/qwen3.8-27b-text-q4k-ehaf16.json +1 -1
- package/src/config/conversion/qwen3/qwen4exp-gguf-source-policy.json +53 -0
- package/src/config/embedding-contract.d.ts +1 -1
- package/src/config/embedding-contract.js +1 -1
- package/src/config/execution-registry-contract.d.ts +5 -5
- package/src/config/execution-registry-contract.js +5 -5
- package/src/config/generation-contract.js +2 -1
- package/src/config/hardware-partition-plan.d.ts +32 -0
- package/src/config/hardware-partition-plan.js +78 -0
- package/src/config/inspection/observation-policies.json +1 -1
- package/src/config/kernel-path-loader.d.ts +1 -1
- package/src/config/kernel-path-loader.js +1 -1
- package/src/config/kernel-registry-contract.d.ts +2 -2
- package/src/config/kernel-registry-contract.js +1 -1
- package/src/config/kernels/kernel-ref-digests.js +509 -0
- package/src/config/kernels/registry.js +2 -2
- package/src/config/kernels/registry.json +2065 -1241
- package/src/config/layer-partition-plan.d.ts +4 -0
- package/src/config/layer-partition-plan.js +10 -4
- package/src/config/merge.js +1 -1
- package/src/config/model-session-declarations.js +85 -0
- package/src/config/param-validator.js +3 -3
- package/src/config/platforms/amd-rdna3.json +2 -2
- package/src/config/production-release-evidence.js +1 -1
- package/src/config/revocation-policy.js +6 -2
- package/src/config/revocation-registry.json +1 -1
- package/src/config/runtime/model-sessions.json +23 -0
- package/src/config/runtime/profiles/flash-next-gguf-control-gemv.json +72 -0
- package/src/config/runtime/profiles/flash-next-gguf-hc-inject-gemv.json +47 -0
- package/src/config/schema/capsule/index.d.ts +14 -0
- package/src/config/schema/capsule/index.js +10 -0
- package/src/config/schema/{program-bundle.schema.d.ts → capsule/program-bundle.schema.d.ts} +3 -3
- package/src/config/schema/{program-bundle.schema.js → capsule/program-bundle.schema.js} +6 -6
- package/src/config/schema/{execution-v1.schema.d.ts → execution/execution-v1.schema.d.ts} +2 -0
- package/src/config/schema/{execution-v1.schema.js → execution/execution-v1.schema.js} +2 -2
- package/src/config/schema/execution/index.d.ts +203 -0
- package/src/config/schema/execution/index.js +94 -0
- package/src/config/schema/{inference.schema.d.ts → execution/inference.schema.d.ts} +1 -1
- package/src/config/schema/{kernel-path.schema.d.ts → execution/kernel-path.schema.d.ts} +3 -0
- package/src/config/schema/{kernel-registry.schema.d.ts → execution/kernel-registry.schema.d.ts} +2 -2
- package/src/config/schema/{kernel-thresholds.schema.d.ts → execution/kernel-thresholds.schema.d.ts} +1 -1
- package/src/config/schema/{kernel-thresholds.schema.js → execution/kernel-thresholds.schema.js} +1 -1
- package/src/config/schema/{kernel-warmup.schema.d.ts → execution/kernel-warmup.schema.d.ts} +14 -0
- package/src/config/schema/{kernel-warmup.schema.js → execution/kernel-warmup.schema.js} +3 -0
- package/src/config/schema/{kvcache.schema.d.ts → execution/kvcache.schema.d.ts} +8 -2
- package/src/config/schema/{kvcache.schema.js → execution/kvcache.schema.js} +17 -1
- package/src/config/schema/{ngram-embedding.schema.d.ts → execution/ngram-embedding.schema.d.ts} +3 -1
- package/src/config/schema/{ngram-embedding.schema.js → execution/ngram-embedding.schema.js} +16 -1
- package/src/config/schema/index.d.ts +8 -897
- package/src/config/schema/index.js +8 -506
- package/src/config/schema/{converter.schema.d.ts → manifest/converter.schema.d.ts} +4 -4
- package/src/config/schema/{converter.schema.js → manifest/converter.schema.js} +1 -1
- package/src/config/schema/manifest/index.d.ts +142 -0
- package/src/config/schema/manifest/index.js +62 -0
- package/src/config/schema/manifest/manifest-inference.schema.d.ts +298 -0
- package/src/config/schema/{manifest.schema.d.ts → manifest/manifest.schema.d.ts} +9 -291
- package/src/config/schema/{manifest.schema.js → manifest/manifest.schema.js} +2 -2
- package/src/config/schema/{emulation.schema.d.ts → platform/emulation.schema.d.ts} +1 -1
- package/src/config/schema/{emulation.schema.js → platform/emulation.schema.js} +6 -6
- package/src/config/schema/{gpu-cache.schema.js → platform/gpu-cache.schema.js} +1 -1
- package/src/config/schema/platform/index.d.ts +72 -0
- package/src/config/schema/platform/index.js +27 -0
- package/src/config/schema/{memory-limits.schema.js → platform/memory-limits.schema.js} +1 -1
- package/src/config/schema/policy/index.d.ts +14 -0
- package/src/config/schema/policy/index.js +9 -0
- package/src/config/schema/{buffer-pool.schema.js → runtime/buffer-pool.schema.js} +2 -2
- package/src/config/schema/{doppler.schema.d.ts → runtime/doppler.schema.d.ts} +3 -3
- package/src/config/schema/{doppler.schema.js → runtime/doppler.schema.js} +12 -12
- package/src/config/schema/runtime/index.d.ts +238 -0
- package/src/config/schema/runtime/index.js +135 -0
- package/src/config/schema/{loading.schema.d.ts → runtime/loading.schema.d.ts} +31 -1
- package/src/config/schema/runtime/loading.schema.js +275 -0
- package/src/config/schema/{shared-runtime.schema.d.ts → runtime/shared-runtime.schema.d.ts} +5 -5
- package/src/config/schema/{shared-runtime.schema.js → runtime/shared-runtime.schema.js} +4 -4
- package/src/config/schema/{tooling.schema.d.ts → runtime/tooling.schema.d.ts} +1 -1
- package/src/config/schema/training/index.d.ts +63 -0
- package/src/config/schema/training/index.js +43 -0
- package/src/config/schema/{lora.schema.js → training/lora.schema.js} +1 -1
- package/src/config/source-packages/litert/gemma-4-e2b-it.json +1 -1
- package/src/config/supported-operations.json +358 -88
- package/src/config/target-plan.d.ts +2 -2
- package/src/config/target-plan.js +2 -2
- package/src/config/training-defaults.d.ts +2 -2
- package/src/config/training-defaults.js +2 -2
- package/src/config/transforms/execution-graph/correctness-fallback.js +86 -0
- package/src/config/transforms/execution-graph/prefill-attention.js +116 -0
- package/src/{converter → config/transforms}/execution-graph-transforms.js +5 -194
- package/src/converter/core.js +18 -164
- package/src/converter/forge-stages.js +2 -2
- package/src/converter/fp8-source.js +1 -1
- package/src/converter/gemma4/validation.js +110 -0
- package/src/converter/linear-norm-gate.js +1 -1
- package/src/converter/manifest-builder.js +2 -9
- package/src/converter/parsers/gguf.js +54 -4
- package/src/converter/quantization-info.js +2 -2
- package/src/converter/qwen4-exp-source-execution.js +94 -0
- package/src/converter/qwen4-exp-source.js +1 -201
- package/src/converter/tensor-transform.js +2 -9
- package/src/converter/tokenizer-contract.js +2 -9
- package/src/converter/w8a8-quantizer.js +2 -2
- package/src/debug/config.js +1 -1
- package/src/debug/perf-guards.js +1 -1
- package/src/errors/doppler-error.d.ts +54 -0
- package/src/errors/doppler-error.js +13 -5
- package/src/experimental/distribution/shard-delivery/receipt.js +1 -1
- package/src/experimental/distribution/shard-delivery/retry.js +4 -4
- package/src/experimental/energy/types.d.ts +1 -1
- package/src/experimental/logits/cpu-output.js +1 -1
- package/src/experimental/training/artifacts.d.ts +2 -2
- package/src/experimental/training/artifacts.js +2 -2
- package/src/experimental/training/autograd.d.ts +1 -1
- package/src/experimental/training/distillation/student-fixture/execution.js +1 -1
- package/src/experimental/training/distillation/student-fixture/model.js +1 -1
- package/src/experimental/training/distillation/student-fixture.js +1 -1
- package/src/experimental/training/loss-scaling.d.ts +1 -1
- package/src/experimental/training/runner/evaluation.js +1 -1
- package/src/experimental/training/runner.js +1 -1
- package/src/experimental/training/trainer.d.ts +1 -1
- package/src/experimental/training/workloads.js +2 -2
- package/src/formats/gguf/block-storage.d.ts +9 -0
- package/src/formats/gguf/block-storage.js +12 -3
- package/src/formats/gguf/qwen-source.js +5 -3
- package/src/formats/gguf/qwen-source.json +8 -1
- package/src/formats/gguf/qwen4-exp-source.js +101 -0
- package/src/formats/q4k-dequant.js +1 -1
- package/src/formats/qwen4-exp-source-contracts.js +204 -0
- package/src/formats/rdrr/fp8-contract.js +1 -1
- package/src/formats/rdrr/fp8-storage.js +1 -1
- package/src/formats/rdrr/q4_0-storage.js +1 -1
- package/src/formats/rdrr/storage-descriptor.js +1 -1
- package/src/formats/rdrr/tensor-config-validator.js +3 -3
- package/src/formats/rdrr/validation.js +1 -1
- package/src/formats/rdrr/w8a8-storage.js +1 -1
- package/src/generation/index.d.ts +1 -1
- package/src/generation/index.js +1 -0
- package/src/gpu/device.d.ts +1 -1
- package/src/gpu/kernels/attention/contiguous-quantized.js +252 -0
- package/src/gpu/kernels/attention/tiered-quantized.js +321 -0
- package/src/gpu/kernels/attention/tiered.js +240 -0
- package/src/gpu/kernels/attention.js +5 -801
- package/src/gpu/kernels/backward/layernorm_backward.wgsl +3 -3
- package/src/gpu/kernels/channel-gain.js +100 -0
- package/src/gpu/kernels/channel-rotation.d.ts +16 -0
- package/src/gpu/kernels/channel-rotation.js +47 -0
- package/src/gpu/kernels/channel_energy.wgsl +17 -0
- package/src/gpu/kernels/channel_energy_pair.wgsl +27 -0
- package/src/gpu/kernels/channel_gain_control.wgsl +60 -0
- package/src/gpu/kernels/channel_gain_derive.wgsl +40 -0
- package/src/gpu/kernels/channel_gain_swiglu.wgsl +64 -0
- package/src/gpu/kernels/channel_rotation_apply.wgsl +38 -0
- package/src/gpu/kernels/channel_rotation_derive.wgsl +76 -0
- package/src/gpu/kernels/check-finiteness-f32.js +22 -0
- package/src/gpu/kernels/check_finiteness.wgsl +2 -2
- package/src/gpu/kernels/clamp.js +2 -2
- package/src/gpu/kernels/dequant.js +1 -1
- package/src/gpu/kernels/dispatch.js +7 -0
- package/src/gpu/kernels/energy_quintel_grad.wgsl +7 -7
- package/src/gpu/kernels/energy_quintel_grad_f16.wgsl +7 -7
- package/src/gpu/kernels/energy_quintel_reduce.wgsl +6 -6
- package/src/gpu/kernels/energy_quintel_reduce_f16.wgsl +6 -6
- package/src/gpu/kernels/energy_quintel_update.wgsl +7 -7
- package/src/gpu/kernels/energy_quintel_update_f16.wgsl +7 -7
- package/src/gpu/kernels/fused_matmul_q4_widetile_f32.wgsl +174 -0
- package/src/gpu/kernels/fused_matmul_rmsnorm.wgsl +0 -49
- package/src/gpu/kernels/gated-delta-norm-gate.js +1 -1
- package/src/gpu/kernels/gated_delta_recurrent_gguf_register_prefill_sigmoid_gate.wgsl +730 -0
- package/src/gpu/kernels/gated_delta_recurrent_gguf_sigmoid_gate.wgsl +499 -0
- package/src/gpu/kernels/gather.js +16 -1
- package/src/gpu/kernels/gather_split4_packed_f16_vec4_f32_out.wgsl +104 -0
- package/src/gpu/kernels/generated/uniform-writers-0.js +7 -11
- package/src/gpu/kernels/generated/uniform-writers-1.js +43 -46
- package/src/gpu/kernels/generated/uniform-writers-2.js +80 -2
- package/src/gpu/kernels/generated/uniform-writers.js +33 -21
- package/src/gpu/kernels/hyper-connection-up.js +64 -0
- package/src/gpu/kernels/hyper-connection.js +15 -4
- package/src/gpu/kernels/hyper_connection_mix_branch.wgsl +105 -0
- package/src/gpu/kernels/hyper_connection_norm.wgsl +3 -3
- package/src/gpu/kernels/hyper_connection_up_mix_branch_q4_0.wgsl +164 -0
- package/src/gpu/kernels/kernel-execution.js +18 -7
- package/src/gpu/kernels/linear-attention-core.js +3 -4
- package/src/gpu/kernels/moe-route.js +14 -0
- package/src/gpu/kernels/moe_gather.wgsl +0 -63
- package/src/gpu/kernels/moe_route_q4_0_block_f32.wgsl +74 -0
- package/src/gpu/kernels/moe_route_q4_0_split_block_f32.wgsl +80 -0
- package/src/gpu/kernels/moe_route_q4_1_block_f32.wgsl +59 -0
- package/src/gpu/kernels/moe_route_remap.wgsl +22 -0
- package/src/gpu/kernels/paired-linear-attention.js +1 -1
- package/src/gpu/kernels/pipeline-cache.js +18 -2
- package/src/gpu/kernels/ple.js +11 -7
- package/src/gpu/kernels/ple_q4_0_to_f32.wgsl +32 -0
- package/src/gpu/kernels/qsa-indexer.d.ts +2 -0
- package/src/gpu/kernels/qsa-indexer.js +6 -4
- package/src/gpu/kernels/qsa_block_keys.wgsl +2 -1
- package/src/gpu/kernels/qsa_prepare.wgsl +2 -1
- package/src/gpu/kernels/relu.js +1 -1
- package/src/gpu/kernels/relu.wgsl +2 -2
- package/src/gpu/kernels/relu_f16.wgsl +2 -2
- package/src/gpu/kernels/residual.js +1 -1
- package/src/gpu/kernels/residual.wgsl +4 -4
- package/src/gpu/kernels/residual_f16.wgsl +2 -2
- package/src/gpu/kernels/residual_f16_vec4.wgsl +2 -2
- package/src/gpu/kernels/residual_vec4.wgsl +2 -2
- package/src/gpu/kernels/rope-precompute.js +1 -1
- package/src/gpu/kernels/scatter_add.wgsl +0 -31
- package/src/gpu/kernels/shader-cache.js +17 -3
- package/src/gpu/kernels/shader-capability-filter.js +120 -0
- package/src/gpu/kernels/softmax.js +10 -10
- package/src/gpu/kernels/topk.wgsl +0 -55
- package/src/gpu/kernels/transpose.js +1 -1
- package/src/gpu/kernels/transpose.wgsl +2 -2
- package/src/gpu/kernels/w8a8.js +1 -1
- package/src/gpu/uniform-cache.d.ts +15 -1
- package/src/gpu/uniform-cache.js +82 -7
- package/src/gpu/weight-buffer.d.ts +1 -1
- package/src/inference/moe-router.d.ts +1 -126
- package/src/inference/moe-router.js +0 -149
- package/src/inference/pipelines/context.js +3 -15
- package/src/inference/pipelines/shader-scoped-pipeline.js +10 -5
- package/src/inference/pipelines/text/attention/interpreter/immediate.js +4 -4
- package/src/inference/pipelines/text/attention/interpreter/output.js +203 -0
- package/src/inference/pipelines/text/attention/interpreter/recorded.js +92 -67
- package/src/inference/pipelines/text/attention/interpreter.js +77 -350
- package/src/inference/pipelines/text/attention/projection-executor.js +2 -2
- package/src/inference/pipelines/text/attention/projection-plan.js +1 -1
- package/src/inference/pipelines/text/attention/projection-precision.js +132 -0
- package/src/inference/pipelines/text/attention/projections.js +13 -151
- package/src/inference/pipelines/text/attention/recorded-health.js +83 -0
- package/src/inference/pipelines/text/attention/rope-observation.js +1 -1
- package/src/inference/pipelines/text/attention/types.js +0 -13
- package/src/inference/pipelines/text/config/normalization.js +2 -2
- package/src/inference/pipelines/text/config.d.ts +4 -4
- package/src/inference/pipelines/text/config.js +17 -135
- package/src/inference/pipelines/text/{config-components.js → configuration/config-components.js} +4 -4
- package/src/inference/pipelines/text/configuration/model-geometry.js +114 -0
- package/src/inference/pipelines/text/{resolved-runtime-session.d.ts → configuration/resolved-runtime-session.d.ts} +4 -4
- package/src/inference/pipelines/text/{resolved-runtime-session.js → configuration/resolved-runtime-session.js} +1 -1
- package/src/inference/pipelines/text/{abort-contract.js → contracts/abort-contract.js} +2 -1
- package/src/inference/pipelines/text/contracts/metrics.d.ts +234 -0
- package/src/inference/pipelines/text/{embed.js → embedding/embed.js} +15 -15
- package/src/inference/pipelines/text/{embedding-batch.js → embedding/embedding-batch.js} +1 -1
- package/src/inference/pipelines/text/{embedding-extraction.js → embedding/embedding-extraction.js} +12 -12
- package/src/inference/pipelines/text/{embedding-normalization.js → embedding/embedding-normalization.js} +4 -4
- package/src/inference/pipelines/text/{execution-plan.d.ts → execution/execution-plan.d.ts} +8 -8
- package/src/inference/pipelines/text/{execution-plan.js → execution/execution-plan.js} +9 -9
- package/src/inference/pipelines/text/{execution-runtime-builders.js → execution/execution-runtime-builders.js} +3 -3
- package/src/inference/pipelines/text/{execution-v1.d.ts → execution/execution-v1.d.ts} +2 -2
- package/src/inference/pipelines/text/{execution-v1.js → execution/execution-v1.js} +13 -183
- package/src/inference/pipelines/text/execution/layer-contract.js +90 -0
- package/src/inference/pipelines/text/{layer-execution-contract.js → execution/layer-execution-contract.js} +2 -2
- package/src/inference/pipelines/text/{layer-input-execution.js → execution/layer-input-execution.js} +27 -36
- package/src/inference/pipelines/text/{layer-partition-contract.d.ts → execution/layer-partition-contract.d.ts} +3 -3
- package/src/inference/pipelines/text/{layer-partition-contract.js → execution/layer-partition-contract.js} +2 -2
- package/src/inference/pipelines/text/{layer-plan-gpu.js → execution/layer-plan-gpu.js} +12 -12
- package/src/inference/pipelines/text/{layer-plan.d.ts → execution/layer-plan.d.ts} +1 -1
- package/src/inference/pipelines/text/{layer-plan.js → execution/layer-plan.js} +1 -1
- package/src/inference/pipelines/text/{layer.js → execution/layer.js} +34 -142
- package/src/inference/pipelines/text/{operator-identity.js → execution/operator-identity.js} +1 -1
- package/src/inference/pipelines/text/{ops.js → execution/ops.js} +26 -26
- package/src/inference/pipelines/text/{paired-execution.js → execution/paired-execution.js} +9 -9
- package/src/inference/pipelines/text/execution-v1/hybrid-projection-contract.js +159 -0
- package/src/inference/pipelines/text/expert-loading-stats.d.ts +16 -0
- package/src/inference/pipelines/text/expert-loading-stats.js +15 -0
- package/src/inference/pipelines/text/ffn/dense-executor.js +5 -4
- package/src/inference/pipelines/text/ffn/dense-plan.js +23 -34
- package/src/inference/pipelines/text/ffn/dense.js +72 -392
- package/src/inference/pipelines/text/ffn/fused-gate-up.js +197 -0
- package/src/inference/pipelines/text/ffn/moe.js +1 -1
- package/src/inference/pipelines/text/ffn/packed-gate-up.js +218 -0
- package/src/inference/pipelines/text/ffn/paired.js +3 -3
- package/src/inference/pipelines/text/ffn/qwen-shared-expert.js +1 -1
- package/src/inference/pipelines/text/ffn/sandwich.js +6 -7
- package/src/inference/pipelines/text/ffn/standard.js +5 -5
- package/src/inference/pipelines/text/ffn/types.js +6 -3
- package/src/inference/pipelines/text/{activity-weights.js → generator/activation/activity-weights.js} +8 -8
- package/src/inference/pipelines/text/generator/activation/channel-gain.d.ts +61 -0
- package/src/inference/pipelines/text/generator/activation/channel-gain.js +352 -0
- package/src/inference/pipelines/text/generator/activation/final-layer-replay.js +43 -0
- package/src/inference/pipelines/text/generator/{capsule-generation.js → adapters/capsule-generation.js} +5 -5
- package/src/inference/pipelines/text/generator/{diffusion.js → adapters/diffusion.js} +19 -19
- package/src/inference/pipelines/text/generator/{training.js → adapters/training.js} +21 -21
- package/src/inference/pipelines/text/generator/decode/advance-token.js +73 -0
- package/src/inference/pipelines/text/{generator-steps.js → generator/decode/batch-generation.js} +25 -288
- package/src/inference/pipelines/text/generator/decode/batch-readback.js +175 -0
- package/src/inference/pipelines/text/generator/{compiled-greedy-window.js → decode/compiled-greedy-window.js} +5 -5
- package/src/inference/pipelines/text/generator/{decode-components.js → decode/decode-components.js} +1 -1
- package/src/inference/pipelines/text/generator/{decode-runtime.js → decode/decode-loop.js} +20 -412
- package/src/inference/pipelines/text/generator/decode/decode-mechanics.js +288 -0
- package/src/inference/pipelines/text/generator/{decode-readback.js → decode/decode-readback.js} +1 -1
- package/src/inference/pipelines/text/generator/decode/decode-runtime.js +2 -0
- package/src/inference/pipelines/text/generator/{decode.js → decode/decode-step.js} +34 -313
- package/src/inference/pipelines/text/generator/decode/decode.js +5 -0
- package/src/inference/pipelines/text/generator/{greedy-head-policy.js → decode/greedy-head-policy.js} +1 -1
- package/src/inference/pipelines/text/generator/{recorded-step-executor.js → decode/recorded-step-executor.js} +2 -2
- package/src/inference/pipelines/text/generator/generator-steps.js +4 -0
- package/src/inference/pipelines/text/{generator.d.ts → generator/generator.d.ts} +13 -11
- package/src/inference/pipelines/text/{generator.js → generator/generator.js} +150 -46
- package/src/inference/pipelines/text/generator/{prefill-runtime.js → prefill/prefill-execution.js} +72 -278
- package/src/inference/pipelines/text/generator/prefill/prefill-logits.js +144 -0
- package/src/inference/pipelines/text/generator/{prefill-observation.js → prefill/prefill-observation.js} +2 -2
- package/src/inference/pipelines/text/generator/{prefill-policy.js → prefill/prefill-policy.js} +6 -1
- package/src/inference/pipelines/text/generator/prefill/prefill-runtime.js +2 -0
- package/src/inference/pipelines/text/generator/prefill/prefill-sequence.js +100 -0
- package/src/inference/pipelines/text/generator/{prefix-embedding.js → prefill/prefix-embedding.js} +7 -7
- package/src/inference/pipelines/text/{generation-request.d.ts → generator/request/generation-request.d.ts} +2 -2
- package/src/inference/pipelines/text/{generation-request.js → generator/request/generation-request.js} +3 -3
- package/src/inference/pipelines/text/generator/{prompt-input.js → request/prompt-input.js} +2 -2
- package/src/inference/pipelines/text/{sampling-config.d.ts → generator/sampling/sampling-config.d.ts} +1 -1
- package/src/inference/pipelines/text/{sampling-config.js → generator/sampling/sampling-config.js} +2 -2
- package/src/inference/pipelines/text/{sampling.d.ts → generator/sampling/sampling.d.ts} +15 -0
- package/src/inference/pipelines/text/{sampling.js → generator/sampling/sampling.js} +22 -3
- package/src/inference/pipelines/text/{stopping.js → generator/sampling/stopping.js} +3 -3
- package/src/inference/pipelines/text/generator/{token-selection.d.ts → sampling/token-selection.d.ts} +1 -1
- package/src/inference/pipelines/text/generator/{token-selection.js → sampling/token-selection.js} +8 -8
- package/src/inference/pipelines/text/generator/{activation-storage.js → session/activation-storage.js} +2 -2
- package/src/inference/pipelines/text/generator/{attention-lifecycle.js → session/attention-lifecycle.js} +1 -1
- package/src/inference/pipelines/text/generator/{diagnostic-sequence.d.ts → session/diagnostic-sequence.d.ts} +3 -3
- package/src/inference/pipelines/text/generator/{diagnostic-sequence.js → session/diagnostic-sequence.js} +3 -3
- package/src/inference/pipelines/text/generator/session/generation-runtime.js +278 -0
- package/src/inference/pipelines/text/{generator-runtime.d.ts → generator/session/generator-runtime.d.ts} +7 -7
- package/src/inference/pipelines/text/{generator-runtime.js → generator/session/generator-runtime.js} +8 -8
- package/src/inference/pipelines/text/generator/{recovery.js → session/recovery.js} +6 -6
- package/src/inference/pipelines/text/{sequence-state.d.ts → generator/session/sequence-state.d.ts} +5 -5
- package/src/inference/pipelines/text/{sequence-state.js → generator/session/sequence-state.js} +26 -18
- package/src/inference/pipelines/text/generator/{session-context.js → session/session-context.js} +12 -11
- package/src/inference/pipelines/text/generator/{speculative-runtime.js → session/speculative-runtime.js} +1 -1
- package/src/inference/pipelines/text/generator/{text.js → session/text.js} +9 -9
- package/src/inference/pipelines/text/{warm-generation.d.ts → generator/session/warm-generation.d.ts} +1 -1
- package/src/inference/pipelines/text/{warm-generation.js → generator/session/warm-generation.js} +1 -1
- package/src/inference/pipelines/text/generator/{whole-token-lane.js → whole-token/whole-token-lane.js} +10 -10
- package/src/inference/pipelines/text/generator/{whole-token-observation.d.ts → whole-token/whole-token-observation.d.ts} +2 -2
- package/src/inference/pipelines/text/generator/{whole-token-observation.js → whole-token/whole-token-observation.js} +5 -5
- package/src/inference/pipelines/text/generator/{whole-token-program.d.ts → whole-token/whole-token-program.d.ts} +3 -3
- package/src/inference/pipelines/text/generator/{whole-token-program.js → whole-token/whole-token-program.js} +5 -5
- package/src/inference/pipelines/text/hyper-connection/decoder-layer.d.ts +17 -4
- package/src/inference/pipelines/text/hyper-connection/decoder-layer.js +38 -14
- package/src/inference/pipelines/text/hyper-connection/layer-executor.js +14 -9
- package/src/inference/pipelines/text/hyper-connection/logits-input.js +5 -3
- package/src/inference/pipelines/text/hyper-connection/projection.js +52 -2
- package/src/inference/pipelines/text/hyper-connection/sequence-transaction.js +3 -3
- package/src/inference/pipelines/text/hyper-connection/session.d.ts +1 -1
- package/src/inference/pipelines/text/hyper-connection/session.js +9 -5
- package/src/inference/pipelines/text/indexed-attention/qsa-cache.d.ts +2 -1
- package/src/inference/pipelines/text/indexed-attention/qsa-cache.js +3 -3
- package/src/inference/pipelines/text/init.d.ts +4 -2
- package/src/inference/pipelines/text/init.js +10 -934
- package/src/inference/pipelines/text/initialization/init-emulation.js +76 -0
- package/src/inference/pipelines/text/initialization/init-kv-cache.js +547 -0
- package/src/inference/pipelines/text/initialization/init-moe.js +34 -0
- package/src/inference/pipelines/text/{init-qkv-fusion.js → initialization/init-qkv-fusion.js} +7 -7
- package/src/inference/pipelines/text/{init-rope.d.ts → initialization/init-rope.d.ts} +4 -4
- package/src/inference/pipelines/text/{init-rope.js → initialization/init-rope.js} +9 -11
- package/src/inference/pipelines/text/initialization/init-speculative.js +31 -0
- package/src/inference/pipelines/text/{init-tokenizer.js → initialization/init-tokenizer.js} +2 -2
- package/src/inference/pipelines/text/lifecycle.js +18 -18
- package/src/inference/pipelines/text/{linear-attention-ab-fusion.js → linear-attention/linear-attention-ab-fusion.js} +3 -3
- package/src/inference/pipelines/text/{linear-attention.d.ts → linear-attention/linear-attention.d.ts} +12 -12
- package/src/inference/pipelines/text/{linear-attention.js → linear-attention/linear-attention.js} +17 -17
- package/src/inference/pipelines/text/linear-attention/plan.js +4 -4
- package/src/inference/pipelines/text/linear-attention/state-checkpoint.js +1 -1
- package/src/inference/pipelines/text/linear-attention/state-clone.js +1 -1
- package/src/inference/pipelines/text/linear-attention/state-layout.js +2 -2
- package/src/inference/pipelines/text/loading/decode-session-config.js +295 -0
- package/src/inference/pipelines/text/loading/init-weights.js +254 -0
- package/src/inference/pipelines/text/{load-errors.js → loading/load-errors.js} +3 -3
- package/src/inference/pipelines/text/{model-load.js → loading/model-load.js} +14 -306
- package/src/inference/pipelines/text/{weight-loading-config.js → loading/weight-loading-config.js} +6 -6
- package/src/inference/pipelines/text/{weights.d.ts → loading/weights.d.ts} +13 -13
- package/src/inference/pipelines/text/{weights.js → loading/weights.js} +5 -5
- package/src/inference/pipelines/text/logits/gpu-executor.js +2 -2
- package/src/inference/pipelines/text/logits/gpu.d.ts +4 -4
- package/src/inference/pipelines/text/logits/gpu.js +2 -2
- package/src/inference/pipelines/text/{generator → logits}/hidden-projection.js +1 -1
- package/src/inference/pipelines/text/logits/index.d.ts +1 -1
- package/src/inference/pipelines/text/logits/index.js +3 -3
- package/src/inference/pipelines/text/{generator → logits}/logits-config.js +1 -1
- package/src/inference/pipelines/text/logits/output-transform.js +1 -1
- package/src/inference/pipelines/text/logits/precision-policy.js +1 -1
- package/src/inference/pipelines/text/logits/types.d.ts +2 -2
- package/src/inference/pipelines/text/lora-types.d.ts +1 -1
- package/src/inference/pipelines/text/moe/expert-loading.js +4 -1
- package/src/inference/pipelines/text/moe/gpu-executor.js +44 -101
- package/src/inference/pipelines/text/moe/grouped-prefill.js +6 -2
- package/src/inference/pipelines/text/moe/qwen-route-executor.d.ts +2 -2
- package/src/inference/pipelines/text/moe/qwen-route-executor.js +31 -2
- package/src/inference/pipelines/text/moe/routed-expert-chunks.js +99 -15
- package/src/inference/pipelines/text/moe-cache.js +11 -77
- package/src/inference/pipelines/text/moe-gpu.js +33 -75
- package/src/inference/pipelines/text/moe-impl.d.ts +3 -38
- package/src/inference/pipelines/text/moe-impl.js +1 -7
- package/src/inference/pipelines/text/ngram/ngram-addresses.d.ts +1 -1
- package/src/inference/pipelines/text/ngram/ngram-addresses.js +2 -2
- package/src/inference/pipelines/text/ngram/ngram-table.d.ts +3 -3
- package/src/inference/pipelines/text/ngram/ngram-table.js +5 -6
- package/src/inference/pipelines/text/ngram/ple-layer.d.ts +3 -3
- package/src/inference/pipelines/text/ngram/ple-layer.js +3 -3
- package/src/inference/pipelines/text/{generator-logits-observation.js → observation/generator-logits-observation.js} +4 -4
- package/src/inference/pipelines/text/{kernel-trace.js → observation/kernel-trace.js} +1 -1
- package/src/inference/pipelines/text/{operator-events.js → observation/operator-events.js} +1 -1
- package/src/inference/pipelines/text/{probes.js → observation/probes.js} +13 -13
- package/src/inference/pipelines/text/{tsir-fixture-writer.js → observation/tsir-fixture-writer.js} +3 -3
- package/src/inference/pipelines/text/{execution.js → operations/execution.js} +14 -14
- package/src/inference/pipelines/text/{image-transcription.js → operations/image-transcription.js} +8 -8
- package/src/inference/pipelines/text/{partition-attempt.js → partitions/partition-attempt.js} +7 -7
- package/src/inference/pipelines/text/{partition-execution.js → partitions/partition-execution.js} +107 -35
- package/src/inference/pipelines/text/partitions/partition-wire.js +50 -0
- package/src/inference/pipelines/text/{resident-partition-contract.d.ts → partitions/resident-partition-contract.d.ts} +10 -8
- package/src/inference/pipelines/text/{resident-partition-contract.js → partitions/resident-partition-contract.js} +6 -6
- package/src/inference/pipelines/text/partitions/resident-partition-warmup.js +67 -0
- package/src/inference/pipelines/text/{resident-partition.js → partitions/resident-partition.js} +27 -15
- package/src/inference/pipelines/text/{resident-pipeline-preparation.js → partitions/resident-pipeline-preparation.js} +32 -32
- package/src/inference/pipelines/text/per-layer/materialize.js +4 -4
- package/src/inference/pipelines/text/per-layer/per-layer-inputs.js +5 -0
- package/src/inference/pipelines/text/per-layer/plan.js +4 -4
- package/src/inference/pipelines/text/{per-layer-inputs.js → per-layer/prepare.js} +14 -280
- package/src/inference/pipelines/text/per-layer/projection-contract.js +79 -0
- package/src/inference/pipelines/text/per-layer/runtime-cache.js +183 -0
- package/src/inference/pipelines/text/pipeline.d.ts +13 -8
- package/src/inference/pipelines/text/pipeline.js +78 -22
- package/src/inference/pipelines/text/{drift-policy.js → policy/drift-policy.js} +1 -1
- package/src/inference/pipelines/text/{dtype-contract.js → policy/dtype-contract.js} +2 -2
- package/src/inference/pipelines/text/{finiteness-guard-status.js → policy/finiteness-guard-status.js} +2 -1
- package/src/inference/pipelines/text/{generator-decode-policy.js → policy/generator-decode-policy.js} +5 -5
- package/src/inference/pipelines/text/state.d.ts +9 -8
- package/src/inference/pipelines/text/state.js +1 -0
- package/src/inference/pipelines/text/types.d.ts +14 -218
- package/src/inference/pipelines/text.js +2 -0
- package/src/inference/tokenizers/bundled/merge.js +130 -0
- package/src/inference/tokenizers/bundled.js +2 -36
- package/src/loader/doppler-loader.d.ts +17 -7
- package/src/loader/doppler-loader.js +27 -29
- package/src/loader/embedding-limit-preflight.js +29 -1
- package/src/loader/embedding-loader.js +18 -4
- package/src/loader/experts/compact-expert-cache.js +508 -0
- package/src/loader/experts/expert-cache.d.ts +1 -1
- package/src/loader/experts/expert-dequant-cache.d.ts +9 -0
- package/src/loader/experts/expert-dequant-cache.js +75 -0
- package/src/loader/experts/expert-frequency-index.js +129 -0
- package/src/loader/experts/expert-loader.d.ts +4 -0
- package/src/loader/experts/expert-loader.js +13 -1
- package/src/loader/experts/expert-preparation.d.ts +1 -1
- package/src/loader/experts/expert-preparation.js +1 -1
- package/src/loader/experts/loader-bindings.js +129 -14
- package/src/loader/experts/mapped-expert-staging.js +393 -0
- package/src/loader/experts/mapped-expert-upload-cache.d.ts +61 -0
- package/src/loader/experts/mapped-expert-upload-cache.js +382 -0
- package/src/loader/experts/qwen-paging.d.ts +1 -1
- package/src/loader/experts/qwen-paging.js +1 -1
- package/src/loader/experts/routed-expert-cache.d.ts +70 -3
- package/src/loader/experts/routed-expert-cache.js +452 -99
- package/src/loader/experts/routed-expert-plan.d.ts +2 -0
- package/src/loader/experts/routed-expert-plan.js +14 -8
- package/src/loader/experts/routed-expert-prefill.js +158 -0
- package/src/loader/experts/routed-expert-source.js +234 -0
- package/src/loader/experts/routed-expert-trace.js +77 -0
- package/src/loader/experts/tiered-expert-cache.d.ts +12 -0
- package/src/loader/experts/tiered-expert-cache.js +327 -0
- package/src/loader/experts/tiered-expert-plan.js +74 -0
- package/src/loader/final-weights-loader.d.ts +1 -1
- package/src/loader/hyper-connection-loader.d.ts +2 -2
- package/src/loader/hyper-connection-loader.js +85 -13
- package/src/loader/layer-loader.d.ts +3 -3
- package/src/loader/layer-loader.js +1 -1
- package/src/loader/layer-tensor-aliases.js +14 -11
- package/src/loader/loader-types.d.ts +6 -1
- package/src/loader/manifest-config.js +4 -0
- package/src/loader/model-executor.js +18 -1
- package/src/loader/model-load.d.ts +1 -1
- package/src/loader/model-load.js +1 -1
- package/src/loader/ngram-table-bindings.d.ts +3 -3
- package/src/loader/ngram-table-bindings.js +7 -6
- package/src/loader/q4k-reference-weight.js +1 -1
- package/src/loader/shard-cache.d.ts +8 -1
- package/src/loader/shard-cache.js +33 -0
- package/src/loader/tensors/bf16-loader.js +2 -1
- package/src/loader/tensors/fp8-loader.js +1 -1
- package/src/loader/tensors/load-executor.js +10 -2
- package/src/loader/tensors/tensor-loader.js +1 -1
- package/src/loader/tensors/tensor-reader.js +18 -5
- package/src/loader/tensors/w8a8-loader.js +1 -1
- package/src/loader/weight-representation-plan.d.ts +1 -1
- package/src/memory/buffer-pool/readback.js +82 -0
- package/src/memory/buffer-pool.d.ts +11 -2
- package/src/memory/buffer-pool.js +12 -73
- package/src/partitions.d.ts +4 -3
- package/src/partitions.js +3 -2
- package/src/{converter → rules}/capability-transform-resolver.js +3 -3
- package/src/rules/generated/rule-bundle.json +1 -1
- package/src/rules/inference/capability-transforms.rules.json +95 -28
- package/src/rules/rule-registry.js +2 -0
- package/src/storage/artifact-range-loaders.js +42 -0
- package/src/storage/artifact-storage-context.d.ts +4 -0
- package/src/storage/artifact-storage-context.js +8 -18
- package/src/storage/backends/idb-store.js +1 -1
- package/src/storage/distribution-transport.js +1 -1
- package/src/storage/download/http-contract.js +1 -1
- package/src/storage/download/http-transport.js +4 -4
- package/src/storage/inventory.js +1 -1
- package/src/storage/model-cache.d.ts +2 -1
- package/src/storage/model-cache.js +12 -8
- package/src/storage/node-file-access.js +4 -0
- package/src/storage/quota.d.ts +2 -1
- package/src/storage/quota.js +2 -1
- package/src/storage/shard-manager.d.ts +1 -1
- package/src/storage/shards/integrity.js +17 -0
- package/src/storage/source-artifact-store.d.ts +1 -1
- package/src/storage/source-artifact-store.js +1 -0
- package/src/storage/source-storage-context.d.ts +4 -0
- package/src/storage/source-storage-context.js +14 -0
- package/src/storage/verified-piece-direct.js +153 -0
- package/src/storage/verified-piece-gather-lane.js +210 -0
- package/src/storage/verified-piece-gather-worker.js +739 -0
- package/src/storage/verified-piece-gather.d.ts +160 -0
- package/src/storage/verified-piece-gather.js +315 -0
- package/src/storage/verified-piece-reader.d.ts +15 -1
- package/src/storage/verified-piece-reader.js +85 -15
- package/src/storage/verified-piece-state-cache.d.ts +75 -0
- package/src/storage/verified-piece-state-cache.js +511 -0
- package/src/storage/verified-piece-storage.d.ts +22 -1
- package/src/storage/verified-piece-storage.js +299 -11
- package/src/storage/verified-piece-timings.js +25 -0
- package/src/tooling/browser-command-runner.js +1 -1
- package/src/tooling/browser-harness/benchmark-execution.js +2 -2
- package/src/tooling/browser-harness/benchmark-rerank.js +2 -2
- package/src/tooling/browser-harness/inference-execution.js +1 -1
- package/src/tooling/browser-harness/model-resolution.js +1 -1
- package/src/tooling/browser-harness/runtime-config.js +1 -1
- package/src/tooling/browser-harness/runtime-isolation.js +2 -5
- package/src/tooling/browser-harness/text-evidence.js +1 -1
- package/src/tooling/browser-harness-contract-helpers.js +1 -1
- package/src/tooling/browser-harness.js +2 -2
- package/src/tooling/capsule-v2.js +2 -2
- package/src/tooling/capsule.js +2 -2
- package/src/tooling/command-api-constants.js +4 -2
- package/src/tooling/command-api.d.ts +1 -1
- package/src/tooling/command-api.js +15 -21
- package/src/tooling/command-runtime-execution.js +1 -8
- package/src/tooling/diagnose-runner.js +3 -41
- package/src/tooling/node-browser/static-server.js +2 -2
- package/src/tooling/node-command-execution.js +2 -34
- package/src/tooling/node-converter/input.js +1 -1
- package/src/tooling/node-converter/output.js +2 -2
- package/src/tooling/node-converter.js +1 -1
- package/src/tooling/node-runtime-bridge.js +43 -0
- package/src/tooling/node-source-runtime.js +34 -35
- package/src/tooling/production-release.js +1 -1
- package/src/tooling/program-bundle/materialize.js +1 -1
- package/src/tooling/program-bundle-parity.d.ts +1 -1
- package/src/tooling/program-bundle-parity.js +1 -1
- package/src/tooling/program-bundle.d.ts +1 -1
- package/src/tooling/program-bundle.js +2 -2
- package/src/tooling/runtime-intent.js +1 -1
- package/src/tooling/source-artifact-adapter.js +3 -3
- package/src/tooling/source-runtime/node-loader.js +32 -5
- package/src/tooling/source-runtime-converter-config.js +28 -6
- package/src/tooling-exports.shared.d.ts +3 -3
- package/src/tooling-exports.shared.js +2 -2
- package/src/version.js +1 -1
- package/tests/kernels/browser/kernel-suite.js +1 -1
- package/tests/kernels/harness/benchmark.js +1 -1
- package/tools/README.md +74 -0
- package/src/config/runtime/profiles/qwen3-6-35b-a3b-q4_0-pooled-q8-common-diagnostic.json +0 -4407
- package/src/config/schema/loading.schema.js +0 -164
- package/src/inference/pipelines/text/moe-cpu.js +0 -20
- /package/src/config/{capsule-loading.d.ts → capsule/capsule-loading.d.ts} +0 -0
- /package/src/config/{capsule-release-contract.d.ts → capsule/capsule-release-contract.d.ts} +0 -0
- /package/src/config/{capsule-signature.d.ts → capsule/capsule-signature.d.ts} +0 -0
- /package/src/config/{capsule-signature.js → capsule/capsule-signature.js} +0 -0
- /package/src/config/{capsule-v3.d.ts → capsule/capsule-v3.d.ts} +0 -0
- /package/src/config/{capsule.d.ts → capsule/capsule.d.ts} +0 -0
- /package/src/config/schema/{diffusion.schema.d.ts → execution/diffusion.schema.d.ts} +0 -0
- /package/src/config/schema/{diffusion.schema.js → execution/diffusion.schema.js} +0 -0
- /package/src/config/schema/{energy.schema.d.ts → execution/energy.schema.d.ts} +0 -0
- /package/src/config/schema/{energy.schema.js → execution/energy.schema.js} +0 -0
- /package/src/config/schema/{hyper-connection-decoder.schema.d.ts → execution/hyper-connection-decoder.schema.d.ts} +0 -0
- /package/src/config/schema/{hyper-connection-decoder.schema.js → execution/hyper-connection-decoder.schema.js} +0 -0
- /package/src/config/schema/{inference.schema.js → execution/inference.schema.js} +0 -0
- /package/src/config/schema/{kernel-path.schema.js → execution/kernel-path.schema.js} +0 -0
- /package/src/config/schema/{kernel-registry.schema.js → execution/kernel-registry.schema.js} +0 -0
- /package/src/config/schema/{speculation-self.schema.d.ts → execution/speculation-self.schema.d.ts} +0 -0
- /package/src/config/schema/{speculation-self.schema.js → execution/speculation-self.schema.js} +0 -0
- /package/src/config/schema/{speculative.schema.d.ts → execution/speculative.schema.d.ts} +0 -0
- /package/src/config/schema/{speculative.schema.js → execution/speculative.schema.js} +0 -0
- /package/src/config/schema/{conversion-report.schema.d.ts → manifest/conversion-report.schema.d.ts} +0 -0
- /package/src/config/schema/{conversion-report.schema.js → manifest/conversion-report.schema.js} +0 -0
- /package/src/config/schema/{conversion.schema.d.ts → manifest/conversion.schema.d.ts} +0 -0
- /package/src/config/schema/{conversion.schema.js → manifest/conversion.schema.js} +0 -0
- /package/src/config/schema/{fp8.schema.d.ts → manifest/fp8.schema.d.ts} +0 -0
- /package/src/config/schema/{fp8.schema.js → manifest/fp8.schema.js} +0 -0
- /package/src/config/schema/{manifest-rope.schema.d.ts → manifest/manifest-rope.schema.d.ts} +0 -0
- /package/src/config/schema/{quantization-defaults.schema.d.ts → manifest/quantization-defaults.schema.d.ts} +0 -0
- /package/src/config/schema/{quantization-defaults.schema.js → manifest/quantization-defaults.schema.js} +0 -0
- /package/src/config/schema/{quantization.schema.d.ts → manifest/quantization.schema.d.ts} +0 -0
- /package/src/config/schema/{quantization.schema.js → manifest/quantization.schema.js} +0 -0
- /package/src/config/schema/{w8a8.schema.d.ts → manifest/w8a8.schema.d.ts} +0 -0
- /package/src/config/schema/{w8a8.schema.js → manifest/w8a8.schema.js} +0 -0
- /package/src/config/schema/{gpu-cache.schema.d.ts → platform/gpu-cache.schema.d.ts} +0 -0
- /package/src/config/schema/{memory-limits.schema.d.ts → platform/memory-limits.schema.d.ts} +0 -0
- /package/src/config/schema/{platform.schema.d.ts → platform/platform.schema.d.ts} +0 -0
- /package/src/config/schema/{support-tiers.schema.d.ts → policy/support-tiers.schema.d.ts} +0 -0
- /package/src/config/schema/{support-tiers.schema.js → policy/support-tiers.schema.js} +0 -0
- /package/src/config/schema/{adapter.schema.d.ts → runtime/adapter.schema.d.ts} +0 -0
- /package/src/config/schema/{adapter.schema.js → runtime/adapter.schema.js} +0 -0
- /package/src/config/schema/{benchmark.schema.d.ts → runtime/benchmark.schema.d.ts} +0 -0
- /package/src/config/schema/{benchmark.schema.js → runtime/benchmark.schema.js} +0 -0
- /package/src/config/schema/{bridge.schema.d.ts → runtime/bridge.schema.d.ts} +0 -0
- /package/src/config/schema/{bridge.schema.js → runtime/bridge.schema.js} +0 -0
- /package/src/config/schema/{browser-suite-metrics.schema.d.ts → runtime/browser-suite-metrics.schema.d.ts} +0 -0
- /package/src/config/schema/{browser-suite-metrics.schema.js → runtime/browser-suite-metrics.schema.js} +0 -0
- /package/src/config/schema/{buffer-pool.schema.d.ts → runtime/buffer-pool.schema.d.ts} +0 -0
- /package/src/config/schema/{debug.schema.d.ts → runtime/debug.schema.d.ts} +0 -0
- /package/src/config/schema/{debug.schema.js → runtime/debug.schema.js} +0 -0
- /package/src/config/schema/{distribution.schema.d.ts → runtime/distribution.schema.d.ts} +0 -0
- /package/src/config/schema/{distribution.schema.js → runtime/distribution.schema.js} +0 -0
- /package/src/config/schema/{ecosystem.schema.d.ts → runtime/ecosystem.schema.d.ts} +0 -0
- /package/src/config/schema/{ecosystem.schema.js → runtime/ecosystem.schema.js} +0 -0
- /package/src/config/schema/{harness.schema.d.ts → runtime/harness.schema.d.ts} +0 -0
- /package/src/config/schema/{harness.schema.js → runtime/harness.schema.js} +0 -0
- /package/src/config/schema/{hotswap.schema.d.ts → runtime/hotswap.schema.d.ts} +0 -0
- /package/src/config/schema/{hotswap.schema.js → runtime/hotswap.schema.js} +0 -0
- /package/src/config/schema/{intent-bundle.schema.d.ts → runtime/intent-bundle.schema.d.ts} +0 -0
- /package/src/config/schema/{intent-bundle.schema.js → runtime/intent-bundle.schema.js} +0 -0
- /package/src/config/schema/{runtime-profile.schema.d.ts → runtime/runtime-profile.schema.d.ts} +0 -0
- /package/src/config/schema/{runtime-profile.schema.js → runtime/runtime-profile.schema.js} +0 -0
- /package/src/config/schema/{storage.schema.d.ts → runtime/storage.schema.d.ts} +0 -0
- /package/src/config/schema/{storage.schema.js → runtime/storage.schema.js} +0 -0
- /package/src/config/schema/{tooling.schema.js → runtime/tooling.schema.js} +0 -0
- /package/src/config/schema/{units.schema.d.ts → runtime/units.schema.d.ts} +0 -0
- /package/src/config/schema/{units.schema.js → runtime/units.schema.js} +0 -0
- /package/src/config/schema/{backward-registry.schema.d.ts → training/backward-registry.schema.d.ts} +0 -0
- /package/src/config/schema/{backward-registry.schema.js → training/backward-registry.schema.js} +0 -0
- /package/src/config/schema/{distill-training.schema.d.ts → training/distill-training.schema.d.ts} +0 -0
- /package/src/config/schema/{distill-training.schema.js → training/distill-training.schema.js} +0 -0
- /package/src/config/schema/{lora.schema.d.ts → training/lora.schema.d.ts} +0 -0
- /package/src/config/schema/{training-metrics.schema.d.ts → training/training-metrics.schema.d.ts} +0 -0
- /package/src/config/schema/{training-metrics.schema.js → training/training-metrics.schema.js} +0 -0
- /package/src/config/schema/{training.schema.d.ts → training/training.schema.d.ts} +0 -0
- /package/src/config/schema/{training.schema.js → training/training.schema.js} +0 -0
- /package/src/config/schema/{ul-training.schema.d.ts → training/ul-training.schema.d.ts} +0 -0
- /package/src/config/schema/{ul-training.schema.js → training/ul-training.schema.js} +0 -0
- /package/src/{converter → config/transforms}/execution-graph/dtype.js +0 -0
- /package/src/{converter → config/transforms}/execution-graph/fusion.js +0 -0
- /package/src/{converter → config/transforms}/execution-graph/kv-storage.js +0 -0
- /package/src/{converter → config/transforms}/execution-graph/validation.js +0 -0
- /package/src/{converter → formats}/qwen4-exp-source-defaults.json +0 -0
- /package/src/inference/pipelines/text/{buffer-types.d.ts → contracts/buffer-types.d.ts} +0 -0
- /package/src/inference/pipelines/text/{embedding-contract.d.ts → embedding/embedding-contract.d.ts} +0 -0
- /package/src/inference/pipelines/text/{embedding-contract.js → embedding/embedding-contract.js} +0 -0
- /package/src/inference/pipelines/text/{precomputed-input-norm.js → execution/precomputed-input-norm.js} +0 -0
- /package/src/inference/pipelines/text/generator/{batch-retry-policy.js → decode/batch-retry-policy.js} +0 -0
- /package/src/inference/pipelines/text/{chat-format.d.ts → generator/request/chat-format.d.ts} +0 -0
- /package/src/inference/pipelines/text/{chat-format.js → generator/request/chat-format.js} +0 -0
- /package/src/inference/pipelines/text/generator/{recorder-stats.js → session/recorder-stats.js} +0 -0
- /package/src/inference/pipelines/text/generator/{whole-token-template.js → whole-token/whole-token-template.js} +0 -0
- /package/src/inference/pipelines/text/{init-chat-templates.js → initialization/init-chat-templates.js} +0 -0
- /package/src/inference/pipelines/text/{load-policy.js → loading/load-policy.js} +0 -0
- /package/src/inference/pipelines/text/{pipeline-load-timing.js → loading/pipeline-load-timing.js} +0 -0
- /package/src/inference/pipelines/text/{generator → logits}/hidden-projection.d.ts +0 -0
- /package/src/inference/pipelines/text/{operator-events.d.ts → observation/operator-events.d.ts} +0 -0
- /package/src/inference/pipelines/text/{stage-names.d.ts → observation/stage-names.d.ts} +0 -0
- /package/src/inference/pipelines/text/{stage-names.js → observation/stage-names.js} +0 -0
- /package/src/inference/pipelines/text/{tsir-fixture-writer.d.ts → observation/tsir-fixture-writer.d.ts} +0 -0
- /package/src/inference/pipelines/text/{generator → per-layer}/per-layer-buffer.js +0 -0
- /package/src/inference/pipelines/text/{finiteness-policy.js → policy/finiteness-policy.js} +0 -0
|
@@ -1,47 +1,130 @@
|
|
|
1
|
+
import { isRoutedExpertTraceActive, readRoutedExpertSelection } from '../../../../loader/experts/routed-expert-trace.js';
|
|
2
|
+
import { getRuntimeConfig } from '../../../../config/runtime.js';
|
|
1
3
|
import { acquireBuffer, readBufferSlice, releaseBuffer } from '../../../../memory/buffer-pool.js';
|
|
2
4
|
import { getDevice } from '../../../../gpu/device.js';
|
|
3
5
|
import { createTensor } from '../../../../gpu/tensor.js';
|
|
4
|
-
import {
|
|
6
|
+
import { runRouteRemap } from '../../../../gpu/kernels/moe-route.js';
|
|
7
|
+
import { runQwenRouteExperts, recordQwenRouteExperts } from './qwen-route-executor.js';
|
|
8
|
+
|
|
9
|
+
/** @param {import('../../../../loader/experts/routed-expert-cache.js').RoutedExpertWeights} weights
|
|
10
|
+
* @param {GPUBuffer} indices @param {number} numTokens @param {number} topK
|
|
11
|
+
* @param {number[]} ids @param {boolean} allowPrefill @param {string | undefined} phase
|
|
12
|
+
* @param {import('../../../../gpu/command-recorder.js').CommandRecorder | null} recorder */
|
|
13
|
+
async function remapCompactIndices(weights, indices, numTokens, topK, ids, allowPrefill, phase, recorder) {
|
|
14
|
+
if (!weights.sourceExpertToSlot) return null;
|
|
15
|
+
const sourceNumExperts = weights.sourceNumExperts;
|
|
16
|
+
if (!((phase === 'decode' && numTokens === 1) || (allowPrefill && phase === 'prefill'))
|
|
17
|
+
|| !Number.isSafeInteger(sourceNumExperts) || !sourceNumExperts
|
|
18
|
+
|| ids.some(id => id >= sourceNumExperts)) {
|
|
19
|
+
throw new Error('Compact expert slots require valid source IDs and an admitted decode or tiered-prefill chunk.');
|
|
20
|
+
}
|
|
21
|
+
return runRouteRemap(indices, weights.sourceExpertToSlot, numTokens * topK,
|
|
22
|
+
sourceNumExperts, weights.numExperts, recorder);
|
|
23
|
+
}
|
|
5
24
|
|
|
6
25
|
// Actual GPU selections, not predictions. The next metadata read precedes the
|
|
7
26
|
// current chunk's dispatch, so known-route I/O can overlap that dispatch.
|
|
8
27
|
/** @type {typeof import('./routed-expert-chunks.js').executeRoutedExpertChunks} */
|
|
9
|
-
export async function executeRoutedExpertChunks(options, loader, policy, signal) {
|
|
10
|
-
if (options.
|
|
11
|
-
throw new Error('Routed expert loading requires
|
|
28
|
+
export async function executeRoutedExpertChunks(options, loader, policy, signal, recorder = null) {
|
|
29
|
+
if (options.activationDtype !== 'f32') {
|
|
30
|
+
throw new Error('Routed expert loading requires an F32 route executor.');
|
|
31
|
+
}
|
|
32
|
+
if (options.groupedPrefill && (options.phase !== 'prefill' || options.sharedSlot || recorder)) {
|
|
33
|
+
throw new Error('Grouped routed experts require immediate prefill without shared-slot execution.');
|
|
34
|
+
}
|
|
35
|
+
// Shared slots keep their layer-owned shared bank. Only the routed banks and
|
|
36
|
+
// expert IDs are compacted. Multi-token auxiliary gate/residual slicing is
|
|
37
|
+
// deliberately excluded; this path hands over the complete single token.
|
|
38
|
+
if (options.sharedSlot && (options.phase !== 'decode' || options.numTokens !== 1 || options.sharedSlot.residual)) {
|
|
39
|
+
throw new Error('Paged shared slots require single-token decode without an outer residual.');
|
|
40
|
+
}
|
|
41
|
+
if (recorder && (options.phase !== 'decode' || options.numTokens !== 1)) {
|
|
42
|
+
throw new Error('Recorded routed experts require single-token decode.');
|
|
12
43
|
}
|
|
13
44
|
const { numTokens, topK, hiddenSize, indicesBuffer, weightsBuffer, inputTensor, layerIdx } = options;
|
|
14
45
|
if (typeof layerIdx !== 'number' || !Number.isSafeInteger(layerIdx)) throw new Error('Routed expert loading requires an explicit layer.');
|
|
15
46
|
const device = getDevice();
|
|
16
47
|
if (!device) throw new Error('Routed expert loading requires the owning GPU device.');
|
|
48
|
+
const tiered = policy.storageMode === 'tiered-compact';
|
|
49
|
+
signal?.throwIfAborted();
|
|
50
|
+
const resident = tiered ? loader.getResidentRoutedExpertWeights?.(layerIdx) : null;
|
|
51
|
+
if (resident) {
|
|
52
|
+
if (resident.sourceExpertToSlot) throw new Error('Resident routed experts require a complete direct-index bank.');
|
|
53
|
+
if (isRoutedExpertTraceActive(loader)) for (let start = 0; start < numTokens; start += policy.routeChunkTokens) {
|
|
54
|
+
const count = Math.min(policy.routeChunkTokens, numTokens - start);
|
|
55
|
+
await readRoutedExpertSelection(loader, { layer: layerIdx, phase: options.phase ?? 'unspecified',
|
|
56
|
+
tokenOffset: start, numTokens: count, topK, source: 'resident' },
|
|
57
|
+
() => readBufferSlice(indicesBuffer, start * topK * 4, count * topK * 4, recorder), false);
|
|
58
|
+
signal?.throwIfAborted();
|
|
59
|
+
}
|
|
60
|
+
|
|
61
|
+
return recorder ? recordQwenRouteExperts(recorder, { ...options, weights: resident })
|
|
62
|
+
: runQwenRouteExperts({ ...options, weights: resident });
|
|
63
|
+
}
|
|
64
|
+
const capacity = tiered || (policy.storageMode === 'compact-decode' && options.phase === 'decode')
|
|
65
|
+
? loader.getRoutedExpertCapacity?.(layerIdx, options.phase) : undefined;
|
|
66
|
+
if (tiered && capacity === undefined) throw new Error('Tiered routed experts require an explicit admitted capacity.');
|
|
67
|
+
if (capacity !== undefined && (!Number.isSafeInteger(capacity) || capacity < topK)) {
|
|
68
|
+
throw new Error('Routed expert capacity cannot fit one complete top-K selection.');
|
|
69
|
+
}
|
|
70
|
+
const chunkTokens = capacity === undefined ? policy.routeChunkTokens
|
|
71
|
+
: Math.min(policy.routeChunkTokens, Math.floor(capacity / topK));
|
|
17
72
|
/** @param {number} start */
|
|
18
73
|
const selections = async start => {
|
|
19
74
|
signal?.throwIfAborted();
|
|
20
|
-
const count = Math.min(
|
|
21
|
-
const
|
|
75
|
+
const count = Math.min(chunkTokens, numTokens - start);
|
|
76
|
+
const profile = getRuntimeConfig().shared.debug.profiler.enabled;
|
|
77
|
+
const started = profile ? performance.now() : 0;
|
|
78
|
+
// Append the selection copy to routing's encoder and submit that prefix once.
|
|
79
|
+
// Mapping still settles all earlier bank consumers before cache reuse. The
|
|
80
|
+
// outer owner keeps temporaries until its final submission.
|
|
81
|
+
const data = isRoutedExpertTraceActive(loader) ? await readRoutedExpertSelection(loader, { layer: layerIdx, phase: options.phase ?? 'unspecified',
|
|
82
|
+
tokenOffset: start, numTokens: count, topK, source: 'paging' },
|
|
83
|
+
() => readBufferSlice(indicesBuffer, start * topK * 4, count * topK * 4, recorder), true)
|
|
84
|
+
: await readBufferSlice(indicesBuffer, start * topK * 4, count * topK * 4, recorder);
|
|
85
|
+
if (!data) throw new Error('Routed paging requires selection readback.');
|
|
86
|
+
if (profile) loader.recordRoutedReadback?.(options.phase, performance.now() - started, count * topK * 4);
|
|
22
87
|
signal?.throwIfAborted();
|
|
23
88
|
return [...new Set(new Uint32Array(data))];
|
|
24
89
|
};
|
|
25
90
|
let ids = await selections(0);
|
|
26
|
-
if (numTokens <=
|
|
27
|
-
const weights = await loader.prepareRoutedExperts(layerIdx, ids, signal);
|
|
91
|
+
if (numTokens <= chunkTokens) {
|
|
92
|
+
const weights = await loader.prepareRoutedExperts(layerIdx, ids, signal, options.phase);
|
|
28
93
|
signal?.throwIfAborted();
|
|
29
|
-
|
|
94
|
+
if (!weights.sourceExpertToSlot) return recorder
|
|
95
|
+
? recordQwenRouteExperts(recorder, { ...options, weights })
|
|
96
|
+
: runQwenRouteExperts({ ...options, weights });
|
|
97
|
+
const remapped = await remapCompactIndices(weights, indicesBuffer, numTokens, topK, ids, tiered, options.phase, recorder);
|
|
98
|
+
if (!remapped) throw new Error('Compact expert remap was not produced.');
|
|
99
|
+
if (recorder) {
|
|
100
|
+
recorder.trackTemporaryBuffer(remapped);
|
|
101
|
+
return recordQwenRouteExperts(recorder, { ...options, weights, indicesBuffer: remapped });
|
|
102
|
+
}
|
|
103
|
+
let submitted = false;
|
|
104
|
+
try {
|
|
105
|
+
const result = await runQwenRouteExperts({ ...options, weights, indicesBuffer: remapped });
|
|
106
|
+
submitted = true;
|
|
107
|
+
return result;
|
|
108
|
+
} finally {
|
|
109
|
+
try { if (!submitted) await device.queue.onSubmittedWorkDone(); }
|
|
110
|
+
finally { releaseBuffer(remapped); }
|
|
111
|
+
}
|
|
30
112
|
}
|
|
31
113
|
const output = acquireBuffer(numTokens * hiddenSize * 4, undefined, 'routed_prefill_output');
|
|
32
114
|
let completed = false;
|
|
33
115
|
let failure;
|
|
34
116
|
try {
|
|
35
|
-
for (let start = 0; start < numTokens; start +=
|
|
117
|
+
for (let start = 0; start < numTokens; start += chunkTokens) {
|
|
36
118
|
signal?.throwIfAborted();
|
|
37
|
-
const count = Math.min(
|
|
119
|
+
const count = Math.min(chunkTokens, numTokens - start);
|
|
38
120
|
const next = start + count < numTokens ? selections(start + count) : null;
|
|
39
121
|
// A rejection is observed immediately even if weight preparation fails.
|
|
40
122
|
next?.catch(() => undefined);
|
|
41
|
-
const
|
|
123
|
+
const selectedIds = ids;
|
|
124
|
+
const preparing = loader.prepareRoutedExperts(layerIdx, selectedIds, signal, options.phase);
|
|
42
125
|
preparing.catch(() => undefined);
|
|
43
126
|
/** @type {GPUBuffer | null} */
|
|
44
|
-
let input = null, indices = null, routes = null;
|
|
127
|
+
let input = null, indices = null, routes = null, remapped = null;
|
|
45
128
|
let chunkOutput = null;
|
|
46
129
|
let chunkComplete = false;
|
|
47
130
|
let chunkFailure;
|
|
@@ -57,9 +140,10 @@ export async function executeRoutedExpertChunks(options, loader, policy, signal)
|
|
|
57
140
|
const weights = await preparing;
|
|
58
141
|
ids = next ? await next : [];
|
|
59
142
|
signal?.throwIfAborted();
|
|
143
|
+
remapped = await remapCompactIndices(weights, indices, count, topK, selectedIds, tiered, options.phase, null);
|
|
60
144
|
chunkOutput = await runQwenRouteExperts({ ...options, numTokens: count, weights,
|
|
61
145
|
inputTensor: createTensor(input, 'f32', [count, hiddenSize], 'routed_chunk'),
|
|
62
|
-
indicesBuffer: indices, weightsBuffer: routes });
|
|
146
|
+
indicesBuffer: remapped ?? indices, weightsBuffer: routes });
|
|
63
147
|
const combine = device.createCommandEncoder({ label: 'routed_chunk_output' });
|
|
64
148
|
combine.copyBufferToBuffer(chunkOutput.buffer, 0, output, start * hiddenSize * 4, count * hiddenSize * 4);
|
|
65
149
|
device.queue.submit([combine.finish()]);
|
|
@@ -73,7 +157,7 @@ export async function executeRoutedExpertChunks(options, loader, policy, signal)
|
|
|
73
157
|
if (!chunkComplete) {
|
|
74
158
|
try { await device.queue.onSubmittedWorkDone(); } catch (error) { cleanup.push(error); }
|
|
75
159
|
}
|
|
76
|
-
for (const buffer of [chunkOutput?.buffer, input, indices, routes]) {
|
|
160
|
+
for (const buffer of [chunkOutput?.buffer, remapped, input, indices, routes]) {
|
|
77
161
|
if (buffer) { try { releaseBuffer(buffer); } catch (error) { cleanup.push(error); } }
|
|
78
162
|
}
|
|
79
163
|
if (cleanup.length) throw new AggregateError([chunkFailure, ...cleanup].filter(Boolean), 'Routed chunk cleanup failed.');
|
|
@@ -1,20 +1,21 @@
|
|
|
1
1
|
import { getRuntimeConfig } from '../../../config/runtime.js';
|
|
2
2
|
import { QK_K } from '../../../config/schema/index.js';
|
|
3
|
-
import { releaseBuffer } from '../../../memory/buffer-pool.js';
|
|
4
|
-
|
|
5
|
-
const dequantCache = new Map();
|
|
6
|
-
let dequantCacheMaxEntriesOverride = null;
|
|
7
|
-
let dequantCacheHits = 0;
|
|
8
|
-
let dequantCacheMisses = 0;
|
|
9
3
|
|
|
10
4
|
export function resolveMaxTokensPerExpert(numTokens, numExperts, topK, hiddenSize, activationDtype) {
|
|
11
5
|
const routingConfig = getRuntimeConfig().inference.moe.routing;
|
|
12
6
|
const {
|
|
13
|
-
maxTokensPerExpert
|
|
14
|
-
maxTokensPerExpertHeadroom
|
|
15
|
-
maxTokensPerExpertMin
|
|
16
|
-
maxTokensPerExpertCap
|
|
7
|
+
maxTokensPerExpert,
|
|
8
|
+
maxTokensPerExpertHeadroom,
|
|
9
|
+
maxTokensPerExpertMin,
|
|
10
|
+
maxTokensPerExpertCap,
|
|
17
11
|
} = routingConfig;
|
|
12
|
+
for (const [name, value] of Object.entries({
|
|
13
|
+
maxTokensPerExpert, maxTokensPerExpertHeadroom, maxTokensPerExpertMin, maxTokensPerExpertCap,
|
|
14
|
+
})) {
|
|
15
|
+
if (!Number.isFinite(value) || value < 0) {
|
|
16
|
+
throw new Error(`runtime.inference.moe.routing.${name} must be a finite non-negative number, got ${String(value)}.`);
|
|
17
|
+
}
|
|
18
|
+
}
|
|
18
19
|
|
|
19
20
|
let target = maxTokensPerExpert > 0
|
|
20
21
|
? maxTokensPerExpert
|
|
@@ -39,70 +40,3 @@ export function resolveMaxTokensPerExpert(numTokens, numExperts, topK, hiddenSiz
|
|
|
39
40
|
}
|
|
40
41
|
return Math.min(target, numTokens);
|
|
41
42
|
}
|
|
42
|
-
|
|
43
|
-
function getDequantCacheMaxEntries() {
|
|
44
|
-
return dequantCacheMaxEntriesOverride ?? getRuntimeConfig().inference.moe.cache.dequantCacheMaxEntries;
|
|
45
|
-
}
|
|
46
|
-
|
|
47
|
-
function getDequantCacheKey(layerIdx, expertIdx, outputDtype) {
|
|
48
|
-
return `${layerIdx}_${expertIdx}_${outputDtype}`;
|
|
49
|
-
}
|
|
50
|
-
|
|
51
|
-
export function getCachedDequant(layerIdx, expertIdx, outputDtype) {
|
|
52
|
-
const key = getDequantCacheKey(layerIdx, expertIdx, outputDtype);
|
|
53
|
-
const cached = dequantCache.get(key);
|
|
54
|
-
if (cached) {
|
|
55
|
-
cached.lastUsed = performance.now();
|
|
56
|
-
dequantCacheHits++;
|
|
57
|
-
}
|
|
58
|
-
return cached;
|
|
59
|
-
}
|
|
60
|
-
|
|
61
|
-
export function setCachedDequant(layerIdx, expertIdx, outputDtype, gateUp, down) {
|
|
62
|
-
const key = getDequantCacheKey(layerIdx, expertIdx, outputDtype);
|
|
63
|
-
dequantCacheMisses++;
|
|
64
|
-
|
|
65
|
-
if (dequantCache.size >= getDequantCacheMaxEntries()) {
|
|
66
|
-
let oldestKey = '';
|
|
67
|
-
let oldestTime = Infinity;
|
|
68
|
-
for (const [k, v] of dequantCache.entries()) {
|
|
69
|
-
if (v.lastUsed < oldestTime) {
|
|
70
|
-
oldestTime = v.lastUsed;
|
|
71
|
-
oldestKey = k;
|
|
72
|
-
}
|
|
73
|
-
}
|
|
74
|
-
if (oldestKey) {
|
|
75
|
-
const evicted = dequantCache.get(oldestKey);
|
|
76
|
-
if (evicted) {
|
|
77
|
-
releaseBuffer(evicted.gateUp);
|
|
78
|
-
releaseBuffer(evicted.down);
|
|
79
|
-
}
|
|
80
|
-
dequantCache.delete(oldestKey);
|
|
81
|
-
}
|
|
82
|
-
}
|
|
83
|
-
|
|
84
|
-
dequantCache.set(key, { gateUp, down, lastUsed: performance.now() });
|
|
85
|
-
}
|
|
86
|
-
|
|
87
|
-
export function clearDequantCache() {
|
|
88
|
-
for (const cached of dequantCache.values()) {
|
|
89
|
-
releaseBuffer(cached.gateUp);
|
|
90
|
-
releaseBuffer(cached.down);
|
|
91
|
-
}
|
|
92
|
-
dequantCache.clear();
|
|
93
|
-
dequantCacheHits = 0;
|
|
94
|
-
dequantCacheMisses = 0;
|
|
95
|
-
}
|
|
96
|
-
|
|
97
|
-
export function getDequantCacheStats() {
|
|
98
|
-
return {
|
|
99
|
-
hits: dequantCacheHits,
|
|
100
|
-
misses: dequantCacheMisses,
|
|
101
|
-
size: dequantCache.size,
|
|
102
|
-
maxEntries: getDequantCacheMaxEntries(),
|
|
103
|
-
};
|
|
104
|
-
}
|
|
105
|
-
|
|
106
|
-
export function setDequantCacheMaxEntries(maxEntries) {
|
|
107
|
-
dequantCacheMaxEntriesOverride = maxEntries;
|
|
108
|
-
}
|
|
@@ -8,7 +8,7 @@ import { runSoftmaxTopK, recordSoftmaxTopK } from '../../../gpu/kernels/softmax.
|
|
|
8
8
|
import { runMoEGather, runMoEBuildTokenOffsets, runScatterAddDynamic } from '../../../gpu/kernels/moe.js';
|
|
9
9
|
import { trace, isTraceEnabled } from '../../../debug/index.js';
|
|
10
10
|
import { f16ToF32Array } from '../../kv-cache/types.js';
|
|
11
|
-
import { resolveMaxTokensPerExpert
|
|
11
|
+
import { resolveMaxTokensPerExpert } from './moe-cache.js';
|
|
12
12
|
import { ensureExpertLoaded } from './moe/expert-loading.js';
|
|
13
13
|
import { selectRuleValue } from '../../../rules/rule-registry.js';
|
|
14
14
|
import { getRuntimeConfig } from '../../../config/runtime.js';
|
|
@@ -33,7 +33,7 @@ import {
|
|
|
33
33
|
} from './moe/plan.js';
|
|
34
34
|
import { runQwenRouteExperts, recordQwenRouteExperts } from './moe/qwen-route-executor.js';
|
|
35
35
|
import { executeRoutedExpertChunks } from './moe/routed-expert-chunks.js';
|
|
36
|
-
import { releaseOrTrack } from
|
|
36
|
+
import { releaseOrTrack } from "./execution/ops.js";
|
|
37
37
|
|
|
38
38
|
export { buildActiveExpertScheduleFromIndices } from './moe/plan.js';
|
|
39
39
|
|
|
@@ -75,10 +75,10 @@ const MOE_RECORDED_ROUTE_EXECUTORS = Object.freeze({
|
|
|
75
75
|
});
|
|
76
76
|
|
|
77
77
|
const MOE_EXPERT_EXECUTORS = Object.freeze({
|
|
78
|
-
'gpt-oss':
|
|
79
|
-
'gemma4-packed':
|
|
80
|
-
'qwen35-packed':
|
|
81
|
-
mixtral:
|
|
78
|
+
'gpt-oss': runGptOssExpert,
|
|
79
|
+
'gemma4-packed': runGemma4Expert,
|
|
80
|
+
'qwen35-packed': runGemma4Expert,
|
|
81
|
+
mixtral: runMixtralExpert,
|
|
82
82
|
});
|
|
83
83
|
|
|
84
84
|
function requireMoeExecutor(registry, id, label) {
|
|
@@ -101,64 +101,6 @@ function assertMoeExpertWeights(moeProfile, weights, expertKey) {
|
|
|
101
101
|
}
|
|
102
102
|
}
|
|
103
103
|
|
|
104
|
-
async function runGptOssProfileExpert(args) {
|
|
105
|
-
return runGptOssExpert(
|
|
106
|
-
args.gathered,
|
|
107
|
-
args.expertOutputs,
|
|
108
|
-
args.weights,
|
|
109
|
-
args.layerIdx,
|
|
110
|
-
args.expertIdx,
|
|
111
|
-
args.count,
|
|
112
|
-
args.inputOffset,
|
|
113
|
-
args.outputOffset,
|
|
114
|
-
args.hiddenSize,
|
|
115
|
-
args.intermediateSize,
|
|
116
|
-
args.numExperts,
|
|
117
|
-
args.activationDtype,
|
|
118
|
-
args.swigluLimit,
|
|
119
|
-
args.kernelPath,
|
|
120
|
-
args.executionPolicies,
|
|
121
|
-
args.modelType,
|
|
122
|
-
args.vendorProfile,
|
|
123
|
-
args.moeKernelPathProfile
|
|
124
|
-
);
|
|
125
|
-
}
|
|
126
|
-
|
|
127
|
-
async function runGemma4ProfileExpert(args) {
|
|
128
|
-
return runGemma4Expert(
|
|
129
|
-
args.gathered,
|
|
130
|
-
args.expertOutputs,
|
|
131
|
-
args.weights,
|
|
132
|
-
args.count,
|
|
133
|
-
args.inputOffset,
|
|
134
|
-
args.outputOffset,
|
|
135
|
-
args.hiddenSize,
|
|
136
|
-
args.intermediateSize,
|
|
137
|
-
args.activationDtype,
|
|
138
|
-
args.swigluLimit,
|
|
139
|
-
args.kernelPath,
|
|
140
|
-
args.hiddenActivation,
|
|
141
|
-
args.moeProfile.matmulRoles
|
|
142
|
-
);
|
|
143
|
-
}
|
|
144
|
-
|
|
145
|
-
async function runMixtralProfileExpert(args) {
|
|
146
|
-
return runMixtralExpert(
|
|
147
|
-
args.gathered,
|
|
148
|
-
args.expertOutputs,
|
|
149
|
-
args.weights,
|
|
150
|
-
args.count,
|
|
151
|
-
args.inputOffset,
|
|
152
|
-
args.outputOffset,
|
|
153
|
-
args.hiddenSize,
|
|
154
|
-
args.intermediateSize,
|
|
155
|
-
args.hiddenActivation,
|
|
156
|
-
args.activationDtype,
|
|
157
|
-
args.swigluLimit,
|
|
158
|
-
args.kernelPath
|
|
159
|
-
);
|
|
160
|
-
}
|
|
161
|
-
|
|
162
104
|
export async function moeFeedForwardGPU(
|
|
163
105
|
inputBuffer,
|
|
164
106
|
numTokens,
|
|
@@ -194,7 +136,18 @@ export async function moeFeedForwardGPU(
|
|
|
194
136
|
const moeProfile = resolveMoeExecutionProfile(config, { modelType });
|
|
195
137
|
const activeExpertSelection = config.activeExpertSelection === undefined
|
|
196
138
|
? resolveConfiguredMoEActiveExpertSelection() : resolveMoEActiveExpertSelection(config.activeExpertSelection);
|
|
197
|
-
|
|
139
|
+
const segmentedRoute = requestedRecorder && expertLoader.getRoutedExpertPolicy?.()?.enabled;
|
|
140
|
+
if (segmentedRoute) {
|
|
141
|
+
if (config.phase !== 'decode' || numTokens !== 1 || config.activationDtype !== 'f32'
|
|
142
|
+
|| config.groupedPrefill || config.sharedSlot?.residual || config.normRouter || activeExpertSelection !== 'topk-route'
|
|
143
|
+
|| moeProfile.topkRouteExecutor !== 'qwen35-route') {
|
|
144
|
+
throw new Error('Recorded paged experts require single-token F32 Qwen decode without outer-residual or norm fusion.');
|
|
145
|
+
}
|
|
146
|
+
// Routed loading's preparation hook is admission-only. The supported recorded
|
|
147
|
+
// path flushes routing and prepares actual banks at its dependency boundary.
|
|
148
|
+
} else {
|
|
149
|
+
await expertLoader.prepareExpertLayer?.(layerIdx, expertWeights, requestedRecorder, config.signal, activeExpertSelection);
|
|
150
|
+
}
|
|
198
151
|
const recordRouteExecutor = activeExpertSelection === 'topk-route'
|
|
199
152
|
? MOE_RECORDED_ROUTE_EXECUTORS[moeProfile.topkRouteExecutor] : null;
|
|
200
153
|
const recorder = recordRouteExecutor ? requestedRecorder : null;
|
|
@@ -252,14 +205,18 @@ export async function moeFeedForwardGPU(
|
|
|
252
205
|
let activeExpertSchedule = null;
|
|
253
206
|
|
|
254
207
|
const layerRouter = layerRouterWeights?.get(layerIdx) || null;
|
|
255
|
-
if (layerRouter) {
|
|
256
|
-
|
|
257
|
-
|
|
258
|
-
|
|
259
|
-
layerRouter.scale || null,
|
|
260
|
-
layerRouter.perExpertScale || null
|
|
208
|
+
if (!layerRouter) {
|
|
209
|
+
throw new Error(
|
|
210
|
+
`[MoE] Layer ${layerIdx} is an MoE layer but has no router weight; ` +
|
|
211
|
+
'refusing to reuse another layer\'s router. Re-convert the model with per-layer router tensors.'
|
|
261
212
|
);
|
|
262
213
|
}
|
|
214
|
+
moeRouter.loadWeights(
|
|
215
|
+
layerRouter.weight,
|
|
216
|
+
layerRouter.bias || null,
|
|
217
|
+
layerRouter.scale || null,
|
|
218
|
+
layerRouter.perExpertScale || null
|
|
219
|
+
);
|
|
263
220
|
|
|
264
221
|
// Immediate-only adapters must execute after their recorded producers.
|
|
265
222
|
// Keep the parent recorder open for following residuals and layers.
|
|
@@ -491,7 +448,7 @@ export async function moeFeedForwardGPU(
|
|
|
491
448
|
hiddenActivation,
|
|
492
449
|
};
|
|
493
450
|
outputTensor = routedPolicy?.enabled
|
|
494
|
-
? await executeRoutedExpertChunks(routeOptions, expertLoader, routedPolicy, config.signal)
|
|
451
|
+
? await executeRoutedExpertChunks(routeOptions, expertLoader, routedPolicy, config.signal, recorder)
|
|
495
452
|
: recorder
|
|
496
453
|
? await recordRouteExecutor(recorder, routeOptions)
|
|
497
454
|
: await routeExecutor(routeOptions);
|
|
@@ -623,6 +580,7 @@ export async function moeFeedForwardGPU(
|
|
|
623
580
|
vendorProfile,
|
|
624
581
|
moeKernelPathProfile,
|
|
625
582
|
hiddenActivation,
|
|
583
|
+
dequantCache: expertLoader.expertDequantCache ?? null,
|
|
626
584
|
});
|
|
627
585
|
perfLog(`MoE L${layerIdx} expert_exec`, stepStart, { expertIdx, count });
|
|
628
586
|
}
|
|
@@ -665,7 +623,7 @@ export async function moeFeedForwardGPU(
|
|
|
665
623
|
weightsBuffer = null;
|
|
666
624
|
|
|
667
625
|
if (perfEnabled) {
|
|
668
|
-
const cacheStats =
|
|
626
|
+
const cacheStats = expertLoader.expertDequantCache?.getStats() ?? null;
|
|
669
627
|
trace.perf(`MoE L${layerIdx} done`, {
|
|
670
628
|
numTokens,
|
|
671
629
|
topK,
|
|
@@ -674,8 +632,8 @@ export async function moeFeedForwardGPU(
|
|
|
674
632
|
activeExpertSelection: activeExpertSchedule?.selection ?? null,
|
|
675
633
|
rowsPerExpert,
|
|
676
634
|
maxTokensPerExpert,
|
|
677
|
-
dequantCacheHits: cacheStats
|
|
678
|
-
dequantCacheMisses: cacheStats
|
|
635
|
+
dequantCacheHits: cacheStats?.hits ?? null,
|
|
636
|
+
dequantCacheMisses: cacheStats?.misses ?? null,
|
|
679
637
|
expertCache: typeof expertLoader?.getExpertCacheStats === 'function'
|
|
680
638
|
? expertLoader.getExpertCacheStats()
|
|
681
639
|
: null,
|
|
@@ -19,29 +19,9 @@ import type { Tensor, TensorDtype } from '../../../gpu/tensor.js';
|
|
|
19
19
|
import type { MoERouter } from '../../moe-router.js';
|
|
20
20
|
import type { ExpertWeights } from './types.js';
|
|
21
21
|
import type { KernelPathSchema } from '../../../config/schema/index.js';
|
|
22
|
-
import type { ExecutionV1PoliciesSchema } from
|
|
22
|
+
import type { ExecutionV1PoliciesSchema } from "../../../config/schema/execution/execution-v1.schema.js";
|
|
23
23
|
import type { WeightBuffer } from '../../../gpu/weight-buffer.js';
|
|
24
24
|
|
|
25
|
-
/**
|
|
26
|
-
* Clear the dequantization cache (call on model unload).
|
|
27
|
-
*/
|
|
28
|
-
export function clearDequantCache(): void;
|
|
29
|
-
|
|
30
|
-
/**
|
|
31
|
-
* Get cache stats for debugging.
|
|
32
|
-
*/
|
|
33
|
-
export function getDequantCacheStats(): {
|
|
34
|
-
hits: number;
|
|
35
|
-
misses: number;
|
|
36
|
-
size: number;
|
|
37
|
-
maxEntries: number;
|
|
38
|
-
};
|
|
39
|
-
|
|
40
|
-
/**
|
|
41
|
-
* Configure dequant cache max entries at runtime.
|
|
42
|
-
*/
|
|
43
|
-
export function setDequantCacheMaxEntries(maxEntries: number): void;
|
|
44
|
-
|
|
45
25
|
/**
|
|
46
26
|
* Configuration for MoE feed-forward.
|
|
47
27
|
*/
|
|
@@ -113,21 +93,10 @@ export interface ExpertLoader {
|
|
|
113
93
|
recorder: import('../../../gpu/command-recorder.js').CommandRecorder | null,
|
|
114
94
|
signal?: AbortSignal): Promise<unknown>;
|
|
115
95
|
loadExpert(layerIdx: number, expertIdx: number): Promise<MoEExpertWeights | null>;
|
|
96
|
+
/** False once the loader's LRU cache evicted (and released) `weights`. */
|
|
97
|
+
isExpertResident(layerIdx: number, expertIdx: number, weights: unknown): boolean;
|
|
116
98
|
}
|
|
117
99
|
|
|
118
|
-
/**
|
|
119
|
-
* MoE feed-forward with CPU routing.
|
|
120
|
-
*/
|
|
121
|
-
export function moeFeedForwardCPU(
|
|
122
|
-
hiddenStates: Float32Array,
|
|
123
|
-
numTokens: number,
|
|
124
|
-
config: MoEConfig,
|
|
125
|
-
moeRouter: MoERouter,
|
|
126
|
-
expertWeights: Map<string, MoEExpertWeights>,
|
|
127
|
-
expertLoader: ExpertLoader,
|
|
128
|
-
layerIdx: number
|
|
129
|
-
): Promise<Float32Array>;
|
|
130
|
-
|
|
131
100
|
/**
|
|
132
101
|
* MoE feed-forward fully on GPU.
|
|
133
102
|
*/
|
|
@@ -143,7 +112,3 @@ export function moeFeedForwardGPU(
|
|
|
143
112
|
recorder?: import('../../../gpu/command-recorder.js').CommandRecorder | null
|
|
144
113
|
): Promise<GPUBuffer>;
|
|
145
114
|
|
|
146
|
-
/**
|
|
147
|
-
* Check if layer is MoE layer (some models have dense layers too).
|
|
148
|
-
*/
|
|
149
|
-
export function isMoELayer(layerIdx: number): boolean;
|
|
@@ -1,9 +1,3 @@
|
|
|
1
|
-
import { clearDequantCache, getDequantCacheStats, setDequantCacheMaxEntries } from './moe-cache.js';
|
|
2
|
-
import { moeFeedForwardCPU } from './moe-cpu.js';
|
|
3
1
|
import { moeFeedForwardGPU } from './moe-gpu.js';
|
|
4
2
|
|
|
5
|
-
export {
|
|
6
|
-
|
|
7
|
-
export function isMoELayer(_layerIdx) {
|
|
8
|
-
return true;
|
|
9
|
-
}
|
|
3
|
+
export { moeFeedForwardGPU };
|
|
@@ -1,7 +1,7 @@
|
|
|
1
1
|
import type {
|
|
2
2
|
ResolvedNgramEmbedding,
|
|
3
3
|
ResolvedNgramEmbeddingLayer,
|
|
4
|
-
} from
|
|
4
|
+
} from "../../../../config/schema/execution/ngram-embedding.schema.js";
|
|
5
5
|
|
|
6
6
|
export interface NgramHistoryCheckpoint {
|
|
7
7
|
readonly epoch: number;
|
|
@@ -25,8 +25,8 @@ export class NgramTokenHistory {
|
|
|
25
25
|
#epoch = 0;
|
|
26
26
|
|
|
27
27
|
/**
|
|
28
|
-
* @param {import(
|
|
29
|
-
* @param {import(
|
|
28
|
+
* @param {import("../../../../config/schema/execution/ngram-embedding.schema.js").ResolvedNgramEmbedding} contract
|
|
29
|
+
* @param {import("../../../../config/schema/execution/ngram-embedding.schema.js").ResolvedNgramEmbeddingLayer} layer
|
|
30
30
|
* @param {number} [initialCapacity]
|
|
31
31
|
*/
|
|
32
32
|
constructor(contract, layer, initialCapacity = 256) {
|
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import type { ResolvedNgramEmbeddingSession } from
|
|
1
|
+
import type { ResolvedNgramEmbeddingSession } from "../../../../config/schema/execution/ngram-embedding.schema.js";
|
|
2
2
|
|
|
3
3
|
export type NgramTablePartSource = string | number;
|
|
4
4
|
|
|
@@ -22,7 +22,7 @@ export type NgramTableRangePort = (
|
|
|
22
22
|
export interface NgramTableReaderOptions {
|
|
23
23
|
rows: number;
|
|
24
24
|
rowWidth: number;
|
|
25
|
-
dtype: 'bf16';
|
|
25
|
+
dtype: 'bf16' | 'q4_0';
|
|
26
26
|
/** Must tile [0, rows) in order. */
|
|
27
27
|
parts: NgramTablePart[];
|
|
28
28
|
readRange: NgramTableRangePort;
|
|
@@ -48,7 +48,7 @@ export declare class NgramTableReader {
|
|
|
48
48
|
constructor(options: NgramTableReaderOptions);
|
|
49
49
|
readonly rows: number;
|
|
50
50
|
readonly rowWidth: number;
|
|
51
|
-
readonly dtype: 'bf16';
|
|
51
|
+
readonly dtype: 'bf16' | 'q4_0';
|
|
52
52
|
readonly rowBytes: number;
|
|
53
53
|
readonly parts: readonly Readonly<Omit<NgramTablePart, 'byteLength'>>[];
|
|
54
54
|
readonly rowBytesVerified: boolean;
|
|
@@ -1,3 +1,5 @@
|
|
|
1
|
+
import { ngramTableRowBytes } from "../../../../config/schema/execution/ngram-embedding.schema.js";
|
|
2
|
+
|
|
1
3
|
// Host-resident reader of one qwen4_exp n-gram embedding table.
|
|
2
4
|
//
|
|
3
5
|
// The table is one logical [rows, rowWidth] matrix stored as row-contiguous
|
|
@@ -10,10 +12,9 @@
|
|
|
10
12
|
// maxBytes; one read batch stages at most staging.maxRows rows; at most
|
|
11
13
|
// reads.maxConcurrent range reads are in flight. Cancellation aborts pending
|
|
12
14
|
// reads and inserts nothing from an incomplete batch. Row bytes are returned
|
|
13
|
-
// as stored (bf16); this reader authenticates the row-to-part mapping, not the
|
|
15
|
+
// as stored (bf16 or block-packed q4_0); this reader authenticates the row-to-part mapping, not the
|
|
14
16
|
// bytes of individual rows (see `rowBytesVerified` in stats).
|
|
15
17
|
|
|
16
|
-
const ROW_ELEMENT_BYTES = Object.freeze({ bf16: 2 });
|
|
17
18
|
|
|
18
19
|
/** @param {string} message @returns {never} */
|
|
19
20
|
function fail(message) {
|
|
@@ -33,8 +34,6 @@ export class NgramTableReader {
|
|
|
33
34
|
|
|
34
35
|
/** @param {import('./ngram-table.js').NgramTableReaderOptions} options */
|
|
35
36
|
constructor({ rows, rowWidth, dtype, parts, readRange, session, rowBytesVerified = false }) {
|
|
36
|
-
const elementBytes = ROW_ELEMENT_BYTES[/** @type {'bf16'} */ (dtype)];
|
|
37
|
-
if (!elementBytes) fail(`unsupported table dtype ${JSON.stringify(dtype)}.`);
|
|
38
37
|
if (!Number.isSafeInteger(rows) || rows < 1 || rows > 2 ** 32) fail('rows must be a positive count within u32 addressing.');
|
|
39
38
|
if (!Number.isSafeInteger(rowWidth) || rowWidth < 1) fail('rowWidth must be positive.');
|
|
40
39
|
if (typeof readRange !== 'function') fail('a readRange(part, offset, length, signal) port is required.');
|
|
@@ -42,7 +41,7 @@ export class NgramTableReader {
|
|
|
42
41
|
this.rows = rows;
|
|
43
42
|
this.rowWidth = rowWidth;
|
|
44
43
|
this.dtype = dtype;
|
|
45
|
-
this.rowBytes = rowWidth
|
|
44
|
+
this.rowBytes = ngramTableRowBytes(rowWidth, dtype);
|
|
46
45
|
this.readRange = readRange;
|
|
47
46
|
this.session = session;
|
|
48
47
|
this.rowBytesVerified = rowBytesVerified === true;
|
|
@@ -214,7 +213,7 @@ export class NgramTableReader {
|
|
|
214
213
|
export function ngramTablePartsFromCheckpointHeaders(receipt) {
|
|
215
214
|
const table = receipt?.ngramTable;
|
|
216
215
|
if (!table || !Array.isArray(table.shards)) fail('checkpoint header receipt lacks ngramTable.shards.');
|
|
217
|
-
const rowBytes = table.rowWidth
|
|
216
|
+
const rowBytes = ngramTableRowBytes(table.rowWidth, String(table.dtype).toLowerCase());
|
|
218
217
|
return table.shards.map((shard) => {
|
|
219
218
|
if (shard.fileByteEnd - shard.fileByteBegin !== (shard.rowEnd - shard.rowBegin) * rowBytes) {
|
|
220
219
|
fail(`shard ${shard.shard} byte span does not match its rows.`);
|
|
@@ -1,7 +1,7 @@
|
|
|
1
1
|
import type { CommandRecorder } from '../../../../gpu/command-recorder.js';
|
|
2
2
|
import type { Tensor } from '../../../../gpu/tensor.js';
|
|
3
|
-
import type { ResolvedNgramEmbedding } from
|
|
4
|
-
import type {
|
|
3
|
+
import type { ResolvedNgramEmbedding } from "../../../../config/schema/execution/ngram-embedding.schema.js";
|
|
4
|
+
import type { RuntimeHyperConnections } from '../hyper-connection/decoder-layer.js';
|
|
5
5
|
|
|
6
6
|
export declare class PleConvState {
|
|
7
7
|
constructor(options: { channels: number; stateLength: number });
|
|
@@ -42,7 +42,7 @@ export interface PleLayerInput {
|
|
|
42
42
|
weights: PleLayerWeights;
|
|
43
43
|
state: PleConvState;
|
|
44
44
|
ngram: ResolvedNgramEmbedding;
|
|
45
|
-
hyper:
|
|
45
|
+
hyper: RuntimeHyperConnections;
|
|
46
46
|
numTokens: number;
|
|
47
47
|
project: PleProjection;
|
|
48
48
|
}
|
|
@@ -75,15 +75,15 @@ export async function runPleLayer(target, { streams, rows, weights, state, ngram
|
|
|
75
75
|
};
|
|
76
76
|
let retainedGate = null;
|
|
77
77
|
let returned = false;
|
|
78
|
-
const normOptions = { numTokens, streamCount: hyper.streamCount, streamWidth: hyper.streamWidth, eps: ngram.normEps };
|
|
78
|
+
const normOptions = { numTokens, streamCount: hyper.streamCount, streamWidth: hyper.streamWidth, eps: ngram.normEps, normWeightOffset: hyper.normWeightOffset };
|
|
79
79
|
/** @param {import('../../../../gpu/tensor.js').Tensor} input @param {import('../../../../gpu/tensor.js').Tensor} weight */
|
|
80
80
|
const norm = (input, weight) => (recorded
|
|
81
81
|
? recordHyperConnectionNorm(target, input, weight, normOptions)
|
|
82
82
|
: runHyperConnectionNorm(input, weight, normOptions));
|
|
83
83
|
try {
|
|
84
84
|
const embedding = keep(recorded
|
|
85
|
-
? await recordPleRowDecode(target, rows, { numTokens, embedDim: ngram.embedDim })
|
|
86
|
-
: await runPleRowDecode(rows, { numTokens, embedDim: ngram.embedDim }));
|
|
85
|
+
? await recordPleRowDecode(target, rows, { numTokens, embedDim: ngram.embedDim, tableDtype: ngram.tableDtype })
|
|
86
|
+
: await runPleRowDecode(rows, { numTokens, embedDim: ngram.embedDim, tableDtype: ngram.tableDtype }));
|
|
87
87
|
const key = keep(await project('ple_key_proj', embedding, weights.keyProj, channels, ngram.embedDim));
|
|
88
88
|
const keyNormed = keep(await norm(key, weights.normKey));
|
|
89
89
|
const value = keep(await project('ple_value_proj', embedding, weights.valueProj, hyper.streamWidth, ngram.embedDim));
|