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
|
@@ -0,0 +1,197 @@
|
|
|
1
|
+
import { shouldMarkWideTileResidualFused, resolveDenseFFNFusedPathDtypes, resolveFusedGateUpPipelineConstants, resolveFusedGateUpVariant } from "./dense-plan.js";
|
|
2
|
+
import { getWeightBuffer } from "../loading/weights.js";
|
|
3
|
+
import { coerceTensorDtype, dispatchFusedGateUp, resolveMatmulStepDtype } from "./dense-executor.js";
|
|
4
|
+
import { borrowDecodeFfnIntermediate } from "./decode-intermediate.js";
|
|
5
|
+
import { runProbes } from "../observation/probes.js";
|
|
6
|
+
import { captureTrainingActivation, captureProjectionInput, enqueueRecordedDenseHealth } from "./activation-observation.js";
|
|
7
|
+
import { isGpuBufferInstance, isWeightBuffer } from "../../../../gpu/weight-buffer.js";
|
|
8
|
+
import { doMatmul, releaseOrTrack } from "../execution/ops.js";
|
|
9
|
+
import { getLoRAModule } from "../lora.js";
|
|
10
|
+
import { getRuntimeConfig } from "../../../../config/runtime.js";
|
|
11
|
+
import { applyLoRA } from "../lora-apply.js";
|
|
12
|
+
import { releaseBuffer } from "../../../../memory/buffer-pool.js";
|
|
13
|
+
|
|
14
|
+
export async function runFusedGateUpFFNGPU({
|
|
15
|
+
phase,
|
|
16
|
+
layerIdx,
|
|
17
|
+
kernelPath,
|
|
18
|
+
ffnStepPrecision,
|
|
19
|
+
inputTensor,
|
|
20
|
+
defaultMatmulOutputDtype,
|
|
21
|
+
fusedGateUpWeights,
|
|
22
|
+
layerWeights,
|
|
23
|
+
context,
|
|
24
|
+
recorder,
|
|
25
|
+
gateDtype,
|
|
26
|
+
hiddenSize,
|
|
27
|
+
intermediateSize,
|
|
28
|
+
numTokens,
|
|
29
|
+
hiddenActivation,
|
|
30
|
+
swigluLimit,
|
|
31
|
+
lora
|
|
32
|
+
}) {
|
|
33
|
+
const {
|
|
34
|
+
fusedGateUpInputDtype,
|
|
35
|
+
downInputDtype,
|
|
36
|
+
} = resolveDenseFFNFusedPathDtypes({
|
|
37
|
+
phase,
|
|
38
|
+
layerIdx,
|
|
39
|
+
kernelPath,
|
|
40
|
+
ffnStepPrecision,
|
|
41
|
+
fallbackInputDtype: inputTensor.dtype,
|
|
42
|
+
fallbackOutputDtype: defaultMatmulOutputDtype,
|
|
43
|
+
});
|
|
44
|
+
const gateWeight = getWeightBuffer(fusedGateUpWeights.gate ?? layerWeights.gate, 'ffn_gate');
|
|
45
|
+
const upWeight = getWeightBuffer(fusedGateUpWeights.up ?? layerWeights.up, 'ffn_up');
|
|
46
|
+
const downWeight = getWeightBuffer(layerWeights.down, 'ffn_down');
|
|
47
|
+
const fusedRuntimeSession = context.resolvedRuntimeSession?.runtime?.session;
|
|
48
|
+
const fusedGateUpPipelineConstants = resolveFusedGateUpPipelineConstants({
|
|
49
|
+
runtimeSession: fusedRuntimeSession,
|
|
50
|
+
kernelPath,
|
|
51
|
+
phase,
|
|
52
|
+
layerIdx,
|
|
53
|
+
});
|
|
54
|
+
const fusedGateUpVariant = resolveFusedGateUpVariant({ phase, runtimeSession: fusedRuntimeSession });
|
|
55
|
+
const fusedDownOutputDtype = resolveMatmulStepDtype(
|
|
56
|
+
'ffn_down',
|
|
57
|
+
phase,
|
|
58
|
+
layerIdx,
|
|
59
|
+
kernelPath,
|
|
60
|
+
'f32',
|
|
61
|
+
'outputDtype',
|
|
62
|
+
ffnStepPrecision
|
|
63
|
+
);
|
|
64
|
+
let fusedInput = inputTensor;
|
|
65
|
+
let fusedInputOwned = false;
|
|
66
|
+
if (fusedGateUpInputDtype && fusedGateUpInputDtype !== inputTensor.dtype) {
|
|
67
|
+
fusedInput = await coerceTensorDtype(inputTensor, fusedGateUpInputDtype, recorder, {
|
|
68
|
+
executionPolicies: context.executionPolicies ?? null,
|
|
69
|
+
op: 'ffn_gate_up_input',
|
|
70
|
+
transitionDeclaredBy: 'step_precision',
|
|
71
|
+
});
|
|
72
|
+
fusedInputOwned = fusedInput !== inputTensor;
|
|
73
|
+
}
|
|
74
|
+
const fusedOutput = await dispatchFusedGateUp({
|
|
75
|
+
inputTensor: fusedInput, gateWeight, upWeight, gateDtype,
|
|
76
|
+
hiddenSize, intermediateSize, numTokens,
|
|
77
|
+
hiddenActivation, swigluLimit, recorder, layerIdx,
|
|
78
|
+
executionPolicies: context.executionPolicies ?? null,
|
|
79
|
+
normStats: context.__pendingFfnInputNormStats ?? null,
|
|
80
|
+
pipelineConstants: fusedGateUpPipelineConstants,
|
|
81
|
+
variant: fusedGateUpVariant,
|
|
82
|
+
outputBuffer: borrowDecodeFfnIntermediate(context, numTokens, intermediateSize, fusedInput.dtype),
|
|
83
|
+
});
|
|
84
|
+
await runProbes('ffn_act', fusedOutput.buffer, {
|
|
85
|
+
layerIdx,
|
|
86
|
+
numTokens,
|
|
87
|
+
hiddenSize: intermediateSize,
|
|
88
|
+
probes: context.debugProbes,
|
|
89
|
+
recorder,
|
|
90
|
+
operatorDiagnostics: context.operatorDiagnostics,
|
|
91
|
+
dtype: fusedOutput.dtype,
|
|
92
|
+
});
|
|
93
|
+
enqueueRecordedDenseHealth(context, layerIdx, 'ffn_fused_gate_up', fusedOutput, numTokens * intermediateSize);
|
|
94
|
+
await captureTrainingActivation(context, layerIdx, fusedOutput, numTokens, intermediateSize);
|
|
95
|
+
|
|
96
|
+
let downInput = fusedOutput;
|
|
97
|
+
if (downInputDtype && fusedOutput.dtype !== downInputDtype) {
|
|
98
|
+
downInput = await coerceTensorDtype(fusedOutput, downInputDtype, recorder, {
|
|
99
|
+
executionPolicies: context.executionPolicies ?? null,
|
|
100
|
+
op: 'ffn_down_input',
|
|
101
|
+
transitionDeclaredBy: 'step_precision',
|
|
102
|
+
});
|
|
103
|
+
if (recorder) {
|
|
104
|
+
recorder.trackTemporaryBuffer(downInput.buffer);
|
|
105
|
+
}
|
|
106
|
+
}
|
|
107
|
+
|
|
108
|
+
if (!isGpuBufferInstance(layerWeights.gate) && !isWeightBuffer(layerWeights.gate)) {
|
|
109
|
+
releaseOrTrack(recorder, isWeightBuffer(gateWeight) ? gateWeight.buffer : gateWeight);
|
|
110
|
+
}
|
|
111
|
+
if (!isGpuBufferInstance(layerWeights.up) && !isWeightBuffer(layerWeights.up)) {
|
|
112
|
+
releaseOrTrack(recorder, isWeightBuffer(upWeight) ? upWeight.buffer : upWeight);
|
|
113
|
+
}
|
|
114
|
+
|
|
115
|
+
// Opt-in WideTile+residual fusion (fused-gate-up path).
|
|
116
|
+
const pendingResidualFused = context.__pendingFfnResidualTensor;
|
|
117
|
+
const downLoraProbeFused = getLoRAModule(lora, layerIdx, 'down_proj');
|
|
118
|
+
const mergedSessionFused = getRuntimeConfig().inference?.session;
|
|
119
|
+
const tryFuseDownResidualFused = pendingResidualFused != null
|
|
120
|
+
&& !downLoraProbeFused
|
|
121
|
+
&& downInput.dtype === 'f32'
|
|
122
|
+
&& fusedDownOutputDtype === 'f32'
|
|
123
|
+
&& pendingResidualFused.dtype === 'f32';
|
|
124
|
+
await captureProjectionInput(context, layerIdx, downInput, numTokens, intermediateSize);
|
|
125
|
+
let output = await doMatmul(
|
|
126
|
+
downInput,
|
|
127
|
+
downWeight,
|
|
128
|
+
numTokens,
|
|
129
|
+
hiddenSize,
|
|
130
|
+
intermediateSize,
|
|
131
|
+
{
|
|
132
|
+
transposeB: 'auto',
|
|
133
|
+
label: `L${layerIdx}.ffn_down`,
|
|
134
|
+
layerIdx,
|
|
135
|
+
kernelPath,
|
|
136
|
+
outputDtype: fusedDownOutputDtype,
|
|
137
|
+
role: 'ffn_down',
|
|
138
|
+
executionPolicies: context.executionPolicies ?? null,
|
|
139
|
+
residualTensor: tryFuseDownResidualFused ? pendingResidualFused : null,
|
|
140
|
+
},
|
|
141
|
+
recorder
|
|
142
|
+
);
|
|
143
|
+
enqueueRecordedDenseHealth(context, layerIdx, 'ffn_down', output, numTokens * hiddenSize);
|
|
144
|
+
{
|
|
145
|
+
if (tryFuseDownResidualFused
|
|
146
|
+
&& shouldMarkWideTileResidualFused(mergedSessionFused, phase)
|
|
147
|
+
) {
|
|
148
|
+
context.__ffnResidualFusedFired = true;
|
|
149
|
+
}
|
|
150
|
+
}
|
|
151
|
+
|
|
152
|
+
const loraDown = getLoRAModule(lora, layerIdx, 'down_proj');
|
|
153
|
+
if (loraDown) {
|
|
154
|
+
const combined = await applyLoRA(
|
|
155
|
+
downInput,
|
|
156
|
+
output,
|
|
157
|
+
loraDown,
|
|
158
|
+
{ M: numTokens, N: hiddenSize, K: intermediateSize },
|
|
159
|
+
getWeightBuffer,
|
|
160
|
+
recorder,
|
|
161
|
+
{ kernelPath }
|
|
162
|
+
);
|
|
163
|
+
if (combined.buffer !== output.buffer) {
|
|
164
|
+
if (recorder) {
|
|
165
|
+
recorder.trackTemporaryBuffer(output.buffer);
|
|
166
|
+
} else {
|
|
167
|
+
releaseBuffer(output.buffer);
|
|
168
|
+
}
|
|
169
|
+
output = combined;
|
|
170
|
+
}
|
|
171
|
+
}
|
|
172
|
+
|
|
173
|
+
if (!isGpuBufferInstance(layerWeights.down) && !isWeightBuffer(layerWeights.down)) {
|
|
174
|
+
releaseOrTrack(recorder, isWeightBuffer(downWeight) ? downWeight.buffer : downWeight);
|
|
175
|
+
}
|
|
176
|
+
|
|
177
|
+
if (recorder) {
|
|
178
|
+
if (downInput !== fusedOutput) {
|
|
179
|
+
recorder.trackTemporaryBuffer(downInput.buffer);
|
|
180
|
+
}
|
|
181
|
+
if (fusedInputOwned) {
|
|
182
|
+
recorder.trackTemporaryBuffer(fusedInput.buffer);
|
|
183
|
+
}
|
|
184
|
+
releaseOrTrack(recorder, fusedOutput.buffer, context.decodeBuffers);
|
|
185
|
+
} else {
|
|
186
|
+
if (downInput !== fusedOutput) {
|
|
187
|
+
releaseBuffer(downInput.buffer);
|
|
188
|
+
}
|
|
189
|
+
if (fusedInputOwned) {
|
|
190
|
+
releaseBuffer(fusedInput.buffer);
|
|
191
|
+
}
|
|
192
|
+
releaseOrTrack(null, fusedOutput.buffer, context.decodeBuffers);
|
|
193
|
+
}
|
|
194
|
+
|
|
195
|
+
return output;
|
|
196
|
+
|
|
197
|
+
}
|
|
@@ -1,7 +1,7 @@
|
|
|
1
1
|
import { resolveGroupedPrefill } from '../moe/grouped-prefill.js';
|
|
2
2
|
|
|
3
3
|
|
|
4
|
-
import { releaseOrTrack } from
|
|
4
|
+
import { releaseOrTrack } from "../execution/ops.js";
|
|
5
5
|
import { runQwenSharedExpert, resolveDeclaredTopK, resolveQwenNormRouter, resolveQwenSharedSlot, resolveQwenSharedSlotVariants } from './qwen-shared-expert.js';
|
|
6
6
|
import { createTensor } from '../../../../gpu/tensor.js';
|
|
7
7
|
|
|
@@ -0,0 +1,218 @@
|
|
|
1
|
+
import { getWeightBuffer } from "../loading/weights.js";
|
|
2
|
+
import { selectRuleValue } from "../../../../rules/rule-registry.js";
|
|
3
|
+
import { resolveActivationOp, resolveMatmulStepDtype } from "./dense-executor.js";
|
|
4
|
+
import { doMatmul, doSiLURowSplit, releaseOrTrack } from "../execution/ops.js";
|
|
5
|
+
import { getLoRAModule } from "../lora.js";
|
|
6
|
+
import { applyLoRA } from "../lora-apply.js";
|
|
7
|
+
import { releaseBuffer } from "../../../../memory/buffer-pool.js";
|
|
8
|
+
import { isKernelDebugEnabled, dumpTokenVector } from "../debug-utils/tensor.js";
|
|
9
|
+
import { captureTrainingActivation, captureProjectionInput, enqueueRecordedDenseHealth } from "./activation-observation.js";
|
|
10
|
+
import { isGpuBufferInstance, isWeightBuffer } from "../../../../gpu/weight-buffer.js";
|
|
11
|
+
import { getRuntimeConfig } from "../../../../config/runtime.js";
|
|
12
|
+
import { shouldMarkWideTileResidualFused } from "./dense-plan.js";
|
|
13
|
+
|
|
14
|
+
export async function runPackedGateUpFFNGPU({
|
|
15
|
+
layerWeights,
|
|
16
|
+
inputTensor,
|
|
17
|
+
phase,
|
|
18
|
+
layerIdx,
|
|
19
|
+
kernelPath,
|
|
20
|
+
ffnStepPrecision,
|
|
21
|
+
numTokens,
|
|
22
|
+
intermediateSize,
|
|
23
|
+
hiddenSize,
|
|
24
|
+
context,
|
|
25
|
+
recorder,
|
|
26
|
+
lora,
|
|
27
|
+
lastTokenIdx,
|
|
28
|
+
hiddenActivation,
|
|
29
|
+
swigluLimit
|
|
30
|
+
}) {
|
|
31
|
+
const gateUpWeight = getWeightBuffer(layerWeights.gateUp, 'ffn_gate_up');
|
|
32
|
+
const downWeight = getWeightBuffer(layerWeights.down, 'ffn_down');
|
|
33
|
+
|
|
34
|
+
const useF16 = inputTensor.dtype === 'f16';
|
|
35
|
+
const defaultMatmulOutputDtype = selectRuleValue('shared', 'dtype', 'f16OrFallbackByFlag', {
|
|
36
|
+
useF16,
|
|
37
|
+
fallback: inputTensor.dtype,
|
|
38
|
+
});
|
|
39
|
+
const matmulOutputDtype = resolveMatmulStepDtype(
|
|
40
|
+
'ffn_gate_up',
|
|
41
|
+
phase,
|
|
42
|
+
layerIdx,
|
|
43
|
+
kernelPath,
|
|
44
|
+
defaultMatmulOutputDtype,
|
|
45
|
+
'outputDtype',
|
|
46
|
+
ffnStepPrecision
|
|
47
|
+
);
|
|
48
|
+
const downOutputDtype = resolveMatmulStepDtype(
|
|
49
|
+
'ffn_down',
|
|
50
|
+
phase,
|
|
51
|
+
layerIdx,
|
|
52
|
+
kernelPath,
|
|
53
|
+
'f32',
|
|
54
|
+
'outputDtype',
|
|
55
|
+
ffnStepPrecision
|
|
56
|
+
);
|
|
57
|
+
let gateUpOutput = await doMatmul(
|
|
58
|
+
inputTensor, gateUpWeight,
|
|
59
|
+
numTokens, intermediateSize * 2, hiddenSize,
|
|
60
|
+
{
|
|
61
|
+
transposeB: 'auto',
|
|
62
|
+
label: `L${layerIdx}.ffn_gate_up`,
|
|
63
|
+
layerIdx,
|
|
64
|
+
kernelPath,
|
|
65
|
+
outputDtype: matmulOutputDtype,
|
|
66
|
+
role: 'ffn_gate_up',
|
|
67
|
+
executionPolicies: context.executionPolicies ?? null,
|
|
68
|
+
},
|
|
69
|
+
recorder
|
|
70
|
+
);
|
|
71
|
+
|
|
72
|
+
const loraGateUp = getLoRAModule(lora, layerIdx, 'gate_up_proj');
|
|
73
|
+
if (loraGateUp) {
|
|
74
|
+
const combined = await applyLoRA(
|
|
75
|
+
inputTensor,
|
|
76
|
+
gateUpOutput,
|
|
77
|
+
loraGateUp,
|
|
78
|
+
{ M: numTokens, N: intermediateSize * 2, K: hiddenSize },
|
|
79
|
+
getWeightBuffer,
|
|
80
|
+
recorder,
|
|
81
|
+
{ kernelPath }
|
|
82
|
+
);
|
|
83
|
+
if (combined.buffer !== gateUpOutput.buffer) {
|
|
84
|
+
if (recorder) {
|
|
85
|
+
recorder.trackTemporaryBuffer(gateUpOutput.buffer);
|
|
86
|
+
} else {
|
|
87
|
+
releaseBuffer(gateUpOutput.buffer);
|
|
88
|
+
}
|
|
89
|
+
gateUpOutput = combined;
|
|
90
|
+
}
|
|
91
|
+
}
|
|
92
|
+
|
|
93
|
+
if (isKernelDebugEnabled(layerIdx) && !recorder) {
|
|
94
|
+
await dumpTokenVector(gateUpOutput.buffer, 'ffn_gate_up', {
|
|
95
|
+
layerIdx,
|
|
96
|
+
tokenIdx: lastTokenIdx,
|
|
97
|
+
rowSize: intermediateSize * 2,
|
|
98
|
+
dtype: gateUpOutput.dtype,
|
|
99
|
+
});
|
|
100
|
+
}
|
|
101
|
+
enqueueRecordedDenseHealth(context, layerIdx, 'ffn_gate_up', gateUpOutput, numTokens * intermediateSize * 2);
|
|
102
|
+
|
|
103
|
+
if (!isGpuBufferInstance(layerWeights.gateUp) && !isWeightBuffer(layerWeights.gateUp)) {
|
|
104
|
+
releaseOrTrack(recorder, isWeightBuffer(gateUpWeight) ? gateUpWeight.buffer : gateUpWeight);
|
|
105
|
+
}
|
|
106
|
+
|
|
107
|
+
const activatedOutput = await doSiLURowSplit(gateUpOutput, {
|
|
108
|
+
numTokens,
|
|
109
|
+
dim: intermediateSize,
|
|
110
|
+
activation: resolveActivationOp(hiddenActivation),
|
|
111
|
+
swigluLimit,
|
|
112
|
+
label: `L${layerIdx}.ffn_activation`,
|
|
113
|
+
layerIdx,
|
|
114
|
+
}, recorder);
|
|
115
|
+
|
|
116
|
+
if (isKernelDebugEnabled(layerIdx) && !recorder) {
|
|
117
|
+
await dumpTokenVector(activatedOutput.buffer, 'ffn_activated', {
|
|
118
|
+
layerIdx,
|
|
119
|
+
tokenIdx: lastTokenIdx,
|
|
120
|
+
rowSize: intermediateSize,
|
|
121
|
+
dtype: activatedOutput.dtype,
|
|
122
|
+
});
|
|
123
|
+
}
|
|
124
|
+
enqueueRecordedDenseHealth(context, layerIdx, 'ffn_act', activatedOutput, numTokens * intermediateSize);
|
|
125
|
+
await captureTrainingActivation(context, layerIdx, activatedOutput, numTokens, intermediateSize);
|
|
126
|
+
|
|
127
|
+
if (recorder) {
|
|
128
|
+
recorder.trackTemporaryBuffer(gateUpOutput.buffer);
|
|
129
|
+
} else {
|
|
130
|
+
releaseBuffer(gateUpOutput.buffer);
|
|
131
|
+
}
|
|
132
|
+
|
|
133
|
+
// Opt-in WideTile+residual fusion: if the caller (processFFNStandard)
|
|
134
|
+
// staged a residual tensor on the context AND no LoRA is present on
|
|
135
|
+
// down_proj (LoRA requires the pre-residual output for its add), route
|
|
136
|
+
// this matmul to the q4_fused_widetile_residual variant which produces
|
|
137
|
+
// (ffn_down_result + residual) in one dispatch. Tell the caller via a
|
|
138
|
+
// context flag so processFFNStandard skips its downstream doResidualAdd.
|
|
139
|
+
const pendingResidual = context.__pendingFfnResidualTensor;
|
|
140
|
+
const downLoraProbe = getLoRAModule(lora, layerIdx, 'down_proj');
|
|
141
|
+
const mergedSession = getRuntimeConfig().inference?.session;
|
|
142
|
+
const tryFuseDownResidual = pendingResidual != null
|
|
143
|
+
&& !downLoraProbe
|
|
144
|
+
&& activatedOutput.dtype === 'f32'
|
|
145
|
+
&& downOutputDtype === 'f32'
|
|
146
|
+
&& pendingResidual.dtype === 'f32';
|
|
147
|
+
await captureProjectionInput(context, layerIdx, activatedOutput, numTokens, intermediateSize);
|
|
148
|
+
let output = await doMatmul(
|
|
149
|
+
activatedOutput, downWeight,
|
|
150
|
+
numTokens, hiddenSize, intermediateSize,
|
|
151
|
+
{
|
|
152
|
+
transposeB: 'auto',
|
|
153
|
+
label: `L${layerIdx}.ffn_down`,
|
|
154
|
+
layerIdx,
|
|
155
|
+
kernelPath,
|
|
156
|
+
outputDtype: downOutputDtype,
|
|
157
|
+
role: 'ffn_down',
|
|
158
|
+
executionPolicies: context.executionPolicies ?? null,
|
|
159
|
+
residualTensor: tryFuseDownResidual ? pendingResidual : null,
|
|
160
|
+
},
|
|
161
|
+
recorder
|
|
162
|
+
);
|
|
163
|
+
// Detect whether the fusion fired: if the selected variant for ffn_down
|
|
164
|
+
// matmul was q4_fused_widetile_residual, then output IS post-residual.
|
|
165
|
+
// We infer this cheaply by re-checking the conditions the selector uses.
|
|
166
|
+
// (A cleaner signal would require a return-shape change across all
|
|
167
|
+
// dense.js paths; this local signal is enough for correctness.)
|
|
168
|
+
{
|
|
169
|
+
if (tryFuseDownResidual
|
|
170
|
+
&& shouldMarkWideTileResidualFused(mergedSession, phase)
|
|
171
|
+
) {
|
|
172
|
+
context.__ffnResidualFusedFired = true;
|
|
173
|
+
}
|
|
174
|
+
}
|
|
175
|
+
|
|
176
|
+
const loraDown = getLoRAModule(lora, layerIdx, 'down_proj');
|
|
177
|
+
if (loraDown) {
|
|
178
|
+
const combined = await applyLoRA(
|
|
179
|
+
activatedOutput,
|
|
180
|
+
output,
|
|
181
|
+
loraDown,
|
|
182
|
+
{ M: numTokens, N: hiddenSize, K: intermediateSize },
|
|
183
|
+
getWeightBuffer,
|
|
184
|
+
recorder,
|
|
185
|
+
{ kernelPath }
|
|
186
|
+
);
|
|
187
|
+
if (combined.buffer !== output.buffer) {
|
|
188
|
+
if (recorder) {
|
|
189
|
+
recorder.trackTemporaryBuffer(output.buffer);
|
|
190
|
+
} else {
|
|
191
|
+
releaseBuffer(output.buffer);
|
|
192
|
+
}
|
|
193
|
+
output = combined;
|
|
194
|
+
}
|
|
195
|
+
}
|
|
196
|
+
|
|
197
|
+
if (isKernelDebugEnabled(layerIdx) && !recorder) {
|
|
198
|
+
await dumpTokenVector(output.buffer, 'ffn_down_out', {
|
|
199
|
+
layerIdx,
|
|
200
|
+
tokenIdx: lastTokenIdx,
|
|
201
|
+
rowSize: hiddenSize,
|
|
202
|
+
dtype: output.dtype,
|
|
203
|
+
});
|
|
204
|
+
}
|
|
205
|
+
enqueueRecordedDenseHealth(context, layerIdx, 'ffn_down', output, numTokens * hiddenSize);
|
|
206
|
+
|
|
207
|
+
if (!isGpuBufferInstance(layerWeights.down) && !isWeightBuffer(layerWeights.down)) {
|
|
208
|
+
releaseOrTrack(recorder, isWeightBuffer(downWeight) ? downWeight.buffer : downWeight);
|
|
209
|
+
}
|
|
210
|
+
if (recorder) {
|
|
211
|
+
recorder.trackTemporaryBuffer(activatedOutput.buffer);
|
|
212
|
+
} else {
|
|
213
|
+
releaseBuffer(activatedOutput.buffer);
|
|
214
|
+
}
|
|
215
|
+
|
|
216
|
+
return output;
|
|
217
|
+
|
|
218
|
+
}
|
|
@@ -1,11 +1,11 @@
|
|
|
1
|
-
import { doMatmul, doSiLU, releaseOrTrack } from
|
|
2
|
-
import { getWeightBuffer } from
|
|
1
|
+
import { doMatmul, doSiLU, releaseOrTrack } from "../execution/ops.js";
|
|
2
|
+
import { getWeightBuffer } from "../loading/weights.js";
|
|
3
3
|
import { resolveLayerIntermediateSize } from '../config.js';
|
|
4
4
|
import { getLayerSteps, getKernelPathMatmulVariant } from '../../../../config/kernel-path-loader.js';
|
|
5
5
|
import { getKernelConfigs } from '../../../../config/kernel-registry-contract.js';
|
|
6
6
|
import { isGpuBufferInstance, isWeightBuffer } from '../../../../gpu/weight-buffer.js';
|
|
7
7
|
import { tensorBytes } from '../../../../gpu/tensor.js';
|
|
8
|
-
import { runProbes } from
|
|
8
|
+
import { runProbes } from "../observation/probes.js";
|
|
9
9
|
|
|
10
10
|
/** @type {typeof import('./paired.js').validatePairedFFN} */
|
|
11
11
|
export function validatePairedFFN(layerIdx, input, numTokens, context, weights) {
|
|
@@ -5,7 +5,7 @@ import { runMatmul, recordMatmul } from '../../../../gpu/kernels/matmul.js';
|
|
|
5
5
|
import { runSiLU, recordSiLU } from '../../../../gpu/kernels/silu.js';
|
|
6
6
|
import { runRepeatChannels, recordRepeatChannels } from '../../../../gpu/kernels/repeat_channels.js';
|
|
7
7
|
import { runResidualAdd, recordResidualAdd } from '../../../../gpu/kernels/residual.js';
|
|
8
|
-
import { releaseOrTrack } from
|
|
8
|
+
import { releaseOrTrack } from "../execution/ops.js";
|
|
9
9
|
import { getKernelConfigs } from '../../../../config/kernel-registry-contract.js';
|
|
10
10
|
import { getLayerSteps, getKernelPathStepConfig } from '../../../../config/kernel-path-loader.js';
|
|
11
11
|
import { runSigmoidGatedAdd, recordSigmoidGatedAdd } from '../../../../gpu/kernels/sigmoid-gated-add.js';
|
|
@@ -1,15 +1,15 @@
|
|
|
1
1
|
|
|
2
2
|
|
|
3
|
-
import { doRMSNorm, doResidualAdd, doSandwichRMSNormPair, releaseOrTrack } from
|
|
3
|
+
import { doRMSNorm, doResidualAdd, doSandwichRMSNormPair, releaseOrTrack } from "../execution/ops.js";
|
|
4
4
|
import { getLayout, getWeightDtype, isCpuWeightBuffer, isGpuBufferInstance, isWeightBuffer } from '../../../../gpu/weight-buffer.js';
|
|
5
5
|
import { trace } from '../../../../debug/index.js';
|
|
6
6
|
import { isKernelDebugEnabled, dumpTokenVector } from '../debug-utils/tensor.js';
|
|
7
7
|
import { logFFN } from '../debug-utils/logging.js';
|
|
8
8
|
import { getBufferStats } from '../debug-utils/health-metrics.js';
|
|
9
9
|
import { shouldDebugLayerOutput } from '../debug-utils/config.js';
|
|
10
|
-
import { getNormWeightBuffer } from
|
|
11
|
-
import { runProbes } from
|
|
12
|
-
import {
|
|
10
|
+
import { getNormWeightBuffer } from "../loading/weights.js";
|
|
11
|
+
import { runProbes } from "../observation/probes.js";
|
|
12
|
+
import { isMoELayer, hasLoggedFusedDownNorm, setLoggedFusedDownNorm } from './types.js';
|
|
13
13
|
import { runDenseFFNGPU, runDenseFFNWithFusedPostNormGPU } from './dense.js';
|
|
14
14
|
import { runMoEFFNGPU } from './moe.js';
|
|
15
15
|
import { resolveLayerIntermediateSize } from '../config.js';
|
|
@@ -27,14 +27,14 @@ async function debugFFNBuffer(context, layerIdx, label, tensor, numTokens, hidde
|
|
|
27
27
|
function shouldUseDensePlusMoeFFN(layerIdx, config, layerWeights) {
|
|
28
28
|
return config.ffnBranchMode === 'dense_plus_moe'
|
|
29
29
|
&& config.useMoE
|
|
30
|
-
&&
|
|
30
|
+
&& isMoELayer(layerIdx, config);
|
|
31
31
|
}
|
|
32
32
|
|
|
33
33
|
function shouldUseLegacyMoeOnlyFFN(layerIdx, config, layerWeights) {
|
|
34
34
|
return config.ffnBranchMode !== 'dense'
|
|
35
35
|
&& config.ffnBranchMode !== 'dense_plus_moe'
|
|
36
36
|
&& config.useMoE
|
|
37
|
-
&&
|
|
37
|
+
&& isMoELayer(layerIdx, config);
|
|
38
38
|
}
|
|
39
39
|
|
|
40
40
|
function requireNormWeight(layerIdx, layerWeights, key) {
|
|
@@ -233,7 +233,6 @@ export async function processFFNWithSandwichNorm(
|
|
|
233
233
|
|
|
234
234
|
let canUseFusedDownNorm = numTokens === 1
|
|
235
235
|
&& !config.useMoE
|
|
236
|
-
&& !isMoELayerLocal(layerIdx, config, layerWeights)
|
|
237
236
|
&& sandwichNorm.hasPostFeedforwardNorm
|
|
238
237
|
&& layerWeights?.postFeedforwardNorm
|
|
239
238
|
&& layerWeights?.down
|
|
@@ -8,10 +8,10 @@ import {
|
|
|
8
8
|
doResidualAdd,
|
|
9
9
|
doResidualNextRMSNormPair,
|
|
10
10
|
releaseOrTrack,
|
|
11
|
-
} from
|
|
12
|
-
import { getNormWeightBuffer, getVectorTensor } from
|
|
13
|
-
import { runProbes } from
|
|
14
|
-
import {
|
|
11
|
+
} from "../execution/ops.js";
|
|
12
|
+
import { getNormWeightBuffer, getVectorTensor } from "../loading/weights.js";
|
|
13
|
+
import { runProbes } from "../observation/probes.js";
|
|
14
|
+
import { isMoELayer } from './types.js';
|
|
15
15
|
import { runDenseFFNGPU } from './dense.js';
|
|
16
16
|
import { runMoEFFNGPU } from './moe.js';
|
|
17
17
|
import { resolveQwenNormRouter } from './qwen-shared-expert.js';
|
|
@@ -95,7 +95,7 @@ export async function processFFNStandard(
|
|
|
95
95
|
const useLegacyMoeOnlyFFN = config.ffnBranchMode !== 'dense'
|
|
96
96
|
&& config.ffnBranchMode !== 'dense_plus_moe'
|
|
97
97
|
&& config.useMoE
|
|
98
|
-
&&
|
|
98
|
+
&& isMoELayer(layerIdx, config);
|
|
99
99
|
// A declared moe_norm_router step moves this norm into the MoE's router dispatch.
|
|
100
100
|
const normRouterDeclared = useLegacyMoeOnlyFFN
|
|
101
101
|
&& resolveQwenNormRouter({ kernelPath: context.kernelPath ?? null, phase: context.phase, layerIdx });
|
|
@@ -1,11 +1,14 @@
|
|
|
1
1
|
|
|
2
2
|
|
|
3
3
|
|
|
4
|
-
|
|
4
|
+
// The single MoE-layer test. layerTypes names the attention mixer of each layer
|
|
5
|
+
// (full_attention, linear_attention, ...); it names the FFN kind only when the
|
|
6
|
+
// manifest declares explicit 'moe' entries. Otherwise every layer of a useMoE model
|
|
7
|
+
// is an MoE layer, and a missing router is a load error, not a dense fallback.
|
|
8
|
+
export function isMoELayer(layerIdx, config) {
|
|
5
9
|
if (!config.useMoE) return false;
|
|
6
|
-
if (layerWeights?.routerWeight) return true;
|
|
7
10
|
const layerTypes = config.layerTypes;
|
|
8
|
-
if (Array.isArray(layerTypes) &&
|
|
11
|
+
if (Array.isArray(layerTypes) && layerTypes.includes('moe')) {
|
|
9
12
|
return layerTypes[layerIdx] === 'moe';
|
|
10
13
|
}
|
|
11
14
|
return true;
|
|
@@ -1,8 +1,8 @@
|
|
|
1
|
-
import { isWeightBuffer, isSplitWeightBuffer, isGpuBufferInstance, getWeightDtype } from
|
|
2
|
-
import { readBufferSlice } from
|
|
3
|
-
import { dequantizeQ4KMRowWise } from
|
|
4
|
-
import { QK_K, QK4_K_BLOCK_SIZE } from
|
|
5
|
-
import { decodeSnapshotData } from
|
|
1
|
+
import { isWeightBuffer, isSplitWeightBuffer, isGpuBufferInstance, getWeightDtype } from "../../../../../gpu/weight-buffer.js";
|
|
2
|
+
import { readBufferSlice } from "../../../../../memory/buffer-pool.js";
|
|
3
|
+
import { dequantizeQ4KMRowWise } from "../../../../../formats/q4k-dequant.js";
|
|
4
|
+
import { QK_K, QK4_K_BLOCK_SIZE } from "../../../../../config/schema/manifest/quantization.schema.js";
|
|
5
|
+
import { decodeSnapshotData } from "../../../../../debug/tensor.js";
|
|
6
6
|
|
|
7
7
|
/** @type {typeof import('./activity-weights.js').captureActivityWeights} */
|
|
8
8
|
export async function captureActivityWeights(weights, limits, signal, tiedReference = false) {
|
|
@@ -10,7 +10,7 @@ export async function captureActivityWeights(weights, limits, signal, tiedRefere
|
|
|
10
10
|
if (![weightBlockValues, weightReadValues, weightRowGroupSize, weightRawRows, weightMaxBytes].every(n => Number.isSafeInteger(n) && n > 0)) {
|
|
11
11
|
throw new Error('Weight observation needs explicit tile and memory budgets.');
|
|
12
12
|
}
|
|
13
|
-
/** @type {import(
|
|
13
|
+
/** @type {import("../../../../../debug/activity-trace.js").ActivityWeight[]} */
|
|
14
14
|
const result = [];
|
|
15
15
|
const seen = new Set();
|
|
16
16
|
let retained = 0, haveContribution = false;
|
|
@@ -30,7 +30,7 @@ export async function captureActivityWeights(weights, limits, signal, tiedRefere
|
|
|
30
30
|
const dtype = getWeightDtype(weight) ?? 'unknown';
|
|
31
31
|
const byteLength = isWeightBuffer(weight) ? weight.buffer.size : isGpuBufferInstance(weight) ? weight.size
|
|
32
32
|
: weight.sections.reduce((n, section) => n + section.buffer.size, 0);
|
|
33
|
-
/** @type {import(
|
|
33
|
+
/** @type {import("../../../../../debug/activity-trace.js").ActivityWeight} */
|
|
34
34
|
const entry = { id, layer, role, shape, dtype, layout: shaped ? weight.layout : 'unknown',
|
|
35
35
|
byteLength, resident: true, status: 'unavailable', reason: null,
|
|
36
36
|
blockSize: weightBlockValues, rms: [], raw: [], rawOffset: 0, columnEnergy: [], rowGroupSize: weightRowGroupSize };
|
|
@@ -103,7 +103,7 @@ export async function prepareActivityRecording(state, opts) {
|
|
|
103
103
|
activity.configure({ numLayers: config.numLayers, hiddenSize: config.hiddenSize,
|
|
104
104
|
intermediateSize: config.intermediateSize, numHeads: config.numHeads, numKVHeads: config.numKVHeads,
|
|
105
105
|
headDim: config.headDim, layerTypes: config.layerTypes, vocabSize: config.vocabSize, useTiedEmbeddings: config.useTiedEmbeddings });
|
|
106
|
-
activity.setWeights(await captureActivityWeights(state.weights, /** @type {import(
|
|
106
|
+
activity.setWeights(await captureActivityWeights(state.weights, /** @type {import("../../../../../debug/activity-trace.js").ActivityLimits} */ (opts.__activationTrace.snapshot('recording').capturePolicy), opts.signal, config.useTiedEmbeddings === true));
|
|
107
107
|
activity.unavailable('quantization-error', 'No pinned source-reference tensors were supplied with this recording.');
|
|
108
108
|
activity.unavailable('expert-routing', 'No expert-routing boundary was captured for this execution.');
|
|
109
109
|
}
|
|
@@ -0,0 +1,61 @@
|
|
|
1
|
+
import type { Tensor } from "../../../../../gpu/tensor.js";
|
|
2
|
+
import type { CommandRecorder } from "../../../../../gpu/command-recorder.js";
|
|
3
|
+
import type { LayerContext } from "../../types.js";
|
|
4
|
+
export interface ChannelGainSpec {
|
|
5
|
+
/** Rotation selects split gate/up + post-SwiGLU rotation on selected dense layers only. */
|
|
6
|
+
transformKind?: 'gain' | 'rotation';
|
|
7
|
+
rotationEpsilon?: number;
|
|
8
|
+
/** Opt-in capture: retain ordinary final-layer FFN inputs for guarded same-sequence replay. */
|
|
9
|
+
retainFinalLayerReplay?: boolean;
|
|
10
|
+
mode: 'capture' | 'apply'; requestId: string; gainHash: string;
|
|
11
|
+
cacheArm: 'base-cache' | 're-prefill'; tokenCount: number;
|
|
12
|
+
span: { start: number; stop: number; templatePositions: number[] };
|
|
13
|
+
layers: { layerIdx: number; channels: number }[];
|
|
14
|
+
/** Historical field name; rotation uses the opaque handle from freezeChannelRotations. */
|
|
15
|
+
frozenGains?: FrozenChannelGains;
|
|
16
|
+
/** Capture-only: original full-prompt content span. Enables same-pass complement reference. */
|
|
17
|
+
captureContentSpan?: { start: number; stop: number };
|
|
18
|
+
}
|
|
19
|
+
export interface ChannelGainRequest {
|
|
20
|
+
readonly transformKind: 'gain' | 'rotation';
|
|
21
|
+
readonly requestId: string; readonly mode: 'capture' | 'apply'; readonly gainHash: string;
|
|
22
|
+
readonly cacheArm: 'base-cache' | 're-prefill';
|
|
23
|
+
has(layerIdx: number): boolean;
|
|
24
|
+
validateLayer(layerIdx: number, context: LayerContext, tokens: number, channels: number): void;
|
|
25
|
+
activate(layerIdx: number, up: Tensor, gate: Tensor, tokens: number, channels: number,
|
|
26
|
+
swigluLimit: number | null, recorder: CommandRecorder | null): Promise<Tensor>;
|
|
27
|
+
transform(layerIdx: number, tensor: Tensor, tokens: number, channels: number,
|
|
28
|
+
recorder: CommandRecorder | null): Promise<Tensor>;
|
|
29
|
+
captureFinalLayer(layerIdx: number, postAttn: Tensor, residual: Tensor | null, tokens: number,
|
|
30
|
+
context: LayerContext, weights: import("../../types.js").LayerWeights | undefined, supported: boolean): void;
|
|
31
|
+
replayFinalLayer(request: ChannelGainRequest): Promise<Tensor>;
|
|
32
|
+
takeRotations(): Map<number, import("../../../../../gpu/kernels/channel-rotation.js").ChannelRotation>;
|
|
33
|
+
capture(layerIdx: number, tensor: Tensor, tokens: number, channels: number,
|
|
34
|
+
recorder: CommandRecorder | null): Promise<void>;
|
|
35
|
+
/** After submitted prefill completion, transfers energy-buffer ownership to caller. */
|
|
36
|
+
takeEnergies(): Map<number, Tensor>;
|
|
37
|
+
/** Transfers optional same-pass reference buffers independently of instruction buffers. */
|
|
38
|
+
takeReferenceEnergies(): Map<number, Tensor>;
|
|
39
|
+
/** Caller must finish/discard recorders first. Frozen gain buffers are borrowed, never released. */
|
|
40
|
+
close(): void;
|
|
41
|
+
}
|
|
42
|
+
export interface FrozenChannelGains {
|
|
43
|
+
readonly gainHash: string;
|
|
44
|
+
/** Sum of private GPUBuffer.size at creation, including pool rounding; remains historical after close. */
|
|
45
|
+
readonly gpuBytes: number;
|
|
46
|
+
/** Sum of selected channel counts times four f32 bytes, excluding allocation padding. */
|
|
47
|
+
readonly logicalBytes: number;
|
|
48
|
+
readonly layers: readonly { readonly layerIdx: number; readonly channels: number }[];
|
|
49
|
+
close(): void;
|
|
50
|
+
}
|
|
51
|
+
/** Copies gains GPU-to-GPU into private buffers, validates positive finite values, and hashes copied bytes.
|
|
52
|
+
* Call only after source-producing recorders settle. Readback is never modified or uploaded. */
|
|
53
|
+
export function freezeChannelGains(layers: { layerIdx: number; channels: number; gains: Tensor }[]): Promise<FrozenChannelGains>;
|
|
54
|
+
/** Internal research seam; no source weights change. Caller owns frozen gains and their verified hash. */
|
|
55
|
+
export function createChannelGainRequest(spec: ChannelGainSpec): ChannelGainRequest;
|
|
56
|
+
|
|
57
|
+
export function assertChannelGainIdle(state: { isGenerating?: boolean }): void;
|
|
58
|
+
export function withChannelGainOperation<T>(state: object, operation: () => Promise<T>): Promise<T>;
|
|
59
|
+
|
|
60
|
+
/** Copies/validates plane+status and seals angle in opaque immutable request-local ownership. */
|
|
61
|
+
export function freezeChannelRotations(layers: { layerIdx: number; channels: number; theta: number; rotation: import("../../../../../gpu/kernels/channel-rotation.js").ChannelRotation }[]): Promise<FrozenChannelGains>;
|