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 +1,327 @@
|
|
|
1
|
+
import { createWeightBuffer } from '../../gpu/weight-buffer.js';
|
|
2
|
+
import { isDeviceLost, observeDeviceLoss } from '../../memory/device-residency.js';
|
|
3
|
+
import { CompactExpertCache } from './compact-expert-cache.js';
|
|
4
|
+
import { MappedExpertUploadCache } from './mapped-expert-upload-cache.js';
|
|
5
|
+
import { planTieredExpertPlacement } from './tiered-expert-plan.js';
|
|
6
|
+
import { assertExpertReadShape, createRoutedExpertCounters, readRoutedExpertLease, readRoutedExpertPayload } from './routed-expert-source.js';
|
|
7
|
+
|
|
8
|
+
// Placement is explicit and model-bound. Setup owns resident banks; streamed
|
|
9
|
+
// layers keep compact slots in both phases, never a second full prefill bank.
|
|
10
|
+
export class TieredExpertCache {
|
|
11
|
+
/** @type {import('./routed-expert-cache.js').RoutedExpertCacheOptions} */
|
|
12
|
+
#options;
|
|
13
|
+
/** @type {import('./tiered-expert-plan.js').TieredExpertPlacement} */
|
|
14
|
+
#placement;
|
|
15
|
+
#stats = createRoutedExpertCounters();
|
|
16
|
+
/** @type {CompactExpertCache | null} */
|
|
17
|
+
#compact = null;
|
|
18
|
+
/** @type {MappedExpertUploadCache | null} */
|
|
19
|
+
#mapped = null;
|
|
20
|
+
/** @type {Map<number, import('./routed-expert-cache.js').RoutedExpertWeights>} */
|
|
21
|
+
#resident = new Map();
|
|
22
|
+
/** @type {Set<GPUBuffer>} */
|
|
23
|
+
#owned = new Set();
|
|
24
|
+
#shutdown = new AbortController();
|
|
25
|
+
#closed = false;
|
|
26
|
+
#ready = false;
|
|
27
|
+
#uploadingBytes = 0;
|
|
28
|
+
/** @type {'prefill' | 'decode' | null} */
|
|
29
|
+
#phase = null;
|
|
30
|
+
/** @type {Promise<unknown> | null} */
|
|
31
|
+
#active = null;
|
|
32
|
+
|
|
33
|
+
/** @param {import('./routed-expert-cache.js').RoutedExpertCacheOptions} options
|
|
34
|
+
* @param {readonly import('./routed-expert-plan.js').RoutedExpertBankPlan[]} plans @param {number} minimumSlots */
|
|
35
|
+
constructor(options, plans, minimumSlots) {
|
|
36
|
+
this.#options = options;
|
|
37
|
+
observeDeviceLoss(options.device);
|
|
38
|
+
this.#placement = planTieredExpertPlacement(plans, options.policy, options.device.limits, minimumSlots);
|
|
39
|
+
for (const plan of plans) assertExpertReadShape(options, plan);
|
|
40
|
+
const lease = options.rangeLease;
|
|
41
|
+
if (options.policy.sourceReadMode === 'expert-parallel') {
|
|
42
|
+
if (lease || !options.readRangeInto
|
|
43
|
+
|| options.policy.stagingBytes < Math.max(...plans.map(plan => plan.expertBytes)) * 2 * options.policy.readConcurrency) {
|
|
44
|
+
throw new Error('Tiered parallel expert reads require bounded direct scratch within stagingBytes.');
|
|
45
|
+
}
|
|
46
|
+
} else {
|
|
47
|
+
const mappedLimit = options.policy.stagingBytes - (lease?.memoryBytes ?? 0);
|
|
48
|
+
const lanes = Math.min(options.policy.readConcurrency, lease?.concurrency ?? 1);
|
|
49
|
+
const leaseLanes = lease?.concurrency ?? 1;
|
|
50
|
+
if (!lease || !Number.isSafeInteger(leaseLanes) || leaseLanes < 1
|
|
51
|
+
|| !Number.isSafeInteger(lease.maxPartBytes) || lease.maxPartBytes < 4 || lease.maxPartBytes % 4
|
|
52
|
+
|| !Number.isSafeInteger(lease.memoryBytes) || !Number.isSafeInteger(lease.maxPartBytes * leaseLanes)
|
|
53
|
+
|| lease.memoryBytes < lease.maxPartBytes * leaseLanes
|
|
54
|
+
|| mappedLimit < Math.max(...plans.map(plan => plan.expertBytes)) * lanes) {
|
|
55
|
+
throw new Error('Tiered expert leases and concurrent mapped writers must fit aggregate stagingBytes.');
|
|
56
|
+
}
|
|
57
|
+
}
|
|
58
|
+
const mappedLimit = options.policy.stagingBytes - (lease?.memoryBytes ?? 0);
|
|
59
|
+
if (options.otherGpuBytes() + this.#placement.cacheUpperBytes > options.policy.gpuBudgetBytes - options.policy.reserveBytes) {
|
|
60
|
+
throw new Error('Tiered expert placement and existing weights exceed the GPU reserve.');
|
|
61
|
+
}
|
|
62
|
+
const mappedBytes = this.#placement.parkedBytes + this.#placement.mappedCacheBytes;
|
|
63
|
+
if (mappedBytes) this.#mapped = new MappedExpertUploadCache({ device: options.device,
|
|
64
|
+
maxBytes: mappedBytes, maxMappedBytes: Math.min(mappedLimit, mappedBytes),
|
|
65
|
+
maxEntries: plans.reduce((sum, plan) => sum + plan.numExperts, 0),
|
|
66
|
+
admission: options.policy.hostCacheAdmission,
|
|
67
|
+
protectedShare: options.policy.tiered?.mappedProtectedShare,
|
|
68
|
+
maxGhostEntries: Math.max(1, Math.floor(this.#placement.mappedCacheBytes / Math.min(...plans.map(plan => plan.expertBytes)))),
|
|
69
|
+
allocate: descriptor => this.#allocate(descriptor), dispose: buffer => this.#retire(buffer) });
|
|
70
|
+
const compactLayerCount = [...this.#placement.layers.values()].filter(row => row.mode !== 'resident').length;
|
|
71
|
+
if (compactLayerCount) {
|
|
72
|
+
const policy = { ...options.policy, storageMode: /** @type {const} */ ('compact-decode'), tiered: null,
|
|
73
|
+
cacheBytes: this.#placement.compactBytes, compactLayerBytes: this.#placement.compactLayerBytes };
|
|
74
|
+
this.#compact = new CompactExpertCache({ ...options, policy, compactLayerCount,
|
|
75
|
+
copyCachedExpert: (plan, expert, planes, slot, signal) => this.#copy(plan, expert, planes, slot, signal),
|
|
76
|
+
rememberExpert: (plan, expert, bytes, signal) => this.#remember(plan, expert, bytes, signal) }, this.#stats);
|
|
77
|
+
}
|
|
78
|
+
}
|
|
79
|
+
|
|
80
|
+
/** @param {GPUBufferDescriptor} descriptor */
|
|
81
|
+
#allocate(descriptor) {
|
|
82
|
+
const bytes = Number(descriptor.size), { device, policy } = this.#options;
|
|
83
|
+
if (!Number.isSafeInteger(bytes) || bytes < 4 || bytes > device.limits.maxBufferSize
|
|
84
|
+
|| this.#stats.residentBytes + bytes > this.#placement.cacheUpperBytes
|
|
85
|
+
|| this.#options.otherGpuBytes() + this.#stats.residentBytes + bytes > policy.gpuBudgetBytes - policy.reserveBytes) {
|
|
86
|
+
throw new Error('Tiered expert allocation exceeds admitted cache or device budget.');
|
|
87
|
+
}
|
|
88
|
+
const buffer = device.createBuffer(descriptor);
|
|
89
|
+
this.#owned.add(buffer);
|
|
90
|
+
this.#stats.residentBytes += buffer.size;
|
|
91
|
+
this.#stats.peakResidentBytes = Math.max(this.#stats.peakResidentBytes, this.#stats.residentBytes);
|
|
92
|
+
this.#stats.peakAccountedGpuBytes = Math.max(this.#stats.peakAccountedGpuBytes,
|
|
93
|
+
this.#options.otherGpuBytes() + this.#stats.residentBytes);
|
|
94
|
+
return buffer;
|
|
95
|
+
}
|
|
96
|
+
|
|
97
|
+
/** @param {GPUBuffer} buffer */
|
|
98
|
+
#retire(buffer) {
|
|
99
|
+
buffer.destroy();
|
|
100
|
+
if (this.#owned.delete(buffer)) this.#stats.residentBytes -= buffer.size;
|
|
101
|
+
}
|
|
102
|
+
|
|
103
|
+
/** @template T @param {AbortSignal | undefined} external @param {(signal: AbortSignal) => Promise<T>} action */
|
|
104
|
+
#run(external, action) {
|
|
105
|
+
if (this.#closed || this.#active || isDeviceLost(this.#options.device)) throw new Error('Tiered expert cache is closed, lost or already preparing.');
|
|
106
|
+
const controller = new AbortController();
|
|
107
|
+
const abort = () => controller.abort(external?.reason);
|
|
108
|
+
const shutdown = () => controller.abort(this.#shutdown.signal.reason);
|
|
109
|
+
external?.addEventListener('abort', abort, { once: true });
|
|
110
|
+
this.#shutdown.signal.addEventListener('abort', shutdown, { once: true });
|
|
111
|
+
if (external?.aborted) abort();
|
|
112
|
+
const task = Promise.resolve().then(() => { controller.signal.throwIfAborted(); return action(controller.signal); });
|
|
113
|
+
this.#active = task;
|
|
114
|
+
return task.finally(() => {
|
|
115
|
+
external?.removeEventListener('abort', abort);
|
|
116
|
+
this.#shutdown.signal.removeEventListener('abort', shutdown);
|
|
117
|
+
this.#active = null;
|
|
118
|
+
});
|
|
119
|
+
}
|
|
120
|
+
|
|
121
|
+
/** @param {import('./routed-expert-plan.js').RoutedExpertBankPlan} plan @param {number} expert */
|
|
122
|
+
#key(plan, expert) { return `${plan.layer}:${expert}`; }
|
|
123
|
+
|
|
124
|
+
/** @param {import('./routed-expert-plan.js').RoutedExpertBankPlan} plan @param {number} expert
|
|
125
|
+
* @param {Uint8Array} bytes @param {AbortSignal} signal */
|
|
126
|
+
async #remember(plan, expert, bytes, signal) {
|
|
127
|
+
if (!this.#mapped) return;
|
|
128
|
+
const pinned = this.#placement.layers.get(plan.layer)?.mode === 'parked';
|
|
129
|
+
if (!pinned && this.#options.policy.hostCacheAdmissionPhase === 'decode' && this.#phase !== 'decode') return;
|
|
130
|
+
const accepted = await this.#mapped.admit(this.#key(plan, expert), plan.expertBytes, destination => {
|
|
131
|
+
signal.throwIfAborted();
|
|
132
|
+
this.#sampleStaging();
|
|
133
|
+
destination.set(bytes);
|
|
134
|
+
return bytes.byteLength;
|
|
135
|
+
}, { pinned, signal });
|
|
136
|
+
if (pinned && !accepted) throw new Error('Declared parked expert could not be admitted.');
|
|
137
|
+
this.#sampleStaging();
|
|
138
|
+
}
|
|
139
|
+
|
|
140
|
+
/** @param {import('./routed-expert-plan.js').RoutedExpertBankPlan} plan @param {number} expert
|
|
141
|
+
* @param {ReadonlyMap<string, GPUBuffer>} planes @param {number} slot @param {AbortSignal} signal */
|
|
142
|
+
#copy(plan, expert, planes, slot, signal) {
|
|
143
|
+
let offset = 0;
|
|
144
|
+
const ranges = plan.planes.map(plane => {
|
|
145
|
+
const destination = planes.get(plane.role);
|
|
146
|
+
if (!destination) throw new Error('Tiered compact copy requires all declared destination planes.');
|
|
147
|
+
const range = { sourceOffset: offset, destination, destinationOffset: slot * plane.expertBytes, bytes: plane.expertBytes };
|
|
148
|
+
offset += plane.expertBytes;
|
|
149
|
+
return range;
|
|
150
|
+
});
|
|
151
|
+
const copy = this.#mapped?.copyInto(this.#key(plan, expert), ranges, signal);
|
|
152
|
+
if (!copy) return false;
|
|
153
|
+
// Queue ordering makes the data ready for subsequent kernels. The mapped
|
|
154
|
+
// cache independently retains source ownership until this actual fence settles.
|
|
155
|
+
this.#stats.uploadBytes += copy.bytes;
|
|
156
|
+
this.#stats.expertCopySubmissions++;
|
|
157
|
+
return true;
|
|
158
|
+
}
|
|
159
|
+
|
|
160
|
+
#sampleStaging() {
|
|
161
|
+
this.#stats.peakStagingBytes = Math.max(this.#stats.peakStagingBytes,
|
|
162
|
+
(this.#options.rangeLease?.memoryBytes ?? 0) + (this.#mapped?.getStats().mappedBytes ?? 0) + this.#uploadingBytes);
|
|
163
|
+
}
|
|
164
|
+
|
|
165
|
+
/** @type {import('./tiered-expert-cache.js').TieredExpertCache['preload']} */
|
|
166
|
+
preload(external) {
|
|
167
|
+
return this.#run(external, signal => this.#preload(signal));
|
|
168
|
+
}
|
|
169
|
+
|
|
170
|
+
/** @param {AbortSignal} signal */
|
|
171
|
+
async #preload(signal) {
|
|
172
|
+
if (this.#ready) return;
|
|
173
|
+
// Failed setup is retired before retry, never reused as a complete bank.
|
|
174
|
+
if (this.#owned.size) throw new Error('Failed tiered setup requires unload before retry.');
|
|
175
|
+
for (const { plan, mode } of this.#placement.layers.values()) {
|
|
176
|
+
if (mode === 'streamed') continue;
|
|
177
|
+
/** @type {Map<string, GPUBuffer>} */
|
|
178
|
+
const planes = new Map();
|
|
179
|
+
if (mode === 'resident') {
|
|
180
|
+
for (const plane of plan.planes) planes.set(plane.role, this.#allocate({ size: plane.bytes,
|
|
181
|
+
usage: GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST | GPUBufferUsage.COPY_SRC,
|
|
182
|
+
label: `tiered_resident_${plan.layer}_${plane.role}` }));
|
|
183
|
+
}
|
|
184
|
+
let cursor = 0;
|
|
185
|
+
const controller = new AbortController();
|
|
186
|
+
const abort = () => controller.abort(signal.reason);
|
|
187
|
+
signal.addEventListener('abort', abort, { once: true });
|
|
188
|
+
if (signal.aborted) abort();
|
|
189
|
+
try {
|
|
190
|
+
const directParallel = this.#options.policy.sourceReadMode === 'expert-parallel';
|
|
191
|
+
const concurrency = directParallel
|
|
192
|
+
? this.#options.policy.readConcurrency
|
|
193
|
+
: Math.min(this.#options.policy.readConcurrency, this.#options.rangeLease?.concurrency ?? 1);
|
|
194
|
+
const results = await Promise.allSettled(Array.from({ length: concurrency }, async () => {
|
|
195
|
+
const directScratch = directParallel ? new Uint8Array(plan.expertBytes) : null;
|
|
196
|
+
if (directScratch) {
|
|
197
|
+
this.#uploadingBytes += directScratch.byteLength;
|
|
198
|
+
this.#sampleStaging();
|
|
199
|
+
}
|
|
200
|
+
try {
|
|
201
|
+
while (cursor < plan.numExperts) {
|
|
202
|
+
const expert = cursor++;
|
|
203
|
+
try {
|
|
204
|
+
if (directScratch) {
|
|
205
|
+
await readRoutedExpertPayload(this.#options, this.#stats, plan, expert, controller.signal, directScratch, null);
|
|
206
|
+
controller.signal.throwIfAborted();
|
|
207
|
+
this.#uploadingBytes += plan.expertBytes;
|
|
208
|
+
this.#sampleStaging();
|
|
209
|
+
try {
|
|
210
|
+
let offset = 0;
|
|
211
|
+
for (const plane of plan.planes) {
|
|
212
|
+
const buffer = planes.get(plane.role);
|
|
213
|
+
if (!buffer) throw new Error('Tiered resident plane is missing.');
|
|
214
|
+
this.#options.device.queue.writeBuffer(buffer, expert * plane.expertBytes,
|
|
215
|
+
directScratch.subarray(offset, offset + plane.expertBytes));
|
|
216
|
+
this.#stats.uploadBytes += plane.expertBytes;
|
|
217
|
+
offset += plane.expertBytes;
|
|
218
|
+
}
|
|
219
|
+
await this.#options.device.queue.onSubmittedWorkDone();
|
|
220
|
+
controller.signal.throwIfAborted();
|
|
221
|
+
} finally { this.#uploadingBytes -= plan.expertBytes; }
|
|
222
|
+
} else {
|
|
223
|
+
const lease = await readRoutedExpertLease(this.#options, this.#stats, plan, expert, controller.signal, null);
|
|
224
|
+
try {
|
|
225
|
+
controller.signal.throwIfAborted();
|
|
226
|
+
if (mode === 'parked') await this.#remember(plan, expert, lease.bytes, controller.signal);
|
|
227
|
+
else {
|
|
228
|
+
this.#uploadingBytes += plan.expertBytes;
|
|
229
|
+
this.#sampleStaging();
|
|
230
|
+
try {
|
|
231
|
+
let offset = 0;
|
|
232
|
+
for (const plane of plan.planes) {
|
|
233
|
+
const buffer = planes.get(plane.role);
|
|
234
|
+
if (!buffer) throw new Error('Tiered resident plane is missing.');
|
|
235
|
+
this.#options.device.queue.writeBuffer(buffer, expert * plane.expertBytes,
|
|
236
|
+
lease.bytes.subarray(offset, offset + plane.expertBytes));
|
|
237
|
+
this.#stats.uploadBytes += plane.expertBytes;
|
|
238
|
+
offset += plane.expertBytes;
|
|
239
|
+
}
|
|
240
|
+
// writeBuffer copies its source immediately, but its native
|
|
241
|
+
// upload allocation may survive until queue completion. Keep
|
|
242
|
+
// this lane charged and stop further reads until that upload
|
|
243
|
+
// settles; model-sized setup must not queue model-sized staging.
|
|
244
|
+
await this.#options.device.queue.onSubmittedWorkDone();
|
|
245
|
+
controller.signal.throwIfAborted();
|
|
246
|
+
} finally { this.#uploadingBytes -= plan.expertBytes; }
|
|
247
|
+
}
|
|
248
|
+
} finally { lease.release(); }
|
|
249
|
+
}
|
|
250
|
+
} catch (error) { controller.abort(error); throw error; }
|
|
251
|
+
}
|
|
252
|
+
} finally {
|
|
253
|
+
if (directScratch) this.#uploadingBytes -= directScratch.byteLength;
|
|
254
|
+
}
|
|
255
|
+
}));
|
|
256
|
+
const failure = results.find(result => result.status === 'rejected');
|
|
257
|
+
if (failure) throw failure.reason;
|
|
258
|
+
signal.throwIfAborted();
|
|
259
|
+
if (mode === 'resident') {
|
|
260
|
+
const weights = Object.fromEntries(plan.planes.map(plane => [plane.role,
|
|
261
|
+
createWeightBuffer(/** @type {GPUBuffer} */ (planes.get(plane.role)), plane.dtype, 'row', plane.location.shape, plane.name)]));
|
|
262
|
+
this.#resident.set(plan.layer, { expertIdx: 0, expertFormat: 'qwen3_5', numExperts: plan.numExperts,
|
|
263
|
+
expertIntermediateSize: plan.intermediateSize, ...weights, down: weights.down });
|
|
264
|
+
}
|
|
265
|
+
} finally { signal.removeEventListener('abort', abort); }
|
|
266
|
+
}
|
|
267
|
+
await this.#options.device.queue.onSubmittedWorkDone();
|
|
268
|
+
signal.throwIfAborted();
|
|
269
|
+
this.#ready = true;
|
|
270
|
+
this.#sampleStaging();
|
|
271
|
+
}
|
|
272
|
+
|
|
273
|
+
/** @type {import('./tiered-expert-cache.js').TieredExpertCache['getResidentWeights']} */
|
|
274
|
+
getResidentWeights(layer) {
|
|
275
|
+
return this.#ready && !this.#closed && !isDeviceLost(this.#options.device) ? this.#resident.get(layer) ?? null : null;
|
|
276
|
+
}
|
|
277
|
+
|
|
278
|
+
/** @type {import('./tiered-expert-cache.js').TieredExpertCache['capacity']} */
|
|
279
|
+
capacity(layer) { return this.#placement.layers.get(layer)?.capacity ?? null; }
|
|
280
|
+
|
|
281
|
+
/** @type {import('./tiered-expert-cache.js').TieredExpertCache['prepare']} */
|
|
282
|
+
prepare(plan, ids, external, phase) {
|
|
283
|
+
return this.#run(external, async signal => {
|
|
284
|
+
if (!this.#ready || (phase !== 'prefill' && phase !== 'decode')) throw new Error('Tiered experts require completed setup and an explicit phase.');
|
|
285
|
+
const row = this.#placement.layers.get(plan.layer);
|
|
286
|
+
if (!row || row.plan !== plan || !ids.length || ids.some(id => !Number.isSafeInteger(id) || id < 0 || id >= plan.numExperts)) {
|
|
287
|
+
throw new Error('Tiered selection must bind its admitted source plan and expert IDs.');
|
|
288
|
+
}
|
|
289
|
+
const resident = this.#resident.get(plan.layer);
|
|
290
|
+
if (resident) return resident;
|
|
291
|
+
if (!this.#compact) throw new Error('Tiered compact storage is absent.');
|
|
292
|
+
this.#phase = phase;
|
|
293
|
+
try { return await this.#compact.prepare(plan, [...new Set(ids)], signal); }
|
|
294
|
+
finally { this.#phase = null; }
|
|
295
|
+
});
|
|
296
|
+
}
|
|
297
|
+
|
|
298
|
+
getStats() {
|
|
299
|
+
const compact = this.#compact?.getStats();
|
|
300
|
+
return { ...this.#stats, evictionPolicy: /** @type {const} */ ('lru'), frequencyHistoryEntries: 0,
|
|
301
|
+
frequencyResidentEntries: 0, phaseProfiles: {}, experts: (compact?.experts ?? 0)
|
|
302
|
+
+ [...this.#resident.values()].reduce((sum, row) => sum + row.numExperts, 0),
|
|
303
|
+
cacheBytes: this.#stats.residentBytes, workingBytes: 0, compactLayers: compact?.compactLayers ?? {},
|
|
304
|
+
hostCacheBytes: 0, hostReservedBytes: 0, hostExperts: 0,
|
|
305
|
+
tiered: { residentLayers: [...this.#resident.keys()], ready: this.#ready, mapped: this.#mapped?.getStats() ?? null } };
|
|
306
|
+
}
|
|
307
|
+
|
|
308
|
+
async destroy() {
|
|
309
|
+
this.#closed = true;
|
|
310
|
+
this.#ready = false;
|
|
311
|
+
this.#shutdown.abort(new DOMException('Tiered experts unloaded.', 'AbortError'));
|
|
312
|
+
await this.#active?.catch(() => undefined);
|
|
313
|
+
const errors = [];
|
|
314
|
+
try { await this.#mapped?.close(); } catch (error) { errors.push(error); }
|
|
315
|
+
try { await this.#compact?.destroy(); } catch (error) { errors.push(error); }
|
|
316
|
+
try { await this.#options.device.queue.onSubmittedWorkDone(); }
|
|
317
|
+
catch (error) { if (!isDeviceLost(this.#options.device)) errors.push(error); }
|
|
318
|
+
// Mapped allocations retain their own retryable ownership if close failed.
|
|
319
|
+
if (!errors.length) {
|
|
320
|
+
for (const buffer of this.#owned) {
|
|
321
|
+
try { this.#retire(buffer); } catch (error) { errors.push(error); }
|
|
322
|
+
}
|
|
323
|
+
}
|
|
324
|
+
if (!this.#owned.size) this.#resident.clear();
|
|
325
|
+
if (errors.length) throw new AggregateError(errors, 'Tiered expert cleanup failed; ownership retained for retry.');
|
|
326
|
+
}
|
|
327
|
+
}
|
|
@@ -0,0 +1,74 @@
|
|
|
1
|
+
import { validateQwen35RoutedConfig } from '../../config/schema/runtime/loading.schema.js';
|
|
2
|
+
|
|
3
|
+
/** @type {typeof import('./tiered-expert-plan.js').planTieredExpertPlacement} */
|
|
4
|
+
export function planTieredExpertPlacement(plans, policy, limits, minimumSlots) {
|
|
5
|
+
validateQwen35RoutedConfig(policy);
|
|
6
|
+
const tiered = policy.tiered;
|
|
7
|
+
if (policy.storageMode !== 'tiered-compact' || !tiered || !plans.length
|
|
8
|
+
|| !Number.isSafeInteger(minimumSlots) || minimumSlots < 1
|
|
9
|
+
|| ![limits.maxBufferSize, limits.maxStorageBufferBindingSize].every(value => Number.isSafeInteger(value) && value > 0)) {
|
|
10
|
+
throw new Error('Tiered expert placement requires assigned plans, declared top-K and device limits.');
|
|
11
|
+
}
|
|
12
|
+
const numLayers = plans[0].numLayers;
|
|
13
|
+
if (new Set(plans.map(plan => plan.layer)).size !== plans.length
|
|
14
|
+
|| plans.some(plan => plan.numLayers !== numLayers || !Number.isSafeInteger(plan.layer)
|
|
15
|
+
|| plan.layer < 0 || plan.layer >= numLayers || !Number.isSafeInteger(plan.bankBytes) || plan.bankBytes < 1
|
|
16
|
+
|| !Number.isSafeInteger(plan.expertBytes) || plan.expertBytes < 1 || plan.expertBytes % 4
|
|
17
|
+
|| !Number.isSafeInteger(plan.numExperts) || plan.numExperts < minimumSlots
|
|
18
|
+
|| plan.bankBytes !== plan.numExperts * plan.expertBytes
|
|
19
|
+
|| !plan.planes.length || plan.planes.some(plane => !Number.isSafeInteger(plane.expertBytes)
|
|
20
|
+
|| plane.expertBytes < 1 || plane.expertBytes % 4 || plane.bytes !== plane.expertBytes * plan.numExperts)
|
|
21
|
+
|| plan.planes.reduce((sum, plane) => sum + plane.bytes, 0) !== plan.bankBytes)
|
|
22
|
+
|| [...tiered.residentLayers, ...tiered.parkedLayers].some(layer => layer >= numLayers)) {
|
|
23
|
+
throw new Error('Tiered expert geometry must preserve assigned original layers and complete packed planes.');
|
|
24
|
+
}
|
|
25
|
+
const resident = new Set(tiered.residentLayers), parked = new Set(tiered.parkedLayers);
|
|
26
|
+
const residentBytes = plans.filter(plan => resident.has(plan.layer)).reduce((sum, plan) => sum + plan.bankBytes, 0);
|
|
27
|
+
const parkedBytes = plans.filter(plan => parked.has(plan.layer)).reduce((sum, plan) => sum + plan.bankBytes, 0);
|
|
28
|
+
const offloaded = plans.filter(plan => !resident.has(plan.layer));
|
|
29
|
+
const remaining = policy.cacheBytes - residentBytes - parkedBytes - tiered.mappedCacheBytes;
|
|
30
|
+
if (!Number.isSafeInteger(remaining) || remaining < 0
|
|
31
|
+
|| plans.some(plan => resident.has(plan.layer)
|
|
32
|
+
? plan.planes.some(plane => plane.bytes > Math.min(limits.maxBufferSize, limits.maxStorageBufferBindingSize))
|
|
33
|
+
: plan.expertBytes > limits.maxBufferSize)) {
|
|
34
|
+
throw new Error('Resident, parked and mapped expert storage exceeds the declared cache or device bounds.');
|
|
35
|
+
}
|
|
36
|
+
const supplied = policy.compactLayerBytes;
|
|
37
|
+
if (supplied !== null && (Object.keys(supplied).length !== offloaded.length
|
|
38
|
+
|| offloaded.some(plan => !Object.hasOwn(supplied, plan.layer)))) {
|
|
39
|
+
throw new Error('Tiered compact quotas must cover exactly the assigned nonresident layers.');
|
|
40
|
+
}
|
|
41
|
+
if (supplied !== null && Object.values(supplied).reduce((sum, bytes) => sum + bytes, 0) > remaining) {
|
|
42
|
+
throw new Error('Tiered compact quotas exceed the remaining model cache budget.');
|
|
43
|
+
}
|
|
44
|
+
/** @type {Map<number, import('./tiered-expert-plan.js').TieredExpertLayer>} */
|
|
45
|
+
const layers = new Map();
|
|
46
|
+
/** @type {Record<string, number>} */
|
|
47
|
+
const compactLayerBytes = {};
|
|
48
|
+
let compactBytes = 0;
|
|
49
|
+
for (const plan of plans) {
|
|
50
|
+
if (resident.has(plan.layer)) {
|
|
51
|
+
layers.set(plan.layer, { plan, mode: 'resident', capacity: plan.numExperts, quotaBytes: plan.bankBytes });
|
|
52
|
+
continue;
|
|
53
|
+
}
|
|
54
|
+
const quota = supplied === null ? Math.floor(remaining / offloaded.length / 4) * 4 : supplied[plan.layer];
|
|
55
|
+
const mapBytes = plan.numExperts * 4;
|
|
56
|
+
const capacity = Math.min(plan.numExperts, Math.floor((quota - mapBytes) / plan.expertBytes),
|
|
57
|
+
...plan.planes.map(plane => Math.floor(Math.min(limits.maxBufferSize, limits.maxStorageBufferBindingSize) / plane.expertBytes)));
|
|
58
|
+
if (!Number.isSafeInteger(capacity) || capacity < minimumSlots
|
|
59
|
+
|| mapBytes > Math.min(limits.maxBufferSize, limits.maxStorageBufferBindingSize)) {
|
|
60
|
+
throw new Error(`Tiered expert layer ${plan.layer} cannot hold one complete top-K selection.`);
|
|
61
|
+
}
|
|
62
|
+
const actual = mapBytes + capacity * plan.expertBytes;
|
|
63
|
+
compactBytes += actual;
|
|
64
|
+
compactLayerBytes[plan.layer] = actual;
|
|
65
|
+
layers.set(plan.layer, { plan, mode: parked.has(plan.layer) ? 'parked' : 'streamed', capacity, quotaBytes: actual });
|
|
66
|
+
}
|
|
67
|
+
if (compactBytes > remaining) throw new Error('Tiered compact quotas exceed the remaining model cache budget.');
|
|
68
|
+
const cacheUpperBytes = residentBytes + parkedBytes + tiered.mappedCacheBytes + compactBytes;
|
|
69
|
+
if (!Number.isSafeInteger(cacheUpperBytes) || cacheUpperBytes > policy.cacheBytes) {
|
|
70
|
+
throw new Error('Tiered expert allocation sum exceeds cacheBytes.');
|
|
71
|
+
}
|
|
72
|
+
return { layers, residentBytes, parkedBytes, mappedCacheBytes: tiered.mappedCacheBytes,
|
|
73
|
+
compactBytes, cacheUpperBytes, compactLayerBytes };
|
|
74
|
+
}
|
|
@@ -45,7 +45,7 @@ export interface LoadedEmbeddingPostprocessor
|
|
|
45
45
|
* Context required for final weights loading.
|
|
46
46
|
*/
|
|
47
47
|
export interface FinalWeightsContext {
|
|
48
|
-
hyperConnections?: import(
|
|
48
|
+
hyperConnections?: import("../config/schema/execution/hyper-connection-decoder.schema.js").ResolvedHyperConnections | null;
|
|
49
49
|
/** Tensor locations map */
|
|
50
50
|
tensorLocations: Map<string, TensorLocation>;
|
|
51
51
|
weightRepresentationPlan?: import('./weight-representation-plan.js').WeightRepresentationPlan | null;
|
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
import type { LayerLoaderContext } from './layer-loader.js';
|
|
2
|
-
import type { ResolvedHyperConnections, ResolvedIndexedAttention } from
|
|
3
|
-
import type { ResolvedNgramEmbedding } from
|
|
2
|
+
import type { ResolvedHyperConnections, ResolvedIndexedAttention } from "../config/schema/execution/hyper-connection-decoder.schema.js";
|
|
3
|
+
import type { ResolvedNgramEmbedding } from "../config/schema/execution/ngram-embedding.schema.js";
|
|
4
4
|
import type { GatedResidualWeights } from '../inference/pipelines/text/hyper-connection/decoder-layer.js';
|
|
5
5
|
import type { PleLayerWeights } from '../inference/pipelines/text/ngram/ple-layer.js';
|
|
6
6
|
import type { Tensor } from '../gpu/tensor.js';
|
|
@@ -1,5 +1,8 @@
|
|
|
1
|
+
import { acquireBuffer, releaseBuffer } from '../memory/buffer-pool.js';
|
|
2
|
+
import { ggufBlockBytes, isGgufBlockDtype } from '../formats/gguf/block-storage.js';
|
|
3
|
+
import { getDevice } from '../gpu/device.js';
|
|
1
4
|
import { createTensor } from '../gpu/tensor.js';
|
|
2
|
-
import { getWeightDtype, isWeightBuffer } from '../gpu/weight-buffer.js';
|
|
5
|
+
import { createWeightBuffer, getWeightDtype, isGpuBufferInstance, isWeightBuffer } from '../gpu/weight-buffer.js';
|
|
3
6
|
import { castF16ToF32 } from '../gpu/kernels/cast.js';
|
|
4
7
|
import {
|
|
5
8
|
GATED_RESIDUAL_SUFFIXES, HYPER_CONNECTION_MIXER_PREFIXES, HYPER_CONNECTION_SUFFIXES,
|
|
@@ -15,13 +18,16 @@ function fail(message) {
|
|
|
15
18
|
* @param {readonly string[]} prefixes @param {readonly string[]} suffixes
|
|
16
19
|
* @param {readonly number[]} shape @param {string | null} role @param {boolean} tensor */
|
|
17
20
|
function binding(ctx, prefixes, suffixes, shape, role, tensor) {
|
|
18
|
-
const candidates = prefixes.flatMap(prefix => suffixes.map(suffix => `${prefix}.${suffix}`));
|
|
21
|
+
const candidates = prefixes.flatMap(prefix => suffixes.map(suffix => prefix ? `${prefix}.${suffix}` : suffix));
|
|
19
22
|
const matches = candidates.filter(name => ctx.tensorLocations.has(name));
|
|
20
23
|
if (matches.length !== 1) fail(`Expected one ${candidates[0]}; found ${matches.length}.`);
|
|
21
24
|
const name = matches[0];
|
|
22
25
|
const location = ctx.tensorLocations.get(name);
|
|
23
26
|
if (!location) fail(`Missing location for ${name}.`);
|
|
24
|
-
const
|
|
27
|
+
const storedShape = location.originalShape ?? location.shape;
|
|
28
|
+
// GGUF removes the unit group axis from the depthwise convolution.
|
|
29
|
+
const logicalShape = name.endsWith('.ple_conv1d.weight') && storedShape.length === 2
|
|
30
|
+
? [storedShape[0], 1, storedShape[1]] : storedShape;
|
|
25
31
|
if (logicalShape.length !== shape.length || logicalShape.some((value, index) => value !== shape[index])) {
|
|
26
32
|
fail(`${name} requires shape [${shape}]; got [${logicalShape}].`);
|
|
27
33
|
}
|
|
@@ -55,19 +61,36 @@ async function materialize(ctx, bindings) {
|
|
|
55
61
|
result[key] = converted;
|
|
56
62
|
} else result[key] = vector;
|
|
57
63
|
} else {
|
|
58
|
-
|
|
59
|
-
|
|
64
|
+
let projection = loaded;
|
|
65
|
+
// Small GGUF projections can carry role=other. Their float loader returns
|
|
66
|
+
// a tagged GPU buffer after BF16 expansion, rather than a WeightBuffer.
|
|
67
|
+
// Bind that already-owned materialization; never infer dtype or layout.
|
|
68
|
+
if (!isWeightBuffer(projection) && isGpuBufferInstance(projection)
|
|
69
|
+
&& (dtype === 'f16' || dtype === 'f32') && spec.location.layout === 'row') {
|
|
70
|
+
if (!(ctx.gpuBuffers instanceof Set) || !ctx.gpuBuffers.has(projection)) {
|
|
71
|
+
fail(`${spec.name} requires its loaded GPU projection owner.`);
|
|
72
|
+
}
|
|
73
|
+
projection = createWeightBuffer(projection, dtype, spec.location.layout, [...spec.shape], spec.name);
|
|
60
74
|
}
|
|
61
|
-
if (
|
|
62
|
-
|
|
75
|
+
if (!isWeightBuffer(projection) || (dtype !== 'f16' && dtype !== 'f32' && !isGgufBlockDtype(dtype))) {
|
|
76
|
+
fail(`${spec.name} requires a WeightBuffer with supported floating or GGUF block projection storage and layout.`);
|
|
77
|
+
}
|
|
78
|
+
if (projection.shape.length !== spec.shape.length || projection.shape.some((value, index) => value !== spec.shape[index])
|
|
79
|
+
|| buffer.size < projectionBytes(projection) || (isGgufBlockDtype(dtype) && projection.layout !== 'row')) {
|
|
63
80
|
fail(`${spec.name} has an incomplete or mismatched projection representation.`);
|
|
64
81
|
}
|
|
65
|
-
result[key] =
|
|
82
|
+
result[key] = projection;
|
|
66
83
|
}
|
|
67
84
|
}
|
|
68
85
|
return result;
|
|
69
86
|
}
|
|
70
87
|
|
|
88
|
+
/** @param {import('../gpu/weight-buffer.js').WeightBuffer} weight */
|
|
89
|
+
function projectionBytes(weight) {
|
|
90
|
+
if (isGgufBlockDtype(weight.dtype)) return ggufBlockBytes(weight.shape, weight.dtype);
|
|
91
|
+
return weight.shape.reduce((a, b) => a * b, 1) * (weight.dtype === 'f16' ? 2 : 4);
|
|
92
|
+
}
|
|
93
|
+
|
|
71
94
|
/** @param {Awaited<ReturnType<typeof materialize>>} weights @param {string} key */
|
|
72
95
|
function vectorWeight(weights, key) {
|
|
73
96
|
const value = weights[key];
|
|
@@ -108,6 +131,56 @@ function gatedBindings(ctx, prefixes, suffixes, role, inject) {
|
|
|
108
131
|
return bindings;
|
|
109
132
|
}
|
|
110
133
|
|
|
134
|
+
/** @param {import('./hyper-connection-loader.js').MechanismLoaderContext} ctx
|
|
135
|
+
* @param {readonly string[]} prefixes
|
|
136
|
+
* @param {import("../config/schema/execution/hyper-connection-decoder.schema.js").ResolvedIndexedAttention} indexed
|
|
137
|
+
* @returns {Record<string, ReturnType<typeof binding>>} */
|
|
138
|
+
function indexerBindings(ctx, prefixes, indexed) {
|
|
139
|
+
const has = (/** @type {readonly string[]} */ suffixes) => prefixes.some(prefix => suffixes.some(suffix => ctx.tensorLocations.has(`${prefix}.${suffix}`)));
|
|
140
|
+
const fused = has(QSA_SUFFIXES.projection);
|
|
141
|
+
const split = has(QSA_SUFFIXES.qProjection) || has(QSA_SUFFIXES.kProjection);
|
|
142
|
+
if (fused && split) fail('Indexer cannot declare both fused and separate q/k projections.');
|
|
143
|
+
if (!ctx.hyperConnections) fail('Indexer requires hyperConnections.');
|
|
144
|
+
const width = ctx.hyperConnections.streamWidth;
|
|
145
|
+
if (!split) return { projection: binding(ctx, prefixes, QSA_SUFFIXES.projection,
|
|
146
|
+
[(indexed.queryHeads + indexed.keyHeads) * indexed.headDim, width], 'qsa_index_qk', false) };
|
|
147
|
+
return {
|
|
148
|
+
qProjection: binding(ctx, prefixes, QSA_SUFFIXES.qProjection, [indexed.queryHeads * indexed.headDim, width], 'qsa_index_qk', false),
|
|
149
|
+
kProjection: binding(ctx, prefixes, QSA_SUFFIXES.kProjection, [indexed.keyHeads * indexed.headDim, width], 'qsa_index_qk', false),
|
|
150
|
+
};
|
|
151
|
+
}
|
|
152
|
+
|
|
153
|
+
/** GGUF stores Q then K as separate row-major matrices. Join their already
|
|
154
|
+
* materialized rows once on the GPU, retaining the existing fused runtime op.
|
|
155
|
+
* @param {import('./hyper-connection-loader.js').MechanismLoaderContext} ctx
|
|
156
|
+
* @param {Awaited<ReturnType<typeof materialize>>} weights */
|
|
157
|
+
function fuseIndexer(ctx, weights) {
|
|
158
|
+
if (weights.projection) return projectionWeight(weights, 'projection');
|
|
159
|
+
const q = projectionWeight(weights, 'qProjection');
|
|
160
|
+
const k = projectionWeight(weights, 'kProjection');
|
|
161
|
+
if (q.layout !== 'row' || k.layout !== 'row' || q.dtype !== k.dtype || q.shape[1] !== k.shape[1]) {
|
|
162
|
+
fail('Separate indexer projections require matching row-major storage.');
|
|
163
|
+
}
|
|
164
|
+
if (!(ctx.gpuBuffers instanceof Set)) fail('Separate indexer projections require a GPU buffer owner.');
|
|
165
|
+
const device = getDevice();
|
|
166
|
+
if (!device) fail('Separate indexer projections require a GPU device.');
|
|
167
|
+
const qBytes = projectionBytes(q);
|
|
168
|
+
const kBytes = projectionBytes(k);
|
|
169
|
+
if (qBytes % 4 || kBytes % 4) fail('Separate indexer projection rows require copy-aligned byte lengths.');
|
|
170
|
+
const buffer = acquireBuffer(qBytes + kBytes, undefined, 'qsa_index_qk');
|
|
171
|
+
try {
|
|
172
|
+
const encoder = device.createCommandEncoder();
|
|
173
|
+
encoder.copyBufferToBuffer(q.buffer, 0, buffer, 0, qBytes);
|
|
174
|
+
encoder.copyBufferToBuffer(k.buffer, 0, buffer, qBytes, kBytes);
|
|
175
|
+
device.queue.submit([encoder.finish()]);
|
|
176
|
+
ctx.gpuBuffers.add(buffer);
|
|
177
|
+
return createWeightBuffer(buffer, q.dtype, 'row', [q.shape[0] + k.shape[0], q.shape[1]], 'qsa_index_qk');
|
|
178
|
+
} catch (error) {
|
|
179
|
+
releaseBuffer(buffer);
|
|
180
|
+
throw error;
|
|
181
|
+
}
|
|
182
|
+
}
|
|
183
|
+
|
|
111
184
|
/** @type {typeof import('./hyper-connection-loader.js').loadHyperConnectionLayerWeights} */
|
|
112
185
|
export async function loadHyperConnectionLayerWeights(ctx, layerIdx) {
|
|
113
186
|
if (ctx.hyperConnections == null) return {};
|
|
@@ -118,8 +191,7 @@ export async function loadHyperConnectionLayerWeights(ctx, layerIdx) {
|
|
|
118
191
|
const ngram = ctx.ngramEmbedding;
|
|
119
192
|
const width = ctx.hyperConnections.streamCount * ctx.hyperConnections.streamWidth;
|
|
120
193
|
const qsa = indexed?.layers.includes(layerIdx) ? {
|
|
121
|
-
|
|
122
|
-
[(indexed.queryHeads + indexed.keyHeads) * indexed.headDim, ctx.hyperConnections.streamWidth], 'qsa_index_qk', false),
|
|
194
|
+
...indexerBindings(ctx, prefixes, indexed),
|
|
123
195
|
qNorm: binding(ctx, prefixes, QSA_SUFFIXES.qNorm, [indexed.headDim], null, true),
|
|
124
196
|
kNorm: binding(ctx, prefixes, QSA_SUFFIXES.kNorm, [indexed.headDim], null, true),
|
|
125
197
|
} : null;
|
|
@@ -135,7 +207,7 @@ export async function loadHyperConnectionLayerWeights(ctx, layerIdx) {
|
|
|
135
207
|
const qsaWeights = qsa ? await materialize(ctx, qsa) : null;
|
|
136
208
|
const pleWeights = ple ? await materialize(ctx, ple) : null;
|
|
137
209
|
return { hyperConnections,
|
|
138
|
-
qsaIndexer: qsaWeights ? { projection:
|
|
210
|
+
qsaIndexer: qsaWeights ? { projection: fuseIndexer(ctx, qsaWeights),
|
|
139
211
|
qNorm: vectorWeight(qsaWeights, 'qNorm'), kNorm: vectorWeight(qsaWeights, 'kNorm') } : null,
|
|
140
212
|
ple: pleWeights ? { keyProj: projectionWeight(pleWeights, 'keyProj'), valueProj: projectionWeight(pleWeights, 'valueProj'),
|
|
141
213
|
normKey: vectorWeight(pleWeights, 'normKey'), normQuery: vectorWeight(pleWeights, 'normQuery'),
|
|
@@ -145,6 +217,6 @@ export async function loadHyperConnectionLayerWeights(ctx, layerIdx) {
|
|
|
145
217
|
/** @type {typeof import('./hyper-connection-loader.js').loadHyperConnectionMixerWeights} */
|
|
146
218
|
export async function loadHyperConnectionMixerWeights(ctx) {
|
|
147
219
|
if (ctx.hyperConnections == null) return null;
|
|
148
|
-
const suffixes = Object.fromEntries(Object.entries(GATED_RESIDUAL_SUFFIXES).map(([key, value]) => [key, [value]]));
|
|
149
|
-
return materializeGated(ctx, gatedBindings(ctx, HYPER_CONNECTION_MIXER_PREFIXES, suffixes, 'hc_mixer', false));
|
|
220
|
+
const suffixes = Object.fromEntries(Object.entries(GATED_RESIDUAL_SUFFIXES).map(([key, value]) => [key, [value, `output_hc_${key}.weight`]]));
|
|
221
|
+
return materializeGated(ctx, gatedBindings(ctx, [...HYPER_CONNECTION_MIXER_PREFIXES, ''], suffixes, 'hc_mixer', false));
|
|
150
222
|
}
|
|
@@ -25,9 +25,9 @@ export type TensorLoader = (
|
|
|
25
25
|
* Context required for layer loading.
|
|
26
26
|
*/
|
|
27
27
|
export interface LayerLoaderContext {
|
|
28
|
-
hyperConnections?: import(
|
|
29
|
-
indexedAttention?: import(
|
|
30
|
-
ngramEmbedding?: import(
|
|
28
|
+
hyperConnections?: import("../config/schema/execution/hyper-connection-decoder.schema.js").ResolvedHyperConnections | null;
|
|
29
|
+
indexedAttention?: import("../config/schema/execution/hyper-connection-decoder.schema.js").ResolvedIndexedAttention | null;
|
|
30
|
+
ngramEmbedding?: import("../config/schema/execution/ngram-embedding.schema.js").ResolvedNgramEmbedding | null;
|
|
31
31
|
/** Tensor locations map */
|
|
32
32
|
tensorLocations: Map<string, TensorLocation>;
|
|
33
33
|
weightRepresentationPlan?: import('./weight-representation-plan.js').WeightRepresentationPlan | null;
|
|
@@ -8,7 +8,7 @@ import {
|
|
|
8
8
|
import { dequantize, dequantizeRowwise } from '../gpu/kernels/dequant.js';
|
|
9
9
|
import { releaseBuffer } from '../memory/buffer-pool.js';
|
|
10
10
|
import { batchDowncastWeights } from './weight-downcast.js';
|
|
11
|
-
import { QK_K } from
|
|
11
|
+
import { QK_K } from "../config/schema/manifest/quantization.schema.js";
|
|
12
12
|
import { log, trace as debugTrace } from '../debug/index.js';
|
|
13
13
|
import { createTensor } from '../gpu/tensor.js';
|
|
14
14
|
import { castF16ToF32 } from '../gpu/kernels/cast.js';
|