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
package/src/inference/pipelines/text/{generator-runtime.js → generator/session/generator-runtime.js}
RENAMED
|
@@ -1,11 +1,11 @@
|
|
|
1
|
-
import { readBuffer } from
|
|
2
|
-
import { resolveSamplingConfig } from
|
|
3
|
-
import { resolveTextGenerationRequest, resolveChatTemplateSetting } from
|
|
4
|
-
import { isGpuBufferInstance, isWeightBuffer, isCpuWeightBuffer, getBufferDtype } from
|
|
5
|
-
import { decodeReadback } from
|
|
6
|
-
import { resolveExecutionSessionPlan } from
|
|
7
|
-
import { selectRuleValue } from
|
|
8
|
-
import { assertSpeculativeCapability } from
|
|
1
|
+
import { readBuffer } from "../../../../../memory/buffer-pool.js";
|
|
2
|
+
import { resolveSamplingConfig } from "../sampling/sampling-config.js";
|
|
3
|
+
import { resolveTextGenerationRequest, resolveChatTemplateSetting } from "../request/generation-request.js";
|
|
4
|
+
import { isGpuBufferInstance, isWeightBuffer, isCpuWeightBuffer, getBufferDtype } from "../../../../../gpu/weight-buffer.js";
|
|
5
|
+
import { decodeReadback } from "../../debug-utils/readback-decoding.js";
|
|
6
|
+
import { resolveExecutionSessionPlan } from "../../execution/execution-plan.js";
|
|
7
|
+
import { selectRuleValue } from "../../../../../rules/rule-registry.js";
|
|
8
|
+
import { assertSpeculativeCapability } from "./speculative-runtime.js";
|
|
9
9
|
|
|
10
10
|
const UNKNOWN_TOKENIZER_VOCAB_SIZE = 'unknown';
|
|
11
11
|
const INVALID_DTYPE_SENTINEL = '__invalid_dtype__';
|
|
@@ -1,5 +1,5 @@
|
|
|
1
|
-
import { log, trace, isTraceEnabled } from
|
|
2
|
-
import { createKVCache, isStopToken } from
|
|
1
|
+
import { log, trace, isTraceEnabled } from "../../../../../debug/index.js";
|
|
2
|
+
import { createKVCache, isStopToken } from "../../init.js";
|
|
3
3
|
import {
|
|
4
4
|
decodeStep,
|
|
5
5
|
decodeStepLogits,
|
|
@@ -14,7 +14,7 @@ import {
|
|
|
14
14
|
advanceDecodeStepCount,
|
|
15
15
|
createTsirFixtureState,
|
|
16
16
|
drainPendingTsirReads,
|
|
17
|
-
} from
|
|
17
|
+
} from "../../observation/tsir-fixture-writer.js";
|
|
18
18
|
import {
|
|
19
19
|
activateFallbackExecutionPlan,
|
|
20
20
|
hasFallbackExecutionPlan,
|
|
@@ -24,13 +24,13 @@ import {
|
|
|
24
24
|
resolvePrefillRecorderChunkLayers,
|
|
25
25
|
resolveActiveExecutionPlan,
|
|
26
26
|
setActiveExecutionPlan,
|
|
27
|
-
} from
|
|
27
|
+
} from "../../execution/execution-plan.js";
|
|
28
28
|
import {
|
|
29
29
|
cloneLinearAttentionRuntime,
|
|
30
30
|
hasLinearAttentionLayers,
|
|
31
31
|
resetLinearAttentionRuntime,
|
|
32
32
|
restoreLinearAttentionRuntime,
|
|
33
|
-
} from
|
|
33
|
+
} from "../../linear-attention/linear-attention.js";
|
|
34
34
|
import {
|
|
35
35
|
shouldDisableBatchDecodeAfterShortBatch,
|
|
36
36
|
resolveHotVocabularyBatchDecodeAvailability,
|
|
@@ -45,7 +45,7 @@ import {
|
|
|
45
45
|
resolveTargetPlanKVDtype,
|
|
46
46
|
resolveCurrentKVCacheDtype,
|
|
47
47
|
cloneRuntimeInferenceWithKVDtype,
|
|
48
|
-
} from
|
|
48
|
+
} from "../../policy/generator-decode-policy.js";
|
|
49
49
|
|
|
50
50
|
export async function _replayPrefillDecodeLogits(currentIds, opts) {
|
|
51
51
|
// Guard: cap replay-prefill sequence length to the config-owned maxSeqLen.
|
package/src/inference/pipelines/text/{sequence-state.d.ts → generator/session/sequence-state.d.ts}
RENAMED
|
@@ -1,6 +1,6 @@
|
|
|
1
|
-
import type { PipelineState } from
|
|
2
|
-
import type { AnyLayerPartitionPlan } from
|
|
3
|
-
export type ContextOptions = import('./generator-runtime.js').GenerateOptionsResolved & import(
|
|
1
|
+
import type { PipelineState } from "../../state.js";
|
|
2
|
+
import type { AnyLayerPartitionPlan } from "../../execution/layer-partition-contract.js";
|
|
3
|
+
export type ContextOptions = import('./generator-runtime.js').GenerateOptionsResolved & import("../../types.js").GenerateOptions;
|
|
4
4
|
export type ContextPipelineState = PipelineState & {
|
|
5
5
|
modelPartition?: { plan: AnyLayerPartitionPlan; index: number } | null;
|
|
6
6
|
};
|
|
@@ -49,7 +49,7 @@ export function closeGenerationContexts(state: ContextPipelineState): void;
|
|
|
49
49
|
export function beginSequenceMemoryBoundary(): Promise<boolean>;
|
|
50
50
|
export function beginGenerationContext(state: ContextPipelineState, context: GenerationContext | undefined, options: ContextOptions): void;
|
|
51
51
|
export function generationContextInput(state: ContextPipelineState, inputIds: number[]): number[];
|
|
52
|
-
export function generationContextCheckpointLength(state: ContextPipelineState, prompt: import(
|
|
52
|
+
export function generationContextCheckpointLength(state: ContextPipelineState, prompt: import("../generator.js").PromptInput,
|
|
53
53
|
inputIds: number[], options: ContextOptions): number | null;
|
|
54
54
|
export function checkpointGenerationContext(state: ContextPipelineState, tokens: number[]): Promise<void>;
|
|
55
55
|
export function prepareGenerationContextPrefill(generator: {
|
|
@@ -57,7 +57,7 @@ export function prepareGenerationContextPrefill(generator: {
|
|
|
57
57
|
_resolvePrefillTokenChunkSize(ids: number[]): number | null;
|
|
58
58
|
_prefillToHidden(ids: number[], options: ContextOptions): Promise<unknown>;
|
|
59
59
|
_commitPrefillHiddenChunk(value: unknown): Promise<void>;
|
|
60
|
-
}, prompt: import(
|
|
60
|
+
}, prompt: import("../generator.js").PromptInput, inputIds: number[], options: ContextOptions): Promise<number | null>;
|
|
61
61
|
export function finishGenerationContext(state: ContextPipelineState, generatedIds: number[], completed: boolean): Promise<void>;
|
|
62
62
|
|
|
63
63
|
export interface SequenceState {
|
package/src/inference/pipelines/text/{sequence-state.js → generator/session/sequence-state.js}
RENAMED
|
@@ -1,14 +1,14 @@
|
|
|
1
|
-
import { getDevice } from
|
|
2
|
-
import { acquireBuffer, releaseBuffer } from
|
|
3
|
-
import { retireIdleBuffersUnderDeviceBudget } from
|
|
4
|
-
import { createKVCache } from
|
|
5
|
-
import { createLinearAttentionRuntime, resetLinearAttentionRuntime, resetLinearAttentionSequence } from
|
|
6
|
-
import { isConvLayerType } from
|
|
7
|
-
import { resolveLayerPartition } from
|
|
8
|
-
import { withHyperConnectionTransaction } from
|
|
9
|
-
import { KVCache } from
|
|
10
|
-
import { cloneLinearAttentionRuntimeOnGPU } from
|
|
11
|
-
import { resolvePromptInput } from
|
|
1
|
+
import { getDevice } from "../../../../../gpu/device.js";
|
|
2
|
+
import { acquireBuffer, releaseBuffer } from "../../../../../memory/buffer-pool.js";
|
|
3
|
+
import { retireIdleBuffersUnderDeviceBudget } from "../../../../../memory/sequence-retirement.js";
|
|
4
|
+
import { createKVCache } from "../../init.js";
|
|
5
|
+
import { createLinearAttentionRuntime, resetLinearAttentionRuntime, resetLinearAttentionSequence } from "../../linear-attention/linear-attention.js";
|
|
6
|
+
import { isConvLayerType } from "../../execution/layer-input-execution.js";
|
|
7
|
+
import { resolveLayerPartition } from "../../execution/layer-partition-contract.js";
|
|
8
|
+
import { withHyperConnectionTransaction } from "../../hyper-connection/sequence-transaction.js";
|
|
9
|
+
import { KVCache } from "../../../../kv-cache/base.js";
|
|
10
|
+
import { cloneLinearAttentionRuntimeOnGPU } from "../../linear-attention/state-clone.js";
|
|
11
|
+
import { resolvePromptInput } from "../request/prompt-input.js";
|
|
12
12
|
|
|
13
13
|
// Handles are local to one loaded pipeline. Applications own naming, admission,
|
|
14
14
|
// retention and eviction; this module owns complete decoder memory and settlement.
|
|
@@ -53,7 +53,7 @@ function releaseHyperSnapshot(hyper) {
|
|
|
53
53
|
}
|
|
54
54
|
}
|
|
55
55
|
|
|
56
|
-
/** @param {import(
|
|
56
|
+
/** @param {import("../../state.js").PipelineState['convLayerStates']} convLayerStates */
|
|
57
57
|
async function snapshotConvOnGPU(convLayerStates) {
|
|
58
58
|
if (!convLayerStates?.size) return null;
|
|
59
59
|
const device = getDevice();
|
|
@@ -80,7 +80,7 @@ async function snapshotConvOnGPU(convLayerStates) {
|
|
|
80
80
|
}
|
|
81
81
|
}
|
|
82
82
|
|
|
83
|
-
/** @param {import(
|
|
83
|
+
/** @param {import("../../state.js").PipelineState['convLayerStates']} convLayerStates
|
|
84
84
|
* @param {Map<number, GPUBuffer> | null | undefined} snapshot */
|
|
85
85
|
async function restoreConvOnGPU(convLayerStates, snapshot) {
|
|
86
86
|
if (!convLayerStates?.size || !snapshot?.size) return;
|
|
@@ -96,7 +96,7 @@ async function restoreConvOnGPU(convLayerStates, snapshot) {
|
|
|
96
96
|
await device.queue.onSubmittedWorkDone();
|
|
97
97
|
}
|
|
98
98
|
|
|
99
|
-
/** @param {import(
|
|
99
|
+
/** @param {import("../../state.js").PipelineState['hyperConnectionSession']} session */
|
|
100
100
|
async function snapshotHyperConnectionOnGPU(session) {
|
|
101
101
|
if (!session) return null;
|
|
102
102
|
const device = getDevice();
|
|
@@ -142,7 +142,7 @@ async function snapshotHyperConnectionOnGPU(session) {
|
|
|
142
142
|
}
|
|
143
143
|
}
|
|
144
144
|
|
|
145
|
-
/** @param {import(
|
|
145
|
+
/** @param {import("../../state.js").PipelineState['hyperConnectionSession']} session
|
|
146
146
|
* @param {import('./sequence-state.js').HyperConnectionSnapshot | null | undefined} snapshot
|
|
147
147
|
* @param {readonly number[]} tokens */
|
|
148
148
|
async function restoreHyperConnectionOnGPU(session, snapshot, tokens) {
|
|
@@ -237,8 +237,8 @@ function release(entry) {
|
|
|
237
237
|
entry.signature = null;
|
|
238
238
|
}
|
|
239
239
|
|
|
240
|
-
/** @param {NonNullable<import(
|
|
241
|
-
* @param {import(
|
|
240
|
+
/** @param {NonNullable<import("../../state.js").PipelineState['kvCache']>} cache
|
|
241
|
+
* @param {import("../../state.js").PipelineState['linearAttentionRuntime']} linear
|
|
242
242
|
* @param {Map<number, GPUBuffer> | null} conv
|
|
243
243
|
* @param {import('./sequence-state.js').HyperConnectionSnapshot | null} hyper
|
|
244
244
|
* @param {number} tokenCount */
|
|
@@ -301,7 +301,15 @@ export function beginGenerationContext(state, context, options) {
|
|
|
301
301
|
// change the representation of the already validated token prefix.
|
|
302
302
|
const signature = JSON.stringify({ runtime: state.runtimeConfig, model: state.modelConfig,
|
|
303
303
|
partition: state.modelPartition ?? null,
|
|
304
|
-
plan: options.executionPlan, speculation: options.speculation, adapter: state.lora?.identity?.digest ?? null }
|
|
304
|
+
plan: options.executionPlan, speculation: options.speculation, adapter: state.lora?.identity?.digest ?? null },
|
|
305
|
+
(_key, value) => {
|
|
306
|
+
// Resolved contracts can contain exact 64-bit integers. Encode them without
|
|
307
|
+
// rounding; tag strings too so a literal string cannot alias an integer.
|
|
308
|
+
// This signature is local cache state, not a serialized execution contract.
|
|
309
|
+
if (typeof value === 'bigint') return `bigint:${value}`;
|
|
310
|
+
if (typeof value === 'string') return `string:${value}`;
|
|
311
|
+
return value;
|
|
312
|
+
});
|
|
305
313
|
if (entry.signature !== signature || (state.lora && !state.lora.identity?.digest && entry.adapter !== state.lora)) release(entry);
|
|
306
314
|
entry.reusedTokens = 0;
|
|
307
315
|
const estimated = state.kvCache.getMemoryStats().allocated;
|
package/src/inference/pipelines/text/generator/{session-context.js → session/session-context.js}
RENAMED
|
@@ -1,17 +1,17 @@
|
|
|
1
|
-
import { snapshotElementBytes } from
|
|
2
|
-
import { mergeRuntimeValues } from
|
|
1
|
+
import { snapshotElementBytes } from "../../../../../debug/tensor.js";
|
|
2
|
+
import { mergeRuntimeValues } from "../../../../../config/runtime-merge.js";
|
|
3
3
|
import {
|
|
4
4
|
assertBundledAdapterAuthorized,
|
|
5
5
|
assertKnownResolutionNotRevoked,
|
|
6
|
-
} from
|
|
7
|
-
import { log } from
|
|
8
|
-
import { allowReadback } from
|
|
9
|
-
import { readBuffer } from
|
|
10
|
-
import { selectRuleValue } from
|
|
11
|
-
import { isPlainObject } from
|
|
12
|
-
import { decodeReadback } from
|
|
13
|
-
import { resolveActiveExecutionPlan } from
|
|
14
|
-
import { resolveRangeAwareSelectiveWideningConfig } from
|
|
6
|
+
} from "../../../../../config/revocation-policy.js";
|
|
7
|
+
import { log } from "../../../../../debug/index.js";
|
|
8
|
+
import { allowReadback } from "../../../../../debug/perf-guards.js";
|
|
9
|
+
import { readBuffer } from "../../../../../memory/buffer-pool.js";
|
|
10
|
+
import { selectRuleValue } from "../../../../../rules/rule-registry.js";
|
|
11
|
+
import { isPlainObject } from "../../../../../formats/plain-object.js";
|
|
12
|
+
import { decodeReadback } from "../../debug-utils/readback-decoding.js";
|
|
13
|
+
import { resolveActiveExecutionPlan } from "../../execution/execution-plan.js";
|
|
14
|
+
import { resolveRangeAwareSelectiveWideningConfig } from "../../policy/finiteness-policy.js";
|
|
15
15
|
|
|
16
16
|
export function resolvePerLayerInputsSession(manifestSession, runtimeSession) {
|
|
17
17
|
if (!isPlainObject(runtimeSession)) {
|
|
@@ -164,6 +164,7 @@ export function buildLayerContext(
|
|
|
164
164
|
layerRouterWeights: state.layerRouterWeights,
|
|
165
165
|
recorder,
|
|
166
166
|
lora: state.lora,
|
|
167
|
+
channelGainRequest: state.channelGainRequest,
|
|
167
168
|
decodeBuffers: isDecodeMode && state.decodeBuffers?.hasBuffers() ? state.decodeBuffers : null,
|
|
168
169
|
runtimeComputeConfig: effectiveComputeConfig,
|
|
169
170
|
runtimeMoERoutingConfig: config.useMoE ? { ...state.runtimeConfig.inference.moe?.routing } : undefined,
|
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import { checkpointLinearAttentionRuntime } from
|
|
1
|
+
import { checkpointLinearAttentionRuntime } from "../../linear-attention/state-checkpoint.js";
|
|
2
2
|
|
|
3
3
|
/** @type {import('./speculative-runtime.js').lookupDraft} */
|
|
4
4
|
export function lookupDraft(context, maxTokens, maxMatchTokens, minMatchTokens) {
|
|
@@ -1,22 +1,22 @@
|
|
|
1
1
|
import { mergeRecordOpLabelCounts } from './recorder-stats.js';
|
|
2
2
|
export { mergeRecordOpLabelCounts } from './recorder-stats.js';
|
|
3
|
-
import
|
|
4
|
-
import { releaseBuffer, readBuffer } from
|
|
5
|
-
import { trace, isTraceEnabled } from
|
|
3
|
+
import "../../../../../gpu/device.js"; // Preserve declared device initialization side effects.
|
|
4
|
+
import { releaseBuffer, readBuffer } from "../../../../../memory/buffer-pool.js";
|
|
5
|
+
import { trace, isTraceEnabled } from "../../../../../debug/index.js";
|
|
6
6
|
import {
|
|
7
7
|
isWeightBuffer,
|
|
8
8
|
isGpuBufferInstance,
|
|
9
9
|
isSplitWeightBuffer,
|
|
10
10
|
getWeightDtype,
|
|
11
11
|
getLayout,
|
|
12
|
-
} from
|
|
13
|
-
import { decodeReadback } from
|
|
14
|
-
import { getLogitsHealth } from
|
|
15
|
-
import { tensorElementBytes } from
|
|
12
|
+
} from "../../../../../gpu/weight-buffer.js";
|
|
13
|
+
import { decodeReadback } from "../../debug-utils/readback-decoding.js";
|
|
14
|
+
import { getLogitsHealth } from "../../debug-utils/health-metrics.js";
|
|
15
|
+
import { tensorElementBytes } from "../../../../../gpu/tensor.js";
|
|
16
16
|
import {
|
|
17
17
|
getWeightBuffer as getPipelineWeightBuffer,
|
|
18
18
|
getNormWeightBuffer as getPipelineNormWeightBuffer,
|
|
19
|
-
} from
|
|
19
|
+
} from "../../loading/weights.js";
|
|
20
20
|
|
|
21
21
|
export const SPECIAL_LIKE_TOKEN_RE = /^(<pad>|<unused\d*>|<eos>|<bos>|<s>|<\/s>|\[PAD\]|\[UNK\]|\[SEP\]|\[CLS\]|<[^>\n]{1,32}>)$/i;
|
|
22
22
|
|
|
@@ -311,6 +311,6 @@ export function normalizeSelfConditioningSoftEmbeddingState(state, canvasLength,
|
|
|
311
311
|
export let intentBundleModulePromise = null;
|
|
312
312
|
|
|
313
313
|
export async function getExperimentalIntentBundleModule() {
|
|
314
|
-
intentBundleModulePromise ??= import(
|
|
314
|
+
intentBundleModulePromise ??= import("../../../../../experimental/hotswap/intent-bundle.js");
|
|
315
315
|
return intentBundleModulePromise;
|
|
316
316
|
}
|
package/src/inference/pipelines/text/{warm-generation.d.ts → generator/session/warm-generation.d.ts}
RENAMED
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import type { KernelWarmupGenerationSchema } from
|
|
1
|
+
import type { KernelWarmupGenerationSchema } from "../../../../../config/schema/execution/kernel-warmup.schema.js";
|
|
2
2
|
|
|
3
3
|
export interface WarmGenerationResult {
|
|
4
4
|
generatedTokens: number;
|
package/src/inference/pipelines/text/{warm-generation.js → generator/session/warm-generation.js}
RENAMED
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import { log } from
|
|
1
|
+
import { log } from "../../../../../debug/index.js";
|
|
2
2
|
|
|
3
3
|
// runtime.shared.kernelWarmup.warmGeneration: one greedy generation at the end of a text
|
|
4
4
|
// pipeline load. It compiles the pipelines this model's plan binds, and with
|
|
@@ -1,9 +1,9 @@
|
|
|
1
|
-
import { getDevice } from
|
|
2
|
-
import { isGreedySampling } from
|
|
3
|
-
import { FINITENESS_RESET_WORDS, FinitenessError, parseFinitenessStatusWords } from
|
|
4
|
-
import { advanceDecodeStepCount } from
|
|
5
|
-
import { queueLinearAttentionCheckpoint } from
|
|
6
|
-
import { resolveActiveExecutionPlan } from
|
|
1
|
+
import { getDevice } from "../../../../../gpu/device.js";
|
|
2
|
+
import { isGreedySampling } from "../../../../../rules/greedy-sampling.js";
|
|
3
|
+
import { FINITENESS_RESET_WORDS, FinitenessError, parseFinitenessStatusWords } from "../../policy/finiteness-guard-status.js";
|
|
4
|
+
import { advanceDecodeStepCount } from "../../observation/tsir-fixture-writer.js";
|
|
5
|
+
import { queueLinearAttentionCheckpoint } from "../../linear-attention/state-checkpoint.js";
|
|
6
|
+
import { resolveActiveExecutionPlan } from "../../execution/execution-plan.js";
|
|
7
7
|
import { WholeTokenProgram } from './whole-token-program.js';
|
|
8
8
|
|
|
9
9
|
// Whole-token program lane of single-token decode (decodeLoop.wholeTokenProgram):
|
|
@@ -17,7 +17,7 @@ import { WholeTokenProgram } from './whole-token-program.js';
|
|
|
17
17
|
// (verifyWholeTokenProgramAhead); a match replays step 1 itself. See
|
|
18
18
|
// whole-token-program.js.
|
|
19
19
|
|
|
20
|
-
/** @typedef {import(
|
|
20
|
+
/** @typedef {import("../../state.js").PipelineState} PipelineState */
|
|
21
21
|
/** @typedef {{ seqLen?: unknown, keysGPU?: GPUBuffer | null, valuesGPU?: GPUBuffer | null }} KVLayerView */
|
|
22
22
|
/** @typedef {{ seqLen?: unknown, convStateGPU?: GPUBuffer | null, recurrentStateGPU?: GPUBuffer | null }} LinearLayerView */
|
|
23
23
|
/**
|
|
@@ -191,9 +191,9 @@ function kvCapacityLimits(state) {
|
|
|
191
191
|
/**
|
|
192
192
|
* One decode step of the lane. `runRecorded` runs the ordinary recorded step; given
|
|
193
193
|
* an `attach` function, it hands that the step's recorder first.
|
|
194
|
-
* @param {import(
|
|
194
|
+
* @param {import("../../state.js").PipelineState} state
|
|
195
195
|
* @param {number[]} currentIds
|
|
196
|
-
* @param {(attach: ((recorder: import(
|
|
196
|
+
* @param {(attach: ((recorder: import("../../../../../gpu/command-recorder.js").CommandRecorder) => void) | null) => Promise<number>} runRecorded
|
|
197
197
|
* @returns {Promise<number>}
|
|
198
198
|
*/
|
|
199
199
|
export async function decodeWholeTokenStep(state, currentIds, runRecorded) {
|
|
@@ -368,4 +368,4 @@ export async function prepareWholeTokenBatch(state, currentIds, requestedCount)
|
|
|
368
368
|
}
|
|
369
369
|
// Recorded generation mechanisms share this lane owner; each caller supplies its
|
|
370
370
|
// declared execution program and retains exclusive model/cache ownership.
|
|
371
|
-
export { createCompiledGreedyWindow } from
|
|
371
|
+
export { createCompiledGreedyWindow } from "../decode/compiled-greedy-window.js";
|
|
@@ -1,5 +1,5 @@
|
|
|
1
|
-
import type { CommandRecorder } from
|
|
2
|
-
import type { StepContractDeclaration } from
|
|
1
|
+
import type { CommandRecorder } from "../../../../../gpu/command-recorder.js";
|
|
2
|
+
import type { StepContractDeclaration } from "../../../../../gpu/step-contract.js";
|
|
3
3
|
|
|
4
4
|
/** A u32 word as a function of the decode position and input token. */
|
|
5
5
|
export type WordModel =
|
|
@@ -1,6 +1,6 @@
|
|
|
1
|
-
import { peekUniformCacheBytes } from
|
|
2
|
-
import { observeStepContract } from
|
|
3
|
-
import { getBufferPool, isBufferActive, isPersistentBuffer } from
|
|
1
|
+
import { peekUniformCacheBytes } from "../../../../../gpu/uniform-cache.js";
|
|
2
|
+
import { observeStepContract } from "../../../../../gpu/step-contract.js";
|
|
3
|
+
import { getBufferPool, isBufferActive, isPersistentBuffer } from "../../../../../memory/buffer-pool.js";
|
|
4
4
|
|
|
5
5
|
// Observation of one ordinary recorded decode step for the whole-token program
|
|
6
6
|
// (whole-token-program.js): every GPU command the step records (dispatches, bind
|
|
@@ -10,7 +10,7 @@ import { getBufferPool, isBufferActive, isPersistentBuffer } from '../../../../m
|
|
|
10
10
|
// input token id, and the step positions over which their commands stay fixed or
|
|
11
11
|
// affine in the position.
|
|
12
12
|
|
|
13
|
-
/** @typedef {import(
|
|
13
|
+
/** @typedef {import("../../../../../gpu/command-recorder.js").CommandRecorder} CommandRecorder */
|
|
14
14
|
|
|
15
15
|
/**
|
|
16
16
|
* A u32 word as a function of the decode position and input token.
|
|
@@ -41,7 +41,7 @@ import { getBufferPool, isBufferActive, isPersistentBuffer } from '../../../../m
|
|
|
41
41
|
/** @typedef {{ kind: 'copy', src: GPUBuffer, srcOffset: number, dst: GPUBuffer, dstOffset: number, size: number }} CopyOp */
|
|
42
42
|
/** @typedef {{ kind: 'clear', buffer: GPUBuffer, offset: number, size: number }} ClearOp */
|
|
43
43
|
/** @typedef {DispatchOp | IndirectOp | WriteOp | CopyOp | ClearOp} ObservedOp */
|
|
44
|
-
/** @typedef {import(
|
|
44
|
+
/** @typedef {import("../../../../../gpu/step-contract.js").StepContractDeclaration} StepContractDeclaration */
|
|
45
45
|
/** A declared range of step positions (inclusive) and the operation that declared it. @typedef {{ owner: string, min: number, max: number }} PositionLimit */
|
|
46
46
|
/** @typedef {{ uniform: Uint8Array, known: Uint8Array }} UniformContents */
|
|
47
47
|
|
|
@@ -44,7 +44,7 @@ export declare class WholeTokenProgram {
|
|
|
44
44
|
observe<T>(
|
|
45
45
|
pos: number,
|
|
46
46
|
token: number,
|
|
47
|
-
run: (attach: (recorder: import(
|
|
47
|
+
run: (attach: (recorder: import("../../../../../gpu/command-recorder.js").CommandRecorder | null) => void) => Promise<T>,
|
|
48
48
|
identityOf: () => readonly unknown[],
|
|
49
49
|
limits?: readonly import('./whole-token-observation.js').PositionLimit[]
|
|
50
50
|
): Promise<T>;
|
|
@@ -58,7 +58,7 @@ export declare class WholeTokenProgram {
|
|
|
58
58
|
verify<T>(
|
|
59
59
|
pos: number,
|
|
60
60
|
token: number,
|
|
61
|
-
run: (attach: (recorder: import(
|
|
61
|
+
run: (attach: (recorder: import("../../../../../gpu/command-recorder.js").CommandRecorder | null) => void) => Promise<T>,
|
|
62
62
|
identityOf: () => readonly unknown[],
|
|
63
63
|
limits?: readonly import('./whole-token-observation.js').PositionLimit[]
|
|
64
64
|
): Promise<T>;
|
|
@@ -70,7 +70,7 @@ export declare class WholeTokenProgram {
|
|
|
70
70
|
verifyAhead(
|
|
71
71
|
pos: number,
|
|
72
72
|
token: number,
|
|
73
|
-
run: (attach: (recorder: import(
|
|
73
|
+
run: (attach: (recorder: import("../../../../../gpu/command-recorder.js").CommandRecorder | null) => void) => Promise<unknown>,
|
|
74
74
|
identityOf: () => readonly unknown[],
|
|
75
75
|
limits: readonly import('./whole-token-observation.js').PositionLimit[],
|
|
76
76
|
stopSignal: unknown
|
|
@@ -1,9 +1,9 @@
|
|
|
1
|
-
import { isBufferActive } from
|
|
2
|
-
import { log } from
|
|
1
|
+
import { isBufferActive } from "../../../../../memory/buffer-pool.js";
|
|
2
|
+
import { log } from "../../../../../debug/index.js";
|
|
3
3
|
import { StepObservation, observeStep, patchMethod, restorePatches, U32 } from './whole-token-observation.js';
|
|
4
4
|
import { affineLastPosition, buildTemplate } from './whole-token-template.js';
|
|
5
|
-
import { encodeRecordedStep, writeRecordedStepInputs } from
|
|
6
|
-
import { isNativeGpuBuffer } from
|
|
5
|
+
import { encodeRecordedStep, writeRecordedStepInputs } from "../decode/recorded-step-executor.js";
|
|
6
|
+
import { isNativeGpuBuffer } from "../../../../../gpu/buffer-identity.js";
|
|
7
7
|
|
|
8
8
|
|
|
9
9
|
// Whole-token decode program (runtime.inference.session.decodeLoop.wholeTokenProgram).
|
|
@@ -37,7 +37,7 @@ import { isNativeGpuBuffer } from '../../../../gpu/buffer-identity.js';
|
|
|
37
37
|
// buffers and dispatch order.
|
|
38
38
|
|
|
39
39
|
|
|
40
|
-
/** @typedef {import(
|
|
40
|
+
/** @typedef {import("../../../../../gpu/command-recorder.js").CommandRecorder} CommandRecorder */
|
|
41
41
|
/** @typedef {import('./whole-token-observation.js').WordModel} WordModel */
|
|
42
42
|
/** @typedef {import('./whole-token-observation.js').DynamicWord} DynamicWord */
|
|
43
43
|
/** @typedef {import('./whole-token-observation.js').BindGroupDescriptors} BindGroupDescriptors */
|
|
@@ -1,6 +1,8 @@
|
|
|
1
1
|
import type { CommandRecorder } from '../../../../gpu/command-recorder.js';
|
|
2
2
|
import type { Tensor } from '../../../../gpu/tensor.js';
|
|
3
|
-
import type { ResolvedHyperConnections } from
|
|
3
|
+
import type { ResolvedHyperConnections } from "../../../../config/schema/execution/hyper-connection-decoder.schema.js";
|
|
4
|
+
|
|
5
|
+
export type RuntimeHyperConnections = ResolvedHyperConnections & { normWeightOffset: boolean };
|
|
4
6
|
|
|
5
7
|
/** Caller-owned projection; roles are `${attn_hc|mlp_hc|hc_mixer}_{down|up|inject}`. */
|
|
6
8
|
export type HyperConnectionProjection = (
|
|
@@ -23,9 +25,15 @@ export interface GatedResidualWeights {
|
|
|
23
25
|
export interface GatedResidualInput {
|
|
24
26
|
streams: Tensor;
|
|
25
27
|
weights: GatedResidualWeights;
|
|
26
|
-
hyper:
|
|
28
|
+
hyper: RuntimeHyperConnections;
|
|
27
29
|
numTokens: number;
|
|
28
30
|
project: HyperConnectionProjection;
|
|
31
|
+
/** Resolved explicit hc_mix_branch declaration. */
|
|
32
|
+
mixBranch?: boolean;
|
|
33
|
+
/** Explicit Q4_0 up+mix+branch declaration for this attention/MLP site. */
|
|
34
|
+
upMixBranch?: boolean;
|
|
35
|
+
/** Required for prepared single-token output reuse; root mixer uses layer 0. */
|
|
36
|
+
layerIdx?: number;
|
|
29
37
|
role: string;
|
|
30
38
|
}
|
|
31
39
|
|
|
@@ -39,7 +47,7 @@ export interface HyperConnectionDecoderLayerInput {
|
|
|
39
47
|
/** Caller-owned streams [numTokens, streamCount * streamWidth]; not released. */
|
|
40
48
|
streams: Tensor;
|
|
41
49
|
numTokens: number;
|
|
42
|
-
hyper:
|
|
50
|
+
hyper: RuntimeHyperConnections;
|
|
43
51
|
weights: { attention: GatedResidualWeights; mlp: GatedResidualWeights };
|
|
44
52
|
/** PLE on the incoming streams for PLE layers; returns new streams owned by the layer. */
|
|
45
53
|
ple?: ((streams: Tensor) => Promise<Tensor>) | null;
|
|
@@ -48,12 +56,17 @@ export interface HyperConnectionDecoderLayerInput {
|
|
|
48
56
|
/** MoE on the branch input [numTokens, streamWidth]. */
|
|
49
57
|
mlpBranch: (branchInput: Tensor) => Promise<Tensor>;
|
|
50
58
|
project: HyperConnectionProjection;
|
|
59
|
+
/** Resolved explicit hc_mix_branch declaration. */
|
|
60
|
+
mixBranch?: boolean;
|
|
61
|
+
upMixBranch?: { attention: boolean; mlp: boolean };
|
|
62
|
+
/** Required for prepared single-token output reuse; root mixer uses layer 0. */
|
|
63
|
+
layerIdx?: number;
|
|
51
64
|
}
|
|
52
65
|
|
|
53
66
|
export declare function replicateEmbeddings(
|
|
54
67
|
target: CommandRecorder | null,
|
|
55
68
|
embeds: Tensor,
|
|
56
|
-
options: { hyper:
|
|
69
|
+
options: { hyper: RuntimeHyperConnections; numTokens: number }
|
|
57
70
|
): Promise<Tensor>;
|
|
58
71
|
|
|
59
72
|
/** Branch input [numTokens, streamWidth] and injection gate [numTokens, streamCount] (null without inject weights). */
|
|
@@ -1,8 +1,9 @@
|
|
|
1
|
+
import { recordHyperConnectionUpMixBranch, runHyperConnectionUpMixBranch } from '../../../../gpu/kernels/hyper-connection-up.js';
|
|
1
2
|
import { releaseBuffer } from '../../../../memory/buffer-pool.js';
|
|
2
3
|
import {
|
|
3
|
-
recordHyperConnectionActivation, recordHyperConnectionBranch, recordHyperConnectionInject, recordHyperConnectionNorm,
|
|
4
|
+
recordHyperConnectionActivation, recordHyperConnectionBranch, recordHyperConnectionInject, recordHyperConnectionNorm, recordHyperConnectionMixBranch,
|
|
4
5
|
recordHyperConnectionReplicate, runHyperConnectionActivation, runHyperConnectionBranch, runHyperConnectionInject,
|
|
5
|
-
runHyperConnectionNorm, runHyperConnectionReplicate,
|
|
6
|
+
runHyperConnectionNorm, runHyperConnectionReplicate, runHyperConnectionMixBranch,
|
|
6
7
|
} from '../../../../gpu/kernels/hyper-connection.js';
|
|
7
8
|
|
|
8
9
|
// qwen4_exp decoder-layer composition (Qwen4ExpTextDecoderLayer.forward):
|
|
@@ -34,12 +35,14 @@ function operators(target) {
|
|
|
34
35
|
norm: recordHyperConnectionNorm.bind(null, target),
|
|
35
36
|
activate: recordHyperConnectionActivation.bind(null, target),
|
|
36
37
|
branch: recordHyperConnectionBranch.bind(null, target),
|
|
38
|
+
mixBranch: recordHyperConnectionMixBranch.bind(null, target),
|
|
37
39
|
inject: recordHyperConnectionInject.bind(null, target),
|
|
38
40
|
replicate: recordHyperConnectionReplicate.bind(null, target),
|
|
39
41
|
};
|
|
40
42
|
}
|
|
41
43
|
return {
|
|
42
44
|
norm: runHyperConnectionNorm, activate: runHyperConnectionActivation, branch: runHyperConnectionBranch,
|
|
45
|
+
mixBranch: runHyperConnectionMixBranch,
|
|
43
46
|
inject: runHyperConnectionInject, replicate: runHyperConnectionReplicate,
|
|
44
47
|
};
|
|
45
48
|
}
|
|
@@ -50,7 +53,10 @@ export function replicateEmbeddings(target, embeds, { hyper, numTokens }) {
|
|
|
50
53
|
}
|
|
51
54
|
|
|
52
55
|
/** @type {typeof import('./decoder-layer.js').runGatedResidual} */
|
|
53
|
-
export async function runGatedResidual(target, { streams, weights, hyper, numTokens, project, role }) {
|
|
56
|
+
export async function runGatedResidual(target, { streams, weights, hyper, numTokens, project, role, mixBranch = false, upMixBranch = false, layerIdx }) {
|
|
57
|
+
if (upMixBranch && (role !== 'attn_hc' && role !== 'mlp_hc')) {
|
|
58
|
+
throw new Error('[HyperConnection] Up-mix fusion is limited to declared attention/MLP sites, never the root mixer.');
|
|
59
|
+
}
|
|
54
60
|
const ops = operators(target);
|
|
55
61
|
const { streamCount, streamWidth, projectionRank, normEps } = hyper;
|
|
56
62
|
const width = streamCount * streamWidth;
|
|
@@ -59,12 +65,29 @@ export async function runGatedResidual(target, { streams, weights, hyper, numTok
|
|
|
59
65
|
const temporaries = [];
|
|
60
66
|
const keep = (/** @type {import('../../../../gpu/tensor.js').Tensor} */ tensor) => { temporaries.push(tensor); return tensor; };
|
|
61
67
|
try {
|
|
62
|
-
const normed = keep(await ops.norm(streams, weights.norm, { ...geometry, eps: normEps }));
|
|
68
|
+
const normed = keep(await ops.norm(streams, weights.norm, { ...geometry, eps: normEps, normWeightOffset: hyper.normWeightOffset }));
|
|
63
69
|
const down = keep(await project(`${role}_down`, normed, weights.down, projectionRank, width));
|
|
64
70
|
const lowRank = keep(await ops.activate(down, { mode: 'low_rank_silu', rows: numTokens, columns: projectionRank, streamCount }));
|
|
65
|
-
const
|
|
66
|
-
|
|
67
|
-
|
|
71
|
+
const workspace = (mixBranch || upMixBranch) && numTokens === 1 && isRecorder(target) ? target.decodeWorkspace : null;
|
|
72
|
+
if (workspace && (!Number.isSafeInteger(layerIdx) || /** @type {number} */ (layerIdx) < 0)) {
|
|
73
|
+
throw new Error('[HyperConnection] Prepared mix-branch output requires an explicit layer index.');
|
|
74
|
+
}
|
|
75
|
+
const branchOptions = workspace
|
|
76
|
+
? { ...geometry, outputBuffer: workspace.getBuffer(`hc:${layerIdx}:${role}:branch`, streamWidth * 4) }
|
|
77
|
+
: geometry;
|
|
78
|
+
let branchInput;
|
|
79
|
+
if (upMixBranch) {
|
|
80
|
+
const upBranch = isRecorder(target)
|
|
81
|
+
? recordHyperConnectionUpMixBranch.bind(null, target) : runHyperConnectionUpMixBranch;
|
|
82
|
+
branchInput = await upBranch(lowRank, weights.up, normed, { ...branchOptions, projectionRank });
|
|
83
|
+
} else {
|
|
84
|
+
const up = keep(await project(`${role}_up`, lowRank, weights.up, width, projectionRank));
|
|
85
|
+
if (mixBranch) branchInput = await ops.mixBranch(up, normed, branchOptions);
|
|
86
|
+
else {
|
|
87
|
+
const mix = keep(await ops.activate(up, { mode: 'mix_sigmoid', rows: numTokens, columns: width, streamCount }));
|
|
88
|
+
branchInput = await ops.branch(mix, normed, branchOptions);
|
|
89
|
+
}
|
|
90
|
+
}
|
|
68
91
|
if (!weights.inject) return { branchInput, gate: null };
|
|
69
92
|
try {
|
|
70
93
|
const logits = keep(await project(`${role}_inject`, normed, weights.inject, streamCount, width));
|
|
@@ -85,8 +108,8 @@ export async function runGatedResidual(target, { streams, weights, hyper, numTok
|
|
|
85
108
|
* @param {import('../../../../gpu/command-recorder.js').CommandRecorder | null} target
|
|
86
109
|
* @param {import('./decoder-layer.js').HyperConnectionBranchStep} step
|
|
87
110
|
*/
|
|
88
|
-
async function runBranchStep(target, { streams, ownsStreams, weights, hyper, numTokens, project, role, branch }) {
|
|
89
|
-
const { branchInput, gate } = await runGatedResidual(target, { streams, weights, hyper, numTokens, project, role });
|
|
111
|
+
async function runBranchStep(target, { streams, ownsStreams, weights, hyper, numTokens, project, role, branch, mixBranch, upMixBranch, layerIdx }) {
|
|
112
|
+
const { branchInput, gate } = await runGatedResidual(target, { streams, weights, hyper, numTokens, project, role, mixBranch, upMixBranch, layerIdx });
|
|
90
113
|
/** @type {import('../../../../gpu/tensor.js').Tensor | null} */
|
|
91
114
|
let output = null;
|
|
92
115
|
try {
|
|
@@ -105,7 +128,7 @@ async function runBranchStep(target, { streams, ownsStreams, weights, hyper, num
|
|
|
105
128
|
|
|
106
129
|
/** @type {typeof import('./decoder-layer.js').runHyperConnectionDecoderLayer} */
|
|
107
130
|
export async function runHyperConnectionDecoderLayer(target, input) {
|
|
108
|
-
const { streams, numTokens, hyper, weights, ple = null, attentionBranch, mlpBranch, project } = input;
|
|
131
|
+
const { streams, numTokens, hyper, weights, ple = null, attentionBranch, mlpBranch, project, mixBranch = false, upMixBranch, layerIdx } = input;
|
|
109
132
|
let current = streams;
|
|
110
133
|
let owns = false;
|
|
111
134
|
try {
|
|
@@ -114,10 +137,10 @@ export async function runHyperConnectionDecoderLayer(target, input) {
|
|
|
114
137
|
owns = true;
|
|
115
138
|
}
|
|
116
139
|
current = await runBranchStep(target, { streams: current, ownsStreams: owns, weights: weights.attention, hyper, numTokens,
|
|
117
|
-
project, role: 'attn_hc', branch: attentionBranch });
|
|
140
|
+
project, mixBranch, upMixBranch: upMixBranch?.attention, layerIdx, role: 'attn_hc', branch: attentionBranch });
|
|
118
141
|
owns = true;
|
|
119
142
|
return await runBranchStep(target, { streams: current, ownsStreams: true, weights: weights.mlp, hyper, numTokens,
|
|
120
|
-
project, role: 'mlp_hc', branch: mlpBranch });
|
|
143
|
+
project, mixBranch, upMixBranch: upMixBranch?.mlp, layerIdx, role: 'mlp_hc', branch: mlpBranch });
|
|
121
144
|
} catch (error) {
|
|
122
145
|
if (owns && current !== streams) release(target, current);
|
|
123
146
|
throw error;
|
|
@@ -125,8 +148,9 @@ export async function runHyperConnectionDecoderLayer(target, input) {
|
|
|
125
148
|
}
|
|
126
149
|
|
|
127
150
|
/** @type {typeof import('./decoder-layer.js').runRootMixer} */
|
|
128
|
-
export async function runRootMixer(target, { streams, weights, hyper, numTokens, project }) {
|
|
151
|
+
export async function runRootMixer(target, { streams, weights, hyper, numTokens, project, mixBranch = false, upMixBranch = false, layerIdx }) {
|
|
152
|
+
if (upMixBranch) throw new Error('[HyperConnection] Q4_0 up-mix fusion cannot replace the root mixer program.');
|
|
129
153
|
if (weights.inject) throw new Error('[HyperConnection] the root mixer has no injection weights.');
|
|
130
|
-
const { branchInput } = await runGatedResidual(target, { streams, weights, hyper, numTokens, project, role: 'hc_mixer' });
|
|
154
|
+
const { branchInput } = await runGatedResidual(target, { streams, weights, hyper, numTokens, project, mixBranch, layerIdx, role: 'hc_mixer' });
|
|
131
155
|
return branchInput;
|
|
132
156
|
}
|