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/gpu/kernels/ple.js
CHANGED
|
@@ -1,9 +1,10 @@
|
|
|
1
|
+
import { selectRuleValue } from './rule-registry.js';
|
|
1
2
|
import { getDevice } from '../device.js';
|
|
2
3
|
import { createTensor } from '../tensor.js';
|
|
3
4
|
import { acquireBuffer, releaseBuffer } from '../../memory/buffer-pool.js';
|
|
4
5
|
import { unifiedKernelWrapper, withKernelOutput } from './kernel-execution.js';
|
|
5
6
|
|
|
6
|
-
// qwen4_exp PLE operators: BF16 row decode, the signed-square-root sigmoid
|
|
7
|
+
// qwen4_exp PLE operators: BF16/Q4_0 row decode, the signed-square-root sigmoid
|
|
7
8
|
// gate, and the dilated depthwise short convolution with owned state and the
|
|
8
9
|
// residual into the streams. Projections and grouped norms are separate ops.
|
|
9
10
|
|
|
@@ -41,17 +42,20 @@ function releaseFor(target, buffer) {
|
|
|
41
42
|
* @param {GPUBuffer} rows
|
|
42
43
|
* @param {import('./ple.js').PleRowDecodeOptions} options
|
|
43
44
|
*/
|
|
44
|
-
async function executeDecode(target, rows, { numTokens, embedDim }) {
|
|
45
|
+
async function executeDecode(target, rows, { numTokens, embedDim, tableDtype }) {
|
|
45
46
|
integer(numTokens, 'numTokens');
|
|
46
47
|
integer(embedDim, 'embedDim', 2);
|
|
47
|
-
|
|
48
|
+
const plan = selectRuleValue('dequant', 'pleRowDecode', { tableDtype });
|
|
49
|
+
if (!plan) fail('tableDtype must be bf16 or q4_0.');
|
|
50
|
+
if (embedDim % plan.blockElements !== 0) fail('embedDim must align to the table encoding block.');
|
|
48
51
|
const elements = numTokens * embedDim;
|
|
49
|
-
|
|
50
|
-
|
|
51
|
-
|
|
52
|
+
const bytes = elements / plan.blockElements * plan.blockBytes;
|
|
53
|
+
if (!Number.isSafeInteger(elements) || elements > 0xffffffff) fail('decoded element count exceeds u32.');
|
|
54
|
+
if (!Number.isSafeInteger(rows?.size) || rows.size < Math.ceil(bytes / 4) * 4) fail('row bytes buffer is too small.');
|
|
55
|
+
const groups = Math.ceil(elements / plan.elementsPerThread / WORKGROUP_SIZE);
|
|
52
56
|
const grid = groups > MAX_X_GROUPS ? [MAX_X_GROUPS, Math.ceil(groups / MAX_X_GROUPS), 1] : [groups, 1, 1];
|
|
53
57
|
return withKernelOutput(target, null, elements * 4, 'ple_embedding', async (output) => {
|
|
54
|
-
await unifiedKernelWrapper(
|
|
58
|
+
await unifiedKernelWrapper(plan.operation, target, plan.variant, [rows, output],
|
|
55
59
|
{ num_elements: elements, input_offset: 0, output_offset: 0, _pad: 0 }, grid, { WORKGROUP_SIZE });
|
|
56
60
|
return createTensor(output, 'f32', [numTokens, embedDim], 'ple_embedding');
|
|
57
61
|
});
|
|
@@ -0,0 +1,32 @@
|
|
|
1
|
+
// GGML Q4_0: each 32-value block is an fp16 scale and 16 packed bytes.
|
|
2
|
+
// Blocks and 160-value rows may start halfway through a storage u32.
|
|
3
|
+
override WORKGROUP_SIZE: u32 = 256u;
|
|
4
|
+
const MAX_WORKGROUPS_X: u32 = 65535u;
|
|
5
|
+
|
|
6
|
+
struct Uniforms {
|
|
7
|
+
num_elements: u32,
|
|
8
|
+
input_offset: u32,
|
|
9
|
+
output_offset: u32,
|
|
10
|
+
_pad: u32,
|
|
11
|
+
}
|
|
12
|
+
|
|
13
|
+
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
14
|
+
@group(0) @binding(1) var<storage, read> input: array<u32>;
|
|
15
|
+
@group(0) @binding(2) var<storage, read_write> output: array<f32>;
|
|
16
|
+
|
|
17
|
+
fn byte_at(offset: u32) -> u32 {
|
|
18
|
+
return (input[offset / 4u] >> ((offset % 4u) * 8u)) & 255u;
|
|
19
|
+
}
|
|
20
|
+
|
|
21
|
+
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
22
|
+
fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {
|
|
23
|
+
let i = global_id.x + global_id.y * MAX_WORKGROUPS_X * WORKGROUP_SIZE;
|
|
24
|
+
if (i >= u.num_elements) { return; }
|
|
25
|
+
let element = u.input_offset + i;
|
|
26
|
+
let base = (element / 32u) * 18u;
|
|
27
|
+
let lane = element % 32u;
|
|
28
|
+
let scale = unpack2x16float(byte_at(base) | (byte_at(base + 1u) << 8u)).x;
|
|
29
|
+
let packed = byte_at(base + 2u + lane % 16u);
|
|
30
|
+
let quant = (packed >> ((lane / 16u) * 4u)) & 15u;
|
|
31
|
+
output[u.output_offset + i] = scale * f32(i32(quant) - 8);
|
|
32
|
+
}
|
|
@@ -18,6 +18,7 @@ export interface QsaPrepareOptions extends QsaIndexerGeometry {
|
|
|
18
18
|
/** Cache row and position of the first token; positions are contiguous cache rows. */
|
|
19
19
|
startPos: number;
|
|
20
20
|
eps: number;
|
|
21
|
+
normWeightOffset: boolean;
|
|
21
22
|
/** Raw key cache capacity in tokens. */
|
|
22
23
|
capacity: number;
|
|
23
24
|
outputBuffer?: GPUBuffer | null;
|
|
@@ -30,6 +31,7 @@ export interface QsaBlockKeyOptions {
|
|
|
30
31
|
rotaryDim: number;
|
|
31
32
|
compressRatio: number;
|
|
32
33
|
eps: number;
|
|
34
|
+
normWeightOffset: boolean;
|
|
33
35
|
capacity: number;
|
|
34
36
|
}
|
|
35
37
|
|
|
@@ -82,12 +82,13 @@ function validateIndexerGeometry({ queryHeads, headDim, rotaryDim }) {
|
|
|
82
82
|
* @param {import('./qsa-indexer.js').QsaPrepareOptions} options
|
|
83
83
|
*/
|
|
84
84
|
async function executePrepare(target, projection, qNormWeight, rope, rawKeys, options) {
|
|
85
|
-
const { numTokens, startPos, queryHeads, headDim, rotaryDim, eps, capacity, outputBuffer = null } = options;
|
|
85
|
+
const { numTokens, startPos, queryHeads, headDim, rotaryDim, eps, normWeightOffset, capacity, outputBuffer = null } = options;
|
|
86
86
|
validateIndexerGeometry({ queryHeads, headDim, rotaryDim });
|
|
87
87
|
integer(numTokens, 'numTokens');
|
|
88
88
|
integer(startPos, 'startPos', 0);
|
|
89
89
|
integer(capacity, 'capacity');
|
|
90
90
|
positive(eps, 'eps');
|
|
91
|
+
if (typeof normWeightOffset !== 'boolean') fail('normWeightOffset must be explicit boolean.');
|
|
91
92
|
if (startPos + numTokens > capacity) fail(`append of ${numTokens} rows at ${startPos} exceeds cache capacity ${capacity}.`);
|
|
92
93
|
assertTensor(projection, [numTokens, (queryHeads + 1) * headDim], 'index_qk projection');
|
|
93
94
|
assertTensor(qNormWeight, [headDim], 'q_layernorm weight');
|
|
@@ -102,7 +103,7 @@ async function executePrepare(target, projection, qNormWeight, rope, rawKeys, op
|
|
|
102
103
|
await unifiedKernelWrapper('qsa_prepare', target, 'f32',
|
|
103
104
|
[projection, qNormWeight, rope.cos, rope.sin, output, rawKeys],
|
|
104
105
|
{ num_tokens: numTokens, start_pos: startPos, group_stride: columns, eps }, grid,
|
|
105
|
-
{ WORKGROUP_SIZE: NORM_WORKGROUP, QUERY_HEADS: queryHeads, HEAD_DIM: headDim, ROTARY_DIM: rotaryDim });
|
|
106
|
+
{ WORKGROUP_SIZE: NORM_WORKGROUP, QUERY_HEADS: queryHeads, HEAD_DIM: headDim, ROTARY_DIM: rotaryDim, NORM_WEIGHT_OFFSET: Number(normWeightOffset) });
|
|
106
107
|
return createTensor(output, 'f32', [numTokens, queryHeads * headDim], 'qsa_query');
|
|
107
108
|
});
|
|
108
109
|
}
|
|
@@ -116,13 +117,14 @@ async function executePrepare(target, projection, qNormWeight, rope, rawKeys, op
|
|
|
116
117
|
* @param {import('./qsa-indexer.js').QsaBlockKeyOptions} options
|
|
117
118
|
*/
|
|
118
119
|
async function executeBlockKeys(target, rawKeys, kNormWeight, rope, blockKeys, options) {
|
|
119
|
-
const { firstBlock, blockCount, headDim, rotaryDim, compressRatio, eps, capacity } = options;
|
|
120
|
+
const { firstBlock, blockCount, headDim, rotaryDim, compressRatio, eps, normWeightOffset, capacity } = options;
|
|
120
121
|
validateIndexerGeometry({ queryHeads: 1, headDim, rotaryDim });
|
|
121
122
|
integer(firstBlock, 'firstBlock', 0);
|
|
122
123
|
integer(blockCount, 'blockCount', 0);
|
|
123
124
|
integer(compressRatio, 'compressRatio');
|
|
124
125
|
integer(capacity, 'capacity');
|
|
125
126
|
positive(eps, 'eps');
|
|
127
|
+
if (typeof normWeightOffset !== 'boolean') fail('normWeightOffset must be explicit boolean.');
|
|
126
128
|
if (blockCount === 0) return;
|
|
127
129
|
const lastToken = (firstBlock + blockCount) * compressRatio;
|
|
128
130
|
if (lastToken > capacity) fail(`blocks through ${firstBlock + blockCount} exceed cache capacity ${capacity}.`);
|
|
@@ -134,7 +136,7 @@ async function executeBlockKeys(target, rawKeys, kNormWeight, rope, blockKeys, o
|
|
|
134
136
|
await unifiedKernelWrapper('qsa_block_keys', target, 'f32',
|
|
135
137
|
[rawKeys, kNormWeight, rope.cos, rope.sin, blockKeys],
|
|
136
138
|
{ first_block: firstBlock, block_count: blockCount, group_stride: columns, eps }, grid,
|
|
137
|
-
{ WORKGROUP_SIZE: NORM_WORKGROUP, HEAD_DIM: headDim, ROTARY_DIM: rotaryDim, COMPRESS_RATIO: compressRatio });
|
|
139
|
+
{ WORKGROUP_SIZE: NORM_WORKGROUP, HEAD_DIM: headDim, ROTARY_DIM: rotaryDim, COMPRESS_RATIO: compressRatio, NORM_WEIGHT_OFFSET: Number(normWeightOffset) });
|
|
138
140
|
}
|
|
139
141
|
|
|
140
142
|
/** @param {import('./qsa-indexer.js').QsaSelectionGeometry} options */
|
|
@@ -5,6 +5,7 @@
|
|
|
5
5
|
// the position of the block's first token. Positions equal cache rows: the
|
|
6
6
|
// owning cache admits only contiguous text positions.
|
|
7
7
|
|
|
8
|
+
override NORM_WEIGHT_OFFSET: f32;
|
|
8
9
|
override WORKGROUP_SIZE: u32 = 128u;
|
|
9
10
|
override HEAD_DIM: u32;
|
|
10
11
|
override ROTARY_DIM: u32;
|
|
@@ -86,7 +87,7 @@ fn main(
|
|
|
86
87
|
}
|
|
87
88
|
let inverse_rms = inverse_root_refined(divide_refined(partial[0], f32(HEAD_DIM)) + u.eps);
|
|
88
89
|
for (var d = local; d < HEAD_DIM; d = d + WORKGROUP_SIZE) {
|
|
89
|
-
normalized[d] = (normalized[d] * inverse_rms) * (
|
|
90
|
+
normalized[d] = (normalized[d] * inverse_rms) * (NORM_WEIGHT_OFFSET + k_norm_weight[d]);
|
|
90
91
|
}
|
|
91
92
|
workgroupBarrier();
|
|
92
93
|
|
|
@@ -5,6 +5,7 @@
|
|
|
5
5
|
// The key head is appended raw (pre-norm, pre-RoPE) to the owned key cache at
|
|
6
6
|
// cache row start_pos + token; every call appends, dense or sparse.
|
|
7
7
|
|
|
8
|
+
override NORM_WEIGHT_OFFSET: f32;
|
|
8
9
|
override WORKGROUP_SIZE: u32 = 128u;
|
|
9
10
|
override QUERY_HEADS: u32;
|
|
10
11
|
override HEAD_DIM: u32;
|
|
@@ -92,7 +93,7 @@ fn main(
|
|
|
92
93
|
}
|
|
93
94
|
let inverse_rms = inverse_root_refined(divide_refined(partial[0], f32(HEAD_DIM)) + u.eps);
|
|
94
95
|
for (var d = local; d < HEAD_DIM; d = d + WORKGROUP_SIZE) {
|
|
95
|
-
normalized[d] = (projection[source + d] * inverse_rms) * (
|
|
96
|
+
normalized[d] = (projection[source + d] * inverse_rms) * (NORM_WEIGHT_OFFSET + q_norm_weight[d]);
|
|
96
97
|
}
|
|
97
98
|
workgroupBarrier();
|
|
98
99
|
|
package/src/gpu/kernels/relu.js
CHANGED
|
@@ -42,7 +42,7 @@ async function _relu(target, input, options = {}) {
|
|
|
42
42
|
target,
|
|
43
43
|
variant,
|
|
44
44
|
[input, output],
|
|
45
|
-
{ size,
|
|
45
|
+
{ size, dispatch_stride: dispatchPlan.dispatchStride, _pad1: 0, _pad2: 0 },
|
|
46
46
|
dispatchPlan.workgroups, null, null, null, options.signal
|
|
47
47
|
);
|
|
48
48
|
|
|
@@ -2,7 +2,7 @@ override WORKGROUP_SIZE: u32 = 256u;
|
|
|
2
2
|
|
|
3
3
|
struct Uniforms {
|
|
4
4
|
size: u32,
|
|
5
|
-
|
|
5
|
+
dispatch_stride: u32,
|
|
6
6
|
_pad1: u32,
|
|
7
7
|
_pad2: u32,
|
|
8
8
|
}
|
|
@@ -13,7 +13,7 @@ struct Uniforms {
|
|
|
13
13
|
|
|
14
14
|
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
15
15
|
fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
|
|
16
|
-
let dispatch_stride = max(u.
|
|
16
|
+
let dispatch_stride = max(u.dispatch_stride, 1u);
|
|
17
17
|
let idx = gid.y * dispatch_stride + gid.x;
|
|
18
18
|
if (idx >= u.size) {
|
|
19
19
|
return;
|
|
@@ -4,7 +4,7 @@ override WORKGROUP_SIZE: u32 = 256u;
|
|
|
4
4
|
|
|
5
5
|
struct Uniforms {
|
|
6
6
|
size: u32,
|
|
7
|
-
|
|
7
|
+
dispatch_stride: u32,
|
|
8
8
|
_pad1: u32,
|
|
9
9
|
_pad2: u32,
|
|
10
10
|
}
|
|
@@ -15,7 +15,7 @@ struct Uniforms {
|
|
|
15
15
|
|
|
16
16
|
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
17
17
|
fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
|
|
18
|
-
let dispatch_stride = max(u.
|
|
18
|
+
let dispatch_stride = max(u.dispatch_stride, 1u);
|
|
19
19
|
let idx = gid.y * dispatch_stride + gid.x;
|
|
20
20
|
if (idx >= u.size) {
|
|
21
21
|
return;
|
|
@@ -152,7 +152,7 @@ async function _residualAdd(target, a, b, size, options = {}) {
|
|
|
152
152
|
await unifiedKernelWrapper(
|
|
153
153
|
'residual', target, variant,
|
|
154
154
|
[aAligned, bAligned, output],
|
|
155
|
-
{ size, scale: outputScale,
|
|
155
|
+
{ size, scale: outputScale, dispatch_stride: dispatchPlan.dispatchStride, _pad2: 0 },
|
|
156
156
|
dispatchPlan.workgroups
|
|
157
157
|
);
|
|
158
158
|
return createTensor(output, outputDtype, [size], 'residual_output');
|
|
@@ -10,7 +10,7 @@
|
|
|
10
10
|
struct Uniforms {
|
|
11
11
|
size: u32, // Total number of elements
|
|
12
12
|
scale: f32, // Output scale for main/add_vec4; residual scale for add_scaled
|
|
13
|
-
|
|
13
|
+
dispatch_stride: u32,
|
|
14
14
|
_pad2: u32,
|
|
15
15
|
}
|
|
16
16
|
|
|
@@ -23,7 +23,7 @@ override WORKGROUP_SIZE: u32 = 256u;
|
|
|
23
23
|
|
|
24
24
|
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
25
25
|
fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
|
|
26
|
-
let dispatch_stride = max(u.
|
|
26
|
+
let dispatch_stride = max(u.dispatch_stride, 1u);
|
|
27
27
|
let idx = gid.y * dispatch_stride + gid.x;
|
|
28
28
|
if (idx >= u.size) {
|
|
29
29
|
return;
|
|
@@ -36,7 +36,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
|
|
|
36
36
|
// This avoids requiring a different bind group layout with read_write on 'a'
|
|
37
37
|
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
38
38
|
fn add_inplace(@builtin(global_invocation_id) gid: vec3<u32>) {
|
|
39
|
-
let dispatch_stride = max(u.
|
|
39
|
+
let dispatch_stride = max(u.dispatch_stride, 1u);
|
|
40
40
|
let idx = gid.y * dispatch_stride + gid.x;
|
|
41
41
|
if (idx >= u.size) {
|
|
42
42
|
return;
|
|
@@ -47,7 +47,7 @@ fn add_inplace(@builtin(global_invocation_id) gid: vec3<u32>) {
|
|
|
47
47
|
// Fused residual + scale: output = a + scale * b
|
|
48
48
|
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
49
49
|
fn add_scaled(@builtin(global_invocation_id) gid: vec3<u32>) {
|
|
50
|
-
let dispatch_stride = max(u.
|
|
50
|
+
let dispatch_stride = max(u.dispatch_stride, 1u);
|
|
51
51
|
let idx = gid.y * dispatch_stride + gid.x;
|
|
52
52
|
if (idx >= u.size) {
|
|
53
53
|
return;
|
|
@@ -16,7 +16,7 @@ enable f16;
|
|
|
16
16
|
struct Uniforms {
|
|
17
17
|
size: u32, // Total number of elements
|
|
18
18
|
scale: f32, // Output scale for main/add_vec4; residual scale for add_scaled
|
|
19
|
-
|
|
19
|
+
dispatch_stride: u32,
|
|
20
20
|
_pad2: u32,
|
|
21
21
|
}
|
|
22
22
|
|
|
@@ -29,7 +29,7 @@ override WORKGROUP_SIZE: u32 = 256u;
|
|
|
29
29
|
|
|
30
30
|
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
31
31
|
fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
|
|
32
|
-
let dispatch_stride = max(u.
|
|
32
|
+
let dispatch_stride = max(u.dispatch_stride, 1u);
|
|
33
33
|
let idx = gid.y * dispatch_stride + gid.x;
|
|
34
34
|
if (idx >= u.size) {
|
|
35
35
|
return;
|
|
@@ -11,7 +11,7 @@ enable f16;
|
|
|
11
11
|
struct Uniforms {
|
|
12
12
|
size: u32, // Total number of elements
|
|
13
13
|
scale: f32, // Output scale
|
|
14
|
-
|
|
14
|
+
dispatch_stride: u32,
|
|
15
15
|
_pad2: u32,
|
|
16
16
|
}
|
|
17
17
|
|
|
@@ -25,7 +25,7 @@ override WORKGROUP_SIZE_VEC4: u32 = 64u;
|
|
|
25
25
|
// Vectorized version for better throughput
|
|
26
26
|
@compute @workgroup_size(WORKGROUP_SIZE_VEC4, 1, 1)
|
|
27
27
|
fn add_vec4(@builtin(global_invocation_id) gid: vec3<u32>) {
|
|
28
|
-
let dispatch_stride = max(u.
|
|
28
|
+
let dispatch_stride = max(u.dispatch_stride, 4u);
|
|
29
29
|
let idx = gid.y * dispatch_stride + gid.x * 4u;
|
|
30
30
|
let size = u.size;
|
|
31
31
|
|
|
@@ -9,7 +9,7 @@
|
|
|
9
9
|
struct Uniforms {
|
|
10
10
|
size: u32, // Total number of elements
|
|
11
11
|
scale: f32, // Output scale
|
|
12
|
-
|
|
12
|
+
dispatch_stride: u32,
|
|
13
13
|
_pad2: u32,
|
|
14
14
|
}
|
|
15
15
|
|
|
@@ -23,7 +23,7 @@ override WORKGROUP_SIZE_VEC4: u32 = 64u;
|
|
|
23
23
|
// Vectorized version for better throughput
|
|
24
24
|
@compute @workgroup_size(WORKGROUP_SIZE_VEC4, 1, 1)
|
|
25
25
|
fn add_vec4(@builtin(global_invocation_id) gid: vec3<u32>) {
|
|
26
|
-
let dispatch_stride = max(u.
|
|
26
|
+
let dispatch_stride = max(u.dispatch_stride, 4u);
|
|
27
27
|
let idx = gid.y * dispatch_stride + gid.x * 4u;
|
|
28
28
|
let size = u.size;
|
|
29
29
|
|
|
@@ -1,7 +1,7 @@
|
|
|
1
1
|
import { planRoPEPrecomputeDispatch } from './rope-dispatch.js';
|
|
2
2
|
import { runPairedRoPEPrecompute } from './paired-rope.js';
|
|
3
3
|
import { validateRoPEInverseFrequencies } from '../../config/rope-frequencies.js';
|
|
4
|
-
import { DEFAULT_MANIFEST_INFERENCE } from
|
|
4
|
+
import { DEFAULT_MANIFEST_INFERENCE } from "../../config/schema/manifest/manifest.schema.js";
|
|
5
5
|
import { getDevice } from '../device.js';
|
|
6
6
|
import { acquireBuffer, releaseBuffer } from '../../memory/buffer-pool.js';
|
|
7
7
|
import { unifiedKernelWrapper } from './kernel-execution.js';
|
|
@@ -55,34 +55,3 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
|
|
|
55
55
|
|
|
56
56
|
output[tid] = sum;
|
|
57
57
|
}
|
|
58
|
-
|
|
59
|
-
// In-place accumulation version (adds to existing output)
|
|
60
|
-
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
61
|
-
fn scatter_add_accumulate(@builtin(global_invocation_id) gid: vec3<u32>) {
|
|
62
|
-
let tid = gid.x;
|
|
63
|
-
let total_elements = u.num_tokens * u.hidden_size;
|
|
64
|
-
|
|
65
|
-
if (tid >= total_elements) {
|
|
66
|
-
return;
|
|
67
|
-
}
|
|
68
|
-
|
|
69
|
-
let token_idx = tid / u.hidden_size;
|
|
70
|
-
let dim_idx = tid % u.hidden_size;
|
|
71
|
-
let top_k = u.top_k;
|
|
72
|
-
let hidden_size = u.hidden_size;
|
|
73
|
-
let num_tokens = u.num_tokens;
|
|
74
|
-
|
|
75
|
-
var sum: f32 = 0.0;
|
|
76
|
-
let routing_base = token_idx * top_k;
|
|
77
|
-
|
|
78
|
-
for (var k: u32 = 0u; k < top_k; k = k + 1u) {
|
|
79
|
-
let expert_idx = indices[routing_base + k];
|
|
80
|
-
let weight = weights[routing_base + k];
|
|
81
|
-
|
|
82
|
-
let expert_offset = expert_idx * num_tokens * hidden_size + token_idx * hidden_size + dim_idx;
|
|
83
|
-
sum = sum + weight * expert_outputs[expert_offset];
|
|
84
|
-
}
|
|
85
|
-
|
|
86
|
-
// Accumulate to existing value
|
|
87
|
-
output[tid] = output[tid] + sum;
|
|
88
|
-
}
|
|
@@ -4,6 +4,8 @@ import { log } from '../../debug/index.js';
|
|
|
4
4
|
import { isDeviceLost, observeDeviceLoss } from '../../memory/device-residency.js';
|
|
5
5
|
import { getScopedShaderSource } from './shader-source-scope.js';
|
|
6
6
|
import { hashBytesSha256 } from '../../formats/canonical-hash.js';
|
|
7
|
+
import { filterShaderForCapabilities } from './shader-capability-filter.js';
|
|
8
|
+
import { DOPPLER_VERSION } from '../../version.js';
|
|
7
9
|
|
|
8
10
|
// ============================================================================
|
|
9
11
|
// Caches
|
|
@@ -219,8 +221,14 @@ export async function compileShader(
|
|
|
219
221
|
log.debug('compileShader', `${label}: ${msg.message} (line ${msg.lineNum}:${msg.linePos})`);
|
|
220
222
|
}
|
|
221
223
|
}
|
|
222
|
-
|
|
223
|
-
|
|
224
|
+
const errors = compilationInfo.messages.filter(m => m.type === 'error');
|
|
225
|
+
if (errors.length > 0) {
|
|
226
|
+
// The compiler's own words, where it stopped, and what the device offers:
|
|
227
|
+
// without these a failure on one browser cannot be told from another.
|
|
228
|
+
const details = errors.slice(0, 3).map(m => `${m.message} (line ${m.lineNum}:${m.linePos})`).join('; ');
|
|
229
|
+
const features = [...(device.features ?? [])].sort().join(', ') || 'none';
|
|
230
|
+
throw new Error(`Shader compilation failed for ${label}: ${details}`
|
|
231
|
+
+ `${errors.length > 3 ? `; and ${errors.length - 3} more` : ''}. Device features: ${features}. Doppler ${DOPPLER_VERSION}.`);
|
|
224
232
|
}
|
|
225
233
|
}
|
|
226
234
|
|
|
@@ -234,7 +242,13 @@ export async function getShaderModule(
|
|
|
234
242
|
label
|
|
235
243
|
) {
|
|
236
244
|
const shaderModuleCache = moduleCacheFor(device);
|
|
237
|
-
const
|
|
245
|
+
const loadedSource = await loadShaderSource(shaderFile);
|
|
246
|
+
// A device without subgroups is given the module it can compile: see
|
|
247
|
+
// shader-capability-filter.js. The identity below is of the source compiled.
|
|
248
|
+
const shaderSource = filterShaderForCapabilities(loadedSource, {
|
|
249
|
+
hasSubgroups: device.features?.has?.('subgroups') === true,
|
|
250
|
+
label: shaderFile,
|
|
251
|
+
}).source;
|
|
238
252
|
const digest = sourceIdentity(shaderFile, shaderSource);
|
|
239
253
|
const cacheKey = `${getDeviceId(device)}:${digest}`;
|
|
240
254
|
const cached = shaderModuleCache.get(cacheKey);
|
|
@@ -0,0 +1,120 @@
|
|
|
1
|
+
// A WGSL `enable` directive applies to the whole module. Several kernels keep a
|
|
2
|
+
// subgroup entry point beside plain ones in one file and declare
|
|
3
|
+
// `enable subgroups;` at the top. A device without the subgroups feature then
|
|
4
|
+
// rejects the whole module, including the entry points that never touch a
|
|
5
|
+
// subgroup builtin, before variant selection has any say.
|
|
6
|
+
//
|
|
7
|
+
// This filter gives such a device the module it can compile: the directive is
|
|
8
|
+
// removed together with every function that uses subgroup builtins, directly
|
|
9
|
+
// or through a function it calls. Devices that offer subgroups receive the
|
|
10
|
+
// source unchanged. A module left with no compute entry point needs the feature
|
|
11
|
+
// outright, and that is reported as such instead of as a compiler failure.
|
|
12
|
+
|
|
13
|
+
const ENABLE_SUBGROUPS = /^[ \t]*enable[ \t]+subgroups[ \t]*;[ \t]*\r?\n?/m;
|
|
14
|
+
// The compute stage attribute, written escaped: this file matches shader text and holds none.
|
|
15
|
+
const COMPUTE_STAGE = /\x40compute\b/;
|
|
16
|
+
const SUBGROUP_USE = /\bsubgroup[A-Z]\w*\s*\(|@builtin\(\s*subgroup_\w+\s*\)/;
|
|
17
|
+
|
|
18
|
+
// Comments are blanked, not removed, so offsets into the stripped text are
|
|
19
|
+
// offsets into the source.
|
|
20
|
+
/** @param {string} source */
|
|
21
|
+
function blankComments(source) {
|
|
22
|
+
let out = '';
|
|
23
|
+
let index = 0;
|
|
24
|
+
while (index < source.length) {
|
|
25
|
+
if (source.startsWith('//', index)) {
|
|
26
|
+
const end = source.indexOf('\n', index);
|
|
27
|
+
const stop = end === -1 ? source.length : end;
|
|
28
|
+
out += ' '.repeat(stop - index);
|
|
29
|
+
index = stop;
|
|
30
|
+
} else if (source.startsWith('/*', index)) {
|
|
31
|
+
let depth = 1;
|
|
32
|
+
let cursor = index + 2;
|
|
33
|
+
while (cursor < source.length && depth > 0) {
|
|
34
|
+
if (source.startsWith('/*', cursor)) { depth += 1; cursor += 2; }
|
|
35
|
+
else if (source.startsWith('*/', cursor)) { depth -= 1; cursor += 2; }
|
|
36
|
+
else cursor += 1;
|
|
37
|
+
}
|
|
38
|
+
out += source.slice(index, cursor).replace(/[^\n]/g, ' ');
|
|
39
|
+
index = cursor;
|
|
40
|
+
} else {
|
|
41
|
+
out += source[index];
|
|
42
|
+
index += 1;
|
|
43
|
+
}
|
|
44
|
+
}
|
|
45
|
+
return out;
|
|
46
|
+
}
|
|
47
|
+
|
|
48
|
+
// Every module-scope function: its name, the span from its first attribute to
|
|
49
|
+
// its closing brace, and its text with comments blanked.
|
|
50
|
+
/** @param {string} source */
|
|
51
|
+
function moduleFunctions(source) {
|
|
52
|
+
const code = blankComments(source);
|
|
53
|
+
/** @type {{ name: string, start: number, end: number, text: string }[]} */
|
|
54
|
+
const functions = [];
|
|
55
|
+
const header = /(^|\n)((?:[ \t]*@[^\n]*\n)*[ \t]*(?:@[^\n{]*?\s)?fn\s+([A-Za-z_]\w*)\s*\()/g;
|
|
56
|
+
let match;
|
|
57
|
+
while ((match = header.exec(code)) !== null) {
|
|
58
|
+
const start = match.index + match[1].length;
|
|
59
|
+
const open = code.indexOf('{', match.index + match[0].length);
|
|
60
|
+
if (open === -1) break;
|
|
61
|
+
let depth = 0;
|
|
62
|
+
let end = -1;
|
|
63
|
+
for (let cursor = open; cursor < code.length; cursor += 1) {
|
|
64
|
+
if (code[cursor] === '{') depth += 1;
|
|
65
|
+
else if (code[cursor] === '}') { depth -= 1; if (depth === 0) { end = cursor + 1; break; } }
|
|
66
|
+
}
|
|
67
|
+
if (end === -1) break;
|
|
68
|
+
functions.push({ name: match[3], start, end, text: code.slice(start, end) });
|
|
69
|
+
header.lastIndex = end;
|
|
70
|
+
}
|
|
71
|
+
return functions;
|
|
72
|
+
}
|
|
73
|
+
|
|
74
|
+
/** @param {string} source */
|
|
75
|
+
export function shaderEnablesSubgroups(source) {
|
|
76
|
+
return ENABLE_SUBGROUPS.test(blankComments(String(source ?? '')));
|
|
77
|
+
}
|
|
78
|
+
|
|
79
|
+
// The source a device can compile, whether it was filtered, and the functions removed.
|
|
80
|
+
/**
|
|
81
|
+
* @param {string} source
|
|
82
|
+
* @param {{ hasSubgroups: boolean, label?: string }} capabilities
|
|
83
|
+
*/
|
|
84
|
+
export function filterShaderForCapabilities(source, { hasSubgroups, label = 'shader' }) {
|
|
85
|
+
const text = String(source ?? '');
|
|
86
|
+
if (hasSubgroups || !shaderEnablesSubgroups(text)) {
|
|
87
|
+
return { source: text, filtered: false, removedFunctions: [] };
|
|
88
|
+
}
|
|
89
|
+
const functions = moduleFunctions(text);
|
|
90
|
+
const removed = new Set(functions.filter((fn) => SUBGROUP_USE.test(fn.text)).map((fn) => fn.name));
|
|
91
|
+
// A function that calls a removed function goes too.
|
|
92
|
+
for (let changed = true; changed;) {
|
|
93
|
+
changed = false;
|
|
94
|
+
for (const fn of functions) {
|
|
95
|
+
if (removed.has(fn.name)) continue;
|
|
96
|
+
if ([...removed].some((name) => new RegExp(`\\b${name}\\s*\\(`).test(fn.text.slice(fn.text.indexOf('{'))))) {
|
|
97
|
+
removed.add(fn.name);
|
|
98
|
+
changed = true;
|
|
99
|
+
}
|
|
100
|
+
}
|
|
101
|
+
}
|
|
102
|
+
let filtered = '';
|
|
103
|
+
let cursor = 0;
|
|
104
|
+
for (const fn of functions) {
|
|
105
|
+
if (!removed.has(fn.name)) continue;
|
|
106
|
+
filtered += text.slice(cursor, fn.start);
|
|
107
|
+
cursor = fn.end;
|
|
108
|
+
}
|
|
109
|
+
filtered += text.slice(cursor);
|
|
110
|
+
const directive = blankComments(filtered).match(ENABLE_SUBGROUPS);
|
|
111
|
+
if (directive) {
|
|
112
|
+
const at = directive.index ?? 0;
|
|
113
|
+
filtered = filtered.slice(0, at) + filtered.slice(at + directive[0].length);
|
|
114
|
+
}
|
|
115
|
+
const remaining = blankComments(filtered);
|
|
116
|
+
if (SUBGROUP_USE.test(remaining) || !COMPUTE_STAGE.test(remaining)) {
|
|
117
|
+
throw new Error(`${label} requires the WebGPU "subgroups" feature, which this device does not offer.`);
|
|
118
|
+
}
|
|
119
|
+
return { source: filtered, filtered: true, removedFunctions: [...removed] };
|
|
120
|
+
}
|
|
@@ -1,5 +1,5 @@
|
|
|
1
1
|
|
|
2
|
-
import { getKernelCapabilities } from '../device.js';
|
|
2
|
+
import { getKernelCapabilities, getDevice } from '../device.js';
|
|
3
3
|
import { acquireBuffer, releaseBuffer } from '../../memory/buffer-pool.js';
|
|
4
4
|
import { createTensor } from '../tensor.js';
|
|
5
5
|
import { unifiedKernelWrapper } from './kernel-execution.js';
|
|
@@ -10,7 +10,7 @@ import { dispatchKernel } from './dispatch.js';
|
|
|
10
10
|
import { trace, isTraceEnabled } from '../../debug/index.js';
|
|
11
11
|
import { getKernelThresholdSnapshot } from '../../config/schema/index.js';
|
|
12
12
|
import { selectRuleValue } from './rule-registry.js';
|
|
13
|
-
|
|
13
|
+
|
|
14
14
|
import { getKernelConfig } from '../../config/kernel-registry-contract.js';
|
|
15
15
|
|
|
16
16
|
function selectSoftmaxVariant(innerSize) {
|
|
@@ -63,10 +63,10 @@ export async function recordSoftmax(recorder, input, axis, options = {}) {
|
|
|
63
63
|
|
|
64
64
|
async function softmaxTopK(target, logits, numTokens, numExperts, topK, options = {}) {
|
|
65
65
|
const device = target?.device ?? getDevice();
|
|
66
|
-
const {
|
|
67
|
-
|
|
68
|
-
|
|
69
|
-
}
|
|
66
|
+
const { normalize, modelType = null } = options;
|
|
67
|
+
if (typeof normalize !== 'boolean') {
|
|
68
|
+
throw new Error(`SoftmaxTopK requires options.normalize from the router's normalizeWeights, got ${String(normalize)}.`);
|
|
69
|
+
}
|
|
70
70
|
const inputDtype = options.inputDtype;
|
|
71
71
|
const weightsDtype = options.weightsDtype;
|
|
72
72
|
|
|
@@ -100,12 +100,12 @@ async function softmaxTopK(target, logits, numTokens, numExperts, topK, options
|
|
|
100
100
|
throw new Error(`SoftmaxTopK fused_logit_rank needs subgroups of at least top-${topK} lanes; got ${minSubgroup}.`);
|
|
101
101
|
}
|
|
102
102
|
}
|
|
103
|
+
// Every top-k variant declares the expert count its workgroup storage holds.
|
|
103
104
|
const limits = getKernelConfig('topk', variant).variantMetadata;
|
|
104
|
-
if (limits?.maxExperts
|
|
105
|
-
|
|
106
|
-
throw new Error(`SoftmaxTopK ${variant} requires declared integer expert and top-k limits.`);
|
|
105
|
+
if (!Number.isSafeInteger(limits?.maxExperts) || !Number.isSafeInteger(limits?.maxTopK)) {
|
|
106
|
+
throw new Error(`SoftmaxTopK ${variant} requires declared integer variantMetadata.maxExperts and maxTopK.`);
|
|
107
107
|
}
|
|
108
|
-
if (
|
|
108
|
+
if (numExperts > limits.maxExperts || topK > limits.maxTopK || topK > numExperts) {
|
|
109
109
|
throw new Error(`SoftmaxTopK ${variant} supports at most ${limits.maxExperts} experts and top-${limits.maxTopK}; got experts=${numExperts} topK=${topK}.`);
|
|
110
110
|
}
|
|
111
111
|
const pipeline = await createPipeline('topk', variant);
|
|
@@ -101,61 +101,6 @@ fn main(
|
|
|
101
101
|
}
|
|
102
102
|
}
|
|
103
103
|
|
|
104
|
-
// Optimized version for topK=2, numExperts<=8
|
|
105
|
-
// Each thread handles one token
|
|
106
|
-
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
107
|
-
fn topk_2_small(@builtin(global_invocation_id) gid: vec3<u32>) {
|
|
108
|
-
let token_idx = gid.x;
|
|
109
|
-
|
|
110
|
-
if (token_idx >= u.num_tokens) {
|
|
111
|
-
return;
|
|
112
|
-
}
|
|
113
|
-
|
|
114
|
-
let num_experts = u.num_experts;
|
|
115
|
-
let base_offset = token_idx * num_experts;
|
|
116
|
-
|
|
117
|
-
// Find top 2 in a single pass
|
|
118
|
-
var top1_idx: u32 = 0u;
|
|
119
|
-
var top1_val: f32 = probs[base_offset];
|
|
120
|
-
var top2_idx: u32 = 1u;
|
|
121
|
-
var top2_val: f32 = probs[base_offset + 1u];
|
|
122
|
-
|
|
123
|
-
// Ensure top1 >= top2
|
|
124
|
-
if (top2_val > top1_val) {
|
|
125
|
-
let tmp_idx = top1_idx;
|
|
126
|
-
let tmp_val = top1_val;
|
|
127
|
-
top1_idx = top2_idx;
|
|
128
|
-
top1_val = top2_val;
|
|
129
|
-
top2_idx = tmp_idx;
|
|
130
|
-
top2_val = tmp_val;
|
|
131
|
-
}
|
|
132
|
-
|
|
133
|
-
// Scan remaining experts
|
|
134
|
-
for (var i: u32 = 2u; i < num_experts; i = i + 1u) {
|
|
135
|
-
let val = probs[base_offset + i];
|
|
136
|
-
if (val > top1_val) {
|
|
137
|
-
top2_idx = top1_idx;
|
|
138
|
-
top2_val = top1_val;
|
|
139
|
-
top1_idx = i;
|
|
140
|
-
top1_val = val;
|
|
141
|
-
} else if (val > top2_val) {
|
|
142
|
-
top2_idx = i;
|
|
143
|
-
top2_val = val;
|
|
144
|
-
}
|
|
145
|
-
}
|
|
146
|
-
|
|
147
|
-
// Renormalize weights
|
|
148
|
-
let weight_sum = top1_val + top2_val;
|
|
149
|
-
let inv_sum = select(1.0, 1.0 / weight_sum, u.normalize == 1u && weight_sum > 0.0);
|
|
150
|
-
|
|
151
|
-
// Write output
|
|
152
|
-
let out_base = token_idx * 2u;
|
|
153
|
-
out_indices[out_base] = top1_idx;
|
|
154
|
-
out_indices[out_base + 1u] = top2_idx;
|
|
155
|
-
out_weights[out_base] = top1_val * inv_sum;
|
|
156
|
-
out_weights[out_base + 1u] = top2_val * inv_sum;
|
|
157
|
-
}
|
|
158
|
-
|
|
159
104
|
// Fused softmax + top-k for efficiency
|
|
160
105
|
// Avoids separate softmax kernel call
|
|
161
106
|
// Workgroup size WORKGROUP_SIZE to support up to WORKGROUP_SIZE experts
|
|
@@ -29,7 +29,7 @@ async function _transpose(target, input, rows, cols, options = {}) {
|
|
|
29
29
|
target,
|
|
30
30
|
'default',
|
|
31
31
|
[input, outputBuf],
|
|
32
|
-
{ rows, cols,
|
|
32
|
+
{ rows, cols, dispatch_stride: dispatchPlan.dispatchStride, _pad1: 0 },
|
|
33
33
|
[dispatchPlan.workgroups[0], rows, 1]
|
|
34
34
|
);
|
|
35
35
|
|