doppler-gpu 0.7.2 → 0.7.7
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 +50 -0
- package/package.json +5 -48
- package/src/capsule-runtime.d.ts +2 -2
- package/src/cli/doppler-cli.js +3 -1
- package/src/cli/doppler-serve.js +3 -1
- 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 +508 -0
- package/src/config/kernels/registry.js +2 -2
- package/src/config/kernels/registry.json +1892 -1096
- 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/{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/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/{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} +26 -1
- package/src/config/schema/{loading.schema.js → runtime/loading.schema.js} +88 -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_split_block_f32.wgsl +63 -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} +79 -34
- 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/{resident-partition.js → partitions/resident-partition.js} +19 -13
- 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 +488 -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 +117 -13
- 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 +62 -3
- package/src/loader/experts/routed-expert-cache.js +451 -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 +82 -0
- package/src/loader/experts/routed-expert-source.js +233 -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 +287 -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-gather-lane.js +192 -0
- package/src/storage/verified-piece-gather-worker.js +520 -0
- package/src/storage/verified-piece-gather.d.ts +113 -0
- package/src/storage/verified-piece-gather.js +199 -0
- package/src/storage/verified-piece-reader.d.ts +7 -0
- package/src/storage/verified-piece-reader.js +69 -12
- package/src/storage/verified-piece-storage.d.ts +3 -1
- package/src/storage/verified-piece-storage.js +101 -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/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/{kernel-warmup.schema.d.ts → execution/kernel-warmup.schema.d.ts} +0 -0
- /package/src/config/schema/{kernel-warmup.schema.js → execution/kernel-warmup.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/{gpu-cache.schema.js → platform/gpu-cache.schema.js} +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/{buffer-pool.schema.js → runtime/buffer-pool.schema.js} +0 -0
- /package/src/config/schema/{debug.schema.d.ts → runtime/debug.schema.d.ts} +0 -0
- /package/src/config/schema/{debug.schema.js → runtime/debug.schema.js} +0 -0
- /package/src/config/schema/{distribution.schema.d.ts → runtime/distribution.schema.d.ts} +0 -0
- /package/src/config/schema/{distribution.schema.js → runtime/distribution.schema.js} +0 -0
- /package/src/config/schema/{ecosystem.schema.d.ts → runtime/ecosystem.schema.d.ts} +0 -0
- /package/src/config/schema/{ecosystem.schema.js → runtime/ecosystem.schema.js} +0 -0
- /package/src/config/schema/{harness.schema.d.ts → runtime/harness.schema.d.ts} +0 -0
- /package/src/config/schema/{harness.schema.js → runtime/harness.schema.js} +0 -0
- /package/src/config/schema/{hotswap.schema.d.ts → runtime/hotswap.schema.d.ts} +0 -0
- /package/src/config/schema/{hotswap.schema.js → runtime/hotswap.schema.js} +0 -0
- /package/src/config/schema/{intent-bundle.schema.d.ts → runtime/intent-bundle.schema.d.ts} +0 -0
- /package/src/config/schema/{intent-bundle.schema.js → runtime/intent-bundle.schema.js} +0 -0
- /package/src/config/schema/{runtime-profile.schema.d.ts → runtime/runtime-profile.schema.d.ts} +0 -0
- /package/src/config/schema/{runtime-profile.schema.js → runtime/runtime-profile.schema.js} +0 -0
- /package/src/config/schema/{storage.schema.d.ts → runtime/storage.schema.d.ts} +0 -0
- /package/src/config/schema/{storage.schema.js → runtime/storage.schema.js} +0 -0
- /package/src/config/schema/{tooling.schema.js → runtime/tooling.schema.js} +0 -0
- /package/src/config/schema/{units.schema.d.ts → runtime/units.schema.d.ts} +0 -0
- /package/src/config/schema/{units.schema.js → runtime/units.schema.js} +0 -0
- /package/src/config/schema/{backward-registry.schema.d.ts → training/backward-registry.schema.d.ts} +0 -0
- /package/src/config/schema/{backward-registry.schema.js → training/backward-registry.schema.js} +0 -0
- /package/src/config/schema/{distill-training.schema.d.ts → training/distill-training.schema.d.ts} +0 -0
- /package/src/config/schema/{distill-training.schema.js → training/distill-training.schema.js} +0 -0
- /package/src/config/schema/{lora.schema.d.ts → training/lora.schema.d.ts} +0 -0
- /package/src/config/schema/{training-metrics.schema.d.ts → training/training-metrics.schema.d.ts} +0 -0
- /package/src/config/schema/{training-metrics.schema.js → training/training-metrics.schema.js} +0 -0
- /package/src/config/schema/{training.schema.d.ts → training/training.schema.d.ts} +0 -0
- /package/src/config/schema/{training.schema.js → training/training.schema.js} +0 -0
- /package/src/config/schema/{ul-training.schema.d.ts → training/ul-training.schema.d.ts} +0 -0
- /package/src/config/schema/{ul-training.schema.js → training/ul-training.schema.js} +0 -0
- /package/src/{converter → config/transforms}/execution-graph/dtype.js +0 -0
- /package/src/{converter → config/transforms}/execution-graph/fusion.js +0 -0
- /package/src/{converter → config/transforms}/execution-graph/kv-storage.js +0 -0
- /package/src/{converter → config/transforms}/execution-graph/validation.js +0 -0
- /package/src/{converter → formats}/qwen4-exp-source-defaults.json +0 -0
- /package/src/inference/pipelines/text/{buffer-types.d.ts → contracts/buffer-types.d.ts} +0 -0
- /package/src/inference/pipelines/text/{embedding-contract.d.ts → embedding/embedding-contract.d.ts} +0 -0
- /package/src/inference/pipelines/text/{embedding-contract.js → embedding/embedding-contract.js} +0 -0
- /package/src/inference/pipelines/text/{precomputed-input-norm.js → execution/precomputed-input-norm.js} +0 -0
- /package/src/inference/pipelines/text/generator/{batch-retry-policy.js → decode/batch-retry-policy.js} +0 -0
- /package/src/inference/pipelines/text/{chat-format.d.ts → generator/request/chat-format.d.ts} +0 -0
- /package/src/inference/pipelines/text/{chat-format.js → generator/request/chat-format.js} +0 -0
- /package/src/inference/pipelines/text/generator/{recorder-stats.js → session/recorder-stats.js} +0 -0
- /package/src/inference/pipelines/text/generator/{whole-token-template.js → whole-token/whole-token-template.js} +0 -0
- /package/src/inference/pipelines/text/{init-chat-templates.js → initialization/init-chat-templates.js} +0 -0
- /package/src/inference/pipelines/text/{load-policy.js → loading/load-policy.js} +0 -0
- /package/src/inference/pipelines/text/{pipeline-load-timing.js → loading/pipeline-load-timing.js} +0 -0
- /package/src/inference/pipelines/text/{generator → logits}/hidden-projection.d.ts +0 -0
- /package/src/inference/pipelines/text/{operator-events.d.ts → observation/operator-events.d.ts} +0 -0
- /package/src/inference/pipelines/text/{stage-names.d.ts → observation/stage-names.d.ts} +0 -0
- /package/src/inference/pipelines/text/{stage-names.js → observation/stage-names.js} +0 -0
- /package/src/inference/pipelines/text/{tsir-fixture-writer.d.ts → observation/tsir-fixture-writer.d.ts} +0 -0
- /package/src/inference/pipelines/text/{generator → per-layer}/per-layer-buffer.js +0 -0
- /package/src/inference/pipelines/text/{finiteness-policy.js → policy/finiteness-policy.js} +0 -0
|
@@ -0,0 +1,164 @@
|
|
|
1
|
+
// HC Q4_0 up + refined sigmoid + branch; exact scalar gemv32 reduction.
|
|
2
|
+
// Cooperative quarter-block GEMV adapted from Pooled engine/wgsl/coop.js (6b4a24f7).
|
|
3
|
+
// Native GGUF blocks remain interleaved; no Pooled split-scale repacking.
|
|
4
|
+
// MIT License
|
|
5
|
+
//
|
|
6
|
+
// Copyright (c) 2026 Nehanth Narendrula
|
|
7
|
+
//
|
|
8
|
+
// Permission is hereby granted, free of charge, to any person obtaining a copy
|
|
9
|
+
// of this software and associated documentation files (the "Software"), to deal
|
|
10
|
+
// in the Software without restriction, including without limitation the rights
|
|
11
|
+
// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
|
12
|
+
// copies of the Software, and to permit persons to whom the Software is
|
|
13
|
+
// furnished to do so, subject to the following conditions:
|
|
14
|
+
//
|
|
15
|
+
// The above copyright notice and this permission notice shall be included in all
|
|
16
|
+
// copies or substantial portions of the Software.
|
|
17
|
+
//
|
|
18
|
+
// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
|
19
|
+
// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
20
|
+
// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
|
21
|
+
// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
22
|
+
// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
|
23
|
+
// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
|
24
|
+
// SOFTWARE.
|
|
25
|
+
override WORKGROUP_SIZE: u32 = 128u;
|
|
26
|
+
override STREAM_COUNT: u32 = 4u;
|
|
27
|
+
override THREADS_PER_COL_GEMV: u32 = 32u;
|
|
28
|
+
override BLOCK_VALUES: u32 = 32u;
|
|
29
|
+
override BLOCK_BYTES: u32 = 18u;
|
|
30
|
+
override QUANT_FORMAT: u32 = 0u;
|
|
31
|
+
|
|
32
|
+
struct Uniforms {
|
|
33
|
+
stream_width: u32,
|
|
34
|
+
rank: u32,
|
|
35
|
+
blocks: u32,
|
|
36
|
+
_pad0: u32,
|
|
37
|
+
}
|
|
38
|
+
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
39
|
+
@group(0) @binding(1) var<storage, read> input: array<f32>;
|
|
40
|
+
@group(0) @binding(2) var<storage, read> weights: array<u32>;
|
|
41
|
+
@group(0) @binding(3) var<storage, read> normed: array<f32>;
|
|
42
|
+
@group(0) @binding(4) var<storage, read_write> output: array<f32>;
|
|
43
|
+
|
|
44
|
+
fn q40_byte(offset: u32) -> u32 {
|
|
45
|
+
return (weights[offset / 4u] >> ((offset % 4u) * 8u)) & 255u;
|
|
46
|
+
}
|
|
47
|
+
fn packed_half(offset: u32) -> f32 {
|
|
48
|
+
return unpack2x16float(q40_byte(offset) | (q40_byte(offset + 1u) << 8u)).x;
|
|
49
|
+
}
|
|
50
|
+
fn packed_signed_byte(offset: u32) -> i32 {
|
|
51
|
+
return bitcast<i32>(q40_byte(offset) << 24u) >> 24;
|
|
52
|
+
}
|
|
53
|
+
fn q40_value(row: u32, k: u32, blocks: u32) -> f32 {
|
|
54
|
+
let offset = (row * blocks + k / BLOCK_VALUES) * BLOCK_BYTES;
|
|
55
|
+
let local = k % BLOCK_VALUES;
|
|
56
|
+
if (QUANT_FORMAT == 3u) {
|
|
57
|
+
let section = (local % 128u) / 32u;
|
|
58
|
+
let lane = local % 32u;
|
|
59
|
+
let low = q40_byte(offset + (local / 128u) * 64u + (section % 2u) * 32u + lane);
|
|
60
|
+
let high = q40_byte(offset + 128u + (local / 128u) * 32u + lane);
|
|
61
|
+
let value = ((low >> ((section / 2u) * 4u)) & 15u) | (((high >> (section * 2u)) & 3u) << 4u);
|
|
62
|
+
return packed_half(offset + 208u) * f32(packed_signed_byte(offset + 192u + local / 16u)) * f32(i32(value) - 32);
|
|
63
|
+
}
|
|
64
|
+
let scale = packed_half(offset);
|
|
65
|
+
if (QUANT_FORMAT == 2u) { return scale * f32(packed_signed_byte(offset + 2u + local)); }
|
|
66
|
+
if (QUANT_FORMAT == 4u) {
|
|
67
|
+
// Q5_0: f16 scale, 32 high bits (one per value), 16 bytes of low nibbles.
|
|
68
|
+
let high_bits = q40_byte(offset + 2u) | (q40_byte(offset + 3u) << 8u)
|
|
69
|
+
| (q40_byte(offset + 4u) << 16u) | (q40_byte(offset + 5u) << 24u);
|
|
70
|
+
let low = (q40_byte(offset + 6u + local % 16u) >> select(0u, 4u, local >= 16u)) & 15u;
|
|
71
|
+
return scale * f32(i32(low | (((high_bits >> local) & 1u) << 4u)) - 16);
|
|
72
|
+
}
|
|
73
|
+
let data_offset = select(2u,4u,QUANT_FORMAT == 1u);
|
|
74
|
+
let packed = q40_byte(offset + data_offset + local % 16u);
|
|
75
|
+
let shift = select(0u,4u,local >= 16u);
|
|
76
|
+
let value = (packed >> shift) & 15u;
|
|
77
|
+
if (QUANT_FORMAT == 1u) { return scale * f32(value) + packed_half(offset + 2u); }
|
|
78
|
+
return scale * f32(i32(value) - 8);
|
|
79
|
+
}
|
|
80
|
+
var<workgroup> partial: array<f32, WORKGROUP_SIZE>;
|
|
81
|
+
var<workgroup> projected: array<f32, STREAM_COUNT>;
|
|
82
|
+
var<workgroup> mixed: array<f32, STREAM_COUNT>;
|
|
83
|
+
var<workgroup> products: array<f32, STREAM_COUNT>;
|
|
84
|
+
|
|
85
|
+
fn exp_refined(x: f32) -> f32 {
|
|
86
|
+
if (abs(x) > 80.0) {
|
|
87
|
+
return exp(x);
|
|
88
|
+
}
|
|
89
|
+
let n = round(x * 1.4426950408889634);
|
|
90
|
+
var r = x;
|
|
91
|
+
if (abs(x) >= 0.000000059604644775390625) {
|
|
92
|
+
let bits = bitcast<u32>(abs(x));
|
|
93
|
+
let mantissa = (bits & 0x007fffffu) | 0x00800000u;
|
|
94
|
+
let shift = i32((bits >> 23u) & 255u) - 118;
|
|
95
|
+
var input_fixed = 0u;
|
|
96
|
+
if (shift >= 0) {
|
|
97
|
+
input_fixed = mantissa << u32(shift);
|
|
98
|
+
} else {
|
|
99
|
+
let right = u32(-shift);
|
|
100
|
+
input_fixed = (mantissa + (1u << (right - 1u))) >> right;
|
|
101
|
+
}
|
|
102
|
+
let power_fixed = u32(abs(n)) * 2977044472u;
|
|
103
|
+
let residual = select(power_fixed - input_fixed, input_fixed - power_fixed, x >= 0.0);
|
|
104
|
+
r = f32(bitcast<i32>(residual)) * 0.00000000023283064365386962890625;
|
|
105
|
+
}
|
|
106
|
+
var p = 0.0001984126984126984;
|
|
107
|
+
p = fma(p, r, 0.001388888888888889);
|
|
108
|
+
p = fma(p, r, 0.008333333333333333);
|
|
109
|
+
p = fma(p, r, 0.041666666666666664);
|
|
110
|
+
p = fma(p, r, 0.16666666666666666);
|
|
111
|
+
p = fma(p, r, 0.5);
|
|
112
|
+
p = fma(p, r, 1.0);
|
|
113
|
+
p = fma(p, r, 1.0);
|
|
114
|
+
return p * bitcast<f32>(u32(i32(n) + 127) << 23u);
|
|
115
|
+
}
|
|
116
|
+
|
|
117
|
+
fn reciprocal_refined(x: f32) -> f32 {
|
|
118
|
+
let estimate = 1.0 / x;
|
|
119
|
+
return fma(estimate, fma(-x, estimate, 1.0), estimate);
|
|
120
|
+
}
|
|
121
|
+
|
|
122
|
+
fn divide_refined(x: f32, y: f32) -> f32 {
|
|
123
|
+
let inverse = reciprocal_refined(y);
|
|
124
|
+
let quotient = x * inverse;
|
|
125
|
+
return fma(fma(-quotient, y, x), inverse, quotient);
|
|
126
|
+
}
|
|
127
|
+
|
|
128
|
+
fn sigmoid_refined(x: f32) -> f32 {
|
|
129
|
+
let z = exp_refined(-abs(x));
|
|
130
|
+
let reciprocal = reciprocal_refined(1.0 + z);
|
|
131
|
+
return select(z * reciprocal, reciprocal, x >= 0.0);
|
|
132
|
+
}
|
|
133
|
+
|
|
134
|
+
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
135
|
+
fn main(@builtin(workgroup_id) group: vec3<u32>, @builtin(num_workgroups) grid: vec3<u32>,
|
|
136
|
+
@builtin(local_invocation_index) lane: u32) {
|
|
137
|
+
let column = group.x + group.y * grid.x;
|
|
138
|
+
if (column >= u.stream_width) { return; }
|
|
139
|
+
let stream = lane / THREADS_PER_COL_GEMV;
|
|
140
|
+
let thread = lane % THREADS_PER_COL_GEMV;
|
|
141
|
+
let row = stream * u.stream_width + column;
|
|
142
|
+
var sum = 0.0f;
|
|
143
|
+
for (var k = thread; k < u.rank; k += THREADS_PER_COL_GEMV) {
|
|
144
|
+
sum = fma(input[k], q40_value(row, k, u.blocks), sum);
|
|
145
|
+
}
|
|
146
|
+
partial[lane] = sum;
|
|
147
|
+
workgroupBarrier();
|
|
148
|
+
for (var stride = THREADS_PER_COL_GEMV / 2u; stride > 0u; stride /= 2u) {
|
|
149
|
+
if (thread < stride) { partial[lane] += partial[lane + stride]; }
|
|
150
|
+
workgroupBarrier();
|
|
151
|
+
}
|
|
152
|
+
// The original GEMV's output store is observable to another invocation.
|
|
153
|
+
if (thread == 0u) { projected[stream] = partial[lane] * 1.0; }
|
|
154
|
+
workgroupBarrier();
|
|
155
|
+
if (thread == 1u) { mixed[stream] = sigmoid_refined(projected[stream]); }
|
|
156
|
+
workgroupBarrier();
|
|
157
|
+
if (thread == 2u) { products[stream] = mixed[stream] * normed[row]; }
|
|
158
|
+
workgroupBarrier();
|
|
159
|
+
if (lane == 3u) {
|
|
160
|
+
var branch = 0.0;
|
|
161
|
+
for (var s = 0u; s < STREAM_COUNT; s += 1u) { branch = branch + products[s]; }
|
|
162
|
+
output[column] = divide_refined(branch, f32(STREAM_COUNT));
|
|
163
|
+
}
|
|
164
|
+
}
|
|
@@ -2,7 +2,8 @@ import { getKernelRegistry } from './kernel-configs.js';
|
|
|
2
2
|
import { getCachedPipeline, getPipelineFast, getPipelineBindGroupLayout } from './pipeline-cache.js';
|
|
3
3
|
import { getDevice } from '../device.js';
|
|
4
4
|
import { dispatchKernel, dispatchIndirect, recordDispatchIndirect } from './dispatch.js';
|
|
5
|
-
import {
|
|
5
|
+
import { getCompletedUniformCache, releaseUniformBuffer } from '../uniform-cache.js';
|
|
6
|
+
import { createKernelUniformBuffer, getUniformByteLength, writeUniformsFromObject } from './uniform-utils.js';
|
|
6
7
|
import { acquireBuffer, releaseBuffer } from '../../memory/buffer-pool.js';
|
|
7
8
|
|
|
8
9
|
const dataBindingsByConfig = new WeakMap();
|
|
@@ -98,9 +99,15 @@ export async function unifiedKernelWrapper(
|
|
|
98
99
|
}
|
|
99
100
|
|
|
100
101
|
let uniformBuffer = null;
|
|
102
|
+
let dispatchAttempted = false;
|
|
101
103
|
try {
|
|
102
104
|
if (config.uniforms !== null) {
|
|
103
|
-
uniformBuffer = createKernelUniformBuffer(`${opName}_uniforms`, config, uniforms, recorder, device);
|
|
105
|
+
if (recorder) uniformBuffer = createKernelUniformBuffer(`${opName}_uniforms`, config, uniforms, recorder, device);
|
|
106
|
+
else {
|
|
107
|
+
const data = new ArrayBuffer(getUniformByteLength(config));
|
|
108
|
+
writeUniformsFromObject(new DataView(data), config, uniforms);
|
|
109
|
+
uniformBuffer = getCompletedUniformCache(device).getOrCreate(data, `${opName}_uniforms`);
|
|
110
|
+
}
|
|
104
111
|
const uniformBinding = config.bindings.find(binding => binding.type === 'uniform');
|
|
105
112
|
if (!uniformBinding) throw new Error(`Kernel ${opName}/${variant} declares uniforms without a binding.`);
|
|
106
113
|
bindGroupEntries.unshift({ binding: uniformBinding.index, resource: { buffer: uniformBuffer } });
|
|
@@ -122,6 +129,7 @@ export async function unifiedKernelWrapper(
|
|
|
122
129
|
const label = typeof dispatchLabel === 'string' && dispatchLabel.length > 0
|
|
123
130
|
? dispatchLabel
|
|
124
131
|
: opName;
|
|
132
|
+
dispatchAttempted = true;
|
|
125
133
|
if (workgroups && typeof workgroups === 'object' && workgroups.indirectBuffer) {
|
|
126
134
|
const indirectOffset = workgroups.indirectOffset ?? 0;
|
|
127
135
|
if (recorder) {
|
|
@@ -133,8 +141,11 @@ export async function unifiedKernelWrapper(
|
|
|
133
141
|
dispatchKernel(recorder ?? device, pipeline, bindGroup, workgroups, label);
|
|
134
142
|
}
|
|
135
143
|
} catch (error) {
|
|
136
|
-
if (!recorder) {
|
|
137
|
-
uniformBuffer
|
|
144
|
+
if (!recorder && uniformBuffer) {
|
|
145
|
+
const leased = uniformBuffer;
|
|
146
|
+
if (dispatchAttempted) device.queue.onSubmittedWorkDone().then(
|
|
147
|
+
() => releaseUniformBuffer(leased), () => releaseUniformBuffer(leased));
|
|
148
|
+
else releaseUniformBuffer(leased);
|
|
138
149
|
}
|
|
139
150
|
throw error;
|
|
140
151
|
}
|
|
@@ -145,16 +156,16 @@ export async function unifiedKernelWrapper(
|
|
|
145
156
|
try {
|
|
146
157
|
await device.queue.onSubmittedWorkDone();
|
|
147
158
|
} finally {
|
|
148
|
-
uniformBuffer
|
|
159
|
+
if (uniformBuffer) releaseUniformBuffer(uniformBuffer);
|
|
149
160
|
}
|
|
150
161
|
signal?.throwIfAborted();
|
|
151
162
|
} else if (!recorder && uniformBuffer) {
|
|
152
163
|
device.queue.onSubmittedWorkDone()
|
|
153
164
|
.then(() => {
|
|
154
|
-
uniformBuffer
|
|
165
|
+
releaseUniformBuffer(uniformBuffer);
|
|
155
166
|
})
|
|
156
167
|
.catch(() => {
|
|
157
|
-
uniformBuffer
|
|
168
|
+
releaseUniformBuffer(uniformBuffer);
|
|
158
169
|
});
|
|
159
170
|
}
|
|
160
171
|
|
|
@@ -255,9 +255,8 @@ export async function runLinearAttentionCoreGPU(qkvTensor, zTensor, aTensor, bTe
|
|
|
255
255
|
throw new Error('GPU prefix recovery requires recorded native GGUF F32 state and activations.');
|
|
256
256
|
}
|
|
257
257
|
if (layerState.sourceLayout === 'gguf-v1' && (qkvTensor?.precision === 'two-f32'
|
|
258
|
-
|| qkvTensor?.dtype !== 'f32' || layerState.recurrentStatePrecision !== 'f32'
|
|
259
|
-
|
|
260
|
-
throw new Error('Native GGUF GDN requires scalar F32 activations/state and SiLU norm gating.');
|
|
258
|
+
|| qkvTensor?.dtype !== 'f32' || layerState.recurrentStatePrecision !== 'f32')) {
|
|
259
|
+
throw new Error('Native GGUF GDN requires scalar F32 activations and state.');
|
|
261
260
|
}
|
|
262
261
|
if (qkvTensor?.precision === 'two-f32') {
|
|
263
262
|
const session = options.runtimeSession ?? getRuntimeConfig().inference.session;
|
|
@@ -338,7 +337,7 @@ export async function runLinearAttentionCoreGPU(qkvTensor, zTensor, aTensor, bTe
|
|
|
338
337
|
const shape = selected.variantMetadata;
|
|
339
338
|
if (!Number.isSafeInteger(shape?.headKDim) || layerState.headKDim !== shape.headKDim
|
|
340
339
|
|| stateComponents !== shape.stateComponents || layerState.headVDim > shape.maxHeadVDim) {
|
|
341
|
-
throw new Error('Register prefill requires its registered head width, F32 state/inputs and
|
|
340
|
+
throw new Error('Register prefill requires its registered head width, F32 state/inputs and declared norm gating.');
|
|
342
341
|
}
|
|
343
342
|
if (shape.sourceLayout === 'gguf-v1'
|
|
344
343
|
&& (!Number.isSafeInteger(numTokens) || !Number.isSafeInteger(layerState.numKHeads)
|
|
@@ -133,3 +133,17 @@ export function runRouteScatter(input, weights, options) {
|
|
|
133
133
|
export function recordRouteScatter(recorder, input, weights, options) {
|
|
134
134
|
return scatter(recorder, input, weights, options);
|
|
135
135
|
}
|
|
136
|
+
|
|
137
|
+
/** @type {typeof import('./moe-route.js').runRouteRemap} */
|
|
138
|
+
export async function runRouteRemap(indices, sourceToSlot, numRoutes, sourceNumExperts, slotCount, recorder = null) {
|
|
139
|
+
for (const [name, value] of Object.entries({ numRoutes, sourceNumExperts, slotCount })) positive(value, name);
|
|
140
|
+
extent(indices, numRoutes * 4, 'source route indices');
|
|
141
|
+
extent(sourceToSlot, sourceNumExperts * 4, 'source expert slot map');
|
|
142
|
+
const config = getKernelConfig('moe_route_remap', 'u32');
|
|
143
|
+
return withKernelOutput(recorder, null, numRoutes * 4, 'moe_route_remap', async output => {
|
|
144
|
+
await unifiedKernelWrapper('moe_route_remap', recorder, 'u32', [indices, sourceToSlot, output], {
|
|
145
|
+
num_routes: numRoutes, source_experts: sourceNumExperts, slot_count: slotCount,
|
|
146
|
+
}, [Math.ceil(numRoutes / config.workgroupSize[0]), 1, 1]);
|
|
147
|
+
return output;
|
|
148
|
+
});
|
|
149
|
+
}
|
|
@@ -105,66 +105,3 @@ fn gather_tokens(@builtin(global_invocation_id) gid: vec3<u32>) {
|
|
|
105
105
|
let src_idx = token_idx * hidden_size + dim_idx;
|
|
106
106
|
gathered[tid] = hidden_states[src_idx];
|
|
107
107
|
}
|
|
108
|
-
|
|
109
|
-
// Combined single-pass version for small models
|
|
110
|
-
// Each workgroup handles one expert
|
|
111
|
-
@compute @workgroup_size(WORKGROUP_SIZE_MAIN, 1, 1)
|
|
112
|
-
fn gather_single_pass(
|
|
113
|
-
@builtin(local_invocation_id) local_id: vec3<u32>,
|
|
114
|
-
@builtin(workgroup_id) wg_id: vec3<u32>
|
|
115
|
-
) {
|
|
116
|
-
let expert_idx = wg_id.x;
|
|
117
|
-
let thread_idx = local_id.x;
|
|
118
|
-
let hidden_size = u.hidden_size;
|
|
119
|
-
let num_tokens = u.num_tokens;
|
|
120
|
-
let top_k = u.top_k;
|
|
121
|
-
let max_tokens_per_expert = u.max_tokens_per_expert;
|
|
122
|
-
|
|
123
|
-
if (expert_idx >= u.num_experts) {
|
|
124
|
-
return;
|
|
125
|
-
}
|
|
126
|
-
|
|
127
|
-
// Phase 1: Count tokens for this expert (thread 0 only)
|
|
128
|
-
var token_count: u32 = 0u;
|
|
129
|
-
if (thread_idx == 0u) {
|
|
130
|
-
for (var t: u32 = 0u; t < num_tokens; t = t + 1u) {
|
|
131
|
-
for (var k: u32 = 0u; k < top_k; k = k + 1u) {
|
|
132
|
-
if (expert_indices[t * top_k + k] == expert_idx) {
|
|
133
|
-
if (token_count < max_tokens_per_expert) {
|
|
134
|
-
let map_base = expert_idx * max_tokens_per_expert * 2u + token_count * 2u;
|
|
135
|
-
token_map[map_base] = t;
|
|
136
|
-
token_map[map_base + 1u] = k;
|
|
137
|
-
token_count = token_count + 1u;
|
|
138
|
-
}
|
|
139
|
-
}
|
|
140
|
-
}
|
|
141
|
-
}
|
|
142
|
-
atomicStore(&token_counts[expert_idx], token_count);
|
|
143
|
-
}
|
|
144
|
-
|
|
145
|
-
workgroupBarrier();
|
|
146
|
-
|
|
147
|
-
// Phase 2: Gather (all threads participate)
|
|
148
|
-
let actual_count = atomicLoad(&token_counts[expert_idx]);
|
|
149
|
-
let elements_per_slot = hidden_size;
|
|
150
|
-
let total_work = actual_count * elements_per_slot;
|
|
151
|
-
let work_per_thread = (total_work + WORKGROUP_SIZE_MAIN - 1u) / WORKGROUP_SIZE_MAIN;
|
|
152
|
-
|
|
153
|
-
for (var i: u32 = 0u; i < work_per_thread; i = i + 1u) {
|
|
154
|
-
let work_idx = thread_idx * work_per_thread + i;
|
|
155
|
-
if (work_idx >= total_work) {
|
|
156
|
-
break;
|
|
157
|
-
}
|
|
158
|
-
|
|
159
|
-
let slot_idx = work_idx / hidden_size;
|
|
160
|
-
let dim_idx = work_idx % hidden_size;
|
|
161
|
-
|
|
162
|
-
let map_base = expert_idx * max_tokens_per_expert * 2u + slot_idx * 2u;
|
|
163
|
-
let token_idx = token_map[map_base];
|
|
164
|
-
|
|
165
|
-
let src_idx = token_idx * hidden_size + dim_idx;
|
|
166
|
-
let dst_idx = expert_idx * max_tokens_per_expert * hidden_size + slot_idx * hidden_size + dim_idx;
|
|
167
|
-
|
|
168
|
-
gathered[dst_idx] = hidden_states[src_idx];
|
|
169
|
-
}
|
|
170
|
-
}
|
|
@@ -0,0 +1,63 @@
|
|
|
1
|
+
// Native Q4_0 separate gate/up projection. One lane consumes a full 32-value block.
|
|
2
|
+
// Preserves packed scales and quantization; output remains [routes, gate then up].
|
|
3
|
+
override WORKGROUP_SIZE: u32 = 128u;
|
|
4
|
+
override COLS_PER_WG: u32 = 4u;
|
|
5
|
+
override THREADS_PER_COL_GEMV: u32 = 32u;
|
|
6
|
+
override TILE_SIZE: u32 = 16u;
|
|
7
|
+
override BLOCK_VALUES: u32 = 32u;
|
|
8
|
+
override BLOCK_BYTES: u32 = 18u;
|
|
9
|
+
override QUANT_FORMAT: u32 = 0u;
|
|
10
|
+
struct Uniforms {
|
|
11
|
+
M: u32, N: u32, K: u32, alpha: f32,
|
|
12
|
+
blocks: u32, num_experts: u32, top_k: u32, input_mode: u32,
|
|
13
|
+
}
|
|
14
|
+
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
15
|
+
@group(0) @binding(1) var<storage, read> input: array<f32>;
|
|
16
|
+
@group(0) @binding(2) var<storage, read> weights: array<u32>;
|
|
17
|
+
@group(0) @binding(3) var<storage, read_write> output: array<f32>;
|
|
18
|
+
@group(0) @binding(4) var<storage, read> route_indices: array<u32>;
|
|
19
|
+
@group(0) @binding(5) var<storage, read> up_weights: array<u32>;
|
|
20
|
+
var<workgroup> partial: array<f32, WORKGROUP_SIZE>;
|
|
21
|
+
fn source_byte(offset: u32, up: bool) -> u32 {
|
|
22
|
+
if (up) { return (up_weights[offset / 4u] >> ((offset % 4u) * 8u)) & 255u; }
|
|
23
|
+
return (weights[offset / 4u] >> ((offset % 4u) * 8u)) & 255u;
|
|
24
|
+
}
|
|
25
|
+
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
26
|
+
fn main(@builtin(workgroup_id) group: vec3<u32>, @builtin(num_workgroups) grid: vec3<u32>,
|
|
27
|
+
@builtin(local_invocation_index) lane: u32) {
|
|
28
|
+
let tiles_n = (u.N + COLS_PER_WG - 1u) / COLS_PER_WG;
|
|
29
|
+
let tile = group.x + group.y * grid.x;
|
|
30
|
+
let m = tile / tiles_n;
|
|
31
|
+
if (m >= u.M) { return; }
|
|
32
|
+
let expert = route_indices[m];
|
|
33
|
+
let token = select(m / u.top_k, m, u.input_mode == 1u);
|
|
34
|
+
let column = lane / THREADS_PER_COL_GEMV;
|
|
35
|
+
let thread = lane % THREADS_PER_COL_GEMV;
|
|
36
|
+
let n = (tile % tiles_n) * COLS_PER_WG + column;
|
|
37
|
+
var sum = 0.0f;
|
|
38
|
+
if (n < u.N && expert < u.num_experts) {
|
|
39
|
+
let row = expert * (u.N / 2u) + n % (u.N / 2u);
|
|
40
|
+
let up = n >= u.N / 2u;
|
|
41
|
+
for (var block = thread; block < u.blocks; block += THREADS_PER_COL_GEMV) {
|
|
42
|
+
let offset = (row * u.blocks + block) * BLOCK_BYTES;
|
|
43
|
+
let scale = unpack2x16float(source_byte(offset, up) | (source_byte(offset + 1u, up) << 8u)).x;
|
|
44
|
+
let base = token * u.K + block * BLOCK_VALUES;
|
|
45
|
+
for (var j = 0u; j < 16u; j += 4u) {
|
|
46
|
+
let packed = vec4<u32>(source_byte(offset + 2u + j, up), source_byte(offset + 3u + j, up),
|
|
47
|
+
source_byte(offset + 4u + j, up), source_byte(offset + 5u + j, up));
|
|
48
|
+
let low = (vec4<f32>(packed & vec4<u32>(15u)) - vec4<f32>(8.0)) * scale;
|
|
49
|
+
let high = (vec4<f32>(packed >> vec4<u32>(4u)) - vec4<f32>(8.0)) * scale;
|
|
50
|
+
let xLow = vec4<f32>(input[base + j], input[base + j + 1u], input[base + j + 2u], input[base + j + 3u]);
|
|
51
|
+
let xHigh = vec4<f32>(input[base + j + 16u], input[base + j + 17u], input[base + j + 18u], input[base + j + 19u]);
|
|
52
|
+
sum += dot(xLow, low) + dot(xHigh, high);
|
|
53
|
+
}
|
|
54
|
+
}
|
|
55
|
+
}
|
|
56
|
+
partial[lane] = sum;
|
|
57
|
+
workgroupBarrier();
|
|
58
|
+
for (var stride = THREADS_PER_COL_GEMV / 2u; stride > 0u; stride /= 2u) {
|
|
59
|
+
if (thread < stride) { partial[lane] += partial[lane + stride]; }
|
|
60
|
+
workgroupBarrier();
|
|
61
|
+
}
|
|
62
|
+
if (thread == 0u && n < u.N) { output[m * u.N + n] = partial[lane] * u.alpha; }
|
|
63
|
+
}
|
|
@@ -0,0 +1,59 @@
|
|
|
1
|
+
// Native Q4_1 routed projection. One lane consumes a full 32-value block.
|
|
2
|
+
// Preserves original scale and minimum per weight; route reduction remains separate.
|
|
3
|
+
override WORKGROUP_SIZE: u32 = 128u;
|
|
4
|
+
override COLS_PER_WG: u32 = 4u;
|
|
5
|
+
override THREADS_PER_COL_GEMV: u32 = 32u;
|
|
6
|
+
override TILE_SIZE: u32 = 16u;
|
|
7
|
+
override BLOCK_VALUES: u32 = 32u;
|
|
8
|
+
override BLOCK_BYTES: u32 = 20u;
|
|
9
|
+
override QUANT_FORMAT: u32 = 1u;
|
|
10
|
+
struct Uniforms {
|
|
11
|
+
M: u32, N: u32, K: u32, alpha: f32,
|
|
12
|
+
blocks: u32, num_experts: u32, top_k: u32, input_mode: u32,
|
|
13
|
+
}
|
|
14
|
+
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
15
|
+
@group(0) @binding(1) var<storage, read> input: array<f32>;
|
|
16
|
+
@group(0) @binding(2) var<storage, read> weights: array<u32>;
|
|
17
|
+
@group(0) @binding(3) var<storage, read_write> output: array<f32>;
|
|
18
|
+
@group(0) @binding(4) var<storage, read> route_indices: array<u32>;
|
|
19
|
+
var<workgroup> partial: array<f32, WORKGROUP_SIZE>;
|
|
20
|
+
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
21
|
+
fn main(@builtin(workgroup_id) group: vec3<u32>, @builtin(num_workgroups) grid: vec3<u32>,
|
|
22
|
+
@builtin(local_invocation_index) lane: u32) {
|
|
23
|
+
let tiles_n = (u.N + COLS_PER_WG - 1u) / COLS_PER_WG;
|
|
24
|
+
let tile = group.x + group.y * grid.x;
|
|
25
|
+
let m = tile / tiles_n;
|
|
26
|
+
if (m >= u.M) { return; }
|
|
27
|
+
let expert = route_indices[m];
|
|
28
|
+
let token = select(m / u.top_k, m, u.input_mode == 1u);
|
|
29
|
+
let column = lane / THREADS_PER_COL_GEMV;
|
|
30
|
+
let thread = lane % THREADS_PER_COL_GEMV;
|
|
31
|
+
let n = (tile % tiles_n) * COLS_PER_WG + column;
|
|
32
|
+
var sum = 0.0f;
|
|
33
|
+
if (n < u.N && expert < u.num_experts) {
|
|
34
|
+
let row = expert * u.N + n;
|
|
35
|
+
for (var block = thread; block < u.blocks; block += THREADS_PER_COL_GEMV) {
|
|
36
|
+
let offset = (row * u.blocks + block) * BLOCK_BYTES;
|
|
37
|
+
let parameters = unpack2x16float(weights[offset / 4u]);
|
|
38
|
+
let scale = parameters.x;
|
|
39
|
+
let minimum = parameters.y;
|
|
40
|
+
let base = token * u.K + block * BLOCK_VALUES;
|
|
41
|
+
for (var j = 0u; j < 16u; j += 4u) {
|
|
42
|
+
let packed_word = weights[(offset + 4u + j) / 4u];
|
|
43
|
+
let packed = vec4<u32>(packed_word, packed_word >> 8u, packed_word >> 16u, packed_word >> 24u) & vec4<u32>(255u);
|
|
44
|
+
let low = vec4<f32>(packed & vec4<u32>(15u)) * scale + vec4<f32>(minimum);
|
|
45
|
+
let high = vec4<f32>(packed >> vec4<u32>(4u)) * scale + vec4<f32>(minimum);
|
|
46
|
+
let xLow = vec4<f32>(input[base + j], input[base + j + 1u], input[base + j + 2u], input[base + j + 3u]);
|
|
47
|
+
let xHigh = vec4<f32>(input[base + j + 16u], input[base + j + 17u], input[base + j + 18u], input[base + j + 19u]);
|
|
48
|
+
sum += dot(xLow, low) + dot(xHigh, high);
|
|
49
|
+
}
|
|
50
|
+
}
|
|
51
|
+
}
|
|
52
|
+
partial[lane] = sum;
|
|
53
|
+
workgroupBarrier();
|
|
54
|
+
for (var stride = THREADS_PER_COL_GEMV / 2u; stride > 0u; stride /= 2u) {
|
|
55
|
+
if (thread < stride) { partial[lane] += partial[lane + stride]; }
|
|
56
|
+
workgroupBarrier();
|
|
57
|
+
}
|
|
58
|
+
if (thread == 0u && n < u.N) { output[m * u.N + n] = partial[lane] * u.alpha; }
|
|
59
|
+
}
|
|
@@ -0,0 +1,22 @@
|
|
|
1
|
+
override WORKGROUP_SIZE: u32 = 256u;
|
|
2
|
+
struct RemapUniforms {
|
|
3
|
+
num_routes: u32,
|
|
4
|
+
source_experts: u32,
|
|
5
|
+
slot_count: u32,
|
|
6
|
+
_pad0: u32,
|
|
7
|
+
}
|
|
8
|
+
@group(0) @binding(0) var<uniform> u: RemapUniforms;
|
|
9
|
+
@group(0) @binding(1) var<storage, read> source_indices: array<u32>;
|
|
10
|
+
@group(0) @binding(2) var<storage, read> source_to_slot: array<u32>;
|
|
11
|
+
@group(0) @binding(3) var<storage, read_write> output: array<u32>;
|
|
12
|
+
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
13
|
+
fn remap_routes(@builtin(global_invocation_id) gid: vec3<u32>) {
|
|
14
|
+
if (gid.x >= u.num_routes) { return; }
|
|
15
|
+
let source = source_indices[gid.x];
|
|
16
|
+
var slot = 0xffffffffu;
|
|
17
|
+
if (source < u.source_experts) {
|
|
18
|
+
let candidate = source_to_slot[source];
|
|
19
|
+
if (candidate < u.slot_count) { slot = candidate; }
|
|
20
|
+
}
|
|
21
|
+
output[gid.x] = slot;
|
|
22
|
+
}
|
|
@@ -13,7 +13,7 @@ const GATE_CONSTANTS = Object.freeze({ WORKGROUP_SIZE: 64 });
|
|
|
13
13
|
const RECURRENT_CONSTANTS = Object.freeze({ WORKGROUP_SIZE: 128 });
|
|
14
14
|
|
|
15
15
|
/** The paired gate phase is selected by rule and must declare the model's activation.
|
|
16
|
-
* @param {import(
|
|
16
|
+
* @param {import("../../config/schema/execution/inference.schema.js").LinearNormGateActivation} normGateActivation */
|
|
17
17
|
function resolvePairedGateVariant(normGateActivation) {
|
|
18
18
|
return resolveGatedDeltaNormGateKernel('gated_delta_recurrent', 'gatedDeltaRecurrent', 'pairedGateVariant',
|
|
19
19
|
{ normGateActivation }).variant;
|
|
@@ -363,7 +363,13 @@ export async function createPipeline(
|
|
|
363
363
|
);
|
|
364
364
|
|
|
365
365
|
// Compile or reuse shader module
|
|
366
|
-
|
|
366
|
+
let shaderModule;
|
|
367
|
+
try {
|
|
368
|
+
shaderModule = await getShaderModule(device, config.shaderFile, `${operation}_${variant}`);
|
|
369
|
+
} catch (error) {
|
|
370
|
+
// Name the kernel the module was being compiled for; the cause keeps the compiler's report.
|
|
371
|
+
throw new Error(`${error.message} Kernel ${operation}/${variant}, file ${config.shaderFile}, entry point ${config.entryPoint}.`, { cause: error });
|
|
372
|
+
}
|
|
367
373
|
assertDeviceValid(device, epoch, requireCurrentDevice);
|
|
368
374
|
if (generation !== cacheGeneration) throw new Error('Pipeline preparation invalidated by cache reset.');
|
|
369
375
|
cacheKey = buildPipelineCacheKey(operation, variant, resolvedConstants, bindGroupLayout, device, getShaderModuleIdentity(shaderModule));
|
|
@@ -392,7 +398,17 @@ export async function createPipeline(
|
|
|
392
398
|
};
|
|
393
399
|
const task = Promise.resolve().then(async () => {
|
|
394
400
|
assertDeviceValid(device, epoch, requireCurrentDevice);
|
|
395
|
-
|
|
401
|
+
let pipeline;
|
|
402
|
+
try {
|
|
403
|
+
pipeline = await device.createComputePipelineAsync(pipelineDescriptor);
|
|
404
|
+
} catch (cause) {
|
|
405
|
+
throw new Error(
|
|
406
|
+
`GPU pipeline creation failed for ${operation}/${variant}: file ${config.shaderFile}, ` +
|
|
407
|
+
`entry point ${config.entryPoint}, source ${provenance.shaderHash}, ` +
|
|
408
|
+
`constants ${JSON.stringify(provenance.constants)}. ${cause?.message ?? String(cause)}`,
|
|
409
|
+
{ cause }
|
|
410
|
+
);
|
|
411
|
+
}
|
|
396
412
|
assertDeviceValid(device, epoch, requireCurrentDevice);
|
|
397
413
|
pipelineProvenance.set(pipeline, provenance);
|
|
398
414
|
if (generation === cacheGeneration && pipelineTasks.get(cacheKey) === task) pipelineCache.set(cacheKey, pipeline);
|
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
|
|