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 +1,86 @@
|
|
|
1
|
+
import { cloneGraph } from './dtype.js';
|
|
2
|
+
import { ATTENTION_F16KV_TO_F32_ENTRIES, deriveMappedKernelEntry, deriveQ4DecodeF32ActivationKernelEntry, deriveQ4PrefillF32ActivationKernelEntry, hasKernelFile } from './validation.js';
|
|
3
|
+
|
|
4
|
+
const F16_TO_F32_CORRECTNESS_FALLBACK_MAP = new Map([
|
|
5
|
+
['rmsnorm_f16.wgsl', 'rmsnorm.wgsl'],
|
|
6
|
+
['rope_f16.wgsl', 'rope.wgsl'],
|
|
7
|
+
['residual_f16.wgsl', 'residual.wgsl'],
|
|
8
|
+
['gelu_f16.wgsl', 'gelu.wgsl'],
|
|
9
|
+
['silu_f16.wgsl', 'silu.wgsl'],
|
|
10
|
+
['sample_f16.wgsl', 'sample.wgsl'],
|
|
11
|
+
['gather_f16.wgsl', 'gather.wgsl'],
|
|
12
|
+
['gather_f16_f16_out.wgsl', 'gather.wgsl'],
|
|
13
|
+
['gather_f16_vec4_f16_out.wgsl', 'gather.wgsl'],
|
|
14
|
+
['matmul_gemv_subgroup_f16a.wgsl', 'matmul_gemv_subgroup.wgsl'],
|
|
15
|
+
['matmul_f16.wgsl', 'matmul_f16w_f32a.wgsl'],
|
|
16
|
+
['matmul_f16_tiled.wgsl', 'matmul_f16w_f32a_tiled.wgsl'],
|
|
17
|
+
['attention_decode_online_f16.wgsl', 'attention_streaming.wgsl'],
|
|
18
|
+
['attention_decode_chunked_f16.wgsl', 'attention_streaming.wgsl'],
|
|
19
|
+
['attention_small_f16.wgsl', 'attention_small.wgsl'],
|
|
20
|
+
['attention_streaming_f16.wgsl', 'attention_streaming.wgsl'],
|
|
21
|
+
...ATTENTION_F16KV_TO_F32_ENTRIES,
|
|
22
|
+
]);
|
|
23
|
+
|
|
24
|
+
/*
|
|
25
|
+
* Full f32 widening: every shader that uses `enable f16;` is replaced with a
|
|
26
|
+
* pure-f32 equivalent. Used when the GPU cannot compile any f16 WGSL at all.
|
|
27
|
+
* Covers f16-activation, f16-weight (f16w), and f16-KV (f16kv) kernels.
|
|
28
|
+
*/
|
|
29
|
+
|
|
30
|
+
/*
|
|
31
|
+
* Widen all f16-activation shaders to f32-activation equivalents.
|
|
32
|
+
*
|
|
33
|
+
* Returns null if the graph contains fused_ffn_f16.wgsl (no direct f32
|
|
34
|
+
* equivalent exists) or if no f16 activation shaders are present.
|
|
35
|
+
*
|
|
36
|
+
* NOTE: The caller is responsible for also updating session.activationDtype
|
|
37
|
+
* to reflect the widened dtype.
|
|
38
|
+
*
|
|
39
|
+
*/
|
|
40
|
+
|
|
41
|
+
/*
|
|
42
|
+
* Widen an f16 execution graph onto the stable f32 correctness lane used for
|
|
43
|
+
* alternate-plan recovery after finiteness failure.
|
|
44
|
+
*
|
|
45
|
+
*/
|
|
46
|
+
/** @type {typeof import('../execution-graph-transforms.js').widenToF32CorrectnessFallback} */
|
|
47
|
+
export function widenToF32CorrectnessFallback(graph, ctx) {
|
|
48
|
+
if (hasKernelFile(graph, 'fused_ffn_f16.wgsl')) {
|
|
49
|
+
return null;
|
|
50
|
+
}
|
|
51
|
+
|
|
52
|
+
const hasTargetShader = Object.values(graph.kernels).some(
|
|
53
|
+
(entry) => F16_TO_F32_CORRECTNESS_FALLBACK_MAP.has(entry.kernel)
|
|
54
|
+
);
|
|
55
|
+
if (!hasTargetShader) {
|
|
56
|
+
return null;
|
|
57
|
+
}
|
|
58
|
+
|
|
59
|
+
const result = cloneGraph(graph);
|
|
60
|
+
for (const [key, entry] of Object.entries(result.kernels)) {
|
|
61
|
+
const q4FallbackEntry = deriveQ4DecodeF32ActivationKernelEntry(entry)
|
|
62
|
+
?? deriveQ4PrefillF32ActivationKernelEntry(entry);
|
|
63
|
+
if (q4FallbackEntry) {
|
|
64
|
+
result.kernels[key] = q4FallbackEntry;
|
|
65
|
+
continue;
|
|
66
|
+
}
|
|
67
|
+
const replacement = F16_TO_F32_CORRECTNESS_FALLBACK_MAP.get(entry.kernel);
|
|
68
|
+
if (replacement !== undefined) {
|
|
69
|
+
result.kernels[key] = deriveMappedKernelEntry(entry, replacement);
|
|
70
|
+
}
|
|
71
|
+
}
|
|
72
|
+
return result;
|
|
73
|
+
}
|
|
74
|
+
|
|
75
|
+
/*
|
|
76
|
+
* Narrow f32-activation shaders back onto their f16-activation equivalents.
|
|
77
|
+
*
|
|
78
|
+
* Returns null if the graph has no supported f32-activation kernels to swap or
|
|
79
|
+
* if the runtime did not explicitly request f16 activations on an f16-capable
|
|
80
|
+
* GPU.
|
|
81
|
+
*
|
|
82
|
+
*/
|
|
83
|
+
|
|
84
|
+
// =============================================================================
|
|
85
|
+
// Transform: swapPrefillAttention
|
|
86
|
+
// =============================================================================
|
|
@@ -0,0 +1,116 @@
|
|
|
1
|
+
import { cloneGraph, deriveKernelEntry } from './dtype.js';
|
|
2
|
+
|
|
3
|
+
const PREFILL_ATTENTION_PAIRS = new Map([
|
|
4
|
+
['attention_streaming_f16kv.wgsl', 'attention_small_f16kv.wgsl'],
|
|
5
|
+
['attention_small_f16kv.wgsl', 'attention_streaming_f16kv.wgsl'],
|
|
6
|
+
['attention_streaming_f16.wgsl', 'attention_small_f16.wgsl'],
|
|
7
|
+
['attention_small_f16.wgsl', 'attention_streaming_f16.wgsl'],
|
|
8
|
+
]);
|
|
9
|
+
|
|
10
|
+
/** @param {Parameters<typeof import('../execution-graph-transforms.js').swapPrefillAttention>[0]} graph
|
|
11
|
+
* @param {string} kernelKey */
|
|
12
|
+
function graphUsesKernelKeyInPrefill(graph, kernelKey) {
|
|
13
|
+
for (const entry of graph.prefill || []) {
|
|
14
|
+
if (Array.isArray(entry)) {
|
|
15
|
+
if (entry[1] === kernelKey) {
|
|
16
|
+
return true;
|
|
17
|
+
}
|
|
18
|
+
continue;
|
|
19
|
+
}
|
|
20
|
+
if (!entry || typeof entry !== 'object' || !Array.isArray(entry.steps)) {
|
|
21
|
+
continue;
|
|
22
|
+
}
|
|
23
|
+
for (const step of entry.steps) {
|
|
24
|
+
if (Array.isArray(step) && step[1] === kernelKey) {
|
|
25
|
+
return true;
|
|
26
|
+
}
|
|
27
|
+
}
|
|
28
|
+
}
|
|
29
|
+
return false;
|
|
30
|
+
}
|
|
31
|
+
|
|
32
|
+
/*
|
|
33
|
+
* Swap prefill attention kernel between streaming and small variants.
|
|
34
|
+
*
|
|
35
|
+
* The `opts` parameter specifies the direction:
|
|
36
|
+
* { from: 'attention_streaming_f16kv.wgsl', to: 'attention_small_f16kv.wgsl' }
|
|
37
|
+
*
|
|
38
|
+
* If `from`/`to` are not provided, uses the bidirectional pair map.
|
|
39
|
+
* Returns null if no matching prefill attention kernel is found.
|
|
40
|
+
*
|
|
41
|
+
*/
|
|
42
|
+
/** @type {typeof import('../execution-graph-transforms.js').swapPrefillAttention} */
|
|
43
|
+
export function swapPrefillAttention(graph, ctx, opts) {
|
|
44
|
+
const from = opts?.from;
|
|
45
|
+
const to = opts?.to;
|
|
46
|
+
|
|
47
|
+
const result = cloneGraph(graph);
|
|
48
|
+
let changed = false;
|
|
49
|
+
|
|
50
|
+
for (const [key, entry] of Object.entries(result.kernels)) {
|
|
51
|
+
let target;
|
|
52
|
+
|
|
53
|
+
if (from && to) {
|
|
54
|
+
// Explicit direction: only swap if the kernel matches `from`
|
|
55
|
+
if (entry.kernel === from) {
|
|
56
|
+
target = to;
|
|
57
|
+
}
|
|
58
|
+
} else {
|
|
59
|
+
// Bidirectional: use pair map
|
|
60
|
+
target = PREFILL_ATTENTION_PAIRS.get(entry.kernel);
|
|
61
|
+
}
|
|
62
|
+
|
|
63
|
+
if (target !== undefined) {
|
|
64
|
+
const usedInPrefill = graphUsesKernelKeyInPrefill(graph, key);
|
|
65
|
+
if (usedInPrefill) {
|
|
66
|
+
result.kernels[key] = deriveKernelEntry(entry, target, entry.entry);
|
|
67
|
+
changed = true;
|
|
68
|
+
}
|
|
69
|
+
}
|
|
70
|
+
}
|
|
71
|
+
|
|
72
|
+
return changed ? result : null;
|
|
73
|
+
}
|
|
74
|
+
|
|
75
|
+
// =============================================================================
|
|
76
|
+
// Transform: useHead256PrefillAttention
|
|
77
|
+
// =============================================================================
|
|
78
|
+
|
|
79
|
+
/*
|
|
80
|
+
* Promote small-tile prefill attention onto the fixed 256-dim shared-block kernel.
|
|
81
|
+
*
|
|
82
|
+
*/
|
|
83
|
+
/** @type {typeof import('../execution-graph-transforms.js').useHead256SmallPrefillAttention} */
|
|
84
|
+
export function useHead256SmallPrefillAttention(graph, ctx) {
|
|
85
|
+
return swapPrefillAttention(graph, ctx, {
|
|
86
|
+
from: 'attention_small_f16kv.wgsl',
|
|
87
|
+
to: 'attention_head256_f16kv.wgsl',
|
|
88
|
+
});
|
|
89
|
+
}
|
|
90
|
+
|
|
91
|
+
/*
|
|
92
|
+
* Promote prefill attention onto the fixed 256-dim shared-block kernel.
|
|
93
|
+
*
|
|
94
|
+
*/
|
|
95
|
+
/** @type {typeof import('../execution-graph-transforms.js').useHead256PrefillAttention} */
|
|
96
|
+
export function useHead256PrefillAttention(graph, ctx) {
|
|
97
|
+
let current = graph;
|
|
98
|
+
let changed = false;
|
|
99
|
+
|
|
100
|
+
const smallRemap = useHead256SmallPrefillAttention(current, ctx);
|
|
101
|
+
if (smallRemap) {
|
|
102
|
+
current = smallRemap;
|
|
103
|
+
changed = true;
|
|
104
|
+
}
|
|
105
|
+
|
|
106
|
+
const streamingRemap = swapPrefillAttention(current, ctx, {
|
|
107
|
+
from: 'attention_streaming_f16kv.wgsl',
|
|
108
|
+
to: 'attention_head256_f16kv.wgsl',
|
|
109
|
+
});
|
|
110
|
+
if (streamingRemap) {
|
|
111
|
+
current = streamingRemap;
|
|
112
|
+
changed = true;
|
|
113
|
+
}
|
|
114
|
+
|
|
115
|
+
return changed ? current : null;
|
|
116
|
+
}
|
|
@@ -1,8 +1,12 @@
|
|
|
1
|
+
import { widenToF32CorrectnessFallback } from './execution-graph/correctness-fallback.js';
|
|
2
|
+
export { widenToF32CorrectnessFallback } from './execution-graph/correctness-fallback.js';
|
|
3
|
+
import { swapPrefillAttention, useHead256SmallPrefillAttention, useHead256PrefillAttention } from './execution-graph/prefill-attention.js';
|
|
4
|
+
export { swapPrefillAttention, useHead256SmallPrefillAttention, useHead256PrefillAttention } from './execution-graph/prefill-attention.js';
|
|
1
5
|
import { widenAttentionKvToF32 } from './execution-graph/kv-storage.js';
|
|
2
6
|
export { widenAttentionKvToF32 } from './execution-graph/kv-storage.js';
|
|
3
7
|
import { F16_TO_F32_ACTIVATION_MAP, F32_TO_F16_ACTIVATION_MAP, KERNEL_FILE_PRECISION_PATCHES, cloneGraph, deriveKernelEntry, deriveKernelKey, deriveLinearDecodeF16KernelEntry, narrowToF16Activations, useQwen36F16Activations, useQwenDecodeF16Matmuls, useQwenF16PrimaryMatmuls } from './execution-graph/dtype.js';
|
|
4
8
|
import { LAYER_PROJECTION_OPS, useGemma4Int4PleAf16Activations, useGemma4Int4PleSelectiveF16Decode, useGemma4TextF16ActivationsForLane } from './execution-graph/fusion.js';
|
|
5
|
-
import {
|
|
9
|
+
import { findKernelKeyByFileAndEntry, remapDenseQ4KPrefillToQ4Native, remapStepKeys, removeSubgroups, widenProjectionWeightsToF32, widenToF32Activations } from './execution-graph/validation.js';
|
|
6
10
|
export { remapDenseQ4KPrefillToQ4Native, removeSubgroups, widenProjectionWeightsToF32, widenToF32Activations } from './execution-graph/validation.js';
|
|
7
11
|
export { useGemma4Int4PleAf16Activations, useGemma4Int4PleSelectiveF16Decode } from './execution-graph/fusion.js';
|
|
8
12
|
export { narrowToF16Activations, useQwen36F16Activations, useQwenDecodeF16Matmuls, useQwenF16PrimaryMatmuls } from './execution-graph/dtype.js';// =============================================================================
|
|
@@ -128,199 +132,6 @@ F32_TO_F16_ACTIVATION_MAP.delete('attention_head512_f16kv.wgsl');
|
|
|
128
132
|
* activations and KV-cache interactions onto the stable f32 execution lane.
|
|
129
133
|
* Used for alternate-plan recovery after finiteness failure.
|
|
130
134
|
*/
|
|
131
|
-
const F16_TO_F32_CORRECTNESS_FALLBACK_MAP = new Map([
|
|
132
|
-
['rmsnorm_f16.wgsl', 'rmsnorm.wgsl'],
|
|
133
|
-
['rope_f16.wgsl', 'rope.wgsl'],
|
|
134
|
-
['residual_f16.wgsl', 'residual.wgsl'],
|
|
135
|
-
['gelu_f16.wgsl', 'gelu.wgsl'],
|
|
136
|
-
['silu_f16.wgsl', 'silu.wgsl'],
|
|
137
|
-
['sample_f16.wgsl', 'sample.wgsl'],
|
|
138
|
-
['gather_f16.wgsl', 'gather.wgsl'],
|
|
139
|
-
['gather_f16_f16_out.wgsl', 'gather.wgsl'],
|
|
140
|
-
['gather_f16_vec4_f16_out.wgsl', 'gather.wgsl'],
|
|
141
|
-
['matmul_gemv_subgroup_f16a.wgsl', 'matmul_gemv_subgroup.wgsl'],
|
|
142
|
-
['matmul_f16.wgsl', 'matmul_f16w_f32a.wgsl'],
|
|
143
|
-
['matmul_f16_tiled.wgsl', 'matmul_f16w_f32a_tiled.wgsl'],
|
|
144
|
-
['attention_decode_online_f16.wgsl', 'attention_streaming.wgsl'],
|
|
145
|
-
['attention_decode_chunked_f16.wgsl', 'attention_streaming.wgsl'],
|
|
146
|
-
['attention_small_f16.wgsl', 'attention_small.wgsl'],
|
|
147
|
-
['attention_streaming_f16.wgsl', 'attention_streaming.wgsl'],
|
|
148
|
-
...ATTENTION_F16KV_TO_F32_ENTRIES,
|
|
149
|
-
]);
|
|
150
|
-
|
|
151
|
-
/*
|
|
152
|
-
* Full f32 widening: every shader that uses `enable f16;` is replaced with a
|
|
153
|
-
* pure-f32 equivalent. Used when the GPU cannot compile any f16 WGSL at all.
|
|
154
|
-
* Covers f16-activation, f16-weight (f16w), and f16-KV (f16kv) kernels.
|
|
155
|
-
*/
|
|
156
|
-
|
|
157
|
-
/*
|
|
158
|
-
* Widen all f16-activation shaders to f32-activation equivalents.
|
|
159
|
-
*
|
|
160
|
-
* Returns null if the graph contains fused_ffn_f16.wgsl (no direct f32
|
|
161
|
-
* equivalent exists) or if no f16 activation shaders are present.
|
|
162
|
-
*
|
|
163
|
-
* NOTE: The caller is responsible for also updating session.activationDtype
|
|
164
|
-
* to reflect the widened dtype.
|
|
165
|
-
*
|
|
166
|
-
*/
|
|
167
|
-
|
|
168
|
-
/*
|
|
169
|
-
* Widen an f16 execution graph onto the stable f32 correctness lane used for
|
|
170
|
-
* alternate-plan recovery after finiteness failure.
|
|
171
|
-
*
|
|
172
|
-
*/
|
|
173
|
-
export function widenToF32CorrectnessFallback(graph, ctx) {
|
|
174
|
-
if (hasKernelFile(graph, 'fused_ffn_f16.wgsl')) {
|
|
175
|
-
return null;
|
|
176
|
-
}
|
|
177
|
-
|
|
178
|
-
const hasTargetShader = Object.values(graph.kernels).some(
|
|
179
|
-
(entry) => F16_TO_F32_CORRECTNESS_FALLBACK_MAP.has(entry.kernel)
|
|
180
|
-
);
|
|
181
|
-
if (!hasTargetShader) {
|
|
182
|
-
return null;
|
|
183
|
-
}
|
|
184
|
-
|
|
185
|
-
const result = cloneGraph(graph);
|
|
186
|
-
for (const [key, entry] of Object.entries(result.kernels)) {
|
|
187
|
-
const q4FallbackEntry = deriveQ4DecodeF32ActivationKernelEntry(entry)
|
|
188
|
-
?? deriveQ4PrefillF32ActivationKernelEntry(entry);
|
|
189
|
-
if (q4FallbackEntry) {
|
|
190
|
-
result.kernels[key] = q4FallbackEntry;
|
|
191
|
-
continue;
|
|
192
|
-
}
|
|
193
|
-
const replacement = F16_TO_F32_CORRECTNESS_FALLBACK_MAP.get(entry.kernel);
|
|
194
|
-
if (replacement !== undefined) {
|
|
195
|
-
result.kernels[key] = deriveMappedKernelEntry(entry, replacement);
|
|
196
|
-
}
|
|
197
|
-
}
|
|
198
|
-
return result;
|
|
199
|
-
}
|
|
200
|
-
|
|
201
|
-
/*
|
|
202
|
-
* Narrow f32-activation shaders back onto their f16-activation equivalents.
|
|
203
|
-
*
|
|
204
|
-
* Returns null if the graph has no supported f32-activation kernels to swap or
|
|
205
|
-
* if the runtime did not explicitly request f16 activations on an f16-capable
|
|
206
|
-
* GPU.
|
|
207
|
-
*
|
|
208
|
-
*/
|
|
209
|
-
|
|
210
|
-
// =============================================================================
|
|
211
|
-
// Transform: swapPrefillAttention
|
|
212
|
-
// =============================================================================
|
|
213
|
-
|
|
214
|
-
const PREFILL_ATTENTION_PAIRS = new Map([
|
|
215
|
-
['attention_streaming_f16kv.wgsl', 'attention_small_f16kv.wgsl'],
|
|
216
|
-
['attention_small_f16kv.wgsl', 'attention_streaming_f16kv.wgsl'],
|
|
217
|
-
['attention_streaming_f16.wgsl', 'attention_small_f16.wgsl'],
|
|
218
|
-
['attention_small_f16.wgsl', 'attention_streaming_f16.wgsl'],
|
|
219
|
-
]);
|
|
220
|
-
|
|
221
|
-
function graphUsesKernelKeyInPrefill(graph, kernelKey) {
|
|
222
|
-
for (const entry of graph.prefill || []) {
|
|
223
|
-
if (Array.isArray(entry)) {
|
|
224
|
-
if (entry[1] === kernelKey) {
|
|
225
|
-
return true;
|
|
226
|
-
}
|
|
227
|
-
continue;
|
|
228
|
-
}
|
|
229
|
-
if (!entry || typeof entry !== 'object' || !Array.isArray(entry.steps)) {
|
|
230
|
-
continue;
|
|
231
|
-
}
|
|
232
|
-
for (const step of entry.steps) {
|
|
233
|
-
if (Array.isArray(step) && step[1] === kernelKey) {
|
|
234
|
-
return true;
|
|
235
|
-
}
|
|
236
|
-
}
|
|
237
|
-
}
|
|
238
|
-
return false;
|
|
239
|
-
}
|
|
240
|
-
|
|
241
|
-
/*
|
|
242
|
-
* Swap prefill attention kernel between streaming and small variants.
|
|
243
|
-
*
|
|
244
|
-
* The `opts` parameter specifies the direction:
|
|
245
|
-
* { from: 'attention_streaming_f16kv.wgsl', to: 'attention_small_f16kv.wgsl' }
|
|
246
|
-
*
|
|
247
|
-
* If `from`/`to` are not provided, uses the bidirectional pair map.
|
|
248
|
-
* Returns null if no matching prefill attention kernel is found.
|
|
249
|
-
*
|
|
250
|
-
*/
|
|
251
|
-
export function swapPrefillAttention(graph, ctx, opts) {
|
|
252
|
-
const from = opts?.from;
|
|
253
|
-
const to = opts?.to;
|
|
254
|
-
|
|
255
|
-
const result = cloneGraph(graph);
|
|
256
|
-
let changed = false;
|
|
257
|
-
|
|
258
|
-
for (const [key, entry] of Object.entries(result.kernels)) {
|
|
259
|
-
let target;
|
|
260
|
-
|
|
261
|
-
if (from && to) {
|
|
262
|
-
// Explicit direction: only swap if the kernel matches `from`
|
|
263
|
-
if (entry.kernel === from) {
|
|
264
|
-
target = to;
|
|
265
|
-
}
|
|
266
|
-
} else {
|
|
267
|
-
// Bidirectional: use pair map
|
|
268
|
-
target = PREFILL_ATTENTION_PAIRS.get(entry.kernel);
|
|
269
|
-
}
|
|
270
|
-
|
|
271
|
-
if (target !== undefined) {
|
|
272
|
-
const usedInPrefill = graphUsesKernelKeyInPrefill(graph, key);
|
|
273
|
-
if (usedInPrefill) {
|
|
274
|
-
result.kernels[key] = deriveKernelEntry(entry, target, entry.entry);
|
|
275
|
-
changed = true;
|
|
276
|
-
}
|
|
277
|
-
}
|
|
278
|
-
}
|
|
279
|
-
|
|
280
|
-
return changed ? result : null;
|
|
281
|
-
}
|
|
282
|
-
|
|
283
|
-
// =============================================================================
|
|
284
|
-
// Transform: useHead256PrefillAttention
|
|
285
|
-
// =============================================================================
|
|
286
|
-
|
|
287
|
-
/*
|
|
288
|
-
* Promote small-tile prefill attention onto the fixed 256-dim shared-block kernel.
|
|
289
|
-
*
|
|
290
|
-
*/
|
|
291
|
-
export function useHead256SmallPrefillAttention(graph, ctx) {
|
|
292
|
-
return swapPrefillAttention(graph, ctx, {
|
|
293
|
-
from: 'attention_small_f16kv.wgsl',
|
|
294
|
-
to: 'attention_head256_f16kv.wgsl',
|
|
295
|
-
});
|
|
296
|
-
}
|
|
297
|
-
|
|
298
|
-
/*
|
|
299
|
-
* Promote prefill attention onto the fixed 256-dim shared-block kernel.
|
|
300
|
-
*
|
|
301
|
-
*/
|
|
302
|
-
export function useHead256PrefillAttention(graph, ctx) {
|
|
303
|
-
let current = graph;
|
|
304
|
-
let changed = false;
|
|
305
|
-
|
|
306
|
-
const smallRemap = useHead256SmallPrefillAttention(current, ctx);
|
|
307
|
-
if (smallRemap) {
|
|
308
|
-
current = smallRemap;
|
|
309
|
-
changed = true;
|
|
310
|
-
}
|
|
311
|
-
|
|
312
|
-
const streamingRemap = swapPrefillAttention(current, ctx, {
|
|
313
|
-
from: 'attention_streaming_f16kv.wgsl',
|
|
314
|
-
to: 'attention_head256_f16kv.wgsl',
|
|
315
|
-
});
|
|
316
|
-
if (streamingRemap) {
|
|
317
|
-
current = streamingRemap;
|
|
318
|
-
changed = true;
|
|
319
|
-
}
|
|
320
|
-
|
|
321
|
-
return changed ? current : null;
|
|
322
|
-
}
|
|
323
|
-
|
|
324
135
|
// =============================================================================
|
|
325
136
|
// Transform: widenProjectionWeightsToF32
|
|
326
137
|
// =============================================================================
|
package/src/converter/core.js
CHANGED
|
@@ -1,52 +1,21 @@
|
|
|
1
|
-
import { prepareFP8SourceTensors, readFP8CompanionScales } from
|
|
2
|
-
import { appendRetainedSourceShards, partitionRetainedSourceTensors, preflightRetainedSourceShards } from
|
|
3
|
-
|
|
4
|
-
import {
|
|
5
|
-
|
|
6
|
-
|
|
7
|
-
|
|
8
|
-
|
|
9
|
-
|
|
10
|
-
|
|
11
|
-
} from
|
|
12
|
-
|
|
13
|
-
import {
|
|
14
|
-
|
|
15
|
-
|
|
16
|
-
|
|
17
|
-
|
|
18
|
-
|
|
19
|
-
import { log } from '../debug/index.js';
|
|
20
|
-
import {
|
|
21
|
-
getInferenceLayerPatternContractArtifact,
|
|
22
|
-
selectRuleValue,
|
|
23
|
-
} from '../rules/rule-registry.js';
|
|
24
|
-
import { createConverterConfig } from '../config/schema/index.js';
|
|
25
|
-
import { buildExecutionContractArtifact } from '../config/execution-contract-check.js';
|
|
26
|
-
import { buildManifestRequiredInferenceFieldsArtifact } from '../config/required-inference-fields-contract-check.js';
|
|
27
|
-
import { resolveEosTokenId } from './tokenizer-special-tokens.js';
|
|
28
|
-
import { inferBundledTokenizerBehaviorFlags } from '../inference/tokenizers/behavior-flags.js';
|
|
29
|
-
import {
|
|
30
|
-
normalizeQ4KLayout,
|
|
31
|
-
resolveManifestQuantization,
|
|
32
|
-
resolveEffectiveQuantizationInfo,
|
|
33
|
-
} from './quantization-info.js';
|
|
34
|
-
import {
|
|
35
|
-
float16ToFloat32,
|
|
36
|
-
float32ToFloat16,
|
|
37
|
-
quantizeToQ4KM,
|
|
38
|
-
quantizeToQ4KMRowWise,
|
|
39
|
-
quantizeToQ4KMColumnWise,
|
|
40
|
-
quantizeToInt4PerRowSymmetric,
|
|
41
|
-
} from './quantizer.js';
|
|
42
|
-
import { cloneJsonValue } from '../formats/clone-json.js';
|
|
43
|
-
import { SOURCE_PACKED_QUANT_DTYPES, isGemma4PerLayerEmbedTensor, normalizeModulesToNotConvert, normalizeStorageQuant, resolveQuantizeEmbeddings, transformTensorBytes } from './tensor-transform.js';
|
|
44
|
-
import { buildArtifactIdentity, extractArchitecture, getNestedTextConfig, sanitizeModelId } from './artifact-identity.js';
|
|
45
|
-
import { RDRR_VERSION, createManifest, modelHasMoETensors } from './manifest-builder.js';
|
|
46
|
-
import { normalizeCompressedTensorsW4A16, normalizeTensorName } from './tokenizer-contract.js';
|
|
47
|
-
export { RDRR_VERSION, buildBundledTokenizer, buildSentencepieceTokenizer, createManifest, inferEmbeddingOutputConfig, resolveBundledTokenizerVocabSize, resolveConvertedAt, resolveManifestMoEConfig, resolveManifestMultimodalConfig } from './manifest-builder.js';
|
|
48
|
-
export { extractArchitecture, sanitizeModelId } from './artifact-identity.js';
|
|
49
|
-
export { isGemma4PerLayerEmbedTensor, normalizeStorageQuant, resolveTensorTargetQuant, shouldQuantize, transformTensorBytes } from './tensor-transform.js';
|
|
1
|
+
import { prepareFP8SourceTensors, readFP8CompanionScales } from "./fp8-source.js";
|
|
2
|
+
import { appendRetainedSourceShards, partitionRetainedSourceTensors, preflightRetainedSourceShards } from "./retained-source-shards.js";
|
|
3
|
+
import { SHARD_SIZE as SCHEMA_SHARD_SIZE, ConversionStage as SchemaConversionStage, formatBytes, createConverterConfig } from "../config/schema/index.js";
|
|
4
|
+
import { generateShardFilename, resolveTensorGroup, resolveTensorRole } from "../formats/rdrr/index.js";
|
|
5
|
+
import { getInferenceLayerPatternContractArtifact } from "../rules/rule-registry.js";
|
|
6
|
+
import { buildExecutionContractArtifact } from "../config/execution-contract-check.js";
|
|
7
|
+
import { buildManifestRequiredInferenceFieldsArtifact } from "../config/required-inference-fields-contract-check.js";
|
|
8
|
+
import { normalizeQ4KLayout, resolveManifestQuantization, resolveEffectiveQuantizationInfo } from "./quantization-info.js";
|
|
9
|
+
import { SOURCE_PACKED_QUANT_DTYPES, normalizeModulesToNotConvert, normalizeStorageQuant, resolveQuantizeEmbeddings, transformTensorBytes } from "./tensor-transform.js";
|
|
10
|
+
import { buildArtifactIdentity, sanitizeModelId } from "./artifact-identity.js";
|
|
11
|
+
import { createManifest } from "./manifest-builder.js";
|
|
12
|
+
import { normalizeCompressedTensorsW4A16, normalizeTensorName } from "./tokenizer-contract.js";
|
|
13
|
+
import { assertSupportedGemma4Conversion, validateInt4PleMaterializationContract } from "./gemma4/validation.js";
|
|
14
|
+
export { RDRR_VERSION, buildBundledTokenizer, buildSentencepieceTokenizer, createManifest, inferEmbeddingOutputConfig, resolveBundledTokenizerVocabSize, resolveConvertedAt, resolveManifestMoEConfig, resolveManifestMultimodalConfig } from "./manifest-builder.js";
|
|
15
|
+
export { extractArchitecture, sanitizeModelId } from "./artifact-identity.js";
|
|
16
|
+
export { isGemma4PerLayerEmbedTensor, normalizeStorageQuant, resolveTensorTargetQuant, shouldQuantize, transformTensorBytes } from "./tensor-transform.js";
|
|
17
|
+
export { formatBytes };
|
|
18
|
+
export { generateShardFilename };
|
|
50
19
|
|
|
51
20
|
// ============================================================================
|
|
52
21
|
// Re-exports for Backward Compatibility
|
|
@@ -226,115 +195,6 @@ function materializeTiedLmHeadTensor(tensors, options) {
|
|
|
226
195
|
];
|
|
227
196
|
}
|
|
228
197
|
|
|
229
|
-
// Re-export formatBytes from schema for backward compatibility
|
|
230
|
-
export { formatBytes };
|
|
231
|
-
|
|
232
|
-
function validateInt4PleMaterializationContract(tensorLocations, inference, modelId) {
|
|
233
|
-
const materialization = inference?.session?.perLayerInputs?.materialization;
|
|
234
|
-
if (materialization !== 'gpu_split_tables') {
|
|
235
|
-
return;
|
|
236
|
-
}
|
|
237
|
-
for (const [name, location] of Object.entries(tensorLocations ?? {})) {
|
|
238
|
-
const sourceTransform = location?.sourceTransform ?? null;
|
|
239
|
-
if (
|
|
240
|
-
isGemma4PerLayerEmbedTensor(name)
|
|
241
|
-
&& sourceTransform?.kind === 'litert_axis_dequant'
|
|
242
|
-
&& String(sourceTransform?.sourceDtype ?? '').toUpperCase() === 'INT4'
|
|
243
|
-
) {
|
|
244
|
-
throw new Error(
|
|
245
|
-
`Manifest "${modelId}" Gemma 4 INT4 PLE tensor "${name}" cannot use ` +
|
|
246
|
-
'inference.session.perLayerInputs.materialization="gpu_split_tables". ' +
|
|
247
|
-
'Use materialization="range_backed" or disable INT4 PLE quantization.'
|
|
248
|
-
);
|
|
249
|
-
}
|
|
250
|
-
}
|
|
251
|
-
}
|
|
252
|
-
|
|
253
|
-
function resolveGemma4TextConfig(rawConfig) {
|
|
254
|
-
const textConfig = getNestedTextConfig(rawConfig);
|
|
255
|
-
const modelType = String(textConfig?.model_type ?? rawConfig?.model_type ?? '').trim().toLowerCase();
|
|
256
|
-
if (
|
|
257
|
-
modelType !== 'gemma4'
|
|
258
|
-
&& modelType !== 'gemma4_text'
|
|
259
|
-
&& modelType !== 'gemma4_unified'
|
|
260
|
-
&& modelType !== 'gemma4_unified_text'
|
|
261
|
-
) {
|
|
262
|
-
return null;
|
|
263
|
-
}
|
|
264
|
-
return textConfig ?? rawConfig ?? null;
|
|
265
|
-
}
|
|
266
|
-
|
|
267
|
-
function collectGemma4UnsupportedTensorFlags(tensors) {
|
|
268
|
-
const names = Array.isArray(tensors) ? tensors.map((tensor) => normalizeTensorName(tensor).toLowerCase()) : [];
|
|
269
|
-
const flags = [];
|
|
270
|
-
if (names.some((name) => name.includes('.experts.gate_up_proj'))) {
|
|
271
|
-
flags.push('experts.gate_up_proj');
|
|
272
|
-
}
|
|
273
|
-
if (names.some((name) => name.includes('.experts.down_proj'))) {
|
|
274
|
-
flags.push('experts.down_proj');
|
|
275
|
-
}
|
|
276
|
-
if (names.some((name) => name.includes('.router.per_expert_scale'))) {
|
|
277
|
-
flags.push('router.per_expert_scale');
|
|
278
|
-
}
|
|
279
|
-
if (names.some((name) => name.includes('.router.scale'))) {
|
|
280
|
-
flags.push('router.scale');
|
|
281
|
-
}
|
|
282
|
-
if (names.some((name) => name.includes('.post_feedforward_layernorm_1.'))) {
|
|
283
|
-
flags.push('post_feedforward_layernorm_1');
|
|
284
|
-
}
|
|
285
|
-
if (names.some((name) => name.includes('.post_feedforward_layernorm_2.'))) {
|
|
286
|
-
flags.push('post_feedforward_layernorm_2');
|
|
287
|
-
}
|
|
288
|
-
if (names.some((name) => name.includes('.pre_feedforward_layernorm_2.'))) {
|
|
289
|
-
flags.push('pre_feedforward_layernorm_2');
|
|
290
|
-
}
|
|
291
|
-
return flags;
|
|
292
|
-
}
|
|
293
|
-
|
|
294
|
-
function assertSupportedGemma4Conversion(model, tensors, modelId) {
|
|
295
|
-
const rawConfig = model?.config ?? null;
|
|
296
|
-
const textConfig = resolveGemma4TextConfig(rawConfig);
|
|
297
|
-
if (!textConfig) return;
|
|
298
|
-
|
|
299
|
-
const hiddenSizePerLayerInput = Number(textConfig.hidden_size_per_layer_input ?? 0);
|
|
300
|
-
if (Number.isFinite(hiddenSizePerLayerInput) && hiddenSizePerLayerInput > 0) {
|
|
301
|
-
const names = Array.isArray(tensors)
|
|
302
|
-
? tensors.map((tensor) => String(tensor?.name ?? '').trim())
|
|
303
|
-
: [];
|
|
304
|
-
const requiredNames = [
|
|
305
|
-
'embed_tokens_per_layer.weight',
|
|
306
|
-
'per_layer_input_gate.weight',
|
|
307
|
-
'per_layer_projection.weight',
|
|
308
|
-
'post_per_layer_input_norm.weight',
|
|
309
|
-
'per_layer_model_projection.weight',
|
|
310
|
-
'per_layer_projection_norm.weight',
|
|
311
|
-
];
|
|
312
|
-
const missing = requiredNames.filter((suffix) => !names.some((name) => name.endsWith(suffix)));
|
|
313
|
-
if (missing.length > 0) {
|
|
314
|
-
throw new Error(
|
|
315
|
-
`Gemma 4 model "${modelId}" declares hidden_size_per_layer_input=${hiddenSizePerLayerInput}, ` +
|
|
316
|
-
`but the checkpoint is missing required per-layer input tensors: ${missing.join(', ')}.`
|
|
317
|
-
);
|
|
318
|
-
}
|
|
319
|
-
}
|
|
320
|
-
|
|
321
|
-
if (textConfig.enable_moe_block !== true) {
|
|
322
|
-
return;
|
|
323
|
-
}
|
|
324
|
-
|
|
325
|
-
const unsupportedFlags = collectGemma4UnsupportedTensorFlags(tensors);
|
|
326
|
-
if (unsupportedFlags.length === 0 && !modelHasMoETensors({ tensors })) {
|
|
327
|
-
return;
|
|
328
|
-
}
|
|
329
|
-
|
|
330
|
-
throw new Error(
|
|
331
|
-
`Gemma 4 model "${modelId}" is not supported yet: Gemma 4 MoE decoder blocks require ` +
|
|
332
|
-
'Gemma-specific router scaling and dual dense+MoE FFN execution, but current Doppler MoE runtime ' +
|
|
333
|
-
'only supports Mixtral/GPT-OSS semantics. ' +
|
|
334
|
-
`Detected: ${unsupportedFlags.length > 0 ? unsupportedFlags.join(', ') : 'Gemma 4 MoE tensors'}.`
|
|
335
|
-
);
|
|
336
|
-
}
|
|
337
|
-
|
|
338
198
|
export function buildTensorMap(tensors, shardSize) {
|
|
339
199
|
if (!shardSize || shardSize <= 0) {
|
|
340
200
|
throw new Error('Missing shard size for tensor map');
|
|
@@ -838,9 +698,3 @@ export async function convertModel(model, io, options = {}) {
|
|
|
838
698
|
requiredInferenceFieldsArtifact,
|
|
839
699
|
};
|
|
840
700
|
}
|
|
841
|
-
|
|
842
|
-
// ============================================================================
|
|
843
|
-
// Utility Exports
|
|
844
|
-
// ============================================================================
|
|
845
|
-
|
|
846
|
-
export { generateShardFilename };
|
|
@@ -18,11 +18,11 @@ import {
|
|
|
18
18
|
CAPSULE_V2_PROGRAM_SCHEMA_ID,
|
|
19
19
|
buildCapsuleV2,
|
|
20
20
|
signCapsuleV2,
|
|
21
|
-
} from
|
|
21
|
+
} from "../config/capsule/capsule-v2.js";
|
|
22
22
|
import { sha256Hex } from '../formats/sha256.js';
|
|
23
23
|
import { buildQualificationRecords, promoteQualifiedModelIRV2, normalizeQualificationEvidence } from './forge-qualification.js';
|
|
24
24
|
import { resolveCapsuleEmbeddingContract } from '../config/embedding-contract.js';
|
|
25
|
-
import { assertPairedLinearMechanismClosure, assertPairedQKNormClosure } from
|
|
25
|
+
import { assertPairedLinearMechanismClosure, assertPairedQKNormClosure } from "../inference/pipelines/text/execution/paired-execution.js";
|
|
26
26
|
import { evaluateForgeCandidates } from './forge-candidate-evaluation.js';
|
|
27
27
|
|
|
28
28
|
export const RIG_PIPELINE_VERSION = '2.0.0';
|
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import { validateFP8Policy } from
|
|
1
|
+
import { validateFP8Policy } from "../config/schema/manifest/fp8.schema.js";
|
|
2
2
|
import { float8ToFloat32 } from '../formats/float8.js';
|
|
3
3
|
import { createFP8StorageDescriptor, fp8StorageBytes, validateFP8StorageBytes } from '../formats/rdrr/fp8-storage.js';
|
|
4
4
|
|