doppler-gpu 0.7.3 → 0.7.8
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +60 -0
- package/README.md +9 -17
- package/package.json +5 -48
- package/src/capsule-runtime.d.ts +2 -2
- package/src/client/capsule-host.browser.js +4 -4
- package/src/client/capsule-host.d.ts +8 -8
- package/src/client/capsule-host.js +4 -4
- package/src/client/doppler-api.d.ts +2 -2
- package/src/client/inspection.d.ts +3 -0
- package/src/client/inspection.js +29 -0
- package/src/client/model-host/index.d.ts +5 -4
- package/src/client/model-host/index.js +28 -8
- package/src/client/model-host/model-evidence.js +2 -0
- package/src/client/model-host/model-session.d.ts +7 -5
- package/src/client/model-host/model-session.js +10 -3
- package/src/client/model-host/privacy-receipt.js +22 -0
- package/src/client/resident-partitions.d.ts +3 -3
- package/src/client/resident-partitions.js +2 -2
- package/src/client/runtime/capsule-acquisition.d.ts +3 -3
- package/src/client/runtime/capsule-acquisition.js +1 -1
- package/src/client/runtime/capsule-adapter-execution.d.ts +2 -2
- package/src/client/runtime/capsule-adapter-execution.js +3 -3
- package/src/client/runtime/capsule-embedding.d.ts +2 -2
- package/src/client/runtime/capsule-embedding.js +1 -1
- package/src/client/runtime/capsule-forecast.d.ts +2 -2
- package/src/client/runtime/capsule-forecast.js +1 -1
- package/src/client/runtime/capsule-operation-adapters.d.ts +1 -1
- package/src/client/runtime/capsule-operation-adapters.js +1 -1
- package/src/client/runtime/capsule-operation-deltas.js +1 -1
- package/src/client/runtime/capsule-operation-executor.d.ts +1 -1
- package/src/client/runtime/capsule-operation-executor.js +2 -2
- package/src/client/runtime/capsule-operation-stream.d.ts +1 -1
- package/src/client/runtime/capsule-operation-stream.js +2 -2
- package/src/client/runtime/capsule-piece-reader.js +1 -1
- package/src/client/runtime/capsule-program-adapter.js +1 -1
- package/src/client/runtime/capsule-release-authorization.js +1 -1
- package/src/client/runtime/capsule-rerank.d.ts +2 -2
- package/src/client/runtime/composition-root.d.ts +6 -6
- package/src/client/runtime/composition-root.js +6 -6
- package/src/client/runtime/fetch-capsule-artifact-store.d.ts +1 -1
- package/src/client/runtime/fetch-capsule-artifact-store.js +1 -1
- package/src/client/runtime/model-source.d.ts +4 -2
- package/src/client/runtime/model-source.js +5 -5
- package/src/client/runtime/scoped-session.d.ts +2 -2
- package/src/client/runtime/verified-capsule-artifact-store.d.ts +2 -2
- package/src/client/runtime/verified-capsule-artifact-store.js +1 -1
- package/src/config/backward-registry-loader.js +1 -1
- package/src/config/{capsule-adapter-policy.js → capsule/capsule-adapter-policy.js} +2 -2
- package/src/config/{capsule-adapters.d.ts → capsule/capsule-adapters.d.ts} +1 -1
- package/src/config/{capsule-adapters.js → capsule/capsule-adapters.js} +2 -2
- package/src/config/{capsule-loading.js → capsule/capsule-loading.js} +1 -1
- package/src/config/{capsule-operation.d.ts → capsule/capsule-operation.d.ts} +1 -1
- package/src/config/{capsule-operation.js → capsule/capsule-operation.js} +3 -3
- package/src/config/{capsule-release-contract.js → capsule/capsule-release-contract.js} +2 -2
- package/src/config/{capsule-release-events.d.ts → capsule/capsule-release-events.d.ts} +2 -1
- package/src/config/{capsule-release-events.js → capsule/capsule-release-events.js} +3 -2
- package/src/config/{capsule-sequence-receipt.js → capsule/capsule-sequence-receipt.js} +1 -1
- package/src/config/{capsule-token-selection.d.ts → capsule/capsule-token-selection.d.ts} +1 -1
- package/src/config/{capsule-token-selection.js → capsule/capsule-token-selection.js} +2 -2
- package/src/config/{capsule-v2.d.ts → capsule/capsule-v2.d.ts} +5 -4
- package/src/config/{capsule-v2.js → capsule/capsule-v2.js} +6 -6
- package/src/config/{capsule-v3.js → capsule/capsule-v3.js} +1 -1
- package/src/config/{capsule.js → capsule/capsule.js} +26 -7
- package/src/config/choice-scoring.js +1 -1
- package/src/config/conversion/amplify/amplify-120m-f16-af32.json +1 -1
- package/src/config/conversion/diffusiongemma/diffusiongemma-26b-a4b-it-q4k-ehf16-af16.json +1 -1
- package/src/config/conversion/embeddinggemma/google-embeddinggemma-300m-q4k-ehf16-af32.json +1 -1
- package/src/config/conversion/esm/esm2-t12-35m-ur50d-f32-af32.json +2 -2
- package/src/config/conversion/esmc/esmc-300m-f32-af32.json +1 -1
- package/src/config/conversion/gemma3/gemma-3-1b-it-f16-af32.json +1 -1
- package/src/config/conversion/gemma3/gemma-3-1b-it-q4k-ehf16-af32.json +1 -1
- package/src/config/conversion/gemma3/gemma-3-270m-it-f16-af32.json +1 -1
- package/src/config/conversion/gemma3/gemma-3-270m-it-q4k-ehf16-af16.json +1 -1
- package/src/config/conversion/gemma3/gemma-3-270m-it-q4k-ehf16-af32.json +1 -1
- package/src/config/conversion/gemma3/gemma-3-270m-it-w8a8-ehf16-af32.json +1 -1
- package/src/config/conversion/gemma3/translategemma-4b-1b-enes-q4k-ehf16-af32.json +1 -1
- package/src/config/conversion/gemma3/translategemma-4b-it-q4k-ehf16-af32.json +1 -1
- package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-af16.json +8 -8
- package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-af32.json +8 -8
- package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-hq4k-af16.json +8 -8
- package/src/config/conversion/gemma4/gemma-4-12b-it-text-w4a16-ct-ehf16-af16.json +8 -8
- package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af16.json +8 -8
- package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af32.json +8 -8
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af16-int4ple.json +8 -8
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32-int4ple.json +8 -8
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32.json +8 -8
- package/src/config/conversion/gemma4/gemma-4-moe-q4k-ehf16-af32.json +1 -1
- package/src/config/conversion/glmocr/glm-ocr-f16-af32.json +1 -1
- package/src/config/conversion/gpt-oss-20b-f16-xmxfp4.json +1 -1
- package/src/config/conversion/janus/janus-pro-1b-text-q4k-ehaf16.json +1 -1
- package/src/config/conversion/lfm2/lfm2.5-1.2b-instruct-q4k-ehf16-af32.json +1 -1
- package/src/config/conversion/minicpm/minicpm4-0-5b-f16-af32.json +1 -1
- package/src/config/conversion/muse-glimmer/muse-glimmer-30b-text-bf16-af32.json +1 -1
- package/src/config/conversion/muse-glimmer/muse-glimmer-30b-text-f16-af32.json +1 -1
- package/src/config/conversion/nucleotide-transformer/nucleotide-transformer-v2-50m-f32-af32.json +1 -1
- package/src/config/conversion/qwen3/qwen-3-4b-instruct-2507-q4k-ehf16-af32.json +2 -2
- package/src/config/conversion/qwen3/qwen-3-5-0-8b-canonical-f32.json +1 -1
- package/src/config/conversion/qwen3/qwen-3-5-0-8b-q4k-ehaf16.json +1 -1
- package/src/config/conversion/qwen3/qwen-3-5-2b-q4k-ehaf16.json +1 -1
- package/src/config/conversion/qwen3/qwen-3-5-9b-f16-af32.json +1 -1
- package/src/config/conversion/qwen3/qwen-3-5-9b-q4k-ehaf16.json +1 -1
- package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-eaf16.json +1 -1
- package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-ehaf16.json +1 -1
- package/src/config/conversion/qwen3/qwen-3-6-35b-a3b-q4_0-passthrough.json +1 -1
- package/src/config/conversion/qwen3/qwen-3-6-35b-a3b-text-q4k-ehf16-af32.json +1 -1
- package/src/config/conversion/qwen3/qwen-3-8-flash-next-text-q4k-ehf16-af32.json +25 -22
- package/src/config/conversion/qwen3/qwen-3-embedding-0-6b-f16-af32.json +2 -2
- package/src/config/conversion/qwen3/qwen-3-embedding-0-6b-q4k-ehf16-af32.json +1 -1
- package/src/config/conversion/qwen3/qwen-3-reranker-0-6b-f16-af32.json +1 -1
- package/src/config/conversion/qwen3/qwen-3-reranker-0-6b-f16-true-logit-af32.json +2 -2
- package/src/config/conversion/qwen3/qwen-3-reranker-0-6b-q4k-ehf16-af32.json +2 -2
- package/src/config/conversion/qwen3/qwen3.8-27b-q4_0-passthrough.json +1 -1
- package/src/config/conversion/qwen3/qwen3.8-27b-text-f16-af32.json +1 -1
- package/src/config/conversion/qwen3/qwen3.8-27b-text-q4k-ehaf16.json +1 -1
- package/src/config/conversion/qwen3/qwen4exp-gguf-source-policy.json +53 -0
- package/src/config/embedding-contract.d.ts +1 -1
- package/src/config/embedding-contract.js +1 -1
- package/src/config/execution-registry-contract.d.ts +5 -5
- package/src/config/execution-registry-contract.js +5 -5
- package/src/config/generation-contract.js +2 -1
- package/src/config/hardware-partition-plan.d.ts +32 -0
- package/src/config/hardware-partition-plan.js +78 -0
- package/src/config/inspection/observation-policies.json +1 -1
- package/src/config/kernel-path-loader.d.ts +1 -1
- package/src/config/kernel-path-loader.js +1 -1
- package/src/config/kernel-registry-contract.d.ts +2 -2
- package/src/config/kernel-registry-contract.js +1 -1
- package/src/config/kernels/kernel-ref-digests.js +509 -0
- package/src/config/kernels/registry.js +2 -2
- package/src/config/kernels/registry.json +2065 -1241
- package/src/config/layer-partition-plan.d.ts +4 -0
- package/src/config/layer-partition-plan.js +10 -4
- package/src/config/merge.js +1 -1
- package/src/config/model-session-declarations.js +85 -0
- package/src/config/param-validator.js +3 -3
- package/src/config/platforms/amd-rdna3.json +2 -2
- package/src/config/production-release-evidence.js +1 -1
- package/src/config/revocation-policy.js +6 -2
- package/src/config/revocation-registry.json +1 -1
- package/src/config/runtime/model-sessions.json +23 -0
- package/src/config/runtime/profiles/flash-next-gguf-control-gemv.json +72 -0
- package/src/config/runtime/profiles/flash-next-gguf-hc-inject-gemv.json +47 -0
- package/src/config/schema/capsule/index.d.ts +14 -0
- package/src/config/schema/capsule/index.js +10 -0
- package/src/config/schema/{program-bundle.schema.d.ts → capsule/program-bundle.schema.d.ts} +3 -3
- package/src/config/schema/{program-bundle.schema.js → capsule/program-bundle.schema.js} +6 -6
- package/src/config/schema/{execution-v1.schema.d.ts → execution/execution-v1.schema.d.ts} +2 -0
- package/src/config/schema/{execution-v1.schema.js → execution/execution-v1.schema.js} +2 -2
- package/src/config/schema/execution/index.d.ts +203 -0
- package/src/config/schema/execution/index.js +94 -0
- package/src/config/schema/{inference.schema.d.ts → execution/inference.schema.d.ts} +1 -1
- package/src/config/schema/{kernel-path.schema.d.ts → execution/kernel-path.schema.d.ts} +3 -0
- package/src/config/schema/{kernel-registry.schema.d.ts → execution/kernel-registry.schema.d.ts} +2 -2
- package/src/config/schema/{kernel-thresholds.schema.d.ts → execution/kernel-thresholds.schema.d.ts} +1 -1
- package/src/config/schema/{kernel-thresholds.schema.js → execution/kernel-thresholds.schema.js} +1 -1
- package/src/config/schema/{kernel-warmup.schema.d.ts → execution/kernel-warmup.schema.d.ts} +14 -0
- package/src/config/schema/{kernel-warmup.schema.js → execution/kernel-warmup.schema.js} +3 -0
- package/src/config/schema/{kvcache.schema.d.ts → execution/kvcache.schema.d.ts} +8 -2
- package/src/config/schema/{kvcache.schema.js → execution/kvcache.schema.js} +17 -1
- package/src/config/schema/{ngram-embedding.schema.d.ts → execution/ngram-embedding.schema.d.ts} +3 -1
- package/src/config/schema/{ngram-embedding.schema.js → execution/ngram-embedding.schema.js} +16 -1
- package/src/config/schema/index.d.ts +8 -897
- package/src/config/schema/index.js +8 -506
- package/src/config/schema/{converter.schema.d.ts → manifest/converter.schema.d.ts} +4 -4
- package/src/config/schema/{converter.schema.js → manifest/converter.schema.js} +1 -1
- package/src/config/schema/manifest/index.d.ts +142 -0
- package/src/config/schema/manifest/index.js +62 -0
- package/src/config/schema/manifest/manifest-inference.schema.d.ts +298 -0
- package/src/config/schema/{manifest.schema.d.ts → manifest/manifest.schema.d.ts} +9 -291
- package/src/config/schema/{manifest.schema.js → manifest/manifest.schema.js} +2 -2
- package/src/config/schema/{emulation.schema.d.ts → platform/emulation.schema.d.ts} +1 -1
- package/src/config/schema/{emulation.schema.js → platform/emulation.schema.js} +6 -6
- package/src/config/schema/{gpu-cache.schema.js → platform/gpu-cache.schema.js} +1 -1
- package/src/config/schema/platform/index.d.ts +72 -0
- package/src/config/schema/platform/index.js +27 -0
- package/src/config/schema/{memory-limits.schema.js → platform/memory-limits.schema.js} +1 -1
- package/src/config/schema/policy/index.d.ts +14 -0
- package/src/config/schema/policy/index.js +9 -0
- package/src/config/schema/{buffer-pool.schema.js → runtime/buffer-pool.schema.js} +2 -2
- package/src/config/schema/{doppler.schema.d.ts → runtime/doppler.schema.d.ts} +3 -3
- package/src/config/schema/{doppler.schema.js → runtime/doppler.schema.js} +12 -12
- package/src/config/schema/runtime/index.d.ts +238 -0
- package/src/config/schema/runtime/index.js +135 -0
- package/src/config/schema/{loading.schema.d.ts → runtime/loading.schema.d.ts} +31 -1
- package/src/config/schema/runtime/loading.schema.js +275 -0
- package/src/config/schema/{shared-runtime.schema.d.ts → runtime/shared-runtime.schema.d.ts} +5 -5
- package/src/config/schema/{shared-runtime.schema.js → runtime/shared-runtime.schema.js} +4 -4
- package/src/config/schema/{tooling.schema.d.ts → runtime/tooling.schema.d.ts} +1 -1
- package/src/config/schema/training/index.d.ts +63 -0
- package/src/config/schema/training/index.js +43 -0
- package/src/config/schema/{lora.schema.js → training/lora.schema.js} +1 -1
- package/src/config/source-packages/litert/gemma-4-e2b-it.json +1 -1
- package/src/config/supported-operations.json +358 -88
- package/src/config/target-plan.d.ts +2 -2
- package/src/config/target-plan.js +2 -2
- package/src/config/training-defaults.d.ts +2 -2
- package/src/config/training-defaults.js +2 -2
- package/src/config/transforms/execution-graph/correctness-fallback.js +86 -0
- package/src/config/transforms/execution-graph/prefill-attention.js +116 -0
- package/src/{converter → config/transforms}/execution-graph-transforms.js +5 -194
- package/src/converter/core.js +18 -164
- package/src/converter/forge-stages.js +2 -2
- package/src/converter/fp8-source.js +1 -1
- package/src/converter/gemma4/validation.js +110 -0
- package/src/converter/linear-norm-gate.js +1 -1
- package/src/converter/manifest-builder.js +2 -9
- package/src/converter/parsers/gguf.js +54 -4
- package/src/converter/quantization-info.js +2 -2
- package/src/converter/qwen4-exp-source-execution.js +94 -0
- package/src/converter/qwen4-exp-source.js +1 -201
- package/src/converter/tensor-transform.js +2 -9
- package/src/converter/tokenizer-contract.js +2 -9
- package/src/converter/w8a8-quantizer.js +2 -2
- package/src/debug/config.js +1 -1
- package/src/debug/perf-guards.js +1 -1
- package/src/errors/doppler-error.d.ts +54 -0
- package/src/errors/doppler-error.js +13 -5
- package/src/experimental/distribution/shard-delivery/receipt.js +1 -1
- package/src/experimental/distribution/shard-delivery/retry.js +4 -4
- package/src/experimental/energy/types.d.ts +1 -1
- package/src/experimental/logits/cpu-output.js +1 -1
- package/src/experimental/training/artifacts.d.ts +2 -2
- package/src/experimental/training/artifacts.js +2 -2
- package/src/experimental/training/autograd.d.ts +1 -1
- package/src/experimental/training/distillation/student-fixture/execution.js +1 -1
- package/src/experimental/training/distillation/student-fixture/model.js +1 -1
- package/src/experimental/training/distillation/student-fixture.js +1 -1
- package/src/experimental/training/loss-scaling.d.ts +1 -1
- package/src/experimental/training/runner/evaluation.js +1 -1
- package/src/experimental/training/runner.js +1 -1
- package/src/experimental/training/trainer.d.ts +1 -1
- package/src/experimental/training/workloads.js +2 -2
- package/src/formats/gguf/block-storage.d.ts +9 -0
- package/src/formats/gguf/block-storage.js +12 -3
- package/src/formats/gguf/qwen-source.js +5 -3
- package/src/formats/gguf/qwen-source.json +8 -1
- package/src/formats/gguf/qwen4-exp-source.js +101 -0
- package/src/formats/q4k-dequant.js +1 -1
- package/src/formats/qwen4-exp-source-contracts.js +204 -0
- package/src/formats/rdrr/fp8-contract.js +1 -1
- package/src/formats/rdrr/fp8-storage.js +1 -1
- package/src/formats/rdrr/q4_0-storage.js +1 -1
- package/src/formats/rdrr/storage-descriptor.js +1 -1
- package/src/formats/rdrr/tensor-config-validator.js +3 -3
- package/src/formats/rdrr/validation.js +1 -1
- package/src/formats/rdrr/w8a8-storage.js +1 -1
- package/src/generation/index.d.ts +1 -1
- package/src/generation/index.js +1 -0
- package/src/gpu/device.d.ts +1 -1
- package/src/gpu/kernels/attention/contiguous-quantized.js +252 -0
- package/src/gpu/kernels/attention/tiered-quantized.js +321 -0
- package/src/gpu/kernels/attention/tiered.js +240 -0
- package/src/gpu/kernels/attention.js +5 -801
- package/src/gpu/kernels/backward/layernorm_backward.wgsl +3 -3
- package/src/gpu/kernels/channel-gain.js +100 -0
- package/src/gpu/kernels/channel-rotation.d.ts +16 -0
- package/src/gpu/kernels/channel-rotation.js +47 -0
- package/src/gpu/kernels/channel_energy.wgsl +17 -0
- package/src/gpu/kernels/channel_energy_pair.wgsl +27 -0
- package/src/gpu/kernels/channel_gain_control.wgsl +60 -0
- package/src/gpu/kernels/channel_gain_derive.wgsl +40 -0
- package/src/gpu/kernels/channel_gain_swiglu.wgsl +64 -0
- package/src/gpu/kernels/channel_rotation_apply.wgsl +38 -0
- package/src/gpu/kernels/channel_rotation_derive.wgsl +76 -0
- package/src/gpu/kernels/check-finiteness-f32.js +22 -0
- package/src/gpu/kernels/check_finiteness.wgsl +2 -2
- package/src/gpu/kernels/clamp.js +2 -2
- package/src/gpu/kernels/dequant.js +1 -1
- package/src/gpu/kernels/dispatch.js +7 -0
- package/src/gpu/kernels/energy_quintel_grad.wgsl +7 -7
- package/src/gpu/kernels/energy_quintel_grad_f16.wgsl +7 -7
- package/src/gpu/kernels/energy_quintel_reduce.wgsl +6 -6
- package/src/gpu/kernels/energy_quintel_reduce_f16.wgsl +6 -6
- package/src/gpu/kernels/energy_quintel_update.wgsl +7 -7
- package/src/gpu/kernels/energy_quintel_update_f16.wgsl +7 -7
- package/src/gpu/kernels/fused_matmul_q4_widetile_f32.wgsl +174 -0
- package/src/gpu/kernels/fused_matmul_rmsnorm.wgsl +0 -49
- package/src/gpu/kernels/gated-delta-norm-gate.js +1 -1
- package/src/gpu/kernels/gated_delta_recurrent_gguf_register_prefill_sigmoid_gate.wgsl +730 -0
- package/src/gpu/kernels/gated_delta_recurrent_gguf_sigmoid_gate.wgsl +499 -0
- package/src/gpu/kernels/gather.js +16 -1
- package/src/gpu/kernels/gather_split4_packed_f16_vec4_f32_out.wgsl +104 -0
- package/src/gpu/kernels/generated/uniform-writers-0.js +7 -11
- package/src/gpu/kernels/generated/uniform-writers-1.js +43 -46
- package/src/gpu/kernels/generated/uniform-writers-2.js +80 -2
- package/src/gpu/kernels/generated/uniform-writers.js +33 -21
- package/src/gpu/kernels/hyper-connection-up.js +64 -0
- package/src/gpu/kernels/hyper-connection.js +15 -4
- package/src/gpu/kernels/hyper_connection_mix_branch.wgsl +105 -0
- package/src/gpu/kernels/hyper_connection_norm.wgsl +3 -3
- package/src/gpu/kernels/hyper_connection_up_mix_branch_q4_0.wgsl +164 -0
- package/src/gpu/kernels/kernel-execution.js +18 -7
- package/src/gpu/kernels/linear-attention-core.js +3 -4
- package/src/gpu/kernels/moe-route.js +14 -0
- package/src/gpu/kernels/moe_gather.wgsl +0 -63
- package/src/gpu/kernels/moe_route_q4_0_block_f32.wgsl +74 -0
- package/src/gpu/kernels/moe_route_q4_0_split_block_f32.wgsl +80 -0
- package/src/gpu/kernels/moe_route_q4_1_block_f32.wgsl +59 -0
- package/src/gpu/kernels/moe_route_remap.wgsl +22 -0
- package/src/gpu/kernels/paired-linear-attention.js +1 -1
- package/src/gpu/kernels/pipeline-cache.js +18 -2
- package/src/gpu/kernels/ple.js +11 -7
- package/src/gpu/kernels/ple_q4_0_to_f32.wgsl +32 -0
- package/src/gpu/kernels/qsa-indexer.d.ts +2 -0
- package/src/gpu/kernels/qsa-indexer.js +6 -4
- package/src/gpu/kernels/qsa_block_keys.wgsl +2 -1
- package/src/gpu/kernels/qsa_prepare.wgsl +2 -1
- package/src/gpu/kernels/relu.js +1 -1
- package/src/gpu/kernels/relu.wgsl +2 -2
- package/src/gpu/kernels/relu_f16.wgsl +2 -2
- package/src/gpu/kernels/residual.js +1 -1
- package/src/gpu/kernels/residual.wgsl +4 -4
- package/src/gpu/kernels/residual_f16.wgsl +2 -2
- package/src/gpu/kernels/residual_f16_vec4.wgsl +2 -2
- package/src/gpu/kernels/residual_vec4.wgsl +2 -2
- package/src/gpu/kernels/rope-precompute.js +1 -1
- package/src/gpu/kernels/scatter_add.wgsl +0 -31
- package/src/gpu/kernels/shader-cache.js +17 -3
- package/src/gpu/kernels/shader-capability-filter.js +120 -0
- package/src/gpu/kernels/softmax.js +10 -10
- package/src/gpu/kernels/topk.wgsl +0 -55
- package/src/gpu/kernels/transpose.js +1 -1
- package/src/gpu/kernels/transpose.wgsl +2 -2
- package/src/gpu/kernels/w8a8.js +1 -1
- package/src/gpu/uniform-cache.d.ts +15 -1
- package/src/gpu/uniform-cache.js +82 -7
- package/src/gpu/weight-buffer.d.ts +1 -1
- package/src/inference/moe-router.d.ts +1 -126
- package/src/inference/moe-router.js +0 -149
- package/src/inference/pipelines/context.js +3 -15
- package/src/inference/pipelines/shader-scoped-pipeline.js +10 -5
- package/src/inference/pipelines/text/attention/interpreter/immediate.js +4 -4
- package/src/inference/pipelines/text/attention/interpreter/output.js +203 -0
- package/src/inference/pipelines/text/attention/interpreter/recorded.js +92 -67
- package/src/inference/pipelines/text/attention/interpreter.js +77 -350
- package/src/inference/pipelines/text/attention/projection-executor.js +2 -2
- package/src/inference/pipelines/text/attention/projection-plan.js +1 -1
- package/src/inference/pipelines/text/attention/projection-precision.js +132 -0
- package/src/inference/pipelines/text/attention/projections.js +13 -151
- package/src/inference/pipelines/text/attention/recorded-health.js +83 -0
- package/src/inference/pipelines/text/attention/rope-observation.js +1 -1
- package/src/inference/pipelines/text/attention/types.js +0 -13
- package/src/inference/pipelines/text/config/normalization.js +2 -2
- package/src/inference/pipelines/text/config.d.ts +4 -4
- package/src/inference/pipelines/text/config.js +17 -135
- package/src/inference/pipelines/text/{config-components.js → configuration/config-components.js} +4 -4
- package/src/inference/pipelines/text/configuration/model-geometry.js +114 -0
- package/src/inference/pipelines/text/{resolved-runtime-session.d.ts → configuration/resolved-runtime-session.d.ts} +4 -4
- package/src/inference/pipelines/text/{resolved-runtime-session.js → configuration/resolved-runtime-session.js} +1 -1
- package/src/inference/pipelines/text/{abort-contract.js → contracts/abort-contract.js} +2 -1
- package/src/inference/pipelines/text/contracts/metrics.d.ts +234 -0
- package/src/inference/pipelines/text/{embed.js → embedding/embed.js} +15 -15
- package/src/inference/pipelines/text/{embedding-batch.js → embedding/embedding-batch.js} +1 -1
- package/src/inference/pipelines/text/{embedding-extraction.js → embedding/embedding-extraction.js} +12 -12
- package/src/inference/pipelines/text/{embedding-normalization.js → embedding/embedding-normalization.js} +4 -4
- package/src/inference/pipelines/text/{execution-plan.d.ts → execution/execution-plan.d.ts} +8 -8
- package/src/inference/pipelines/text/{execution-plan.js → execution/execution-plan.js} +9 -9
- package/src/inference/pipelines/text/{execution-runtime-builders.js → execution/execution-runtime-builders.js} +3 -3
- package/src/inference/pipelines/text/{execution-v1.d.ts → execution/execution-v1.d.ts} +2 -2
- package/src/inference/pipelines/text/{execution-v1.js → execution/execution-v1.js} +13 -183
- package/src/inference/pipelines/text/execution/layer-contract.js +90 -0
- package/src/inference/pipelines/text/{layer-execution-contract.js → execution/layer-execution-contract.js} +2 -2
- package/src/inference/pipelines/text/{layer-input-execution.js → execution/layer-input-execution.js} +27 -36
- package/src/inference/pipelines/text/{layer-partition-contract.d.ts → execution/layer-partition-contract.d.ts} +3 -3
- package/src/inference/pipelines/text/{layer-partition-contract.js → execution/layer-partition-contract.js} +2 -2
- package/src/inference/pipelines/text/{layer-plan-gpu.js → execution/layer-plan-gpu.js} +12 -12
- package/src/inference/pipelines/text/{layer-plan.d.ts → execution/layer-plan.d.ts} +1 -1
- package/src/inference/pipelines/text/{layer-plan.js → execution/layer-plan.js} +1 -1
- package/src/inference/pipelines/text/{layer.js → execution/layer.js} +34 -142
- package/src/inference/pipelines/text/{operator-identity.js → execution/operator-identity.js} +1 -1
- package/src/inference/pipelines/text/{ops.js → execution/ops.js} +26 -26
- package/src/inference/pipelines/text/{paired-execution.js → execution/paired-execution.js} +9 -9
- package/src/inference/pipelines/text/execution-v1/hybrid-projection-contract.js +159 -0
- package/src/inference/pipelines/text/expert-loading-stats.d.ts +16 -0
- package/src/inference/pipelines/text/expert-loading-stats.js +15 -0
- package/src/inference/pipelines/text/ffn/dense-executor.js +5 -4
- package/src/inference/pipelines/text/ffn/dense-plan.js +23 -34
- package/src/inference/pipelines/text/ffn/dense.js +72 -392
- package/src/inference/pipelines/text/ffn/fused-gate-up.js +197 -0
- package/src/inference/pipelines/text/ffn/moe.js +1 -1
- package/src/inference/pipelines/text/ffn/packed-gate-up.js +218 -0
- package/src/inference/pipelines/text/ffn/paired.js +3 -3
- package/src/inference/pipelines/text/ffn/qwen-shared-expert.js +1 -1
- package/src/inference/pipelines/text/ffn/sandwich.js +6 -7
- package/src/inference/pipelines/text/ffn/standard.js +5 -5
- package/src/inference/pipelines/text/ffn/types.js +6 -3
- package/src/inference/pipelines/text/{activity-weights.js → generator/activation/activity-weights.js} +8 -8
- package/src/inference/pipelines/text/generator/activation/channel-gain.d.ts +61 -0
- package/src/inference/pipelines/text/generator/activation/channel-gain.js +352 -0
- package/src/inference/pipelines/text/generator/activation/final-layer-replay.js +43 -0
- package/src/inference/pipelines/text/generator/{capsule-generation.js → adapters/capsule-generation.js} +5 -5
- package/src/inference/pipelines/text/generator/{diffusion.js → adapters/diffusion.js} +19 -19
- package/src/inference/pipelines/text/generator/{training.js → adapters/training.js} +21 -21
- package/src/inference/pipelines/text/generator/decode/advance-token.js +73 -0
- package/src/inference/pipelines/text/{generator-steps.js → generator/decode/batch-generation.js} +25 -288
- package/src/inference/pipelines/text/generator/decode/batch-readback.js +175 -0
- package/src/inference/pipelines/text/generator/{compiled-greedy-window.js → decode/compiled-greedy-window.js} +5 -5
- package/src/inference/pipelines/text/generator/{decode-components.js → decode/decode-components.js} +1 -1
- package/src/inference/pipelines/text/generator/{decode-runtime.js → decode/decode-loop.js} +20 -412
- package/src/inference/pipelines/text/generator/decode/decode-mechanics.js +288 -0
- package/src/inference/pipelines/text/generator/{decode-readback.js → decode/decode-readback.js} +1 -1
- package/src/inference/pipelines/text/generator/decode/decode-runtime.js +2 -0
- package/src/inference/pipelines/text/generator/{decode.js → decode/decode-step.js} +34 -313
- package/src/inference/pipelines/text/generator/decode/decode.js +5 -0
- package/src/inference/pipelines/text/generator/{greedy-head-policy.js → decode/greedy-head-policy.js} +1 -1
- package/src/inference/pipelines/text/generator/{recorded-step-executor.js → decode/recorded-step-executor.js} +2 -2
- package/src/inference/pipelines/text/generator/generator-steps.js +4 -0
- package/src/inference/pipelines/text/{generator.d.ts → generator/generator.d.ts} +13 -11
- package/src/inference/pipelines/text/{generator.js → generator/generator.js} +150 -46
- package/src/inference/pipelines/text/generator/{prefill-runtime.js → prefill/prefill-execution.js} +72 -278
- package/src/inference/pipelines/text/generator/prefill/prefill-logits.js +144 -0
- package/src/inference/pipelines/text/generator/{prefill-observation.js → prefill/prefill-observation.js} +2 -2
- package/src/inference/pipelines/text/generator/{prefill-policy.js → prefill/prefill-policy.js} +6 -1
- package/src/inference/pipelines/text/generator/prefill/prefill-runtime.js +2 -0
- package/src/inference/pipelines/text/generator/prefill/prefill-sequence.js +100 -0
- package/src/inference/pipelines/text/generator/{prefix-embedding.js → prefill/prefix-embedding.js} +7 -7
- package/src/inference/pipelines/text/{generation-request.d.ts → generator/request/generation-request.d.ts} +2 -2
- package/src/inference/pipelines/text/{generation-request.js → generator/request/generation-request.js} +3 -3
- package/src/inference/pipelines/text/generator/{prompt-input.js → request/prompt-input.js} +2 -2
- package/src/inference/pipelines/text/{sampling-config.d.ts → generator/sampling/sampling-config.d.ts} +1 -1
- package/src/inference/pipelines/text/{sampling-config.js → generator/sampling/sampling-config.js} +2 -2
- package/src/inference/pipelines/text/{sampling.d.ts → generator/sampling/sampling.d.ts} +15 -0
- package/src/inference/pipelines/text/{sampling.js → generator/sampling/sampling.js} +22 -3
- package/src/inference/pipelines/text/{stopping.js → generator/sampling/stopping.js} +3 -3
- package/src/inference/pipelines/text/generator/{token-selection.d.ts → sampling/token-selection.d.ts} +1 -1
- package/src/inference/pipelines/text/generator/{token-selection.js → sampling/token-selection.js} +8 -8
- package/src/inference/pipelines/text/generator/{activation-storage.js → session/activation-storage.js} +2 -2
- package/src/inference/pipelines/text/generator/{attention-lifecycle.js → session/attention-lifecycle.js} +1 -1
- package/src/inference/pipelines/text/generator/{diagnostic-sequence.d.ts → session/diagnostic-sequence.d.ts} +3 -3
- package/src/inference/pipelines/text/generator/{diagnostic-sequence.js → session/diagnostic-sequence.js} +3 -3
- package/src/inference/pipelines/text/generator/session/generation-runtime.js +278 -0
- package/src/inference/pipelines/text/{generator-runtime.d.ts → generator/session/generator-runtime.d.ts} +7 -7
- package/src/inference/pipelines/text/{generator-runtime.js → generator/session/generator-runtime.js} +8 -8
- package/src/inference/pipelines/text/generator/{recovery.js → session/recovery.js} +6 -6
- package/src/inference/pipelines/text/{sequence-state.d.ts → generator/session/sequence-state.d.ts} +5 -5
- package/src/inference/pipelines/text/{sequence-state.js → generator/session/sequence-state.js} +26 -18
- package/src/inference/pipelines/text/generator/{session-context.js → session/session-context.js} +12 -11
- package/src/inference/pipelines/text/generator/{speculative-runtime.js → session/speculative-runtime.js} +1 -1
- package/src/inference/pipelines/text/generator/{text.js → session/text.js} +9 -9
- package/src/inference/pipelines/text/{warm-generation.d.ts → generator/session/warm-generation.d.ts} +1 -1
- package/src/inference/pipelines/text/{warm-generation.js → generator/session/warm-generation.js} +1 -1
- package/src/inference/pipelines/text/generator/{whole-token-lane.js → whole-token/whole-token-lane.js} +10 -10
- package/src/inference/pipelines/text/generator/{whole-token-observation.d.ts → whole-token/whole-token-observation.d.ts} +2 -2
- package/src/inference/pipelines/text/generator/{whole-token-observation.js → whole-token/whole-token-observation.js} +5 -5
- package/src/inference/pipelines/text/generator/{whole-token-program.d.ts → whole-token/whole-token-program.d.ts} +3 -3
- package/src/inference/pipelines/text/generator/{whole-token-program.js → whole-token/whole-token-program.js} +5 -5
- package/src/inference/pipelines/text/hyper-connection/decoder-layer.d.ts +17 -4
- package/src/inference/pipelines/text/hyper-connection/decoder-layer.js +38 -14
- package/src/inference/pipelines/text/hyper-connection/layer-executor.js +14 -9
- package/src/inference/pipelines/text/hyper-connection/logits-input.js +5 -3
- package/src/inference/pipelines/text/hyper-connection/projection.js +52 -2
- package/src/inference/pipelines/text/hyper-connection/sequence-transaction.js +3 -3
- package/src/inference/pipelines/text/hyper-connection/session.d.ts +1 -1
- package/src/inference/pipelines/text/hyper-connection/session.js +9 -5
- package/src/inference/pipelines/text/indexed-attention/qsa-cache.d.ts +2 -1
- package/src/inference/pipelines/text/indexed-attention/qsa-cache.js +3 -3
- package/src/inference/pipelines/text/init.d.ts +4 -2
- package/src/inference/pipelines/text/init.js +10 -934
- package/src/inference/pipelines/text/initialization/init-emulation.js +76 -0
- package/src/inference/pipelines/text/initialization/init-kv-cache.js +547 -0
- package/src/inference/pipelines/text/initialization/init-moe.js +34 -0
- package/src/inference/pipelines/text/{init-qkv-fusion.js → initialization/init-qkv-fusion.js} +7 -7
- package/src/inference/pipelines/text/{init-rope.d.ts → initialization/init-rope.d.ts} +4 -4
- package/src/inference/pipelines/text/{init-rope.js → initialization/init-rope.js} +9 -11
- package/src/inference/pipelines/text/initialization/init-speculative.js +31 -0
- package/src/inference/pipelines/text/{init-tokenizer.js → initialization/init-tokenizer.js} +2 -2
- package/src/inference/pipelines/text/lifecycle.js +18 -18
- package/src/inference/pipelines/text/{linear-attention-ab-fusion.js → linear-attention/linear-attention-ab-fusion.js} +3 -3
- package/src/inference/pipelines/text/{linear-attention.d.ts → linear-attention/linear-attention.d.ts} +12 -12
- package/src/inference/pipelines/text/{linear-attention.js → linear-attention/linear-attention.js} +17 -17
- package/src/inference/pipelines/text/linear-attention/plan.js +4 -4
- package/src/inference/pipelines/text/linear-attention/state-checkpoint.js +1 -1
- package/src/inference/pipelines/text/linear-attention/state-clone.js +1 -1
- package/src/inference/pipelines/text/linear-attention/state-layout.js +2 -2
- package/src/inference/pipelines/text/loading/decode-session-config.js +295 -0
- package/src/inference/pipelines/text/loading/init-weights.js +254 -0
- package/src/inference/pipelines/text/{load-errors.js → loading/load-errors.js} +3 -3
- package/src/inference/pipelines/text/{model-load.js → loading/model-load.js} +14 -306
- package/src/inference/pipelines/text/{weight-loading-config.js → loading/weight-loading-config.js} +6 -6
- package/src/inference/pipelines/text/{weights.d.ts → loading/weights.d.ts} +13 -13
- package/src/inference/pipelines/text/{weights.js → loading/weights.js} +5 -5
- package/src/inference/pipelines/text/logits/gpu-executor.js +2 -2
- package/src/inference/pipelines/text/logits/gpu.d.ts +4 -4
- package/src/inference/pipelines/text/logits/gpu.js +2 -2
- package/src/inference/pipelines/text/{generator → logits}/hidden-projection.js +1 -1
- package/src/inference/pipelines/text/logits/index.d.ts +1 -1
- package/src/inference/pipelines/text/logits/index.js +3 -3
- package/src/inference/pipelines/text/{generator → logits}/logits-config.js +1 -1
- package/src/inference/pipelines/text/logits/output-transform.js +1 -1
- package/src/inference/pipelines/text/logits/precision-policy.js +1 -1
- package/src/inference/pipelines/text/logits/types.d.ts +2 -2
- package/src/inference/pipelines/text/lora-types.d.ts +1 -1
- package/src/inference/pipelines/text/moe/expert-loading.js +4 -1
- package/src/inference/pipelines/text/moe/gpu-executor.js +44 -101
- package/src/inference/pipelines/text/moe/grouped-prefill.js +6 -2
- package/src/inference/pipelines/text/moe/qwen-route-executor.d.ts +2 -2
- package/src/inference/pipelines/text/moe/qwen-route-executor.js +31 -2
- package/src/inference/pipelines/text/moe/routed-expert-chunks.js +99 -15
- package/src/inference/pipelines/text/moe-cache.js +11 -77
- package/src/inference/pipelines/text/moe-gpu.js +33 -75
- package/src/inference/pipelines/text/moe-impl.d.ts +3 -38
- package/src/inference/pipelines/text/moe-impl.js +1 -7
- package/src/inference/pipelines/text/ngram/ngram-addresses.d.ts +1 -1
- package/src/inference/pipelines/text/ngram/ngram-addresses.js +2 -2
- package/src/inference/pipelines/text/ngram/ngram-table.d.ts +3 -3
- package/src/inference/pipelines/text/ngram/ngram-table.js +5 -6
- package/src/inference/pipelines/text/ngram/ple-layer.d.ts +3 -3
- package/src/inference/pipelines/text/ngram/ple-layer.js +3 -3
- package/src/inference/pipelines/text/{generator-logits-observation.js → observation/generator-logits-observation.js} +4 -4
- package/src/inference/pipelines/text/{kernel-trace.js → observation/kernel-trace.js} +1 -1
- package/src/inference/pipelines/text/{operator-events.js → observation/operator-events.js} +1 -1
- package/src/inference/pipelines/text/{probes.js → observation/probes.js} +13 -13
- package/src/inference/pipelines/text/{tsir-fixture-writer.js → observation/tsir-fixture-writer.js} +3 -3
- package/src/inference/pipelines/text/{execution.js → operations/execution.js} +14 -14
- package/src/inference/pipelines/text/{image-transcription.js → operations/image-transcription.js} +8 -8
- package/src/inference/pipelines/text/{partition-attempt.js → partitions/partition-attempt.js} +7 -7
- package/src/inference/pipelines/text/{partition-execution.js → partitions/partition-execution.js} +107 -35
- package/src/inference/pipelines/text/partitions/partition-wire.js +50 -0
- package/src/inference/pipelines/text/{resident-partition-contract.d.ts → partitions/resident-partition-contract.d.ts} +10 -8
- package/src/inference/pipelines/text/{resident-partition-contract.js → partitions/resident-partition-contract.js} +6 -6
- package/src/inference/pipelines/text/partitions/resident-partition-warmup.js +67 -0
- package/src/inference/pipelines/text/{resident-partition.js → partitions/resident-partition.js} +27 -15
- package/src/inference/pipelines/text/{resident-pipeline-preparation.js → partitions/resident-pipeline-preparation.js} +32 -32
- package/src/inference/pipelines/text/per-layer/materialize.js +4 -4
- package/src/inference/pipelines/text/per-layer/per-layer-inputs.js +5 -0
- package/src/inference/pipelines/text/per-layer/plan.js +4 -4
- package/src/inference/pipelines/text/{per-layer-inputs.js → per-layer/prepare.js} +14 -280
- package/src/inference/pipelines/text/per-layer/projection-contract.js +79 -0
- package/src/inference/pipelines/text/per-layer/runtime-cache.js +183 -0
- package/src/inference/pipelines/text/pipeline.d.ts +13 -8
- package/src/inference/pipelines/text/pipeline.js +78 -22
- package/src/inference/pipelines/text/{drift-policy.js → policy/drift-policy.js} +1 -1
- package/src/inference/pipelines/text/{dtype-contract.js → policy/dtype-contract.js} +2 -2
- package/src/inference/pipelines/text/{finiteness-guard-status.js → policy/finiteness-guard-status.js} +2 -1
- package/src/inference/pipelines/text/{generator-decode-policy.js → policy/generator-decode-policy.js} +5 -5
- package/src/inference/pipelines/text/state.d.ts +9 -8
- package/src/inference/pipelines/text/state.js +1 -0
- package/src/inference/pipelines/text/types.d.ts +14 -218
- package/src/inference/pipelines/text.js +2 -0
- package/src/inference/tokenizers/bundled/merge.js +130 -0
- package/src/inference/tokenizers/bundled.js +2 -36
- package/src/loader/doppler-loader.d.ts +17 -7
- package/src/loader/doppler-loader.js +27 -29
- package/src/loader/embedding-limit-preflight.js +29 -1
- package/src/loader/embedding-loader.js +18 -4
- package/src/loader/experts/compact-expert-cache.js +508 -0
- package/src/loader/experts/expert-cache.d.ts +1 -1
- package/src/loader/experts/expert-dequant-cache.d.ts +9 -0
- package/src/loader/experts/expert-dequant-cache.js +75 -0
- package/src/loader/experts/expert-frequency-index.js +129 -0
- package/src/loader/experts/expert-loader.d.ts +4 -0
- package/src/loader/experts/expert-loader.js +13 -1
- package/src/loader/experts/expert-preparation.d.ts +1 -1
- package/src/loader/experts/expert-preparation.js +1 -1
- package/src/loader/experts/loader-bindings.js +129 -14
- package/src/loader/experts/mapped-expert-staging.js +393 -0
- package/src/loader/experts/mapped-expert-upload-cache.d.ts +61 -0
- package/src/loader/experts/mapped-expert-upload-cache.js +382 -0
- package/src/loader/experts/qwen-paging.d.ts +1 -1
- package/src/loader/experts/qwen-paging.js +1 -1
- package/src/loader/experts/routed-expert-cache.d.ts +70 -3
- package/src/loader/experts/routed-expert-cache.js +452 -99
- package/src/loader/experts/routed-expert-plan.d.ts +2 -0
- package/src/loader/experts/routed-expert-plan.js +14 -8
- package/src/loader/experts/routed-expert-prefill.js +158 -0
- package/src/loader/experts/routed-expert-source.js +234 -0
- package/src/loader/experts/routed-expert-trace.js +77 -0
- package/src/loader/experts/tiered-expert-cache.d.ts +12 -0
- package/src/loader/experts/tiered-expert-cache.js +327 -0
- package/src/loader/experts/tiered-expert-plan.js +74 -0
- package/src/loader/final-weights-loader.d.ts +1 -1
- package/src/loader/hyper-connection-loader.d.ts +2 -2
- package/src/loader/hyper-connection-loader.js +85 -13
- package/src/loader/layer-loader.d.ts +3 -3
- package/src/loader/layer-loader.js +1 -1
- package/src/loader/layer-tensor-aliases.js +14 -11
- package/src/loader/loader-types.d.ts +6 -1
- package/src/loader/manifest-config.js +4 -0
- package/src/loader/model-executor.js +18 -1
- package/src/loader/model-load.d.ts +1 -1
- package/src/loader/model-load.js +1 -1
- package/src/loader/ngram-table-bindings.d.ts +3 -3
- package/src/loader/ngram-table-bindings.js +7 -6
- package/src/loader/q4k-reference-weight.js +1 -1
- package/src/loader/shard-cache.d.ts +8 -1
- package/src/loader/shard-cache.js +33 -0
- package/src/loader/tensors/bf16-loader.js +2 -1
- package/src/loader/tensors/fp8-loader.js +1 -1
- package/src/loader/tensors/load-executor.js +10 -2
- package/src/loader/tensors/tensor-loader.js +1 -1
- package/src/loader/tensors/tensor-reader.js +18 -5
- package/src/loader/tensors/w8a8-loader.js +1 -1
- package/src/loader/weight-representation-plan.d.ts +1 -1
- package/src/memory/buffer-pool/readback.js +82 -0
- package/src/memory/buffer-pool.d.ts +11 -2
- package/src/memory/buffer-pool.js +12 -73
- package/src/partitions.d.ts +4 -3
- package/src/partitions.js +3 -2
- package/src/{converter → rules}/capability-transform-resolver.js +3 -3
- package/src/rules/generated/rule-bundle.json +1 -1
- package/src/rules/inference/capability-transforms.rules.json +95 -28
- package/src/rules/rule-registry.js +2 -0
- package/src/storage/artifact-range-loaders.js +42 -0
- package/src/storage/artifact-storage-context.d.ts +4 -0
- package/src/storage/artifact-storage-context.js +8 -18
- package/src/storage/backends/idb-store.js +1 -1
- package/src/storage/distribution-transport.js +1 -1
- package/src/storage/download/http-contract.js +1 -1
- package/src/storage/download/http-transport.js +4 -4
- package/src/storage/inventory.js +1 -1
- package/src/storage/model-cache.d.ts +2 -1
- package/src/storage/model-cache.js +12 -8
- package/src/storage/node-file-access.js +4 -0
- package/src/storage/quota.d.ts +2 -1
- package/src/storage/quota.js +2 -1
- package/src/storage/shard-manager.d.ts +1 -1
- package/src/storage/shards/integrity.js +17 -0
- package/src/storage/source-artifact-store.d.ts +1 -1
- package/src/storage/source-artifact-store.js +1 -0
- package/src/storage/source-storage-context.d.ts +4 -0
- package/src/storage/source-storage-context.js +14 -0
- package/src/storage/verified-piece-direct.js +153 -0
- package/src/storage/verified-piece-gather-lane.js +210 -0
- package/src/storage/verified-piece-gather-worker.js +739 -0
- package/src/storage/verified-piece-gather.d.ts +160 -0
- package/src/storage/verified-piece-gather.js +315 -0
- package/src/storage/verified-piece-reader.d.ts +15 -1
- package/src/storage/verified-piece-reader.js +85 -15
- package/src/storage/verified-piece-state-cache.d.ts +75 -0
- package/src/storage/verified-piece-state-cache.js +511 -0
- package/src/storage/verified-piece-storage.d.ts +22 -1
- package/src/storage/verified-piece-storage.js +299 -11
- package/src/storage/verified-piece-timings.js +25 -0
- package/src/tooling/browser-command-runner.js +1 -1
- package/src/tooling/browser-harness/benchmark-execution.js +2 -2
- package/src/tooling/browser-harness/benchmark-rerank.js +2 -2
- package/src/tooling/browser-harness/inference-execution.js +1 -1
- package/src/tooling/browser-harness/model-resolution.js +1 -1
- package/src/tooling/browser-harness/runtime-config.js +1 -1
- package/src/tooling/browser-harness/runtime-isolation.js +2 -5
- package/src/tooling/browser-harness/text-evidence.js +1 -1
- package/src/tooling/browser-harness-contract-helpers.js +1 -1
- package/src/tooling/browser-harness.js +2 -2
- package/src/tooling/capsule-v2.js +2 -2
- package/src/tooling/capsule.js +2 -2
- package/src/tooling/command-api-constants.js +4 -2
- package/src/tooling/command-api.d.ts +1 -1
- package/src/tooling/command-api.js +15 -21
- package/src/tooling/command-runtime-execution.js +1 -8
- package/src/tooling/diagnose-runner.js +3 -41
- package/src/tooling/node-browser/static-server.js +2 -2
- package/src/tooling/node-command-execution.js +2 -34
- package/src/tooling/node-converter/input.js +1 -1
- package/src/tooling/node-converter/output.js +2 -2
- package/src/tooling/node-converter.js +1 -1
- package/src/tooling/node-runtime-bridge.js +43 -0
- package/src/tooling/node-source-runtime.js +34 -35
- package/src/tooling/production-release.js +1 -1
- package/src/tooling/program-bundle/materialize.js +1 -1
- package/src/tooling/program-bundle-parity.d.ts +1 -1
- package/src/tooling/program-bundle-parity.js +1 -1
- package/src/tooling/program-bundle.d.ts +1 -1
- package/src/tooling/program-bundle.js +2 -2
- package/src/tooling/runtime-intent.js +1 -1
- package/src/tooling/source-artifact-adapter.js +3 -3
- package/src/tooling/source-runtime/node-loader.js +32 -5
- package/src/tooling/source-runtime-converter-config.js +28 -6
- package/src/tooling-exports.shared.d.ts +3 -3
- package/src/tooling-exports.shared.js +2 -2
- package/src/version.js +1 -1
- package/tests/kernels/browser/kernel-suite.js +1 -1
- package/tests/kernels/harness/benchmark.js +1 -1
- package/tools/README.md +74 -0
- package/src/config/runtime/profiles/qwen3-6-35b-a3b-q4_0-pooled-q8-common-diagnostic.json +0 -4407
- package/src/config/schema/loading.schema.js +0 -164
- package/src/inference/pipelines/text/moe-cpu.js +0 -20
- /package/src/config/{capsule-loading.d.ts → capsule/capsule-loading.d.ts} +0 -0
- /package/src/config/{capsule-release-contract.d.ts → capsule/capsule-release-contract.d.ts} +0 -0
- /package/src/config/{capsule-signature.d.ts → capsule/capsule-signature.d.ts} +0 -0
- /package/src/config/{capsule-signature.js → capsule/capsule-signature.js} +0 -0
- /package/src/config/{capsule-v3.d.ts → capsule/capsule-v3.d.ts} +0 -0
- /package/src/config/{capsule.d.ts → capsule/capsule.d.ts} +0 -0
- /package/src/config/schema/{diffusion.schema.d.ts → execution/diffusion.schema.d.ts} +0 -0
- /package/src/config/schema/{diffusion.schema.js → execution/diffusion.schema.js} +0 -0
- /package/src/config/schema/{energy.schema.d.ts → execution/energy.schema.d.ts} +0 -0
- /package/src/config/schema/{energy.schema.js → execution/energy.schema.js} +0 -0
- /package/src/config/schema/{hyper-connection-decoder.schema.d.ts → execution/hyper-connection-decoder.schema.d.ts} +0 -0
- /package/src/config/schema/{hyper-connection-decoder.schema.js → execution/hyper-connection-decoder.schema.js} +0 -0
- /package/src/config/schema/{inference.schema.js → execution/inference.schema.js} +0 -0
- /package/src/config/schema/{kernel-path.schema.js → execution/kernel-path.schema.js} +0 -0
- /package/src/config/schema/{kernel-registry.schema.js → execution/kernel-registry.schema.js} +0 -0
- /package/src/config/schema/{speculation-self.schema.d.ts → execution/speculation-self.schema.d.ts} +0 -0
- /package/src/config/schema/{speculation-self.schema.js → execution/speculation-self.schema.js} +0 -0
- /package/src/config/schema/{speculative.schema.d.ts → execution/speculative.schema.d.ts} +0 -0
- /package/src/config/schema/{speculative.schema.js → execution/speculative.schema.js} +0 -0
- /package/src/config/schema/{conversion-report.schema.d.ts → manifest/conversion-report.schema.d.ts} +0 -0
- /package/src/config/schema/{conversion-report.schema.js → manifest/conversion-report.schema.js} +0 -0
- /package/src/config/schema/{conversion.schema.d.ts → manifest/conversion.schema.d.ts} +0 -0
- /package/src/config/schema/{conversion.schema.js → manifest/conversion.schema.js} +0 -0
- /package/src/config/schema/{fp8.schema.d.ts → manifest/fp8.schema.d.ts} +0 -0
- /package/src/config/schema/{fp8.schema.js → manifest/fp8.schema.js} +0 -0
- /package/src/config/schema/{manifest-rope.schema.d.ts → manifest/manifest-rope.schema.d.ts} +0 -0
- /package/src/config/schema/{quantization-defaults.schema.d.ts → manifest/quantization-defaults.schema.d.ts} +0 -0
- /package/src/config/schema/{quantization-defaults.schema.js → manifest/quantization-defaults.schema.js} +0 -0
- /package/src/config/schema/{quantization.schema.d.ts → manifest/quantization.schema.d.ts} +0 -0
- /package/src/config/schema/{quantization.schema.js → manifest/quantization.schema.js} +0 -0
- /package/src/config/schema/{w8a8.schema.d.ts → manifest/w8a8.schema.d.ts} +0 -0
- /package/src/config/schema/{w8a8.schema.js → manifest/w8a8.schema.js} +0 -0
- /package/src/config/schema/{gpu-cache.schema.d.ts → platform/gpu-cache.schema.d.ts} +0 -0
- /package/src/config/schema/{memory-limits.schema.d.ts → platform/memory-limits.schema.d.ts} +0 -0
- /package/src/config/schema/{platform.schema.d.ts → platform/platform.schema.d.ts} +0 -0
- /package/src/config/schema/{support-tiers.schema.d.ts → policy/support-tiers.schema.d.ts} +0 -0
- /package/src/config/schema/{support-tiers.schema.js → policy/support-tiers.schema.js} +0 -0
- /package/src/config/schema/{adapter.schema.d.ts → runtime/adapter.schema.d.ts} +0 -0
- /package/src/config/schema/{adapter.schema.js → runtime/adapter.schema.js} +0 -0
- /package/src/config/schema/{benchmark.schema.d.ts → runtime/benchmark.schema.d.ts} +0 -0
- /package/src/config/schema/{benchmark.schema.js → runtime/benchmark.schema.js} +0 -0
- /package/src/config/schema/{bridge.schema.d.ts → runtime/bridge.schema.d.ts} +0 -0
- /package/src/config/schema/{bridge.schema.js → runtime/bridge.schema.js} +0 -0
- /package/src/config/schema/{browser-suite-metrics.schema.d.ts → runtime/browser-suite-metrics.schema.d.ts} +0 -0
- /package/src/config/schema/{browser-suite-metrics.schema.js → runtime/browser-suite-metrics.schema.js} +0 -0
- /package/src/config/schema/{buffer-pool.schema.d.ts → runtime/buffer-pool.schema.d.ts} +0 -0
- /package/src/config/schema/{debug.schema.d.ts → runtime/debug.schema.d.ts} +0 -0
- /package/src/config/schema/{debug.schema.js → runtime/debug.schema.js} +0 -0
- /package/src/config/schema/{distribution.schema.d.ts → runtime/distribution.schema.d.ts} +0 -0
- /package/src/config/schema/{distribution.schema.js → runtime/distribution.schema.js} +0 -0
- /package/src/config/schema/{ecosystem.schema.d.ts → runtime/ecosystem.schema.d.ts} +0 -0
- /package/src/config/schema/{ecosystem.schema.js → runtime/ecosystem.schema.js} +0 -0
- /package/src/config/schema/{harness.schema.d.ts → runtime/harness.schema.d.ts} +0 -0
- /package/src/config/schema/{harness.schema.js → runtime/harness.schema.js} +0 -0
- /package/src/config/schema/{hotswap.schema.d.ts → runtime/hotswap.schema.d.ts} +0 -0
- /package/src/config/schema/{hotswap.schema.js → runtime/hotswap.schema.js} +0 -0
- /package/src/config/schema/{intent-bundle.schema.d.ts → runtime/intent-bundle.schema.d.ts} +0 -0
- /package/src/config/schema/{intent-bundle.schema.js → runtime/intent-bundle.schema.js} +0 -0
- /package/src/config/schema/{runtime-profile.schema.d.ts → runtime/runtime-profile.schema.d.ts} +0 -0
- /package/src/config/schema/{runtime-profile.schema.js → runtime/runtime-profile.schema.js} +0 -0
- /package/src/config/schema/{storage.schema.d.ts → runtime/storage.schema.d.ts} +0 -0
- /package/src/config/schema/{storage.schema.js → runtime/storage.schema.js} +0 -0
- /package/src/config/schema/{tooling.schema.js → runtime/tooling.schema.js} +0 -0
- /package/src/config/schema/{units.schema.d.ts → runtime/units.schema.d.ts} +0 -0
- /package/src/config/schema/{units.schema.js → runtime/units.schema.js} +0 -0
- /package/src/config/schema/{backward-registry.schema.d.ts → training/backward-registry.schema.d.ts} +0 -0
- /package/src/config/schema/{backward-registry.schema.js → training/backward-registry.schema.js} +0 -0
- /package/src/config/schema/{distill-training.schema.d.ts → training/distill-training.schema.d.ts} +0 -0
- /package/src/config/schema/{distill-training.schema.js → training/distill-training.schema.js} +0 -0
- /package/src/config/schema/{lora.schema.d.ts → training/lora.schema.d.ts} +0 -0
- /package/src/config/schema/{training-metrics.schema.d.ts → training/training-metrics.schema.d.ts} +0 -0
- /package/src/config/schema/{training-metrics.schema.js → training/training-metrics.schema.js} +0 -0
- /package/src/config/schema/{training.schema.d.ts → training/training.schema.d.ts} +0 -0
- /package/src/config/schema/{training.schema.js → training/training.schema.js} +0 -0
- /package/src/config/schema/{ul-training.schema.d.ts → training/ul-training.schema.d.ts} +0 -0
- /package/src/config/schema/{ul-training.schema.js → training/ul-training.schema.js} +0 -0
- /package/src/{converter → config/transforms}/execution-graph/dtype.js +0 -0
- /package/src/{converter → config/transforms}/execution-graph/fusion.js +0 -0
- /package/src/{converter → config/transforms}/execution-graph/kv-storage.js +0 -0
- /package/src/{converter → config/transforms}/execution-graph/validation.js +0 -0
- /package/src/{converter → formats}/qwen4-exp-source-defaults.json +0 -0
- /package/src/inference/pipelines/text/{buffer-types.d.ts → contracts/buffer-types.d.ts} +0 -0
- /package/src/inference/pipelines/text/{embedding-contract.d.ts → embedding/embedding-contract.d.ts} +0 -0
- /package/src/inference/pipelines/text/{embedding-contract.js → embedding/embedding-contract.js} +0 -0
- /package/src/inference/pipelines/text/{precomputed-input-norm.js → execution/precomputed-input-norm.js} +0 -0
- /package/src/inference/pipelines/text/generator/{batch-retry-policy.js → decode/batch-retry-policy.js} +0 -0
- /package/src/inference/pipelines/text/{chat-format.d.ts → generator/request/chat-format.d.ts} +0 -0
- /package/src/inference/pipelines/text/{chat-format.js → generator/request/chat-format.js} +0 -0
- /package/src/inference/pipelines/text/generator/{recorder-stats.js → session/recorder-stats.js} +0 -0
- /package/src/inference/pipelines/text/generator/{whole-token-template.js → whole-token/whole-token-template.js} +0 -0
- /package/src/inference/pipelines/text/{init-chat-templates.js → initialization/init-chat-templates.js} +0 -0
- /package/src/inference/pipelines/text/{load-policy.js → loading/load-policy.js} +0 -0
- /package/src/inference/pipelines/text/{pipeline-load-timing.js → loading/pipeline-load-timing.js} +0 -0
- /package/src/inference/pipelines/text/{generator → logits}/hidden-projection.d.ts +0 -0
- /package/src/inference/pipelines/text/{operator-events.d.ts → observation/operator-events.d.ts} +0 -0
- /package/src/inference/pipelines/text/{stage-names.d.ts → observation/stage-names.d.ts} +0 -0
- /package/src/inference/pipelines/text/{stage-names.js → observation/stage-names.js} +0 -0
- /package/src/inference/pipelines/text/{tsir-fixture-writer.d.ts → observation/tsir-fixture-writer.d.ts} +0 -0
- /package/src/inference/pipelines/text/{generator → per-layer}/per-layer-buffer.js +0 -0
- /package/src/inference/pipelines/text/{finiteness-policy.js → policy/finiteness-policy.js} +0 -0
|
@@ -0,0 +1,110 @@
|
|
|
1
|
+
import { isGemma4PerLayerEmbedTensor } from "../tensor-transform.js";
|
|
2
|
+
import { getNestedTextConfig } from "../artifact-identity.js";
|
|
3
|
+
import { modelHasMoETensors } from "../manifest-builder.js";
|
|
4
|
+
import { normalizeTensorName } from "../tokenizer-contract.js";
|
|
5
|
+
|
|
6
|
+
export function validateInt4PleMaterializationContract(tensorLocations, inference, modelId) {
|
|
7
|
+
const materialization = inference?.session?.perLayerInputs?.materialization;
|
|
8
|
+
if (materialization !== 'gpu_split_tables') {
|
|
9
|
+
return;
|
|
10
|
+
}
|
|
11
|
+
for (const [name, location] of Object.entries(tensorLocations ?? {})) {
|
|
12
|
+
const sourceTransform = location?.sourceTransform ?? null;
|
|
13
|
+
if (
|
|
14
|
+
isGemma4PerLayerEmbedTensor(name)
|
|
15
|
+
&& sourceTransform?.kind === 'litert_axis_dequant'
|
|
16
|
+
&& String(sourceTransform?.sourceDtype ?? '').toUpperCase() === 'INT4'
|
|
17
|
+
) {
|
|
18
|
+
throw new Error(
|
|
19
|
+
`Manifest "${modelId}" Gemma 4 INT4 PLE tensor "${name}" cannot use ` +
|
|
20
|
+
'inference.session.perLayerInputs.materialization="gpu_split_tables". ' +
|
|
21
|
+
'Use materialization="range_backed" or disable INT4 PLE quantization.'
|
|
22
|
+
);
|
|
23
|
+
}
|
|
24
|
+
}
|
|
25
|
+
}
|
|
26
|
+
|
|
27
|
+
function resolveGemma4TextConfig(rawConfig) {
|
|
28
|
+
const textConfig = getNestedTextConfig(rawConfig);
|
|
29
|
+
const modelType = String(textConfig?.model_type ?? rawConfig?.model_type ?? '').trim().toLowerCase();
|
|
30
|
+
if (
|
|
31
|
+
modelType !== 'gemma4'
|
|
32
|
+
&& modelType !== 'gemma4_text'
|
|
33
|
+
&& modelType !== 'gemma4_unified'
|
|
34
|
+
&& modelType !== 'gemma4_unified_text'
|
|
35
|
+
) {
|
|
36
|
+
return null;
|
|
37
|
+
}
|
|
38
|
+
return textConfig ?? rawConfig ?? null;
|
|
39
|
+
}
|
|
40
|
+
|
|
41
|
+
function collectGemma4UnsupportedTensorFlags(tensors) {
|
|
42
|
+
const names = Array.isArray(tensors) ? tensors.map((tensor) => normalizeTensorName(tensor).toLowerCase()) : [];
|
|
43
|
+
const flags = [];
|
|
44
|
+
if (names.some((name) => name.includes('.experts.gate_up_proj'))) {
|
|
45
|
+
flags.push('experts.gate_up_proj');
|
|
46
|
+
}
|
|
47
|
+
if (names.some((name) => name.includes('.experts.down_proj'))) {
|
|
48
|
+
flags.push('experts.down_proj');
|
|
49
|
+
}
|
|
50
|
+
if (names.some((name) => name.includes('.router.per_expert_scale'))) {
|
|
51
|
+
flags.push('router.per_expert_scale');
|
|
52
|
+
}
|
|
53
|
+
if (names.some((name) => name.includes('.router.scale'))) {
|
|
54
|
+
flags.push('router.scale');
|
|
55
|
+
}
|
|
56
|
+
if (names.some((name) => name.includes('.post_feedforward_layernorm_1.'))) {
|
|
57
|
+
flags.push('post_feedforward_layernorm_1');
|
|
58
|
+
}
|
|
59
|
+
if (names.some((name) => name.includes('.post_feedforward_layernorm_2.'))) {
|
|
60
|
+
flags.push('post_feedforward_layernorm_2');
|
|
61
|
+
}
|
|
62
|
+
if (names.some((name) => name.includes('.pre_feedforward_layernorm_2.'))) {
|
|
63
|
+
flags.push('pre_feedforward_layernorm_2');
|
|
64
|
+
}
|
|
65
|
+
return flags;
|
|
66
|
+
}
|
|
67
|
+
|
|
68
|
+
export function assertSupportedGemma4Conversion(model, tensors, modelId) {
|
|
69
|
+
const rawConfig = model?.config ?? null;
|
|
70
|
+
const textConfig = resolveGemma4TextConfig(rawConfig);
|
|
71
|
+
if (!textConfig) return;
|
|
72
|
+
|
|
73
|
+
const hiddenSizePerLayerInput = Number(textConfig.hidden_size_per_layer_input ?? 0);
|
|
74
|
+
if (Number.isFinite(hiddenSizePerLayerInput) && hiddenSizePerLayerInput > 0) {
|
|
75
|
+
const names = Array.isArray(tensors)
|
|
76
|
+
? tensors.map((tensor) => String(tensor?.name ?? '').trim())
|
|
77
|
+
: [];
|
|
78
|
+
const requiredNames = [
|
|
79
|
+
'embed_tokens_per_layer.weight',
|
|
80
|
+
'per_layer_input_gate.weight',
|
|
81
|
+
'per_layer_projection.weight',
|
|
82
|
+
'post_per_layer_input_norm.weight',
|
|
83
|
+
'per_layer_model_projection.weight',
|
|
84
|
+
'per_layer_projection_norm.weight',
|
|
85
|
+
];
|
|
86
|
+
const missing = requiredNames.filter((suffix) => !names.some((name) => name.endsWith(suffix)));
|
|
87
|
+
if (missing.length > 0) {
|
|
88
|
+
throw new Error(
|
|
89
|
+
`Gemma 4 model "${modelId}" declares hidden_size_per_layer_input=${hiddenSizePerLayerInput}, ` +
|
|
90
|
+
`but the checkpoint is missing required per-layer input tensors: ${missing.join(', ')}.`
|
|
91
|
+
);
|
|
92
|
+
}
|
|
93
|
+
}
|
|
94
|
+
|
|
95
|
+
if (textConfig.enable_moe_block !== true) {
|
|
96
|
+
return;
|
|
97
|
+
}
|
|
98
|
+
|
|
99
|
+
const unsupportedFlags = collectGemma4UnsupportedTensorFlags(tensors);
|
|
100
|
+
if (unsupportedFlags.length === 0 && !modelHasMoETensors({ tensors })) {
|
|
101
|
+
return;
|
|
102
|
+
}
|
|
103
|
+
|
|
104
|
+
throw new Error(
|
|
105
|
+
`Gemma 4 model "${modelId}" is not supported yet: Gemma 4 MoE decoder blocks require ` +
|
|
106
|
+
'Gemma-specific router scaling and dual dense+MoE FFN execution, but current Doppler MoE runtime ' +
|
|
107
|
+
'only supports Mixtral/GPT-OSS semantics. ' +
|
|
108
|
+
`Detected: ${unsupportedFlags.length > 0 ? unsupportedFlags.join(', ') : 'Gemma 4 MoE tensors'}.`
|
|
109
|
+
);
|
|
110
|
+
}
|
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import { LINEAR_NORM_GATE_ACTIVATIONS } from
|
|
1
|
+
import { LINEAR_NORM_GATE_ACTIVATIONS } from "../config/schema/manifest/manifest.schema.js";
|
|
2
2
|
|
|
3
3
|
// Source semantics of the linear-attention gated RMSNorm activation, as pinned
|
|
4
4
|
// in artifacts/qwen38-flash-next-20261013/source (Transformers e0eb2810):
|
|
@@ -1,12 +1,5 @@
|
|
|
1
1
|
import { validateReasoningOutput } from '../config/reasoning-output.js';
|
|
2
|
-
import {
|
|
3
|
-
// Constants
|
|
4
|
-
SHARD_SIZE as SCHEMA_SHARD_SIZE,
|
|
5
|
-
RDRR_VERSION as SCHEMA_RDRR_VERSION,
|
|
6
|
-
ConversionStage as SchemaConversionStage,
|
|
7
|
-
DEFAULT_MANIFEST_INFERENCE,
|
|
8
|
-
formatBytes,
|
|
9
|
-
} from '../config/schema/index.js';
|
|
2
|
+
import { SHARD_SIZE as SCHEMA_SHARD_SIZE, RDRR_VERSION as SCHEMA_RDRR_VERSION, ConversionStage as SchemaConversionStage, DEFAULT_MANIFEST_INFERENCE, formatBytes, createConverterConfig } from '../config/schema/index.js';
|
|
10
3
|
import {
|
|
11
4
|
classifyTensorRole,
|
|
12
5
|
generateShardFilename,
|
|
@@ -18,7 +11,7 @@ import {
|
|
|
18
11
|
getInferenceLayerPatternContractArtifact,
|
|
19
12
|
selectRuleValue,
|
|
20
13
|
} from '../rules/rule-registry.js';
|
|
21
|
-
|
|
14
|
+
|
|
22
15
|
import { buildExecutionContractArtifact } from '../config/execution-contract-check.js';
|
|
23
16
|
import { buildManifestRequiredInferenceFieldsArtifact } from '../config/required-inference-fields-contract-check.js';
|
|
24
17
|
import { resolveEosTokenId } from './tokenizer-special-tokens.js';
|
|
@@ -1,3 +1,4 @@
|
|
|
1
|
+
import { bindQwen4ExpGGUFTable } from '../../formats/gguf/qwen4-exp-source.js';
|
|
1
2
|
import { ConvertStage } from '../core.js';
|
|
2
3
|
import { extractArchitecture } from '../artifact-identity.js';
|
|
3
4
|
import { normalizeQwenGGUFSource } from '../../formats/gguf/qwen-source.js';
|
|
@@ -28,10 +29,58 @@ export async function parseGGUFModel(adapter) {
|
|
|
28
29
|
|
|
29
30
|
if (signal?.aborted) throw toAbortError();
|
|
30
31
|
|
|
31
|
-
const
|
|
32
|
-
|
|
32
|
+
const sources = (adapter.files ?? [file]).map(normalizeTensorSource);
|
|
33
|
+
if (sources.length === 0) throw new Error('GGUF input requires at least one source.');
|
|
34
|
+
const pieces = [];
|
|
35
|
+
for (const source of sources) {
|
|
36
|
+
if (signal?.aborted) throw toAbortError();
|
|
37
|
+
pieces.push({ source, info: await parseGGUFHeaderFromSource(source) });
|
|
38
|
+
}
|
|
39
|
+
const split = pieces.length > 1 || Number(pieces[0].info.metadata?.['split.count'] ?? 1) > 1;
|
|
40
|
+
if (split) {
|
|
41
|
+
const count = pieces[0].info.metadata?.['split.count'];
|
|
42
|
+
const total = pieces[0].info.metadata?.['split.tensors.count'];
|
|
43
|
+
if (!Number.isSafeInteger(count) || count < 2 || count !== pieces.length
|
|
44
|
+
|| !Number.isSafeInteger(total) || total < 1) {
|
|
45
|
+
throw new Error('GGUF split requires every declared part and a valid total tensor count.');
|
|
46
|
+
}
|
|
47
|
+
const indices = new Set();
|
|
48
|
+
for (const { info } of pieces) {
|
|
49
|
+
const index = info.metadata?.['split.no'];
|
|
50
|
+
if (!Number.isSafeInteger(index) || index < 0 || index >= count || indices.has(index)
|
|
51
|
+
|| info.metadata?.['split.count'] !== count || info.metadata?.['split.tensors.count'] !== total) {
|
|
52
|
+
throw new Error('GGUF split has duplicate, missing or inconsistent part metadata.');
|
|
53
|
+
}
|
|
54
|
+
indices.add(index);
|
|
55
|
+
}
|
|
56
|
+
pieces.sort((a, b) => a.info.metadata['split.no'] - b.info.metadata['split.no']);
|
|
57
|
+
if (pieces.reduce((n, piece) => n + piece.info.tensors.length, 0) !== total) {
|
|
58
|
+
throw new Error('GGUF split tensor count differs from its declared total.');
|
|
59
|
+
}
|
|
60
|
+
for (const { info } of pieces.slice(1)) {
|
|
61
|
+
for (const [key, value] of Object.entries(info.metadata)) {
|
|
62
|
+
if (key.startsWith('split.')) continue;
|
|
63
|
+
if (JSON.stringify(value) !== JSON.stringify(pieces[0].info.metadata[key])) {
|
|
64
|
+
throw new Error(`GGUF split metadata differs from part zero: ${key}.`);
|
|
65
|
+
}
|
|
66
|
+
}
|
|
67
|
+
}
|
|
68
|
+
}
|
|
69
|
+
const { source, info: primary } = pieces[0];
|
|
70
|
+
const names = new Set();
|
|
71
|
+
const allTensors = pieces.flatMap(piece => piece.info.tensors.map(tensor => {
|
|
72
|
+
if (names.has(tensor.name)) throw new Error(`GGUF split contains duplicate tensor ${tensor.name}.`);
|
|
73
|
+
if (!Number.isSafeInteger(tensor.offset) || tensor.offset < 0 || !Number.isSafeInteger(tensor.size)
|
|
74
|
+
|| tensor.size < 1 || tensor.offset > piece.source.size - tensor.size) {
|
|
75
|
+
throw new Error(`GGUF tensor ${tensor.name} exceeds its source byte extent.`);
|
|
76
|
+
}
|
|
77
|
+
names.add(tensor.name);
|
|
78
|
+
return { ...tensor, source: piece.source };
|
|
79
|
+
}));
|
|
80
|
+
const ggufInfo = { ...primary, tensors: allTensors };
|
|
33
81
|
const metadata = ggufInfo.metadata ?? {};
|
|
34
82
|
const native = normalizeQwenGGUFSource(ggufInfo.architecture, metadata);
|
|
83
|
+
const ngramTable = ggufInfo.architecture === 'qwen4exp' ? bindQwen4ExpGGUFTable(native.config, allTensors) : null;
|
|
35
84
|
const tokenizer = native ? buildGGUFTokenizer(metadata) : null;
|
|
36
85
|
const tensors = ggufInfo.tensors.filter(tensor => {
|
|
37
86
|
const layer = /^blk\.(\d+)\./u.exec(tensor.name);
|
|
@@ -47,12 +96,13 @@ export async function parseGGUFModel(adapter) {
|
|
|
47
96
|
tokenizerConfig: tokenizer?.tokenizerConfig ?? null,
|
|
48
97
|
tensors: tensors.map((tensor) => ({
|
|
49
98
|
...tensor,
|
|
99
|
+
...(tensor === ngramTable?.table ? { name: ngramTable.name, originalName: tensor.name, role: 'ngram_table', group: 'ngram_table' } : {}),
|
|
50
100
|
sourceShape: [...tensor.shape],
|
|
51
101
|
shape: native && /ffn_gate_inp_shexp\.weight$/u.test(tensor.name) && tensor.shape.length === 1
|
|
52
102
|
? [1, ...tensor.shape] : [...tensor.shape].reverse(),
|
|
53
103
|
...(tensor.shape.length >= 2 || (native && /ffn_gate_inp_shexp\.weight$/u.test(tensor.name)) ? { layout: 'row' } : {}),
|
|
54
|
-
file: source.file,
|
|
55
|
-
source,
|
|
104
|
+
file: tensor.source.file,
|
|
105
|
+
source: tensor.source,
|
|
56
106
|
offset: tensor.offset,
|
|
57
107
|
})),
|
|
58
108
|
config: native ? { ...ggufInfo.config, ...native.config } : ggufInfo.config,
|
|
@@ -1,5 +1,5 @@
|
|
|
1
|
-
import { validateFP8Policy } from
|
|
2
|
-
import { validateW8A8Policy } from
|
|
1
|
+
import { validateFP8Policy } from "../config/schema/manifest/fp8.schema.js";
|
|
2
|
+
import { validateW8A8Policy } from "../config/schema/manifest/w8a8.schema.js";
|
|
3
3
|
|
|
4
4
|
import { DEFAULT_QUANTIZATION_DEFAULTS, DEFAULT_Q4K_LAYOUT } from '../config/schema/index.js';
|
|
5
5
|
import { classifyTensorRole } from '../formats/rdrr/index.js';
|
|
@@ -0,0 +1,94 @@
|
|
|
1
|
+
import profile from '../config/conversion/qwen3/qwen-3-8-flash-next-text-q4k-ehf16-af32.json' with { type: 'json' };
|
|
2
|
+
import policy from '../config/conversion/qwen3/qwen4exp-gguf-source-policy.json' with { type: 'json' };
|
|
3
|
+
import registry from '../config/kernels/registry.json' with { type: 'json' };
|
|
4
|
+
import { KERNEL_REF_CONTENT_DIGESTS } from '../config/kernels/kernel-ref-digests.js';
|
|
5
|
+
import { ggufBlockFormat, isGgufBlockSourceDtype } from '../formats/gguf/block-storage.js';
|
|
6
|
+
import { selectRuleValue } from '../rules/rule-registry.js';
|
|
7
|
+
import { cloneJsonValue } from '../formats/clone-json.js';
|
|
8
|
+
import { expandExecutionV1 } from '../config/schema/execution/execution-v1.schema.js';
|
|
9
|
+
import { SUPPORTED_EXECUTION_V1_OPS } from '../config/supported-operations.js';
|
|
10
|
+
|
|
11
|
+
/** Compile the GGUF source's physical projection types, never the HF Q4K recipe.
|
|
12
|
+
* @type {typeof import('./qwen4-exp-source-execution.js').createQwen4ExpSourceExecution} */
|
|
13
|
+
export function createQwen4ExpSourceExecution(tensors, numLayers) {
|
|
14
|
+
if (!Array.isArray(tensors) || tensors.length === 0 || !Number.isInteger(numLayers) || numLayers < 1) {
|
|
15
|
+
throw new Error('Qwen4Exp source execution requires tensor storage facts and a positive layer count.');
|
|
16
|
+
}
|
|
17
|
+
const byName = new Map(tensors.map(t => [t.name, t]));
|
|
18
|
+
if (byName.size !== tensors.length) throw new Error('Duplicate Qwen4Exp source tensor names.');
|
|
19
|
+
const graph = /** @type {import("../config/schema/execution/execution-v1.schema.js").ExecutionV1GraphSchema} */ (/** @type {unknown} */ (cloneJsonValue(profile.execution)));
|
|
20
|
+
for (const spec of Object.values(graph.kernels)) {
|
|
21
|
+
const digest = KERNEL_REF_CONTENT_DIGESTS[`${spec.kernel}#${spec.entry}`];
|
|
22
|
+
if (!digest) throw new Error(`Missing family kernel digest for ${spec.kernel}.`);
|
|
23
|
+
spec.digest = `sha256:${digest}`;
|
|
24
|
+
}
|
|
25
|
+
const operations = /** @type {import("../config/schema/execution/kernel-registry.schema.js").KernelRegistrySchema} */ (/** @type {unknown} */ (registry)).operations;
|
|
26
|
+
const floats = /** @type {Record<string, string>} */ (policy.floatVariants);
|
|
27
|
+
const phases = /** @type {Record<string, number>} */ (policy.phases);
|
|
28
|
+
/** @param {string} operation @param {string} variant */
|
|
29
|
+
const install = (operation, variant) => {
|
|
30
|
+
const spec = operations[operation]?.variants?.[variant];
|
|
31
|
+
if (!spec) throw new Error(`Unsupported Qwen4Exp source kernel ${operation}:${variant}.`);
|
|
32
|
+
const key = `source_${operation}_${variant}`;
|
|
33
|
+
const digest = KERNEL_REF_CONTENT_DIGESTS[`${spec.wgsl}#${spec.entryPoint}`];
|
|
34
|
+
if (!digest) throw new Error(`Missing source kernel digest for ${key}.`);
|
|
35
|
+
graph.kernels[key] = { kernel: spec.wgsl, entry: spec.entryPoint,
|
|
36
|
+
digest: `sha256:${digest}`, ...(spec.wgslOverrides ? { constants: cloneJsonValue(spec.wgslOverrides) } : {}) };
|
|
37
|
+
return key;
|
|
38
|
+
};
|
|
39
|
+
/** @param {string} role @param {string} name @param {string} phase
|
|
40
|
+
* @returns {import("../config/schema/execution/execution-v1.schema.js").ExecutionV1StepTuple} */
|
|
41
|
+
const projection = (role, name, phase) => {
|
|
42
|
+
const tensor = byName.get(name);
|
|
43
|
+
if (!tensor) throw new Error(`Missing Qwen4Exp projection ${name}.`);
|
|
44
|
+
const variant = isGgufBlockSourceDtype(tensor.dtype)
|
|
45
|
+
? selectRuleValue('kernels', 'matmul', 'q40Variant', {
|
|
46
|
+
M: phases[phase], weightDtype: ggufBlockFormat(tensor.dtype).variant,
|
|
47
|
+
}) : floats[tensor.dtype];
|
|
48
|
+
if (!variant) throw new Error(`Unsupported Qwen4Exp projection ${name} dtype ${tensor.dtype}.`);
|
|
49
|
+
return [role, install('matmul', variant), name];
|
|
50
|
+
};
|
|
51
|
+
for (const phase of /** @type {const} */ (['decode', 'prefill'])) {
|
|
52
|
+
// Retain non-projection semantics. Each physical projection is then bound
|
|
53
|
+
// explicitly per layer, including mixed attention and shared-expert types.
|
|
54
|
+
const unweighted = graph[phase].filter(step => Array.isArray(step) && step.length < 3);
|
|
55
|
+
/** @type {Map<string, import("../config/schema/execution/execution-v1.schema.js").ExecutionV1LayerGroupSchema>} */
|
|
56
|
+
const groups = new Map();
|
|
57
|
+
for (let layer = 0; layer < numLayers; layer++) {
|
|
58
|
+
const steps = [
|
|
59
|
+
...Object.entries(policy.layerRoles).flatMap(([role, suffix]) => {
|
|
60
|
+
const name = `blk.${layer}.${suffix}`;
|
|
61
|
+
if (!byName.has(name)) return [];
|
|
62
|
+
const step = projection(role, name, phase);
|
|
63
|
+
step[2] = name.replace(`blk.${layer}.`, 'blk.{L}.');
|
|
64
|
+
return [step];
|
|
65
|
+
}),
|
|
66
|
+
];
|
|
67
|
+
const key = JSON.stringify(steps);
|
|
68
|
+
let group = groups.get(key);
|
|
69
|
+
if (!group) { group = { layers: [], steps }; groups.set(key, group); }
|
|
70
|
+
group.layers.push(layer);
|
|
71
|
+
}
|
|
72
|
+
graph[phase] = [...cloneJsonValue(unweighted), ...groups.values()];
|
|
73
|
+
}
|
|
74
|
+
graph.postLayer = Object.entries(policy.rootRoles).map(([role, name]) =>
|
|
75
|
+
projection(role, name, role === 'lm_head_prefill' ? 'prefill' : 'decode'));
|
|
76
|
+
graph.postLayer.push(['sample', 'sample']);
|
|
77
|
+
const embedding = byName.get(policy.embedding);
|
|
78
|
+
if (!embedding || !isGgufBlockSourceDtype(embedding.dtype)) {
|
|
79
|
+
throw new Error('Qwen4Exp GGUF source execution requires a supported packed token embedding.');
|
|
80
|
+
}
|
|
81
|
+
graph.preLayer = [['embed', install('gather', ggufBlockFormat(embedding.dtype).variant), policy.embedding]];
|
|
82
|
+
const pleTable = byName.get(policy.pleRowDecode.tensor);
|
|
83
|
+
if (pleTable) {
|
|
84
|
+
if (pleTable.dtype !== policy.pleRowDecode.dtype) throw new Error('Unsupported Qwen4Exp source PLE row storage.');
|
|
85
|
+
const key = install(policy.pleRowDecode.operation, policy.pleRowDecode.variant);
|
|
86
|
+
graph.kernels.ple_row_decode = graph.kernels[key];
|
|
87
|
+
delete graph.kernels[key];
|
|
88
|
+
}
|
|
89
|
+
const recurrent = install(policy.linearRecurrent.operation, policy.linearRecurrent.variant);
|
|
90
|
+
graph.kernels.gated_delta_recurrent = graph.kernels[recurrent];
|
|
91
|
+
delete graph.kernels[recurrent];
|
|
92
|
+
expandExecutionV1(graph, { knownOps: SUPPORTED_EXECUTION_V1_OPS, strict: true });
|
|
93
|
+
return graph;
|
|
94
|
+
}
|
|
@@ -1,207 +1,7 @@
|
|
|
1
|
-
import sourceDefaults from './qwen4-exp-source-defaults.json' with { type: 'json' };
|
|
2
1
|
import { sha256BytesHex } from '../formats/sha256.js';
|
|
3
|
-
import {
|
|
4
|
-
deriveNgramLayerTable,
|
|
5
|
-
deriveNgramMultipliers,
|
|
6
|
-
} from '../config/schema/ngram-embedding.schema.js';
|
|
2
|
+
import { resolveQwen4ExpSourceContracts } from '../formats/qwen4-exp-source-contracts.js';
|
|
7
3
|
|
|
8
|
-
// Lowers a qwen4_exp (Qwen3.8 Flash-Next) source config into Doppler's
|
|
9
|
-
// hyper-connection decoder contracts. Omitted source fields resolve from the
|
|
10
|
-
// pinned Qwen4ExpTextConfig class defaults (qwen4-exp-source-defaults.json),
|
|
11
|
-
// never from another family. Upstream validate_architecture rules are applied
|
|
12
|
-
// here so a malformed source fails before conversion.
|
|
13
|
-
|
|
14
|
-
/** @type {Readonly<Record<string, unknown>>} */
|
|
15
|
-
const DEFAULTS = sourceDefaults.defaults;
|
|
16
|
-
const QWEN4_EXP_FAMILIES = new Set(sourceDefaults.sourceModelTypes);
|
|
17
|
-
const QSA_FIELDS = Object.freeze([
|
|
18
|
-
'indexer_n_heads', 'indexer_kv_heads', 'indexer_head_dim', 'indexer_budget', 'indexer_compress_ratio',
|
|
19
|
-
]);
|
|
20
4
|
const CONTRACT_FIELDS = Object.freeze(['hyperConnections', 'indexedAttention', 'ngramEmbedding']);
|
|
21
|
-
/** @type {Readonly<Record<string, 'bf16'>>} */
|
|
22
|
-
const SOURCE_TABLE_DTYPES = Object.freeze({ bfloat16: 'bf16' });
|
|
23
|
-
|
|
24
|
-
/** @param {Record<string, any> | null | undefined} rawConfig @returns {Record<string, any> | null} */
|
|
25
|
-
function textConfigOf(rawConfig) {
|
|
26
|
-
if (!rawConfig || typeof rawConfig !== 'object') return null;
|
|
27
|
-
const text = rawConfig.text_config;
|
|
28
|
-
return text && typeof text === 'object' ? text : rawConfig;
|
|
29
|
-
}
|
|
30
|
-
|
|
31
|
-
/** Whether the source config names the qwen4_exp family. @type {typeof import('./qwen4-exp-source.js').isQwen4ExpSource} */
|
|
32
|
-
export function isQwen4ExpSource(rawConfig) {
|
|
33
|
-
const top = String(rawConfig?.model_type ?? '').trim().toLowerCase();
|
|
34
|
-
const text = String(textConfigOf(rawConfig)?.model_type ?? '').trim().toLowerCase();
|
|
35
|
-
return QWEN4_EXP_FAMILIES.has(top) || QWEN4_EXP_FAMILIES.has(text);
|
|
36
|
-
}
|
|
37
|
-
|
|
38
|
-
/** @param {Record<string, any>} text @param {string} key @returns {any} */
|
|
39
|
-
function sourceField(text, key) {
|
|
40
|
-
if (text[key] !== undefined) return text[key];
|
|
41
|
-
if (!Object.hasOwn(DEFAULTS, key)) {
|
|
42
|
-
throw new Error(`qwen4_exp source omits "${key}", which has no pinned Qwen4ExpTextConfig default.`);
|
|
43
|
-
}
|
|
44
|
-
return DEFAULTS[key];
|
|
45
|
-
}
|
|
46
|
-
|
|
47
|
-
/** @param {unknown} value @param {string} label @returns {number} */
|
|
48
|
-
function requireInteger(value, label) {
|
|
49
|
-
if (typeof value !== 'number' || !Number.isSafeInteger(value)) {
|
|
50
|
-
throw new Error(`qwen4_exp source ${label} must be an integer; got ${JSON.stringify(value)}.`);
|
|
51
|
-
}
|
|
52
|
-
return value;
|
|
53
|
-
}
|
|
54
|
-
|
|
55
|
-
/** @param {Record<string, any>} text @returns {string[]} */
|
|
56
|
-
function sourceLayerTypes(text) {
|
|
57
|
-
if (Array.isArray(text.layer_types)) return text.layer_types.map(String);
|
|
58
|
-
const numLayers = requireInteger(sourceField(text, 'num_hidden_layers'), 'num_hidden_layers');
|
|
59
|
-
const interval = requireInteger(text.full_attention_interval ?? 4, 'full_attention_interval');
|
|
60
|
-
return Array.from({ length: numLayers }, (_, i) => ((i + 1) % interval ? 'linear_attention' : 'indexed_attention'));
|
|
61
|
-
}
|
|
62
|
-
|
|
63
|
-
/** @param {Record<string, any>} text @returns {number} */
|
|
64
|
-
function resolveRotaryDim(text) {
|
|
65
|
-
const factor = text.rope_parameters?.partial_rotary_factor ?? 1.0;
|
|
66
|
-
return Math.trunc(requireInteger(sourceField(text, 'head_dim'), 'head_dim') * factor);
|
|
67
|
-
}
|
|
68
|
-
|
|
69
|
-
/** @param {Record<string, any>} text @returns {number} */
|
|
70
|
-
function resolveBoundaryToken(text) {
|
|
71
|
-
const eos = sourceField(text, 'eos_token_id');
|
|
72
|
-
const value = Array.isArray(eos) ? eos[0] : eos;
|
|
73
|
-
if (!Number.isSafeInteger(value)) {
|
|
74
|
-
throw new Error('qwen4_exp PLE requires text_config.eos_token_id (the n-gram boundary token).');
|
|
75
|
-
}
|
|
76
|
-
return value;
|
|
77
|
-
}
|
|
78
|
-
|
|
79
|
-
/**
|
|
80
|
-
* @param {Record<string, any>} text @param {string[]} layerTypes
|
|
81
|
-
* @returns {import('../config/schema/ngram-embedding.schema.js').NgramEmbeddingSchema | null}
|
|
82
|
-
*/
|
|
83
|
-
function lowerNgramEmbedding(text, layerTypes) {
|
|
84
|
-
const rawIds = sourceField(text, 'ple_layer_ids') ?? [];
|
|
85
|
-
const sourceLayerIds = [...new Set(rawIds.map((/** @type {unknown} */ id) => requireInteger(id, 'ple_layer_ids[]')))].sort((a, b) => a - b);
|
|
86
|
-
if (sourceLayerIds.length === 0) return null;
|
|
87
|
-
const hiddenSize = requireInteger(sourceField(text, 'hidden_size'), 'hidden_size');
|
|
88
|
-
const embedDim = requireInteger(sourceField(text, 'ple_embed_dim') ?? hiddenSize, 'ple_embed_dim');
|
|
89
|
-
const ngramSize = requireInteger(sourceField(text, 'ngram_size'), 'ngram_size');
|
|
90
|
-
const headsPerNgram = requireInteger(sourceField(text, 'heads_per_ngram'), 'heads_per_ngram');
|
|
91
|
-
const headCount = (ngramSize - 1) * headsPerNgram;
|
|
92
|
-
if (headCount <= 0 || embedDim <= 0 || embedDim % headCount !== 0) {
|
|
93
|
-
throw new Error(`qwen4_exp ple_embed_dim ${embedDim} must be positive and divisible by ${headCount} n-gram heads.`);
|
|
94
|
-
}
|
|
95
|
-
const unigramVocabSize = requireInteger(sourceField(text, 'vocab_size'), 'vocab_size');
|
|
96
|
-
const vocabSizeBase = requireInteger(sourceField(text, 'ngram_vocab_size_base'), 'ngram_vocab_size_base');
|
|
97
|
-
const vocabPadMultiple = requireInteger(sourceField(text, 'make_ngram_vocab_size_divisible_by'), 'make_ngram_vocab_size_divisible_by');
|
|
98
|
-
const seed = requireInteger(sourceField(text, 'seed'), 'seed');
|
|
99
|
-
const convKernelSize = requireInteger(sourceField(text, 'ple_conv_kernel_size'), 'ple_conv_kernel_size');
|
|
100
|
-
const sourceDtype = String(text.dtype ?? text.torch_dtype ?? '');
|
|
101
|
-
const tableDtype = SOURCE_TABLE_DTYPES[sourceDtype];
|
|
102
|
-
if (!tableDtype) throw new Error(`qwen4_exp n-gram table dtype ${JSON.stringify(sourceDtype)} is not a supported stored dtype.`);
|
|
103
|
-
const layers = sourceLayerIds.map((sourceLayerId, pleLayerIndex) => {
|
|
104
|
-
if (sourceLayerId < 1 || sourceLayerId > layerTypes.length) {
|
|
105
|
-
throw new Error(`qwen4_exp ple_layer_ids must be one-indexed ids in [1, ${layerTypes.length}]; got ${sourceLayerId}.`);
|
|
106
|
-
}
|
|
107
|
-
if (layerTypes[sourceLayerId - 1] !== 'linear_attention') {
|
|
108
|
-
throw new Error(`qwen4_exp PLE is only supported on linear_attention layers; got layer id ${sourceLayerId}.`);
|
|
109
|
-
}
|
|
110
|
-
const table = deriveNgramLayerTable({ vocabSizeBase, vocabPadMultiple, headCount, pleLayerIndex });
|
|
111
|
-
return {
|
|
112
|
-
sourceLayerId,
|
|
113
|
-
decoderLayerIndex: sourceLayerId - 1,
|
|
114
|
-
pleLayerIndex,
|
|
115
|
-
multipliers: deriveNgramMultipliers(unigramVocabSize, ngramSize, pleLayerIndex, seed).map(String),
|
|
116
|
-
headVocabSizes: table.headVocabSizes,
|
|
117
|
-
headOffsets: table.headOffsets,
|
|
118
|
-
unpaddedRows: table.unpaddedRows,
|
|
119
|
-
rows: table.rows,
|
|
120
|
-
};
|
|
121
|
-
});
|
|
122
|
-
return {
|
|
123
|
-
semantics: 'qwen4_exp_ple',
|
|
124
|
-
unigramVocabSize,
|
|
125
|
-
ngramSize,
|
|
126
|
-
headsPerNgram,
|
|
127
|
-
headCount,
|
|
128
|
-
embedDim,
|
|
129
|
-
headDim: embedDim / headCount,
|
|
130
|
-
vocabSizeBase,
|
|
131
|
-
vocabPadMultiple,
|
|
132
|
-
seed,
|
|
133
|
-
boundaryTokenId: resolveBoundaryToken(text),
|
|
134
|
-
convKernelSize,
|
|
135
|
-
convDilation: ngramSize,
|
|
136
|
-
convStateLength: (convKernelSize - 1) * ngramSize,
|
|
137
|
-
normEps: sourceField(text, 'rms_norm_eps'),
|
|
138
|
-
tableDtype,
|
|
139
|
-
layers,
|
|
140
|
-
};
|
|
141
|
-
}
|
|
142
|
-
|
|
143
|
-
/**
|
|
144
|
-
* @param {Record<string, any>} text @param {string[]} layerTypes
|
|
145
|
-
* @returns {import('../config/schema/hyper-connection-decoder.schema.js').IndexedAttentionSchema | null}
|
|
146
|
-
*/
|
|
147
|
-
function lowerIndexedAttention(text, layerTypes) {
|
|
148
|
-
const values = Object.fromEntries(QSA_FIELDS.map((field) => [field, sourceField(text, field)]));
|
|
149
|
-
if (QSA_FIELDS.every((field) => values[field] === null)) return null;
|
|
150
|
-
const missing = QSA_FIELDS.filter((field) => values[field] === null);
|
|
151
|
-
if (missing.length > 0) throw new Error(`qwen4_exp QSA config is missing required fields: ${missing.join(', ')}.`);
|
|
152
|
-
QSA_FIELDS.forEach((field) => {
|
|
153
|
-
if (requireInteger(values[field], field) <= 0) throw new Error(`qwen4_exp ${field} must be positive.`);
|
|
154
|
-
});
|
|
155
|
-
if (values.indexer_kv_heads !== 1) throw new Error('qwen4_exp QSA requires indexer_kv_heads=1.');
|
|
156
|
-
if (values.indexer_budget % values.indexer_compress_ratio !== 0) {
|
|
157
|
-
throw new Error('qwen4_exp indexer_budget must be divisible by indexer_compress_ratio.');
|
|
158
|
-
}
|
|
159
|
-
const rotaryDim = resolveRotaryDim(text);
|
|
160
|
-
if (rotaryDim > values.indexer_head_dim) {
|
|
161
|
-
throw new Error(`qwen4_exp attention rotary_dim ${rotaryDim} must fit indexer_head_dim ${values.indexer_head_dim}.`);
|
|
162
|
-
}
|
|
163
|
-
return {
|
|
164
|
-
semantics: 'qwen4_exp_qsa',
|
|
165
|
-
layers: layerTypes.flatMap((type, index) => (type === 'full_attention' || type === 'indexed_attention' ? [index] : [])),
|
|
166
|
-
queryHeads: values.indexer_n_heads,
|
|
167
|
-
keyHeads: 1,
|
|
168
|
-
headDim: values.indexer_head_dim,
|
|
169
|
-
rotaryDim,
|
|
170
|
-
tokenBudget: values.indexer_budget,
|
|
171
|
-
compressRatio: values.indexer_compress_ratio,
|
|
172
|
-
blockTopK: values.indexer_budget / values.indexer_compress_ratio,
|
|
173
|
-
denseVisibleLimit: values.indexer_budget + values.indexer_compress_ratio - 1,
|
|
174
|
-
normEps: sourceField(text, 'rms_norm_eps'),
|
|
175
|
-
tieOrder: 'lower_block_index',
|
|
176
|
-
forceSelfToken: false,
|
|
177
|
-
};
|
|
178
|
-
}
|
|
179
|
-
|
|
180
|
-
/**
|
|
181
|
-
* Source-derived contracts for a qwen4_exp config, or null for other families.
|
|
182
|
-
* Layer types use the checkpoint spelling ("full_attention" layers are indexed).
|
|
183
|
-
* @type {typeof import('./qwen4-exp-source.js').resolveQwen4ExpSourceContracts}
|
|
184
|
-
*/
|
|
185
|
-
export function resolveQwen4ExpSourceContracts(rawConfig) {
|
|
186
|
-
if (!isQwen4ExpSource(rawConfig)) return null;
|
|
187
|
-
const text = /** @type {Record<string, any>} */ (textConfigOf(rawConfig));
|
|
188
|
-
const layerTypes = sourceLayerTypes(text);
|
|
189
|
-
const unsupported = [...new Set(layerTypes)].filter((type) => !['linear_attention', 'full_attention', 'indexed_attention'].includes(type));
|
|
190
|
-
if (unsupported.length > 0) throw new Error(`Unsupported qwen4_exp layer types: ${unsupported.join(', ')}.`);
|
|
191
|
-
const streamCount = requireInteger(sourceField(text, 'hc_count'), 'hc_count');
|
|
192
|
-
if (streamCount <= 1) throw new Error(`qwen4_exp requires hc_count > 1; got ${streamCount}.`);
|
|
193
|
-
return {
|
|
194
|
-
hyperConnections: {
|
|
195
|
-
semantics: 'qwen4_exp_gated_residual',
|
|
196
|
-
streamCount,
|
|
197
|
-
streamWidth: requireInteger(sourceField(text, 'hidden_size'), 'hidden_size'),
|
|
198
|
-
projectionRank: requireInteger(sourceField(text, 'hc_lowrank'), 'hc_lowrank'),
|
|
199
|
-
normEps: sourceField(text, 'rms_norm_eps'),
|
|
200
|
-
},
|
|
201
|
-
indexedAttention: lowerIndexedAttention(text, layerTypes),
|
|
202
|
-
ngramEmbedding: lowerNgramEmbedding(text, layerTypes),
|
|
203
|
-
};
|
|
204
|
-
}
|
|
205
5
|
|
|
206
6
|
/** @param {unknown} value @returns {string} */
|
|
207
7
|
function canonicalJson(value) {
|
|
@@ -2,14 +2,7 @@ import { ggufBlockBytes, isGgufBlockSourceDtype } from '../formats/gguf/block-st
|
|
|
2
2
|
import { validateQ40Storage } from '../formats/rdrr/q4_0-storage.js';
|
|
3
3
|
import { fp8SourceEncoding, packFP8Source, materializeFP8Source } from './fp8-source.js';
|
|
4
4
|
import { quantizeToW8A8Rows } from './w8a8-quantizer.js';
|
|
5
|
-
import {
|
|
6
|
-
// Constants
|
|
7
|
-
SHARD_SIZE as SCHEMA_SHARD_SIZE,
|
|
8
|
-
RDRR_VERSION as SCHEMA_RDRR_VERSION,
|
|
9
|
-
ConversionStage as SchemaConversionStage,
|
|
10
|
-
DEFAULT_MANIFEST_INFERENCE,
|
|
11
|
-
formatBytes,
|
|
12
|
-
} from '../config/schema/index.js';
|
|
5
|
+
import { SHARD_SIZE as SCHEMA_SHARD_SIZE, RDRR_VERSION as SCHEMA_RDRR_VERSION, ConversionStage as SchemaConversionStage, DEFAULT_MANIFEST_INFERENCE, formatBytes, createConverterConfig } from '../config/schema/index.js';
|
|
13
6
|
import {
|
|
14
7
|
classifyTensorRole,
|
|
15
8
|
generateShardFilename,
|
|
@@ -21,7 +14,7 @@ import {
|
|
|
21
14
|
getInferenceLayerPatternContractArtifact,
|
|
22
15
|
selectRuleValue,
|
|
23
16
|
} from '../rules/rule-registry.js';
|
|
24
|
-
|
|
17
|
+
|
|
25
18
|
import { buildExecutionContractArtifact } from '../config/execution-contract-check.js';
|
|
26
19
|
import { buildManifestRequiredInferenceFieldsArtifact } from '../config/required-inference-fields-contract-check.js';
|
|
27
20
|
import { resolveEosTokenId } from './tokenizer-special-tokens.js';
|
|
@@ -1,11 +1,4 @@
|
|
|
1
|
-
import {
|
|
2
|
-
// Constants
|
|
3
|
-
SHARD_SIZE as SCHEMA_SHARD_SIZE,
|
|
4
|
-
RDRR_VERSION as SCHEMA_RDRR_VERSION,
|
|
5
|
-
ConversionStage as SchemaConversionStage,
|
|
6
|
-
DEFAULT_MANIFEST_INFERENCE,
|
|
7
|
-
formatBytes,
|
|
8
|
-
} from '../config/schema/index.js';
|
|
1
|
+
import { SHARD_SIZE as SCHEMA_SHARD_SIZE, RDRR_VERSION as SCHEMA_RDRR_VERSION, ConversionStage as SchemaConversionStage, DEFAULT_MANIFEST_INFERENCE, formatBytes, createConverterConfig } from '../config/schema/index.js';
|
|
9
2
|
import {
|
|
10
3
|
classifyTensorRole,
|
|
11
4
|
generateShardFilename,
|
|
@@ -17,7 +10,7 @@ import {
|
|
|
17
10
|
getInferenceLayerPatternContractArtifact,
|
|
18
11
|
selectRuleValue,
|
|
19
12
|
} from '../rules/rule-registry.js';
|
|
20
|
-
|
|
13
|
+
|
|
21
14
|
import { buildExecutionContractArtifact } from '../config/execution-contract-check.js';
|
|
22
15
|
import { buildManifestRequiredInferenceFieldsArtifact } from '../config/required-inference-fields-contract-check.js';
|
|
23
16
|
import { resolveEosTokenId } from './tokenizer-special-tokens.js';
|
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import { validateW8A8Policy } from
|
|
1
|
+
import { validateW8A8Policy } from "../config/schema/manifest/w8a8.schema.js";
|
|
2
2
|
import { createW8A8StorageDescriptor } from '../formats/rdrr/w8a8-storage.js';
|
|
3
3
|
|
|
4
4
|
// All operands are F32 dyadics. Multiplication by a half-integer <=126.5
|
|
@@ -21,7 +21,7 @@ export function roundW8A8Code(value, scale) {
|
|
|
21
21
|
return value < 0 ? -lower : lower;
|
|
22
22
|
}
|
|
23
23
|
|
|
24
|
-
/** @param {Float32Array} values @param {number[]} shape @param {import(
|
|
24
|
+
/** @param {Float32Array} values @param {number[]} shape @param {import("../config/schema/manifest/w8a8.schema.js").W8A8Policy} policy */
|
|
25
25
|
export function quantizeToW8A8Rows(values, shape, policy) {
|
|
26
26
|
validateW8A8Policy(policy);
|
|
27
27
|
if (!(values instanceof Float32Array) || !Array.isArray(shape) || shape.length !== 2
|
package/src/debug/config.js
CHANGED
package/src/debug/perf-guards.js
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
import { log } from './log.js';
|
|
2
2
|
import { trace } from './trace.js';
|
|
3
|
-
import { DEFAULT_PERF_GUARDS_CONFIG } from
|
|
3
|
+
import { DEFAULT_PERF_GUARDS_CONFIG } from "../config/schema/runtime/debug.schema.js";
|
|
4
4
|
|
|
5
5
|
let config = { ...DEFAULT_PERF_GUARDS_CONFIG };
|
|
6
6
|
|