@simulatte/doppler 0.1.0
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/BRANDING.md +14 -0
- package/LICENSE +201 -0
- package/NOTICE +5 -0
- package/README.md +85 -0
- package/SECURITY.md +19 -0
- package/package.json +144 -0
- package/src/adapters/adapter-manager.d.ts +200 -0
- package/src/adapters/adapter-manager.js +509 -0
- package/src/adapters/adapter-manifest.d.ts +290 -0
- package/src/adapters/adapter-manifest.js +320 -0
- package/src/adapters/adapter-registry.d.ts +192 -0
- package/src/adapters/adapter-registry.js +466 -0
- package/src/adapters/index.d.ts +89 -0
- package/src/adapters/index.js +42 -0
- package/src/adapters/lora-loader.d.ts +105 -0
- package/src/adapters/lora-loader.js +397 -0
- package/src/bootstrap.d.ts +1 -0
- package/src/bootstrap.js +30 -0
- package/src/bridge/extension/background.d.ts +14 -0
- package/src/bridge/extension/background.js +168 -0
- package/src/bridge/extension/manifest.json +34 -0
- package/src/bridge/extension-client.d.ts +109 -0
- package/src/bridge/extension-client.js +369 -0
- package/src/bridge/index.d.ts +68 -0
- package/src/bridge/index.js +51 -0
- package/src/bridge/protocol.d.ts +96 -0
- package/src/bridge/protocol.js +130 -0
- package/src/browser/browser-converter.d.ts +71 -0
- package/src/browser/browser-converter.js +947 -0
- package/src/browser/file-picker.d.ts +63 -0
- package/src/browser/file-picker.js +275 -0
- package/src/browser/gguf-importer.d.ts +136 -0
- package/src/browser/gguf-importer.js +532 -0
- package/src/browser/gguf-parser-browser.d.ts +14 -0
- package/src/browser/gguf-parser-browser.js +17 -0
- package/src/browser/quantization.d.ts +69 -0
- package/src/browser/quantization.js +328 -0
- package/src/browser/safetensors-parser-browser.d.ts +193 -0
- package/src/browser/safetensors-parser-browser.js +264 -0
- package/src/browser/shard-io-browser.d.ts +57 -0
- package/src/browser/shard-io-browser.js +89 -0
- package/src/browser/tensor-source-download.d.ts +27 -0
- package/src/browser/tensor-source-download.js +239 -0
- package/src/browser/tensor-source-file.d.ts +26 -0
- package/src/browser/tensor-source-file.js +53 -0
- package/src/browser/tensor-source-http.d.ts +28 -0
- package/src/browser/tensor-source-http.js +126 -0
- package/src/client/doppler-provider/generation.d.ts +25 -0
- package/src/client/doppler-provider/generation.js +114 -0
- package/src/client/doppler-provider/index.d.ts +2 -0
- package/src/client/doppler-provider/index.js +3 -0
- package/src/client/doppler-provider/model-manager.d.ts +61 -0
- package/src/client/doppler-provider/model-manager.js +667 -0
- package/src/client/doppler-provider/provider.d.ts +5 -0
- package/src/client/doppler-provider/provider.js +102 -0
- package/src/client/doppler-provider/source-runtime.d.ts +22 -0
- package/src/client/doppler-provider/source-runtime.js +522 -0
- package/src/client/doppler-provider/types.d.ts +127 -0
- package/src/client/doppler-provider/types.js +17 -0
- package/src/client/doppler-provider.d.ts +46 -0
- package/src/client/doppler-provider.js +36 -0
- package/src/config/README.md +69 -0
- package/src/config/backward-registry-loader.d.ts +3 -0
- package/src/config/backward-registry-loader.js +8 -0
- package/src/config/index.d.ts +63 -0
- package/src/config/index.js +31 -0
- package/src/config/kernel-path-loader.d.ts +149 -0
- package/src/config/kernel-path-loader.js +534 -0
- package/src/config/kernels/backward-registry.json +99 -0
- package/src/config/kernels/kernel-ref-digests.d.ts +1 -0
- package/src/config/kernels/kernel-ref-digests.js +214 -0
- package/src/config/kernels/kernel-ref.d.ts +17 -0
- package/src/config/kernels/kernel-ref.js +75 -0
- package/src/config/kernels/moe/gpt-oss.paths.json +49 -0
- package/src/config/kernels/registry.d.ts +86 -0
- package/src/config/kernels/registry.js +103 -0
- package/src/config/kernels/registry.json +6771 -0
- package/src/config/loader.d.ts +57 -0
- package/src/config/loader.js +513 -0
- package/src/config/merge.d.ts +142 -0
- package/src/config/merge.js +389 -0
- package/src/config/param-categories.d.ts +17 -0
- package/src/config/param-categories.js +72 -0
- package/src/config/param-validator.d.ts +26 -0
- package/src/config/param-validator.js +235 -0
- package/src/config/platforms/amd-rdna3.json +16 -0
- package/src/config/platforms/apple-m1.json +16 -0
- package/src/config/platforms/apple-m2.json +16 -0
- package/src/config/platforms/apple-m3.json +16 -0
- package/src/config/platforms/generic.json +14 -0
- package/src/config/platforms/loader.d.ts +65 -0
- package/src/config/platforms/loader.js +153 -0
- package/src/config/platforms/nvidia-rtx30.json +16 -0
- package/src/config/platforms/nvidia-rtx40.json +16 -0
- package/src/config/presets/kernel-paths/embeddinggemma-f16-f32a.json +60 -0
- package/src/config/presets/kernel-paths/embeddinggemma-f32-f32a.json +60 -0
- package/src/config/presets/kernel-paths/embeddinggemma-q4k-dequant-f32a.json +60 -0
- package/src/config/presets/kernel-paths/gemma2-f16-f16a.json +61 -0
- package/src/config/presets/kernel-paths/gemma2-f16-f32a.json +60 -0
- package/src/config/presets/kernel-paths/gemma2-q4k-dequant-f16a.json +61 -0
- package/src/config/presets/kernel-paths/gemma2-q4k-dequant-f32a.json +60 -0
- package/src/config/presets/kernel-paths/gemma2-q4k-fused-f32a.json +57 -0
- package/src/config/presets/kernel-paths/gemma3-f16-fused-f16a-online.json +200 -0
- package/src/config/presets/kernel-paths/gemma3-f16-fused-f32a-online.json +223 -0
- package/src/config/presets/kernel-paths/gemma3-q4k-dequant-f16a-online.json +60 -0
- package/src/config/presets/kernel-paths/gemma3-q4k-dequant-f32a-online.json +61 -0
- package/src/config/presets/kernel-paths/gemma3-q4k-dequant-f32a.json +61 -0
- package/src/config/presets/kernel-paths/lfm2-q4k-dequant-f32a-online.json +61 -0
- package/src/config/presets/kernel-paths/registry.json +103 -0
- package/src/config/presets/models/deepseek.json +20 -0
- package/src/config/presets/models/diffusion.json +10 -0
- package/src/config/presets/models/embeddinggemma.json +74 -0
- package/src/config/presets/models/functiongemma.json +31 -0
- package/src/config/presets/models/gemma2.json +59 -0
- package/src/config/presets/models/gemma3.json +75 -0
- package/src/config/presets/models/gpt-oss.json +68 -0
- package/src/config/presets/models/kimi-k2.json +25 -0
- package/src/config/presets/models/lfm2.json +83 -0
- package/src/config/presets/models/llama3.json +40 -0
- package/src/config/presets/models/mamba.json +34 -0
- package/src/config/presets/models/mixtral.json +37 -0
- package/src/config/presets/models/modernbert.json +32 -0
- package/src/config/presets/models/qwen3.json +41 -0
- package/src/config/presets/models/transformer.json +73 -0
- package/src/config/presets/models/translategemma.json +30 -0
- package/src/config/presets/platforms/nvidia-gb200-8gpu.json +45 -0
- package/src/config/presets/platforms/nvidia-gb200-nvl72.json +45 -0
- package/src/config/presets/platforms/nvidia-gh200-nvl2.json +44 -0
- package/src/config/presets/platforms/nvidia-gh200.json +44 -0
- package/src/config/presets/runtime/compute/f16-activations.json +30 -0
- package/src/config/presets/runtime/compute/f16-batched.json +32 -0
- package/src/config/presets/runtime/default.json +101 -0
- package/src/config/presets/runtime/diagnostics/debug-logits.json +53 -0
- package/src/config/presets/runtime/experiments/bench/gemma3-bench-q4k.json +53 -0
- package/src/config/presets/runtime/experiments/debug/gemma3-debug-q4k.json +210 -0
- package/src/config/presets/runtime/experiments/verify/gemma3-verify.json +39 -0
- package/src/config/presets/runtime/kernels/dequant-f16-q4k.json +20 -0
- package/src/config/presets/runtime/kernels/dequant-f32-q4k.json +20 -0
- package/src/config/presets/runtime/kernels/fused-q4k.json +20 -0
- package/src/config/presets/runtime/kernels/safe-q4k.json +20 -0
- package/src/config/presets/runtime/model/gemma2-debug.json +77 -0
- package/src/config/presets/runtime/model/gemma2-pipeline-debug.json +66 -0
- package/src/config/presets/runtime/model/gemma2-pipeline.json +75 -0
- package/src/config/presets/runtime/model/gemma3-layer-probe.json +85 -0
- package/src/config/presets/runtime/modes/bench.json +37 -0
- package/src/config/presets/runtime/modes/debug.json +39 -0
- package/src/config/presets/runtime/modes/default.json +10 -0
- package/src/config/presets/runtime/modes/embedding-bench.json +28 -0
- package/src/config/presets/runtime/modes/embedding.json +54 -0
- package/src/config/presets/runtime/modes/low-memory.json +40 -0
- package/src/config/presets/runtime/modes/production.json +48 -0
- package/src/config/presets/runtime/modes/simulation.json +30 -0
- package/src/config/presets/runtime/modes/trace-layers.json +126 -0
- package/src/config/presets/runtime/platform/metal-apple-q4k.json +11 -0
- package/src/config/runtime-merge.d.ts +5 -0
- package/src/config/runtime-merge.js +21 -0
- package/src/config/runtime.d.ts +28 -0
- package/src/config/runtime.js +56 -0
- package/src/config/schema/adapter.schema.d.ts +53 -0
- package/src/config/schema/adapter.schema.js +60 -0
- package/src/config/schema/backward-registry.schema.d.ts +14 -0
- package/src/config/schema/backward-registry.schema.js +46 -0
- package/src/config/schema/benchmark.schema.d.ts +54 -0
- package/src/config/schema/benchmark.schema.js +74 -0
- package/src/config/schema/bridge.schema.d.ts +25 -0
- package/src/config/schema/bridge.schema.js +22 -0
- package/src/config/schema/buffer-pool.schema.d.ts +92 -0
- package/src/config/schema/buffer-pool.schema.js +50 -0
- package/src/config/schema/conversion.schema.d.ts +183 -0
- package/src/config/schema/conversion.schema.js +13 -0
- package/src/config/schema/converter.schema.d.ts +123 -0
- package/src/config/schema/converter.schema.js +136 -0
- package/src/config/schema/debug.schema.d.ts +245 -0
- package/src/config/schema/debug.schema.js +106 -0
- package/src/config/schema/diffusion.schema.d.ts +88 -0
- package/src/config/schema/diffusion.schema.js +62 -0
- package/src/config/schema/distill-training.schema.d.ts +48 -0
- package/src/config/schema/distill-training.schema.js +139 -0
- package/src/config/schema/distribution.schema.d.ts +155 -0
- package/src/config/schema/distribution.schema.js +81 -0
- package/src/config/schema/doppler.schema.d.ts +75 -0
- package/src/config/schema/doppler.schema.js +352 -0
- package/src/config/schema/ecosystem.schema.d.ts +255 -0
- package/src/config/schema/ecosystem.schema.js +534 -0
- package/src/config/schema/emulation.schema.d.ts +351 -0
- package/src/config/schema/emulation.schema.js +299 -0
- package/src/config/schema/energy.schema.d.ts +102 -0
- package/src/config/schema/energy.schema.js +72 -0
- package/src/config/schema/execution-v0.schema.d.ts +187 -0
- package/src/config/schema/execution-v0.schema.js +55 -0
- package/src/config/schema/gpu-cache.schema.d.ts +26 -0
- package/src/config/schema/gpu-cache.schema.js +8 -0
- package/src/config/schema/harness.schema.d.ts +32 -0
- package/src/config/schema/harness.schema.js +20 -0
- package/src/config/schema/hotswap.schema.d.ts +55 -0
- package/src/config/schema/hotswap.schema.js +18 -0
- package/src/config/schema/index.d.ts +863 -0
- package/src/config/schema/index.js +471 -0
- package/src/config/schema/inference-defaults.schema.d.ts +276 -0
- package/src/config/schema/inference-defaults.schema.js +185 -0
- package/src/config/schema/inference.schema.d.ts +289 -0
- package/src/config/schema/inference.schema.js +39 -0
- package/src/config/schema/intent-bundle.schema.d.ts +28 -0
- package/src/config/schema/intent-bundle.schema.js +12 -0
- package/src/config/schema/kernel-path.schema.d.ts +173 -0
- package/src/config/schema/kernel-path.schema.js +9 -0
- package/src/config/schema/kernel-registry.schema.d.ts +199 -0
- package/src/config/schema/kernel-registry.schema.js +46 -0
- package/src/config/schema/kernel-thresholds.schema.d.ts +302 -0
- package/src/config/schema/kernel-thresholds.schema.js +187 -0
- package/src/config/schema/kernel-warmup.schema.d.ts +19 -0
- package/src/config/schema/kernel-warmup.schema.js +5 -0
- package/src/config/schema/kvcache.schema.d.ts +131 -0
- package/src/config/schema/kvcache.schema.js +31 -0
- package/src/config/schema/loading.schema.d.ts +153 -0
- package/src/config/schema/loading.schema.js +84 -0
- package/src/config/schema/lora.schema.d.ts +12 -0
- package/src/config/schema/lora.schema.js +12 -0
- package/src/config/schema/manifest.schema.d.ts +500 -0
- package/src/config/schema/manifest.schema.js +130 -0
- package/src/config/schema/memory-limits.schema.d.ts +107 -0
- package/src/config/schema/memory-limits.schema.js +57 -0
- package/src/config/schema/moe.schema.d.ts +78 -0
- package/src/config/schema/moe.schema.js +31 -0
- package/src/config/schema/platform.schema.d.ts +121 -0
- package/src/config/schema/platform.schema.js +1 -0
- package/src/config/schema/preset.schema.d.ts +124 -0
- package/src/config/schema/preset.schema.js +1 -0
- package/src/config/schema/quantization-defaults.schema.d.ts +34 -0
- package/src/config/schema/quantization-defaults.schema.js +5 -0
- package/src/config/schema/quantization.schema.d.ts +10 -0
- package/src/config/schema/quantization.schema.js +33 -0
- package/src/config/schema/shared-runtime.schema.d.ts +75 -0
- package/src/config/schema/shared-runtime.schema.js +45 -0
- package/src/config/schema/speculative.schema.d.ts +21 -0
- package/src/config/schema/speculative.schema.js +11 -0
- package/src/config/schema/storage.schema.d.ts +123 -0
- package/src/config/schema/storage.schema.js +66 -0
- package/src/config/schema/tooling.schema.d.ts +29 -0
- package/src/config/schema/tooling.schema.js +12 -0
- package/src/config/schema/training-metrics.schema.d.ts +89 -0
- package/src/config/schema/training-metrics.schema.js +374 -0
- package/src/config/schema/training.schema.d.ts +88 -0
- package/src/config/schema/training.schema.js +106 -0
- package/src/config/schema/tuner.schema.d.ts +39 -0
- package/src/config/schema/tuner.schema.js +13 -0
- package/src/config/schema/ul-training.schema.d.ts +61 -0
- package/src/config/schema/ul-training.schema.js +140 -0
- package/src/config/schema/units.schema.d.ts +27 -0
- package/src/config/schema/units.schema.js +26 -0
- package/src/config/training-defaults.d.ts +24 -0
- package/src/config/training-defaults.js +91 -0
- package/src/converter/conversion-plan.d.ts +64 -0
- package/src/converter/conversion-plan.js +472 -0
- package/src/converter/core.d.ts +247 -0
- package/src/converter/core.js +1329 -0
- package/src/converter/execution-v0-manifest.d.ts +15 -0
- package/src/converter/execution-v0-manifest.js +146 -0
- package/src/converter/index.d.ts +98 -0
- package/src/converter/index.js +59 -0
- package/src/converter/manifest-inference.d.ts +20 -0
- package/src/converter/manifest-inference.js +492 -0
- package/src/converter/parsers/diffusion.d.ts +50 -0
- package/src/converter/parsers/diffusion.js +270 -0
- package/src/converter/parsers/gguf.d.ts +22 -0
- package/src/converter/parsers/gguf.js +46 -0
- package/src/converter/parsers/index.d.ts +21 -0
- package/src/converter/parsers/index.js +12 -0
- package/src/converter/parsers/transformer.d.ts +16 -0
- package/src/converter/parsers/transformer.js +25 -0
- package/src/converter/quantization-info.d.ts +37 -0
- package/src/converter/quantization-info.js +398 -0
- package/src/converter/quantizer.d.ts +96 -0
- package/src/converter/quantizer.js +422 -0
- package/src/converter/rope-config.d.ts +15 -0
- package/src/converter/rope-config.js +218 -0
- package/src/converter/shard-packer.d.ts +138 -0
- package/src/converter/shard-packer.js +422 -0
- package/src/converter/tokenizer-utils.d.ts +11 -0
- package/src/converter/tokenizer-utils.js +87 -0
- package/src/debug/config.d.ts +78 -0
- package/src/debug/config.js +235 -0
- package/src/debug/history.d.ts +65 -0
- package/src/debug/history.js +71 -0
- package/src/debug/index.d.ts +268 -0
- package/src/debug/index.js +192 -0
- package/src/debug/log.d.ts +46 -0
- package/src/debug/log.js +132 -0
- package/src/debug/perf.d.ts +33 -0
- package/src/debug/perf.js +51 -0
- package/src/debug/reference/README.md +114 -0
- package/src/debug/reference/hf_attn_debug.py +114 -0
- package/src/debug/reference/hf_embed_check.py +89 -0
- package/src/debug/reference/hf_layer_out.py +100 -0
- package/src/debug/reference/hf_rope_check.py +116 -0
- package/src/debug/reference/hf_weights.py +75 -0
- package/src/debug/signals.d.ts +63 -0
- package/src/debug/signals.js +33 -0
- package/src/debug/stats.d.ts +47 -0
- package/src/debug/stats.js +160 -0
- package/src/debug/tensor.d.ts +123 -0
- package/src/debug/tensor.js +257 -0
- package/src/debug/trace.d.ts +17 -0
- package/src/debug/trace.js +167 -0
- package/src/diffusion/image-regression.d.ts +31 -0
- package/src/diffusion/image-regression.js +107 -0
- package/src/diffusion/index.d.ts +8 -0
- package/src/diffusion/index.js +8 -0
- package/src/distribution/p2p-control-plane.d.ts +52 -0
- package/src/distribution/p2p-control-plane.js +232 -0
- package/src/distribution/p2p-observability.d.ts +116 -0
- package/src/distribution/p2p-observability.js +267 -0
- package/src/distribution/p2p-transport-contract.d.ts +57 -0
- package/src/distribution/p2p-transport-contract.js +310 -0
- package/src/distribution/p2p-webrtc-browser.d.ts +37 -0
- package/src/distribution/p2p-webrtc-browser.js +434 -0
- package/src/distribution/shard-delivery.d.ts +251 -0
- package/src/distribution/shard-delivery.js +2096 -0
- package/src/energy/index.d.ts +2 -0
- package/src/energy/index.js +2 -0
- package/src/errors/doppler-error.d.ts +21 -0
- package/src/errors/doppler-error.js +25 -0
- package/src/errors/index.d.ts +1 -0
- package/src/errors/index.js +1 -0
- package/src/formats/gguf/index.d.ts +8 -0
- package/src/formats/gguf/index.js +4 -0
- package/src/formats/gguf/types.d.ts +137 -0
- package/src/formats/gguf/types.js +443 -0
- package/src/formats/index.d.ts +51 -0
- package/src/formats/index.js +13 -0
- package/src/formats/rdrr/classification.d.ts +39 -0
- package/src/formats/rdrr/classification.js +275 -0
- package/src/formats/rdrr/groups.d.ts +27 -0
- package/src/formats/rdrr/groups.js +76 -0
- package/src/formats/rdrr/index.d.ts +25 -0
- package/src/formats/rdrr/index.js +19 -0
- package/src/formats/rdrr/manifest.d.ts +32 -0
- package/src/formats/rdrr/manifest.js +108 -0
- package/src/formats/rdrr/parsing.d.ts +23 -0
- package/src/formats/rdrr/parsing.js +101 -0
- package/src/formats/rdrr/tensor-config-validator.d.ts +42 -0
- package/src/formats/rdrr/tensor-config-validator.js +156 -0
- package/src/formats/rdrr/types.d.ts +200 -0
- package/src/formats/rdrr/types.js +16 -0
- package/src/formats/rdrr/validation.d.ts +9 -0
- package/src/formats/rdrr/validation.js +200 -0
- package/src/formats/safetensors/index.d.ts +8 -0
- package/src/formats/safetensors/index.js +4 -0
- package/src/formats/safetensors/types.d.ts +67 -0
- package/src/formats/safetensors/types.js +102 -0
- package/src/formats/tokenizer/index.d.ts +5 -0
- package/src/formats/tokenizer/index.js +3 -0
- package/src/formats/tokenizer/types.d.ts +9 -0
- package/src/formats/tokenizer/types.js +22 -0
- package/src/generation/index.d.ts +18 -0
- package/src/generation/index.js +12 -0
- package/src/gpu/command-recorder.d.ts +175 -0
- package/src/gpu/command-recorder.js +473 -0
- package/src/gpu/device.d.ts +141 -0
- package/src/gpu/device.js +350 -0
- package/src/gpu/kernel-runtime.d.ts +20 -0
- package/src/gpu/kernel-runtime.js +37 -0
- package/src/gpu/kernel-selection-cache.d.ts +13 -0
- package/src/gpu/kernel-selection-cache.js +13 -0
- package/src/gpu/kernel-selection-log.d.ts +12 -0
- package/src/gpu/kernel-selection-log.js +28 -0
- package/src/gpu/kernel-selector.d.ts +11 -0
- package/src/gpu/kernel-selector.js +10 -0
- package/src/gpu/kernel-tuner/benchmarks.d.ts +144 -0
- package/src/gpu/kernel-tuner/benchmarks.js +892 -0
- package/src/gpu/kernel-tuner/cache.d.ts +55 -0
- package/src/gpu/kernel-tuner/cache.js +66 -0
- package/src/gpu/kernel-tuner/index.d.ts +59 -0
- package/src/gpu/kernel-tuner/index.js +38 -0
- package/src/gpu/kernel-tuner/tuner.d.ts +82 -0
- package/src/gpu/kernel-tuner/tuner.js +229 -0
- package/src/gpu/kernel-tuner/types.d.ts +101 -0
- package/src/gpu/kernel-tuner/types.js +4 -0
- package/src/gpu/kernel-tuner.d.ts +33 -0
- package/src/gpu/kernel-tuner.js +12 -0
- package/src/gpu/kernels/README.md +127 -0
- package/src/gpu/kernels/attention.d.ts +236 -0
- package/src/gpu/kernels/attention.js +1359 -0
- package/src/gpu/kernels/attention.wgsl +249 -0
- package/src/gpu/kernels/attention_bdpa_decode_f16.wgsl +246 -0
- package/src/gpu/kernels/attention_decode.wgsl +233 -0
- package/src/gpu/kernels/attention_decode_chunked_f16.wgsl +183 -0
- package/src/gpu/kernels/attention_decode_chunked_f16kv.wgsl +208 -0
- package/src/gpu/kernels/attention_decode_f16.wgsl +202 -0
- package/src/gpu/kernels/attention_decode_f16kv.wgsl +224 -0
- package/src/gpu/kernels/attention_decode_online_f16.wgsl +223 -0
- package/src/gpu/kernels/attention_decode_online_f16kv.wgsl +225 -0
- package/src/gpu/kernels/attention_decode_optimized.wgsl +445 -0
- package/src/gpu/kernels/attention_decode_paged_f16.wgsl +172 -0
- package/src/gpu/kernels/attention_decode_paged_f16kv.wgsl +174 -0
- package/src/gpu/kernels/attention_decode_subgroup.wgsl +233 -0
- package/src/gpu/kernels/attention_decode_tiered_f16.wgsl +218 -0
- package/src/gpu/kernels/attention_decode_tiered_f16kv.wgsl +220 -0
- package/src/gpu/kernels/attention_decode_tiered_int4_f16kv.wgsl +242 -0
- package/src/gpu/kernels/attention_decode_tiered_int8_f16kv.wgsl +242 -0
- package/src/gpu/kernels/attention_f16.wgsl +214 -0
- package/src/gpu/kernels/attention_f16kv.wgsl +242 -0
- package/src/gpu/kernels/attention_small.wgsl +260 -0
- package/src/gpu/kernels/attention_small_f16.wgsl +240 -0
- package/src/gpu/kernels/attention_small_f16kv.wgsl +266 -0
- package/src/gpu/kernels/attention_streaming.wgsl +149 -0
- package/src/gpu/kernels/attention_streaming_f16.wgsl +147 -0
- package/src/gpu/kernels/attention_streaming_f16kv.wgsl +151 -0
- package/src/gpu/kernels/backward/adam.d.ts +28 -0
- package/src/gpu/kernels/backward/adam.js +199 -0
- package/src/gpu/kernels/backward/adam.wgsl +50 -0
- package/src/gpu/kernels/backward/attention_backward.d.ts +22 -0
- package/src/gpu/kernels/backward/attention_backward.js +276 -0
- package/src/gpu/kernels/backward/attention_backward.wgsl +49 -0
- package/src/gpu/kernels/backward/bias_add_backward.d.ts +17 -0
- package/src/gpu/kernels/backward/bias_add_backward.js +24 -0
- package/src/gpu/kernels/backward/bias_add_backward.wgsl +33 -0
- package/src/gpu/kernels/backward/conv2d_backward.d.ts +31 -0
- package/src/gpu/kernels/backward/conv2d_backward.js +135 -0
- package/src/gpu/kernels/backward/conv2d_backward_input.wgsl +83 -0
- package/src/gpu/kernels/backward/conv2d_backward_weight.wgsl +70 -0
- package/src/gpu/kernels/backward/cross_entropy_backward.d.ts +23 -0
- package/src/gpu/kernels/backward/cross_entropy_backward.js +29 -0
- package/src/gpu/kernels/backward/cross_entropy_backward.wgsl +39 -0
- package/src/gpu/kernels/backward/embed_backward.d.ts +29 -0
- package/src/gpu/kernels/backward/embed_backward.js +118 -0
- package/src/gpu/kernels/backward/embed_backward.wgsl +73 -0
- package/src/gpu/kernels/backward/gelu_backward.d.ts +16 -0
- package/src/gpu/kernels/backward/gelu_backward.js +39 -0
- package/src/gpu/kernels/backward/gelu_backward.wgsl +38 -0
- package/src/gpu/kernels/backward/groupnorm_backward.d.ts +24 -0
- package/src/gpu/kernels/backward/groupnorm_backward.js +29 -0
- package/src/gpu/kernels/backward/groupnorm_backward.wgsl +143 -0
- package/src/gpu/kernels/backward/index.d.ts +17 -0
- package/src/gpu/kernels/backward/index.js +23 -0
- package/src/gpu/kernels/backward/layernorm_backward.d.ts +22 -0
- package/src/gpu/kernels/backward/layernorm_backward.js +135 -0
- package/src/gpu/kernels/backward/layernorm_backward.wgsl +194 -0
- package/src/gpu/kernels/backward/matmul_backward.d.ts +32 -0
- package/src/gpu/kernels/backward/matmul_backward.js +124 -0
- package/src/gpu/kernels/backward/matmul_backward.wgsl +90 -0
- package/src/gpu/kernels/backward/matmul_transpose_a.wgsl +84 -0
- package/src/gpu/kernels/backward/pixel_shuffle_backward.d.ts +22 -0
- package/src/gpu/kernels/backward/pixel_shuffle_backward.js +30 -0
- package/src/gpu/kernels/backward/pixel_shuffle_backward.wgsl +54 -0
- package/src/gpu/kernels/backward/rmsnorm_backward.d.ts +24 -0
- package/src/gpu/kernels/backward/rmsnorm_backward.js +101 -0
- package/src/gpu/kernels/backward/rmsnorm_backward.wgsl +78 -0
- package/src/gpu/kernels/backward/rope_backward.d.ts +25 -0
- package/src/gpu/kernels/backward/rope_backward.js +109 -0
- package/src/gpu/kernels/backward/rope_backward.wgsl +59 -0
- package/src/gpu/kernels/backward/scale_backward.d.ts +16 -0
- package/src/gpu/kernels/backward/scale_backward.js +84 -0
- package/src/gpu/kernels/backward/scale_backward.wgsl +27 -0
- package/src/gpu/kernels/backward/silu_backward.d.ts +16 -0
- package/src/gpu/kernels/backward/silu_backward.js +39 -0
- package/src/gpu/kernels/backward/silu_backward.wgsl +31 -0
- package/src/gpu/kernels/backward/softmax_backward.d.ts +16 -0
- package/src/gpu/kernels/backward/softmax_backward.js +43 -0
- package/src/gpu/kernels/backward/softmax_backward.wgsl +44 -0
- package/src/gpu/kernels/backward/upsample2d_backward.d.ts +21 -0
- package/src/gpu/kernels/backward/upsample2d_backward.js +30 -0
- package/src/gpu/kernels/backward/upsample2d_backward.wgsl +59 -0
- package/src/gpu/kernels/backward/utils.d.ts +45 -0
- package/src/gpu/kernels/backward/utils.js +371 -0
- package/src/gpu/kernels/bf16_to_f16.wgsl +54 -0
- package/src/gpu/kernels/bf16_to_f32.wgsl +70 -0
- package/src/gpu/kernels/bias_add.wgsl +40 -0
- package/src/gpu/kernels/bias_add_f16.wgsl +44 -0
- package/src/gpu/kernels/cast.d.ts +67 -0
- package/src/gpu/kernels/cast.js +422 -0
- package/src/gpu/kernels/cast_f16_to_f32.wgsl +31 -0
- package/src/gpu/kernels/cast_f32_to_f16.wgsl +36 -0
- package/src/gpu/kernels/check-finiteness.d.ts +15 -0
- package/src/gpu/kernels/check-finiteness.js +149 -0
- package/src/gpu/kernels/check-stop.d.ts +31 -0
- package/src/gpu/kernels/check-stop.js +181 -0
- package/src/gpu/kernels/clamp.d.ts +22 -0
- package/src/gpu/kernels/clamp.js +42 -0
- package/src/gpu/kernels/clamp.wgsl +24 -0
- package/src/gpu/kernels/constants.d.ts +168 -0
- package/src/gpu/kernels/constants.js +129 -0
- package/src/gpu/kernels/conv2d.d.ts +34 -0
- package/src/gpu/kernels/conv2d.js +81 -0
- package/src/gpu/kernels/conv2d.wgsl +71 -0
- package/src/gpu/kernels/conv2d_f16.wgsl +73 -0
- package/src/gpu/kernels/cross_entropy_loss.d.ts +21 -0
- package/src/gpu/kernels/cross_entropy_loss.js +54 -0
- package/src/gpu/kernels/cross_entropy_loss.wgsl +39 -0
- package/src/gpu/kernels/dequant.d.ts +108 -0
- package/src/gpu/kernels/dequant.js +524 -0
- package/src/gpu/kernels/dequant_f16_out.wgsl +151 -0
- package/src/gpu/kernels/dequant_f16_out_vec4.wgsl +149 -0
- package/src/gpu/kernels/dequant_f16_rowwise.wgsl +139 -0
- package/src/gpu/kernels/dequant_f32_rowwise.wgsl +133 -0
- package/src/gpu/kernels/dequant_mxfp4.wgsl +120 -0
- package/src/gpu/kernels/dequant_mxfp4_expert.wgsl +129 -0
- package/src/gpu/kernels/dequant_mxfp4_expert_f16.wgsl +105 -0
- package/src/gpu/kernels/dequant_mxfp4_vec4.wgsl +116 -0
- package/src/gpu/kernels/dequant_q6k.wgsl +140 -0
- package/src/gpu/kernels/dequant_q8_0.wgsl +98 -0
- package/src/gpu/kernels/dequant_shared.wgsl +202 -0
- package/src/gpu/kernels/dequant_shared_vec4.wgsl +153 -0
- package/src/gpu/kernels/dequant_subgroup.wgsl +202 -0
- package/src/gpu/kernels/dispatch.d.ts +157 -0
- package/src/gpu/kernels/dispatch.js +235 -0
- package/src/gpu/kernels/energy.d.ts +131 -0
- package/src/gpu/kernels/energy.js +425 -0
- package/src/gpu/kernels/energy_eval.wgsl +26 -0
- package/src/gpu/kernels/energy_eval_f16.wgsl +30 -0
- package/src/gpu/kernels/energy_quintel_grad.wgsl +92 -0
- package/src/gpu/kernels/energy_quintel_grad_f16.wgsl +96 -0
- package/src/gpu/kernels/energy_quintel_reduce.wgsl +112 -0
- package/src/gpu/kernels/energy_quintel_reduce_f16.wgsl +116 -0
- package/src/gpu/kernels/energy_quintel_update.wgsl +92 -0
- package/src/gpu/kernels/energy_quintel_update_f16.wgsl +96 -0
- package/src/gpu/kernels/energy_update.wgsl +25 -0
- package/src/gpu/kernels/energy_update_f16.wgsl +30 -0
- package/src/gpu/kernels/feature-check.d.ts +42 -0
- package/src/gpu/kernels/feature-check.js +70 -0
- package/src/gpu/kernels/fused_ffn.d.ts +65 -0
- package/src/gpu/kernels/fused_ffn.js +318 -0
- package/src/gpu/kernels/fused_ffn.wgsl +420 -0
- package/src/gpu/kernels/fused_ffn_f16.wgsl +213 -0
- package/src/gpu/kernels/fused_ffn_q4k.wgsl +375 -0
- package/src/gpu/kernels/fused_matmul_q4.wgsl +404 -0
- package/src/gpu/kernels/fused_matmul_q4_batched.wgsl +194 -0
- package/src/gpu/kernels/fused_matmul_q4_batched_f16.wgsl +170 -0
- package/src/gpu/kernels/fused_matmul_q4_batched_f16a.wgsl +154 -0
- package/src/gpu/kernels/fused_matmul_q4_f16a.wgsl +219 -0
- package/src/gpu/kernels/fused_matmul_q4_multicol_f16.wgsl +216 -0
- package/src/gpu/kernels/fused_matmul_q4_multicol_f16a.wgsl +204 -0
- package/src/gpu/kernels/fused_matmul_residual.d.ts +46 -0
- package/src/gpu/kernels/fused_matmul_residual.js +152 -0
- package/src/gpu/kernels/fused_matmul_rmsnorm.d.ts +64 -0
- package/src/gpu/kernels/fused_matmul_rmsnorm.js +273 -0
- package/src/gpu/kernels/fused_matmul_rmsnorm.wgsl +324 -0
- package/src/gpu/kernels/fused_matmul_rmsnorm_f16.wgsl +303 -0
- package/src/gpu/kernels/fused_swiglu.wgsl +63 -0
- package/src/gpu/kernels/fused_swiglu_f16.wgsl +57 -0
- package/src/gpu/kernels/gather.d.ts +64 -0
- package/src/gpu/kernels/gather.js +119 -0
- package/src/gpu/kernels/gather.wgsl +61 -0
- package/src/gpu/kernels/gather_f16.wgsl +65 -0
- package/src/gpu/kernels/gather_f16_f16_out.wgsl +55 -0
- package/src/gpu/kernels/gather_f16_out.wgsl +55 -0
- package/src/gpu/kernels/gather_f16_vec4.wgsl +76 -0
- package/src/gpu/kernels/gather_f16_vec4_f16_out.wgsl +68 -0
- package/src/gpu/kernels/gather_vec4.wgsl +74 -0
- package/src/gpu/kernels/gather_vec4_f16_out.wgsl +68 -0
- package/src/gpu/kernels/gelu.d.ts +33 -0
- package/src/gpu/kernels/gelu.js +47 -0
- package/src/gpu/kernels/gelu.wgsl +64 -0
- package/src/gpu/kernels/gelu_f16.wgsl +66 -0
- package/src/gpu/kernels/gptoss_mxfp4_expert_fused.wgsl +127 -0
- package/src/gpu/kernels/gptoss_router_topk.wgsl +119 -0
- package/src/gpu/kernels/groupnorm.d.ts +31 -0
- package/src/gpu/kernels/groupnorm.js +91 -0
- package/src/gpu/kernels/groupnorm_apply.wgsl +41 -0
- package/src/gpu/kernels/groupnorm_apply_f16.wgsl +46 -0
- package/src/gpu/kernels/groupnorm_stats.wgsl +76 -0
- package/src/gpu/kernels/groupnorm_stats_f16.wgsl +79 -0
- package/src/gpu/kernels/index.d.ts +336 -0
- package/src/gpu/kernels/index.js +284 -0
- package/src/gpu/kernels/kernel-base.d.ts +33 -0
- package/src/gpu/kernels/kernel-base.js +46 -0
- package/src/gpu/kernels/kernel-configs.d.ts +65 -0
- package/src/gpu/kernels/kernel-configs.js +50 -0
- package/src/gpu/kernels/kernel-tuning.d.ts +42 -0
- package/src/gpu/kernels/kernel-tuning.js +149 -0
- package/src/gpu/kernels/kv-quantize.d.ts +37 -0
- package/src/gpu/kernels/kv-quantize.js +138 -0
- package/src/gpu/kernels/kv_quantize_int4.wgsl +119 -0
- package/src/gpu/kernels/kv_quantize_int8.wgsl +119 -0
- package/src/gpu/kernels/layernorm.d.ts +37 -0
- package/src/gpu/kernels/layernorm.js +80 -0
- package/src/gpu/kernels/layernorm.wgsl +121 -0
- package/src/gpu/kernels/layernorm_f16.wgsl +103 -0
- package/src/gpu/kernels/linear-attention-core.d.ts +39 -0
- package/src/gpu/kernels/linear-attention-core.js +535 -0
- package/src/gpu/kernels/logit-merge.d.ts +110 -0
- package/src/gpu/kernels/logit-merge.js +392 -0
- package/src/gpu/kernels/matmul-dispatch.d.ts +38 -0
- package/src/gpu/kernels/matmul-dispatch.js +155 -0
- package/src/gpu/kernels/matmul-selection.d.ts +87 -0
- package/src/gpu/kernels/matmul-selection.js +474 -0
- package/src/gpu/kernels/matmul.d.ts +109 -0
- package/src/gpu/kernels/matmul.js +271 -0
- package/src/gpu/kernels/matmul_f16.wgsl +170 -0
- package/src/gpu/kernels/matmul_f16_tiled.wgsl +165 -0
- package/src/gpu/kernels/matmul_f16w_f32a.wgsl +89 -0
- package/src/gpu/kernels/matmul_f16w_f32a_tiled.wgsl +154 -0
- package/src/gpu/kernels/matmul_f32.wgsl +100 -0
- package/src/gpu/kernels/matmul_gemv.wgsl +80 -0
- package/src/gpu/kernels/matmul_gemv_f16a.wgsl +81 -0
- package/src/gpu/kernels/matmul_gemv_residual.wgsl +119 -0
- package/src/gpu/kernels/matmul_gemv_residual_f16.wgsl +78 -0
- package/src/gpu/kernels/matmul_gemv_subgroup.wgsl +345 -0
- package/src/gpu/kernels/matmul_gemv_subgroup_f16a.wgsl +514 -0
- package/src/gpu/kernels/modulate.d.ts +29 -0
- package/src/gpu/kernels/modulate.js +49 -0
- package/src/gpu/kernels/modulate.wgsl +40 -0
- package/src/gpu/kernels/modulate_f16.wgsl +43 -0
- package/src/gpu/kernels/moe.d.ts +164 -0
- package/src/gpu/kernels/moe.js +496 -0
- package/src/gpu/kernels/moe_gather.wgsl +170 -0
- package/src/gpu/kernels/moe_gather_f16.wgsl +82 -0
- package/src/gpu/kernels/moe_gather_vec4.wgsl +74 -0
- package/src/gpu/kernels/moe_offsets.wgsl +48 -0
- package/src/gpu/kernels/pipeline-cache.d.ts +88 -0
- package/src/gpu/kernels/pipeline-cache.js +305 -0
- package/src/gpu/kernels/pixel_shuffle.d.ts +27 -0
- package/src/gpu/kernels/pixel_shuffle.js +49 -0
- package/src/gpu/kernels/pixel_shuffle.wgsl +44 -0
- package/src/gpu/kernels/pixel_shuffle_f16.wgsl +47 -0
- package/src/gpu/kernels/residual.d.ts +74 -0
- package/src/gpu/kernels/residual.js +127 -0
- package/src/gpu/kernels/residual.wgsl +53 -0
- package/src/gpu/kernels/residual_f16.wgsl +35 -0
- package/src/gpu/kernels/residual_f16_vec4.wgsl +47 -0
- package/src/gpu/kernels/residual_vec4.wgsl +46 -0
- package/src/gpu/kernels/rmsnorm.d.ts +53 -0
- package/src/gpu/kernels/rmsnorm.js +140 -0
- package/src/gpu/kernels/rmsnorm.wgsl +417 -0
- package/src/gpu/kernels/rmsnorm_f16.wgsl +164 -0
- package/src/gpu/kernels/rope.d.ts +48 -0
- package/src/gpu/kernels/rope.js +53 -0
- package/src/gpu/kernels/rope.wgsl +328 -0
- package/src/gpu/kernels/rope_f16.wgsl +271 -0
- package/src/gpu/kernels/rule-matcher.d.ts +30 -0
- package/src/gpu/kernels/rule-matcher.js +42 -0
- package/src/gpu/kernels/rule-registry.d.ts +7 -0
- package/src/gpu/kernels/rule-registry.js +41 -0
- package/src/gpu/kernels/sample.d.ts +75 -0
- package/src/gpu/kernels/sample.js +578 -0
- package/src/gpu/kernels/sample.wgsl +377 -0
- package/src/gpu/kernels/sample_f16.wgsl +331 -0
- package/src/gpu/kernels/scale.d.ts +35 -0
- package/src/gpu/kernels/scale.js +37 -0
- package/src/gpu/kernels/scale.wgsl +38 -0
- package/src/gpu/kernels/scatter_add.wgsl +88 -0
- package/src/gpu/kernels/scatter_add_dynamic.wgsl +59 -0
- package/src/gpu/kernels/scatter_add_dynamic_f16.wgsl +52 -0
- package/src/gpu/kernels/scatter_add_dynamic_f16_weights.wgsl +50 -0
- package/src/gpu/kernels/scatter_add_vec4.wgsl +70 -0
- package/src/gpu/kernels/shader-cache.d.ts +56 -0
- package/src/gpu/kernels/shader-cache.js +206 -0
- package/src/gpu/kernels/silu.d.ts +75 -0
- package/src/gpu/kernels/silu.js +340 -0
- package/src/gpu/kernels/silu.wgsl +99 -0
- package/src/gpu/kernels/silu_f16.wgsl +98 -0
- package/src/gpu/kernels/softmax.d.ts +57 -0
- package/src/gpu/kernels/softmax.js +106 -0
- package/src/gpu/kernels/softmax.wgsl +388 -0
- package/src/gpu/kernels/softmax_subgroup.wgsl +175 -0
- package/src/gpu/kernels/split_qkv.d.ts +51 -0
- package/src/gpu/kernels/split_qkv.js +41 -0
- package/src/gpu/kernels/split_qkv.wgsl +71 -0
- package/src/gpu/kernels/split_qkv_f16.wgsl +75 -0
- package/src/gpu/kernels/topk.wgsl +243 -0
- package/src/gpu/kernels/topk_f16.wgsl +108 -0
- package/src/gpu/kernels/topk_f16_weights.wgsl +101 -0
- package/src/gpu/kernels/transpose.d.ts +21 -0
- package/src/gpu/kernels/transpose.js +30 -0
- package/src/gpu/kernels/transpose.wgsl +32 -0
- package/src/gpu/kernels/types.d.ts +21 -0
- package/src/gpu/kernels/types.js +4 -0
- package/src/gpu/kernels/uniform-utils.d.ts +48 -0
- package/src/gpu/kernels/uniform-utils.js +94 -0
- package/src/gpu/kernels/upsample2d.d.ts +25 -0
- package/src/gpu/kernels/upsample2d.js +58 -0
- package/src/gpu/kernels/upsample2d.wgsl +37 -0
- package/src/gpu/kernels/upsample2d_f16.wgsl +41 -0
- package/src/gpu/kernels/utils.d.ts +106 -0
- package/src/gpu/kernels/utils.js +224 -0
- package/src/gpu/multi-model-recorder.d.ts +21 -0
- package/src/gpu/multi-model-recorder.js +31 -0
- package/src/gpu/partitioned-buffer-pool.d.ts +28 -0
- package/src/gpu/partitioned-buffer-pool.js +49 -0
- package/src/gpu/perf-guards.d.ts +25 -0
- package/src/gpu/perf-guards.js +140 -0
- package/src/gpu/profiler.d.ts +114 -0
- package/src/gpu/profiler.js +391 -0
- package/src/gpu/submit-tracker.d.ts +111 -0
- package/src/gpu/submit-tracker.js +229 -0
- package/src/gpu/tensor.d.ts +69 -0
- package/src/gpu/tensor.js +75 -0
- package/src/gpu/uniform-cache.d.ts +108 -0
- package/src/gpu/uniform-cache.js +242 -0
- package/src/gpu/weight-buffer.d.ts +115 -0
- package/src/gpu/weight-buffer.js +118 -0
- package/src/hotswap/intent-bundle.d.ts +37 -0
- package/src/hotswap/intent-bundle.js +123 -0
- package/src/hotswap/manifest.d.ts +33 -0
- package/src/hotswap/manifest.js +114 -0
- package/src/hotswap/runtime.d.ts +31 -0
- package/src/hotswap/runtime.js +128 -0
- package/src/index-browser.d.ts +47 -0
- package/src/index-browser.js +53 -0
- package/src/index-internal.d.ts +2 -0
- package/src/index-internal.js +2 -0
- package/src/index.d.ts +102 -0
- package/src/index.js +75 -0
- package/src/inference/README.md +593 -0
- package/src/inference/browser-harness.d.ts +234 -0
- package/src/inference/browser-harness.js +2665 -0
- package/src/inference/decode-buffers.d.ts +108 -0
- package/src/inference/decode-buffers.js +181 -0
- package/src/inference/decode-ring.d.ts +52 -0
- package/src/inference/decode-ring.js +273 -0
- package/src/inference/expert-router.d.ts +27 -0
- package/src/inference/expert-router.js +55 -0
- package/src/inference/functiongemma.d.ts +15 -0
- package/src/inference/functiongemma.js +1 -0
- package/src/inference/kv-cache/base.d.ts +150 -0
- package/src/inference/kv-cache/base.js +1037 -0
- package/src/inference/kv-cache/basis-decomposed-paged.d.ts +50 -0
- package/src/inference/kv-cache/basis-decomposed-paged.js +276 -0
- package/src/inference/kv-cache/index.d.ts +35 -0
- package/src/inference/kv-cache/index.js +20 -0
- package/src/inference/kv-cache/sliding-window.d.ts +72 -0
- package/src/inference/kv-cache/sliding-window.js +243 -0
- package/src/inference/kv-cache/tiered.d.ts +89 -0
- package/src/inference/kv-cache/tiered.js +574 -0
- package/src/inference/kv-cache/types.d.ts +188 -0
- package/src/inference/kv-cache/types.js +80 -0
- package/src/inference/kv-cache.d.ts +36 -0
- package/src/inference/kv-cache.js +18 -0
- package/src/inference/moe-router.d.ts +212 -0
- package/src/inference/moe-router.js +553 -0
- package/src/inference/multi-model-network.d.ts +139 -0
- package/src/inference/multi-model-network.js +769 -0
- package/src/inference/multi-pipeline-pool.d.ts +62 -0
- package/src/inference/multi-pipeline-pool.js +161 -0
- package/src/inference/network-evolution.d.ts +46 -0
- package/src/inference/network-evolution.js +80 -0
- package/src/inference/pipelines/context.d.ts +18 -0
- package/src/inference/pipelines/context.js +44 -0
- package/src/inference/pipelines/diffusion/helpers.d.ts +29 -0
- package/src/inference/pipelines/diffusion/helpers.js +112 -0
- package/src/inference/pipelines/diffusion/index.d.ts +3 -0
- package/src/inference/pipelines/diffusion/index.js +3 -0
- package/src/inference/pipelines/diffusion/init.d.ts +24 -0
- package/src/inference/pipelines/diffusion/init.js +124 -0
- package/src/inference/pipelines/diffusion/pipeline.d.ts +38 -0
- package/src/inference/pipelines/diffusion/pipeline.js +632 -0
- package/src/inference/pipelines/diffusion/scheduler.d.ts +19 -0
- package/src/inference/pipelines/diffusion/scheduler.js +65 -0
- package/src/inference/pipelines/diffusion/sd3-transformer.d.ts +20 -0
- package/src/inference/pipelines/diffusion/sd3-transformer.js +1194 -0
- package/src/inference/pipelines/diffusion/sd3-weights.d.ts +21 -0
- package/src/inference/pipelines/diffusion/sd3-weights.js +287 -0
- package/src/inference/pipelines/diffusion/text-encoder-gpu.d.ts +80 -0
- package/src/inference/pipelines/diffusion/text-encoder-gpu.js +935 -0
- package/src/inference/pipelines/diffusion/text-encoder.d.ts +29 -0
- package/src/inference/pipelines/diffusion/text-encoder.js +178 -0
- package/src/inference/pipelines/diffusion/types.d.ts +112 -0
- package/src/inference/pipelines/diffusion/types.js +1 -0
- package/src/inference/pipelines/diffusion/vae.d.ts +20 -0
- package/src/inference/pipelines/diffusion/vae.js +675 -0
- package/src/inference/pipelines/diffusion/weights.d.ts +40 -0
- package/src/inference/pipelines/diffusion/weights.js +150 -0
- package/src/inference/pipelines/dream/energy-head-pipeline.d.ts +29 -0
- package/src/inference/pipelines/dream/energy-head-pipeline.js +6 -0
- package/src/inference/pipelines/dream/pipeline.d.ts +17 -0
- package/src/inference/pipelines/dream/pipeline.js +8 -0
- package/src/inference/pipelines/energy/index.d.ts +1 -0
- package/src/inference/pipelines/energy/index.js +1 -0
- package/src/inference/pipelines/energy/pipeline.d.ts +27 -0
- package/src/inference/pipelines/energy/pipeline.js +680 -0
- package/src/inference/pipelines/energy/quintel.d.ts +87 -0
- package/src/inference/pipelines/energy/quintel.js +207 -0
- package/src/inference/pipelines/energy/types.d.ts +63 -0
- package/src/inference/pipelines/energy/types.js +1 -0
- package/src/inference/pipelines/energy-head/index.d.ts +6 -0
- package/src/inference/pipelines/energy-head/index.js +6 -0
- package/src/inference/pipelines/energy-head/row-head-pipeline.d.ts +103 -0
- package/src/inference/pipelines/energy-head/row-head-pipeline.js +487 -0
- package/src/inference/pipelines/factory.d.ts +10 -0
- package/src/inference/pipelines/factory.js +6 -0
- package/src/inference/pipelines/index.d.ts +22 -0
- package/src/inference/pipelines/index.js +19 -0
- package/src/inference/pipelines/registry.d.ts +15 -0
- package/src/inference/pipelines/registry.js +23 -0
- package/src/inference/pipelines/rng.d.ts +2 -0
- package/src/inference/pipelines/rng.js +17 -0
- package/src/inference/pipelines/structured/index.d.ts +8 -0
- package/src/inference/pipelines/structured/index.js +8 -0
- package/src/inference/pipelines/structured/json-head-pipeline.d.ts +58 -0
- package/src/inference/pipelines/structured/json-head-pipeline.js +181 -0
- package/src/inference/pipelines/text/attention/index.d.ts +24 -0
- package/src/inference/pipelines/text/attention/index.js +17 -0
- package/src/inference/pipelines/text/attention/projections.d.ts +101 -0
- package/src/inference/pipelines/text/attention/projections.js +435 -0
- package/src/inference/pipelines/text/attention/record.d.ts +36 -0
- package/src/inference/pipelines/text/attention/record.js +613 -0
- package/src/inference/pipelines/text/attention/run.d.ts +38 -0
- package/src/inference/pipelines/text/attention/run.js +826 -0
- package/src/inference/pipelines/text/attention/types.d.ts +98 -0
- package/src/inference/pipelines/text/attention/types.js +67 -0
- package/src/inference/pipelines/text/attention.d.ts +23 -0
- package/src/inference/pipelines/text/attention.js +12 -0
- package/src/inference/pipelines/text/bdpa-steamroller.d.ts +22 -0
- package/src/inference/pipelines/text/bdpa-steamroller.js +158 -0
- package/src/inference/pipelines/text/buffer-types.d.ts +7 -0
- package/src/inference/pipelines/text/buffer-types.js +4 -0
- package/src/inference/pipelines/text/chat-format.d.ts +46 -0
- package/src/inference/pipelines/text/chat-format.js +366 -0
- package/src/inference/pipelines/text/config.d.ts +235 -0
- package/src/inference/pipelines/text/config.js +623 -0
- package/src/inference/pipelines/text/debug-utils/config.d.ts +144 -0
- package/src/inference/pipelines/text/debug-utils/config.js +156 -0
- package/src/inference/pipelines/text/debug-utils/index.d.ts +53 -0
- package/src/inference/pipelines/text/debug-utils/index.js +44 -0
- package/src/inference/pipelines/text/debug-utils/logging.d.ts +106 -0
- package/src/inference/pipelines/text/debug-utils/logging.js +152 -0
- package/src/inference/pipelines/text/debug-utils/tensor.d.ts +119 -0
- package/src/inference/pipelines/text/debug-utils/tensor.js +268 -0
- package/src/inference/pipelines/text/debug-utils/utils.d.ts +77 -0
- package/src/inference/pipelines/text/debug-utils/utils.js +139 -0
- package/src/inference/pipelines/text/debug-utils.d.ts +42 -0
- package/src/inference/pipelines/text/debug-utils.js +34 -0
- package/src/inference/pipelines/text/embed.d.ts +67 -0
- package/src/inference/pipelines/text/embed.js +461 -0
- package/src/inference/pipelines/text/execution-plan.d.ts +116 -0
- package/src/inference/pipelines/text/execution-plan.js +314 -0
- package/src/inference/pipelines/text/execution-v0.d.ts +66 -0
- package/src/inference/pipelines/text/execution-v0.js +1139 -0
- package/src/inference/pipelines/text/ffn/dense.d.ts +40 -0
- package/src/inference/pipelines/text/ffn/dense.js +759 -0
- package/src/inference/pipelines/text/ffn/index.d.ts +23 -0
- package/src/inference/pipelines/text/ffn/index.js +16 -0
- package/src/inference/pipelines/text/ffn/moe.d.ts +21 -0
- package/src/inference/pipelines/text/ffn/moe.js +49 -0
- package/src/inference/pipelines/text/ffn/sandwich.d.ts +25 -0
- package/src/inference/pipelines/text/ffn/sandwich.js +196 -0
- package/src/inference/pipelines/text/ffn/standard.d.ts +23 -0
- package/src/inference/pipelines/text/ffn/standard.js +84 -0
- package/src/inference/pipelines/text/ffn/types.d.ts +30 -0
- package/src/inference/pipelines/text/ffn/types.js +25 -0
- package/src/inference/pipelines/text/ffn.d.ts +31 -0
- package/src/inference/pipelines/text/ffn.js +18 -0
- package/src/inference/pipelines/text/finiteness-guard-status.d.ts +11 -0
- package/src/inference/pipelines/text/finiteness-guard-status.js +21 -0
- package/src/inference/pipelines/text/finiteness-policy.d.ts +35 -0
- package/src/inference/pipelines/text/finiteness-policy.js +45 -0
- package/src/inference/pipelines/text/generator-helpers.d.ts +34 -0
- package/src/inference/pipelines/text/generator-helpers.js +175 -0
- package/src/inference/pipelines/text/generator-runtime.d.ts +93 -0
- package/src/inference/pipelines/text/generator-runtime.js +373 -0
- package/src/inference/pipelines/text/generator-steps.d.ts +75 -0
- package/src/inference/pipelines/text/generator-steps.js +1078 -0
- package/src/inference/pipelines/text/generator.d.ts +41 -0
- package/src/inference/pipelines/text/generator.js +1345 -0
- package/src/inference/pipelines/text/index.d.ts +5 -0
- package/src/inference/pipelines/text/index.js +6 -0
- package/src/inference/pipelines/text/init.d.ts +295 -0
- package/src/inference/pipelines/text/init.js +965 -0
- package/src/inference/pipelines/text/kernel-path-auto-select.d.ts +12 -0
- package/src/inference/pipelines/text/kernel-path-auto-select.js +90 -0
- package/src/inference/pipelines/text/kernel-trace.d.ts +150 -0
- package/src/inference/pipelines/text/kernel-trace.js +324 -0
- package/src/inference/pipelines/text/layer-plan.d.ts +65 -0
- package/src/inference/pipelines/text/layer-plan.js +249 -0
- package/src/inference/pipelines/text/layer.d.ts +56 -0
- package/src/inference/pipelines/text/layer.js +916 -0
- package/src/inference/pipelines/text/linear-attention.d.ts +94 -0
- package/src/inference/pipelines/text/linear-attention.js +803 -0
- package/src/inference/pipelines/text/logits/cpu.d.ts +81 -0
- package/src/inference/pipelines/text/logits/cpu.js +91 -0
- package/src/inference/pipelines/text/logits/gpu.d.ts +113 -0
- package/src/inference/pipelines/text/logits/gpu.js +406 -0
- package/src/inference/pipelines/text/logits/index.d.ts +57 -0
- package/src/inference/pipelines/text/logits/index.js +305 -0
- package/src/inference/pipelines/text/logits/types.d.ts +46 -0
- package/src/inference/pipelines/text/logits/types.js +4 -0
- package/src/inference/pipelines/text/logits/utils.d.ts +49 -0
- package/src/inference/pipelines/text/logits/utils.js +59 -0
- package/src/inference/pipelines/text/logits.d.ts +27 -0
- package/src/inference/pipelines/text/logits.js +16 -0
- package/src/inference/pipelines/text/lora-apply.d.ts +28 -0
- package/src/inference/pipelines/text/lora-apply.js +58 -0
- package/src/inference/pipelines/text/lora-types.d.ts +39 -0
- package/src/inference/pipelines/text/lora-types.js +18 -0
- package/src/inference/pipelines/text/lora.d.ts +18 -0
- package/src/inference/pipelines/text/lora.js +12 -0
- package/src/inference/pipelines/text/model-load.d.ts +58 -0
- package/src/inference/pipelines/text/model-load.js +561 -0
- package/src/inference/pipelines/text/moe-cache.d.ts +32 -0
- package/src/inference/pipelines/text/moe-cache.js +107 -0
- package/src/inference/pipelines/text/moe-cpu-gptoss.d.ts +9 -0
- package/src/inference/pipelines/text/moe-cpu-gptoss.js +110 -0
- package/src/inference/pipelines/text/moe-cpu.d.ts +13 -0
- package/src/inference/pipelines/text/moe-cpu.js +116 -0
- package/src/inference/pipelines/text/moe-gpu.d.ts +13 -0
- package/src/inference/pipelines/text/moe-gpu.js +611 -0
- package/src/inference/pipelines/text/moe-helpers.d.ts +12 -0
- package/src/inference/pipelines/text/moe-helpers.js +21 -0
- package/src/inference/pipelines/text/moe-impl.d.ts +117 -0
- package/src/inference/pipelines/text/moe-impl.js +9 -0
- package/src/inference/pipelines/text/moe-shape-validator.d.ts +31 -0
- package/src/inference/pipelines/text/moe-shape-validator.js +78 -0
- package/src/inference/pipelines/text/ops.d.ts +167 -0
- package/src/inference/pipelines/text/ops.js +367 -0
- package/src/inference/pipelines/text/probes.d.ts +31 -0
- package/src/inference/pipelines/text/probes.js +170 -0
- package/src/inference/pipelines/text/sampling.d.ts +54 -0
- package/src/inference/pipelines/text/sampling.js +203 -0
- package/src/inference/pipelines/text/state.d.ts +112 -0
- package/src/inference/pipelines/text/state.js +152 -0
- package/src/inference/pipelines/text/types.d.ts +627 -0
- package/src/inference/pipelines/text/types.js +4 -0
- package/src/inference/pipelines/text/weights.d.ts +110 -0
- package/src/inference/pipelines/text/weights.js +163 -0
- package/src/inference/pipelines/text.d.ts +157 -0
- package/src/inference/pipelines/text.js +586 -0
- package/src/inference/speculative.d.ts +239 -0
- package/src/inference/speculative.js +416 -0
- package/src/inference/test-harness.d.ts +178 -0
- package/src/inference/test-harness.js +349 -0
- package/src/inference/tokenizer.d.ts +77 -0
- package/src/inference/tokenizer.js +258 -0
- package/src/inference/tokenizers/base.d.ts +39 -0
- package/src/inference/tokenizers/base.js +69 -0
- package/src/inference/tokenizers/bpe.d.ts +27 -0
- package/src/inference/tokenizers/bpe.js +171 -0
- package/src/inference/tokenizers/bundled.d.ts +63 -0
- package/src/inference/tokenizers/bundled.js +866 -0
- package/src/inference/tokenizers/sentencepiece.d.ts +28 -0
- package/src/inference/tokenizers/sentencepiece.js +389 -0
- package/src/inference/tokenizers/types.d.ts +166 -0
- package/src/inference/tokenizers/types.js +7 -0
- package/src/loader/doppler-loader.d.ts +134 -0
- package/src/loader/doppler-loader.js +1036 -0
- package/src/loader/dtype-utils.d.ts +40 -0
- package/src/loader/dtype-utils.js +102 -0
- package/src/loader/embedding-loader.d.ts +56 -0
- package/src/loader/embedding-loader.js +207 -0
- package/src/loader/experts/expert-cache.d.ts +156 -0
- package/src/loader/experts/expert-cache.js +375 -0
- package/src/loader/experts/expert-loader.d.ts +108 -0
- package/src/loader/experts/expert-loader.js +384 -0
- package/src/loader/final-weights-loader.d.ts +68 -0
- package/src/loader/final-weights-loader.js +262 -0
- package/src/loader/index.d.ts +150 -0
- package/src/loader/index.js +124 -0
- package/src/loader/layer-loader.d.ts +63 -0
- package/src/loader/layer-loader.js +417 -0
- package/src/loader/loader-state.d.ts +51 -0
- package/src/loader/loader-state.js +142 -0
- package/src/loader/loader-types.d.ts +236 -0
- package/src/loader/loader-types.js +4 -0
- package/src/loader/manifest-config.d.ts +97 -0
- package/src/loader/manifest-config.js +132 -0
- package/src/loader/memory-monitor.d.ts +112 -0
- package/src/loader/memory-monitor.js +276 -0
- package/src/loader/multi-model-loader.d.ts +37 -0
- package/src/loader/multi-model-loader.js +87 -0
- package/src/loader/quantization-constants.d.ts +23 -0
- package/src/loader/quantization-constants.js +14 -0
- package/src/loader/shard-cache.d.ts +60 -0
- package/src/loader/shard-cache.js +568 -0
- package/src/loader/shard-resolver.d.ts +12 -0
- package/src/loader/shard-resolver.js +83 -0
- package/src/loader/tensors/tensor-loader.d.ts +154 -0
- package/src/loader/tensors/tensor-loader.js +427 -0
- package/src/loader/tensors/tensor-reader.d.ts +22 -0
- package/src/loader/tensors/tensor-reader.js +56 -0
- package/src/loader/tensors/tensor-role.d.ts +7 -0
- package/src/loader/tensors/tensor-role.js +12 -0
- package/src/loader/weight-downcast.d.ts +62 -0
- package/src/loader/weight-downcast.js +213 -0
- package/src/loader/weights.d.ts +22 -0
- package/src/loader/weights.js +4 -0
- package/src/memory/address-table.d.ts +104 -0
- package/src/memory/address-table.js +114 -0
- package/src/memory/buffer-pool.d.ts +196 -0
- package/src/memory/buffer-pool.js +756 -0
- package/src/memory/capability.d.ts +49 -0
- package/src/memory/capability.js +95 -0
- package/src/memory/heap-manager.d.ts +104 -0
- package/src/memory/heap-manager.js +264 -0
- package/src/memory/unified-detect.d.ts +59 -0
- package/src/memory/unified-detect.js +192 -0
- package/src/rules/converter/execution.rules.json +20 -0
- package/src/rules/converter/tensor-roles.rules.json +13 -0
- package/src/rules/converter/tokenizer.rules.json +7 -0
- package/src/rules/inference/attention.rules.json +54 -0
- package/src/rules/inference/config.rules.json +58 -0
- package/src/rules/inference/dtype.rules.json +94 -0
- package/src/rules/inference/execution.rules.json +45 -0
- package/src/rules/inference/ffn.rules.json +35 -0
- package/src/rules/inference/kernel-path.rules.json +76 -0
- package/src/rules/inference/layer-pattern.rules.json +16 -0
- package/src/rules/inference/layer.rules.json +7 -0
- package/src/rules/inference/moe.rules.json +48 -0
- package/src/rules/kernels/attention.rules.json +61 -0
- package/src/rules/kernels/conv2d.rules.json +6 -0
- package/src/rules/kernels/dequant.rules.json +58 -0
- package/src/rules/kernels/energy.rules.json +22 -0
- package/src/rules/kernels/fused-ffn.rules.json +13 -0
- package/src/rules/kernels/fused-matmul-residual.rules.json +6 -0
- package/src/rules/kernels/fused-matmul-rmsnorm.rules.json +8 -0
- package/src/rules/kernels/gather.rules.json +12 -0
- package/src/rules/kernels/gelu.rules.json +11 -0
- package/src/rules/kernels/groupnorm.rules.json +10 -0
- package/src/rules/kernels/kernel-validator.d.ts +24 -0
- package/src/rules/kernels/kernel-validator.js +160 -0
- package/src/rules/kernels/kv_quantize.rules.json +7 -0
- package/src/rules/kernels/layernorm.rules.json +6 -0
- package/src/rules/kernels/matmul.rules.json +60 -0
- package/src/rules/kernels/modulate.rules.json +6 -0
- package/src/rules/kernels/moe.rules.gptoss.json +105 -0
- package/src/rules/kernels/moe.rules.json +11 -0
- package/src/rules/kernels/pixel_shuffle.rules.json +6 -0
- package/src/rules/kernels/residual.rules.json +12 -0
- package/src/rules/kernels/rmsnorm.rules.json +11 -0
- package/src/rules/kernels/rope.rules.json +6 -0
- package/src/rules/kernels/sample.rules.json +6 -0
- package/src/rules/kernels/scale.rules.json +6 -0
- package/src/rules/kernels/silu.rules.json +21 -0
- package/src/rules/kernels/softmax.rules.json +23 -0
- package/src/rules/kernels/split-qkv.rules.json +6 -0
- package/src/rules/kernels/upsample2d.rules.json +6 -0
- package/src/rules/loader/tensor-loader.rules.json +15 -0
- package/src/rules/loader/weights.rules.json +41 -0
- package/src/rules/rule-registry.d.ts +48 -0
- package/src/rules/rule-registry.js +177 -0
- package/src/rules/tooling/command-runtime.rules.json +38 -0
- package/src/storage/backends/idb-store.d.ts +52 -0
- package/src/storage/backends/idb-store.js +590 -0
- package/src/storage/backends/memory-store.d.ts +36 -0
- package/src/storage/backends/memory-store.js +242 -0
- package/src/storage/backends/opfs-store.d.ts +41 -0
- package/src/storage/backends/opfs-store.js +429 -0
- package/src/storage/blake3.d.ts +17 -0
- package/src/storage/blake3.js +269 -0
- package/src/storage/download-types.d.ts +157 -0
- package/src/storage/download-types.js +48 -0
- package/src/storage/downloader.d.ts +103 -0
- package/src/storage/downloader.js +839 -0
- package/src/storage/emulated-vram.d.ts +264 -0
- package/src/storage/emulated-vram.js +576 -0
- package/src/storage/export.d.ts +20 -0
- package/src/storage/export.js +159 -0
- package/src/storage/index.d.ts +253 -0
- package/src/storage/index.js +185 -0
- package/src/storage/inventory.d.ts +26 -0
- package/src/storage/inventory.js +218 -0
- package/src/storage/preflight.d.ts +144 -0
- package/src/storage/preflight.js +294 -0
- package/src/storage/quickstart-downloader.d.ts +154 -0
- package/src/storage/quickstart-downloader.js +265 -0
- package/src/storage/quota.d.ts +150 -0
- package/src/storage/quota.js +304 -0
- package/src/storage/registry.d.ts +28 -0
- package/src/storage/registry.js +125 -0
- package/src/storage/reports.d.ts +20 -0
- package/src/storage/reports.js +94 -0
- package/src/storage/shard-manager.d.ts +137 -0
- package/src/storage/shard-manager.js +801 -0
- package/src/sw.d.ts +1 -0
- package/src/sw.js +187 -0
- package/src/tooling/browser-command-runner.d.ts +28 -0
- package/src/tooling/browser-command-runner.js +82 -0
- package/src/tooling/command-api.d.ts +147 -0
- package/src/tooling/command-api.js +523 -0
- package/src/tooling/command-envelope.d.ts +81 -0
- package/src/tooling/command-envelope.js +195 -0
- package/src/tooling/command-runner-shared.d.ts +73 -0
- package/src/tooling/command-runner-shared.js +146 -0
- package/src/tooling/command-runner.html +45 -0
- package/src/tooling/node-browser-command-runner.d.ts +30 -0
- package/src/tooling/node-browser-command-runner.js +868 -0
- package/src/tooling/node-command-runner.d.ts +36 -0
- package/src/tooling/node-command-runner.js +127 -0
- package/src/tooling/node-convert-worker-pool.d.ts +16 -0
- package/src/tooling/node-convert-worker-pool.js +186 -0
- package/src/tooling/node-convert-worker.d.ts +1 -0
- package/src/tooling/node-convert-worker.js +60 -0
- package/src/tooling/node-convert.d.ts +44 -0
- package/src/tooling/node-converter.d.ts +1 -0
- package/src/tooling/node-converter.js +1227 -0
- package/src/tooling/node-file-fetch.d.ts +1 -0
- package/src/tooling/node-file-fetch.js +38 -0
- package/src/tooling/node-source-runtime.d.ts +19 -0
- package/src/tooling/node-source-runtime.js +469 -0
- package/src/tooling/node-webgpu.d.ts +6 -0
- package/src/tooling/node-webgpu.js +321 -0
- package/src/tooling/opfs-cache.d.ts +11 -0
- package/src/tooling/opfs-cache.js +174 -0
- package/src/tooling/source-runtime-bundle.d.ts +102 -0
- package/src/tooling/source-runtime-bundle.js +484 -0
- package/src/tooling-exports.browser.d.ts +7 -0
- package/src/tooling-exports.browser.js +2 -0
- package/src/tooling-exports.d.ts +22 -0
- package/src/tooling-exports.js +7 -0
- package/src/tooling-exports.shared.d.ts +105 -0
- package/src/tooling-exports.shared.js +92 -0
- package/src/training/README.md +153 -0
- package/src/training/artifacts.d.ts +160 -0
- package/src/training/artifacts.js +896 -0
- package/src/training/attention-backward.d.ts +30 -0
- package/src/training/attention-backward.js +217 -0
- package/src/training/attention-forward.d.ts +22 -0
- package/src/training/attention-forward.js +82 -0
- package/src/training/autograd.d.ts +51 -0
- package/src/training/autograd.js +380 -0
- package/src/training/checkpoint.d.ts +31 -0
- package/src/training/checkpoint.js +238 -0
- package/src/training/clip.d.ts +9 -0
- package/src/training/clip.js +54 -0
- package/src/training/dataloader.d.ts +8 -0
- package/src/training/dataloader.js +44 -0
- package/src/training/datasets/index.d.ts +12 -0
- package/src/training/datasets/index.js +6 -0
- package/src/training/datasets/jsonl.d.ts +11 -0
- package/src/training/datasets/jsonl.js +50 -0
- package/src/training/datasets/reploid.d.ts +3 -0
- package/src/training/datasets/reploid.js +36 -0
- package/src/training/datasets/text-pairs.d.ts +21 -0
- package/src/training/datasets/text-pairs.js +42 -0
- package/src/training/datasets/token-batch.d.ts +21 -0
- package/src/training/datasets/token-batch.js +40 -0
- package/src/training/datasets/translation-pairs.d.ts +34 -0
- package/src/training/datasets/translation-pairs.js +49 -0
- package/src/training/export.d.ts +32 -0
- package/src/training/export.js +112 -0
- package/src/training/index.d.ts +52 -0
- package/src/training/index.js +41 -0
- package/src/training/lora.d.ts +19 -0
- package/src/training/lora.js +57 -0
- package/src/training/loss-scaling.d.ts +21 -0
- package/src/training/loss-scaling.js +80 -0
- package/src/training/loss.d.ts +10 -0
- package/src/training/loss.js +41 -0
- package/src/training/objectives/base.d.ts +58 -0
- package/src/training/objectives/base.js +38 -0
- package/src/training/objectives/cross_entropy.d.ts +18 -0
- package/src/training/objectives/cross_entropy.js +37 -0
- package/src/training/objectives/distill_kd.d.ts +16 -0
- package/src/training/objectives/distill_kd.js +369 -0
- package/src/training/objectives/distill_triplet.d.ts +16 -0
- package/src/training/objectives/distill_triplet.js +412 -0
- package/src/training/objectives/index.d.ts +12 -0
- package/src/training/objectives/index.js +6 -0
- package/src/training/objectives/ul_stage1_joint.d.ts +16 -0
- package/src/training/objectives/ul_stage1_joint.js +188 -0
- package/src/training/objectives/ul_stage2_base.d.ts +16 -0
- package/src/training/objectives/ul_stage2_base.js +222 -0
- package/src/training/optimizer.d.ts +22 -0
- package/src/training/optimizer.js +115 -0
- package/src/training/runner.d.ts +196 -0
- package/src/training/runner.js +1194 -0
- package/src/training/suite.d.ts +187 -0
- package/src/training/suite.js +3156 -0
- package/src/training/trainer.d.ts +89 -0
- package/src/training/trainer.js +301 -0
- package/src/training/ul_dataset.d.ts +47 -0
- package/src/training/ul_dataset.js +153 -0
- package/src/training/ul_schedule.d.ts +6 -0
- package/src/training/ul_schedule.js +29 -0
- package/src/types/chrome.d.ts +36 -0
- package/src/types/chrome.js +1 -0
- package/src/types/gpu.d.ts +185 -0
- package/src/types/gpu.js +5 -0
- package/src/types/index.d.ts +3 -0
- package/src/types/index.js +3 -0
- package/src/types/inference.d.ts +197 -0
- package/src/types/inference.js +5 -0
- package/src/types/model.d.ts +125 -0
- package/src/types/model.js +5 -0
- package/src/utils/index.d.ts +7 -0
- package/src/utils/index.js +7 -0
- package/src/utils/load-json.d.ts +5 -0
- package/src/utils/load-json.js +23 -0
- package/src/utils/plain-object.d.ts +1 -0
- package/src/utils/plain-object.js +3 -0
- package/src/utils/sha256.d.ts +4 -0
- package/src/utils/sha256.js +135 -0
- package/tools/convert-safetensors-node.js +180 -0
- package/tools/doppler-cli.js +1170 -0
|
@@ -0,0 +1,144 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Debug configuration management for pipeline tracing.
|
|
3
|
+
*
|
|
4
|
+
* Manages debug categories, layer filters, and decode step tracking.
|
|
5
|
+
* Enable via: setDebugCategories({ embed: true, layer: true })
|
|
6
|
+
*
|
|
7
|
+
* Categories:
|
|
8
|
+
* - embed: Embedding layer output
|
|
9
|
+
* - layer: Per-layer entry/exit, hidden state stats
|
|
10
|
+
* - attn: Attention computation details
|
|
11
|
+
* - ffn: FFN computation details
|
|
12
|
+
* - kv: KV cache operations
|
|
13
|
+
* - logits: Logits computation and top-k
|
|
14
|
+
* - sample: Sampling decisions
|
|
15
|
+
* - io: GPU buffer read/writes
|
|
16
|
+
* - perf: Timing and benchmarks
|
|
17
|
+
* - kernel: Kernel step debugging - inspect tensor state after each kernel (SLOW!)
|
|
18
|
+
* - all: Enable everything (use sparingly)
|
|
19
|
+
*
|
|
20
|
+
* @module inference/pipelines/text/debug-utils/config
|
|
21
|
+
*/
|
|
22
|
+
|
|
23
|
+
import type { PipelineDebugConfigSchema } from '../../../../config/schema/index.js';
|
|
24
|
+
|
|
25
|
+
// ============================================================================
|
|
26
|
+
// Debug Configuration Types
|
|
27
|
+
// ============================================================================
|
|
28
|
+
|
|
29
|
+
export type DebugCategory =
|
|
30
|
+
| 'embed'
|
|
31
|
+
| 'layer'
|
|
32
|
+
| 'attn'
|
|
33
|
+
| 'ffn'
|
|
34
|
+
| 'kv'
|
|
35
|
+
| 'logits'
|
|
36
|
+
| 'sample'
|
|
37
|
+
| 'io'
|
|
38
|
+
| 'perf'
|
|
39
|
+
| 'kernel'
|
|
40
|
+
| 'all';
|
|
41
|
+
|
|
42
|
+
export interface DebugConfig {
|
|
43
|
+
/** Which categories are enabled */
|
|
44
|
+
categories: Partial<Record<DebugCategory, boolean>>;
|
|
45
|
+
/** Only log these layer indices (empty = all) */
|
|
46
|
+
layers?: number[];
|
|
47
|
+
/** Only log first N decode steps (0 = all) */
|
|
48
|
+
maxDecodeSteps?: number;
|
|
49
|
+
/** Warn if maxAbs exceeds this */
|
|
50
|
+
maxAbsThreshold?: number;
|
|
51
|
+
/** Log GPU buffer stats (expensive - requires readback) */
|
|
52
|
+
bufferStats?: boolean;
|
|
53
|
+
}
|
|
54
|
+
|
|
55
|
+
// ============================================================================
|
|
56
|
+
// Configuration API
|
|
57
|
+
// ============================================================================
|
|
58
|
+
|
|
59
|
+
/**
|
|
60
|
+
* Set debug categories. Merges with existing config.
|
|
61
|
+
*
|
|
62
|
+
* @example
|
|
63
|
+
* setDebugCategories({ embed: true, layer: true });
|
|
64
|
+
* setDebugCategories({ all: true }); // Enable everything
|
|
65
|
+
* setDebugCategories({ all: true, io: false }); // All except io
|
|
66
|
+
*/
|
|
67
|
+
export function setDebugCategories(
|
|
68
|
+
categories: Partial<Record<DebugCategory, boolean>>,
|
|
69
|
+
options?: Partial<Omit<DebugConfig, 'categories'>>
|
|
70
|
+
): void;
|
|
71
|
+
|
|
72
|
+
/**
|
|
73
|
+
* Reset debug config to defaults (all off).
|
|
74
|
+
*/
|
|
75
|
+
export function resetDebugConfig(): void;
|
|
76
|
+
|
|
77
|
+
/**
|
|
78
|
+
* Apply pipeline debug config (runtime.shared.debug.pipeline) to debug-utils.
|
|
79
|
+
*/
|
|
80
|
+
export function applyPipelineDebugConfig(pipeline?: PipelineDebugConfigSchema | null): void;
|
|
81
|
+
|
|
82
|
+
/**
|
|
83
|
+
* Get current debug config (for inspection).
|
|
84
|
+
*/
|
|
85
|
+
export function getDebugConfig(): DebugConfig;
|
|
86
|
+
|
|
87
|
+
// ============================================================================
|
|
88
|
+
// Decode Step Tracking
|
|
89
|
+
// ============================================================================
|
|
90
|
+
|
|
91
|
+
/**
|
|
92
|
+
* Increment decode step counter.
|
|
93
|
+
*/
|
|
94
|
+
export function incrementDecodeStep(): number;
|
|
95
|
+
|
|
96
|
+
/**
|
|
97
|
+
* Reset decode step counter (call at start of generation).
|
|
98
|
+
*/
|
|
99
|
+
export function resetDecodeStep(): void;
|
|
100
|
+
|
|
101
|
+
/**
|
|
102
|
+
* Get current decode step.
|
|
103
|
+
*/
|
|
104
|
+
export function getDecodeStep(): number;
|
|
105
|
+
|
|
106
|
+
// ============================================================================
|
|
107
|
+
// Layer Filtering
|
|
108
|
+
// ============================================================================
|
|
109
|
+
|
|
110
|
+
/**
|
|
111
|
+
* Check if a layer should be debugged based on debugLayers config.
|
|
112
|
+
* @param layerIdx - The layer index to check
|
|
113
|
+
* @param debugLayers - Array of layer indices to debug, null means none, undefined/empty means hardcoded defaults
|
|
114
|
+
* @returns true if the layer should be debugged
|
|
115
|
+
*/
|
|
116
|
+
export function shouldDebugLayerOutput(layerIdx: number, debugLayers: number[] | null | undefined): boolean;
|
|
117
|
+
|
|
118
|
+
// ============================================================================
|
|
119
|
+
// Internal Helpers (exported for use by other debug-utils modules)
|
|
120
|
+
// ============================================================================
|
|
121
|
+
|
|
122
|
+
/**
|
|
123
|
+
* Check if a debug category is enabled for the given layer.
|
|
124
|
+
* @internal
|
|
125
|
+
*/
|
|
126
|
+
export function isEnabled(category: DebugCategory, layerIdx?: number): boolean;
|
|
127
|
+
|
|
128
|
+
/**
|
|
129
|
+
* Format a debug tag with category, layer, and step info.
|
|
130
|
+
* @internal
|
|
131
|
+
*/
|
|
132
|
+
export function formatTag(category: string, layerIdx?: number, step?: number): string;
|
|
133
|
+
|
|
134
|
+
/**
|
|
135
|
+
* Check if buffer stats collection is enabled.
|
|
136
|
+
* @internal
|
|
137
|
+
*/
|
|
138
|
+
export function isBufferStatsEnabled(): boolean;
|
|
139
|
+
|
|
140
|
+
/**
|
|
141
|
+
* Get max abs threshold for explosion warnings.
|
|
142
|
+
* @internal
|
|
143
|
+
*/
|
|
144
|
+
export function getMaxAbsThreshold(): number;
|
|
@@ -0,0 +1,156 @@
|
|
|
1
|
+
|
|
2
|
+
|
|
3
|
+
import { getRuntimeConfig } from '../../../../config/runtime.js';
|
|
4
|
+
|
|
5
|
+
// ============================================================================
|
|
6
|
+
// Module State
|
|
7
|
+
// ============================================================================
|
|
8
|
+
|
|
9
|
+
|
|
10
|
+
function readDefaultConfig() {
|
|
11
|
+
const pipeline = getRuntimeConfig().shared.debug.pipeline;
|
|
12
|
+
return {
|
|
13
|
+
categories: {},
|
|
14
|
+
layers: pipeline.layers ?? null,
|
|
15
|
+
maxDecodeSteps: pipeline.maxDecodeSteps,
|
|
16
|
+
maxAbsThreshold: pipeline.maxAbsThreshold,
|
|
17
|
+
bufferStats: pipeline.bufferStats,
|
|
18
|
+
};
|
|
19
|
+
}
|
|
20
|
+
|
|
21
|
+
|
|
22
|
+
let config = readDefaultConfig();
|
|
23
|
+
let decodeStep = 0;
|
|
24
|
+
|
|
25
|
+
// ============================================================================
|
|
26
|
+
// Configuration API
|
|
27
|
+
// ============================================================================
|
|
28
|
+
|
|
29
|
+
|
|
30
|
+
export function setDebugCategories(categories, options) {
|
|
31
|
+
config = {
|
|
32
|
+
...config,
|
|
33
|
+
...options,
|
|
34
|
+
categories: { ...config.categories, ...categories },
|
|
35
|
+
};
|
|
36
|
+
}
|
|
37
|
+
|
|
38
|
+
|
|
39
|
+
export function resetDebugConfig() {
|
|
40
|
+
config = readDefaultConfig();
|
|
41
|
+
decodeStep = 0;
|
|
42
|
+
}
|
|
43
|
+
|
|
44
|
+
|
|
45
|
+
export function applyPipelineDebugConfig(pipeline) {
|
|
46
|
+
if (!pipeline) return;
|
|
47
|
+
|
|
48
|
+
const shouldEnable = pipeline.enabled || (pipeline.categories && pipeline.categories.length > 0);
|
|
49
|
+
if (!shouldEnable) {
|
|
50
|
+
resetDebugConfig();
|
|
51
|
+
return;
|
|
52
|
+
}
|
|
53
|
+
|
|
54
|
+
const categories = pipeline.categories && pipeline.categories.length > 0
|
|
55
|
+
? pipeline.categories
|
|
56
|
+
: ['all'];
|
|
57
|
+
|
|
58
|
+
|
|
59
|
+
const categoryMap = {};
|
|
60
|
+
if (categories.includes('all')) {
|
|
61
|
+
categoryMap.all = true;
|
|
62
|
+
} else {
|
|
63
|
+
for (const cat of categories) {
|
|
64
|
+
categoryMap[cat] = true;
|
|
65
|
+
}
|
|
66
|
+
}
|
|
67
|
+
|
|
68
|
+
setDebugCategories(categoryMap, {
|
|
69
|
+
layers: pipeline.layers ?? undefined,
|
|
70
|
+
maxDecodeSteps: pipeline.maxDecodeSteps ?? undefined,
|
|
71
|
+
maxAbsThreshold: pipeline.maxAbsThreshold ?? undefined,
|
|
72
|
+
bufferStats: pipeline.bufferStats ?? undefined,
|
|
73
|
+
});
|
|
74
|
+
}
|
|
75
|
+
|
|
76
|
+
|
|
77
|
+
export function getDebugConfig() {
|
|
78
|
+
return { ...config };
|
|
79
|
+
}
|
|
80
|
+
|
|
81
|
+
// ============================================================================
|
|
82
|
+
// Decode Step Tracking
|
|
83
|
+
// ============================================================================
|
|
84
|
+
|
|
85
|
+
|
|
86
|
+
export function incrementDecodeStep() {
|
|
87
|
+
return ++decodeStep;
|
|
88
|
+
}
|
|
89
|
+
|
|
90
|
+
|
|
91
|
+
export function resetDecodeStep() {
|
|
92
|
+
decodeStep = 0;
|
|
93
|
+
}
|
|
94
|
+
|
|
95
|
+
|
|
96
|
+
export function getDecodeStep() {
|
|
97
|
+
return decodeStep;
|
|
98
|
+
}
|
|
99
|
+
|
|
100
|
+
// ============================================================================
|
|
101
|
+
// Layer Filtering
|
|
102
|
+
// ============================================================================
|
|
103
|
+
|
|
104
|
+
|
|
105
|
+
export function shouldDebugLayerOutput(layerIdx, debugLayers) {
|
|
106
|
+
if (debugLayers === null) return false;
|
|
107
|
+
if (!Array.isArray(debugLayers) || debugLayers.length === 0) return true;
|
|
108
|
+
return debugLayers.includes(layerIdx);
|
|
109
|
+
}
|
|
110
|
+
|
|
111
|
+
// ============================================================================
|
|
112
|
+
// Internal Helpers (exported for use by other debug-utils modules)
|
|
113
|
+
// ============================================================================
|
|
114
|
+
|
|
115
|
+
|
|
116
|
+
export function isEnabled(category, layerIdx) {
|
|
117
|
+
// Check if category is enabled
|
|
118
|
+
if (!config.categories.all && !config.categories[category]) {
|
|
119
|
+
return false;
|
|
120
|
+
}
|
|
121
|
+
|
|
122
|
+
// Check layer filter
|
|
123
|
+
if (layerIdx !== undefined && config.layers?.length) {
|
|
124
|
+
if (!config.layers.includes(layerIdx)) {
|
|
125
|
+
return false;
|
|
126
|
+
}
|
|
127
|
+
}
|
|
128
|
+
|
|
129
|
+
// Check decode step limit
|
|
130
|
+
if (config.maxDecodeSteps && decodeStep > config.maxDecodeSteps) {
|
|
131
|
+
// Only apply to non-prefill logs
|
|
132
|
+
if (decodeStep > 0) {
|
|
133
|
+
return false;
|
|
134
|
+
}
|
|
135
|
+
}
|
|
136
|
+
|
|
137
|
+
return true;
|
|
138
|
+
}
|
|
139
|
+
|
|
140
|
+
|
|
141
|
+
export function formatTag(category, layerIdx, step) {
|
|
142
|
+
let tag = `[${category.toUpperCase()}]`;
|
|
143
|
+
if (layerIdx !== undefined) tag += `[L${layerIdx}]`;
|
|
144
|
+
if (step !== undefined) tag += `[S${step}]`;
|
|
145
|
+
return tag;
|
|
146
|
+
}
|
|
147
|
+
|
|
148
|
+
|
|
149
|
+
export function isBufferStatsEnabled() {
|
|
150
|
+
return config.bufferStats;
|
|
151
|
+
}
|
|
152
|
+
|
|
153
|
+
|
|
154
|
+
export function getMaxAbsThreshold() {
|
|
155
|
+
return config.maxAbsThreshold;
|
|
156
|
+
}
|
|
@@ -0,0 +1,53 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Debug utilities for pipeline tracing.
|
|
3
|
+
*
|
|
4
|
+
* Re-exports all debug utilities from submodules.
|
|
5
|
+
*
|
|
6
|
+
* @module inference/pipelines/text/debug-utils
|
|
7
|
+
*/
|
|
8
|
+
|
|
9
|
+
// Configuration
|
|
10
|
+
export {
|
|
11
|
+
type DebugCategory,
|
|
12
|
+
type DebugConfig,
|
|
13
|
+
setDebugCategories,
|
|
14
|
+
resetDebugConfig,
|
|
15
|
+
applyPipelineDebugConfig,
|
|
16
|
+
getDebugConfig,
|
|
17
|
+
incrementDecodeStep,
|
|
18
|
+
resetDecodeStep,
|
|
19
|
+
getDecodeStep,
|
|
20
|
+
shouldDebugLayerOutput,
|
|
21
|
+
} from './config.js';
|
|
22
|
+
|
|
23
|
+
// Logging
|
|
24
|
+
export {
|
|
25
|
+
logEmbed,
|
|
26
|
+
logLayer,
|
|
27
|
+
logAttn,
|
|
28
|
+
logFFN,
|
|
29
|
+
logKV,
|
|
30
|
+
logLogits,
|
|
31
|
+
logSample,
|
|
32
|
+
logIO,
|
|
33
|
+
logPerf,
|
|
34
|
+
} from './logging.js';
|
|
35
|
+
|
|
36
|
+
// Tensor inspection
|
|
37
|
+
export {
|
|
38
|
+
type TensorStats,
|
|
39
|
+
dumpTensor,
|
|
40
|
+
dumpTokenVector,
|
|
41
|
+
dumpKVCache,
|
|
42
|
+
logKernelStep,
|
|
43
|
+
isKernelDebugEnabled,
|
|
44
|
+
} from './tensor.js';
|
|
45
|
+
|
|
46
|
+
// Utilities
|
|
47
|
+
export {
|
|
48
|
+
f16ToF32,
|
|
49
|
+
decodeReadback,
|
|
50
|
+
getLogitsHealth,
|
|
51
|
+
getBufferStats,
|
|
52
|
+
DEBUG_PRESETS,
|
|
53
|
+
} from './utils.js';
|
|
@@ -0,0 +1,44 @@
|
|
|
1
|
+
|
|
2
|
+
|
|
3
|
+
// Configuration
|
|
4
|
+
export {
|
|
5
|
+
setDebugCategories,
|
|
6
|
+
resetDebugConfig,
|
|
7
|
+
applyPipelineDebugConfig,
|
|
8
|
+
getDebugConfig,
|
|
9
|
+
incrementDecodeStep,
|
|
10
|
+
resetDecodeStep,
|
|
11
|
+
getDecodeStep,
|
|
12
|
+
shouldDebugLayerOutput,
|
|
13
|
+
} from './config.js';
|
|
14
|
+
|
|
15
|
+
// Logging
|
|
16
|
+
export {
|
|
17
|
+
logEmbed,
|
|
18
|
+
logLayer,
|
|
19
|
+
logAttn,
|
|
20
|
+
logFFN,
|
|
21
|
+
logKV,
|
|
22
|
+
logLogits,
|
|
23
|
+
logSample,
|
|
24
|
+
logIO,
|
|
25
|
+
logPerf,
|
|
26
|
+
} from './logging.js';
|
|
27
|
+
|
|
28
|
+
// Tensor inspection
|
|
29
|
+
export {
|
|
30
|
+
dumpTensor,
|
|
31
|
+
dumpTokenVector,
|
|
32
|
+
dumpKVCache,
|
|
33
|
+
logKernelStep,
|
|
34
|
+
isKernelDebugEnabled,
|
|
35
|
+
} from './tensor.js';
|
|
36
|
+
|
|
37
|
+
// Utilities
|
|
38
|
+
export {
|
|
39
|
+
f16ToF32,
|
|
40
|
+
decodeReadback,
|
|
41
|
+
getLogitsHealth,
|
|
42
|
+
getBufferStats,
|
|
43
|
+
DEBUG_PRESETS,
|
|
44
|
+
} from './utils.js';
|
|
@@ -0,0 +1,106 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Logging functions for pipeline debug tracing.
|
|
3
|
+
*
|
|
4
|
+
* Provides category-specific logging for each pipeline stage:
|
|
5
|
+
* embedding, layers, attention, FFN, KV cache, logits, sampling, I/O, and performance.
|
|
6
|
+
*
|
|
7
|
+
* Log format: [CATEGORY] message
|
|
8
|
+
* This enables post-filtering: grep -E "^\[LAYER\]|\[ATTN\]"
|
|
9
|
+
*
|
|
10
|
+
* @module inference/pipelines/text/debug-utils/logging
|
|
11
|
+
*/
|
|
12
|
+
|
|
13
|
+
// ============================================================================
|
|
14
|
+
// Logging Functions
|
|
15
|
+
// ============================================================================
|
|
16
|
+
|
|
17
|
+
/**
|
|
18
|
+
* Log embedding info.
|
|
19
|
+
*/
|
|
20
|
+
export function logEmbed(
|
|
21
|
+
tokenIds: number[],
|
|
22
|
+
info: { maxAbs?: number; nonZero?: number; total?: number; sample?: number[] }
|
|
23
|
+
): void;
|
|
24
|
+
|
|
25
|
+
/**
|
|
26
|
+
* Log layer entry/exit.
|
|
27
|
+
*/
|
|
28
|
+
export function logLayer(
|
|
29
|
+
layerIdx: number,
|
|
30
|
+
phase: 'enter' | 'exit',
|
|
31
|
+
isPrefill: boolean,
|
|
32
|
+
info: { numTokens?: number; maxAbs?: number; sample?: number[] }
|
|
33
|
+
): void;
|
|
34
|
+
|
|
35
|
+
/**
|
|
36
|
+
* Log attention details.
|
|
37
|
+
*/
|
|
38
|
+
export function logAttn(
|
|
39
|
+
layerIdx: number,
|
|
40
|
+
isPrefill: boolean,
|
|
41
|
+
info: {
|
|
42
|
+
numTokens: number;
|
|
43
|
+
kvLen: number;
|
|
44
|
+
startPos?: number;
|
|
45
|
+
maxAbsQ?: number;
|
|
46
|
+
maxAbsK?: number;
|
|
47
|
+
maxAbsV?: number;
|
|
48
|
+
maxAbsOut?: number;
|
|
49
|
+
}
|
|
50
|
+
): void;
|
|
51
|
+
|
|
52
|
+
/**
|
|
53
|
+
* Log FFN details.
|
|
54
|
+
*/
|
|
55
|
+
export function logFFN(
|
|
56
|
+
layerIdx: number,
|
|
57
|
+
info: { maxAbsGate?: number; maxAbsUp?: number; maxAbsOut?: number }
|
|
58
|
+
): void;
|
|
59
|
+
|
|
60
|
+
/**
|
|
61
|
+
* Log KV cache operations.
|
|
62
|
+
*/
|
|
63
|
+
export function logKV(
|
|
64
|
+
layerIdx: number,
|
|
65
|
+
op: 'write' | 'read' | 'init' | 'clear',
|
|
66
|
+
info: { seqLen?: number; kvLen?: number; startPos?: number }
|
|
67
|
+
): void;
|
|
68
|
+
|
|
69
|
+
/**
|
|
70
|
+
* Log logits computation.
|
|
71
|
+
*/
|
|
72
|
+
export function logLogits(
|
|
73
|
+
phase: 'prefill' | 'decode',
|
|
74
|
+
info: {
|
|
75
|
+
min: number;
|
|
76
|
+
max: number;
|
|
77
|
+
topK?: Array<{ token: number | string; prob: number; text?: string }>;
|
|
78
|
+
}
|
|
79
|
+
): void;
|
|
80
|
+
|
|
81
|
+
/**
|
|
82
|
+
* Log sampling decision.
|
|
83
|
+
*/
|
|
84
|
+
export function logSample(
|
|
85
|
+
tokenId: number,
|
|
86
|
+
tokenText: string,
|
|
87
|
+
info: { prob?: number; temperature?: number; topK?: number }
|
|
88
|
+
): void;
|
|
89
|
+
|
|
90
|
+
/**
|
|
91
|
+
* Log GPU buffer I/O.
|
|
92
|
+
*/
|
|
93
|
+
export function logIO(
|
|
94
|
+
op: 'read' | 'write' | 'copy',
|
|
95
|
+
label: string,
|
|
96
|
+
bytes: number
|
|
97
|
+
): void;
|
|
98
|
+
|
|
99
|
+
/**
|
|
100
|
+
* Log performance timing.
|
|
101
|
+
*/
|
|
102
|
+
export function logPerf(
|
|
103
|
+
phase: string,
|
|
104
|
+
ms: number,
|
|
105
|
+
extra?: Record<string, number | string>
|
|
106
|
+
): void;
|
|
@@ -0,0 +1,152 @@
|
|
|
1
|
+
|
|
2
|
+
|
|
3
|
+
import { log } from '../../../../debug/index.js';
|
|
4
|
+
import {
|
|
5
|
+
isEnabled,
|
|
6
|
+
formatTag,
|
|
7
|
+
getDecodeStep,
|
|
8
|
+
getMaxAbsThreshold,
|
|
9
|
+
} from './config.js';
|
|
10
|
+
|
|
11
|
+
// ============================================================================
|
|
12
|
+
// Logging Functions
|
|
13
|
+
// ============================================================================
|
|
14
|
+
|
|
15
|
+
|
|
16
|
+
export function logEmbed(tokenIds, info) {
|
|
17
|
+
if (!isEnabled('embed')) return;
|
|
18
|
+
|
|
19
|
+
const tag = formatTag('embed');
|
|
20
|
+
const tokens = tokenIds.length > 3
|
|
21
|
+
? `[${tokenIds.slice(0, 3).join(',')},...] (${tokenIds.length} total)`
|
|
22
|
+
: `[${tokenIds.join(',')}]`;
|
|
23
|
+
|
|
24
|
+
let msg = `${tag} tokens=${tokens}`;
|
|
25
|
+
if (info.maxAbs !== undefined) msg += ` maxAbs=${info.maxAbs.toFixed(2)}`;
|
|
26
|
+
if (info.nonZero !== undefined) msg += ` nonZero=${info.nonZero}/${info.total}`;
|
|
27
|
+
if (info.sample?.length) msg += ` sample=[${info.sample.map(v => v.toFixed(3)).join(',')}]`;
|
|
28
|
+
|
|
29
|
+
log.debug('Debug', msg);
|
|
30
|
+
}
|
|
31
|
+
|
|
32
|
+
|
|
33
|
+
export function logLayer(layerIdx, phase, isPrefill, info) {
|
|
34
|
+
if (!isEnabled('layer', layerIdx)) return;
|
|
35
|
+
|
|
36
|
+
const tag = formatTag('layer', layerIdx);
|
|
37
|
+
const decodeStep = getDecodeStep();
|
|
38
|
+
const mode = isPrefill ? 'prefill' : `decode:${decodeStep}`;
|
|
39
|
+
|
|
40
|
+
let msg = `${tag} ${phase} ${mode}`;
|
|
41
|
+
if (info.numTokens !== undefined) msg += ` n=${info.numTokens}`;
|
|
42
|
+
if (info.maxAbs !== undefined) msg += ` maxAbs=${info.maxAbs.toFixed(2)}`;
|
|
43
|
+
if (info.sample?.length) msg += ` sample=[${info.sample.map(v => v.toFixed(3)).join(',')}]`;
|
|
44
|
+
|
|
45
|
+
// Warn on explosion
|
|
46
|
+
const maxAbsThreshold = getMaxAbsThreshold();
|
|
47
|
+
if (info.maxAbs !== undefined && info.maxAbs > maxAbsThreshold) {
|
|
48
|
+
msg += ` △ EXPLOSION`;
|
|
49
|
+
}
|
|
50
|
+
|
|
51
|
+
log.debug('Debug', msg);
|
|
52
|
+
}
|
|
53
|
+
|
|
54
|
+
|
|
55
|
+
export function logAttn(layerIdx, isPrefill, info) {
|
|
56
|
+
if (!isEnabled('attn', layerIdx)) return;
|
|
57
|
+
|
|
58
|
+
const tag = formatTag('attn', layerIdx);
|
|
59
|
+
const decodeStep = getDecodeStep();
|
|
60
|
+
const mode = isPrefill ? 'prefill' : `decode:${decodeStep}`;
|
|
61
|
+
|
|
62
|
+
let msg = `${tag} ${mode} n=${info.numTokens} kvLen=${info.kvLen}`;
|
|
63
|
+
if (info.startPos !== undefined) msg += ` startPos=${info.startPos}`;
|
|
64
|
+
if (info.maxAbsQ !== undefined) msg += ` Q=${info.maxAbsQ.toFixed(1)}`;
|
|
65
|
+
if (info.maxAbsK !== undefined) msg += ` K=${info.maxAbsK.toFixed(1)}`;
|
|
66
|
+
if (info.maxAbsV !== undefined) msg += ` V=${info.maxAbsV.toFixed(1)}`;
|
|
67
|
+
if (info.maxAbsOut !== undefined) msg += ` out=${info.maxAbsOut.toFixed(1)}`;
|
|
68
|
+
|
|
69
|
+
log.debug('Debug', msg);
|
|
70
|
+
}
|
|
71
|
+
|
|
72
|
+
|
|
73
|
+
export function logFFN(layerIdx, info) {
|
|
74
|
+
if (!isEnabled('ffn', layerIdx)) return;
|
|
75
|
+
|
|
76
|
+
const tag = formatTag('ffn', layerIdx);
|
|
77
|
+
let msg = tag;
|
|
78
|
+
if (info.maxAbsGate !== undefined) msg += ` gate=${info.maxAbsGate.toFixed(1)}`;
|
|
79
|
+
if (info.maxAbsUp !== undefined) msg += ` up=${info.maxAbsUp.toFixed(1)}`;
|
|
80
|
+
if (info.maxAbsOut !== undefined) msg += ` out=${info.maxAbsOut.toFixed(1)}`;
|
|
81
|
+
|
|
82
|
+
log.debug('Debug', msg);
|
|
83
|
+
}
|
|
84
|
+
|
|
85
|
+
|
|
86
|
+
export function logKV(layerIdx, op, info) {
|
|
87
|
+
if (!isEnabled('kv', layerIdx)) return;
|
|
88
|
+
|
|
89
|
+
const tag = formatTag('kv', layerIdx);
|
|
90
|
+
let msg = `${tag} ${op}`;
|
|
91
|
+
if (info.seqLen !== undefined) msg += ` seqLen=${info.seqLen}`;
|
|
92
|
+
if (info.kvLen !== undefined) msg += ` kvLen=${info.kvLen}`;
|
|
93
|
+
if (info.startPos !== undefined) msg += ` startPos=${info.startPos}`;
|
|
94
|
+
|
|
95
|
+
log.debug('Debug', msg);
|
|
96
|
+
}
|
|
97
|
+
|
|
98
|
+
|
|
99
|
+
export function logLogits(phase, info) {
|
|
100
|
+
if (!isEnabled('logits')) return;
|
|
101
|
+
|
|
102
|
+
const decodeStep = getDecodeStep();
|
|
103
|
+
const tag = phase === 'prefill' ? '[LOGITS][PREFILL]' : `[LOGITS][S${decodeStep}]`;
|
|
104
|
+
let msg = `${tag} min=${info.min.toFixed(2)} max=${info.max.toFixed(2)}`;
|
|
105
|
+
|
|
106
|
+
if (info.topK?.length) {
|
|
107
|
+
const topStr = info.topK
|
|
108
|
+
.slice(0, 5)
|
|
109
|
+
.map(t => `"${t.text || t.token}"(${(t.prob * 100).toFixed(1)}%)`)
|
|
110
|
+
.join(', ');
|
|
111
|
+
msg += ` top-5: ${topStr}`;
|
|
112
|
+
}
|
|
113
|
+
|
|
114
|
+
log.debug('Debug', msg);
|
|
115
|
+
}
|
|
116
|
+
|
|
117
|
+
|
|
118
|
+
export function logSample(tokenId, tokenText, info) {
|
|
119
|
+
if (!isEnabled('sample')) return;
|
|
120
|
+
|
|
121
|
+
const decodeStep = getDecodeStep();
|
|
122
|
+
const tag = decodeStep === 0 ? '[SAMPLE][PREFILL]' : `[SAMPLE][S${decodeStep}]`;
|
|
123
|
+
let msg = `${tag} -> ${tokenId} "${tokenText}"`;
|
|
124
|
+
if (info.prob !== undefined) msg += ` p=${(info.prob * 100).toFixed(1)}%`;
|
|
125
|
+
if (info.temperature !== undefined) msg += ` T=${info.temperature}`;
|
|
126
|
+
|
|
127
|
+
log.debug('Debug', msg);
|
|
128
|
+
}
|
|
129
|
+
|
|
130
|
+
|
|
131
|
+
export function logIO(op, label, bytes) {
|
|
132
|
+
if (!isEnabled('io')) return;
|
|
133
|
+
|
|
134
|
+
const tag = '[IO]';
|
|
135
|
+
const kb = (bytes / 1024).toFixed(1);
|
|
136
|
+
log.debug('Debug', `${tag} ${op} ${label}: ${kb}KB`);
|
|
137
|
+
}
|
|
138
|
+
|
|
139
|
+
|
|
140
|
+
export function logPerf(phase, ms, extra) {
|
|
141
|
+
if (!isEnabled('perf')) return;
|
|
142
|
+
|
|
143
|
+
const tag = '[PERF]';
|
|
144
|
+
let msg = `${tag} ${phase}: ${ms.toFixed(1)}ms`;
|
|
145
|
+
if (extra) {
|
|
146
|
+
for (const [k, v] of Object.entries(extra)) {
|
|
147
|
+
msg += ` ${k}=${typeof v === 'number' ? v.toFixed(1) : v}`;
|
|
148
|
+
}
|
|
149
|
+
}
|
|
150
|
+
|
|
151
|
+
log.debug('Debug', msg);
|
|
152
|
+
}
|