doppler-gpu 0.2.0 → 0.2.1
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 +21 -1
- package/README.md +97 -29
- package/package.json +40 -27
- package/src/cli/config/doppler-quickstart.json +10 -0
- package/src/cli/doppler-cli.js +66 -22
- package/src/cli/doppler-quickstart.d.ts +61 -0
- package/src/cli/doppler-quickstart.js +299 -0
- package/src/cli/doppler-serve.js +395 -0
- package/src/client/doppler-api.browser.js +13 -377
- package/src/client/doppler-api.d.ts +35 -56
- package/src/client/doppler-api.js +12 -363
- package/src/client/doppler-registry.json +55 -2
- package/src/client/runtime/index.d.ts +63 -0
- package/src/client/runtime/index.js +173 -0
- package/src/client/runtime/lora.d.ts +36 -0
- package/src/client/runtime/lora.js +165 -0
- package/src/client/{doppler-provider → runtime}/model-manager.d.ts +1 -1
- package/src/client/{doppler-provider → runtime}/model-manager.js +56 -30
- package/src/client/runtime/model-session.d.ts +72 -0
- package/src/client/runtime/model-session.js +122 -0
- package/src/client/runtime/model-source.d.ts +43 -0
- package/src/client/runtime/model-source.js +101 -0
- package/src/client/{doppler-provider → runtime}/source-runtime.d.ts +8 -5
- package/src/client/runtime/source-runtime.js +919 -0
- package/src/client/runtime/storage.d.ts +5 -0
- package/src/client/runtime/storage.js +51 -0
- package/src/client/{doppler-provider → runtime}/types.d.ts +14 -2
- package/src/config/README.md +2 -0
- package/src/config/conversion/README.md +47 -7
- package/src/config/conversion/embeddinggemma/google-embeddinggemma-300m-q4k-ehf16-af32.json +67 -5
- package/src/config/conversion/gemma3/gemma-3-1b-it-f16-af32.json +62 -11
- package/src/config/conversion/gemma3/gemma-3-1b-it-q4k-ehf16-af32.json +79 -18
- package/src/config/conversion/gemma3/gemma-3-270m-it-q4k-ehf16-af32.json +61 -13
- package/src/config/conversion/gemma3/translategemma-4b-1b-enes-q4k-ehf16-af32.json +53 -10
- package/src/config/conversion/gemma3/translategemma-4b-it-q4k-ehf16-af32.json +66 -18
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32.json +256 -0
- package/src/config/conversion/gemma4/gemma-4-moe-q4k-ehf16-af32.json +49 -8
- package/src/config/conversion/gpt-oss-20b-f16-xmxfp4.json +45 -6
- package/src/config/conversion/janus/janus-pro-1b-text-q4k-ehaf16.json +45 -6
- package/src/config/conversion/lfm2/lfm2.5-1.2b-instruct-q4k-ehf16-af32.json +57 -9
- package/src/config/conversion/qwen3/qwen-3-5-0-8b-q4k-ehaf16.json +105 -26
- package/src/config/conversion/qwen3/qwen-3-5-2b-q4k-ehaf16.json +83 -33
- package/src/config/conversion/sana/sana-sprint-0.6b-f16.json +33 -5
- package/src/config/execution-contract-check.d.ts +2 -1
- package/src/config/execution-contract-check.js +187 -53
- package/src/config/kernel-path-loader.d.ts +36 -0
- package/src/config/kernel-path-loader.js +97 -10
- package/src/config/kernels/kernel-ref-digests.js +40 -24
- package/src/config/kernels/registry.js +0 -1
- package/src/config/kernels/registry.json +772 -66
- package/src/config/merge-contract-check.js +3 -0
- package/src/config/merge.d.ts +18 -1
- package/src/config/merge.js +56 -0
- package/src/config/param-categories.js +5 -0
- package/src/config/param-validator.js +42 -26
- package/src/config/platforms/apple-m3.json +1 -1
- package/src/config/platforms/loader.js +21 -16
- package/src/config/required-inference-fields-contract-check.js +16 -0
- package/src/config/runtime/compute/f16-batched.json +2 -0
- package/src/config/runtime/default.json +1 -0
- package/src/config/runtime/experiments/bench/gemma3-bench-q4k-alt.json +57 -0
- package/src/config/runtime/experiments/bench/gemma3-bench-q4k-batched.json +57 -0
- package/src/config/runtime/experiments/bench/gemma3-bench-q4k.json +2 -0
- package/src/config/runtime/experiments/bench/qwen3-bench-q4k-alt.json +57 -0
- package/src/config/runtime/experiments/bench/qwen3-bench-q4k-batched.json +57 -0
- package/src/config/runtime/experiments/debug/gemma3-browser-diagnose-q4k.json +81 -0
- package/src/config/runtime/experiments/debug/gemma3-debug-q4k.json +5 -0
- package/src/config/runtime/experiments/debug/qwen35-browser-diagnose-q4k.json +81 -0
- package/src/config/runtime/experiments/verify/gemma4-e2b-verify.json +38 -0
- package/src/config/runtime/profiles/gemma4-e2b-low-memory.json +63 -0
- package/src/config/runtime/profiles/gemma4-e2b-throughput.json +39 -0
- package/src/config/runtime/profiles/low-memory.json +1 -0
- package/src/config/runtime/profiles/turboquant-contiguous-prod.json +29 -0
- package/src/config/runtime/profiles/turboquant-contiguous.json +29 -0
- package/src/config/runtime/profiles/turboquant.json +36 -0
- package/src/config/runtime/profiles/vector-stability.json +1 -0
- package/src/config/runtime/tiers/gemma4-128gb.json +79 -0
- package/src/config/runtime/tiers/gemma4-16gb.json +15 -2
- package/src/config/runtime/tiers/gemma4-24gb.json +15 -2
- package/src/config/runtime/tiers/gemma4-32gb.json +15 -2
- package/src/config/runtime.js +11 -4
- package/src/config/schema/converter.schema.d.ts +4 -2
- package/src/config/schema/converter.schema.js +7 -5
- package/src/config/schema/doppler.schema.d.ts +26 -2
- package/src/config/schema/doppler.schema.js +77 -28
- package/src/config/schema/execution-v1.schema.d.ts +78 -5
- package/src/config/schema/execution-v1.schema.js +62 -1
- package/src/config/schema/index.d.ts +22 -29
- package/src/config/schema/index.js +15 -17
- package/src/config/schema/inference.schema.d.ts +26 -4
- package/src/config/schema/kernel-path.schema.d.ts +30 -3
- package/src/config/schema/kernel-registry.schema.d.ts +1 -1
- package/src/config/schema/kernel-registry.schema.js +1 -0
- package/src/config/schema/kvcache.schema.js +7 -0
- package/src/config/schema/loading.schema.d.ts +28 -0
- package/src/config/schema/loading.schema.js +9 -0
- package/src/config/schema/manifest.schema.d.ts +79 -5
- package/src/config/schema/manifest.schema.js +7 -1
- package/src/config/schema/platform.schema.d.ts +10 -8
- package/src/config/schema/speculation-self.schema.js +0 -5
- package/src/config/schema/support-tiers.schema.d.ts +53 -0
- package/src/config/schema/support-tiers.schema.js +128 -0
- package/src/config/source-packages/litert/gemma-4-e2b-it.json +622 -0
- package/src/config/source-packages/registry.json +15 -0
- package/src/config/support-tiers/subsystems.json +502 -0
- package/src/config/transforms/capability-transform-resolver.d.ts +16 -5
- package/src/config/transforms/capability-transform-resolver.js +44 -9
- package/src/config/transforms/execution-graph-transforms.d.ts +113 -2
- package/src/config/transforms/execution-graph-transforms.js +1220 -8
- package/src/converter/conversion-plan.js +20 -8
- package/src/converter/core.d.ts +48 -1
- package/src/converter/core.js +249 -62
- package/src/converter/parsers/gguf.d.ts +10 -1
- package/src/converter/rope-config.d.ts +1 -0
- package/src/converter/rope-config.js +75 -5
- package/src/converter/shard-packer.js +2 -2
- package/src/debug/tensor.js +13 -1
- package/src/{adapters → experimental/adapters}/adapter-manager.d.ts +1 -1
- package/src/{adapters → experimental/adapters}/adapter-manager.js +2 -2
- package/src/{adapters → experimental/adapters}/adapter-manifest.d.ts +1 -1
- package/src/{adapters → experimental/adapters}/adapter-manifest.js +1 -1
- package/src/{adapters → experimental/adapters}/adapter-registry.d.ts +1 -1
- package/src/{adapters → experimental/adapters}/adapter-registry.js +1 -1
- package/src/{adapters → experimental/adapters}/lora-loader.d.ts +1 -1
- package/src/{adapters → experimental/adapters}/lora-loader.js +2 -2
- package/src/{bridge → experimental/bridge}/extension-client.js +2 -2
- package/src/{bridge → experimental/bridge}/node-bridge.js +2 -4
- package/src/{bridge → experimental/bridge}/protocol.js +1 -1
- package/src/{browser → experimental/browser}/browser-converter.d.ts +4 -4
- package/src/{browser → experimental/browser}/browser-converter.js +12 -12
- package/src/{browser → experimental/browser}/gguf-importer.d.ts +1 -1
- package/src/{browser → experimental/browser}/gguf-importer.js +8 -8
- package/src/{browser → experimental/browser}/gguf-parser-browser.d.ts +2 -2
- package/src/{browser → experimental/browser}/gguf-parser-browser.js +3 -3
- package/src/{browser → experimental/browser}/quantization.d.ts +3 -3
- package/src/{browser → experimental/browser}/quantization.js +4 -4
- package/src/{browser → experimental/browser}/safetensors-parser-browser.d.ts +3 -3
- package/src/{browser → experimental/browser}/safetensors-parser-browser.js +3 -3
- package/src/{browser → experimental/browser}/shard-io-browser.d.ts +1 -1
- package/src/{browser → experimental/browser}/shard-io-browser.js +2 -2
- package/src/{browser → experimental/browser}/tensor-source-download.js +3 -3
- package/src/experimental/diffusion/index.d.ts +8 -0
- package/src/experimental/diffusion/index.js +8 -0
- package/src/{distribution → experimental/distribution}/p2p-transport-contract.js +1 -1
- package/src/{distribution → experimental/distribution}/p2p-transport.js +1 -1
- package/src/{distribution → experimental/distribution}/shard-delivery.js +5 -5
- package/src/experimental/energy/index.d.ts +2 -0
- package/src/experimental/energy/index.js +2 -0
- package/src/{hotswap → experimental/hotswap}/intent-bundle.d.ts +1 -1
- package/src/{hotswap → experimental/hotswap}/intent-bundle.js +2 -2
- package/src/{hotswap → experimental/hotswap}/manifest.d.ts +2 -2
- package/src/{hotswap → experimental/hotswap}/manifest.js +1 -1
- package/src/{hotswap → experimental/hotswap}/model-swap.js +1 -1
- package/src/{hotswap → experimental/hotswap}/runtime.d.ts +1 -1
- package/src/experimental/orchestration/index.d.ts +12 -0
- package/src/experimental/orchestration/index.js +12 -0
- package/src/{training → experimental/training}/README.md +1 -1
- package/src/{training → experimental/training}/artifacts.d.ts +3 -3
- package/src/{training → experimental/training}/artifacts.js +3 -3
- package/src/{training → experimental/training}/attention-backward.d.ts +1 -1
- package/src/{training → experimental/training}/attention-backward.js +3 -3
- package/src/{training → experimental/training}/attention-forward.d.ts +2 -2
- package/src/{training → experimental/training}/attention-forward.js +1 -1
- package/src/{training → experimental/training}/autograd.d.ts +2 -2
- package/src/{training → experimental/training}/autograd.js +7 -7
- package/src/{training → experimental/training}/backward-kernels.js +1 -1
- package/src/{training → experimental/training}/checkpoint.js +1 -1
- package/src/{training → experimental/training}/clip.d.ts +2 -2
- package/src/{training → experimental/training}/clip.js +3 -3
- package/src/{training → experimental/training}/export.d.ts +2 -2
- package/src/{training → experimental/training}/export.js +2 -2
- package/src/{training → experimental/training}/lora-pipeline.js +4 -4
- package/src/{training → experimental/training}/lora.d.ts +1 -1
- package/src/{training → experimental/training}/lora.js +4 -4
- package/src/{training → experimental/training}/loss-scaling.d.ts +2 -2
- package/src/{training → experimental/training}/loss-scaling.js +3 -3
- package/src/{training → experimental/training}/loss.d.ts +2 -2
- package/src/{training → experimental/training}/loss.js +1 -1
- package/src/{training → experimental/training}/operator-artifacts.js +1 -1
- package/src/{training → experimental/training}/optimizer.d.ts +2 -2
- package/src/{training → experimental/training}/optimizer.js +3 -3
- package/src/{training → experimental/training}/runner.d.ts +2 -2
- package/src/{training → experimental/training}/runner.js +5 -5
- package/src/{training → experimental/training}/suite.js +16 -16
- package/src/{training → experimental/training}/tensor-factory.d.ts +1 -1
- package/src/{training → experimental/training}/tensor-factory.js +2 -2
- package/src/{training → experimental/training}/trainer.d.ts +3 -3
- package/src/{training → experimental/training}/trainer.js +3 -3
- package/src/{training → experimental/training}/ul_dataset.d.ts +2 -2
- package/src/{training → experimental/training}/ul_dataset.js +1 -1
- package/src/{training → experimental/training}/ul_schedule.d.ts +1 -1
- package/src/{training → experimental/training}/workloads.js +4 -4
- package/src/formats/index.d.ts +24 -0
- package/src/formats/index.js +6 -0
- package/src/formats/litert/index.d.ts +1 -0
- package/src/formats/litert/index.js +1 -0
- package/src/formats/litert/types.d.ts +87 -0
- package/src/formats/litert/types.js +568 -0
- package/src/formats/rdrr/classification.js +16 -8
- package/src/formats/rdrr/parsing.d.ts +3 -2
- package/src/formats/rdrr/types.d.ts +33 -0
- package/src/formats/rdrr/validation.js +16 -5
- package/src/formats/tflite/index.d.ts +1 -0
- package/src/formats/tflite/index.js +1 -0
- package/src/formats/tflite/types.d.ts +84 -0
- package/src/formats/tflite/types.js +633 -0
- package/src/generation/index.d.ts +16 -14
- package/src/generation/index.js +1 -32
- package/src/gpu/command-recorder.d.ts +17 -1
- package/src/gpu/command-recorder.js +47 -10
- package/src/gpu/device.d.ts +3 -1
- package/src/gpu/device.js +17 -11
- package/src/gpu/kernel-selector.js +3 -1
- package/src/gpu/kernels/attention.d.ts +2 -0
- package/src/gpu/kernels/attention.js +359 -14
- package/src/gpu/kernels/attention.wgsl +20 -1
- package/src/gpu/kernels/attention_decode_contiguous_turboquant_f16kv.wgsl +180 -0
- package/src/gpu/kernels/attention_decode_contiguous_turboquant_prod_f16kv.wgsl +192 -0
- package/src/gpu/kernels/attention_decode_online_f16.wgsl +36 -18
- package/src/gpu/kernels/attention_decode_online_f16kv.wgsl +36 -18
- package/src/gpu/kernels/attention_decode_tiered_turboquant_f16kv.wgsl +269 -0
- package/src/gpu/kernels/attention_decode_tiered_turboquant_prod_f16kv.wgsl +281 -0
- package/src/gpu/kernels/attention_head256_f16kv.wgsl +269 -0
- package/src/gpu/kernels/attention_streaming.wgsl +22 -3
- package/src/gpu/kernels/attention_streaming_f16.wgsl +1 -1
- package/src/gpu/kernels/attention_streaming_f16kv.wgsl +1 -1
- package/src/gpu/kernels/backward/attention_backward.d.ts +14 -1
- package/src/gpu/kernels/check-hot-vocab-stop.d.ts +18 -0
- package/src/gpu/kernels/check-hot-vocab-stop.js +108 -0
- package/src/gpu/kernels/codegen/patches/attention_streaming_f16.from.attention_streaming.diff +1 -1
- package/src/gpu/kernels/codegen/patches/attention_streaming_f16kv.from.attention_streaming.diff +1 -1
- package/src/gpu/kernels/codegen/patches/rope_f16.from.rope.diff +126 -194
- package/src/gpu/kernels/codegen/wgsl-variants.js +24 -6
- package/src/gpu/kernels/depthwise_conv1d.d.ts +23 -0
- package/src/gpu/kernels/depthwise_conv1d.js +51 -0
- package/src/gpu/kernels/depthwise_conv1d.wgsl +46 -0
- package/src/gpu/kernels/feature-check.js +0 -1
- package/src/gpu/kernels/fused_ffn.js +67 -22
- package/src/gpu/kernels/fused_ffn_q4k.wgsl +48 -43
- package/src/gpu/kernels/fused_ffn_q4k_f16.wgsl +329 -0
- package/src/gpu/kernels/fused_matmul_q4.wgsl +417 -0
- package/src/gpu/kernels/fused_matmul_q4_batched_multicol_shared.wgsl +174 -0
- package/src/gpu/kernels/gather.d.ts +4 -0
- package/src/gpu/kernels/gather.js +13 -2
- package/src/gpu/kernels/gather.wgsl +5 -3
- package/src/gpu/kernels/gather_f16.wgsl +5 -3
- package/src/gpu/kernels/index.d.ts +7 -0
- package/src/gpu/kernels/index.js +6 -0
- package/src/gpu/kernels/kernel-configs.js +1 -0
- package/src/gpu/kernels/kv-quantize.js +87 -55
- package/src/gpu/kernels/kv_quantize_turboquant.wgsl +153 -0
- package/src/gpu/kernels/kv_quantize_turboquant_prod.wgsl +230 -0
- package/src/gpu/kernels/linear-attention-core.d.ts +3 -0
- package/src/gpu/kernels/linear-attention-core.js +145 -56
- package/src/gpu/kernels/matmul-dispatch.js +3 -0
- package/src/gpu/kernels/matmul-selection.js +31 -42
- package/src/gpu/kernels/matmul.d.ts +3 -0
- package/src/gpu/kernels/matmul.js +72 -34
- package/src/gpu/kernels/residual.d.ts +2 -0
- package/src/gpu/kernels/residual.js +43 -4
- package/src/gpu/kernels/rope.d.ts +2 -0
- package/src/gpu/kernels/rope.js +19 -6
- package/src/gpu/kernels/rope.wgsl +24 -12
- package/src/gpu/kernels/rope_f16.wgsl +97 -38
- package/src/gpu/kernels/scale.d.ts +5 -0
- package/src/gpu/kernels/scale.js +6 -1
- package/src/gpu/kernels/scale.wgsl +3 -1
- package/src/gpu/kernels/scale_f16.wgsl +32 -0
- package/src/gpu/kernels/turboquant-codebook.js +564 -0
- package/src/gpu/kernels/uniform-utils.js +26 -0
- package/src/gpu/kernels/utils.js +6 -1
- package/src/gpu/submit-probe.d.ts +12 -0
- package/src/gpu/submit-probe.js +77 -0
- package/src/gpu/weight-buffer.d.ts +26 -3
- package/src/gpu/weight-buffer.js +62 -2
- package/src/index-browser.d.ts +1 -91
- package/src/index-browser.js +0 -66
- package/src/index.d.ts +1 -102
- package/src/index.js +0 -74
- package/src/inference/browser-harness-model-helpers.d.ts +3 -0
- package/src/inference/browser-harness-model-helpers.js +134 -11
- package/src/inference/browser-harness-runtime-helpers.js +1 -9
- package/src/inference/browser-harness-text-helpers.js +441 -70
- package/src/inference/browser-harness.js +61 -4
- package/src/inference/decode-ring.d.ts +2 -0
- package/src/inference/decode-ring.js +20 -0
- package/src/inference/kv-cache/index.d.ts +3 -0
- package/src/inference/kv-cache/index.js +2 -0
- package/src/inference/kv-cache/mixed-geometry.d.ts +84 -0
- package/src/inference/kv-cache/mixed-geometry.js +422 -0
- package/src/inference/kv-cache/quantized.d.ts +56 -0
- package/src/inference/kv-cache/quantized.js +383 -0
- package/src/inference/kv-cache/tiered.d.ts +20 -3
- package/src/inference/kv-cache/tiered.js +203 -65
- package/src/inference/kv-cache/types.d.ts +49 -6
- package/src/inference/kv-cache.d.ts +3 -0
- package/src/inference/kv-cache.js +2 -0
- package/src/inference/multi-model-network.js +2 -1
- package/src/inference/pipelines/audio/gemma4.js +655 -0
- package/src/inference/pipelines/audio/index.js +39 -0
- package/src/inference/pipelines/audio/mel.js +171 -0
- package/src/inference/pipelines/context.js +9 -1
- package/src/inference/pipelines/diffusion/text-encoder-gpu.js +2 -0
- package/src/inference/pipelines/shared/clipped-linear.js +46 -0
- package/src/inference/pipelines/text/attention/dispatch-params.d.ts +2 -1
- package/src/inference/pipelines/text/attention/dispatch-params.js +52 -8
- package/src/inference/pipelines/text/attention/output-projection.js +3 -3
- package/src/inference/pipelines/text/attention/precision-contract.d.ts +36 -0
- package/src/inference/pipelines/text/attention/precision-contract.js +115 -0
- package/src/inference/pipelines/text/attention/projections.d.ts +23 -0
- package/src/inference/pipelines/text/attention/projections.js +182 -15
- package/src/inference/pipelines/text/attention/quant-options.js +54 -0
- package/src/inference/pipelines/text/attention/record.js +257 -69
- package/src/inference/pipelines/text/attention/run.js +306 -94
- package/src/inference/pipelines/text/attention/types.d.ts +11 -0
- package/src/inference/pipelines/text/chat-format.d.ts +8 -1
- package/src/inference/pipelines/text/chat-format.js +89 -5
- package/src/inference/pipelines/text/config.d.ts +78 -26
- package/src/inference/pipelines/text/config.js +722 -28
- package/src/inference/pipelines/text/dtype-contract.d.ts +15 -0
- package/src/inference/pipelines/text/dtype-contract.js +38 -0
- package/src/inference/pipelines/text/embed.d.ts +4 -0
- package/src/inference/pipelines/text/embed.js +165 -172
- package/src/inference/pipelines/text/execution-plan.d.ts +19 -0
- package/src/inference/pipelines/text/execution-plan.js +112 -20
- package/src/inference/pipelines/text/execution-runtime-builders.d.ts +8 -4
- package/src/inference/pipelines/text/execution-runtime-builders.js +90 -49
- package/src/inference/pipelines/text/execution-v1.d.ts +23 -4
- package/src/inference/pipelines/text/execution-v1.js +411 -33
- package/src/inference/pipelines/text/ffn/dense.d.ts +59 -0
- package/src/inference/pipelines/text/ffn/dense.js +575 -63
- package/src/inference/pipelines/text/ffn/moe.js +1 -0
- package/src/inference/pipelines/text/ffn/sandwich.js +20 -4
- package/src/inference/pipelines/text/ffn/standard.js +28 -4
- package/src/inference/pipelines/text/finiteness-policy.d.ts +4 -0
- package/src/inference/pipelines/text/finiteness-policy.js +14 -11
- package/src/inference/pipelines/text/generator-helpers.d.ts +7 -0
- package/src/inference/pipelines/text/generator-helpers.js +39 -4
- package/src/inference/pipelines/text/generator-runtime.js +45 -8
- package/src/inference/pipelines/text/generator-steps.js +407 -75
- package/src/inference/pipelines/text/generator.d.ts +6 -0
- package/src/inference/pipelines/text/generator.js +1179 -305
- package/src/inference/pipelines/text/init.d.ts +39 -7
- package/src/inference/pipelines/text/init.js +525 -110
- package/src/inference/pipelines/text/layer-plan.d.ts +1 -0
- package/src/inference/pipelines/text/layer-plan.js +2 -0
- package/src/inference/pipelines/text/layer.js +402 -19
- package/src/inference/pipelines/text/linear-attention.d.ts +4 -0
- package/src/inference/pipelines/text/linear-attention.js +112 -41
- package/src/inference/pipelines/text/logits/gpu.d.ts +2 -1
- package/src/inference/pipelines/text/logits/gpu.js +334 -15
- package/src/inference/pipelines/text/logits/index.d.ts +11 -0
- package/src/inference/pipelines/text/logits/index.js +227 -29
- package/src/inference/pipelines/text/logits/types.d.ts +3 -0
- package/src/inference/pipelines/text/logits/utils.js +3 -1
- package/src/inference/pipelines/text/lora-apply.js +3 -3
- package/src/inference/pipelines/text/model-load.d.ts +2 -1
- package/src/inference/pipelines/text/model-load.js +205 -178
- package/src/inference/pipelines/text/moe-gpu.js +8 -0
- package/src/inference/pipelines/text/moe-impl.d.ts +2 -0
- package/src/inference/pipelines/text/ops.d.ts +16 -2
- package/src/inference/pipelines/text/ops.js +11 -2
- package/src/inference/pipelines/text/per-layer-inputs.d.ts +107 -0
- package/src/inference/pipelines/text/per-layer-inputs.js +1581 -0
- package/src/inference/pipelines/text/probes.js +6 -0
- package/src/inference/pipelines/text/sampling-config.js +9 -1
- package/src/inference/pipelines/text/stage-names.js +3 -0
- package/src/inference/pipelines/text/state.d.ts +12 -3
- package/src/inference/pipelines/text/state.js +19 -0
- package/src/inference/pipelines/text/types.d.ts +79 -7
- package/src/inference/pipelines/text/weights.d.ts +2 -2
- package/src/inference/pipelines/text/weights.js +4 -0
- package/src/inference/pipelines/text.d.ts +6 -0
- package/src/inference/pipelines/text.js +860 -11
- package/src/inference/pipelines/video/frame-sampler.js +36 -0
- package/src/inference/pipelines/video/index.js +87 -0
- package/src/inference/pipelines/vision/encoder.js +27 -6
- package/src/inference/pipelines/vision/gemma4.js +731 -0
- package/src/inference/pipelines/vision/image-preprocess.js +22 -7
- package/src/inference/pipelines/vision/index.js +35 -10
- package/src/inference/pipelines/vision/ops.js +31 -26
- package/src/inference/pipelines/vision/patch-embed.js +22 -10
- package/src/inference/runtime-model.d.ts +89 -0
- package/src/inference/runtime-model.js +152 -0
- package/src/inference/test-harness.d.ts +8 -1
- package/src/inference/test-harness.js +54 -20
- package/src/inference/tokenizer.d.ts +5 -0
- package/src/inference/tokenizer.js +43 -0
- package/src/inference/tokenizers/base.d.ts +5 -0
- package/src/inference/tokenizers/base.js +5 -0
- package/src/inference/tokenizers/behavior-flags.d.ts +9 -0
- package/src/inference/tokenizers/behavior-flags.js +100 -0
- package/src/inference/tokenizers/bundled.d.ts +4 -0
- package/src/inference/tokenizers/bundled.js +110 -0
- package/src/inference/tokenizers/sentencepiece.d.ts +2 -0
- package/src/inference/tokenizers/sentencepiece.js +16 -0
- package/src/inference/tokenizers/types.d.ts +2 -0
- package/src/loader/doppler-loader.d.ts +10 -3
- package/src/loader/doppler-loader.js +103 -5
- package/src/loader/embedding-loader.d.ts +5 -1
- package/src/loader/embedding-loader.js +57 -6
- package/src/loader/final-weights-loader.js +44 -5
- package/src/loader/layer-loader.js +210 -8
- package/src/loader/loader-state.js +2 -2
- package/src/loader/loader-types.d.ts +17 -1
- package/src/loader/memory-monitor.d.ts +8 -0
- package/src/loader/memory-monitor.js +21 -0
- package/src/loader/multi-model-loader.d.ts +9 -1
- package/src/loader/multi-model-loader.js +9 -1
- package/src/loader/per-layer-input-loader.d.ts +30 -0
- package/src/loader/per-layer-input-loader.js +478 -0
- package/src/loader/shard-cache.d.ts +4 -3
- package/src/loader/shard-resolver.js +2 -0
- package/src/loader/tensors/source-transform.d.ts +50 -0
- package/src/loader/tensors/source-transform.js +776 -0
- package/src/loader/tensors/tensor-loader.d.ts +12 -0
- package/src/loader/tensors/tensor-loader.js +208 -123
- package/src/loader/tensors/tensor-reader.d.ts +8 -0
- package/src/loader/tensors/tensor-reader.js +283 -15
- package/src/loaders/index.d.ts +4 -0
- package/src/loaders/index.js +4 -0
- package/src/memory/buffer-pool.d.ts +4 -0
- package/src/memory/buffer-pool.js +36 -0
- package/src/rules/execution-rules-contract-check.d.ts +4 -0
- package/src/rules/execution-rules-contract-check.js +150 -24
- package/src/rules/inference/attention.rules.json +15 -0
- package/src/rules/inference/capability-transforms.rules.json +30 -0
- package/src/rules/inference/execution.rules.json +54 -2
- package/src/rules/inference/ffn.rules.json +33 -0
- package/src/rules/kernels/attention.rules.json +7 -0
- package/src/rules/kernels/fused-ffn.rules.json +4 -2
- package/src/rules/kernels/kv_quantize.rules.json +2 -0
- package/src/rules/kernels/matmul.rules.json +14 -0
- package/src/rules/kernels/scale.rules.json +2 -0
- package/src/rules/loader/tensor-loader.rules.json +1 -0
- package/src/storage/artifact-storage-context.d.ts +53 -0
- package/src/storage/artifact-storage-context.js +560 -0
- package/src/storage/downloader.js +7 -1
- package/src/storage/index.d.ts +18 -0
- package/src/storage/index.js +13 -0
- package/src/storage/preflight.js +22 -0
- package/src/storage/quickstart-downloader.js +35 -2
- package/src/storage/registry.d.ts +1 -0
- package/src/storage/registry.js +26 -4
- package/src/storage/shard-manager.d.ts +4 -0
- package/src/storage/shard-manager.js +25 -6
- package/src/storage/source-artifact-store.d.ts +9 -0
- package/src/storage/source-artifact-store.js +134 -0
- package/src/tooling/browser-command-runner.js +12 -5
- package/src/tooling/command-api-family-normalizers.js +144 -0
- package/src/tooling/command-api-helpers.js +1 -0
- package/src/tooling/command-api.d.ts +32 -9
- package/src/tooling/command-api.js +0 -25
- package/src/tooling/command-runner-shared.js +1 -2
- package/src/tooling/command-runner.html +5 -2
- package/src/tooling/diagnose-runner.js +1 -1
- package/src/tooling/kernel-path-builder/index.js +48 -32
- package/src/tooling/lean-execution-contract.d.ts +1 -1
- package/src/tooling/lean-execution-contract.js +2 -149
- package/src/tooling/litert-package-runtime.d.ts +46 -0
- package/src/tooling/litert-package-runtime.js +950 -0
- package/src/tooling/node-browser-command-runner.d.ts +13 -0
- package/src/tooling/node-browser-command-runner.js +163 -3
- package/src/tooling/node-command-runner.js +14 -5
- package/src/tooling/node-converter.js +226 -63
- package/src/tooling/node-source-runtime.d.ts +8 -3
- package/src/tooling/node-source-runtime.js +561 -236
- package/src/tooling/node-webgpu.js +7 -30
- package/src/tooling/opfs-cache.d.ts +18 -1
- package/src/tooling/opfs-cache.js +126 -9
- package/src/tooling/runtime-input-composition.d.ts +0 -1
- package/src/tooling/runtime-input-composition.js +0 -5
- package/src/tooling/source-artifact-adapter.d.ts +134 -0
- package/src/tooling/source-artifact-adapter.js +366 -0
- package/src/tooling/source-package-profiles.d.ts +33 -0
- package/src/tooling/source-package-profiles.js +57 -0
- package/src/tooling/source-runtime-bundle.d.ts +46 -6
- package/src/tooling/source-runtime-bundle.js +371 -60
- package/src/tooling/source-runtime-converter-config.d.ts +24 -0
- package/src/tooling/source-runtime-converter-config.js +297 -0
- package/src/tooling-experimental-exports.browser.js +2 -0
- package/src/tooling-experimental-exports.d.ts +7 -0
- package/src/tooling-experimental-exports.js +7 -0
- package/src/tooling-experimental-exports.shared.d.ts +26 -0
- package/src/tooling-experimental-exports.shared.js +29 -0
- package/src/tooling-exports.shared.d.ts +0 -23
- package/src/tooling-exports.shared.js +6 -28
- package/src/version.js +1 -1
- package/src/client/doppler-provider/generation.d.ts +0 -25
- package/src/client/doppler-provider/generation.js +0 -126
- package/src/client/doppler-provider/index.d.ts +0 -2
- package/src/client/doppler-provider/index.js +0 -3
- package/src/client/doppler-provider/provider.d.ts +0 -5
- package/src/client/doppler-provider/provider.js +0 -102
- package/src/client/doppler-provider/source-runtime.js +0 -677
- package/src/client/doppler-provider.d.ts +0 -46
- package/src/client/doppler-provider.js +0 -36
- package/src/config/schema/inference-defaults.schema.d.ts +0 -276
- package/src/config/schema/inference-defaults.schema.js +0 -189
- package/src/diffusion/index.d.ts +0 -8
- package/src/diffusion/index.js +0 -8
- package/src/energy/index.d.ts +0 -2
- package/src/energy/index.js +0 -2
- package/src/index-internal.d.ts +0 -2
- package/src/index-internal.js +0 -2
- /package/src/client/{doppler-provider → runtime}/types.js +0 -0
- /package/src/{adapters → experimental/adapters}/index.d.ts +0 -0
- /package/src/{adapters → experimental/adapters}/index.js +0 -0
- /package/src/{bridge → experimental/bridge}/extension/background.d.ts +0 -0
- /package/src/{bridge → experimental/bridge}/extension/background.js +0 -0
- /package/src/{bridge → experimental/bridge}/extension/manifest.json +0 -0
- /package/src/{bridge → experimental/bridge}/extension-client.d.ts +0 -0
- /package/src/{bridge → experimental/bridge}/index.d.ts +0 -0
- /package/src/{bridge → experimental/bridge}/index.js +0 -0
- /package/src/{bridge → experimental/bridge}/protocol.d.ts +0 -0
- /package/src/{browser → experimental/browser}/file-picker.d.ts +0 -0
- /package/src/{browser → experimental/browser}/file-picker.js +0 -0
- /package/src/{browser → experimental/browser}/tensor-source-download.d.ts +0 -0
- /package/src/{browser → experimental/browser}/tensor-source-file.d.ts +0 -0
- /package/src/{browser → experimental/browser}/tensor-source-file.js +0 -0
- /package/src/{browser → experimental/browser}/tensor-source-http.d.ts +0 -0
- /package/src/{browser → experimental/browser}/tensor-source-http.js +0 -0
- /package/src/{diffusion → experimental/diffusion}/config/expansion-lanes.json +0 -0
- /package/src/{diffusion → experimental/diffusion}/image-regression.d.ts +0 -0
- /package/src/{diffusion → experimental/diffusion}/image-regression.js +0 -0
- /package/src/{distribution → experimental/distribution}/p2p-control-plane.d.ts +0 -0
- /package/src/{distribution → experimental/distribution}/p2p-control-plane.js +0 -0
- /package/src/{distribution → experimental/distribution}/p2p-observability.d.ts +0 -0
- /package/src/{distribution → experimental/distribution}/p2p-observability.js +0 -0
- /package/src/{distribution → experimental/distribution}/p2p-transport-contract.d.ts +0 -0
- /package/src/{distribution → experimental/distribution}/p2p-webrtc-browser.d.ts +0 -0
- /package/src/{distribution → experimental/distribution}/p2p-webrtc-browser.js +0 -0
- /package/src/{distribution → experimental/distribution}/shard-delivery.d.ts +0 -0
- /package/src/{hotswap → experimental/hotswap}/runtime.js +0 -0
- /package/src/{training → experimental/training}/checkpoint-watch.d.ts +0 -0
- /package/src/{training → experimental/training}/checkpoint-watch.js +0 -0
- /package/src/{training → experimental/training}/checkpoint.d.ts +0 -0
- /package/src/{training → experimental/training}/dataloader.d.ts +0 -0
- /package/src/{training → experimental/training}/dataloader.js +0 -0
- /package/src/{training → experimental/training}/datasets/index.d.ts +0 -0
- /package/src/{training → experimental/training}/datasets/index.js +0 -0
- /package/src/{training → experimental/training}/datasets/jsonl.d.ts +0 -0
- /package/src/{training → experimental/training}/datasets/jsonl.js +0 -0
- /package/src/{training → experimental/training}/datasets/reploid.d.ts +0 -0
- /package/src/{training → experimental/training}/datasets/reploid.js +0 -0
- /package/src/{training → experimental/training}/datasets/text-pairs.d.ts +0 -0
- /package/src/{training → experimental/training}/datasets/text-pairs.js +0 -0
- /package/src/{training → experimental/training}/datasets/token-batch.d.ts +0 -0
- /package/src/{training → experimental/training}/datasets/token-batch.js +0 -0
- /package/src/{training → experimental/training}/datasets/translation-pairs.d.ts +0 -0
- /package/src/{training → experimental/training}/datasets/translation-pairs.js +0 -0
- /package/src/{training → experimental/training}/distillation/artifacts.d.ts +0 -0
- /package/src/{training → experimental/training}/distillation/artifacts.js +0 -0
- /package/src/{training → experimental/training}/distillation/checkpoint-watch.d.ts +0 -0
- /package/src/{training → experimental/training}/distillation/checkpoint-watch.js +0 -0
- /package/src/{training → experimental/training}/distillation/dataset.d.ts +0 -0
- /package/src/{training → experimental/training}/distillation/dataset.js +0 -0
- /package/src/{training → experimental/training}/distillation/eval.d.ts +0 -0
- /package/src/{training → experimental/training}/distillation/eval.js +0 -0
- /package/src/{training → experimental/training}/distillation/index.d.ts +0 -0
- /package/src/{training → experimental/training}/distillation/index.js +0 -0
- /package/src/{training → experimental/training}/distillation/runtime.d.ts +0 -0
- /package/src/{training → experimental/training}/distillation/runtime.js +0 -0
- /package/src/{training → experimental/training}/distillation/scoreboard.d.ts +0 -0
- /package/src/{training → experimental/training}/distillation/scoreboard.js +0 -0
- /package/src/{training → experimental/training}/distillation/stage-a.d.ts +0 -0
- /package/src/{training → experimental/training}/distillation/stage-a.js +0 -0
- /package/src/{training → experimental/training}/distillation/stage-b.d.ts +0 -0
- /package/src/{training → experimental/training}/distillation/stage-b.js +0 -0
- /package/src/{training → experimental/training}/distillation/student-fixture.d.ts +0 -0
- /package/src/{training → experimental/training}/distillation/student-fixture.js +0 -0
- /package/src/{training → experimental/training}/distillation/suite-data.d.ts +0 -0
- /package/src/{training → experimental/training}/distillation/suite-data.js +0 -0
- /package/src/{training → experimental/training}/index.d.ts +0 -0
- /package/src/{training → experimental/training}/index.js +0 -0
- /package/src/{training → experimental/training}/lora-pipeline.d.ts +0 -0
- /package/src/{training → experimental/training}/objectives/base.d.ts +0 -0
- /package/src/{training → experimental/training}/objectives/base.js +0 -0
- /package/src/{training → experimental/training}/objectives/cross_entropy.d.ts +0 -0
- /package/src/{training → experimental/training}/objectives/cross_entropy.js +0 -0
- /package/src/{training → experimental/training}/objectives/distill_kd.d.ts +0 -0
- /package/src/{training → experimental/training}/objectives/distill_kd.js +0 -0
- /package/src/{training → experimental/training}/objectives/distill_triplet.d.ts +0 -0
- /package/src/{training → experimental/training}/objectives/distill_triplet.js +0 -0
- /package/src/{training → experimental/training}/objectives/index.d.ts +0 -0
- /package/src/{training → experimental/training}/objectives/index.js +0 -0
- /package/src/{training → experimental/training}/objectives/ul_stage1_joint.d.ts +0 -0
- /package/src/{training → experimental/training}/objectives/ul_stage1_joint.js +0 -0
- /package/src/{training → experimental/training}/objectives/ul_stage2_base.d.ts +0 -0
- /package/src/{training → experimental/training}/objectives/ul_stage2_base.js +0 -0
- /package/src/{training → experimental/training}/operator-artifacts.d.ts +0 -0
- /package/src/{training → experimental/training}/operator-command.d.ts +0 -0
- /package/src/{training → experimental/training}/operator-command.js +0 -0
- /package/src/{training → experimental/training}/operator-eval.d.ts +0 -0
- /package/src/{training → experimental/training}/operator-eval.js +0 -0
- /package/src/{training → experimental/training}/operator-scoreboard.d.ts +0 -0
- /package/src/{training → experimental/training}/operator-scoreboard.js +0 -0
- /package/src/{training → experimental/training}/suite.d.ts +0 -0
- /package/src/{training → experimental/training}/ul_schedule.js +0 -0
- /package/src/{training → experimental/training}/workload-packs/distill-translategemma-medium.json +0 -0
- /package/src/{training → experimental/training}/workload-packs/distill-translategemma-tiny.json +0 -0
- /package/src/{training → experimental/training}/workload-packs/lora-toy-medium.json +0 -0
- /package/src/{training → experimental/training}/workload-packs/lora-toy-tiny.json +0 -0
- /package/src/{training → experimental/training}/workload-packs/registry.json +0 -0
- /package/src/{training → experimental/training}/workload-packs/ul-training-medium.json +0 -0
- /package/src/{training → experimental/training}/workload-packs/ul-training-tiny.json +0 -0
- /package/src/{training → experimental/training}/workloads.d.ts +0 -0
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
|
|
2
2
|
|
|
3
|
-
import { isWeightBuffer, getWeightDtype } from '../../../../gpu/weight-buffer.js';
|
|
3
|
+
import { isGpuBufferInstance, isWeightBuffer, getWeightDtype } from '../../../../gpu/weight-buffer.js';
|
|
4
4
|
import { acquireBuffer } from '../../../../memory/buffer-pool.js';
|
|
5
5
|
import {
|
|
6
6
|
recordMatmul,
|
|
@@ -9,6 +9,7 @@ import {
|
|
|
9
9
|
recordAttention,
|
|
10
10
|
recordAttentionTiered,
|
|
11
11
|
recordAttentionTieredQuant,
|
|
12
|
+
recordAttentionContiguousQuant,
|
|
12
13
|
recordAttentionBDPA,
|
|
13
14
|
recordSiLU,
|
|
14
15
|
recordCastF16ToF32,
|
|
@@ -27,19 +28,43 @@ import {
|
|
|
27
28
|
resolveAttentionProjectionOutputDtype,
|
|
28
29
|
projectAttentionQKV,
|
|
29
30
|
applyAttentionQKNorm,
|
|
31
|
+
applyAttentionValueNorm,
|
|
30
32
|
} from './projections.js';
|
|
31
33
|
import { prepareAttentionProjectionInput } from './output-projection.js';
|
|
32
34
|
|
|
33
35
|
import { releaseOrTrack, shouldDebugLayer } from './types.js';
|
|
34
|
-
import {
|
|
36
|
+
import {
|
|
37
|
+
getKernelPathMatmulPrecision,
|
|
38
|
+
getKernelPathMatmulVariant,
|
|
39
|
+
} from '../../../../config/kernel-path-loader.js';
|
|
35
40
|
import {
|
|
36
41
|
resolveKVCacheState,
|
|
37
42
|
buildAttentionDispatchParams,
|
|
38
43
|
buildAttentionInputsData,
|
|
39
44
|
} from './dispatch-params.js';
|
|
45
|
+
import {
|
|
46
|
+
buildTieredQuantAttentionOptions,
|
|
47
|
+
buildContiguousQuantAttentionOptions,
|
|
48
|
+
} from './quant-options.js';
|
|
49
|
+
import { assertImplicitDtypeTransitionAllowed } from '../dtype-contract.js';
|
|
50
|
+
import {
|
|
51
|
+
resolveAttentionPrecisionContract,
|
|
52
|
+
isAttentionKvDtypeExplicit,
|
|
53
|
+
} from './precision-contract.js';
|
|
40
54
|
|
|
41
55
|
const ATTENTION_DTYPE_LOGGED = new Set();
|
|
42
56
|
|
|
57
|
+
function assertAttentionDtypeTransitionAllowed(state, fromDtype, toDtype, detail, transitionDeclaredBy = null) {
|
|
58
|
+
assertImplicitDtypeTransitionAllowed({
|
|
59
|
+
executionPolicies: state?.executionPolicies ?? null,
|
|
60
|
+
fromDtype,
|
|
61
|
+
toDtype,
|
|
62
|
+
op: 'attention',
|
|
63
|
+
detail,
|
|
64
|
+
transitionDeclaredBy,
|
|
65
|
+
});
|
|
66
|
+
}
|
|
67
|
+
|
|
43
68
|
|
|
44
69
|
export async function recordLayerAttentionGPU(
|
|
45
70
|
recorder,
|
|
@@ -73,16 +98,30 @@ export async function recordLayerAttentionGPU(
|
|
|
73
98
|
tokenIds = null,
|
|
74
99
|
kernelPath = null,
|
|
75
100
|
disableRoPE = false,
|
|
101
|
+
multimodalBidirectionalSpan = null,
|
|
102
|
+
sharedKVSourceLayerIdx = null,
|
|
103
|
+
storeSharedKV = false,
|
|
76
104
|
} = config;
|
|
77
105
|
|
|
78
|
-
const
|
|
79
|
-
|
|
106
|
+
const phase = isPrefill ? 'prefill' : 'decode';
|
|
107
|
+
const attentionPrecisionContract = resolveAttentionPrecisionContract(config, state);
|
|
108
|
+
const attentionActivationDtype = selectRuleValue('shared', 'dtype', 'f16OrF32FromDtype', {
|
|
109
|
+
dtype: attentionPrecisionContract.resolvedActivationDtype ?? config.activationDtype,
|
|
110
|
+
});
|
|
111
|
+
const oProjPrecision = getKernelPathMatmulPrecision('o_proj', phase, layerIdx, kernelPath);
|
|
112
|
+
const oProjInputDtype = selectRuleValue('shared', 'dtype', 'f16OrF32FromDtype', {
|
|
113
|
+
dtype: oProjPrecision?.inputDtype ?? attentionActivationDtype,
|
|
80
114
|
});
|
|
81
|
-
const
|
|
82
|
-
|
|
83
|
-
|
|
84
|
-
|
|
85
|
-
|
|
115
|
+
const oProjOutputDtype = selectRuleValue('shared', 'dtype', 'f16OrF32FromDtype', {
|
|
116
|
+
dtype: oProjPrecision?.outputDtype
|
|
117
|
+
?? attentionPrecisionContract.resolvedOutputDtype
|
|
118
|
+
?? config.activationDtype,
|
|
119
|
+
});
|
|
120
|
+
const wantsF16Output = oProjOutputDtype === 'f16';
|
|
121
|
+
const useF16Activations = attentionActivationDtype === 'f16';
|
|
122
|
+
const kvCacheFallback = selectRuleValue('inference', 'dtype', 'f16OrF32', { useF16: useF16Activations });
|
|
123
|
+
const kvCacheDtype = attentionPrecisionContract.resolvedKvCacheDtype ?? state.kvCache?.kvDtype ?? kvCacheFallback;
|
|
124
|
+
const allowF16Attention = useF16Activations && kvCacheDtype === 'f16';
|
|
86
125
|
let attentionInput = input;
|
|
87
126
|
let attentionInputTemp = false;
|
|
88
127
|
let normed = attentionInput;
|
|
@@ -95,8 +134,17 @@ export async function recordLayerAttentionGPU(
|
|
|
95
134
|
let output = null;
|
|
96
135
|
let finalOutput = null;
|
|
97
136
|
let oProjInputTemp = null;
|
|
98
|
-
|
|
99
|
-
|
|
137
|
+
let retainSharedKvBuffers = false;
|
|
138
|
+
if (!allowF16Attention && input.dtype !== attentionActivationDtype) {
|
|
139
|
+
assertAttentionDtypeTransitionAllowed(
|
|
140
|
+
state,
|
|
141
|
+
input.dtype,
|
|
142
|
+
attentionActivationDtype,
|
|
143
|
+
'The attention kernel selection would widen the input implicitly.'
|
|
144
|
+
);
|
|
145
|
+
attentionInput = attentionActivationDtype === 'f16'
|
|
146
|
+
? await recordCastF32ToF16(recorder, input)
|
|
147
|
+
: await recordCastF16ToF32(recorder, input);
|
|
100
148
|
attentionInputTemp = true;
|
|
101
149
|
normed = attentionInput;
|
|
102
150
|
}
|
|
@@ -104,8 +152,7 @@ export async function recordLayerAttentionGPU(
|
|
|
104
152
|
if (!layerWeights) {
|
|
105
153
|
const bytesPerElement = wantsF16Output ? 2 : 4;
|
|
106
154
|
const outputBuf = acquireBuffer(numTokens * hiddenSize * bytesPerElement, undefined, 'attn_output');
|
|
107
|
-
const
|
|
108
|
-
const output = createTensor(outputBuf, outputDtype, [numTokens, hiddenSize], 'attn_output');
|
|
155
|
+
const output = createTensor(outputBuf, oProjOutputDtype, [numTokens, hiddenSize], 'attn_output');
|
|
109
156
|
return { output, residualFused: false };
|
|
110
157
|
}
|
|
111
158
|
|
|
@@ -122,7 +169,7 @@ export async function recordLayerAttentionGPU(
|
|
|
122
169
|
hiddenSize,
|
|
123
170
|
rmsNormWeightOffset: config.rmsNormWeightOffset,
|
|
124
171
|
});
|
|
125
|
-
if (!(layerWeights.inputNorm
|
|
172
|
+
if (!isGpuBufferInstance(layerWeights.inputNorm) && !isWeightBuffer(layerWeights.inputNorm)) releaseOrTrack(recorder, normWeightBuf);
|
|
126
173
|
}
|
|
127
174
|
|
|
128
175
|
const debugLayers = debugFlags.debugLayers;
|
|
@@ -132,16 +179,16 @@ export async function recordLayerAttentionGPU(
|
|
|
132
179
|
const logKey = `L${layerIdx}_${phase}_dtypes`;
|
|
133
180
|
if (!ATTENTION_DTYPE_LOGGED.has(logKey)) {
|
|
134
181
|
ATTENTION_DTYPE_LOGGED.add(logKey);
|
|
135
|
-
trace.attn(layerIdx, `dtypes: activation=${
|
|
182
|
+
trace.attn(layerIdx, `dtypes: activation=${attentionActivationDtype}, input=${input.dtype}, normed=${normed.dtype}`);
|
|
136
183
|
}
|
|
137
184
|
}
|
|
138
185
|
|
|
139
186
|
// 2. Q/K/V projections
|
|
140
|
-
const qProjVariant = getKernelPathMatmulVariant('q_proj',
|
|
187
|
+
const qProjVariant = getKernelPathMatmulVariant('q_proj', phase, layerIdx, kernelPath);
|
|
141
188
|
const kernelPathIsF16 = qProjVariant != null && qProjVariant.includes('f16') && !qProjVariant.includes('f32');
|
|
142
|
-
const matmulOutputDtype = resolveAttentionProjectionOutputDtype(
|
|
189
|
+
const matmulOutputDtype = resolveAttentionProjectionOutputDtype(attentionActivationDtype, {
|
|
143
190
|
forceF32: shouldForceF32AttentionProjectionForRoPE({
|
|
144
|
-
attentionInputDtype:
|
|
191
|
+
attentionInputDtype: attentionActivationDtype,
|
|
145
192
|
headDim,
|
|
146
193
|
rotaryDim: config.ropeRotaryDim,
|
|
147
194
|
interleaved: config.ropeInterleaved,
|
|
@@ -149,6 +196,27 @@ export async function recordLayerAttentionGPU(
|
|
|
149
196
|
}),
|
|
150
197
|
});
|
|
151
198
|
let usedFusedQKV = false;
|
|
199
|
+
const sharedKVEntry = sharedKVSourceLayerIdx == null
|
|
200
|
+
? null
|
|
201
|
+
: (state.sharedAttentionState?.get(sharedKVSourceLayerIdx) ?? null);
|
|
202
|
+
if (sharedKVSourceLayerIdx != null && !sharedKVEntry) {
|
|
203
|
+
throw new Error(
|
|
204
|
+
`Layer ${layerIdx} requires shared K/V from layer ${sharedKVSourceLayerIdx}, ` +
|
|
205
|
+
'but no shared K/V state was stored for that source layer.'
|
|
206
|
+
);
|
|
207
|
+
}
|
|
208
|
+
if (sharedKVEntry && (
|
|
209
|
+
sharedKVEntry.headDim !== headDim
|
|
210
|
+
|| sharedKVEntry.numKVHeads !== numKVHeads
|
|
211
|
+
)) {
|
|
212
|
+
throw new Error(
|
|
213
|
+
`Layer ${layerIdx} shared K/V geometry mismatch. ` +
|
|
214
|
+
`Expected numKVHeads=${numKVHeads}, headDim=${headDim}; ` +
|
|
215
|
+
`got numKVHeads=${sharedKVEntry.numKVHeads}, headDim=${sharedKVEntry.headDim}.`
|
|
216
|
+
);
|
|
217
|
+
}
|
|
218
|
+
const reusesSharedKV = sharedKVEntry != null;
|
|
219
|
+
retainSharedKvBuffers = reusesSharedKV || storeSharedKV;
|
|
152
220
|
({ qTensor, qGateTensor, kTensor, vTensor, usedFusedQKV } = await projectAttentionQKV({
|
|
153
221
|
recorder,
|
|
154
222
|
normed,
|
|
@@ -165,6 +233,9 @@ export async function recordLayerAttentionGPU(
|
|
|
165
233
|
lora,
|
|
166
234
|
matmulDebug: state.runtimeConfig?.shared?.debug?.matmul ?? null,
|
|
167
235
|
attentionOutputGate: config.attentionOutputGate === true,
|
|
236
|
+
sharedKTensor: sharedKVEntry?.kTensor ?? null,
|
|
237
|
+
sharedVTensor: sharedKVEntry?.vTensor ?? null,
|
|
238
|
+
executionPolicies: state.executionPolicies ?? null,
|
|
168
239
|
releaseTemporary: (buffer) => releaseOrTrack(recorder, buffer),
|
|
169
240
|
onFusedQKV: layerIdx === 0 && isPrefill
|
|
170
241
|
? ({ qSize: qSizeFused, kSize: kSizeFused, vSize: vSizeFused, totalSize }) => {
|
|
@@ -192,8 +263,21 @@ export async function recordLayerAttentionGPU(
|
|
|
192
263
|
headDim,
|
|
193
264
|
rmsNormWeightOffset: config.rmsNormWeightOffset,
|
|
194
265
|
releaseTemporary: (buffer) => releaseOrTrack(recorder, buffer),
|
|
266
|
+
skipKNorm: reusesSharedKV,
|
|
195
267
|
}));
|
|
196
268
|
|
|
269
|
+
if (config.valueNorm === true && !reusesSharedKV) {
|
|
270
|
+
vTensor = await applyAttentionValueNorm({
|
|
271
|
+
recorder,
|
|
272
|
+
vTensor,
|
|
273
|
+
rmsNormEps,
|
|
274
|
+
numTokens,
|
|
275
|
+
numKVHeads,
|
|
276
|
+
headDim,
|
|
277
|
+
releaseTemporary: (buffer) => releaseOrTrack(recorder, buffer),
|
|
278
|
+
});
|
|
279
|
+
}
|
|
280
|
+
|
|
197
281
|
if (normed !== attentionInput) releaseOrTrack(recorder, normed.buffer);
|
|
198
282
|
if (attentionInputTemp) recorder.trackTemporaryBuffer(attentionInput.buffer);
|
|
199
283
|
|
|
@@ -205,22 +289,41 @@ export async function recordLayerAttentionGPU(
|
|
|
205
289
|
rotaryDim: config.ropeRotaryDim,
|
|
206
290
|
interleaved: config.ropeInterleaved,
|
|
207
291
|
startPos: currentSeqLen,
|
|
292
|
+
executionPolicies: state.executionPolicies ?? null,
|
|
208
293
|
});
|
|
209
|
-
|
|
210
|
-
|
|
294
|
+
if (!reusesSharedKV) {
|
|
295
|
+
await recordRoPE(recorder, kTensor, state.ropeFreqsCos, state.ropeFreqsSin, numTokens, {
|
|
296
|
+
numHeads: numKVHeads,
|
|
297
|
+
headDim,
|
|
298
|
+
rotaryDim: config.ropeRotaryDim,
|
|
299
|
+
interleaved: config.ropeInterleaved,
|
|
300
|
+
startPos: currentSeqLen,
|
|
301
|
+
executionPolicies: state.executionPolicies ?? null,
|
|
302
|
+
});
|
|
303
|
+
}
|
|
304
|
+
}
|
|
305
|
+
|
|
306
|
+
if (storeSharedKV && state.sharedAttentionState) {
|
|
307
|
+
state.sharedAttentionState.set(layerIdx, {
|
|
308
|
+
kTensor,
|
|
309
|
+
vTensor,
|
|
211
310
|
headDim,
|
|
212
|
-
|
|
213
|
-
interleaved: config.ropeInterleaved,
|
|
214
|
-
startPos: currentSeqLen,
|
|
311
|
+
numKVHeads,
|
|
215
312
|
});
|
|
216
313
|
}
|
|
217
314
|
|
|
218
315
|
// 4. Update KV cache (cache stores raw GPUBuffers for memory efficiency)
|
|
219
|
-
|
|
220
316
|
if (state.kvCache?.hasGPUCache?.()) {
|
|
221
317
|
// Use recordUpdateFromGPU to record copy operations to the recorder's encoder
|
|
222
318
|
// This ensures K/V buffers are populated before copying (all ops submitted together)
|
|
223
319
|
if (state.kvCache.kvDtype === 'f16') {
|
|
320
|
+
const hasExplicitF16KvContract = isAttentionKvDtypeExplicit(attentionPrecisionContract, 'f16');
|
|
321
|
+
if (kTensor.dtype !== 'f16' && !hasExplicitF16KvContract) {
|
|
322
|
+
assertAttentionDtypeTransitionAllowed(state, kTensor.dtype, 'f16', 'K would be narrowed implicitly for KV cache storage.');
|
|
323
|
+
}
|
|
324
|
+
if (vTensor.dtype !== 'f16' && !hasExplicitF16KvContract) {
|
|
325
|
+
assertAttentionDtypeTransitionAllowed(state, vTensor.dtype, 'f16', 'V would be narrowed implicitly for KV cache storage.');
|
|
326
|
+
}
|
|
224
327
|
const kCasted = kTensor.dtype === 'f16' ? kTensor : await recordCastF32ToF16(recorder, kTensor);
|
|
225
328
|
const vCasted = vTensor.dtype === 'f16' ? vTensor : await recordCastF32ToF16(recorder, vTensor);
|
|
226
329
|
|
|
@@ -239,14 +342,14 @@ export async function recordLayerAttentionGPU(
|
|
|
239
342
|
layerIdx, numTokens, isPrefill, numHeads, numKVHeads, headDim, hiddenSize,
|
|
240
343
|
slidingWindow, layerType, layerTypes: config.layerTypes,
|
|
241
344
|
queryPreAttnScalar, causalAttention: config.causalAttention,
|
|
242
|
-
activationDtype:
|
|
243
|
-
kvCacheDtype: state.kvCache?.kvDtype ?? null,
|
|
345
|
+
activationDtype: attentionActivationDtype,
|
|
346
|
+
kvCacheDtype: attentionPrecisionContract.resolvedKvCacheDtype ?? state.kvCache?.kvDtype ?? null,
|
|
244
347
|
};
|
|
245
348
|
const dispatchParams = buildAttentionDispatchParams(dispatchConfig, state, kTensor, vTensor, kvState);
|
|
246
349
|
const {
|
|
247
350
|
effectiveSlidingWindow, attentionKernelVariant, attnScale,
|
|
248
351
|
cachedKDtype, cachedVDtype, cachedKTensor, cachedVTensor,
|
|
249
|
-
causalForAttention,
|
|
352
|
+
prefillFallbackNeedsCast, causalForAttention,
|
|
250
353
|
} = dispatchParams;
|
|
251
354
|
|
|
252
355
|
// 5. Attention
|
|
@@ -267,6 +370,7 @@ export async function recordLayerAttentionGPU(
|
|
|
267
370
|
|
|
268
371
|
let qForBDPA = qTensor;
|
|
269
372
|
if (qForBDPA.dtype !== 'f16') {
|
|
373
|
+
assertAttentionDtypeTransitionAllowed(state, qForBDPA.dtype, 'f16', 'BDPA attention would narrow Q implicitly.');
|
|
270
374
|
qForBDPA = await recordCastF32ToF16(recorder, qTensor);
|
|
271
375
|
recorder.trackTemporaryBuffer(qForBDPA.buffer);
|
|
272
376
|
}
|
|
@@ -288,6 +392,7 @@ export async function recordLayerAttentionGPU(
|
|
|
288
392
|
tieredQuant: async () => {
|
|
289
393
|
let qForAttention = qTensor;
|
|
290
394
|
if (kvState.coldQuantMode !== 'none' && qTensor.dtype !== 'f32') {
|
|
395
|
+
assertAttentionDtypeTransitionAllowed(state, qTensor.dtype, 'f32', 'Tiered quant attention would widen Q implicitly.');
|
|
291
396
|
qForAttention = await recordCastF16ToF32(recorder, qTensor);
|
|
292
397
|
recorder.trackTemporaryBuffer(qForAttention.buffer);
|
|
293
398
|
}
|
|
@@ -300,21 +405,63 @@ export async function recordLayerAttentionGPU(
|
|
|
300
405
|
|
|
301
406
|
const cachedHotKTensor = createTensor(kvState.cachedKHot, cachedKDtype, [kvState.hotLen, numKVHeads * headDim], 'cached_K_hot');
|
|
302
407
|
const cachedHotVTensor = createTensor(kvState.cachedVHot, cachedVDtype, [kvState.hotLen, numKVHeads * headDim], 'cached_V_hot');
|
|
303
|
-
return recordAttentionTieredQuant(
|
|
304
|
-
|
|
305
|
-
|
|
306
|
-
|
|
307
|
-
|
|
308
|
-
|
|
309
|
-
|
|
310
|
-
|
|
311
|
-
|
|
312
|
-
|
|
313
|
-
|
|
314
|
-
|
|
315
|
-
|
|
316
|
-
|
|
317
|
-
|
|
408
|
+
return recordAttentionTieredQuant(
|
|
409
|
+
recorder,
|
|
410
|
+
qForAttention,
|
|
411
|
+
cachedHotKTensor,
|
|
412
|
+
cachedHotVTensor,
|
|
413
|
+
kvState.cachedKCold,
|
|
414
|
+
kvState.cachedVCold,
|
|
415
|
+
kvState.coldScalesK,
|
|
416
|
+
kvState.coldScalesV,
|
|
417
|
+
numHeads,
|
|
418
|
+
headDim,
|
|
419
|
+
buildTieredQuantAttentionOptions(kvState, {
|
|
420
|
+
seqLen: numTokens,
|
|
421
|
+
numKVHeads,
|
|
422
|
+
causal: causalForAttention,
|
|
423
|
+
startPos: kvState.startPosForMask,
|
|
424
|
+
slidingWindow: effectiveSlidingWindow ?? 0,
|
|
425
|
+
attnSoftcap,
|
|
426
|
+
scale: attnScale,
|
|
427
|
+
})
|
|
428
|
+
);
|
|
429
|
+
},
|
|
430
|
+
contiguousQuant: async () => {
|
|
431
|
+
let qForAttention = qTensor;
|
|
432
|
+
if (qTensor.dtype !== 'f32') {
|
|
433
|
+
assertAttentionDtypeTransitionAllowed(state, qTensor.dtype, 'f32', 'Contiguous quant attention would widen Q implicitly.');
|
|
434
|
+
qForAttention = await recordCastF16ToF32(recorder, qTensor);
|
|
435
|
+
recorder.trackTemporaryBuffer(qForAttention.buffer);
|
|
436
|
+
}
|
|
437
|
+
|
|
438
|
+
if (!kvState.coldScalesK || !kvState.coldScalesV) {
|
|
439
|
+
throw new Error('Contiguous quant attention requires scale buffers.');
|
|
440
|
+
}
|
|
441
|
+
if (!kvState.rotationMatrixBuffer || !kvState.codebookCentroidsBuffer) {
|
|
442
|
+
throw new Error('Contiguous quant attention requires TurboQuant shared buffers.');
|
|
443
|
+
}
|
|
444
|
+
|
|
445
|
+
return recordAttentionContiguousQuant(
|
|
446
|
+
recorder,
|
|
447
|
+
qForAttention,
|
|
448
|
+
kvState.cachedKCold,
|
|
449
|
+
kvState.cachedVCold,
|
|
450
|
+
kvState.coldScalesK,
|
|
451
|
+
kvState.coldScalesV,
|
|
452
|
+
numHeads,
|
|
453
|
+
headDim,
|
|
454
|
+
buildContiguousQuantAttentionOptions(kvState, {
|
|
455
|
+
seqLen: numTokens,
|
|
456
|
+
kvLen: kvState.kvLenForAttention,
|
|
457
|
+
numKVHeads,
|
|
458
|
+
causal: causalForAttention,
|
|
459
|
+
startPos: kvState.startPosForMask,
|
|
460
|
+
slidingWindow: effectiveSlidingWindow ?? 0,
|
|
461
|
+
attnSoftcap,
|
|
462
|
+
scale: attnScale,
|
|
463
|
+
})
|
|
464
|
+
);
|
|
318
465
|
},
|
|
319
466
|
tiered: async () => {
|
|
320
467
|
const cachedHotKTensor = createTensor(kvState.cachedKHot, cachedKDtype, [kvState.hotLen, numKVHeads * headDim], 'cached_K_hot');
|
|
@@ -339,22 +486,46 @@ export async function recordLayerAttentionGPU(
|
|
|
339
486
|
hotLayout: kvState.hotWindow > 0 ? 1 : 0,
|
|
340
487
|
});
|
|
341
488
|
},
|
|
342
|
-
contiguous: async () =>
|
|
343
|
-
|
|
344
|
-
|
|
345
|
-
|
|
346
|
-
|
|
347
|
-
|
|
348
|
-
|
|
349
|
-
|
|
350
|
-
|
|
351
|
-
|
|
352
|
-
|
|
353
|
-
|
|
354
|
-
|
|
355
|
-
|
|
356
|
-
|
|
357
|
-
|
|
489
|
+
contiguous: async () => {
|
|
490
|
+
// Prefill fallback: quantized/tiered layouts use raw K/V for prefill, cast to f16 to match kernel path
|
|
491
|
+
let kForAttn = cachedKTensor;
|
|
492
|
+
let vForAttn = cachedVTensor;
|
|
493
|
+
if (prefillFallbackNeedsCast) {
|
|
494
|
+
const hasExplicitF16KvContract = isAttentionKvDtypeExplicit(attentionPrecisionContract, 'f16');
|
|
495
|
+
if (cachedKDtype === 'f16' && kTensor.dtype !== 'f16' && !hasExplicitF16KvContract) {
|
|
496
|
+
assertAttentionDtypeTransitionAllowed(state, kTensor.dtype, 'f16', 'Prefill fallback attention would narrow K implicitly.');
|
|
497
|
+
}
|
|
498
|
+
if (cachedVDtype === 'f16' && vTensor.dtype !== 'f16' && !hasExplicitF16KvContract) {
|
|
499
|
+
assertAttentionDtypeTransitionAllowed(state, vTensor.dtype, 'f16', 'Prefill fallback attention would narrow V implicitly.');
|
|
500
|
+
}
|
|
501
|
+
const kCasted = cachedKDtype === 'f16' && kTensor.dtype !== 'f16'
|
|
502
|
+
? await recordCastF32ToF16(recorder, kTensor) : kTensor;
|
|
503
|
+
const vCasted = cachedVDtype === 'f16' && vTensor.dtype !== 'f16'
|
|
504
|
+
? await recordCastF32ToF16(recorder, vTensor) : vTensor;
|
|
505
|
+
kForAttn = createTensor(kCasted.buffer, kCasted.dtype, [kvState.kvLenForAttention, numKVHeads * headDim], 'cached_K');
|
|
506
|
+
vForAttn = createTensor(vCasted.buffer, vCasted.dtype, [kvState.kvLenForAttention, numKVHeads * headDim], 'cached_V');
|
|
507
|
+
if (kTensor.dtype !== 'f16') recorder.trackTemporaryBuffer(kCasted.buffer);
|
|
508
|
+
if (vTensor.dtype !== 'f16') recorder.trackTemporaryBuffer(vCasted.buffer);
|
|
509
|
+
}
|
|
510
|
+
return recordAttention(recorder, qTensor, kForAttn, vForAttn, null, numHeads, headDim, {
|
|
511
|
+
seqLen: numTokens,
|
|
512
|
+
kvLen: kvState.kvLenForAttention,
|
|
513
|
+
numKVHeads,
|
|
514
|
+
causal: causalForAttention,
|
|
515
|
+
bidirectionalSpanStart: multimodalBidirectionalSpan?.start ?? 0,
|
|
516
|
+
bidirectionalSpanLength: multimodalBidirectionalSpan?.length ?? 0,
|
|
517
|
+
startPos: kvState.startPosForMask,
|
|
518
|
+
layerIdx,
|
|
519
|
+
slidingWindow: effectiveSlidingWindow,
|
|
520
|
+
attnSoftcap,
|
|
521
|
+
scale: attnScale,
|
|
522
|
+
kvStart: kvState.kvStart,
|
|
523
|
+
kvLayout: kvState.kvLayout,
|
|
524
|
+
kvPageTable: kvState.kvPageTable,
|
|
525
|
+
kvPageSize: kvState.kvPageSize,
|
|
526
|
+
kernelPath,
|
|
527
|
+
});
|
|
528
|
+
},
|
|
358
529
|
};
|
|
359
530
|
const runAttentionKernel = attentionKernelRunners[attentionKernelVariant];
|
|
360
531
|
if (!runAttentionKernel) {
|
|
@@ -384,8 +555,19 @@ export async function recordLayerAttentionGPU(
|
|
|
384
555
|
if (layerWeights.oProj && getWeightBuffer) {
|
|
385
556
|
({ oProjInput, oProjInputTemp } = await prepareAttentionProjectionInput(
|
|
386
557
|
attnForProjection,
|
|
387
|
-
|
|
388
|
-
(tensor) =>
|
|
558
|
+
oProjInputDtype,
|
|
559
|
+
(tensor) => {
|
|
560
|
+
assertAttentionDtypeTransitionAllowed(
|
|
561
|
+
state,
|
|
562
|
+
tensor.dtype,
|
|
563
|
+
oProjInputDtype,
|
|
564
|
+
'Attention output projection would change activations implicitly.',
|
|
565
|
+
'step_precision'
|
|
566
|
+
);
|
|
567
|
+
return oProjInputDtype === 'f16'
|
|
568
|
+
? recordCastF32ToF16(recorder, tensor)
|
|
569
|
+
: recordCastF16ToF32(recorder, tensor);
|
|
570
|
+
}
|
|
389
571
|
));
|
|
390
572
|
const oProjBuf = getWeightBuffer(layerWeights.oProj, 'o_proj');
|
|
391
573
|
const loraO = getLoRAModule(lora, layerIdx, 'o_proj');
|
|
@@ -417,11 +599,12 @@ export async function recordLayerAttentionGPU(
|
|
|
417
599
|
role: 'o_proj',
|
|
418
600
|
layerIdx,
|
|
419
601
|
kernelPath,
|
|
420
|
-
outputDtype:
|
|
602
|
+
outputDtype: oProjOutputDtype,
|
|
603
|
+
executionPolicies: state.executionPolicies ?? null,
|
|
421
604
|
});
|
|
422
605
|
}
|
|
423
606
|
// Release temporary buffer if we created it (original was not already on GPU)
|
|
424
|
-
if (!(layerWeights.oProj
|
|
607
|
+
if (!isGpuBufferInstance(layerWeights.oProj) && !isWeightBuffer(layerWeights.oProj)) {
|
|
425
608
|
releaseOrTrack(recorder, isWeightBuffer(oProjBuf) ? oProjBuf.buffer : oProjBuf);
|
|
426
609
|
}
|
|
427
610
|
} else {
|
|
@@ -457,10 +640,13 @@ export async function recordLayerAttentionGPU(
|
|
|
457
640
|
if (oProjInputTemp && oProjInputTemp.buffer !== attnForProjection.buffer) {
|
|
458
641
|
buffersToTrack.push(oProjInputTemp.buffer);
|
|
459
642
|
}
|
|
460
|
-
if (
|
|
461
|
-
|
|
643
|
+
if (output.dtype !== oProjOutputDtype) {
|
|
644
|
+
assertAttentionDtypeTransitionAllowed(state, output.dtype, oProjOutputDtype, 'Attention output would change implicitly before leaving the layer.');
|
|
645
|
+
const coercedOutput = oProjOutputDtype === 'f16'
|
|
646
|
+
? await recordCastF32ToF16(recorder, output)
|
|
647
|
+
: await recordCastF16ToF32(recorder, output);
|
|
462
648
|
buffersToTrack.push(output.buffer);
|
|
463
|
-
finalOutput =
|
|
649
|
+
finalOutput = coercedOutput;
|
|
464
650
|
}
|
|
465
651
|
|
|
466
652
|
// Track intermediate buffers for cleanup after submit (not release!)
|
|
@@ -471,8 +657,10 @@ export async function recordLayerAttentionGPU(
|
|
|
471
657
|
if (qGateTensor) {
|
|
472
658
|
recorder.trackTemporaryBuffer(qGateTensor.buffer);
|
|
473
659
|
}
|
|
474
|
-
|
|
475
|
-
|
|
660
|
+
if (!retainSharedKvBuffers) {
|
|
661
|
+
recorder.trackTemporaryBuffer(kTensor.buffer);
|
|
662
|
+
recorder.trackTemporaryBuffer(vTensor.buffer);
|
|
663
|
+
}
|
|
476
664
|
for (const buffer of buffersToTrack) {
|
|
477
665
|
recorder.trackTemporaryBuffer(buffer);
|
|
478
666
|
}
|
|
@@ -506,10 +694,10 @@ export async function recordLayerAttentionGPU(
|
|
|
506
694
|
if (qTensor?.buffer) {
|
|
507
695
|
trackOnce(qTensor.buffer);
|
|
508
696
|
}
|
|
509
|
-
if (kTensor?.buffer) {
|
|
697
|
+
if (kTensor?.buffer && !retainSharedKvBuffers) {
|
|
510
698
|
trackOnce(kTensor.buffer);
|
|
511
699
|
}
|
|
512
|
-
if (vTensor?.buffer) {
|
|
700
|
+
if (vTensor?.buffer && !retainSharedKvBuffers) {
|
|
513
701
|
trackOnce(vTensor.buffer);
|
|
514
702
|
}
|
|
515
703
|
if (normed?.buffer && normed.buffer !== attentionInput?.buffer) {
|