doppler-gpu 0.2.0 → 0.2.2
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 +29 -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
|
@@ -402,3 +402,420 @@ fn main_multicol(
|
|
|
402
402
|
C[col] = final_sum * u.alpha;
|
|
403
403
|
}
|
|
404
404
|
}
|
|
405
|
+
|
|
406
|
+
// ============================================================================
|
|
407
|
+
// Fast-dequant variant: replaces per-element get_q4() calls (~10 ALU each)
|
|
408
|
+
// with direct word extraction (~6 ALU per 4 elements).
|
|
409
|
+
// Q4_K sub-block layout: even sub-blocks use lower nibbles, odd use upper.
|
|
410
|
+
// 4 consecutive bytes (1 u32 word) yield 4 q4 values via shift+mask.
|
|
411
|
+
// ============================================================================
|
|
412
|
+
var<workgroup> multicol_sums_fd: array<f32, MAX_WORKGROUP_SIZE>;
|
|
413
|
+
|
|
414
|
+
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
415
|
+
fn main_multicol_fast(
|
|
416
|
+
@builtin(local_invocation_id) lid: vec3<u32>,
|
|
417
|
+
@builtin(workgroup_id) wg_id: vec3<u32>,
|
|
418
|
+
@builtin(subgroup_invocation_id) sg_id: u32,
|
|
419
|
+
@builtin(subgroup_size) sg_size: u32
|
|
420
|
+
) {
|
|
421
|
+
let local_id = lid.x;
|
|
422
|
+
|
|
423
|
+
let col_in_wg = local_id / THREADS_PER_COL_GEMV;
|
|
424
|
+
let tid_in_col = local_id % THREADS_PER_COL_GEMV;
|
|
425
|
+
|
|
426
|
+
let col = wg_id.x * COLS_PER_WG + col_in_wg;
|
|
427
|
+
|
|
428
|
+
let is_valid = col < u.N;
|
|
429
|
+
|
|
430
|
+
var partial_sum: f32 = 0.0;
|
|
431
|
+
|
|
432
|
+
if (is_valid) {
|
|
433
|
+
let num_blocks = u.num_blocks_per_row;
|
|
434
|
+
let tail_size = u.K & 255u;
|
|
435
|
+
let full_blocks = num_blocks - select(0u, 1u, tail_size > 0u);
|
|
436
|
+
|
|
437
|
+
for (var b: u32 = tid_in_col; b < full_blocks; b = b + THREADS_PER_COL_GEMV) {
|
|
438
|
+
let block = B_q4k[col * num_blocks + b];
|
|
439
|
+
let d = unpack_f16_lo(block.d_dmin);
|
|
440
|
+
let dmin = unpack_f16_hi(block.d_dmin);
|
|
441
|
+
let k_base = b * QK_K;
|
|
442
|
+
|
|
443
|
+
for (var sb: u32 = 0u; sb < 8u; sb = sb + 1u) {
|
|
444
|
+
let sm = get_scale_min_k4(block.scales, sb);
|
|
445
|
+
let scale = d * f32(sm.x);
|
|
446
|
+
let min_val = dmin * f32(sm.y);
|
|
447
|
+
let sb_base = sb * SUBBLOCK_SIZE;
|
|
448
|
+
|
|
449
|
+
// Direct word extraction: even sb = lower nibble, odd sb = upper
|
|
450
|
+
let chunk = sb >> 1u;
|
|
451
|
+
let nibble_shift = (sb & 1u) * 4u;
|
|
452
|
+
let word_base = chunk * 8u;
|
|
453
|
+
|
|
454
|
+
for (var i: u32 = 0u; i < SUBBLOCK_SIZE; i = i + 4u) {
|
|
455
|
+
let k0 = k_base + sb_base + i;
|
|
456
|
+
|
|
457
|
+
let a0 = A[k0];
|
|
458
|
+
let a1 = A[k0 + 1u];
|
|
459
|
+
let a2 = A[k0 + 2u];
|
|
460
|
+
let a3 = A[k0 + 3u];
|
|
461
|
+
|
|
462
|
+
let word = block.qs[word_base + (i >> 2u)];
|
|
463
|
+
let q0 = (word >> nibble_shift) & 0xFu;
|
|
464
|
+
let q1 = (word >> (nibble_shift + 8u)) & 0xFu;
|
|
465
|
+
let q2 = (word >> (nibble_shift + 16u)) & 0xFu;
|
|
466
|
+
let q3 = (word >> (nibble_shift + 24u)) & 0xFu;
|
|
467
|
+
|
|
468
|
+
let w0 = scale * f32(q0) - min_val;
|
|
469
|
+
let w1 = scale * f32(q1) - min_val;
|
|
470
|
+
let w2 = scale * f32(q2) - min_val;
|
|
471
|
+
let w3 = scale * f32(q3) - min_val;
|
|
472
|
+
|
|
473
|
+
partial_sum = partial_sum + a0 * w0 + a1 * w1 + a2 * w2 + a3 * w3;
|
|
474
|
+
}
|
|
475
|
+
}
|
|
476
|
+
}
|
|
477
|
+
|
|
478
|
+
if (tail_size > 0u) {
|
|
479
|
+
let tail_block = full_blocks;
|
|
480
|
+
if (tail_block % THREADS_PER_COL_GEMV == tid_in_col) {
|
|
481
|
+
let block = B_q4k[col * num_blocks + tail_block];
|
|
482
|
+
let d = unpack_f16_lo(block.d_dmin);
|
|
483
|
+
let dmin = unpack_f16_hi(block.d_dmin);
|
|
484
|
+
let k_base = tail_block * QK_K;
|
|
485
|
+
|
|
486
|
+
for (var sb: u32 = 0u; sb < 8u; sb = sb + 1u) {
|
|
487
|
+
let sb_base = sb * SUBBLOCK_SIZE;
|
|
488
|
+
if (sb_base >= tail_size) {
|
|
489
|
+
break;
|
|
490
|
+
}
|
|
491
|
+
let sm = get_scale_min_k4(block.scales, sb);
|
|
492
|
+
let scale = d * f32(sm.x);
|
|
493
|
+
let min_val = dmin * f32(sm.y);
|
|
494
|
+
|
|
495
|
+
let chunk = sb >> 1u;
|
|
496
|
+
let nibble_shift = (sb & 1u) * 4u;
|
|
497
|
+
let word_base = chunk * 8u;
|
|
498
|
+
|
|
499
|
+
for (var i: u32 = 0u; i < SUBBLOCK_SIZE; i = i + 4u) {
|
|
500
|
+
let k0 = k_base + sb_base + i;
|
|
501
|
+
|
|
502
|
+
var a0: f32 = 0.0;
|
|
503
|
+
var a1: f32 = 0.0;
|
|
504
|
+
var a2: f32 = 0.0;
|
|
505
|
+
var a3: f32 = 0.0;
|
|
506
|
+
if (k0 < u.K) { a0 = A[k0]; }
|
|
507
|
+
if (k0 + 1u < u.K) { a1 = A[k0 + 1u]; }
|
|
508
|
+
if (k0 + 2u < u.K) { a2 = A[k0 + 2u]; }
|
|
509
|
+
if (k0 + 3u < u.K) { a3 = A[k0 + 3u]; }
|
|
510
|
+
|
|
511
|
+
let word = block.qs[word_base + (i >> 2u)];
|
|
512
|
+
let q0 = (word >> nibble_shift) & 0xFu;
|
|
513
|
+
let q1 = (word >> (nibble_shift + 8u)) & 0xFu;
|
|
514
|
+
let q2 = (word >> (nibble_shift + 16u)) & 0xFu;
|
|
515
|
+
let q3 = (word >> (nibble_shift + 24u)) & 0xFu;
|
|
516
|
+
|
|
517
|
+
let w0 = scale * f32(q0) - min_val;
|
|
518
|
+
let w1 = scale * f32(q1) - min_val;
|
|
519
|
+
let w2 = scale * f32(q2) - min_val;
|
|
520
|
+
let w3 = scale * f32(q3) - min_val;
|
|
521
|
+
|
|
522
|
+
partial_sum = partial_sum + a0 * w0 + a1 * w1 + a2 * w2 + a3 * w3;
|
|
523
|
+
}
|
|
524
|
+
}
|
|
525
|
+
}
|
|
526
|
+
}
|
|
527
|
+
}
|
|
528
|
+
|
|
529
|
+
multicol_sums_fd[local_id] = partial_sum;
|
|
530
|
+
workgroupBarrier();
|
|
531
|
+
|
|
532
|
+
if (tid_in_col == 0u && is_valid) {
|
|
533
|
+
var final_sum: f32 = 0.0;
|
|
534
|
+
let base = col_in_wg * THREADS_PER_COL_GEMV;
|
|
535
|
+
for (var i: u32 = 0u; i < THREADS_PER_COL_GEMV; i = i + 1u) {
|
|
536
|
+
final_sum = final_sum + multicol_sums_fd[base + i];
|
|
537
|
+
}
|
|
538
|
+
C[col] = final_sum * u.alpha;
|
|
539
|
+
}
|
|
540
|
+
}
|
|
541
|
+
|
|
542
|
+
// ============================================================================
|
|
543
|
+
// Shared-A variant: loads the A vector into workgroup memory once,
|
|
544
|
+
// eliminating redundant global reads across columns.
|
|
545
|
+
// Use when K <= SHARED_A_MAX (compile-time override).
|
|
546
|
+
// Total shared memory: SHARED_A_MAX*4 + 256*4 = e.g. 2048*4+1024 = 9216 bytes.
|
|
547
|
+
// ============================================================================
|
|
548
|
+
override SHARED_A_MAX: u32 = 3584u;
|
|
549
|
+
|
|
550
|
+
var<workgroup> shared_A_buf: array<f32, SHARED_A_MAX>;
|
|
551
|
+
var<workgroup> multicol_sums_sa: array<f32, MAX_WORKGROUP_SIZE>;
|
|
552
|
+
|
|
553
|
+
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
554
|
+
fn main_multicol_shared(
|
|
555
|
+
@builtin(local_invocation_id) lid: vec3<u32>,
|
|
556
|
+
@builtin(workgroup_id) wg_id: vec3<u32>,
|
|
557
|
+
@builtin(subgroup_invocation_id) sg_id: u32,
|
|
558
|
+
@builtin(subgroup_size) sg_size: u32
|
|
559
|
+
) {
|
|
560
|
+
let local_id = lid.x;
|
|
561
|
+
|
|
562
|
+
// Cooperative load: all 256 threads fill shared_A_buf from global A
|
|
563
|
+
for (var idx: u32 = local_id; idx < u.K; idx = idx + WORKGROUP_SIZE) {
|
|
564
|
+
shared_A_buf[idx] = A[idx];
|
|
565
|
+
}
|
|
566
|
+
workgroupBarrier();
|
|
567
|
+
|
|
568
|
+
// Which column within this workgroup (0..31)
|
|
569
|
+
let col_in_wg = local_id / THREADS_PER_COL_GEMV;
|
|
570
|
+
// Which thread within the column (0..7)
|
|
571
|
+
let tid_in_col = local_id % THREADS_PER_COL_GEMV;
|
|
572
|
+
|
|
573
|
+
// Global column index
|
|
574
|
+
let col = wg_id.x * COLS_PER_WG + col_in_wg;
|
|
575
|
+
|
|
576
|
+
// Track validity
|
|
577
|
+
let is_valid = col < u.N;
|
|
578
|
+
|
|
579
|
+
var partial_sum: f32 = 0.0;
|
|
580
|
+
|
|
581
|
+
if (is_valid) {
|
|
582
|
+
let num_blocks = u.num_blocks_per_row;
|
|
583
|
+
let tail_size = u.K & 255u;
|
|
584
|
+
let full_blocks = num_blocks - select(0u, 1u, tail_size > 0u);
|
|
585
|
+
|
|
586
|
+
for (var b: u32 = tid_in_col; b < full_blocks; b = b + THREADS_PER_COL_GEMV) {
|
|
587
|
+
let block = B_q4k[col * num_blocks + b];
|
|
588
|
+
let d = unpack_f16_lo(block.d_dmin);
|
|
589
|
+
let dmin = unpack_f16_hi(block.d_dmin);
|
|
590
|
+
let k_base = b * QK_K;
|
|
591
|
+
|
|
592
|
+
for (var sb: u32 = 0u; sb < 8u; sb = sb + 1u) {
|
|
593
|
+
let sm = get_scale_min_k4(block.scales, sb);
|
|
594
|
+
let scale = d * f32(sm.x);
|
|
595
|
+
let min_val = dmin * f32(sm.y);
|
|
596
|
+
let sb_base = sb * SUBBLOCK_SIZE;
|
|
597
|
+
|
|
598
|
+
for (var i: u32 = 0u; i < SUBBLOCK_SIZE; i = i + 4u) {
|
|
599
|
+
let k0 = k_base + sb_base + i;
|
|
600
|
+
|
|
601
|
+
let a0 = shared_A_buf[k0];
|
|
602
|
+
let a1 = shared_A_buf[k0 + 1u];
|
|
603
|
+
let a2 = shared_A_buf[k0 + 2u];
|
|
604
|
+
let a3 = shared_A_buf[k0 + 3u];
|
|
605
|
+
|
|
606
|
+
let q0 = get_q4(block.qs, sb_base + i);
|
|
607
|
+
let q1 = get_q4(block.qs, sb_base + i + 1u);
|
|
608
|
+
let q2 = get_q4(block.qs, sb_base + i + 2u);
|
|
609
|
+
let q3 = get_q4(block.qs, sb_base + i + 3u);
|
|
610
|
+
|
|
611
|
+
let w0 = scale * f32(q0) - min_val;
|
|
612
|
+
let w1 = scale * f32(q1) - min_val;
|
|
613
|
+
let w2 = scale * f32(q2) - min_val;
|
|
614
|
+
let w3 = scale * f32(q3) - min_val;
|
|
615
|
+
|
|
616
|
+
partial_sum = partial_sum + a0 * w0 + a1 * w1 + a2 * w2 + a3 * w3;
|
|
617
|
+
}
|
|
618
|
+
}
|
|
619
|
+
}
|
|
620
|
+
|
|
621
|
+
if (tail_size > 0u) {
|
|
622
|
+
let tail_block = full_blocks;
|
|
623
|
+
if (tail_block % THREADS_PER_COL_GEMV == tid_in_col) {
|
|
624
|
+
let block = B_q4k[col * num_blocks + tail_block];
|
|
625
|
+
let d = unpack_f16_lo(block.d_dmin);
|
|
626
|
+
let dmin = unpack_f16_hi(block.d_dmin);
|
|
627
|
+
let k_base = tail_block * QK_K;
|
|
628
|
+
|
|
629
|
+
for (var sb: u32 = 0u; sb < 8u; sb = sb + 1u) {
|
|
630
|
+
let sb_base = sb * SUBBLOCK_SIZE;
|
|
631
|
+
if (sb_base >= tail_size) {
|
|
632
|
+
break;
|
|
633
|
+
}
|
|
634
|
+
let sm = get_scale_min_k4(block.scales, sb);
|
|
635
|
+
let scale = d * f32(sm.x);
|
|
636
|
+
let min_val = dmin * f32(sm.y);
|
|
637
|
+
|
|
638
|
+
for (var i: u32 = 0u; i < SUBBLOCK_SIZE; i = i + 4u) {
|
|
639
|
+
let k0 = k_base + sb_base + i;
|
|
640
|
+
|
|
641
|
+
var a0: f32 = 0.0;
|
|
642
|
+
var a1: f32 = 0.0;
|
|
643
|
+
var a2: f32 = 0.0;
|
|
644
|
+
var a3: f32 = 0.0;
|
|
645
|
+
if (k0 < u.K) { a0 = shared_A_buf[k0]; }
|
|
646
|
+
if (k0 + 1u < u.K) { a1 = shared_A_buf[k0 + 1u]; }
|
|
647
|
+
if (k0 + 2u < u.K) { a2 = shared_A_buf[k0 + 2u]; }
|
|
648
|
+
if (k0 + 3u < u.K) { a3 = shared_A_buf[k0 + 3u]; }
|
|
649
|
+
|
|
650
|
+
let q0 = get_q4(block.qs, sb_base + i);
|
|
651
|
+
let q1 = get_q4(block.qs, sb_base + i + 1u);
|
|
652
|
+
let q2 = get_q4(block.qs, sb_base + i + 2u);
|
|
653
|
+
let q3 = get_q4(block.qs, sb_base + i + 3u);
|
|
654
|
+
|
|
655
|
+
let w0 = scale * f32(q0) - min_val;
|
|
656
|
+
let w1 = scale * f32(q1) - min_val;
|
|
657
|
+
let w2 = scale * f32(q2) - min_val;
|
|
658
|
+
let w3 = scale * f32(q3) - min_val;
|
|
659
|
+
|
|
660
|
+
partial_sum = partial_sum + a0 * w0 + a1 * w1 + a2 * w2 + a3 * w3;
|
|
661
|
+
}
|
|
662
|
+
}
|
|
663
|
+
}
|
|
664
|
+
}
|
|
665
|
+
}
|
|
666
|
+
|
|
667
|
+
// Reduction within each column's 8 threads
|
|
668
|
+
multicol_sums_sa[local_id] = partial_sum;
|
|
669
|
+
workgroupBarrier();
|
|
670
|
+
|
|
671
|
+
if (tid_in_col == 0u && is_valid) {
|
|
672
|
+
var final_sum: f32 = 0.0;
|
|
673
|
+
let base = col_in_wg * THREADS_PER_COL_GEMV;
|
|
674
|
+
for (var i: u32 = 0u; i < THREADS_PER_COL_GEMV; i = i + 1u) {
|
|
675
|
+
final_sum = final_sum + multicol_sums_sa[base + i];
|
|
676
|
+
}
|
|
677
|
+
C[col] = final_sum * u.alpha;
|
|
678
|
+
}
|
|
679
|
+
}
|
|
680
|
+
|
|
681
|
+
// ============================================================================
|
|
682
|
+
// Optimised GEMV: shared-A cooperative load + fast nibble extraction.
|
|
683
|
+
// Combines the shared A buffer from main_multicol_shared with the direct
|
|
684
|
+
// word extraction from main_multicol_fast for maximum M=1 decode throughput.
|
|
685
|
+
// Use when K <= SHARED_A_MAX (compile-time override, default 3584).
|
|
686
|
+
// ============================================================================
|
|
687
|
+
var<workgroup> gemv_shared_A: array<f32, SHARED_A_MAX>;
|
|
688
|
+
var<workgroup> gemv_sums: array<f32, MAX_WORKGROUP_SIZE>;
|
|
689
|
+
|
|
690
|
+
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
691
|
+
fn main_gemv(
|
|
692
|
+
@builtin(local_invocation_id) lid: vec3<u32>,
|
|
693
|
+
@builtin(workgroup_id) wg_id: vec3<u32>,
|
|
694
|
+
@builtin(subgroup_invocation_id) sg_id: u32,
|
|
695
|
+
@builtin(subgroup_size) sg_size: u32
|
|
696
|
+
) {
|
|
697
|
+
let local_id = lid.x;
|
|
698
|
+
|
|
699
|
+
// Cooperative load: all 256 threads fill shared A from global memory once.
|
|
700
|
+
for (var idx: u32 = local_id; idx < u.K; idx = idx + WORKGROUP_SIZE) {
|
|
701
|
+
gemv_shared_A[idx] = A[idx];
|
|
702
|
+
}
|
|
703
|
+
workgroupBarrier();
|
|
704
|
+
|
|
705
|
+
let col_in_wg = local_id / THREADS_PER_COL_GEMV;
|
|
706
|
+
let tid_in_col = local_id % THREADS_PER_COL_GEMV;
|
|
707
|
+
let col = wg_id.x * COLS_PER_WG + col_in_wg;
|
|
708
|
+
let is_valid = col < u.N;
|
|
709
|
+
|
|
710
|
+
var partial_sum: f32 = 0.0;
|
|
711
|
+
|
|
712
|
+
if (is_valid) {
|
|
713
|
+
let num_blocks = u.num_blocks_per_row;
|
|
714
|
+
let tail_size = u.K & 255u;
|
|
715
|
+
let full_blocks = num_blocks - select(0u, 1u, tail_size > 0u);
|
|
716
|
+
|
|
717
|
+
for (var b: u32 = tid_in_col; b < full_blocks; b = b + THREADS_PER_COL_GEMV) {
|
|
718
|
+
let block = B_q4k[col * num_blocks + b];
|
|
719
|
+
let d = unpack_f16_lo(block.d_dmin);
|
|
720
|
+
let dmin = unpack_f16_hi(block.d_dmin);
|
|
721
|
+
let k_base = b * QK_K;
|
|
722
|
+
|
|
723
|
+
for (var sb: u32 = 0u; sb < 8u; sb = sb + 1u) {
|
|
724
|
+
let sm = get_scale_min_k4(block.scales, sb);
|
|
725
|
+
let scale = d * f32(sm.x);
|
|
726
|
+
let min_val = dmin * f32(sm.y);
|
|
727
|
+
let sb_base = sb * SUBBLOCK_SIZE;
|
|
728
|
+
|
|
729
|
+
// Fast nibble extraction: even sub-blocks use lower nibble,
|
|
730
|
+
// odd sub-blocks use upper nibble.
|
|
731
|
+
let chunk = sb >> 1u;
|
|
732
|
+
let nibble_shift = (sb & 1u) * 4u;
|
|
733
|
+
let word_base = chunk * 8u;
|
|
734
|
+
|
|
735
|
+
for (var i: u32 = 0u; i < SUBBLOCK_SIZE; i = i + 4u) {
|
|
736
|
+
let k0 = k_base + sb_base + i;
|
|
737
|
+
|
|
738
|
+
let a0 = gemv_shared_A[k0];
|
|
739
|
+
let a1 = gemv_shared_A[k0 + 1u];
|
|
740
|
+
let a2 = gemv_shared_A[k0 + 2u];
|
|
741
|
+
let a3 = gemv_shared_A[k0 + 3u];
|
|
742
|
+
|
|
743
|
+
let word = block.qs[word_base + (i >> 2u)];
|
|
744
|
+
let q0 = (word >> nibble_shift) & 0xFu;
|
|
745
|
+
let q1 = (word >> (nibble_shift + 8u)) & 0xFu;
|
|
746
|
+
let q2 = (word >> (nibble_shift + 16u)) & 0xFu;
|
|
747
|
+
let q3 = (word >> (nibble_shift + 24u)) & 0xFu;
|
|
748
|
+
|
|
749
|
+
let w0 = scale * f32(q0) - min_val;
|
|
750
|
+
let w1 = scale * f32(q1) - min_val;
|
|
751
|
+
let w2 = scale * f32(q2) - min_val;
|
|
752
|
+
let w3 = scale * f32(q3) - min_val;
|
|
753
|
+
|
|
754
|
+
partial_sum = partial_sum + a0 * w0 + a1 * w1 + a2 * w2 + a3 * w3;
|
|
755
|
+
}
|
|
756
|
+
}
|
|
757
|
+
}
|
|
758
|
+
|
|
759
|
+
if (tail_size > 0u) {
|
|
760
|
+
let tail_block = full_blocks;
|
|
761
|
+
if (tail_block % THREADS_PER_COL_GEMV == tid_in_col) {
|
|
762
|
+
let block = B_q4k[col * num_blocks + tail_block];
|
|
763
|
+
let d = unpack_f16_lo(block.d_dmin);
|
|
764
|
+
let dmin = unpack_f16_hi(block.d_dmin);
|
|
765
|
+
let k_base = tail_block * QK_K;
|
|
766
|
+
|
|
767
|
+
for (var sb: u32 = 0u; sb < 8u; sb = sb + 1u) {
|
|
768
|
+
let sb_base = sb * SUBBLOCK_SIZE;
|
|
769
|
+
if (sb_base >= tail_size) {
|
|
770
|
+
break;
|
|
771
|
+
}
|
|
772
|
+
let sm = get_scale_min_k4(block.scales, sb);
|
|
773
|
+
let scale = d * f32(sm.x);
|
|
774
|
+
let min_val = dmin * f32(sm.y);
|
|
775
|
+
|
|
776
|
+
let chunk = sb >> 1u;
|
|
777
|
+
let nibble_shift = (sb & 1u) * 4u;
|
|
778
|
+
let word_base = chunk * 8u;
|
|
779
|
+
|
|
780
|
+
for (var i: u32 = 0u; i < SUBBLOCK_SIZE; i = i + 4u) {
|
|
781
|
+
let k0 = k_base + sb_base + i;
|
|
782
|
+
|
|
783
|
+
var a0: f32 = 0.0;
|
|
784
|
+
var a1: f32 = 0.0;
|
|
785
|
+
var a2: f32 = 0.0;
|
|
786
|
+
var a3: f32 = 0.0;
|
|
787
|
+
if (k0 < u.K) { a0 = gemv_shared_A[k0]; }
|
|
788
|
+
if (k0 + 1u < u.K) { a1 = gemv_shared_A[k0 + 1u]; }
|
|
789
|
+
if (k0 + 2u < u.K) { a2 = gemv_shared_A[k0 + 2u]; }
|
|
790
|
+
if (k0 + 3u < u.K) { a3 = gemv_shared_A[k0 + 3u]; }
|
|
791
|
+
|
|
792
|
+
let word = block.qs[word_base + (i >> 2u)];
|
|
793
|
+
let q0 = (word >> nibble_shift) & 0xFu;
|
|
794
|
+
let q1 = (word >> (nibble_shift + 8u)) & 0xFu;
|
|
795
|
+
let q2 = (word >> (nibble_shift + 16u)) & 0xFu;
|
|
796
|
+
let q3 = (word >> (nibble_shift + 24u)) & 0xFu;
|
|
797
|
+
|
|
798
|
+
let w0 = scale * f32(q0) - min_val;
|
|
799
|
+
let w1 = scale * f32(q1) - min_val;
|
|
800
|
+
let w2 = scale * f32(q2) - min_val;
|
|
801
|
+
let w3 = scale * f32(q3) - min_val;
|
|
802
|
+
|
|
803
|
+
partial_sum = partial_sum + a0 * w0 + a1 * w1 + a2 * w2 + a3 * w3;
|
|
804
|
+
}
|
|
805
|
+
}
|
|
806
|
+
}
|
|
807
|
+
}
|
|
808
|
+
}
|
|
809
|
+
|
|
810
|
+
gemv_sums[local_id] = partial_sum;
|
|
811
|
+
workgroupBarrier();
|
|
812
|
+
|
|
813
|
+
if (tid_in_col == 0u && is_valid) {
|
|
814
|
+
var final_sum: f32 = 0.0;
|
|
815
|
+
let base = col_in_wg * THREADS_PER_COL_GEMV;
|
|
816
|
+
for (var i: u32 = 0u; i < THREADS_PER_COL_GEMV; i = i + 1u) {
|
|
817
|
+
final_sum = final_sum + gemv_sums[base + i];
|
|
818
|
+
}
|
|
819
|
+
C[col] = final_sum * u.alpha;
|
|
820
|
+
}
|
|
821
|
+
}
|
|
@@ -0,0 +1,174 @@
|
|
|
1
|
+
// Fused Q4_K prefill matmul.
|
|
2
|
+
//
|
|
3
|
+
// Workgroup topology:
|
|
4
|
+
// - a small tile of rows (TILE_M)
|
|
5
|
+
// - multiple output columns per workgroup (COLS_PER_WG)
|
|
6
|
+
// - a few threads cooperate on each column (THREADS_PER_COL)
|
|
7
|
+
//
|
|
8
|
+
// This preserves Q4_K block traversal while loading each row's activation
|
|
9
|
+
// block once into workgroup memory for reuse across columns.
|
|
10
|
+
|
|
11
|
+
const QK_K: u32 = 256u;
|
|
12
|
+
const SUBBLOCK_SIZE: u32 = 32u;
|
|
13
|
+
|
|
14
|
+
const MAX_TILE_M: u32 = 4u;
|
|
15
|
+
const MAX_COLS_PER_WG: u32 = 8u;
|
|
16
|
+
const MAX_THREADS_PER_COL: u32 = 4u;
|
|
17
|
+
const MAX_WORKGROUP_X: u32 = MAX_COLS_PER_WG * MAX_THREADS_PER_COL;
|
|
18
|
+
const MAX_SHARED_A: u32 = MAX_TILE_M * QK_K;
|
|
19
|
+
const MAX_PARTIALS: u32 = MAX_TILE_M * MAX_COLS_PER_WG * MAX_THREADS_PER_COL;
|
|
20
|
+
|
|
21
|
+
override WORKGROUP_X: u32 = 32u;
|
|
22
|
+
override TILE_M: u32 = 4u;
|
|
23
|
+
override COLS_PER_WG: u32 = 8u;
|
|
24
|
+
override THREADS_PER_COL: u32 = 4u;
|
|
25
|
+
|
|
26
|
+
struct Uniforms {
|
|
27
|
+
M: u32,
|
|
28
|
+
N: u32,
|
|
29
|
+
K: u32,
|
|
30
|
+
alpha: f32,
|
|
31
|
+
num_blocks_per_row: u32,
|
|
32
|
+
_pad0: u32,
|
|
33
|
+
_pad1: u32,
|
|
34
|
+
_pad2: u32,
|
|
35
|
+
}
|
|
36
|
+
|
|
37
|
+
struct Q4KBlock {
|
|
38
|
+
d_dmin: u32,
|
|
39
|
+
scales: array<u32, 3>,
|
|
40
|
+
qs: array<u32, 32>,
|
|
41
|
+
}
|
|
42
|
+
|
|
43
|
+
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
44
|
+
@group(0) @binding(1) var<storage, read> A: array<f32>;
|
|
45
|
+
@group(0) @binding(2) var<storage, read> B_q4k: array<Q4KBlock>;
|
|
46
|
+
@group(0) @binding(3) var<storage, read_write> C: array<f32>;
|
|
47
|
+
|
|
48
|
+
var<workgroup> shared_A: array<f32, MAX_SHARED_A>;
|
|
49
|
+
var<workgroup> partial_sums: array<f32, MAX_PARTIALS>;
|
|
50
|
+
|
|
51
|
+
fn unpack_f16_lo(packed: u32) -> f32 {
|
|
52
|
+
return unpack2x16float(packed).x;
|
|
53
|
+
}
|
|
54
|
+
|
|
55
|
+
fn unpack_f16_hi(packed: u32) -> f32 {
|
|
56
|
+
return unpack2x16float(packed).y;
|
|
57
|
+
}
|
|
58
|
+
|
|
59
|
+
fn get_scale_byte(scales: array<u32, 3>, byte_idx: u32) -> u32 {
|
|
60
|
+
let word_idx = byte_idx / 4u;
|
|
61
|
+
let byte_in_word = byte_idx % 4u;
|
|
62
|
+
return (scales[word_idx] >> (byte_in_word * 8u)) & 0xFFu;
|
|
63
|
+
}
|
|
64
|
+
|
|
65
|
+
fn get_scale_min_k4(scales: array<u32, 3>, j: u32) -> vec2<u32> {
|
|
66
|
+
var sc: u32;
|
|
67
|
+
var mn: u32;
|
|
68
|
+
|
|
69
|
+
if (j < 4u) {
|
|
70
|
+
sc = get_scale_byte(scales, j) & 63u;
|
|
71
|
+
mn = get_scale_byte(scales, j + 4u) & 63u;
|
|
72
|
+
} else {
|
|
73
|
+
let q_j = get_scale_byte(scales, j + 4u);
|
|
74
|
+
let q_lo = get_scale_byte(scales, j - 4u);
|
|
75
|
+
let q_hi = get_scale_byte(scales, j);
|
|
76
|
+
sc = (q_j & 0xFu) | ((q_lo >> 6u) << 4u);
|
|
77
|
+
mn = (q_j >> 4u) | ((q_hi >> 6u) << 4u);
|
|
78
|
+
}
|
|
79
|
+
return vec2<u32>(sc, mn);
|
|
80
|
+
}
|
|
81
|
+
|
|
82
|
+
fn get_q4(qs: array<u32, 32>, idx: u32) -> u32 {
|
|
83
|
+
let chunk = idx / 64u;
|
|
84
|
+
let pos_in_chunk = idx % 64u;
|
|
85
|
+
let use_upper = pos_in_chunk >= 32u;
|
|
86
|
+
let byte_in_range = select(pos_in_chunk, pos_in_chunk - 32u, use_upper);
|
|
87
|
+
let byte_idx = chunk * 32u + byte_in_range;
|
|
88
|
+
|
|
89
|
+
let word_idx = byte_idx / 4u;
|
|
90
|
+
let byte_in_word = byte_idx % 4u;
|
|
91
|
+
let byte_val = (qs[word_idx] >> (byte_in_word * 8u)) & 0xFFu;
|
|
92
|
+
|
|
93
|
+
if (use_upper) {
|
|
94
|
+
return (byte_val >> 4u) & 0xFu;
|
|
95
|
+
}
|
|
96
|
+
return byte_val & 0xFu;
|
|
97
|
+
}
|
|
98
|
+
|
|
99
|
+
@compute @workgroup_size(WORKGROUP_X, TILE_M, 1)
|
|
100
|
+
fn main(
|
|
101
|
+
@builtin(local_invocation_id) lid: vec3<u32>,
|
|
102
|
+
@builtin(workgroup_id) wg_id: vec3<u32>
|
|
103
|
+
) {
|
|
104
|
+
if (
|
|
105
|
+
TILE_M > MAX_TILE_M
|
|
106
|
+
|| COLS_PER_WG > MAX_COLS_PER_WG
|
|
107
|
+
|| THREADS_PER_COL > MAX_THREADS_PER_COL
|
|
108
|
+
|| WORKGROUP_X > MAX_WORKGROUP_X
|
|
109
|
+
|| COLS_PER_WG * THREADS_PER_COL != WORKGROUP_X
|
|
110
|
+
) {
|
|
111
|
+
return;
|
|
112
|
+
}
|
|
113
|
+
|
|
114
|
+
let row_in_tile = lid.y;
|
|
115
|
+
let row = wg_id.y * TILE_M + row_in_tile;
|
|
116
|
+
let row_valid = row < u.M;
|
|
117
|
+
let col_in_wg = lid.x / THREADS_PER_COL;
|
|
118
|
+
let tid_in_col = lid.x % THREADS_PER_COL;
|
|
119
|
+
let col = wg_id.x * COLS_PER_WG + col_in_wg;
|
|
120
|
+
let col_valid = col < u.N;
|
|
121
|
+
let is_valid = row_valid && col_valid;
|
|
122
|
+
let row_shared_base = row_in_tile * QK_K;
|
|
123
|
+
|
|
124
|
+
var partial_sum: f32 = 0.0;
|
|
125
|
+
|
|
126
|
+
for (var block_idx: u32 = 0u; block_idx < u.num_blocks_per_row; block_idx = block_idx + 1u) {
|
|
127
|
+
let block_k_base = block_idx * QK_K;
|
|
128
|
+
|
|
129
|
+
for (var elem_idx: u32 = lid.x; elem_idx < QK_K; elem_idx = elem_idx + WORKGROUP_X) {
|
|
130
|
+
let k = block_k_base + elem_idx;
|
|
131
|
+
var a_value: f32 = 0.0;
|
|
132
|
+
if (row_valid && k < u.K) {
|
|
133
|
+
a_value = A[row * u.K + k];
|
|
134
|
+
}
|
|
135
|
+
shared_A[row_shared_base + elem_idx] = a_value;
|
|
136
|
+
}
|
|
137
|
+
workgroupBarrier();
|
|
138
|
+
|
|
139
|
+
if (is_valid) {
|
|
140
|
+
let block = B_q4k[col * u.num_blocks_per_row + block_idx];
|
|
141
|
+
let d = unpack_f16_lo(block.d_dmin);
|
|
142
|
+
let dmin = unpack_f16_hi(block.d_dmin);
|
|
143
|
+
|
|
144
|
+
for (var sb: u32 = 0u; sb < 8u; sb = sb + 1u) {
|
|
145
|
+
let sm = get_scale_min_k4(block.scales, sb);
|
|
146
|
+
let scale = d * f32(sm.x);
|
|
147
|
+
let min_val = dmin * f32(sm.y);
|
|
148
|
+
let sb_base = sb * SUBBLOCK_SIZE;
|
|
149
|
+
|
|
150
|
+
for (var i: u32 = tid_in_col; i < SUBBLOCK_SIZE; i = i + THREADS_PER_COL) {
|
|
151
|
+
let elem = sb_base + i;
|
|
152
|
+
let a_value = shared_A[row_shared_base + elem];
|
|
153
|
+
let q = get_q4(block.qs, elem);
|
|
154
|
+
let w = scale * f32(q) - min_val;
|
|
155
|
+
partial_sum = partial_sum + a_value * w;
|
|
156
|
+
}
|
|
157
|
+
}
|
|
158
|
+
}
|
|
159
|
+
|
|
160
|
+
workgroupBarrier();
|
|
161
|
+
}
|
|
162
|
+
|
|
163
|
+
let partial_base = (row_in_tile * MAX_COLS_PER_WG + col_in_wg) * MAX_THREADS_PER_COL;
|
|
164
|
+
partial_sums[partial_base + tid_in_col] = partial_sum;
|
|
165
|
+
workgroupBarrier();
|
|
166
|
+
|
|
167
|
+
if (tid_in_col == 0u && is_valid) {
|
|
168
|
+
var final_sum: f32 = 0.0;
|
|
169
|
+
for (var i: u32 = 0u; i < THREADS_PER_COL; i = i + 1u) {
|
|
170
|
+
final_sum = final_sum + partial_sums[partial_base + i];
|
|
171
|
+
}
|
|
172
|
+
C[row * u.N + col] = final_sum * u.alpha;
|
|
173
|
+
}
|
|
174
|
+
}
|
|
@@ -25,6 +25,10 @@ export interface GatherOptions extends OutputBufferOptions {
|
|
|
25
25
|
transpose?: boolean;
|
|
26
26
|
/** Optional index offset into the token indices buffer. */
|
|
27
27
|
indexOffset?: number;
|
|
28
|
+
/** Total hidden size stored in the source embedding table (defaults to hiddenSize). */
|
|
29
|
+
inputHiddenSize?: number;
|
|
30
|
+
/** Hidden-dimension offset inside the source embedding row (defaults to 0). */
|
|
31
|
+
hiddenOffset?: number;
|
|
28
32
|
/** Optional indirect dispatch buffer for GPU-driven workgroup counts. */
|
|
29
33
|
indirectBuffer?: GPUBuffer | null;
|
|
30
34
|
/** Byte offset into indirect dispatch buffer (default: 0). */
|
|
@@ -31,6 +31,8 @@ async function _gather(
|
|
|
31
31
|
outputDtype,
|
|
32
32
|
transpose = false,
|
|
33
33
|
indexOffset = 0,
|
|
34
|
+
inputHiddenSize = hiddenSize,
|
|
35
|
+
hiddenOffset = 0,
|
|
34
36
|
indirectBuffer = null,
|
|
35
37
|
indirectOffset = 0,
|
|
36
38
|
} = options;
|
|
@@ -49,11 +51,18 @@ async function _gather(
|
|
|
49
51
|
throw new Error('[Gather] outputDtype=f16 requires shader-f16 support.');
|
|
50
52
|
}
|
|
51
53
|
|
|
52
|
-
const
|
|
54
|
+
const usesHiddenSlice = inputHiddenSize !== hiddenSize || hiddenOffset !== 0;
|
|
55
|
+
const requestedVec4 = usesHiddenSlice ? false : options.useVec4;
|
|
53
56
|
const wantsVec4 = requestedVec4 ?? true;
|
|
54
57
|
if (requestedVec4 === true && hiddenSize % 4 !== 0) {
|
|
55
58
|
throw new Error('[Gather] useVec4=true requires hiddenSize to be divisible by 4.');
|
|
56
59
|
}
|
|
60
|
+
if (!Number.isFinite(inputHiddenSize) || inputHiddenSize < hiddenSize) {
|
|
61
|
+
throw new Error('[Gather] inputHiddenSize must be >= hiddenSize.');
|
|
62
|
+
}
|
|
63
|
+
if (!Number.isFinite(hiddenOffset) || hiddenOffset < 0 || (hiddenOffset + hiddenSize) > inputHiddenSize) {
|
|
64
|
+
throw new Error('[Gather] hiddenOffset must select a valid hidden slice inside inputHiddenSize.');
|
|
65
|
+
}
|
|
57
66
|
|
|
58
67
|
const useF16Input = embeddingDtype === 'f16';
|
|
59
68
|
const useF16Output = outputDtype === 'f16';
|
|
@@ -62,6 +71,7 @@ async function _gather(
|
|
|
62
71
|
trace.embed(
|
|
63
72
|
`Gather: numTokens=${numTokens}, hiddenSize=${hiddenSize}, vocabSize=${vocabSize}, ` +
|
|
64
73
|
`transpose=${transpose}, indexOffset=${indexOffset}, ` +
|
|
74
|
+
`inputHiddenSize=${inputHiddenSize}, hiddenOffset=${hiddenOffset}, ` +
|
|
65
75
|
`embeddingDtype=${embeddingDtype}, outputDtype=${outputDtype}, ` +
|
|
66
76
|
`useF16Input=${useF16Input}, useF16Output=${useF16Output}`
|
|
67
77
|
);
|
|
@@ -84,9 +94,10 @@ async function _gather(
|
|
|
84
94
|
vocab_size: vocabSize,
|
|
85
95
|
transpose: transpose ? 1 : 0,
|
|
86
96
|
index_offset: indexOffset,
|
|
97
|
+
input_hidden_size: inputHiddenSize,
|
|
98
|
+
hidden_offset: hiddenOffset,
|
|
87
99
|
_pad0: 0,
|
|
88
100
|
_pad1: 0,
|
|
89
|
-
_pad2: 0,
|
|
90
101
|
};
|
|
91
102
|
|
|
92
103
|
const workgroups = indirectBuffer
|
|
@@ -16,9 +16,10 @@ struct Uniforms {
|
|
|
16
16
|
vocab_size: u32, // Vocabulary size (for bounds checking)
|
|
17
17
|
transpose: u32, // 1 if embeddings are [hidden_size, vocab_size] (GGUF layout), 0 otherwise
|
|
18
18
|
index_offset: u32, // Starting index into indices buffer
|
|
19
|
+
input_hidden_size: u32, // Source embedding row width before hidden slicing
|
|
20
|
+
hidden_offset: u32, // Starting hidden dimension inside each embedding row
|
|
19
21
|
_pad0: u32,
|
|
20
22
|
_pad1: u32,
|
|
21
|
-
_pad2: u32,
|
|
22
23
|
}
|
|
23
24
|
|
|
24
25
|
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
@@ -52,10 +53,11 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
|
|
|
52
53
|
// For GGUF layout [hidden_size, vocab_size]: offset = dim_idx * vocab_size + token_id
|
|
53
54
|
// For standard layout [vocab_size, hidden_size]: offset = token_id * hidden_size + dim_idx
|
|
54
55
|
var embed_offset: u32;
|
|
56
|
+
let source_dim = u.hidden_offset + dim_idx;
|
|
55
57
|
if (u.transpose == 1u) {
|
|
56
|
-
embed_offset =
|
|
58
|
+
embed_offset = source_dim * u.vocab_size + token_id;
|
|
57
59
|
} else {
|
|
58
|
-
embed_offset = token_id * u.
|
|
60
|
+
embed_offset = token_id * u.input_hidden_size + source_dim;
|
|
59
61
|
}
|
|
60
62
|
output[tid] = embeddings[embed_offset];
|
|
61
63
|
}
|