ONNX Runtime 发布: ONNX Runtime WebGPU Plugin EP v0.4.0

来源摘要

This release brings kernel performance improvements, expanded operator support, and reliability fixes to the ONNX Runtime WebGPU Plugin EP. ## Highlights ### Performance - Optimized `MatMulNBits` wide-tile execution using subgroup shuffle ([#31703](https://github.com/microsoft/onnxruntime/pull/31703)) - Supplied fused activation parameters to Conv/MatMul as uniforms, and fused eight more activations into Conv ([#32116](https://github.com/microsoft/onnxruntime/pull/32116), [#32117](https://github.com/microsoft/onnxruntime/pull/32117)) - Added fused activation support to the im2col Conv path ([#32185](https://github.com/microsoft/onnxruntime/pull/32185)) - Vectorized `Split` when every output segment is vec4-aligned ([#32251](https://github.com/microsoft/onnxruntime/pull/32251)) - Shared subgroup matrix MatMul with pointwise Conv, and pinned subgroup size to 32 for subgroup-matrix MatMul/Gemm ([#32304](https://github.com/microsoft/onnxruntime/pull/32304), [#32306](https://github.com/microsoft/onnxruntime/pull/32306)) - Selected the pooling path by occupancy rather than output size ([#32313](https://github.com/microsoft/onnxruntime/pull/32313)) - Prepacked Conv weights for the im2col-matmul path ([#32420](https://github.com/microsoft/onnxruntime/pull/32420)) - Scaled Dawn pipeline compilation workers with CPU count ([#29820](https://github.com/microsoft/onnxruntime/pull/29820)) - Added PagedAttention metadata and GPT-OSS support, and continued optimizing PagedAttention (2/n) ([#32277](https://github.com/microsoft/onnxruntime/pull/32277), [#31727](https://github.com/microsoft/onnxruntime/pull/31727)) - Added int8 KV cache block quantization ([#32284](https://github.com/microsoft/onnxruntime/pull/32284)) ### New Operator and Model Support - Implemented WebGPU subgroup-size-c