ONNX Runtime 发布: ONNX Runtime WebGPU Plugin EP v0.2.1

来源摘要

## Highlights - **Major performance work for attention-heavy LLMs.** - FlashAttention decode kernels were fused and extended for any sequence length ([#28389](https://github.com/microsoft/onnxruntime/pull/28389)). - FlashAttention prefill shared-memory path was generalized ([#28520](https://github.com/microsoft/onnxruntime/pull/28520)). - Dynamic `max_k_step` was enabled for NVIDIA ([#28511](https://github.com/microsoft/onnxruntime/pull/28511)). - QKV bias support was added for FlashAttention in MultiHeadAttention ([#28380](https://github.com/microsoft/onnxruntime/pull/28380)). - M4 Max-specific FlashAttention optimization landed ([#27780](https://github.com/microsoft/onnxruntime/pull/27780)). - **Qwen3 and Gemma 4 model-path improvements.** - QKV and MLP fusions for Qwen3-style models ([#28280](https://github.com/microsoft/onnxruntime/pull/28280)). - Q/K RMSNorm fusion into GroupQueryAttention for Qwen3-style models ([#28484](https://github.com/microsoft/onnxruntime/pull/28484)). - Opset 24 and KV-shared decoder layer support for Gemma 4 ([#28501](https://github.com/microsoft/onnxruntime/pull/28501)). - GroupQueryAttention now supports optional present-key/value outputs ([#28242](https://github.com/microsoft/onnxruntime/pull/28242)). - **LinearAttention and quantized-path optimizations.** - LinearAttention subgroup optimizations and larger `tile_v` with subgroup support ([#28412](https://github.com/microsoft/onnxruntime/pull/28412), [#28519](https://github.com/microsoft/onnxruntime/pull/28519)). - GatherBlockQuantized gained 2-bit support ([#28530](https://github.com/microsoft/onnxruntime/pull/28530)). - **Reliability and hardening fixes.** - Fixes for out-of-bounds read risks in GatherBlockQuantized and Pad ([#28718](https://github.com/microsoft/onnxruntime/pull/28718