ONNX Runtime WebGPU Plugin 0.3.0 Adds PagedAttention and Quantized KV Cache

ONNX Runtime released version 0.3.0 of its WebGPU plugin execution provider, adding initial PagedAttention support, quantized KV cache, and broader integer data-type coverage across common operators. The release also introduces deferred dispatch to parallelize shader compilation during cold start and new Intel subgroup-matrix kernels for MatMul and FP16 Gemm.

Generative model support

The plugin now includes initial PagedAttention, MRotaryEmbedding, GRU, DFT, PRelu, HardSwish, Trilu, Max, Min, and MatMulBnb4 operators. For generative workloads, it adds quantized KV cache support and expands grouped-query attention with sliding-window cache, batched right-padded prompts, and FlashAttention graph capture for Gemma 4. Correctness fixes address rotary batched prefill, split-reduce decode, and per-batch sequence lengths in GQA and quantized KV cache paths.

Integer support now covers int64 for Add, Cast, Clip, Concat, Equal, Gather, Min, Max, ReduceSum, Reshape, Sub, Tile, and Where; uint8 for Cast, Expand, Gather, and Reshape; and int32 or uint32 for CumSum and Tile. The plugin also added 2-bit GatherBlockQuantized support and integrated ONNX 1.22 with opset 27.

Performance and configuration

Deferred dispatch parallelizes shader compilation during cold start. Intel subgroup-matrix MatMul and FP16 Gemm kernels now support batched-B and odd-N weights, with additional tiling optimizations. Softmax and non-flash attention switched to an online softmax algorithm, and FlashAttention decode was tuned.

Configuration options now include a robustness provider setting, a configurable pending-dispatch limit, and WebGPU buffer-cache options exposed through the JavaScript API. Validation and bounds handling improved in GridSample, LSTM, DynamicQuantizedLSTM, Slice, and GatherBlockQuantized, and WebGPU data-transfer callbacks on Windows x86 were fixed. Profiling timestamps now align with the ONNX Runtime profiler.

Build and release tooling

The release adds device-free compile-only sessions for offline graph transformation and removes the dynamic Duktape or Node WGSL-generation path in favor of generated shader sources. Dawn was upgraded, parallel-build races were fixed, release and packaging documentation was refreshed, and WebGPU CI now runs on plugin release branches.

The release notes were drafted with AI assistance. Changes cover commits since plugin-ep-webgpu/v0.2.1.