# ONNX Runtime plugin-ep-webgpu/v0.3.0 — ONNX Runtime WebGPU Plugin EP v0.3.0

- Product: ONNX Runtime (https://whatsnew.fyi/product/onnx-runtime)
- Vendor: Microsoft
- Date: 2026-08-24
- Version: plugin-ep-webgpu/v0.3.0
- Original notes: https://github.com/microsoft/onnxruntime/releases/tag/plugin-ep-webgpu/v0.3.0
- Permalink: https://whatsnew.fyi/product/onnx-runtime/releases/plugin-ep-webgpu-v0.3.0

What's New is an index, not a publisher: every entry below links to the vendor's own release notes, which are the authoritative source. Entries are labelled where they are hand-curated sample data, pre-releases, or drawn from a secondary source such as a developer blog.

Reuse: the summaries, labels and curation here are © What's New. Quote freely with attribution and a link back; wholesale republication of the corpus is not permitted — terms: https://whatsnew.fyi/terms. The vendors' own release notes remain their publishers'.

---

- **added** — Initial PagedAttention support, MRotaryEmbedding, GRU, DFT, PRelu, HardSwish, Trilu, Max and Min, and MatMulBnb4 operators
- **added** — Integer support across common operators, including 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/uint32 for CumSum and Tile
- **added** — 2-bit GatherBlockQuantized support and ONNX 1.22 with opset 27 integration
- **added** — Quantized KV cache support and expanded GQA with sliding-window cache, batched right-padded prompts, and FlashAttention graph capture for Gemma 4
- **fixed** — GQA and quantized KV cache correctness for rotary batched prefill, split-reduce decode, and per-batch sequence lengths
- **added** — Deferred dispatch to parallelize shader compilation during cold start
- **added** — Intel subgroup-matrix MatMul and FP16 Gemm kernels with batched-B and odd-N weight support and additional tiling optimizations
- **changed** — Softmax and non-flash Attention to use an online softmax algorithm and tuned FlashAttention decode
- **added** — Robustness provider option, configurable pending-dispatch limit, and WebGPU buffer-cache options exposed through the JavaScript API
- **changed** — Validation and bounds handling in GridSample, LSTM, DynamicQuantizedLSTM, Slice, and GatherBlockQuantized
- **fixed** — WebGPU data-transfer callbacks on Windows x86
- **fixed** — Profiling timestamp alignment with the ONNX Runtime profiler
- **added** — Device-free compile-only sessions for offline graph transformation
- **removed** — Dynamic Duktape/Node WGSL-generation path in favor of generated shader sources
- **changed** — Upgraded Dawn and fixed parallel-build races
- **changed** — Refreshed release and packaging documentation and enabled WebGPU CI on plugin release branches

ONNX Runtime WebGPU Plugin EP 0.3.0 expands model and data-type coverage, improves generative-model performance, and strengthens configuration, reliability, and release tooling.

These release notes were drafted with AI assistance.

##### Highlights

###### Model and operator coverage

- Added initial PagedAttention support, MRotaryEmbedding, GRU, DFT, PRelu, HardSwish, Trilu, Max and Min, and MatMulBnb4. ([#31611](https://github.com/microsoft/onnxruntime/pull/31611), [#31976](https://github.com/microsoft/onnxruntime/pull/31976), [#29840](https://github.com/microsoft/onnxruntime/pull/29840), [#29454](https://github.com/microsoft/onnxruntime/pull/29454), [#30512](https://github.com/microsoft/onnxruntime/pull/30512), [#29828](https://github.com/microsoft/onnxruntime/pull/29828), [#29845](https://github.com/microsoft/onnxruntime/pull/29845), [#29833](https://github.com/microsoft/onnxruntime/pull/29833), [#29587](https://github.com/microsoft/onnxruntime/pull/29587))
- Expanded integer support across common operators, including `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`/`uint32` for CumSum and Tile. ([#28804](https://github.com/microsoft/onnxruntime/pull/28804), [#29392](https://github.com/microsoft/onnxruntime/pull/29392), [#29830](https://github.com/microsoft/onnxruntime/pull/29830), [#29834](https://github.com/microsoft/onnxruntime/pull/29834), [#29839](https://github.com/microsoft/onnxruntime/pull/29839), [#29844](https://github.com/microsoft/onnxruntime/pull/29844), [#29847](https://github.com/microsoft/onnxruntime/pull/29847), [#29854](https://github.com/microsoft/onnxruntime/pull/29854), [#29861](https://github.com/microsoft/onnxruntime/pull/29861), [#29897](https://github.com/microsoft/onnxruntime/pull/29897), [#31049](https://github.com/microsoft/onnxruntime/pull/31049), [#31702](https://github.com/microsoft/onnxruntime/pull/31702), [#31709](https://github.com/microsoft/onnxruntime/pull/31709), [#31714](https://github.com/microsoft/onnxruntime/pull/31714))
- Added 2-bit `GatherBlockQuantized` support and integrated ONNX 1.22 with opset 27. ([#29054](https://github.com/microsoft/onnxruntime/pull/29054), [#28754](https://github.com/microsoft/onnxruntime/pull/28754))

###### Generative models

- Added quantized KV cache support and expanded GQA with sliding-window cache, batched right-padded prompts, and FlashAttention graph capture for Gemma 4. ([#28059](https://github.com/microsoft/onnxruntime/pull/28059), [#29236](https://github.com/microsoft/onnxruntime/pull/29236), [#29247](https://github.com/microsoft/onnxruntime/pull/29247), [#29904](https://github.com/microsoft/onnxruntime/pull/29904))
- Fixed GQA and quantized KV cache correctness for rotary batched prefill, split-reduce decode, and per-batch sequence lengths. ([#29002](https://github.com/microsoft/onnxruntime/pull/29002), [#29593](https://github.com/microsoft/onnxruntime/pull/29593), [#29752](https://github.com/microsoft/onnxruntime/pull/29752))

###### Performance

- Added deferred dispatch to parallelize shader compilation during cold start. ([#29557](https://github.com/microsoft/onnxruntime/pull/29557))
- Added Intel subgroup-matrix MatMul and FP16 Gemm kernels, then extended them with batched-B and odd-N weight support and additional tiling optimizations. ([#29271](https://github.com/microsoft/onnxruntime/pull/29271), [#29592](https://github.com/microsoft/onnxruntime/pull/29592), [#29749](https://github.com/microsoft/onnxruntime/pull/29749), [#29813](https://github.com/microsoft/onnxruntime/pull/29813), [#29893](https://github.com/microsoft/onnxruntime/pull/29893), [#30514](https://github.com/microsoft/onnxruntime/pull/30514))
- Switched Softmax and non-flash Attention to an online softmax algorithm and tuned FlashAttention decode. ([#29694](https://github.com/microsoft/onnxruntime/pull/29694), [#29724](https://github.com/m

_[Truncated at 4000 characters — full notes: https://github.com/microsoft/onnxruntime/releases/tag/plugin-ep-webgpu/v0.3.0]_