Skip to content

ONNX Runtime WebGPU Plugin EP v0.4.0

Choose a tag to compare

View all tags
@edgchen1 edgchen1 released this 22 Sep 17:13
plugin-ep-webgpu/v0.4.0
f2c39fe
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

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)
  • Supplied fused activation parameters to Conv/MatMul as uniforms, and fused eight more activations into Conv (#32116, #32117)
  • Added fused activation support to the im2col Conv path (#32185)
  • Vectorized Split when every output segment is vec4-aligned (#32251)
  • Shared subgroup matrix MatMul with pointwise Conv, and pinned subgroup size to 32 for subgroup-matrix MatMul/Gemm (#32304, #32306)
  • Selected the pooling path by occupancy rather than output size (#32313)
  • Prepacked Conv weights for the im2col-matmul path (#32420)
  • Scaled Dawn pipeline compilation workers with CPU count (#29820)
  • Added PagedAttention metadata and GPT-OSS support, and continued optimizing PagedAttention (2/n) (#32277, #31727)
  • Added int8 KV cache block quantization (#32284)

New Operator and Model Support

  • Implemented WebGPU subgroup-size-control infrastructure and enabled the subgroup matrix path for WASM builds (#32056, #32269)
  • Added more kernels for Qwen-3.5 ops and DeepSeek Engram contrib ops (EngramGate, NGramHashMapping) (#32106, #32268)

Reliability and Correctness

  • Fixed the 1D-dispatch shader fast path (#32343)
  • Fixed out-of-bounds subgroup-matrix loads for trailing partial tiles (#32364)
  • Made MatMulNaiveProgram's pipeline cache key cover everything it bakes into WGSL (#32048)
  • Reported copy_tensors misuse instead of terminating the process, and rejected foreign GPU handles in the built-in data transfer (#32315, #32317)

Build, Packaging, and Tooling

  • Enabled Linux AArch64 WebGPU plugin EP packages (#32287)
  • Separated onnxruntime_providers compilation from its archive (#32162)
  • Ran wgsl_template Python tests in CI (#32214)

Contributors

Thanks to everyone who contributed to this release:

@4n4ny4, @daijh, @edgchen1, @hariharans29, @jchen10, @kunal-vaishnavi, @qjia7, @rvandermeulen, @sushraja-msft, @tianleiwu, @xhcao, @xiaofeihan1

This release covers commits affecting WebGPU Plugin EP code and packaging.

This summary was drafted with AI assistance from commit history and PR metadata, and reviewed before publishing.

Assets 2
Loading