b10172

ggml-org/llama.cppb10172Jul 28, 2026by github-actions[bot]same day as b10171

What changed since b10171

low risk

WebGPU backend gains cross-architecture binding alias fixes plus corrections for >4GB buffer alignment and view offsets.

New features

  • Added an overlapped GLU variant in the WebGPU backend so the path works on all architectures rather than a subset.

Improvements

  • WebGPU binding overlap handling was reworked so aliased ranges are computed correctly across all architectures.
  • Buffer offset alignment is now correct for buffers larger than 4GB.
  • View offset handling was corrected, and the set_rows edge case is now handled.
  • Trace/diagnostic infrastructure used while debugging the WebGPU issues was cleaned up and simplified.

Bug fixes

  • Fixed binding alias issues that prevented the WebGPU backend from working on all architectures.
  • Fixed the recurrent-state-rollback test failure.
  • Fixed incorrect alignment of offsets into buffers exceeding 4GB.
  • Fixed incorrect view offsets in the WebGPU backend.
  • Fixed an edge case in set_rows handling.

Upgrade notes

This build is a WebGPU-focused bug-fix release and carries no config or API changes, so a straight binary swap is sufficient. WebGPU users on previously unsupported architectures should see the backend start working; note that deepseek32 is temporarily skipped in WebGPU testing, so validate that model separately. The OpenCL Adreno multi-stream fix from b10171 is unrelated and still required if you run Adreno GPUs.

Analyzed Oct 31, 2026

Upgrading?

Release notes summary

Fixed critical WebGPU binding alias issues that prevented support on all architectures and resolved a recurrent-state-rollback test failure, alongside fixes for large buffer alignment and view offsets.

Key Highlights

  • Fixed WebGPU binding alias issues to support all architectures
  • Fixed recurrent-state-rollback test
  • Fixed buffer alignment for buffers larger than 4GB
  • Fixed view offsets

Release Notes (excerpt)

<details open> ggml-webgpu: Fix some binding alias issues to support all archs, fix recurrent-state-rollback test (#25931) * Add overlap glu variant to support all archs, fix recurrent-state-rollback test * format * Fix all arch overlapped ranges * format * diagnose bus error on apple ci * More testing * more testing * more targeted testing * Fix bug in alignment for > 4gb buffer offsets * Fix bug in view offsets * Try avoiding multi_buffers * not fixed yet, more logging :(…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories