Issue evidence · Serve · ggml-org/llama.cpp

vulkan: coopmat FA silently falls back to scalar on Intel Xe2

GitHub issue: https://github.com/ggml-org/llama.cpp/issues/30503

Not run

On Intel Xe2 with a given driver, Vulkan flash attention silently falls back from coopmat to the scalar path.

What Badgr ran

Nothing was run: Intel Xe2 hardware is not available from Badgr's providers.

What came back

  • The report has no commands to run.

A different GPU would not test the Intel coopmat selection.

Checked 2026-10-14 in Badgr’s local development environment.

← All issue evidence