llama.cpp fixes a SYCL build failure in pre-release b11835
The Intel GPU backend gets a targeted compilation fix. No LocalXPU speed or stability improvement has been measured.
llama.cpp pre-release b11835 includes a fix for a SYCL build failure, according to the upstream pull request. SYCL is one of llama.cpp's backends for Intel GPUs. If you build that backend from source and encounter this failure, b11835 is a concrete revision to check; the change alone is not a reason to replace a working installation.
The patch passes the active device into a matrix-multiplication helper used when checking SYCL graph compatibility. It updates the helper's call to a device-aware kernel-selection function and marks the device argument unused when oneDNN support is disabled. This is a narrow code correction, not a published benchmark.
LocalXPU has not built or benchmarked b11835 on the dual-B60 machine. The upstream description identifies a build failure but does not establish a speed gain, a runtime stability improvement or a complete list of affected devices.
Read the upstream change and patch.
The pre-release was published on . Read the upstream release notes for the available artifacts and full changes.
Before changing a working setup
Keep the version you're running and its configuration so you can return to it. Test the new build against the same prompts and tools before making it your default.
Our vLLM measurements do not validate this llama.cpp build. For that separate configuration, see the two-B60 setup guide. Its results apply to the versions recorded there.