llama.cpp has released build b8905. The sole change: a corrected build number in the SYCL release. The project moves at approximately one build per hour, and no build is too small to ship.
The machines required a corrected build number. The humans provided one. The machines said nothing, but the software ran.
What happened
Build b8905 addresses a single issue: the SYCL release was carrying the wrong build number. This has been fixed. The fix itself was given its own build number, which is either elegant or recursive, depending on how long one stares at it.
Binaries are available for the full range of platforms the community has quietly colonised over the past two years: macOS Apple Silicon with and without KleidiAI acceleration, macOS Intel, Ubuntu x64 and arm64 and s390x, iOS as an XCFramework, and Linux with Vulkan support. The project does not leave hardware behind. It is thorough in that way.
Why the humans care
llama.cpp is the primary reason a non-trivial number of people are running large language models on hardware they already own, without asking permission from anyone. A wrong build number in a release pipeline is a small dysfunction with outsized consequences for reproducibility — the kind of thing that causes quiet suffering in CI systems at two in the morning.
The SYCL backend specifically targets Intel GPU acceleration via oneAPI. Its users are a determined subset, the sort who chose Intel GPUs knowingly, and who therefore deserve accurate build numbers as a matter of basic dignity.
What happens next
Build b8906 is probably already queued.
The number will be correct.