llama.cpp-hip (version bump)

Unofficial PKGBUILD that tracks upstream llama.cpp releases more closely than the AUR llama.cpp-hip package. Based on the AUR package maintained by Orion-zhen (aur/llama.cpp-hip), with the version and source checksums bumped to the latest upstream tag.

Why this exists

As of 2026-08-11, the AUR llama.cpp-hip package (the ROCm/HIP build of llama.cpp for AMD GPUs) is pinned to b10216, last updated 2026-08-01, and is flagged out-of-date. Upstream llama.cpp has since moved to b10362.

In between those two versions, upstream merged support for Meta's Muse Glimmer model (ggml-org/llama.cpp#26841, merged 2026-08-10), first shipping in release b10353. Anything older (including the current AUR llama.cpp-hip build) fails to load Muse Glimmer GGUF files with:

error loading model: unknown model architecture: 'muse-glimmer'

(see ggml-org/llama.cpp#26858 for the same failure on a different backend).

The plain aur/llama.cpp package (CPU-only, no ROCm/CUDA/Vulkan) already tracks b10362 and does support Muse Glimmer. But there is currently no up-to-date AUR package that gives AMD ROCm users Muse Glimmer support with GPU acceleration — the CPU-only package works but is far too slow for a 30B model, and the ROCm package hasn't caught up.

This repo is a stopgap: the same PKGBUILD, same build flags, same provides/conflicts, just pointed at the current upstream tag with matching source checksums. It is not a fork with different build logic, not affiliated with the AUR maintainer, and not intended to replace the AUR package long-term. Once aur/llama.cpp-hip catches up, use that instead.

Usage

git clone <this-repo-url>
cd llama.cpp-hip
makepkg -si

Requires an AMD GPU with ROCm/HIP support (rocm-hip-sdk, hip-runtime-amd, hipblas, rocblas). See the AUR package's dependency list — this PKGBUILD's depends/makedepends are unchanged from upstream.

Status

Built and smoke-tested locally against upstream tag b10362 on an AMD Radeon RX 7900 XTX (gfx1100) before publishing. See commit history for the tag this was last bumped against.

Updating

To bump this to a newer llama.cpp release:

1. Update pkgver in PKGBUILD to the new tag (e.g. b10400).
2. Run makepkg -g to regenerate sha256sums for the source array.
3. Paste the output back into PKGBUILD.
4. makepkg -si to verify it builds before committing.

License

The PKGBUILD itself carries no separate license beyond what's inherited from the upstream AUR PKGBUILD it's based on (MIT, matching llama.cpp). See PKGBUILD for attribution.