The llama.cpp project released version b11847 with a fix for HIP builds targeting AMD RDNA3 and RDNA4 GPUs. The change uses 16-bit `short` vectors for BF16 operands passed to AMD’s matrix instructions; the values remain bit-identical when reinterpreted.
The release listing includes Ubuntu x64 builds for ROCm 10.0. It does not say whether those prebuilt packages include the specific fix, so users may need to build from the updated source.
Why it matters
Developers building llama.cpp with HIP for the listed AMD GPU generations should no longer hit this specific compilation error; the release listing does not confirm which prebuilt packages contain the fix.
