Misc. bug: HIP CI is currently broken, likely by 23862

[ORippler]

Name and Version

#23862 seems to break CI, see

https://github.com/ggml-org/llama.cpp/actions/runs/28798959490/job/85396827596
https://github.com/ggml-org/llama.cpp/actions/runs/28775765114/job/85414604884

Operating systems

Linux, Other? (Please let us know in description)

Which llama.cpp modules do you know to be affected?

Other (Please specify in the next section)

Command line

Problem description & steps to reproduce

I don't have AMD GPUs to test, but AI tells me this is the most likely culprit, and the PR in question itself did not trigger ubuntu-22-hip-quality-check. Was surprised about CI failing on #24481 after I did some unrelated changes

Edit: see #25361 (comment)

First Bad Commit

d06ddd3

Relevant log output

Logs

[ORippler]

CC @IMbackK @am17an

[ORippler]

Given that it's a build issue, I took the time to repro locally:

  docker run --rm -it \
    -v "$PWD":/src \
    -w /src \
    rocm/dev-ubuntu-22.04:7.2.1 \
    bash

Inside the container, install and build with:

  apt-get update
  apt-get install -y build-essential git cmake rocblas-dev hipblas-dev libssl-dev python3

  cmake -B build -S . \
    -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" \
    -DGPU_TARGETS=gfx942 \
    -DGGML_HIP=ON \
    -DGGML_HIP_EXPORT_METRICS=Off \
    -DCMAKE_HIP_FLAGS="-Werror -Wno-tautological-compare" \
    -DCMAKE_BUILD_TYPE=Release

  cmake --build build --target ggml-hip -j"$(nproc)"

This fails for d06ddd3 with the same errors about INFINITY, but passes for 898b088

[asf0]

Adding -fno-finite-math-only alongside -ffast-math in ggml/src/ggml-hip/CMakeLists.txt fixes the compile. Fix in #25373 . Verified on gfx1151, ROCm 7.2.4.