setup (--backend hip): a runtime-only system ROCm 7 in /opt/rocm is picked for the engine build, and cmake fails on the missing hip-lang package instead of falling back to the TheRock wheels #446

What happens

./setup.sh --backend hip on a PC whose /opt/rocm has the ROCm 7.14 runtime packages but not the development ones stops in step 4:

  The ROCm root directory:

   /opt/rocm

  does not contain the HIP runtime CMake package, expected at one of:

   /opt/rocm/lib/cmake/hip-lang/hip-lang-config.cmake
   /opt/rocm/lib64/cmake/hip-lang/hip-lang-config.cmake
   /opt/rocm/lib/x86_64-unknown-linux-gnu/cmake/hip-lang/hip-lang-config.cmake

Call Stack (most recent call first):
  cmake/hip_backend.cmake:38 (enable_language)
  CMakeLists.txt:50 (include)

-- Configuring incomplete, errors occurred!

  [X]  command failed (exit 1): cmake

Why

rocm_root() in setup.py takes the system ROCm when bin/hipcc and lib/libhipblas.so* exist and the version is 7.0 or newer. A runtime-only install has both, so the check passes:

• /opt/rocm/bin/hipcc exists (on this install it comes from amdrocm-llvm7.14)
• /opt/rocm/lib/libhipblas.so* and libamdhip64.so* exist
• include/rocm-core/rocm_version.h says 7.14

The parts the build needs are missing, though: lib/cmake/hip-lang/, lib/cmake/hip/ and include/hip/hip_runtime.h. On AMD's 26.13 repo the HIP ones ship in amdrocm-runtime-dev7.14 (checked with dpkg -c), which, like amdrocm-blas-dev7.14, comes with amdrocm-core-dev7.14-gfx1201; none of the -dev packages are installed here. So setup never tries the TheRock wheels path, which works on this PC without sudo.

Workaround

Point ROCM_PATH somewhere without ROCm, so setup takes the TheRock wheels path:

ROCM_PATH=$HOME/.no-system-rocm ./setup.sh --backend hip

With that, the setup ran to the end (ROCm 7.10.0a20251120 wheels from gfx120X-all, engine compiled, IQ3_XXS ready), and the server answers normally, tool calls included.

Suggested fix

Have the system-ROCm check also look for the files the build needs (for example lib/cmake/hip-lang/hip-lang-config.cmake and include/hip/hip_runtime.h). When they are missing, warn and use the wheels, the same way an older ROCm is handled now:

[!] the ROCm in /opt/rocm has no HIP development files (lib/cmake/hip-lang): using AMD's wheels in .venv instead
    (or install them, e.g. sudo apt install amdrocm-core-dev7.14-gfx1201)

(I have not tried the build with the -dev packages installed; the wheels path is the one verified here.)

Note: open PR #426 (Windows AMD) also changes rocm_root(), so this fix may be easier on top of it.

Environment

• Strata main @ aeb35be (the code in question is the same on 1678de3)
• Ubuntu 24.04.5, kernel 7.0.0-31-generic
• AMD Radeon RX 9070 XT 16 GB (gfx1201), Ryzen 9 3950X, 63 GB RAM
• amdgpu-install 31.40 + AMD's repo rocmradeon/apt/26.13: amdrocm-*7.14 7.14.0~pre3 runtime packages (amdrocm-core7.14-gfx1201, amdrocm-blas7.14-gfx1201, amdrocm-llvm7.14, ...), no -dev packages
• Command: ./setup.sh --backend hip --yes --no-start