- Use --preset 'ROCm 6' for HIP build step (enables enable_language(HIP))
- Remove /opt/rocm from PATH for CPU build to prevent check_language(HIP)
- Add CMAKE_PREFIX_PATH=/opt/rocm so find_package(hip) finds hip-config.cmake
- cmake --install --component HIP now works correctly with OLLAMA_RUNNER_DIR=rocm
gfx940/gfx1010/gfx1030/gfx1100 cause C++ narrowing errors in ollama's
mma.cuh with hipcc. Since we only have MI50 (gfx906) cards, compile
for gfx906 only. Reduces build time and avoids upstream code bugs.
CPU backends compiled with GCC (fixes AVX intrinsic errors from hipcc).
HIP backend compiled with hipcc (legacy mode skips enable_language(HIP)).
Go binary built with GCC for CGo linking.
This avoids both CMAKE_HIP_COMPILER rejection and CXX=hipcc CPU failures.
The old Dockerfile used the deprecated llama.cpp/ subdirectory approach
which doesn't exist in ollama v0.23.2. Now using the official CMake
presets (ROCm 6 preset) with AMDGPU_TARGETS including gfx906:xnack-.
- Add ollama/Dockerfile that builds ollama from source with AMDGPU_TARGETS=gfx906
- Uses ROCm 6.1 (rocm/dev-ubuntu-22.04:6.1.2-complete) for MI50 support
- Builds llama.cpp runner with HIPBLAS for gfx906 architecture
- Updates compose.yml to build from this Dockerfile instead of pulling ollama/ollama:latest
