How to use the ROCm build ?

[tupo2]

I have ROCm installed but when I start SD, it use the CPU only : load_backend: loaded CPU backend from /home/tupo2/sd-Linux-Ubuntu-24.04-x86_64-rocm-7.14.0/libggml-cpu-haswell.so

[tupo2]

Setting export HSA_OVERRIDE_GFX_VERSION=11.0.0 has no effect. Any ideas ?

[tk1475]

The Linux ROCm zip ships libggml-hip.so next to sd-cli, but it does not bundle the ROCm runtime it links against (libamdhip64, hipBLAS, rocBLAS and friends). The binaries are built with GGML_BACKEND_DL=ON, so backends are dlopened at startup, and in release builds a backend that fails to load is skipped silently (ggml-backend-reg.cpp, silent = true under NDEBUG). That's why you only see loaded CPU backend and no error. HSA_OVERRIDE_GFX_VERSION can't help until the HIP backend actually loads.

Check what's missing from the folder you unzipped:

cd sd-Linux-Ubuntu-24.04-x86_64-rocm-7.14.0
ldd ./libggml-hip.so | grep "not found"

Whatever shows up there needs to come from a ROCm install that matches the build (the release is built against ROCm 7.14.0, see the ubuntu-latest-rocm job in build.yml). An older system ROCm usually has different sonames, so it won't satisfy it. The CI gets ROCm from AMD's pip wheels, and you can do the same:

python3 -m venv ~/rocm714 && source ~/rocm714/bin/activate
# replace gfx1100 with your GPU target (rocminfo | grep gfx)
pip install --index-url https://repo.amd.com/rocm/whl-multi-arch/ "rocm[libraries,devel,device-gfx1100]==7.14.0"
rocm-sdk init
export LD_LIBRARY_PATH="$(rocm-sdk path --root)/lib:$LD_LIBRARY_PATH"
ldd ./libggml-hip.so | grep "not found"   # should print nothing now
./sd-cli ...

(The device-gfx* extra matters: in AMD's multi-arch wheels the GPU-specific kernels for rocBLAS etc. live in separate per-target packages, see TheRock RELEASES.md.) A system-wide ROCm 7.x install with /opt/rocm/lib on LD_LIBRARY_PATH works too, as long as ldd comes back clean.

Once it loads you should see loaded ROCm backend (or similar) in the log instead of only the CPU one. Also check that your GPU's gfx target is in the build's list (gfx900 through gfx1201, e.g. gfx1030, gfx1100, gfx1151, gfx1201; rocminfo | grep gfx shows yours) and that your user is in the render and video groups.