Make Ollama use an AMD GPU that ROCm silently ignores
Read the Arch wiki Ollama page (raw): it lists ollama-vulkan, ollama-cuda and ollama-rocm as GPU add-ons, has the tip that Vulkan is usually faster than ROCm on RDNA 3+ and avoids unsupported-GPU problems, says integrated GPUs are skipped without OLLAMA_IGPU_ENABLE, and gives the rocminfo | grep amdhsa, find Kernels.so-* and X.Y.Z matching rules the record repeats. Arch package JSON confirms all three backends exist in extra and each depends on ollama. ollama-rocm pulls rocminfo transitively (hipblas -> hip-runtime-amd -> rocminfo), so /opt/rocm/bin/rocminfo exists after the install step in the fix. Confirmed on this 4.0.4-1 machine that /usr/share/omarchy/default/omarchy/omarchy-menu.jsonc entry install.ai.ollama picks ollama-cuda when nvidia-smi is present, ollama-rocm only when rocminfo is present, else ollama, matching the cause. omarchy#10380 is open and says the same. pacman -S without -u passes the ALPM guard. Nothing was installed or exercised.
Symptom
Ollama answers slowly and amdgpu_top or ollama ps shows the model running on CPU (100% CPU) even though the machine has a Radeon card. On Omarchy, Install > AI > Ollama may also have installed plain ollama or ollama-cuda instead of a ROCm or Vulkan backend.
Cause
ROCm skips GPUs that are not on its supported list without an error, so ollama-rocm falls back to the CPU on many consumer and integrated Radeons. Separately, Omarchy's Ollama menu entry only picks ollama-rocm when the rocminfo command exists, and a stock Omarchy AMD install never has it, so AMD owners get CPU-only ollama, or ollama-cuda if nvidia-utils was pulled in (omarchy#10380, unchanged in 4.0.4). The Arch backends are add-on packages that depend on ollama: ollama-rocm, ollama-vulkan and ollama-cuda.
Fix
**Easiest on RDNA 3 and newer: use the Vulkan backend.** The Arch wiki notes it is usually faster than ROCm there and avoids the unsupported-GPU problem. No environment variable is needed:
sudo pacman -S --needed ollama-vulkan
sudo systemctl restart ollama
Integrated GPUs are skipped unless you set OLLAMA_IGPU_ENABLE=1 in the service.
**Staying on ROCm: override the GFX version.**
sudo pacman -S --needed ollama-rocm
/opt/rocm/bin/rocminfo | grep amdhsa # e.g. gfx1103
find /opt/rocm/lib/rocblas/library -name 'Kernels.so-*'
Pick a version from the rocblas list whose major part equals yours and whose other parts are not higher (a 4-digit gfx1103 reads as 11.0.3):
sudo systemctl edit ollama
[Service]
Environment="HSA_OVERRIDE_GFX_VERSION=11.0.0"
sudo systemctl restart ollama
If the menu installed ollama-cuda on an AMD machine, remove it with sudo pacman -Rns ollama-cuda.
Verify
Run a model, then ollama ps shows 100% GPU and amdgpu_top shows GPU load and VRAM in use. journalctl -u ollama -b | grep -i -E 'rocm|vulkan|gfx' names the backend and device.