### System details

omarchy 4.0.2-1, Hyprland 0.56.2-1 - GPU: AMD Radeon RX 7900 XT — Navi 31 `[1002:744c]`, `amdgpu`, gfx1100 - mesa 1:26.2.1-1, vulkan-radeon 1:26.2.1-1 - nvidia-utils 610.57.04-1 present (pulled in as a dependency of   hyprland / aquamarine / steam / lib32-nvidia-utils / umu-launcher); no NVIDIA   hardware; no ROCm packages installed

### What's wrong?

## What happened

On an AMD-only machine (Radeon RX 7900 XT, `amdgpu`), **Menu → Install → AI → Ollama**
installed `ollama-cuda` plus the full `cuda` dependency tree (~3 GB installed)
instead of `ollama-rocm`.

## Root cause

`default/omarchy/omarchy-menu.jsonc`, entry `install.ai.ollama`:

```sh
if omarchy-cmd-present nvidia-smi; then ollama_pkg=ollama-cuda;
elif omarchy-cmd-present rocminfo; then ollama_pkg=ollama-rocm;
else ollama_pkg=ollama; fi
```

Two problems:

1. **`nvidia-smi` present ≠ NVIDIA GPU.** `nvidia-utils` ships `/usr/bin/nvidia-smi`
   and is routinely pulled onto AMD/Intel machines as an `opengl-driver` /
   `vulkan-driver` provider. On this box it is a dependency of `hyprland`,
   `aquamarine`, `steam`, `lib32-nvidia-utils`, and `umu-launcher`.
   `omarchy-cmd-present nvidia-smi` (i.e. `command -v nvidia-smi`) therefore
   succeeds on hardware with no NVIDIA GPU, and the menu installs `ollama-cuda`.
   Note `nvidia-smi -L` itself fails here with *"NVIDIA-SMI has failed because it
   couldn't communicate with the NVIDIA driver"* — so even checking the command's
   exit status, not just its presence, would avoid this.

2. **The `rocminfo` branch is effectively dead.** `rocminfo` comes from the
   `rocminfo` / `rocm-core` package, which a stock Omarchy install never places on
   disk — AMD systems get `mesa` + `vulkan-radeon` (see
   `install/hardware/vulkan.sh`), not ROCm. `ollama-rocm` does not depend on
   `rocminfo` either. So via this menu an AMD GPU owner gets `ollama-cuda` (if any
   `nvidia-utils` is present) or the CPU-only `ollama` — but never `ollama-rocm`.

## Expected

AMD GPU → `ollama-rocm`; NVIDIA GPU → `ollama-cuda`; otherwise → `ollama`.

## Suggested fix

Detect by hardware vendor, the way `install/hardware/nvidia.sh` and
`install/hardware/vulkan.sh` already do, instead of probing for userspace
binaries:

```sh
if lspci | grep -qiE '(VGA|3D|Display).*NVIDIA'; then
  ollama_pkg=ollama-cuda
elif lspci | grep -qiE '(VGA|3D|Display).*(AMD|ATI)'; then
  ollama_pkg=ollama-rocm
else
  ollama_pkg=ollama
fi
```

Gating on the loaded kernel module (`nvidia` vs `amdgpu`) would work equally well.
A machine with both GPUs should probably prefer whichever the compositor is
actually running on.

## Repro

1. Any AMD-GPU machine that also has `nvidia-utils` installed — common, since
   Steam or anything that resolved the virtual `opengl-driver` / `vulkan-driver`
   to `nvidia-utils` drags it in.
2. Menu → Install → AI → Ollama.
3. `ollama-cuda` + `cuda` are installed instead of `ollama-rocm`.