ROCm: HSA_STATUS_ERROR_INVALID_DEVICE — GPU not detected

ROCm can't see your AMD GPU. Causes:

• GPU not in the ROCm-supported list (Polaris/Vega: not supported in 2026 ROCm)
• User not in render and video groups
• amdgpu kernel module not loaded
• Conflicting open-source vs proprietary AMD driver
• Wrong ROCm version for your GPU (RX 7900 XTX needs ROCm 6.x)

Common mistake: trying ROCm on Windows. Linux has the supported path; Windows ROCm is improving but still trails.

1. Verify GPU is in the support matrix. rx-7900-xtx, mi300, mi250, w7900: yes. Polaris (RX 580 / Vega 64): no.

2. Verify kernel module:

lsmod | grep amdgpu
# Should show amdgpu loaded

3. Add yourself to render + video groups:

sudo usermod -a -G render,video $USER
# Log out and back in (or reboot)

4. Verify ROCm sees the GPU:

rocm-smi
rocminfo | grep "Marketing Name"

5. If still failing, reinstall ROCm cleanly:

sudo amdgpu-install --uninstall
sudo apt autoremove
# Reboot
sudo amdgpu-install -y --usecase=rocm

6. Test inference:

HSA_OVERRIDE_GFX_VERSION=11.0.0 python -c "import torch; print(torch.cuda.is_available())"
# True = ROCm is providing the CUDA-compat path

For full Linux + AMD setup, see Linux local AI guide.