Everything you need to know about running AI workloads on AMD GPUs and APUs with ROCm.

Check Your AMD GPU

Before installing anything, confirm your GPU is supported. Open a terminal and run lspci | grep VGA. You are looking for an AMD Radeon RX 6000, 7000, or 7000-series APU. Thumper-Run's GPU detection will also surface this information automatically in the Settings > Hardware panel.

Not every AMD GPU works with ROCm. The official support matrix covers RDNA 2 and RDNA 3 discrete cards plus select APUs. If your card is not listed, you may still succeed with environment overrides, but expect rough edges.

ROCm Prerequisites

ROCm is AMD's open-source GPU compute stack—the equivalent of NVIDIA's CUDA. You need:

Linux kernel 6.x (Ubuntu 22.04+, Fedora 38+, Arch)
ROCm 6.x packages (amdgpu-dkms, rocm-hip-sdk)
Your user in the video and render groups

Thumper checks these prerequisites at launch and will warn you if something is missing. See the troubleshooting guide for step-by-step fixes.

Driver Installation

On Ubuntu:

`bash

sudo apt install amdgpu-dkms rocm-hip-sdk

sudo usermod -aG video,render $USER

Reboot after installing. Verify with rocminfo—you should see your GPU listed as an agent.

On Arch Linux, install rocm-hip-sdk from the AUR or the official extra repo. The DKMS module builds automatically against your running kernel.

MIOpen Find-DB Warmup

The first time you run an AI model on AMD, MIOpen benchmarks convolution kernels for your specific GPU. This warmup can take 5–15 minutes and writes results to ~/.config/miopen/. Subsequent launches skip this step.

Set MIOPEN_FIND_MODE=2 (FAST) in production to use cached results without re-benchmarking. Never leave MIOPEN_FIND_ENFORCE=4 enabled permanently—it forces a full benchmark every launch.

APU-Specific Tips

AMD APUs (like the Ryzen 7 8845HS) share system RAM between CPU and GPU. Key rules:

Always use BF16, never FP16. FP16's max value of ~65504 causes NaN in DiT attention layers. BF16 has the same dynamic range as FP32 and is safe.
Keep smart memory enabled unless a model requires a GPU-only text encoder. Disabling it costs 50–70 seconds in reload overhead.
The comgr cache at ~/.cache/comgr/ can reach 650 MB. Do not delete it between runs—it caches compiled GPU kernels.
Never use `--lowvram` on an APU. Per-layer offload shuffles data between the same physical RAM and tanks utilization to ~5%.

TheRock Nightly Wheels

For bleeding-edge GPU support (like gfx1150), use AMD's TheRock nightly PyTorch wheels:

`bash

pip install torch --index-url https://rocm.nightlies.amd.com/v2-staging/gfx1150/

These wheels include native GPU kernels so you avoid runtime recompilation. Add --disable-offload-from-vram to prevent kernel recompilation at inference time.

Common Errors

"hipErrorNoBinaryForGpu" – Your GPU architecture is not in the wheel. Set HSA_OVERRIDE_GFX_VERSION to the nearest supported arch.
"MIOpen: No suitable algorithm found" – The Find-DB warmup was interrupted. Delete ~/.config/miopen/ and re-run.
OOM during VAE decode – Move the text encoder to CPU. See the ComfyUI guide for per-node device assignment.
Process stops immediately (SIGTSTP) – If launching from a terminal, ensure the app uses setsid() not process_group(0). Thumper handles this automatically.