LoRA Training on AMD RX 9060 XT with Native Linux ROCm 7.2.3 – A Complete Guide
Patch: install ROCm 7.2.3, add user to render,video groups, compile bitsandbytes for gfx1200, use cupertinomiranda fork, download SDXL fp16 and symlink missing files.
Patch: install ROCm 7.2.3, add user to render,video groups, compile bitsandbytes for gfx1200, use cupertinomiranda fork, download SDXL fp16 and symlink missing files.
Summary
On Kubuntu 24.04.4 with kernel 6.17.0-23-generic, training a LoRA on an AMD RX 9060 XT (RDNA4, gfx1200, 16 GB GDDR6, 150 W TDP) works only on native Linux; WSL2 fails due to a closed‑source DXG bridge bug (librocdxg issue #22) that forces training to run on CPU or hang. The native ROCm 7.2.3 stack bypasses the DXG bridge, delivering 3–4 seconds per SDXL step versus 50 minutes on WSL2, and requires adding the user to the render and video groups for /dev/kfd and /dev/dri access. The recommended training fork is cupertinomiranda/ai‑toolkit‑amd‑rocm‑support, which explicitly supports gfx1200/gfx1201 and ROCm 7.x, and includes bitsandbytes build instructions. Bitsandbytes must be compiled from source with -DBNB_ROCM_ARCH=gfx1200 and the full path to ROCm clang++; nightly 2.13.0.dev crashes, so use the stable 2.11.0+rocm7.2 wheel.
For SDXL, download only the fp16 safetensors and create symlinks for the missing files; otherwise the pipeline fails to load. GPU monitoring works with rocm‑smi on native Linux, whereas WSL2 rocm‑smi is broken. The guide also warns against using float32 on WSL2, which can OOM the GPU and trigger a BSOD on Windows. Finally, avoid installing on NTFS mounts because Linux file permissions are not preserved, which would break chmod operations during setup.
Key changes
- WSL2 GPU passthrough via DXG bridge fails for RDNA4 due to librocdxg issue #22, causing training to run on CPU or hang.
- Native Linux ROCm 7.2.3 bypasses the DXG bridge, enabling 3–4 seconds per SDXL step versus 50 minutes on WSL2.
- ROCm installation requires manual addition of the user to render and video groups; otherwise permission errors occur.
- cupertinomiranda/ai‑toolkit‑amd‑rocm‑support is the only fork that supports gfx1200/gfx1201 and ROCm 7.x, with bitsandbytes build instructions.
- Bitsandbytes must be compiled from source with -DBNB_ROCM_ARCH=gfx1200 and the full path to ROCm clang++; nightly 2.13.0.dev crashes, use stable 2.11.0+rocm7.2.
- SDXL fp16 safetensors must be downloaded and symlinked for missing files; otherwise the pipeline fails to load.
- GPU monitoring works with rocm‑smi on native Linux; WSL2 rocm‑smi is broken.