RX 9060 XT trains SDXL and Flux LoRAs on native Linux; WSL2 blocked by AMD driver bug

AMD's RX 9060 XT — a 16GB RDNA4 card with the gfx1200 architecture ID — successfully trains portrait LoRAs for both SDXL and Flux.1 Dev on native Linux. The setup runs Kubuntu 24.04.4 with ROCm 7.2.3 and a Ryzen 5 5600G, and the author documents the full path from hardware setup through completed training runs, including what failed and why.

The native Linux route matters because WSL2 remains broken for RDNA4 training as of May 2026. Windows 10 users attempting LoRA training on the RX 9060 XT hit a confirmed bug in AMD's closed-source DXG bridge library (libthunk_proxy.a). Training pipelines load successfully and fill 8–10GB of VRAM, but GPU compute stays at zero percent while the CPU climbs to 30 percent and RAM to 28GB. The first SDXL training step either runs entirely on CPU — taking roughly 50 minutes for a single step at batch size 1 — or the process hangs indefinitely. The error [GetSegmentId] Failed to get segment id for type 1 appears in logs, tied to librocdxg Issue #22, opened in April 2026 and still unfixed. Only AMD can resolve it by shipping an updated Adrenalin driver.

On native Linux, the same hardware completes SDXL LoRA training without the DXG bridge. The author trained a 1500-step Flux.1 Dev LoRA and used it in ComfyUI to generate lifestyle portrait photos. The post includes architecture notes — native Linux ROCm and amd-smi report the card as gfx1200, not the gfx1201 identifier that appeared under WSL2's incorrect reporting — and covers cmake flags, arch-specific builds, and the full toolchain from dataset prep through inference. The writeup confirms the 16GB RX 9060 XT's viability in open-weight workflows: SDXL and Flux.1 Dev LoRAs are working on a sub-$500 card with native Linux tooling. Whether AMD will backport the WSL2 fix to older driver branches or leave Windows users waiting for a future Adrenalin release remains to be seen.