SD.Next ROCm build installer stuck at "Exiting..." and shows "Unknown Package" after restart

[Panki1212]

Package

SD.Next ROCm

When did the issue occur?

Installing the Package

What GPU / hardware type are you using?

AMD Radeon RX 7900XTX

What happened?

Environment

StabilityMatrix: 2.15.2 App: SD.Next ROCm build OS: Windows 11 (build 26200.6725) GPU: AMD Radeon RX 7900 XTX (gfx1100) ROCm: 6.4.4 (Windows preview) Python: 3.12.11 in venv (no global Python installed) VGA Driver: Adrenalin 25.9.1

What happens From SM, install SD.Next ROCm. Installer runs, torch-rocm wheels are downloaded/installed, ROCm agent detected, then the log ends with: INFO Exiting...

The installer window never closes; I must close StabilityMatrix manually. After restarting SM, the app card appears as “Unknown Package”. Clicking Import and selecting the correct SD.Next commit does not restore the ROCm build: SM reconfigures the app to Python 3.10 and ZLUDA (CUDA fallback), because that’s the “known working” profile. So the ROCm environment I just installed is effectively discarded/overwritten.

(The other SD.Next card is the older ZLUDA version)

Why this is a problem The ROCm install itself is actually successful (venv=3.12, torch-rocm loads, HIP 6.4, device gfx1100). The installer doesn’t terminate cleanly → SM never records the successful state → card shows Unknown Package. The Import path then forces the older Python 3.10 + ZLUDA profile (historical “working default”), instead of preserving/using the ROCm (3.12) setup.

Expected behavior

Installer should exit cleanly (no hang) and persist success state (card should not read “Unknown Package”). Import / profile restore should respect the selected ROCm build and Python 3.12 interpreter, not force ZLUDA/3.10.

Repro steps Clean install of SD.Next ROCm from SM. Observe hang at Exiting.... Restart SM → card shows “Unknown Package”. Use Import → SM switches the app to Python 3.10 + ZLUDA.

Workarounds (confirmed) Kill the stuck installer, then launch SD.Next ROCm manually: venv\Scripts\python.exe launch.py --use-rocm (Or via a wrapper .bat that pins the venv’s python.exe and --use-rocm.) Avoid SM’s Import/auto-repair path (it reverts to 3.10 + ZLUDA).

Suggestions for a fix

Ensure the installer removes --test/dry-run flags and returns control properly; update app status so the card isn’t “Unknown Package”.

For SD.Next ROCm, pin the interpreter to Python 3.12 in app metadata and don’t auto-fallback to ZLUDA when ROCm wheels succeeded.

Keep the chosen compute backend (rocm) and venv path upon Import/Repair if a valid ROCm environment is detected.

Notes

This reproduces 100% on clean setups.

ROCm pipeline runs fine when launched outside the installer (so the stack is good; it’s an installer/state-persistence issue).

Console output

No response

Version

2.15.2

What Operating System are you using?

Windows