ROCm Stable Diffusion

AMD ROCm-10 (AMD's ROCm / therock) SD.Next (Stable Diffusion web UI) on top of the rocm-torch-install foundation — one script to go from a proven AMD GPU to a running image engine.

What this does

setup-ai-distro.sh installs SD.Next — the actively-maintained A1111-style web UI (same look, same port 7860, same API) — and points it at your model library.

It does not reinstall ROCm or PyTorch. That's the job of the sibling project, kept deliberately separate so the GPU foundation is reusable for any app (Stable Diffusion here, LLM serving there, etc.):

Project	Responsibility
rocm-torch-install	ROCm + PyTorch venv + smoke test (the reusable foundation)
rocm-stable-diffusion (this repo)	SD.Next on top of that foundation

Both are idempotent — safe to re-run; finished steps are detected and skipped.

Quick start (two steps)

# 1) Install & verify the foundation (once):
bash <(curl -fsSL https://raw.githubusercontent.com/IP050/rocm-torch-install/main/setup-rocm.sh) --gpu rx9070xt

# 2) Build the diffusion UI on top:
git clone https://github.com/IP050/rocm-stable-diffusion.git
cd rocm-stable-diffusion
bash setup-ai-distro.sh --models ~/models   # point at your model library

Step 2 refuses to run (with a clear message) until step 1 has put the ROCm+torch venv at ~/ai-rocm-venv. That's on purpose — it avoids guessing at a GPU arch you may not have.

Pick your card with a preset name — rx9070xt, rx7900xtx, rx7800xt, rx7600, mi300, and more (see the base project's matrix).

Run both as your normal user (not root, not sudo).

Options

Flag	What it does
--models <dir>	dir holding your models/ subfolders (symlinked in)
--venv <dir>	the ROCm+torch venv from the base (default ~/ai-rocm-venv)
--engine <sdnext>	web UI; today only sdnext is wired
--port <n>	port to serve on (default: 7860)
--no-boot	install everything but don't launch the UI
--dry-run	print the plan, install nothing
--help	full usage

What gets installed, where

Path	What
~/ai-rocm-venv/ (from base)	Python venv with the ROCm PyTorch wheels
~/sdnext/	the SD.Next web UI
~/sdnext/start.sh	launcher (re-run any time to relaunch on :7860)
~/.cache/rocm-sd/models/ (default)	where your models/ subfolders are symlinked from
~/sdnext-setup.log	full rolling log

The model models/ subfolders are symlinked into SD.Next (nothing is copied), so you can keep a large model library on a different drive or a shared location and just point --models at it.

WSL2 vs native Linux

Handled entirely by the base project, whose script auto-detects WSL2 (/dev/dxg + the librocdxg bridge) vs native Linux (/dev/kfd). This app layer sits on top of whichever you chose.

Why split the install across two projects?

ROCm's install docs present a big matrix (device family × GPU × OS × distro × use case). The foundation — ROCm + the matching PyTorch — is identical whether you then run Stable Diffusion, an LLM, or fine-tuning. Splitting it out means:

• the foundation is a small, focused, testable project people can trust for any AMD-GPU ML work, and
• this app stays small and only concerns itself with SD.Next.

Native Windows?

ROCm is also pip-installable directly on Windows. See docs/windows-native.md.

License

MIT. The script installs SD.Next under its own license.

---

Not affiliated with AMD, PyTorch, or the SD.Next project. Trademarks belong to their owners.