rocm-wsl-ai 发布 5.0，支持 ROCm 10.1 的 WSL2 一键安装与升级

What changed in 5.0

ROCm 10.1 is not a version bump — it is a different packaging system. AMD moved ROCm to a build system called TheRock, and the installer changed shape accordingly:

The practical effect: no Windows SDK, no CMake, no compiling anything. Five stages of the old installer simply do not exist any more.

Verified on a real machine — RX 7900 XTX, WSL2 Ubuntu 22.04, Python 3.10:

ROCm 10.1.0 (core channel) · librocdxg 1.1.0 · torch 2.14.0+rocm10.1.0
torch.cuda.is_available() = True · AMD Radeon RX 7900 XTX
rocminfo: "WSL environment detected."

Already on 4.x? One command.

cd ~/rocm-wsl-ai && ./upgrade.sh

It discovers what is available, prints a plan, and only then touches anything. ./upgrade.sh --check shows the plan and changes nothing.

Read this before you upgrade. ROCm 10.x WSL support is a technical preview, and AMD requires ROCm 7.2.x to be uninstalled first — the two package trees both claim /opt/rocm. The upgrade therefore removes your working 7.2.x stack, and asks you to type REMOVE before it does. Your /opt/rocm-7.2.x directories are renamed, never deleted, so a rollback stays possible. If you would rather stay on 7.2.x, use ./upgrade.sh --target legacy — that path is still maintained.

Full detail: docs/UPGRADING.md

Why this exists

Install

Windows prerequisites — only one, now:

• Windows 11
• AMD Software: Adrenalin Edition 26.10.41.05 or newer (for WSL2)
• WSL2 with Ubuntu 22.04, 24.04 or 26.04

Then, inside WSL:

curl -fsSL https://raw.githubusercontent.com/daMustermann/rocm-wsl-ai/main/install.sh | bash

Already have WSL2? Install_WSL_Ubuntu.bat sets it up from Windows.

From a clone:

git clone https://github.com/daMustermann/rocm-wsl-ai.git
cd rocm-wsl-ai && ./install.sh

Then run ./menu.sh and pick Quick start.

What you get

A GPU tuner that measures instead of guessing

./menu.sh → Performance → Auto-tune runs four shaped workloads on your actual GPU — convolutions, attention, and a multi-step denoising loop — and keeps the configuration that genuinely wins. It refuses to declare a winner when the numbers overlap, and it discards a run whose VRAM or clocks drifted mid-measurement.

Measured on an RX 7900 XTX, ROCm 10.1 / torch 2.14:

These numbers come from the engine itself; run ./scripts/utils/perf_engine.py bench to get your own. See docs/PERFORMANCE.md.

One launcher for every tool

Every supported tool goes through the same path: environment applied before Python starts, launch flags validated against the tool's own --help, a cached ~4 ms preflight, and idle hibernation that frees VRAM when you walk away.

Any other git repository can be added from Install → Add a third-party tool, or by appending one line to ~/.config/rocm-wsl-ai/tools.local.

Diagnostics that name the cause

The single most common support question is "PyTorch cannot see my GPU". This answers it directly — including the trap that HSA_OVERRIDE_GFX_VERSION is not a fallback under WSL but the thing that makes the runtime reject your device, so the diagnostics now report it as a problem when it is set.

The GPU-visibility trap

Under WSL the GPU is reached through DXCore, not through /dev/kfd. ROCr detects this by looking for /dev/dxg, and ROCm 10.x ships the bridge library that does it. rocminfo printing WSL environment detected. is the confirmation.

rocminfo | grep -i wsl          # should say "WSL environment detected."
~/genai_env/bin/python -c "import torch; print(torch.cuda.is_available())"

If either fails, run Settings → GPU Diagnostics. It checks the driver version, the DXCore device, the bridge library, group membership, the environment variables and the venv, in that order, and prints the fix for whichever one is wrong.

Requirements

• GPU — RDNA3 / RDNA4 / RDNA4 Radeon, or Ryzen AI. gfx1100 and newer. RDNA2 (gfx1030 and below) is not supported.
• OS — WSL2 on Windows 11, with Ubuntu 22.04, 24.04 or 26.04.
• Driver — AMD Adrenalin 26.10.41.05+ (for WSL2).
• Disk — ~20 GB, plus your models.
• Python — 3.10 to 3.14. The installer picks the interpreter that matches your Ubuntu release.

Nothing is version-pinned in this repository. ROCm, PyTorch, Triton and the framework wheels are resolved from AMD's index when you run, so a new ROCm release works without a code change.

Command reference

Where your settings live

~/.config/rocm-wsl-ai/
├── user.env           your settings (ports, timeouts, GPU target)
├── perf.env           the profile the tuner selected
├── logs/              every upgrade writes a log here
└── cache/             version lookups, 24h TTL

Troubleshooting

Full guide: docs/TROUBLESHOOTING.md

Documentation

Roadmap

• Per-model VRAM presets
• MIOpen find-database baking per GPU
• Shareable performance reports
• Model browser from inside the menu

Acknowledgements

ROCm and librocdxg are AMD's; PyTorch is the PyTorch Foundation's; ComfyUI, SD.Next, Automatic1111, kohya_ss and Text Generation WebUI are theirs. The terminal UI uses gum by Charm, and is optional — every screen falls back to plain ANSI when gum is not installed.

License

MIT — see LICENSE.