Automatic venvs make it not work

[lainlives]

Describe the bug

There must be a way for the user to control the software being installed it keeps installing incompatible torch versions.

Steps to reproduce

.....

Modly version

mainline

Operating system

Fedora44

Screenshots & logs

[ext-setup] Downloading https://download-r2.pytorch.org/whl/cu118/torch-2.5.1%2Bcu118-cp312-cp312-linux_x86_64.whl (838.3 MB)

Does not work for my nvidias, and definitely doesnt work for my AMDs. the python code theres no reasson it wont owrk on AMDs other than the fact this program sloppily creates its own venvs

Confirmation

• I have searched existing issues and this has not been reported yet.
• I can reproduce this on the latest version of Modly.

[Lorchie]

Thanks for the report! To properly diagnose this we need a few more details:

System info

• Full nvidia-smi output (if you have an NVIDIA GPU)
• GPU model(s) — NVIDIA and/or AMD
• For AMD: ROCm version installed (rocm-smi --version or apt show rocm-libs)
• Linux kernel version (uname -r)
• Python version inside the venv (python3 --version)

Extension info

• Which extension triggered the torch-2.5.1+cu118 download? (name or GitHub URL)
• Is it a built-in extension or one you installed manually?

Error details

• Does the install itself fail, or does it install successfully but fail at runtime?
• If it fails at runtime: full error traceback from the Modly logs (Settings → Logs)
• The exact torch version that ended up installed (pip show torch inside the extension venv, located at ~/.config/Modly/<extension>/venv)

AMD specifics

• Does /dev/kfd exist on your system?
• Output of rocm-smi if available

This will let us determine whether the issue is wrong CUDA version selection, missing ROCm support, or something else entirely.