[nvMolKit] Default pip/uv install nvmolkit pulls a CUDA-13 torch the host driver cannot run

[xinyu-dev]

Summary

The default pip/uv install nvmolkit pulls torch 2.12.0 + nvidia-*-cu13 (CUDA-13 wheels). On a CUDA-12.x driver this makes torch unusable and every nvMolKit call fails. (Note: Python 3.10 has no wheels — only cp311–cp314.)

Severity: High · Status: workaround verified; docs gap open

Steps to reproduce

uv venv --python 3.12 .venv          # py3.10 has no wheels (cp311–cp314 only)
uv pip install nvmolkit              # pulls torch 2.12.0 + nvidia-*-cu13
python -c "import torch; print(torch.cuda.is_available())"

Actual outcome

torch.cuda.is_available() → False; torch._C._cuda_init() raises "NVIDIA driver too old (found 12070)"; all nvMolKit calls fail.

Expected outcome

A default install should yield a torch whose CUDA backend matches the host driver (CUDA 12.x here), or the docs/skill should warn and point to the --torch-backend fix.

Fix

Reinstall torch with a CUDA-12 backend in the same env (--reinstall-package is required; uv otherwise sees torch as satisfied and no-ops):

uv pip install --torch-backend=cu126 --reinstall-package torch torch

→ torch 2.12.0+cu126, cuda.is_available()=True. nvMolKit itself is already CUDA-12 (nvidia-cuda-runtime-cu12); only torch's default backend was wrong. The runtime-requirements docs say only "CUDA 12.6+ driver" / "working torch with CUDA" and don't warn that the default pip/uv install grabs CUDA-13 torch, nor mention --torch-backend (the README warns about the conda path only).

Environment

A100 80GB PCIe (sm_80) · driver 565.57.01 · CUDA 12.6 · nvMolKit 0.5.0 @ 6f967ed · Python 3.12 · Linux x86_64.