`--torch-backend=auto` fails on systems with multiple GPUs and without `/proc/driver/nvidia/version`

[coezbek]

Summary

This only affects systems which don't have /sys/module/nvidia/version or /proc/driver/nvidia/version, i.e. for instance on WSL2.

When running uv pip install with the new --torch-backend=auto on such a system with multiple GPUs, the detection will fallback to using nvidia-smi. Unfortunately, nvidia-smi will return a separate line for each graphics card which isn't considered in the code:

$ nvidia-smi --query-gpu=driver_version --format=csv,noheader
572.60
572.60

This will make uv pip install with --torch-backend=auto fail with the following error:

uv pip install -U "vllm[audio]" --torch-backend=auto
error: after parsing `572.60
`, found `572.60
`, which is not part of a valid version

nvidia-smi does not respect NVIDIA_VISIBLE_DEVICES, so there is no way from the outside to use --torch-backend=auto with two graphics cards at the moment.

Workaround is to run nvidia-smi, identify CUDA version there and run with --torch-backend=cuXXX as indicated by nvidia-smi.

I think the wrongly implemented line of code is the one I marked in this pullrequestreview:

#12070 (review)

Or in the current code base:

uv/crates/uv-torch/src/accelerator.rs

Line 129 in 8d6d067

let driver_version = Version::from_str(&String::from_utf8(output.stdout)?)?;

The easiest fix would be to take just the first line of the output of nvidia-smi. More elaborate would be a way to select the device to query or respecting NVIDIA_VISIBLE_DEVICE environment variable.

Platform

Linux 6.6.87.2-microsoft-standard-WSL2 x86_64 GNU/Linux

Version

uv 0.7.21

Python version

Python 3.12.3

[charliermarsh]

Sorry, I thought I fixed this!