[BUG] Warp CUDA error

[robotNo201]

Bug Description

isn't that funny?

System Information

[shi-eric]

I'm not sure what's going on and I can't reproduce an issue on my system, but have you tried installing warp-lang from PyPI? From the image you posted, it seems you installed warp_lang-1.5.0+cu11-py3-none-manylinux2014_x86_64.whl, which works on my system with driver 560.35.03. But I guess we might get more information if you run into the same issue with the warp-lang on PyPI.

FYI, Warp actually doesn't use the nvcc in your system path unless you are building Warp from source.

There was a recent commit from @nvlukasz on the main branch that fixed an issue with driver entry points, but it was meant to fix something else: 2df89ed Might be worth building Warp from the main branch on your local system and seeing what happens.

[nvlukasz]

Interesting. Looks like we fail to get the cuGetProcAddress() function from the driver. This function is needed to get all required driver entry points, including cuDriverGetVersion(). Since we fail to get that function, the detected driver version falls back to 0.0.

It's unclear why cuGetProcAddress() is not found, it should be present in the driver.

Based on nvidia-smi, the driver is 561.09, which I believe is a Windows driver. The kernel cache dir looks Linux-like (/home/.../.cache/warp/1.5.0). Is it running on WSL?

[nvlukasz]

FWIW, I tried to reproduce this with WSL but didn't see a problem.

The nvidia-smi and driver versions match yours:

There might be an issue with your driver installation or WSL setup. I suggest to update your driver to latest (try custom installation option with clean install). And like Eric suggested, maybe try the latest PyPI build of Warp:

pip install --upgrade --force-reinstall warp-lang

If that still fails then the issue might be related to WSL somehow.

[nvlukasz]

Just to get another data point, you could try running some Torch CUDA code in that environment.

test_torch.py:

import torch
a = torch.arange(10, device="cuda:0")
print(a)

$ pip install torch
$ python test_torch.py

See if that works correctly.