[BUG][CuTeDSL] segfault when calling cutlass.cuda.initialize_cuda_context() on latest cute dsl

[sfc-gh-lpaille]

Which component has the problem?

CuTe DSL

Bug Report

Describe the bug
I get a segmentation fault when initializing cuda context.

Steps/Code to reproduce bug
use nvidia-cutlass-dsl==4.2.1 AND cuda-python>=13.0.0 (which is the default on latest cute dsl).

import cutlass
cutlass.cuda.initialize_cuda_context()

Expected behavior
I expect to not get a segfault like it does on nvidia-cutlass-dsl==4.2.0 with cuda-python==12.9.2.

Environment details (please complete the following information):
nvidia-cutlass-dsl==4.2.1 + cuda-python>=13.0.1

[sfc-gh-lpaille]

workaround for now is just to downgrade to cuda-python==12.9.2 but it's not a long term solution
pip install cuda-python==12.9.2

[fengxie]

@brandon-yujie-sun

[brandon-yujie-sun]

@sfc-gh-lpaille would you please provide more info on your setup? Which driver was used in the setup?

[Chtholly-Boss]

@sfc-gh-lpaille would you please provide more info on your setup? Which driver was used in the setup?

I encountered the same problem, here is my driver version:

NVIDIA-SMI 570.86.10              Driver Version: 570.86.10      CUDA Version: 12.8 

Downgrade to 4.2.0 fixed the issue

[brandon-yujie-sun]

@sfc-gh-lpaille would you please provide more info on your setup? Which driver was used in the setup?

I encountered the same problem, here is my driver version:

NVIDIA-SMI 570.86.10              Driver Version: 570.86.10      CUDA Version: 12.8 

Downgrade to 4.2.0 fixed the issue

Thanks @Chtholly-Boss Did you also install cuda-python>=13.0.1? If so, that explained why - cuda-python 13.0.1 would need a more recent driver (580+) to work correctly instead of driver with version 570, see https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html#id5

[Chtholly-Boss]

You're right. Thanks