Check cuDF lazily

[grafail]

In some cases the "import cudf" can raise a RuntimeError or other errors while loading. Doing a lazy check is safer for these cases. Similiar to #7752.

Example of RuntimeError that could be generated by import:

Traceback (most recent call last):
  File "cuda/_cuda/ccuda.pyx", line 3671, in cuda._cuda.ccuda._cuInit
  File "cuda/_cuda/ccuda.pyx", line 435, in cuda._cuda.ccuda.cuPythonInit
RuntimeError: Failed to dlopen libcuda.so

[bdice]

Hi @grafail, would you be able to share a full traceback, cudf version, and also the version of cuda-python? I’d like to dig into this further. This is being raised by cuda-python and we try to guard against this RuntimeError in cudf but might be missing a case. Thanks!

[trivialfis]

Thank you for the PR! It would be great if you can help share a reproducible example with cuDF developers.

[grafail]

Hi @bdice. This issue can by recreated by just importing cudf on rapidsai/rapidsai-core:22.06-cuda11.2-base-ubuntu18.04-py3.8, while not using the nvidia runtime in docker.

docker run --rm -it rapidsai/rapidsai-core:22.06-cuda11.2-base-ubuntu18.04-py3.8 python3 -c "import cudf"

cuDF version: 22.6.0
cuda-python version: 11.7.0

This is the entire traceback:

Traceback (most recent call last):
  File "cuda/_cuda/ccuda.pyx", line 3671, in cuda._cuda.ccuda._cuInit
  File "cuda/_cuda/ccuda.pyx", line 435, in cuda._cuda.ccuda.cuPythonInit
RuntimeError: Failed to dlopen libcuda.so
Exception ignored in: 'cuda._lib.ccudart.utils.cudaPythonGlobal.lazyInitGlobal'
Traceback (most recent call last):
  File "cuda/_cuda/ccuda.pyx", line 3671, in cuda._cuda.ccuda._cuInit
  File "cuda/_cuda/ccuda.pyx", line 435, in cuda._cuda.ccuda.cuPythonInit
RuntimeError: Failed to dlopen libcuda.so
Segmentation fault (core dumped)

It seems to not create an issue on rapidsai/rapidsai-core:21.10-cuda11.2-base-ubuntu18.04-py3.8 though.

[bdice]

@grafail Thanks for the additional information!