[Bug] error: [Errno 2] No such file or directory: '/bin/nvcc'

[gpastal24]

Prerequisite

• I have searched Issues and Discussions but cannot get the expected help.
• The bug has not been fixed in the latest version(https://github.com/open-mmlab/mmcv).

Environment

Pytorch -> pip
CUDA_HOME -> ~/miniconda3/envs/env_name/lib/python3.10/site-packages/nvidia

Reproduces the problem - code sample

CUDA_HOME=~/miniconda3/envs/env_name/lib/python3.10/site-packages/nvidia pip install mmcv-full

Reproduces the problem - command or script

CUDA_HOME=~/miniconda3/envs/env_name/lib/python3.10/site-packages/nvidia mim install mmcv-full

Reproduces the problem - error message

error: [Errno 2] No such file or directory: '~/miniconda3/envs/env_name/lib/python3.10/site-packages/nvidia/bin/nvcc'

Additional information

Hello I am trying to install mccv in a virtual environment.
I set the CUDA_HOME env variable before installing mmcv.
The CUDA path in this case in the one that is shipped with torch, not the one installed in the OS.
I cannot progress since nvcc is needed, why is that? I thought the basic cuda library shipped with torch is sufficient for pretty much anything CV related.
Thanks

[samskalicky]

Hi @gpastal24, the CUDA path to nvcc is usually installed when you install Nvidia's CUDA tools at /usr/local/cuda/bin/nvcc. Its not installed as part of PyTorch -- that just has pre-compiled libraries to run on GPUs.

[gpastal24]

@samskalicky I dont get why we need nvcc. Are we compiling cuda kernels or what?

[samskalicky]

Sorry, im new to MMCV. But it looks like on PyPI mmcv-full is a source distribution so it needs to be built (its not a prebuilt wheel) https://pypi.org/project/mmcv-full/1.7.2/#files