In Python <3.8 it is possible to use the cupy-cudaXXX inside a conda environment in combination with cudatoolkit from anaconda repositories. The cuda binaries are installed into the conda environment and no CUDA_PATH has to be set.
Under Python 3.8 the following lines make the import fail, although CUDA is present:
|
if sys.platform.startswith('win32') and (3, 8) <= sys.version_info: # NOQA |
|
_environment._setup_win32_dll_directory() # NOQA |
Uncommenting these lines make the import work again and cupy works also as expected. So my question: Is it really necessary to enforce users to set the CUDA_PATH variable? And wouldn't it be better to explicitely check if the needed dll is present rather than the environment variable?
In Python <3.8 it is possible to use the cupy-cudaXXX inside a conda environment in combination with cudatoolkit from anaconda repositories. The cuda binaries are installed into the conda environment and no CUDA_PATH has to be set.
Under Python 3.8 the following lines make the import fail, although CUDA is present:
cupy/cupy/__init__.py
Lines 11 to 12 in 1654d38
Uncommenting these lines make the import work again and cupy works also as expected. So my question: Is it really necessary to enforce users to set the CUDA_PATH variable? And wouldn't it be better to explicitely check if the needed dll is present rather than the environment variable?