Satellite repository of https://github.com/PetrGlad/statmoments and https://github.com/akochepasov/statmoments.
Used for GPU experimentation to reduce dependencies and avoid handling unnecessary details.
Using cupy to access CUDA functionality.
Assuming Linux environment.
Install CUDA toolkit (Libux/Debian)
apt install nvidia-cuda-toolkit
Also may help: nvidia-cuda-dev, nvidia-cuda-gdb
Get installed CUDA version
nvcc --version
Get CUDA capabilities supported by the hardware
nvidia-smi
Install cupy wheel (package name suffix depends on the installed CUDA version):
pip install cupy-cuda11x
Ensure the CUDA driver can be actually used
import cupy
cupy.cuda.get_local_runtime_version()
cupy.cuda.device.Device().compute_capability
Reinstallation, library changes, or drive upgrade may cause CUDA to load. For example, you may get an "CUDA error: unknown error" or "Error: cudaErrorUnknown: unknown error". In that case cleaning compiled kernel cache (and in some cases also rebooting the system) may help:
rm -r "$HOME/.nv"