The notebook currently assumes/requires CUDA 12.x to be installed and the version at /usr/local/cuda (see #3), which is not always the case.
FLAME GPU 2 currerntly supports CUDA 12 and CUDA 13, so optionally using the version found via env variables / path would be nice.
This would require selecting an alternate wheelhouse, based on the version found.
Potentially, it would also be possible to depend on pip-provided packages which provide parts of the CUDA toolkit, (i.e. nvrtc). However that might not currently be feasible, depending on if the cuda runtime headers are available (rather than just the .so files)
The notebook currently assumes/requires CUDA 12.x to be installed and the version at
/usr/local/cuda(see #3), which is not always the case.FLAME GPU 2 currerntly supports CUDA 12 and CUDA 13, so optionally using the version found via env variables / path would be nice.
This would require selecting an alternate wheelhouse, based on the version found.
Potentially, it would also be possible to depend on pip-provided packages which provide parts of the CUDA toolkit, (i.e. nvrtc). However that might not currently be feasible, depending on if the cuda runtime headers are available (rather than just the .so files)