One of the longest running tasks is generating the dump file before training, which currently uses CPU only. I tried changing it to use CuPy, but it turns out to be non-trivial because of the optimization tricks you use, so I am creating an issue to propose using it to speed it up for applicable GPUs.
I understand that this is designed for CPUs, would be nice to have a GPU option!
One of the longest running tasks is generating the dump file before training, which currently uses CPU only. I tried changing it to use CuPy, but it turns out to be non-trivial because of the optimization tricks you use, so I am creating an issue to propose using it to speed it up for applicable GPUs.
I understand that this is designed for CPUs, would be nice to have a GPU option!
cupy.dev/
rapids.ai