enable-gpu about draco HOT 1 OPEN

Hi @innerop, you're right that with respect to GPU support of this repo. The current version dose not have that functionality currently.

One key reason is that it seems to be the case that it's tricky to make the current implementation (PyTorch+MPI4PY) compatible with NCCL + PyTorch. Thus, to enable GPU, we need to let worker nodes compute gradients on GPUs, then copy the gradients from GPUs to CPUs, then communicate among CPUs. That would obviously incur non-negligible data movement overheads.

However, we're actively working on moving the current version to PyTorch's communication library completely. Please stay tuned.

Hope this can be helpful.

from draco.