Disable Peer2Peer Access in PyTorch

Hi guys,

I wonder whether there is a way to disable multi-GPU Peer2Peer Access in Pytorch.


In case you are using NCCL, you could use NCCL_P2P_DISABLE=1.


Any available solutions for cudaMemcpy (device to device)?

What is your use case? Are you using UVA with cudaMemcpyDeviceToDevice?
If you don’t want to use p2p access, you could copy the data to the host and from the host to another device.

Thanks! That’s just a benchmark testing for research purpose.