Closed Ed4everU closed 11 months ago
mpi4py already supports this since v3.1.0, I am not sure how it is relevant for CUDA Python? CUDA Python is unaware of MPI and it shouldn't be.
Ah, I believe I understand now. So, if I allocate the objects on the CUDA device and transmit them using MPI, the mpi4py library will automatically handle the device-to-device transmission, isn't it?
It'd be handled by the MPI library, not mpi4py which just forwards the pointers to the underlying MPI.
Thank you so much for your informative answer. I see it clearly now. 😁
Hi Developers,
As a newcomer to cuda-python, I'm eager to learn how to implement CUDA-aware MPI transmission. I'd appreciate any guidance or resources on this topic. Thank you!