ah-gus
commented
2 years ago well, it seems as if this doesn't actually quite work , and we're either allocating host memory (not gpu) or we're making things crash. so we withraw the beta status back to pre-alpha, still got some work to do , but we're getting there
port using nv_peer_mem on NVIDIA jetson Xavier based on 1.1.0 version so it supports OFED 4.9-4.0.8.0 LTS and CUDA 10.2 changes would probably work for newer releases as well code quality isn't that great, paths were changed manually need to add proper #ifdef`s and change documentation. it's still a beta , but it works (with a connectx3 pro) i'd be happy if this was available to the community, for now as a separate branch