this line dgl/nccl_api.cu at master · dmlc/dgl · GitHub
why not use other CUDA stream like sparsePull method directly ?
this line dgl/nccl_api.cu at master · dmlc/dgl · GitHub
why not use other CUDA stream like sparsePull method directly ?
Both of them are nullptr I think, so there’s no difference here. as the comments said, it’s TODO yet. Maybe it should be replaced by DeviceAPIManager::Get(ctx)->GetStream()
.
thx, I will try unit test it in 8gpu-v100
This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.