hustvl/SparseInst

The problem of two gpus training

cici-999 opened this issue · 2 comments

Hello, I want to train the model with two gpus, I changed os.environ['CUDA_VISIBLE_DEVICES']='0,1' and gpus=2, there is a RuntimeError:Distributed package doesn't have NCCL built in problem, please have any suggestions?

Hi @cici-999, how about trying CUDA_VISIBLE_DEVICES=0,1 python tools/train_net.py ...

Hi @cici-999, how about trying CUDA_VISIBLE_DEVICES=0,1 python tools/train_net.py ...

I tried the method you mentioned and I don't get RuntimeError, but it still shows 1 gpu for training.