Aterminate called after throwing an instance of 'std::runtime_error' what(): NCCL Error 1: unhandled cuda error

Question

This error occurs when using DataParallel. but it works when using only 1 GPU. May I ask why this problem occurs and how can I solve it？

Aterminate called after throwing an instance of 'std::runtime_error'
what():  NCCL Error 1: unhandled cuda error
Aborted (core dumped)

My code is:

gpus = [0, 1, 2]
my_model.to(pytorch_device)
my_model = DataParallel(my_model, device_ids=gpus, output_device=gpus[0])