idist initialize with nccl backend without GPU

[sdesrozis]

🐛 Bug description

Call idist.initialize("nccl", ...) without GPU (but pytorch providing nccl) raises an error

Floating point exception

Exception is thrown in barrier following dist.initialize.

Case nccl + no GPU should raise an explicit exception.

Environment

• PyTorch Version (e.g., 1.4): 1.5
• Ignite Version (e.g., 0.3.0): 0.4
• OS (e.g., Linux): Linux
• How you installed Ignite (conda, pip, source): conda
• Python version: 3.7
• Any other relevant information:

[vfdev-5]

Similar code on pure pytorch

> CUDA_VISIBLE_DEVICES="" python -m torch.distributed.launch --use_env --nproc_per_node=1 ddp-no-gpu.py 
import torch.distributed as dist

if __name__ == '__main__':
    dist.init_process_group("nccl", init_method="env://")
    dist.barrier()

just say

subprocess.CalledProcessError: Command '['/opt/conda/bin/python', '-u', 'ddp-no-gpu.py']' died with <Signals.SIGFPE: 8>.

Anyway, good catch!