Moving a graph to GPU will change the default CUDA device

[chwan1016]

🐛 Bug

To Reproduce

import torch
import dgl

torch.cuda.set_device(1)
print(torch.cuda.current_device())  # print 1
device = 'cuda'  # 'cuda:1'
g = dgl.graph((torch.tensor([0, 1, 2]), torch.tensor([1, 2, 3]))).to(device)
print(torch.cuda.current_device())  # print 0

Expected behavior

The index of the current device should not be changed.

Environment

• DGL Version (e.g., 1.0): 0.6
• Backend Library & Version (e.g., PyTorch 0.4.1, MXNet/Gluon 1.3): PyTorch 1.9.0a0+gitaeaa91b
• OS (e.g., Linux): RHEL
• How you installed DGL (conda, pip, source): source
• Build command you used (if compiling from source):
• Python version: 3.8
• CUDA/cuDNN version (if applicable): 11.0
• GPU models and configuration (e.g. V100): NVIDIA GeForce RTX 2080 Ti
• Any other relevant information:

Additional context

[chwan1016]

This issue should be reopened as it has not been completely solved. @BarclayII @yzh119

import torch
import dgl

torch.cuda.set_device(1)
print(torch.cuda.current_device())  # print 1
device = 'cuda:2'
g = dgl.graph((torch.tensor([0, 1, 2]), torch.tensor([1, 2, 3]))).to(device)
print(torch.cuda.current_device())  # print 2