How does pytorch check which GPU the tensor and model are on and specify the GPU device

1. View the device where the tensor is located:

data = data.cuda()#将数据转移到gpu上
 
print(data.device)  # 输出：cuda:0
 
data = data.cpu()#将数据转移到cpu上
 
print(data.device)  # 输出：cpu

2. View the device where the model is located

model = model.cuda()#将模型转移到gpu上
 
print(next(model.parameters()).device)  # 输出：cuda:0
 
model = model.cpu()#将模型转移到cpu上
 
print(next(model.parameters()).device)  # 输出：cpu

3. There are two common ways to load models and tensors to GPU in Pytorch.

Method 1:

# 如果GPU可用，将模型和张量加载到GPU上
if torch.cuda.is_available():
    model = model.cuda()
    x = x.cuda()
    y = y.cuda()

Method 2:

# 分配到的GPU或CPU
device=torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# 将模型加到GPU
model=model.to(device)
# 将张量加到GPU
x=x.to(device)
y=y.to(device)

4. Specify the GPU code

# 代码1：
torch.cuda.set_device(1)

# 代码2：
device = torch.device("cuda:1")

# 代码3：（官方推荐使用）,
os.environ["CUDA_VISIBLE_DEVICES"] = '1'

（如果你想同时调用两块GPU的话）
os.environ["CUDA_VISIBLE_DEVICES"] = '1,2'

Reference link: Select the specified GPU in PyTorch

Note that the specified GPU code needs to be placed at the beginning of the program segment, as shown in the following figure:

5. View the number of GPUs

torch.cuda.device_count()