1、–nproc_per_node=设置错误，比如就2块可见卡，设置3，那么代码中这行torch.cuda.set_device(args.local_rank)
就会报以下错误

Traceback (most recent call last):
  File "trainDDP.py", line 32, in <module>
    torch.cuda.set_device(args.local_rank)
  File "/opt/conda/envs/torch1.10/lib/python3.8/site-packages/torch/cuda/__init__.py", line 311, in set_device
    torch._C._cuda_setDevice(device)
RuntimeError: CUDA error: invalid device ordinal

2、在使用“DistributedDataParallel”时，model没有先放在GPU上，而是默认的CPU。所以会报如下错误

ValueError: DistributedDataParallel device_ids and output_device arguments only work with single-device/multiple-device GPU modules or CPU modules, but got device_ids [1], output_device 1, and module parameters {device(type='cpu')}.

代码，注意这个args.local_rank 是所有分布式训练中都需要加入的一个args，这个args会被自动传入参数

device = torch.device('cuda:{}'.format(args.local_rank))
#错误
model = VGG16_3D.VGG163D(num_classes=num_classes)
DDPmodel= torch.nn.parallel.DistributedDataParallel(model,device_ids=[args.local_rank],
                                                    output_device=args.local_rank)
#正确
model = VGG16_3D.VGG163D(num_classes=num_classes)
model = model.to(device)
DDPmodel= torch.nn.parallel.DistributedDataParallel(model,device_ids=[args.local_rank],
                                                    output_device=args.local_rank)

3、在优化器中添加参数时，获取参数应该用未DDP封装的model，如果有DDP封装的model就会报如下错误：

optimizer = optim.SGD(train_params, lr=lr, momentum=0.9, weight_decay=5e-4)
  File "/opt/conda/envs/torch1.10/lib/python3.8/site-packages/torch/optim/sgd.py", line 95, in __init__
    super(SGD, self).__init__(params, defaults)
  File "/opt/conda/envs/torch1.10/lib/python3.8/site-packages/torch/optim/optimizer.py", line 54, in __init__
    super(SGD, self).__init__(params, defaults)
  File "/opt/conda/envs/torch1.10/lib/python3.8/site-packages/torch/optim/optimizer.py", line 54, in __init__
    self.add_param_group(param_group)
    self.add_param_group(param_group)  File "/opt/conda/envs/torch1.10/lib/python3.8/site-packages/torch/optim/optimizer.py", line 251, in add_param_group

  File "/opt/conda/envs/torch1.10/lib/python3.8/site-packages/torch/optim/optimizer.py", line 251, in add_param_group
    param_group['params'] = list(params)
  File "/workspace/ws/C3D-main/network/VGG16_3D.py", line 106, in get_1x_lr_params
    param_group['params'] = list(params)
  File "/workspace/ws/C3D-main/network/VGG16_3D.py", line 106, in get_1x_lr_params
    b = [model.conv1a,model.conv1b, model.conv2a,model.conv2b, model.conv3a, model.conv3b,model.conv3c, model.conv4a, model.conv4b,model.conv4c,
  File "/opt/conda/envs/torch1.10/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1177, in __getattr__
    b = [model.conv1a,model.conv1b, model.conv2a,model.conv2b, model.conv3a, model.conv3b,model.conv3c, model.conv4a, model.conv4b,model.conv4c,
  File "/opt/conda/envs/torch1.10/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1177, in __getattr__
        raise AttributeError("'{}' object has no attribute '{}'".format(raise AttributeError("'{}' object has no attribute '{}'".format(

AttributeErrorAttributeError: 'DistributedDataParallel' object has no attribute 'conv1a': 
'DistributedDataParallel' object has no attribute 'conv1a

代码

#错误
model = VGG16_3D.VGG163D(num_classes=num_classes)
model = model.to(device)
DDPmodel= torch.nn.parallel.DistributedDataParallel(model,device_ids=[args.local_rank],
                                                    output_device=args.local_rank)
train_params = [{'params': VGG16_3D.get_1x_lr_params(DDPmodel), 'lr': lr},
                {'params': VGG16_3D.get_10x_lr_params(DDPmodel), 'lr': lr * 10}]
criterion = nn.CrossEntropyLoss()  # standard crossentropy loss for classification
optimizer = optim.SGD(train_params, lr=lr, momentum=0.9, weight_decay=5e-4)
#正确
    model = VGG16_3D.VGG163D(num_classes=num_classes)
    model = model.to(device)
    DDPmodel= torch.nn.parallel.DistributedDataParallel(model,device_ids=[args.local_rank],
                                                    output_device=args.local_rank)
    
    train_params = [{'params': VGG16_3D.get_1x_lr_params(model), 'lr': lr},
                        {'params': VGG16_3D.get_10x_lr_params(model), 'lr': lr * 10}]
    criterion = nn.CrossEntropyLoss()  # standard crossentropy loss for classification
    optimizer = optim.SGD(train_params, lr=lr, momentum=0.9, weight_decay=5e-4)

4、因为分布式是多进程，而且我们一般都是用的all_reduce 即所有的GPU上张量相加，然后再把这些数值都放到这些GPU上，所以在用tensorboardX等工具时，需要注意，不用每个进程都写，因为每个进程都往文件里写入，需要设置锁机制，就会很复杂同时因为用的all_reduce也没必要，所以选定一个进程就行
reduce操作：

all_reduce操作：

#错误方式
log_dir = os.path.join(save_dir, 'models', datetime.now().strftime('%b%d_%H-%M-%S') + '_' + socket.gethostname())
writer = SummaryWriter(log_dir=log_dir)
writer.add_scalar('data/train_loss_epoch', epoch_loss, epoch)
writer.close()
#正确方式
log_dir = os.path.join(save_dir, 'models', datetime.now().strftime('%b%d_%H-%M-%S') + '_' + socket.gethostname())
    os.makedirs(log_dir, exist_ok=True) #exist_ok=True表示如果存在就不创建了
    global_rank=torch.distributed.get_rank()
    if global_rank==0:
        writer = SummaryWriter(log_dir=log_dir)
    else:
        writer  = None

if writer is not None：
     writer.add_scalar('data/train_loss_epoch', epoch_loss, epoch)
if writer is not None：
     writer.close()

5、因为分布式是多进程，而且我们一般都是用的all_reduce 即所有的GPU上张量相加，然后再把这些数值都放到这些GPU上， 所以保存模型的时候，一个进程保存就行了

if torch.distributed.get_rank() == 0:  #一般用0，当然，可以选任意的rank保存。
    torch.save(net, "net.pth")

6、

	model
	print('823看看分布式前后模型名称的变化')
    for name, layer in model.named_modules():
        print(name)  # 比如 model.0.conv
    # DP mode
    if cuda and RANK == -1 and torch.cuda.device_count() > 1:
        model = torch.nn.DataParallel(model)
    for name, layer in model.named_modules():
        print(name) #比如 module.model.0.conv