并行训练模型，模型的参数修改过程是累积的。

import torch
import torch.nn as nn
import os
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = "0, 1,2,3"
class MyModel(nn.Module):
    def __init__(self):
        super(MyModel, self).__init__()
        self.counter = torch.tensor(0)

    def forward(self, x):
        self.counter += 1
        return x

# 假设有输入数据x
model = MyModel()

# 使用四个GPU进行数据并行训练
model = nn.DataParallel(model, device_ids=[0, 1,2,3])
x = torch.rand((128,))
x = x.to("cuda:0")

# 第一次调用forward
output1 = model(x)

# 获取counter的值
counter_value = model.module.counter.item()

print("Counter value:", counter_value)

输出为4