diff --git "a/AscendPyTorch\346\250\241\345\236\213\344\274\227\346\231\272\346\226\207\346\241\243-\350\256\255\347\273\203.md" "b/AscendPyTorch\346\250\241\345\236\213\344\274\227\346\231\272\346\226\207\346\241\243-\350\256\255\347\273\203.md"
index 54a541bd99e3338c38d2d1e607b27056c1335748..353c0b87150bd36e843c3f6fc3191344e2216650 100644
--- "a/AscendPyTorch\346\250\241\345\236\213\344\274\227\346\231\272\346\226\207\346\241\243-\350\256\255\347\273\203.md"
+++ "b/AscendPyTorch\346\250\241\345\236\213\344\274\227\346\231\272\346\226\207\346\241\243-\350\256\255\347\273\203.md"
@@ -201,8 +201,43 @@ print(prof.key_averages().table(sort_by="self_cpu_time_total"))
     > **实际操作中推荐优先使用 ```opt_level='O2', loss_scale=128.0``` 的配置进行amp.initialize**
 - 若源代码为单P训练，则需要修改为8P训练
     1. 初始化group
+       ```python
+       dist.init_process_group(backend=args.dist_backend, world_size=args.world_size, rank=args.rank)
+       ```
+       - 参数说明：
+         - dist_backend：GPU环境默认值nccl, NPU环境默认值hccl
+         - world_size：取值1,2,4,8。8p训练参数配置为8
+         - rank：进程id，每个进程对应一个id
+
     2. 添加ddp
+       ```python
+       model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu])
+       ```
+       - 8p训练中，args.rank和args.gpu值相等
+
     3. 添加DistributedSampler
+       ```python
+        if is_distributed == 0:
+            self.train_sampler = None
+        else:
+            self.train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset)
+       ```
+        - 一般只需要对训练数据进行DistributedSampler，load数据时也要同步适配
+        ```python
+        self.train_dataloader = data.DataLoader(dataset=train_dataset,
+                                                batch_size=cfg["train"]["train_batch_size"],
+                                                shuffle=(self.train_sampler is None) ,
+                                                num_workers=workers,
+                                                pin_memory=False,
+                                                sampler=self.train_sampler ,
+                                                drop_last=False)
+        ```
+        - 同时在train epoch循环过程中，需要添加set_epoch步骤
+        ```python
+        for epoch in range(self.epochs):
+            if self.dataparallel:
+                 self.train_sampler.set_epoch(epoch) 
+        ```
 - GPU需要保存的数据
     >![](https://gitee.com/wangjiangben_hw/ascend-pytorch-crowdintelligence-doc/raw/master/public_sys-resources/icon-note.gif) 
     **说明：**