grad is none when run gpt2 with pipeline parallelism only

### 🐛 Describe the bug

<img width="1184" alt="image" src="https://user-images.githubusercontent.com/26682403/209071634-85b4ff19-ab62-4633-a9e0-6b3894446347.png">


### Environment

torch==1.12.0a0+8a1a93a
num_gpu=4
pipeline=4
model = gpt2-small