memory leakage problem when loss.backward()

Hi.

How can i solve memory leakage problem on loss.backward?
Actually my code is a bit complex, which makes it hard to provide whole.

The main memory burden comes from here. Exactly after the execution of loss.backward(). The memory increases for every iteration, which leads to OOM.
```python
loss = bce_extended(logits, y).sum()
with backpack(BatchGrad()):
if real_sample:
loss.backward(inputs = list(model.parameters()))
```

I also tried ```with disable():``` which prevent the memory leak problem. However, it cannot be implemented with ```with backpack(BatchGrad()):``` when i want to get the per-sample gradient.