What are the best practices for GPU memory management in PyTorch?

I keep running into OOM errors when training large models. I've tried: - Gradient checkpointing - Mixed precision training - Reducing batch size Any other techniques people recommend?