Closed feifeibear closed 2 years ago
Hi, sorry for the late reply!
It does seem that there's no way as of now to make this compatible with checkpointing. Perhaps in the future I'll figure a way to do this. I'll leave this issue open in case something comes up.
Thanks! good job anyway. I closed the issue.
Hello, I found this project is interesting. However, I found the lazy tensor mechanism is impossible to work with the PyTorch backward hooks, which makes it difficult to be used in combination with PyTorch checkpointing (https://pytorch.org/docs/stable/checkpoint.html). Checkpointing is a common way to avoid OOM in training.