Resuming from checkpoint, mid epoch gives a very distorted time estimate
#18 220 ouverte le 3 août 2023
Métriques du dépôt
- Stars
- (26 687 stars)
- Métriques de merge PR
- (Merge moyen 9j 15h) (3 PRs mergées en 30 j)
Description
Bug description
This might be related to : https://github.com/Lightning-AI/lightning/issues/13124
Currently when resuming from a deepspeed checkpoint, it seems like the time estimate, uses the "current running time" against the "total dataset steps". This gives incredibly warped numbers, when resuming mid epoch for long 1 day+ runs.
Where you can see estimates for runs which would have taken hours, being in minutes.
This can be observed even in small dataset / models, where you can see crazy it/s rates - and remaining time estimate - at the start, which improves over time, but never fall back inline with a more realistic estimate (especially if it resumed > 50% mark)
I do not have full repro steps here, but i am filing it so that others might be able to confirm / follow up on it.
What version are you seeing the problem on?
v2.0
How to reproduce the bug
No response
Error messages and logs
No response
Environment
#- Lightning Component (e.g. Trainer, LightningModule, LightningApp, LightningWork, LightningFlow):
#- PyTorch Lightning Version (e.g., 1.5.0):
#- Lightning App Version (e.g., 0.5.2):
#- PyTorch Version (e.g., 2.0):
#- Python version (e.g., 3.9):
#- OS (e.g., Linux):
#- CUDA/cuDNN version:
#- GPU models and configuration:
#- How you installed Lightning(`conda`, `pip`, source):
#- Running environment of LightningApp (e.g. local, cloud):
More info
No response
cc @tchaton @awaelchli