composer
composer copied to clipboard
eval_batch_size=auto
Oops meant to open draft PR -- not ready for review yet!
I think this works -- tested on resnet with large batch size, but I'd love to get some help testing this on more workloads. I honestly don't have the bandwidth to test this across a good amount of workloads, so I'm going to leave this to people who've been requesting this feature @moinnadeem @abhi-mosaic.
Otherwise, I'll get to this sometime mid-late next week (OOO for rest of this week)
Is this ready for review?
Is this ready for review?
Yes! :)
Do we clear the already computed gradients if we run out of memory while backward is partially executed? PyTorch will do += for such already computed gradients on retry.