The question about training speed

[fxcdl]

Hi, thank u for u code, I trained GRL base model on dn task, but the speed of training is a bit slow, is this normal?

training config
Batch = 1
gpus = 2
workers = 2
patch_size = 128
stripe_size1 = 32
stripe_size2 = 64

[PyoJunCode]

I have same issue, It tooks about 90sec to 10iteration with 288 patch size.

[QMoQ]

the same. i trained grl-s and it cost much more memory than i thought.

[fxcdl]

the same. i trained grl-s and it cost much more memory than i thought.
i trained grl-s cost 34G memory on A100, but grl-base is out of memory

[Kiteretsu77]

Is it possible for the author to publish an official training code in ReadMe.md? Thanks!