GPU load is severely unbalanced about lama

[windj007]

Hi! Sorry for the leate reply!

You can try to turn off the validation - then the pipeline would use less gpu memory and load will be a bit more balanced. However, as far as I know, this is a usual issue with DDP - it always occupies more memory on gpu #0 than on others. If you want to achieve equal memory consumption, try horovod - it should be pretty easy to set up with pytorch-lightning (however, I did not try it myself)

[windj007]

This link might be also useful

[jxbs]

You can try to turn off the validation
how to turn off the validation? we should update .yaml?

[windj007]

Just set very long period of validation, by adjusting .yaml or by adding a command line argument to train.py, e.g. trainer.kwargs.check_val_every_n_epoch=10000

However, when turning validation off you have to adjust parameters for saving the checkpoints (see configs/trainer/*.yaml)

Instead of this

checkpoint_kwargs:
  verbose: True
  save_top_k: 5
  save_last: True
  period: 1
  monitor: val_lpips_fid100_f1_total_mean
  mode: max

Write this:

checkpoint_kwargs
  verbose: True
  period: 1

As a result, you'll have a model dumped after each epoch - so you'll have to score them afterwards and choose the best one.