Some question about out of memory?

[upczxy]

Hello, In the paper you say the code can run under a 16G GPU when batch is under 4 . but It still showed out of the memory when I tested. What should I do

[gaozhihan]

Thanks for reporting it. You could try either

1. using smaller micro_batch_size, e.g. set micro_batch_size: 2 or micro_batch_size: 1 in cfg.yaml config files.
2. using gradient checkpointing by setting checkpoint_level: 1 or checkpoint_level: 2 in cfg.yaml config files.

[sxjscience]

Also, @upczxy can you point us to the training code that you cannot run? We are happy to help.

[upczxy]

thank you very much,the code can run now

[sxjscience]

@upczxy Let me close the issue for now. Feel free to reopen if you find any problems.