Output effective batch size and GPU memory usage in logs when using auto_find_batch_size

  1. Is there a metric definition in evaluate that uses pynvml to monitor the GPU RAM usage while we run trainer.train()?

  2. Is there a logging function that will automatically output the effective batch size by computing the the accumulations + batch size when using auto_find_batch_size with trainer.train()?

Seems related to auto_find_batch_size should say what batch size it is using · Issue #21950 · huggingface/transformers · GitHub