I note that max_steps is meant to be a multiple of save_steps so that a checkpoint is saved at the end of training. I just wanted to confirm that for e.g. max_steps = 150, save_strategy='steps', save_steps=100 load_best_model_at_end=True
the final 50 steps of training would essentially be thrown away, as there’s no checkpoint saved at that point and the ‘best’ model that’s reloaded (i.e. checkpoint-100) will just write-over the model state after 150 steps (i.e. save_model just saves checkpoint-100’s weights). Thanks!
That’s very true! Do you want to add a warning to the doc?