I refactored out the logic to adjust the batch size and number of GPU's into their own functions. It reproduces the logic we had before and also coves the case where the user explicitly passes a --batch_size_reduction or --gpu_multiplier flag. I did some spot-checks that the auto-generated configs don't change for models with the name 7B, 13B, 30B, 65B.
I refactored out the logic to adjust the batch size and number of GPU's into their own functions. It reproduces the logic we had before and also coves the case where the user explicitly passes a
--batch_size_reductionor--gpu_multiplierflag. I did some spot-checks that the auto-generated configs don't change for models with the name7B,13B,30B,65B.