flash-attention/training/configs/experiment/pile/gpt3m-flash-8k.yaml

# @package _global_
defaults:
  - /experiment/pile/gpt3m-flash.yaml

datamodule:
  max_length: 8192
  batch_size: ${eval:"2 if ${train.gpu_mem} < 24 else (4 if ${train.gpu_mem} < 40 else 8)"}

train:
  global_batch_size: 64