Assuming we will be doing large amounts of data filtering dynamically on JZ, we may need to do this we will need to adapt this: https://github.com/bigscience-workshop/Megatron-DeepSpeed/pull/60 To run the filtering pipeline.