Closed warner-benjamin closed 3 months ago
LGTM I do not really have time to test the code right now but this is how I imagined it should be done and seems coherent with my findings.
I've extensively tested it, both with a full training run and the smoketests. It should be good to go.
This PR enables setting the
device_eval_microbatch_size
via config or cli arg by wrapping the eval dataloader in a Composer Evaluator.I also removed the
device_eval_batch_size
config option since it didn't do anything.