From 7f75e47f8421bf4abb476f2bef89c7d7859889fb Mon Sep 17 00:00:00 2001 From: Terry Kong Date: Thu, 14 Nov 2024 00:09:04 -0800 Subject: [PATCH] require grad_sync to be true now for this ckpt Signed-off-by: Terry Kong --- tests/functional/ppo.sh | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tests/functional/ppo.sh b/tests/functional/ppo.sh index 46f4974b5..fb8e7957d 100755 --- a/tests/functional/ppo.sh +++ b/tests/functional/ppo.sh @@ -82,7 +82,7 @@ python -u ${GPFS}/examples/nlp/gpt/serve_ppo_critic.py \ ++model.global_batch_size=1 \ ++model.tensor_model_parallel_size=1 \ ++model.optim.bucket_cap_mb=200 \ - ++model.optim.overlap_grad_sync=False \ + ++model.optim.overlap_grad_sync=True \ ++model.optim.contiguous_grad_buffer=True \ ++trainer.ppo.pad_sequence_length_to_multiple=32 \ model.reward_standardization.enable=True \