pytorch gradient checkpointing is much better than deepspeed ! about deepspeed HOT 4 CLOSED

You can train a larger batch size in two ways:

1. Use gradient accumulation. If you want to train a batch size of 32 but can only fit batch size of 4, you can use a micro_batch_per_gpu of size 4 and gradient_accumulation_step of 8. Was the original code using gradient accumulation? Are the batch sizes consistent between deepscale config file and the data loader you are using?

2. Use activation check-pointing/re-materialization/re-computation. I assume this is what you mean by gradient check-pointing. With this you should be able to run a significantly larger batch size for the model you are describing.

What is the sequence length you are using, and what are your hidden dimensions?

Please take a look at the line 390-398 in the following file for example of activation checkpointing.

https://github.com/microsoft/DeepSpeedExamples/blob/07d1ce0d26044602a7b8bb289e590a980f14aded/Megatron-LM/mpu/transformer.py

from deepspeed.

Thanks @samyam for the feedback.

In the original code, I was using both gradient accumulation and activation check-pointing/re-materialization/re-computation.
The total batch size was: 32 batch size * 8 gradient accumulation * 6 GPUs
Input length: 1024
embedding dimensions: 768
hidden dimensions: 3072
Vocab: 22

The gradient accumulation is not an issue between my code and deepspeed code.
The batch size is consistent between my custom data loader and deepspeed config file.

The problem is in the activation check-pointing/re-materialization/re-computation. I assumed if I disabled it and activate deepspeed, I can get the same results. Apparently, this is not the case.

I will re-integrate gradient checkpoint and see if there is a benefit of using deepspeed in this case or not.

from deepspeed.

Yes, DeepSpeed does not do activation check-pointing automatically. You have to add that in the model. Please keep us posted on if this solves your issue

from deepspeed.

Closing the issue. Please let us know if there are further questions or issues!

from deepspeed.