Optimizer step being called 2 times when using deepspeed

[harsh2912]

System Info

In version transformers==4.57.3, and deepspeed==0.18.3,

in below screenshot, when accelerator.backward is called, the deepspeed backward internally calls engine.step which is performing optimizer step at gradient accumulation step

The below snapshot is from trainer.py in transformers library

Also, inside trainer as well optimizer.step is called again post this backward at gradient accumulation step, attaching the below SS for reference.

So inherently in a single iteration it is doing two optimizer step which is wrong. Please update this bug.

Who can help?

No response

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

This bug is currently working as a feature

Expected behavior

There should pe single step, but there are inherently two steps for optimizer which is wrong

[Rocketknight1]

I don't think this bug is real! I think the second step being called is this one, which has no effect. So it look a bit strange if you're not familiar with the code, but it's only a single optimizer step.

[harsh2912]

Yes you are correct