[Bug] Unclear what "<model name> does not accept `num_items_in_batch`." means

[RonanKMcGovern]

This error seems to appear with certain models when training with gradient accumulation. Full error:

<model name> does not accept `num_items_in_batch
Using gradient accumulation will be very slightly less accurate.
Read more on gradient accumulation issues here: https://unsloth.ai/blog/gradient

There was a (now) well-known bug with gradient accumulation whereby losses were incorrectly aggregated.

At the time, it was necessary to wrap the trainer for the fix, but transformers then fixed their trainer. If that's the case, what does this error here mean?

I have version:

Name: unsloth
Version: 2025.8.10

I see this for Llama 3.2 1b, and qwen3-4b. Probably other models too.

[mmathew23]

Hi @RonanKMcGovern, could you let us know what version of transformers and trl you're using?

And if you could also let us know the exact model name, and ideally how you're loading the model. I just ran one of our qwen notebooks on colab and it worked, so I think it may be due to a dependency version mismatch.