What's the usage of num_logits_to_keep?

[FuRuF-11]

I was trying to run train.py with Qwen2.5 1.5B. And a error occurred.

File "/home/comp/xxxxx/miniconda3/envs/rex/lib/python3.12/site-packages/torch/amp/autocast_mode.py", line 44, in decorate_autocast
return func(args, *kwargs)
^^^^^^^^^^^^^^^^^^^^^
File "/home/comp/xxxxx/PathIndependent/OpenCoconut/opencoconut/models/qwen2.py", line 368, in forward
outputs = forward_fn(
^^^^^^^^^^^
File "/home/comp/xxxxx/PathIndependent/OpenCoconut/opencoconut/models/qwen2.py", line 329, in train_forward
outputs = super().forward(
^^^^^^^^^^^^^^^^
TypeError: Qwen2ForCausalLM.forward() got an unexpected keyword argument 'num_logits_to_keep'

So I just deleted num_logits_to_keep in qwen2.py. Then the program ran.

But I am still confused about the usage of num_logits_to_keep.

And I am also worried that simply delete this arg will lead to another bug.

I would be very grateful if you could relieve my confusion.

[FuRuF-11]

Fixed now.

num_logits_to_keep is not available for transformers 4.40.1.

After updated my package, everything is good.