AttributeError: 'Qwen2_5_VLModel' object has no attribute 'embed_tokens'

[gray114514]

VLM-R1-main/src/open-r1-multimodal/src/open_r1/qwen2_5vl_monkey_patch.py", line 79, in qwen2_5vl_forward
[rank0]: inputs_embeds = self.model.embed_tokens(input_ids)

I encountered the following error when trying to access model.embed_tokens:

AttributeError: 'Qwen2_5_VLModel' object has no attribute 'embed_tokens'.

Is it possible that the embedding layer has been renamed or moved in this model? How can I correctly access the token embeddings for the Qwen2.5-VL model?

Thanks for your help!

[xLalalala]

same problem

[XuZhengzhuo]

it works for qwen2.5VL:

  is_embed_trainable = any(
      param.requires_grad for param in self.get_input_embeddings().parameters()
  )

[CuteClaire]

Did you change the transformers version? I ran into this issue too and changing transformers versions worked for me.

[gray114514]

Did you change the transformers version? I ran into this issue too and changing transformers versions worked for me.

Which transformers version did you use?

[CuteClaire]

Did you change the transformers version? I ran into this issue too and changing transformers versions worked for me.

Which transformers version did you use?

Using transformers version 4.51.3 or less worked for me (both 4.51.3 and 4.49.0 worked for me)

[LizzyEw]

Did you change the transformers version? I ran into this issue too and changing transformers versions worked for me.

Which transformers version did you use?

Using transformers version 4.51.3 or less worked for me (both 4.51.3 and 4.49.0 worked for me)

Thanks for your share! It works again now