Grpotrainer cannot find the "pad_token_id" when using Qwen2.5-VL-72B-Instruct model

[doggystyle-star]

I try to I tried to run the (Qwen2.5_(3B)-GRPO.ipynb) replace the model Qwen2.5-3B to Qwen2.5-VL-72B, but it encountered this error

"qwen2vlprocessor object has no attribute 'pad_token_id' ".

but in the preprocesser_config.json the "pad_token_id" is set 151643 how can I fix it? here is my package vision: huggingface-hub==0.28.1 pillow==10.4.0 transformers==4.49.0 triton==3.1.0 trl==0.15.1 unsloth==2025.2.12 unsloth-zoo=2025.2.5 VLLM==0.7.2

[danielhanchen]

Oh my - will work on a fix!

[jia-zhen-yu]

Have you encountered a situation where the merged fine-tuned model can't find the local model? If so, how did you resolve it?"

[doggystyle-star]

Have you encountered a situation where the merged fine-tuned model can't find the local model? If so, how did you resolve it?"

No, I havent yet, maybe you should add before: model, tokenizer = FastVisionModel.from_pretrained( "{YOUR_MODEL}", load_in_4bit = True, # Use 4bit to reduce memory use. False for 16bit LoRA. use_gradient_checkpointing = "unsloth", # True or "unsloth" for long context )

[doggystyle-star]

maybe the source of the problem is trl(version==0.15.1), the above is I replace the pad_token_id in Qwen-VL-72B but the new problem is occur:

also in trl==0.15.1

[shimmyshimmer]

@doggystyle-star we fixed some issues for qwen2.5vl let us know if youre sitll having the issue. you can reopen!