RuntimeError: Expected there to be 1 prompt updates corresponding to 1 image items, but instead found 0 prompt updates

[FerryHuang]

RuntimeError: Expected 1 prompt updates but found 0 prompt updates

Error

RuntimeError: Expected there to be 1 prompt updates corresponding to 1 image items, but instead found 0 prompt updates! Either the prompt text has missing/incorrect tokens for multi-modal inputs, or there is a problem with your implementation of merged multi-modal processor for this model (usually arising from an inconsistency between `_call_hf_processor` and `_get_prompt_updates`).

Model

• Qwen-2.5-VL-7B model

Data Format

All verified correct:

• Every sample has exactly 1 image and 1 <image> tag
• Image format: {"image": "/path/to/image.jpg"}
• Prompt format: "<image>Math problem text..."

Training

Using DAPO training with VERL framework.

Data format is confirmed correct - seems like a model processor compatibility issue rather than data problem.

[lzcomeon]

Have you solved this problem? I encountered the same issue, and I've verified the data multiple times. I'm using Swift for the GRPO training on my end.

MODEL：Qwen-2.5-VL-3B pip install transformers==4.51.3 accelerate pip install trl==0.17.0 pip install vllm==0.8.5

[Y-L-LIU]

Have you includeddata.image_key=images in the training script

[cui36]

I think the prompt <image> can be modified to <im_start>, like this: vision_language.py

[disperaller]

does verl support DAPO with each sample contain ONLY ONE image? But the qwen 2.5 vl is able to perform sft with each sample contain multiple images