Error When Testing Icon Classification with BLIP-2 Model

Description:
When running the command below to test icon classification with the BLIP-2 model, I encountered a tensor mismatch error. Despite following the instructions referenced in the error message, the issue persists.

python gradio_demo.py --icon_detect_model weights/icon_detect_v1_5/model_v1_5.pt --icon_caption_model blip2

Steps to Reproduce:

1. Run the following command:

   python gradio_demo.py --icon_detect_model weights/icon_detect_v1_5/model_v1_5.pt --icon_caption_model blip2
   

2. Observe the output log:

   Using MPS device.
   Loading checkpoint shards: 100%|...| 2/2 [00:36<00:00, 18.20s/it]
   * Running on local URL: http://127.0.0.1:7860
   
   To create a public link, set share=True in launch().
   start processing datetime: 2025-01-15 16:02:40.503170
   
   image 1/1 ...binary_image_demo.png: 576x1280 21 icons, 429.1ms
   Speed: 3.4ms preprocess, 429.1ms inference, 18.8ms postprocess per image at shape (1, 3, 576, 1280)
   Expanding inputs for image tokens in BLIP-2 should be done in processing. Please follow instructions here: (https://gist.github.com/zucchini-nlp/e9f20b054fa322f84ac9311d9ab67042)...
   Sizes of tensors must match except in dimension 1. Expected size 19 but got size 1 for tensor number 1 in the list.
   

Environment:

• OS: macOS
• Device: MPS (Metal Performance Shaders)
• Python Version: 3.12
• Gradio Version: 5.12.0
• BLIP-2 Model: Salesforce/blip2-opt-2.7b
• Icon Detection Model Path: weights/icon_detect_v1_5/model_v1_5.pt

Additional Notes:

1. The error mentions following a specific [gist](https://gist.github.com/zucchini-nlp/e9f20b054fa322f84ac9311d9ab67042), but even after following its instructions, the issue persists.

Let me know if additional information or clarifications are needed!
thanks

Start by running the command in gradio_demo.py with the stated macOS, MPS, Python, and BLIP-2 setup, then review the linked BLIP-2 processing guidance. Trace the tensor construction around the BLIP-2 inference path and confirm the command completes without the size-mismatch error.