Why blip2_feature_extraction does not align with demo

[Roberyan]

I am using blip2's feature extraction, however for the feature_text, the text_embedding shape is [1,6,768] instead of [1,12,768], I am confused if this is the case or I did some part wrong?

I am using salesforce-lavis ==1.0.2 , and the model related part is

model, vis_processors, txt_processors = load_model_and_preprocess(
        name="blip2_feature_extractor", 
        model_type="pretrain_vitL", 
        is_eval=True, device=device)