How to retrive the raw attention scores or logits from blip model ( image captioning)

[umme17]

from PIL import Image
import requests
url = "https://encrypted-tbn0.gstatic.com/images?q=tbn:ANd9GcTrp7yKuY1NxcXlHQX10JtTlECuna-xWv-jetxnv73WBw&s"
image = Image.open(requests.get(url, stream=True).raw)

prepare image for the model

inputs_base= processor_base(images=image, return_tensors="pt")
pixel_values = inputs_base.pixel_values
outputs_base = model_base.generate(**inputs_base, renormalize_logits=True, max_length=50 )
generated_caption_base = processor_base.batch_decode(outputs_base[0], skip_special_tokens=True)[0]
print(f"Generated caption: {generated_caption_base}")

I want to retrive the raw attention scores or logits from outputs_base. then make some change to the logits and then generate the new caption for new logits. how to do that?