app.py

import gradio as gr
from transformers import AutoProcessor, PaliGemmaForConditionalGeneration
from peft import PeftModel
import spaces

device = "cuda"
@spaces.GPU
def greet(image, prompt):
    base_model = PaliGemmaForConditionalGeneration.from_pretrained("google/paligemma-3b-mix-224").to(device)
    processor = AutoProcessor.from_pretrained("google/paligemma-3b-mix-224")
    model = PeftModel(base_model, "/folders").to(device)
    inputs = processor(prompt, raw_image, return_tensors="pt")
    output = model.generate(**inputs, max_new_tokens=20)
    return output

demo = gr.Interface(fn=greet, inputs=[gr.Image(label="Upload image", sources=['upload', 'webcam'], type="pil"), gr.Text()], outputs="text")
demo.launch()