app.py

import gradio as gr
import spaces
import os
import torch
from transformers import AutoProcessor, MllamaForConditionalGeneration
from PIL import Image, ImageOps
import whisper

# Hugging Face token
hf_token = os.getenv("HUGGING_FACE_HUB_TOKEN")
if not hf_token:
    raise ValueError("HUGGING_FACE_HUB_TOKEN not found.")

# Model
model_name = "meta-llama/Llama-3.2-11B-Vision-Instruct"
model = MllamaForConditionalGeneration.from_pretrained(
    model_name,
    token=hf_token,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
processor = AutoProcessor.from_pretrained(model_name, token=hf_token)

@spaces.GPU
def predict(image, text):
    messages = [
        {"role": "user", "content": [
            {"type": "image"},
            {"type": "text", "text": text}
        ]}
    ]
    
    input_text = processor.apply_chat_template(messages, add_generation_prompt=True)
    inputs = processor(image, input_text, return_tensors="pt").to(model.device)
    outputs = model.generate(**inputs, max_new_tokens=250)
    response = processor.decode(outputs[0], skip_special_tokens=True)
    # Split the response at the first occurrence of "assistant" and return only the part after it
    response = response.split("assistant", 1)[1].strip()
    return response

# Whisper STT optional model
#@spaces.GPU
#def transcribe_audio(audio):
#    result = whisper.transcribe(audio, model="base")
#    return result["text"]

# Example photos and prompts
example_images = [
    ImageOps.exif_transpose(Image.open("Illustration by @twentyone21___.jpg")),
    ImageOps.exif_transpose(Image.open("Kynda Coffee.jpg")),
    ImageOps.exif_transpose(Image.open("Cowboy Hat.jpg")),
    ImageOps.exif_transpose(Image.open("Norway.JPG"))
]
example_prompts = ["Describe the photo", 
                   "Search for the business name on his t-shirt to get a description of where the person is in Texas.", 
                   "Describe the photo", 
                   "Where do you think this photo was taken based on the architecture?"
]

# Gradio
demo = gr.Blocks()

with demo:
    gr.Markdown("# Image Question Answering and Optional (WIP) Audio Transcription")
    
    with gr.Tab("Image & Text Prompt"):
        image_input = gr.Image(type="pil", label="Image Input")
        text_input = gr.Textbox(label="Text Input")
        output = gr.Textbox(label="Output")
        gr.Button("Submit").click(predict, inputs=[image_input, text_input], outputs=output)
        
        gr.Examples(examples=[[image, prompt] for image, prompt in zip(example_images, example_prompts)], inputs=[image_input, text_input])
    
#    with gr.Tab("Audio Transcription (WIP) Prompt"):
#        gr.load("models/openai/whisper-large-v3")
#        audio_input = gr.Audio(label="Audio Input")
#        text_output = gr.Textbox(label="Transcribed Text")
#        gr.Button("Transcribe").click(transcribe_audio, inputs=audio_input, outputs=text_output)
#        gr.Button("Submit").click(predict, inputs=[image_input, text_output], outputs=output)

demo.launch()