app.py

import gradio as gr
from transformers import AutoProcessor, Llama4ForConditionalGeneration
import torch

model_id = "meta-llama/Llama-4-Maverick-17B-128E-Instruct"

processor = AutoProcessor.from_pretrained(model_id)
model = Llama4ForConditionalGeneration.from_pretrained(
    model_id,
    attn_implementation="flex_attention",
    device_map="auto",
    torch_dtype=torch.bfloat16,
)

def analyze_images(image1, image2, question):
    messages = [
        {
            "role": "user",
            "content": [
                {"type": "image", "image": image1},
                {"type": "image", "image": image2},
                {"type": "text", "text": question}
            ]
        }
    ]

    inputs = processor.apply_chat_template(
        messages,
        add_generation_prompt=True,
        tokenize=True,
        return_dict=True,
        return_tensors="pt",
    ).to(model.device)

    outputs = model.generate(
        **inputs,
        max_new_tokens=256,
    )

    response = processor.batch_decode(outputs[:, inputs["input_ids"].shape[-1]:])[0]
    return response.strip()

gr.Interface(
    fn=analyze_images,
    inputs=[
        gr.Image(type="pil", label="Image 1"),
        gr.Image(type="pil", label="Image 2"),
        gr.Textbox(lines=2, label="Your Question"),
    ],
    outputs="text",
    title="LLaMA 4 Multimodal Visual Q&A",
    description="Upload two images and ask a question — powered by LLaMA 4"
).launch()