app.py

import gradio as gr
import torch
from transformers import Qwen2VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info

# Configuration
MODEL_ID = "Qwen/Qwen2-VL-2B-Instruct"

print(f"Loading {MODEL_ID}...")

# 1. Load Model
# We use bfloat16 (half precision) which is faster than 4-bit for small models
# and fits easily in 16GB or even 8GB VRAM.
try:
    model = Qwen2VLForConditionalGeneration.from_pretrained(
        MODEL_ID,
        torch_dtype=torch.bfloat16,
        device_map="auto",
    )
    
    # The min_pixels and max_pixels arguments help control resolution for speed
    processor = AutoProcessor.from_pretrained(MODEL_ID, min_pixels=256*28*28, max_pixels=1280*28*28)
    print("Model loaded successfully!")
except Exception as e:
    print(f"Error loading model: {e}")
    print("Ensure you have a GPU available.")
    exit()

def chat_response(message, history, image_input):
    """
    Main generation function called by Gradio.
    """
    if image_input is None:
        return "Please upload an image first to chat about it!"

    # 2. Prepare the messages for Qwen2-VL
    # Qwen expects a specific format: a list of messages with specific 'type' keys
    messages = [
        {
            "role": "user",
            "content": [
                {
                    "type": "image",
                    "image": image_input, # Pass the PIL image directly
                },
                {"type": "text", "text": message},
            ],
        }
    ]

    # 3. Process inputs
    # qwen_vl_utils helps process the image and text into tensors
    text = processor.apply_chat_template(
        messages, tokenize=False, add_generation_prompt=True
    )
    
    image_inputs, video_inputs = process_vision_info(messages)
    
    inputs = processor(
        text=[text],
        images=image_inputs,
        videos=video_inputs,
        padding=True,
        return_tensors="pt",
    )
    
    # Move inputs to the same device as the model
    inputs = inputs.to(model.device)

    # 4. Generate Response
    # We limit max_new_tokens to 200 for speed
    generated_ids = model.generate(
        **inputs,
        max_new_tokens=200,
        do_sample=True,
        temperature=0.7,
        top_p=0.9
    )

    # 5. Decode output
    # We trim the input tokens from the output to get only the new response
    generated_ids_trimmed = [
        out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
    ]
    
    response = processor.batch_decode(
        generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
    )[0]

    return response

# --- Gradio UI Setup ---
with gr.Blocks(title="Qwen2-VL Chat", theme=gr.themes.Soft()) as demo:
    gr.Markdown("# 🚀 Qwen2-VL-2B: Fast Image Chat")
    gr.Markdown("Upload an image and ask questions. This 2B model is significantly faster than LLaVA-7B.")
    
    with gr.Row():
        with gr.Column(scale=1):
            image_box = gr.Image(type="pil", label="Upload Image")
        
        with gr.Column(scale=2):
            chatbot = gr.ChatInterface(
                fn=chat_response,
                additional_inputs=[image_box],
                title="Chat",
                description="Ask about the uploaded image.",
                examples=[
                    ["What is in this image?", None],
                    ["Describe the lighting.", None],
                    ["Read the text in the image.", None],
                ],
            )

if __name__ == "__main__":
    demo.queue().launch()