app.py

import gradio as gr
from PIL import Image
import torch
from transformers import AutoProcessor, AutoModelForVision2Seq
import os

# Determine the device to use
# Using os.environ.get to allow device override from Space hardware config if needed
# Defaults to CUDA if available, else CPU.
device_choice = os.environ.get("DEVICE", "auto")
if device_choice == "auto":
    device = "cuda" if torch.cuda.is_available() else "cpu"
else:
    device = device_choice

print(f"Using device: {device}")

# Load the model and processor
model_id = "lusxvr/nanoVLM-222M"
try:
    processor = AutoProcessor.from_pretrained(model_id)
    model = AutoModelForVision2Seq.from_pretrained(model_id).to(device)
    print("Model and processor loaded successfully.")
except Exception as e:
    print(f"Error loading model/processor: {e}")
    # If loading fails, we'll have the Gradio app display an error.
    # This helps in debugging if the Space doesn't start correctly.
    processor = None
    model = None

def generate_text_for_image(image_input, prompt_input):
    """
    Generates text based on an image and a text prompt.
    """
    if model is None or processor is None:
        return "Error: Model or processor not loaded. Check the Space logs for details."

    if image_input is None:
        return "Please upload an image."
    if not prompt_input:
        return "Please provide a prompt (e.g., 'Describe this image' or 'What color is the car?')."

    try:
        # Ensure the image is in PIL format and RGB
        if not isinstance(image_input, Image.Image):
            pil_image = Image.fromarray(image_input)
        else:
            pil_image = image_input
        
        if pil_image.mode != "RGB":
            pil_image = pil_image.convert("RGB")

        # Prepare inputs for the model
        # The prompt for nanoVLM is typically a question or an instruction.
        inputs = processor(text=[prompt_input], images=[pil_image], return_tensors="pt").to(device)

        # Generate text
        # You can adjust max_new_tokens, temperature, top_k, etc.
        generated_ids = model.generate(
            **inputs,
            max_new_tokens=150,  # Increased for potentially longer descriptions
            num_beams=3,         # Example of adding beam search
            no_repeat_ngram_size=2,
            early_stopping=True
        )
        
        # Decode the generated tokens
        generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
        
        # The output might sometimes include the prompt itself, depending on the model.
        # Simple heuristic to remove prompt if it appears at the beginning:
        if generated_text.startswith(prompt_input):
             cleaned_text = generated_text[len(prompt_input):].lstrip(" ,.:")
        else:
            cleaned_text = generated_text

        return cleaned_text.strip()

    except Exception as e:
        print(f"Error during generation: {e}")
        return f"An error occurred: {str(e)}"

# Create the Gradio interface
description = """
Upload an image and provide a text prompt (e.g., "What is in this image?", "Describe the animal in detail.").
The model will generate a textual response based on the visual content and your query.
This Space uses the `lusxvr/nanoVLM-222M` model.
"""

# Example image from COCO dataset
example_image_url = "http://images.cocodataset.org/val2017/000000039769.jpg" # A cat and a remote

iface = gr.Interface(
    fn=generate_text_for_image,
    inputs=[
        gr.Image(type="pil", label="Upload Image"),
        gr.Textbox(label="Your Prompt/Question", info="e.g., 'What is this a picture of?', 'Describe the main subject.', 'How many animals are there?'")
    ],
    outputs=gr.Textbox(label="Generated Text", show_copy_button=True),
    title="Interactive nanoVLM-222M Demo",
    description=description,
    examples=[
        [example_image_url, "a photo of a"],
        [example_image_url, "Describe the image in detail."],
        [example_image_url, "What objects are on the sofa?"],
    ],
    cache_examples=True # Cache results for examples to load faster
)

if __name__ == "__main__":
    # For Hugging Face Spaces, it's common to launch with server_name="0.0.0.0"
    # The Space infrastructure handles the public URL and port mapping.
    iface.launch(server_name="0.0.0.0", server_port=7860)