app.py

import spaces
import gradio as gr
import torch
from transformers import AutoProcessor, AutoModelForImageTextToText
import os

hf_token = os.environ.get("HF_TOKEN")
model_id = "CohereForAI/aya-vision-8b"

# Load the model and processor on startup.
try:
    processor = AutoProcessor.from_pretrained(model_id)
    model = AutoModelForImageTextToText.from_pretrained(
        model_id, device_map="auto", torch_dtype=torch.float16, use_auth_token=hf_token
    )
    model_status = "Model loaded successfully!"
except Exception as e:
    processor = None
    model = None
    model_status = (
        f"Error loading model: {e}\nMake sure to install the correct version of transformers with: "
        "pip install 'git+https://github.com/huggingface/transformers.git@v4.49.0-AyaVision'"
    )

@spaces.GPU
def process_image_and_prompt(uploaded_image, image_url, prompt, temperature=0.3, max_tokens=300):
    global processor, model

    if processor is None or model is None:
        return "Model failed to load. Please check the logs."
    
    # Determine which image input to use:
    if uploaded_image:
        # If an image is uploaded, use the image directly.
        messages = [{
            "role": "user",
            "content": [
                {"type": "image", "image": uploaded_image},
                {"type": "text", "text": prompt},
            ],
        }]
    elif image_url and image_url.strip():
        # Otherwise, use the provided image URL.
        img_url = image_url.strip()
        messages = [{
            "role": "user",
            "content": [
                {"type": "image", "url": img_url},
                {"type": "text", "text": prompt},
            ],
        }]
    else:
        return "Please provide either an image upload or an image URL."
    
    try:
        inputs = processor.apply_chat_template(
            messages,
            padding=True,
            add_generation_prompt=True,
            tokenize=True,
            return_dict=True,
            return_tensors="pt"
        ).to(model.device)

        gen_tokens = model.generate(
            **inputs,
            max_new_tokens=int(max_tokens),
            do_sample=True,
            temperature=float(temperature),
        )
    
        response = processor.tokenizer.decode(
            gen_tokens[0][inputs.input_ids.shape[1]:],
            skip_special_tokens=True
        )
        return response
    except Exception as e:
        return f"Error generating response: {e}"

# Example inputs for testing.
examples = [
    [None, "https://media.istockphoto.com/id/458012057/photo/istanbul-turkey.jpg?s=612x612&w=0&k=20&c=qogAOVvkpfUyqLUMr_XJQyq-HkACXyYUSZbKhBlPrxo=", "What landmark is shown in this image?", 0.3, 300],
    [None, "https://pbs.twimg.com/media/Fx7YvfQWYAIp6rZ?format=jpg&name=medium", "What does the text in this image say?", 0.3, 300],
    [None, "https://upload.wikimedia.org/wikipedia/commons/d/da/The_Parthenon_in_Athens.jpg", "Describe esta imagen en español", 0.3, 300]
]

# Build the Gradio interface.
with gr.Blocks(title="Aya Vision 8B Demo") as demo:
    gr.Markdown("# Aya Vision 8B Model Demo")
    gr.Markdown(
        """
This app demonstrates the Aya Vision 8B model. You can either upload an image or provide an image URL. Enter a prompt along with the image.
        """
    )
    gr.Markdown(f"**Model Status:** {model_status}")

    gr.Markdown("### Provide an Image")
    with gr.Tab("Upload Image"):
        # Using type="filepath" returns the local file path which is then passed directly.
        image_upload = gr.Image(label="Upload Image", type="filepath")
    with gr.Tab("Image URL"):
        image_url_input = gr.Textbox(label="Image URL", placeholder="Enter a direct image URL")
    
    prompt = gr.Textbox(label="Prompt", placeholder="Enter your prompt here", lines=3)
    
    with gr.Accordion("Generation Settings", open=False):
        temperature_slider = gr.Slider(minimum=0.0, maximum=1.0, step=0.1, value=0.3, label="Temperature")
        max_tokens_slider = gr.Slider(minimum=50, maximum=1000, step=50, value=300, label="Max Tokens")
    
    generate_btn = gr.Button("Generate Response", variant="primary")
    output = gr.Textbox(label="Model Response", lines=10)
    
    gr.Markdown("### Examples")
    gr.Examples(
        examples=examples,
        inputs=[image_upload, image_url_input, prompt, temperature_slider, max_tokens_slider],
        outputs=output,
        fn=process_image_and_prompt
    )
    
    def generate_response(uploaded_image, image_url, prompt, temperature, max_tokens):
        return process_image_and_prompt(uploaded_image, image_url, prompt, temperature, max_tokens)
    
    generate_btn.click(
        generate_response,
        inputs=[image_upload, image_url_input, prompt, temperature_slider, max_tokens_slider],
        outputs=output
    )

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