app.py

import gradio as gr
import torch
from transformers import AutoModelForCausalLM
from deepseek_vl.models import VLChatProcessor, MultiModalityCausalLM
from deepseek_vl.utils.io import load_pil_images
from io import BytesIO
from PIL import Image
import spaces  # Import spaces for ZeroGPU support

# Load the model and processor
model_path = "deepseek-ai/deepseek-vl-1.3b-chat"
vl_chat_processor = VLChatProcessor.from_pretrained(model_path)
tokenizer = vl_chat_processor.tokenizer

# Define the function for image description with ZeroGPU support
@spaces.GPU  # Ensures GPU allocation for this function
def describe_image(image, user_question="Describe this image in great detail."):
    try:
        # Convert the PIL Image to a BytesIO object for compatibility
        image_byte_arr = BytesIO()
        image.save(image_byte_arr, format="PNG")  # Save image in PNG format
        image_byte_arr.seek(0)  # Move pointer to the start

        # Define the conversation, using the user's question
        conversation = [
            {
                "role": "User",
                "content": f"<image_placeholder>{user_question}",
                "images": [image_byte_arr]  # Pass the image byte array instead of an object
            },
            {
                "role": "Assistant",
                "content": ""
            }
        ]

        # Convert image byte array back to a PIL image for processing
        pil_images = [Image.open(BytesIO(image_byte_arr.read()))]  # Convert byte back to PIL Image
        image_byte_arr.seek(0)  # Reset the byte stream again for reuse

        # Load images and prepare the inputs
        prepare_inputs = vl_chat_processor(
            conversations=conversation,
            images=pil_images,
            force_batchify=True
        ).to('cuda')

        # Load and prepare the model
        vl_gpt = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True).to(torch.bfloat16).cuda().eval()

        # Generate embeddings from the image input
        inputs_embeds = vl_gpt.prepare_inputs_embeds(**prepare_inputs)

        # Generate the model's response
        outputs = vl_gpt.language_model.generate(
            inputs_embeds=inputs_embeds,
            attention_mask=prepare_inputs.attention_mask,
            pad_token_id=tokenizer.eos_token_id,
            bos_token_id=tokenizer.bos_token_id,
            eos_token_id=tokenizer.eos_token_id,
            max_new_tokens=512,
            do_sample=False,
            use_cache=True
        )

        # Decode the generated tokens into text
        answer = tokenizer.decode(outputs[0].cpu().tolist(), skip_special_tokens=True)
        return answer

    except Exception as e:
        # Provide detailed error information
        return f"Error: {str(e)}"

# Gradio interface
def gradio_app():
    with gr.Blocks() as demo:
        gr.Markdown("# Image Description with DeepSeek VL 1.3b 🐬\n### Upload an image and ask a question about it.")

        with gr.Row():
            image_input = gr.Image(type="pil", label="Upload an Image")
            question_input = gr.Textbox(
                label="Question (optional)", 
                placeholder="Ask a question about the image (e.g., 'What is happening in this image?')",
                lines=2
            )

        output_text = gr.Textbox(label="Image Description", interactive=False)

        submit_btn = gr.Button("Generate Description")

        submit_btn.click(
            fn=describe_image,
            inputs=[image_input, question_input],  # Pass both image and question as inputs
            outputs=output_text
        )

    demo.launch()

# Launch the Gradio app
gradio_app()