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
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
@spaces.GPU  # Ensures GPU allocation for this function
def describe_image(image):
    # Define the conversation
    conversation = [
        {
            "role": "User",
            "content": "<image_placeholder>Describe this image in great detail.",
            "images": [image]
        },
        {
            "role": "Assistant",
            "content": ""
        }
    ]

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

    # Run the image encoder to get embeddings
    vl_gpt = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True).to(torch.bfloat16).cuda().eval()
    inputs_embeds = vl_gpt.prepare_inputs_embeds(**prepare_inputs)

    # Generate response from the model
    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

# Gradio interface
def gradio_app():
    with gr.Blocks() as demo:
        gr.Markdown("# Image Description with DeepSeek VL 1.3b\n### Upload an image to receive a detailed description.")

        with gr.Row():
            image_input = gr.Image(type="pil", label="Upload an Image")

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

        submit_btn = gr.Button("Generate Description")

        submit_btn.click(
            fn=describe_image,
            inputs=[image_input],
            outputs=output_text
        )

    demo.launch()

# Launch the Gradio app
gradio_app()