app.py

import gradio as gr
from transformers import AutoProcessor, AutoModelForVision2Seq
import torch

class DeepSeekVL:
    def __init__(self, model_path="deepseek-ai/deepseek-vl-7b", device="cpu"):
        self.device = device
        self.processor = AutoProcessor.from_pretrained(model_path)
        self.model = AutoModelForVision2Seq.from_pretrained(
            model_path,
            torch_dtype=torch.float32
        ).to(device)

    def generate(self, image, question, max_new_tokens=128):
        inputs = self.processor(text=question, images=image, return_tensors="pt").to(self.device)
        with torch.no_grad():
            output_ids = self.model.generate(**inputs, max_new_tokens=max_new_tokens)
        return self.processor.batch_decode(output_ids, skip_special_tokens=True)[0]

# Initialize DeepSeek-VL model (CPU for free Spaces)
model = DeepSeekVL(model_path="deepseek-ai/deepseek-vl-7b", device="cpu")

def qa(image, question):
    # Run DeepSeek-VL inference: image + question -> answer
    return model.generate(image, question)

demo = gr.Interface(
    fn=qa,
    inputs=[
        gr.Image(type="pil", label="Upload Image"),
        gr.Textbox(label="Enter your question")
    ],
    outputs="text",
    title="DeepSeek-VL Multimodal QA Demo",
    description="Upload an image and enter a question. Experience DeepSeek-VL's vision-language capabilities."
)

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