app.py

import gradio as gr
import torch
from transformers import AutoModel, AutoTokenizer
import spaces
import os
import tempfile

# Load model and tokenizer
model_name = "deepseek-ai/DeepSeek-OCR"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModel.from_pretrained(
    model_name,
    _attn_implementation="flash_attention_2",
    trust_remote_code=True,
    use_safetensors=True,
)
model = model.eval()


@spaces.GPU
def process_image(image, model_size, task_type):
    """
    Process image with DeepSeek-OCR

    Args:
        image: PIL Image or file path
        model_size: Model size configuration
        task_type: OCR task type
    """
    # 在 GPU 函数内部移动模型到 GPU
    model_gpu = model.cuda().to(torch.bfloat16)

    # Create temporary directory for output
    with tempfile.TemporaryDirectory() as output_path:
        # Set prompt based on task type
        if task_type == "Free OCR":
            prompt = "<image>\nFree OCR. "
        elif task_type == "Convert to Markdown":
            prompt = "<image>\n<|grounding|>Convert the document to markdown. "
        else:
            prompt = "<image>\nFree OCR. "

        # Save uploaded image temporarily
        temp_image_path = os.path.join(output_path, "temp_image.jpg")
        image.save(temp_image_path)

        # Configure model size parameters
        size_configs = {
            "Tiny": {"base_size": 512, "image_size": 512, "crop_mode": False},
            "Small": {"base_size": 640, "image_size": 640, "crop_mode": False},
            "Base": {"base_size": 1024, "image_size": 1024, "crop_mode": False},
            "Large": {"base_size": 1280, "image_size": 1280, "crop_mode": False},
            "Gundam (Recommended)": {
                "base_size": 1024,
                "image_size": 640,
                "crop_mode": True,
            },
        }

        config = size_configs.get(model_size, size_configs["Gundam (Recommended)"])

        # Run inference
        result = model_gpu.infer(
            tokenizer,
            prompt=prompt,
            image_file=temp_image_path,
            output_path=output_path,
            base_size=config["base_size"],
            image_size=config["image_size"],
            crop_mode=config["crop_mode"],
            save_results=True,
            test_compress=True,
            eval_mode=True,
        )

        print(f"====\nresult: {result}\n====\n")
        return result


# Create Gradio interface
with gr.Blocks(title="DeepSeek-OCR") as demo:
    gr.Markdown(
        """
        # DeepSeek-OCR Document Recognition
        
        Upload an image to extract text using DeepSeek-OCR model. 
        Supports various document types and handwriting recognition.
        
        **Model Sizes:**
        - **Tiny**: Fastest, lower accuracy (512x512)
        - **Small**: Fast, good accuracy (640x640)
        - **Base**: Balanced performance (1024x1024)
        - **Large**: Best accuracy, slower (1280x1280)
        - **Gundam (Recommended)**: Optimized for documents (1024 base, 640 image, crop mode)
        """
    )

    with gr.Row():
        with gr.Column():
            image_input = gr.Image(
                type="pil", label="Upload Image", sources=["upload", "clipboard"]
            )

            model_size = gr.Dropdown(
                choices=["Tiny", "Small", "Base", "Large", "Gundam (Recommended)"],
                value="Gundam (Recommended)",
                label="Model Size",
            )

            task_type = gr.Dropdown(
                choices=["Free OCR", "Convert to Markdown"],
                value="Convert to Markdown",
                label="Task Type",
            )

            submit_btn = gr.Button("Process Image", variant="primary")

        with gr.Column():
            output_text = gr.Textbox(
                label="OCR Result", lines=20, show_copy_button=True
            )

    # Examples
    gr.Examples(
        examples=[
            ["examples/math.png", "Gundam (Recommended)", "Convert to Markdown"],
            ["examples/receipt.jpg", "Base", "Free OCR"],
        ],
        inputs=[image_input, model_size, task_type],
        outputs=output_text,
        fn=process_image,
        cache_examples=False,
    )

    submit_btn.click(
        fn=process_image,
        inputs=[image_input, model_size, task_type],
        outputs=output_text,
    )

# Launch the app
if __name__ == "__main__":
    demo.queue(max_size=20)
    demo.launch()