Document OCR job queue

This tutorial shows you how to use Modal as an infinitely scalable job queue that can service async tasks from a web app. For the purpose of this tutorial, we’ve also built a React + FastAPI web app on Modal that works together with it, but note that you don’t need a web app running on Modal to use this pattern. You can submit async tasks to Modal from any Python application (for example, a regular Django app running on Kubernetes).

Our job queue will handle a single task: running OCR transcription for images. We’ll make use of a pre-trained Document Understanding model using the donut package to accomplish this. Try it out for yourself here.

receipt parser frontend

Define a Stub

Let’s first import modal and define a Stub. Later, we’ll use the name provided for our Stub to find it from our web app, and submit tasks to it.

import modal

stub = modal.Stub("doc_ocr_jobs")

Model cache

donut downloads the weights for pre-trained models to a local directory, if those weights don’t already exist. To decrease start-up time, we want this download to happen just once, even across separate function invocations. To accomplish this, we use a SharedVolume, a writable volume that can be attached to Modal functions and persisted across function runs.

volume = modal.SharedVolume().persist("doc_ocr_model_vol")
CACHE_PATH = "/root/model_cache"

Handler function

Now let’s define our handler function. Using the @stub.function decorator, we set up a Modal Function that uses GPUs, has a SharedVolume mount, runs on a custom container image, and automatically retries failures up to 3 times.

    shared_volumes={CACHE_PATH: volume},
def parse_receipt(image: bytes):
    from PIL import Image
    from donut import DonutModel
    import torch
    import io

    # Use donut fine-tuned on an OCR dataset.
    task_prompt = "<s_cord-v2>"
    pretrained_model = DonutModel.from_pretrained(
        "naver-clova-ix/donut-base-finetuned-cord-v2", cache_dir=CACHE_PATH

    # Initialize model.
    device = torch.device("cuda")

    # Run inference.
    input_img =
    output = pretrained_model.inference(image=input_img, prompt=task_prompt)[
    print("Result: ", output)

    return output


Now that we have a function, we can publish it by deploying the app:

modal app deploy

Once it’s published, we can look up this function from another Python process and submit tasks to it:

fn = modal.lookup("doc_ocr_jobs", "parse_receipt")

Modal will auto-scale to handle all the tasks queued, and then scale back down to 0 when there’s no work left. To see how you could use this from a Python web app, take a look at the receipt parser frontend tutorial.

Run manually

We can also trigger parse_receipt manually for easier debugging. To try it out, you can find some example receipts here.

if __name__ == "__main__":
        with open("./receipt.png", "rb") as f:
            image =

The raw source code for this example can be found on GitHub.