Serving a Stable Diffusion Model with Ray Serve

Try it out

Run this example in the Anyscale Console or view it on GitHub.

⏱️ Time to complete: 5 min (15 on GCP)

This template shows you how to:

1. Develop and run a Ray Serve application running the SDXL diffusion model.
2. Send test requests to the application running locally.
3. Deploy the application to production as a service.
4. Send requests to the application running in production as a service.

Step 1: Install python dependencies

The application requires a few extra Python dependencies. Install them using pip and they'll be saved in the workspace and picked up when deploying to production.

!pip install -q diffusers==0.25.0 transformers==4.36.2 accelerate==0.25.0 && echo 'Install complete!'

Step 2: Run the model locally

• Run the command below in a VSCode terminal (Ctrl-`).
• The model will be available at http://localhost:8000.
• The command will block and print logs for the application.

# Run the following in a VSCode terminal because it's a blocking command.
$ serve run main:stable_diffusion_app

Step 3: Send a test request to the model running locally

The generate_image function sends an HTTP request to the model and saves the response as a local image.

As the request to generate the image runs, logs will be printed to the terminal that is running serve run.

import requests

HOST = "http://localhost:8000"

def generate_image(prompt: str, image_size: int) -> bytes:
    response: requests.Response = requests.get(
        f"{HOST}/imagine",
        params={"prompt": prompt, "img_size": image_size},
    )
    response.raise_for_status()
    return response.content

image = generate_image("anyscale logo valentines day card, professional quality art, surrounded by flowers, white envelope", 640)

filename = "image.png"
with open(filename, "wb") as f:
    f.write(image)

from IPython.display import Image
Image(filename=filename)

Step 4: Deploy the model to production as a service

Deploy the model to production using the anyscale service deploy command.

This creates a long-running service with a stable endpoint to query the application.

Local files and dependencies installed in the workspace are automatically included when the service is deployed.

!anyscale service deploy --name stable_diffusion_service main:stable_diffusion_app

Step 5: Send a test request to the model running in the service

Query the service using the same logic as when testing it locally, with two changes:

1. Update the HOST to the service endpoint.
2. Add the authorization token as a header in the HTTP request.

Both of these values are printed when you run anyscale service deploy. You can also find them on the service page. For example, if the output looks like:

(anyscale +4.0s) You can query the service endpoint using the curl request below:
(anyscale +4.0s) curl -H 'Authorization: Bearer 26hTWi2kZwEz0Tdi1_CKRep4NLXbuuaSTDb3WMXK9DM' https://stable_diffusion_app-4rq8m.cld-ltw6mi8dxaebc3yf.s.anyscaleuserdata-staging.com

Then:

• The authorization token is 26hTWi2kZwEz0Tdi1_CKRep4NLXbuuaSTDb3WMXK9DM.
• The service endpoint is https://stable_diffusion_app-4rq8m.cld-ltw6mi8dxaebc3yf.s.anyscaleuserdata-staging.com.

import requests

HOST = "TODO_INSERT_YOUR_SERVICE_HOST"
TOKEN = "TODO_INSERT_YOUR_SERVICE_TOKEN"

def generate_image(prompt: str, image_size: int) -> bytes:
    response: requests.Response = requests.get(
        f"{HOST}/imagine",
        params={"prompt": prompt, "img_size": image_size},
        headers={
            "Authorization": f"Bearer {TOKEN}",
        },
    )
    response.raise_for_status()
    return response.content

image = generate_image("twin peaks sf in basquiat painting style", 640)

filename = "image.png"
with open(filename, "wb") as f:
    f.write(image)

from IPython.display import Image
Image(filename=filename)

Summary

This notebook:

• Developed and ran a model serving application locally.
• Sent a test request to the application locally.
• Deployed the application to production as a service.
• Sent another test request to the application running as a service.