Using model weights in Tigris anywhere with SkyPilot

The most common way to deploy AI models in production is by using “serverless” inference. This means that every time you get a request, you don’t know what state the underlying hardware is in. You don’t know if you have your models cached, and in the worst case you need to do a cold start and download your model weights from scratch.

A couple fixable problems arise when running your models on serverless or any frequently changing infrastructure:

• Model distribution that's not optimized for latency causes needless GPU idle time as the model weights are downloaded to the machine on cold start. Tigris behaves like a content delivery network by default and is designed for low latency, saving idle time on cold start.
• Compliance restrictions like data sovereignty and GDPR increase complexity quickly. Tigris makes regional restrictions a one-line configuration, guide here.
• Reliance on third party caches for distributing models creates an upstream dependency and leaves your system vulnerable to downtime. Tigris guarantees 99.99% availability with public availability data.

SkyPilot

SkyPilot is a tool that lets you route GPU compute to the cheapest possible locale based on your requirements. The same configuration lets you control AWS, Azure, Google Cloud, Oracle Cloud, Kubernetes, Runpod, Fluidstack, or more. For more information about Skypilot, check out their documentation.

To get started, you'll need to install SkyPilot following their directions. Be sure to have Conda installed.

You will need to configure your cloud of choice for this example. See SkyPilot's documentation on how to do this. We have tested this against a few clouds:

• AWS
• Lambda
• Runpod

However the other providers should work fine.

Usecase

You can put AI model weights into Tigris so that they are cached and fast no matter where you’re inferencing from. This allows you to have cold starts be faster and you can take advantage of Tigris' globally distributed architecture, enabling your workloads to start quickly no matter where they are in the world.

For this example, we’ll set up SDXL Lightning by ByteDance for inference with the weights stored in Tigris.

Getting Started

Download the sdxl-in-tigris template from GitHub:

git clone https://github.com/tigrisdata-community/sdxl-in-tigris

Prerequisite tools

In order to run this example locally, you need these tools installed:

• Python 3.11
• pipenv
• The AWS CLI

Also be sure to configure the AWS CLI for use with Tigris: Configuring the AWS CLI.

To build a custom variant of the image, you need these tools installed:

• Mac/Windows: Docker Desktop app, alternatives such as Podman Desktop will not work.
• Linux: Docker daemon, alternatives such as Podman will not work.
• Replicate's cog tool
• jq

To install all of the tool depedencies at once, clone the template repo and run brew bundle.

Create a new bucket for generated images, it’ll be called generated-images in this article.

aws s3 create-bucket --acl private generated-images

Optional: upload your own model

If you want to upload your own models, create a bucket for this. It'll be referred to as model-storage-demo in this tutorial.

Both of these buckets should be private.

Then activate the virtual environment with pipenv shell and install the dependencies for uploading a model:

pipenv shell --python 3.11
pip install -r requirements.txt

Run the prepare_model script to massage and upload a Stable Diffusion XL model or finetune to Tigris:

python scripts/prepare_model.py ByteDance/SDXL-Lightning model-storage

info

Want differently styled images? Try finetunes like Kohaku XL! Pass the Hugging Face repo name to the prepare_model script like this:

python scripts/prepare_model.py KBlueLeaf/Kohaku-XL-Zeta model-storage

Access keys

Create a new access key in the Tigris Dashboard. Don't assign any permissions to it.

Copy the access key ID and secret access keys into either your notes or a password manager, you will not be able to see them again. These credentials will be used later to deploy your app in the cloud. This keypair will be referred to as the workload-keypair in this tutorial.

Limit the scope of this access key to only the model-storage-demo (or a custom bucket if you're uploading your own models) and generated-images buckets.

Customizing the skypilot.yaml file

Open skypilot.yaml in your favorite text editor. Customize the environment variables in the envs: key:

envs:
  # Tigris config
  AWS_ACCESS_KEY_ID: tid_AzureDiamond # workload access key ID
  AWS_SECRET_ACCESS_KEY: tsec_hunter2 # workload secret access key
  AWS_ENDPOINT_URL_S3: https://fly.storage.tigris.dev
  AWS_REGION: auto

  # Bucket names
  MODEL_BUCKET_NAME: model-storage-demo
  PUBLIC_BUCKET_NAME: generated-images

  # Model to load
  MODEL_PATH: ByteDance/SDXL-Lightning

Envvar name	Value
AWS_ACCESS_KEY_ID	The access key ID from the workload keypair
AWS_SECRET_ACCESS_KEY	The secret access key from the workload keypair
AWS_ENDPOINT_URL_S3	https://fly.storage.tigris.dev
AWS_REGION	auto
MODEL_PATH	ByteDance/SDXL-Lightning
MODEL_BUCKET_NAME	model-storage-demo (Optional: replace with your own bucket name)
PUBLIC_BUCKET_NAME	generated-images (replace with your own bucket name)

Launching it in a cloud

Run sky serve up to start the image in a cloud:

sky serve up skypilot.yaml -n sdxl

Wait a few minutes for everything to converge, and then you can use the endpoint URL to poke it:

⚙︎ Service registered.

Service name: sdxl
Endpoint URL: 3.84.60.169:30001

note

You can run sky serve status to find out if your endpoint is ready:

$ sky serve status
<...>
Service Replicas
SERVICE_NAME  ID  VERSION  ENDPOINT                  LAUNCHED     RESOURCES                   STATUS  REGION
sdxl          1   1        http://69.30.85.69:22112  47 secs ago  1x RunPod({'RTXA4000': 1})  READY   CA

Finally, run a test generation with this curl command:

curl "http://ip:port/predictions/$(uuidgen)" \
  -X PUT \
  -H "Content-Type: application/json" \
  --data-binary '{
    "input": {
        "prompt": "The space needle in Seattle, best quality, masterpiece",
        "aspect_ratio": "1:1",
        "guidance_scale": 3.5,
        "num_inference_steps": 4,
        "max_sequence_length": 512,
        "output_format": "png",
        "num_outputs": 1
    }
}'

If all goes well, you should get an image like this:

You can destroy the machine with this command:

sky serve down sdxl