samlhuillier / learning-to-banana

🍌 Banana Serverless

This repo gives a framework to serve ML models in production using simple HTTP servers.

To generalize this to deploy anything on Banana, see the guide here.

1. Download

Choose 1 of 3 options

• Fork this repo to create a public repo
• Click Use this Template, which creates a private or public repo
• Create your own repo and copy the template files from here this repo into yours

Then clone the repo to a machine with a GPU to install + test it locally

2. Install

The repo default runs a HuggingFace BERT model.

1. Run pip3 install -r requirements.txt to download dependencies.
2. Run python3 server.py to start the server.
3. Run python3 test.py in a different terminal session to test an inference against it.

3. Make it your own

1. Edit app.py to load and run your model.
2. Make sure to test with test.py!
3. When ready to deploy:

• edit download.py (or the Dockerfile itself) with scripts download your custom model weights at build time
• edit requirements.txt with your pip packages. Don't delete the "sanic" line, as it's a banana dependency.

4. Ship to prod

You now have a functioning http server that should work using Docker and complete inferences on a GPU

To deploy on banana

• Log in to the Banana App
• Connect your github
• Select this repository

It'll then be built from the dockerfile, optimized, then deployed on our Serverless GPU cluster! You can then call it with any of our SDKs

• Python
• Node JS / Typescript
• Go

Use Banana for scale.