Model packaging

High-level information about the packaging of models using Ersilia Pack

To package models in Ersilia, there are two essential components, namely the Ersilia model template, and the Ersilia Pack tool. Ersilia Pack creates a FastAPI app based on the information available in a given model repository, based on the template.

Ersilia Pack is a lightweight, plug-and-play framework to serve AI models from the Ersilia Model Hub via an API.

It allows developers to:

• Package models with their dependencies

• Run them as web services

• Submit data and retrieve predictions

• Monitor jobs and system performance

• Access detailed model metadata

Detailed documentation on Ersilia Pack is available in the corresponding GitHub repository. Here, we simply provide a high-level overview of the tool.

Typical Workflow

1. Lint a model repository to check it has the correct structure

2. Package the model into a self-contained bundle

3. Serve the model via API locally or in the cloud

4. Interact with the model through documented endpoints

Typical Ersilia Pack Workflow Test

• Each run lints, packs, serves, and validates an Ersilia model.

• Endpoints checked:

  • Sync → /run

  • Async → /job/submit, /job/status/{id}, /job/result/{id}

• Ensures: server starts, health passes, responses are valid JSON (contents are valid and match the predefined example output results of the corresponding model), jobs complete successfully.

Python / Model Matrix

Python	Model
3.8	eos43at
3.9	eos6oli
3.10	eos3b5e
3.11	eos5guo
3.12	eos5cl7

In summary: 5 Ersilia models are smoke-tested (sync + async) across Python 3.8–3.12.