Given an Ersilia model repository, this package allows you to serve it using FastAPI.
pip install git+https://github.com/ersilia-os/ersilia-pack.gitYou can check that the model repository contains all the expected files as follows:
ersilia_model_lint --repo_path $REPO_PATHYou can pack models that either require conda install or only pip installs, or both:
ersilia_model_pack --repo_path $REPO_PATH --bundles_repo_path $BUNDLE_PATHIf there are conda dependencies within the provided install instructions, the command above will install them in the base conda environment. To install conda dependencies in a specific conda environment, modify the above command as follows:
ersilia_model_pack --repo_path $REPO_PATH --bundles_repo_path $BUNDLE_PATH --conda_env_name $CONDA_ENVThis will create a folder with the packed model and the app available. The app can be served as follows:
ersilia_model_serve --bundle_path $BUNDLE_PATH --port $PORTThe FastAPI APIs are divided into two blocks: Metadata and App. Generally, Metadata includes GET requests, and App includes POST requests.
- GET /card: Retrieves metadata about the model, such as its name, title, and description.
- GET /card/model_id: Gets the model identifier.
- POST /run: Takes input data and returns predictions from the model.
Outputs are always provided in JSON format. We use the orient syntax from Pandas, including:
records: Each row is a dictionary.split: Dictionary containing separate lists forindex,columns, anddata.columns: Dictionary where each key is a column name, and the value is a list of column values.index: Dictionary where each key is an index value, and the value is a dictionary of row data.values: List of lists with row data.
- Currently, only
"Compound"and"Single"inputs are accepted. If more input types are needed, they should be included insidetemplates/input_schemas. - Ensure to follow best practices for FastAPI development and maintain proper documentation and tests for any new features added to the package.