pymlup 0.3.1

Introduction

MLup is a library for easy and fast running of ML models in production.

All you need is to deliver the model file to the server (a config is optional) — pymlup turns it into a FastAPI web application with one CLI command. No web app code to write or maintain.

• Pure Python, no required framework-specific glue code;
• Uses FastAPI for the web layer;
• Works with any Python object that exposes a predict-like method, plus native (de)serialization support for scikit-learn, lightgbm, tensorflow, torch and onnx models.

Requirements

Python 3.8+ (3.12, 3.13, 3.14 supported; the tensorflow extra requires Python <3.14 until TensorFlow publishes 3.14 wheels).

Installation

pip install pymlup

With an ML backend extra:

pip install "pymlup[scikit-learn]"  # For scikit-learn
pip install "pymlup[lightgbm]"      # For microsoft lightgbm
pip install "pymlup[tensorflow]"    # For tensorflow
pip install "pymlup[torch]"         # For torch
pip install "pymlup[onnx]"          # For onnx models: torch, tensorflow, sklearn, etc...

Quick start

From a clean environment to a working prediction API in about five minutes, using scikit-learn as the example.

1. Create a virtual environment and install pymlup with the scikit-learn extra:

python3 -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install "pymlup[scikit-learn]"

2. Train and save a tiny model. Save this as train_model.py and run it:

# train_model.py
import pickle
from sklearn.tree import DecisionTreeClassifier

X = [[0, 0], [1, 1], [2, 2], [3, 3]]
y = [0, 0, 1, 1]

model = DecisionTreeClassifier().fit(X, y)

with open("model.pkl", "wb") as f:
    pickle.dump(model, f)

python train_model.py

3. Start the API:

mlup run -m ./model.pkl

4. Check it's alive (in another terminal):

curl http://localhost:8009/health

{"status":200}

5. Call the model:

curl -X POST http://localhost:8009/predict \
  -H "Content-Type: application/json" \
  -d '{"X": [[1, 1], [3, 3]]}'

{"predict_result":[0,1]}

6. Explore the interactive API docs. Open http://localhost:8009/docs in your browser for the Swagger UI, auto-generated from your model's predict signature — you can try /predict right there.

Stop the server with Ctrl+C.

Already have a serialized model (onnx, joblib, lightgbm, torch, tensorflow, ...) instead of training a new one? Point mlup run -m at it directly (install the matching extra from Installation) — see the full documentation for every supported format and config option.

Any Python object, no ML framework required

pymlup doesn't require a serialized model file at all — any Python object with a predict-like method works:

import mlup

class MyAnyModelForExample:
    def predict(self, X):
        return X

up = mlup.UP(ml_model=MyAnyModelForExample())
up.ml.load()
up.run_web_app(daemon=True)

Open http://localhost:8009/docs to try it, or call it from the same script — this needs pip install requests separately, it's not a pymlup dependency:

import requests
response = requests.post('http://localhost:8009/predict', json={'X': [[1, 2, 3], [4, 5, 6]]})
print(response.json())

up.stop_web_app()

Supported ML frameworks

Work tested with machine learning model frameworks (links to tests):

• scikit-learn>=1.2.0,<2.0.0
• tensorflow>=2.0.0,<3.0.0, Python<3.14
• lightgbm>=4.0.0,<5.0.0
• torch>=2.0.0,<3.0.0
• onnx>=1.0.0,<2.0.0
• onnxruntime>=1.14.0,<1.26.0

Support and tested with machine learning libraries:

• numpy>=1.0.0,<3.0.0
• pandas>=2.0.0,<3.0.0
• joblib>=1.2.0,<2.0.0

Documentation

The full documentation — Python API, CLI reference, config file format, storages, binarizers, data transformers, web app architectures/API, application life cycle — lives at mlup.org (source: docs/).

Useful links

• Full documentation
• Examples
• Test models

Metrics

MLup PyPi download statistics: https://pepy.tech/project/pymlup