Skip to content

MLup logo

MLup

MLup turns Python objects and serialized machine learning models into self-hosted FastAPI REST APIs without writing serving boilerplate.

Serve sklearn, PyTorch, TensorFlow, ONNX, and custom Python models through a generated FastAPI application with validation, OpenAPI docs, and health endpoints.

Linters and testing PyPI version Downloads

MLup demo: mlup run, curl /predict, Swagger UI

Quick start

pip install "pymlup[scikit-learn]"

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

model = DecisionTreeClassifier().fit([[0, 0], [1, 1], [2, 2], [3, 3]], [0, 0, 1, 1])
with open("model.pkl", "wb") as f:
    pickle.dump(model, f)

python model.py
mlup run -m model.pkl

curl -X POST http://localhost:8009/predict \
  -H "Content-Type: application/json" \
  -d '{"X": [[1, 1], [3, 3]]}'
# {"predict_result":[0,1]}

Interactive API docs: http://localhost:8009/docs. Full walkthrough and every option: Quickstart.

What MLup provides

  • A generated FastAPI application — up.web.app is a real fastapi.FastAPI instance;
  • POST /predict, with request validation built from your model's signature or a column config;
  • Auto-generated OpenAPI schema and Swagger UI at /docs;
  • GET /health and GET /info;
  • Loading from serialized models (pickle, joblib, and framework-native formats) or from a plain Python object with a predict-like method — pip install pymlup alone, no extra, covers the latter;
  • Optional worker_and_queue and batching execution modes for the prediction call.

Supported frameworks

Type / framework Typical formats Installation extra
Any Python object in-memory object with a predict-like method — no file needed (none, core install)
scikit-learn pickle / joblib — via mlup's generic binarizers, not a dedicated adapter pymlup[scikit-learn]
LightGBM native LightGBM format, pickle, joblib pymlup[lightgbm]
PyTorch native torch formats (including TorchScript), pickle pymlup[torch]
TensorFlow SavedModel, .h5, .keras, pickle — Python <3.14 only pymlup[tensorflow]
ONNX .onnx pymlup[onnx]

See Binarizers for how model-format detection works.

When to use MLup

  • You already have a trained scikit-learn, LightGBM, PyTorch, TensorFlow, or ONNX model, or a plain Python object with a predict-like method;
  • You need an internal API or a small, self-hosted model service;
  • You're moving a prototype out of a notebook and just need it reachable over HTTP;
  • You don't need a full MLOps platform for this — just a clean serving layer.

When MLup probably isn't the right fit

  • Kubernetes-native autoscaling;
  • Highly optimized multi-GPU inference;
  • Distributed serving graphs;
  • Managed cloud deployment;
  • A full MLOps lifecycle / model registry;
  • A complex, application-specific HTTP API with substantial custom logic beyond serving a model.

Where to go next

Downloads

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

Downloads Downloads Downloads