mlstac 0.4.8

A machine learning model-sharing specification based on STAC MLM and Safetensors.

MLSTAC

Machine learning model sharing, built on the STAC MLM extension and Safetensors.

---

Overview

MLSTAC gives you a single, consistent way to publish and consume machine learning models. Each model is described by a STAC Item using the MLM extension, and its weights travel as Safetensors. From that metadata you can load a model from many backends (HTTP, local disk, S3, Google Cloud Storage), download it, inspect it, and turn it into a usable PyTorch model.

Status: experimental. The API may change between minor versions.

Installation

pip install mlstac

Quickstart

import mlstac

# Load only the metadata first (no weights are downloaded yet)
model = mlstac.load("https://example.com/my-model/mlm.json")

# Inspect what you have
print(model.get_model_summary())
model.print_schema()

# Download every file into a local folder; you get back a loader
# pointing at the local copy
local = model.download("./my-model")

# Build a usable model from the local files
net = local.compiled_model()

Loading from different sources

mlstac.load is the main entry point. It accepts a URL, a local path, or a list of .pt2 files for an ad-hoc ensemble.

# Remote metadata
mlstac.load("https://example.com/model/mlm.json")

# A local directory that contains an mlm.json
mlstac.load("./my-model")

# An ensemble built directly from .pt2 files
ensemble = mlstac.load(["model_a.pt2", "model_b.pt2"])
ensemble.is_ensemble  # True

Downloading

You can download from the loader instance, or call the module-level function directly.

# From an existing loader (reuses the source it was created from)
local = model.download("./my-model")

# Or as a one-shot call
local = mlstac.download("https://example.com/model/mlm.json", "./my-model")

Supported backends

Scheme	Example
http / https / ftp	https://example.com/model/mlm.json
local	./my-model or /abs/path/mlm.json
s3	s3://bucket/key/mlm.json (needs boto3)
gs	gs://bucket/key/mlm.json (needs google-cloud-storage)

Documentation

Full docs live at tacofoundation.github.io/mlstac.

Development

# install with dev dependencies
poetry install --with dev

# run the test suite
poetry run pytest tests/ -v

License

See LICENSE.