Model Versioning

Every artifact. Every lineage edge. Immutable.

Git tracks code. Inferpathio tracks model weights, feature schemas, hyperparameters, and training data hashes in a directed acyclic graph you can traverse in any direction.

Start versioning for free SDK reference

Abstract visualization of a model version directed acyclic graph with branching paths and glowing violet nodes on dark background

Capabilities

What model versioning actually means

Not just storing model files — capturing every dependency, every decision, every path that led to the artifact.

Content-addressed storage

Model artifacts are identified by SHA-256 hash of their content. Two identical models produce one storage entry — deduplication is automatic.

Full lineage DAG

Every model version links back to the training run, dataset version, feature schema, and code commit hash that produced it.

One-call rollback

Rolling back means re-registering a previous version as the production candidate. The promotion event is logged. The old artifact is never deleted.

Stage management

Move versions through Staging, Shadow, Production, and Archived states with policy-enforced promotion gates. Prevent accidental production deployments.

SDK example

Register a version in three lines

The SDK instruments your training script at the run level. After training completes, call run.register() to commit the artifact to the registry with full provenance.

register_model.py

import inferpathio as ifp

with ifp.track("fraud-detection") as run:
    model.fit(X_train, y_train)
    run.log_metric("f1", 0.941)
    run.register(
        stage="staging",
        tags=["q4-retrain", "v3.2"]
    )

▶ Registered: fraud-detection@sha256:a4f9
▶ Stage: staging | Run: run-8832