Skip to main content

Packages and composition

cloudposterior composes two independently usable Python packages. Its cp.cloud(model) context manager remains the only global PyMC interceptor.

Package Owns Public documentation
pymc-monitor Sampling engine, adaptive diagnostics, notebook/terminal displays and authenticated localhost dashboard Monitor API
pymc-cache Versioned model/data/kwargs fingerprints, policy-aware preflight, memory and disk storage Cache API
cloudposterior Modal image/source delivery, volumes, progress transport, hosted dashboard, ntfy, model-map fanout and PyMC interception This API

The monitor never imports the cache, and neither package imports Modal or cloudposterior. Each is on PyPI and usable on its own (pip install pymc-monitor, pip install pymc-cache); pip install cloudposterior pulls in both, as pymc-monitor[notebook,dashboard]>=0.1,<0.2 and pymc-cache>=0.1,<0.2. Modal images install only pinned third-party dependencies and mount all three package sources and the monitor assets, so the remote worker runs exactly the code installed locally.

Local sampling uses the monitor engine with the captured original pm.sample, including until= and init="prior". A local dashboard=True binds an authenticated loopback-only server; remote dashboard URLs are Modal-hosted and support phone access. PyMC/nutpie provide live draws and early stop; JAX samplers provide phase-level progress only. PyMC 5 and 6 are supported on their respective Python versions. Diagnostic thresholds are heuristics rather than guarantees.