Migration after the package extraction
The public cp.cloud, cp.sample, cp.map, cp.cleanup_cache, cp.cleanup_volumes and cp.check_convergence names remain. Existing paired notebook examples do not need edits. The two new standalone packages are optional when accessing cloudposterior’s API: cloudposterior’s declared requirements include them.
Cache identity and old files
pymc-cache uses a versioned canonical graph/data/kwargs/environment fingerprint. Previous cloudposterior cache keys are intentionally invalidated because some representations could miss changes in a model or its data. Existing .cloudposterior/ files are preserved, not silently read or deleted. New disk entries have full SHA-256 suffixes and full-key metadata. This means the first call after upgrade recomputes the posterior; subsequent equivalent calls can hit the new entry. A custom callable, step method or model operation without a safely representable identity bypasses automatic caching rather than risking a false hit.
cache=True reuses an unseeded result even with random_seed=None. Use cache=False to skip caching altogether or overwrite=True to run again and replace an eligible entry. Only completed fixed-length runs and adaptive runs with met targets are reusable. Manually stopped, futile, capped-but-unmet, failed and indeterminate adaptive results are returned to the caller but not stored for reuse. This is not a sampler checkpoint/resume mechanism.
cp.cleanup_cache(base_dir=...) removes verified new-format entries and matching legacy .nc/.key pairs under the selected cloudposterior root. It preserves the root, unrelated files and symlinks. It does not remove Modal volumes; cp.cleanup_volumes(project=...) is separate.
Local monitoring and report attributes
cp.cloud(model, until="draft") now checks targets and stops locally with the same monitor engine as the remote worker. dashboard=True with remote=False opens an authenticated localhost dashboard rather than warning that it has no effect. Local return_inferencedata=False remains a native PyMC passthrough, disables enhanced monitoring/caching for that call, and cannot be combined with adaptive stopping.
The canonical report is stored in posterior.attrs["pymc_monitor_convergence"]; cloudposterior also writes the historical cloudposterior_convergence alias. Superchain membership similarly has canonical pymc_monitor_superchains and legacy cloudposterior_superchains aliases. Use cp.check_convergence(idata) to assess any InferenceData independently. For precise interpretation, see the monitor statistical policy and cache policy.
The supported floor is PyMC 5.28 and PyTensor 2.38 (with Python >=3.11). PyMC 6 requires Python >=3.12. Remote numpyro/blackjax still warn and ignore until= for compatibility; standalone pymc-monitor rejects unsupported adaptive requests. Native callback= is still unavailable across Modal; it works locally through the captured original sampler.