Getting started

Install from a pinned commit, while the package is not yet on PyPI:

pip install 'pymc-cache @ git+https://github.com/justmytwospence/pymc-cache.git@<commit-sha>'

import pymc as pm
import pymc_cache as pc

with pm.Model() as model:
    mu = pm.Normal("mu", 0, 1)
    options = dict(model=model, draws=10, tune=10, chains=1,
                   cores=1, random_seed=12, cache="disk", progressbar=False)
    trace = pc.sample(**options)
    again = pc.sample(**options)

cache=True uses an eight-entry in-memory LRU. cache="disk" uses .pymc-cache/; a custom path selects a private cache directory. cache=False neither hashes the model nor reads/writes cache data. overwrite=True skips reading and atomically replaces a matching entry. Use pc.sample_smc(model=model, ...) for a separate SMC identity.

For remote transports that must avoid provisioning on a cache hit:

session = pc.CacheSession(cache="disk", model=model)
lookup = session.lookup({"draws": 10, "tune": 10, "chains": 1, "cores": 1,
                         "random_seed": 12, "progressbar": False})
if lookup.hit:
    trace = lookup.result
else:
    trace = pm.sample(model=model, **lookup.normalized_kwargs)
    session.store(lookup, trace, status="complete")

A runner integrated with pc.sample(..., sample_fn=runner) can define __pymc_cache_normalize__(model, call_kwargs) returning a dict with exactly call_kwargs, key_kwargs, sampler_identity, and versions. The caller receives call_kwargs; the key uses key_kwargs and a sampler-specific identity. A trusted caller without a normalization hook needs a unique sampler_key. For manual adaptive results, store() only accepts an attested met convergence report; stopped/capped/futile results are misses.