bayesmith.dispatch.sampling

Source-derived signatures and contracts. See the reference index for coverage and conventions.

Module contract

One MCMC state across fixed-budget or diagnostic-checkpoint collection.

SamplingControl prepare_initial_params checkpoint_diagnostics run_mcmc

SamplingControl

Defined in bayesmith.dispatch.sampling, line 22.

class SamplingControl
initialization: InitializationPolicy | None = None
stopping: StoppingPolicy = field(default_factory=StoppingPolicy)
max_seconds: float | None = None
initial_values: tuple[NamedArray, ...] = ()
details: dict = field(default_factory=dict)
checkpoints: list = field(default_factory=list)

Runtime attachment: inputs are artifact policies; outputs are run facts.

SamplingControl.record

record(self)

prepare_initial_params

Defined in bayesmith.dispatch.sampling, line 48.

prepare_initial_params(mcmc, graph, key, control, gibbs_names=())

Encode validated model coordinates for NumPyro's mixed state contract.

checkpoint_diagnostics

Defined in bayesmith.dispatch.sampling, line 124.

checkpoint_diagnostics(grouped, divergences, policy)

Every coordinate is monitored; missing/nonfinite evidence cannot pass.

run_mcmc

Defined in bayesmith.dispatch.sampling, line 172.

run_mcmc(mcmc, key, graph, control, gibbs_names=())

Collect complete batches and resume the full warmed state, once per batch.

The draw cap is per chain. Time limits are checked at batch boundaries,
including the first warmup/production batch; they cannot interrupt XLA.