v0.42.0

DynamicPPL v0.42.0

Diff since v0.41.8

LogDensityFunction now performs AD preparation through AbstractPPL's prepare / value_and_gradient!! interface instead of calling DifferentiationInterface directly. Internally this removes the _use_closure heuristic and the explicit DI.Constant plumbing; the choice between closure and constants now lives in AbstractPPL.

logdensity_at has been renamed to logdensity_internal. The old name is kept as a const alias so external callers do not break.

LogDensityAt is now a deprecation shim that emits a warning and returns an AbstractPPL.Evaluators.VectorEvaluator whose call forwards to logdensity_internal. New code should call AbstractPPL.prepare(logdensity_internal, x; context=...) directly.

Breaking changes

DifferentiationInterface is no longer a hard dependency of DynamicPPL. With AbstractPPL 0.15.2, the following backends now have native AbstractPPL extensions and only need the concrete AD package loaded:

• AutoForwardDiff — load ForwardDiff
• AutoMooncake, AutoMooncakeForward — load Mooncake

For other DI-routed backends like AutoReverseDiff, users must load DifferentiationInterface together with the concrete AD package:

using DynamicPPL, ADTypes, DifferentiationInterface, ReverseDiff
ldf = LogDensityFunction(model; adtype=AutoReverseDiff())

For distributed sampling the same packages must be loaded on every worker.

Compatibility bounds bumped:

• AbstractPPL 0.14 → 0.15
• Bijectors 0.15.17 → 0.16

The integration test suites for MarginalLogDensities, ReverseDiff, and Enzyme now live in their own environments under test/ext/DynamicPPL*Ext/ and run as separate CI jobs.

Merged pull requests:

• Use AbstractPPL AD interface (#1363) (@yebai)
• Benchmark improvements (#1396) (@yebai)
• Update HISTORY.md for 0.42.0 (#1408) (@shravanngoswamii)

Closed issues:

• Improve DynamicPPL benchmarks (#1374)