v0.2.0

ParallelMCMC v0.2.0

Diff since v0.1.0

Breaking

• FlexiChains is now the default (and only built-in) chain type. sample(model, sampler, N; chain_type=...) returns a FlexiChains.FlexiChain instead of an MCMCChains.Chains. Use chain_type=SymChain for Symbol-keyed chains or chain_type=VNChain for VarName-keyed chains.
• DynamicPPL-backed models must use VNChain; requesting SymChain for a DynamicPPL model now throws an ArgumentError.
• The MCMCChains dependency has been removed i.e., there is no longer any MCMCChains output.

Changed

• param_names handling is more forgiving: user-supplied names are wrapped in FlexiChains.Parameter automatically, Symbol names are upgraded to VarNames when a VNChain is requested, and better default names are generated when none are supplied.
• DynamicPPL compat bumped to 0.41.6, 0.42.

Added

• FlexiChains dependency; DynamicPPLExt now also loads on FlexiChains.

Merged pull requests:

• Bump codecov/codecov-action from 6 to 7 (#33) (@dependabot[bot])
• docs: clarify Enzyme GPU gradient abort is a gc-transition bug, not b… (#43) (@rsenne)
• FlexiChains by default (#44) (@penelopeysm)
• Bump actions/checkout from 6 to 7 (#46) (@dependabot[bot])
• Use [sources] in docs (#47) (@penelopeysm)
• Update Project.toml (#50) (@rsenne)
• Document 0.2.0 in changelog (#51) (@rsenne)

Closed issues:

• Docs for v0.1.0 didn't build (#45)
• Add docs link in repo description (#48)

v0.1.0

ParallelMCMC v0.1.0

Diff since v0.0.1

ParallelMCMC v0.1.0

Diff since v0.0.1

Breaking changes

• backend is now a required keyword argument on ParallelMALASampler;
  DEER.DEFAULT_BACKEND / DEFAULT_HVP_BACKEND and the old default-Enzyme
  machinery were removed. To keep the old behaviour, load Enzyme and pass
  backend=AutoEnzyme(; mode=Enzyme.Forward, function_annotation=Enzyme.Duplicated).
• Enzyme is now an optional dependency loaded via EnzymeExt rather than a
  hard dependency.
• The DynamicPPL convenience constructor no longer populates param_names, and
  the heuristic prior-based name extraction was removed.

Merged pull requests:

• GPU fixes (#32) (@rsenne)
• Map DynamicPPL samples back to original space (#20) (@penelopeysm)
• Use DynamicPPL adtype path for Turing models (#16) (@rsenne)

Merged pull requests:

• Map DynamicPPL samples back to original space (#20) (@penelopeysm)
• Bump actions/checkout from 4 to 6 (#22) (@dependabot[bot])
• Bump actions/upload-artifact from 4 to 7 (#23) (@dependabot[bot])
• Bump julia-actions/setup-julia from 2 to 3 (#24) (@dependabot[bot])
• Simplify installation command for ParallelMCMC (#27) (@rsenne)
• docs: add rsenne as a contributor for code, maintenance, and 4 more (#30) (@allcontributors[bot])
• docs: add penelopeysm as a contributor for code, test, and 3 more (#31) (@allcontributors[bot])
• GPU fixes (#32) (@rsenne)
• docs: add gdalle as a contributor for review, and ideas (#34) (@allcontributors[bot])
• docs: add wsmoses as a contributor for review (#35) (@allcontributors[bot])
• Update changelog for PR #32, release as 0.1.0 (#42) (@rsenne)

Closed issues:

• Add more worked examples (#28)
• Remove default backend (Enzyme) and make DI the main UI (#29)

v0.0.1

ParallelMCMC v0.0.1

v0.0.1 (2026-05-02) — Initial Release

Features:

• DEER solver — Parallel-across-the-sequence MALA with Newton iteration. Solves an entire trajectory of $T$ correlated steps simultaneously via affine scan (prefix scan) with $O(\log T)$ work.
• ParallelMALASampler — Primary parallel sampler supporting stochastic diagonal Jacobian estimation via pushforward (Jacobian-vector products), with jacobian options for diagonal and dense modes.
• Sequential baselines — MALASampler and AdaptiveMALASampler for comparison / fallback use.
• GPU support — CUDA-accelerated DEER solver with GPU-compatible indexing and passing GPU tests.
• DynamicPPL / Turing.jl integration — Extension module ext/DynamicPPLExt provides interoperability with Turing.jl via the LogDensityProblemsAD backend.
• Multi-AD-backend support — Uses DifferentiationInterface and Enzyme for autodiff, with support for multiple backends including pullbacks and pushforwards.
• Preconditioning — Support for preconditioned MALA proposals.
• AbstractMCMC interface — All samplers implement AbstractMCMC and return MCMCChains.Chains objects.
• Benchmarking suite — benchmarks/ module for performance evaluation.
• BlockMALA sampler — Blocked version of the MALA sampler included.
  Dependencies
  Requires Julia >= 1.10, AbstractMCMC, CUDA, DifferentiationInterface, Enzyme, and MCMCChains, with optional DynamicPPL / LogDensityProblems / LogDensityProblemsAD for Turing integration.