1.6

What's Changed

• Add adjoint-differentiated Laplace HMC (blackjax.laplace_hmc) by @junpenglao in #856
• Code quality cleanup: bugs, naming, typos, and modernization by @junpenglao in #857
• Add missing docstrings and document magic numbers by @junpenglao in #858
• Code quality: protocol tests, deferred cleanups, and namespace encapsulation by @junpenglao in #859
• Refactor lambda-assigned run_steps to named function in mclmc_adaptation by @junpenglao in #860
• Split util.py into cohesive modules (Phase 7.1) by @junpenglao in #861
• Refactor mclmc build_kernel to standard pattern (Phase 7.2) by @junpenglao in #862
• Simplify random_walk.py indirection (Phase 7.3) by @junpenglao in #863
• Add build_sampling_algorithm helper to reduce boilerplate (7.7) by @junpenglao in #864
• Refactor adjusted_mclmc_dynamic build_kernel to standard pattern (7.8) by @junpenglao in #865
• Deduplicate adjusted_mclmc shared code (7.9) by @junpenglao in #867
• Fix incorrect citation: Sountsov → Seyboldt et al. by @junpenglao in #868
• Deduplicate rescale() between dynamic_hmc and adjusted_mclmc (7.10) by @junpenglao in #869
• Replace premature ravel_pytree with jax.tree.map (7.11/7.12) by @junpenglao in #870
• Modernize jax.lax.cond: remove legacy operand=None (4.5) by @junpenglao in #872
• docs: fix broken notebooks, replace Aesara with PyTensor, add Design Principles by @junpenglao in #873
• Add guidelines for new algorithms (docs and skeletons) by @albcab in #485
• Move Design Principles to DEVELOPER DOCUMENTATION section by @junpenglao in #874
• Fix flaky TestLaplaceAdjointAnalytical: cap Poisson observation range by @junpenglao in #875
• Feature/multinomal hmc sampler by @Idg772 in #854
• Add laplace_multinomial_hmc and dynamic_multinomial_hmc aliases by @junpenglao in #876
• Rename HMC aliases to short names (mhmc, dhmc, dmhmc, laplace_mhmc) by @junpenglao in #877
• Add laplace_dhmc and laplace_dmhmc samplers by @junpenglao in #878
• Update all exports for all public samplers and algorithms by @junpenglao in #879
• Unify integration_steps_fn to accept *integration_steps_params by @junpenglao in #880
• Add speed-up guide and horseshoe NUTS benchmark by @junpenglao in #883
• Fix JAX nightly CI silently swallowing test failures by @junpenglao in #884
• Fix missing dependencies in JAX nightly CI by @junpenglao in #885
• fix: refactor flaky SMC tuning test by @junpenglao in #887
• tests: relax atol in Laplace-HMC funnel test for robustness by @junpenglao in #888
• tests: improve SMC test stability by @junpenglao in #889
• Standardize JAX patterns (jax.tree) by @junpenglao in #891
• Add MCMC diagnostics guide by @junpenglao in #892
• Restructure and consolidate developer documentation by @junpenglao in #893
• Fix progress bar DeprecationWarning (JAX 0.10.0 compatibility) by @junpenglao in #895
• docs: add Funsor PPL integration tutorial (GMM discrete marginalisation) by @junpenglao in #896
• docs: fix funsor tutorial plot — replace az.plot_dist with matplotlib by @junpenglao in #898
• docs: improve Funsor tutorial (idiomatic API, NumPyro section, ArviZ fix) by @junpenglao in #899
• docs: fix label switching in Funsor GMM posterior plots by @junpenglao in #900
• docs: add algorithm index and standard inference loop (issue #890) by @junpenglao in #901
• docs: add Laplace-preconditioned HMC how-to guide (issue #890) by @junpenglao in #902
• build: migrate package management to uv by @junpenglao in #903
• docs: update CONTRIBUTING.md for uv dev workflow by @junpenglao in #904
• Minor update post uv update by @junpenglao in #905
• docs(mclmc): document 3-tuple return of mclmc_find_L_and_step_size by @junpenglao in #906
• docs(smc): correct adaptive_persistent_sampling step signature in docstring by @junpenglao in #908
• tests(vi): robustify test_elbo_decreases_over_steps against PRNG noise by @junpenglao in #909
• fix(smc): clamp natural parameters to >= 1 in pretuning by @junpenglao in #910
• fix(smc): correct sign in ess_solver to match tempered SMC weight update by @junpenglao in #915
• fix: make low_rank_window_adaptation vmap-compatible (#916) by @junpenglao in #917
• window_adaptation: add initial_inverse_mass_matrix kwarg by @junpenglao in #918
• tests: extend std_normal_logdensity with scale kwarg + drop orphan _fixtures.py by @junpenglao in #920
• pathfinder_adaptation: add num_chains + n_paths for multi-chain via multipathfinder dispatch by @junpenglao in #919
• window_adaptation: add imm_shrinkage_to_previous pseudo-count kwarg by @junpenglao in #921
• fix(pathfinder_adaptation): ravel pytree positions before PSIS mixture-covariance einsum by @junpenglao in #922
• rename low_rank_window_adaptation → window_adaptation_low_rank by @junpenglao in #923
• tests(vi): stabilize Schrödinger-Föllmer recover-posterior test under date-rotated seed by @junpenglao in #924
• Surface L-BFGS diagnostics in laplace kernel Info (#925) by @junpenglao in #926
• feat(laplace): laplace_lbfgs_grad_evals() utility for benchmark grad-count denominators by @junpenglao in #928
• fix(eca): replace keys_sampling.T with jnp.swapaxes to fix old-style PRNGKey shape corruption by @junpenglao in #931
• feat(smc): add batch_size for memory-bounded SMC execution via jax.lax.map by @thomasckng in #929
• test(smc): trim redundant non-divisor batch test to a map_fn/map_kernel unit test by @junpenglao in #933
• feat(mclmc): LRD preconditioned ESH integrator — O(dk) Low-Rank+Diagonal MCLMC by @junpenglao in #936
• feat(adaptation): mclmc_lrd_adaptation — Scheme A pilot-free LRD warmup by @junpenglao in #937
• feat(adaptation): gated √d warm-start for mclmc_lrd_warmup by @junpenglao in #938
• fix(adaptation): default avg=2 trajectory length for adjusted_mclmc_dynamic (escape MALA collapse) by @junpenglao in #939
• fix(adaptation): avg-preserving trajectory-length calibration for adjusted_mclmc_dynamic (fix high-d acceptance collapse) by @junpenglao in #940
• fix(tests): reduce MC error in Laplace flaky tests via larger N by @junpenglao in #942
• fix(adaptation): mclmc_lrd adjusted-path step/L mismatch — drop L_init re-pin (option ii) by @junpenglao in #941
• Add slice sampling kernel by @dirmeier in #943
• Slice integration by @yallup in #946
• Nested sampling by @yallup in #947
• fix(ns/tests): shrink heavy end-to-end test configs to fix CI OOM by @junpenglao in #948
• Fix low-rank Fisher mass-matrix adaptation: invert score covariance, fix regularization scale by @junpenglao in #949
• Generalize ghmc to accept dense / low-rank momentum metrics by @junpenglao in #950
• Slim ghmc backward-compat metric dispatch (behavior-preserving) by @junpenglao in #953
• Add nutpie-style schedule option to low-rank window adaptation (proportional window sizing + gradient-based init) by @junpenglao in #951
• Add GIST sampler family: self-tuning step size (autoStep-style) and trajectory length (no-U-turn) by @junpenglao in #952
• MEADS-LRD: opt-in low-rank momentum metric for meads_adaptation by @junpenglao in #954
• ChEES-HMC: opt-in ensemble diagonal metric + metric-consistent criterion + slow-direction trajectory-length floor by @junpenglao in #955
• Reduce MC flakiness in GIST trajectory-length funnel-neck test by @junpenglao in #957
• Harden low-rank Fisher metric: scale-relative PD guard + f64 promotion by @junpenglao in #956
• Add opt-in nutpie-style continuous warmup schedule for window_adaptation_low_rank by @junpenglao in #958
• Make GIST funnel-neck moment tests robust across JAX versions (self-calibrating ESS-gated tolerance) by @junpenglao in #959
• fix(eca): pass args through shard_map instead of closing over it by @Sumu004 in #961
• feat(ci): dedicated multi-device test job + eca.py cleanup by @junpenglao in #962
• fix(deps): bump 5 vulnerable transitive deps (idna/msgpack/starlette/tornado/urllib3) by @junpenglao in #963
• Replace progress-bar mechanism with a scan-instrumenting context manager by @junpenglao in #964
• Add howto: show progress during sampling (progress_bar context manager) by @junpenglao in #965
• docs: verified-frontends line in the progress_bar howto by @junpenglao in #966
• Auto-provision >=2 devices for multi-device tests via conftest by @junpenglao in #967
• Skip flaky GIST funnel-neck marginal test (tracking: #970) by @junpenglao in #971
• Add jax-tap to the progress extra by @junpenglao in #968
• Re-platform progress_bar emission onto jax-tap (M1) by @junpenglao in #972
• Fix MCLMC NaN handling: truthful divergence flag, revived step-size shrink, no silent sampling corruption by @junpenglao in #974
• Expose MCLMC warmup divergence flags as tuning-scan ys (jaxtap-observable) + degenerate-warmup checking docs by @junpenglao in #975

New Contributors

• @Idg772 made their first contribution in #854
• @dirmeier made their first contribution in #943
• @yallup made their first contribution in #946
• @Sumu004 made their first contribution in #961

Full Changelog: 1.5...1.6