v5.22.0

What's Changed

Major Changes 🛠

• Bump PyTensor dependency by @ricardoV94 in #7740
  • This release changes how JAX dispatch of random variables are implemented. Custom dispatches will have to be updated

Documentation 📖

• Add return type overload for sample_posterior_predictive by @nataziel in #7710
• Replaced "log-likelihood" with "distribution" in for all the classes where distributions is more appropriate rather than log-likelihood. by @tanishy7777 in #7573

Maintenance 🔧

• Allow passing graph level attributes to graphviz by @zaxtax in #7741
• Add warning if a Minibatched variable is used without total_size by @zaxtax in #7742

New Contributors

• @tanishy7777 made their first contribution in #7573

Full Changelog: v5.21.1...v5.22.0

v5.21.2

Important

This was an erroneous duplicate release of v5.22.0.

v5.21.1

What's Changed

New Features 🎉

• Support PyTensor deterministic operations as observations by @wd60622 in #7656
• Allow draws from Weibull, MvStudentT, LKJCorr and LKJCholeskyCovRV in alternative backends by @ricardoV94 in #7685

Maintenance 🔧

• Make do interventions shared variables by default by @ricardoV94 in #7596
• Remove setupegg.py by @Armavica in #7697
• Expand logging test cases for sample_prior_predictive and add return type overloads by @nataziel in #7707
• Allow compile_kwargs in sample_smc by @jessegrabowski in #7702

Full Changelog: v5.21.0...v5.21.1

v5.21.0

What's Changed

Major Changes 🛠

• Bump PyTensor and support numpy>2.0 and Python 3.13 by @ricardoV94 in #7688
  • Note: Checking if variable:, equivalent to bool(variable) now raises for PyMC variables. Use if variable is not None or whatever is appropriate in your context.
• Remove deprecated generator data by @wd60622 in #7664

Bugfixes 🪲

• Fix bug with chained CustomSymbolicDists by @ricardoV94 in #7690
• Fix bug when reusing jax logp for initial point generation by @ricardoV94 in #7695

Maintenance 🔧

• Relax observe to allow observing already observed variables by @zaxtax in #7679
• Reuse jaxified logp when sampling via jax by @nataziel in #7681

Full Changelog: v5.20.1...v5.21.0

v5.20.1

What's Changed

New Features 🎉

• Add ZarrTrace by @lucianopaz in #7540

Bugfixes 🪲

• fix: deep copy nuts_sampler_kwarg to prevent .pop side effects by @inclinedadarsh in #7652

Documentation 📖

• Adds shape/rate info to Gamma docs by @Dpananos in #7625
• Probability distributions guide update by @fonnesbeck in #7671

Maintenance 🔧

• Foward compile_kwargs to ADVI when init = "advi+..." by @jessegrabowski in #7640
• Check for observed variables in the trace by @zaxtax in #7641
• Use jaxified logp for initial point evaluation when sampling via Jax by @nataziel in #7610
• Show one progress bar per chain when sampling by @jessegrabowski in #7634
• Ignore inner unused RNG inputs in collect_default_updates by @ricardoV94 in #7658

New Contributors

• @nataziel made their first contribution in #7610
• @inclinedadarsh made their first contribution in #7652

Full Changelog: v5.20.0...v5.20.1

v5.20.0

What's Changed

Major Changes 🛠

• Rename compile_pymc to compile by @ricardoV94 in #7606

Bugfixes 🪲

• Fix MCMC non-deterministic seeding with Generators by @ricardoV94 in #7637

Documentation 📖

• Fix error in docstring example for MatrixNormal by @roesta07 in #7599

New Contributors

• @roesta07 made their first contribution in #7599

Full Changelog: v5.19.1...v5.19.2

v5.19.1

What's Changed

Bugfixes 🪲

• Bump numpy version due to use of Generator.spawn only available in >=1.25 by @ricardoV94 in #7607

Maintenance 🔧

• Exponential scale default to 1.0 by @ricardoV94 in #7604

Full Changelog: v5.19.0...v5.19.1

v5.19.0

What's Changed

Major Changes 🛠

• Speedup sample and allow specifying compile_kwargs by @ricardoV94 in #7578. This introduces several major changes related to step samplers:
  • internal uses of logp_dlogp_function now work with raveled inputs. External use will issue a warning unless ravel_inputs is specified explicitly. Eventually it will only be possible to use ravel_inputs=True.
  • Step samplers arguments besides vars must be passed by keyword
  • RaveledVars.point_map_info is now a 4-n tuple, with size introduced.
  • assign_step_method does not call instantiate_steppers, but returns arguments needed for the latter.
  • Allow passing compile_kwargs to sample which is then forwarded to the step samplers functions

Bugfixes 🪲

• Fix error in find_measurable_bitwise by @ricardoV94 in #7585

Documentation 📖

• Explain difference between BinaryMetropolis and BinaryGibbsMetropolis by @ricardoV94 in #7586
• Add example on freeze_data_and_dims by @ricardoV94 in #7594

Maintenance 🔧

• Register the overloads added by CustomDist so it works with multiprocess with SMC by @EliasRas in #7241

New Contributors

• @EliasRas made their first contribution in #7241

Full Changelog: v5.18.2...v5.19.0

v5.18.2

What's Changed

Maintenance 🔧

• Allow interdependent initial points from same OpFromGraph node by @ricardoV94 in #7569

Full Changelog: v5.18.1...v5.18.2

v5.18.1

What's Changed

Bugfixes 🪲

• Harmonize HSGP.prior dimension names and order by @juanitorduz in #7562
• Fix bug in implicit_size_from_params by @ricardoV94 in #7564
• Do not mutate Scan inner graph when deriving logprob by @ricardoV94 in #7575

Maintenance 🔧

• Make pytensorf.constant_fold unconditional by @ricardoV94 in #7568

Full Changelog: v5.18.0...v5.18.1