rel-2.28.0

What's Changed

Major Changes 🛠

• Support Python 3.13 by @maresb in #1199
• Make PyTensor compatible with numpy 2.0 by @brendan-m-murphy in #1194
  • Due to changes in the Numpy C-API AdvancedIncSubtensor1 no longer has a C-implementation when paired with numpy>=2.0. This operator shows up in the gradient graph of advanced indexing operations, which could now be slower when using the C backend.
• Raise explicitly on Python methods that are incompatible with lazy variables by @ricardoV94 in #1190

New Features 🎉

• Speedup implementation of multivariate_normal and allow method of covariance decomposition by @ricardoV94 in #1203
• Compute pushforward (Rop) via double application of pullback (Lop) and fix Scan and Max gradient bugs by @ricardoV94 in #1207
• Add LAPACK overloads for all variants of solve in Numba backend by @jessegrabowski in #1146
• Implement multinomial and inverse gamma in numba backend by @ricardoV94 in #1223
• Implement gradient for vector repetitions by @ricardoV94 in #1192

Bugfixes 🐛

• Rewrite reshapes that only expand or squeeze dims by @ricardoV94 in #1200
• Fix Blockwise and RandomVariable in Numba with repeated arguments by @ricardoV94 in #1222

Maintenance 🔧

• Simplify implementation of tile by @ricardoV94 in #1187
• Remove accidental print statements by @ricardoV94 in #1193
• PyTorch inline constants in dispatch to avoid graph breaks by @ricardoV94 in #1118
• Speedup scalar Op python implementations by @ricardoV94 in #1169
• Only do reshapes in tensordot when needed by @ricardoV94 in #1202

New Contributors

• @brendan-m-murphy made their first contribution in #1194

Full Changelog: rel-2.27.1...rel-2.28.0