rel-3.1.1

What's Changed

Major Changes 🛠

• Drop support for Python 3.11 by @ricardoV94 in #2211
• Rewrite ShapeFeature to not hold live variables by @ricardoV94 in #2056
• Remove symbolic axis from join/split by @jessegrabowski in #2144
• Deprecate implicit shared-variable capture in OpFromGraph by @alessandrogentili001 in #2047
• Make inner graph Ops immutable by @ricardoV94 in #2243

New Features 🎉

• Add asumption for unique indices by @ricardoV94 in #2225

Bugfixes 🐛

• Fix until scan gradient step count with recurring outputs by @jessegrabowski in #2217
• Fix assumption bug with (1,1) eye by @jessegrabowski in #2224
• Fix ScanMerge cycle by @ricardoV94 in #2226
• Fix numba blockwise layout inference by @ricardoV94 in #2234
• Fix MLX variadic Elemwise crash on broadcast-shaped inputs by @cetagostini in #2235
• Make exp/log rewrites dtype-safe under cast_policy="numpy+floatX" by @maresb in #2247
• Fix MLX IncSubtensor with slice index by @cetagostini in #2240
• Keep _make_scalar_loop updates dtype-stable by @maresb in #2249
• Accumulate duplicate indices in MLX AdvancedIncSubtensor by @cetagostini in #2242
• Numba scan: fix invalid cross alias and inplace by @ricardoV94 in #2263

Documentation 📖

• Expand documentation on broadcasting limitations by @ricardoV94 in #2084
• Tiny transformer LLM example by @cetagostini in #2163

Maintenance 🔧

• Numba: Cache cython-special functions by @jessegrabowski in #2212
• Add MLX GEMV dispatch by @drbenvincent in #2214
• Add MLX GER dispatch by @drbenvincent in #2216
• Numba: Speedup RV eval by casting float parameters to float64 by @ricardoV94 in #2223
• Do not recompile identical OFG in same graph by @ricardoV94 in #2209
• Exclude general fusion from JAX and MLX modes by @ricardoV94 in #2236
• Fix local_log_div float64 precision loss for integer-literal operands by @maresb in #2245
• Fix MLX Shape_i output type by @gaoflow in #2230
• Refactor BLAS code by @jessegrabowski in #2049
• Generalize local_sum_prod_alloc by @ricardoV94 in #2198
• Fold variadic Add/Mul in JAX and PyTorch backends instead of stacking by @cetagostini in #2238
• Add gradient (pullback) to Min Op by @hermesdt in #2253
• Fix MLX Metal codegen crash on NaN scalar constants by @cetagostini in #2239
• Enforce AdvancedIncSubtensor1 runtime-broadcast check on MLX backend by @cetagostini in #2241
• Numba scan: Allow inplacing on discarded mit-mot by @ricardoV94 in #2266

New Contributors

• @drbenvincent made their first contribution in #2214
• @gaoflow made their first contribution in #2230
• @hermesdt made their first contribution in #2253

Full Changelog: rel-3.0.5...rel-3.1.0