rel-3.0.3

What's Changed

New Features 🎉

• More rewrites for ExtractDiag by @jessegrabowski in #2045
• System for algebraic reasoning about linear alegbra by @jessegrabowski in #2032
• Add numba dispatch for expm by @jessegrabowski in #2148

Bugfixes 🐛

• Fix local_sqrt_sqr rewrite logic bug by @WHOIM1205 in #1952
• Do not duplicate outer inputs when inlining OFG by @ricardoV94 in #2147
• Fix fusion replacement ordering by @ricardoV94 in #2150

Maintenance 🔧

• Support single string dims in XTensorType by @williambdean in #2129
• Add @ (matmul) operator to XTensorVariable by @williambdean in #2128
• Make exceptions less verbose by default by @ricardoV94 in #1330
• Numba CAReduce: reorder loops based on strides by @ricardoV94 in #1961
• Stop subtensor merge bonanza by @ricardoV94 in #2098
• as_tensor_variable dispatch for range objects by @williambdean in #2141
• Normalize string dims across all xtensor paths by @williambdean in #2146
• Extend inv(A) @ b -> solve(A, b) to batched matmul by @ricardoV94 in #2157
• Add numba dispatch for SVD by @jessegrabowski in #2136

Full Changelog: rel-3.0.2...rel-3.0.3

rel-3.0.2

What's Changed

Bugfixes 🐛

• Fix Softmax handling of non-tensor inputs regression by @ricardoV94 in #2122

Full Changelog: rel-3.0.1...rel-3.0.2

rel-3.0.1

What's Changed

New Features 🎉

• Optimize solve of block_diag by @jessegrabowski in #2109
• Add custom jax jvp for solve_sylvester by @jessegrabowski in #2116

Maintenance 🔧

• Replace several Ops by OpFromGraph by @ricardoV94 in #2110
• Rewrite solve(matrix_inverse(X), b) → X @ b by @alessandrogentili001 in #2114
• Fix numba support of cython functions with ssize by @ricardoV94 in #2113
• Lazy scipy import by @ricardoV94 in #2112

Full Changelog: rel-3.0.0...rel-3.0.1

rel-3.0.0

What's Changed

Major Changes 🛠

• Make Numba the default linker by @ricardoV94 in #1862
• Remove test_value functionality by @ricardoV94 in #1972
• Reorganize compile module and remove deprecated functionality by @ricardoV94 in #1351
• Change user-facing random variable API by @ricardoV94 in #1707
• Cleanup root objects by @ricardoV94 in #1863
• Rename autodiff L_op and R_op methods to pull_back and push_forward by @ricardoV94 in #1983
• Reorganize linear algebra by @jessegrabowski in #2037
• XTensor module no longer experimental by @ricardoV94 in #2048
• Refactor advanced subtensor by @jaanerik in #1756

New Features 🎉

• Implement isfinite helper by @nirdesho6o in #1957
• Implement rich debugprint format via file="rich" by @williambdean in #2042
• Add xtensor.where as a method and root operation by @ricardoV94 in #1985

Bugfixes 🐛

• Make numba caching safe across internal name collisions by @ricardoV94 in #2051
• Python Scalar Pow: do not return complex values for fractional powers of negative floats by @vybhav72954 in #1945
• Fix scan_save_mem with 0 steps by @ricardoV94 in #1977
• Fix Scan higher order derivatives by @ricardoV94 in #1975
• Scan: Fix untraced_sit_sot when n_steps=0 by @ricardoV94 in #2001
• Fix issues with RNGs in Scan by @ricardoV94 in #2039
• Add empty input guards to numba linalg by @jessegrabowski in #2057
• Make input output_grads to OFG pullback function consistent at runtime by @jessegrabowski in #2065
• Fix error in local_reshape_to_dimshuffle by @ricardoV94 in #2107
• Improve Feature hook API and pickle round-trip by @jessegrabowski in #2099

Documentation 📖

• make scan dosctring render legibly by @elc45 in #1897
• Add examples to function and Function.copy docstrings by @alessandrogentili001 in #2046
• docs(tutorial): change stale alias at -> pt by @williambdean in #2068

Maintenance 🔧

• Einsum: Compatibility with newer numpy by @ricardoV94 in #1916
• Graph replace: Handle redundant replacements by @ricardoV94 in #1942
• Allow freezing of FunctionGraph for hashing by @jessegrabowski in #1908
• Validate that output_subset indices are within bounds in CVM by @WHOIM1205 in #1903
• Rewrites for consecutive advanced read-write operations by @ricardoV94 in #2061
• Better static shape inference in linalg.kron by @ayulockedin in #1898
• Remove complex objectmode in linalg Ops by @jessegrabowski in #2062
• Add pullback for chosolve by @jessegrabowski in #2066
• Update Eigh and Eigvalsh Ops by @jessegrabowski in #2059
• Do not include constant branches in minimization arguments by @ricardoV94 in #1959
• Inline constants and merge duplicate inputs in OptimizeOp by @ricardoV94 in #1934
• Remove Iv core Op in favor of Ive and add rewrite rule for log of it by @alessandrogentili001 in #1929
• Rewrite a / exp(b) -> a * exp(-b) by @jessegrabowski in #1965
• Scan: remove unused outputs by @ricardoV94 in #2067
• Scan: Inplace on untraced_sit_sot (like RNGs) by @ricardoV94 in #2003
• Convert unit buffer sit sot to untraced sit sot by @ricardoV94 in #2005
• Canonicalize JoinDims and SplitDims to simpler operations than reshape by @mengxingbw in #1847
• XTensorType: Add values_eq and values_eq_approx by @ricardoV94 in #1914
• Allow complex inputs to numba linalg functions by @jessegrabowski in #1974
• Numba: Add several remaining sparse ops by @tomicapretto in #1896
• Add MLX dispatch for OpFromGraph by @jessegrabowski in #2012
• Add MLX dispatch for IfElse by @jessegrabowski in #2013
• MLX Elementwise and scalar dispatches by @jessegrabowski in #2007
• Add MLX dispatch for BatchedDot by @jessegrabowski in #2009
• Add einsum dispatch for mlx by @jessegrabowski in #2017
• Add MLX dispatch for Pad by @jessegrabowski in #2019
• More MLX linalg by @jessegrabowski in #2011
• Facilitate graph pattern matching by @ricardoV94 in #1786
• Add ScalarOp mathematical props by @jessegrabowski in #2053
• Remove errant dprint in local_lift_transpose_through_dot by @jessegrabowski in #2036

New Contributors

• @jaanerik made their first contribution in #1756
• @elc45 made their first contribution in #1897
• @alessandrogentili001 made their first contribution in #1929
• @nirdesho6o made their first contribution in #1957
• @vybhav72954 made their first contribution in #1945
• @mengxingbw made their first contribution in #1847
• @github-actions[bot] made their first contribution in #2076

Full Changelog: rel-2.38.0...rel-3.0.0

rel-2.38.3

What's Changed

Bugfixes 🐛

• Fix optimize gradient bug with repeated inputs by @ricardoV94 in #1933
• Fix scalar optimize gradient bug with repeated inputs by @ricardoV94 in #1958
• Fix numba compilation crashes from functions cached in fork processes by @ricardoV94 in #1992
• Fix numba caching in multiprocess contexts by @ricardoV94 in #2079

Maintenance 🔧

• Handle vectorization of Alloc nodes by @ricardoV94 in #1943

Full Changelog: rel-2.38.2...rel-2.38.3

rel-2.38.2

What's Changed

Maintenance 🔧

• Improve static shape of shaped xtensor Ops by @ricardoV94 in #1937

Full Changelog: rel-2.38.1...rel-2.38.2

rel-2.38.1

What's Changed

Bugfixes 🐛

• Validate that output_subset indices are within bounds in CVM by @WHOIM1205 in #1903

Documentation 📖

• make scan dosctring render legibly by @elc45 in #1897

Maintenance 🔧

• Fix static shape inference in pt.linalg.kron and add regression test by @ayulockedin in #1898
• Implement vectorize for xtensor Dot by @ricardoV94 in #1915
• XTensorType: Add values_eq and values_eq_approx by @ricardoV94 in #1914
• Einsum: Compatibility with newer numpy by @ricardoV94 in #1916
• Numba: Add several remaining sparse ops by @tomicapretto in #1896

New Contributors

• @jaanerik made their first contribution in #1756
• @ayulockedin made their first contribution in #1898
• @elc45 made their first contribution in #1897

Full Changelog: rel-2.38.0...rel-2.38.1

rel-2.38.0

What's Changed

New Features 🎉

• Add QZ Op and support for Numba solve_discrete_are by @jessegrabowski in #1859
• Implement xtensor.signal.convolve1d by @ricardoV94 in #1868
• Implement StructuredDotGradCSR and StructuredDotGradCSC in numba backend by @tomicapretto in #1860
• Implement sparse HStack and VStack in numba backend by @tomicapretto in #1889
• Implement shared XTensorVariables by @ricardoV94 in #1882
• Implement vectorize_graph for XTensor Ops by @ricardoV94 in #1876

Bugfixes 🐛

• Abort local_block_diag_dot_to_dot_block_diag when n_inputs < 2 by @jessegrabowski in #1865
• Fix xtensor broadcast dtypes by @ricardoV94 in #1877
• Fix xtensor discrete reduction by @ricardoV94 in #1879
• Fix empty xtensor indexing by @ricardoV94 in #1880
• Respect late introduction of ShapeFeature by @ricardoV94 in #1881
• Fix incorrect static output shape in MatrixPinv for non-square inputs by @WHOIM1205 in #1885

Maintenance 🔧

• Numba sparse: optimize csc @ csc by @ricardoV94 in #1866
• Numba sparse dot: reduce number of conversions for non-csr inputs by @tomicapretto in #1873
• Stabilize logdiffexp_to_log1mexpdiff -inf, -inf case by @ricardoV94 in #1883
• Allow infer_shape to work with xtensor lowering by @ricardoV94 in #1891

New Contributors

• @WHOIM1205 made their first contribution in #1885
• @arpanjha234 made their first contribution in #1890

Full Changelog: rel-2.37.0...rel-2.38.0

rel-2.37.0

What's Changed

Major Changes 🛠

• Do not raise in linalg Ops by @ricardoV94 in #1834

New Features 🎉

• Add rewrite to fuse nested BlockDiag Ops by @eby0303 in #1671
• Fix gradient of minimize and root wrt higher-dimensional arguments by @jessegrabowski in #1806
• Determinant of factorized matrices by @ricardoV94 in #1785
• Improved Linear Control Ops, with Numba Dispatches by @jessegrabowski in #1840
• Basic Sparse functionality in Numba by @ricardoV94 in #1676
• Implement sparse dot product in numba backend by @tomicapretto in #1854

Bugfixes 🐛

• Do not coerce gradients to TensorVariable by @ricardoV94 in #1685
• Fix bug in local_block_diag_dot_to_dot_block_diag by @ricardoV94 in #1839
• Fix issues with split and split_dims by @ricardoV94 in #1828
• Tweaks to reshape Ops by @ricardoV94 in #1842
• Fix scalar Blockwise in numba backend by @ricardoV94 in #1846

Maintenance 🔧

• local_[div|mul]_switch_sink: Allow zero inside a DimShuffle/Alloc by @ricardoV94 in #1850

New Contributors

• @eby0303 made their first contribution in #1671
• @tomicapretto made their first contribution in #1854

Full Changelog: rel-2.36.3...rel-2.37.0

rel-2.36.3

What's Changed

Major Changes 🛠

• Remove deprecated AbstractConv Ops and tests by @jessegrabowski in #1817

New Features 🎉

• Implement L_Op for join_dims and split_dims by @jessegrabowski in #1812

Bugfixes 🐛

• Fix local_subtensor_of_squeeze when squeezing multiple dimensions by @jessegrabowski in #1819
• Fix unboxing of numpy boolean scalars in C-backend by @ricardoV94 in #1790
• Exclude numba rewrites from JAX Scan rewrites by @ricardoV94 in #1825
• JAX Scan: Fix bug when recurring trace matches final requested size by @ricardoV94 in #1826
• Fix BlockDiag failure to compile in numba backend by @ricardoV94 in #1830
• Numba DimShuffle: special case for 0d input by @ricardoV94 in #1832

Maintenance 🔧

• Allow caching of cython functions in numba backend by @jessegrabowski in #1807
• Eager check for invalid ndim in TensorType by @ricardoV94 in #1813
• Implement tri symbolically by @ricardoV94 in #1814
• Cache tridiagonal solve in numba mode by @jessegrabowski in #1816
• Switch to new Scan API in jacobian by @ricardoV94 in #1824
• Rewrite log1mexp(log1mexp(x)) to x by @maresb in #1831

Full Changelog: rel-2.36.2...rel-2.36.3