Fix incorrect static output shape in MatrixPinv for non-square inputs

[WHOIM1205]

Fix static output shape in MatrixPinv for non-square inputs

Summary

MatrixPinv.make_node was incorrectly assigning the output static shape to match the input (m, n), even though the pseudoinverse returns (n, m).

Although gufunc_signature and infer_shape were correct, the static TensorType.shape created in make_node was not transposed. This caused incorrect shape reporting when input dimensions were known.

Fix

Updated pytensor/tensor/nlinalg.py:

python

Before

matrix(shape=x.type.shape, dtype=out_dtype)

After

matrix(shape=(x.type.shape[1], x.type.shape[0]), dtype=out_dtype)

• all the test cases are passed locally

[WHOIM1205]

hey @ricardoV94
This fixes a static shape mismatch in MatrixPinv.make_node for non-square inputs. Runtime behavior was correct, but the symbolic shape wasn’t transposed, which could break gradients. The change is minimal and all related tests pass.

[WHOIM1205]

pre-commit.ci autofix

[ricardoV94]

Interesting, I didn't know that

[ricardoV94]

Thanks @WHOIM1205

[ricardoV94]

@WHOIM1205 Do you mind me asking how did you find this issue?

[WHOIM1205]

@WHOIM1205 Do you mind me asking how did you find this issue?

Thanks 🙂

I was experimenting with pinv on non square matrices and noticed that the reported type.shape didn’t match the expected transposed output.
That felt odd, so I followed it through the Op definition and saw that make_node was using the input shape directly.
From there it was pretty clear that the static shape just needed to be swapped to stay consistent with the signature and infer_shape.

[jessegrabowski]

Great work!