Fix einsum failing with repeated inputs (#1260)

pytensor/tensor/einsum.py

@@ -417,6 +417,18 @@ def _right_to_left_path(n: int) -> tuple[tuple[int, int], ...]:
     return tuple(pairwise(reversed(range(n))))
 
 
+def _ensure_not_equal(elements):
+    """
+    Ensures that any pair in a list of elements are not the same object. If a pair of elements is found to be equal, then one of them is converted to a copy.
+    """
+    elements = list(elements)
+    for i, elem1 in enumerate(elements[:-1]):
+        for j, elem2 in enumerate(elements[i + 1 :], start=i + 1):
+            if elem1 is elem2:
+                elements[j] = elem1.copy()
+    return elements
+
+
 def einsum(subscripts: str, *operands: "TensorLike", optimize=None) -> TensorVariable:
     """
     Multiplication and summation of tensors using the Einstein summation convention.
@@ -553,7 +565,7 @@ def einsum(subscripts: str, *operands: "TensorLike", optimize=None) -> TensorVar
             "If you need this functionality open an issue in https://github.com/pymc-devs/pytensor/issues to let us know. "
         )
 
-    tensor_operands = [as_tensor(operand) for operand in operands]
+    tensor_operands = _ensure_not_equal([as_tensor(operand) for operand in operands])
     shapes = [operand.type.shape for operand in tensor_operands]
 
     path: PATH

tests/tensor/test_einsum.py

@@ -8,6 +8,7 @@ import pytensor
 from pytensor import Mode, config, function
 from pytensor.graph import FunctionGraph
 from pytensor.graph.op import HasInnerGraph
+from pytensor.tensor import matrix
 from pytensor.tensor.basic import moveaxis
 from pytensor.tensor.blockwise import Blockwise
 from pytensor.tensor.einsum import _delta, _general_dot, _iota, einsum
@@ -281,3 +282,15 @@ def test_threeway_mul(static_length):
         out.eval({x: x_test, y: y_test, z: z_test}),
         np.full((3,), fill_value=6),
     )
+
+
+def test_repeated_inputs():
+    x = matrix("x")
+    out_repeated = einsum("ij,ij->i", x, x)
+    out_copy = einsum("ij,ij->i", x, x.copy())
+
+    x_test = np.array([[1, 2], [3, 4]]).astype(x.dtype)
+
+    np.testing.assert_allclose(
+        out_repeated.eval({x: x_test}), out_copy.eval({x: x_test})
+    )