Eig impls: Keep complex outputs in Blockwise and Numba

pytensor/link/numba/dispatch/nlinalg.py

@@ -81,9 +81,11 @@ def numba_funcify_Eig(op, node, **kwargs):
 
     @numba_basic.numba_njit
     def eig(x):
-        return np.linalg.eig(inputs_cast(x))
+        w, v = np.linalg.eig(inputs_cast(x))
+        return w.astype(w_dtype), v.astype(w_dtype)
 
-    return eig
+    cache_version = 1
+    return eig, cache_version
 
 
 @register_funcify_default_op_cache_key(Eigh)

pytensor/tensor/nlinalg.py

@@ -325,7 +325,8 @@ class Eig(Op):
     """
 
     __props__: tuple[str, ...] = ()
-    gufunc_spec = ("numpy.linalg.eig", 1, 2)
+    # Can't use numpy directly in Blockwise, because of the dynamic dtype
+    # gufunc_spec = ("numpy.linalg.eig", 1, 2)
     gufunc_signature = "(m,m)->(m),(m,m)"
 
     def make_node(self, x):

tests/tensor/test_blockwise.py

@@ -24,7 +24,7 @@ from pytensor.tensor import (
     vector,
 )
 from pytensor.tensor.blockwise import Blockwise, vectorize_node_fallback
-from pytensor.tensor.nlinalg import MatrixInverse
+from pytensor.tensor.nlinalg import MatrixInverse, eig
 from pytensor.tensor.rewriting.blas import specialize_matmul_to_batched_dot
 from pytensor.tensor.signal import convolve1d
 from pytensor.tensor.slinalg import (
@@ -763,3 +763,22 @@ def test_partial_inplace():
     add_supervisor_to_fgraph(fgraph, [In(inp, mutable=True) for inp in fgraph.inputs])
     rewrite_graph(fgraph, include=("inplace",))
     assert fgraph.outputs[0].owner.op.destroy_map == {1: [1]}
+
+
+def test_eig_blockwise():
+    x = tensor("x", shape=(2, 3, 3), dtype="float64")
+    eigen_values, eigen_vectors = eig(x)
+    assert eigen_values.dtype == "complex128"
+    assert eigen_vectors.dtype == "complex128"
+    fn = function([x], [eigen_values, eigen_vectors])
+    eigen_values_res, eigen_vectors_res = fn(np.full((2, 3, 3), np.eye(3)))
+    np.testing.assert_allclose(
+        eigen_values_res,
+        np.ones((2, 3), dtype="complex128"),
+        strict=True,
+    )
+    np.testing.assert_allclose(
+        eigen_vectors_res,
+        np.full((2, 3, 3), np.eye(3), dtype="complex128"),
+        strict=True,
+    )