Fix mistaken output types in aesara.tensor.nlinalg

aesara/tensor/nlinalg.py

@@ -270,7 +270,7 @@ class Eigh(Eig):
         # input.
         w_dtype = self._numop([[np.dtype(x.dtype).type()]])[0].dtype.name
         w = vector(dtype=w_dtype)
-        v = matrix(dtype=x.dtype)
+        v = matrix(dtype=w_dtype)
         return Apply(self, [x], [w, v])
 
     def perform(self, node, inputs, outputs):
@@ -407,15 +407,18 @@ class QRFull(Op):
 
     def make_node(self, x):
         x = as_tensor_variable(x)
+
         assert x.ndim == 2, "The input of qr function should be a matrix."
 
         in_dtype = x.type.numpy_dtype
         out_dtype = np.dtype(f"f{in_dtype.itemsize}")
 
+        q = matrix(dtype=out_dtype)
+
         if self.mode != "raw":
-            r = matrix(dtype=x.dtype)
+            r = matrix(dtype=out_dtype)
         else:
-            r = vector(dtype=x.dtype)
+            r = vector(dtype=out_dtype)
 
         if self.mode != "r":
             q = matrix(dtype=out_dtype)
@@ -507,10 +510,15 @@ class SVD(Op):
     def make_node(self, x):
         x = as_tensor_variable(x)
         assert x.ndim == 2, "The input of svd function should be a matrix."
-        s = vector(dtype=x.dtype)
+
+        in_dtype = x.type.numpy_dtype
+        out_dtype = np.dtype(f"f{in_dtype.itemsize}")
+
+        s = vector(dtype=out_dtype)
+
         if self.compute_uv:
-            u = matrix(dtype=x.dtype)
-            vt = matrix(dtype=x.dtype)
+            u = matrix(dtype=out_dtype)
+            vt = matrix(dtype=out_dtype)
             return Apply(self, [x], [u, s, vt])
         else:
             return Apply(self, [x], [s])