Apply casting in as_tensor_variable in normalize_size_param

This allows PyTensor to infer more broadcastable patterns, by placing the casting inside the MakeVector Op

Author: ricardoV94

pytensor/tensor/random/utils.py

@@ -134,7 +134,7 @@ def normalize_size_param(
             "Parameter size must be None, an integer, or a sequence with integers."
         )
     else:
-        size = cast(as_tensor_variable(size, ndim=1), "int64")
+        size = cast(as_tensor_variable(size, ndim=1, dtype="int64"), "int64")
 
         if not isinstance(size, Constant):
             # This should help ensure that the length of non-constant `size`s

tests/tensor/random/test_op.py

@@ -148,6 +148,9 @@ def test_RandomVariable_bcast():
     res = rv(0, 1, size=at.as_tensor(1, dtype=np.int64))
     assert res.broadcastable == (True,)
 
+    res = rv(0, 1, size=(at.as_tensor(1, dtype=np.int32), s3))
+    assert res.broadcastable == (True, False)
+
 
 def test_RandomVariable_bcast_specify_shape():
     rv = RandomVariable("normal", 0, [0, 0], config.floatX, inplace=True)