Do the same fix for all opt path and for prod.

theano/tensor/opt.py

@@ -5210,11 +5210,12 @@ def local_opt_alloc(node):
                     val = get_scalar_constant_value(input)
                     assert val.size == 1
                     # check which type of op
+                    casted = T.mul(*shapes).astype(str(val.dtype))
                     if isinstance(node.op, T.Sum):
-                        val = val.reshape(1)[0] * T.mul(*shapes)
+                        val = val.reshape(1)[0] * casted
                     else:
-                        val = val.reshape(1)[0] ** T.mul(*shapes)
-                    return [T.cast(val, dtype=node.outputs[0].dtype)]
+                        val = val.reshape(1)[0] ** casted
+                    return [val]
 
                 except NotScalarConstantError:
                     pass
@@ -5226,11 +5227,12 @@ def local_opt_alloc(node):
                     to_prod = [shapes[i] for i in xrange(len(shapes))
                                if i in node.op.axis]
                     if to_prod:
+                        casted = T.mul(*to_prod).astype(str(val.dtype))
                         if isinstance(node.op, T.Sum):
-                            val *= T.mul(*to_prod).astype(str(val.dtype))
+                            val *= casted
                         else:
-                            val = val ** T.mul(*to_prod)
-                    return [T.alloc(T.cast(val, dtype=node.outputs[0].dtype),
+                            val = val ** casted
+                    return [T.alloc(val,
                                     *[shapes[i] for i in xrange(len(shapes))
                                       if i not in node.op.axis])]
                 except NotScalarConstantError:

theano/tensor/tests/test_opt.py

@@ -5271,8 +5271,8 @@ class T_local_sum_prod(unittest.TestCase):
         assert numpy.allclose(f(input), input.sum())
         assert len(f.maker.fgraph.apply_nodes) == 1
 
-
     def test_local_sum_prod_alloc(self):
+        # test local_opt_alloc
         a = T.dtensor3()
         input = numpy.asarray(numpy.arange(2 * 3 * 4).reshape(2, 3, 4),
                               dtype='float64')
@@ -5376,6 +5376,30 @@ class T_local_sum_prod(unittest.TestCase):
             config.on_opt_error = backup
 
 
+class T_local_opt_alloc(unittest.TestCase):
+    def test_sum_upcast(self):
+        s = theano.tensor.lscalar()
+        a = theano.tensor.alloc(numpy.asarray(5, dtype='float32'), s, s)
+        orig = theano.config.warn_float64
+        theano.config.warn_float64 = "raise"
+        try:
+            f = theano.function([s], a.sum())
+            f(5)
+        finally:
+            theano.config.warn_float64 = orig
+
+    def test_prod_upcast(self):
+        s = theano.tensor.lscalar()
+        a = theano.tensor.alloc(numpy.asarray(5, dtype='float32'), s, s)
+        orig = theano.config.warn_float64
+        theano.config.warn_float64 = "raise"
+        try:
+            f = theano.function([s], a.prod())
+            f(5)
+        finally:
+            theano.config.warn_float64 = orig
+
+
 class T_local_reduce(unittest.TestCase):
     def setUp(self):
         self.mode = theano.compile.get_default_mode().including(

theano/tests/test_raise_on_float64.py

-from __future__ import absolute_import, print_function, division
-
-__author__ = "deepworx"
-__contact__ = "deepworx@GitHub"
-
-import unittest
-import theano, numpy
-
-class RaiseOnFloat64(unittest.TestCase):
-    s = theano.tensor.lscalar()
-    a = theano.tensor.alloc(numpy.asarray(5, dtype='float32'), s, s)
-    orig = theano.config.warn_float64
-    theano.config.warn_float64 = "raise"
-    try:
-        f = theano.function([s], a.sum())
-        theano.printing.debugprint(f)
-        f(5)
-    finally:
-        theano.config.warn_float64 = orig
\ No newline at end of file