Fix mixed dtype bug in gammaincc_grad

pytensor/scalar/math.py

@@ -854,7 +854,7 @@ def gammaincc_grad(k, x, skip_loops=constant(False, dtype="bool")):
         log_s = np.array(0.0, dtype=dtype)
         s_sign = np.array(1, dtype="int8")
         n = np.array(1, dtype="int32")
-        log_delta = log_s - 2 * log(k)
+        log_delta = log_s - 2 * log(k).astype(dtype)
 
         def inner_loop_b(sum_b, log_s, s_sign, log_delta, n, k, log_x):
             delta = exp(log_delta)

tests/scalar/test_math.py

+import itertools
+
 import numpy as np
+import pytest
 import scipy.special as sp
 
 import pytensor.tensor as at
 from pytensor import function
 from pytensor.compile.mode import Mode
+from pytensor.graph import ancestors
 from pytensor.graph.fg import FunctionGraph
 from pytensor.link.c.basic import CLinker
+from pytensor.scalar import ScalarLoop, float32, float64, int32
 from pytensor.scalar.math import (
     betainc,
     betainc_grad,
@@ -13,6 +18,7 @@ from pytensor.scalar.math import (
     gammaincc,
     gammal,
     gammau,
+    hyp2f1,
 )
 from tests.link.test_link import make_function
 
@@ -89,3 +95,32 @@ def test_betainc_derivative_nan():
     assert np.isnan(test_func(1, 1, 2))
     assert np.isnan(test_func(1, -1, 1))
     assert np.isnan(test_func(1, 1, -1))
+
+
+@pytest.mark.parametrize(
+    "op, scalar_loop_grads",
+    [
+        (gammainc, [0]),
+        (gammaincc, [0]),
+        (betainc, [0, 1]),
+        (hyp2f1, [0, 1, 2]),
+    ],
+)
+def test_scalarloop_grad_mixed_dtypes(op, scalar_loop_grads):
+    nin = op.nin
+    for types in itertools.product((float32, float64, int32), repeat=nin):
+        inputs = [type() for type in types]
+        out = op(*inputs)
+        wrt = [
+            inp
+            for idx, inp in enumerate(inputs)
+            if idx in scalar_loop_grads and inp.type.dtype.startswith("float")
+        ]
+        if not wrt:
+            continue
+        # The ScalarLoop in the graph will fail if the input types are different from the updates
+        grad = at.grad(out, wrt=wrt)
+        assert any(
+            (var.owner and isinstance(var.owner.op, ScalarLoop))
+            for var in ancestors(grad)
+        )