Fix missing optional field of maximum_double_grad (#66308)

HydrogenSulfate · web-flow · commit 58a31336e9da · 2024-07-23T14:07:05.000+08:00
diff --git a/paddle/phi/ops/yaml/inconsistent/dygraph_backward.yaml b/paddle/phi/ops/yaml/inconsistent/dygraph_backward.yaml
@@ -382,6 +382,7 @@
   args: (Tensor x, Tensor y, Tensor grad_x_grad, Tensor grad_y_grad)
   output: Tensor(grad_out_grad)
   composite: maximum_double_grad(x, y, grad_x_grad, grad_y_grad, grad_out_grad)
+  optional : grad_x_grad, grad_y_grad
 
 - backward_op: minimum_double_grad
   forward: minimum_grad(Tensor x, Tensor y, Tensor grad_out) -> Tensor(grad_x), Tensor(grad_y)
diff --git a/test/prim/prim/vjp/test_comp_high_grad.py b/test/prim/prim/vjp/test_comp_high_grad.py
@@ -20,6 +20,7 @@
 import numpy as np
 import parameterized as param
 from decorator_helper import prog_scope
+from utils import dygraph_guard
 
 import paddle
 from paddle import base
@@ -801,5 +802,68 @@ def test_high_grad(self):
             self.func_double(p)
 
 
+@param.parameterized_class(
+    ('shape1', 'shape2'),
+    [
+        ([2], [2], True),
+        ([2, 3], [2, 3], True),
+        ([2, 3, 4], [2, 3, 4], True),
+        ([2, 3, 3, 4], [2, 3, 3, 4], True),
+    ],
+)
+class TestMaximumHighGradCheck2(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.shape1 = cls.shape1
+        cls.shape2 = cls.shape2
+
+    def _grad(self, y, x, order):
+        u = y
+        dx = paddle.ones_like(x)
+        for _ in range(order):
+            dx = paddle.grad(u, x, create_graph=True)[0]
+            u = dx
+        return dx
+
+    def func_double(self, place, x_stop, y_stop):
+        x = paddle.randn(self.shape1).astype("float32")
+        x.stop_gradient = x_stop
+        y = paddle.randn(self.shape2).astype("float32")
+        y.stop_gradient = y_stop
+
+        # wraping with tanh to enable high order gradient
+        z = paddle.maximum(paddle.tanh(x), paddle.tanh(y))
+
+        if not x.stop_gradient:
+            dzdx = self._grad(z, x, 2)
+        if not y.stop_gradient:
+            dzdy = self._grad(z, y, 2)
+
+    def func_triple(self, place, x_stop, y_stop):
+        x = paddle.randn(self.shape1).astype("float32")
+        x.stop_gradient = x_stop
+        y = paddle.randn(self.shape2).astype("float32")
+        y.stop_gradient = y_stop
+
+        # wraping with tanh to enable high order gradient
+        z = paddle.maximum(paddle.tanh(x), paddle.tanh(y))
+
+        if not x.stop_gradient:
+            dzdx = self._grad(z, x, 3)
+        if not y.stop_gradient:
+            dzdy = self._grad(z, y, 3)
+
+    def test_high_grad(self):
+        places = [base.CPUPlace()]
+        if core.is_compiled_with_cuda():
+            places.append(base.CUDAPlace(0))
+        for p in places:
+            for x_stop in [False, True]:
+                for y_stop in [False, True]:
+                    with dygraph_guard():
+                        self.func_double(p, x_stop, y_stop)
+                        self.func_triple(p, x_stop, y_stop)
+
+
 if __name__ == '__main__':
     unittest.main()
