register ATen fallback for instance_norm2d gradient

orttraining/orttraining/python/training/ortmodule/_custom_gradient_registry.py

@@ -114,6 +114,18 @@ def diagonal_gradient():
     ]
 
 
+@register_gradient("org.pytorch.aten", "ATen", "instance_norm2d", "")
+def instance_norm2d_gradient():
+    return [
+        (
+            ("ATen", "org.pytorch.aten"),
+            ["GO(0)", "I(0)", "I(1)", "I(2)", "I(3)", "I(4)", "I(5)", "O(1)"],
+            ["GI(0)"],
+            {"operator": {"value": "instance_norm2d_backward", "dtype": "string"}},
+        ),
+    ]
+
+
 @register_gradient("org.pytorch.aten", "ATen", "max_pool2d_with_indices", "")
 def max_pool2d_gradient():
     return [
@@ -125,7 +137,6 @@ def max_pool2d_gradient():
         ),
     ]
 
-
 def minmax_gradient():
     # Gradient of torch.min(input) (and max)
     # In PyTorch, when there are multiple maxima/minima, the gradient is evenly distributed among them.