diff --git a/onnxruntime/python/tools/symbolic_shape_infer.py b/onnxruntime/python/tools/symbolic_shape_infer.py
index 31f3a3a2b3..e7b7074783 100755
--- a/onnxruntime/python/tools/symbolic_shape_infer.py
+++ b/onnxruntime/python/tools/symbolic_shape_infer.py
@@ -241,6 +241,7 @@ class SymbolicShapeInference:
             "upsample_nearest1d": self._infer_aten_upsample,
             "upsample_nearest2d": self._infer_aten_upsample,
             "upsample_nearest3d": self._infer_aten_upsample,
+            "upsample_bicubic2d": self._infer_aten_upsample,
         }
         self.run_ = True
         self.suggested_merge_ = {}
diff --git a/orttraining/orttraining/python/training/ortmodule/_custom_gradient_registry.py b/orttraining/orttraining/python/training/ortmodule/_custom_gradient_registry.py
index 7731724272..4883075112 100644
--- a/orttraining/orttraining/python/training/ortmodule/_custom_gradient_registry.py
+++ b/orttraining/orttraining/python/training/ortmodule/_custom_gradient_registry.py
@@ -241,7 +241,7 @@ def native_group_norm_gradient():
 # are available for all versions, though they are not that convienent to use.
 def _upsample_gradient(backward_fn, dims):
     scales = ["" for _ in range(dims)]
-    if "bilinear" in backward_fn:
+    if "bicubic" in backward_fn:
         scales = ["I(2)", *scales]
     return [
         ("Shape", ["I(0)"], ["Shape_X"]),
@@ -271,3 +271,8 @@ def upsample_nearest2d_gradient():
 @register_gradient("org.pytorch.aten", "ATen", "upsample_nearest3d", "vec")
 def upsample_nearest3d_gradient():
     return _upsample_gradient("upsample_nearest3d_backward", 3)
+
+
+@register_gradient("org.pytorch.aten", "ATen", "upsample_bicubic2d", "vec")
+def upsample_bicubic2d_gradient():
+    return _upsample_gradient("upsample_bicubic2d_backward", 2)
diff --git a/orttraining/orttraining/python/training/ortmodule/_custom_op_symbolic_registry.py b/orttraining/orttraining/python/training/ortmodule/_custom_op_symbolic_registry.py
index 99e8851b6a..9288027f01 100644
--- a/orttraining/orttraining/python/training/ortmodule/_custom_op_symbolic_registry.py
+++ b/orttraining/orttraining/python/training/ortmodule/_custom_op_symbolic_registry.py
@@ -808,3 +808,16 @@ def upsample_nearest2d(g, input, output_size, scale_factors):
 @register_symbolic("upsample_nearest3d")
 def upsample_nearest3d(g, input, output_size, scale_factors):
     return _upsample_nearest(g, input, output_size, scale_factors, "upsample_nearest3d")
+
+
+@register_symbolic("upsample_bicubic2d")
+def upsample_bicubic2d(g, input, output_size, align_corners, scale_factors):
+    return g.op(
+        "org.pytorch.aten::ATen",
+        input,
+        output_size,
+        align_corners,
+        scale_factors,
+        operator_s="upsample_bicubic2d",
+        overload_name_s="vec",
+    )
diff --git a/orttraining/orttraining/test/python/orttraining_test_ortmodule_api.py b/orttraining/orttraining/test/python/orttraining_test_ortmodule_api.py
index 938d33cc9a..6a6832e063 100644
--- a/orttraining/orttraining/test/python/orttraining_test_ortmodule_api.py
+++ b/orttraining/orttraining/test/python/orttraining_test_ortmodule_api.py
@@ -1805,6 +1805,34 @@ def test_resize_grad_correctness_bilinear_2d(interpolate_size_scale, align_corne
     _test_helpers.assert_values_are_close(ort_input.grad, pt_input.grad)
 
 
+def test_aten_upsample_bicubic():
+    class _NeuralNetUpsampleBicubic(torch.nn.Module):
+        def __init__(self):
+            super().__init__()
+
+        def forward(self, input):
+            return torch.nn.functional.interpolate(input, size=(8, 12), mode="bicubic")
+
+    device = "cuda"
+    pt_model = _NeuralNetUpsampleBicubic().to(device)
+    ort_model = ORTModule(copy.deepcopy(pt_model))
+
+    def run_step(model, input):
+        prediction = model(input)
+        prediction.sum().backward()
+        return prediction
+
+    # reset manual seed to reset the generator
+    torch.manual_seed(2333)
+    pt_input = torch.randn([2, 4, 6, 8], dtype=torch.float, device=device, requires_grad=True)
+    ort_input = copy.deepcopy(pt_input)
+    pt_prediction = run_step(pt_model, pt_input)
+    ort_prediction = run_step(ort_model, ort_input)
+
+    _test_helpers.assert_values_are_close(ort_prediction, pt_prediction)
+    _test_helpers.assert_values_are_close(ort_input.grad, pt_input.grad)
+
+
 def test_gradient_correctness_cast_chain():
     class NeuralNetCast(torch.nn.Module):
         def __init__(self, D):