diff --git a/src/ATen/native/xpu/sycl/GridSampler.cpp b/src/ATen/native/xpu/sycl/GridSampler.cpp
index 4e4c094b3f9..b31154c85bf 100644
--- a/src/ATen/native/xpu/sycl/GridSampler.cpp
+++ b/src/ATen/native/xpu/sycl/GridSampler.cpp
@@ -682,8 +682,7 @@ void grid_sampler_2d_backward_kernel(
     int64_t padding_mode,
     bool align_corners,
     std::array<bool, 2> output_mask) {
-  check_grid_sampler_common(input, grid);
-  check_grid_sampler_2d(input, grid);
+  check_grid_sampler_2d_backward(input, grid, grad_output);
 
   globalContext().alertNotDeterministic("grid_sampler_2d_backward_xpu");
   auto N = input.size(0);
@@ -1491,8 +1490,13 @@ void grid_sampler_3d_backward_kernel(
     std::array<bool, 2> output_mask) {
   // See NOTE [ grid_sampler Native Functions ].
   // Add checks here in case this is called instead of grid_sampler.
-  check_grid_sampler_common(input, grid);
-  check_grid_sampler_3d(input, grid, interpolation_mode);
+  check_grid_sampler_3d_backward(input, grid, grad_output);
+  TORCH_CHECK_NOT_IMPLEMENTED(
+      static_cast<GridSamplerInterpolation>(interpolation_mode) !=
+          GridSamplerInterpolation::Bicubic,
+      "grid_sampler(): bicubic interpolation with 5D input is not "
+      "implemented for ",
+      input.device().type());
 
   globalContext().alertNotDeterministic("grid_sampler_3d_backward_xpu");
   auto input_requires_grad = output_mask[0];
@@ -1506,7 +1510,7 @@ void grid_sampler_3d_backward_kernel(
         at::ScalarType::BFloat16,
         at::ScalarType::Half,
         input.scalar_type(),
-        "grid_sampler_2d_backward_xpu",
+        "grid_sampler_3d_backward_xpu",
         [&] {
           if (canUse32BitIndexMath(input) && canUse32BitIndexMath(grid) &&
               canUse32BitIndexMath(grad_output)) {
diff --git a/test/xpu/test_nn_xpu.py b/test/xpu/test_nn_xpu.py
index 80491dbb1af..35ac0066dff 100644
--- a/test/xpu/test_nn_xpu.py
+++ b/test/xpu/test_nn_xpu.py
@@ -14201,6 +14201,37 @@ def test_save_lstm_compatibility(self, device):
         result = loaded_model(x)
         self.assertEqual(result, expected)
 
+    def test_grid_sample_backward_error_checking(self, device):
+        input = torch.empty(1, 1, 2, 2, device=device)
+        grid = torch.empty(1, 1, 1, 2, device=device)
+        grad_output = torch.empty(1, 1, 1, 1, device=device)
+
+        # assert no error
+        torch.ops.aten.grid_sampler_2d_backward(
+            grad_output, input, grid, 0, 0, False, (True, True)
+        )
+
+        with self.assertRaisesRegex(ValueError, "expected grad_output to have sizes"):
+            invalid_grad_output = torch.empty(2, 1, 1, 1, device=device)
+            torch.ops.aten.grid_sampler_2d_backward(
+                invalid_grad_output, input, grid, 0, 0, False, (True, True)
+            )
+
+        input = torch.empty(1, 1, 2, 2, 2, device=device)
+        grid = torch.empty(1, 1, 1, 1, 3, device=device)
+        grad_output = torch.empty(1, 1, 1, 1, 1, device=device)
+
+        # assert no error
+        torch.ops.aten.grid_sampler_3d_backward(
+            grad_output, input, grid, 0, 0, False, (True, True)
+        )
+
+        with self.assertRaisesRegex(ValueError, "expected grad_output to have sizes"):
+            invalid_grad_output = torch.empty(2, 1, 1, 1, 1, device=device)
+            torch.ops.aten.grid_sampler_3d_backward(
+                invalid_grad_output, input, grid, 0, 0, False, (True, True)
+            )
+
     @onlyOn(["cuda", "xpu"])
     @tf32_on_and_off(0.005)
     def test_grid_sample_large(self, device):