[ONNX] Fix circular padding to support dynamic axes (#95647)

This commit fixes a bug where the ONNX exporter for circular padding queried the input tensor shape in order to get the correct 'end' index for a slice node. This doesn't work when the axis in question is has dynamic size. The commit fixes this by setting the 'end' index to INT_MAX, which is the recommended way of slicing to the end of a dimension with unknown size per ONNX spec.

See https://onnx.ai/onnx/operators/onnx__Slice.html

Also adds a regression test.

Pull Request resolved: #95647
Approved by: https://github.com/BowenBao

test/onnx/test_pytorch_onnx_onnxruntime.py

@@ -7585,6 +7585,21 @@ def forward(self, x):
         x = torch.randn(2, 3, 6)
         self.run_test(PadModel(), (x))
 
+    @skipIfUnsupportedMinOpsetVersion(11)
+    def test_pad_circular_dynamic_axes(self):
+        class PadModel(torch.nn.Module):
+            def forward(self, x):
+                out = torch.nn.functional.pad(x, (2, 1, 2, 1), mode="circular")
+                return out
+
+        x = torch.randn(4, 3, 5, 6)
+        self.run_test(
+            PadModel(),
+            x,
+            input_names=["input_1"],
+            dynamic_axes={"input_1": [0, 1, 2, 3]},
+        )
+
     @skipIfUnsupportedMaxOpsetVersion(10)
     @skipScriptTest()  # TODO: the logic in symbolic_opset9 doesn't handle script
     def test_unsupported_pad(self):

torch/onnx/symbolic_opset9.py

@@ -1915,12 +1915,10 @@ def _pad_circular(g: jit_utils.GraphContext, input: _C.Value, pad: _C.Value):
     for idx in range(ndim):
         pad_r = padding[-(2 * idx + 1)]
         pad_l = padding[-(2 * idx + 2)]
-        # get size for targeting the last idx, as Slice don't take start=[-1], end=[-1]
-        size = symbolic_helper._get_tensor_sizes(input)
         tensors = []
         if pad_l > 0:
             left = symbolic_helper._slice_helper(
-                g, cur, axes=[2 + idx], starts=[-(pad_l)], ends=[size[2 + idx]]
+                g, cur, axes=[2 + idx], starts=[-(pad_l)], ends=[_constants.INT64_MAX]
             )
             tensors.append(left)