[XPU] Validate replication_pad2d_backward channel dimension

src/ATen/native/xpu/sycl/ReplicationPaddingKernels.cpp

@@ -530,19 +530,35 @@ void replication_pad2d_backward_kernel(
   const auto padR = padding[1];
   const auto padT = padding[2];
   const auto padB = padding[3];
+  int dimc = 0;
   int dimh = 1;
   int dimw = 2;
 
   int numInputDims = input.dim();
   if (numInputDims == 4) {
+    dimc++;
     dimh++;
     dimw++;
   }
+  const auto ichannel = input.size(dimc);
   const auto iheight = input.size(dimh);
   const auto iwidth = input.size(dimw);
   const auto oheight = iheight + padT + padB;
   const auto owidth = iwidth + padL + padR;
 
+  // Validate the channel (plane) dimension in addition to the spatial dims.
+  // Without this check a grad_output whose channel count does not match the
+  // input (e.g. 0 channels) passes the width/height checks, a non-empty
+  // grad_input is allocated from the input shape, and the kernel reads
+  // grad_output out of bounds -> SIGSEGV. Mirrors the 3d backward check and
+  // the CPU/CUDA fix in pytorch/pytorch#189463.
+  TORCH_CHECK(
+      ichannel == grad_output.size(dimc),
+      "gradOutput channel unexpected. Expected: ",
+      ichannel,
+      ", Got: ",
+      grad_output.size(dimc));
+
   TORCH_CHECK(
       owidth == grad_output.size(dimw),
       "grad_output width unexpected. Expected: ",

test/regressions/test_replication_pad_channel_mismatch.py

@@ -0,0 +1,51 @@
+# Copyright 2020-2026 Intel Corporation
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+
+# Owner(s): ["module: intel"]
+#
+# Regression test for the XPU replication_pad2d_backward channel-mismatch bug.
+#
+# replication_pad2d_backward validated the spatial (width/height) dims of
+# grad_output but not the channel/plane dim.  A grad_output with a mismatched
+# channel count (e.g. 0 channels) passed the checks, a non-empty grad_input was
+# allocated from the input shape, and the kernel read grad_output out of bounds
+# -> SIGSEGV.  The fix mirrors the 3d backward channel check (and the CPU/CUDA
+# fix in pytorch/pytorch#189463) into the 2d XPU backward, raising a clear
+# "gradOutput channel unexpected" error instead.  See intel/torch-xpu-ops#4754.
+import torch
+from torch.testing._internal.common_utils import TestCase
+
+xpu_device = torch.device("xpu")
+
+
+class TestReplicationPadChannelMismatch(TestCase):
+    def test_replication_pad2d_backward_channel_mismatch_xpu(self):
+        # channel dim of grad_output (0) does not match input's (2).
+        grad_output = torch.ones(2, 0, 6, 8, device=xpu_device)
+        inp = torch.ones(2, 2, 4, 4, device=xpu_device)
+        with self.assertRaisesRegex(RuntimeError, "gradOutput channel unexpected"):
+            torch.ops.aten.replication_pad2d_backward(grad_output, inp, [2, 2, 1, 1])
+
+    def test_replication_pad2d_backward_valid_xpu(self):
+        # A matching-channel grad_output must still work (no false rejection).
+        inp = torch.randn(2, 3, 4, 4, device=xpu_device, requires_grad=True)
+        out = torch.nn.functional.pad(inp, [2, 2, 1, 1], mode="replicate")
+        out.sum().backward()
+        self.assertEqual(inp.grad.shape, inp.shape)
+
+    def test_replication_pad2d_backward_matches_cpu_xpu(self):
+        inp = torch.randn(2, 3, 4, 4)
+        inp_cpu = inp.clone().requires_grad_()
+        out_cpu = torch.nn.functional.pad(inp_cpu, [2, 2, 1, 1], mode="replicate")
+        out_cpu.sum().backward()
+
+        inp_xpu = inp.clone().to(xpu_device).requires_grad_()
+        out_xpu = torch.nn.functional.pad(inp_xpu, [2, 2, 1, 1], mode="replicate")
+        out_xpu.sum().backward()
+
+        self.assertEqual(inp_cpu.grad, inp_xpu.grad.cpu())