From 070df8bdc44b8951116940ec05ef14c016579b77 Mon Sep 17 00:00:00 2001
From: Jake Stevens <jrstevens@meta.com>
Date: Tue, 27 Oct 2026 14:09:16 -0700
Subject: [PATCH] Accept empty tensor bindings in CPUBackend execute (#24195)

Summary:
Pull Request resolved: https://github.com/pytorch/executorch/pull/24195

The CPU delegate rejected any binding whose strides/dim-order were not canonical-contiguous, including empty (0-byte) tensors whose layout is meaningless: PyTorch does not canonicalize strides for 0-size dims, so e.g. a (0,0,0,0,0) input arrives with dim-order != identity and fails with 0x12 (InvalidArgument).

Skip the contiguity check for 0-byte bindings; the input/output memcpys already skip empty buffers, and providers handle 0-size execution.

Reviewed By: digantdesai

Differential Revision: D126930703
---
 backends/cpu/runtime/CPUBackend.cpp | 23 ++++++++++++++---------
 backends/cpu/test/test_runtime.py   | 11 +++++++++++
 2 files changed, 25 insertions(+), 9 deletions(-)

diff --git a/backends/cpu/runtime/CPUBackend.cpp b/backends/cpu/runtime/CPUBackend.cpp
index 963bf3e7f13..3ff1396c71d 100644
--- a/backends/cpu/runtime/CPUBackend.cpp
+++ b/backends/cpu/runtime/CPUBackend.cpp
@@ -183,15 +183,20 @@ class CPUExecution {
           InvalidArgument,
           "CPU static shape/dtype/layout/storage mismatch: %s",
           method_->graph.value(id).name.c_str());
-      int64_t stride = 1;
-      for (size_t reverse = expected.sizes.size(); reverse > 0; --reverse) {
-        const size_t dim = reverse - 1;
-        ET_CHECK_OR_RETURN_ERROR(
-            tensor.dim_order()[dim] == dim && tensor.strides()[dim] == stride,
-            InvalidArgument,
-            "CPU binding must be contiguous: %s",
-            method_->graph.value(id).name.c_str());
-        stride *= expected.sizes[dim];
+      // Empty bindings carry no data, so their strides and dim order are
+      // meaningless (PyTorch does not canonicalize them); skip the contiguity
+      // check rather than rejecting valid empty tensors.
+      if (tensor.nbytes() != 0) {
+        int64_t stride = 1;
+        for (size_t reverse = expected.sizes.size(); reverse > 0; --reverse) {
+          const size_t dim = reverse - 1;
+          ET_CHECK_OR_RETURN_ERROR(
+              tensor.dim_order()[dim] == dim && tensor.strides()[dim] == stride,
+              InvalidArgument,
+              "CPU binding must be contiguous: %s",
+              method_->graph.value(id).name.c_str());
+          stride *= expected.sizes[dim];
+        }
       }
     }
     // EValue exposes the logical span, but no backing capacity or readable
diff --git a/backends/cpu/test/test_runtime.py b/backends/cpu/test/test_runtime.py
index b3d1df05982..4e821d76eec 100644
--- a/backends/cpu/test/test_runtime.py
+++ b/backends/cpu/test/test_runtime.py
@@ -95,6 +95,17 @@ def forward(self, x, mask):
                 for output, expected in zip(actual, model(x, current), strict=True):
                     torch.testing.assert_close(output, expected, rtol=0, atol=0)
 
+    def test_empty_tensors_execute(self):
+        class Add(torch.nn.Module):
+            def forward(self, x):
+                return x + 1
+
+        model = Add()
+        x = torch.empty(2, 0, 3)
+        with self._lower(model, (x,)) as runtime:
+            actual = runtime.forward((x,))[0]
+            torch.testing.assert_close(actual, model(x))
+
     def test_neg_view_copy_routes_to_neg(self):
         class NegView(torch.nn.Module):
             def forward(self, x):