diff --git a/runtime/kernel/targets.bzl b/runtime/kernel/targets.bzl
index e67f76728b8..5c95f10276d 100644
--- a/runtime/kernel/targets.bzl
+++ b/runtime/kernel/targets.bzl
@@ -56,6 +56,7 @@ def define_common_targets():
         exported_headers = ["thread_parallel_interface.h"],
         exported_deps = [
             "//executorch/runtime/core:core",
+            "//executorch/runtime/core/portable_type/c10/c10:c10",
             "//executorch/runtime/platform:platform",
         ],
         visibility = [
diff --git a/runtime/kernel/thread_parallel_interface.h b/runtime/kernel/thread_parallel_interface.h
index 1e79acc75a4..52100475c7b 100644
--- a/runtime/kernel/thread_parallel_interface.h
+++ b/runtime/kernel/thread_parallel_interface.h
@@ -11,6 +11,7 @@
 #include <cstdint>
 #include <functional>
 
+#include <c10/util/irange.h>
 #include <executorch/runtime/core/error.h>
 #include <executorch/runtime/platform/assert.h>
 
@@ -29,7 +30,17 @@ inline bool parallel_for_no_threadpool(
       begin,
       end);
   ET_CHECK_OR_RETURN_FALSE(grain_size > 0, "grain_size = %" PRId64, grain_size);
+#ifndef NDEBUG
+  // Go backwards through the range elementwise to catch code that
+  // assumes parallel_for is in order like a regular for loop.
+  for (const auto i : c10::irange(begin, end)) {
+    const auto offset = i - begin;
+    const auto idx = end - offset - 1;
+    f(idx, idx + 1);
+  }
+#else // NDEBUG
   f(begin, end);
+#endif
   return true;
 }