diff --git a/src/op/parallel.cc b/src/op/parallel.cc
index 8a4790fd4..96d0dfda7 100644
--- a/src/op/parallel.cc
+++ b/src/op/parallel.cc
@@ -422,6 +422,7 @@ LayoutMap ParallelOpNode::InferLayout(const LayoutInferArgs &T,
 
   // Step 1: try to infer loop's partition from a source fragment
   Buffer source_buffer, read_source_buffer;
+  bool source_buffer_is_write = false;
   Buffer replicated_write_buffer; // Backup: fully replicated write buffer
 
   for (const auto &buffer : access_order_) {
@@ -437,6 +438,7 @@ LayoutMap ParallelOpNode::InferLayout(const LayoutInferArgs &T,
 
       if (access.is_write) {
         source_buffer = buffer;
+        source_buffer_is_write = true;
       } else {
         // Keep the buffer with largest number of indices
         // (which means the inference based on that buffer is more accurate)
@@ -455,6 +457,7 @@ LayoutMap ParallelOpNode::InferLayout(const LayoutInferArgs &T,
         if (frag.has_value() && is_one(frag.value()->ReplicateExtent()) &&
             !source_buffer.defined()) {
           source_buffer = buffer;
+          source_buffer_is_write = false;
         }
       }
     }
@@ -476,7 +479,7 @@ LayoutMap ParallelOpNode::InferLayout(const LayoutInferArgs &T,
       predicate_ = annotated_predicate_.value();
     }
   } else if (!loop_layout_.defined() && source_buffer.defined() &&
-             allow_layout_propgate) {
+             (allow_layout_propgate || source_buffer_is_write)) {
     loop_layout_ = ComputeLoopLayoutFromBuffer(source_buffer, T);
   } else if (!loop_layout_.defined() && level == InferLevel::kFree) {
     // For free layout inference
diff --git a/testing/python/layout/test_tilelang_layout_inference.py b/testing/python/layout/test_tilelang_layout_inference.py
index 13b58cf23..ad8f03ce0 100644
--- a/testing/python/layout/test_tilelang_layout_inference.py
+++ b/testing/python/layout/test_tilelang_layout_inference.py
@@ -60,5 +60,28 @@ def test_reject_fragment_write_from_non_owner_threads():
         _invalid_fragment_write_owner_layout()
 
 
+@tilelang.jit
+def _scalar_reduce_accumulator_owner_layout():
+    @T.prim_func
+    def main(out: T.Tensor((1,), T.float32)):
+        with T.Kernel(1, threads=128):
+            acc = T.alloc_fragment((128,), T.float32)
+            total = T.alloc_fragment((1,), T.float32)
+            total[0] = T.float32(0.0)
+            for k in T.Parallel(128):
+                acc[k] = T.float32(1.0)
+            T.reduce_sum(acc, total, clear=False)
+            if T.get_thread_binding() == 0:
+                out[0] = total[0]
+
+    return main
+
+
+def test_scalar_reduce_accumulator_owner_layout():
+    kernel = _scalar_reduce_accumulator_owner_layout()
+    source = kernel.get_kernel_source()
+    assert "AllReduce" in source
+
+
 if __name__ == "__main__":
     tilelang.testing.main()