From 407489d2c041abfdb65e9f21f8548470e7e4417c Mon Sep 17 00:00:00 2001
From: "Ding, Yi" <yi.ding@amd.com>
Date: Tue, 3 Jun 2025 09:29:50 +0000
Subject: [PATCH] Fix ThreadwiseTensorSliceTransfer_v4::Run (Fuse scale)

---
 .../gpu/thread/threadwise_tensor_slice_transfer.hpp           | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp
index 063bd71de6..ad91b65fb8 100644
--- a/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp
+++ b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp
@@ -1496,7 +1496,7 @@ struct ThreadwiseTensorSliceTransfer_v4
             }
         });
     }
-#if 0
+
     // Fuse scale
     template <typename SrcRefToOriginDisplacement,
               typename DstOriginIdx,
@@ -1696,7 +1696,7 @@ struct ThreadwiseTensorSliceTransfer_v4
             }
         });
     }
-#endif
+
     template <typename SrcSliceMoveStepIdx>
     __device__ void MoveSrcSliceWindow(const SrcDesc&,
                                        const SrcSliceMoveStepIdx& src_slice_move_step_idx)