From 91e3d1542e545c9429b86ad72c526f291ed2b4dc Mon Sep 17 00:00:00 2001
From: Baizhou Zhang <sobereddiezhang@gmail.com>
Date: Sun, 27 Jul 2025 00:36:15 -0700
Subject: [PATCH] Update Cutlass in sgl-kernel to v4.1 (#8392)

---
 sgl-kernel/CMakeLists.txt                                       | 2 +-
 .../sm90_mma_array_tma_gmma_rs_warpspecialized_mixed_input_.hpp | 1 -
 2 files changed, 1 insertion(+), 2 deletions(-)

diff --git a/sgl-kernel/CMakeLists.txt b/sgl-kernel/CMakeLists.txt
index 739782372..b359c76c1 100644
--- a/sgl-kernel/CMakeLists.txt
+++ b/sgl-kernel/CMakeLists.txt
@@ -45,7 +45,7 @@ include(FetchContent)
 FetchContent_Declare(
     repo-cutlass
     GIT_REPOSITORY https://github.com/NVIDIA/cutlass
-    GIT_TAG        f115c3f85467d5d9619119d1dbeb9c03c3d73864
+    GIT_TAG        664c4f7b3ed1959414905025728eef5568209479
     GIT_SHALLOW    OFF
 )
 FetchContent_Populate(repo-cutlass)
diff --git a/sgl-kernel/csrc/cutlass_extensions/gemm/collective/sm90_mma_array_tma_gmma_rs_warpspecialized_mixed_input_.hpp b/sgl-kernel/csrc/cutlass_extensions/gemm/collective/sm90_mma_array_tma_gmma_rs_warpspecialized_mixed_input_.hpp
index 6e1a01e22..13e890e35 100644
--- a/sgl-kernel/csrc/cutlass_extensions/gemm/collective/sm90_mma_array_tma_gmma_rs_warpspecialized_mixed_input_.hpp
+++ b/sgl-kernel/csrc/cutlass_extensions/gemm/collective/sm90_mma_array_tma_gmma_rs_warpspecialized_mixed_input_.hpp
@@ -36,7 +36,6 @@
 #include "cute/arch/copy_sm90.hpp"
 #include "cute/atom/mma_atom.hpp"
 #include "cute/numeric/arithmetic_tuple.hpp"
-#include "cute/tensor_predicate.hpp"
 #include "cutlass/cuda_host_adapter.hpp"
 #include "cutlass/cutlass.h"
 #include "cutlass/gemm/dispatch_policy.hpp"