diff --git a/.gitmodules b/.gitmodules
index 3a14f6297..c588176e7 100644
--- a/.gitmodules
+++ b/.gitmodules
@@ -1,3 +1,6 @@
 [submodule "sgl-kernel/3rdparty/cutlass"]
 	path = sgl-kernel/3rdparty/cutlass
 	url = https://github.com/NVIDIA/cutlass.git
+[submodule "sgl-kernel/3rdparty/cub"]
+	path = sgl-kernel/3rdparty/cub
+	url = https://github.com/NVIDIA/cub.git
diff --git a/sgl-kernel/3rdparty/cub b/sgl-kernel/3rdparty/cub
new file mode 160000
index 000000000..0fc3c3701
--- /dev/null
+++ b/sgl-kernel/3rdparty/cub
@@ -0,0 +1 @@
+Subproject commit 0fc3c3701632a4be906765b73be20a9ad0da603d
diff --git a/sgl-kernel/CMakeLists.txt b/sgl-kernel/CMakeLists.txt
index 15818d289..623984f2f 100644
--- a/sgl-kernel/CMakeLists.txt
+++ b/sgl-kernel/CMakeLists.txt
@@ -9,6 +9,7 @@ set(CMAKE_CUDA_STANDARD 17)
 set(CMAKE_CUDA_STANDARD_REQUIRED ON)
 
 set(CUTLASS_DIR "3rdparty/cutlass")
+set(CUB_DIR "3rdparty/cub")
 
 # Set CUDA architectures
 set(CMAKE_CUDA_ARCHITECTURES "75;80;86;89;90")
@@ -43,6 +44,7 @@ target_include_directories(_kernels
         ${TORCH_INCLUDE_DIRS}
         ${CUTLASS_DIR}/include
         ${CUTLASS_DIR}/tools/util/include
+        ${CUB_DIR}/cub
 )
 
 target_link_libraries(_kernels