diff --git a/sgl-kernel/cmake/flashmla.cmake b/sgl-kernel/cmake/flashmla.cmake
index a1b1353a8..7387f3e5d 100644
--- a/sgl-kernel/cmake/flashmla.cmake
+++ b/sgl-kernel/cmake/flashmla.cmake
@@ -1,8 +1,9 @@
 # flash_mla
+# sm90 dense decode HEAD_DIM_K=512 support (sgl-project/FlashMLA#9, merged).
 FetchContent_Declare(
     repo-flashmla
-    URL      https://${GITHUB_ARTIFACTORY}/sgl-project/FlashMLA/archive/df022ebafb88578eab9f0300606ee765608d8b5c.tar.gz
-    URL_HASH SHA256=45992d7de7d051dc897aff33156a2b7515d745fc489008adc7924ab884578d52
+    URL      https://${GITHUB_ARTIFACTORY}/sgl-project/FlashMLA/archive/05e26647fe840b8baedae486c2d86d5ce4efeb7c.tar.gz
+    URL_HASH SHA256=ce369489bbfc42cdfbba9aa949de0270e64469d530748dea9f4f60b3c69dea9b
 )
 FetchContent_Populate(repo-flashmla)
 
@@ -156,6 +157,16 @@ target_compile_options(flashmla_ops PRIVATE
 if(FLASHMLA_ENABLE_SM100)
     target_compile_definitions(flashmla_ops PRIVATE FLASHMLA_ENABLE_SM100)
 endif()
+
+# CUDA 13 moved cuda/std/* under cccl/cuda/std/*. The vendored cutlass routes
+# <cuda/std/...> to <cccl/cuda/std/...> when __CUDACC_VER_MAJOR__ >= 13, so the
+# host C++ TU (compiled by g++, where that macro is unset for the legacy path)
+# needs the cccl include root on the search path.
+if(CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL "13.0")
+    find_path(FLASHMLA_CCCL_INCLUDE NAMES cuda/std/utility
+        HINTS ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES}
+              ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES}/cccl)
+endif()
 target_include_directories(flashmla_ops PRIVATE
     ${repo-flashmla_SOURCE_DIR}/csrc
     ${repo-flashmla_SOURCE_DIR}/csrc/kerutils/include
@@ -163,6 +174,7 @@ target_include_directories(flashmla_ops PRIVATE
     ${repo-flashmla_SOURCE_DIR}/csrc/extension/sm90/dense_fp8/
     ${repo-flashmla_SOURCE_DIR}/csrc/cutlass/include
     ${repo-flashmla_SOURCE_DIR}/csrc/cutlass/tools/util/include
+    ${FLASHMLA_CCCL_INCLUDE}
 )
 
 target_link_libraries(flashmla_ops PRIVATE ${TORCH_LIBRARIES} c10 cuda)