fixed

include/ck/utility/static_buffer.hpp

@@ -114,7 +114,10 @@ struct StaticBufferTupleOfVector
 
     // Get X
     // i is offset of S, not X. i should be aligned to X
-    template <typename X, index_t I>
+    template <typename X,
+              index_t I,
+              typename enable_if<has_same_scalar_type<S, X>::value || !is_native_type<S>(),
+                                 bool>::type = false>
     __host__ __device__ constexpr auto GetAsType(Number<I> i) const
     {
         constexpr auto s_per_x = Number<scalar_type<remove_cvref_t<X>>::vector_size>{};
@@ -130,7 +133,10 @@ struct StaticBufferTupleOfVector
 
     // Set X
     // i is offset of S, not X. i should be aligned to X
-    template <typename X, index_t I>
+    template <typename X,
+              index_t I,
+              typename enable_if<has_same_scalar_type<S, X>::value || !is_native_type<S>(),
+                                 bool>::type = false>
     __host__ __device__ constexpr void SetAsType(Number<I> i, X x)
     {
         constexpr auto s_per_x = Number<scalar_type<remove_cvref_t<X>>::vector_size>{};

profiler/src/CMakeLists.txt

@@ -43,7 +43,6 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
     list(APPEND PROFILER_SOURCES profile_gemm_add_silu.cpp)
     list(APPEND PROFILER_SOURCES profile_gemm_add_relu_add_layernorm.cpp)
     list(APPEND PROFILER_SOURCES profile_grouped_gemm_fixed_nk.cpp)
-    list(APPEND PROFILER_SOURCES profile_grouped_gemm_two_stage.cpp)
     list(APPEND PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp)
     list(APPEND PROFILER_SOURCES profile_grouped_gemm_tile_loop.cpp)
     list(APPEND PROFILER_SOURCES profile_grouped_gemm_multiply_tile_loop.cpp)