diff --git a/example/ck_tile/38_block_scale_gemm/gemm_utils.hpp b/example/ck_tile/38_block_scale_gemm/gemm_utils.hpp
index d64297cb35..930cdefb7e 100644
--- a/example/ck_tile/38_block_scale_gemm/gemm_utils.hpp
+++ b/example/ck_tile/38_block_scale_gemm/gemm_utils.hpp
@@ -18,7 +18,7 @@
 template <typename PrecType, ck_tile::index_t M_Warp_Tile>
 constexpr ck_tile::index_t get_k_warp_tile()
 {
-#if defined(__gfx950__)
+#if defined(CK_GFX950_SUPPORT)
     constexpr bool is_8bit_float =
         std::is_same_v<PrecType, ck_tile::fp8_t> || std::is_same_v<PrecType, ck_tile::bf8_t>;
     if constexpr(M_Warp_Tile == 32)
@@ -35,7 +35,7 @@ constexpr ck_tile::index_t get_k_warp_tile()
 template <typename PrecType, ck_tile::index_t M_Warp_Tile>
 constexpr ck_tile::index_t get_k_from_preshuffled_warp_tile()
 {
-#if defined(__gfx950__)
+#if defined(CK_GFX950_SUPPORT)
     if constexpr(M_Warp_Tile == 32)
         return sizeof(PrecType) == 2 ? 16 : 64;
     else
diff --git a/include/ck_tile/ops/gemm/warp/warp_gemm.hpp b/include/ck_tile/ops/gemm/warp/warp_gemm.hpp
index 87772f78fc..f83bbc2a18 100644
--- a/include/ck_tile/ops/gemm/warp/warp_gemm.hpp
+++ b/include/ck_tile/ops/gemm/warp/warp_gemm.hpp
@@ -301,6 +301,30 @@ using WarpGemmMfma_f32_16x16x128_bf8_bf8 = WarpGemmImpl<
     WarpGemmAttributeMfma<WarpGemmAttributeMfmaImpl_f32_16x16x128_bf8_bf8<WGAttrCtlEnum::Default_>,
                           AttrNumAccess>>;
 
+template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
+using WarpGemmMfma_f32_16x16x128_fp8_fp8_CTransposed =
+    WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution<
+        WarpGemmAttributeMfmaImpl_f32_16x16x128_fp8_fp8<WGAttrCtlEnum::Default_>,
+        AttrNumAccess>>;
+
+template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
+using WarpGemmMfma_f32_16x16x128_fp8_bf8_CTransposed =
+    WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution<
+        WarpGemmAttributeMfmaImpl_f32_16x16x128_fp8_bf8<WGAttrCtlEnum::Default_>,
+        AttrNumAccess>>;
+
+template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
+using WarpGemmMfma_f32_16x16x128_bf8_fp8_CTransposed =
+    WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution<
+        WarpGemmAttributeMfmaImpl_f32_16x16x128_bf8_fp8<WGAttrCtlEnum::Default_>,
+        AttrNumAccess>>;
+
+template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
+using WarpGemmMfma_f32_16x16x128_bf8_bf8_CTransposed =
+    WarpGemmImpl<WarpGemmAttributeMfmaTransposedCDistribution<
+        WarpGemmAttributeMfmaImpl_f32_16x16x128_bf8_bf8<WGAttrCtlEnum::Default_>,
+        AttrNumAccess>>;
+
 template <WGAttrNumAccessEnum AttrNumAccess = WGAttrNumAccessEnum::Single>
 using WarpGemmMfma_f32_32x32x64_fp8_fp8 = WarpGemmImpl<
     WarpGemmAttributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x64_fp8_fp8<WGAttrCtlEnum::Default_>,
diff --git a/include/ck_tile/ops/gemm/warp/warp_gemm_dispatcher.hpp b/include/ck_tile/ops/gemm/warp/warp_gemm_dispatcher.hpp
index 5021fb9907..1d3dd2ae6f 100644
--- a/include/ck_tile/ops/gemm/warp/warp_gemm_dispatcher.hpp
+++ b/include/ck_tile/ops/gemm/warp/warp_gemm_dispatcher.hpp
@@ -112,6 +112,10 @@ template<> struct WarpGemmDispatcher<ck_tile::fp8_t, ck_tile::fp8_t, float, 16,
 template<> struct WarpGemmDispatcher<ck_tile::fp8_t, ck_tile::bf8_t, float, 16, 16,  128, false> { using Type = WarpGemmMfma_f32_16x16x128_fp8_bf8<>; };
 template<> struct WarpGemmDispatcher<ck_tile::bf8_t, ck_tile::fp8_t, float, 16, 16,  128, false> { using Type = WarpGemmMfma_f32_16x16x128_bf8_fp8<>; };
 template<> struct WarpGemmDispatcher<ck_tile::bf8_t, ck_tile::bf8_t, float, 16, 16,  128, false> { using Type = WarpGemmMfma_f32_16x16x128_bf8_bf8<>; };
+template<> struct WarpGemmDispatcher<ck_tile::fp8_t, ck_tile::fp8_t, float, 16, 16,  128, true> { using Type = WarpGemmMfma_f32_16x16x128_fp8_fp8_CTransposed<>; };
+template<> struct WarpGemmDispatcher<ck_tile::fp8_t, ck_tile::bf8_t, float, 16, 16,  128, true> { using Type = WarpGemmMfma_f32_16x16x128_fp8_bf8_CTransposed<>; };
+template<> struct WarpGemmDispatcher<ck_tile::bf8_t, ck_tile::fp8_t, float, 16, 16,  128, true> { using Type = WarpGemmMfma_f32_16x16x128_bf8_fp8_CTransposed<>; };
+template<> struct WarpGemmDispatcher<ck_tile::bf8_t, ck_tile::bf8_t, float, 16, 16,  128, true> { using Type = WarpGemmMfma_f32_16x16x128_bf8_bf8_CTransposed<>; };
 
 template<> struct WarpGemmDispatcher<ck_tile::fp8_t, ck_tile::fp8_t, float, 32, 32,  64, false> { using Type = WarpGemmMfma_f32_32x32x64_fp8_fp8<>; };
 template<> struct WarpGemmDispatcher<ck_tile::fp8_t, ck_tile::bf8_t, float, 32, 32,  64, false> { using Type = WarpGemmMfma_f32_32x32x64_fp8_bf8<>; };
diff --git a/test/ck_tile/gemm_block_scale/test_gemm_aquant_utils.hpp b/test/ck_tile/gemm_block_scale/test_gemm_aquant_utils.hpp
index cf9bf18c5a..61bb1a8bdd 100644
--- a/test/ck_tile/gemm_block_scale/test_gemm_aquant_utils.hpp
+++ b/test/ck_tile/gemm_block_scale/test_gemm_aquant_utils.hpp
@@ -18,7 +18,7 @@
 template <typename PrecType, ck_tile::index_t M_Warp_Tile>
 constexpr ck_tile::index_t get_k_warp_tile()
 {
-#if defined(__gfx950__)
+#if defined(CK_GFX950_SUPPORT)
     constexpr bool is_8bit_float =
         std::is_same_v<PrecType, ck_tile::fp8_t> || std::is_same_v<PrecType, ck_tile::bf8_t>;
     if constexpr(M_Warp_Tile == 32)