diff --git a/include/ck_tile/core/arch/amd_buffer_addressing_builtins.hpp b/include/ck_tile/core/arch/amd_buffer_addressing_builtins.hpp
index 5c7ffefc6a..4e0a86119a 100644
--- a/include/ck_tile/core/arch/amd_buffer_addressing_builtins.hpp
+++ b/include/ck_tile/core/arch/amd_buffer_addressing_builtins.hpp
@@ -2570,6 +2570,60 @@ CK_TILE_DEVICE void amd_buffer_atomic_max(const thread_buffer<T, N>& src_thread_
 #endif
 }
 
+// amd_wave_read_first_lane is the SGPR function from AMD GPU device to load 1 or a series of the
+// memory to the SGPR registers.
+__device__ inline uint32_t amd_wave_read_first_lane(uint16_t v)
+{
+    return __builtin_amdgcn_readfirstlane(static_cast<uint32_t>(v));
+}
+
+__device__ inline uint32_t amd_wave_read_first_lane(uint8_t v)
+{
+    return __builtin_amdgcn_readfirstlane(static_cast<uint32_t>(v));
+}
+
+__device__ inline uint32_t amd_wave_read_first_lane(uint32_t value)
+{
+    return __builtin_amdgcn_readfirstlane(value);
+}
+
+__device__ inline int32_t amd_wave_read_first_lane(int32_t value)
+{
+    return __builtin_amdgcn_readfirstlane(value);
+}
+
+template <typename Object, std::enable_if_t<std::is_trivially_copyable_v<Object>, int> = 0>
+__device__ inline auto amd_wave_read_first_lane(const Object& obj)
+{
+    constexpr size_t ObjectSize = sizeof(Object);
+    constexpr size_t SGPR_size  = 4;
+    constexpr size_t NumFull    = ObjectSize / SGPR_size;
+    constexpr size_t Tail       = ObjectSize % SGPR_size;
+
+    const unsigned char* src = reinterpret_cast<const unsigned char*>(&obj);
+    alignas(Object) unsigned char dst[ObjectSize];
+
+    static_for<0, NumFull, 1>{}([&](auto Ic) {
+        constexpr size_t offset = Ic * SGPR_size;
+        uint32_t read_src;
+        __builtin_memcpy(&read_src, src + offset, SGPR_size);
+        read_src = __builtin_amdgcn_readfirstlane(read_src);
+        __builtin_memcpy(dst + offset, &read_src, SGPR_size);
+    });
+
+    if constexpr(Tail != 0)
+    {
+        constexpr size_t offset = NumFull * SGPR_size;
+        uint32_t tail_loc       = 0;
+        __builtin_memcpy(&tail_loc, src + offset, Tail);
+        tail_loc = __builtin_amdgcn_readfirstlane(tail_loc);
+        __builtin_memcpy(dst + offset, &tail_loc, Tail);
+    }
+    Object out;
+    __builtin_memcpy(&out, dst, ObjectSize);
+    return out;
+}
+
 template <typename T, index_t NumElemsPerThread>
 CK_TILE_DEVICE void amd_direct_load_global_to_lds(const T* global_base_ptr,
                                                   const index_t global_offset,
diff --git a/include/ck_tile/core/tensor/load_tile.hpp b/include/ck_tile/core/tensor/load_tile.hpp
index a3620453b4..2e9ab0f5c6 100644
--- a/include/ck_tile/core/tensor/load_tile.hpp
+++ b/include/ck_tile/core/tensor/load_tile.hpp
@@ -158,7 +158,4 @@ CK_TILE_DEVICE auto load_tile_raw(T& /*null_tile*/, const null_tile_window<Windo
 {
 }
 
-template <typename Tile>
-concept IsLoadableTile = requires { load_tile(std::declval<Tile>()); };
-
 } // namespace ck_tile
diff --git a/include/ck_tile/ops/epilogue/cshuffle_epilogue.hpp b/include/ck_tile/ops/epilogue/cshuffle_epilogue.hpp
index 6c815d804d..585a5f5b42 100644
--- a/include/ck_tile/ops/epilogue/cshuffle_epilogue.hpp
+++ b/include/ck_tile/ops/epilogue/cshuffle_epilogue.hpp
@@ -481,13 +481,10 @@ struct CShuffleEpilogue
         auto sm_tile = make_static_distributed_tensor<SMType>(dram_tile_distribution);
         auto sn_tile = make_static_distributed_tensor<SNType>(dram_tile_distribution);
 
-        // Build windows only if scales are provided
+        // Build windows only if non-scalar scales are provided
         auto scale_m_window = [&]() {
             if constexpr(has_scales && !has_scalar_scales)
             {
-                static_assert(
-                    IsLoadableTile<decltype(make_tile_window(scale_m, dram_tile_distribution))>,
-                    "ScaleM must be a loadable tile");
                 return make_tile_window(scale_m, dram_tile_distribution);
             }
             else
@@ -498,9 +495,6 @@ struct CShuffleEpilogue
         auto scale_n_window = [&]() {
             if constexpr(has_scales && !has_scalar_scales)
             {
-                static_assert(
-                    IsLoadableTile<decltype(make_tile_window(scale_n, dram_tile_distribution))>,
-                    "ScaleN must be a loadable tile");
                 return make_tile_window(scale_n, dram_tile_distribution);
             }
             else
@@ -515,8 +509,8 @@ struct CShuffleEpilogue
                 merge_sequences(sequence<mIter, 0>{}, c_warp_y_index_zeros),
                 merge_sequences(sequence<1, NRepeat>{}, c_warp_y_lengths));
 
-            // If scales provided, load them with identical distribution
-            if constexpr(has_scales && IsLoadableTile<ScaleM> && IsLoadableTile<ScaleN>)
+            // If non-scalar scales provided, load them with identical distribution
+            if constexpr(has_scales && !has_scalar_scales)
             {
                 sm_tile = load_tile(scale_m_window); // row scales in permuted layout
                 sn_tile = load_tile(scale_n_window); // col scales in permuted layout
@@ -535,7 +529,7 @@ struct CShuffleEpilogue
                     {
                         v = static_cast<AccDataType>(v * scale_m * scale_n);
                     }
-                    else if constexpr(has_scales)
+                    else if constexpr(has_scales && !has_scalar_scales)
                     {
                         // same linear index mapping on the permuted distribution
                         const auto s_m = static_cast<float>(sm_tile.get_thread_buffer()[out_idx]);
@@ -636,9 +630,6 @@ struct CShuffleEpilogue
             }
             else if constexpr(has_scales)
             {
-                static_assert(
-                    IsLoadableTile<decltype(make_tile_window(scale_m, dram_tile_distribution))>,
-                    "ScaleM must be a loadable tile");
                 return make_tile_window(scale_m, lds_tile.get_tile_distribution());
             }
             else
@@ -653,9 +644,6 @@ struct CShuffleEpilogue
             }
             else if constexpr(has_scales)
             {
-                static_assert(
-                    IsLoadableTile<decltype(make_tile_window(scale_n, dram_tile_distribution))>,
-                    "ScaleN must be a loadable tile");
                 return make_tile_window(scale_n, lds_tile.get_tile_distribution());
             }
             else