diff --git a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_bpreshuffle.cpp b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_bpreshuffle.cpp
index d5db3199ca..bd9c358b57 100644
--- a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_bpreshuffle.cpp
+++ b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_bpreshuffle.cpp
@@ -140,14 +140,14 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_Xdl_CShu
     // clang-format off
     <   Row, Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,
         AElementOp,  BElementOp, CDEElementOp, GemmSpec, 256,
-        32,   128,    128,
+        128,   128,    128,
         16,   16,
         32,   32,
-        1,    1,
+        4,    1,
         S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
         S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
         1,    1,   S<1, 32, 1, 8>, S<8, 8, 1>,
-        ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v2, FP8>;
+        ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, FP8>;
 // clang-format on
 
 int main(int argc, char* argv[])
diff --git a/example/65_gemm_multiply_multiply/moe_gemm2_xdl_fp8.cpp b/example/65_gemm_multiply_multiply/moe_gemm2_xdl_fp8.cpp
index 04f10b53ae..12a6f4bcb2 100644
--- a/example/65_gemm_multiply_multiply/moe_gemm2_xdl_fp8.cpp
+++ b/example/65_gemm_multiply_multiply/moe_gemm2_xdl_fp8.cpp
@@ -185,8 +185,8 @@ int main(int argc, char* argv[])
     ck::index_t N               = 4096;
     ck::index_t K               = 4096;
     ck::index_t experts         = 8;
-    ck::index_t sorted_tile_num = 6;
-    ck::index_t valid_tile_num  = 6;
+    ck::index_t sorted_tile_num = 18;
+    ck::index_t valid_tile_num  = 16;
     ck::index_t sorted_size     = sorted_tile_num * MPerBlock;
     ck::index_t valid_size      = valid_tile_num * MPerBlock;
     ck::index_t tokens          = 128;
@@ -237,10 +237,10 @@ int main(int argc, char* argv[])
     // max_token_id.mData = {valid_size, 0, 2, 3, 4, 6, 8, 10, 12, 13};
     // int eids[]         = {0, 0, 1, 2, 3, 3, 4, 4, 5, 5, 6, 7, 7, 3, 3, 3};
     max_token_id.mData = {valid_size, 0, 1, 2, 3, 4, 5, 6, 7, 8};
-    int eids[]         = {0, 1, 2, 3, 4, 5, 6, 7, 3, 3, 3}; // {2, 1, 1, 2, 2, 2, 1, 2}
+    // int eids[]         = {0, 1, 2, 3, 4, 5, 6, 7, 3, 3, 3}; // {2, 1, 1, 2, 2, 2, 1, 2}
     for(int i = 0; i < sorted_tile_num; i++)
     {
-        expert_ids.mData[i] = eids[i];
+        expert_ids.mData[i] = i / ((valid_tile_num + experts - 1) / experts);
     }
     if(tokens * topk > valid_size)
     {
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_moe_gemm.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_moe_gemm.hpp
index a2d1114bbe..fba46d4ac6 100644
--- a/include/ck/tensor_operation/gpu/grid/gridwise_moe_gemm.hpp
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_moe_gemm.hpp
@@ -186,7 +186,9 @@ struct GridwiseMoeGemm
         math::max(math::lcm(AK1Number, BK1Number), mfma_selector::selected_mfma.k_per_blk);
     static constexpr index_t KLane =
         mfma_selector::GetKPerXdlops() / mfma_selector::GetK1PerXdlops();
-    static constexpr index_t KRepeat = KPerBlock / KLane / KPack;
+
+    static constexpr index_t KGroup = mfma_selector::selected_mfma.k_per_blk == 32 ? 2 : 1;
+    static constexpr index_t KRepeat = KPerBlock / KLane / (KPack / KGroup);
     static constexpr index_t NLane   = NPerXdl;
     static constexpr index_t NWave   = NPerBlock / NPerXdl / NXdlPerWave;
     // static constexpr index_t NumTokens = 1;
@@ -246,7 +248,7 @@ struct GridwiseMoeGemm
     }
     __host__ __device__ static auto CalculateBK0Shuffled(index_t K)
     {
-        return math::integer_divide_ceil(K, KLane * KPack);
+        return math::integer_divide_ceil(K, KLane * KPack / KGroup);
     }
 
     __host__ __device__ static auto CalculateKPadded(index_t K)
@@ -388,7 +390,7 @@ struct GridwiseMoeGemm
 
     __host__ __device__ static auto MakeBGridDescriptor_Preshuffled(index_t N0, index_t K0)
     {
-        constexpr index_t NkSwizzleNumber = Number<warpSize * KPack>{};
+        constexpr index_t NkSwizzleNumber = Number<warpSize * KPack / KGroup>{};
         return make_naive_tensor_descriptor(
             make_tuple(N0 / NWave, NWave, K0, NkSwizzleNumber),
             make_tuple(NWave * K0 * NkSwizzleNumber, K0 * NkSwizzleNumber, NkSwizzleNumber, I1));
@@ -1297,7 +1299,7 @@ struct GridwiseMoeGemm
                   make_multi_index(n_block_data_idx_on_grid,
                                    get_warp_local_1d_id() % NWave,
                                    0,
-                                   KPack * (get_thread_local_1d_id() % warpSize)));
+                                   KPack / KGroup * (get_thread_local_1d_id() % warpSize)));
 
         // LDS allocation for A and B: be careful of alignment
         // Cast after lds
@@ -1804,7 +1806,7 @@ struct GridwiseMoeGemm
                   make_multi_index(n_block_data_idx_on_grid,
                                    get_warp_local_1d_id() % NWave,
                                    0,
-                                   KPack * (get_thread_local_1d_id() % warpSize)));
+                                   KPack / KGroup * (get_thread_local_1d_id() % warpSize)));
 
         // LDS allocation for A and B: be careful of alignment
         // Cast after lds