diff --git a/example/65_gemm_multiply_multiply/moe_gemm1_xdl_fp8.cpp b/example/65_gemm_multiply_multiply/moe_gemm1_xdl_fp8.cpp
index bc94162cd7..186416904c 100644
--- a/example/65_gemm_multiply_multiply/moe_gemm1_xdl_fp8.cpp
+++ b/example/65_gemm_multiply_multiply/moe_gemm1_xdl_fp8.cpp
@@ -121,12 +121,12 @@ using AElementOp = PassThrough;
 using BElementOp = PassThrough;
 
 static constexpr auto GemmSpec         = ck::tensor_operation::device::GemmSpecialization::Default;
-static constexpr ck::index_t MPerBlock = 32;
-static constexpr ck::index_t MXDLPerWave = 1;
-static constexpr ck::index_t NXDLPerWave = 1;
+static constexpr ck::index_t MPerBlock = 128;
+static constexpr ck::index_t MXDLPerWave = 4;
+static constexpr ck::index_t NXDLPerWave = 2;
 static constexpr ck::index_t BLOCKSIZE   = 256;
-static constexpr ck::index_t NPerBlock   = 128;
-static constexpr ck::index_t MNPerXDL    = 32;
+static constexpr ck::index_t NPerBlock   = 64;
+static constexpr ck::index_t MNPerXDL    = 16;
 static constexpr ck::index_t KPerBlock   = 128 / sizeof(A0DataType);
 static constexpr ck::index_t Nswizzle    = false;
 static constexpr ck::index_t AK1         = 16 / sizeof(A0DataType);
@@ -134,7 +134,7 @@ static constexpr ck::index_t BK1         = 16 / sizeof(B0DataType);
 static constexpr ck::index_t EVec        = 16 / sizeof(EDataType);
 static constexpr ck::index_t D0Vec       = 1;
 static constexpr ck::index_t D1Vec       = 1;
-static constexpr ck::index_t ActOP       = 0; // 0: gelu, 1: silu, 2: swiglu
+static constexpr ck::index_t ActOP       = 2; // 0: gelu, 1: silu, 2: swiglu
 // using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_Xdl_CShuffle_V3
 using DeviceOpInstance = ck::tensor_operation::device::DeviceMoeGemm
     // clang-format off
@@ -154,7 +154,7 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceMoeGemm
                //    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
                //    MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
                 //  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
-                1,    1,   S<1, 32, 1, 8>, S<EVec, D0Vec, D1Vec>,
+                2,    2,   S<1, 32, 1, 8>, S<EVec, D0Vec, D1Vec>,
                ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, ActOP, Nswizzle, true, true, int32_t, A0DataType>;
 
 // clang-format on
@@ -169,8 +169,8 @@ int main(int argc, char* argv[])
     ck::index_t N               = 4096;
     ck::index_t K               = 6144;
     ck::index_t experts         = 8;
-    ck::index_t sorted_tile_num = 8;
-    ck::index_t valid_tile_num  = 8;
+    ck::index_t sorted_tile_num = 16;
+    ck::index_t valid_tile_num  = 13;
     ck::index_t tokens          = 64;
     ck::index_t topk            = 2;
 
@@ -232,9 +232,9 @@ int main(int argc, char* argv[])
     // max_token_id.mData =  {valid_size, 2, 2, 1, 1, 2, 2, 2,2, 2, 2, 2, 2,1,0,0,0};
     // 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, 6, 7, 3, 3, 3}; // {2, 1, 1, 2, 2, 2, 1, 2}
-    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, 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}
+    max_token_id.mData = {valid_size};
+    int eids[] = {0, 0, 1, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 3, 3, 3}; // {2, 1, 1, 2, 2, 2, 1, 2}
     // max_token_id.mData = {valid_size};
     for(int i = 0; i < sorted_tile_num; i++)
     {
@@ -285,9 +285,9 @@ int main(int argc, char* argv[])
         d1_e_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
         break;
     case 2:
-        a0_t_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
-        b0_e_n_k.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
-        d0_t_n.GenerateTensorValue(GeneratorTensor_1<D0DataType>{});
+        a0_t_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
+        b0_e_n_k.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
+        d0_t_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0, 1});
         d1_e_n.GenerateTensorValue(GeneratorTensor_1<D1DataType>{});
         break;
     case 3:
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 4c7b438bea..5a183321cd 100644
--- a/include/ck/tensor_operation/gpu/grid/gridwise_moe_gemm.hpp
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_moe_gemm.hpp
@@ -1421,7 +1421,7 @@ struct GridwiseMoeGemm
             vector_type<int32_t, 4> scale_token_ids;
             vector_type<float, 4> topk_weights; // for gemm2 only
             static_for<0, NXdlPerWave, 1>{}([&](auto n0) {
-                const float scale_b = p_scale_b[n0 * NWave * PerTokenQuant];
+                const float scale_b = p_scale_b[n0 * NWave * NPerXdl * PerTokenQuant];
                 static_for<0, MXdlPerWave, 1>{}([&](auto m0) { // MXDLPerWave
                     static_for<0, M2, 1>{}([&](auto m2) {      // m_inst_num_groups_per_blk
                         const index_t m_pos = block_m_id * MPerBlock + m0 * M1 * M2 * M3 * M4 +
@@ -1471,7 +1471,8 @@ struct GridwiseMoeGemm
                                 else if(ActivationOperation == Activation::swiglu)
                                 {
                                     const float scale_up =
-                                        p_scale_b[(n0 * NPerXdl + problem.N) * PerTokenQuant];
+                                        p_scale_b[(n0 * NWave * NPerXdl + problem.N) *
+                                                  PerTokenQuant];
                                     auto gate          = scale_a * scale_b * c_thread_buf[cidx];
                                     auto up            = scale_a * scale_up * c_thread_buf_up[cidx];
                                     gate               = gate * math::rcp(1.0 + math::exp(-gate));