diff --git a/example/ck_tile/36_quant/example_dynamic_per_tensor_quant.cpp b/example/ck_tile/36_quant/example_dynamic_per_tensor_quant.cpp
index 06a9247e31..15d00995c5 100644
--- a/example/ck_tile/36_quant/example_dynamic_per_tensor_quant.cpp
+++ b/example/ck_tile/36_quant/example_dynamic_per_tensor_quant.cpp
@@ -36,20 +36,20 @@ auto get_elimit<ck_tile::int8_t>()
 int main()
 {
 
-    static constexpr ck_tile::index_t Repeat_M_ = 1;
+    static constexpr ck_tile::index_t Repeat_M_ = 8;
     static constexpr ck_tile::index_t Repeat_N_ = 1;
     
-    static constexpr ck_tile::index_t ThreadPerBlock_M_ = 4;
+    static constexpr ck_tile::index_t ThreadPerBlock_M_ = 8;
     static constexpr ck_tile::index_t ThreadPerBlock_N_ = 64;
 
-    static constexpr ck_tile::index_t Vector_N_ = 2;
+    static constexpr ck_tile::index_t Vector_N_ = 1;
     
     static constexpr bool is_warp_per_row = ThreadPerBlock_N_ <= warpSize;
     static_assert((ThreadPerBlock_M_ * ThreadPerBlock_N_) % warpSize == 0);
     static constexpr ck_tile::index_t total_warps =
         (ThreadPerBlock_M_ * ThreadPerBlock_N_) / warpSize;
-
-    // num of warps along mreference_static_per_tensor_quantization2d
+    // std::cout<<"total_warps: "<<total_warps<<std::endl;
+    // num of warps along m
     static constexpr ck_tile::index_t BlockWarps_M = []() {
         if constexpr(is_warp_per_row)
         {
@@ -107,9 +107,9 @@ int main()
 
     using Kernel = ck_tile::PerTensorQuant<Pipeline>;
 
-    int m = 256;
-    int n = 256;
-    int x_stride = 256;
+    int m = 64;
+    int n = 64;
+    int x_stride = 64;
     ck_tile::HostTensor<XDataType> x_host({m, n}, {x_stride, 1});
     ck_tile::HostTensor<ScaleDataType> scale_host({1}, {1});
 
@@ -144,7 +144,7 @@ int main()
     scale_buf.FromDevice(scale_host.data());
     ck_tile::reference_per_tensor_quantization2d<XDataType, ScaleDataType, QXDataType>(
         x_host, scale_host, qx_host_ref);
-
+    // std::cout<<scale_host(0)<<std::endl;
     qx_buf.FromDevice(qx_host_dev.data());
     
     auto [rtol, atol] = get_elimit<QXDataType>();
diff --git a/include/ck_tile/ops/quant/pipeline/quant_pipeline.hpp b/include/ck_tile/ops/quant/pipeline/quant_pipeline.hpp
index fac15254ff..756282623a 100644
--- a/include/ck_tile/ops/quant/pipeline/quant_pipeline.hpp
+++ b/include/ck_tile/ops/quant/pipeline/quant_pipeline.hpp
@@ -9,7 +9,6 @@
 #include <type_traits>
 
 namespace ck_tile {
-
 template <typename Problem_, typename Policy_ = PerTensorQuantPipelineDefaultPolicy>
 struct StaticPerTensorQuantPipeline
 {
@@ -92,7 +91,7 @@ struct DynamicPerTensorQuantPipeline
     {
         auto x_window =
             make_tile_window(x_window_, Policy::template MakeXBlockTileDistribution<Problem>());
-
+        auto origin = x_window.get_window_origin();
         static constexpr index_t Block_N = Problem::BlockShape::Block_N;
         index_t num_n_tile_iteration =
             __builtin_amdgcn_readfirstlane(integer_divide_ceil(row_size, Block_N));
@@ -130,12 +129,8 @@ struct DynamicPerTensorQuantPipeline
         block_reduce2d_cross_warp_sync(absmax, smem, reduce_max_func);
         *scale = absmax.get_thread_buffer()[0] / ck_tile::numeric<QXDataType>::max();
 
-        ck_tile::index_t stride_to_right_most_window =
-            row_size % Block_N == 0 ? row_size - Block_N : row_size - row_size % Block_N;
-        move_tile_window(x_window, {0, -Block_N});
-        move_tile_window(qx_window, {0, stride_to_right_most_window});
-        for(int iN = __builtin_amdgcn_readfirstlane(0); iN < num_n_tile_iteration; ++iN)
-        {
+        x_window.set_window_origin(origin);
+        for(int iN = __builtin_amdgcn_readfirstlane(0); iN < num_n_tile_iteration; ++iN){
             const auto x       = load_tile(x_window);
             const auto qx      = tile_elementwise_in(
                 [&](const auto& a) {
@@ -143,8 +138,8 @@ struct DynamicPerTensorQuantPipeline
                 },
                 x);
             store_tile(qx_window, qx);
-            move_tile_window(x_window, {0, -Block_N});
-            move_tile_window(qx_window, {0, -Block_N});
+            move_tile_window(x_window, {0, Block_N});
+            move_tile_window(qx_window, {0, Block_N});
         }
     }
 };