diff --git a/example/ck_tile/38_block_scale_gemm/CMakeLists.txt b/example/ck_tile/38_block_scale_gemm/CMakeLists.txt
index d9a6c2bac0..7358d4d749 100644
--- a/example/ck_tile/38_block_scale_gemm/CMakeLists.txt
+++ b/example/ck_tile/38_block_scale_gemm/CMakeLists.txt
@@ -8,12 +8,6 @@ list(APPEND EXAMPLE_GEMM_COMPILE_OPTIONS -mllvm -enable-noalias-to-md-conversion
 if(GPU_TARGETS MATCHES "gfx94" OR GPU_TARGETS MATCHES "gfx95")
     add_executable(tile_example_gemm_quant_basic EXCLUDE_FROM_ALL gemm_quant_basic.cpp)
     target_compile_options(tile_example_gemm_quant_basic PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
-
-    add_executable(tile_example_gemm_aquant_preshuffle EXCLUDE_FROM_ALL gemm_aquant_preshuffle.cpp)
-    target_compile_options(tile_example_gemm_aquant_preshuffle PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
-
-    add_executable(tile_example_gemm_bquant_basic EXCLUDE_FROM_ALL gemm_bquant_basic.cpp)
-    target_compile_options(tile_example_gemm_bquant_basic PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
 else()
     message(DEBUG "Skipping ck_tile quant gemm tests for current target")
 endif()
diff --git a/example/ck_tile/38_block_scale_gemm/README.md b/example/ck_tile/38_block_scale_gemm/README.md
index 36f7995c14..9acc4f9bfc 100644
--- a/example/ck_tile/38_block_scale_gemm/README.md
+++ b/example/ck_tile/38_block_scale_gemm/README.md
@@ -3,6 +3,7 @@
 This folder contains examples of quant GEMMs using the ck_tile tile-programming implementation.
 
 - AQuant kernel with blocks of A matrix sharing scales: custom GEMM pipeline
+- BQuant kernel with blocks of B matrix sharing scales: custom GEMM pipeline
 - Row and Column-wise scaled: scaling implemented in Epilogue
 
 ## build
@@ -38,3 +39,13 @@ args:
       -timer    gpu:gpu timer, cpu:cpu timer (default:gpu)
  -quant_mode    Which quant method to use (aquant, rowcol)
 ```
+
+User need to select correct mapping of config for each quant mode:
+
+|  | quant_mode as runtime argument | Config in cpp file |
+|:--------|:-----:|-------|
+| For selecting AQuant  | aquant  | GemmConfigQuant    |
+| For selecting Aquant with Preshuffle   | aquant  | GemmConfigPreshuffleQuant    |
+| For selecting BQuant  | bquant  | GemmConfigQuant    |
+| For selecting RowCol quant  | rowcolquant  | GemmConfigRowColQuant    |
+
diff --git a/example/ck_tile/38_block_scale_gemm/gemm_aquant_preshuffle.cpp b/example/ck_tile/38_block_scale_gemm/gemm_aquant_preshuffle.cpp
deleted file mode 100644
index 799e4a0e73..0000000000
--- a/example/ck_tile/38_block_scale_gemm/gemm_aquant_preshuffle.cpp
+++ /dev/null
@@ -1,234 +0,0 @@
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
-
-#include <cstring>
-#include <iostream>
-#include <ostream>
-#include <stdexcept>
-#include <string>
-#include <tuple>
-
-#include "gemm_utils.hpp"
-
-template <typename GemmConfig,
-          typename ADataType,
-          typename AQDataType,
-          typename BDataType,
-          typename AccDataType,
-          typename CDataType,
-          typename ComputeDataType,
-          typename ALayout,
-          typename BLayout,
-          typename CLayout,
-          uint32_t QuantGroupSize>
-float gemm_calc_aquant(const ck_tile::QuantGemmHostArgs& args, const ck_tile::stream_config& s)
-{
-    constexpr bool kPadM = false;
-    constexpr bool kPadN = false;
-    constexpr bool kPadK = false;
-
-    constexpr int kBlockPerCu = 1;
-
-    static_assert(std::is_same_v<CLayout, ck_tile::tensor_layout::gemm::RowMajor>);
-
-    constexpr ck_tile::index_t M_Tile = GemmConfig::M_Tile;
-    constexpr ck_tile::index_t N_Tile = GemmConfig::N_Tile;
-    constexpr ck_tile::index_t K_Tile = GemmConfig::K_Tile;
-
-    constexpr ck_tile::index_t M_Warp = GemmConfig::M_Warp;
-    constexpr ck_tile::index_t N_Warp = GemmConfig::N_Warp;
-    constexpr ck_tile::index_t K_Warp = GemmConfig::K_Warp;
-
-    constexpr ck_tile::index_t M_Warp_Tile = GemmConfig::M_Warp_Tile;
-    constexpr ck_tile::index_t N_Warp_Tile = GemmConfig::N_Warp_Tile;
-    constexpr ck_tile::index_t K_Warp_Tile = GemmConfig::K_Warp_Tile;
-
-    using CodegenGemmShape =
-        ck_tile::TileGemmShape<ck_tile::sequence<M_Tile, N_Tile, K_Tile>,
-                               ck_tile::sequence<M_Warp, N_Warp, K_Warp>,
-                               ck_tile::sequence<M_Warp_Tile, N_Warp_Tile, K_Warp_Tile>>;
-
-    using TilePartitioner = ck_tile::GemmTile1DPartitioner<CodegenGemmShape>;
-
-    using CodegenGemmTraits = ck_tile::TileGemmQuantTraits<kPadM,
-                                                           kPadN,
-                                                           kPadK,
-                                                           GemmConfig::PreshuffleQuant,
-                                                           ALayout,
-                                                           BLayout,
-                                                           CLayout,
-                                                           ck_tile::QuantType::AQuantGrouped>;
-
-    using GemmPipelineProblem = ck_tile::GemmPipelineProblemBase<ADataType,
-                                                                 BDataType,
-                                                                 AccDataType,
-                                                                 CodegenGemmShape,
-                                                                 CodegenGemmTraits,
-                                                                 ComputeDataType>;
-
-    using BaseGemmPipeline = ck_tile::BaseAQuantGemmPipelineAgBgCrCompV3<GemmPipelineProblem>;
-
-    const ck_tile::index_t K_split      = (args.K + K_Tile - 1) / K_Tile * K_Tile;
-    const ck_tile::index_t num_loop     = TilePartitioner::GetLoopNum(K_split);
-    const bool has_hot_loop             = BaseGemmPipeline::BlockHasHotloop(num_loop);
-    const ck_tile::TailNumber tail_num  = BaseGemmPipeline::GetBlockLoopTailNum(num_loop);
-    constexpr bool transposed_warp_gemm = false;
-
-    const auto Run = [&](const auto has_hot_loop_, const auto tail_number_) {
-        constexpr bool has_hot_loop_v = has_hot_loop_.value;
-        constexpr auto tail_number_v  = tail_number_.value;
-
-        using CodegenPipelineProblem =
-            ck_tile::GemmAQuantPipelineProblem<ADataType,
-                                               AQDataType,
-                                               BDataType,
-                                               AccDataType,
-                                               CodegenGemmShape,
-                                               CodegenGemmTraits,
-                                               QuantGroupSize,
-                                               transposed_warp_gemm,
-                                               ComputeDataType,
-                                               ck_tile::GemmPipelineScheduler::Intrawave,
-                                               has_hot_loop_v,
-                                               tail_number_v>;
-        using CodegenGemmPipeline = ck_tile::AQuantGemmPipelineAgBgCrCompV3<CodegenPipelineProblem>;
-        using GemmEpilogue        = ck_tile::CShuffleEpilogue<
-                   ck_tile::CShuffleEpilogueProblem<ADataType,
-                                                    BDataType,
-                                                    ck_tile::tuple<>,
-                                                    AccDataType,
-                                                    CDataType,
-                                                    ck_tile::tuple<>,
-                                                    CLayout,
-                                                    ck_tile::element_wise::PassThrough,
-                                                    TilePartitioner::MPerBlock,
-                                                    TilePartitioner::NPerBlock,
-                                                    M_Warp,
-                                                    N_Warp,
-                                                    M_Warp_Tile,
-                                                    N_Warp_Tile,
-                                                    K_Warp_Tile,
-                                                    transposed_warp_gemm,
-                                                    ck_tile::memory_operation_enum::set>>;
-        using Kernel = ck_tile::QuantGemmKernel<TilePartitioner,
-                                                CodegenGemmPipeline,
-                                                GemmEpilogue,
-                                                ck_tile::QuantType::AQuantGrouped>;
-
-        auto kargs = Kernel::MakeKernelArgs(args);
-
-        const dim3 grids  = Kernel::GridSize(args.M, args.N, args.k_batch);
-        const dim3 blocks = Kernel::BlockSize();
-
-        if(args.k_batch != 1)
-        {
-            throw std::runtime_error("split-k is not supported yet!");
-        }
-
-        if(!Kernel::IsSupportedArgument(kargs))
-        {
-            throw std::runtime_error("Wrong! Arguments not supported! Skipping gemm!\n");
-        }
-
-        if(s.log_level_ > 0)
-        {
-            std::cout << "Launching kernel with args: " << Kernel::GetName() << '\n'
-                      << "shape: " << CodegenGemmShape::GetName() << '\n'
-                      << "problem: " << CodegenPipelineProblem::GetName() << '\n'
-                      << "pipeline: " << CodegenGemmPipeline::GetName() << '\n'
-                      << "grid: {" << grids.x << ", " << grids.y << ", " << grids.z << "}"
-                      << ", blocks: {" << blocks.x << ", " << blocks.y << ", " << blocks.z << "}"
-                      << std::endl;
-        }
-
-        float ave_time = ck_tile::launch_kernel(
-            s, ck_tile::make_kernel<kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
-
-        return ave_time;
-    };
-    return BaseGemmPipeline::TailHandler(Run, has_hot_loop, tail_num);
-}
-
-#include "run_gemm_aquant_example.inc"
-
-template <typename GemmConfig, typename TypeConfig, uint32_t QuantGroupSize>
-int run_gemm_example_prec_type(std::string a_layout, std::string b_layout, int argc, char* argv[])
-{
-    using Row = ck_tile::tensor_layout::gemm::RowMajor;
-    using Col = ck_tile::tensor_layout::gemm::ColumnMajor;
-
-    if constexpr(std::is_same_v<typename TypeConfig::ADataType, ck_tile::pk_int4_t> ||
-                 std::is_same_v<typename TypeConfig::ADataType, ck_tile::fp8_t> ||
-                 std::is_same_v<typename TypeConfig::ADataType, ck_tile::bf8_t>)
-    {
-        if(a_layout == "R" && b_layout == "C")
-        {
-            return run_gemm_example_with_layouts<GemmConfig, TypeConfig, QuantGroupSize>(
-                argc, argv, Row{}, Row{}, Col{}, Row{});
-        }
-        else
-        {
-            throw std::runtime_error("Unsupported memory layout for the input matrices!");
-        }
-    }
-    else
-    {
-        throw std::runtime_error("Unsupported data type for A.");
-    }
-
-    return 0;
-}
-
-template <template <typename PreType> typename GemmConfig>
-int run_gemm_example(int argc, char* argv[])
-{
-    auto [result, arg_parser] = create_args(argc, argv);
-    if(!result)
-        return -1;
-
-    std::string data_type = arg_parser.get_str("prec");
-    std::string a_layout  = arg_parser.get_str("a_layout");
-    std::string b_layout  = arg_parser.get_str("b_layout");
-
-    if(data_type == "fp8")
-    {
-        using TypeConfig =
-            decltype(GemmQuantTypeConfig<ck_tile::fp8_t, ck_tile::fp8_t, ck_tile::half_t, float>{});
-        return run_gemm_example_prec_type<GemmConfig<ck_tile::fp8_t>, TypeConfig, 128>(
-            a_layout, b_layout, argc, argv);
-    }
-    else if(data_type == "bf8")
-    {
-        using TypeConfig =
-            decltype(GemmQuantTypeConfig<ck_tile::bf8_t, ck_tile::bf8_t, ck_tile::half_t, float>{});
-        return run_gemm_example_prec_type<GemmConfig<ck_tile::bf8_t>, TypeConfig, 128>(
-            a_layout, b_layout, argc, argv);
-    }
-    else if(data_type == "i4fp8")
-    {
-        using TypeConfig = decltype(GemmQuantTypeConfig<ck_tile::pk_int4_t,
-                                                        ck_tile::fp8_t,
-                                                        ck_tile::half_t,
-                                                        ck_tile::fp8_t>{});
-        return run_gemm_example_prec_type<GemmConfig<ck_tile::pk_int4_t>, TypeConfig, 128>(
-            a_layout, b_layout, argc, argv);
-    }
-    else if(data_type == "i4bf8")
-    {
-        using TypeConfig = decltype(GemmQuantTypeConfig<ck_tile::pk_int4_t,
-                                                        ck_tile::bf8_t,
-                                                        ck_tile::half_t,
-                                                        ck_tile::bf8_t>{});
-        return run_gemm_example_prec_type<GemmConfig<ck_tile::pk_int4_t>, TypeConfig, 128>(
-            a_layout, b_layout, argc, argv);
-    }
-    else
-    {
-        throw std::runtime_error("Unsupported data type for this operation !!!");
-    }
-}
-
-int main(int argc, char* argv[])
-{
-    return !run_gemm_example<GemmConfigPreshuffleQuant>(argc, argv);
-}
diff --git a/example/ck_tile/38_block_scale_gemm/gemm_bquant_basic.cpp b/example/ck_tile/38_block_scale_gemm/gemm_bquant_basic.cpp
deleted file mode 100644
index 52a815dfdb..0000000000
--- a/example/ck_tile/38_block_scale_gemm/gemm_bquant_basic.cpp
+++ /dev/null
@@ -1,231 +0,0 @@
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
-
-#include <cstring>
-#include <iostream>
-#include <ostream>
-#include <stdexcept>
-#include <string>
-#include <tuple>
-
-#include "ck_tile/core/config.hpp"
-#include "ck_tile/host.hpp"
-#include "gemm_utils.hpp"
-
-template <typename GemmConfig,
-          typename ADataType,
-          typename BDataType,
-          typename BQDataType,
-          typename AccDataType,
-          typename CDataType,
-          typename ComputeDataType,
-          typename ALayout,
-          typename BLayout,
-          typename CLayout,
-          uint32_t QuantGroupSize>
-float gemm_calc_bquant(const ck_tile::QuantGemmHostArgs& args, const ck_tile::stream_config& s)
-{
-    constexpr bool kPadM = false;
-    constexpr bool kPadN = false;
-    constexpr bool kPadK = false;
-
-    static_assert(std::is_same_v<CLayout, ck_tile::tensor_layout::gemm::RowMajor>);
-
-    constexpr ck_tile::index_t M_Tile = GemmConfig::M_Tile;
-    constexpr ck_tile::index_t N_Tile = GemmConfig::N_Tile;
-    constexpr ck_tile::index_t K_Tile = GemmConfig::K_Tile;
-
-    constexpr ck_tile::index_t M_Warp = GemmConfig::M_Warp;
-    constexpr ck_tile::index_t N_Warp = GemmConfig::N_Warp;
-    constexpr ck_tile::index_t K_Warp = GemmConfig::K_Warp;
-
-    constexpr ck_tile::index_t M_Warp_Tile = GemmConfig::M_Warp_Tile;
-    constexpr ck_tile::index_t N_Warp_Tile = GemmConfig::N_Warp_Tile;
-    constexpr ck_tile::index_t K_Warp_Tile = GemmConfig::K_Warp_Tile;
-
-    using CodegenGemmShape =
-        ck_tile::TileGemmShape<ck_tile::sequence<M_Tile, N_Tile, K_Tile>,
-                               ck_tile::sequence<M_Warp, N_Warp, K_Warp>,
-                               ck_tile::sequence<M_Warp_Tile, N_Warp_Tile, K_Warp_Tile>>;
-
-    using TilePartitioner = ck_tile::GemmTile1DPartitioner<CodegenGemmShape>;
-
-    using CodegenGemmTraits = ck_tile::TileGemmQuantTraits<kPadM,
-                                                           kPadN,
-                                                           kPadK,
-                                                           GemmConfig::PreshuffleQuant,
-                                                           ALayout,
-                                                           BLayout,
-                                                           CLayout,
-                                                           ck_tile::QuantType::BQuantGrouped>;
-
-    using GemmPipelineProblem = ck_tile::GemmPipelineProblemBase<ADataType,
-                                                                 BDataType,
-                                                                 AccDataType,
-                                                                 CodegenGemmShape,
-                                                                 CodegenGemmTraits,
-                                                                 ComputeDataType>;
-
-    using BaseGemmPipeline = ck_tile::BaseBQuantGemmPipelineAgBgCrCompV3<GemmPipelineProblem>;
-
-    const ck_tile::index_t K_split      = (args.K + K_Tile - 1) / K_Tile * K_Tile;
-    const ck_tile::index_t num_loop     = TilePartitioner::GetLoopNum(K_split);
-    const bool has_hot_loop             = BaseGemmPipeline::BlockHasHotloop(num_loop);
-    const ck_tile::TailNumber tail_num  = BaseGemmPipeline::GetBlockLoopTailNum(num_loop);
-    constexpr bool transposed_warp_gemm = false;
-
-    const auto Run = [&](const auto has_hot_loop_, const auto tail_number_) {
-        constexpr bool has_hot_loop_v = has_hot_loop_.value;
-        constexpr auto tail_number_v  = tail_number_.value;
-
-        using CodegenPipelineProblem =
-            ck_tile::GemmBQuantPipelineProblem<ADataType,
-                                               BDataType,
-                                               BQDataType,
-                                               AccDataType,
-                                               CodegenGemmShape,
-                                               CodegenGemmTraits,
-                                               QuantGroupSize,
-                                               ComputeDataType,
-                                               ck_tile::GemmPipelineScheduler::Intrawave,
-                                               has_hot_loop_v,
-                                               tail_number_v>;
-        using CodegenGemmPipeline = ck_tile::BQuantGemmPipelineAgBgCrCompV3<CodegenPipelineProblem>;
-        using GemmEpilogue        = ck_tile::CShuffleEpilogue<
-                   ck_tile::CShuffleEpilogueProblem<ADataType,
-                                                    BDataType,
-                                                    ck_tile::tuple<>,
-                                                    AccDataType,
-                                                    CDataType,
-                                                    ck_tile::tuple<>,
-                                                    CLayout,
-                                                    ck_tile::element_wise::PassThrough,
-                                                    TilePartitioner::MPerBlock,
-                                                    TilePartitioner::NPerBlock,
-                                                    M_Warp,
-                                                    N_Warp,
-                                                    M_Warp_Tile,
-                                                    N_Warp_Tile,
-                                                    K_Warp_Tile,
-                                                    transposed_warp_gemm,
-                                                    ck_tile::memory_operation_enum::set>>;
-        using Kernel = ck_tile::QuantGemmKernel<TilePartitioner,
-                                                CodegenGemmPipeline,
-                                                GemmEpilogue,
-                                                ck_tile::QuantType::BQuantGrouped>;
-
-        auto kargs = Kernel::MakeKernelArgs(args);
-
-        const dim3 grids  = Kernel::GridSize(args.M, args.N, args.k_batch);
-        const dim3 blocks = Kernel::BlockSize();
-
-        if(args.k_batch != 1)
-        {
-            throw std::runtime_error("split-k is not supported yet!");
-        }
-
-        if(!Kernel::IsSupportedArgument(kargs))
-        {
-            throw std::runtime_error("Wrong! Arguments not supported! Skipping gemm!\n");
-        }
-
-        if(s.log_level_ > 0)
-        {
-            std::cout << "Launching kernel with args: " << Kernel::GetName() << '\n'
-                      << "shape: " << CodegenGemmShape::GetName() << '\n'
-                      << "problem: " << CodegenPipelineProblem::GetName() << '\n'
-                      << "pipeline: " << CodegenGemmPipeline::GetName() << '\n'
-                      << "grid: {" << grids.x << ", " << grids.y << ", " << grids.z << "}"
-                      << ", blocks: {" << blocks.x << ", " << blocks.y << ", " << blocks.z << "}"
-                      << std::endl;
-        }
-
-        float ave_time = ck_tile::launch_kernel(
-            s, ck_tile::make_kernel<GemmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
-
-        return ave_time;
-    };
-    return BaseGemmPipeline::TailHandler(Run, has_hot_loop, tail_num);
-    ;
-}
-
-#include "run_gemm_bquant_example.inc"
-
-template <typename GemmConfig, typename TypeConfig, uint32_t QuantGroupSize>
-int run_gemm_example_prec_type(std::string a_layout, std::string b_layout, int argc, char* argv[])
-{
-    using Row = ck_tile::tensor_layout::gemm::RowMajor;
-    using Col = ck_tile::tensor_layout::gemm::ColumnMajor;
-
-    if constexpr(std::is_same_v<typename TypeConfig::BDataType, ck_tile::pk_int4_t> ||
-                 std::is_same_v<typename TypeConfig::BDataType, ck_tile::fp8_t> ||
-                 std::is_same_v<typename TypeConfig::BDataType, ck_tile::bf8_t>)
-    {
-        if(a_layout == "R" && b_layout == "C")
-        {
-            return run_gemm_example_with_layouts<GemmConfig, TypeConfig, QuantGroupSize>(
-                argc, argv, Row{}, Col{}, Col{}, Row{});
-        }
-        else
-        {
-            throw std::runtime_error("Unsupported memory layout for the input matrices!");
-        }
-    }
-    else
-    {
-        throw std::runtime_error("Unsupported data type for B.");
-    }
-
-    return 0;
-}
-
-template <template <typename PreType> typename GemmConfig>
-int run_gemm_example(int argc, char* argv[])
-{
-    auto [result, arg_parser] = create_args(argc, argv);
-    if(!result)
-        return -1;
-
-    std::string data_type = arg_parser.get_str("prec");
-    std::string a_layout  = arg_parser.get_str("a_layout");
-    std::string b_layout  = arg_parser.get_str("b_layout");
-
-    if(data_type == "fp8")
-    {
-        using TypeConfig =
-            decltype(GemmQuantTypeConfig<ck_tile::fp8_t, ck_tile::fp8_t, ck_tile::half_t, float>{});
-        return run_gemm_example_prec_type<GemmConfig<ck_tile::fp8_t>, TypeConfig, 128>(
-            a_layout, b_layout, argc, argv);
-    }
-    else if(data_type == "bf8")
-    {
-        using TypeConfig =
-            decltype(GemmQuantTypeConfig<ck_tile::bf8_t, ck_tile::bf8_t, ck_tile::half_t, float>{});
-        return run_gemm_example_prec_type<GemmConfig<ck_tile::bf8_t>, TypeConfig, 128>(
-            a_layout, b_layout, argc, argv);
-    }
-    else if(data_type == "fp8i4")
-    {
-        using TypeConfig = decltype(GemmQuantTypeConfig<ck_tile::fp8_t,
-                                                        ck_tile::pk_int4_t,
-                                                        ck_tile::half_t,
-                                                        ck_tile::fp8_t>{});
-        return run_gemm_example_prec_type<GemmConfig<ck_tile::fp8_t>, TypeConfig, 128>(
-            a_layout, b_layout, argc, argv);
-    }
-    else if(data_type == "bf8i4")
-    {
-        using TypeConfig = decltype(GemmQuantTypeConfig<ck_tile::bf8_t,
-                                                        ck_tile::pk_int4_t,
-                                                        ck_tile::half_t,
-                                                        ck_tile::bf8_t>{});
-        return run_gemm_example_prec_type<GemmConfig<ck_tile::bf8_t>, TypeConfig, 128>(
-            a_layout, b_layout, argc, argv);
-    }
-    else
-    {
-        throw std::runtime_error("Unsupported data type for this operation !!!");
-    }
-}
-
-int main(int argc, char* argv[]) { return !run_gemm_example<GemmConfigQuant>(argc, argv); }
diff --git a/example/ck_tile/38_block_scale_gemm/gemm_quant_basic.cpp b/example/ck_tile/38_block_scale_gemm/gemm_quant_basic.cpp
index 5345686e20..35ffcf1d56 100644
--- a/example/ck_tile/38_block_scale_gemm/gemm_quant_basic.cpp
+++ b/example/ck_tile/38_block_scale_gemm/gemm_quant_basic.cpp
@@ -18,58 +18,45 @@ template <typename GemmConfig,
           typename BLayout,
           typename CLayout,
           uint32_t QuantGroupSize,
-          ck_tile::QuantType QuantMode>
+          ck_tile::QuantType QuantMode,
+          typename CDEElementWise>
 float gemm_calc_quant(const ck_tile::QuantGemmHostArgs& args, const ck_tile::stream_config& s)
 {
-    constexpr bool kPadM = GemmConfig::kPadM;
-    constexpr bool kPadN = GemmConfig::kPadN;
-    constexpr bool kPadK = GemmConfig::kPadK;
-
-    constexpr int kBlockPerCu = GemmConfig::kBlockPerCu;
-
     static_assert(std::is_same_v<CLayout, ck_tile::tensor_layout::gemm::RowMajor>);
-
-    constexpr ck_tile::index_t M_Tile = GemmConfig::M_Tile;
-    constexpr ck_tile::index_t N_Tile = GemmConfig::N_Tile;
-    constexpr ck_tile::index_t K_Tile = GemmConfig::K_Tile;
-
-    constexpr ck_tile::index_t M_Warp = GemmConfig::M_Warp;
-    constexpr ck_tile::index_t N_Warp = GemmConfig::N_Warp;
-    constexpr ck_tile::index_t K_Warp = GemmConfig::K_Warp;
-
-    constexpr ck_tile::index_t M_Warp_Tile = GemmConfig::M_Warp_Tile;
-    constexpr ck_tile::index_t N_Warp_Tile = GemmConfig::N_Warp_Tile;
-    constexpr ck_tile::index_t K_Warp_Tile = GemmConfig::K_Warp_Tile;
-
     // B datatype is safe to use as compute type as it should be at least fp8
-    using ComputeDataType = typename TypeConfig::BDataType;
+    using ComputeDataType = std::conditional_t<QuantMode == ck_tile::QuantType::AQuantGrouped ||
+                                                   QuantMode == ck_tile::QuantType::RowColQuant,
+                                               typename TypeConfig::BDataType,
+                                               typename TypeConfig::ADataType>;
 
-    using CodegenGemmShape =
-        ck_tile::TileGemmShape<ck_tile::sequence<M_Tile, N_Tile, K_Tile>,
-                               ck_tile::sequence<M_Warp, N_Warp, K_Warp>,
-                               ck_tile::sequence<M_Warp_Tile, N_Warp_Tile, K_Warp_Tile>>;
+    using GemmShape = ck_tile::TileGemmShape<
+        ck_tile::sequence<GemmConfig::M_Tile, GemmConfig::N_Tile, GemmConfig::K_Tile>,
+        ck_tile::sequence<GemmConfig::M_Warp, GemmConfig::N_Warp, GemmConfig::K_Warp>,
+        ck_tile::
+            sequence<GemmConfig::M_Warp_Tile, GemmConfig::N_Warp_Tile, GemmConfig::K_Warp_Tile>>;
 
-    using TilePartitioner = ck_tile::GemmTile1DPartitioner<CodegenGemmShape>;
+    using TilePartitioner = ck_tile::GemmTile1DPartitioner<GemmShape>;
 
-    using CodegenGemmTraits = ck_tile::TileGemmQuantTraits<kPadM,
-                                                           kPadN,
-                                                           kPadK,
-                                                           GemmConfig::PreshuffleQuant,
-                                                           ALayout,
-                                                           BLayout,
-                                                           CLayout,
-                                                           QuantMode>;
+    using GemmTraits = ck_tile::TileGemmQuantTraits<GemmConfig::kPadM,
+                                                    GemmConfig::kPadN,
+                                                    GemmConfig::kPadK,
+                                                    GemmConfig::PreshuffleQuant,
+                                                    ALayout,
+                                                    BLayout,
+                                                    CLayout,
+                                                    QuantMode>;
 
     using GemmPipelineProblem = ck_tile::GemmPipelineProblemBase<typename TypeConfig::ADataType,
                                                                  typename TypeConfig::BDataType,
                                                                  typename TypeConfig::AccDataType,
-                                                                 CodegenGemmShape,
-                                                                 CodegenGemmTraits,
+                                                                 GemmShape,
+                                                                 GemmTraits,
                                                                  ComputeDataType>;
 
     using BaseGemmPipeline = ck_tile::BaseGemmPipelineAgBgCrCompV3<GemmPipelineProblem>;
 
-    const ck_tile::index_t K_split      = (args.K + K_Tile - 1) / K_Tile * K_Tile;
+    const ck_tile::index_t K_split =
+        (args.K + GemmConfig::K_Tile - 1) / GemmConfig::K_Tile * GemmConfig::K_Tile;
     const ck_tile::index_t num_loop     = TilePartitioner::GetLoopNum(K_split);
     const bool has_hot_loop             = BaseGemmPipeline::BlockHasHotloop(num_loop);
     const ck_tile::TailNumber tail_num  = BaseGemmPipeline::GetBlockLoopTailNum(num_loop);
@@ -80,36 +67,50 @@ float gemm_calc_quant(const ck_tile::QuantGemmHostArgs& args, const ck_tile::str
         constexpr auto tail_number_v  = tail_number_.value;
         constexpr bool transpose_c    = false;
 
-        using CodegenPipelineProblem = std::conditional_t<
-            QuantMode == ck_tile::QuantType::AQuantGrouped,
-            ck_tile::GemmAQuantPipelineProblem<typename TypeConfig::ADataType,
-                                               typename TypeConfig::QDataType,
-                                               typename TypeConfig::BDataType,
-                                               typename TypeConfig::AccDataType,
-                                               CodegenGemmShape,
-                                               CodegenGemmTraits,
-                                               QuantGroupSize,
-                                               transpose_c,
-                                               ComputeDataType,
-                                               ck_tile::GemmPipelineScheduler::Intrawave,
-                                               has_hot_loop_v,
-                                               tail_number_v>,
+        using PipelineProblem = std::conditional_t<
+            QuantMode == ck_tile::QuantType::RowColQuant,
             ck_tile::GemmRowColQuantPipelineProblem<typename TypeConfig::ADataType,
                                                     typename TypeConfig::BDataType,
                                                     typename TypeConfig::AccDataType,
                                                     typename TypeConfig::AccDataType,
-                                                    CodegenGemmShape,
-                                                    CodegenGemmTraits,
+                                                    GemmShape,
+                                                    GemmTraits,
                                                     transpose_c,
                                                     ComputeDataType,
-                                                    ck_tile::GemmPipelineScheduler::Intrawave,
+                                                    GemmConfig::Scheduler,
                                                     has_hot_loop_v,
-                                                    tail_number_v>>;
-
-        using CodegenGemmPipeline =
+                                                    tail_number_v>,
             std::conditional_t<QuantMode == ck_tile::QuantType::AQuantGrouped,
-                               ck_tile::AQuantGemmPipelineAgBgCrCompV3<CodegenPipelineProblem>,
-                               ck_tile::GemmPipelineAgBgCrCompV3<CodegenPipelineProblem>>;
+                               ck_tile::GemmAQuantPipelineProblem<typename TypeConfig::ADataType,
+                                                                  typename TypeConfig::QDataType,
+                                                                  typename TypeConfig::BDataType,
+                                                                  typename TypeConfig::AccDataType,
+                                                                  GemmShape,
+                                                                  GemmTraits,
+                                                                  QuantGroupSize,
+                                                                  transpose_c,
+                                                                  ComputeDataType,
+                                                                  GemmConfig::Scheduler,
+                                                                  has_hot_loop_v,
+                                                                  tail_number_v>,
+                               ck_tile::GemmBQuantPipelineProblem<typename TypeConfig::ADataType,
+                                                                  typename TypeConfig::BDataType,
+                                                                  typename TypeConfig::QDataType,
+                                                                  typename TypeConfig::AccDataType,
+                                                                  GemmShape,
+                                                                  GemmTraits,
+                                                                  QuantGroupSize,
+                                                                  ComputeDataType,
+                                                                  GemmConfig::Scheduler,
+                                                                  has_hot_loop_v,
+                                                                  tail_number_v>>>;
+
+        using GemmPipeline = std::conditional_t<
+            QuantMode == ck_tile::QuantType::RowColQuant,
+            ck_tile::GemmPipelineAgBgCrCompV3<PipelineProblem>,
+            std::conditional_t<QuantMode == ck_tile::QuantType::AQuantGrouped,
+                               ck_tile::AQuantGemmPipelineAgBgCrCompV3<PipelineProblem>,
+                               ck_tile::BQuantGemmPipelineAgBgCrCompV3<PipelineProblem>>>;
 
         using GemmEpilogue = ck_tile::CShuffleEpilogue<
             ck_tile::CShuffleEpilogueProblem<typename TypeConfig::ADataType,
@@ -119,18 +120,18 @@ float gemm_calc_quant(const ck_tile::QuantGemmHostArgs& args, const ck_tile::str
                                              typename TypeConfig::CDataType,
                                              ck_tile::tuple<>,
                                              CLayout,
-                                             ck_tile::element_wise::PassThrough,
+                                             CDEElementWise,
                                              TilePartitioner::MPerBlock,
                                              TilePartitioner::NPerBlock,
-                                             M_Warp,
-                                             N_Warp,
-                                             M_Warp_Tile,
-                                             N_Warp_Tile,
-                                             K_Warp_Tile,
+                                             GemmConfig::M_Warp,
+                                             GemmConfig::N_Warp,
+                                             GemmConfig::M_Warp_Tile,
+                                             GemmConfig::N_Warp_Tile,
+                                             GemmConfig::K_Warp_Tile,
                                              transposed_warp_gemm,
                                              ck_tile::memory_operation_enum::set>>;
         using Kernel =
-            ck_tile::QuantGemmKernel<TilePartitioner, CodegenGemmPipeline, GemmEpilogue, QuantMode>;
+            ck_tile::QuantGemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue, QuantMode>;
 
         auto kargs = Kernel::MakeKernelArgs(args);
 
@@ -150,16 +151,16 @@ float gemm_calc_quant(const ck_tile::QuantGemmHostArgs& args, const ck_tile::str
         if(s.log_level_ > 0)
         {
             std::cout << "Launching kernel with args: " << Kernel::GetName() << '\n'
-                      << "shape: " << CodegenGemmShape::GetName() << '\n'
-                      << "problem: " << CodegenPipelineProblem::GetName() << '\n'
-                      << "pipeline: " << CodegenGemmPipeline::GetName() << '\n'
+                      << "shape: " << GemmShape::GetName() << '\n'
+                      << "problem: " << PipelineProblem::GetName() << '\n'
+                      << "pipeline: " << GemmPipeline::GetName() << '\n'
                       << "grid: {" << grids.x << ", " << grids.y << ", " << grids.z << "}"
                       << ", blocks: {" << blocks.x << ", " << blocks.y << ", " << blocks.z << "}"
                       << std::endl;
         }
 
         float ave_time = ck_tile::launch_kernel(
-            s, ck_tile::make_kernel<kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
+            s, ck_tile::make_kernel<GemmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
 
         return ave_time;
     };
@@ -199,33 +200,9 @@ int run_gemm_example_prec_type(std::string a_layout, std::string b_layout, int a
     return 0;
 }
 
-void print_help(const char* program_name)
-{
-    std::cout
-        << "Usage: " << program_name << " [OPTIONS]\n"
-        << "\n Parameters:\n"
-        << "  -quant_mode=MODE     aquant (A quantization) or rowcol (row/column quantization)\n"
-        << "  -v=LEVEL              0=No validation, 1=CPU validation, 2=GPU validation\n"
-        << "  -prec=TYPE           Data types: fp8, bf8, i4fp8, i4bf8, i4f32fp8, i4f32bf8\n"
-        << "  -m, -n, -k=SIZE       Matrix dimensions (M×K) @ (K×N) = (M×N)\n"
-        << "\nExample:\n"
-        << "  " << program_name << " -quant_mode=rowcol -v=1 -prec=fp8\n"
-        << std::endl;
-}
-
 template <template <typename PreType> typename GemmConfig>
 int run_gemm_example(int argc, char* argv[])
 {
-
-    for(int i = 1; i < argc; ++i)
-    {
-        if(strcmp(argv[i], "--help") == 0 || strcmp(argv[i], "-h") == 0)
-        {
-            print_help(argv[0]);
-            return 0;
-        }
-    }
-
     auto [result, arg_parser] = create_args(argc, argv);
     if(!result)
         return -1;
@@ -239,7 +216,7 @@ int run_gemm_example(int argc, char* argv[])
     if(data_type == "fp8")
     {
         using TypeConfig =
-            decltype(GemmQuantTypeConfig<ck_tile::fp8_t, ck_tile::fp8_t, ck_tile::half_t>{});
+            decltype(GemmQuantTypeConfig<ck_tile::fp8_t, ck_tile::fp8_t, ck_tile::half_t, float>{});
 
         if(quant_mode == "aquant")
         {
@@ -249,6 +226,14 @@ int run_gemm_example(int argc, char* argv[])
                                               ck_tile::QuantType::AQuantGrouped>(
                 a_layout, b_layout, argc, argv);
         }
+        else if(quant_mode == "bquant")
+        {
+            return run_gemm_example_prec_type<GemmConfig<ck_tile::fp8_t>,
+                                              TypeConfig,
+                                              128,
+                                              ck_tile::QuantType::BQuantGrouped>(
+                a_layout, b_layout, argc, argv);
+        }
         else if(quant_mode == "rowcol")
         {
             return run_gemm_example_prec_type<GemmConfig<ck_tile::fp8_t>,
@@ -259,12 +244,14 @@ int run_gemm_example(int argc, char* argv[])
         }
         else
         {
-            throw std::runtime_error("Unsupported quantization mode! Use 'aquant' or 'rowcol'");
+            throw std::runtime_error(
+                "Unsupported quantization mode! Use 'aquant', 'bquant' or 'rowcol'");
         }
     }
     else if(data_type == "bf8")
     {
-        using TypeConfig = decltype(GemmQuantTypeConfig<ck_tile::bf8_t, ck_tile::bf8_t, float>{});
+        using TypeConfig =
+            decltype(GemmQuantTypeConfig<ck_tile::bf8_t, ck_tile::bf8_t, ck_tile::half_t, float>{});
 
         if(quant_mode == "aquant")
         {
@@ -274,6 +261,14 @@ int run_gemm_example(int argc, char* argv[])
                                               ck_tile::QuantType::AQuantGrouped>(
                 a_layout, b_layout, argc, argv);
         }
+        else if(quant_mode == "bquant")
+        {
+            return run_gemm_example_prec_type<GemmConfig<ck_tile::bf8_t>,
+                                              TypeConfig,
+                                              128,
+                                              ck_tile::QuantType::BQuantGrouped>(
+                a_layout, b_layout, argc, argv);
+        }
         else if(quant_mode == "rowcol")
         {
             return run_gemm_example_prec_type<GemmConfig<ck_tile::bf8_t>,
@@ -284,19 +279,20 @@ int run_gemm_example(int argc, char* argv[])
         }
         else
         {
-            throw std::runtime_error("Unsupported quantization mode! Use 'aquant' or 'rowcol'");
+            throw std::runtime_error(
+                "Unsupported quantization mode! Use 'aquant', 'bquant' or 'rowcol'");
         }
     }
     else if(data_type == "i4fp8")
     {
         using TypeConfig = decltype(GemmQuantTypeConfig<ck_tile::pk_int4_t,
                                                         ck_tile::fp8_t,
-                                                        float,
+                                                        ck_tile::half_t,
                                                         ck_tile::fp8_t>{});
 
         if(quant_mode == "aquant")
         {
-            return run_gemm_example_prec_type<GemmConfig<ck_tile::pk_int4_t>,
+            return run_gemm_example_prec_type<GemmConfig<ck_tile::fp8_t>,
                                               TypeConfig,
                                               128,
                                               ck_tile::QuantType::AQuantGrouped>(
@@ -312,12 +308,12 @@ int run_gemm_example(int argc, char* argv[])
     {
         using TypeConfig = decltype(GemmQuantTypeConfig<ck_tile::pk_int4_t,
                                                         ck_tile::bf8_t,
-                                                        float,
+                                                        ck_tile::half_t,
                                                         ck_tile::bf8_t>{});
 
         if(quant_mode == "aquant")
         {
-            return run_gemm_example_prec_type<GemmConfig<ck_tile::pk_int4_t>,
+            return run_gemm_example_prec_type<GemmConfig<ck_tile::bf8_t>,
                                               TypeConfig,
                                               128,
                                               ck_tile::QuantType::AQuantGrouped>(
@@ -329,42 +325,46 @@ int run_gemm_example(int argc, char* argv[])
                 "Unsupported quantization mode for this datatype! Use 'aquant'.");
         }
     }
-    else if(data_type == "i4f32fp8")
+    else if(data_type == "fp8i4")
     {
-        using TypeConfig =
-            decltype(GemmQuantTypeConfig<ck_tile::pk_int4_t, ck_tile::fp8_t, float, float>{});
+        using TypeConfig = decltype(GemmQuantTypeConfig<ck_tile::fp8_t,
+                                                        ck_tile::pk_int4_t,
+                                                        ck_tile::half_t,
+                                                        ck_tile::fp8_t>{});
 
-        if(quant_mode == "aquant")
+        if(quant_mode == "bquant")
         {
-            return run_gemm_example_prec_type<GemmConfig<ck_tile::pk_int4_t>,
+            return run_gemm_example_prec_type<GemmConfig<ck_tile::fp8_t>,
                                               TypeConfig,
                                               128,
-                                              ck_tile::QuantType::AQuantGrouped>(
+                                              ck_tile::QuantType::BQuantGrouped>(
                 a_layout, b_layout, argc, argv);
         }
         else
         {
             throw std::runtime_error(
-                "Unsupported quantization mode for this datatype! Use 'aquant'.");
+                "Unsupported quantization mode for this datatype! Use 'bquant'.");
         }
     }
-    else if(data_type == "i4f32bf8")
+    else if(data_type == "bf8i4")
     {
-        using TypeConfig =
-            decltype(GemmQuantTypeConfig<ck_tile::pk_int4_t, ck_tile::bf8_t, float, float>{});
+        using TypeConfig = decltype(GemmQuantTypeConfig<ck_tile::bf8_t,
+                                                        ck_tile::pk_int4_t,
+                                                        ck_tile::half_t,
+                                                        ck_tile::bf8_t>{});
 
-        if(quant_mode == "aquant")
+        if(quant_mode == "bquant")
         {
-            return run_gemm_example_prec_type<GemmConfig<ck_tile::pk_int4_t>,
+            return run_gemm_example_prec_type<GemmConfig<ck_tile::bf8_t>,
                                               TypeConfig,
                                               128,
-                                              ck_tile::QuantType::AQuantGrouped>(
+                                              ck_tile::QuantType::BQuantGrouped>(
                 a_layout, b_layout, argc, argv);
         }
         else
         {
             throw std::runtime_error(
-                "Unsupported quantization mode for this datatype! Use 'aquant'.");
+                "Unsupported quantization mode for this datatype! Use 'bquant'.");
         }
     }
     else
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 555631eaa3..ccf07460fa 100644
--- a/example/ck_tile/38_block_scale_gemm/gemm_utils.hpp
+++ b/example/ck_tile/38_block_scale_gemm/gemm_utils.hpp
@@ -11,10 +11,6 @@
 #include "ck_tile/ops/gemm.hpp"
 #include "ck_tile/ops/gemm_group_quant.hpp"
 
-#define CK_TILE_PIPELINE_PREFILL 1
-#define CK_TILE_PIPELINE_DECODE 2
-#define CK_TILE_PIPELINE_PRESHUFFLEQUANT 3
-
 template <typename PrecType, ck_tile::index_t M_Warp_Tile>
 constexpr ck_tile::index_t get_k_warp_tile()
 {
@@ -48,6 +44,13 @@ constexpr ck_tile::index_t get_k_from_preshuffled_warp_tile()
 #endif
 }
 
+template <typename Layout>
+static constexpr inline auto is_row_major(Layout layout_)
+{
+    return ck_tile::bool_constant<std::is_same_v<ck_tile::remove_cvref_t<decltype(layout_)>,
+                                                 ck_tile::tensor_layout::gemm::RowMajor>>{};
+}
+
 template <typename ADataType, typename BDataType, typename AccDataType, typename CDataType>
 auto calculate_rtol_atol(const ck_tile::index_t K,
                          const ck_tile::index_t kbatch,
@@ -81,6 +84,9 @@ struct GemmConfigBase
     static constexpr bool TransposeC            = false;
     static constexpr bool UseStructuredSparsity = false;
 
+    static constexpr int kBlockPerCu = 1;
+    static constexpr auto Scheduler  = ck_tile::GemmPipelineScheduler::Intrawave;
+
     static constexpr ck_tile::index_t TileParitionerGroupNum = 8;
     static constexpr ck_tile::index_t TileParitionerM01      = 4;
 
@@ -102,8 +108,22 @@ struct GemmConfigQuant : public GemmConfigBase
     static constexpr ck_tile::index_t M_Warp_Tile = 16;
     static constexpr ck_tile::index_t N_Warp_Tile = 16;
     static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile<PrecType, M_Warp_Tile>();
+};
 
-    static constexpr int kBlockPerCu = 1;
+template <typename PrecType>
+struct GemmConfigRowColQuant : public GemmConfigBase
+{
+    static constexpr ck_tile::index_t M_Tile = 16;
+    static constexpr ck_tile::index_t N_Tile = 64;
+    static constexpr ck_tile::index_t K_Tile = 256 / sizeof(PrecType);
+
+    static constexpr ck_tile::index_t M_Warp = 1;
+    static constexpr ck_tile::index_t N_Warp = 4;
+    static constexpr ck_tile::index_t K_Warp = 1;
+
+    static constexpr ck_tile::index_t M_Warp_Tile = 16;
+    static constexpr ck_tile::index_t N_Warp_Tile = 16;
+    static constexpr ck_tile::index_t K_Warp_Tile = get_k_warp_tile<PrecType, M_Warp_Tile>();
 };
 
 template <typename PrecType>
@@ -122,8 +142,6 @@ struct GemmConfigPreshuffleQuant : public GemmConfigBase
     static constexpr ck_tile::index_t K_Warp_Tile =
         get_k_from_preshuffled_warp_tile<PrecType, M_Warp_Tile>();
 
-    static constexpr int kBlockPerCu = 1;
-
     static constexpr bool PreshuffleQuant = true;
 };
 
@@ -206,25 +224,25 @@ auto create_args(int argc, char* argv[])
         .insert("a_layout", "R", "A tensor data layout - Row by default")
         .insert("aq_layout", "R", "Aq tensor data layout - Row by default")
         .insert("b_layout", "C", "B tensor data layout - Column by default")
+        .insert("bq_layout", "C", "Bq tensor data layout - Column by default")
         .insert("c_layout", "R", "C tensor data layout - Row by default")
         .insert("stride_a", "0", "Tensor A stride")
         .insert("stride_q", "0", "Tensor AQ stride")
         .insert("stride_b", "0", "Tensor B stride")
         .insert("stride_c", "0", "Tensor C stride")
         .insert("v", "1", "0. No validation, 1. Validation on CPU, 2. Validation on GPU")
-        .insert("prec", "fp8", "data type. fp8/bf8/i4fp8/i4bf8/i4f32fp8/i4f32bf8")
+        .insert("prec",
+                "fp8",
+                "data type. For AQuant: fp8/bf8/i4fp8/i4bf8, For Bquant: fp8/bf8/fp8i4/bf8i4")
         .insert("warmup", "50", "number of iterations before benchmark the kernel")
         .insert("repeat", "1000", "number of iterations to benchmark the kernel")
         .insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer")
         .insert("split_k", "1", "splitK value")
         .insert("init", "0", "0:random, 1:linear, 2:constant(1)")
-        .insert("persistent", "0", "0:non-persistent, 1:persistent")
-        .insert("as_br_cr", "false", "Choose between as_br_cr and as_bs_cr")
-        .insert("quant_mode", "aquant", "Choose aquant (default) or rowcol");
+        .insert("flush_cache", "true", "flush cache before running the kernel, defaults to true")
+        .insert("rotating_count", "1", "rotating count, defaults to 1")
+        .insert("quant_mode", "aquant", "Choose aquant (default), bquant or rowcol");
 
     bool result = arg_parser.parse(argc, argv);
     return std::make_tuple(result, arg_parser);
 }
-
-// host API
-float gemm_calc_aquant(const ck_tile::QuantGemmHostArgs& args, const ck_tile::stream_config& s);
diff --git a/example/ck_tile/38_block_scale_gemm/run_gemm_aquant_example.inc b/example/ck_tile/38_block_scale_gemm/run_gemm_aquant_example.inc
deleted file mode 100644
index 477cb125be..0000000000
--- a/example/ck_tile/38_block_scale_gemm/run_gemm_aquant_example.inc
+++ /dev/null
@@ -1,302 +0,0 @@
-
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
-
-#pragma once
-#include <random>
-#include <stdexcept>
-#include "../00_shared/host_tensor_utils.hpp"
-
-template <typename Layout>
-static constexpr inline auto is_row_major(Layout layout_)
-{
-    return ck_tile::bool_constant<std::is_same_v<ck_tile::remove_cvref_t<decltype(layout_)>,
-                                                 ck_tile::tensor_layout::gemm::RowMajor>>{};
-}
-
-template <typename T>
-auto shuffle_aq(const ck_tile::HostTensor<T>& t, int block_aq_k)
-{
-    if(t.get_lengths().size() != 2)
-    {
-        throw std::runtime_error("Host tensor is not rank 2 tensor.");
-    }
-    int m_   = t.get_lengths()[0];
-    int aqk_ = t.get_lengths()[1];
-    if(aqk_ % block_aq_k != 0)
-    {
-        throw std::runtime_error("shuffle_aq needs a aqk of multiple times of block_aq_k.");
-    }
-    ck_tile::HostTensor<T> t_view({m_, aqk_ / block_aq_k, block_aq_k});
-    std::copy(t.begin(), t.end(), t_view.begin());
-    return ck_tile::reference_permute(t_view, {1, 0, 2});
-}
-
-template <typename GemmConfig,
-          typename ADataType,
-          typename AQDataType,
-          typename BDataType,
-          typename AccDataType,
-          typename CDataType,
-          typename ALayout,
-          typename AQLayout,
-          typename BLayout,
-          typename CLayout,
-          uint32_t QuantGroupSize>
-float invoke_gemm(ck_tile::DeviceMem& a_m_k_dev_buf,
-                  ck_tile::DeviceMem& aq_m_aqk_dev_buf,
-                  ck_tile::DeviceMem& b_k_n_dev_buf,
-                  ck_tile::DeviceMem& c_m_n_dev_buf,
-                  ck_tile::index_t M,
-                  ck_tile::index_t N,
-                  ck_tile::index_t K,
-                  ck_tile::index_t AQK,
-                  ck_tile::index_t stride_A,
-                  ck_tile::index_t stride_AQ,
-                  ck_tile::index_t stride_B,
-                  ck_tile::index_t stride_C,
-                  ck_tile::index_t kbatch,
-                  int n_warmup,
-                  int n_repeat)
-{
-    ck_tile::QuantGemmHostArgs args;
-    args.a_ptr     = a_m_k_dev_buf.GetDeviceBuffer();
-    args.aq_ptr    = aq_m_aqk_dev_buf.GetDeviceBuffer();
-    args.b_ptr     = b_k_n_dev_buf.GetDeviceBuffer();
-    args.c_ptr     = c_m_n_dev_buf.GetDeviceBuffer();
-    args.k_batch   = kbatch;
-    args.M         = M;
-    args.N         = N;
-    args.K         = K;
-    args.QK_A      = AQK;
-    args.stride_A  = stride_A;
-    args.stride_B  = stride_B;
-    args.stride_C  = stride_C;
-    args.stride_AQ = stride_AQ;
-
-    float ave_time = gemm_calc_aquant<GemmConfig,
-                                      ADataType,
-                                      AQDataType,
-                                      BDataType,
-                                      AccDataType,
-                                      CDataType,
-                                      BDataType,
-                                      ALayout,
-                                      BLayout,
-                                      CLayout,
-                                      QuantGroupSize>(
-        args, ck_tile::stream_config{nullptr, true, 1, n_warmup, n_repeat});
-
-    std::size_t flop     = std::size_t(2) * M * N * K;
-    std::size_t num_byte = sizeof(ADataType) * M * K + sizeof(AQDataType) * M * AQK +
-                           sizeof(BDataType) * N * K + sizeof(CDataType) * M * N;
-    float tflops     = static_cast<float>(flop) / 1.E9 / ave_time;
-    float gb_per_sec = num_byte / 1.E6 / ave_time;
-
-    std::cout << "Run Gemm kernel with M =" << M << " N =" << N << " K =" << K
-              << " StrideA =" << stride_A << " StrideAQ =" << stride_AQ << " StrideB =" << stride_B
-              << " StrideC =" << stride_C << " A_Layout =" << ALayout::name
-              << " B_Layout =" << BLayout::name << " C_Layout =" << CLayout::name
-              << " A_Type = " << DataTypeTraits<ADataType>::name
-              << " AQ_Type = " << DataTypeTraits<AQDataType>::name
-              << " B_Type = " << DataTypeTraits<BDataType>::name
-              << " Acc_Type = " << DataTypeTraits<AccDataType>::name
-              << " C_Type = " << DataTypeTraits<CDataType>::name << " : " << ave_time << " ms, "
-              << tflops << " TFlops, " << gb_per_sec << " GB/s, " << std::endl;
-
-    return ave_time;
-}
-
-template <typename GemmConfig,
-          typename TypeConfig,
-          uint32_t QuantGroupSize,
-          typename ALayout,
-          typename AQLayout,
-          typename BLayout,
-          typename CLayout>
-int run_gemm_example_with_layouts(int argc,
-                                  char* argv[],
-                                  const ALayout a_layout                  = ALayout{},
-                                  const AQLayout aq_layout                = AQLayout{},
-                                  const BLayout b_layout                  = BLayout{},
-                                  [[maybe_unused]] const CLayout c_layout = CLayout{})
-{
-    auto [result, arg_parser] = create_args(argc, argv);
-    if(!result)
-        return -1;
-
-    using ADataType   = typename TypeConfig::ADataType;
-    using AQDataType  = typename TypeConfig::QDataType;
-    using BDataType   = typename TypeConfig::BDataType;
-    using AccDataType = typename TypeConfig::AccDataType;
-    using CDataType   = typename TypeConfig::CDataType;
-
-    ck_tile::index_t M = arg_parser.get_int("m");
-    ck_tile::index_t N = arg_parser.get_int("n");
-    ck_tile::index_t K = arg_parser.get_int("k");
-
-    if(K % QuantGroupSize != 0)
-    {
-        throw std::runtime_error("K must be aligned with QuantGroupSize");
-    }
-
-    ck_tile::index_t AQK = K / QuantGroupSize;
-
-    ck_tile::index_t stride_A  = arg_parser.get_int("stride_a");
-    ck_tile::index_t stride_AQ = arg_parser.get_int("stride_q");
-    ck_tile::index_t stride_B  = arg_parser.get_int("stride_b");
-    ck_tile::index_t stride_C  = arg_parser.get_int("stride_c");
-
-    ck_tile::index_t kbatch      = arg_parser.get_int("split_k");
-    int n_warmup                 = arg_parser.get_int("warmup");
-    int n_repeat                 = arg_parser.get_int("repeat");
-    ck_tile::index_t init_method = arg_parser.get_int("init");
-
-    stride_A  = ck_tile::get_default_stride(M, K, stride_A, is_row_major(a_layout));
-    stride_AQ = ck_tile::get_default_stride(M, AQK, stride_AQ, is_row_major(aq_layout));
-    stride_B  = ck_tile::get_default_stride(K, N, stride_B, is_row_major(b_layout));
-    stride_C  = ck_tile::get_default_stride(M, N, stride_C, is_row_major(CLayout{}));
-
-    ck_tile::HostTensor<ADataType> a_m_k(
-        ck_tile::host_tensor_descriptor(M, K, stride_A, is_row_major(a_layout)));
-    ck_tile::HostTensor<AQDataType> aq_m_aqk(
-        ck_tile::host_tensor_descriptor(M, AQK, stride_AQ, is_row_major(aq_layout)));
-    ck_tile::HostTensor<BDataType> b_k_n(
-        ck_tile::host_tensor_descriptor(K, N, stride_B, is_row_major(b_layout)));
-    ck_tile::HostTensor<CDataType> c_m_n_dev_result(
-        ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
-
-    std::random_device rd;
-    std::mt19937 gen(rd());
-    std::uniform_int_distribution<std::uint32_t> fill_seed(0, 500);
-
-    if(init_method == 0)
-    {
-        if constexpr(std::is_same_v<ADataType, ck_tile::pk_int4_t>)
-        {
-            ck_tile::FillUniformDistribution<ck_tile::pk_int4_t>{-5.0f, 5.0f, fill_seed(gen)}(
-                a_m_k);
-        }
-        else
-        {
-            ck_tile::FillUniformDistribution<ADataType>{-2.0f, 3.0f, fill_seed(gen)}(a_m_k);
-        }
-        ck_tile::FillUniformDistribution<AQDataType>{-2.0f, 2.0f, fill_seed(gen)}(aq_m_aqk);
-        ck_tile::FillUniformDistribution<BDataType>{-5.0f, 5.0f, fill_seed(gen)}(b_k_n);
-    }
-    else if(init_method == 1)
-    {
-        std::cout << "Monotonic initialization is not supported." << std::endl;
-        return 0;
-    }
-    else if(init_method == 2)
-    {
-        ck_tile::FillConstant<ADataType>{static_cast<ADataType>(0x22)}(a_m_k);
-        ck_tile::FillConstant<AQDataType>{static_cast<AQDataType>(0.5f)}(aq_m_aqk);
-        ck_tile::FillConstant<BDataType>{static_cast<BDataType>(0x38)}(b_k_n);
-    }
-    else
-    {
-        a_m_k.SetZero();
-        aq_m_aqk.SetZero();
-        b_k_n.SetZero();
-    }
-
-    ck_tile::DeviceMem a_m_k_dev_buf(a_m_k.get_element_space_size_in_bytes());
-    ck_tile::DeviceMem aq_m_aqk_dev_buf(aq_m_aqk.get_element_space_size_in_bytes());
-    ck_tile::DeviceMem b_k_n_dev_buf(b_k_n.get_element_space_size_in_bytes());
-    ck_tile::DeviceMem c_m_n_dev_buf(c_m_n_dev_result.get_element_space_size_in_bytes());
-
-    if constexpr(GemmConfig::PreshuffleQuant)
-    {
-        ck_tile::HostTensor<AQDataType> aq_shuffle_host =
-            shuffle_aq(aq_m_aqk, GemmConfig::K_Tile / QuantGroupSize);
-        aq_m_aqk_dev_buf.ToDevice(aq_shuffle_host.data());
-    }
-    else
-    {
-        aq_m_aqk_dev_buf.ToDevice(aq_m_aqk.data());
-    }
-
-    if constexpr(std::is_same_v<ADataType, ck_tile::pk_int4_t>)
-    {
-        ck_tile::HostTensor<ADataType> a_m_k_dev = a_m_k;
-        ck_tile::permute_vectors_i4x4_b(a_m_k_dev);
-        a_m_k_dev_buf.ToDevice(a_m_k_dev.data());
-    }
-    else
-    {
-        a_m_k_dev_buf.ToDevice(a_m_k.data());
-    }
-    b_k_n_dev_buf.ToDevice(b_k_n.data());
-    c_m_n_dev_buf.SetZero();
-    c_m_n_dev_result.SetZero();
-
-    invoke_gemm<GemmConfig,
-                ADataType,
-                AQDataType,
-                BDataType,
-                AccDataType,
-                CDataType,
-                ALayout,
-                AQLayout,
-                BLayout,
-                CLayout,
-                QuantGroupSize>(a_m_k_dev_buf,
-                                aq_m_aqk_dev_buf,
-                                b_k_n_dev_buf,
-                                c_m_n_dev_buf,
-                                M,
-                                N,
-                                K,
-                                AQK,
-                                stride_A,
-                                stride_AQ,
-                                stride_B,
-                                stride_C,
-                                kbatch,
-                                n_warmup,
-                                n_repeat);
-
-    c_m_n_dev_buf.FromDevice(c_m_n_dev_result.data());
-    bool pass = true;
-
-    if(arg_parser.get_int("v") == 1)
-    {
-        ck_tile::HostTensor<CDataType> c_m_n_host_ref(
-            ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
-        c_m_n_host_ref.SetZero();
-
-        ck_tile::reference_gemm_quant<ADataType,
-                                      AQDataType,
-                                      BDataType,
-                                      AccDataType,
-                                      CDataType,
-                                      QuantGroupSize,
-                                      true>(a_m_k, aq_m_aqk, b_k_n, c_m_n_host_ref);
-        const float max_accumulated_value =
-            *std::max_element(c_m_n_host_ref.mData.begin(), c_m_n_host_ref.mData.end());
-        const auto rtol_atol = calculate_rtol_atol<ADataType, BDataType, AccDataType, CDataType>(
-            K, kbatch, max_accumulated_value);
-        pass = ck_tile::check_err(c_m_n_dev_result,
-                                  c_m_n_host_ref,
-                                  "Error: Incorrect results!",
-                                  rtol_atol.at(ck_tile::number<0>{}),
-                                  rtol_atol.at(ck_tile::number<1>{}));
-
-        if(!pass)
-        {
-            std::cout << "Relative error threshold: " << rtol_atol.at(ck_tile::number<0>{})
-                      << " Absolute error threshold: " << rtol_atol.at(ck_tile::number<1>{})
-                      << std::endl;
-        }
-        std::cout << "CPU verification " << (pass ? "Passed!" : "Failed ...") << std::endl;
-    }
-    else if(arg_parser.get_int("v") == 2)
-    {
-        std::cout << "GPU verification is not implemented yet. Re-run with -v=1" << std::endl;
-        return false;
-    }
-
-    return pass;
-}
diff --git a/example/ck_tile/38_block_scale_gemm/run_gemm_bquant_example.inc b/example/ck_tile/38_block_scale_gemm/run_gemm_bquant_example.inc
deleted file mode 100644
index f30039b1b2..0000000000
--- a/example/ck_tile/38_block_scale_gemm/run_gemm_bquant_example.inc
+++ /dev/null
@@ -1,296 +0,0 @@
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
-
-#pragma once
-#include <random>
-#include "ck_tile/host/permute_pk_int4.hpp"
-
-template <typename Layout>
-static constexpr inline auto is_row_major(Layout layout_)
-{
-    return ck_tile::bool_constant<std::is_same_v<ck_tile::remove_cvref_t<decltype(layout_)>,
-                                                 ck_tile::tensor_layout::gemm::RowMajor>>{};
-}
-
-template <typename T>
-auto shuffle_bq(const ck_tile::HostTensor<T>& t, int block_bq_k)
-{
-    if(t.get_lengths().size() != 2)
-    {
-        throw std::runtime_error("Host tensor is not rank 2 tensor.");
-    }
-    int n_   = t.get_lengths()[0];
-    int bqk_ = t.get_lengths()[1];
-    if(bqk_ % block_bq_k != 0)
-    {
-        throw std::runtime_error("shuffle_aq needs a bqk of multiple times of block_bq_k.");
-    }
-    ck_tile::HostTensor<T> t_view({n_, bqk_ / block_bq_k, block_bq_k});
-    std::copy(t.begin(), t.end(), t_view.begin());
-    return ck_tile::reference_permute(t_view, {1, 0, 2});
-}
-
-template <typename GemmConfig,
-          typename ADataType,
-          typename BDataType,
-          typename BQDataType,
-          typename DsDataType,
-          typename AccDataType,
-          typename CDataType,
-          typename ALayout,
-          typename BLayout,
-          typename BQLayout,
-          typename DsLayout,
-          typename CLayout,
-          uint32_t QuantGroupSize,
-          typename CDEElementWise = ck_tile::element_wise::PassThrough>
-float invoke_gemm(ck_tile::DeviceMem& a_m_k_dev_buf,
-                  ck_tile::DeviceMem& b_k_n_dev_buf,
-                  ck_tile::DeviceMem& bq_bqk_n_dev_buf,
-                  ck_tile::DeviceMem& c_m_n_dev_buf,
-                  ck_tile::index_t M,
-                  ck_tile::index_t N,
-                  ck_tile::index_t K,
-                  ck_tile::index_t BQK,
-                  ck_tile::index_t stride_A,
-                  ck_tile::index_t stride_B,
-                  ck_tile::index_t stride_BQ,
-                  ck_tile::index_t stride_C,
-                  ck_tile::index_t kbatch,
-                  int n_warmup,
-                  int n_repeat)
-{
-    ck_tile::QuantGemmHostArgs args;
-    args.a_ptr     = a_m_k_dev_buf.GetDeviceBuffer();
-    args.b_ptr     = b_k_n_dev_buf.GetDeviceBuffer();
-    args.bq_ptr    = bq_bqk_n_dev_buf.GetDeviceBuffer();
-    args.c_ptr     = c_m_n_dev_buf.GetDeviceBuffer();
-    args.k_batch   = kbatch;
-    args.M         = M;
-    args.N         = N;
-    args.K         = K;
-    args.QK_B      = BQK;
-    args.stride_A  = stride_A;
-    args.stride_B  = stride_B;
-    args.stride_C  = stride_C;
-    args.stride_BQ = stride_BQ;
-
-    float ave_time = gemm_calc_bquant<GemmConfig,
-                                      ADataType,
-                                      BDataType,
-                                      BQDataType,
-                                      AccDataType,
-                                      CDataType,
-                                      ADataType, // computeDatatype
-                                      ALayout,
-                                      BLayout,
-                                      CLayout,
-                                      QuantGroupSize>(
-        args, ck_tile::stream_config{nullptr, true, 1, n_warmup, n_repeat});
-
-    std::size_t flop     = std::size_t(2) * M * N * K;
-    std::size_t num_byte = sizeof(ADataType) * M * K + sizeof(BDataType) * N * K +
-                           sizeof(BQDataType) * BQK * N + sizeof(CDataType) * M * N;
-    float tflops     = static_cast<float>(flop) / 1.E9 / ave_time;
-    float gb_per_sec = num_byte / 1.E6 / ave_time;
-
-    std::cout << "Run Gemm kernel with M =" << M << " N =" << N << " K =" << K
-              << " StrideA =" << stride_A << " StrideB =" << stride_B << " StrideBQ =" << stride_BQ
-              << " StrideC =" << stride_C << " A_Layout =" << ALayout::name
-              << " B_Layout =" << BLayout::name << " C_Layout =" << CLayout::name
-              << " A_Type = " << DataTypeTraits<ADataType>::name
-              << " B_Type = " << DataTypeTraits<BDataType>::name
-              << " BQ_Type = " << DataTypeTraits<BQDataType>::name
-              << " Acc_Type = " << DataTypeTraits<AccDataType>::name
-              << " C_Type = " << DataTypeTraits<CDataType>::name << " : " << ave_time << " ms, "
-              << tflops << " TFlops, " << gb_per_sec << " GB/s, " << std::endl;
-
-    return ave_time;
-}
-
-template <typename GemmConfig,
-          typename TypeConfig,
-          uint32_t QuantGroupSize,
-          typename ALayout,
-          typename BLayout,
-          typename BQLayout,
-          typename CLayout>
-int run_gemm_example_with_layouts(int argc,
-                                  char* argv[],
-                                  const ALayout a_layout                  = ALayout{},
-                                  const BLayout b_layout                  = BLayout{},
-                                  const BQLayout bq_layout                = BQLayout{},
-                                  [[maybe_unused]] const CLayout c_layout = CLayout{})
-{
-    auto [result, arg_parser] = create_args(argc, argv);
-    if(!result)
-        return -1;
-
-    using ADataType   = typename TypeConfig::ADataType;
-    using BDataType   = typename TypeConfig::BDataType;
-    using BQDataType  = typename TypeConfig::QDataType;
-    using AccDataType = typename TypeConfig::AccDataType;
-    using CDataType   = typename TypeConfig::CDataType;
-
-    ck_tile::index_t M = arg_parser.get_int("m");
-    ck_tile::index_t N = arg_parser.get_int("n");
-    ck_tile::index_t K = arg_parser.get_int("k");
-
-    if(K % QuantGroupSize != 0)
-    {
-        throw std::runtime_error("K must be aligned with QuantGroupSize");
-    }
-
-    ck_tile::index_t BQK = K / QuantGroupSize;
-
-    ck_tile::index_t stride_A  = arg_parser.get_int("stride_a");
-    ck_tile::index_t stride_B  = arg_parser.get_int("stride_b");
-    ck_tile::index_t stride_BQ = arg_parser.get_int("stride_q");
-    ck_tile::index_t stride_C  = arg_parser.get_int("stride_c");
-
-    ck_tile::index_t kbatch      = arg_parser.get_int("split_k");
-    int n_warmup                 = arg_parser.get_int("warmup");
-    int n_repeat                 = arg_parser.get_int("repeat");
-    ck_tile::index_t init_method = arg_parser.get_int("init");
-
-    stride_A  = ck_tile::get_default_stride(M, K, stride_A, is_row_major(a_layout));
-    stride_B  = ck_tile::get_default_stride(K, N, stride_B, is_row_major(b_layout));
-    stride_BQ = ck_tile::get_default_stride(BQK, N, stride_BQ, is_row_major(bq_layout));
-    stride_C  = ck_tile::get_default_stride(M, N, stride_C, is_row_major(CLayout{}));
-
-    ck_tile::HostTensor<ADataType> a_m_k(
-        ck_tile::host_tensor_descriptor(M, K, stride_A, is_row_major(a_layout)));
-    ck_tile::HostTensor<BDataType> b_k_n(
-        ck_tile::host_tensor_descriptor(K, N, stride_B, is_row_major(b_layout)));
-    ck_tile::HostTensor<BQDataType> bq_bqk_n(
-        ck_tile::host_tensor_descriptor(BQK, N, stride_BQ, is_row_major(bq_layout)));
-    ck_tile::HostTensor<CDataType> c_m_n_dev_result(
-        ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
-
-    std::random_device rd;
-    std::mt19937 gen(rd());
-    std::uniform_int_distribution<std::uint32_t> fill_seed(0, 500);
-
-    if(init_method == 0)
-    {
-        if constexpr(std::is_same_v<BDataType, ck_tile::pk_int4_t>)
-        {
-            ck_tile::FillUniformDistribution<ck_tile::pk_int4_t>{-5.0f, 5.0f, fill_seed(gen)}(
-                b_k_n);
-        }
-        else
-        {
-            ck_tile::FillUniformDistribution<BDataType>{-2.0f, 3.0f, fill_seed(gen)}(b_k_n);
-        }
-        ck_tile::FillUniformDistribution<BQDataType>{-2.0f, 2.0f, fill_seed(gen)}(bq_bqk_n);
-        ck_tile::FillUniformDistribution<ADataType>{-5.0f, 5.0f, fill_seed(gen)}(a_m_k);
-    }
-    else if(init_method == 1)
-    {
-        std::cout << "Monotonic initialization is not supported." << std::endl;
-        return 0;
-    }
-    else if(init_method == 2)
-    {
-        ck_tile::FillConstant<ADataType>{static_cast<ADataType>(0x38)}(a_m_k);
-        ck_tile::FillConstant<BDataType>{static_cast<BDataType>(0x22)}(b_k_n);
-        ck_tile::FillConstant<BQDataType>{static_cast<BQDataType>(0.5f)}(bq_bqk_n);
-    }
-    else
-    {
-        a_m_k.SetZero();
-        b_k_n.SetZero();
-        bq_bqk_n.SetZero();
-    }
-
-    ck_tile::DeviceMem a_m_k_dev_buf(a_m_k.get_element_space_size_in_bytes());
-    ck_tile::DeviceMem b_k_n_dev_buf(b_k_n.get_element_space_size_in_bytes());
-    ck_tile::DeviceMem bq_bqk_n_dev_buf(bq_bqk_n.get_element_space_size_in_bytes());
-    ck_tile::DeviceMem c_m_n_dev_buf(c_m_n_dev_result.get_element_space_size_in_bytes());
-
-    a_m_k_dev_buf.ToDevice(a_m_k.data());
-    if constexpr(std::is_same_v<BDataType, ck_tile::pk_int4_t>)
-    {
-        // Permute vector pk_i4x4 data for device implementation
-        ck_tile::HostTensor<BDataType> b_k_n_dev = b_k_n;
-        ck_tile::permute_vectors_i4x4_b(b_k_n_dev);
-        b_k_n_dev_buf.ToDevice(b_k_n_dev.data());
-    }
-    else
-    {
-        b_k_n_dev_buf.ToDevice(b_k_n.data());
-    }
-    bq_bqk_n_dev_buf.ToDevice(bq_bqk_n.data());
-    c_m_n_dev_buf.SetZero();
-    c_m_n_dev_result.SetZero();
-
-    invoke_gemm<GemmConfig,
-                ADataType,
-                BDataType,
-                BQDataType,
-                ck_tile::tuple<>,
-                AccDataType,
-                CDataType,
-                ALayout,
-                BLayout,
-                BQLayout,
-                ck_tile::tuple<>,
-                CLayout,
-                QuantGroupSize>(a_m_k_dev_buf,
-                                b_k_n_dev_buf,
-                                bq_bqk_n_dev_buf,
-                                c_m_n_dev_buf,
-                                M,
-                                N,
-                                K,
-                                BQK,
-                                stride_A,
-                                stride_B,
-                                stride_BQ,
-                                stride_C,
-                                kbatch,
-                                n_warmup,
-                                n_repeat);
-
-    c_m_n_dev_buf.FromDevice(c_m_n_dev_result.data());
-    bool pass = true;
-
-    if(arg_parser.get_int("v") == 1)
-    {
-        ck_tile::HostTensor<CDataType> c_m_n_host_ref(
-            ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
-        c_m_n_host_ref.SetZero();
-
-        ck_tile::reference_gemm_quant<ADataType,
-                                      BQDataType,
-                                      BDataType,
-                                      AccDataType,
-                                      CDataType,
-                                      QuantGroupSize,
-                                      false>(a_m_k, bq_bqk_n, b_k_n, c_m_n_host_ref);
-        const float max_accumulated_value =
-            *std::max_element(c_m_n_host_ref.mData.begin(), c_m_n_host_ref.mData.end());
-        const auto rtol_atol = calculate_rtol_atol<ADataType, BDataType, AccDataType, CDataType>(
-            K, kbatch, max_accumulated_value);
-        pass = ck_tile::check_err(c_m_n_dev_result,
-                                  c_m_n_host_ref,
-                                  "Error: Incorrect results!",
-                                  rtol_atol.at(ck_tile::number<0>{}),
-                                  rtol_atol.at(ck_tile::number<1>{}));
-
-        if(!pass)
-        {
-            std::cout << "Relative error threshold: " << rtol_atol.at(ck_tile::number<0>{})
-                      << " Absolute error threshold: " << rtol_atol.at(ck_tile::number<1>{})
-                      << std::endl;
-        }
-        std::cout << "CPU verification " << (pass ? "Passed!" : "Failed ...") << std::endl;
-    }
-    else if(arg_parser.get_int("v") == 2)
-    {
-        std::cout << "GPU verification is not implemented yet. Re-run with -v=1" << std::endl;
-        return false;
-    }
-
-    return pass;
-}
diff --git a/example/ck_tile/38_block_scale_gemm/run_gemm_quant_example.inc b/example/ck_tile/38_block_scale_gemm/run_gemm_quant_example.inc
index 54f139dd74..0f45811ff3 100644
--- a/example/ck_tile/38_block_scale_gemm/run_gemm_quant_example.inc
+++ b/example/ck_tile/38_block_scale_gemm/run_gemm_quant_example.inc
@@ -6,28 +6,21 @@
 #include <stdexcept>
 #include "ck_tile/host/permute_pk_int4.hpp"
 
-template <typename Layout>
-static constexpr inline auto is_row_major(Layout layout_)
-{
-    return ck_tile::bool_constant<std::is_same_v<ck_tile::remove_cvref_t<decltype(layout_)>,
-                                                 ck_tile::tensor_layout::gemm::RowMajor>>{};
-}
-
 template <typename T>
-auto shuffle_aq(const ck_tile::HostTensor<T>& t, int block_aq_k)
+auto shuffle_aq(const ck_tile::HostTensor<T>* t, int block_aq_k)
 {
-    if(t.get_lengths().size() != 2)
+    if(t->get_lengths().size() != 2)
     {
         throw std::runtime_error("Host tensor is not rank 2 tensor.");
     }
-    int m_   = t.get_lengths()[0];
-    int aqk_ = t.get_lengths()[1];
+    int m_   = t->get_lengths()[0];
+    int aqk_ = t->get_lengths()[1];
     if(aqk_ % block_aq_k != 0)
     {
         throw std::runtime_error("shuffle_aq needs a aqk of multiple times of block_aq_k.");
     }
     ck_tile::HostTensor<T> t_view({m_, aqk_ / block_aq_k, block_aq_k});
-    std::copy(t.begin(), t.end(), t_view.begin());
+    std::copy(t->begin(), t->end(), t_view.begin());
     return ck_tile::reference_permute(t_view, {1, 0, 2});
 }
 
@@ -39,12 +32,13 @@ template <typename GemmConfig,
           typename BQLayout,
           typename CLayout,
           uint32_t QuantGroupSize,
-          ck_tile::QuantType QuantMode>
+          ck_tile::QuantType QuantMode,
+          typename CDEElementWise = ck_tile::element_wise::PassThrough>
 float invoke_gemm(ck_tile::DeviceMem& a_m_k_dev_buf,
-                  ck_tile::DeviceMem& aq_dev_buf,
+                  ck_tile::DeviceMem* aq_dev_buf,
                   ck_tile::DeviceMem& b_k_n_dev_buf,
-                  ck_tile::DeviceMem& c_m_n_dev_buf,
                   ck_tile::DeviceMem* bq_dev_buf,
+                  ck_tile::DeviceMem& c_m_n_dev_buf,
                   ck_tile::index_t M,
                   ck_tile::index_t N,
                   ck_tile::index_t K,
@@ -57,11 +51,13 @@ float invoke_gemm(ck_tile::DeviceMem& a_m_k_dev_buf,
                   ck_tile::index_t stride_C,
                   ck_tile::index_t kbatch,
                   int n_warmup,
-                  int n_repeat)
+                  int n_repeat,
+                  bool flush_cache,
+                  int rotating_count)
 {
     ck_tile::QuantGemmHostArgs args;
     args.a_ptr     = a_m_k_dev_buf.GetDeviceBuffer();
-    args.aq_ptr    = aq_dev_buf.GetDeviceBuffer();
+    args.aq_ptr    = (aq_dev_buf != nullptr) ? aq_dev_buf->GetDeviceBuffer() : nullptr;
     args.b_ptr     = b_k_n_dev_buf.GetDeviceBuffer();
     args.bq_ptr    = (bq_dev_buf != nullptr) ? bq_dev_buf->GetDeviceBuffer() : nullptr;
     args.c_ptr     = c_m_n_dev_buf.GetDeviceBuffer();
@@ -83,17 +79,23 @@ float invoke_gemm(ck_tile::DeviceMem& a_m_k_dev_buf,
                                      BLayout,
                                      CLayout,
                                      QuantGroupSize,
-                                     QuantMode>(
-        args, ck_tile::stream_config{nullptr, true, 1, n_warmup, n_repeat});
+                                     QuantMode,
+                                     CDEElementWise>(
+        args,
+        ck_tile::stream_config{
+            nullptr, true, 1, n_warmup, n_repeat, true, flush_cache, rotating_count});
 
     std::size_t flop     = std::size_t(2) * M * N * K;
     std::size_t num_byte = sizeof(typename TypeConfig::ADataType) * M * K +
-                           sizeof(typename TypeConfig::QDataType) * M * AQK +
                            sizeof(typename TypeConfig::BDataType) * N * K +
                            sizeof(typename TypeConfig::CDataType) * M * N;
+    if(aq_dev_buf != nullptr)
+    {
+        num_byte += sizeof(typename TypeConfig::QDataType) * M * AQK;
+    }
     if(bq_dev_buf != nullptr)
     {
-        num_byte += sizeof(typename TypeConfig::QDataType) * BQK;
+        num_byte += sizeof(typename TypeConfig::QDataType) * N * BQK;
     }
 
     float tflops     = static_cast<float>(flop) / 1.E9 / ave_time;
@@ -103,7 +105,8 @@ float invoke_gemm(ck_tile::DeviceMem& a_m_k_dev_buf,
               << " StrideA =" << stride_A << " StrideAQ =" << stride_AQ << " StrideB =" << stride_B
               << " StrideC =" << stride_C << " A_Layout =" << ALayout::name
               << " B_Layout =" << BLayout::name << " C_Layout =" << CLayout::name;
-    if constexpr(QuantMode == ck_tile::QuantType::RowColQuant)
+    if constexpr(QuantMode == ck_tile::QuantType::BQuantGrouped ||
+                 QuantMode == ck_tile::QuantType::RowColQuant)
     {
         std::cout << " StrideBQ =" << stride_BQ;
     }
@@ -117,8 +120,12 @@ float invoke_gemm(ck_tile::DeviceMem& a_m_k_dev_buf,
     std::cout << " Acc_Type = " << DataTypeTraits<typename TypeConfig::AccDataType>::name
               << " C_Type = " << DataTypeTraits<typename TypeConfig::CDataType>::name
               << " QuantMode = "
-              << (QuantMode == ck_tile::QuantType::AQuantGrouped ? "AQuantGrouped" : "RowColQuant")
-              << " : " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << (QuantMode == ck_tile::QuantType::AQuantGrouped
+                      ? "AQuantGrouped"
+                      : (QuantMode == ck_tile::QuantType::BQuantGrouped ? "BQuantGrouped"
+                                                                        : "RowColQuant"))
+              << " PreshuffleQuant = " << (GemmConfig::PreshuffleQuant ? "true" : "false") << " : "
+              << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
               << std::endl;
 
     return ave_time;
@@ -156,7 +163,8 @@ int run_gemm_example_with_layouts(int argc,
     ck_tile::index_t N = arg_parser.get_int("n");
     ck_tile::index_t K = arg_parser.get_int("k");
 
-    if constexpr(QuantMode == ck_tile::QuantType::AQuantGrouped)
+    if constexpr(QuantMode == ck_tile::QuantType::AQuantGrouped ||
+                 QuantMode == ck_tile::QuantType::BQuantGrouped)
     {
         if(K % QuantGroupSize != 0)
         {
@@ -164,13 +172,17 @@ int run_gemm_example_with_layouts(int argc,
                 "K must be aligned with QuantGroupSize for AQuantGrouped mode");
         }
     }
-
     ck_tile::index_t AQK, BQK;
     if constexpr(QuantMode == ck_tile::QuantType::AQuantGrouped)
     {
         AQK = K / QuantGroupSize; // Group quantization: AQK = K / GroupSize
         BQK = 0;                  // No B quantization
     }
+    else if constexpr(QuantMode == ck_tile::QuantType::BQuantGrouped)
+    {
+        AQK = 0;                  // No A quantization
+        BQK = K / QuantGroupSize; // Group quantization: BQK = K / GroupSize
+    }
     else if constexpr(QuantMode == ck_tile::QuantType::RowColQuant)
     {
         AQK = 1; // Row quantization: tensor shape [M, 1]
@@ -186,12 +198,14 @@ int run_gemm_example_with_layouts(int argc,
     ck_tile::index_t stride_B  = arg_parser.get_int("stride_b");
     ck_tile::index_t stride_C  = arg_parser.get_int("stride_c");
 
-    ck_tile::index_t stride_BQ = 0;
+    ck_tile::index_t stride_BQ = arg_parser.get_int("stride_q");
 
     ck_tile::index_t kbatch      = arg_parser.get_int("split_k");
     int n_warmup                 = arg_parser.get_int("warmup");
     int n_repeat                 = arg_parser.get_int("repeat");
     ck_tile::index_t init_method = arg_parser.get_int("init");
+    bool flush_cache             = arg_parser.get_bool("flush_cache");
+    int rotating_count           = arg_parser.get_int("rotating_count");
 
     stride_A = ck_tile::get_default_stride(M, K, stride_A, is_row_major(a_layout));
     stride_B = ck_tile::get_default_stride(K, N, stride_B, is_row_major(b_layout));
@@ -201,6 +215,12 @@ int run_gemm_example_with_layouts(int argc,
     if constexpr(QuantMode == ck_tile::QuantType::AQuantGrouped)
     {
         stride_AQ = ck_tile::get_default_stride(M, AQK, stride_AQ, is_row_major(aq_layout));
+        stride_BQ = 0; // No B quantization
+    }
+    else if constexpr(QuantMode == ck_tile::QuantType::BQuantGrouped)
+    {
+        stride_AQ = 0; // No A quantization
+        stride_BQ = ck_tile::get_default_stride(BQK, N, stride_BQ, is_row_major(bq_layout));
     }
     else if constexpr(QuantMode == ck_tile::QuantType::RowColQuant)
     {
@@ -216,21 +236,26 @@ int run_gemm_example_with_layouts(int argc,
         ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
 
     // Create AQ tensor with appropriate shape
-    ck_tile::HostTensor<AQDataType> aq_tensor = [&]() {
-        if constexpr(QuantMode == ck_tile::QuantType::AQuantGrouped)
-        {
-            return ck_tile::HostTensor<AQDataType>(
-                ck_tile::host_tensor_descriptor(M, AQK, stride_AQ, is_row_major(aq_layout)));
-        }
-        else
-        {
-            return ck_tile::HostTensor<AQDataType>(
-                ck_tile::host_tensor_descriptor(M, AQK, stride_AQ, is_row_major(aq_layout)));
-        }
-    }();
+    std::unique_ptr<ck_tile::HostTensor<AQDataType>> aq_tensor_ptr = nullptr;
+    if constexpr(QuantMode == ck_tile::QuantType::AQuantGrouped)
+    {
+        aq_tensor_ptr = std::make_unique<ck_tile::HostTensor<AQDataType>>(
+            ck_tile::host_tensor_descriptor(M, AQK, stride_AQ, is_row_major(aq_layout)));
+    }
+    else if(QuantMode == ck_tile::QuantType::RowColQuant)
+    {
+        aq_tensor_ptr = std::make_unique<ck_tile::HostTensor<AQDataType>>(
+            ck_tile::host_tensor_descriptor(M, AQK, stride_AQ, is_row_major(aq_layout)));
+    }
+
     // Create BQ tensor only for RowColQuant mode
     std::unique_ptr<ck_tile::HostTensor<BQDataType>> bq_tensor_ptr = nullptr;
-    if constexpr(QuantMode == ck_tile::QuantType::RowColQuant)
+    if constexpr(QuantMode == ck_tile::QuantType::BQuantGrouped)
+    {
+        bq_tensor_ptr = std::make_unique<ck_tile::HostTensor<BQDataType>>(
+            ck_tile::host_tensor_descriptor(BQK, N, stride_BQ, is_row_major(bq_layout)));
+    }
+    else if constexpr(QuantMode == ck_tile::QuantType::RowColQuant)
     {
         bq_tensor_ptr = std::make_unique<ck_tile::HostTensor<BQDataType>>(
             ck_tile::host_tensor_descriptor(1, N, stride_BQ, is_row_major(bq_layout)));
@@ -242,22 +267,41 @@ int run_gemm_example_with_layouts(int argc,
 
     if(init_method == 0)
     {
-        if constexpr(std::is_same_v<ADataType, ck_tile::pk_int4_t>)
+        if constexpr(QuantMode == ck_tile::QuantType::BQuantGrouped)
         {
-            ck_tile::FillUniformDistribution<ck_tile::pk_int4_t>{-5.0f, 5.0f, fill_seed(gen)}(
-                a_m_k);
+            if constexpr(std::is_same_v<BDataType, ck_tile::pk_int4_t>)
+            {
+                ck_tile::FillUniformDistribution<ck_tile::pk_int4_t>{-5.0f, 5.0f, fill_seed(gen)}(
+                    b_k_n);
+            }
+            else
+            {
+                ck_tile::FillUniformDistribution<BDataType>{-2.0f, 3.0f, fill_seed(gen)}(b_k_n);
+            }
+            ck_tile::FillUniformDistribution<BQDataType>{-2.0f, 2.0f, fill_seed(gen)}(
+                *bq_tensor_ptr);
+            ck_tile::FillUniformDistribution<ADataType>{-5.0f, 5.0f, fill_seed(gen)}(a_m_k);
         }
         else
         {
-            ck_tile::FillUniformDistribution<ADataType>{-2.0f, 3.0f, fill_seed(gen)}(a_m_k);
-        }
-        ck_tile::FillUniformDistribution<AQDataType>{-2.0f, 2.0f, fill_seed(gen)}(aq_tensor);
-        ck_tile::FillUniformDistribution<BDataType>{-5.0f, 5.0f, fill_seed(gen)}(b_k_n);
+            if constexpr(std::is_same_v<ADataType, ck_tile::pk_int4_t>)
+            {
+                ck_tile::FillUniformDistribution<ck_tile::pk_int4_t>{-5.0f, 5.0f, fill_seed(gen)}(
+                    a_m_k);
+            }
+            else
+            {
+                ck_tile::FillUniformDistribution<ADataType>{-2.0f, 3.0f, fill_seed(gen)}(a_m_k);
+            }
+            ck_tile::FillUniformDistribution<AQDataType>{-2.0f, 2.0f, fill_seed(gen)}(
+                *aq_tensor_ptr);
+            ck_tile::FillUniformDistribution<BDataType>{-5.0f, 5.0f, fill_seed(gen)}(b_k_n);
 
-        if constexpr(QuantMode == ck_tile::QuantType::RowColQuant)
-        {
-            ck_tile::FillUniformDistribution<BQDataType>{-2.0f, 2.0f, fill_seed(gen)}(
-                *bq_tensor_ptr);
+            if constexpr(QuantMode == ck_tile::QuantType::RowColQuant)
+            {
+                ck_tile::FillUniformDistribution<BQDataType>{-2.0f, 2.0f, fill_seed(gen)}(
+                    *bq_tensor_ptr);
+            }
         }
     }
     else if(init_method == 1)
@@ -267,46 +311,65 @@ int run_gemm_example_with_layouts(int argc,
     }
     else if(init_method == 2)
     {
-        ck_tile::FillConstant<ADataType>{static_cast<ADataType>(0x22)}(a_m_k);
-        ck_tile::FillConstant<AQDataType>{static_cast<AQDataType>(0.5f)}(aq_tensor);
-        ck_tile::FillConstant<BDataType>{static_cast<BDataType>(0x38)}(b_k_n);
-        if constexpr(QuantMode == ck_tile::QuantType::RowColQuant)
+        if constexpr(QuantMode == ck_tile::QuantType::BQuantGrouped)
         {
+            ck_tile::FillConstant<ADataType>{static_cast<ADataType>(0x38)}(a_m_k);
+            ck_tile::FillConstant<BDataType>{static_cast<BDataType>(0x22)}(b_k_n);
             ck_tile::FillConstant<BQDataType>{static_cast<BQDataType>(0.5f)}(*bq_tensor_ptr);
         }
+        else
+        {
+            ck_tile::FillConstant<ADataType>{static_cast<ADataType>(0x22)}(a_m_k);
+            ck_tile::FillConstant<AQDataType>{static_cast<AQDataType>(0.5f)}(*aq_tensor_ptr);
+            ck_tile::FillConstant<BDataType>{static_cast<BDataType>(0x38)}(b_k_n);
+
+            if constexpr(QuantMode == ck_tile::QuantType::RowColQuant)
+            {
+                ck_tile::FillConstant<BQDataType>{static_cast<BQDataType>(0.5f)}(*bq_tensor_ptr);
+            }
+        }
     }
     else
     {
         a_m_k.SetZero();
-        aq_tensor.SetZero();
+        aq_tensor_ptr->SetZero();
         b_k_n.SetZero();
-        if constexpr(QuantMode == ck_tile::QuantType::RowColQuant)
-        {
-            bq_tensor_ptr->SetZero();
-        }
+        bq_tensor_ptr->SetZero();
     }
 
     ck_tile::DeviceMem a_m_k_dev_buf(a_m_k.get_element_space_size_in_bytes());
-    ck_tile::DeviceMem aq_dev_buf(aq_tensor.get_element_space_size_in_bytes());
     ck_tile::DeviceMem b_k_n_dev_buf(b_k_n.get_element_space_size_in_bytes());
     ck_tile::DeviceMem c_m_n_dev_buf(c_m_n_dev_result.get_element_space_size_in_bytes());
 
+    std::unique_ptr<ck_tile::DeviceMem> aq_dev_buf_ptr = nullptr;
+    if constexpr(QuantMode == ck_tile::QuantType::AQuantGrouped ||
+                 QuantMode == ck_tile::QuantType::RowColQuant)
+    {
+        aq_dev_buf_ptr =
+            std::make_unique<ck_tile::DeviceMem>(aq_tensor_ptr->get_element_space_size_in_bytes());
+    }
+
     std::unique_ptr<ck_tile::DeviceMem> bq_dev_buf_ptr = nullptr;
-    if constexpr(QuantMode == ck_tile::QuantType::RowColQuant)
+    if constexpr(QuantMode == ck_tile::QuantType::BQuantGrouped ||
+                 QuantMode == ck_tile::QuantType::RowColQuant)
     {
         bq_dev_buf_ptr =
             std::make_unique<ck_tile::DeviceMem>(bq_tensor_ptr->get_element_space_size_in_bytes());
     }
 
-    if constexpr(GemmConfig::PreshuffleQuant && QuantMode == ck_tile::QuantType::AQuantGrouped)
+    if constexpr(QuantMode == ck_tile::QuantType::AQuantGrouped ||
+                 QuantMode == ck_tile::QuantType::RowColQuant)
     {
-        ck_tile::HostTensor<AQDataType> aq_shuffle_host =
-            shuffle_aq(aq_tensor, GemmConfig::K_Tile / QuantGroupSize);
-        aq_dev_buf.ToDevice(aq_shuffle_host.data());
-    }
-    else
-    {
-        aq_dev_buf.ToDevice(aq_tensor.data());
+        if constexpr(GemmConfig::PreshuffleQuant)
+        {
+            ck_tile::HostTensor<AQDataType> aq_shuffle_host =
+                shuffle_aq(aq_tensor_ptr.get(), GemmConfig::K_Tile / QuantGroupSize);
+            aq_dev_buf_ptr->ToDevice(aq_shuffle_host.data());
+        }
+        else
+        {
+            aq_dev_buf_ptr->ToDevice(aq_tensor_ptr->data());
+        }
     }
 
     if constexpr(std::is_same_v<ADataType, ck_tile::pk_int4_t>)
@@ -320,11 +383,22 @@ int run_gemm_example_with_layouts(int argc,
     {
         a_m_k_dev_buf.ToDevice(a_m_k.data());
     }
-    b_k_n_dev_buf.ToDevice(b_k_n.data());
+    if constexpr(std::is_same_v<BDataType, ck_tile::pk_int4_t>)
+    {
+        // Permute vector pk_i4x4 data for device implementation
+        ck_tile::HostTensor<BDataType> b_k_n_dev = b_k_n;
+        ck_tile::permute_vectors_i4x4_b(b_k_n_dev);
+        b_k_n_dev_buf.ToDevice(b_k_n_dev.data());
+    }
+    else
+    {
+        b_k_n_dev_buf.ToDevice(b_k_n.data());
+    }
     c_m_n_dev_buf.SetZero();
     c_m_n_dev_result.SetZero();
 
-    if constexpr(QuantMode == ck_tile::QuantType::RowColQuant)
+    if constexpr(QuantMode == ck_tile::QuantType::BQuantGrouped ||
+                 QuantMode == ck_tile::QuantType::RowColQuant)
     {
         bq_dev_buf_ptr->ToDevice(bq_tensor_ptr->data());
     }
@@ -338,11 +412,16 @@ int run_gemm_example_with_layouts(int argc,
                 CLayout,
                 QuantGroupSize,
                 QuantMode>(a_m_k_dev_buf,
-                           aq_dev_buf,
+                           (QuantMode == ck_tile::QuantType::AQuantGrouped ||
+                            QuantMode == ck_tile::QuantType::RowColQuant)
+                               ? aq_dev_buf_ptr.get()
+                               : nullptr,
                            b_k_n_dev_buf,
+                           (QuantMode == ck_tile::QuantType::BQuantGrouped ||
+                            QuantMode == ck_tile::QuantType::RowColQuant)
+                               ? bq_dev_buf_ptr.get()
+                               : nullptr,
                            c_m_n_dev_buf,
-                           (QuantMode == ck_tile::QuantType::RowColQuant) ? bq_dev_buf_ptr.get()
-                                                                          : nullptr,
                            M,
                            N,
                            K,
@@ -355,7 +434,9 @@ int run_gemm_example_with_layouts(int argc,
                            stride_C,
                            kbatch,
                            n_warmup,
-                           n_repeat);
+                           n_repeat,
+                           flush_cache,
+                           rotating_count);
 
     c_m_n_dev_buf.FromDevice(c_m_n_dev_result.data());
     bool pass = true;
@@ -374,7 +455,17 @@ int run_gemm_example_with_layouts(int argc,
                                           AccDataType,
                                           CDataType,
                                           QuantGroupSize,
-                                          true>(a_m_k, aq_tensor, b_k_n, c_m_n_host_ref);
+                                          true>(a_m_k, *aq_tensor_ptr, b_k_n, c_m_n_host_ref);
+        }
+        else if constexpr(QuantMode == ck_tile::QuantType::BQuantGrouped)
+        {
+            ck_tile::reference_gemm_quant<ADataType,
+                                          AQDataType,
+                                          BDataType,
+                                          AccDataType,
+                                          CDataType,
+                                          QuantGroupSize,
+                                          false>(a_m_k, *bq_tensor_ptr, b_k_n, c_m_n_host_ref);
         }
         else
         {
@@ -384,7 +475,7 @@ int run_gemm_example_with_layouts(int argc,
                                                  BQDataType,
                                                  AccDataType,
                                                  CDataType>(
-                a_m_k, aq_tensor, b_k_n, *bq_tensor_ptr, c_m_n_host_ref);
+                a_m_k, *aq_tensor_ptr, b_k_n, *bq_tensor_ptr, c_m_n_host_ref);
         }
 
         const float max_accumulated_value =
diff --git a/include/ck_tile/ops/gemm_group_quant/kernel/gemm_quant_kernel.hpp b/include/ck_tile/ops/gemm_group_quant/kernel/gemm_quant_kernel.hpp
index db41273f11..13fa0b8dfb 100644
--- a/include/ck_tile/ops/gemm_group_quant/kernel/gemm_quant_kernel.hpp
+++ b/include/ck_tile/ops/gemm_group_quant/kernel/gemm_quant_kernel.hpp
@@ -70,17 +70,21 @@ struct get_bq_data_type_or<T, Default>
     using type = typename T::BQDataType;
 };
 
-template <typename T, typename Default>
-struct get_preshuffle_or
-{
-    using type = Default;
+template <typename T>
+concept HasStaticPreshuffleQuant = requires {
+    { T::PreshuffleQuant } -> std::convertible_to<decltype(T::PreshuffleQuant)>;
 };
 
-template <typename T, typename Default>
-    requires requires { typename T::PreshuffleQuant; }
-struct get_preshuffle_or<T, Default>
+template <typename T>
+struct is_quantpreshuffle_enabled
 {
-    using type = typename T::PreshuffleQuant;
+    static constexpr bool value = false;
+};
+
+template <HasStaticPreshuffleQuant T>
+struct is_quantpreshuffle_enabled<T>
+{
+    static constexpr auto value = T::PreshuffleQuant;
 };
 } // namespace detail
 
@@ -198,9 +202,9 @@ struct QuantGemmKernel
     using BQLayout = remove_cvref_t<
         typename detail::get_bq_layout_or<GemmPipeline, typename GemmPipeline::BLayout>::type>;
 
-    static constexpr index_t kBlockSize   = GemmPipeline::BlockSize;
-    static constexpr bool PreshuffleQuant = remove_cvref_t<
-        typename detail::get_preshuffle_or<GemmPipeline, std::false_type>::type>::value;
+    static constexpr index_t kBlockSize = GemmPipeline::BlockSize;
+    static constexpr bool PreshuffleQuant =
+        detail::is_quantpreshuffle_enabled<GemmPipeline_>::value;
 
     using ADataType   = remove_cvref_t<typename GemmPipeline::ADataType>;
     using BDataType   = remove_cvref_t<typename GemmPipeline::BDataType>;