Merge branch 'develop' into universal_streamk

.azuredevops/rocm-ci.yml

@@ -23,20 +23,7 @@ trigger:
     - Jenkinsfile
     - LICENSE
 
-pr:
-  autoCancel: true
-  branches:
-    include:
-    - develop
-  paths:
-    exclude:
-    - .github
-    - docs
-    - '.*.y*ml'
-    - '*.md'
-    - Jenkinsfile
-    - LICENSE
-  drafts: false
+pr: none
 
 jobs:
   - template: ${{ variables.CI_COMPONENT_PATH }}/composable_kernel.yml@pipelines_repo

CMakeLists.txt

@@ -117,7 +117,7 @@ else()
     add_definitions(-DPROFILER_ONLY)
     set(GPU_TARGETS "" CACHE STRING "" FORCE)
     if(GPU_TARGETS)
-        message(FATAL_ERROR "For PROFILE_ONLY build, please do not set GPU_TARGETS, use GPU_ARCH = gfx90, gfx94, gfx10, or gfx11")
+        message(FATAL_ERROR "For PROFILE_ONLY build, please do not set GPU_TARGETS, use GPU_ARCH = gfx90, gfx94, gfx10, gfx11 or gfx12")
     endif()
     if(GPU_ARCH MATCHES "gfx90")
         rocm_check_target_ids(DEFAULT_GPU_TARGETS TARGETS "gfx908;gfx90a")
@@ -127,8 +127,10 @@ else()
         rocm_check_target_ids(DEFAULT_GPU_TARGETS TARGETS "gfx1030")
     elseif(GPU_ARCH MATCHES "gfx11")
         rocm_check_target_ids(DEFAULT_GPU_TARGETS TARGETS "gfx1100;gfx1101;gfx1102")
+    elseif(GPU_ARCH MATCHES "gfx12")
+        rocm_check_target_ids(DEFAULT_GPU_TARGETS TARGETS "gfx1200;gfx1201")
     else()
-        message(FATAL_ERROR "For PROFILE_ONLY build, please specify GPU_ARCH as gfx90, gfx94, gfx10, or gfx11")
+        message(FATAL_ERROR "For PROFILE_ONLY build, please specify GPU_ARCH as gfx90, gfx94, gfx10, gfx11 or gfx12")
     endif()
     set(GPU_TARGETS "${DEFAULT_GPU_TARGETS}" CACHE STRING " " FORCE)
 endif()

Jenkinsfile

@@ -493,6 +493,7 @@ def Build_CK(Map conf=[:]){
 
         def variant = env.STAGE_NAME
         def retimage
+
         gitStatusWrapper(credentialsId: "${env.status_wrapper_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCm', repo: 'composable_kernel') {
             try {
                 (retimage, image) = getDockerImage(conf)
@@ -660,9 +661,6 @@ CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true;ROCM
 
 pipeline {
     agent none
-    triggers {
-        parameterizedCron(CRON_SETTINGS)
-    }
     options {
         parallelsAlwaysFailFast()
     }

cmake/EnableCompilerWarnings.cmake

@@ -66,7 +66,7 @@ else()
             -Wunreachable-code
             -Wunused
             -Wno-reserved-identifier
-            -Werror
+	    -Werror
             -Wno-option-ignored
             -Wsign-compare
             -Wno-extra-semi-stmt

example/01_gemm/gemm_wmma_fp16.cpp

@@ -23,45 +23,45 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa
 
 // clang-format off
 using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmWmma_CShuffle
-         < ALayout,             
-           BLayout,             
-           CLayout,             
-           ADataType, 
+         < ALayout,
+           BLayout,
+           CLayout,
+           ADataType,
            BDataType,
-           CDataType, 
-           AccDataType, 
-           CShuffleDataType,  
-           AElementOp,  
-           BElementOp,  
-           CElementOp,    
-           GemmDefault, 
+           CDataType,
+           AccDataType,
+           CShuffleDataType,
+           AElementOp,
+           BElementOp,
+           CElementOp,
+           GemmDefault,
            1,           // Prefetch stage
            128,         // BlockSize
            64,          // MPerBlock
            128,         // NPerBlock
            64,          // KPerBlock
-           8,           // K1
+           2,           // K1
            16,          // MPerWmma
            16,          // NPerWmma
            2,           // M-Repeat // M-PerWmma / M-Repeat = M-Wave
            4,           // N-Repeat // N-PerWmma / N-Repeat = N-Wave
-           S<4, 32, 1>,     
-           S<1, 0, 2>,     
-           S<1, 0, 2>,              
-           2,              
-           8,              
-           8,      
-           true,     
-           S<4, 32, 1>,     
-           S<1, 0, 2>,     
-           S<1, 0, 2>,             
-           2,              
-           8,              
-           8,      
-           true,           
+           S<4, 32, 1>,
+           S<1, 0, 2>,
+           S<1, 0, 2>,
+           2,
+           2,
+           2,
+           true,
+           S<4, 32, 1>,
+           S<1, 0, 2>,
+           S<1, 0, 2>,
+           2,
+           2,
+           2,
+           true,
            1,           // C shuffle (M Repeat) Per store
            1,           // C shuffle (N Repeat) Per store
-           S<1, 32, 1,  4>,               
+           S<1, 32, 1,  4>,
            8>;
 // clang-format on
 

example/01_gemm/run_gemm_example.inc

@@ -159,7 +159,7 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
         ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n);
         break;
     case 4:
-        ck::utils::FillUniformDistributionIntegerValue<ADataType>{1.f, 1.f}(a_m_k);
+        ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k);
         ck::utils::FillUniformDistributionIntegerValue<BDataType>{1.f, 1.f}(b_k_n);
         break;
     case 5:

example/04_gemm_add_add_fastgelu/CMakeLists.txt

@@ -24,4 +24,4 @@ foreach(gpu IN LISTS GPU_TARGETS)
         add_example_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_lds_direct_load_fp32)
         set(target 1)
     endif()
-endforeach()
+endforeach()
\ No newline at end of file

example/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_wmma_fp16.cpp

@@ -83,14 +83,14 @@ using DeviceOpInstanceKKNN =
                                                                                   2,
                                                                                   4,
                                                                                   4,
-                                                                                  true,
+                                                                                  false,
                                                                                   S<4, 32, 1>,
                                                                                   S<1, 0, 2>,
                                                                                   S<1, 0, 2>,
                                                                                   2,
                                                                                   4,
                                                                                   4,
-                                                                                  true,
+                                                                                  false,
                                                                                   1,
                                                                                   1,
                                                                                   S<1, 64, 1, 2>,

example/32_batched_gemm_scale_softmax_gemm/cross_attention_forward_wmma_fp16.cpp

@@ -71,7 +71,7 @@ static constexpr auto TensorSpecC  = ck::tensor_operation::device::TensorSpecial
 #define CK_MHA_USE_WAVE_1
 #define CK_MHA_USE_WAVE_2
 #define CK_MHA_USE_WAVE_4
-#define CK_MHA_USE_WAVE_8
+//#define CK_MHA_USE_WAVE_8
 using DeviceMHAFactory = 
     std::tuple<
 #ifdef CK_MHA_USE_WAVE_1
@@ -277,10 +277,10 @@ using DeviceMHAFactory =
             S<2, 8, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 1, false,
             // CShuffleBlockTransfer MN
             1, 1, S<1, 64, 1, 2>, 8,             
-            MaskingSpec>,
+            MaskingSpec>
 #endif
 #ifdef CK_MHA_USE_WAVE_8
-        ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
+        ,ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
             NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
             ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
             AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,

example/32_batched_gemm_scale_softmax_gemm/self_attention_forward_wmma_fp16.cpp

@@ -71,7 +71,7 @@ static constexpr auto TensorSpecC  = ck::tensor_operation::device::TensorSpecial
 #define CK_MHA_USE_WAVE_1
 #define CK_MHA_USE_WAVE_2
 #define CK_MHA_USE_WAVE_4
-#define CK_MHA_USE_WAVE_8
+//#define CK_MHA_USE_WAVE_8
 using DeviceMHAFactory = 
     std::tuple<
 #ifdef CK_MHA_USE_WAVE_1
@@ -277,10 +277,10 @@ using DeviceMHAFactory =
             S<2, 8, 8>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 1, false,
             // CShuffleBlockTransfer MN
             1, 1, S<1, 64, 1, 2>, 8,             
-            MaskingSpec>,
+            MaskingSpec>
 #endif
 #ifdef CK_MHA_USE_WAVE_8
-        ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
+        ,ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle<
             NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,
             ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType,
             AElementOp, B0ElementOp, Acc0ElementOp, B1ElementOp, CElementOp,

example/CMakeLists.txt

@@ -67,7 +67,7 @@ function(add_example_executable EXAMPLE_NAME FILE_NAME)
     endforeach()
     #Do not build any WMMA examples if gfx11 targets are not on the list
     foreach(source IN LISTS FILE_NAME)
-        if(NOT EX_TARGETS MATCHES "gfx11" AND source MATCHES "_wmma")
+	if(NOT GPU_TARGETS MATCHES "gfx11" AND NOT GPU_TARGETS MATCHES "gfx12" AND source MATCHES "_wmma")
             message("removing wmma example ${source} ")
             list(REMOVE_ITEM FILE_NAME "${source}")
         endif()
@@ -154,7 +154,7 @@ function(add_example_executable_no_testing EXAMPLE_NAME FILE_NAME)
     endforeach()
     #Do not build any WMMA examples if gfx11 targets are not on the list
     foreach(source IN LISTS FILE_NAME)
-        if(NOT EX_TARGETS MATCHES "gfx11" AND source MATCHES "_wmma")
+	if(NOT GPU_TARGETS MATCHES "gfx11" AND NOT GPU_TARGETS MATCHES "gfx12" AND source MATCHES "_wmma")
             message("removing wmma example ${source} ")
             list(REMOVE_ITEM FILE_NAME "${source}")
         endif()

example/ck_tile/01_fmha/CMakeLists.txt

 # generate a list of kernels, but not actually emit files at config stage
 execute_process(
   COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/generate.py
-  --direction fwd --list_blobs ${CMAKE_CURRENT_BINARY_DIR}/fwd_blob_list.txt
+  --api fwd,fwd_splitkv --list_blobs ${CMAKE_CURRENT_BINARY_DIR}/fwd_blob_list.txt
 )
 
 execute_process(
   COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/generate.py
-  --direction bwd --list_blobs ${CMAKE_CURRENT_BINARY_DIR}/bwd_blob_list.txt
+  --api bwd --list_blobs ${CMAKE_CURRENT_BINARY_DIR}/bwd_blob_list.txt
 )
 
 # NOTE: for cmake, the FMHA_FWD_GEN_BLOBS/FMHA_BWD_GEN_BLOBS files must be in the same directory
@@ -17,13 +17,13 @@ file(STRINGS ${CMAKE_CURRENT_BINARY_DIR}/bwd_blob_list.txt FMHA_BWD_GEN_BLOBS)
 add_custom_command(
   OUTPUT ${FMHA_FWD_GEN_BLOBS}
   COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/generate.py
-  --direction fwd --output_dir ${CMAKE_CURRENT_BINARY_DIR}
+  --api fwd,fwd_splitkv --output_dir ${CMAKE_CURRENT_BINARY_DIR}
 )
 
 add_custom_command(
   OUTPUT ${FMHA_BWD_GEN_BLOBS}
   COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/generate.py
-  --direction bwd --output_dir ${CMAKE_CURRENT_BINARY_DIR}
+  --api bwd --output_dir ${CMAKE_CURRENT_BINARY_DIR}
 )
 
 set(EXAMPLE_FMHA_FWD "tile_example_fmha_fwd")

example/ck_tile/01_fmha/codegen/cmake_config.py

+# SPDX-License-Identifier: MIT
+# Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+# generate kernel instances to speed up compilation
+
+GEN_DIR = ""    # in Cmake, have to generate files in same folder
\ No newline at end of file

example/ck_tile/01_fmha/codegen/cpp_symbol_map.py

+# SPDX-License-Identifier: MIT
+# Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+# generate kernel instances to speed up compilation
+
+DTYPE_MAP = {
+    "fp16": "ck_tile::fp16_t",
+    "bf16": "ck_tile::bf16_t",
+    "fp8" : "ck_tile::fp8_t"
+}
+
+MASK_IMPL = {
+    "generic" : "ck_tile::GenericAttentionMask",
+    "simplified"  : "ck_tile::SimplifiedGenericAttentionMask"
+}
+
+_MASK_SIMPLIFIED_MAP = {
+    "s_no" : "ck_tile::SimplifiedGenericAttentionMask<false>",
+    "s_mask" : "ck_tile::SimplifiedGenericAttentionMask<true>",
+}
+
+_MASK_MAP = {
+    "no" : "FmhaMasks::NoMask",
+    "causal" : "FmhaMasks::CausalMask",
+    "generic" : "FmhaMasks::GenericMask"
+}
+
+def get_mask_map(mask : str):
+    if mask == "generic":
+        return _MASK_MAP
+    elif mask == "simplified":
+        return _MASK_SIMPLIFIED_MAP
+    else:
+        assert False
+        return None
+
+_MASK_CHECK_MAP = {
+    "no" : "t.mask_type == mask_enum::no_mask",
+    "causal" : "t.mask_type == mask_enum::mask_top_left || t.mask_type == mask_enum::mask_bottom_right",
+    "generic" : "t.mask_type == mask_enum::window_generic",
+}
+
+_MASK_SIMPLIFIED_CHECK_MAP = {
+    "s_no" : "t.mask_type == mask_enum::no_mask",
+    "s_mask" : "t.mask_type != mask_enum::no_mask",
+}
+
+def get_mask_check_map(mask : str):
+    if mask == "generic":
+        return _MASK_CHECK_MAP
+    elif mask == "simplified":
+        return _MASK_SIMPLIFIED_CHECK_MAP
+    else:
+        assert False
+        return None
+
+BIAS_MAP = {
+    "no" : "ck_tile::BlockAttentionBiasEnum::NO_BIAS",
+    "bias"  : "ck_tile::BlockAttentionBiasEnum::ELEMENTWISE_BIAS",
+    "alibi" : "ck_tile::BlockAttentionBiasEnum::ALIBI"
+}
+
+# TODO: this is ugly
+BIAS_CHECK_MAP = {
+    "no" : "bias_enum::no_bias",
+    "bias"  : "bias_enum::elementwise_bias",
+    "alibi" : "bias_enum::alibi"
+}
+
+MODE_MAP = {
+    "batch" : "false",
+    "group" : "true"
+}
+
+LAYOUT_MAP = {
+    "row" : "true",
+    "col" : "false"
+}
+
+PIPELINE_MAP = {
+    "qr" : "ck_tile::BlockFmhaPipelineQRKSVS",
+    "qr_async" : "ck_tile::BlockFmhaPipelineQRKSVSAsync",
+}
+
+PIPELINE_ENUM_MAP = {
+    "qr" : "ck_tile::BlockFmhaPipelineEnum::QRKSVS",
+    "qr_async" : "ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC",
+}
+
+BOOL_MAP = {
+    "t" : "true",
+    "f" : "false"
+}
\ No newline at end of file

example/ck_tile/01_fmha/codegen/ops/fmha_bwd.py

+# SPDX-License-Identifier: MIT
+# Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+# generate kernel instances to speed up compilation
+
+import copy
+from dataclasses import dataclass
+import fnmatch
+import itertools
+from pathlib import Path
+from typing import List, Optional, Tuple
+
+from codegen.cmake_config import *
+from codegen.cpp_symbol_map import *
+
+
+BWD_DQDKDV_PIPELINE_MAP = {
+    "ks_kts_vr" : "ck_tile::BlockFmhaBwdDQDKDVPipelineKSKTSVR",
+    "qs_ks_vr_dos" : "ck_tile::BlockFmhaBwdDQDKDVPipelineQSKSVROGradS",
+    "ks_vr" : "ck_tile::BlockFmhaBwdDQDKDVPipelineKSVR",
+}
+
+BWD_DQDKDV_PIPELINE_ENUM_MAP = {
+    "ks_kts_vr" : "ck_tile::BlockFmhaBwdPipelineEnum::KSKTSVR",
+    "qs_ks_vr_dos" : "ck_tile::BlockFmhaBwdPipelineEnum::QSKSVROGradS",
+    "ks_vr" : "ck_tile::BlockFmhaBwdPipelineEnum::KSVR",
+}
+
+FMHA_BWD_KERNEL_HEADER = """// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.\n
+// auto generated by generate.py
+#include "fmha_bwd.hpp"
+"""
+
+FMHA_BWD_DQ_DK_DV_KERNEL_BODY="""
+using fmha_dtype_{F_idx} = {F_dtype};
+
+using fmha_block_tile_{F_idx} = ck_tile::sequence<{F_bm0}, {F_bn0}, {F_bk0}, {F_bk1}, {F_bk2}, {F_bk3}, {F_bk4}, {F_bhdq}, {F_bhdv}>;
+using fmha_block_warps0_{F_idx} = ck_tile::sequence<{F_rm0}, {F_rn0}, {F_rk0}>;
+using fmha_block_warps1_{F_idx} = ck_tile::sequence<{F_rm1}, {F_rn1}, {F_rk1}>;
+using fmha_block_warps2_{F_idx} = ck_tile::sequence<{F_rm2}, {F_rn2}, {F_rk2}>;
+using fmha_warp_tile_{F_idx} = ck_tile::sequence<{F_wm}, {F_wn}, {F_wk}>;
+
+// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
+//       G0&G2 -> GSdP
+//       G1&G3 -> GdKV
+//       G4    -> GdQ
+using fmha_bwd_shape_{F_idx} = ck_tile::TileFmhaBwdShape<fmha_block_tile_{F_idx},
+                                      fmha_block_warps0_{F_idx},
+                                      fmha_warp_tile_{F_idx},
+                                      fmha_block_warps1_{F_idx},
+                                      fmha_warp_tile_{F_idx},
+                                      fmha_block_warps0_{F_idx},
+                                      fmha_warp_tile_{F_idx},
+                                      fmha_block_warps1_{F_idx},
+                                      fmha_warp_tile_{F_idx},
+                                      fmha_block_warps2_{F_idx},
+                                      fmha_warp_tile_{F_idx}>;
+
+using fmha_bwd_trait_{F_idx} = ck_tile::TileFmhaTraits<{F_spad},
+                                                    {F_skpad},
+                                                    {F_dpad},
+                                                    {F_dvpad},
+                                                    {F_bias},
+                                                    {F_dbias},
+                                                    false,
+                                                    {F_dropout},
+                                                    false,
+                                                    {F_occupancy}>;
+using fmha_mask_{F_idx} = {F_mask};
+
+using fmha_bwd_pipeline_problem_{F_idx} = ck_tile::BlockFmhaBwdPipelineProblem<
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::QDataType,
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::KDataType,
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::VDataType,
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::GemmDataType,
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::LSEDataType,
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::AccDataType,
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::DDataType,
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::BiasDataType,
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::RandValOutputDataType,
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::ODataType,
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::OGradDataType,
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::QGradDataType,
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::KGradDataType,
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::VGradDataType,
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::BiasGradDataType,
+    fmha_bwd_shape_{F_idx},
+    {F_mode},
+    fmha_mask_{F_idx},
+    fmha_bwd_trait_{F_idx}>;
+
+using fmha_bwd_pipeline_{F_idx} = {F_pipeline}<
+    fmha_bwd_pipeline_problem_{F_idx}>;
+
+using fmha_bwd_dk_epilogue_{F_idx} =
+    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<{F_dtype}>::AccDataType,
+                               typename FmhaBwdTypeConfig<{F_dtype}>::KGradDataType,
+                               false, false>>;
+
+using fmha_bwd_dv_epilogue_{F_idx} =
+    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<{F_dtype}>::AccDataType,
+                               typename FmhaBwdTypeConfig<{F_dtype}>::VGradDataType,
+                               false, false>>;
+
+using fmha_bwd_dq_dk_dv_kernel_{F_idx} =
+    ck_tile::FmhaBwdDQDKDVKernel<ck_tile::FmhaBwdTilePartitioner<fmha_bwd_shape_{F_idx}>,
+                        fmha_bwd_pipeline_{F_idx},
+                        fmha_bwd_dk_epilogue_{F_idx},
+                        fmha_bwd_dv_epilogue_{F_idx}>;
+
+using dq_dk_dv_trait_{F_idx} = fmha_bwd_dq_dk_dv_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_pipeline_enum}, fmha_mask_{F_idx}, {F_bias}, {F_dbias}, {F_dropout}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}>;
+
+#include <iostream>
+
+template<>
+float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_{F_idx}>(const ck_tile::stream_config& s, fmha_bwd_args a)
+{{
+    using k_ = fmha_bwd_dq_dk_dv_kernel_{F_idx};
+    if(s.log_level_ > 0)
+        std::cout << ", " << k_::GetName() << std::flush;
+    auto [kargs, grids] = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
+    constexpr dim3 blocks             = k_::BlockSize();
+    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
+    return ck_tile::launch_kernel(s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs));
+}}
+
+template<>
+void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_{F_idx}>(const ck_tile::stream_config& s, fmha_bwd_args a)
+{{
+    using k_ = fmha_bwd_dq_dk_dv_kernel_{F_idx};
+    auto [kargs, grids] = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
+    constexpr dim3 blocks             = k_::BlockSize();
+    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
+    ck_tile::make_kernel<blocks.x, kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs)(ck_tile::stream_config{{s.stream_id_}});
+}}
+
+template<>
+std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_{F_idx}>()
+{{
+    using k_ = fmha_bwd_dq_dk_dv_kernel_{F_idx};
+    return k_::GetName();
+}}
+"""
+
+FMHA_BWD_API_FILENAME="fmha_bwd_api.cpp"
+FMHA_BWD_API="""
+#include <iostream>
+
+template<typename dot_do_o_trait_, typename dq_dk_dv_trait_>
+float fmha_bwd_(const ck_tile::stream_config& s, fmha_bwd_args a)
+{{
+    if(s.log_level_ > 0)
+        std::cout << ", " << fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_>() << ", " << fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_>() << std::flush;
+    return ck_tile::launch_kernel(s,
+            [=](const ck_tile::stream_config& s_){{ fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_>(s_, a); }},
+            [=](const ck_tile::stream_config& s_){{ fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_>(s_, a); }}
+    );
+}}
+
+float fmha_bwd(fmha_bwd_traits t, fmha_bwd_args a, const ck_tile::stream_config& s){{
+    float r = -1;
+{F_dispatch}
+    return r;
+}}
+"""
+
+FMHA_BWD_API_PER_DTYPE="""    {F_if}(t.data_type.compare(\"{F_dtype}\") == 0){{
+{F_hdim_case}
+    }}
+"""
+FMHA_BWD_API_PER_HDIM_CASE="""        {F_if} (t.hdim_q <= {F_hdim} && t.hdim_v <= {F_hdim}) {{
+{F_inner_dispatch}
+        }}
+"""
+
+FMHA_BWD_API_INNER_DISPATCH="""            {F_if}((t.is_group_mode == {F_mode}) && ({F_mask_check}) && (t.bias_type == {F_bias_check}) && (t.has_dbias == {F_dbias}) && (t.has_dropout == {F_dropout}) &&
+                        ({F_scheck}) && ({F_skcheck}) && ({F_dcheck}) && ({F_dvcheck})) {{
+                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_pipeline_enum}, {F_mask}, {F_bias}, {F_dbias}, {F_dropout}, {F_spad0}, {F_skpad}, {F_dpad}, {F_dvpad}>;
+                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_spad1}, {F_dvpad}>;
+                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_>(s, a);
+                return r;
+            }}
+"""
+
+@dataclass
+class FmhaBwdDQDKDVApiTrait:
+    pipeline  : str
+    # sync with fmha_bwd_traits<>, to generate fallback calls
+    hdim      : str
+    dtype     : str  # data type
+    mode      : str  # value from MODE_MAP
+    bm0       : int  # tile size along q seqlen (block size)
+    bn0       : int  # tile size along k seqlen
+    bhdq      : int  # q head_dim
+    bhdv      : int  # v head_dim
+    mask      : str
+    bias      : str
+    dbias     : str
+    dropout   : str
+    spad      : str
+    skpad     : str
+    dpad      : str
+    dvpad     : str
+
+    @property
+    def name(self) -> str:
+        return f'{self.pipeline}-{self.hdim}-{self.dtype}-{self.mode}-{self.mask}-{self.bias}-{self.dbias}-{self.dropout}-{self.spad}-{self.skpad}-{self.dpad}-{self.dvpad}'
+
+    def scheck(self, spad1 : str) -> str:
+        if self.mode == 'group':
+            return 'true' # always support
+        elif self.spad == 't' and spad1 == 't':
+            return f'a.seqlen_q % {self.bm0} != 0'
+        elif self.spad == 'f' and spad1 == 't':
+            return f'a.seqlen_q % {self.bm0} == 0 and a.seqlen_q % 256 != 0' # BlockSize
+        else: # self.skpad == 'f' and skpad1 == 'f'
+            return f'a.seqlen_q % 256 == 0' # BlockSize
+
+    @property
+    def skcheck(self) -> str:
+        if self.mode == 'group':
+            return 'true' # always support
+        elif self.skpad == 't':
+            return f'a.seqlen_k % {self.bn0} != 0'
+        else:
+            return f'a.seqlen_k % {self.bn0} == 0'
+
+    @property
+    def dcheck(self) -> str:
+        if self.dpad == 't': return f'a.hdim_q % {self.bhdq} != 0'
+        else :               return f'a.hdim_q % {self.bhdq} == 0'
+
+    @property
+    def dvcheck(self) -> str:
+        if self.dvpad == 't': return f'a.hdim_v % {self.bhdv} != 0'
+        else :                return f'a.hdim_v % {self.bhdv} == 0'
+
+class FmhaBwdApiPool:
+    def __init__(self, mask_impl):
+        self.dq_dk_dv_pool = dict()
+        self.mask_impl = mask_impl
+
+    def register_dq_dk_dv_traits(self, trait : FmhaBwdDQDKDVApiTrait) -> None:
+        # TODO: do we need to check duplication?
+        if trait.dtype not in self.dq_dk_dv_pool.keys():
+            self.dq_dk_dv_pool[trait.dtype] = dict()
+        if trait.hdim not in self.dq_dk_dv_pool[trait.dtype].keys():
+            self.dq_dk_dv_pool[trait.dtype][trait.hdim] = list()
+
+        self.dq_dk_dv_pool[trait.dtype][trait.hdim].append(copy.copy(trait))
+
+    @property
+    def api(self) -> str:
+        per_dtypes=str()
+        for i, dtype in enumerate(self.dq_dk_dv_pool.keys()):
+            per_hdim_case=str()
+            for j, hdim in enumerate(self.dq_dk_dv_pool[dtype].keys()):
+                traits=self.dq_dk_dv_pool[dtype][hdim]
+                inners=str()
+                for k, trait in enumerate(traits):
+                    if_k = 'if' if k == 0 else 'else if'
+                    for spad1 in ["t", "f"]:
+                        if ((spad1 == "f" and trait.spad == "t") or (trait.mode == "group" and spad1 == "f")):
+                            continue
+                        inners = inners + FMHA_BWD_API_INNER_DISPATCH.format(F_if=if_k, F_mode=MODE_MAP[trait.mode], F_mask=get_mask_map(self.mask_impl)[trait.mask], F_pipeline_enum=BWD_DQDKDV_PIPELINE_ENUM_MAP[trait.pipeline],
+                                    F_mask_check=get_mask_check_map(self.mask_impl)[trait.mask], F_bias_check=BIAS_CHECK_MAP[trait.bias], F_bias=BIAS_MAP[trait.bias], F_dbias=BOOL_MAP[trait.dbias], F_dropout=BOOL_MAP[trait.dropout],
+                                    F_scheck=trait.scheck(spad1=spad1), F_skcheck=trait.skcheck, F_dcheck=trait.dcheck, F_dvcheck=trait.dvcheck, F_hdim=hdim, F_dtype=DTYPE_MAP[dtype],
+                                    F_spad0=BOOL_MAP[trait.spad], F_spad1=BOOL_MAP[spad1], F_skpad=BOOL_MAP[trait.skpad], F_dpad=BOOL_MAP[trait.dpad], F_dvpad=BOOL_MAP[trait.dvpad])
+
+                if_j = 'if' if j == 0 else 'else if'
+                per_hdim_case = per_hdim_case + FMHA_BWD_API_PER_HDIM_CASE.format(F_if=if_j, F_hdim=hdim, F_inner_dispatch=inners)
+            if_i = 'if' if i == 0 else 'else if'
+            per_dtypes = per_dtypes + FMHA_BWD_API_PER_DTYPE.format(F_if=if_i, F_dtype=dtype, F_hdim_case=per_hdim_case)
+
+        return FMHA_BWD_KERNEL_HEADER + FMHA_BWD_API.format(F_dispatch = per_dtypes)
+
+# GEMM0: Q@K=S^T
+# GEMM1: P^T@dO^T=dV(This was chosen as G1 to match fwd, but N1 must be equal to headdim_v)
+# GEMM2: dO@V=dP^T(This was chosen as G2 because of the calculation order)
+# GEMM3: dS^T@Q^T=dK(Similar to G1, but N3 must be equal to headdim_qk)
+# GEMM4: dS@K^T=dQ(N4 must be equal to headdim_qk)
+# Is it necessary to distinguish between K0~K4?
+@dataclass
+class FmhaBwdDQDKDVTileSize:
+    F_bm0       : int  # tile size along q seqlen (block size)
+    F_bn0       : int  # tile size along k seqlen
+    F_bk0       : int  # tile size along gemm0 unroll(F_bhdq)
+    F_bk1       : int  # tile size along gemm1 unroll(F_bm0)
+    F_bk2       : int  # tile size along gemm2 unroll(F_bhdv)
+    F_bk3       : int  # tile size along gemm3 unroll(F_bm0)
+    F_bk4       : int  # tile size along gemm4 unroll(F_bn0)
+    F_bhdq      : int  # q head_dim
+    F_bhdv      : int  # v head_dim
+    F_rm0       : int  # number of warps along q seqlen (block warps) in gemm0/gemm2
+    F_rn0       : int  # number of warps along k seqlen (block warps) in gemm0/gemm2
+    F_rk0       : int  # number of warps along gemm-k (not used) in gemm0/gemm2
+    F_rm1       : int  # number of warps along k seqlen (block warps) in gemm1/gemm3
+    F_rn1       : int  # number of warps along q seqlen (block warps) in gemm1/gemm3
+    F_rk1       : int  # number of warps along gemm-k (not used) in gemm1/gemm3
+    F_rm2       : int  # number of warps along k seqlen (block warps) in gemm4
+    F_rn2       : int  # number of warps along q seqlen (block warps) in gemm4
+    F_rk2       : int  # number of warps along gemm-k (not used) in gemm4
+    F_wm        : int  # warp size along m (warp size)
+    F_wn        : int  # warp size along n
+    F_wk        : int  # warp size along k
+    F_occupancy : int  # occupancy
+    @property
+    def name(self) -> str:
+        return f"b{self.F_bm0}x{self.F_bn0}x{self.F_bk0}x{self.F_bk1}x{self.F_bk2}x{self.F_bk3}x{self.F_bk4}x{self.F_bhdq}x{self.F_bhdv}" +\
+        f"_r{self.F_rm0}x{self.F_rn0}x{self.F_rk0}_r{self.F_rm1}x{self.F_rn1}x{self.F_rk1}_r{self.F_rm2}x{self.F_rn2}x{self.F_rk2}" +\
+        f"_w{self.F_wm}x{self.F_wn}x{self.F_wk}_o{self.F_occupancy}"
+
+@dataclass
+class FmhaBwdDQDKDVKernel:
+    F_idx       : int  # this is not a tunable, but a counter to differentiate symbol
+    F_hdim      : int  # hdim
+    F_dtype     : str  # data type
+    F_tile      : FmhaBwdDQDKDVTileSize
+    F_spad      : str  # true/false
+    F_skpad     : str  #
+    F_dpad      : str  #
+    F_dvpad     : str  #
+    F_bias      : str  #
+    F_dbias     : str  #
+    F_dropout   : str  #
+    F_mask      : str  # value from MASK_MAP
+    F_mode      : str  # value from MODE_MAP
+    F_pipeline  : str
+    mask_impl   : str
+
+    @property
+    def template(self) -> str:
+        return FMHA_BWD_KERNEL_HEADER + \
+            FMHA_BWD_DQ_DK_DV_KERNEL_BODY.format(
+                F_idx       = self.F_idx,
+                F_hdim      = self.F_hdim,
+                F_dtype     = DTYPE_MAP[self.F_dtype],
+                F_bm0       = self.F_tile.F_bm0,
+                F_bn0       = self.F_tile.F_bn0,
+                F_bk0       = self.F_tile.F_bk0,
+                F_bk1       = self.F_tile.F_bk1,
+                F_bk2       = self.F_tile.F_bk2,
+                F_bk3       = self.F_tile.F_bk3,
+                F_bk4       = self.F_tile.F_bk4,
+                F_bhdq      = self.F_tile.F_bhdq,
+                F_bhdv      = self.F_tile.F_bhdv,
+                F_rm0       = self.F_tile.F_rm0,
+                F_rn0       = self.F_tile.F_rn0,
+                F_rk0       = self.F_tile.F_rk0,
+                F_rm1       = self.F_tile.F_rm1,
+                F_rn1       = self.F_tile.F_rn1,
+                F_rk1       = self.F_tile.F_rk1,
+                F_rm2       = self.F_tile.F_rm2,
+                F_rn2       = self.F_tile.F_rn2,
+                F_rk2       = self.F_tile.F_rk2,
+                F_wm        = self.F_tile.F_wm,
+                F_wn        = self.F_tile.F_wn,
+                F_wk        = self.F_tile.F_wk,
+                F_spad      = BOOL_MAP[self.F_spad],
+                F_skpad     = BOOL_MAP[self.F_skpad],
+                F_dpad      = BOOL_MAP[self.F_dpad],
+                F_dvpad     = BOOL_MAP[self.F_dvpad],
+                F_bias      = BIAS_MAP[self.F_bias],
+                F_dbias     = BOOL_MAP[self.F_dbias],
+                F_dropout   = BOOL_MAP[self.F_dropout],
+                F_occupancy = self.F_tile.F_occupancy,
+                F_mask      = get_mask_map(self.mask_impl)[self.F_mask],
+                F_mode      = MODE_MAP[self.F_mode],
+                F_pipeline_enum = BWD_DQDKDV_PIPELINE_ENUM_MAP[self.F_pipeline],
+                F_pipeline  = BWD_DQDKDV_PIPELINE_MAP[self.F_pipeline])
+
+    @property
+    def name(self) -> str:
+        def pad_name() -> str:
+            n = ''
+            if self.F_spad == 't': n += 's'
+            if self.F_skpad == 't' : n += 'sk'
+            if self.F_dpad == 't' : n += 'd'
+            if self.F_dvpad == 't' : n += 'dv'
+            if n != '' : n = 'p' + n
+            return n
+        pn = pad_name()
+        n = f"fmha_bwd_d{self.F_hdim}_{self.F_dtype}_{self.F_mode}_" + self.F_tile.name
+        if pn != '' : n += f'_{pn}'
+        if self.F_bias != 'no' : n += f'_{self.F_bias}'
+        if self.F_dbias == 't' : n += '_dbias'
+        if self.F_mask[0:2] == 's_':
+            if self.F_mask == 's_mask': n += f'_mask'
+        else:
+            if self.F_mask != 'no' : n += f'_m{self.F_mask[0]}'
+        if self.F_dropout == 't' : n += '_dropout'
+        return n
+
+    @property
+    def filename(self) -> str:
+        return self.name + ".cpp"
+
+    def api_trait(self) -> FmhaBwdDQDKDVApiTrait:
+        return FmhaBwdDQDKDVApiTrait(pipeline=self.F_pipeline,
+                hdim=str(self.F_hdim),
+                dtype=self.F_dtype,
+                mode=self.F_mode,
+                bm0=self.F_tile.F_bm0,
+                bn0=self.F_tile.F_bn0,
+                bhdq=self.F_tile.F_bhdq,
+                bhdv=self.F_tile.F_bhdv,
+                mask=self.F_mask,
+                bias=self.F_bias,
+                dbias=self.F_dbias,
+                dropout=self.F_dropout,
+                spad=self.F_spad,
+                skpad=self.F_skpad,
+                dpad=self.F_dpad,
+                dvpad=self.F_dvpad)
+
+# TODO: design a more practical way to do it
+# this is current supported tile size & pipeline.
+def get_fmha_bwd_dq_dk_dv_tile_ppl_dict_from_dtype(dtype : str) -> Optional[dict]:
+    if dtype == 'fp16' or dtype == 'bf16':
+        return {
+            '32'  : [FmhaBwdDQDKDVTileSize(128, 128, 32, 32, 32, 32, 32,  32,  32, 1, 4, 1, 4, 1, 1, 4, 1, 1, 32, 32, 16, 1),
+                        "qs_ks_vr_dos"],
+            '64'  : [FmhaBwdDQDKDVTileSize( 64, 128, 32, 32, 32, 32, 32,  64,  64, 1, 4, 1, 4, 1, 1, 2, 2, 1, 32, 32, 16, 1),
+                        "qs_ks_vr_dos"],
+            '128' : [FmhaBwdDQDKDVTileSize( 64, 128, 32, 32, 32, 32, 32, 128, 128, 1, 4, 1, 4, 1, 1, 2, 2, 1, 32, 32, 16, 1),
+                        "ks_vr"]
+        }
+    else:
+        return None
+
+def get_bwd_dq_dk_dv_blobs(kernel_filter : Optional[str], receipt, mask_impl) -> Tuple[FmhaBwdApiPool, List[FmhaBwdDQDKDVKernel]]:
+    # TODO: we don't support tuning yet, so pick up one value for pad
+    #       support this in future
+    gen = list()
+    api_pool = FmhaBwdApiPool(mask_impl)
+
+    for dtype in DTYPE_MAP.keys():
+        d = get_fmha_bwd_dq_dk_dv_tile_ppl_dict_from_dtype(dtype)
+        if d == None:
+            continue
+        for hdim_str, mode, mask, bias, dbias, dropout, spad, skpad, dpad, dvpad in itertools.product(d.keys(), MODE_MAP.keys(), get_mask_map(mask_impl).keys(), BIAS_MAP.keys(), ["t", "f"], ["t", "f"], ["t", "f"], ["t", "f"], ["t", "f"], ["t", "f"]):
+            tile = d[hdim_str][0]
+            ppl = d[hdim_str][1]
+            hdim = int(hdim_str)
+            if (mode == "group") and (spad == "f" or skpad == "f"):
+                continue
+            if ((bias == "no" or bias == "alibi") and dbias == "t"):
+                continue
+            k = FmhaBwdDQDKDVKernel(F_idx=0, F_hdim=hdim, F_dtype=dtype, F_tile=tile,
+                                F_spad=spad, F_skpad=skpad, F_dpad=dpad, F_dvpad=dvpad,
+                                F_bias=bias, F_dbias=dbias, F_dropout=dropout, F_mask=mask, F_mode=mode,
+                                F_pipeline=ppl, mask_impl=mask_impl)
+            if kernel_filter != None:
+                if not fnmatch.fnmatch(k.name, kernel_filter):
+                    continue
+            if receipt == 2:
+                    cond = dtype in ['fp16', 'bf16']
+                    cond &= bias in ['no', 'alibi']
+                    if not cond:
+                        continue
+            api_pool.register_dq_dk_dv_traits(k.api_trait())
+            gen.append(k)
+
+    return (api_pool, gen)
+
+FMHA_BWD_DOT_DO_O_KERNEL_BODY="""
+using fmha_dtype_{F_idx} = {F_dtype};
+
+using fmha_bwd_dot_do_o_trait_{F_idx} = ck_tile::TileFmhaBwdOGradDotOTraits<{F_spad},
+                                                    {F_dvpad},
+                                                    {F_occupancy}>;
+
+using fmha_bwd_dot_do_o_pipeline_problem_{F_idx} = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::ODataType,
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::OGradDataType,
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::DDataType,
+    /* BlockSize = */ 256,
+    {F_hdim},
+    {F_mode},
+    fmha_bwd_dot_do_o_trait_{F_idx}>;
+
+using fmha_bwd_dot_do_o_{F_idx} = typename ck_tile::BlockFmhaBwdOGradDotO<
+    fmha_bwd_dot_do_o_pipeline_problem_{F_idx}>;
+
+using fmha_bwd_dot_do_o_kernel_{F_idx} =
+    ck_tile::FmhaBwdOGradDotOKernel<ck_tile::FmhaBwdOGradDotOTilePartitioner</* BlockSize = */ 256>,
+                                    fmha_bwd_dot_do_o_{F_idx}>;
+
+using dot_do_o_trait_{F_idx} = fmha_bwd_dot_do_o_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_spad}, {F_dvpad}>;
+
+#include <iostream>
+
+template<>
+float fmha_bwd_dot_do_o_<dot_do_o_trait_{F_idx}>(const ck_tile::stream_config& s, fmha_bwd_args a)
+{{
+    using k_ = fmha_bwd_dot_do_o_kernel_{F_idx};
+    if(s.log_level_ > 0)
+        std::cout << ", " << k_::GetName() << std::flush;
+    auto [kargs, grids] = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
+    constexpr dim3 blocks             = k_::BlockSize();
+    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
+    return ck_tile::launch_kernel(s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs));
+}}
+
+template<>
+void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_{F_idx}>(const ck_tile::stream_config& s, fmha_bwd_args a)
+{{
+    using k_ = fmha_bwd_dot_do_o_kernel_{F_idx};
+    auto [kargs, grids] = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
+    constexpr dim3 blocks             = k_::BlockSize();
+    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
+    ck_tile::make_kernel<blocks.x, kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs)(ck_tile::stream_config{{s.stream_id_}});
+}}
+
+template<>
+std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_{F_idx}>()
+{{
+    using k_ = fmha_bwd_dot_do_o_kernel_{F_idx};
+    return k_::GetName();
+}}
+"""
+
+@dataclass
+class FmhaBwdOGradDotOKernel:
+    F_idx       : int  # this is not a tunable, but a counter to differentiate symbol
+    F_hdim      : int  # hdim
+    F_dtype     : str  # data type
+    F_spad      : str  # true/false
+    F_dvpad     : str  #
+    F_mode      : str  # value from MODE_MAP
+    F_occupancy : int
+
+    @property
+    def template(self) -> str:
+        return FMHA_BWD_KERNEL_HEADER + \
+            FMHA_BWD_DOT_DO_O_KERNEL_BODY.format(
+                F_idx       = self.F_idx,
+                F_hdim      = self.F_hdim,
+                F_dtype     = DTYPE_MAP[self.F_dtype],
+                F_spad      = BOOL_MAP[self.F_spad],
+                F_dvpad     = BOOL_MAP[self.F_dvpad],
+                F_mode      = MODE_MAP[self.F_mode],
+                F_occupancy = self.F_occupancy)
+
+    @property
+    def name(self) -> str:
+        def pad_name() -> str:
+            n = ''
+            if self.F_spad == 't': n += 's'
+            if self.F_dvpad == 't' : n += 'dv'
+            if n != '' : n = 'p' + n
+            return n
+        pn = pad_name()
+        n = f"fmha_bwd_dot_do_o_d{self.F_hdim}_{self.F_dtype}_{self.F_mode}_o{self.F_occupancy}"
+        if pn != '' : n += f'_{pn}'
+        return n
+
+    @property
+    def filename(self) -> str:
+        return self.name + ".cpp"
+
+def get_bwd_dot_do_o_blobs() -> List[FmhaBwdOGradDotOKernel]:
+    # TODO: we don't support tuning yet, so pick up one value for pad/occupancy
+    #       support this in future
+    def get_occupancy(dtype, hdim):
+        return 2
+
+    gen = list()
+
+    for dtype in DTYPE_MAP.keys():
+        d = get_fmha_bwd_dq_dk_dv_tile_ppl_dict_from_dtype(dtype)
+        if d == None:
+            continue
+        for hdim_str, mode, spad, dvpad in itertools.product(d.keys(), MODE_MAP.keys(), ["t", "f"], ["t", "f"]):
+            hdim = int(hdim_str)
+            if (mode == "group" and spad == "f"):
+                continue
+            k = FmhaBwdOGradDotOKernel(F_idx=0, F_hdim=hdim, F_dtype=dtype,
+                                F_spad=spad, F_dvpad=dvpad, F_mode=mode,
+                                F_occupancy=get_occupancy(dtype, hdim))
+            gen.append(k)
+
+    return gen
+
+def write_single_bwd_dq_dk_dv_kernel(kernel: FmhaBwdDQDKDVKernel, autogen_dir: Path) -> None:
+    (autogen_dir / kernel.filename).write_text(kernel.template)
+
+def write_single_bwd_dot_do_o_kernel(kernel: FmhaBwdOGradDotOKernel, autogen_dir: Path) -> None:
+    (autogen_dir / kernel.filename).write_text(kernel.template)
+
+def write_bwd_api(api_pool : FmhaBwdApiPool, autogen_dir: Path) -> None:
+    (autogen_dir / FMHA_BWD_API_FILENAME).write_text(api_pool.api)
+
+def write_blobs(output_dir : Path, kernel_filter : Optional[str], receipt, mask_impl) -> None:
+    kernels = get_bwd_dot_do_o_blobs()
+    for kernel in kernels:
+        write_single_bwd_dot_do_o_kernel(kernel, output_dir)
+    api_pool, kernels = get_bwd_dq_dk_dv_blobs(kernel_filter, receipt, mask_impl)
+    for kernel in kernels:
+        write_single_bwd_dq_dk_dv_kernel(kernel, output_dir)
+    write_bwd_api(api_pool, output_dir)
+
+def list_blobs(file_path : Path, kernel_filter : Optional[str], receipt, mask_impl) -> None:
+    with file_path.open('a') as f:
+        kernels = get_bwd_dot_do_o_blobs()
+        for kernel in kernels:
+            f.write(str(file_path.parent / GEN_DIR / kernel.filename) + "\n")
+        _, kernels = get_bwd_dq_dk_dv_blobs(kernel_filter, receipt, mask_impl)
+        for kernel in kernels:
+            f.write(str(file_path.parent / GEN_DIR / kernel.filename) + "\n")
+        f.write(str(file_path.parent / GEN_DIR / FMHA_BWD_API_FILENAME) + "\n")
\ No newline at end of file

example/ck_tile/01_fmha/codegen/ops/fmha_fwd.py

+# SPDX-License-Identifier: MIT
+# Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+# generate kernel instances to speed up compilation
+
+import copy
+from dataclasses import dataclass
+import fnmatch
+import itertools
+from pathlib import Path
+from typing import List, Optional, Tuple
+
+from codegen.cmake_config import *
+from codegen.cpp_symbol_map import *
+
+
+DTYPE_BITS = {
+    "fp32": 32,
+    "fp16": 16,
+    "bf16": 16,
+    "fp8" : 8,
+    "bf8" : 8
+}
+
+TILE_PARTITIONER_MAP = {
+    "shb" : "ck_tile::FmhaFwdTilePartitioner_SHB",
+    "hbs" : "ck_tile::FmhaFwdTilePartitioner_HBS",
+}
+
+FMHA_FWD_KERNEL_HEADER = """// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.\n
+// auto generated by generate.py
+#include "fmha_fwd.hpp"
+"""
+
+FMHA_FWD_KERNEL_BODY="""
+using fmha_dtype_{F_idx} = {F_dtype};
+
+using fmha_block_tile_{F_idx} = ck_tile::sequence<{F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0blen}>;
+using fmha_block_warps_{F_idx} = ck_tile::sequence<{F_rm}, {F_rn}, {F_rk}>;
+using fmha_warp_tile_{F_idx} = ck_tile::sequence<{F_wm}, {F_wn}, {F_wk}>;
+
+using fmha_shape_{F_idx} = ck_tile::TileFmhaShape<fmha_block_tile_{F_idx},
+                                      fmha_block_warps_{F_idx},
+                                      fmha_warp_tile_{F_idx},
+                                      fmha_block_warps_{F_idx},
+                                      fmha_warp_tile_{F_idx},
+                                      {F_vlayout}>;
+
+using fmha_trait_{F_idx} = ck_tile::TileFmhaTraits<{F_spad},
+                                                    {F_skpad},
+                                                    {F_dpad},
+                                                    {F_dvpad},
+                                                    {F_bias},
+                                                    false,
+                                                    {F_lse},
+                                                    {F_dropout},
+                                                    {F_squant},
+                                                    {F_occupancy}>;
+using fmha_mask_{F_idx} = {F_mask};
+
+using fmha_pipeline_problem_{F_idx} = ck_tile::BlockFmhaPipelineProblem<
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::QDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::KDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::VDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::SaccDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::SMPLComputeDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::BiasDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::RandValOutputDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::LSEDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::PDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::OaccDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::ODataType,
+    fmha_shape_{F_idx},
+    {F_mode},
+    fmha_mask_{F_idx},
+    fmha_trait_{F_idx}>;
+
+using fmha_pipeline_{F_idx} = {F_pipeline}<
+    fmha_pipeline_problem_{F_idx}>;
+
+using fmha_epilogue_{F_idx} =
+    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<{F_dtype}>::OaccDataType,
+                                           typename FmhaFwdTypeConfig<{F_dtype}>::ODataType,
+                                           {F_spad}, {F_dvpad}>>;
+
+using fmha_kernel_{F_idx} =
+    ck_tile::FmhaFwdKernel<{F_tile_partitioner}<fmha_shape_{F_idx}>,
+                  fmha_pipeline_{F_idx},
+                  fmha_epilogue_{F_idx}>;
+
+using trait_{F_idx} = fmha_fwd_traits_<{F_hdim}, {F_dtype}, {F_mode},{F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0blen}, {F_vlayout},
+                        {F_pipeline_enum}, fmha_mask_{F_idx}, {F_bias}, {F_lse}, {F_dropout}, {F_squant}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}>;
+
+#include <iostream>
+
+template<>
+float fmha_fwd_<trait_{F_idx}>(const ck_tile::stream_config& s, fmha_fwd_args a)
+{{
+    using k_ = fmha_kernel_{F_idx};
+    if(s.log_level_ > 0)
+        std::cout << ", " << k_::GetName() << std::flush;
+    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
+    constexpr dim3 blocks             = k_::BlockSize();
+    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
+    return ck_tile::launch_kernel(s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs));
+}}
+"""
+
+FMHA_FWD_API_FILENAME="fmha_fwd_api.cpp"
+FMHA_FWD_API="""
+float fmha_fwd(fmha_fwd_traits t, fmha_fwd_args a, const ck_tile::stream_config& s){{
+    float r = -1;
+{F_dispatch}
+    return r;
+}}
+"""
+
+FMHA_FWD_API_PER_DTYPE="""    {F_if}(t.data_type.compare(\"{F_dtype}\") == 0){{
+{F_hdim_case}
+    }}
+"""
+FMHA_FWD_API_PER_HDIM_CASE="""        {F_if} (t.hdim_q <= {F_hdim} && t.hdim_v <= {F_hdim}) {{
+{F_inner_dispatch}
+        }}
+"""
+
+FMHA_FWD_API_INNER_DISPATCH="""            {F_if}((t.is_group_mode == {F_mode}) && (t.is_v_rowmajor == {F_vlayout}) && ({F_mask_check}) && (t.bias_type == {F_bias_check}) && (t.has_lse == {F_lse})  && (t.has_dropout == {F_dropout}) && (t.do_fp8_static_quant == {F_squant}) &&
+                        ({F_scheck}) && ({F_skcheck}) && ({F_dcheck}) && ({F_dvcheck})) {{
+                using trait_ = fmha_fwd_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0blen}, {F_vlayout}, {F_pipeline_enum}, {F_mask}, {F_bias}, {F_lse}, {F_dropout}, {F_squant}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}>;
+                return fmha_fwd_<trait_>(s, a);
+            }}
+"""
+
+@dataclass
+class FmhaFwdApiTrait:
+    pipeline_tag : str
+    # sync with fmha_fwd_traits<>, to generate fallback calls
+    hdim      : str
+    dtype     : str  # data type
+    mode      : str  # value from MODE_MAP
+    bm0       : int  # tile size along q seqlen (block size)
+    bn0       : int  # tile size along qk seqlen
+    bk0       : int  # tile size along qk gemm unroll
+    bn1       : int  # tile size along v head_dim
+    bk1       : int  # tile size along kv gemm unroll
+    bk0blen   : int
+    vlayout   : str
+    mask      : str
+    bias      : str  #
+    lse       : str  #
+    dropout   : str
+    squant    : str  #
+    spad      : str
+    skpad     : str
+    dpad      : str
+    dvpad     : str
+
+    @property
+    def name(self) -> str:
+        return f'{self.hdim}-{self.dtype}-{self.mode}-{self.bm0}-{self.bn0}-{self.bk0}-{self.bn0}-{self.bk1}-{self.bk0blen}-'+\
+                    f'{self.vlayout}-{self.mask}-{self.bias}-{self.lse}-{self.dropout}-{self.squant}-{self.spad}-{self.skpad}-{self.dpad}-{self.dvpad}'
+
+    @property
+    def scheck(self) -> str:
+        if self.mode == 'group': return 'true/*group mode spad always true*/'                  # group mode only generate spad/skpad == true
+        if self.pipeline_tag == 'qr_async':
+            if self.spad == 't' : return 'true' # always support
+            else :                return 'true'
+        elif self.pipeline_tag in ['qr']:
+            if self.spad == 't' : return f'true /*a.seqlen_q % {self.bm0} != 0*/'  # TODO: order of get_pipelines() matters! (ugly)
+            else :                return f'a.seqlen_q % {self.bm0} == 0'
+        else: assert False
+
+    @property
+    def skcheck(self) -> str:
+        if self.mode == 'group': return 'true/*group mode skpad always true*/'                  # group mode only generate spad/skpad == true
+        if self.pipeline_tag == 'qr_async':
+            if self.skpad == 't' : return f'a.seqlen_k == 0 || a.seqlen_k % {self.bn0} != 0'
+            else :                 return f'a.seqlen_k != 0 && a.seqlen_k % {self.bn0} == 0'
+        elif self.pipeline_tag in ['qr', 'qr_fp8']:
+            if self.skpad == 't' : return f'true /*a.seqlen_k % {self.bn0} != 0*/' # TODO: order of get_pipelines() matters! (ugly)
+            else :                return f'a.seqlen_k % {self.bn0} == 0'
+        else: assert False
+
+    @property
+    def dcheck(self) -> str:
+        if self.pipeline_tag == 'qr_async':
+            vec = int((32 * 4) / DTYPE_BITS[self.dtype])
+            if self.dpad == 't': return f'a.hdim_q % {vec} == 0'
+            else :               assert False
+        elif self.pipeline_tag in ['qr']:
+            if self.dpad == 't': return f'true /*a.hdim_q % {self.bk0blen} != 0*/' # TODO: order of get_pipelines() matters! (ugly)
+            else :               return f'a.hdim_q % {self.bk0blen} == 0'
+        else:   assert False
+
+    @property
+    def dvcheck(self) -> str:
+        if self.pipeline_tag == 'qr_async':
+            vec = int((32 * 4) / DTYPE_BITS[self.dtype])
+            if self.dvpad == 't': return f'a.hdim_v % {vec} == 0'
+            else :                assert False
+        elif self.pipeline_tag in ['qr']:
+            if self.dvpad == 't': return f'true /*a.hdim_v % {self.bk0blen} != 0*/' # TODO: order of get_pipelines() matters! (ugly)
+            else :                return f'a.hdim_v % {self.bk0blen} == 0'
+        else:   assert False
+
+@dataclass
+class FmhaFwdPipeline:
+    tag : str
+
+    F_vlayout   : str  # row/col
+    F_spad      : str  # true/false
+    F_skpad     : str  #
+    F_dpad      : str  #
+    F_dvpad     : str  #
+    F_bias      : str  # true/false
+    F_lse       : str  #
+    F_dropout   : str  #
+    F_squant    : str  #
+    F_mask      : str  # value from MASK_MAP
+
+    @property
+    def name(self) -> str:
+        def pad_name() -> str:
+            n = ''
+            if self.F_spad == 't': n += 's'
+            if self.F_skpad == 't' : n += 'sk'
+            if self.F_dpad == 't' : n += 'd'
+            if self.F_dvpad == 't' : n += 'dv'
+            if n != '' : n = 'p' + n
+            return n
+        pn = pad_name()
+        n = f'{self.tag}_v{self.F_vlayout[0]}'
+        if pn != '' : n += f'_{pn}'
+        if self.F_bias != 'no' : n += f'_{self.F_bias}'
+        if self.F_mask[0:2] == 's_':
+            if self.F_mask == 's_mask': n += f'_mask'
+        else:
+            if self.F_mask != 'no' : n += f'_m{self.F_mask[0]}'
+        if self.F_lse == 't' : n += '_lse'
+        if self.F_dropout == 't' : n += '_dropout'
+        if self.F_squant == 't' : n += '_squant'
+        return n
+
+class FmhaFwdApiPool:
+    def __init__(self, mask_impl):
+        self.pool = dict()
+        self.mask_impl = mask_impl
+
+    def register_traits(self, trait : FmhaFwdApiTrait) -> None:
+        # TODO: do we need to check duplication?
+        if trait.dtype not in self.pool.keys():
+            self.pool[trait.dtype] = dict()
+        if trait.hdim not in self.pool[trait.dtype].keys():
+            self.pool[trait.dtype][trait.hdim] = list()
+
+        self.pool[trait.dtype][trait.hdim].append(copy.copy(trait))
+
+    @property
+    def api(self) -> str:
+        per_dtypes=str()
+        for i, dtype in enumerate(self.pool.keys()):
+            per_hdim_case=str()
+            for j, hdim in enumerate(self.pool[dtype].keys()):
+                traits=self.pool[dtype][hdim]
+                inners=str()
+                for k, trait in enumerate(traits):
+                    if_k = 'if' if k == 0 else 'else if'
+                    inners = inners + FMHA_FWD_API_INNER_DISPATCH.format(F_if=if_k, F_mode=MODE_MAP[trait.mode], F_vlayout=LAYOUT_MAP[trait.vlayout],
+                                   F_pipeline_enum=PIPELINE_ENUM_MAP[trait.pipeline_tag], F_mask=get_mask_map(self.mask_impl)[trait.mask],
+                                   F_mask_check=get_mask_check_map(self.mask_impl)[trait.mask], F_bias_check=BIAS_CHECK_MAP[trait.bias], F_bias=BIAS_MAP[trait.bias],
+                                   F_lse=BOOL_MAP[trait.lse], F_dropout=BOOL_MAP[trait.dropout] ,
+                                   F_squant=BOOL_MAP[trait.squant], F_scheck=trait.scheck, F_skcheck=trait.skcheck, F_dcheck=trait.dcheck, F_dvcheck=trait.dvcheck,
+                                   F_spad=BOOL_MAP[trait.spad], F_skpad=BOOL_MAP[trait.skpad], F_dpad=BOOL_MAP[trait.dpad], F_dvpad=BOOL_MAP[trait.dvpad],
+                                   F_bm0=trait.bm0, F_bn0=trait.bn0, F_bk0=trait.bk0, F_bn1=trait.bn1, F_bk1=trait.bk1, F_bk0blen=trait.bk0blen,
+                                   F_hdim=hdim, F_dtype=DTYPE_MAP[dtype])
+                if_j = 'if' if j == 0 else 'else if'
+                per_hdim_case = per_hdim_case + FMHA_FWD_API_PER_HDIM_CASE.format(F_if=if_j, F_hdim=hdim, F_inner_dispatch=inners)
+            if_i = 'if' if i == 0 else 'else if'
+            per_dtypes = per_dtypes + FMHA_FWD_API_PER_DTYPE.format(F_if=if_i, F_dtype=dtype, F_hdim_case=per_hdim_case)
+        return FMHA_FWD_KERNEL_HEADER + FMHA_FWD_API.format(F_dispatch = per_dtypes)
+
+@dataclass
+class FmhaFwdTileSize:
+    F_bm0       : int  # tile size along q seqlen (block size)
+    F_bn0       : int  # tile size along k seqlen
+    F_bk0       : int  # tile size along qk gemm unroll
+    F_bn1       : int  # tile size along v head_dim
+    F_bk1       : int  # tile size along kv gemm unroll
+    F_bk0blen   : int  # total length of K0, used for pipeline that need load Q at once (or repeately load Q as a whole tile)
+    F_rm        : int  # number of warps along q seqlen (block warps)
+    F_rn        : int  # number of warps along k seqlen(not used)
+    F_rk        : int  # number of warps along gemm-k(not used)
+    F_wm        : int  # warp size along m (warp size)
+    F_wn        : int  # warp size along n
+    F_wk        : int  # warp size along k
+    F_occupancy : int  # occupancy, -1 will let pipeline decide the occupancy, other value will overwrite occupancy
+    @property
+    def name(self) -> str:
+        return f"b{self.F_bm0}x{self.F_bn0}x{self.F_bk0}x{self.F_bn1}x{self.F_bk1}x{self.F_bk0blen}" +\
+        f"_r{self.F_rm}x{self.F_rn}x{self.F_rk}_w{self.F_wm}x{self.F_wn}x{self.F_wk}" +\
+            ("" if self.F_occupancy == -1 else f"_o{self.F_occupancy}")
+
+@dataclass
+class FmhaFwdKernel:
+    F_idx           : int  # this is not a tunable, but a counter to differentiate symbol
+    F_hdim          : int  # hdim
+    F_dtype         : str  # data type
+    F_mode          : str  # value from MODE_MAP
+    F_tile          : FmhaFwdTileSize
+    F_pipeline      : FmhaFwdPipeline
+    mask_impl       : str
+
+    def get_tp(self) -> str:
+        if self.F_mode == 'group':
+            return 'hbs'
+        else:
+            return 'shb'
+
+    @property
+    def template(self) -> str:
+        kernel_body = str()
+        return FMHA_FWD_KERNEL_HEADER + \
+            FMHA_FWD_KERNEL_BODY.format(
+                F_idx           = self.F_idx,
+                F_hdim          = self.F_hdim,
+                F_dtype         = DTYPE_MAP[self.F_dtype],
+                F_bm0           = self.F_tile.F_bm0,
+                F_bn0           = self.F_tile.F_bn0,
+                F_bk0           = self.F_tile.F_bk0,
+                F_bn1           = self.F_tile.F_bn1,
+                F_bk1           = self.F_tile.F_bk1,
+                F_bk0blen       = self.F_tile.F_bk0blen,
+                F_rm            = self.F_tile.F_rm,
+                F_rn            = self.F_tile.F_rn,
+                F_rk            = self.F_tile.F_rk,
+                F_wm            = self.F_tile.F_wm,
+                F_wn            = self.F_tile.F_wn,
+                F_wk            = self.F_tile.F_wk,
+                F_vlayout       = LAYOUT_MAP[self.F_pipeline.F_vlayout],
+                F_spad          = BOOL_MAP[self.F_pipeline.F_spad],
+                F_skpad         = BOOL_MAP[self.F_pipeline.F_skpad],
+                F_dpad          = BOOL_MAP[self.F_pipeline.F_dpad],
+                F_dvpad         = BOOL_MAP[self.F_pipeline.F_dvpad],
+                F_bias          = BIAS_MAP[self.F_pipeline.F_bias],
+                F_lse           = BOOL_MAP[self.F_pipeline.F_lse],
+                F_dropout       = BOOL_MAP[self.F_pipeline.F_dropout],
+                F_squant        = BOOL_MAP[self.F_pipeline.F_squant],
+                F_occupancy     = self.F_tile.F_occupancy,
+                F_pipeline_enum = PIPELINE_ENUM_MAP[self.F_pipeline.tag],
+                F_mask          = get_mask_map(self.mask_impl)[self.F_pipeline.F_mask],
+                F_mode          = MODE_MAP[self.F_mode],
+                F_pipeline      = PIPELINE_MAP[self.F_pipeline.tag],
+                F_tile_partitioner = TILE_PARTITIONER_MAP[self.get_tp()])
+
+    @property
+    def name(self) -> str:
+        # TODO: we don't encode idx here
+        return f"fmha_fwd_d{self.F_hdim}_{self.F_dtype}_{self.F_mode}_{self.get_tp()}_" + \
+                self.F_tile.name + '_' + self.F_pipeline.name
+
+    @property
+    def filename(self) -> str:
+        return self.name + ".cpp"
+
+    def api_trait(self) -> FmhaFwdApiTrait:
+        return FmhaFwdApiTrait(
+                pipeline_tag=self.F_pipeline.tag,
+                hdim=str(self.F_hdim),
+                dtype=self.F_dtype,
+                mode=self.F_mode,
+                bm0=self.F_tile.F_bm0,
+                bn0=self.F_tile.F_bn0,
+                bk0=self.F_tile.F_bk0,
+                bn1=self.F_tile.F_bn1,
+                bk1=self.F_tile.F_bk1,
+                bk0blen=self.F_tile.F_bk0blen,
+                vlayout=self.F_pipeline.F_vlayout,
+                mask=self.F_pipeline.F_mask,
+                bias=self.F_pipeline.F_bias,
+                lse=self.F_pipeline.F_lse,
+                dropout=self.F_pipeline.F_dropout,
+                squant=self.F_pipeline.F_squant,
+                spad=self.F_pipeline.F_spad,
+                skpad=self.F_pipeline.F_skpad,
+                dpad=self.F_pipeline.F_dpad,
+                dvpad=self.F_pipeline.F_dvpad)
+
+# TODO: design a more practical way to do it
+# this is current supported tile size per hdim
+def get_fmha_fwd_tile_dict_from_dtype(dtype : str) -> Optional[dict]:
+    if dtype == 'fp16' or dtype == 'bf16':
+        return {
+            '32'  : FmhaFwdTileSize(128, 64, 16, 32, 32, 32,     2, 1, 1, 32, 32, 16, -1),
+            '64'  : FmhaFwdTileSize(128, 64, 32, 64, 32, 64,     4, 1, 1, 32, 32, 16, -1),
+            '128' : FmhaFwdTileSize(128, 128, 32, 128, 32, 128,  4, 1, 1, 32, 32, 16, -1),
+            '256' : FmhaFwdTileSize(128, 128, 32, 256, 32, 256,  4, 1, 1, 32, 32, 16, -1),
+        }
+    elif dtype == 'fp8' or dtype == 'bf8':
+        return {
+            '64'  : FmhaFwdTileSize(128, 64, 32, 64, 32, 64,     2, 1, 1, 32, 32, 32, -1),
+            '128' : FmhaFwdTileSize(128, 128, 32, 128, 32, 128,  4, 1, 1, 32, 32, 32, -1),
+            '256' : FmhaFwdTileSize(128, 128, 32, 256, 32, 256,  4, 1, 1, 32, 32, 32, -1)
+        }
+    else:
+        return None
+
+def get_fwd_blobs(kernel_filter : Optional[str], receipt, mask_impl) -> Tuple[FmhaFwdApiPool, List[FmhaFwdKernel]]:
+    # TODO: we don't support tuning yet, so pick up one value for vlayout/pipeline/pad
+    #       support this in future
+    def get_pipelines(dtype, hdim) -> List[FmhaFwdPipeline]:
+        # this function will populate a list possible pipelines
+        # TODO: the order of List matters! the later in this list will be also be checked later
+        # TODO: currently for qr pipeline, let 't' padding to appear later!!
+        # TODO: how to design this more generic?
+        squant = 't' if dtype == 'fp8' else 'f'
+        pipelines = []
+        if dtype in ['fp16', 'bf16']:
+            for mask, bias, lse, dropout in itertools.product(get_mask_map(mask_impl).keys(), BIAS_MAP.keys(), ["t", "f"], ["t", "f"]):
+                if hdim == 256:
+                # if True:
+                    pipelines.append(FmhaFwdPipeline('qr', 'row', 'f', 'f', 'f', 'f', bias, lse, dropout, squant, mask))
+                    pipelines.append(FmhaFwdPipeline('qr', 'col', 'f', 'f', 'f', 'f', bias, lse, dropout, squant, mask))
+
+                    pipelines.append(FmhaFwdPipeline('qr', 'row', 't', 't', 't', 't', bias, lse, dropout, squant, mask))
+                    pipelines.append(FmhaFwdPipeline('qr', 'col', 't', 't', 't', 't', bias, lse, dropout, squant, mask))
+                else:
+                    pipelines.append(FmhaFwdPipeline('qr_async', 'row', 't', 'f', 't', 't', bias, lse, dropout, squant, mask))
+                    pipelines.append(FmhaFwdPipeline('qr_async', 'row', 't', 't', 't', 't', bias, lse, dropout, squant, mask))
+                    pipelines.append(FmhaFwdPipeline('qr_async', 'col', 't', 'f', 't', 't', bias, lse, dropout, squant, mask))
+                    pipelines.append(FmhaFwdPipeline('qr_async', 'col', 't', 't', 't', 't', bias, lse, dropout, squant, mask))
+                    if receipt == 1:
+                        pipelines.append(FmhaFwdPipeline('qr', 'row', 't', 't', 't', 't', bias, lse, dropout, squant, mask)) # TODO: cover arbitraty hdim
+                        pipelines.append(FmhaFwdPipeline('qr', 'col', 't', 'f', 't', 't', bias, lse, dropout, squant, mask)) # TODO: cover arbitraty hdim
+        elif dtype in ['fp8', 'bf8']:
+            # no need lse/dropout kernels
+            for mask, bias in itertools.product(get_mask_map(mask_impl).keys(), BIAS_MAP.keys()):
+                pipelines.append(FmhaFwdPipeline('qr', 'col', 'f', 'f', 'f', 'f', bias, 'f', 'f', squant, mask))
+        else:
+            assert False
+        return pipelines
+
+    gen = list()
+    api_pool = FmhaFwdApiPool(mask_impl)
+
+    for dtype in DTYPE_MAP.keys():
+        d = get_fmha_fwd_tile_dict_from_dtype(dtype)
+        if d == None:
+            continue
+        #for hdim_str, mode, mask, bias, lse in itertools.product(d.keys(), MODE_MAP.keys(), MASK_MAP.keys(), ["t", "f"], ["t", "f"]):
+        for hdim_str, mode in itertools.product(d.keys(), MODE_MAP.keys()):
+            tile = d[hdim_str]
+            hdim = int(hdim_str)
+            for pipeline in get_pipelines(dtype, hdim):
+                if mode == "group":
+                    if pipeline.F_spad != 't' or pipeline.F_skpad != 't':
+                        # in group mode, spad/skpad must be true, since we can't predict if seqlen of current batch need pad or not
+                        continue
+                k = FmhaFwdKernel(F_idx=0,
+                                  F_hdim=hdim,
+                                  F_dtype=dtype,
+                                  F_mode=mode,
+                                  F_tile=tile,
+                                  F_pipeline=pipeline,
+                                  mask_impl=mask_impl)
+                if kernel_filter != None:
+                    if not fnmatch.fnmatch(k.name, kernel_filter):
+                        continue
+                if receipt == 2:
+                    cond = dtype in ['fp16', 'bf16']
+                    cond &= pipeline.F_vlayout == 'row'
+                    cond &= pipeline.F_bias in ['no', 'alibi']
+                    cond &= pipeline.F_squant == 'f'
+                    if not cond:
+                        continue
+                api_pool.register_traits(k.api_trait())
+                gen.append(k)
+
+    return (api_pool, gen)
+
+def write_single_fwd_kernel(kernel: FmhaFwdKernel, autogen_dir: Path) -> None:
+    (autogen_dir / kernel.filename).write_text(kernel.template)
+
+def write_fwd_api(api_pool : FmhaFwdApiPool, autogen_dir: Path) -> None:
+    (autogen_dir / FMHA_FWD_API_FILENAME).write_text(api_pool.api)
+
+def write_blobs(output_dir : Path, kernel_filter : Optional[str], receipt, mask_impl) -> None:
+    api_pool, kernels = get_fwd_blobs(kernel_filter, receipt, mask_impl)
+    for kernel in kernels:
+        write_single_fwd_kernel(kernel, output_dir)
+    write_fwd_api(api_pool, output_dir)
+
+def list_blobs(file_path : Path, kernel_filter : Optional[str], receipt, mask_impl) -> None:
+    with file_path.open('a') as f:
+        _, kernels = get_fwd_blobs(kernel_filter, receipt, mask_impl)
+        for kernel in kernels:
+            f.write(str(file_path.parent / GEN_DIR / kernel.filename) + "\n")
+        f.write(str(file_path.parent / GEN_DIR / FMHA_FWD_API_FILENAME) + "\n")
\ No newline at end of file

example/ck_tile/01_fmha/codegen/ops/fmha_fwd_splitkv.py

+# SPDX-License-Identifier: MIT
+# Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+# generate kernel instances to speed up compilation
+
+import copy
+from dataclasses import dataclass
+import fnmatch
+import itertools
+from pathlib import Path
+from typing import List, Optional, Tuple, Union
+
+from codegen.cmake_config import *
+from codegen.cpp_symbol_map import *
+
+from codegen.ops.fmha_fwd import (
+    FmhaFwdTileSize,
+    FmhaFwdApiTrait,
+    FMHA_FWD_KERNEL_HEADER,
+    FMHA_FWD_API_PER_DTYPE,
+    FMHA_FWD_API_PER_HDIM_CASE,
+)
+
+
+FMHA_FWD_SPLITKV_PIPELINE_MAP = {
+    "qr" : "ck_tile::BlockFmhaFwdSplitKVPipelineQRKSVS",
+    "qr_async" : "ck_tile::BlockFmhaFwdSplitKVPipelineQRKSVSAsync",
+}
+
+FMHA_FWD_SPLITKV_KERNEL_BODY="""
+using fmha_dtype_{F_idx} = {F_dtype};
+using fmha_mask_{F_idx} = {F_mask};
+
+namespace {{
+template <bool kHasUnevenSplits>
+struct kernel_runner {{
+using fmha_block_tile = ck_tile::sequence<{F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0blen}>;
+using fmha_block_warps = ck_tile::sequence<{F_rm}, {F_rn}, {F_rk}>;
+using fmha_warp_tile = ck_tile::sequence<{F_wm}, {F_wn}, {F_wk}>;
+
+using fmha_shape = ck_tile::TileFmhaShape<fmha_block_tile,
+                                          fmha_block_warps,
+                                          fmha_warp_tile,
+                                          fmha_block_warps,
+                                          fmha_warp_tile,
+                                          {F_vlayout}>;
+
+using fmha_trait = ck_tile::TileFmhaFwdSplitKVTraits<{F_spad},
+                                                     {F_skpad},
+                                                     {F_dpad},
+                                                     {F_dvpad},
+                                                     {F_bias},
+                                                     false,
+                                                     {F_lse},
+                                                     {F_dropout},
+                                                     {F_squant},
+                                                     kHasUnevenSplits,
+                                                     {F_occupancy}>;
+
+using fmha_pipeline_problem = ck_tile::BlockFmhaFwdSplitKVPipelineProblem<
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::QDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::KDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::VDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::SaccDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::SMPLComputeDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::BiasDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::RandValOutputDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::LSEDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::PDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::OaccDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::OaccDataType,
+    fmha_shape,
+    {F_mode},
+    fmha_mask_{F_idx},
+    fmha_trait>;
+
+using fmha_pipeline = {F_pipeline}<
+    fmha_pipeline_problem>;
+
+using fmha_epilogue =
+    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<{F_dtype}>::OaccDataType,
+                                           typename FmhaFwdTypeConfig<{F_dtype}>::OaccDataType,
+                                           {F_spad}, {F_dvpad}>>;
+
+using fmha_kernel =
+    ck_tile::FmhaFwdSplitKVKernel<ck_tile::FmhaFwdSplitKVTilePartitioner<fmha_shape>,
+                  fmha_pipeline,
+                  fmha_epilogue>;
+
+static void run(const ck_tile::stream_config& s, fmha_fwd_args a)
+{{
+    using k_ = fmha_kernel;
+    auto [kargs, grids] = fmha_fwd_splitkv_create_kargs_and_grids<k_>(a);
+    constexpr dim3 blocks             = k_::BlockSize();
+    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
+    ck_tile::make_kernel<blocks.x, kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs)(ck_tile::stream_config{{s.stream_id_}});
+}}
+}};
+}}
+
+using trait_{F_idx} = fmha_fwd_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0blen}, {F_vlayout},
+                        {F_pipeline_enum}, fmha_mask_{F_idx}, {F_bias}, {F_lse}, {F_dropout}, {F_squant}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}>;
+
+#include <iostream>
+
+template<>
+void fmha_fwd_splitkv_oneshot_<trait_{F_idx}>(const ck_tile::stream_config& s, fmha_fwd_args a)
+{{
+    if constexpr({F_mode} == false) {{ // batch mode
+        if (a.seqlen_k % (a.num_splits * {F_bn0}) == 0) {{
+            kernel_runner<false>::run(s, a);
+        }} else {{
+            kernel_runner<true>::run(s, a);
+        }}
+    }} else {{
+        kernel_runner<true>::run(s, a);
+    }}
+}}
+
+template<>
+std::string fmha_fwd_splitkv_get_name_<trait_{F_idx}>()
+{{
+    using k_ = kernel_runner<true>::fmha_kernel; /// FIXME: choose real kernel type
+    return k_::GetName();
+}}
+"""
+
+FMHA_FWD_SPLITKV_COMBINE_KERNEL_BODY="""
+using fmha_dtype_{F_idx} = {F_dtype};
+
+namespace {{
+template <ck_tile::index_t kLogMaxSplits>
+struct kernel_runner {{
+using fmha_trait = ck_tile::TileFmhaFwdSplitKVCombineTraits<{F_spad},
+                                                    {F_dvpad},
+                                                    {F_lse},
+                                                    {F_squant},
+                                                    kLogMaxSplits,
+                                                    {F_occupancy}>;
+
+using fmha_pipeline_problem = ck_tile::BlockFmhaSplitKVCombinePipelineProblem<
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::LSEDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::OaccDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::ODataType,
+    {F_hdim},
+    {F_bm0}, 
+    {F_bn1},
+    {F_mode},
+    fmha_trait>;
+
+using fmha_pipeline = ck_tile::BlockFmhaFwdSplitKVCombinePipeline<
+    fmha_pipeline_problem>;
+
+using fmha_epilogue =
+    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<{F_dtype}>::OaccDataType,
+                                           typename FmhaFwdTypeConfig<{F_dtype}>::ODataType,
+                                           {F_spad}, {F_dvpad}>>;
+
+using fmha_kernel =
+    ck_tile::FmhaFwdSplitKVCombineKernel<ck_tile::FmhaFwdSplitKVCombineTilePartitioner<{F_bm0}, {F_bn1}>,
+                  fmha_pipeline,
+                  fmha_epilogue>;
+
+static void run(const ck_tile::stream_config& s, fmha_fwd_args a)
+{{
+    using k_ = fmha_kernel;
+    auto [kargs, grids] = fmha_fwd_splitkv_combine_create_kargs_and_grids<k_>(a);
+    constexpr dim3 blocks             = k_::BlockSize();
+    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
+    ck_tile::make_kernel<blocks.x, kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs)(ck_tile::stream_config{{s.stream_id_}});
+}}
+}};
+}}
+
+using trait_{F_idx} = fmha_fwd_splitkv_combine_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}, {F_bn1},
+                        {F_lse}, {F_squant}, {F_spad}, {F_dvpad}>;
+
+#include <iostream>
+
+template<>
+void fmha_fwd_splitkv_combine_oneshot_<trait_{F_idx}>(const ck_tile::stream_config& s, fmha_fwd_args a)
+{{
+    if (a.num_splits <= 16) {{
+        kernel_runner<4>::run(s, a);
+    }} else if (a.num_splits <= 32) {{
+        kernel_runner<5>::run(s, a);
+    }} else if (a.num_splits <= 64) {{
+        kernel_runner<6>::run(s, a);
+    }} else if (a.num_splits <= 128) {{
+        kernel_runner<7>::run(s, a);
+    }}
+}}
+
+template<>
+std::string fmha_fwd_splitkv_combine_get_name_<trait_{F_idx}>()
+{{
+    using k_ = kernel_runner<6>::fmha_kernel; /// FIXME: choose real kernel type
+    return k_::GetName();
+}}
+"""
+
+FMHA_FWD_SPLITKV_API_FILENAME="fmha_fwd_splitkv_api.cpp"
+FMHA_FWD_SPLITKV_API="""
+#include <iostream>
+
+template<typename fmha_fwd_splitkv_traits_, typename fmha_fwd_splitkv_combine_traits_>
+float fmha_fwd_splitkv_(const ck_tile::stream_config& s, fmha_fwd_args a)
+{{
+    if(s.log_level_ > 0)
+    std::cout
+    << ", " << fmha_fwd_splitkv_get_name_<fmha_fwd_splitkv_traits_>()
+    << ", " << fmha_fwd_splitkv_combine_get_name_<fmha_fwd_splitkv_combine_traits_>() 
+    << std::flush;
+
+    return ck_tile::launch_kernel(s,
+        [=](const ck_tile::stream_config& s_){{ fmha_fwd_splitkv_oneshot_<fmha_fwd_splitkv_traits_>(s_, a); }}, 
+        [=](const ck_tile::stream_config& s_){{ fmha_fwd_splitkv_combine_oneshot_<fmha_fwd_splitkv_combine_traits_>(s_, a); }}
+    );
+}}
+
+float fmha_fwd_splitkv(fmha_fwd_traits t, fmha_fwd_args a, const ck_tile::stream_config& s){{
+    float r = -1;
+{F_dispatch}
+    return r;
+}}
+"""
+
+FMHA_FWD_SPLITKV_API_INNER_DISPATCH="""            {F_if}((t.is_group_mode == {F_mode}) && (t.is_v_rowmajor == {F_vlayout}) && ({F_mask_check}) && (t.bias_type == {F_bias_check}) && (t.has_lse == {F_lse})  && (t.has_dropout == {F_dropout}) && (t.do_fp8_static_quant == {F_squant}) &&
+                        ({F_scheck}) && ({F_skcheck}) && ({F_dcheck}) && ({F_dvcheck})) {{
+                using traits_ = fmha_fwd_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0blen}, {F_vlayout}, {F_pipeline_enum}, {F_mask}, {F_bias}, {F_lse}, {F_dropout}, {F_squant}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}>;
+                using traits2_ = fmha_fwd_splitkv_combine_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}/2, {F_bn1}, {F_lse}, {F_squant}, {F_spad}, {F_dvpad}>;
+
+                return fmha_fwd_splitkv_<traits_, traits2_>(s, a);
+            }}
+"""
+
+@dataclass
+class FmhaFwdSplitKVPipeline:
+    tag : str
+
+    F_vlayout   : str  # row/col
+    F_spad      : str  # true/false
+    F_skpad     : str  #
+    F_dpad      : str  #
+    F_dvpad     : str  #
+    F_bias      : str  # true/false
+    F_lse       : str  #
+    F_dropout   : str  #
+    F_squant    : str  #
+    F_mask      : str  # value from MASK_MAP
+
+    @property
+    def name(self) -> str:
+        def pad_name() -> str:
+            n = ''
+            if self.F_spad == 't': n += 's'
+            if self.F_skpad == 't' : n += 'sk'
+            if self.F_dpad == 't' : n += 'd'
+            if self.F_dvpad == 't' : n += 'dv'
+            if n != '' : n = 'p' + n
+            return n
+        pn = pad_name()
+        n = f'{self.tag}_v{self.F_vlayout[0]}'
+        if pn != '' : n += f'_{pn}'
+        if self.F_bias != 'no' : n += f'_{self.F_bias}'
+        if self.F_mask[0:2] == 's_':
+            if self.F_mask == 's_mask': n += f'_mask'
+        else:
+            if self.F_mask != 'no' : n += f'_m{self.F_mask[0]}'
+        if self.F_lse == 't' : n += '_lse'
+        if self.F_dropout == 't' : n += '_dropout'
+        if self.F_squant == 't' : n += '_squant'
+        return n
+
+@dataclass
+class FmhaFwdSplitKVCombinePipeline:
+    tag : str
+
+    F_spad      : str  # true/false
+    F_dvpad     : str  #
+    F_lse       : str  #
+    F_squant    : str  #
+
+    @property
+    def name(self) -> str:
+        def pad_name() -> str:
+            n = ''
+            if self.F_spad == 't': n += 's'
+            if self.F_dvpad == 't' : n += 'dv'
+            if n != '' : n = 'p' + n
+            return n
+        pn = pad_name()
+        n = f'{self.tag}'
+        if pn != '' : n += f'_{pn}'
+        if self.F_lse == 't' : n += '_lse'
+        if self.F_squant == 't' : n += '_squant'
+        return n
+
+class FmhaFwdSplitKVApiPool:
+    def __init__(self, mask_impl):
+        self.pool = dict()
+        self.mask_impl = mask_impl
+
+    def register_traits(self, trait : FmhaFwdApiTrait) -> None:
+        # TODO: do we need to check duplication?
+        if trait.dtype not in self.pool.keys():
+            self.pool[trait.dtype] = dict()
+        if trait.hdim not in self.pool[trait.dtype].keys():
+            self.pool[trait.dtype][trait.hdim] = list()
+
+        self.pool[trait.dtype][trait.hdim].append(copy.copy(trait))
+
+    @property
+    def api(self) -> str:
+        per_dtypes=str()
+        for i, dtype in enumerate(self.pool.keys()):
+            per_hdim_case=str()
+            for j, hdim in enumerate(self.pool[dtype].keys()):
+                traits=self.pool[dtype][hdim]
+                inners=str()
+                for k, trait in enumerate(traits):
+                    if_k = 'if' if k == 0 else 'else if'
+                    inners = inners + FMHA_FWD_SPLITKV_API_INNER_DISPATCH.format(F_if=if_k, F_mode=MODE_MAP[trait.mode], F_vlayout=LAYOUT_MAP[trait.vlayout],
+                                   F_pipeline_enum=PIPELINE_ENUM_MAP[trait.pipeline_tag], F_mask=get_mask_map(self.mask_impl)[trait.mask],
+                                   F_mask_check=get_mask_check_map(self.mask_impl)[trait.mask], F_bias_check=BIAS_CHECK_MAP[trait.bias], F_bias=BIAS_MAP[trait.bias],
+                                   F_lse=BOOL_MAP[trait.lse], F_dropout=BOOL_MAP[trait.dropout] ,
+                                   F_squant=BOOL_MAP[trait.squant], F_scheck=trait.scheck, F_skcheck=trait.skcheck, F_dcheck=trait.dcheck, F_dvcheck=trait.dvcheck,
+                                   F_spad=BOOL_MAP[trait.spad], F_skpad=BOOL_MAP[trait.skpad], F_dpad=BOOL_MAP[trait.dpad], F_dvpad=BOOL_MAP[trait.dvpad],
+                                   F_bm0=trait.bm0, F_bn0=trait.bn0, F_bk0=trait.bk0, F_bn1=trait.bn1, F_bk1=trait.bk1, F_bk0blen=trait.bk0blen,
+                                   F_hdim=hdim, F_dtype=DTYPE_MAP[dtype])
+                if_j = 'if' if j == 0 else 'else if'
+                per_hdim_case = per_hdim_case + FMHA_FWD_API_PER_HDIM_CASE.format(F_if=if_j, F_hdim=hdim, F_inner_dispatch=inners)
+            if_i = 'if' if i == 0 else 'else if'
+            per_dtypes = per_dtypes + FMHA_FWD_API_PER_DTYPE.format(F_if=if_i, F_dtype=dtype, F_hdim_case=per_hdim_case)
+        return FMHA_FWD_KERNEL_HEADER + FMHA_FWD_SPLITKV_API.format(F_dispatch = per_dtypes)
+
+@dataclass
+class FmhaFwdSplitKVCombineTileSize:
+    F_bm0       : int  # tile size along q seqlen
+    F_bn1       : int  # tile size along v head_dim
+    F_occupancy : int  # occupancy, -1 will let pipeline decide the occupancy, other value will overwrite occupancy
+    @property
+    def name(self) -> str:
+        return f"b{self.F_bm0}x{self.F_bn1}" +\
+            ("" if self.F_occupancy == -1 else f"_o{self.F_occupancy}")
+
+@dataclass
+class FmhaFwdSplitKVKernel:
+    F_idx           : int  # this is not a tunable, but a counter to differentiate symbol
+    F_hdim          : int  # hdim
+    F_dtype         : str  # data type
+    F_mode          : str  # value from MODE_MAP
+    F_tile          : FmhaFwdTileSize
+    F_pipeline      : FmhaFwdSplitKVPipeline
+    mask_impl       : str
+
+    @property
+    def template(self) -> str:
+        kernel_body = str()
+        return FMHA_FWD_KERNEL_HEADER + \
+            FMHA_FWD_SPLITKV_KERNEL_BODY.format(
+                F_idx           = self.F_idx,
+                F_hdim          = self.F_hdim,
+                F_dtype         = DTYPE_MAP[self.F_dtype],
+                F_bm0           = self.F_tile.F_bm0,
+                F_bn0           = self.F_tile.F_bn0,
+                F_bk0           = self.F_tile.F_bk0,
+                F_bn1           = self.F_tile.F_bn1,
+                F_bk1           = self.F_tile.F_bk1,
+                F_bk0blen       = self.F_tile.F_bk0blen,
+                F_rm            = self.F_tile.F_rm,
+                F_rn            = self.F_tile.F_rn,
+                F_rk            = self.F_tile.F_rk,
+                F_wm            = self.F_tile.F_wm,
+                F_wn            = self.F_tile.F_wn,
+                F_wk            = self.F_tile.F_wk,
+                F_vlayout       = LAYOUT_MAP[self.F_pipeline.F_vlayout],
+                F_spad          = BOOL_MAP[self.F_pipeline.F_spad],
+                F_skpad         = BOOL_MAP[self.F_pipeline.F_skpad],
+                F_dpad          = BOOL_MAP[self.F_pipeline.F_dpad],
+                F_dvpad         = BOOL_MAP[self.F_pipeline.F_dvpad],    
+                F_bias          = BIAS_MAP[self.F_pipeline.F_bias],
+                F_lse           = BOOL_MAP[self.F_pipeline.F_lse],
+                F_dropout       = BOOL_MAP[self.F_pipeline.F_dropout],
+                F_squant        = BOOL_MAP[self.F_pipeline.F_squant],
+                F_occupancy     = self.F_tile.F_occupancy,
+                F_pipeline_enum = PIPELINE_ENUM_MAP[self.F_pipeline.tag],
+                F_mask          = get_mask_map(self.mask_impl)[self.F_pipeline.F_mask],
+                F_mode          = MODE_MAP[self.F_mode],
+                F_pipeline      = FMHA_FWD_SPLITKV_PIPELINE_MAP[self.F_pipeline.tag])
+
+    @property
+    def name(self) -> str:
+        # TODO: we don't encode idx here
+        return f"fmha_fwd_splitkv_d{self.F_hdim}_{self.F_dtype}_{self.F_mode}_" + \
+                self.F_tile.name + '_' + self.F_pipeline.name
+
+    @property
+    def filename(self) -> str:
+        return self.name + ".cpp"
+
+    def api_trait(self) -> FmhaFwdApiTrait:
+        return FmhaFwdApiTrait(
+                pipeline_tag=self.F_pipeline.tag,
+                hdim=str(self.F_hdim),
+                dtype=self.F_dtype,
+                mode=self.F_mode,
+                bm0=self.F_tile.F_bm0,
+                bn0=self.F_tile.F_bn0,
+                bk0=self.F_tile.F_bk0,
+                bn1=self.F_tile.F_bn1,
+                bk1=self.F_tile.F_bk1,
+                bk0blen=self.F_tile.F_bk0blen,
+                vlayout=self.F_pipeline.F_vlayout,
+                mask=self.F_pipeline.F_mask,
+                bias=self.F_pipeline.F_bias,
+                lse=self.F_pipeline.F_lse,
+                dropout=self.F_pipeline.F_dropout,
+                squant=self.F_pipeline.F_squant,
+                spad=self.F_pipeline.F_spad,
+                skpad=self.F_pipeline.F_skpad,
+                dpad=self.F_pipeline.F_dpad,
+                dvpad=self.F_pipeline.F_dvpad)
+
+@dataclass
+class FmhaFwdSplitKVCombineKernel:
+    F_idx           : int  # this is not a tunable, but a counter to differentiate symbol
+    F_hdim          : int  # hdim
+    F_dtype         : str  # data type
+    F_mode          : str  # value from MODE_MAP
+    F_tile          : FmhaFwdSplitKVCombineTileSize
+    F_pipeline      : FmhaFwdSplitKVCombinePipeline
+
+    @property
+    def template(self) -> str:
+        kernel_body = str()
+        return FMHA_FWD_KERNEL_HEADER + \
+            FMHA_FWD_SPLITKV_COMBINE_KERNEL_BODY.format(
+                F_idx           = self.F_idx,
+                F_hdim          = self.F_hdim,
+                F_dtype         = DTYPE_MAP[self.F_dtype],
+                F_bm0           = self.F_tile.F_bm0,
+                F_bn1           = self.F_tile.F_bn1,
+                F_spad          = BOOL_MAP[self.F_pipeline.F_spad],
+                F_dvpad         = BOOL_MAP[self.F_pipeline.F_dvpad],
+                F_lse           = BOOL_MAP[self.F_pipeline.F_lse],
+                F_squant        = BOOL_MAP[self.F_pipeline.F_squant],
+                F_occupancy     = self.F_tile.F_occupancy,
+                F_mode          = MODE_MAP[self.F_mode])
+
+    @property
+    def name(self) -> str:
+        # TODO: we don't encode idx here
+        return f"fmha_fwd_splitkv_combine_d{self.F_hdim}_{self.F_dtype}_{self.F_mode}_" + \
+                self.F_tile.name + '_' + self.F_pipeline.name
+
+    @property
+    def filename(self) -> str:
+        return self.name + ".cpp"
+
+    def api_trait(self) -> FmhaFwdApiTrait:
+        return FmhaFwdApiTrait(
+                pipeline_tag=self.F_pipeline.tag,
+                hdim=str(self.F_hdim),
+                dtype=self.F_dtype,
+                mode=self.F_mode,
+                bm0=self.F_tile.F_bm0,
+                bn0=self.F_tile.F_bn0,
+                bk0=self.F_tile.F_bk0,
+                bn1=self.F_tile.F_bn1,
+                bk1=self.F_tile.F_bk1,
+                bk0blen=self.F_tile.F_bk0blen,
+                vlayout=self.F_pipeline.F_vlayout,
+                mask=self.F_pipeline.F_mask,
+                bias=self.F_pipeline.F_bias,
+                lse=self.F_pipeline.F_lse,
+                dropout=self.F_pipeline.F_dropout,
+                squant=self.F_pipeline.F_squant,
+                spad=self.F_pipeline.F_spad,
+                skpad=self.F_pipeline.F_skpad,
+                dpad=self.F_pipeline.F_dpad,
+                dvpad=self.F_pipeline.F_dvpad)
+
+# TODO: design a more practical way to do it
+# this is current supported tile size per hdim
+def get_fmha_fwd_tile_dict_from_dtype(dtype : str) -> Optional[dict]:
+    if dtype == 'fp16' or dtype == 'bf16':
+        return {
+                '32'  : FmhaFwdTileSize(128, 64, 16, 32, 32, 32,     2, 1, 1, 32, 32, 16, -1),
+                '64'  : FmhaFwdTileSize(128, 64, 32, 64, 32, 64,     4, 1, 1, 32, 32, 16, -1),
+                '128' : FmhaFwdTileSize(128, 128, 32, 128, 32, 128,  4, 1, 1, 32, 32, 16, -1),
+                '256' : FmhaFwdTileSize(128, 128, 32, 256, 32, 256,  4, 1, 1, 32, 32, 16, -1),
+        }
+    elif dtype == 'fp8' or dtype == 'bf8':
+        return {
+            '64'  : FmhaFwdTileSize(128, 64, 32, 64, 32, 64,     2, 1, 1, 32, 32, 32, -1),
+            '128' : FmhaFwdTileSize(128, 128, 32, 128, 32, 128,  4, 1, 1, 32, 32, 32, -1),
+            '256' : FmhaFwdTileSize(128, 128, 32, 256, 32, 256,  4, 1, 1, 32, 32, 32, -1)
+        }
+    else:
+        return None
+
+def get_fmha_fwd_splitkv_combine_tile_dict_from_dtype(dtype : str) -> Optional[dict]:
+    if dtype == 'fp16' or dtype == 'bf16':
+        return {
+                '32'  : FmhaFwdSplitKVCombineTileSize(64, 32, -1),
+                '64'  : FmhaFwdSplitKVCombineTileSize(64, 64, -1),
+                '128' : FmhaFwdSplitKVCombineTileSize(64, 128, -1),
+                '256' : FmhaFwdSplitKVCombineTileSize(64, 256, -1),
+    }
+    elif dtype == 'fp8' or dtype == 'bf8':
+        return {
+                '64'  : FmhaFwdSplitKVCombineTileSize(64, 64, -1),
+                '128' : FmhaFwdSplitKVCombineTileSize(64, 128, -1),
+                '256' : FmhaFwdSplitKVCombineTileSize(64, 256, -1),
+        }
+    else:
+        return None
+
+def get_fwd_splitkv_blobs(kernel_filter : Optional[str], receipt, mask_impl) -> Tuple[FmhaFwdSplitKVApiPool, List[FmhaFwdSplitKVKernel]]:
+    Pipeline = FmhaFwdSplitKVPipeline
+    Kernel = FmhaFwdSplitKVKernel
+
+    # TODO: we don't support tuning yet, so pick up one value for vlayout/pipeline/pad
+    #       support this in future
+    def get_pipelines(dtype, hdim) -> List[FmhaFwdSplitKVPipeline]:
+        # this function will populate a list possible pipelines
+        # TODO: the order of List matters! the later in this list will be also be checked later
+        # TODO: currently for qr pipeline, let 't' padding to appear later!!
+        # TODO: how to design this more generic?
+        squant = 't' if dtype == 'fp8' else 'f'
+        pipelines = []
+        if dtype in ['fp16', 'bf16']:
+            # splitkv kernel donot support dropout
+            for mask, bias, lse, dropout in itertools.product(get_mask_map(mask_impl).keys(), BIAS_MAP.keys(), ["t", "f"], ["f"]):
+                if hdim == 256:
+                # if True:
+                    pipelines.append(Pipeline('qr', 'row', 'f', 'f', 'f', 'f', bias, lse, dropout, squant, mask))
+                    pipelines.append(Pipeline('qr', 'col', 'f', 'f', 'f', 'f', bias, lse, dropout, squant, mask))
+
+                    pipelines.append(Pipeline('qr', 'row', 't', 't', 't', 't', bias, lse, dropout, squant, mask))
+                    pipelines.append(Pipeline('qr', 'col', 't', 't', 't', 't', bias, lse, dropout, squant, mask))
+                else:
+                    pipelines.append(Pipeline('qr_async', 'row', 't', 'f', 't', 't', bias, lse, dropout, squant, mask))
+                    pipelines.append(Pipeline('qr_async', 'row', 't', 't', 't', 't', bias, lse, dropout, squant, mask))
+                    pipelines.append(Pipeline('qr_async', 'col', 't', 'f', 't', 't', bias, lse, dropout, squant, mask))
+                    pipelines.append(Pipeline('qr_async', 'col', 't', 't', 't', 't', bias, lse, dropout, squant, mask))
+                    if receipt == 1:
+                        pipelines.append(Pipeline('qr', 'row', 't', 't', 't', 't', bias, lse, dropout, squant, mask)) # TODO: cover arbitraty hdim
+                        pipelines.append(Pipeline('qr', 'col', 't', 'f', 't', 't', bias, lse, dropout, squant, mask)) # TODO: cover arbitraty hdim
+        elif dtype in ['fp8', 'bf8']:
+            # no need lse/dropout kernels
+            for mask, bias in itertools.product(get_mask_map(mask_impl).keys(), BIAS_MAP.keys()):
+                pipelines.append(Pipeline('qr', 'col', 'f', 'f', 'f', 'f', bias, 'f', 'f', squant, mask))
+        else:
+            assert False
+        return pipelines
+
+    gen = list()
+    api_pool = FmhaFwdSplitKVApiPool(mask_impl)
+
+    for dtype in DTYPE_MAP.keys():
+        d = get_fmha_fwd_tile_dict_from_dtype(dtype)
+        if d == None:
+            continue
+        #for hdim_str, mode, mask, bias, lse in itertools.product(d.keys(), MODE_MAP.keys(), MASK_MAP.keys(), ["t", "f"], ["t", "f"]):
+        for hdim_str, mode in itertools.product(d.keys(), MODE_MAP.keys()):
+            tile = d[hdim_str]
+            hdim = int(hdim_str)
+            for pipeline in get_pipelines(dtype, hdim):
+                if mode == "group":
+                    if pipeline.F_spad != 't' or pipeline.F_skpad != 't':
+                        # in group mode, spad/skpad must be true, since we can't predict if seqlen of current batch need pad or not
+                        continue
+                k = Kernel(F_idx=0,
+                           F_hdim=hdim,
+                           F_dtype=dtype,
+                           F_mode=mode,
+                           F_tile=tile,
+                           F_pipeline=pipeline,
+                           mask_impl=mask_impl)
+                if kernel_filter != None:
+                    if not fnmatch.fnmatch(k.name, kernel_filter):
+                        continue
+                if receipt == 2:
+                    cond = dtype in ['fp16', 'bf16']
+                    cond &= pipeline.F_vlayout == 'row'
+                    cond &= pipeline.F_bias in ['no', 'alibi']
+                    cond &= pipeline.F_squant == 'f'
+                    if not cond:
+                        continue
+                api_pool.register_traits(k.api_trait())
+                gen.append(k)
+
+    return (api_pool, gen)
+
+def get_fwd_splitkv_combine_blobs(kernel_filter : Optional[str], receipt) -> List[FmhaFwdSplitKVCombineKernel]:
+    Pipeline = FmhaFwdSplitKVCombinePipeline
+    Kernel = FmhaFwdSplitKVCombineKernel
+
+    # TODO: we don't support tuning yet, so pick up one value for vlayout/pipeline/pad
+    #       support this in future
+    def get_pipelines(dtype, hdim) -> List[FmhaFwdSplitKVCombinePipeline]:
+        # this function will populate a list possible pipelines
+        # TODO: the order of List matters! the later in this list will be also be checked later
+        # TODO: currently for qr pipeline, let 't' padding to appear later!!
+        # TODO: how to design this more generic?
+        squant = 't' if dtype == 'fp8' else 'f'
+        pipelines = []
+        if dtype in ['fp16', 'bf16']:
+            for spad, dvpad, lse in itertools.product(["t", "f"], ["t", "f"], ["t", "f"]):
+                pipelines.append(Pipeline('unused', spad, dvpad, lse, squant))
+        elif dtype in ['fp8', 'bf8']:
+            # no need lse kernels
+            pipelines.append(Pipeline('unused', 'f', 'f', 'f', squant))
+        else:
+            assert False
+        return pipelines
+
+    gen = list()
+
+    for dtype in DTYPE_MAP.keys():
+        d = get_fmha_fwd_splitkv_combine_tile_dict_from_dtype(dtype)
+        if d == None:
+            continue
+        #for hdim_str, mode, mask, bias, lse in itertools.product(d.keys(), MODE_MAP.keys(), MASK_MAP.keys(), ["t", "f"], ["t", "f"]):
+        for hdim_str, mode in itertools.product(d.keys(), MODE_MAP.keys()):
+            tile = d[hdim_str]
+            hdim = int(hdim_str)
+            for pipeline in get_pipelines(dtype, hdim):
+                if mode == "group":
+                    if pipeline.F_spad != 't':
+                        # in group mode, spad/skpad must be true, since we can't predict if seqlen of current batch need pad or not
+                        continue
+                k = Kernel(F_idx=0,
+                           F_hdim=hdim,
+                           F_dtype=dtype,
+                           F_mode=mode,
+                           F_tile=tile,
+                           F_pipeline=pipeline)
+                if kernel_filter != None:
+                    if not fnmatch.fnmatch(k.name, kernel_filter):
+                        continue
+                gen.append(k)
+
+    return gen
+
+def write_single_kernel(kernel: Union[FmhaFwdSplitKVKernel, FmhaFwdSplitKVCombineKernel], autogen_dir: Path) -> None:
+    (autogen_dir / kernel.filename).write_text(kernel.template)
+
+def write_fwd_splitkv_api(api_pool : FmhaFwdSplitKVApiPool, autogen_dir: Path) -> None:
+    file_path = autogen_dir / FMHA_FWD_SPLITKV_API_FILENAME
+    file_path.write_text(api_pool.api)
+
+def write_blobs(output_dir : Path, kernel_filter : Optional[str], receipt, mask_impl) -> None:
+    kernels = get_fwd_splitkv_combine_blobs(kernel_filter, receipt)
+    for kernel in kernels:
+        write_single_kernel(kernel, output_dir)
+    api_pool, kernels = get_fwd_splitkv_blobs(kernel_filter, receipt, mask_impl)
+    for kernel in kernels:
+        write_single_kernel(kernel, output_dir)
+    write_fwd_splitkv_api(api_pool, output_dir)
+
+def list_blobs(file_path : Path, kernel_filter : Optional[str], receipt, mask_impl) -> None:
+    with file_path.open('a') as f:
+        kernels = get_fwd_splitkv_combine_blobs(kernel_filter, receipt)
+        for kernel in kernels:
+            f.write(str(file_path.parent / GEN_DIR / kernel.filename) + "\n")
+        _, kernels = get_fwd_splitkv_blobs(kernel_filter, receipt, mask_impl)
+        for kernel in kernels:
+            f.write(str(file_path.parent / GEN_DIR / kernel.filename) + "\n")
+        f.write(str(file_path.parent / GEN_DIR / FMHA_FWD_SPLITKV_API_FILENAME) + "\n")
\ No newline at end of file

example/ck_tile/01_fmha/fmha_fwd.cpp

@@ -114,6 +114,9 @@ auto create_args(int argc, char* argv[])
         .insert("drop_seed", "1", "seed for random number generator")
         .insert("drop_offset", "0", "offset for random number generator")
         .insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer")
+        .insert("num_splits",
+                "1",
+                "# of splits for key/value. 0 to determine actual number by heuristic")
         .insert("warmup", "5", "number of iterations before benchmark the kernel")
         .insert("repeat", "20", "number of iterations to benchmark the kernel");
 
@@ -155,6 +158,106 @@ auto get_elimit<ck_tile::fp8_t>(std::string init_method)
     }
 }
 
+int num_splits_heuristic(int batch_nhead_mblocks, int num_SMs, int num_n_blocks, int max_splits)
+{
+    // If we have enough to almost fill the SMs, then just use 1 split
+    if(batch_nhead_mblocks >= 0.8f * num_SMs)
+    {
+        return 1;
+    }
+    max_splits           = std::min({max_splits, num_SMs, num_n_blocks});
+    float max_efficiency = 0.f;
+    std::vector<float> efficiency;
+    efficiency.reserve(max_splits);
+    auto ceildiv = [](int a, int b) { return (a + b - 1) / b; };
+    // Some splits are not eligible. For example, if we have 64 blocks and choose 11 splits,
+    // we'll have 6 * 10 + 4 blocks. If we choose 12 splits, we'll have 6 * 11 + (-2) blocks
+    // (i.e. it's 11 splits anyway).
+    // So we check if the number of blocks per split is the same as the previous num_splits.
+    auto is_split_eligible = [&ceildiv, &num_n_blocks](int num_splits) {
+        return num_splits == 1 ||
+               ceildiv(num_n_blocks, num_splits) != ceildiv(num_n_blocks, num_splits - 1);
+    };
+    for(int num_splits = 1; num_splits <= max_splits; num_splits++)
+    {
+        if(!is_split_eligible(num_splits))
+        {
+            efficiency.push_back(0.f);
+        }
+        else
+        {
+            float n_waves = float(batch_nhead_mblocks * num_splits) / num_SMs;
+            float eff     = n_waves / ceil(n_waves);
+            // printf("num_splits = %d, eff = %f\n", num_splits, eff);
+            if(eff > max_efficiency)
+            {
+                max_efficiency = eff;
+            }
+            efficiency.push_back(eff);
+        }
+    }
+    for(int num_splits = 1; num_splits <= max_splits; num_splits++)
+    {
+        if(!is_split_eligible(num_splits))
+        {
+            continue;
+        }
+        if(efficiency[num_splits - 1] >= 0.85 * max_efficiency)
+        {
+            // printf("num_splits chosen = %d\n", num_splits);
+            return num_splits;
+        }
+    }
+    return 1;
+}
+
+int override_num_splits_if_necessary(
+    int batch, int nhead, int max_seqlen_q, int hdim_v, float p_drop, int num_splits)
+{
+    int device;
+    auto status = hipGetDevice(&device);
+    if(status != hipSuccess)
+    {
+        return num_splits;
+    }
+
+    hipDeviceProp_t props{};
+    status = hipGetDeviceProperties(&props, device);
+    if(status != hipSuccess)
+    {
+        return num_splits;
+    }
+
+    // tile size should match the generate.py
+    const int kM0 = 64;
+    const int kN1 = hdim_v;
+
+    const int num_m_blocks = ck_tile::integer_divide_ceil(max_seqlen_q, kM0);
+    const int num_n_blocks = ck_tile::integer_divide_ceil(hdim_v, kN1);
+
+    if(num_splits < 1 && p_drop == 0.0f)
+    {
+        return num_splits_heuristic(
+            batch * nhead * num_m_blocks, props.multiProcessorCount * 2, num_n_blocks, 128);
+    }
+
+    return num_splits;
+}
+
+float fmha_fwd_dispatch(fmha_fwd_traits traits,
+                        fmha_fwd_args args,
+                        const ck_tile::stream_config& config)
+{
+    if(1 < args.num_splits)
+    {
+        return fmha_fwd_splitkv(traits, args, config);
+    }
+    else
+    {
+        return fmha_fwd(traits, args, config);
+    }
+}
+
 template <typename DataType>
 bool run(const ck_tile::ArgParser& arg_parser)
 {
@@ -260,6 +363,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
         seed.reset();
     }
 
+    int num_splits = arg_parser.get_int("num_splits");
+
     int stream_warmup = arg_parser.get_int("warmup");
     int stream_repeat = arg_parser.get_int("repeat");
     bool kname        = arg_parser.get_bool("kname");
@@ -320,6 +425,18 @@ bool run(const ck_tile::ArgParser& arg_parser)
         }
     }
 
+    // legalize num_splits according to other options
+    if(num_splits < 1)
+    {
+        num_splits = override_num_splits_if_necessary(
+            batch, nhead, max_seqlen_q, hdim_v, p_drop, num_splits);
+    }
+    if(128 < num_splits)
+    {
+        std::cerr << "num_splits greater than 128 is not supported" << std::endl;
+        return false;
+    }
+
     auto get_lengths = [&](bool permute,
                            ck_tile::index_t b /*batch*/,
                            ck_tile::index_t h /*nhead*/,
@@ -361,7 +478,15 @@ bool run(const ck_tile::ArgParser& arg_parser)
                                    : std::array<ck_tile::index_t, 2>{batch, nhead})
             : std::array<ck_tile::index_t, 2>{1, 1});
 
-    // self define lse data layout as [shape_batch, nhead, shape_seqlen_q]
+    ck_tile::HostTensor<LSEDataType> lse_acc_host(
+        1 < num_splits ? std::array<ck_tile::index_t, 4>{num_splits, batch, nhead, max_seqlen_q}
+                       : std::array<ck_tile::index_t, 4>{1, 1, 1, 1});
+    ck_tile::HostTensor<OaccDataType> o_acc_host(
+        1 < num_splits
+            ? std::array<ck_tile::index_t, 5>{num_splits, batch, nhead, max_seqlen_q, hdim_v}
+            : std::array<ck_tile::index_t, 5>{1, 1, 1, 1, 1});
+
+    // self define lse data layout as [batch, nhead, max_seqlen_q]
     ck_tile::HostTensor<LSEDataType> lse_host(
         lse ? std::array<ck_tile::index_t, 3>{batch, nhead, max_seqlen_q}
             : std::array<ck_tile::index_t, 3>{1, 1, 1} /* dummy shape for simplifying code */);
@@ -443,6 +568,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
     ck_tile::DeviceMem k_buf(k_host.get_element_space_size_in_bytes());
     ck_tile::DeviceMem v_buf(v_host.get_element_space_size_in_bytes());
     ck_tile::DeviceMem bias_buf(bias_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem lse_acc_buf(lse_acc_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem o_acc_buf(o_acc_host.get_element_space_size_in_bytes());
     ck_tile::DeviceMem lse_buf(lse_host.get_element_space_size_in_bytes());
     ck_tile::DeviceMem o_buf(o_host.get_element_space_size_in_bytes());
     ck_tile::DeviceMem seqstart_q(seqstart_q_host.size() * sizeof(int32_t));
@@ -479,7 +606,12 @@ bool run(const ck_tile::ArgParser& arg_parser)
                                       : (std::string("(") + std::to_string(seqlen_kpads[0]) + ")"))
               << ", d:" << hdim_q << "/" << hdim_v << ", scale_s:" << scale_s << ", bias:" << bias
               << ", p_drop:" << p_drop << ", lse:" << lse << ", squant:" << squant
-              << ", mask:" << mask << ", v:" << vlayout << std::flush;
+              << ", mask:" << mask << ", v:" << vlayout;
+    if(1 < num_splits)
+    {
+        std::cout << ", num_splits:" << num_splits;
+    }
+    std::cout << std::flush;
 
     auto fmha_traits = fmha_fwd_traits{hdim_q,
                                        hdim_v,
@@ -523,6 +655,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
         }();
         const ck_tile::index_t stride_bias    = (i_perm ? shape_seqlen_k : 1 * shape_seqlen_k);
         const ck_tile::index_t stride_randval = (max_seqlen_k);
+        const ck_tile::index_t stride_o_acc   = hdim_v;
         const ck_tile::index_t stride_o       = (o_perm ? hdim_v : nhead * hdim_v);
         // setup nhead_stride_* arguments
         const ck_tile::index_t nhead_stride_q = (i_perm ? shape_seqlen_q * hdim_q : hdim_q);
@@ -537,6 +670,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
             (i_perm ? 0 * shape_seqlen_q * shape_seqlen_k : 0 * shape_seqlen_k);
         const ck_tile::index_t nhead_stride_randval = (shape_seqlen_q * max_seqlen_k);
         const ck_tile::index_t nhead_stride_lse     = max_seqlen_q;
+        const ck_tile::index_t nhead_stride_lse_acc = max_seqlen_q;
+        const ck_tile::index_t nhead_stride_o_acc   = (max_seqlen_q * hdim_v);
         const ck_tile::index_t nhead_stride_o       = (o_perm ? shape_seqlen_q * hdim_v : hdim_v);
         // setup batch_stride_* arguments
         const ck_tile::index_t batch_stride_q       = (nhead * shape_seqlen_q * hdim_q);
@@ -545,7 +680,12 @@ bool run(const ck_tile::ArgParser& arg_parser)
         const ck_tile::index_t batch_stride_bias    = (0 * nhead * shape_seqlen_q * shape_seqlen_k);
         const ck_tile::index_t batch_stride_randval = (nhead * shape_seqlen_q * max_seqlen_k);
         const ck_tile::index_t batch_stride_lse     = (nhead * max_seqlen_q);
+        const ck_tile::index_t batch_stride_lse_acc = (nhead * max_seqlen_q);
+        const ck_tile::index_t batch_stride_o_acc   = (nhead * max_seqlen_q * hdim_v);
         const ck_tile::index_t batch_stride_o       = (nhead * shape_seqlen_q * hdim_v);
+        // setup split_stride_* arguments (only used in split-kv kernel)
+        const ck_tile::index_t split_stride_lse_acc = (batch * nhead * max_seqlen_q);
+        const ck_tile::index_t split_stride_o_acc   = (batch * nhead * max_seqlen_q * hdim_v);
 
         return fmha_fwd_args{q_buf.GetDeviceBuffer(),
                              k_buf.GetDeviceBuffer(),
@@ -553,6 +693,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
                              bias.type == bias_enum::alibi ? alibi_slope_buf.GetDeviceBuffer()
                                                            : bias_buf.GetDeviceBuffer(),
                              randval_buf.GetDeviceBuffer(),
+                             lse_acc_buf.GetDeviceBuffer(),
+                             o_acc_buf.GetDeviceBuffer(),
                              lse_buf.GetDeviceBuffer(),
                              o_buf.GetDeviceBuffer(),
                              seqstart_q.GetDeviceBuffer(),
@@ -566,6 +708,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
                              hdim_v,
                              nhead,
                              nhead_k,
+                             num_splits,
                              scale_s,
                              scale_p,
                              scale_o,
@@ -575,6 +718,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
                              bias.type == bias_enum::alibi ? (bias.rank_info == 0 ? 0 : nhead)
                                                            : stride_bias,
                              stride_randval,
+                             stride_o_acc,
                              stride_o,
                              nhead_stride_q,
                              nhead_stride_k,
@@ -582,6 +726,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
                              nhead_stride_bias,
                              nhead_stride_randval,
                              nhead_stride_lse,
+                             nhead_stride_lse_acc,
+                             nhead_stride_o_acc,
                              nhead_stride_o,
                              batch_stride_q,
                              batch_stride_k,
@@ -589,7 +735,11 @@ bool run(const ck_tile::ArgParser& arg_parser)
                              batch_stride_bias,
                              batch_stride_randval,
                              batch_stride_lse,
+                             batch_stride_lse_acc,
+                             batch_stride_o_acc,
                              batch_stride_o,
+                             split_stride_lse_acc,
+                             split_stride_o_acc,
                              mask.left,
                              mask.right,
                              static_cast<ck_tile::index_t>(mask.type),
@@ -598,7 +748,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
                              {drop_seed, drop_offset}};
     }();
 
-    float ave_time = fmha_fwd(fmha_traits, fmha_args, stream_config);
+    float ave_time = fmha_fwd_dispatch(fmha_traits, fmha_args, stream_config);
 
     if(ave_time < 0)
     {
@@ -849,14 +999,14 @@ bool run(const ck_tile::ArgParser& arg_parser)
             lse_host_result.ForEach(
                 [&](auto& self, auto idx) { self(idx) = lse_host(wb, idx[0], idx[1]); });
 
-            bool lse_pass = ck_tile::check_err(lse_host_result,
-                                               lse_host_ref,
-                                               "LSE Error: Incorrect results!",
-                                               rtol,
-                                               atol,
-                                               /* allow_infinity_ref = */ true);
+            cur_pass = ck_tile::check_err(lse_host_result,
+                                          lse_host_ref,
+                                          "LSE Error: Incorrect results!",
+                                          rtol,
+                                          atol,
+                                          /* allow_infinity_ref = */ true);
 
-            pass &= lse_pass;
+            pass &= cur_pass;
             if(!cur_pass)
             {
                 std::cerr << "LSE mismatch found at batch: " << wb << std::endl