2.5.2-dtk24.04

01a10755 · yuguo-Jack · 63eb0da5 · 01a10755 · 01a10755 · 01a10755
Commit 01a10755 authored Mar 04, 2024 by yuguo-Jack
20 changed files
--- a/cmake/cinn.cmake
+++ b/cmake/cinn.cmake
@@ -70,6 +70,10 @@ if(WITH_GPU)
    message(STATUS "Enable CINN CUDNN")
    add_definitions(-DCINN_WITH_CUDNN)
  endif()
+  if(WITH_CUTLASS)
+    message(STATUS "Enable CINN CUTLASS")
+    add_definitions(-DCINN_WITH_CUTLASS)
+  endif()
  enable_language(CUDA)
  find_package(CUDA REQUIRED)
  include_directories(${CUDA_INCLUDE_DIRS})
@@ -114,10 +118,6 @@ include(cmake/cinn/core.cmake)
 include(cmake/cinn/nvrtc.cmake)
 include(cmake/cinn/nvtx.cmake)
-if(CINN_ONLY)
-  link_libraries(gflags)
-endif()
 set(LINK_FLAGS
    "-Wl,--version-script ${CMAKE_CURRENT_SOURCE_DIR}/cmake/cinn/export.map"
    CACHE INTERNAL "")
@@ -168,8 +168,8 @@ cinn_cc_library(
 add_dependencies(cinnapi GEN_LLVM_RUNTIME_IR_HEADER ZLIB::ZLIB)
 add_dependencies(cinnapi GEN_LLVM_RUNTIME_IR_HEADER ${core_deps})
 if(NOT CINN_ONLY)
-  target_link_libraries(cinnapi phi)
+  target_link_libraries(cinnapi op_dialect_vjp phi)
-  add_dependencies(cinnapi phi)
+  add_dependencies(cinnapi op_dialect_vjp phi)
 endif()
 target_link_libraries(cinnapi ${PYTHON_LIBRARIES})
@@ -183,6 +183,11 @@ if(WITH_MKL)
  endif()
 endif()
+if(CINN_ONLY)
+  target_link_libraries(cinnapi ${flags_dep})
+  add_dependencies(cinnapi ${flags_dep})
+endif()
 if(WITH_GPU)
  target_link_libraries(
    cinnapi
@@ -198,6 +203,11 @@ if(WITH_GPU)
  endif()
 endif()
+if(WITH_CUTLASS)
+  target_link_libraries(cinnapi cutlass)
+  add_dependencies(cinnapi cutlass)
+endif()
 function(gen_cinncore LINKTYPE)
  set(CINNCORE_TARGET cinncore)
  if(${LINKTYPE} STREQUAL "STATIC")
@@ -221,8 +231,8 @@ function(gen_cinncore LINKTYPE)
  add_dependencies(${CINNCORE_TARGET} GEN_LLVM_RUNTIME_IR_HEADER ZLIB::ZLIB)
  add_dependencies(${CINNCORE_TARGET} GEN_LLVM_RUNTIME_IR_HEADER ${core_deps})
  if(NOT CINN_ONLY)
-    target_link_libraries(${CINNCORE_TARGET} phi)
+    target_link_libraries(${CINNCORE_TARGET} op_dialect_vjp phi)
-    add_dependencies(${CINNCORE_TARGET} phi)
+    add_dependencies(${CINNCORE_TARGET} op_dialect_vjp phi)
  endif()
  add_dependencies(${CINNCORE_TARGET} pybind)
@@ -237,6 +247,11 @@ function(gen_cinncore LINKTYPE)
    endif()
  endif()
+  if(CINN_ONLY)
+    target_link_libraries(${CINNCORE_TARGET} ${flags_dep})
+    add_dependencies(${CINNCORE_TARGET} ${flags_dep})
+  endif()
  if(WITH_GPU)
    target_link_libraries(
      ${CINNCORE_TARGET}
@@ -252,6 +267,11 @@ function(gen_cinncore LINKTYPE)
      target_link_libraries(${CINNCORE_TARGET} ${CUDA_NVTX_LIB})
    endif()
  endif()
+  if(WITH_CUTLASS)
+    target_link_libraries(cinnapi cutlass)
+    add_dependencies(cinnapi cutlass)
+  endif()
 endfunction()
 gen_cinncore(STATIC)
@@ -262,6 +282,9 @@ set(PUBLISH_LIBS ON)
 if(PUBLISH_LIBS)
  set(core_includes
      "${core_includes};paddle/cinn/runtime/cuda/cinn_cuda_runtime_source.cuh")
+  set(core_includes
+      "${core_includes};paddle/utils/flags.h;paddle/utils/flags_native.h;paddle/utils/test_macros.h"
+  )
  foreach(header ${core_includes})
    get_filename_component(prefix ${header} DIRECTORY)
    file(COPY ${header}

--- a/cmake/cinn/external/absl.cmake
+++ b/cmake/cinn/external/absl.cmake
@@ -5,7 +5,7 @@ set(ABSL_INSTALL_DIR ${THIRD_PARTY_PATH}/install/absl)
 set(ABSL_PREFIX_DIR ${THIRD_PARTY_PATH}/absl)
 set(ABSL_CMAKE_CXX_FLAGS ${CMAKE_CXX_FLAGS})
-set(ABSL_REPOSITORY "https://github.com/abseil/abseil-cpp.git")
+set(ABSL_REPOSITORY "${GIT_URL}/abseil/abseil-cpp.git")
 set(ABSL_TAG "20210324.2")
 set(OPTIONAL_ARGS

--- a/cmake/cinn/external/jitify.cmake
+++ b/cmake/cinn/external/jitify.cmake
@@ -7,7 +7,7 @@ include(ExternalProject)
 # clone jitify to Paddle/third_party
 set(JITIFY_SOURCE_DIR ${PADDLE_SOURCE_DIR}/third_party/jitify)
-set(JITIFY_URL https://github.com/NVIDIA/jitify.git)
+set(JITIFY_URL ${GIT_URL}/NVIDIA/jitify.git)
 set(JITIFY_TAG 57de649139c866eb83acacfe50c92ad7c6278776)
 ExternalProject_Add(

--- a/cmake/experiments/cuda_module_loading_lazy.cmake
+++ b/cmake/experiments/cuda_module_loading_lazy.cmake
@@ -17,6 +17,12 @@
 # this experiment option makes Paddle supports lazy loading before CUDA 11.7.
 if(LINUX)
+  if(NOT WITH_NVCC_LAZY)
+    message(
+      "EXP_CUDA_MODULE_LOADING_LAZY only works with WITH_NVCC_LAZY=ON on Linux platforms"
+    )
+    return()
+  endif()
  if(NOT ON_INFER)
    message(
      "EXP_CUDA_MODULE_LOADING_LAZY only works with ON_INFER=ON on Linux platforms"

--- a/cmake/external/brpc.cmake
+++ b/cmake/external/brpc.cmake
@@ -13,7 +13,9 @@
 # limitations under the License.
 include(ExternalProject)
+if(NOT WITH_ARM)
+  set(OPENSSL_USE_STATIC_LIBS ON)
+endif()
 find_package(OpenSSL REQUIRED)
 message(STATUS "ssl:" ${OPENSSL_SSL_LIBRARY})
@@ -38,7 +40,7 @@ include_directories(${BRPC_INCLUDE_DIR})
 # clone brpc to Paddle/third_party
 set(BRPC_SOURCE_DIR ${PADDLE_SOURCE_DIR}/third_party/brpc)
-set(BRPC_URL https://github.com/apache/brpc.git)
+set(BRPC_URL ${GIT_URL}/apache/brpc.git)
 set(BRPC_TAG 1.4.0)
 # Reference https://stackoverflow.com/questions/45414507/pass-a-list-of-prefix-paths-to-externalproject-add-in-cmake-args
@@ -91,3 +93,16 @@ add_dependencies(brpc extern_brpc)
 add_definitions(-DBRPC_WITH_GLOG)
 list(APPEND external_project_dependencies brpc)
+set(EXTERNAL_BRPC_DEPS
+    brpc
+    protobuf
+    ssl
+    crypto
+    leveldb
+    glog
+    snappy)
+if(NOT WITH_GFLAGS)
+  set(EXTERNAL_BRPC_DEPS ${EXTERNAL_BRPC_DEPS} gflags)
+endif()
--- a/cmake/external/cccl.cmake
+++ b/cmake/external/cccl.cmake
+include(ExternalProject)
+set(CCCL_PATH
+    "${THIRD_PARTY_PATH}/cccl"
+    CACHE STRING "A path setting for external_cccl path.")
+set(CCCL_PREFIX_DIR ${CCCL_PATH})
+set(CCCL_SOURCE_DIR ${PADDLE_SOURCE_DIR}/third_party/cccl)
+# The latest commit has bugs in windows, so we set a fix commit.
+set(CCCL_TAG 1f6e4bcae0fbf1bbed87f88544d8d2161c490fc1)
+execute_process(COMMAND git --git-dir=${CCCL_SOURCE_DIR}/.git
+                        --work-tree=${CCCL_SOURCE_DIR} checkout ${CCCL_TAG})
+set(CCCL_INCLUDE_DIR ${CCCL_SOURCE_DIR})
+message("CCCL_INCLUDE_DIR is ${CCCL_INCLUDE_DIR}")
+include_directories(${CCCL_INCLUDE_DIR})
+ExternalProject_Add(
+  extern_cccl
+  ${EXTERNAL_PROJECT_LOG_ARGS}
+  SOURCE_DIR ${CCCL_SOURCE_DIR}
+  PREFIX ${CCCL_PREFIX_DIR}
+  UPDATE_COMMAND ""
+  CONFIGURE_COMMAND ""
+  BUILD_COMMAND ""
+  INSTALL_COMMAND ""
+  TEST_COMMAND "")
+# update include dir and set cccl first for using
+include_directories(BEFORE "${CCCL_SOURCE_DIR}/cub")
+include_directories(BEFORE "${CCCL_SOURCE_DIR}/libcudacxx/include")
+include_directories(BEFORE "${CCCL_SOURCE_DIR}/thrust")
--- a/cmake/external/cryptopp.cmake
+++ b/cmake/external/cryptopp.cmake
@@ -19,7 +19,7 @@ set(CRYPTOPP_INSTALL_DIR ${THIRD_PARTY_PATH}/install/cryptopp)
 set(CRYPTOPP_INCLUDE_DIR
    "${CRYPTOPP_INSTALL_DIR}/include"
    CACHE PATH "cryptopp include directory." FORCE)
-set(CRYPTOPP_REPOSITORY ${GIT_URL}/weidai11/cryptopp.git)
+set(CRYPTOPP_REPOSITORY https://gitee.com/PaddleGitee/cryptopp.git)
 set(CRYPTOPP_TAG CRYPTOPP_8_2_0)
 if(WIN32)
@@ -69,7 +69,7 @@ ExternalProject_Add(
  UPDATE_COMMAND ""
  PATCH_COMMAND
  COMMAND ${CMAKE_COMMAND} -E remove_directory "<SOURCE_DIR>/cmake/"
-  COMMAND git clone ${GIT_URL}/noloader/cryptopp-cmake "<SOURCE_DIR>/cmake"
+  COMMAND git clone https://gitee.com/PaddleGitee/cryptopp-cmake.git "<SOURCE_DIR>/cmake"
  COMMAND cd "<SOURCE_DIR>/cmake" && git checkout tags/${CRYPTOPP_TAG} -b
          ${CRYPTOPP_TAG}
  COMMAND ${CMAKE_COMMAND} -E copy_directory "<SOURCE_DIR>/cmake/"

--- a/cmake/external/cudnn-frontend.cmake
+++ b/cmake/external/cudnn-frontend.cmake
@@ -34,7 +34,7 @@ if((NOT DEFINED CUDNN_FRONTEND_NAME) OR (NOT DEFINED CUDNN_FRONTEND_URL))
      "cudnn-frontend"
      CACHE STRING "" FORCE)
  set(CUDNN_FRONTEND_URL
-      "https://github.com/NVIDIA/cudnn-frontend/archive/refs/tags/${CUDNN_FRONTEND_VER}.tar.gz"
+      "${GIT_URL}/NVIDIA/cudnn-frontend/archive/refs/tags/${CUDNN_FRONTEND_VER}.tar.gz"
      CACHE STRING "" FORCE)
 endif()
 set(CUDNN_FRONTEND_CACHE_FILENAME "${CUDNN_FRONTEND_VER}.tar.gz")

--- a/cmake/external/dirent.cmake
+++ b/cmake/external/dirent.cmake
@@ -25,7 +25,7 @@ if((NOT DEFINED DIRENT_NAME) OR (NOT DEFINED DIRENT_URL))
      "dirent"
      CACHE STRING "" FORCE)
  set(DIRENT_URL
-      "https://github.com/tronkko/dirent/archive/refs/tags/1.23.2.tar.gz"
+      "${GIT_URL}/tronkko/dirent/archive/refs/tags/1.23.2.tar.gz"
      CACHE STRING "" FORCE)
  set(DIRENT_CACHE_FILENAME "1.23.2.tar.gz")
 endif()

--- a/cmake/external/flashattn.cmake
+++ b/cmake/external/flashattn.cmake
@@ -14,92 +14,151 @@
 include(ExternalProject)
-add_definitions(-DPADDLE_WITH_FLASHATTN)
+if(WITH_ROCM)
-set(FLASHATTN_PREFIX_DIR ${THIRD_PARTY_PATH}/flashattn)
-set(FLASHATTN_SOURCE_SUBDIR csrc)
-set(FLASHATTN_INSTALL_DIR ${THIRD_PARTY_PATH}/install/flashattn)
-set(SOURCE_DIR ${PADDLE_SOURCE_DIR}/third_party/flashattn)
-set(FLASHATTN_TAG 18106c1ba0ccee81b97ca947397c08a141815a47)
-set(FLASHATTN_INCLUDE_DIR
-    "${FLASHATTN_INSTALL_DIR}/include"
-    CACHE PATH "flash-attn Directory" FORCE)
-set(FLASHATTN_LIB_DIR
-    "${FLASHATTN_INSTALL_DIR}/lib"
-    CACHE PATH "flash-attn Library Directory" FORCE)
-if(WIN32)
-  set(FLASHATTN_LIBRARIES
-      "${FLASHATTN_INSTALL_DIR}/bin/flashattn${CMAKE_SHARED_LIBRARY_SUFFIX}"
-      CACHE FILEPATH "flash-attn Library" FORCE)
-else()
-  set(FLASHATTN_LIBRARIES
-      "${FLASHATTN_INSTALL_DIR}/lib/libflashattn${CMAKE_SHARED_LIBRARY_SUFFIX}"
-      CACHE FILEPATH "flash-attn Library" FORCE)
-endif()
-if(CMAKE_CXX_COMPILER_ID STREQUAL "Clang"
+    set(FLASHATTN_PREFIX_DIR ${THIRD_PARTY_PATH}/flashattn_hip)
-   OR CMAKE_CXX_COMPILER_ID STREQUAL "AppleClang"
+    set(FLASHATTN_INSTALL_DIR ${THIRD_PARTY_PATH}/install/flashattn)
-   OR WIN32)
+    set(SOURCE_DIR ${PADDLE_SOURCE_DIR}/third_party/flashattn_hip)
-  set(USE_OMP OFF)
-else()
+    set(FLASHATTN_INCLUDE_DIR
-  set(USE_OMP ON)
+        "${FLASHATTN_INSTALL_DIR}/include"
-endif()
+        CACHE PATH "flash-attn Directory" FORCE)
+    set(FLASHATTN_LIB_DIR
+        "${FLASHATTN_INSTALL_DIR}/lib"
+        CACHE PATH "flash-attn Library Directory" FORCE)
+    set(FLASHATTN_LIBRARIES
+        "${FLASHATTN_INSTALL_DIR}/lib/libflashattn${CMAKE_SHARED_LIBRARY_SUFFIX}"
+        CACHE FILEPATH "flash-attn Library" FORCE)
+    set(FLASHATTN_C_FLAGS ${CMAKE_C_FLAGS})
+    set(FLASHATTN_C_FLAGS_DEBUG ${CMAKE_C_FLAGS_DEBUG})
+    set(FLASHATTN_C_FLAGS_RELEASE ${CMAKE_C_FLAGS_RELEASE})
+    set(FLASHATTN_CXX_FLAGS "${CMAKE_CXX_FLAGS} -w -fPIC -O3 -std=c++17 -D__HIP_PLATFORM_HCC__=1 --offload-arch=gfx906 --offload-arch=gfx926 -D__gfx90a__")
+    set(FLASHATTN_CXX_FLAGS_RELEASE ${CMAKE_CXX_FLAGS_RELEASE})
+    set(FLASHATTN_CXX_FLAGS_DEBUG ${CMAKE_CXX_FLAGS_DEBUG})
+    ExternalProject_Add(
+      extern_flashattn
+      SOURCE_DIR ${SOURCE_DIR}
+      PREFIX ${FLASHATTN_PREFIX_DIR}
+      UPDATE_COMMAND ""
+      PATCH_COMMAND ""
+      #BUILD_ALWAYS    1
+      CMAKE_ARGS -DCMAKE_CXX_COMPILER=${ROCM_PATH}/bin/hipcc
+                -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
+                -DCMAKE_C_FLAGS=${FLASHATTN_C_FLAGS}
+                -DCMAKE_C_FLAGS_DEBUG=${FLASHATTN_C_FLAGS_DEBUG}
+                -DCMAKE_C_FLAGS_RELEASE=${FLASHATTN_C_FLAGS_RELEASE}
+                -DCMAKE_CXX_FLAGS=${FLASHATTN_CXX_FLAGS}
+                -DCMAKE_CXX_FLAGS_RELEASE=${FLASHATTN_CXX_FLAGS_RELEASE}
+                -DCMAKE_CXX_FLAGS_DEBUG=${FLASHATTN_CXX_FLAGS_DEBUG}
+                -DCMAKE_INSTALL_PREFIX=${FLASHATTN_INSTALL_DIR}
+                -DWITH_GPU=${WITH_GPU}
+                -DCMAKE_CUDA_COMPILER=${CMAKE_CUDA_COMPILER}
+                -DWITH_ROCM=${WITH_ROCM}
+                -DWITH_OMP=${USE_OMP}
+                -DBUILD_SHARED=ON
+                -DCMAKE_POSITION_INDEPENDENT_CODE=ON
+                -DCMAKE_BUILD_TYPE=${THIRD_PARTY_BUILD_TYPE}
+                -DCMAKE_JOB_POOL_COMPILE:STRING=compile
+                -DCMAKE_JOB_POOLS:STRING=compile=4
+                ${EXTERNAL_OPTIONAL_ARGS}
+      CMAKE_CACHE_ARGS
+        -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
+        -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
+        -DCMAKE_INSTALL_PREFIX:PATH=${FLASHATTN_INSTALL_DIR}
+      BUILD_BYPRODUCTS ${FLASHATTN_LIBRARIES})
-if(WIN32)
-  set(FLASHATTN_C_FLAGS $<FILTER:${CMAKE_C_FLAGS},EXCLUDE,/Zc:inline>)
-  set(FLASHATTN_C_FLAGS_DEBUG
-      $<FILTER:${CMAKE_C_FLAGS_DEBUG},EXCLUDE,/Zc:inline>)
-  set(FLASHATTN_C_FLAGS_RELEASE
-      $<FILTER:${CMAKE_C_FLAGS_RELEASE},EXCLUDE,/Zc:inline>)
-  set(FLASHATTN_CXX_FLAGS $<FILTER:${CMAKE_CXX_FLAGS},EXCLUDE,/Zc:inline>)
-  set(FLASHATTN_CXX_FLAGS_RELEASE
-      $<FILTER:${CMAKE_CXX_FLAGS_RELEASE},EXCLUDE,/Zc:inline>)
-  set(FLASHATTN_CXX_FLAGS_DEBUG
-      $<FILTER:${CMAKE_CXX_FLAGS_DEBUG},EXCLUDE,/Zc:inline>)
 else()
-  set(FLASHATTN_C_FLAGS ${CMAKE_C_FLAGS})
-  set(FLASHATTN_C_FLAGS_DEBUG ${CMAKE_C_FLAGS_DEBUG})
-  set(FLASHATTN_C_FLAGS_RELEASE ${CMAKE_C_FLAGS_RELEASE})
-  set(FLASHATTN_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++17")
-  set(FLASHATTN_CXX_FLAGS_RELEASE ${CMAKE_CXX_FLAGS_RELEASE})
-  set(FLASHATTN_CXX_FLAGS_DEBUG ${CMAKE_CXX_FLAGS_DEBUG})
-endif()
-ExternalProject_Add(
+    add_definitions(-DPADDLE_WITH_FLASHATTN)
-  extern_flashattn
-  ${EXTERNAL_PROJECT_LOG_ARGS}
+    set(FLASHATTN_PREFIX_DIR ${THIRD_PARTY_PATH}/flashattn)
-  SOURCE_DIR ${SOURCE_DIR}
+    set(FLASHATTN_SOURCE_SUBDIR csrc)
-  PREFIX ${FLASHATTN_PREFIX_DIR}
+    set(FLASHATTN_INSTALL_DIR ${THIRD_PARTY_PATH}/install/flashattn)
-  SOURCE_SUBDIR ${FLASHATTN_SOURCE_SUBDIR}
+    set(SOURCE_DIR ${PADDLE_SOURCE_DIR}/third_party/flashattn)
-  UPDATE_COMMAND ""
+    set(FLASHATTN_TAG 18106c1ba0ccee81b97ca947397c08a141815a47)
-  PATCH_COMMAND ""
-  #BUILD_ALWAYS    1
+    set(FLASHATTN_INCLUDE_DIR
-  CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
+        "${FLASHATTN_INSTALL_DIR}/include"
-             -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
+        CACHE PATH "flash-attn Directory" FORCE)
-             -DCMAKE_C_FLAGS=${FLASHATTN_C_FLAGS}
+    set(FLASHATTN_LIB_DIR
-             -DCMAKE_C_FLAGS_DEBUG=${FLASHATTN_C_FLAGS_DEBUG}
+        "${FLASHATTN_INSTALL_DIR}/lib"
-             -DCMAKE_C_FLAGS_RELEASE=${FLASHATTN_C_FLAGS_RELEASE}
+        CACHE PATH "flash-attn Library Directory" FORCE)
-             -DCMAKE_CXX_FLAGS=${FLASHATTN_CXX_FLAGS}
-             -DCMAKE_CXX_FLAGS_RELEASE=${FLASHATTN_CXX_FLAGS_RELEASE}
+    if(WIN32)
-             -DCMAKE_CXX_FLAGS_DEBUG=${FLASHATTN_CXX_FLAGS_DEBUG}
+      set(FLASHATTN_LIBRARIES
-             -DCMAKE_INSTALL_PREFIX=${FLASHATTN_INSTALL_DIR}
+          "${FLASHATTN_INSTALL_DIR}/bin/flashattn${CMAKE_SHARED_LIBRARY_SUFFIX}"
-             -DWITH_GPU=${WITH_GPU}
+          CACHE FILEPATH "flash-attn Library" FORCE)
-             -DCMAKE_CUDA_COMPILER=${CMAKE_CUDA_COMPILER}
+    else()
-             -DWITH_ROCM=${WITH_ROCM}
+      set(FLASHATTN_LIBRARIES
-             -DWITH_OMP=${USE_OMP}
+          "${FLASHATTN_INSTALL_DIR}/lib/libflashattn${CMAKE_SHARED_LIBRARY_SUFFIX}"
-             -DBUILD_SHARED=ON
+          CACHE FILEPATH "flash-attn Library" FORCE)
-             -DCMAKE_POSITION_INDEPENDENT_CODE=ON
+    endif()
-             -DCMAKE_BUILD_TYPE=${THIRD_PARTY_BUILD_TYPE}
-             -DCMAKE_JOB_POOL_COMPILE:STRING=compile
+    if(CMAKE_CXX_COMPILER_ID STREQUAL "Clang"
-             -DCMAKE_JOB_POOLS:STRING=compile=4
+      OR CMAKE_CXX_COMPILER_ID STREQUAL "AppleClang"
-             ${EXTERNAL_OPTIONAL_ARGS}
+      OR WIN32)
-  CMAKE_CACHE_ARGS
+      set(USE_OMP OFF)
-    -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
+    else()
-    -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
+      set(USE_OMP ON)
-    -DCMAKE_INSTALL_PREFIX:PATH=${FLASHATTN_INSTALL_DIR}
+    endif()
-  BUILD_BYPRODUCTS ${FLASHATTN_LIBRARIES})
+    if(WIN32)
+      set(FLASHATTN_C_FLAGS $<FILTER:${CMAKE_C_FLAGS},EXCLUDE,/Zc:inline>)
+      set(FLASHATTN_C_FLAGS_DEBUG
+          $<FILTER:${CMAKE_C_FLAGS_DEBUG},EXCLUDE,/Zc:inline>)
+      set(FLASHATTN_C_FLAGS_RELEASE
+          $<FILTER:${CMAKE_C_FLAGS_RELEASE},EXCLUDE,/Zc:inline>)
+      set(FLASHATTN_CXX_FLAGS $<FILTER:${CMAKE_CXX_FLAGS},EXCLUDE,/Zc:inline>)
+      set(FLASHATTN_CXX_FLAGS_RELEASE
+          $<FILTER:${CMAKE_CXX_FLAGS_RELEASE},EXCLUDE,/Zc:inline>)
+      set(FLASHATTN_CXX_FLAGS_DEBUG
+          $<FILTER:${CMAKE_CXX_FLAGS_DEBUG},EXCLUDE,/Zc:inline>)
+    else()
+      set(FLASHATTN_C_FLAGS ${CMAKE_C_FLAGS})
+      set(FLASHATTN_C_FLAGS_DEBUG ${CMAKE_C_FLAGS_DEBUG})
+      set(FLASHATTN_C_FLAGS_RELEASE ${CMAKE_C_FLAGS_RELEASE})
+      set(FLASHATTN_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++17")
+      set(FLASHATTN_CXX_FLAGS_RELEASE ${CMAKE_CXX_FLAGS_RELEASE})
+      set(FLASHATTN_CXX_FLAGS_DEBUG ${CMAKE_CXX_FLAGS_DEBUG})
+    endif()
+    ExternalProject_Add(
+      extern_flashattn
+      ${EXTERNAL_PROJECT_LOG_ARGS}
+      SOURCE_DIR ${SOURCE_DIR}
+      PREFIX ${FLASHATTN_PREFIX_DIR}
+      SOURCE_SUBDIR ${FLASHATTN_SOURCE_SUBDIR}
+      UPDATE_COMMAND ""
+      PATCH_COMMAND ""
+      #BUILD_ALWAYS    1
+      CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
+                -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
+                -DCMAKE_C_FLAGS=${FLASHATTN_C_FLAGS}
+                -DCMAKE_C_FLAGS_DEBUG=${FLASHATTN_C_FLAGS_DEBUG}
+                -DCMAKE_C_FLAGS_RELEASE=${FLASHATTN_C_FLAGS_RELEASE}
+                -DCMAKE_CXX_FLAGS=${FLASHATTN_CXX_FLAGS}
+                -DCMAKE_CXX_FLAGS_RELEASE=${FLASHATTN_CXX_FLAGS_RELEASE}
+                -DCMAKE_CXX_FLAGS_DEBUG=${FLASHATTN_CXX_FLAGS_DEBUG}
+                -DCMAKE_INSTALL_PREFIX=${FLASHATTN_INSTALL_DIR}
+                -DWITH_GPU=${WITH_GPU}
+                -DCMAKE_CUDA_COMPILER=${CMAKE_CUDA_COMPILER}
+                -DWITH_ROCM=${WITH_ROCM}
+                -DWITH_OMP=${USE_OMP}
+                -DBUILD_SHARED=ON
+                -DCMAKE_POSITION_INDEPENDENT_CODE=ON
+                -DCMAKE_BUILD_TYPE=${THIRD_PARTY_BUILD_TYPE}
+                -DCMAKE_JOB_POOL_COMPILE:STRING=compile
+                -DCMAKE_JOB_POOLS:STRING=compile=4
+                ${EXTERNAL_OPTIONAL_ARGS}
+      CMAKE_CACHE_ARGS
+        -DCMAKE_BUILD_TYPE:STRING=${THIRD_PARTY_BUILD_TYPE}
+        -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
+        -DCMAKE_INSTALL_PREFIX:PATH=${FLASHATTN_INSTALL_DIR}
+      BUILD_BYPRODUCTS ${FLASHATTN_LIBRARIES})
+endif()
 message(STATUS "flash-attn library: ${FLASHATTN_LIBRARIES}")
 get_filename_component(FLASHATTN_LIBRARY_PATH ${FLASHATTN_LIBRARIES} DIRECTORY)

--- a/cmake/external/gflags.cmake
+++ b/cmake/external/gflags.cmake
@@ -102,3 +102,15 @@ if(WIN32)
    set_property(GLOBAL PROPERTY OS_DEPENDENCY_MODULES shlwapi.lib)
  endif()
 endif()
+# We have implemented a custom flags tool paddle_flags to replace gflags.
+# User can also choose to use gflags by setting WITH_GFLAGS=ON. But when
+# using paddle_flags, gflags is also needed for other third party libraries
+# including glog and brpc. So we can not remove gflags completely.
+set(flags_dep)
+if(WITH_GFLAGS)
+  list(APPEND flags_dep gflags)
+  add_definitions(-DPADDLE_WITH_GFLAGS)
+else()
+  list(APPEND flags_dep paddle_flags)
+endif()
--- a/cmake/external/gloo.cmake
+++ b/cmake/external/gloo.cmake
@@ -37,7 +37,7 @@ if(WITH_GPU)
    file(TO_NATIVE_PATH ${PADDLE_SOURCE_DIR}/patches/gloo/device.cc.patch
         native_dst)
    set(GLOO_PATCH_COMMAND
-        git checkout -- . && git checkout ${GLOO_TAG} &&patch -Nd
+        git checkout -- . && git checkout ${GLOO_TAG} && patch -Nd
        ${SOURCE_DIR}/gloo/transport/tcp < ${native_dst})
  endif()
 endif()
@@ -61,6 +61,17 @@ if(CMAKE_COMPILER_IS_GNUCC)
        ${SOURCE_DIR}/gloo/ < ${types_header})
  endif()
 endif()
+file(TO_NATIVE_PATH ${PADDLE_SOURCE_DIR}/patches/gloo/linux.cc.patch
+     linux_cc_ethtool)
+if(GLOO_PATCH_COMMAND STREQUAL "")
+  set(GLOO_PATCH_COMMAND git checkout -- . && git checkout ${GLOO_TAG} && patch
+                         -Nd ${SOURCE_DIR}/gloo/common/ < ${linux_cc_ethtool})
+else()
+  set(GLOO_PATCH_COMMAND ${GLOO_PATCH_COMMAND} && patch -Nd
+                         ${SOURCE_DIR}/gloo/common/ < ${linux_cc_ethtool})
+endif()
 include_directories(${GLOO_INCLUDE_DIR})
 ExternalProject_Add(

--- a/cmake/external/jemalloc.cmake
+++ b/cmake/external/jemalloc.cmake
@@ -5,9 +5,7 @@ set(JEMALLOC_DOWNLOAD_DIR
 set(JEMALLOC_PROJECT "extern_jemalloc")
 set(JEMALLOC_BUILD ${THIRD_PARTY_PATH}/jemalloc/src/extern_jemalloc)
 set(JEMALLOC_PREFIX_DIR ${THIRD_PARTY_PATH}/jemalloc)
-set(JEMALLOC_URL
+set(JEMALLOC_URL https://paddle-ci.gz.bcebos.com/jemalloc-5.1.0.tar.bz2)
-    https://github.com/jemalloc/jemalloc/releases/download/5.1.0/jemalloc-5.1.0.tar.bz2
-)
 set(JEMALLOC_INSTALL ${THIRD_PARTY_PATH}/install/jemalloc)
 set(JEMALLOC_INCLUDE_DIR ${JEMALLOC_INSTALL}/include)

--- a/cmake/external/libxsmm.cmake
+++ b/cmake/external/libxsmm.cmake
@@ -31,9 +31,8 @@ set(LIBXSMMNOBLAS_LIB "${LIBXSMM_LIBRARY_DIR}/libxsmmnoblas.a")
 file(GLOB LIBXSMM_SOURCE_FILE_LIST ${LIBXSMM_SOURCE_DIR})
 list(LENGTH LIBXSMM_SOURCE_FILE_LIST RES_LEN)
 if(RES_LEN EQUAL 0)
-  execute_process(
+  execute_process(COMMAND ${GIT_EXECUTABLE} clone -b ${LIBXSMM_TAG}
-    COMMAND ${GIT_EXECUTABLE} clone -b ${LIBXSMM_TAG}
+                          "${GIT_URL}/hfp/libxsmm.git" ${LIBXSMM_SOURCE_DIR})
-            "https://github.com/hfp/libxsmm.git" ${LIBXSMM_SOURCE_DIR})
 else()
  # check git tag
  execute_process(

--- a/cmake/external/onnxruntime.cmake
+++ b/cmake/external/onnxruntime.cmake
@@ -44,19 +44,19 @@ set(ONNXRUNTIME_DOWNLOAD_DIR
 if(WIN32)
  set(ONNXRUNTIME_URL
-      "https://github.com/microsoft/onnxruntime/releases/download/v${ONNXRUNTIME_VERSION}/onnxruntime-win-x64-${ONNXRUNTIME_VERSION}.zip"
+      "${GIT_URL}/microsoft/onnxruntime/releases/download/v${ONNXRUNTIME_VERSION}/onnxruntime-win-x64-${ONNXRUNTIME_VERSION}.zip"
  )
  set(ONNXRUNTIME_URL_MD5 f21d6bd1feef15935a5f4e1007797593)
  set(ONNXRUNTIME_CACHE_EXTENSION "zip")
 elseif(APPLE)
  set(ONNXRUNTIME_URL
-      "https://github.com/microsoft/onnxruntime/releases/download/v${ONNXRUNTIME_VERSION}/onnxruntime-osx-x86_64-${ONNXRUNTIME_VERSION}.tgz"
+      "${GIT_URL}/microsoft/onnxruntime/releases/download/v${ONNXRUNTIME_VERSION}/onnxruntime-osx-x86_64-${ONNXRUNTIME_VERSION}.tgz"
  )
  set(ONNXRUNTIME_URL_MD5 6a6f6b7df97587da59976042f475d3f4)
  set(ONNXRUNTIME_CACHE_EXTENSION "tgz")
 else()
  set(ONNXRUNTIME_URL
-      "https://github.com/microsoft/onnxruntime/releases/download/v${ONNXRUNTIME_VERSION}/onnxruntime-linux-x64-${ONNXRUNTIME_VERSION}.tgz"
+      "${GIT_URL}/microsoft/onnxruntime/releases/download/v${ONNXRUNTIME_VERSION}/onnxruntime-linux-x64-${ONNXRUNTIME_VERSION}.tgz"
  )
  set(ONNXRUNTIME_URL_MD5 ce3f2376854b3da4b483d6989666995a)
  set(ONNXRUNTIME_CACHE_EXTENSION "tgz")

--- a/cmake/external/openblas.cmake
+++ b/cmake/external/openblas.cmake
@@ -46,9 +46,8 @@ endif()
 file(GLOB CBLAS_SOURCE_FILE_LIST ${CBLAS_SOURCE_DIR})
 list(LENGTH CBLAS_SOURCE_FILE_LIST RES_LEN)
 if(RES_LEN EQUAL 0)
-  execute_process(
+  execute_process(COMMAND ${GIT_EXECUTABLE} clone -b ${CBLAS_TAG}
-    COMMAND ${GIT_EXECUTABLE} clone -b ${CBLAS_TAG}
+                          "${GIT_URL}/xianyi/OpenBLAS.git" ${CBLAS_SOURCE_DIR})
-            "https://github.com/xianyi/OpenBLAS.git" ${CBLAS_SOURCE_DIR})
 else()
  # check git tag
  execute_process(
@@ -93,7 +92,7 @@ if(NOT WIN32)
    PREFIX ${CBLAS_PREFIX_DIR}
    INSTALL_DIR ${CBLAS_INSTALL_DIR}
    BUILD_IN_SOURCE 1
-    BUILD_COMMAND make ${ARM_ARGS} -j${NPROC} ${COMMON_ARGS} ${OPTIONAL_ARGS}
+    BUILD_COMMAND make ${ARM_ARGS} -s -j${NPROC} ${COMMON_ARGS} ${OPTIONAL_ARGS}
    INSTALL_COMMAND make install NO_SHARED=1 NO_LAPACK=1 PREFIX=<INSTALL_DIR>
    UPDATE_COMMAND ""
    CONFIGURE_COMMAND ""
@@ -121,6 +120,7 @@ else()
               -DCMAKE_POSITION_INDEPENDENT_CODE=ON
               -DCMAKE_BUILD_TYPE=${THIRD_PARTY_BUILD_TYPE}
               -DBUILD_SHARED_LIBS=ON
+               -DCMAKE_VERBOSE_MAKEFILE=OFF
               -DMSVC_STATIC_CRT=${MSVC_STATIC_CRT}
               ${EXTERNAL_OPTIONAL_ARGS}
    CMAKE_CACHE_ARGS

--- a/cmake/external/paddle2onnx.cmake
+++ b/cmake/external/paddle2onnx.cmake
@@ -71,19 +71,19 @@ endif()
 if(WIN32)
  set(PADDLE2ONNX_URL
-      "https://github.com/PaddlePaddle/Paddle2ONNX/releases/download/v${PADDLE2ONNX_VERSION}/paddle2onnx-win-x64-${PADDLE2ONNX_VERSION}.zip"
+      "${GIT_URL}/PaddlePaddle/Paddle2ONNX/releases/download/v${PADDLE2ONNX_VERSION}/paddle2onnx-win-x64-${PADDLE2ONNX_VERSION}.zip"
  )
  set(PADDLE2ONNX_URL_MD5 "122b864cb57338191a7e9ef5f607c4ba")
  set(PADDLE2ONNX_CACHE_EXTENSION "zip")
 elseif(APPLE)
  set(PADDLE2ONNX_URL
-      "https://github.com/PaddlePaddle/Paddle2ONNX/releases/download/v${PADDLE2ONNX_VERSION}/paddle2onnx-osx-x86_64-${PADDLE2ONNX_VERSION}.tgz"
+      "${GIT_URL}/PaddlePaddle/Paddle2ONNX/releases/download/v${PADDLE2ONNX_VERSION}/paddle2onnx-osx-x86_64-${PADDLE2ONNX_VERSION}.tgz"
  )
  set(PADDLE2ONNX_URL_MD5 "32a4381ff8441b69d58ef0fd6fd919eb")
  set(PADDLE2ONNX_CACHE_EXTENSION "tgz")
 else()
  set(PADDLE2ONNX_URL
-      "https://github.com/PaddlePaddle/Paddle2ONNX/releases/download/v${PADDLE2ONNX_VERSION}/paddle2onnx-linux-x64-${PADDLE2ONNX_VERSION}.tgz"
+      "${GIT_URL}/PaddlePaddle/Paddle2ONNX/releases/download/v${PADDLE2ONNX_VERSION}/paddle2onnx-linux-x64-${PADDLE2ONNX_VERSION}.tgz"
  )
  set(PADDLE2ONNX_URL_MD5 "3fbb074987ba241327797f76514e937f")
  set(PADDLE2ONNX_CACHE_EXTENSION "tgz")

--- a/cmake/external/protobuf.cmake
+++ b/cmake/external/protobuf.cmake
@@ -182,7 +182,7 @@ if(NOT "${PROTOBUF_ROOT}" STREQUAL "")
  endif()
 endif()
-function(build_protobuf TARGET_NAME BUILD_FOR_HOST)
+function(build_protobuf TARGET_NAME)
  string(REPLACE "extern_" "" TARGET_DIR_NAME "${TARGET_NAME}")
  set(PROTOBUF_PREFIX_DIR ${THIRD_PARTY_PATH}/${TARGET_DIR_NAME})
  set(PROTOBUF_SOURCE_DIR
@@ -210,22 +210,18 @@ function(build_protobuf TARGET_NAME BUILD_FOR_HOST)
      PARENT_SCOPE)
  set(OPTIONAL_CACHE_ARGS "")
-  set(OPTIONAL_ARGS "")
+  set(OPTIONAL_ARGS
-  if(BUILD_FOR_HOST)
+      "-DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}"
-    set(OPTIONAL_ARGS "-Dprotobuf_WITH_ZLIB=OFF")
+      "-DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}"
-  else()
+      "-DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}"
-    set(OPTIONAL_ARGS
+      "-DCMAKE_C_FLAGS_DEBUG=${CMAKE_C_FLAGS_DEBUG}"
-        "-DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}"
+      "-DCMAKE_C_FLAGS_RELEASE=${CMAKE_C_FLAGS_RELEASE}"
-        "-DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}"
+      "-DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}"
-        "-DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}"
+      "-DCMAKE_CXX_FLAGS_RELEASE=${CMAKE_CXX_FLAGS_RELEASE}"
-        "-DCMAKE_C_FLAGS_DEBUG=${CMAKE_C_FLAGS_DEBUG}"
+      "-DCMAKE_CXX_FLAGS_DEBUG=${CMAKE_CXX_FLAGS_DEBUG}"
-        "-DCMAKE_C_FLAGS_RELEASE=${CMAKE_C_FLAGS_RELEASE}"
+      "-Dprotobuf_WITH_ZLIB=ON"
-        "-DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}"
+      ${EXTERNAL_OPTIONAL_ARGS})
-        "-DCMAKE_CXX_FLAGS_RELEASE=${CMAKE_CXX_FLAGS_RELEASE}"
+  if(NOT APPLE)
-        "-DCMAKE_CXX_FLAGS_DEBUG=${CMAKE_CXX_FLAGS_DEBUG}"
-        "-Dprotobuf_WITH_ZLIB=ON"
-        "-DZLIB_ROOT:FILEPATH=${ZLIB_ROOT}"
-        ${EXTERNAL_OPTIONAL_ARGS})
    set(OPTIONAL_CACHE_ARGS "-DZLIB_ROOT:STRING=${ZLIB_ROOT}")
  endif()
  if(WIN32)
@@ -244,7 +240,7 @@ function(build_protobuf TARGET_NAME BUILD_FOR_HOST)
    set(PROTOBUF_TAG 01a05a53f40ca2ac5f0af10c6cc0810bee39b792)
  else()
    if(WITH_PSLIB)
-      set(PROTOBUF_REPOSITORY "https://github.com/google/protobuf.git")
+      set(PROTOBUF_REPOSITORY "${GIT_URL}/google/protobuf.git")
      set(PROTOBUF_TAG "9f75c5aa851cd877fb0d93ccc31b8567a6706546")
    else()
      set(PROTOBUF_REPOSITORY ${GIT_URL}/protocolbuffers/protobuf.git)
@@ -339,7 +335,7 @@ else()
 endif()
 if(NOT PROTOBUF_FOUND)
-  build_protobuf(extern_protobuf FALSE)
+  build_protobuf(extern_protobuf)
  set(PROTOBUF_INCLUDE_DIR
      ${extern_protobuf_INCLUDE_DIR}

--- a/cmake/external/xpu.cmake
+++ b/cmake/external/xpu.cmake
@@ -24,9 +24,9 @@ set(XPU_XFT_LIB_NAME "libxft.so")
 set(XPU_XPTI_LIB_NAME "libxpti.so")
 if(NOT DEFINED XPU_BASE_DATE)
-  set(XPU_BASE_DATE "20230810")
+  set(XPU_BASE_DATE "20231114")
 endif()
-set(XPU_XCCL_BASE_VERSION "1.0.53.6")
+set(XPU_XCCL_BASE_VERSION "1.1.6.1")
 if(NOT DEFINED XPU_XFT_BASE_VERSION)
  set(XPU_XFT_BASE_VERSION "20230602")
 endif()

--- a/cmake/flags.cmake
+++ b/cmake/flags.cmake
@@ -141,6 +141,11 @@ if(NOT WIN32)
  set(COMMON_FLAGS
      -fPIC
      -fno-omit-frame-pointer
+      -pipe
+      -ffunction-sections
+      -fdata-sections
+      -Wl
+      -gc-sections
      -Werror
      -Wall
      -Wextra