Fix

PaddlePaddle · Sep 17, 2024 · c7d1473 · c7d1473
2 parents e8a3d78 + ddd4f78
commit c7d1473
Show file tree

Hide file tree

Showing 1,124 changed files with 31,396 additions and 33,474 deletions.
diff --git a/.gitignore b/.gitignore
@@ -110,3 +110,4 @@ paddle/phi/kernels/fusion/cutlass/conv2d/build
 paddle/phi/kernels/fusion/cutlass/conv2d/cutlass
 paddle/phi/kernels/fusion/cutlass/gemm_epilogue/build
 paddle/phi/kernels/fusion/cutlass/gemm_epilogue/cutlass
+python/paddle/_typing/libs/**/*.pyi
diff --git a/cmake/ccache.cmake b/cmake/ccache.cmake
@@ -4,7 +4,7 @@ if(NOT WIN32)
   find_program(CCACHE_PATH ccache)
   if(CCACHE_PATH)
     execute_process(COMMAND ccache -V OUTPUT_VARIABLE ccache_output)
-    execute_process(COMMAND ccache -s cache directory
+    execute_process(COMMAND ccache -v -s cache directory
                     OUTPUT_VARIABLE cache_directory)
     string(REGEX MATCH "[0-9]+.[0-9]+" ccache_version ${ccache_output})
     message(STATUS "ccache is founded, use ccache to speed up compile on Unix.")
@@ -14,6 +14,7 @@ if(NOT WIN32)
     set(CMAKE_C_COMPILER_LAUNCHER ${CCACHE_PATH})
     set(CMAKE_CXX_COMPILER_LAUNCHER ${CCACHE_PATH})
     set(CMAKE_CUDA_COMPILER_LAUNCHER ${CCACHE_PATH})
+    set(CMAKE_HIP_COMPILER_LAUNCHER ${CCACHE_PATH})
   endif()
 elseif("${CMAKE_GENERATOR}" STREQUAL "Ninja")
   # (Note:zhouwei25) Only Ninja Generator can support sccache now

diff --git a/cmake/cuda.cmake b/cmake/cuda.cmake
@@ -244,6 +244,11 @@ function(select_nvcc_arch_flags out_variable out_arch_bin)
   string(REPLACE ";" " " nvcc_archs_readable "${nvcc_archs_readable}")
   string(REGEX MATCHALL "[0-9()]+" nvcc_archs_bin_list "${nvcc_archs_bin_list}")
   string(JOIN "," nvcc_real_archs ${nvcc_archs_bin_list})
+
+  set(COMPILED_CUDA_ARCHS
+      "${nvcc_real_archs}"
+      CACHE STRING "Specify compiled cuda archs.")
+
   set(${out_variable}
       ${nvcc_flags}
       PARENT_SCOPE)

diff --git a/cmake/external/afs_api.cmake b/cmake/external/afs_api.cmake
@@ -24,7 +24,7 @@ if((NOT DEFINED AFSAPI_VER) OR (NOT DEFINED AFSAPI_URL))
       "afs_api"
       CACHE STRING "" FORCE)
   set(AFSAPI_URL
-      "https://pslib.bj.bcebos.com/afs_api_so.tar.gz"
+      "https://fleet.bj.bcebos.com/heterps/afs_api.tar.gz"
       CACHE STRING "" FORCE)
 endif()
 message(STATUS "AFSAPI_NAME: ${AFSAPI_NAME}, AFSAPI_URL: ${AFSAPI_URL}")

diff --git a/cmake/external/flashattn.cmake b/cmake/external/flashattn.cmake
@@ -53,6 +53,7 @@ if(WITH_ROCM)
     #BUILD_ALWAYS    1
     CMAKE_ARGS -DCMAKE_CXX_COMPILER=${ROCM_PATH}/bin/hipcc
                -DAMDGPU_TARGETS=gfx928
+               -DCMAKE_CXX_COMPILER_LAUNCHER=${CCACHE_PATH}
                -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
                -DCMAKE_C_FLAGS=${FLASHATTN_C_FLAGS}
                -DCMAKE_C_FLAGS_DEBUG=${FLASHATTN_C_FLAGS_DEBUG}

diff --git a/cmake/external/xpu.cmake b/cmake/external/xpu.cmake
@@ -30,9 +30,9 @@ if(NOT DEFINED XPU_XRE_BASE_VERSION)
   set(XPU_XRE_BASE_VERSION "4.32.0.1")
 endif()
 if(NOT DEFINED XPU_XHPC_BASE_DATE)
-  set(XPU_XHPC_BASE_DATE "20240818")
+  set(XPU_XHPC_BASE_DATE "eb35/20240906")
 endif()
-set(XPU_XCCL_BASE_VERSION "1.2.9")
+set(XPU_XCCL_BASE_VERSION "1.2.11c")
 if(NOT DEFINED XPU_XFT_BASE_VERSION)
   set(XPU_XFT_BASE_VERSION "20230602")
 endif()
@@ -69,19 +69,9 @@ if(WITH_XPU_XRE5)
   )
 endif()
 
-if(WITH_XCCL_RDMA)
-  set(XPU_XCCL_PREFIX "xccl_rdma-")
-else()
-  set(XPU_XCCL_PREFIX "xccl_socket-")
-  # NOTE(lijin23): socket has not been supported for XPU3, so the xccl output name was changed.
-  if(WITH_XPU_XRE5)
-    set(XPU_XCCL_PREFIX "xccl_")
-  endif()
-endif()
-
 if(WITH_AARCH64)
   set(XPU_XRE_DIR_NAME "xre-kylin_aarch64")
-  set(XPU_XCCL_DIR_NAME "${XPU_XCCL_PREFIX}kylin_aarch64")
+  set(XPU_XCCL_DIR_NAME "") # TODO: xccl has no kylin output now.
   set(XPU_XFT_DIR_NAME "") # TODO: xft has no kylin output at now.
 elseif(WITH_SUNWAY)
   set(XPU_XRE_DIR_NAME "xre-deepin_sw6_64")
@@ -94,11 +84,11 @@ elseif(WITH_BDCENTOS)
   else()
     set(XPU_XRE_DIR_NAME "xre-bdcentos_x86_64")
   endif()
-  set(XPU_XCCL_DIR_NAME "${XPU_XCCL_PREFIX}bdcentos_x86_64")
+  set(XPU_XCCL_DIR_NAME "xccl_bdcentos_x86_64")
   set(XPU_XFT_DIR_NAME "xft_bdcentos6u3_x86_64_gcc82")
 elseif(WITH_CENTOS)
   set(XPU_XRE_DIR_NAME "xre-centos7_x86_64")
-  set(XPU_XCCL_DIR_NAME "${XPU_XCCL_PREFIX}bdcentos_x86_64")
+  set(XPU_XCCL_DIR_NAME "xccl_Linux_x86_64")
   set(XPU_XFT_DIR_NAME "xft_bdcentos6u3_x86_64_gcc82")
 else()
   # Ubuntu as default
@@ -109,7 +99,7 @@ else()
     set(XPU_XRE_DIR_NAME "xre-ubuntu_1604_x86_64")
     set(XPU_XHPC_DIR_NAME "xhpc-ubuntu1604_x86_64")
   endif()
-  set(XPU_XCCL_DIR_NAME "${XPU_XCCL_PREFIX}ubuntu_x86_64")
+  set(XPU_XCCL_DIR_NAME "xccl_Linux_x86_64")
   set(XPU_XFT_DIR_NAME "xft_ubuntu1604_x86_64")
 endif()
 
@@ -129,7 +119,7 @@ if(WITH_XPTI)
 endif()
 
 set(XPU_XHPC_URL
-    "https://klx-sdk-release-public.su.bcebos.com/xhpc/dev/${XPU_XHPC_BASE_DATE}/${XPU_XHPC_DIR_NAME}.tar.gz"
+    "https://klx-sdk-release-public.su.bcebos.com/xhpc/${XPU_XHPC_BASE_DATE}/${XPU_XHPC_DIR_NAME}.tar.gz"
     CACHE STRING "" FORCE)
 
 if(DEFINED XPU_BASE_URL)

diff --git a/cmake/generic.cmake b/cmake/generic.cmake
@@ -463,21 +463,35 @@ function(cc_test_run TARGET_NAME)
     set(multiValueArgs COMMAND ARGS)
     cmake_parse_arguments(cc_test "${options}" "${oneValueArgs}"
                           "${multiValueArgs}" ${ARGN})
+    string(REGEX MATCH "_deprecated$" DEPRECATED_TARGET_NAME "${TARGET_NAME}")
+
     if(cc_test_DIR STREQUAL "")
       set(cc_test_DIR ${CMAKE_CURRENT_BINARY_DIR})
     endif()
     add_test(
       NAME ${TARGET_NAME}
       COMMAND ${cc_test_COMMAND} ${cc_test_ARGS}
       WORKING_DIRECTORY ${cc_test_DIR})
-    set_property(
-      TEST ${TARGET_NAME}
-      PROPERTY
-        ENVIRONMENT
-        FLAGS_init_allocated_mem=true
-        FLAGS_cudnn_deterministic=true
-        LD_LIBRARY_PATH=$LD_LIBRARY_PATH:${PADDLE_BINARY_DIR}/python/paddle/libs:${PADDLE_BINARY_DIR}/python/paddle/base
-    )
+    if(NOT "${DEPRECATED_TARGET_NAME}" STREQUAL "")
+      set_property(
+        TEST ${TARGET_NAME}
+        PROPERTY
+          ENVIRONMENT
+          FLAGS_init_allocated_mem=true
+          FLAGS_cudnn_deterministic=true
+          FLAGS_enable_pir_api=0
+          LD_LIBRARY_PATH=$LD_LIBRARY_PATH:${PADDLE_BINARY_DIR}/python/paddle/libs:${PADDLE_BINARY_DIR}/python/paddle/base
+      )
+    else()
+      set_property(
+        TEST ${TARGET_NAME}
+        PROPERTY
+          ENVIRONMENT
+          FLAGS_init_allocated_mem=true
+          FLAGS_cudnn_deterministic=true
+          LD_LIBRARY_PATH=$LD_LIBRARY_PATH:${PADDLE_BINARY_DIR}/python/paddle/libs:${PADDLE_BINARY_DIR}/python/paddle/base
+      )
+    endif()
     # No unit test should exceed 2 minutes.
     if(WIN32)
       set_tests_properties(${TARGET_NAME} PROPERTIES TIMEOUT 150)
@@ -1147,14 +1161,21 @@ function(py_test TARGET_NAME)
     cmake_parse_arguments(py_test "${options}" "${oneValueArgs}"
                           "${multiValueArgs}" ${ARGN})
 
+    string(REGEX MATCH "_deprecated\.py$" DEPRECATED_MODULES "${py_test_SRCS}")
+    string(REGEX MATCH "_deprecated$" DEPRECATED_TARGET_NAME "${TARGET_NAME}")
+    set(FLAGS_PIR_MODE "")
+    if((NOT "${DEPRECATED_MODULES}" STREQUAL "")
+       OR (NOT "${DEPRECATED_TARGET_NAME}" STREQUAL ""))
+      set(FLAGS_PIR_MODE FLAGS_enable_pir_api=0)
+    endif()
     if(WITH_COVERAGE AND NOT (WITH_INCREMENTAL_COVERAGE
                               AND "$ENV{PADDLE_GIT_DIFF_PY_FILE}" STREQUAL ""))
       add_test(
         NAME ${TARGET_NAME}
         COMMAND
           ${CMAKE_COMMAND} -E env FLAGS_init_allocated_mem=true
-          FLAGS_cudnn_deterministic=true PYTHONPATH=${PADDLE_BINARY_DIR}/python
-          ${py_test_ENVS}
+          FLAGS_cudnn_deterministic=true ${FLAGS_PIR_MODE}
+          PYTHONPATH=${PADDLE_BINARY_DIR}/python ${py_test_ENVS}
           COVERAGE_FILE=${PADDLE_BINARY_DIR}/python-coverage.data
           ${PYTHON_EXECUTABLE} -m coverage run --branch -p ${py_test_SRCS}
           ${py_test_ARGS}
@@ -1164,8 +1185,8 @@ function(py_test TARGET_NAME)
         NAME ${TARGET_NAME}
         COMMAND
           ${CMAKE_COMMAND} -E env FLAGS_init_allocated_mem=true
-          FLAGS_cudnn_deterministic=true ${py_test_ENVS} ${PYTHON_EXECUTABLE} -u
-          ${py_test_SRCS} ${py_test_ARGS}
+          FLAGS_cudnn_deterministic=true ${FLAGS_PIR_MODE} ${py_test_ENVS}
+          ${PYTHON_EXECUTABLE} -u ${py_test_SRCS} ${py_test_ARGS}
         WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
     endif()
 

diff --git a/cmake/hip.cmake b/cmake/hip.cmake
@@ -75,6 +75,7 @@ endmacro()
 
 find_package_and_include(miopen)
 find_package_and_include(rocblas)
+find_package_and_include(hipblaslt)
 find_package_and_include(hiprand)
 find_package_and_include(rocrand)
 find_package_and_include(rccl)
@@ -85,11 +86,16 @@ find_package_and_include(hipsparse)
 find_package_and_include(rocsparse)
 find_package_and_include(rocfft)
 
+if(CCACHE_PATH)
+  set(HIP_HIPCC_EXECUTABLE ${CCACHE_PATH} ${HIP_HIPCC_EXECUTABLE})
+endif()
+
 # set CXX flags for HIP
 set(CMAKE_C_FLAGS
-    "${CMAKE_C_FLAGS} -D__HIP_PLATFORM_HCC__ -DROCM_NO_WRAPPER_HEADER_WARNING")
+    "${CMAKE_C_FLAGS} -D__HIP_PLATFORM_HCC__ -D__HIP_PLATFORM_AMD__ -DROCM_NO_WRAPPER_HEADER_WARNING"
+)
 set(CMAKE_CXX_FLAGS
-    "${CMAKE_CXX_FLAGS} -D__HIP_PLATFORM_HCC__ -DROCM_NO_WRAPPER_HEADER_WARNING"
+    "${CMAKE_CXX_FLAGS} -D__HIP_PLATFORM_HCC__ -D__HIP_PLATFORM_AMD__ -DROCM_NO_WRAPPER_HEADER_WARNING"
 )
 set(CMAKE_CXX_FLAGS
     "${CMAKE_CXX_FLAGS} -DTHRUST_DEVICE_SYSTEM=THRUST_DEVICE_SYSTEM_HIP")
@@ -98,6 +104,7 @@ set(THRUST_DEVICE_SYSTEM THRUST_DEVICE_SYSTEM_HIP)
 # define HIP_CXX_FLAGS
 list(APPEND HIP_CXX_FLAGS -fPIC)
 list(APPEND HIP_CXX_FLAGS -D__HIP_PLATFORM_HCC__=1)
+list(APPEND HIP_CXX_FLAGS -D__HIP_PLATFORM_AMD__=1)
 # Note(qili93): HIP has compile conflicts of float16.h as platform::float16 overload std::is_floating_point and std::is_integer
 list(APPEND HIP_CXX_FLAGS -D__HIP_NO_HALF_CONVERSIONS__=1)
 list(APPEND HIP_CXX_FLAGS -DROCM_NO_WRAPPER_HEADER_WARNING)

diff --git a/cmake/inference_lib.cmake b/cmake/inference_lib.cmake
@@ -399,6 +399,11 @@ copy(
   SRCS ${PADDLE_SOURCE_DIR}/paddle/pir/include/dialect/shape/utils/*.h
   DSTS ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/pir/dialect/shape/utils/
 )
+copy(
+  inference_lib_dist
+  SRCS ${PADDLE_SOURCE_DIR}/paddle/pir/include/dialect/shape/interface/infer_symbolic_shape/*.h
+  DSTS ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/pir/dialect/shape/interface/infer_symbolic_shape/
+)
 copy(
   inference_lib_dist
   SRCS ${PADDLE_SOURCE_DIR}/paddle/pir/include/pass/*.h

diff --git a/cmake/third_party.cmake b/cmake/third_party.cmake
@@ -521,11 +521,9 @@ if(WITH_PSCORE)
 
   include(external/jemalloc) # download, build, install jemalloc
   list(APPEND third_party_deps extern_jemalloc)
-  if(WITH_HETERPS)
-    include(external/afs_api)
-    list(APPEND third_party_deps extern_afs_api)
-    set(WITH_AFSAPI ON)
-  endif()
+
+  include(external/afs_api)
+  list(APPEND third_party_deps extern_afs_api)
 endif()
 
 if(WITH_RPC

diff --git a/paddle/cinn/adt/equation_util.h b/paddle/cinn/adt/equation_util.h
@@ -75,18 +75,18 @@ EquationGraphTopoWalker<VT, FT> GetSubgraph(
       };
   const auto& VisitInputVariables =
       [graph, IsSelected](FT function, const std::function<void(VT)>& Visit) {
-        PADDLE_ENFORCE_EQ(
-            IsSelected(function),
-            true,
-            phi::errors::PreconditionNotMet("The function must be selected."));
+        PADDLE_ENFORCE_EQ(IsSelected(function),
+                          true,
+                          ::common::errors::PreconditionNotMet(
+                              "The function must be selected."));
         graph.VisitInputVariables(function, Visit);
       };
   const auto& VisitOutputVariables =
       [graph, IsSelected](FT function, const std::function<void(VT)>& Visit) {
-        PADDLE_ENFORCE_EQ(
-            IsSelected(function),
-            true,
-            phi::errors::PreconditionNotMet("The function must be selected."));
+        PADDLE_ENFORCE_EQ(IsSelected(function),
+                          true,
+                          ::common::errors::PreconditionNotMet(
+                              "The function must be selected."));
         graph.VisitOutputVariables(function, Visit);
       };
   return EquationGraphTopoWalker<VT, FT>(

diff --git a/paddle/cinn/adt/m_ir.cc b/paddle/cinn/adt/m_ir.cc
@@ -178,8 +178,8 @@ std::unordered_map<Index, LoopIterators> GenerateAnchorIndex2LoopIterators(
     PADDLE_ENFORCE_EQ(
         anchor_index2loop_iters.emplace(anchor_index, anchor_loop_iters).second,
         true,
-        phi::errors::AlreadyExists("The anchor index has already "
-                                   "been associated with loop iters."));
+        ::common::errors::AlreadyExists("The anchor index has already "
+                                        "been associated with loop iters."));
   }
   return anchor_index2loop_iters;
 }

diff --git a/paddle/cinn/adt/naive_op_equation_context.h b/paddle/cinn/adt/naive_op_equation_context.h
@@ -86,13 +86,13 @@ class NaiveOpEquationContext final : public OpEquationContext {
   }
 
   void Equal(const IteratorTuple& lhs, const IteratorTuple& rhs) override {
-    PADDLE_ENFORCE_EQ(
-        lhs->size(),
-        rhs->size(),
-        phi::errors::InvalidArgument("The sizes of lhs and rhs must be equal. "
-                                     "lhs size: %d, rhs size: %d",
-                                     lhs->size(),
-                                     rhs->size()));
+    PADDLE_ENFORCE_EQ(lhs->size(),
+                      rhs->size(),
+                      ::common::errors::InvalidArgument(
+                          "The sizes of lhs and rhs must be equal. "
+                          "lhs size: %d, rhs size: %d",
+                          lhs->size(),
+                          rhs->size()));
     for (std::size_t i = 0; i < lhs->size(); ++i) {
       this->Equal(lhs->at(i), rhs->at(i));
     }
@@ -259,7 +259,7 @@ class NaiveOpEquationContext final : public OpEquationContext {
         const auto& opt_expr = GetSymbolicInDim_(i, j);
         PADDLE_ENFORCE_EQ(opt_expr.has_value(),
                           true,
-                          phi::errors::InvalidArgument(
+                          ::common::errors::InvalidArgument(
                               "The optional expression must have a value."));
         vec->at(i)->emplace_back(opt_expr.value());
       }
@@ -274,7 +274,7 @@ class NaiveOpEquationContext final : public OpEquationContext {
         const auto& opt_expr = GetSymbolicOutDim_(i, j);
         PADDLE_ENFORCE_EQ(opt_expr.has_value(),
                           true,
-                          phi::errors::InvalidArgument(
+                          ::common::errors::InvalidArgument(
                               "The optional expression must have a value at "
                               "tensor index %d and dimension index %d.",
                               i,
@@ -288,7 +288,7 @@ class NaiveOpEquationContext final : public OpEquationContext {
                  const DimTuple& dim_tuple) {
     PADDLE_ENFORCE_EQ(iterator_tuple->size(),
                       dim_tuple->size(),
-                      phi::errors::InvalidArgument(
+                      ::common::errors::InvalidArgument(
                           "The sizes of iterator_tuple and dim_tuple must be "
                           "equal. iterator_tuple size: %d, dim_tuple size: %d",
                           iterator_tuple->size(),
@@ -343,8 +343,8 @@ class NaiveOpEquationContext final : public OpEquationContext {
     PADDLE_ENFORCE_EQ(
         iter != attr_map_type_.end(),
         true,
-        phi::errors::InvalidArgument("Can't find Attribute with this name: %s",
-                                     name.c_str()));
+        ::common::errors::InvalidArgument(
+            "Can't find Attribute with this name: %s", name.c_str()));
     return iter->second;
   }