Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into cycliclr

Author: Asthestarsfalll

CMakeLists.txt

@@ -60,6 +60,7 @@ option(WITH_ONNXRUNTIME         "Compile PaddlePaddle with ONNXRUNTIME"
 # Note(zhouwei): It use option above, so put here
 include(init)
 include(generic)            # simplify cmake module
+include(experimental)       # experimental build options
 
 if (WITH_GPU  AND WITH_XPU)
     message(FATAL_ERROR "Error when compile GPU and XPU at the same time")

cmake/cblas.cmake

@@ -52,6 +52,7 @@ if(NOT DEFINED CBLAS_PROVIDER)
   set(OPENBLAS_INCLUDE_SEARCH_PATHS
           ${OPENBLAS_ROOT}/include
           /usr/include
+          /usr/include/lapacke
           /usr/include/openblas
           /usr/local/opt/openblas/include)
   set(OPENBLAS_LIB_SEARCH_PATHS
@@ -65,15 +66,17 @@ if(NOT DEFINED CBLAS_PROVIDER)
     PATHS ${OPENBLAS_INCLUDE_SEARCH_PATHS} NO_DEFAULT_PATH)
   find_path(OPENBLAS_LAPACKE_INC_DIR NAMES lapacke.h
     PATHS ${OPENBLAS_INCLUDE_SEARCH_PATHS})
+  find_path(OPENBLAS_CONFIG_INC_DIR NAMES openblas_config.h
+    PATHS ${OPENBLAS_INCLUDE_SEARCH_PATHS})
   find_library(OPENBLAS_LIB NAMES openblas
     PATHS ${OPENBLAS_LIB_SEARCH_PATHS})
 
-  if(OPENBLAS_LAPACKE_INC_DIR AND OPENBLAS_INC_DIR AND OPENBLAS_LIB)
-    file(READ "${OPENBLAS_INC_DIR}/openblas_config.h" config_file)
+  if(OPENBLAS_LAPACKE_INC_DIR AND OPENBLAS_INC_DIR AND OPENBLAS_CONFIG_INC_DIR AND OPENBLAS_LIB)
+    file(READ "${OPENBLAS_CONFIG_INC_DIR}/openblas_config.h" config_file)
     string(REGEX MATCH "OpenBLAS ([0-9]+\.[0-9]+\.[0-9]+)" tmp ${config_file})
     string(REGEX MATCH "([0-9]+\.[0-9]+\.[0-9]+)" ver ${tmp})
 
-    if (${ver} VERSION_GREATER_EQUAL "0.3.7")
+    if (${ver} VERSION_GREATER_EQUAL "0.3.5")
       set(CBLAS_PROVIDER OPENBLAS)
       set(CBLAS_INC_DIR ${OPENBLAS_INC_DIR} ${OPENBLAS_LAPACKE_INC_DIR})
       set(CBLAS_LIBRARIES ${OPENBLAS_LIB})
@@ -138,4 +141,3 @@ if(${CBLAS_PROVIDER} STREQUAL REFERENCE_CBLAS)
 elseif(NOT ${CBLAS_PROVIDER} STREQUAL MKLML)
   target_link_libraries(cblas ${CBLAS_LIBRARIES})
 endif()
-

cmake/experimental.cmake

@@ -0,0 +1,17 @@
+# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+# 
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+# 
+# http://www.apache.org/licenses/LICENSE-2.0
+# 
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# this file contains experimental build options
+
+include(experiments/cuda_module_loading_lazy)

cmake/experiments/cuda_module_loading_lazy.cmake

@@ -0,0 +1,40 @@
+# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+# 
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+# 
+# http://www.apache.org/licenses/LICENSE-2.0
+# 
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# this file contains experimental build options for lazy cuda module loading
+# cuda moduel lazy loading is supported by CUDA 11.6+
+# this experiment option makes Paddle supports lazy loading before CUDA 11.6.
+
+option(EXP_CUDA_MODULE_LOADING_LAZY  "enable lazy cuda module loading" OFF)
+if (${EXP_CUDA_MODULE_LOADING_LAZY})
+  if (NOT ${ON_INFER} OR NOT ${LINUX})
+    message("EXP_CUDA_MODULE_LOADING_LAZY only works with ON_INFER=ON on Linux platforms")
+    return()
+  endif ()
+  if (NOT ${CUDA_FOUND})
+    message("EXP_CUDA_MODULE_LOADING_LAZY only works with CUDA")
+    return()
+  endif ()
+  if (${CUDA_VERSION} VERSION_GREATER_EQUAL "11.6")
+    message("cuda 11.6+ already support lazy module loading")
+    return()
+  endif ()
+
+  message("for cuda before 11.6, libcudart.so must be used for the lazy module loading trick to work, instead of libcudart_static.a")
+  set(CUDA_USE_STATIC_CUDA_RUNTIME OFF CACHE BOOL "" FORCE)
+  set(CMAKE_CUDA_FLAGS "--cudart shared")
+  enable_language(CUDA)
+  set(CUDA_NVCC_EXECUTABLE "${CMAKE_SOURCE_DIR}/tools/nvcc_lazy" CACHE FILEPATH "" FORCE)
+  set(CMAKE_CUDA_COMPILER "${CMAKE_SOURCE_DIR}/tools/nvcc_lazy" CACHE FILEPATH "" FORCE)
+endif()

cmake/external/xpu.cmake

@@ -9,15 +9,15 @@ SET(XPU_RT_LIB_NAME             "libxpurt.so")
 
 if(NOT DEFINED XPU_BASE_URL)
   SET(XPU_BASE_URL_WITHOUT_DATE "https://baidu-kunlun-product.cdn.bcebos.com/KL-SDK/klsdk-dev")
-  SET(XPU_BASE_URL "${XPU_BASE_URL_WITHOUT_DATE}/20220511")
+  SET(XPU_BASE_URL "${XPU_BASE_URL_WITHOUT_DATE}/20220520")
 else()
   SET(XPU_BASE_URL "${XPU_BASE_URL}")
 endif()
 
 # ubuntu and centos: use output by XDNN API team
 if(NOT DEFINED XPU_XDNN_BASE_URL)
   SET(XPU_XDNN_BASE_URL_WITHOUT_DATE "https://klx-sdk-release-public.su.bcebos.com/xdnn/dev")
-  SET(XPU_XDNN_BASE_URL "${XPU_XDNN_BASE_URL_WITHOUT_DATE}/20220511")
+  SET(XPU_XDNN_BASE_URL "${XPU_XDNN_BASE_URL_WITHOUT_DATE}/20220520")
 else()
   SET(XPU_XDNN_BASE_URL "${XPU_XDNN_BASE_URL}")
 endif()

cmake/flags.cmake

@@ -142,12 +142,10 @@ set(COMMON_FLAGS
     -Wno-unused-function
     -Wno-error=literal-suffix
     -Wno-error=unused-local-typedefs
-    -Wno-error=parentheses-equality # Warnings in pybind11
     -Wno-error=ignored-attributes  # Warnings in Eigen, gcc 6.3
     -Wno-error=terminate  # Warning in PADDLE_ENFORCE
     -Wno-error=int-in-bool-context # Warning in Eigen gcc 7.2
     -Wimplicit-fallthrough=0 # Warning in tinyformat.h
-    -Wno-error=maybe-uninitialized # Warning in boost gcc 7.2
     ${fsanitize}
 )
 

paddle/fluid/distributed/collective/ProcessGroup.h

@@ -113,6 +113,19 @@ class ProcessGroup {
         "ProcessGroup%s does not support receive", GetBackendName()));
   }
 
+  virtual std::shared_ptr<ProcessGroup::Task> Send_Partial(phi::DenseTensor&,
+                                                           int, int,
+                                                           int) {  // NOLINT
+    PADDLE_THROW(platform::errors::InvalidArgument(
+        "ProcessGroup%s does not support send", GetBackendName()));
+  }
+
+  virtual std::shared_ptr<ProcessGroup::Task> Recv_Partial(
+      phi::DenseTensor& tensors, int, int, int) {  // NOLINT
+    PADDLE_THROW(platform::errors::InvalidArgument(
+        "ProcessGroup%s does not support receive", GetBackendName()));
+  }
+
   virtual std::shared_ptr<ProcessGroup::Task> AllGather(
       std::vector<phi::DenseTensor>&,    // NOLINT
       std::vector<phi::DenseTensor>&) {  // NOLINT

paddle/fluid/distributed/collective/ProcessGroupNCCL.cc

@@ -428,6 +428,53 @@ std::shared_ptr<ProcessGroup::Task> ProcessGroupNCCL::Recv(
   return task;
 }
 
+std::shared_ptr<ProcessGroup::Task> ProcessGroupNCCL::Send_Partial(
+    phi::DenseTensor& tensors, int dst_rank, int offset, int length) {
+  // CheckTensorsInDifferentDevices(tensors, static_cast<size_t>(GetSize()));
+
+  phi::DenseTensor flatten_tensor;
+  flatten_tensor.ShareDataWith(tensors).Resize({tensors.numel()});
+
+  phi::DenseTensor shared_input = flatten_tensor.Slice(offset, offset + length);
+
+  std::vector<phi::DenseTensor> shared_tensors;
+  shared_tensors.push_back(shared_input);
+
+  auto task = PointToPoint(shared_tensors,
+                           [&](phi::DenseTensor& input, ncclComm_t comm,
+                               const gpuStream_t& stream, int dst_rank) {
+                             return platform::dynload::ncclSend(
+                                 input.data(), input.numel(),
+                                 platform::ToNCCLDataType(input.dtype()),
+                                 dst_rank, comm, stream);
+                           },
+                           dst_rank, CommType::SEND);
+  return task;
+}
+
+std::shared_ptr<ProcessGroup::Task> ProcessGroupNCCL::Recv_Partial(
+    phi::DenseTensor& tensors, int src_rank, int offset, int length) {
+  // phi::DenseTensor shared_input = tensors.Slice(offset, offset+length);
+
+  phi::DenseTensor flatten_tensor;
+  flatten_tensor.ShareDataWith(tensors).Resize({tensors.numel()});
+  phi::DenseTensor shared_input = flatten_tensor.Slice(offset, offset + length);
+
+  std::vector<phi::DenseTensor> shared_tensors;
+  shared_tensors.push_back(shared_input);
+
+  auto task = PointToPoint(shared_tensors,
+                           [&](phi::DenseTensor& output, ncclComm_t comm,
+                               const gpuStream_t& stream, int src_rank) {
+                             return platform::dynload::ncclRecv(
+                                 output.data(), output.numel(),
+                                 platform::ToNCCLDataType(output.dtype()),
+                                 src_rank, comm, stream);
+                           },
+                           src_rank, CommType::RECV);
+  return task;
+}
+
 std::shared_ptr<ProcessGroup::Task> ProcessGroupNCCL::AllGather(
     std::vector<phi::DenseTensor>& in_tensors,
     std::vector<phi::DenseTensor>& out_tensors) {

paddle/fluid/distributed/collective/ProcessGroupNCCL.h

@@ -102,6 +102,14 @@ class ProcessGroupNCCL : public ProcessGroup {
   std::shared_ptr<ProcessGroup::Task> Recv(
       std::vector<phi::DenseTensor>& tensors, int src_rank) override;
 
+  std::shared_ptr<ProcessGroup::Task> Send_Partial(phi::DenseTensor& tensors,
+                                                   int dst_rank, int offset,
+                                                   int length) override;
+
+  std::shared_ptr<ProcessGroup::Task> Recv_Partial(phi::DenseTensor& tensors,
+                                                   int src_rank, int offset,
+                                                   int length) override;
+
   std::shared_ptr<ProcessGroup::Task> AllGather(
       std::vector<phi::DenseTensor>& in_tensors,
       std::vector<phi::DenseTensor>& out_tensors) override;

paddle/fluid/distributed/fleet_executor/dist_model.cc

@@ -546,9 +546,9 @@ bool DistModel::Run(const std::vector<DistModelTensor> &input_data,
 
   DistModelTimer timer;
   timer.tic();
-  double feed_elapse;
-  double fleet_exe_elapse;
-  double fetch_elapse;
+  double feed_elapse = 0;
+  double fleet_exe_elapse = 0;
+  double fetch_elapse = 0;
 
   if (!FeedData(input_data, scope_.get())) {
     LOG(ERROR) << "DistModel failed at feeding data.";