Merge From Develop

PaddlePaddle · Nov 3, 2021 · a88ca81 · a88ca81 · paddle-bot-old · Nov 3, 2021
2 parents 2ca5414 + 3121f88
commit a88ca81
Show file tree

Hide file tree

Showing 102 changed files with 2,327 additions and 463 deletions.
diff --git a/cmake/external/xpu.cmake b/cmake/external/xpu.cmake
@@ -35,7 +35,8 @@ ELSE ()
 ENDIF()
 
 SET(XPU_BASE_URL_WITHOUT_DATE "https://baidu-kunlun-product.cdn.bcebos.com/KL-SDK/klsdk-dev")
-SET(XPU_BASE_URL "${XPU_BASE_URL_WITHOUT_DATE}/20211020")
+SET(XPU_BASE_URL "${XPU_BASE_URL_WITHOUT_DATE}/20211029")
+#SET(XPU_BASE_URL "${XPU_BASE_URL_WITHOUT_DATE}/20211020")
 SET(XPU_XRE_URL  "${XPU_BASE_URL}/${XPU_XRE_DIR_NAME}.tar.gz" CACHE STRING "" FORCE)
 SET(XPU_XDNN_URL "${XPU_BASE_URL}/${XPU_XDNN_DIR_NAME}.tar.gz" CACHE STRING "" FORCE)
 SET(XPU_XCCL_URL "${XPU_BASE_URL_WITHOUT_DATE}/20210623/${XPU_XCCL_DIR_NAME}.tar.gz" CACHE STRING "" FORCE)

diff --git a/paddle/fluid/distributed/CMakeLists.txt b/paddle/fluid/distributed/CMakeLists.txt
@@ -1,4 +1,5 @@
 if(NOT WITH_PSCORE)
+    add_subdirectory(fleet_executor)
     return()
 endif()
 
@@ -16,6 +17,7 @@ add_subdirectory(service)
 add_subdirectory(table)
 add_subdirectory(test)
 add_subdirectory(index_dataset)
+add_subdirectory(fleet_executor)
 
 get_property(RPC_DEPS GLOBAL PROPERTY RPC_DEPS)
 

diff --git a/paddle/fluid/distributed/fleet_executor/CMakeLists.txt b/paddle/fluid/distributed/fleet_executor/CMakeLists.txt
@@ -0,0 +1,6 @@
+proto_library(fleet_executor_desc_proto SRCS fleet_executor_desc.proto)
+cc_library(fleet_executor SRCS fleet_executor.cc DEPS fleet_executor_desc_proto)
+
+if(WITH_PYTHON)
+  py_proto_compile(fleet_executor_desc_py_proto SRCS fleet_executor_desc.proto)
+endif()
diff --git a/paddle/fluid/distributed/fleet_executor/fleet_executor.cc b/paddle/fluid/distributed/fleet_executor/fleet_executor.cc
@@ -0,0 +1,43 @@
+// Copyright (c) 2021 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.
+
+#include "paddle/fluid/distributed/fleet_executor/fleet_executor.h"
+#include "paddle/fluid/distributed/fleet_executor/runtime_graph.h"
+#include "paddle/fluid/framework/program_desc.h"
+
+namespace paddle {
+namespace distributed {
+
+FleetExecutor::FleetExecutor(const std::string& exe_desc_str) {
+  // Initialize Executor
+}
+
+FleetExecutor::~FleetExecutor() {
+  // Destroy Executor
+}
+
+void FleetExecutor::Init(const paddle::framework::ProgramDesc& program_desc) {
+  // Compile and Initialize
+}
+
+void FleetExecutor::Run() {
+  // Run
+}
+
+void FleetExecutor::Release() {
+  // Release
+}
+
+}  // namespace distributed
+}  // namespace paddle
diff --git a/paddle/fluid/distributed/fleet_executor/fleet_executor.h b/paddle/fluid/distributed/fleet_executor/fleet_executor.h
@@ -0,0 +1,44 @@
+// Copyright (c) 2021 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.
+
+#pragma once
+#include <memory>
+#include "paddle/fluid/distributed/fleet_executor/fleet_executor_desc.pb.h"
+#include "paddle/fluid/platform/macros.h"
+
+namespace paddle {
+namespace framework {
+class ProgramDesc;
+}
+
+namespace distributed {
+class RuntimeGraph;
+
+class FleetExecutor final {
+ public:
+  FleetExecutor() = delete;
+  FleetExecutor(const std::string& exe_desc_str);
+  ~FleetExecutor();
+  void Init(const paddle::framework::ProgramDesc& program_desc);
+  void Run();
+  void Release();
+
+ private:
+  DISABLE_COPY_AND_ASSIGN(FleetExecutor);
+  FleetExecutorDesc exe_desc_;
+  std::unique_ptr<RuntimeGraph> runtime_graph_;
+};
+
+}  // namespace distributed
+}  // namespace paddle
diff --git a/paddle/fluid/distributed/fleet_executor/fleet_executor_desc.proto b/paddle/fluid/distributed/fleet_executor/fleet_executor_desc.proto
@@ -0,0 +1,21 @@
+// Copyright (c) 2021 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.
+
+syntax = "proto2";
+package paddle.distributed;
+
+message FleetExecutorDesc {
+  optional string grain = 1 [ default = "coarse" ];
+  repeated string addrs = 2; // "ip:port" of all ranks
+}
diff --git a/paddle/pten/hapi/include/creation.h → ...istributed/fleet_executor/runtime_graph.h b/paddle/pten/hapi/include/creation.h → ...istributed/fleet_executor/runtime_graph.h
@@ -14,20 +14,21 @@
 
 #pragma once
 
-#include "paddle/pten/common/data_type.h"
-#include "paddle/pten/common/scalar.h"
-#include "paddle/pten/hapi/include/tensor.h"
-
 namespace paddle {
-namespace experimental {
+namespace framework {
+class ProgramDesc;
+}
 
-Tensor full_like(const Tensor& x,
-                 const Scalar& value,
-                 DataType dtype = DataType::UNDEFINED);
+namespace distributed {
 
-Tensor ones_like(const Tensor& x, DataType dtype = DataType::UNDEFINED);
+class RuntimeGraph final {
+ public:
+  RuntimeGraph() = default;
+  explicit RuntimeGraph(const paddle::framework::ProgramDesc &program) {}
+  ~RuntimeGraph() = default;
 
-Tensor zeros_like(const Tensor& x, DataType dtype = DataType::UNDEFINED);
+  DISABLE_COPY_AND_ASSIGN(RuntimeGraph);
+};
 
-}  // namespace experimental
+}  // namespace distributed
 }  // namespace paddle
diff --git a/paddle/fluid/distributed/table/CMakeLists.txt b/paddle/fluid/distributed/table/CMakeLists.txt
@@ -37,7 +37,9 @@ set_source_files_properties(table.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPI
 
 set_source_files_properties(sparse_sgd_rule.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
 set_source_files_properties(ctr_accessor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
+set_source_files_properties(memory_sparse_table.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
 cc_library(sparse_sgd_rule SRCS sparse_sgd_rule.cc DEPS ${TABLE_DEPS} ps_framework_proto)
 cc_library(ctr_accessor SRCS ctr_accessor.cc DEPS ${TABLE_DEPS} ps_framework_proto sparse_sgd_rule)
+cc_library(memory_sparse_table SRCS memory_sparse_table.cc DEPS ps_framework_proto ${TABLE_DEPS} fs afs_wrapper ctr_accessor common_table)
 
-cc_library(table SRCS table.cc DEPS common_table tensor_accessor tensor_table ps_framework_proto string_helper device_context gflags glog boost ctr_accessor)
+cc_library(table SRCS table.cc DEPS memory_sparse_table common_table tensor_accessor tensor_table ps_framework_proto string_helper device_context gflags glog boost)
diff --git a/paddle/fluid/distributed/table/depends/dense.h b/paddle/fluid/distributed/table/depends/dense.h
@@ -221,15 +221,6 @@ class DAdamD2Sum : public DenseOptimizer {
   void update(const float* update_values, size_t num, int begin,
               int end) override {
     auto update_numel = end - begin;
-
-    /*
-    // for debug
-    std::cout << "before update:\n";
-    for (int i = 0; i < 3; ++ i) {
-      std::cout << "param: " << i << " " << *(param+begin+i) <<
-                   "grad: " << *(update_values+begin+i) << "\n";
-    }*/
-
     std::vector<float> grad, grad2, scale;
     grad.resize(update_numel);
     grad2.resize(update_numel);
@@ -240,88 +231,35 @@ class DAdamD2Sum : public DenseOptimizer {
     blas.VCOPY(update_numel, update_values + begin, grad.data());
     blas.VCOPY(update_numel, update_values + begin, grad2.data());
 
-    /*
-    for (int i = 0; i < end-begin; ++ i) {
-      std::cout << "copy grad: " << i << " " << *(grad.data()+begin+i) <<
-                   "copy grad2: " << *(grad2.data()+begin+i) << "\n";
-    }
-    for (int i = 0; i < 3; ++ i) {
-      std::cout << "d2sum before: " << i << " " << *(ada_d2sum+begin+i) << "\n";
-    }*/
-
     // d2sum
     blas.SCAL(update_numel, ada_decay_rate[0], ada_d2sum + begin);
     ADD<float>(update_numel, ada_d2sum + begin, 1, ada_d2sum + begin);
 
-    /*
-    for (int i = 0; i < end-begin; ++ i) {
-      std::cout << "d2sum update: " << i << " " << *(ada_d2sum+begin+i) << "\n";
-    }
-    for (int i = 0; i < 3; ++ i) {
-      std::cout << "g2sum before: " << i << " " << *(ada_g2sum+begin+i) << "\n";
-    }*/
-
     // g2sum
     blas.SCAL(update_numel, ada_decay_rate[0], ada_g2sum + begin);
     blas.VSQUARE(update_numel, grad2.data(), grad2.data());
     blas.VADD(update_numel, ada_g2sum + begin, grad2.data(), ada_g2sum + begin);
 
-    /*
-    for (int i = 0; i < end-begin; ++ i) {
-      std::cout << "g2sum update: " << i << " " << *(ada_g2sum+begin+i) << "\n";
-    }
-    for (int i = 0; i < 3; ++ i) {
-      std::cout << "mom before: " << i << " " << *(mom_velocity+begin+i) <<
-    "\n";
-    }*/
-
     // mom
     blas.SCAL(update_numel, mom_decay_rate[0], mom_velocity + begin);
     blas.SCAL(update_numel, 1 - mom_decay_rate[0], grad.data());
     blas.VADD(update_numel, mom_velocity + begin, grad.data(),
               mom_velocity + begin);
 
-    /*
-    for (int i = 0; i < end-begin; ++ i) {
-      std::cout << "mom update: " << i << " " << *(mom_velocity+begin+i) <<
-    "\n";
-    }
-    for (int i = 0; i < 3; ++ i) {
-      std::cout << "scale before: " << i << " " << *(scale.data()+begin+i) <<
-    "\n";
-    }*/
-
     // scale
     float* scale_ = scale.data();
     blas.VDIV(update_numel, ada_g2sum + begin, ada_d2sum + begin, scale_);
     ADD<float>(update_numel, scale_, ada_epsilon[0], scale_);
     DIV<float>(update_numel, 1 + ada_epsilon[0], scale_, scale_);
     SQRT<float>(update_numel, scale_, scale_);
 
-    /*
-    for (int i = 0; i < 3; ++ i) {
-      std::cout << "scale update: " << i << " " << *(scale.data()+begin+i) <<
-    "\n";
-    }*/
-
     blas.SCAL(update_numel, learning_rate[0], scale_);
 
     // TODO(zhaocaibei123): check if there exists elementwise_multiply in blas
     // TODO(zhaocaibei123): blas.VMUL
     ELE_MUL<float>(update_numel, scale_, mom_velocity + begin, scale_);
 
-    /*
-    for (int i = 0; i < 3; ++ i) {
-      std::cout << "scale update2: " << i << " " << *(scale.data()+begin+i) <<
-    "\n";
-    }*/
-
     blas.VSUB(update_numel, param + begin, scale_, param + begin);
-
-    /*
-    for (int i = 0; i < end-begin; ++ i) {
-      std::cout << "param update " << i << " " << *(param+begin+i) << "\n";
-    }*/
   }
 
   float* learning_rate;