fleet_wrapper.cc 22.4 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14
// Copyright (c) 2019 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.

15 16 17 18 19 20 21 22 23 24 25 26 27 28 29
/* Copyright (c) 2018 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/framework/fleet/fleet_wrapper.h"
30
#include <algorithm>
X
xujiaqi01 已提交
31
#include <utility>
32
#include "paddle/fluid/framework/data_feed.h"
33
#include "paddle/fluid/framework/op_registry.h"
34
#include "paddle/fluid/framework/scope.h"
35 36 37 38 39 40

namespace paddle {
namespace framework {

const uint32_t MAX_FEASIGN_NUM = 1024 * 100 * 100;
std::shared_ptr<FleetWrapper> FleetWrapper::s_instance_ = NULL;
41 42
bool FleetWrapper::is_initialized_ = false;

43
#ifdef PADDLE_WITH_PSLIB
D
dongdaxiang 已提交
44 45 46
template <class AR>
paddle::ps::Archive<AR>& operator<<(paddle::ps::Archive<AR>& ar,
                                    const MultiSlotType& ins) {
47 48 49 50
  ar << ins.GetType();
  ar << ins.GetOffset();
  ar << ins.GetFloatData();
  ar << ins.GetUint64Data();
X
xujiaqi01 已提交
51
  return ar;
52 53
}

D
dongdaxiang 已提交
54 55 56
template <class AR>
paddle::ps::Archive<AR>& operator>>(paddle::ps::Archive<AR>& ar,
                                    MultiSlotType& ins) {
57 58 59 60
  ar >> ins.MutableType();
  ar >> ins.MutableOffset();
  ar >> ins.MutableFloatData();
  ar >> ins.MutableUint64Data();
X
xujiaqi01 已提交
61
  return ar;
62 63 64
}
#endif

65 66 67
#ifdef PADDLE_WITH_PSLIB
std::shared_ptr<paddle::distributed::PSlib> FleetWrapper::pslib_ptr_ = NULL;
#endif
68 69 70 71

void FleetWrapper::InitServer(const std::string& dist_desc, int index) {
#ifdef PADDLE_WITH_PSLIB
  if (!is_initialized_) {
D
dongdaxiang 已提交
72
    VLOG(3) << "Going to init server";
73 74 75 76 77
    pslib_ptr_ = std::shared_ptr<paddle::distributed::PSlib>(
        new paddle::distributed::PSlib());
    pslib_ptr_->init_server(dist_desc, index);
    is_initialized_ = true;
  } else {
D
dongdaxiang 已提交
78
    VLOG(3) << "Server can be initialized only once";
79 80 81 82 83 84 85 86 87
  }
#endif
}

void FleetWrapper::InitWorker(const std::string& dist_desc,
                              const std::vector<uint64_t>& host_sign_list,
                              int node_num, int index) {
#ifdef PADDLE_WITH_PSLIB
  if (!is_initialized_) {
D
dongdaxiang 已提交
88
    VLOG(3) << "Going to init worker";
89 90 91 92 93 94 95
    pslib_ptr_ = std::shared_ptr<paddle::distributed::PSlib>(
        new paddle::distributed::PSlib());
    pslib_ptr_->init_worker(dist_desc,
                            const_cast<uint64_t*>(host_sign_list.data()),
                            node_num, index);
    is_initialized_ = true;
  } else {
D
dongdaxiang 已提交
96
    VLOG(3) << "Worker can be initialized only once";
97 98 99 100 101 102
  }
#endif
}

void FleetWrapper::StopServer() {
#ifdef PADDLE_WITH_PSLIB
D
dongdaxiang 已提交
103
  VLOG(3) << "Going to stop server";
104 105 106 107 108 109
  pslib_ptr_->stop_server();
#endif
}

uint64_t FleetWrapper::RunServer() {
#ifdef PADDLE_WITH_PSLIB
D
dongdaxiang 已提交
110
  VLOG(3) << "Going to run server";
111 112 113 114 115 116 117 118 119
  return pslib_ptr_->run_server();
#else
  return 0;
#endif
}

void FleetWrapper::GatherServers(const std::vector<uint64_t>& host_sign_list,
                                 int node_num) {
#ifdef PADDLE_WITH_PSLIB
D
dongdaxiang 已提交
120
  VLOG(3) << "Going to gather server ips";
121 122 123 124 125
  pslib_ptr_->gather_servers(const_cast<uint64_t*>(host_sign_list.data()),
                             node_num);
#endif
}

D
dongdaxiang 已提交
126
void FleetWrapper::GatherClients(const std::vector<uint64_t>& host_sign_list) {
X
xjqbest 已提交
127 128 129
#ifdef PADDLE_WITH_PSLIB
  VLOG(3) << "Going to gather client ips";
  size_t len = host_sign_list.size();
D
dongdaxiang 已提交
130
  pslib_ptr_->gather_clients(const_cast<uint64_t*>(host_sign_list.data()), len);
X
xjqbest 已提交
131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148
#endif
}

std::vector<uint64_t> FleetWrapper::GetClientsInfo() {
#ifdef PADDLE_WITH_PSLIB
  VLOG(3) << "Going to get client info";
  return pslib_ptr_->get_client_info();
#endif
  return std::vector<uint64_t>();
}

void FleetWrapper::CreateClient2ClientConnection() {
#ifdef PADDLE_WITH_PSLIB
  VLOG(3) << "Going to create client2client connection";
  pslib_ptr_->create_client2client_connection();
#endif
}

149 150 151 152 153 154 155 156 157 158 159 160
void FleetWrapper::PullSparseVarsSync(
    const Scope& scope, const uint64_t table_id,
    const std::vector<std::string>& var_names, std::vector<uint64_t>* fea_keys,
    std::vector<std::vector<float>>* fea_values, int fea_value_dim) {
#ifdef PADDLE_WITH_PSLIB
  std::vector<::std::future<int32_t>> pull_sparse_status;
  pull_sparse_status.resize(0);
  fea_keys->clear();
  fea_keys->resize(0);
  fea_keys->reserve(MAX_FEASIGN_NUM);
  for (auto name : var_names) {
    Variable* var = scope.FindVar(name);
161 162 163
    if (var == nullptr) {
      continue;
    }
164
    LoDTensor* tensor = var->GetMutable<LoDTensor>();
165
    CHECK(tensor != nullptr) << "tensor of var " << name << " is null";
166 167 168 169 170 171 172 173 174
    int64_t* ids = tensor->data<int64_t>();
    int len = tensor->numel();
    for (auto i = 0u; i < len; ++i) {
      if (ids[i] == 0u) {
        continue;
      }
      fea_keys->push_back(static_cast<uint64_t>(ids[i]));
    }
  }
D
dongdaxiang 已提交
175 176 177 178 179 180 181 182 183 184 185
  fea_values->resize(fea_keys->size() + 1);
  for (auto& t : *fea_values) {
    t.resize(fea_value_dim);
  }
  std::vector<float*> pull_result_ptr;
  for (auto& t : *fea_values) {
    pull_result_ptr.push_back(t.data());
  }
  auto status = pslib_ptr_->_worker_ptr->pull_sparse(
      pull_result_ptr.data(), table_id, fea_keys->data(), fea_keys->size());
  pull_sparse_status.push_back(std::move(status));
186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201
  for (auto& t : pull_sparse_status) {
    t.wait();
    auto status = t.get();
    if (status != 0) {
      LOG(ERROR) << "fleet pull sparse failed, status[" << status << "]";
      exit(-1);
    }
  }
#endif
}

void FleetWrapper::PullDenseVarsAsync(
    const Scope& scope, const uint64_t tid,
    const std::vector<std::string>& var_names,
    std::vector<::std::future<int32_t>>* pull_dense_status) {
#ifdef PADDLE_WITH_PSLIB
X
xujiaqi01 已提交
202 203
  auto& regions = _regions[tid];
  regions.clear();
204 205 206
  regions.resize(var_names.size());
  for (auto i = 0u; i < var_names.size(); ++i) {
    Variable* var = scope.FindVar(var_names[i]);
207 208 209
    LoDTensor* tensor = var->GetMutable<LoDTensor>();
    float* w = tensor->data<float>();
    paddle::ps::Region reg(w, tensor->numel());
210
    regions[i] = std::move(reg);
211 212 213 214 215 216 217 218 219 220 221
  }
  auto status =
      pslib_ptr_->_worker_ptr->pull_dense(regions.data(), regions.size(), tid);
  pull_dense_status->push_back(std::move(status));
#endif
}

void FleetWrapper::PullDenseVarsSync(
    const Scope& scope, const uint64_t tid,
    const std::vector<std::string>& var_names) {
#ifdef PADDLE_WITH_PSLIB
X
xujiaqi01 已提交
222 223
  auto& regions = _regions[tid];
  regions.clear();
224 225 226 227 228 229 230 231 232 233 234 235 236 237
  regions.reserve(var_names.size());
  for (auto& t : var_names) {
    Variable* var = scope.FindVar(t);
    LoDTensor* tensor = var->GetMutable<LoDTensor>();
    float* w = tensor->data<float>();
    paddle::ps::Region reg(w, tensor->numel());
    regions.emplace_back(std::move(reg));
  }
  auto status =
      pslib_ptr_->_worker_ptr->pull_dense(regions.data(), regions.size(), tid);
  status.wait();
#endif
}

238
void FleetWrapper::PushDenseParamSync(
D
dongdaxiang 已提交
239
    const Scope& scope, const uint64_t table_id,
240 241 242 243 244 245
    const std::vector<std::string>& var_names) {
#ifdef PADDLE_WITH_PSLIB
  auto place = platform::CPUPlace();
  std::vector<paddle::ps::Region> regions;
  for (auto& t : var_names) {
    Variable* var = scope.FindVar(t);
X
xjqbest 已提交
246
    CHECK(var != nullptr) << "var[" << t << "] not found";
247
    LoDTensor* tensor = var->GetMutable<LoDTensor>();
248
    float* g = tensor->mutable_data<float>(place);
249 250 251
    paddle::ps::Region reg(g, tensor->numel());
    regions.emplace_back(std::move(reg));
  }
252 253 254 255 256
  auto push_status = pslib_ptr_->_worker_ptr->push_dense_param(
      regions.data(), regions.size(), table_id);
  push_status.wait();
  auto status = push_status.get();
  CHECK(status == 0) << "push dense param failed, status[" << status << "]";
257 258 259
#endif
}

D
dongdaxiang 已提交
260 261 262 263
void FleetWrapper::PushDenseVarsSync(
    Scope* scope, const uint64_t table_id,
    const std::vector<std::string>& var_names) {}

264 265 266
void FleetWrapper::PushDenseVarsAsync(
    const Scope& scope, const uint64_t table_id,
    const std::vector<std::string>& var_names,
267 268
    std::vector<::std::future<int32_t>>* push_sparse_status,
    float scale_datanorm, int batch_size) {
269 270 271 272 273 274 275
#ifdef PADDLE_WITH_PSLIB
  std::vector<paddle::ps::Region> regions;
  for (auto& t : var_names) {
    Variable* var = scope.FindVar(t);
    LoDTensor* tensor = var->GetMutable<LoDTensor>();
    int count = tensor->numel();
    float* g = tensor->data<float>();
276 277 278 279 280 281 282 283 284 285 286 287 288 289
    if (scale_datanorm >= 0) {
      if (t.find(".batch_size@GRAD") != std::string::npos ||
          t.find(".batch_sum@GRAD") != std::string::npos) {
        Eigen::Map<Eigen::MatrixXf> mat(g, 1, count);
        float scale = 1.0 / batch_size;
        mat *= scale;
      } else if (t.find(".batch_square_sum@GRAD") != std::string::npos) {
        VLOG(3) << "epsilon: " << scale_datanorm;
        for (int i = 0; i < count; ++i) {
          g[i] = (g[i] - batch_size * scale_datanorm) / batch_size +
                 batch_size * scale_datanorm;
        }
      }
    }
290 291 292 293 294 295 296 297 298 299 300 301 302 303 304
    paddle::ps::Region reg(g, count);
    regions.emplace_back(std::move(reg));
  }
  auto status = pslib_ptr_->_worker_ptr->push_dense(regions.data(),
                                                    regions.size(), table_id);
  push_sparse_status->push_back(std::move(status));
#endif
}

void FleetWrapper::PushSparseVarsWithLabelAsync(
    const Scope& scope, const uint64_t table_id,
    const std::vector<uint64_t>& fea_keys, const std::vector<float>& fea_labels,
    const std::vector<std::string>& sparse_key_names,
    const std::vector<std::string>& sparse_grad_names, const int emb_dim,
    std::vector<std::vector<float>>* push_values,
305
    std::vector<::std::future<int32_t>>* push_sparse_status,
T
Thunderbrook 已提交
306
    const int batch_size, const bool use_cvm, const bool dump_slot) {
307 308
#ifdef PADDLE_WITH_PSLIB
  int offset = 2;
T
Thunderbrook 已提交
309
  int slot_offset = 0;
310
  int grad_dim = emb_dim;
T
Thunderbrook 已提交
311 312
  int show_index = 0;
  int click_index = 1;
313 314 315 316
  if (use_cvm) {
    offset = 0;
    grad_dim = emb_dim - 2;
  }
T
Thunderbrook 已提交
317 318 319 320 321
  if (dump_slot) {
    slot_offset = 1;
    show_index = 1;
    click_index = 2;
  }
322
  CHECK_GE(grad_dim, 0);
323 324 325

  push_values->resize(fea_keys.size() + 1);
  for (auto& t : *push_values) {
T
Thunderbrook 已提交
326
    t.resize(emb_dim + offset + slot_offset);
327
  }
328 329 330
  uint64_t fea_idx = 0u;
  for (size_t i = 0; i < sparse_key_names.size(); ++i) {
    Variable* var = scope.FindVar(sparse_key_names[i]);
331 332 333
    if (var == nullptr) {
      continue;
    }
334
    LoDTensor* tensor = var->GetMutable<LoDTensor>();
335 336
    if (tensor == nullptr) {
      LOG(ERROR) << "tensor of var[" << sparse_key_names[i] << "] is null";
337 338 339 340
      exit(-1);
    }
    int len = tensor->numel();
    int64_t* ids = tensor->data<int64_t>();
T
Thunderbrook 已提交
341 342 343 344
    int slot = 0;
    if (dump_slot) {
      slot = boost::lexical_cast<int>(sparse_key_names[i]);
    }
345 346 347 348 349 350
    Variable* g_var = scope.FindVar(sparse_grad_names[i]);
    CHECK(g_var != nullptr) << "var[" << sparse_grad_names[i] << "] not found";
    LoDTensor* g_tensor = g_var->GetMutable<LoDTensor>();
    if (g_tensor == nullptr) {
      LOG(ERROR) << "tensor of var[" << sparse_key_names[i] << "] is null";
      exit(-1);
351
    }
352 353
    float* g = g_tensor->data<float>();

354 355 356 357 358 359 360
    if (scale_sparse_gradient_with_batch_size_ && grad_dim > 0) {
      int dim = emb_dim + offset;
      Eigen::Map<
          Eigen::Matrix<float, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor>>
          g_mat(g, g_tensor->numel() / dim, dim);
      g_mat.rightCols(grad_dim) *= batch_size;
    }
361 362 363 364 365
    for (auto id_idx = 0u; id_idx < len; ++id_idx) {
      if (ids[id_idx] == 0) {
        g += emb_dim;
        continue;
      }
366 367
      CHECK(fea_idx < (*push_values).size());
      CHECK(fea_idx < fea_labels.size());
T
Thunderbrook 已提交
368

369
      if (use_cvm) {
T
Thunderbrook 已提交
370
        memcpy((*push_values)[fea_idx].data() + offset + slot_offset, g,
371 372
               sizeof(float) * emb_dim);
      } else {
T
Thunderbrook 已提交
373
        memcpy((*push_values)[fea_idx].data() + offset + slot_offset, g,
374
               sizeof(float) * emb_dim);
T
Thunderbrook 已提交
375 376 377 378 379 380
        (*push_values)[fea_idx][show_index] = 1.0f;
        (*push_values)[fea_idx][click_index] =
            static_cast<float>(fea_labels[fea_idx]);
      }
      if (dump_slot) {
        (*push_values)[fea_idx][0] = static_cast<float>(slot);
381
      }
382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399
      g += emb_dim;
      fea_idx++;
    }
  }
  CHECK(fea_idx == fea_keys.size()) << "fea_idx: " << fea_idx
                                    << "features size: " << fea_keys.size();
  std::vector<float*> push_g_vec;
  for (auto i = 0u; i < fea_keys.size(); ++i) {
    push_g_vec.push_back((*push_values)[i].data());
  }
  auto status = pslib_ptr_->_worker_ptr->push_sparse(
      table_id, fea_keys.data(), (const float**)push_g_vec.data(),
      fea_keys.size());
  push_sparse_status->push_back(std::move(status));

#endif
}

400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472
void FleetWrapper::LoadFromPaddleModel(Scope& scope, const uint64_t table_id,
                                       std::vector<std::string> var_list,
                                       std::string model_path,
                                       std::string model_proto_file,
                                       bool load_combine) {
  // load ProgramDesc from model file
  auto read_proto_func = [](const std::string& filename) -> ProgramDesc {
    std::string contents;
    std::ifstream fin(filename, std::ios::in | std::ios::binary);
    fin.seekg(0, std::ios::end);
    contents.resize(fin.tellg());
    fin.seekg(0, std::ios::beg);
    fin.read(&contents[0], contents.size());
    fin.close();
    ProgramDesc program_desc(contents);
    return program_desc;
  };
  const ProgramDesc old_program = read_proto_func(model_proto_file);
  Scope* old_scope = new Scope();
  auto& old_block = old_program.Block(0);
  auto place = platform::CPUPlace();
  std::vector<std::string> old_param_list;

  for (auto& t : var_list) {
    VarDesc* old_var_desc = old_block.FindVar(t);
    if (old_var_desc == nullptr) {
      continue;
    }
    // init variable in scope
    Variable* old_var = old_scope->Var(old_var_desc->Name());
    InitializeVariable(old_var, old_var_desc->GetType());
    old_param_list.push_back(t);
    if (load_combine) {
      continue;
    }
    // load variable from model
    paddle::framework::AttributeMap attrs;
    attrs.insert({"file_path", model_path + "/" + old_var_desc->Name()});
    auto load_op = paddle::framework::OpRegistry::CreateOp(
        "load", {}, {{"Out", {old_var_desc->Name()}}}, attrs);
    load_op->Run(*old_scope, place);
  }

  if (load_combine) {
    std::sort(old_param_list.begin(), old_param_list.end());
    paddle::framework::AttributeMap attrs;
    attrs.insert({"file_path", model_path});
    auto load_op = paddle::framework::OpRegistry::CreateOp(
        "load_combine", {}, {{"Out", old_param_list}}, attrs);
    load_op->Run(*old_scope, place);
  }

  for (auto& t : old_param_list) {
    Variable* old_var = old_scope->Var(t);
    // old model data, here we assume data type is float
    LoDTensor* old_tensor = old_var->GetMutable<LoDTensor>();
    float* old_data = old_tensor->data<float>();
    // new model data, here we assume data type is float
    Variable* var = scope.FindVar(t);
    CHECK(var != nullptr) << "var[" << t << "] not found";
    LoDTensor* tensor = var->GetMutable<LoDTensor>();
    float* data = tensor->data<float>();
    // copy from old data to new data
    if (old_tensor->numel() > tensor->numel()) {
      memcpy(data, old_data, tensor->numel() * sizeof(float));
    } else {
      memcpy(data, old_data, old_tensor->numel() * sizeof(float));
    }
  }
  delete old_scope;
  PushDenseParamSync(scope, table_id, old_param_list);
}

473 474 475 476 477 478 479 480 481 482 483 484 485
void FleetWrapper::LoadModel(const std::string& path, const int mode) {
#ifdef PADDLE_WITH_PSLIB
  auto ret = pslib_ptr_->_worker_ptr->load(path, std::to_string(mode));
  ret.wait();
  if (ret.get() != 0) {
    LOG(ERROR) << "load model from path:" << path << " failed";
    exit(-1);
  }
#else
  VLOG(0) << "FleetWrapper::LoadModel does nothing when no pslib";
#endif
}

486 487 488 489 490 491 492 493 494 495 496 497 498 499 500
void FleetWrapper::LoadModelOneTable(const uint64_t table_id,
                                     const std::string& path, const int mode) {
#ifdef PADDLE_WITH_PSLIB
  auto ret =
      pslib_ptr_->_worker_ptr->load(table_id, path, std::to_string(mode));
  ret.wait();
  if (ret.get() != 0) {
    LOG(ERROR) << "load model of table id: " << table_id
               << ", from path: " << path << " failed";
  }
#else
  VLOG(0) << "FleetWrapper::LoadModel does nothing when no pslib";
#endif
}

501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523
void FleetWrapper::SaveModel(const std::string& path, const int mode) {
#ifdef PADDLE_WITH_PSLIB
  auto ret = pslib_ptr_->_worker_ptr->save(path, std::to_string(mode));
  ret.wait();
  int32_t feasign_cnt = ret.get();
  if (feasign_cnt == -1) {
    LOG(ERROR) << "save model failed";
    exit(-1);
  }
#else
  VLOG(0) << "FleetWrapper::SaveModel does nothing when no pslib";
#endif
}

void FleetWrapper::ShrinkSparseTable(int table_id) {
#ifdef PADDLE_WITH_PSLIB
  auto ret = pslib_ptr_->_worker_ptr->shrink(table_id);
  ret.wait();
#else
  VLOG(0) << "FleetWrapper::ShrinkSparseTable does nothing when no pslib";
#endif
}

524 525 526 527 528 529 530 531 532
void FleetWrapper::ClearModel() {
#ifdef PADDLE_WITH_PSLIB
  auto ret = pslib_ptr_->_worker_ptr->clear();
  ret.wait();
#else
  VLOG(0) << "FleetWrapper::ClearModel does nothing when no pslib";
#endif
}

533 534
void FleetWrapper::ShrinkDenseTable(int table_id, Scope* scope,
                                    std::vector<std::string> var_list,
535
                                    float decay, int emb_dim) {
536 537 538 539 540 541
#ifdef PADDLE_WITH_PSLIB
  std::vector<paddle::ps::Region> regions;
  for (std::string& name : var_list) {
    if (name.find("batch_sum") != std::string::npos) {
      Variable* var = scope->FindVar(name);
      CHECK(var != nullptr) << "var[" << name << "] not found";
542
      VLOG(0) << "prepare shrink dense batch_sum";
543 544
      LoDTensor* tensor = var->GetMutable<LoDTensor>();
      float* g = tensor->data<float>();
545 546 547 548 549 550 551 552 553 554 555 556 557

      // show_batch_sum += N * log(decay)
      std::string size_name = name;
      size_name.replace(size_name.find("batch_sum"), size_name.length(),
                        "batch_size");
      Variable* var_size = scope->FindVar(size_name);
      CHECK(var_size != nullptr) << "var[" << size_name << "] not found";
      VLOG(3) << "shrink dense batch_sum: " << name << ", " << size_name;
      float* g_size = var_size->GetMutable<LoDTensor>()->data<float>();

      for (int k = 0; k < tensor->numel(); k += emb_dim) {
        g[k] = g[k] + g_size[k] * log(decay);
      }
558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590
      paddle::ps::Region reg(g, tensor->numel());
      regions.emplace_back(std::move(reg));
    } else {
      Variable* var = scope->FindVar(name);
      CHECK(var != nullptr) << "var[" << name << "] not found";
      LoDTensor* tensor = var->GetMutable<LoDTensor>();
      float* g = tensor->data<float>();
      paddle::ps::Region reg(g, tensor->numel());
      regions.emplace_back(std::move(reg));
    }
  }
  auto push_status = pslib_ptr_->_worker_ptr->push_dense_param(
      regions.data(), regions.size(), table_id);
  push_status.wait();
  auto status = push_status.get();
  if (status != 0) {
    LOG(FATAL) << "push shrink dense param failed, status[" << status << "]";
    exit(-1);
  }
#else
  VLOG(0) << "FleetWrapper::ShrinkSparseTable does nothing when no pslib";
#endif
}

void FleetWrapper::ClientFlush() {
#ifdef PADDLE_WITH_PSLIB
  auto ret = pslib_ptr_->_worker_ptr->flush();
  ret.wait();
#else
  VLOG(0) << "FleetWrapper::ServerFlush does nothing when no pslib";
#endif
}

591 592
int FleetWrapper::RegisterClientToClientMsgHandler(int msg_type,
                                                   MsgHandlerFunc handler) {
593
#ifdef PADDLE_WITH_PSLIB
X
xujiaqi01 已提交
594 595 596
  VLOG(3) << "calling FleetWrapper::RegisterClientToClientMsgHandler";
  VLOG(3) << "pslib_ptr_=" << pslib_ptr_;
  VLOG(3) << "_worker_ptr=" << pslib_ptr_->_worker_ptr;
597 598
  return pslib_ptr_->_worker_ptr->registe_client2client_msg_handler(msg_type,
                                                                    handler);
599 600 601 602
#else
  VLOG(0) << "FleetWrapper::RegisterClientToClientMsgHandler"
          << " does nothing when no pslib";
#endif
X
xujiaqi01 已提交
603
  return 0;
604 605
}

606 607
std::future<int32_t> FleetWrapper::SendClientToClientMsg(
    int msg_type, int to_client_id, const std::string& msg) {
608
#ifdef PADDLE_WITH_PSLIB
609 610
  return pslib_ptr_->_worker_ptr->send_client2client_msg(msg_type, to_client_id,
                                                         msg);
611 612 613 614
#else
  VLOG(0) << "FleetWrapper::SendClientToClientMsg"
          << " does nothing when no pslib";
#endif
615
  return std::future<int32_t>();
X
xujiaqi01 已提交
616 617
}

D
dongdaxiang 已提交
618
template <typename T>
619
void FleetWrapper::Serialize(const std::vector<T*>& t, std::string* str) {
620 621
#ifdef PADDLE_WITH_PSLIB
  paddle::ps::BinaryArchive ar;
622 623 624
  for (size_t i = 0; i < t.size(); ++i) {
    ar << *(t[i]);
  }
X
xujiaqi01 已提交
625
  *str = std::string(ar.buffer(), ar.length());
626
#else
627
  VLOG(0) << "FleetWrapper::Serialize does nothing when no pslib";
628 629 630
#endif
}

D
dongdaxiang 已提交
631
template <typename T>
632
void FleetWrapper::Deserialize(std::vector<T>* t, const std::string& str) {
633
#ifdef PADDLE_WITH_PSLIB
634 635 636
  if (str.length() == 0) {
    return;
  }
637 638
  paddle::ps::BinaryArchive ar;
  ar.set_read_buffer(const_cast<char*>(str.c_str()), str.length(), nullptr);
639 640 641 642 643 644 645 646
  if (ar.cursor() == ar.finish()) {
    return;
  }
  while (ar.cursor() < ar.finish()) {
    t->push_back(ar.get<T>());
  }
  CHECK(ar.cursor() == ar.finish());
  VLOG(3) << "Deserialize size " << t->size();
647
#else
648
  VLOG(0) << "FleetWrapper::Deserialize does nothing when no pslib";
649 650 651
#endif
}

652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669
std::default_random_engine& FleetWrapper::LocalRandomEngine() {
  struct engine_wrapper_t {
    std::default_random_engine engine;
#ifdef PADDLE_WITH_PSLIB
    engine_wrapper_t() {
      struct timespec tp;
      clock_gettime(CLOCK_REALTIME, &tp);
      double cur_time = tp.tv_sec + tp.tv_nsec * 1e-9;
      static std::atomic<uint64_t> x(0);
      std::seed_seq sseq = {x++, x++, x++, (uint64_t)(cur_time * 1000)};
      engine.seed(sseq);
    }
#endif
  };
  thread_local engine_wrapper_t r;
  return r.engine;
}

670
template void FleetWrapper::Serialize<std::vector<MultiSlotType>>(
671 672 673
    const std::vector<std::vector<MultiSlotType>*>&, std::string*);
template void FleetWrapper::Deserialize<std::vector<MultiSlotType>>(
    std::vector<std::vector<MultiSlotType>>*, const std::string&);
674

675 676
}  // end namespace framework
}  // end namespace paddle