pybind.cc 115.4 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
2 3 4 5 6

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

7
http://www.apache.org/licenses/LICENSE-2.0
8 9 10 11 12 13

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. */
L
lgone2000 已提交
14
#include <Python.h>
15

C
chengduoZH 已提交
16
#include <algorithm>
17
#include <cstdlib>
C
chengduoZH 已提交
18
#include <map>
S
sneaxiy 已提交
19
#include <memory>
C
chengduoZH 已提交
20 21 22
#include <mutex>  // NOLINT // for call_once
#include <string>
#include <unordered_map>
23
#include <unordered_set>
C
chengduoZH 已提交
24 25
#include <utility>
#include <vector>
26

27
#include "paddle/fluid/framework/custom_operator.h"
28
#include "paddle/fluid/framework/data_layout.h"
Y
Yi Wang 已提交
29 30
#include "paddle/fluid/framework/executor.h"
#include "paddle/fluid/framework/feed_fetch_method.h"
Z
Zhen Wang 已提交
31
#include "paddle/fluid/framework/feed_fetch_type.h"
Y
Yi Wang 已提交
32
#include "paddle/fluid/framework/framework.pb.h"
S
sneaxiy 已提交
33
#include "paddle/fluid/framework/garbage_collector.h"
H
hutuxian 已提交
34
#include "paddle/fluid/framework/io/fs.h"
35
#include "paddle/fluid/framework/ir/coalesce_grad_tensor_pass.h"
36
#include "paddle/fluid/framework/ir/pass_builder.h"
37
#include "paddle/fluid/framework/load_op_lib.h"
Y
Yi Wang 已提交
38 39 40
#include "paddle/fluid/framework/lod_rank_table.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/lod_tensor_array.h"
S
sneaxiy 已提交
41
#include "paddle/fluid/framework/op_info.h"
42
#include "paddle/fluid/framework/op_registry.h"
43
#include "paddle/fluid/framework/op_version_registry.h"
Y
Yu Yang 已提交
44
#include "paddle/fluid/framework/parallel_executor.h"
Y
Yi Wang 已提交
45
#include "paddle/fluid/framework/prune.h"
Y
Refine  
Yu Yang 已提交
46
#include "paddle/fluid/framework/reader.h"
H
hong 已提交
47
#include "paddle/fluid/framework/save_load_util.h"
S
sneaxiy 已提交
48
#include "paddle/fluid/framework/scope_pool.h"
Y
Yi Wang 已提交
49
#include "paddle/fluid/framework/selected_rows.h"
50
#include "paddle/fluid/framework/tensor_util.h"
51
#include "paddle/fluid/framework/trainer.h"
52
#include "paddle/fluid/framework/type_defs.h"
X
Xin Pan 已提交
53
#include "paddle/fluid/framework/version.h"
H
hong 已提交
54
#include "paddle/fluid/imperative/layer.h"
Y
Refine  
Yu Yang 已提交
55
#include "paddle/fluid/memory/allocation/allocator_strategy.h"
56
#include "paddle/fluid/memory/allocation/mmap_allocator.h"
D
dzhwinter 已提交
57
#include "paddle/fluid/operators/activation_op.h"
L
Leo Chen 已提交
58
#include "paddle/fluid/operators/common_infer_shape_functions.h"
S
sneaxiy 已提交
59
#include "paddle/fluid/operators/py_func_op.h"
60
#include "paddle/fluid/platform/cpu_helper.h"
Y
Yu Yang 已提交
61
#include "paddle/fluid/platform/cpu_info.h"
62
#include "paddle/fluid/platform/device_context.h"
63
#include "paddle/fluid/platform/dynload/dynamic_loader.h"
Y
Yi Wang 已提交
64
#include "paddle/fluid/platform/enforce.h"
65
#include "paddle/fluid/platform/init.h"
H
hutuxian 已提交
66
#include "paddle/fluid/platform/monitor.h"
Y
Yi Wang 已提交
67 68
#include "paddle/fluid/platform/place.h"
#include "paddle/fluid/platform/profiler.h"
69 70 71
#ifdef PADDLE_WITH_ASCEND
#include "paddle/fluid/pybind/ascend_wrapper_py.h"
#endif
H
hutuxian 已提交
72
#include "paddle/fluid/pybind/box_helper_py.h"
73
#include "paddle/fluid/pybind/compatible.h"
Y
Yi Wang 已提交
74
#include "paddle/fluid/pybind/const_value.h"
D
dongdaxiang 已提交
75
#include "paddle/fluid/pybind/data_set_py.h"
Y
Yi Wang 已提交
76
#include "paddle/fluid/pybind/exception.h"
D
dongdaxiang 已提交
77
#include "paddle/fluid/pybind/fleet_wrapper_py.h"
Y
yaoxuefeng 已提交
78
#include "paddle/fluid/pybind/generator_py.h"
79
#include "paddle/fluid/pybind/global_value_getter_setter.h"
80
#include "paddle/fluid/pybind/gloo_context_py.h"
81
#include "paddle/fluid/pybind/gloo_wrapper_py.h"
T
Thunderbrook 已提交
82
#include "paddle/fluid/pybind/heter_wrapper_py.h"
83
#include "paddle/fluid/pybind/imperative.h"
F
flame 已提交
84
#include "paddle/fluid/pybind/inference_api.h"
F
flame 已提交
85
#include "paddle/fluid/pybind/ir.h"
T
Thunderbrook 已提交
86
#include "paddle/fluid/pybind/ps_gpu_wrapper_py.h"
87
#include "paddle/fluid/pybind/pybind_boost_headers.h"
88

89
#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
D
dongdaxiang 已提交
90
#include "paddle/fluid/pybind/nccl_wrapper_py.h"
W
wopeizl 已提交
91
#endif
92
#include "paddle/fluid/framework/data_type.h"
93 94
#include "paddle/fluid/pybind/protobuf.h"
#include "paddle/fluid/pybind/pybind.h"  // NOLINT
S
sneaxiy 已提交
95
#include "paddle/fluid/pybind/reader_py.h"
Y
Yi Wang 已提交
96
#include "paddle/fluid/pybind/tensor_py.h"
97
#include "paddle/fluid/string/to_string.h"
98 99
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
Y
Yi Wang 已提交
100
#include "paddle/fluid/operators/nccl/nccl_gpu_common.h"
P
peizhilin 已提交
101
#endif
102
#ifndef PADDLE_WITH_HIP
Y
Yi Wang 已提交
103
#include "paddle/fluid/platform/cuda_profiler.h"
104
#endif
Y
Yi Wang 已提交
105
#include "paddle/fluid/platform/gpu_info.h"
D
Dong Zhihong 已提交
106 107
#endif

108 109 110 111
#ifdef PADDLE_WITH_XPU
#include "paddle/fluid/platform/xpu_info.h"
#endif

Y
Yanghello 已提交
112 113 114 115
#ifdef PADDLE_WITH_CRYPTO
#include "paddle/fluid/pybind/crypto.h"
#endif

T
tangwei12 已提交
116
#if defined PADDLE_WITH_PSCORE
T
tangwei12 已提交
117 118 119
#include "paddle/fluid/pybind/fleet_py.h"
#endif

M
minqiyang 已提交
120 121
#include "pybind11/stl.h"

122
DECLARE_bool(use_mkldnn);
123

Q
Qiao Longfei 已提交
124 125
// disable auto conversion to list in Python
PYBIND11_MAKE_OPAQUE(paddle::framework::LoDTensorArray);
126 127 128
PYBIND11_MAKE_OPAQUE(paddle::framework::FetchUnmergedList);
PYBIND11_MAKE_OPAQUE(paddle::framework::FetchList);
PYBIND11_MAKE_OPAQUE(paddle::framework::FetchType);
Q
Qiao Longfei 已提交
129

130
namespace paddle {
131
namespace pybind {
132
bool IsCompiledWithCUDA() {
133 134 135 136 137 138 139 140 141
#if !defined(PADDLE_WITH_CUDA) && !defined(PADDLE_WITH_HIP)
  return false;
#else
  return true;
#endif
}

bool IsCompiledWithROCM() {
#ifndef PADDLE_WITH_HIP
Q
qijun 已提交
142 143 144 145 146 147
  return false;
#else
  return true;
#endif
}

148 149 150 151 152 153 154 155
bool IsCompiledWithXPU() {
#ifndef PADDLE_WITH_XPU
  return false;
#else
  return true;
#endif
}

156 157 158 159 160 161 162 163
bool IsCompiledWithMKLDNN() {
#ifndef PADDLE_WITH_MKLDNN
  return false;
#else
  return true;
#endif
}

164 165 166 167 168 169 170 171 172 173 174
bool SupportsBfloat16() {
#ifndef PADDLE_WITH_MKLDNN
  return false;
#else
  if (platform::MayIUse(platform::cpu_isa_t::avx512_core))
    return true;
  else
    return false;
#endif
}

175 176 177 178 179 180 181 182 183 184 185
bool SupportsBfloat16FastPerformance() {
#ifndef PADDLE_WITH_MKLDNN
  return false;
#else
  if (platform::MayIUse(platform::cpu_isa_t::avx512_bf16))
    return true;
  else
    return false;
#endif
}

186
bool IsCompiledWithBrpc() {
187
#ifndef PADDLE_WITH_DISTRIBUTE
188 189
  return false;
#endif
190 191 192 193 194 195

#ifdef PADDLE_WITH_GRPC
  return false;
#endif

  return true;
196 197
}

Y
update  
Yancey1989 已提交
198
bool IsCompiledWithDIST() {
Y
Yancey1989 已提交
199
#ifdef PADDLE_WITH_DISTRIBUTE
Y
update  
Yancey1989 已提交
200 201 202 203 204 205
  return true;
#else
  return false;
#endif
}

S
sneaxiy 已提交
206 207 208 209 210 211 212 213 214 215
template <typename PlaceType1, typename PlaceType2>
static inline bool IsSamePlace(const PlaceType1 &p1, const PlaceType2 &p2) {
  return paddle::platform::Place(p1) == paddle::platform::Place(p2);
}

template <typename PlaceType>
static inline int PlaceIndex(const PlaceType &p) {
  return static_cast<int>(paddle::platform::Place(p).which());
}

H
hong 已提交
216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237
static PyObject *GetPythonAttribute(PyObject *obj, const char *attr_name) {
  // NOTE(zjl): PyObject_GetAttrString would return nullptr when attr_name
  // is not inside obj, but it would also set the error flag of Python.
  // If the error flag is set in C++, C++ code would not raise Exception,
  // but Python would raise Exception once C++ call ends.
  // To avoid unexpected Exception raised in Python, we check whether
  // attribute exists before calling PyObject_GetAttrString.
  //
  // Caution: PyObject_GetAttrString would increase reference count of PyObject.
  // Developer should call Py_DECREF manually after the attribute is not used.
  if (PyObject_HasAttrString(obj, attr_name)) {
    return PyObject_GetAttrString(obj, attr_name);
  } else {
    return nullptr;
  }
}

template <typename T>
static T PyObjectCast(PyObject *obj) {
  try {
    return py::cast<T>(py::handle(obj));
  } catch (py::cast_error &) {
238 239 240
    PADDLE_THROW(platform::errors::InvalidArgument(
        "Python object is not type of %s, the real type is %s",
        typeid(T).name(), obj->ob_type->tp_name));
H
hong 已提交
241 242 243 244 245 246 247 248 249 250 251 252 253
  }
}

using PyNameVarBaseMap = std::unordered_map<std::string, py::handle>;

static std::vector<std::shared_ptr<imperative::VarBase>> GetVarBaseList(
    const PyNameVarBaseMap &state_dict) {
  std::vector<std::shared_ptr<imperative::VarBase>> vec_res;
  vec_res.reserve(state_dict.size());

  for (auto &para : state_dict) {
    PyObject *py_obj = para.second.ptr();
    if (!py_obj || py_obj == Py_None) {
254 255
      PADDLE_THROW(platform::errors::InvalidArgument(
          "The parameter [%s] to save is None", para.first));
H
hong 已提交
256 257
    }
    vec_res.emplace_back(
258
        PyObjectCast<std::shared_ptr<imperative::VarBase>>(py_obj));
H
hong 已提交
259 260 261 262 263 264 265 266 267 268 269 270
  }

  return vec_res;
}

static std::vector<std::string> inline GetNameList(
    const py::handle &py_handle) {
  std::vector<std::string> vec_res;

  PyObject *py_obj = py_handle.ptr();  // get underlying PyObject
  // Python None is not nullptr in C++!
  if (!py_obj || py_obj == Py_None) {
271 272
    PADDLE_THROW(platform::errors::InvalidArgument(
        "The parameter list to save is None"));
H
hong 已提交
273 274 275 276 277 278 279 280 281 282 283 284
  }

  if (PyList_Check(py_obj)) {
    size_t len = PyList_GET_SIZE(py_obj);

    vec_res.reserve(len);

    const char *kNameField = "name";

    for (size_t i = 0; i < len; ++i) {
      PyObject *py_name =
          PyObject_GetAttrString(PyList_GET_ITEM(py_obj, i), kNameField);
285 286 287
      PADDLE_ENFORCE_NOT_NULL(py_name,
                              platform::errors::InvalidArgument(
                                  "The name of parameter to save is None"));
H
hong 已提交
288 289 290 291
      vec_res.emplace_back(PyObjectCast<std::string>(py_name));
      Py_DECREF(py_name);
    }
  } else {
292 293
    PADDLE_THROW(platform::errors::InvalidArgument(
        "The parameters to save is not a list"));
H
hong 已提交
294 295 296 297
  }
  return vec_res;
}

298 299 300 301 302 303 304 305
static void inline CreateVariableIfNotExit(
    const py::handle &py_handle, const framework::Scope &scope,
    const framework::Executor *exe = nullptr) {
  std::vector<std::string> vec_res;

  PyObject *py_obj = py_handle.ptr();  // get underlying PyObject
  // Python None is not nullptr in C++!
  if (!py_obj || py_obj == Py_None) {
306 307
    PADDLE_THROW(
        platform::errors::InvalidArgument("The parameter list to set is None"));
308 309 310 311 312 313 314 315 316 317 318 319 320
  }

  if (PyList_Check(py_obj)) {
    size_t len = PyList_GET_SIZE(py_obj);

    vec_res.reserve(len);

    const char *kNameField = "name";
    const char *kVarDescField = "desc";

    for (size_t i = 0; i < len; ++i) {
      PyObject *py_name =
          PyObject_GetAttrString(PyList_GET_ITEM(py_obj, i), kNameField);
321 322 323
      PADDLE_ENFORCE_NOT_NULL(py_name,
                              platform::errors::InvalidArgument(
                                  "The name of parameter to set is None"));
324 325 326 327 328
      auto para_name = PyObjectCast<std::string>(py_name);
      Py_DECREF(py_name);

      auto var = scope.FindVar(para_name);
      if (var == nullptr) {
329 330 331 332 333
        PADDLE_ENFORCE_NOT_NULL(exe,
                                platform::errors::InvalidArgument(
                                    "Parameter not Initialized, "
                                    "Please set argument [executor] not None "
                                    "or run startup program first"));
334 335
        PyObject *py_var_desc =
            PyObject_GetAttrString(PyList_GET_ITEM(py_obj, i), kVarDescField);
336 337 338
        PADDLE_ENFORCE_NOT_NULL(
            py_var_desc, platform::errors::InvalidArgument(
                             "The var_desc of parameter to set is None"));
339 340 341 342 343 344 345 346 347
        auto var_desc = PyObjectCast<framework::VarDesc>(py_var_desc);
        Py_DECREF(py_var_desc);
        var = const_cast<framework::Scope *>(&scope)->Var(para_name);
        auto *tensor_temp = var->GetMutable<framework::LoDTensor>();
        tensor_temp->Resize(framework::make_ddim(var_desc.GetShape()));
        tensor_temp->mutable_data(exe->GetPlace(), var_desc.GetDataType());
      }
    }
  } else {
348 349
    PADDLE_THROW(platform::errors::InvalidArgument(
        "The parameters to set is not a list"));
350 351 352 353 354
  }

  return;
}

355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378
static void AssertStaticGraphAndDygraphGradMakerNoDiff() {
  std::set<std::string> ops;
  for (auto &pair : framework::OpInfoMap::Instance().map()) {
    bool has_static_grad_maker = (pair.second.grad_op_maker_ != nullptr);
    bool has_dygraph_grad_maker =
        (pair.second.dygraph_grad_op_maker_ != nullptr);
    if (has_static_grad_maker ^ has_dygraph_grad_maker) {
      bool has_kernel =
          (framework::OperatorWithKernel::AllOpKernels().count(pair.first) > 0);
      if (has_kernel) {
        ops.insert(pair.first);
      } else {
        VLOG(5) << pair.first << " has no kernels, skip";
      }
    }
  }
  PADDLE_ENFORCE_EQ(ops.empty(), true,
                    platform::errors::Unimplemented(
                        "OperatorWithKernel [%s] have only static graph grad "
                        "maker or have only dygraph grad maker, which is not "
                        "allowed",
                        string::join_strings(ops, ',')));
}

379 380 381 382 383 384
#ifdef PADDLE_WITH_AVX
PYBIND11_MODULE(core_avx, m) {
#else
PYBIND11_MODULE(core_noavx, m) {
#endif

Y
Yu Yang 已提交
385 386 387
  // Not used, just make sure cpu_info.cc is linked.
  paddle::platform::CpuTotalPhysicalMemory();

Y
Refine  
Yu Yang 已提交
388
  paddle::memory::allocation::UseAllocatorStrategyGFlag();
S
sneaxiy 已提交
389

390 391
  AssertStaticGraphAndDygraphGradMakerNoDiff();

392
  m.doc() = "C++ core of PaddlePaddle";
393

394 395 396 397
  // using framework in this function. Since it is inside a function, it will
  // not cause namespace pollution.
  using namespace paddle::framework;  // NOLINT

398
  BindException(&m);
Y
Yu Yang 已提交
399

400 401
  m.def("set_num_threads", &platform::SetNumThreads);

402
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
403 404 405
  m.def("cudnn_version", &platform::CudnnVersion);
#endif

6
633WHU 已提交
406 407 408 409 410
  m.def("from_dlpack", [](py::capsule *dltensor) {
    DLManagedTensor *dmt = reinterpret_cast<DLManagedTensor *>(
        PyCapsule_GetPointer(dltensor->ptr(), "dltensor"));
    PyCapsule_SetName(dltensor->ptr(), "used_dltensor");
    DLTensor dl = dmt->dl_tensor;
411
    framework::Tensor tensor;
6
633WHU 已提交
412 413 414 415

    if (dl.ctx.device_type == kDLCPU) {
      paddle::framework::TensorFromDLPack(dl, &tensor);
    }
416
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
6
633WHU 已提交
417 418 419 420 421 422 423
    if (dl.ctx.device_type == kDLGPU) {
      paddle::framework::TensorFromDLPack(dl, &tensor);
    }
#endif
    return tensor;
  });

H
hong 已提交
424 425 426 427 428 429 430 431 432
  m.def("_save_static_dict",
        [](const std::string &str_file_name, const py::handle &vec_var_list,
           const Scope &scope) {
          std::vector<std::string> vec_name_list = GetNameList(vec_var_list);
          SaveStaticNameListToDisk(str_file_name, vec_name_list, scope);
        });

  m.def("_load_static_dict",
        [](const std::string &str_file_name, const py::handle &vec_var_list,
433
           const Scope &scope, const Executor *executor) {
H
hong 已提交
434
          std::vector<std::string> vec_name_list = GetNameList(vec_var_list);
435
          CreateVariableIfNotExit(vec_var_list, scope, executor);
H
hong 已提交
436 437 438
          LoadStaticNameListFromDisk(str_file_name, vec_name_list, scope);
        });

439 440 441 442 443 444
  m.def("_create_loaded_parameter",
        [](const py::handle &vec_var_list, const Scope &scope,
           const Executor *executor) {
          CreateVariableIfNotExit(vec_var_list, scope, executor);
        });

H
hong 已提交
445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463
  m.def("_save_dygraph_dict", [](const std::string &str_file_name,
                                 const PyNameVarBaseMap &state_dict) {
    auto vec_var_base_list = GetVarBaseList(state_dict);

    SaveDygraphVarBaseListToDisk(str_file_name, vec_var_base_list);
  });

  m.def("_load_dygraph_dict", [](const std::string &str_file_name) {
    auto load_tensor = LoadDygraphVarBaseListFromDisk(str_file_name);

    std::unordered_map<std::string, std::shared_ptr<imperative::VarBase>>
        map_output;

    for (size_t i = 0; i < load_tensor.size(); ++i) {
      map_output.emplace(load_tensor[i]->Name(), load_tensor[i]);
    }

    return map_output;
  });
6
633WHU 已提交
464

465 466 467 468 469 470
  m.def("save_op_version_info", [](framework::ProgramDesc &desc) {
    framework::compatible::pb::OpVersionMap pb_vmap{desc.OpVersionMap()};
    framework::compatible::SaveOpVersions(
        framework::compatible::OpVersionRegistrar::GetInstance()
            .GetVersionMap(),
        &pb_vmap);
471 472
  });

473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497
  m.def("set_printoptions", [](const py::kwargs &kwargs) {
    auto &print_opt = framework::PrintOptions::Instance();
    if (kwargs.contains("precision")) {
      print_opt.precision = kwargs["precision"].cast<int>();
    }
    if (kwargs.contains("threshold")) {
      print_opt.threshold = kwargs["threshold"].cast<int>();
    }
    if (kwargs.contains("edgeitems")) {
      print_opt.edgeitems = kwargs["edgeitems"].cast<int>();
    }
    if (kwargs.contains("linewidth")) {
      print_opt.linewidth = kwargs["linewidth"].cast<int>();
    }
    if (kwargs.contains("sci_mode")) {
      print_opt.sci_mode = kwargs["sci_mode"].cast<bool>();
    }

    VLOG(4) << "Set printoptions: precision=" << print_opt.precision
            << ", threshold=" << print_opt.threshold
            << ", edgeitems=" << print_opt.edgeitems
            << ", linewidth=" << print_opt.linewidth
            << ", sci_mode=" << print_opt.sci_mode;
  });

L
Leo Chen 已提交
498 499 500 501 502 503
  m.def("broadcast_shape", [](const std::vector<int64_t> &x_dim,
                              const std::vector<int64_t> &y_dim) {
    return vectorize(operators::details::BroadcastTwoDims(
        make_ddim(x_dim), make_ddim(y_dim), -1));
  });

S
sneaxiy 已提交
504
  m.def(
S
sneaxiy 已提交
505
      "_append_python_callable_object_and_return_id",
S
sneaxiy 已提交
506 507 508 509
      [](py::object py_obj) -> size_t {
        return paddle::operators::AppendPythonCallableObjectAndReturnId(py_obj);
      });

S
sneaxiy 已提交
510 511 512
  m.def("_get_use_default_grad_op_desc_maker_ops",
        [] { return OpInfoMap::Instance().GetUseDefaultGradOpDescMakerOps(); });

513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528
  m.def("_get_all_register_op_kernels", [] {
    auto &all_kernels = paddle::framework::OperatorWithKernel::AllOpKernels();
    std::unordered_map<std::string, std::vector<std::string>> all_kernels_info;
    for (auto &kernel_pair : all_kernels) {
      auto op_type = kernel_pair.first;
      std::vector<std::string> kernel_types;
      for (auto &info_pair : kernel_pair.second) {
        paddle::framework::OpKernelType kernel_type = info_pair.first;
        kernel_types.push_back(
            paddle::framework::KernelTypeToString(kernel_type));
      }
      all_kernels_info.emplace(op_type, kernel_types);
    }
    return all_kernels_info;
  });

S
sneaxiy 已提交
529 530 531
  // NOTE(zjl): ctest would load environment variables at the beginning even
  // though we have not `import paddle.fluid as fluid`. So we add this API
  // to enable eager deletion mode in unittest.
S
sneaxiy 已提交
532
  m.def("_set_eager_deletion_mode", &paddle::framework::SetEagerDeletionMode);
S
sneaxiy 已提交
533

534
  m.def("_set_fuse_parameter_group_size",
535
        &paddle::framework::ir::SetFuseParameterGroupsSize);
536
  m.def("_set_fuse_parameter_memory_size",
537
        &paddle::framework::ir::SetFuseParameterMemorySize);
538

S
sneaxiy 已提交
539 540 541
  m.add_object("_cleanup",
               py::capsule([]() { ScopePool::Instance().Clear(); }));

542 543
  m.def("_set_paddle_lib_path", &paddle::platform::dynload::SetPaddleLibPath);

544 545 546
  m.def("_promote_types_if_complex_exists",
        &paddle::framework::PromoteTypesIfComplexExists);

547
  BindImperative(&m);
548

549 550 551
  py::class_<framework::Tensor>(m, "Tensor", py::buffer_protocol())
      .def("__array__",
           [](framework::Tensor &self) { return TensorToPyArray(self); })
S
sneaxiy 已提交
552
      .def("_is_initialized",
553
           [](const framework::Tensor &self) { return self.IsInitialized(); })
Y
yuyang18 已提交
554
      .def("_get_dims",
555
           [](const framework::Tensor &self) { return vectorize(self.dims()); })
Y
yuyang18 已提交
556
      .def("_set_dims",
557
           [](framework::Tensor &self, const std::vector<int64_t> &dim) {
Y
Yu Yang 已提交
558
             self.Resize(make_ddim(dim));
Y
Yu Yang 已提交
559
           })
Y
yuyang18 已提交
560
      .def("_set_layout",
561
           [](framework::Tensor &self, const std::string &layout) {
D
dzhwinter 已提交
562 563
             self.set_layout(StringToDataLayout(layout));
           })
Y
yuyang18 已提交
564
      .def("_alloc_float",
565
           [](framework::Tensor &self, paddle::platform::CUDAPlace &place) {
Q
qijun 已提交
566
             self.mutable_data<float>(place);
Y
Yu Yang 已提交
567
           })
568
      .def("_alloc_float",
569
           [](framework::Tensor &self, paddle::platform::XPUPlace &place) {
570 571
             self.mutable_data<float>(place);
           })
Y
yuyang18 已提交
572
      .def("_alloc_float",
573
           [](framework::Tensor &self, paddle::platform::CPUPlace &place) {
Q
qijun 已提交
574
             self.mutable_data<float>(place);
Y
Yu Yang 已提交
575
           })
576
      .def("_alloc_double",
577
           [](framework::Tensor &self, paddle::platform::CPUPlace &place) {
578 579
             self.mutable_data<double>(place);
           })
Y
yuyang18 已提交
580
      .def("_alloc_int",
581
           [](framework::Tensor &self, paddle::platform::CPUPlace &place) {
Q
qijun 已提交
582
             self.mutable_data<int>(place);
Y
Yu Yang 已提交
583
           })
584
      .def("_alloc_int",
585
           [](framework::Tensor &self, paddle::platform::XPUPlace &place) {
586 587
             self.mutable_data<int>(place);
           })
Y
yuyang18 已提交
588
      .def("_alloc_int",
589
           [](framework::Tensor &self, paddle::platform::CUDAPlace &place) {
Q
qijun 已提交
590
             self.mutable_data<int>(place);
Q
qijun 已提交
591
           })
Y
yuyang18 已提交
592
      .def("_alloc_int",
593 594
           [](framework::Tensor &self,
              paddle::platform::CUDAPinnedPlace &place) {
C
chengduoZH 已提交
595 596
             self.mutable_data<int>(place);
           })
Y
yuyang18 已提交
597
      .def("_alloc_float",
598 599
           [](framework::Tensor &self,
              paddle::platform::CUDAPinnedPlace &place) {
C
chengduoZH 已提交
600 601
             self.mutable_data<float>(place);
           })
602
      .def("_mutable_data",
603
           [](framework::Tensor &self, paddle::platform::CPUPlace &place,
604 605 606
              paddle::framework::proto::VarType::Type type) {
             return reinterpret_cast<uintptr_t>(self.mutable_data(place, type));
           })
607
      .def("_mutable_data",
608
           [](framework::Tensor &self, paddle::platform::XPUPlace &place,
609 610 611
              paddle::framework::proto::VarType::Type type) {
             return reinterpret_cast<uintptr_t>(self.mutable_data(place, type));
           })
612
      .def("_mutable_data",
613
           [](framework::Tensor &self, paddle::platform::CUDAPlace &place,
614 615 616 617
              paddle::framework::proto::VarType::Type type) {
             return reinterpret_cast<uintptr_t>(self.mutable_data(place, type));
           })
      .def("_mutable_data",
618
           [](framework::Tensor &self, paddle::platform::CUDAPinnedPlace &place,
619 620 621
              paddle::framework::proto::VarType::Type type) {
             return reinterpret_cast<uintptr_t>(self.mutable_data(place, type));
           })
622
      .def("_clear", &framework::Tensor::clear)
623
      .def("set", SetTensorFromPyArray<paddle::platform::CPUPlace>,
624
           py::arg("array"), py::arg("place"), py::arg("zero_copy") = false)
625 626
      .def("set", SetTensorFromPyArray<paddle::platform::XPUPlace>,
           py::arg("array"), py::arg("place"), py::arg("zero_copy") = false)
627
      .def("set", SetTensorFromPyArray<paddle::platform::CUDAPlace>,
628
           py::arg("array"), py::arg("place"), py::arg("zero_copy") = false)
629
      .def("set", SetTensorFromPyArray<paddle::platform::CUDAPinnedPlace>,
630 631
           py::arg("array"), py::arg("place"), py::arg("zero_copy") = false,
           R"DOC(
L
Leo Chen 已提交
632 633 634 635
        Set the data of LoDTensor on place with given numpy array.
        
        Args:
          lod (numpy.ndarray): The data to set.
636
          place (CPUPlace|CUDAPlace|XPUPlace|CUDAPinnedPlace): The place where the 
L
Leo Chen 已提交
637
          LoDTensor is to be set.
638 639
          zero_copy (bool, optional): Whether to share memory with the input numpy array.
          This parameter only works with CPUPlace. Default: False.
L
Leo Chen 已提交
640 641 642 643 644 645 646 647 648 649 650 651 652

        Returns:
            None.

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
                import numpy as np

                t = fluid.LoDTensor()
                t.set(np.ndarray([5, 30]), fluid.CPUPlace())
          )DOC")
653

654 655 656
      .def("shape",
           [](framework::Tensor &self) { return vectorize(self.dims()); },
           R"DOC(
L
Leo Chen 已提交
657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672
           Return the shape of LoDTensor.

           Returns:
               list[int]: The shape of LoDTensor.


           Examples:
               .. code-block:: python

                  import paddle.fluid as fluid
                  import numpy as np

                  t = fluid.LoDTensor()
                  t.set(np.ndarray([5, 30]), fluid.CPUPlace())
                  print(t.shape())  # [5, 30]
           )DOC")
6
633WHU 已提交
673
      .def("_to_dlpack",
674
           [](framework::Tensor &self) {
6
633WHU 已提交
675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694
             DLPackTensor dlpack_tensor(self, 1);
             DLManagedTensor *dmt =
                 dlpack_tensor.ToCudfCompatibleDLManagedTensor();
             auto capsule = py::capsule(
                 static_cast<void *>(dmt), "dltensor", [](PyObject *ptr) {
                   if (ptr) {
                     auto dltensor = new DLManagedTensor;
                     try {
                       dltensor = reinterpret_cast<DLManagedTensor *>(
                           PyCapsule_GetPointer(ptr, "used_dltensor"));
                       return;
                     } catch (...) {
                       dltensor = reinterpret_cast<DLManagedTensor *>(
                           PyCapsule_GetPointer(ptr, "dltensor"));
                     }
                     dltensor->deleter(dltensor);
                   }
                 });
             return capsule;
           })
Y
yuyang18 已提交
695 696 697 698
      .def("_set_float_element", TensorSetElement<float>)
      .def("_get_float_element", TensorGetElement<float>)
      .def("_set_double_element", TensorSetElement<double>)
      .def("_get_double_element", TensorGetElement<double>)
699 700
      .def("_place", [](framework::Tensor &self) { return self.place(); })
      .def("_dtype", [](framework::Tensor &self) { return self.type(); })
701
      .def("_layout",
702 703 704 705
           [](framework::Tensor &self) {
             return DataLayoutToString(self.layout());
           })
      .def("_share_data_with", &framework::Tensor::ShareDataWith)
706
      .def("__getitem__", PySliceTensor, py::return_value_policy::reference)
707
      .def("__str__", [](const framework::Tensor &self) {
708 709 710 711
        std::stringstream ostr;
        ostr << self;
        return ostr.str();
      });
Y
Yu Yang 已提交
712

L
Leo Chen 已提交
713
  // TODO(cql): add reference: en_user_guide_lod_tensor
714
  py::class_<LoDTensor, framework::Tensor>(m, "LoDTensor", R"DOC(
L
Leo Chen 已提交
715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788
    LoDTensor is a Tensor with optional LoD (Level of Details) information, 
    it can be used for variable-length sequences, 
    see :ref:`user_guide_lod_tensor` for details.

    LoDTensor can be converted to numpy array using :code:`numpy.array(lod_tensor)`.

    You can skip the following explanation if you don't need to know details 
    of LoDTensor.

    The following two examples show how to use LODtensor to represent 
    variable-length sequences.
    
    Example 1:
    
    Suppose x is a LoDTensor representing a variable-length sequence. 
    It contains two logical subsequences, the length of first logical sequence 
    is 2 (e.g., number of samples is 2), the length of second logical sequence 
    is 3, and the total length is 5. The data of the first logical sequence is 
    [1, 2], [3, 4], and the data of the second logical sequence is [5, 6], 
    [7, 8], [9, 10]. The data dimension of each sample is 2. So, the final 
    shape of the LoDTensor is [5, 2], of which 5 is the total length and 2 is 
    the dimension of each sample.
    
    Logically, we can represent the variable-length sequence in two ways: one 
    is in the form of recursive sequence lengths, that is, 
    x.recursive_sequence_lengths=[[2, 3]]; the other is in the form of offsets, 
    that is, x.lod=[[0, 2, 2+3]]. These two representations are equivalent, and 
    you can set and retrieve recursive_sequence_lengths or LoD through the 
    corresponding interfaces of LoDTensor introduced later.

    Actually, in order to access sequence faster, Paddle uses offset to store 
    different lengths of sequences. 
    Therefore, the operations on recursive_sequence_lengths will be converted 
    to the operations on LoD eventually.
    
    .. code-block:: python

      y.data = [[1, 2], [3, 4],
                [5, 6], [7, 8],
                [9, 10], [11, 12], [13, 14]]

      y.shape = [2+2+3, 2]

      y.recursive_sequence_lengths = [[2, 1], [2, 2, 3]]

      y.lod = [[0, 2, 3], [0, 2, 4, 7]]

    Example 2:

    LoD may have more than one level (for example, a paragraph may have more 
    than one sentence and a sentence may have more than one word). Suppose y 
    is a LoDTensor and its lod_level is 2. 
    From level = 0, there are two logical sequences, the length of which is 
    2 and 1, respectively, indicating that the first logical sequence contains 
    two sub-sequences and the second logical sequence contains one sub-sequence. 
    From level = 1, the lengths of two sub-sequences contained by the first 
    logical sequence is 2 and 2, and the length of sub-sequence contained by 
    the second logical sequence is 3.
      
    Therefore, the LoDTensor is represented in the form of recursive sequence 
    lengths as y.recursive_sequence_lengths=[[2,1], [2,2,3]]; and equally, in 
    the form of offset, it is represented as y.lod=[[0,2,3], [0,2,4,7]].

    .. code-block:: python

      y.data = [[1, 2], [3, 4],
                [5, 6], [7, 8],
                [9, 10], [11, 12], [13, 14]]

      y.shape = [2+2+3, 2]

      y.recursive_sequence_lengths = [[2, 1], [2, 2, 3]]

      y.lod = [[0, 2, 3], [0, 2, 4, 7]]
Z
Zeng Jinle 已提交
789 790 791 792 793 794 795

    Examples:
        .. code-block:: python

          import paddle.fluid as fluid

          t = fluid.LoDTensor()
X
Xin Pan 已提交
796 797

        )DOC")
798 799
      .def("__array__",
           [](framework::Tensor &self) { return TensorToPyArray(self); })
800 801 802 803 804 805 806 807 808
      .def("__init__",
           [](LoDTensor &instance, const std::vector<std::vector<size_t>>
                                       &recursive_sequence_lengths) {
             LoD new_lod;
             new_lod.reserve(recursive_sequence_lengths.size());
             std::copy(recursive_sequence_lengths.begin(),
                       recursive_sequence_lengths.end(),
                       std::back_inserter(new_lod));
             LoD new_offset_lod = ConvertToOffsetBasedLoD(new_lod);
C
chengduo 已提交
809 810
             PADDLE_ENFORCE_EQ(
                 CheckLoD(new_offset_lod, -1), true,
811 812 813 814
                 platform::errors::InvalidArgument(
                     "The provided recursive_sequence_lengths info is invalid, "
                     "the LoD converted by recursive_sequence_lengths is %s",
                     new_lod));
815 816
             new (&instance) LoDTensor(new_offset_lod);
           })
Y
Yu Yang 已提交
817
      .def("__init__", [](LoDTensor &instance) { new (&instance) LoDTensor(); })
G
gongweibao 已提交
818
      // We implement offset based LOD in C++ while we use length based with
H
hong 已提交
819 820
      // Python API. So we changed set_lod to set_recursive_sequence_lengths
      // to
G
gongweibao 已提交
821 822 823
      // avoid misuse.
      // The discussion is here:
      // https://github.com/PaddlePaddle/Paddle/issues/10855
D
dangqingqing 已提交
824
      .def("set_lod",
825
           [](LoDTensor &self, const std::vector<std::vector<size_t>> &lod) {
826
             // the input lod is offset-based level-of-detail info
Y
Yu Yang 已提交
827
             LoD new_lod;
828 829
             new_lod.reserve(lod.size());
             std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod));
C
chengduo 已提交
830 831
             PADDLE_ENFORCE_EQ(
                 CheckLoD(new_lod, vectorize(self.dims()).front()), true,
832 833
                 platform::errors::InvalidArgument(
                     "The provided LoD is invalid, the LoD is %s", new_lod));
834
             self.set_lod(new_lod);
S
sneaxiy 已提交
835 836 837 838 839
           },
           py::arg("lod"), R"DOC(
           Set LoD of the LoDTensor.

           Args:
L
Leo Chen 已提交
840 841 842 843
               lod (list[list[int]]): The lod to set.

           Returns:
                None.
Z
Zeng Jinle 已提交
844 845 846 847 848 849 850 851 852 853

           Examples:
               .. code-block:: python

                 import paddle.fluid as fluid
                 import numpy as np

                 t = fluid.LoDTensor()
                 t.set(np.ndarray([5, 30]), fluid.CPUPlace())
                 t.set_lod([[0, 2, 5]])
L
Leo Chen 已提交
854
                 print(t.lod()) # [[0, 2, 5]]
S
sneaxiy 已提交
855
           )DOC")
856 857 858 859 860 861 862 863 864 865 866
      .def("set_recursive_sequence_lengths",
           [](LoDTensor &self, const std::vector<std::vector<size_t>>
                                   &recursive_sequence_lengths) {
             // the input recursive_sequence_lengths is length-based
             // level-of-detail info
             LoD new_lod;
             new_lod.reserve(recursive_sequence_lengths.size());
             std::copy(recursive_sequence_lengths.begin(),
                       recursive_sequence_lengths.end(),
                       std::back_inserter(new_lod));
             LoD new_offset_lod = ConvertToOffsetBasedLoD(new_lod);
C
chengduo 已提交
867 868
             PADDLE_ENFORCE_EQ(
                 CheckLoD(new_offset_lod, vectorize(self.dims()).front()), true,
869 870 871 872 873
                 platform::errors::InvalidArgument(
                     "The provided recursive_sequence_lengths info is invalid, "
                     "the LoD converted by recursive_sequence_lengths is "
                     "%s",
                     new_lod));
874
             self.set_lod(new_offset_lod);
S
sneaxiy 已提交
875 876
           },
           py::arg("recursive_sequence_lengths"), R"DOC(
L
Leo Chen 已提交
877
           Set LoD of the LoDTensor according to recursive sequence lengths.
S
sneaxiy 已提交
878

L
Leo Chen 已提交
879
           For example, if recursive_sequence_lengths=[[2, 3]], which means
880
           there are two sequences with length 2 and 3 respectively, the
L
Leo Chen 已提交
881
           corresponding lod would be [[0, 2, 2+3]], i.e., [[0, 2, 5]].
S
sneaxiy 已提交
882 883

           Args:
L
Leo Chen 已提交
884 885 886 887
                recursive_sequence_lengths (list[list[int]]): The recursive sequence lengths.
           
           Returns:
                None.
Z
Zeng Jinle 已提交
888 889 890 891 892 893 894 895 896 897

           Examples:
               .. code-block:: python

                 import paddle.fluid as fluid
                 import numpy as np

                 t = fluid.LoDTensor()
                 t.set(np.ndarray([5, 30]), fluid.CPUPlace())
                 t.set_recursive_sequence_lengths([[2, 3]])
L
Leo Chen 已提交
898 899
                 print(t.recursive_sequence_length())  # [[2, 3]]
                 print(t.lod())  # [[0, 2, 5]]
S
sneaxiy 已提交
900
           )DOC")
901 902 903 904 905 906 907 908
      .def("lod",
           [](LoDTensor &self) -> std::vector<std::vector<size_t>> {
             // output the offset-based lod info
             LoD lod = self.lod();
             std::vector<std::vector<size_t>> new_lod;
             new_lod.reserve(lod.size());
             std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod));
             return new_lod;
S
sneaxiy 已提交
909 910 911 912 913
           },
           R"DOC(
           Return the LoD of the LoDTensor.

           Returns:
L
Leo Chen 已提交
914 915
               list[list[int]]: The lod of the LoDTensor.
           
Z
Zeng Jinle 已提交
916 917 918 919 920 921 922 923 924 925
           Examples:
               .. code-block:: python

                 import paddle.fluid as fluid
                 import numpy as np

                 t = fluid.LoDTensor()
                 t.set(np.ndarray([5, 30]), fluid.CPUPlace())
                 t.set_lod([[0, 2, 5]])
                 print(t.lod()) # [[0, 2, 5]]
S
sneaxiy 已提交
926
           )DOC")
G
gongweibao 已提交
927
      // Set above comments of set_lod.
928 929 930 931 932 933 934 935
      .def("recursive_sequence_lengths",
           [](LoDTensor &self) -> std::vector<std::vector<size_t>> {
             // output the length-based lod info
             LoD lod = ConvertToLengthBasedLoD(self.lod());
             std::vector<std::vector<size_t>> new_lod;
             new_lod.reserve(lod.size());
             std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod));
             return new_lod;
S
sneaxiy 已提交
936 937
           },
           R"DOC(
L
Leo Chen 已提交
938 939
           Return the recursive sequence lengths corresponding to of the LodD 
           of the LoDTensor.
S
sneaxiy 已提交
940 941

           Returns:
L
Leo Chen 已提交
942
                list[list[int]]: The recursive sequence lengths.
Z
Zeng Jinle 已提交
943 944 945 946 947 948 949 950 951 952 953

           Examples:
               .. code-block:: python

                 import paddle.fluid as fluid
                 import numpy as np

                 t = fluid.LoDTensor()
                 t.set(np.ndarray([5, 30]), fluid.CPUPlace())
                 t.set_recursive_sequence_lengths([[2, 3]])
                 print(t.recursive_sequence_lengths()) # [[2, 3]]
S
sneaxiy 已提交
954 955 956 957 958 959 960 961
           )DOC")
      .def("has_valid_recursive_sequence_lengths",
           [](LoDTensor &self) -> bool {
             // Check that the lod info is valid and match the outermost
             // dimension of the LoDTensor data
             return CheckLoD(self.lod(), vectorize(self.dims()).front());
           },
           R"DOC(
L
Leo Chen 已提交
962
           Check whether the LoD of the LoDTensor is valid.
S
sneaxiy 已提交
963 964

           Returns:
L
Leo Chen 已提交
965
               bool: Whether the LoD is valid.
Z
Zeng Jinle 已提交
966 967 968 969 970 971 972 973 974 975 976

           Examples:
               .. code-block:: python

                 import paddle.fluid as fluid
                 import numpy as np

                 t = fluid.LoDTensor()
                 t.set(np.ndarray([5, 30]), fluid.CPUPlace())
                 t.set_recursive_sequence_lengths([[2, 3]])
                 print(t.has_valid_recursive_sequence_lengths()) # True
W
wopeizl 已提交
977 978 979 980 981 982 983
           )DOC")
      .def("__getitem__", PySliceTensor, py::return_value_policy::reference,
           R"DOC(
           Slice the original Tensor, and remove the LoD information.

           Returns:
               out (Tensor): new Tensor(NOT LoDTensor).
984
           )DOC")
985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002
      .def("__str__",
           [](const LoDTensor &self) {
             std::stringstream ostr;
             ostr << self;
             return ostr.str();
           })
      .def("_copy", [](const LoDTensor &self, const platform::Place &place) {
        // follow fetch_op's inplementation
        LoDTensor dst;
        if (self.IsInitialized() && self.numel() > 0) {
          TensorCopySync(self, place, &dst);
        } else {
          // Not copy, if the src tensor is empty.
          dst.clear();
          dst.Resize({0});
        }
        dst.set_lod(self.lod());
        return dst;
1003
#ifdef _WIN32
1004
      });
1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054
#else
           })
      .def(py::pickle(
          [](const LoDTensor &t) {  // __getstate__
            auto holder = t.Holder();
            PADDLE_ENFORCE_EQ(
              platform::is_cpu_place(holder->place()), true,
              platform::errors::PreconditionNotMet(
                  "LoDTensor is not on CPU."
                  "Now only LoDTensor on CPU can be serialized."));
            auto* mmap_writer_allocation =
              dynamic_cast<memory::allocation::MemoryMapWriterAllocation *>(
                holder.get());
            PADDLE_ENFORCE_NOT_NULL(mmap_writer_allocation,
              platform::errors::PreconditionNotMet(
                "LoDTensor is not in shared memory."
                "Now only LoDTensor on shared memory can be serialized."));
            int type_idx = static_cast<int>(t.type());

            return py::make_tuple(mmap_writer_allocation->ipc_name(),
                                  mmap_writer_allocation->size(),
                                  type_idx, vectorize(t.dims()), t.lod());
          },
          [](py::tuple t) {  // __setstate__
            if (t.size() != 5)
              throw std::runtime_error("Invalid LoDTensor state!");

            // 1. Create a new C++ instance
            LoDTensor tensor;

            // 2. Rebuild Allocation
            const std::string &ipc_name = t[0].cast<std::string>();
            size_t size = t[1].cast<size_t>();
            auto shared_reader_holder =
              memory::allocation::RebuildMemoryMapReaderAllocation(
                ipc_name, size);

            // 3. Maintain global fd set
            VLOG(3) << "LoDTensor ipc name: " << ipc_name;
            memory::allocation::MemoryMapFdSet::Instance().Insert(ipc_name);

            // 4. Rebuild LoDTensor
            tensor.ResetHolderWithType(shared_reader_holder,
              static_cast<proto::VarType::Type>(t[2].cast<int>()));
            tensor.Resize(make_ddim(t[3].cast<std::vector<int>>()));
            tensor.set_lod(t[4].cast<framework::LoD>());

            return tensor;
          }));
#endif
D
dangqingqing 已提交
1055

Q
qijun 已提交
1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066
  py::class_<SelectedRows>(m, "SelectedRows")
      .def("__init__",
           [](SelectedRows &instance) { new (&instance) SelectedRows(); })
      .def("__init__",
           [](SelectedRows &instance, const std::vector<int64_t> rows,
              const int64_t &height) {
             new (&instance) SelectedRows(rows, height);
           })
      .def("get_tensor",
           [](SelectedRows &self) { return self.mutable_value(); },
           py::return_value_policy::reference)
1067 1068
      .def("numel",
           [](SelectedRows &self) -> int64_t { return self.value().numel(); })
Q
qijun 已提交
1069 1070
      .def("set_height", &SelectedRows::set_height)
      .def("height", &SelectedRows::height)
Q
qijun 已提交
1071 1072
      .def("set_rows",
           [](SelectedRows &self, std::vector<int64_t> rows) {
1073
#if !defined(PADDLE_WITH_CUDA) && !defined(PADDLE_WITH_HIP)
Q
qijun 已提交
1074 1075 1076 1077 1078 1079
             self.set_rows(rows);
#else
        Vector<int64_t> new_rows(rows);
        self.set_rows(new_rows);
#endif
           })
1080
      .def("sync_index", [](SelectedRows &instance) { instance.SyncIndex(); })
1081
      .def("rows", [](SelectedRows &self) {
1082 1083 1084 1085 1086
        auto rows = self.rows();
        std::vector<int64_t> new_rows;
        new_rows.reserve(rows.size());
        std::copy(rows.begin(), rows.end(), std::back_inserter(new_rows));
        return new_rows;
1087
      });
Q
qijun 已提交
1088

1089
  py::class_<Variable>(m, "Variable", R"DOC(Variable Class.
1090 1091 1092

All parameter, weight, gradient are variables in Paddle.
)DOC")
S
sneaxiy 已提交
1093
      .def(py::init<>())
1094
      .def("is_int", [](const Variable &var) { return var.IsType<int>(); })
1095
      .def("set_int",
1096 1097
           [](Variable &var, int val) -> void { *var.GetMutable<int>() = val; })
      .def("get_int", [](const Variable &var) -> int { return var.Get<int>(); })
1098 1099 1100 1101 1102 1103 1104
      .def("is_float", [](const Variable &var) { return var.IsType<float>(); })
      .def("set_float",
           [](Variable &var, float val) -> void {
             *var.GetMutable<float>() = val;
           })
      .def("get_float",
           [](const Variable &var) -> float { return var.Get<float>(); })
Y
Yu Yang 已提交
1105
      .def("get_tensor",
1106 1107
           [](Variable &self) -> LoDTensor * {
             return self.GetMutable<LoDTensor>();
D
dangqingqing 已提交
1108 1109
           },
           py::return_value_policy::reference)
1110 1111 1112 1113
      .def("get_bytes",
           [](Variable &self) {
             return py::bytes(*self.GetMutable<std::string>());
           })
Y
Yu Yang 已提交
1114 1115 1116
      .def("get_lod_rank_table",
           [](Variable &self) { return self.GetMutable<LoDRankTable>(); },
           py::return_value_policy::reference)
Q
qijun 已提交
1117 1118 1119 1120 1121
      .def("get_selected_rows",
           [](Variable &self) -> SelectedRows * {
             return self.GetMutable<SelectedRows>();
           },
           py::return_value_policy::reference)
Y
Yu Yang 已提交
1122 1123 1124
      .def("get_lod_tensor_array",
           [](Variable &self) { return self.GetMutable<LoDTensorArray>(); },
           py::return_value_policy::reference)
1125 1126 1127
      .def("get_fetch_list",
           [](Variable &self) { return self.GetMutable<FetchList>(); },
           py::return_value_policy::reference)
1128
#if (defined(PADDLE_WITH_NCCL))
D
Dong Zhihong 已提交
1129 1130 1131 1132 1133
      .def("get_communicator",
           [](Variable &self) -> platform::Communicator * {
             return self.GetMutable<platform::Communicator>();
           },
           py::return_value_policy::reference)
Y
Yu Yang 已提交
1134
#endif
Y
Refine  
Yu Yang 已提交
1135 1136
      .def("get_reader",
           [](Variable &self) -> framework::ReaderHolder * {
1137 1138 1139 1140
             PADDLE_ENFORCE_EQ(
                 self.IsType<framework::ReaderHolder>(), true,
                 platform::errors::InvalidArgument(
                     "The variable is not type of ReaderHolder."));
Y
Refine  
Yu Yang 已提交
1141 1142
             return self.GetMutable<framework::ReaderHolder>();
           },
1143 1144 1145 1146 1147
           py::return_value_policy::reference)
      .def("set_scope", [](Variable &self, Scope &scope) {
        auto scope_vec = self.GetMutable<std::vector<framework::Scope *>>();
        scope_vec->emplace_back(&scope);
      });
1148

S
sneaxiy 已提交
1149
  BindReader(&m);
Y
Refine  
Yu Yang 已提交
1150

S
sneaxiy 已提交
1151
  py::class_<Scope>(m, "_Scope", R"DOC(
Q
Qiao Longfei 已提交
1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164
    Scope is an association of a name to Variable. All variables belong to Scope.

    Variables in a parent scope can be retrieved from local scope.

    You need to specify a scope to run a Net, i.e., `exe.Run(&scope)`.
    One net can run in different scopes and update different variable in the
    scope.

    You can create var in a scope and get it from the scope.

    Examples:
        .. code-block:: python

1165
          import paddle.fluid as fluid
Q
Qiao Longfei 已提交
1166 1167 1168 1169 1170 1171
          # create tensor from a scope and set value to it.
          param = scope.var('Param').get_tensor()
          param_array = np.full((height, row_numel), 5.0).astype("float32")
          param.set(param_array, place)

        )DOC")
S
sneaxiy 已提交
1172 1173
      .def("_remove_from_pool",
           [](Scope &self) { ScopePool::Instance().Remove(&self); })
D
dongzhihong 已提交
1174
      .def("var",
1175
           [](Scope &self, const std::string &name) -> Variable * {
D
dongzhihong 已提交
1176
             return self.Var(name);
Y
Yu Yang 已提交
1177
           },
S
sneaxiy 已提交
1178 1179
           py::arg("name"),
           R"DOC(
1180
           Find or create variable named :code:`name` in the current scope.
S
sneaxiy 已提交
1181

1182
           If the variable named :code:`name` does not exist in the
S
sneaxiy 已提交
1183
           current scope, the variable would be created. Otherwise,
1184
           return the existing variable.
S
sneaxiy 已提交
1185 1186

           Args:
1187 1188
               name (str): the variable name.

S
sneaxiy 已提交
1189
           Returns:
1190
               out (core.Variable): the found or created variable.
S
sneaxiy 已提交
1191 1192 1193 1194
           )DOC",
           py::return_value_policy::reference)
      .def("find_var", &Scope::FindVar, py::arg("name"),
           R"DOC(
1195
           Find variable named :code:`name` in the current scope or
1196
           its parent scope. Return None if not found. 
1197

S
sneaxiy 已提交
1198 1199
           Args:
               name (str): the variable name.
1200

S
sneaxiy 已提交
1201
           Returns:
1202
               out (core.Variable|None): the found variable or None.
S
sneaxiy 已提交
1203
           )DOC",
1204
           py::return_value_policy::reference)
1205
      .def("new_scope", [](Scope &self) -> Scope * { return &self.NewScope(); },
S
sneaxiy 已提交
1206 1207 1208 1209 1210 1211
           R"DOC(
           Create a new sub-scope of the current scope.

           Returns:
               out (core._Scope): the created sub-scope.
           )DOC",
1212
           py::return_value_policy::reference)
S
sneaxiy 已提交
1213 1214 1215
      .def("drop_kids", &Scope::DropKids,
           R"DOC(
           Delete all sub-scopes of the current scope.
S
sneaxiy 已提交
1216 1217
           )DOC")
      .def("_kids", &Scope::kids);
1218

S
sneaxiy 已提交
1219 1220 1221 1222 1223 1224
  m.def("Scope",
        []() -> Scope * {
          auto *s = new Scope();
          ScopePool::Instance().Insert(std::unique_ptr<Scope>(s));
          return s;
        },
S
sneaxiy 已提交
1225 1226
        R"DOC(
        Create a new scope.
1227

S
sneaxiy 已提交
1228 1229 1230
        Returns:
            out (core._Scope): the created scope.
        )DOC",
S
sneaxiy 已提交
1231 1232
        py::return_value_policy::reference);

Y
Yu Yang 已提交
1233 1234
  //! @note: Be careful! PyBind will return std::string as an unicode, not
  //! Python str. If you want a str object, you should cast them in Python.
Y
Yu Yang 已提交
1235 1236
  m.def("get_all_op_protos", []() -> std::vector<py::bytes> {
    std::vector<py::bytes> ret_values;
1237 1238 1239 1240
    for (auto &iter : OpInfoMap::Instance().map()) {
      auto &info = iter.second;
      if (info.HasOpProtoAndChecker()) {
        std::string str;
C
chengduo 已提交
1241 1242
        PADDLE_ENFORCE_EQ(
            info.Proto().SerializeToString(&str), true,
1243 1244
            platform::errors::Fatal(
                "Serialize OpProto Error. This could be a bug of Paddle."));
1245 1246 1247
        ret_values.emplace_back(str);
      }
    }
Y
Yu Yang 已提交
1248 1249
    return ret_values;
  });
1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262
  m.def("get_op_attrs_default_value",
        [](py::bytes byte_name) -> paddle::framework::AttributeMap {
          std::string op_type = byte_name;
          paddle::framework::AttributeMap res;
          auto info = OpInfoMap::Instance().GetNullable(op_type);
          if (info != nullptr) {
            if (info->HasOpProtoAndChecker()) {
              auto op_checker = info->Checker();
              res = op_checker->GetAttrsDefaultValuesMap();
            }
          }
          return res;
        });
1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278
  m.def(
      "get_grad_op_desc", [](const OpDesc &op_desc,
                             const std::unordered_set<std::string> &no_grad_set,
                             const std::vector<BlockDesc *> &grad_sub_block) {
        std::unordered_map<std::string, std::string> grad_to_var;
        std::vector<std::unique_ptr<OpDesc>> grad_op_descs =
            framework::OpInfoMap::Instance()
                .Get(op_desc.Type())
                .GradOpMaker()(op_desc, no_grad_set, &grad_to_var,
                               grad_sub_block);
        std::vector<OpDesc *> grad_op_desc_ptrs(grad_op_descs.size());
        std::transform(grad_op_descs.begin(), grad_op_descs.end(),
                       grad_op_desc_ptrs.begin(),
                       [](std::unique_ptr<OpDesc> &p) { return p.release(); });
        return std::make_pair(grad_op_desc_ptrs, grad_to_var);
      });
1279 1280 1281
  m.def("has_grad_op_maker", [](const std::string op_type) {
    return framework::OpInfoMap::Instance().Get(op_type).HasGradOpMaker();
  });
1282 1283 1284 1285 1286
  m.def("has_non_empty_grad_op_maker", [](const std::string op_type) {
    return framework::OpInfoMap::Instance()
        .Get(op_type)
        .HasNonEmptyGradOpMaker();
  });
1287 1288 1289
  m.def("has_infer_inplace", [](const std::string op_type) {
    return framework::OpInfoMap::Instance().Get(op_type).HasInferInplace();
  });
1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303
  m.def("infer_no_need_buffer_slots",
        [](const std::string op_type, const framework::VariableNameMap &inputs,
           const framework::VariableNameMap &outputs,
           const framework::AttributeMap &attrs) {
          auto infer_func = framework::OpInfoMap::Instance()
                                .Get(op_type)
                                .NoNeedBufferVarsInferer();
          if (infer_func) {
            return infer_func(inputs, outputs, attrs);
          } else {
            std::unordered_set<std::string> empty = {};
            return empty;
          }
        });
Y
Yu Yang 已提交
1304
  m.def("prune", [](const ProgramDesc &origin,
1305
                    const std::set<std::string> &feeded_var_names,
1306
                    const std::vector<std::array<size_t, 2>> &targets) {
Y
Yu Yang 已提交
1307
    ProgramDesc prog_with_targets(origin);
1308

1309
    for (const auto &t : targets) {
1310
      prog_with_targets.MutableBlock(t[0])->Op(t[1])->SetIsTarget(true);
1311
    }
1312
    proto::ProgramDesc pruned_desc;
1313 1314 1315 1316
    auto pruned_origin_block_id_map =
        Prune(*prog_with_targets.Proto(), feeded_var_names, &pruned_desc);
    return std::make_tuple(ProgramDesc(pruned_desc),
                           pruned_origin_block_id_map);
1317
  });
1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334
  m.def("prune_backward",
        [](const framework::ProgramDesc &program) {
          return PruneBackward(program);
        },
        R"DOC(
             Prune the backward part of a program, mostly called in
             program.clone(for_test=True).
              
             Args:
                   program (ProgramDesc): The original program.

             Returns:
                   tuple(ProgramDesc, map<int, int>): The first part is 
                   the pruned program desc, and the second part is a map
                   which contains the id pair of pruned block and corresponding
                   origin block.
           )DOC");
1335 1336 1337 1338
  m.def("empty_var_name",
        []() { return std::string(framework::kEmptyVarName); });
  m.def("grad_var_suffix",
        []() { return std::string(framework::kGradVarSuffix); });
1339 1340 1341
  m.def_submodule(
       "var_names",
       "The module will return special predefined variable name in Paddle")
Y
Yi Wang 已提交
1342 1343
      .def("empty", []() { return kEmptyVarName; })
      .def("temp", []() { return kTempVarName; });
1344

Q
qijun 已提交
1345
  // clang-format off
Y
Yu Yang 已提交
1346
  py::class_<paddle::platform::DeviceContext>(m, "DeviceContext")
Q
qijun 已提交
1347 1348
      .def_static("create",
                  [](paddle::platform::CPUPlace& place)
Q
qijun 已提交
1349
                      -> paddle::platform::DeviceContext* {
Q
qijun 已提交
1350 1351
                    return new paddle::platform::CPUDeviceContext();
                  })
1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363
      .def_static("create",
                  [](paddle::platform::XPUPlace& place)
                      -> paddle::platform::DeviceContext* {
#ifndef PADDLE_WITH_XPU
             PADDLE_THROW(
                 platform::errors::PermissionDenied(
                 "Cannot use XPUPlace in CPU/GPU version, "
                 "Please recompile or reinstall Paddle with XPU support."));
#else
                    return new paddle::platform::XPUDeviceContext(place);
#endif
                  })
Q
qijun 已提交
1364
      .def_static("create",
D
dzhwinter 已提交
1365
                  [](paddle::platform::CUDAPlace& place)
Q
qijun 已提交
1366
                      -> paddle::platform::DeviceContext* {
1367
#if !defined(PADDLE_WITH_CUDA) && !defined(PADDLE_WITH_HIP)
1368 1369 1370 1371
             PADDLE_THROW(
                 platform::errors::PermissionDenied(
                 "Cannot use CUDAPlace in CPU only version, "
                 "Please recompile or reinstall Paddle with CUDA support."));
Q
qijun 已提交
1372
#else
Q
qijun 已提交
1373
                    return new paddle::platform::CUDADeviceContext(place);
Q
qijun 已提交
1374
#endif
C
chengduoZH 已提交
1375 1376 1377 1378
                  })
          .def_static("create",
                [](paddle::platform::CUDAPinnedPlace& place)
                        -> paddle::platform::DeviceContext* {
1379
#if !defined(PADDLE_WITH_CUDA) && !defined(PADDLE_WITH_HIP)
1380 1381 1382 1383
             PADDLE_THROW(
                 platform::errors::PermissionDenied(
                 "Cannot use CUDAPinnedPlace in CPU only version, "
                 "Please recompile or reinstall Paddle with CUDA support."));
C
chengduoZH 已提交
1384 1385 1386 1387
#else
                  return new paddle::platform::CUDAPinnedDeviceContext(place);
#endif
                });;
D
Dong Zhihong 已提交
1388
// clang-format on
1389
#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
D
Dong Zhihong 已提交
1390 1391
  py::class_<platform::Communicator>(m, "Communicator").def(py::init<>());
#endif
1392
  py::class_<platform::CUDAPlace>(m, "CUDAPlace", R"DOC(
1393 1394 1395 1396 1397

    CUDAPlace is a descriptor of a device.
    It represents a GPU device allocated or to be allocated with Tensor or LoDTensor.
    Each CUDAPlace has a dev_id to indicate the graphics card ID represented by the current CUDAPlace,
    staring from 0.
1398
    The memory of CUDAPlace with different dev_id is not accessible.
1399 1400 1401 1402 1403 1404 1405 1406
    Numbering here refers to the logical ID of the visible graphics card, not the actual ID of the graphics card.
    You can set visible GPU devices by setting the `CUDA_VISIBLE_DEVICES` environment variable.
    When the program starts, visible GPU devices will be numbered from 0.
    If `CUDA_VISIBLE_DEVICES` is not set, all devices are visible by default,
    and the logical ID is the same as the actual ID.

    Parameters:
        id (int): GPU device ID.
L
lujun 已提交
1407 1408 1409 1410

    Examples:
        .. code-block:: python

1411 1412 1413
          import paddle

          place = paddle.CUDAPlace(0)
L
lujun 已提交
1414

1415
        )DOC")
S
sneaxiy 已提交
1416 1417
      .def("__init__",
           [](platform::CUDAPlace &self, int dev_id) {
1418
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442
             if (UNLIKELY(dev_id < 0)) {
               LOG(ERROR) << string::Sprintf(
                   "Invalid CUDAPlace(%d), device id must be 0 or "
                   "positive integer",
                   dev_id);
               std::exit(-1);
             }

             if (UNLIKELY(dev_id >= platform::GetCUDADeviceCount())) {
               if (platform::GetCUDADeviceCount() == 0) {
                 LOG(ERROR) << "Cannot use GPU because there is no GPU "
                               "detected on your "
                               "machine.";
                 std::exit(-1);
               } else {
                 LOG(ERROR) << string::Sprintf(
                     "Invalid CUDAPlace(%d), must inside [0, %d), because GPU "
                     "number on your machine is %d",
                     dev_id, platform::GetCUDADeviceCount(),
                     platform::GetCUDADeviceCount());
                 std::exit(-1);
               }
             }

S
sneaxiy 已提交
1443 1444
             new (&self) platform::CUDAPlace(dev_id);
#else
1445 1446 1447 1448 1449 1450 1451 1452 1453
             LOG(ERROR) << string::Sprintf(
                 "Cannot use GPU because you have installed CPU version "
                 "PaddlePaddle.\n"
                 "If you want to use GPU, please try to install GPU version "
                 "PaddlePaddle by: pip install paddlepaddle-gpu\n"
                 "If you only have CPU, please change CUDAPlace(%d) to be "
                 "CPUPlace().\n",
                 dev_id);
             std::exit(-1);
S
sneaxiy 已提交
1454 1455
#endif
           })
1456
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
1457 1458
      .def("get_device_id",
           [](const platform::CUDAPlace &self) { return self.GetDeviceId(); })
S
sneaxiy 已提交
1459 1460 1461 1462
      .def("_type", &PlaceIndex<platform::CUDAPlace>)
      .def("_equals", &IsSamePlace<platform::CUDAPlace, platform::Place>)
      .def("_equals", &IsSamePlace<platform::CUDAPlace, platform::CUDAPlace>)
      .def("_equals", &IsSamePlace<platform::CUDAPlace, platform::CPUPlace>)
1463
      .def("_equals", &IsSamePlace<platform::CUDAPlace, platform::XPUPlace>)
S
sneaxiy 已提交
1464 1465
      .def("_equals",
           &IsSamePlace<platform::CUDAPlace, platform::CUDAPinnedPlace>)
1466 1467 1468
      .def("_get_device_id",
           [](platform::CUDAPlace &self) -> int { return self.GetDeviceId(); })
#endif
1469
      .def("__repr__", string::to_string<const platform::CUDAPlace &>)
D
dzhwinter 已提交
1470
      .def("__str__", string::to_string<const platform::CUDAPlace &>);
Q
qijun 已提交
1471

1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516
  py::class_<platform::XPUPlace>(m, "XPUPlace", R"DOC(
    **Note**:
    Examples:
        .. code-block:: python
          import paddle.fluid as fluid
          xpu_place = fluid.XPUPlace(0)
        )DOC")
      .def("__init__",
           [](platform::XPUPlace &self, int dev_id) {
#ifdef PADDLE_WITH_XPU
             if (UNLIKELY(dev_id < 0)) {
               LOG(ERROR) << string::Sprintf(
                   "Invalid XPUPlace(%d), device id must be 0 or "
                   "positive integer",
                   dev_id);
               std::exit(-1);
             }
             if (UNLIKELY(dev_id >= platform::GetXPUDeviceCount())) {
               if (platform::GetXPUDeviceCount() == 0) {
                 LOG(ERROR) << "Cannot use XPU because there is no XPU "
                               "detected on your "
                               "machine.";
                 std::exit(-1);
               } else {
                 LOG(ERROR) << string::Sprintf(
                     "Invalid XPUPlace(%d), must inside [0, %d), because XPU "
                     "number on your machine is %d",
                     dev_id, platform::GetXPUDeviceCount(),
                     platform::GetXPUDeviceCount());
                 std::exit(-1);
               }
             }
             new (&self) platform::XPUPlace(dev_id);
#else
             LOG(ERROR) << string::Sprintf(
                 "Cannot use XPU because you have installed CPU/GPU version "
                 "PaddlePaddle.\n"
                 "If you want to use XPU, please try to install XPU version "
                 "PaddlePaddle by: pip install paddlepaddle-xpu\n"
                 "If you only have CPU, please change XPUPlace(%d) to be "
                 "CPUPlace().\n",
                 dev_id);
             std::exit(-1);
#endif
           })
1517
#ifdef PADDLE_WITH_XPU
1518 1519 1520 1521 1522 1523 1524
      .def("_type", &PlaceIndex<platform::XPUPlace>)
      .def("_equals", &IsSamePlace<platform::XPUPlace, platform::Place>)
      .def("_equals", &IsSamePlace<platform::XPUPlace, platform::CUDAPlace>)
      .def("_equals", &IsSamePlace<platform::XPUPlace, platform::CPUPlace>)
      .def("_equals", &IsSamePlace<platform::XPUPlace, platform::XPUPlace>)
      .def("_equals",
           &IsSamePlace<platform::XPUPlace, platform::CUDAPinnedPlace>)
1525 1526 1527
      .def("get_device_id",
           [](const platform::XPUPlace &self) { return self.GetDeviceId(); })
#endif
1528
      .def("__repr__", string::to_string<const platform::XPUPlace &>)
1529
      .def("__str__", string::to_string<const platform::XPUPlace &>);
1530 1531 1532
#ifdef PADDLE_WITH_XPU
  m.def("get_xpu_device_count", platform::GetXPUDeviceCount);
#endif
1533
  py::class_<paddle::platform::CPUPlace>(m, "CPUPlace", R"DOC(
1534
    CPUPlace is a descriptor of a device.
1535
    It represents a CPU device on which a tensor will be allocated and a model will run.
L
lujun 已提交
1536 1537 1538 1539

    Examples:
        .. code-block:: python

1540 1541
          import paddle
          cpu_place = paddle.CPUPlace()
L
lujun 已提交
1542

1543
        )DOC")
1544
      .def(py::init<>())
S
sneaxiy 已提交
1545 1546
      .def("_type", &PlaceIndex<platform::CPUPlace>)
      .def("_equals", &IsSamePlace<platform::CPUPlace, platform::Place>)
1547
      .def("_equals", &IsSamePlace<platform::CPUPlace, platform::XPUPlace>)
S
sneaxiy 已提交
1548 1549 1550 1551
      .def("_equals", &IsSamePlace<platform::CPUPlace, platform::CUDAPlace>)
      .def("_equals", &IsSamePlace<platform::CPUPlace, platform::CPUPlace>)
      .def("_equals",
           &IsSamePlace<platform::CPUPlace, platform::CUDAPinnedPlace>)
1552
      .def("__repr__", string::to_string<const platform::CPUPlace &>)
1553
      .def("__str__", string::to_string<const platform::CPUPlace &>);
Y
Yu Yang 已提交
1554

1555
  py::class_<paddle::platform::CUDAPinnedPlace>(m, "CUDAPinnedPlace", R"DOC(
1556 1557 1558 1559 1560 1561
    CUDAPinnedPlace is a descriptor of a device.
    It refers to the page locked memory allocated by the CUDA function `cudaHostAlloc()` in the host memory.
    The host operating system will not paging and exchanging the memory.
    It can be accessed through direct memory access technology to speed up the copy of data between the host and GPU.
    For more information on CUDA data transfer and `pinned memory`,
    please refer to `official document <https://docs.nvidia.com/cuda/cuda-c-best-practices-guide/index.html#pinned-memory>`_ .
L
lujun 已提交
1562 1563 1564 1565

    Examples:
        .. code-block:: python

1566 1567
          import paddle
          place = paddle.CUDAPinnedPlace()
L
lujun 已提交
1568

1569
        )DOC")
S
sneaxiy 已提交
1570
      .def("__init__",
S
sneaxiy 已提交
1571
           [](platform::CUDAPinnedPlace &self) {
1572
#if !defined(PADDLE_WITH_CUDA) && !defined(PADDLE_WITH_HIP)
1573 1574 1575
             PADDLE_THROW(platform::errors::PermissionDenied(
                 "Cannot use CUDAPinnedPlace in CPU only version, "
                 "Please recompile or reinstall Paddle with CUDA support."));
S
sneaxiy 已提交
1576
#endif
S
sneaxiy 已提交
1577
             new (&self) platform::CUDAPinnedPlace();
S
sneaxiy 已提交
1578
           })
S
sneaxiy 已提交
1579 1580 1581 1582
      .def("_type", &PlaceIndex<platform::CUDAPinnedPlace>)
      .def("_equals", &IsSamePlace<platform::CUDAPinnedPlace, platform::Place>)
      .def("_equals",
           &IsSamePlace<platform::CUDAPinnedPlace, platform::CUDAPlace>)
1583 1584
      .def("_equals",
           &IsSamePlace<platform::CUDAPinnedPlace, platform::XPUPlace>)
S
sneaxiy 已提交
1585 1586 1587 1588
      .def("_equals",
           &IsSamePlace<platform::CUDAPinnedPlace, platform::CPUPlace>)
      .def("_equals",
           &IsSamePlace<platform::CUDAPinnedPlace, platform::CUDAPinnedPlace>)
1589
      .def("__repr__", string::to_string<const platform::CUDAPinnedPlace &>)
C
chengduoZH 已提交
1590 1591
      .def("__str__", string::to_string<const platform::CUDAPinnedPlace &>);

Y
Yu Yang 已提交
1592 1593
  py::class_<platform::Place>(m, "Place")
      .def(py::init<>())
S
sneaxiy 已提交
1594 1595 1596 1597
      .def("_type", &PlaceIndex<platform::Place>)
      .def("_equals", &IsSamePlace<platform::Place, platform::Place>)
      .def("_equals", &IsSamePlace<platform::Place, platform::CUDAPlace>)
      .def("_equals", &IsSamePlace<platform::Place, platform::CPUPlace>)
1598
      .def("_equals", &IsSamePlace<platform::Place, platform::XPUPlace>)
S
sneaxiy 已提交
1599
      .def("_equals", &IsSamePlace<platform::Place, platform::CUDAPinnedPlace>)
X
xuezhong 已提交
1600 1601
      .def("is_gpu_place",
           [](platform::Place &self) { return platform::is_gpu_place(self); })
S
sneaxiy 已提交
1602 1603
      .def("is_cpu_place",
           [](platform::Place &self) { return platform::is_cpu_place(self); })
1604 1605
      .def("is_xpu_place",
           [](platform::Place &self) { return platform::is_xpu_place(self); })
S
sneaxiy 已提交
1606 1607 1608 1609
      .def("is_cuda_pinned_place",
           [](platform::Place &self) {
             return platform::is_cuda_pinned_place(self);
           })
X
xuezhong 已提交
1610 1611
      .def("gpu_device_id",
           [](platform::Place &self) {
1612
             return BOOST_GET_CONST(platform::CUDAPlace, self).device;
X
xuezhong 已提交
1613
           })
1614 1615 1616 1617
      .def("xpu_device_id",
           [](platform::Place &self) {
             return BOOST_GET_CONST(platform::XPUPlace, self).device;
           })
S
sneaxiy 已提交
1618 1619
      .def("set_place", [](platform::Place &self,
                           const platform::Place &other) { self = other; })
Y
Yu Yang 已提交
1620 1621 1622 1623
      .def("set_place",
           [](platform::Place &self, const platform::CPUPlace &cpu_place) {
             self = cpu_place;
           })
1624 1625 1626 1627
      .def("set_place",
           [](platform::Place &self, const platform::XPUPlace &xpu_place) {
             self = xpu_place;
           })
Y
Yu Yang 已提交
1628
      .def("set_place",
D
dzhwinter 已提交
1629
           [](platform::Place &self, const platform::CUDAPlace &gpu_place) {
Y
Yu Yang 已提交
1630
             self = gpu_place;
C
chengduoZH 已提交
1631
           })
1632 1633 1634 1635 1636 1637 1638
      .def("set_place",
           [](platform::Place &self,
              const platform::CUDAPinnedPlace &cuda_pinned_place) {
             self = cuda_pinned_place;
           })
      .def("__repr__", string::to_string<const platform::Place &>)
      .def("__str__", string::to_string<const platform::Place &>);
Y
Yu Yang 已提交
1639

Y
Yu Yang 已提交
1640
  py::class_<OperatorBase>(m, "Operator")
C
chengduo 已提交
1641 1642 1643 1644 1645
      .def_static(
          "create",
          [](py::bytes protobin) {
            proto::OpDesc desc;
            PADDLE_ENFORCE_EQ(desc.ParsePartialFromString(protobin), true,
1646 1647 1648 1649 1650 1651 1652
                              platform::errors::InvalidArgument(
                                  "Cannot parse user input to OpDesc"));
            PADDLE_ENFORCE_EQ(
                desc.IsInitialized(), true,
                platform::errors::InvalidArgument(
                    "The provided OpDesc is not initialized, the reason is: %s",
                    desc.InitializationErrorString()));
C
chengduo 已提交
1653 1654
            return OpRegistry::CreateOp(desc);
          })
1655
      .def("run",
1656
           [](OperatorBase &self, const Scope &scope,
D
dzhwinter 已提交
1657
              const platform::CPUPlace &place) { self.Run(scope, place); })
1658 1659 1660
      .def("run",
           [](OperatorBase &self, const Scope &scope,
              const platform::XPUPlace &place) { self.Run(scope, place); })
D
dzhwinter 已提交
1661 1662
      .def("run",
           [](OperatorBase &self, const Scope &scope,
D
dzhwinter 已提交
1663
              const platform::CUDAPlace &place) { self.Run(scope, place); })
C
chengduoZH 已提交
1664 1665 1666 1667 1668
      .def("run",
           [](OperatorBase &self, const Scope &scope,
              const platform::CUDAPinnedPlace &place) {
             self.Run(scope, place);
           })
Y
Yu Yang 已提交
1669 1670 1671 1672 1673 1674 1675
      .def("type",
           [](const OperatorBase &op) -> std::string { return op.Type(); })
      .def("outputs",
           [](const OperatorBase &op)
               -> std::map<std::string, std::vector<std::string>> {
                 return op.Outputs();
               })
Q
qijun 已提交
1676 1677
      .def("output_vars",
           [](const OperatorBase &op) { return op.OutputVars(true); })
Y
Yu Yang 已提交
1678
      .def("inputs", [](const OperatorBase &op) { return op.Inputs(); })
Q
qijun 已提交
1679
      .def("input_vars", [](const OperatorBase &op) { return op.InputVars(); })
Y
Yu Yang 已提交
1680 1681 1682 1683
      .def("__str__", &OperatorBase::DebugString)
      .def("no_intermediate_outputs",
           [](const OperatorBase &op) { return op.OutputVars(false); })
      .def("support_gpu", &OperatorBase::SupportGPU);
Y
Yu Yang 已提交
1684

1685 1686 1687
  py::class_<framework::ExecutorPrepareContext>(m, "ExecutorPrepareContext")
      .def(py::init<const ProgramDesc &, size_t>());

1688 1689 1690 1691 1692 1693 1694 1695 1696
  py::class_<framework::TrainerBase, std::shared_ptr<framework::TrainerBase>>(
      m, "TrainerBase")
      .def("get_worker_scope",
           [](TrainerBase &self, int thread_id) -> Scope * {
             return self.GetWorkerScope(thread_id);
           },
           py::return_value_policy::reference)
      .def("finalize", &TrainerBase::Finalize);

F
fengjiayi 已提交
1697
  py::class_<framework::Executor>(m, "Executor")
D
dzhwinter 已提交
1698
      .def(py::init<const platform::Place &>())
Y
Yancey1989 已提交
1699
      .def("close", &Executor::Close)
1700 1701
      .def("run_from_dataset", &Executor::RunFromDataset,
           py::call_guard<py::gil_scoped_release>())
D
Dong Daxiang 已提交
1702 1703
      .def("release_trainer", &Executor::ReleaseTrainer,
           py::call_guard<py::gil_scoped_release>())
1704 1705 1706 1707
      .def("init_for_dataset",
           [](Executor &self, const ProgramDesc &prog,
              const std::string &trainer_desc, Scope *scope,
              Dataset *dataset) -> std::shared_ptr<TrainerBase> {
D
Dong Daxiang 已提交
1708
             pybind11::gil_scoped_release release;
1709 1710 1711 1712 1713 1714 1715
             return self.InitForDataset(prog, trainer_desc, scope, dataset);
           })
      .def("run_from_dataset",
           [](Executor &self, std::shared_ptr<TrainerBase> trainer) {
             pybind11::gil_scoped_release release;
             self.RunFromDataset(trainer);
           })
1716 1717 1718
      .def("run_prepared_ctx",
           [](Executor &self, ExecutorPrepareContext *ctx, Scope *scope,
              std::map<std::string, const LoDTensor *> *feed_targets,
1719
              std::map<std::string, FetchType *> *fetch_targets,
1720 1721 1722 1723 1724 1725 1726 1727
              bool create_local_scope = true, bool create_vars = true,
              const std::string &feed_holder_name = "feed",
              const std::string &fetch_holder_name = "fetch") {
             pybind11::gil_scoped_release release;
             self.RunPreparedContext(ctx, scope, feed_targets, fetch_targets,
                                     create_local_scope, create_vars,
                                     feed_holder_name, fetch_holder_name);
           })
1728
      .def("run_prepared_ctx",
G
guru4elephant 已提交
1729 1730 1731 1732 1733 1734 1735
           [](Executor &self, ExecutorPrepareContext *ctx, Scope *scope,
              bool create_local_scope = true, bool create_vars = true,
              bool keep_kids = false) {
             pybind11::gil_scoped_release release;
             self.RunPreparedContext(ctx, scope, create_local_scope,
                                     create_vars, keep_kids);
           })
1736 1737 1738 1739 1740 1741 1742 1743 1744 1745
      .def("prepare",
           [](Executor &self, const ProgramDesc &program, int block_id,
              const std::vector<std::string> &skip_ref_cnt_vars =
                  std::vector<std::string>(),
              bool force_disable_gc = false) {
             pybind11::gil_scoped_release release;
             return self.Prepare(program, block_id, skip_ref_cnt_vars,
                                 force_disable_gc);
           })
      .def("create_variables", &Executor::CreateVariables)
S
sneaxiy 已提交
1746
      .def("run", [](Executor &self, const ProgramDesc &prog, Scope *scope,
S
sneaxiy 已提交
1747 1748
                     int block_id, bool create_local_scope, bool create_vars,
                     const std::vector<std::string> &fetch_vars) {
S
sneaxiy 已提交
1749
        pybind11::gil_scoped_release release;
S
sneaxiy 已提交
1750 1751
        self.Run(prog, scope, block_id, create_local_scope, create_vars,
                 fetch_vars);
S
sneaxiy 已提交
1752
      });
S
sneaxiy 已提交
1753

D
dzhwinter 已提交
1754
  m.def("init_gflags", framework::InitGflags);
Y
Yang Yu 已提交
1755
  m.def("init_glog", framework::InitGLOG);
1756
  m.def("load_op_library", framework::LoadOpLib);
1757 1758
  m.def("load_op_meta_info_and_register_op",
        framework::LoadOpMetaInfoAndRegisterOp);
1759
  m.def("init_devices", []() { framework::InitDevices(); });
1760

1761
  m.def("is_compiled_with_cuda", IsCompiledWithCUDA);
1762
  m.def("is_compiled_with_rocm", IsCompiledWithROCM);
1763
  m.def("is_compiled_with_xpu", IsCompiledWithXPU);
1764
  m.def("is_compiled_with_mkldnn", IsCompiledWithMKLDNN);
1765
  m.def("supports_bfloat16", SupportsBfloat16);
1766
  m.def("supports_bfloat16_fast_performance", SupportsBfloat16FastPerformance);
1767
  m.def("is_compiled_with_brpc", IsCompiledWithBrpc);
Y
update  
Yancey1989 已提交
1768
  m.def("is_compiled_with_dist", IsCompiledWithDIST);
1769 1770 1771
  m.def("_cuda_synchronize", [](const platform::CUDAPlace &place) {
    platform::DeviceContextPool::Instance().Get(place)->Wait();
  });
H
hutuxian 已提交
1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790

  m.def("get_float_stats", []() {
    std::vector<paddle::platform::ExportedStatValue<float>> float_stats;
    paddle::platform::StatRegistry<float>::Instance().publish(float_stats);
    std::unordered_map<std::string, float> stats_map;
    for (const auto &stat : float_stats) {
      stats_map[stat.key] = stat.value;
    }
    return stats_map;
  });
  m.def("get_int_stats", []() {
    std::vector<paddle::platform::ExportedStatValue<int64_t>> int_stats;
    paddle::platform::StatRegistry<int64_t>::Instance().publish(int_stats);
    std::unordered_map<std::string, int64_t> stats_map;
    for (const auto &stat : int_stats) {
      stats_map[stat.key] = stat.value;
    }
    return stats_map;
  });
H
hutuxian 已提交
1791 1792 1793 1794 1795 1796 1797
  m.def("run_cmd",
        [](const std::string &cmd, int time_out = -1,
           int sleep_inter = -1) -> const std::string {
          return paddle::framework::shell_get_command_output(cmd, time_out,
                                                             sleep_inter);
        },
        py::arg("cmd"), py::arg("time_out") = -1, py::arg("sleep_inter") = -1);
G
gongweibao 已提交
1798 1799 1800 1801 1802 1803 1804 1805 1806
  m.def("shell_execute_cmd",
        [](const std::string &cmd, int time_out = 0, int sleep_inter = 0,
           bool redirect_stderr = false) -> std::vector<std::string> {
          return paddle::framework::shell_execute_cmd(
              cmd, time_out, sleep_inter, redirect_stderr);
        },
        py::arg("cmd"), py::arg("time_out") = 0, py::arg("sleep_inter") = 0,
        py::arg("redirect_stderr") = false);

1807
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
1808 1809 1810 1811 1812
  m.def("is_float16_supported", [](const platform::CUDAPlace &place) -> bool {
    // Only GPUs with Compute Capability >= 53 support float16
    return platform::GetCUDAComputeCapability(place.device) >= 53;
  });
#endif
1813

1814
  m.def("set_feed_variable", framework::SetFeedVariable);
1815 1816 1817 1818 1819
  m.def("get_fetch_variable",
        [](const Scope &scope, const std::string &var_name,
           size_t index) -> py::object {
          auto &var = framework::GetFetchVariable(scope, var_name, index);
          if (data_is_lod_tensor(var)) {
1820
            return py::cast(BOOST_GET(LoDTensor, var));
1821
          } else {
1822
            return py::cast(BOOST_GET(LoDTensorArray, var));
1823 1824
          }
        });
1825
  m.def("get_variable_tensor", framework::GetVariableTensor);
Q
qijun 已提交
1826

X
Xin Pan 已提交
1827 1828
  m.def("_is_program_version_supported", IsProgramVersionSupported);

1829 1830 1831 1832 1833
  BindProgramDesc(&m);
  BindBlockDesc(&m);
  BindVarDsec(&m);
  BindOpDesc(&m);
  BindConstValue(&m);
1834
  BindGlobalValueGetterSetter(&m);
Y
Yu Yang 已提交
1835

Y
Yu Yang 已提交
1836 1837 1838 1839 1840 1841 1842 1843 1844
  py::class_<framework::LoDRankTable>(m, "LodRankTable")
      .def("items", [](framework::LoDRankTable &table) {
        std::vector<std::pair<size_t, size_t>> res;
        for (auto &item : table.items()) {
          res.push_back({item.index, item.length});
        }
        return res;
      });

Z
Zeng Jinle 已提交
1845
  py::class_<LoDTensorArray>(m, "LoDTensorArray", R"DOC(
1846
    LoDTensorArray is array of LoDTensor, it supports operator[], len() and for-loop iteration.
Z
Zeng Jinle 已提交
1847 1848 1849

    Examples:
        .. code-block:: python
1850

Z
Zeng Jinle 已提交
1851 1852 1853 1854
          import paddle.fluid as fluid

          arr = fluid.LoDTensorArray()
)DOC")
S
sneaxiy 已提交
1855 1856
      .def("__init__",
           [](LoDTensorArray &instance) { new (&instance) LoDTensorArray(); })
Y
Yu Yang 已提交
1857 1858 1859 1860 1861 1862
      .def("__getitem__",
           [](LoDTensorArray &self, size_t i) { return &self.at(i); },
           py::return_value_policy::reference)
      .def("__len__", [](LoDTensorArray &self) { return self.size(); })
      .def("__setitem__",
           [](LoDTensorArray &self, size_t i, const LoDTensor &t) {
1863 1864 1865 1866
             PADDLE_ENFORCE_LT(i, self.size(),
                               platform::errors::InvalidArgument(
                                   "The index to set is larger than the size "
                                   "of LoDTensorArray."));
Y
Yu Yang 已提交
1867 1868 1869
             self[i].ShareDataWith(t);
             self[i].set_lod(t.lod());
           })
S
sneaxiy 已提交
1870 1871 1872 1873 1874 1875
      .def("append",
           [](LoDTensorArray &self, const LoDTensor &t) {
             self.emplace_back();
             self.back().ShareDataWith(t);
             self.back().set_lod(t.lod());
           },
Z
Zeng Jinle 已提交
1876 1877
           py::arg("tensor"), R"DOC(
             Append a LoDensor to LoDTensorArray.
1878 1879 1880 1881 1882 1883
              
             Args:
                   tensor (LoDTensor): The LoDTensor to be appended.

             Returns:
                   None.
Z
Zeng Jinle 已提交
1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894

             Examples:
                 .. code-block:: python

                   import paddle.fluid as fluid
                   import numpy as np

                   arr = fluid.LoDTensorArray()
                   t = fluid.LoDTensor()
                   t.set(np.ndarray([5, 30]), fluid.CPUPlace())
                   arr.append(t)
1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905
           )DOC")
      .def("_move_to_list",
           [](LoDTensorArray &self) -> py::list {
             py::list res(self.size());
             for (size_t i = 0; i < self.size(); ++i) {
               res[i] = py::cast(std::move(self[i]));
             }
             self.clear();
             return res;
           },
           py::return_value_policy::take_ownership);
Y
Yu Yang 已提交
1906

1907 1908 1909 1910 1911 1912 1913 1914
  py::class_<FetchList>(m, "FetchList", R"DOC( FetchList is a
        vector of boost::variant<LoDTensor, LoDTensorArray>.
        )DOC")
      .def("_move_to_list",
           [](FetchList &self) -> py::list {
             py::list res(self.size());
             for (size_t i = 0; i < self.size(); ++i) {
               if (data_is_lod_tensor(self[i])) {
1915
                 auto &data = BOOST_GET(LoDTensor, self[i]);
1916 1917
                 res[i] = py::cast(std::move(data));
               } else {
1918
                 auto &data = BOOST_GET(LoDTensorArray, self[i]);
1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933
                 py::list tmp(data.size());
                 for (size_t j = 0; j < data.size(); ++j) {
                   tmp[j] = py::cast(std::move(data[j]));
                 }
                 res[i] = std::move(tmp);
               }
             }
             self.clear();
             return res;
           },
           py::return_value_policy::take_ownership)

      .def("append",
           [](FetchList &self, const LoDTensor &t) {
             self.emplace_back();
1934
             auto &lod_tensor = BOOST_GET(LoDTensor, self.back());
1935 1936 1937 1938 1939 1940 1941 1942
             lod_tensor.ShareDataWith(t);
             lod_tensor.set_lod(t.lod());
           },
           py::arg("var"))

      .def("append",
           [](FetchList &self, const LoDTensorArray &t) {
             self.emplace_back();
1943
             auto &lod_tensor_array = BOOST_GET(LoDTensorArray, self.back());
1944 1945 1946 1947 1948 1949 1950 1951 1952
             for (size_t i = 0; i < t.size(); ++i) {
               lod_tensor_array[i].ShareDataWith(t[i]);
               lod_tensor_array[i].set_lod(t[i].lod());
             }
           },
           py::arg("var"));

  py::class_<FetchUnmergedList>(m, "FetchUnmergedList", R"DOC(
        FetchUnmergedList is 2-D array of FetchType(boost::variant(LoDTensor, LoDTensorArray)).
Z
Zhen Wang 已提交
1953 1954
        )DOC")
      .def("_move_to_list",
1955
           [](FetchUnmergedList &self) -> py::list {
Z
Zhen Wang 已提交
1956 1957 1958 1959
             py::list res(self.size());
             for (size_t i = 0; i < self.size(); ++i) {
               py::list tmp(self[i].size());
               for (size_t j = 0; j < self[i].size(); ++j) {
1960
                 if (data_is_lod_tensor(self[i][j])) {
1961
                   auto &var = BOOST_GET(LoDTensor, self[i][j]);
1962 1963
                   tmp[j] = py::cast(std::move(var));
                 } else {
1964
                   auto &var = BOOST_GET(LoDTensorArray, self[i][j]);
1965 1966 1967 1968 1969 1970
                   py::list tmp_array(var.size());
                   for (size_t k = 0; k < var.size(); ++k) {
                     tmp_array[k] = std::move(var[k]);
                   }
                   tmp[j] = std::move(tmp_array);
                 }
Z
Zhen Wang 已提交
1971 1972 1973 1974 1975 1976 1977 1978 1979
               }
               res[i] = std::move(tmp);
               self[i].clear();
             }
             self.clear();
             return res;
           },
           py::return_value_policy::take_ownership);

Y
Yu Yang 已提交
1980
  m.def("op_support_gpu", OpSupportGPU);
1981
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
D
Dong Zhihong 已提交
1982
  m.def("get_cuda_device_count", platform::GetCUDADeviceCount);
D
dangqingqing 已提交
1983

1984
#if !defined(PADDLE_WITH_HIP) && !defined(_WIN32)
D
dangqingqing 已提交
1985 1986 1987
  m.def("nvprof_init", platform::CudaProfilerInit);
  m.def("nvprof_start", platform::CudaProfilerStart);
  m.def("nvprof_stop", platform::CudaProfilerStop);
1988 1989 1990 1991
  m.def("nvprof_nvtx_push", platform::CudaNvtxRangePush);
  m.def("nvprof_nvtx_pop", platform::CudaNvtxRangePop);
  m.def("nvprof_enable_record_event", platform::NvprofEnableRecordEvent);
  m.def("nvprof_disable_record_event", platform::NvprofDisableRecordEvent);
D
Dong Zhihong 已提交
1992
#endif
P
peizhilin 已提交
1993
#endif
Y
Yu Yang 已提交
1994

1995 1996 1997 1998 1999 2000
  py::enum_<platform::TracerOption>(m, "TracerOption", py::arithmetic())
      .value("kDefault", platform::TracerOption::kDefault)
      .value("kOpDetail", platform::TracerOption::kOpDetail)
      .value("kAllOpDetail", platform::TracerOption::kAllOpDetail)
      .export_values();

2001 2002 2003 2004
  py::enum_<platform::ProfilerState>(m, "ProfilerState", py::arithmetic())
      .value("kDisabled", platform::ProfilerState::kDisabled)
      .value("kCPU", platform::ProfilerState::kCPU)
      .value("kCUDA", platform::ProfilerState::kCUDA)
2005
      .value("kAll", platform::ProfilerState::kAll)
2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016
      .export_values();

  py::enum_<platform::EventSortingKey>(m, "EventSortingKey", py::arithmetic())
      .value("kDefault", platform::EventSortingKey::kDefault)
      .value("kCalls", platform::EventSortingKey::kCalls)
      .value("kTotal", platform::EventSortingKey::kTotal)
      .value("kMin", platform::EventSortingKey::kMin)
      .value("kMax", platform::EventSortingKey::kMax)
      .value("kAve", platform::EventSortingKey::kAve)
      .export_values();

2017
  m.def("set_tracer_option", platform::SetTracerOption);
2018 2019
  m.def("enable_profiler", platform::EnableProfiler);
  m.def("disable_profiler", platform::DisableProfiler);
X
Xin Pan 已提交
2020
  m.def("is_profiler_enabled", platform::IsProfileEnabled);
2021
  m.def("reset_profiler", platform::ResetProfiler);
2022
  m.def("get_pass", [](const std::string &pass_type) {
W
WangZhen 已提交
2023 2024 2025
    auto pass = framework::ir::PassRegistry::Instance().Get(pass_type);
    return std::shared_ptr<framework::ir::Pass>(std::move(pass));
  });
Y
Yu Yang 已提交
2026

2027 2028
  m.def("size_of_dtype", framework::SizeOfType);

2029
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
2030 2031
  m.def("set_cublas_switch", platform::SetAllowTF32Cublas);
  m.def("get_cublas_switch", platform::AllowTF32Cublas);
A
AshburnLee 已提交
2032 2033
  m.def("set_cudnn_switch", platform::SetAllowTF32Cudnn);
  m.def("get_cudnn_switch", platform::AllowTF32Cudnn);
2034 2035
#endif  // PADDLE_WITH_CUDA

2036 2037 2038
  using VarQuantScale =
      std::unordered_map<std::string, std::pair<bool, LoDTensor>>;

2039 2040
  py::class_<ir::Pass, std::shared_ptr<ir::Pass>> pass(m, "Pass");
  pass.def(py::init())
W
WangZhen 已提交
2041
      .def("has", &ir::Pass::Has)
2042 2043 2044
      .def("set_not_owned",
           [](ir::Pass &self, const std::string &attr_name, ProgramDesc &attr) {
             self.SetNotOwned<ProgramDesc>(attr_name, &attr);
W
WangZhen 已提交
2045
           })
2046
      .def(
2047
          "set",
2048 2049 2050
          [](ir::Pass &self, const std::string &name, const std::string &attr) {
            self.Set<std::string>(name, new std::string(attr));
          })
2051 2052
      .def("set", [](ir::Pass &self, const std::string &name,
                     bool val) { self.Set<bool>(name, new bool(val)); })
2053 2054
      .def("set", [](ir::Pass &self, const std::string &name,
                     int val) { self.Set<const int>(name, new int(val)); })
2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068
      .def("set",
           [](ir::Pass &self, const std::string &name,
              std::unordered_set<std::string> set) {
             self.Set(name, new std::unordered_set<std::string>(set));
           })
      .def("set",
           [](ir::Pass &self, const std::string &name,
              std::unordered_set<int> set) {
             self.Set(name, new std::unordered_set<int>(set));
           })
      .def("set",
           [](ir::Pass &self, const std::string &name, VarQuantScale scales) {
             self.Set(name, new VarQuantScale(scales));
           })
F
flame 已提交
2069 2070
      .def("type", &ir::Pass::Type)
      .def("apply", [](ir::Pass &self, std::shared_ptr<ir::Graph> graph) {
2071
        self.Apply(graph.get());
F
flame 已提交
2072
      });
2073

X
fix  
Xin Pan 已提交
2074 2075
  py::class_<ir::PassBuilder, std::shared_ptr<ir::PassBuilder>> pb(
      m, "PassBuilder");
2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089
  pb.def(py::init())
      .def("append_pass",
           [](ir::PassBuilder &self,
              const std::string &pass_type) -> std::shared_ptr<ir::Pass> {
             return self.AppendPass(pass_type);
           })
      .def("all_passes", [](ir::PassBuilder &self) { return self.AllPasses(); })
      .def("insert_pass",
           [](ir::PassBuilder &self, size_t idx, const std::string &pass_type) {
             return self.InsertPass(idx, pass_type);
           })
      .def("remove_pass",
           [](ir::PassBuilder &self, size_t idx) { self.RemovePass(idx); });

Y
yuyang18 已提交
2090
  // -- python binds for parallel executor.
X
Xin Pan 已提交
2091

Y
yuyang18 已提交
2092
  py::class_<ParallelExecutor> pe(m, "ParallelExecutor");
C
chengduo 已提交
2093 2094 2095 2096
  py::class_<ExecutionStrategy> exec_strategy(pe, "ExecutionStrategy", R"DOC(
    ExecutionStrategy allows the user to more preciously control how to run
    the program in ParallelExecutor by setting the property.

2097 2098 2099
    Returns:
        ExecutionStrategy: An ExecutionStrategy object.

C
chengduo 已提交
2100 2101 2102
    Examples:
        .. code-block:: python

2103 2104 2105 2106 2107 2108 2109 2110 2111
          import paddle
          import paddle.static as static
          import paddle.nn.functional as F

          paddle.enable_static()

          x = static.data(name='x', shape=[None, 13], dtype='float32')
          y = static.data(name='y', shape=[None, 1], dtype='float32')
          y_predict = static.nn.fc(input=x, size=1, act=None)
2112

2113 2114
          cost = F.square_error_cost(input=y_predict, label=y)
          avg_loss = paddle.mean(cost)
2115

2116
          sgd_optimizer = paddle.optimizer.SGD(learning_rate=0.001)
2117 2118
          sgd_optimizer.minimize(avg_loss)

2119
          exec_strategy = static.ExecutionStrategy()
C
chengduo 已提交
2120 2121
          exec_strategy.num_threads = 4

2122 2123 2124
          train_exe = static.ParallelExecutor(use_cuda=False,
                                              loss_name=avg_loss.name,
                                              exec_strategy=exec_strategy)
C
chengduo 已提交
2125 2126
        )DOC");

2127 2128 2129 2130
  py::enum_<paddle::platform::DeviceType>(m, "DeviceType", py::arithmetic())
      .value("CPU", paddle::platform::DeviceType::CPU)
      .value("CUDA", paddle::platform::DeviceType::CUDA)
      .value("XPU", paddle::platform::DeviceType::XPU);
2131

Y
yuyang18 已提交
2132
  exec_strategy.def(py::init())
Y
yuyang18 已提交
2133 2134 2135 2136 2137
      .def_property(
          "num_threads",
          [](const ExecutionStrategy &self) { return self.num_threads_; },
          [](ExecutionStrategy &self, size_t num_threads) {
            self.num_threads_ = num_threads;
C
chengduo 已提交
2138
          },
2139 2140
          R"DOC(
            The type is INT, num_threads represents the size of thread pool that
C
chengduo 已提交
2141 2142 2143 2144 2145 2146 2147
            used to run the operators of the current program in ParallelExecutor.
            If :math:`num\_threads=1`, all the operators will execute one by one,
            but the order maybe difference between iterations.
            If it is not set, it will be set in ParallelExecutor according to the
            device type and device count, for GPU, :math:`num\_threads=device\_count*4`, for CPU,
            :math:`num\_threads=CPU\_NUM*4`, the explanation of:math:`CPU\_NUM` is in ParallelExecutor.
            if it is not set, ParallelExecutor will get the cpu count by calling
2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160
            `multiprocessing.cpu_count()`. Default 0.

            Examples:
                .. code-block:: python

                    import paddle
                    import paddle.static as static

                    paddle.enable_static()

                    exec_strategy = static.ExecutionStrategy()
                    exec_strategy.num_threads = 4
            )DOC")
Y
yuyang18 已提交
2161
      .def_property(
2162 2163
          "_use_device",
          [](const ExecutionStrategy &self) { return self.use_device_; },
2164
          [](ExecutionStrategy &self, paddle::platform::DeviceType use_device) {
2165 2166 2167
            self.use_device_ = use_device;
          })  // NOTE(liuyuhui): Doesn't add doc for 'use_device', because
              // use_device isn‘t exposed to users.
Y
yuyang18 已提交
2168 2169 2170 2171 2172
      .def_property(
          "allow_op_delay",
          [](const ExecutionStrategy &self) { return self.allow_op_delay_; },
          [](ExecutionStrategy &self, bool allow_op_delay) {
            self.allow_op_delay_ = allow_op_delay;
C
chengduo 已提交
2173 2174 2175
          },
          R"DOC(The type is BOOL, allow_op_delay represents whether to delay the
                communication operators to run, it may make the execution faster.
2176 2177
                Note that this option is invalid now, and it will be removed in
                next version. Default False.)DOC")
Y
yuyang18 已提交
2178 2179 2180 2181 2182 2183 2184
      .def_property(
          "num_iteration_per_drop_scope",
          [](const ExecutionStrategy &self) {
            return self.num_iteration_per_drop_scope_;
          },
          [](ExecutionStrategy &self, size_t num_iteration_per_drop_scope) {
            self.num_iteration_per_drop_scope_ = num_iteration_per_drop_scope;
C
chengduo 已提交
2185 2186 2187 2188
          },
          R"DOC(The type is INT, num_iteration_per_drop_scope indicates how
                many iterations to clean up the temp variables which
                is generated during execution. It may make the execution faster,
2189
                because the temp variable's shape maybe the same between two iterations.
2190 2191 2192 2193 2194 2195 2196 2197 2198 2199
                Default 100.

                .. note::
                    1. If you fetch data when calling the 'run', the ParallelExecutor 
                    will clean up the temp variables at the end of the current iteration. 
                    2. In some NLP model, it may cause the GPU memory is insufficient, 
                    in this case, you should reduce `num_iteration_per_drop_scope`.

                Examples:
                    .. code-block:: python
C
chengduo 已提交
2200

2201 2202 2203 2204 2205 2206 2207
                        import paddle
                        import paddle.static as static

                        paddle.enable_static()

                        exec_strategy = static.ExecutionStrategy()
                        exec_strategy.num_iteration_per_drop_scope = 10
2208
              )DOC")
Q
Qiao Longfei 已提交
2209 2210 2211 2212 2213 2214 2215 2216 2217
      .def_property(
          "num_iteration_per_run",
          [](const ExecutionStrategy &self) {
            return self.num_iteration_per_run_;
          },
          [](ExecutionStrategy &self, size_t num_iteration_per_run) {
            self.num_iteration_per_run_ = num_iteration_per_run;
          },
          R"DOC(This config that how many iteration the executor will run when
2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229
                user call exe.run() in python。Default: 1.

                Examples:
                    .. code-block:: python

                        import paddle
                        import paddle.static as static

                        paddle.enable_static()

                        exec_strategy = static.ExecutionStrategy()
                        exec_strategy.num_iteration_per_run = 10
Q
Qiao Longfei 已提交
2230
              )DOC")
2231 2232 2233 2234 2235 2236 2237 2238
      .def_property(
          "use_thread_barrier",
          [](const ExecutionStrategy &self) { return self.thread_barrier_; },
          [](ExecutionStrategy &self, bool use_thread_barrier) {
            self.thread_barrier_ = use_thread_barrier;
          },
          R"DOC(This config that the this is distributed training with parameter server
              )DOC")
2239 2240 2241 2242 2243
      .def_property("_dry_run",
                    [](const ExecutionStrategy &self) { return self.dry_run_; },
                    [](ExecutionStrategy &self, bool dry_run) {
                      self.dry_run_ = dry_run;
                    });
C
chengduo 已提交
2244

Y
yuyang18 已提交
2245
  exec_strategy.def_property(
Y
yuyang18 已提交
2246 2247 2248 2249 2250 2251 2252
      "use_experimental_executor",
      [](const ExecutionStrategy &self) {
        return self.type_ == ExecutionStrategy::kExperimental;
      },
      [](ExecutionStrategy &self, bool experimental) {
        self.type_ = experimental ? ExecutionStrategy::kExperimental
                                  : ExecutionStrategy::kDefault;
Y
yuyang18 已提交
2253 2254
      });

C
chengduo 已提交
2255 2256 2257 2258
  py::class_<BuildStrategy> build_strategy(pe, "BuildStrategy", R"DOC(
    BuildStrategy allows the user to more preciously control how to
    build the SSA Graph in ParallelExecutor by setting the property.

2259 2260 2261
    Returns:
        BuildStrategy: An BuildStrategy object.

C
chengduo 已提交
2262 2263 2264
    Examples:
        .. code-block:: python

2265
            import os
2266 2267 2268 2269
            import paddle
            import paddle.static as static

            paddle.enable_static()
2270

2271 2272
            os.environ['CPU_NUM'] = str(2)
            places = static.cpu_places()
2273

2274 2275 2276 2277
            data = static.data(name="x", shape=[None, 1], dtype="float32")
            hidden = static.nn.fc(input=data, size=10)
            loss = paddle.mean(hidden)
            paddle.optimizer.SGD(learning_rate=0.01).minimize(loss)
2278

2279
            build_strategy = static.BuildStrategy()
2280 2281
            build_strategy.enable_inplace = True
            build_strategy.memory_optimize = True
2282 2283
            build_strategy.reduce_strategy = static.BuildStrategy.ReduceStrategy.Reduce
            program = static.CompiledProgram(static.default_main_program())
2284
            program = program.with_data_parallel(loss_name=loss.name,
2285 2286
                                                  build_strategy=build_strategy,
                                                  places=places)
C
chengduo 已提交
2287
)DOC");
Y
yuyang18 已提交
2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303

  py::enum_<BuildStrategy::ReduceStrategy>(build_strategy, "ReduceStrategy")
      .value("Reduce", BuildStrategy::ReduceStrategy::kReduce)
      .value("AllReduce", BuildStrategy::ReduceStrategy::kAllReduce);
  py::enum_<BuildStrategy::GradientScaleStrategy>(build_strategy,
                                                  "GradientScaleStrategy")
      .value("CoeffNumDevice",
             BuildStrategy::GradientScaleStrategy::kCoeffNumDevice)
      .value("One", BuildStrategy::GradientScaleStrategy::kOne)
      .value("Customized", BuildStrategy::GradientScaleStrategy::kCustomized);

  build_strategy.def(py::init())
      .def_property(
          "reduce_strategy",
          [](const BuildStrategy &self) { return self.reduce_; },
          [](BuildStrategy &self, BuildStrategy::ReduceStrategy strategy) {
2304 2305 2306 2307
            PADDLE_ENFORCE_NE(self.IsFinalized(), true,
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
Y
yuyang18 已提交
2308
            self.reduce_ = strategy;
C
chengduo 已提交
2309
          },
2310
          R"DOC((fluid.BuildStrategy.ReduceStrategy, optional): there are two reduce
C
chengduo 已提交
2311 2312
                strategies in ParallelExecutor, AllReduce and Reduce. If you want
                that all the parameters' optimization are done on all devices independently,
2313
                you should choose AllReduce; otherwise, if you choose Reduce, all the parameters'
C
chengduo 已提交
2314 2315
                optimization will be evenly distributed to different devices, and then
                broadcast the optimized parameter to other devices.
2316
                Default is 'AllReduce'.
F
flame 已提交
2317 2318 2319 2320

                Examples:
                    .. code-block:: python

2321 2322 2323 2324 2325 2326 2327
                        import paddle
                        import paddle.static as static

                        paddle.enable_static()

                        build_strategy = static.BuildStrategy()
                        build_strategy.reduce_strategy = static.BuildStrategy.ReduceStrategy.Reduce
F
flame 已提交
2328
                  )DOC")
Y
yuyang18 已提交
2329 2330 2331 2332 2333
      .def_property(
          "gradient_scale_strategy",
          [](const BuildStrategy &self) { return self.gradient_scale_; },
          [](BuildStrategy &self,
             BuildStrategy::GradientScaleStrategy strategy) {
2334 2335 2336 2337
            PADDLE_ENFORCE_NE(self.IsFinalized(), true,
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
Y
yuyang18 已提交
2338
            self.gradient_scale_ = strategy;
C
chengduo 已提交
2339
          },
2340
          R"DOC((paddle.static.BuildStrategy.GradientScaleStrategy, optional): there are three
2341
                ways of defining :math:`loss@grad` in ParallelExecutor, that is, CoeffNumDevice,
C
chengduo 已提交
2342 2343
                One and Customized. By default, ParallelExecutor sets the :math:`loss@grad`
                according to the number of devices. If you want to customize :math:`loss@grad`,
2344
                you can choose Customized. Default is 'CoeffNumDevice'.
F
flame 已提交
2345 2346 2347 2348

                Examples:
                    .. code-block:: python

C
chengduo 已提交
2349 2350
                        import numpy
                        import os
2351 2352 2353 2354
                        import paddle
                        import paddle.static as static

                        paddle.enable_static()
C
chengduo 已提交
2355 2356

                        use_cuda = True
2357 2358
                        place = paddle.CUDAPlace(0) if use_cuda else paddle.CPUPlace()
                        exe = static.Executor(place)
C
chengduo 已提交
2359 2360

                        # NOTE: If you use CPU to run the program, you need
2361
                        # to specify the CPU_NUM, otherwise, paddle will use
C
chengduo 已提交
2362 2363 2364 2365 2366 2367
                        # all the number of the logic core as the CPU_NUM,
                        # in that case, the batch size of the input should be
                        # greater than CPU_NUM, if not, the process will be
                        # failed by an exception.
                        if not use_cuda:
                            os.environ['CPU_NUM'] = str(2)
2368
                            places = static.cpu_places()
C
chengduo 已提交
2369
                        else:
2370
                            places = static.cuda_places()
C
chengduo 已提交
2371

2372 2373 2374 2375
                        data = static.data(name='X', shape=[None, 1], dtype='float32')
                        hidden = static.nn.fc(input=data, size=10)
                        loss = paddle.mean(hidden)
                        paddle.optimizer.SGD(learning_rate=0.01).minimize(loss)
C
chengduo 已提交
2376

2377
                        exe.run(static.default_startup_program())
C
chengduo 已提交
2378

2379
                        build_strategy = static.BuildStrategy()
C
chengduo 已提交
2380
                        build_strategy.gradient_scale_strategy = \
2381 2382 2383
                                  static.BuildStrategy.GradientScaleStrategy.Customized
                        compiled_prog = static.CompiledProgram(
                                  static.default_main_program()).with_data_parallel(
C
chengduo 已提交
2384
                                          loss_name=loss.name, build_strategy=build_strategy,
2385
                                          places=places)
C
chengduo 已提交
2386 2387 2388 2389 2390 2391

                        dev_count =  len(places)
                        x = numpy.random.random(size=(10, 1)).astype('float32')
                        loss_grad = numpy.ones((dev_count)).astype("float32") * 0.01
                        loss_grad_name = loss.name+"@GRAD"
                        loss_data = exe.run(compiled_prog,
2392 2393
                                              feed={"X": x, loss_grad_name : loss_grad},
                                              fetch_list=[loss.name, loss_grad_name])
F
flame 已提交
2394
                   )DOC")
Y
yuyang18 已提交
2395 2396 2397 2398
      .def_property(
          "debug_graphviz_path",
          [](const BuildStrategy &self) { return self.debug_graphviz_path_; },
          [](BuildStrategy &self, const std::string &path) {
2399 2400 2401 2402
            PADDLE_ENFORCE_NE(self.IsFinalized(), true,
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
Y
yuyang18 已提交
2403
            self.debug_graphviz_path_ = path;
C
chengduo 已提交
2404
          },
2405
          R"DOC((str, optional): debug_graphviz_path indicates the path that
F
flame 已提交
2406
                writing the SSA Graph to file in the form of graphviz.
2407
                It is useful for debugging. Default is empty string, that is, ""
F
flame 已提交
2408 2409 2410 2411

                Examples:
                    .. code-block:: python

2412 2413 2414 2415
                        import paddle
                        import paddle.static as static

                        paddle.enable_static()
C
chengduo 已提交
2416

2417 2418
                        build_strategy = static.BuildStrategy()
                        build_strategy.debug_graphviz_path = "./graph"
F
flame 已提交
2419
                    )DOC")
S
sneaxiy 已提交
2420 2421 2422 2423 2424 2425
      .def_property(
          "enable_sequential_execution",
          [](const BuildStrategy &self) {
            return self.enable_sequential_execution_;
          },
          [](BuildStrategy &self, bool b) {
2426 2427 2428 2429
            PADDLE_ENFORCE_NE(self.IsFinalized(), true,
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
S
sneaxiy 已提交
2430 2431
            self.enable_sequential_execution_ = b;
          },
2432 2433
          R"DOC((bool, optional): If set True, the execution order of ops would
                be the same as what is in the program. Default is False.
F
flame 已提交
2434 2435 2436 2437

                Examples:
                    .. code-block:: python

2438 2439 2440 2441 2442 2443
                        import paddle
                        import paddle.static as static

                        paddle.enable_static()

                        build_strategy = static.BuildStrategy()
F
flame 已提交
2444 2445
                        build_strategy.enable_sequential_execution = True
          )DOC")
S
sneaxiy 已提交
2446 2447 2448 2449 2450 2451
      .def_property(
          "remove_unnecessary_lock",
          [](const BuildStrategy &self) {
            return self.remove_unnecessary_lock_;
          },
          [](BuildStrategy &self, bool b) {
2452 2453 2454 2455
            PADDLE_ENFORCE_NE(self.IsFinalized(), true,
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
S
sneaxiy 已提交
2456 2457
            self.remove_unnecessary_lock_ = b;
          },
2458 2459
          R"DOC((bool, optional): If set True, some locks in GPU ops would be
                released and ParallelExecutor would run faster. Default is True.
F
flame 已提交
2460 2461 2462 2463

                Examples:
                    .. code-block:: python

2464 2465 2466 2467 2468 2469
                        import paddle
                        import paddle.static as static

                        paddle.enable_static()

                        build_strategy = static.BuildStrategy()
F
flame 已提交
2470 2471
                        build_strategy.remove_unnecessary_lock = True
          )DOC")
2472 2473 2474 2475
      .def_property(
          "num_trainers",
          [](const BuildStrategy &self) { return self.num_trainers_; },
          [](BuildStrategy &self, int num_trainers) {
2476
#ifdef WIN32
2477
            PADDLE_THROW(platform::errors::Unavailable(
2478
                "Distribution mode is not supported on Windows platform."));
2479
#endif
2480 2481
            self.num_trainers_ = num_trainers;
          })
2482 2483 2484 2485 2486 2487 2488 2489 2490 2491 2492 2493
      .def_property(
          "trainers_endpoints",
          [](const BuildStrategy &self) { return self.trainers_endpoints_; },
          [](BuildStrategy &self,
             const std::vector<std::string> &trainers_endpoints) {
            self.trainers_endpoints_ = trainers_endpoints;
          })
      .def_property("trainer_id",
                    [](const BuildStrategy &self) { return self.trainer_id_; },
                    [](BuildStrategy &self, int trainer_id) {
                      self.trainer_id_ = trainer_id;
                    })
2494 2495 2496 2497 2498 2499
      .def_property(
          "nccl_comm_num",
          [](const BuildStrategy &self) { return self.nccl_comm_num_; },
          [](BuildStrategy &self, int nccl_comm_num) {
            self.nccl_comm_num_ = nccl_comm_num;
          })
2500
      .def_property("use_hierarchical_allreduce",
2501 2502 2503 2504 2505 2506
                    [](const BuildStrategy &self) {
                      return self.use_hierarchical_allreduce_;
                    },
                    [](BuildStrategy &self, bool use) {
                      self.use_hierarchical_allreduce_ = use;
                    })
2507
      .def_property("hierarchical_allreduce_inter_nranks",
2508 2509 2510 2511 2512 2513 2514
                    [](const BuildStrategy &self) {
                      return self.hierarchical_allreduce_inter_nranks_;
                    },
                    [](BuildStrategy &self, int nranks) {
                      self.hierarchical_allreduce_inter_nranks_ = nranks;
                    })

C
chengduo 已提交
2515 2516 2517 2518 2519 2520
      .def_property(
          "fuse_elewise_add_act_ops",
          [](const BuildStrategy &self) {
            return self.fuse_elewise_add_act_ops_;
          },
          [](BuildStrategy &self, bool b) {
2521 2522 2523 2524
            PADDLE_ENFORCE_NE(self.IsFinalized(), true,
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
C
chengduo 已提交
2525 2526
            self.fuse_elewise_add_act_ops_ = b;
          },
2527
          R"DOC((bool, optional): fuse_elewise_add_act_ops indicate whether
F
flame 已提交
2528
                to fuse elementwise_add_op and activation_op,
2529
                it may make the execution faster. Default is False.
F
flame 已提交
2530 2531 2532 2533

                Examples:
                    .. code-block:: python

2534 2535 2536 2537 2538 2539
                        import paddle
                        import paddle.static as static

                        paddle.enable_static()

                        build_strategy = static.BuildStrategy()
F
flame 已提交
2540 2541
                        build_strategy.fuse_elewise_add_act_ops = True
                     )DOC")
Z
Zhen Wang 已提交
2542 2543 2544 2545
      .def_property(
          "fuse_bn_act_ops",
          [](const BuildStrategy &self) { return self.fuse_bn_act_ops_; },
          [](BuildStrategy &self, bool b) {
2546
            PADDLE_ENFORCE_NE(self.IsFinalized(), true,
Z
Zhen Wang 已提交
2547
                              platform::errors::PreconditionNotMet(
2548 2549
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
Z
Zhen Wang 已提交
2550 2551 2552 2553 2554 2555 2556 2557 2558
            self.fuse_bn_act_ops_ = b;
          },
          R"DOC((bool, optional): fuse_bn_act_ops indicate whether
                to fuse batch_norm and activation_op,
                it may make the execution faster. Default is False.

                Examples:
                    .. code-block:: python

2559 2560 2561 2562 2563 2564
                        import paddle
                        import paddle.static as static

                        paddle.enable_static()

                        build_strategy = static.BuildStrategy()
Z
Zhen Wang 已提交
2565 2566
                        build_strategy.fuse_bn_act_ops = True
                     )DOC")
Z
Zhang Ting 已提交
2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591
      .def_property(
          "fuse_bn_add_act_ops",
          [](const BuildStrategy &self) { return self.fuse_bn_add_act_ops_; },
          [](BuildStrategy &self, bool b) {
            PADDLE_ENFORCE_NE(self.IsFinalized(), true,
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
            self.fuse_bn_add_act_ops_ = b;
          },
          R"DOC((bool, optional): fuse_bn_add_act_ops indicate whether
                to fuse batch_norm, elementwise_add and activation_op,
                it may make the execution faster. Default is True

                Examples:
                    .. code-block:: python

                        import paddle
                        import paddle.static as static

                        paddle.enable_static()

                        build_strategy = static.BuildStrategy()
                        build_strategy.fuse_bn_add_act_ops = True
                     )DOC")
2592 2593 2594 2595
      .def_property(
          "enable_auto_fusion",
          [](const BuildStrategy &self) { return self.enable_auto_fusion_; },
          [](BuildStrategy &self, bool b) {
2596
            PADDLE_ENFORCE_NE(self.IsFinalized(), true,
2597
                              platform::errors::PreconditionNotMet(
2598 2599
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
2600 2601 2602 2603 2604 2605 2606 2607 2608 2609
            self.enable_auto_fusion_ = b;
          },
          R"DOC((bool, optional): Whether to enable fusing subgraph to a
                fusion_group. Now we only support fusing subgraph that composed
                of elementwise-like operators, such as elementwise_add/mul
                without broadcast and activations.

                Examples:
                    .. code-block:: python

2610 2611 2612 2613 2614 2615
                        import paddle
                        import paddle.static as static

                        paddle.enable_static()

                        build_strategy = static.BuildStrategy()
2616 2617
                        build_strategy.enable_auto_fusion = True
                    )DOC")
2618 2619 2620 2621 2622 2623
      .def_property(
          "fuse_relu_depthwise_conv",
          [](const BuildStrategy &self) {
            return self.fuse_relu_depthwise_conv_;
          },
          [](BuildStrategy &self, bool b) {
2624 2625 2626 2627
            PADDLE_ENFORCE_NE(self.IsFinalized(), true,
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
2628 2629
            self.fuse_relu_depthwise_conv_ = b;
          },
2630
          R"DOC((bool, optional): fuse_relu_depthwise_conv indicate whether
F
flame 已提交
2631 2632 2633
                to fuse relu and depthwise_conv2d,
                it will save GPU memory and may make the execution faster.
                This options is only available in GPU devices.
2634
                Default is False.
F
flame 已提交
2635 2636 2637 2638

                Examples:
                    .. code-block:: python

2639 2640 2641 2642 2643 2644
                        import paddle
                        import paddle.static as static

                        paddle.enable_static()

                        build_strategy = static.BuildStrategy()
F
flame 已提交
2645 2646
                        build_strategy.fuse_relu_depthwise_conv = True
          )DOC")
C
chengduo 已提交
2647 2648 2649 2650 2651 2652
      .def_property("fuse_broadcast_ops",
                    [](const BuildStrategy &self) {
                      return self.fuse_broadcast_ops_ == true ||
                             self.fuse_broadcast_ops_ == boost::none;
                    },
                    [](BuildStrategy &self, bool b) {
2653 2654 2655 2656
                      PADDLE_ENFORCE_NE(self.IsFinalized(), true,
                                        platform::errors::PreconditionNotMet(
                                            "BuildStrategy has been finlaized, "
                                            "cannot be configured again."));
C
chengduo 已提交
2657 2658
                      self.fuse_broadcast_ops_ = b;
                    },
2659
                    R"DOC((bool, optional): fuse_broadcast_op indicates whether
2660 2661 2662 2663
                      to fuse the broadcast ops. Note that, in Reduce mode,
                      fusing broadcast ops may make the program faster. Because
                      fusing broadcast OP equals delaying the execution of all
                      broadcast Ops, in this case, all nccl streams are used only
2664 2665 2666 2667 2668
                      for NCCLReduce operations for a period of time. Default False.

                      Examples:
                          .. code-block:: python

2669 2670 2671 2672 2673 2674
                              import paddle
                              import paddle.static as static

                              paddle.enable_static()

                              build_strategy = static.BuildStrategy()
2675 2676
                              build_strategy.fuse_broadcast_ops = True
                    )DOC")
C
chengduo 已提交
2677 2678
      .def_property("fuse_all_optimizer_ops",
                    [](const BuildStrategy &self) {
C
chengduo 已提交
2679 2680
                      return self.fuse_all_optimizer_ops_ == true ||
                             self.fuse_all_optimizer_ops_ == boost::none;
C
chengduo 已提交
2681 2682
                    },
                    [](BuildStrategy &self, bool b) {
2683 2684 2685 2686
                      PADDLE_ENFORCE_NE(self.IsFinalized(), true,
                                        platform::errors::PreconditionNotMet(
                                            "BuildStrategy has been finlaized, "
                                            "cannot be configured again."));
C
chengduo 已提交
2687 2688
                      self.fuse_all_optimizer_ops_ = b;
                    })
Q
qingqing01 已提交
2689 2690 2691 2692
      .def_property(
          "sync_batch_norm",
          [](const BuildStrategy &self) { return self.sync_batch_norm_; },
          [](BuildStrategy &self, bool b) {
2693 2694 2695 2696
            PADDLE_ENFORCE_NE(self.IsFinalized(), true,
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
Q
qingqing01 已提交
2697 2698
            self.sync_batch_norm_ = b;
          },
2699
          R"DOC((bool, optional): sync_batch_norm indicates whether to use
Q
qingqing01 已提交
2700 2701 2702
                synchronous batch normalization which synchronizes the mean
                and variance through multi-devices in training phase.
                Current implementation doesn't support FP16 training and CPU.
2703 2704
                And only synchronous on one machine, not all machines. 
                Default is False.
F
flame 已提交
2705 2706 2707 2708

                Examples:
                    .. code-block:: python

2709 2710 2711 2712 2713 2714
                        import paddle
                        import paddle.static as static

                        paddle.enable_static()

                        build_strategy = static.BuildStrategy()
F
flame 已提交
2715 2716
                        build_strategy.sync_batch_norm = True
                )DOC")
D
dzhwinter 已提交
2717 2718
      .def_property(
          "memory_optimize",
2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732
          [](const BuildStrategy &self) -> py::object {
            if (self.memory_optimize_) {
              return py::cast(self.memory_optimize_.get());
            } else {
              return py::cast(nullptr);
            }
          },
          [](BuildStrategy &self, const py::handle &value) {
            auto *py_obj = value.ptr();
            if (py_obj == nullptr || py_obj == Py_None) {
              self.memory_optimize_ = boost::none;
            } else if (PyBool_Check(py_obj)) {
              self.memory_optimize_ = (py_obj == Py_True);
            } else {
2733 2734 2735
              PADDLE_THROW(platform::errors::InvalidArgument(
                  "BuildStrategy.memory_optimize must be set to None, False or "
                  "True"));
2736 2737
            }
          },
2738
          R"DOC((bool, optional): memory opitimize aims to save total memory
2739
                consumption, set to True to enable it.
2740

2741 2742 2743
                Default None. None means framework would choose to use or not use 
                this strategy automatically. Currently, None means that it is 
                enabled when GC is disabled, and disabled when GC is enabled. 
2744 2745 2746 2747 2748 2749 2750 2751 2752 2753 2754 2755 2756 2757
                True means enabling and False means disabling. Default is None.

                Examples:
                    .. code-block:: python

                        import paddle
                        import paddle.static as static

                        paddle.enable_static()

                        build_strategy = static.BuildStrategy()
                        build_strategy.memory_optimize = True
                
                )DOC")
2758 2759 2760
      .def_property(
          "is_distribution",
          [](const BuildStrategy &self) { return self.is_distribution_; },
2761 2762 2763
          [](BuildStrategy &self, bool b) {
#ifdef WIN32
            if (b) {
2764
              PADDLE_THROW(platform::errors::Unavailable(
2765
                  "Distribution mode is not supported on Windows platform."));
2766 2767 2768 2769 2770
            }
#else
            self.is_distribution_ = b;
#endif
          })
Q
can run  
Qiao Longfei 已提交
2771 2772 2773
      .def_property("async_mode",
                    [](const BuildStrategy &self) { return self.async_mode_; },
                    [](BuildStrategy &self, bool b) { self.async_mode_ = b; })
D
dzhwinter 已提交
2774
      .def_property(
D
dzhwinter 已提交
2775 2776 2777
          "enable_inplace",
          [](const BuildStrategy &self) { return self.enable_inplace_; },
          [](BuildStrategy &self, bool b) { self.enable_inplace_ = b; })
2778 2779 2780 2781
      .def_property(
          "enable_addto",
          [](const BuildStrategy &self) { return self.enable_addto_; },
          [](BuildStrategy &self, bool b) { self.enable_addto_ = b; })
C
chengduo 已提交
2782 2783
      .def_property(
          "fuse_all_reduce_ops",
C
chengduo 已提交
2784 2785 2786 2787
          [](const BuildStrategy &self) {
            return self.fuse_all_reduce_ops_ == true ||
                   self.fuse_all_reduce_ops_ == boost::none;
          },
C
chengduo 已提交
2788
          [](BuildStrategy &self, bool b) { self.fuse_all_reduce_ops_ = b; })
2789 2790 2791 2792 2793 2794 2795
      .def_property("enable_backward_optimizer_op_deps",
                    [](const BuildStrategy &self) {
                      return self.enable_backward_optimizer_op_deps_;
                    },
                    [](BuildStrategy &self, bool b) {
                      self.enable_backward_optimizer_op_deps_ = b;
                    })
2796 2797 2798 2799
      .def_property(
          "cache_runtime_context",
          [](const BuildStrategy &self) { return self.cache_runtime_context_; },
          [](BuildStrategy &self, bool b) { self.cache_runtime_context_ = b; })
2800 2801 2802 2803 2804 2805 2806 2807 2808
      .def_property(
          "mkldnn_enabled_op_types",
          [](const BuildStrategy &self) {
            return self.mkldnn_enabled_op_types_;
          },
          [](BuildStrategy &self,
             const std::unordered_set<std::string> &mkldnn_enabled_op_types) {
            self.mkldnn_enabled_op_types_ = mkldnn_enabled_op_types;
          })
2809
      .def("_finalize_strategy_and_create_passes",
X
fix  
Xin Pan 已提交
2810
           [](BuildStrategy &self) -> std::shared_ptr<ir::PassBuilder> {
2811 2812 2813 2814 2815
             return self.CreatePassesFromStrategy(true);
           },
           R"DOC(Allow user to customized passes. Normally model-specific
                optimization passes should be defined in this way. BuildStrategy
                cannot be updated after being finalized.)DOC");
Y
yuyang18 已提交
2816 2817

  pe.def(py::init<const std::vector<platform::Place> &,
Y
Yan Xu 已提交
2818
                  const std::vector<std::string> &, const std::string &,
X
Xin Pan 已提交
2819
                  Scope *, std::vector<Scope *> &, const ExecutionStrategy &,
X
Xin Pan 已提交
2820
                  const BuildStrategy &, ir::Graph *>())
Y
Yu Yang 已提交
2821 2822 2823 2824
      // NOTE: even we return a vec<Scope*>* to Python use reference policy.
      // We still cannot get local_scope from this vector, since the element
      // of vec<Scope*> will be freed by Python GC. We can only return Scope*
      // one by one and mark them as reference.
2825 2826 2827 2828 2829
      .def("local_scopes",
           [](ParallelExecutor &self) -> std::vector<Scope *> * {
             return &self.GetLocalScopes();
           },
           py::return_value_policy::reference)
2830 2831 2832
      .def("drop_local_exe_scopes", &ParallelExecutor::DropLocalExeScopes)
      .def("_need_create_local_exe_scopes",
           &ParallelExecutor::NeedCreateLocalExeScope)
Y
Yu Yang 已提交
2833 2834 2835 2836
      .def("feed_tensors_into_local_scopes",
           &ParallelExecutor::FeedTensorsIntoLocalScopes)
      .def("feed_and_split_tensor_into_local_scopes",
           &ParallelExecutor::FeedAndSplitTensorIntoLocalScopes)
2837 2838
      .def("run",
           [](ParallelExecutor &self,
Z
Zhen Wang 已提交
2839 2840 2841 2842 2843 2844 2845 2846
              const std::vector<std::string> &fetch_tensors,
              bool return_merged) -> py::object {
             paddle::framework::FetchResultType ret;
             {
               pybind11::gil_scoped_release release;
               ret = self.Run(fetch_tensors, return_merged);
             }
             if (return_merged) {
2847
               return py::cast(
2848
                   std::move(BOOST_GET(paddle::framework::FetchList, ret)));
Z
Zhen Wang 已提交
2849 2850
             } else {
               return py::cast(std::move(
2851
                   BOOST_GET(paddle::framework::FetchUnmergedList, ret)));
Z
Zhen Wang 已提交
2852
             }
2853 2854
           })
      .def("device_count", &ParallelExecutor::DeviceCount);
Y
Yu Yang 已提交
2855

D
dongdaxiang 已提交
2856
  BindFleetWrapper(&m);
T
Thunderbrook 已提交
2857

T
Thunderbrook 已提交
2858 2859
#ifdef PADDLE_WITH_PSLIB
  BindHeterWrapper(&m);
T
Thunderbrook 已提交
2860
#endif
2861 2862
#if (defined PADDLE_WITH_NCCL || defined PADDLE_WITH_RCCL) && \
    (defined PADDLE_WITH_PSLIB)
T
Thunderbrook 已提交
2863
  BindPSGPUWrapper(&m);
T
Thunderbrook 已提交
2864
#endif
2865
  BindGlooWrapper(&m);
H
hutuxian 已提交
2866
  BindBoxHelper(&m);
H
hutuxian 已提交
2867 2868 2869
#ifdef PADDLE_WITH_BOX_PS
  BindBoxWrapper(&m);
#endif
2870
#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
D
dongdaxiang 已提交
2871
  BindNCCLWrapper(&m);
2872 2873 2874
#endif
#ifdef PADDLE_WITH_GLOO
  BindGlooContext(&m);
W
wopeizl 已提交
2875
#endif
F
flame 已提交
2876 2877
  BindGraph(&m);
  BindNode(&m);
F
flame 已提交
2878
  BindInferenceApi(&m);
2879
  BindCompatible(&m);
2880
  BindDataset(&m);
Y
yaoxuefeng 已提交
2881
  BindGenerator(&m);
2882 2883 2884 2885
#ifdef PADDLE_WITH_ASCEND
  BindAscendWrapper(&m);
  BindAscendGraph(&m);
#endif
Y
Yanghello 已提交
2886 2887 2888
#ifdef PADDLE_WITH_CRYPTO
  BindCrypto(&m);
#endif
T
tangwei12 已提交
2889

T
tangwei12 已提交
2890
#if defined PADDLE_WITH_PSCORE
T
tangwei12 已提交
2891 2892
  BindDistFleetWrapper(&m);
  BindPSHost(&m);
2893
  BindCommunicatorContext(&m);
T
tangwei12 已提交
2894 2895
  BindDistCommunicator(&m);
  BindHeterClient(&m);
2896
#endif
L
Luo Tao 已提交
2897
}
2898
}  // namespace pybind
2899
}  // namespace paddle