pybind.cc 190.9 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
2
Copyright (c) 2022 NVIDIA Authors. All Rights Reserved.
3 4 5 6 7

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

8
http://www.apache.org/licenses/LICENSE-2.0
9 10 11 12 13 14

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
15
#include <Python.h>
16

C
chengduoZH 已提交
17
#include <algorithm>
18
#include <cctype>
19
#include <cstdlib>
20
#include <iterator>
C
chengduoZH 已提交
21
#include <map>
S
sneaxiy 已提交
22
#include <memory>
C
chengduoZH 已提交
23 24
#include <mutex>  // NOLINT // for call_once
#include <string>
25 26
#include <tuple>
#include <type_traits>
C
chengduoZH 已提交
27
#include <unordered_map>
28
#include <unordered_set>
C
chengduoZH 已提交
29 30
#include <utility>
#include <vector>
31

32
#include "paddle/fluid/framework/convert_utils.h"
33
#include "paddle/fluid/framework/custom_operator.h"
34
#include "paddle/fluid/framework/data_layout.h"
L
Leo Chen 已提交
35
#include "paddle/fluid/framework/data_type_transform.h"
Y
Yi Wang 已提交
36
#include "paddle/fluid/framework/executor.h"
37
#include "paddle/fluid/framework/executor_cache.h"
38
#include "paddle/fluid/framework/executor_gc_helper.h"
Y
Yi Wang 已提交
39
#include "paddle/fluid/framework/feed_fetch_method.h"
Z
Zhen Wang 已提交
40
#include "paddle/fluid/framework/feed_fetch_type.h"
S
sneaxiy 已提交
41
#include "paddle/fluid/framework/garbage_collector.h"
H
hutuxian 已提交
42
#include "paddle/fluid/framework/io/fs.h"
43
#include "paddle/fluid/framework/ir/coalesce_grad_tensor_pass.h"
H
Huihuang Zheng 已提交
44
#include "paddle/fluid/framework/ir/cost_model.h"
45
#include "paddle/fluid/framework/ir/generate_pass.h"
46
#include "paddle/fluid/framework/ir/pass_builder.h"
Y
Yi Wang 已提交
47 48
#include "paddle/fluid/framework/lod_rank_table.h"
#include "paddle/fluid/framework/lod_tensor_array.h"
L
liutiexing 已提交
49
#include "paddle/fluid/framework/new_executor/executor_statistics.h"
50
#include "paddle/fluid/framework/new_executor/standalone_executor.h"
S
sneaxiy 已提交
51
#include "paddle/fluid/framework/op_info.h"
52
#include "paddle/fluid/framework/op_registry.h"
53
#include "paddle/fluid/framework/op_version_registry.h"
Y
Yu Yang 已提交
54
#include "paddle/fluid/framework/parallel_executor.h"
55
#include "paddle/fluid/framework/phi_utils.h"
Y
Yi Wang 已提交
56
#include "paddle/fluid/framework/prune.h"
Y
Refine  
Yu Yang 已提交
57
#include "paddle/fluid/framework/reader.h"
H
hong 已提交
58
#include "paddle/fluid/framework/save_load_util.h"
S
sneaxiy 已提交
59
#include "paddle/fluid/framework/scope_pool.h"
60
#include "paddle/fluid/framework/selected_rows_utils.h"
61
#include "paddle/fluid/framework/tensor_util.h"
62
#include "paddle/fluid/framework/trainer.h"
63
#include "paddle/fluid/framework/type_defs.h"
X
Xin Pan 已提交
64
#include "paddle/fluid/framework/version.h"
L
Leo Chen 已提交
65
#include "paddle/fluid/imperative/amp_auto_cast.h"
H
hong 已提交
66
#include "paddle/fluid/imperative/layer.h"
Y
Refine  
Yu Yang 已提交
67
#include "paddle/fluid/memory/allocation/allocator_strategy.h"
68 69 70
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/memory/allocation/cuda_ipc_allocator.h"
#endif
71
#include "paddle/fluid/memory/allocation/mmap_allocator.h"
D
dzhwinter 已提交
72
#include "paddle/fluid/operators/activation_op.h"
L
Leo Chen 已提交
73
#include "paddle/fluid/operators/common_infer_shape_functions.h"
S
sneaxiy 已提交
74
#include "paddle/fluid/operators/py_func_op.h"
75
#include "paddle/fluid/platform/cpu_helper.h"
Y
Yu Yang 已提交
76
#include "paddle/fluid/platform/cpu_info.h"
77
#include "paddle/fluid/platform/device/device_wrapper.h"
78
#include "paddle/fluid/platform/device_context.h"
79
#include "paddle/fluid/platform/dynload/dynamic_loader.h"
Y
Yi Wang 已提交
80
#include "paddle/fluid/platform/enforce.h"
81
#include "paddle/fluid/platform/init.h"
H
hutuxian 已提交
82
#include "paddle/fluid/platform/monitor.h"
Y
Yi Wang 已提交
83 84
#include "paddle/fluid/platform/place.h"
#include "paddle/fluid/platform/profiler.h"
C
chenjian 已提交
85 86 87
#include "paddle/fluid/platform/profiler/event_python.h"
#include "paddle/fluid/platform/profiler/event_tracing.h"
#include "paddle/fluid/platform/profiler/profiler.h"
88
#include "paddle/fluid/pybind/cuda_streams_py.h"
89
#include "paddle/fluid/pybind/distributed_py.h"
90
#include "paddle/fluid/pybind/eager.h"
J
Jiabin Yang 已提交
91
#include "paddle/fluid/pybind/imperative.h"
92
#include "paddle/fluid/pybind/io.h"
93 94
#include "paddle/phi/core/compat/convert_utils.h"
#include "paddle/phi/core/lod_utils.h"
95
#include "paddle/utils/none.h"
96 97 98
#ifdef PADDLE_WITH_ASCEND
#include "paddle/fluid/pybind/ascend_wrapper_py.h"
#endif
H
Huihuang Zheng 已提交
99
#include "paddle/fluid/pybind/bind_cost_model.h"
L
LiYuRio 已提交
100
#include "paddle/fluid/pybind/bind_fleet_executor.h"
H
hutuxian 已提交
101
#include "paddle/fluid/pybind/box_helper_py.h"
102
#include "paddle/fluid/pybind/communication.h"
103
#include "paddle/fluid/pybind/compatible.h"
Y
Yi Wang 已提交
104
#include "paddle/fluid/pybind/const_value.h"
D
dongdaxiang 已提交
105
#include "paddle/fluid/pybind/data_set_py.h"
Y
Yi Wang 已提交
106
#include "paddle/fluid/pybind/exception.h"
D
dongdaxiang 已提交
107
#include "paddle/fluid/pybind/fleet_wrapper_py.h"
Y
yaoxuefeng 已提交
108
#include "paddle/fluid/pybind/generator_py.h"
109
#include "paddle/fluid/pybind/global_value_getter_setter.h"
110
#include "paddle/fluid/pybind/gloo_context_py.h"
111
#include "paddle/fluid/pybind/gloo_wrapper_py.h"
T
Thunderbrook 已提交
112
#include "paddle/fluid/pybind/heter_wrapper_py.h"
F
flame 已提交
113
#include "paddle/fluid/pybind/inference_api.h"
F
flame 已提交
114
#include "paddle/fluid/pybind/ir.h"
115
#include "paddle/fluid/pybind/metrics_py.h"
T
Thunderbrook 已提交
116
#include "paddle/fluid/pybind/ps_gpu_wrapper_py.h"
117
#include "paddle/fluid/pybind/pybind_boost_headers.h"
118
#include "paddle/phi/backends/device_manager.h"
119

120
#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
D
dongdaxiang 已提交
121
#include "paddle/fluid/pybind/nccl_wrapper_py.h"
W
wopeizl 已提交
122
#endif
123
#include "paddle/fluid/framework/data_type.h"
124 125
#include "paddle/fluid/pybind/protobuf.h"
#include "paddle/fluid/pybind/pybind.h"  // NOLINT
S
sneaxiy 已提交
126
#include "paddle/fluid/pybind/reader_py.h"
Y
Yi Wang 已提交
127
#include "paddle/fluid/pybind/tensor_py.h"
128
#include "paddle/fluid/string/to_string.h"
129 130
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
Y
Yi Wang 已提交
131
#include "paddle/fluid/operators/nccl/nccl_gpu_common.h"
P
peizhilin 已提交
132
#endif
133
#ifndef PADDLE_WITH_HIP
134
#include "paddle/fluid/platform/device/gpu/cuda/cuda_profiler.h"
135
#endif
136
#include "paddle/fluid/platform/device/gpu/gpu_info.h"
D
Dong Zhihong 已提交
137 138
#endif

139
#ifdef PADDLE_WITH_ASCEND_CL
140
#include "paddle/fluid/platform/collective_helper.h"
141 142
#include "paddle/fluid/platform/device/npu/npu_info.h"
#include "paddle/fluid/platform/device/npu/npu_profiler.h"
143 144
#endif

145
#ifdef PADDLE_WITH_XPU
146
#include "paddle/fluid/platform/device/xpu/xpu_info.h"
T
TTerror 已提交
147
#include "paddle/fluid/platform/device/xpu/xpu_op_list.h"
148 149
#endif

150 151 152 153
#ifdef PADDLE_WITH_CUSTOM_DEVICE
#include "paddle/phi/capi/capi.h"
#endif

154
#include "paddle/fluid/platform/cuda_graph_with_memory_pool.h"
A
Allen Guo 已提交
155

J
jianghaicheng 已提交
156
#ifdef PADDLE_WITH_IPU
A
Allen Guo 已提交
157 158
#include "paddle/fluid/platform/device/ipu/ipu_backend.h"
#include "paddle/fluid/platform/device/ipu/ipu_info.h"
J
jianghaicheng 已提交
159
#endif
160

161 162 163 164
#ifdef PADDLE_WITH_MLU
#include "paddle/fluid/platform/device/mlu/mlu_info.h"
#endif

Y
Yanghello 已提交
165 166 167 168
#ifdef PADDLE_WITH_CRYPTO
#include "paddle/fluid/pybind/crypto.h"
#endif

T
tangwei12 已提交
169
#if defined PADDLE_WITH_PSCORE
T
tangwei12 已提交
170 171 172
#include "paddle/fluid/pybind/fleet_py.h"
#endif

173 174 175 176
#ifdef PADDLE_WITH_CINN
#include "paddle/fluid/framework/paddle2cinn/cinn_compiler.h"
#endif

177
#include "paddle/fluid/eager/api/utils/global_utils.h"
178
#include "paddle/fluid/imperative/layout_autotune.h"
179 180
#include "paddle/fluid/pybind/eager_utils.h"
#include "paddle/phi/api/ext/op_meta_info.h"
181 182
#include "paddle/phi/kernels/autotune/cache.h"
#include "paddle/phi/kernels/autotune/switch_autotune.h"
M
minqiyang 已提交
183 184
#include "pybind11/stl.h"

185
DECLARE_bool(use_mkldnn);
186

Q
Qiao Longfei 已提交
187 188
// disable auto conversion to list in Python
PYBIND11_MAKE_OPAQUE(paddle::framework::LoDTensorArray);
189 190 191
PYBIND11_MAKE_OPAQUE(paddle::framework::FetchUnmergedList);
PYBIND11_MAKE_OPAQUE(paddle::framework::FetchList);
PYBIND11_MAKE_OPAQUE(paddle::framework::FetchType);
Q
Qiao Longfei 已提交
192

193
namespace paddle {
194
namespace pybind {
195 196

PyTypeObject *g_place_pytype = nullptr;
0
0x45f 已提交
197
PyTypeObject *g_framework_scope_pytype = nullptr;
198 199 200 201 202
PyTypeObject *g_cudaplace_pytype = nullptr;
PyTypeObject *g_cpuplace_pytype = nullptr;
PyTypeObject *g_xpuplace_pytype = nullptr;
PyTypeObject *g_npuplace_pytype = nullptr;
PyTypeObject *g_cudapinnedplace_pytype = nullptr;
203
PyTypeObject *g_mluplace_pytype = nullptr;
204
PyTypeObject *g_customplace_pytype = nullptr;
205
PyTypeObject *g_framework_tensor_pytype = nullptr;
206
PyTypeObject *g_framework_lodtensorarray_pytype = nullptr;
207
PyTypeObject *g_custom_op_kernel_ctx_pytype = nullptr;
208

209
bool IsCompiledWithCUDA() {
210 211 212 213 214 215 216
#if !defined(PADDLE_WITH_CUDA) && !defined(PADDLE_WITH_HIP)
  return false;
#else
  return true;
#endif
}

217 218 219 220 221 222 223 224
bool IsCompiledWithNCCL() {
#ifdef PADDLE_WITH_NCCL
  return true;
#else
  return false;
#endif
}

225 226
bool IsCompiledWithROCM() {
#ifndef PADDLE_WITH_HIP
Q
qijun 已提交
227 228 229 230 231 232
  return false;
#else
  return true;
#endif
}

233 234 235 236 237 238 239 240
bool IsCompiledWithAscend() {
#ifndef PADDLE_WITH_ASCEND
  return false;
#else
  return true;
#endif
}

241 242 243 244 245 246 247 248
bool IsCompiledWithXPU() {
#ifndef PADDLE_WITH_XPU
  return false;
#else
  return true;
#endif
}

249 250 251 252 253 254 255 256
bool IsCompiledWithNPU() {
#ifndef PADDLE_WITH_ASCEND_CL
  return false;
#else
  return true;
#endif
}

J
jianghaicheng 已提交
257 258 259 260 261 262 263 264
bool IsCompiledWithIPU() {
#ifndef PADDLE_WITH_IPU
  return false;
#else
  return true;
#endif
}

265 266 267 268 269 270 271 272
bool IsCompiledWithMKLDNN() {
#ifndef PADDLE_WITH_MKLDNN
  return false;
#else
  return true;
#endif
}

273 274 275 276 277 278 279 280
bool IsCompiledWithCINN() {
#ifndef PADDLE_WITH_CINN
  return false;
#else
  return true;
#endif
}

281 282 283 284 285 286 287 288
bool IsCompiledWithMLU() {
#ifndef PADDLE_WITH_MLU
  return false;
#else
  return true;
#endif
}

289 290 291 292 293 294 295 296
bool IsCompiledWithHETERPS() {
#ifndef PADDLE_WITH_HETERPS
  return false;
#else
  return true;
#endif
}

297 298 299 300 301 302 303 304 305 306 307
bool SupportsBfloat16() {
#ifndef PADDLE_WITH_MKLDNN
  return false;
#else
  if (platform::MayIUse(platform::cpu_isa_t::avx512_core))
    return true;
  else
    return false;
#endif
}

308 309 310 311 312 313 314 315 316 317 318
bool SupportsBfloat16FastPerformance() {
#ifndef PADDLE_WITH_MKLDNN
  return false;
#else
  if (platform::MayIUse(platform::cpu_isa_t::avx512_bf16))
    return true;
  else
    return false;
#endif
}

319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335
bool SupportsInt8() {
#ifndef PADDLE_WITH_MKLDNN
  return false;
#else
  return (platform::MayIUse(platform::cpu_isa_t::avx2) ||
          platform::MayIUse(platform::cpu_isa_t::avx512f));
#endif
}

bool SupportsVNNI() {
#ifndef PADDLE_WITH_MKLDNN
  return false;
#else
  return platform::MayIUse(platform::cpu_isa_t::avx512_core_vnni);
#endif
}

336
bool IsCompiledWithBrpc() {
337
#ifndef PADDLE_WITH_DISTRIBUTE
338 339
  return false;
#endif
340
  return true;
341 342
}

Y
update  
Yancey1989 已提交
343
bool IsCompiledWithDIST() {
Y
Yancey1989 已提交
344
#ifdef PADDLE_WITH_DISTRIBUTE
Y
update  
Yancey1989 已提交
345 346 347 348 349 350
  return true;
#else
  return false;
#endif
}

S
sneaxiy 已提交
351 352 353 354 355 356 357
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) {
358
  return static_cast<int>(paddle::platform::Place(p).GetType());
S
sneaxiy 已提交
359 360
}

H
hong 已提交
361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382
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 &) {
383 384
    PADDLE_THROW(platform::errors::InvalidArgument(
        "Python object is not type of %s, the real type is %s",
385 386
        typeid(T).name(),
        obj->ob_type->tp_name));
H
hong 已提交
387 388 389 390 391 392 393 394 395 396 397 398 399
  }
}

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) {
400 401
      PADDLE_THROW(platform::errors::InvalidArgument(
          "The parameter [%s] to save is None", para.first));
H
hong 已提交
402 403
    }
    vec_res.emplace_back(
404
        PyObjectCast<std::shared_ptr<imperative::VarBase>>(py_obj));
H
hong 已提交
405 406 407 408 409 410 411 412 413 414 415 416
  }

  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) {
417 418
    PADDLE_THROW(platform::errors::InvalidArgument(
        "The parameter list to save is None"));
H
hong 已提交
419 420 421 422 423 424 425 426 427 428 429 430
  }

  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);
431 432 433
      PADDLE_ENFORCE_NOT_NULL(py_name,
                              platform::errors::InvalidArgument(
                                  "The name of parameter to save is None"));
H
hong 已提交
434 435 436 437
      vec_res.emplace_back(PyObjectCast<std::string>(py_name));
      Py_DECREF(py_name);
    }
  } else {
438 439
    PADDLE_THROW(platform::errors::InvalidArgument(
        "The parameters to save is not a list"));
H
hong 已提交
440 441 442 443
  }
  return vec_res;
}

444
static void inline CreateVariableIfNotExit(
445 446
    const py::handle &py_handle,
    const framework::Scope &scope,
447 448 449 450 451 452
    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) {
453 454
    PADDLE_THROW(
        platform::errors::InvalidArgument("The parameter list to set is None"));
455 456 457 458 459 460 461 462 463 464 465 466 467
  }

  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);
468 469 470
      PADDLE_ENFORCE_NOT_NULL(py_name,
                              platform::errors::InvalidArgument(
                                  "The name of parameter to set is None"));
471 472 473 474 475
      auto para_name = PyObjectCast<std::string>(py_name);
      Py_DECREF(py_name);

      auto var = scope.FindVar(para_name);
      if (var == nullptr) {
476 477 478 479 480
        PADDLE_ENFORCE_NOT_NULL(exe,
                                platform::errors::InvalidArgument(
                                    "Parameter not Initialized, "
                                    "Please set argument [executor] not None "
                                    "or run startup program first"));
481 482
        PyObject *py_var_desc =
            PyObject_GetAttrString(PyList_GET_ITEM(py_obj, i), kVarDescField);
483
        PADDLE_ENFORCE_NOT_NULL(
484 485 486
            py_var_desc,
            platform::errors::InvalidArgument(
                "The var_desc of parameter to set is None"));
487 488 489 490
        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>();
491
        tensor_temp->Resize(phi::make_ddim(var_desc.GetShape()));
492 493
        tensor_temp->mutable_data(
            exe->GetPlace(),
494
            framework::TransToPhiDataType(var_desc.GetDataType()));
495 496 497
      }
    }
  } else {
498 499
    PADDLE_THROW(platform::errors::InvalidArgument(
        "The parameters to set is not a list"));
500 501 502 503 504
  }

  return;
}

505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520
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";
      }
    }
  }
521 522
  PADDLE_ENFORCE_EQ(ops.empty(),
                    true,
523 524 525 526 527 528 529
                    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, ',')));
}

Z
Zeng Jinle 已提交
530 531 532 533
#ifdef PADDLE_WITH_NCCL
static int GetNCCLVersion() {
#if NCCL_VERSION_CODE >= 2304
  int ver;
534
  PADDLE_ENFORCE_GPU_SUCCESS(platform::dynload::ncclGetVersion(&ver));
Z
Zeng Jinle 已提交
535 536 537 538 539 540 541 542
  return ver;
#else
  PADDLE_THROW(platform::errors::External(
      "Cannot get NCCL version successfully when nccl version < 2.3.4"));
#endif
}
#endif

Z
Zeng Jinle 已提交
543
template <typename PlaceType>
544 545 546 547
static void TensorCopyFrom(framework::Tensor *dst,
                           const framework::Tensor &src,
                           const PlaceType &place,
                           int64_t batch_size) {
Z
Zeng Jinle 已提交
548 549 550 551 552 553 554 555
  if (batch_size < 0) {
    framework::TensorCopy(src, place, dst);
  } else {
    auto sliced = src.Slice(0, batch_size);
    framework::TensorCopy(sliced, place, dst);
  }
}

556 557 558 559 560 561
#ifdef PADDLE_WITH_AVX
PYBIND11_MODULE(core_avx, m) {
#else
PYBIND11_MODULE(core_noavx, m) {
#endif

J
Jiabin Yang 已提交
562
  BindImperative(&m);
563
  BindEager(&m);
J
Jack Zhou 已提交
564
  BindEagerStringTensor(&m);
565 566
  BindCudaStream(&m);

Y
Yu Yang 已提交
567 568 569
  // Not used, just make sure cpu_info.cc is linked.
  paddle::platform::CpuTotalPhysicalMemory();

Y
Refine  
Yu Yang 已提交
570
  paddle::memory::allocation::UseAllocatorStrategyGFlag();
S
sneaxiy 已提交
571

572 573
  AssertStaticGraphAndDygraphGradMakerNoDiff();

574
  m.doc() = "C++ core of PaddlePaddle";
575

576 577 578 579
  // using framework in this function. Since it is inside a function, it will
  // not cause namespace pollution.
  using namespace paddle::framework;  // NOLINT

580
  BindException(&m);
Y
Yu Yang 已提交
581

582 583
  m.def("set_num_threads", &platform::SetNumThreads);

584 585
  m.def("disable_signal_handler", &DisableSignalHandler);

586 587 588 589 590 591 592 593
  m.def("clear_gradients",
        [](std::vector<std::shared_ptr<imperative::VarBase>> param_list,
           bool set_to_zero) {
          for (auto param : param_list) {
            param->ClearGradient(set_to_zero);
          }
        });

594
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
595
  m.def("cudnn_version", &platform::DnnVersion);
596 597 598 599 600 601
  m.def("gpu_memory_available", []() {
    size_t available = 0;
    size_t total = 0;
    paddle::platform::GpuMemoryUsage(&available, &total);
    return available;
  });
602
#endif
603

Z
Zeng Jinle 已提交
604 605 606 607
#ifdef PADDLE_WITH_NCCL
  m.def("nccl_version", &GetNCCLVersion);
#endif

608 609 610 611 612 613 614 615 616
  m.def("is_cuda_graph_capturing", &platform::IsCUDAGraphCapturing);
#ifdef PADDLE_WITH_CUDA
  py::class_<platform::CUDAGraph>(m, "CUDAGraph")
      .def_static("begin_capture",
                  [](platform::CUDAPlace place, int mode) {
                    platform::BeginCUDAGraphCapture(
                        place, static_cast<cudaStreamCaptureMode>(mode));
                  })
      .def_static("end_capture", &platform::EndCUDAGraphCapture)
617 618
      .def_static("gen_new_memory_pool_id",
                  &platform::CUDAGraph::UniqueMemoryPoolID)
619
      .def("replay", &platform::CUDAGraph::Replay)
620 621
      .def("reset", &platform::CUDAGraph::Reset)
      .def("print_to_dot_files", &platform::CUDAGraph::PrintToDotFiles);
622 623
#endif

Z
Zeng Jinle 已提交
624 625 626 627
  m.def("wait_device", [](const platform::Place &place) {
    platform::DeviceContextPool::Instance().Get(place)->Wait();
  });

6
633WHU 已提交
628 629 630
  m.def("from_dlpack", [](py::capsule *dltensor) {
    DLManagedTensor *dmt = reinterpret_cast<DLManagedTensor *>(
        PyCapsule_GetPointer(dltensor->ptr(), "dltensor"));
631 632

    PADDLE_ENFORCE_NOT_NULL(
633 634 635 636
        dmt,
        platform::errors::InvalidArgument(
            "from_dlpack received an invalid capsule. "
            "Note that a DLPack tensor can be consumed only once."));
637

6
633WHU 已提交
638 639
    PyCapsule_SetName(dltensor->ptr(), "used_dltensor");
    DLTensor dl = dmt->dl_tensor;
640
    framework::Tensor tensor;
6
633WHU 已提交
641

S
Siming Dai 已提交
642
    if (dl.device.device_type == kDLCPU) {
6
633WHU 已提交
643 644
      paddle::framework::TensorFromDLPack(dl, &tensor);
    }
645
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
S
Siming Dai 已提交
646
    if (dl.device.device_type == kDLGPU) {
6
633WHU 已提交
647 648 649 650 651
      paddle::framework::TensorFromDLPack(dl, &tensor);
    }
#endif
    return tensor;
  });
H
hong 已提交
652

653
  m.def("_create_loaded_parameter",
654 655
        [](const py::handle &vec_var_list,
           const Scope &scope,
656 657 658 659
           const Executor *executor) {
          CreateVariableIfNotExit(vec_var_list, scope, executor);
        });

660 661 662 663 664 665
  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);
666 667
  });

668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692
  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;
  });

693 694 695 696 697 698
  m.def(
      "broadcast_shape",
      [](const std::vector<int64_t> &x_dim, const std::vector<int64_t> &y_dim) {
        return phi::vectorize(operators::details::BroadcastTwoDims(
            phi::make_ddim(x_dim), phi::make_ddim(y_dim), -1));
      });
L
Leo Chen 已提交
699

S
sneaxiy 已提交
700
  m.def(
S
sneaxiy 已提交
701
      "_append_python_callable_object_and_return_id",
S
sneaxiy 已提交
702 703 704 705
      [](py::object py_obj) -> size_t {
        return paddle::operators::AppendPythonCallableObjectAndReturnId(py_obj);
      });

S
sneaxiy 已提交
706 707 708
  m.def("_get_use_default_grad_op_desc_maker_ops",
        [] { return OpInfoMap::Instance().GetUseDefaultGradOpDescMakerOps(); });

709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724
  m.def(
      "_get_all_register_op_kernels",
      [](const std::string &lib) {
        std::unordered_map<std::string, std::vector<std::string>>
            all_kernels_info;
        if (lib == "fluid" || lib == "all") {
          auto &all_kernels =
              paddle::framework::OperatorWithKernel::AllOpKernels();

          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.emplace_back(
                  paddle::framework::KernelTypeToString(kernel_type));
725
            }
726
            all_kernels_info.emplace(op_type, kernel_types);
727
          }
728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743
        }
        if (lib == "phi" || lib == "all") {
          auto phi_kernels = phi::KernelFactory::Instance().kernels();
          for (auto &kernel_pair : phi_kernels) {
            auto op_type = phi::TransToFluidOpName(kernel_pair.first);
            std::vector<std::string> kernel_types;
            for (auto &info_pair : kernel_pair.second) {
              framework::OpKernelType kernel_type =
                  framework::TransPhiKernelKeyToOpKernelType(info_pair.first);
              auto kernel_type_str = framework::KernelTypeToString(kernel_type);
              if (all_kernels_info.count(op_type)) {
                if (std::find(all_kernels_info[op_type].begin(),
                              all_kernels_info[op_type].end(),
                              kernel_type_str) ==
                    all_kernels_info[op_type].end()) {
                  all_kernels_info[op_type].emplace_back(kernel_type_str);
744
                }
745 746
              } else {
                kernel_types.emplace_back(kernel_type_str);
747
              }
748
            }
749 750 751
            if (!kernel_types.empty()) {
              all_kernels_info.emplace(op_type, kernel_types);
            }
752
          }
753
        }
754

755 756 757 758
        return all_kernels_info;
      },
      py::arg("lib") = "all",
      R"DOC(
759 760 761
           Return the registered kernels in paddle.

           Args:
762
               lib[string]: the libarary, could be 'phi', 'fluid' and 'all'.
763
           )DOC");
764

765 766 767 768 769 770
  // NOTE(Aganlengzi): KernelFactory static instance is initialized BEFORE
  // plugins are loaded for custom kernels, but de-initialized AFTER they are
  // unloaded. We need manually clear symbols(may contain plugins' symbols)
  // stored in this static instance to avoid illegal memory access.
  m.def("clear_kernel_factory",
        []() { phi::KernelFactory::Instance().kernels().clear(); });
771 772 773 774 775
  m.def("clear_device_manager", []() {
#ifdef PADDLE_WITH_CUSTOM_DEVICE
    phi::DeviceManager::Clear();
#endif
  });
776

S
sneaxiy 已提交
777 778 779
  // 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
780
  m.def("_set_eager_deletion_mode", &paddle::framework::SetEagerDeletionMode);
S
sneaxiy 已提交
781

782
  m.def("_set_fuse_parameter_group_size",
783
        &paddle::framework::ir::SetFuseParameterGroupsSize);
784
  m.def("_set_fuse_parameter_memory_size",
785
        &paddle::framework::ir::SetFuseParameterMemorySize);
786

S
sneaxiy 已提交
787 788 789
  m.add_object("_cleanup",
               py::capsule([]() { ScopePool::Instance().Clear(); }));

790 791
  m.def("_set_paddle_lib_path", &paddle::platform::dynload::SetPaddleLibPath);

792 793 794
  m.def("_promote_types_if_complex_exists",
        &paddle::framework::PromoteTypesIfComplexExists);

795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819
  py::class_<paddle::CustomOpKernelContext> custom_op_kernel_ctx(
      m, "CustomOpKernelContext", R"DOC()DOC");
  g_custom_op_kernel_ctx_pytype =
      reinterpret_cast<PyTypeObject *>(custom_op_kernel_ctx.ptr());
  custom_op_kernel_ctx.def(py::init<>())
      .def("add_inputs",
           [](paddle::CustomOpKernelContext &self, const py::handle &input) {
             PyObject *obj = input.ptr();
             if (PyList_Check(obj) || PyTuple_Check(obj)) {
               self.EmplaceBackInputs(
                   std::move(CastPyArg2VectorOfTensor(obj, 1)));
             } else {
               self.EmplaceBackInput(std::move(CastPyArg2Tensor(obj, 1)));
             }
           })
      .def("add_outputs",
           [](paddle::CustomOpKernelContext &self, py::handle &outputs) {
             PyObject *obj = outputs.ptr();
             if (PyList_Check(obj) || PyTuple_Check(obj)) {
               self.EmplaceBackOutputs(
                   std::move(CastPyArg2VectorOfTensor(obj, 1)));
             } else {
               self.EmplaceBackOutput(std::move(CastPyArg2Tensor(obj, 1)));
             }
           })
820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835
      .def("add_attr",
           [](paddle::CustomOpKernelContext &self, bool attr) {
             self.EmplaceBackAttr(attr);
           })
      .def("add_attr",
           [](paddle::CustomOpKernelContext &self, int attr) {
             self.EmplaceBackAttr(attr);
           })
      .def("add_attr",
           [](paddle::CustomOpKernelContext &self, float attr) {
             self.EmplaceBackAttr(attr);
           })
      .def("add_attr",
           [](paddle::CustomOpKernelContext &self, int64_t attr) {
             self.EmplaceBackAttr(attr);
           })
836 837 838 839 840 841 842 843 844 845 846 847 848
      .def("add_attr",
           [](paddle::CustomOpKernelContext &self, const std::string &attr) {
             self.EmplaceBackAttr(attr);
           })
      .def("add_attr",
           [](paddle::CustomOpKernelContext &self,
              const std::vector<int> &attr) { self.EmplaceBackAttr(attr); })
      .def("add_attr",
           [](paddle::CustomOpKernelContext &self,
              const std::vector<float> &attr) { self.EmplaceBackAttr(attr); })
      .def("add_attr",
           [](paddle::CustomOpKernelContext &self,
              const std::vector<int64_t> &attr) { self.EmplaceBackAttr(attr); })
849 850 851 852 853
      .def("add_attr",
           [](paddle::CustomOpKernelContext &self,
              const std::vector<std::string> &attr) {
             self.EmplaceBackAttr(attr);
           });
854

855 856
  py::class_<framework::Tensor> framework_tensor(
      m, "Tensor", py::buffer_protocol());
857 858 859
  g_framework_tensor_pytype =
      reinterpret_cast<PyTypeObject *>(framework_tensor.ptr());
  framework_tensor
860 861
      .def("__array__",
           [](framework::Tensor &self) { return TensorToPyArray(self); })
S
sneaxiy 已提交
862 863 864 865
      .def("_ptr",
           [](const framework::Tensor &self) {
             return reinterpret_cast<uintptr_t>(self.data());
           })
J
Jiabin Yang 已提交
866 867
      .def("_slice", &framework::Tensor::Slice)
      .def("_numel", &framework::Tensor::numel)
S
sneaxiy 已提交
868
      .def("_is_initialized",
869
           [](const framework::Tensor &self) { return self.IsInitialized(); })
Y
yuyang18 已提交
870
      .def("_get_dims",
871
           [](const framework::Tensor &self) { return vectorize(self.dims()); })
Y
yuyang18 已提交
872
      .def("_set_dims",
873
           [](framework::Tensor &self, const std::vector<int64_t> &dim) {
874
             self.Resize(phi::make_ddim(dim));
Y
Yu Yang 已提交
875
           })
Y
yuyang18 已提交
876
      .def("_set_layout",
877
           [](framework::Tensor &self, const std::string &layout) {
D
dzhwinter 已提交
878 879
             self.set_layout(StringToDataLayout(layout));
           })
R
ronnywang 已提交
880 881 882 883
      .def("_alloc_float",
           [](framework::Tensor &self, paddle::platform::CustomPlace &place) {
             self.mutable_data<float>(place);
           })
Y
yuyang18 已提交
884
      .def("_alloc_float",
885
           [](framework::Tensor &self, paddle::platform::CUDAPlace &place) {
Q
qijun 已提交
886
             self.mutable_data<float>(place);
Y
Yu Yang 已提交
887
           })
888
      .def("_alloc_float",
889
           [](framework::Tensor &self, paddle::platform::XPUPlace &place) {
890 891
             self.mutable_data<float>(place);
           })
Y
yuyang18 已提交
892
      .def("_alloc_float",
893
           [](framework::Tensor &self, paddle::platform::CPUPlace &place) {
Q
qijun 已提交
894
             self.mutable_data<float>(place);
Y
Yu Yang 已提交
895
           })
896 897 898 899
      .def("_alloc_float",
           [](framework::Tensor &self, paddle::platform::NPUPlace &place) {
             self.mutable_data<float>(place);
           })
900 901 902 903
      .def("_alloc_float",
           [](framework::Tensor &self, paddle::platform::MLUPlace &place) {
             self.mutable_data<float>(place);
           })
904
      .def("_alloc_double",
905
           [](framework::Tensor &self, paddle::platform::CPUPlace &place) {
906 907
             self.mutable_data<double>(place);
           })
Y
yuyang18 已提交
908
      .def("_alloc_int",
909
           [](framework::Tensor &self, paddle::platform::CPUPlace &place) {
Q
qijun 已提交
910
             self.mutable_data<int>(place);
Y
Yu Yang 已提交
911
           })
R
ronnywang 已提交
912 913 914 915
      .def("_alloc_int",
           [](framework::Tensor &self, paddle::platform::CustomPlace &place) {
             self.mutable_data<int>(place);
           })
916
      .def("_alloc_int",
917
           [](framework::Tensor &self, paddle::platform::XPUPlace &place) {
918 919
             self.mutable_data<int>(place);
           })
Y
yuyang18 已提交
920
      .def("_alloc_int",
921
           [](framework::Tensor &self, paddle::platform::CUDAPlace &place) {
Q
qijun 已提交
922
             self.mutable_data<int>(place);
Q
qijun 已提交
923
           })
924 925 926 927
      .def("_alloc_int",
           [](framework::Tensor &self, paddle::platform::MLUPlace &place) {
             self.mutable_data<int>(place);
           })
Y
yuyang18 已提交
928
      .def("_alloc_int",
929 930
           [](framework::Tensor &self,
              paddle::platform::CUDAPinnedPlace &place) {
C
chengduoZH 已提交
931 932
             self.mutable_data<int>(place);
           })
Y
yuyang18 已提交
933
      .def("_alloc_float",
934 935
           [](framework::Tensor &self,
              paddle::platform::CUDAPinnedPlace &place) {
C
chengduoZH 已提交
936 937
             self.mutable_data<float>(place);
           })
938
      .def("_mutable_data",
939 940
           [](framework::Tensor &self,
              paddle::platform::CPUPlace &place,
941
              paddle::framework::proto::VarType::Type type) {
942 943
             return reinterpret_cast<uintptr_t>(
                 self.mutable_data(place, framework::TransToPhiDataType(type)));
944
           })
R
ronnywang 已提交
945
      .def("_mutable_data",
946 947
           [](framework::Tensor &self,
              paddle::platform::CustomPlace &place,
R
ronnywang 已提交
948 949 950 951
              paddle::framework::proto::VarType::Type type) {
             return reinterpret_cast<uintptr_t>(
                 self.mutable_data(place, framework::TransToPhiDataType(type)));
           })
952
      .def("_mutable_data",
953 954
           [](framework::Tensor &self,
              paddle::platform::XPUPlace &place,
955
              paddle::framework::proto::VarType::Type type) {
956 957
             return reinterpret_cast<uintptr_t>(
                 self.mutable_data(place, framework::TransToPhiDataType(type)));
958
           })
959
      .def("_mutable_data",
960 961
           [](framework::Tensor &self,
              paddle::platform::CUDAPlace &place,
962
              paddle::framework::proto::VarType::Type type) {
963 964
             return reinterpret_cast<uintptr_t>(
                 self.mutable_data(place, framework::TransToPhiDataType(type)));
965 966
           })
      .def("_mutable_data",
967 968
           [](framework::Tensor &self,
              paddle::platform::CUDAPinnedPlace &place,
969
              paddle::framework::proto::VarType::Type type) {
970 971
             return reinterpret_cast<uintptr_t>(
                 self.mutable_data(place, framework::TransToPhiDataType(type)));
972
           })
973
      .def("_mutable_data",
974 975
           [](framework::Tensor &self,
              paddle::platform::MLUPlace &place,
976
              paddle::framework::proto::VarType::Type type) {
977 978
             return reinterpret_cast<uintptr_t>(
                 self.mutable_data(place, framework::TransToPhiDataType(type)));
979
           })
980
      .def("_clear", &framework::Tensor::clear)
981
      .def("_mutable_data",
982 983
           [](framework::Tensor &self,
              paddle::platform::NPUPlace &place,
984
              paddle::framework::proto::VarType::Type type) {
985 986
             return reinterpret_cast<uintptr_t>(
                 self.mutable_data(place, framework::TransToPhiDataType(type)));
987
           })
988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 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 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067
      .def("_copy_from",
           &TensorCopyFrom<paddle::platform::CPUPlace>,
           py::arg("tensor"),
           py::arg("place"),
           py::arg("batch_size") = -1)
      .def("_copy_from",
           &TensorCopyFrom<paddle::platform::CustomPlace>,
           py::arg("tensor"),
           py::arg("place"),
           py::arg("batch_size") = -1)
      .def("_copy_from",
           &TensorCopyFrom<paddle::platform::XPUPlace>,
           py::arg("tensor"),
           py::arg("place"),
           py::arg("batch_size") = -1)
      .def("_copy_from",
           &TensorCopyFrom<paddle::platform::CUDAPlace>,
           py::arg("tensor"),
           py::arg("place"),
           py::arg("batch_size") = -1)
      .def("_copy_from",
           &TensorCopyFrom<paddle::platform::NPUPlace>,
           py::arg("tensor"),
           py::arg("place"),
           py::arg("batch_size") = -1)
      .def("_copy_from",
           &TensorCopyFrom<paddle::platform::CUDAPinnedPlace>,
           py::arg("tensor"),
           py::arg("place"),
           py::arg("batch_size") = -1)
      .def("_copy_from",
           &TensorCopyFrom<paddle::platform::MLUPlace>,
           py::arg("tensor"),
           py::arg("place"),
           py::arg("batch_size") = -1)
      .def("_copy_from",
           &TensorCopyFrom<paddle::platform::Place>,
           py::arg("tensor"),
           py::arg("place"),
           py::arg("batch_size") = -1)
      .def("set",
           SetTensorFromPyArray<paddle::platform::CPUPlace>,
           py::arg("array"),
           py::arg("place"),
           py::arg("zero_copy") = false)
      .def("set",
           SetTensorFromPyArray<paddle::platform::CustomPlace>,
           py::arg("array"),
           py::arg("place"),
           py::arg("zero_copy") = false)
      .def("set",
           SetTensorFromPyArray<paddle::platform::XPUPlace>,
           py::arg("array"),
           py::arg("place"),
           py::arg("zero_copy") = false)
      .def("set",
           SetTensorFromPyArray<paddle::platform::CUDAPlace>,
           py::arg("array"),
           py::arg("place"),
           py::arg("zero_copy") = false)
      .def("set",
           SetTensorFromPyArray<paddle::platform::NPUPlace>,
           py::arg("array"),
           py::arg("place"),
           py::arg("zero_copy") = false)
      .def("set",
           SetTensorFromPyArray<paddle::platform::IPUPlace>,
           py::arg("array"),
           py::arg("place"),
           py::arg("zero_copy") = false)
      .def("set",
           SetTensorFromPyArray<paddle::platform::MLUPlace>,
           py::arg("array"),
           py::arg("place"),
           py::arg("zero_copy") = false)
      .def("set",
           SetTensorFromPyArray<paddle::platform::CUDAPinnedPlace>,
           py::arg("array"),
           py::arg("place"),
           py::arg("zero_copy") = false,
1068
           R"DOC(
1069
        Set the data of Tensor on place with given numpy array.
L
Leo Chen 已提交
1070 1071 1072
        
        Args:
          lod (numpy.ndarray): The data to set.
1073
          place (CPUPlace|CUDAPlace|XPUPlace|IPUPlace|CUDAPinnedPlace|NPUPlace|MLUPlace): The place where the
1074
          Tensor is to be set.
1075 1076
          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
1077 1078 1079 1080 1081 1082 1083 1084 1085 1086

        Returns:
            None.

        Examples:
            .. code-block:: python

                import paddle.fluid as fluid
                import numpy as np

1087
                t = fluid.Tensor()
L
Leo Chen 已提交
1088 1089
                t.set(np.ndarray([5, 30]), fluid.CPUPlace())
          )DOC")
1090

1091 1092 1093 1094
      .def(
          "shape",
          [](framework::Tensor &self) { return vectorize(self.dims()); },
          R"DOC(
1095
           Return the shape of Tensor.
L
Leo Chen 已提交
1096 1097

           Returns:
1098
               list[int]: The shape of Tensor.
L
Leo Chen 已提交
1099 1100 1101 1102 1103 1104 1105 1106


           Examples:
               .. code-block:: python

                  import paddle.fluid as fluid
                  import numpy as np

1107
                  t = fluid.Tensor()
L
Leo Chen 已提交
1108 1109 1110
                  t.set(np.ndarray([5, 30]), fluid.CPUPlace())
                  print(t.shape())  # [5, 30]
           )DOC")
6
633WHU 已提交
1111
      .def("_to_dlpack",
1112
           [](framework::Tensor &self) {
6
633WHU 已提交
1113
             DLPackTensor dlpack_tensor(self, 1);
S
Siming Dai 已提交
1114
             DLManagedTensor *dmt = dlpack_tensor.ToDLManagedTensor();
6
633WHU 已提交
1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131
             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
1132 1133 1134 1135
      .def("_set_float_element", TensorSetElement<float>)
      .def("_get_float_element", TensorGetElement<float>)
      .def("_set_double_element", TensorSetElement<double>)
      .def("_get_double_element", TensorGetElement<double>)
1136
      .def("_place", [](framework::Tensor &self) { return self.place(); })
1137 1138 1139 1140
      .def("_dtype",
           [](framework::Tensor &self) {
             return framework::TransToProtoVarType(self.type());
           })
1141
      .def("_layout",
1142 1143 1144 1145
           [](framework::Tensor &self) {
             return DataLayoutToString(self.layout());
           })
      .def("_share_data_with", &framework::Tensor::ShareDataWith)
1146
      .def("__getitem__", PySliceTensor, py::return_value_policy::reference)
1147 1148 1149 1150 1151 1152
      .def("__str__",
           [](const framework::Tensor &self) {
             std::stringstream ostr;
             ostr << self;
             return ostr.str();
           }) /* ------ End of original Tensor ------ */
1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172
      .def("__init__",
           [](framework::Tensor &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);
             PADDLE_ENFORCE_EQ(
                 CheckLoD(new_offset_lod, -1),
                 true,
                 platform::errors::InvalidArgument(
                     "The provided recursive_sequence_lengths info is "
                     "invalid, "
                     "the LoD converted by recursive_sequence_lengths is %s",
                     new_lod));
             new (&instance) framework::Tensor(new_offset_lod);
           })
1173
      .def("__init__",
1174 1175
           [](framework::Tensor &instance) {
             new (&instance) framework::Tensor();
1176
           })
G
gongweibao 已提交
1177
      // We implement offset based LOD in C++ while we use length based with
H
hong 已提交
1178 1179
      // Python API. So we changed set_lod to set_recursive_sequence_lengths
      // to
G
gongweibao 已提交
1180 1181 1182
      // avoid misuse.
      // The discussion is here:
      // https://github.com/PaddlePaddle/Paddle/issues/10855
1183 1184 1185 1186 1187 1188 1189 1190 1191
      .def(
          "set_lod",
          [](framework::Tensor &self,
             const std::vector<std::vector<size_t>> &lod) {
            // the input lod is offset-based level-of-detail info
            LoD new_lod;
            new_lod.reserve(lod.size());
            std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod));
            PADDLE_ENFORCE_EQ(
1192 1193
                CheckLoD(new_lod, vectorize(self.dims()).front()),
                true,
1194 1195 1196 1197
                platform::errors::InvalidArgument(
                    "The provided LoD is invalid, the LoD is %s", new_lod));
            self.set_lod(new_lod);
          },
1198 1199
          py::arg("lod"),
          R"DOC(
1200
           Set LoD of the Tensor.
S
sneaxiy 已提交
1201 1202

           Args:
L
Leo Chen 已提交
1203 1204 1205 1206
               lod (list[list[int]]): The lod to set.

           Returns:
                None.
Z
Zeng Jinle 已提交
1207 1208 1209 1210 1211 1212 1213

           Examples:
               .. code-block:: python

                 import paddle.fluid as fluid
                 import numpy as np

1214
                 t = fluid.Tensor()
Z
Zeng Jinle 已提交
1215 1216
                 t.set(np.ndarray([5, 30]), fluid.CPUPlace())
                 t.set_lod([[0, 2, 5]])
L
Leo Chen 已提交
1217
                 print(t.lod()) # [[0, 2, 5]]
S
sneaxiy 已提交
1218
           )DOC")
1219 1220
      .def(
          "set_recursive_sequence_lengths",
1221 1222 1223
          [](framework::Tensor &self,
             const std::vector<std::vector<size_t>>
                 &recursive_sequence_lengths) {
1224 1225 1226 1227 1228 1229 1230 1231 1232
            // 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);
            PADDLE_ENFORCE_EQ(
1233 1234
                CheckLoD(new_offset_lod, vectorize(self.dims()).front()),
                true,
1235 1236 1237 1238 1239 1240 1241 1242
                platform::errors::InvalidArgument(
                    "The provided recursive_sequence_lengths info is "
                    "invalid, "
                    "the LoD converted by recursive_sequence_lengths is "
                    "%s",
                    new_lod));
            self.set_lod(new_offset_lod);
          },
1243 1244
          py::arg("recursive_sequence_lengths"),
          R"DOC(
1245
           Set LoD of the Tensor according to recursive sequence lengths.
S
sneaxiy 已提交
1246

L
Leo Chen 已提交
1247
           For example, if recursive_sequence_lengths=[[2, 3]], which means
1248
           there are two sequences with length 2 and 3 respectively, the
L
Leo Chen 已提交
1249
           corresponding lod would be [[0, 2, 2+3]], i.e., [[0, 2, 5]].
S
sneaxiy 已提交
1250 1251

           Args:
L
Leo Chen 已提交
1252 1253 1254 1255
                recursive_sequence_lengths (list[list[int]]): The recursive sequence lengths.
           
           Returns:
                None.
Z
Zeng Jinle 已提交
1256 1257 1258 1259 1260 1261 1262

           Examples:
               .. code-block:: python

                 import paddle.fluid as fluid
                 import numpy as np

1263
                 t = fluid.Tensor()
Z
Zeng Jinle 已提交
1264 1265
                 t.set(np.ndarray([5, 30]), fluid.CPUPlace())
                 t.set_recursive_sequence_lengths([[2, 3]])
1266
                 print(t.recursive_sequence_lengths())  # [[2, 3]]
L
Leo Chen 已提交
1267
                 print(t.lod())  # [[0, 2, 5]]
S
sneaxiy 已提交
1268
           )DOC")
1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279
      .def(
          "lod",
          [](framework::Tensor &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;
          },
          R"DOC(
1280
           Return the LoD of the Tensor.
S
sneaxiy 已提交
1281 1282

           Returns:
1283
               list[list[int]]: The lod of the Tensor.
L
Leo Chen 已提交
1284
           
Z
Zeng Jinle 已提交
1285 1286 1287 1288 1289 1290
           Examples:
               .. code-block:: python

                 import paddle.fluid as fluid
                 import numpy as np

1291
                 t = fluid.Tensor()
Z
Zeng Jinle 已提交
1292 1293 1294
                 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
1295
           )DOC")
G
gongweibao 已提交
1296
      // Set above comments of set_lod.
1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307
      .def(
          "recursive_sequence_lengths",
          [](framework::Tensor &self) -> std::vector<std::vector<size_t>> {
            // output the length-based lod info
            LoD lod = phi::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;
          },
          R"DOC(
L
Leo Chen 已提交
1308
           Return the recursive sequence lengths corresponding to of the LodD 
1309
           of the Tensor.
S
sneaxiy 已提交
1310 1311

           Returns:
L
Leo Chen 已提交
1312
                list[list[int]]: The recursive sequence lengths.
Z
Zeng Jinle 已提交
1313 1314 1315 1316 1317 1318 1319

           Examples:
               .. code-block:: python

                 import paddle.fluid as fluid
                 import numpy as np

1320
                 t = fluid.Tensor()
Z
Zeng Jinle 已提交
1321 1322 1323
                 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
1324
           )DOC")
1325 1326 1327 1328 1329 1330 1331 1332
      .def(
          "has_valid_recursive_sequence_lengths",
          [](framework::Tensor &self) -> bool {
            // Check that the lod info is valid and match the outermost
            // dimension of the Tensor data
            return CheckLoD(self.lod(), vectorize(self.dims()).front());
          },
          R"DOC(
1333
           Check whether the LoD of the Tensor is valid.
S
sneaxiy 已提交
1334 1335

           Returns:
L
Leo Chen 已提交
1336
               bool: Whether the LoD is valid.
Z
Zeng Jinle 已提交
1337 1338 1339 1340 1341 1342 1343

           Examples:
               .. code-block:: python

                 import paddle.fluid as fluid
                 import numpy as np

1344
                 t = fluid.Tensor()
Z
Zeng Jinle 已提交
1345 1346 1347
                 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
1348
           )DOC")
L
Leo Chen 已提交
1349
      .def("_as_type",
1350
           [](const framework::Tensor &self,
L
Leo Chen 已提交
1351
              paddle::framework::proto::VarType::Type type) {
1352
             framework::Tensor dst;
L
Leo Chen 已提交
1353 1354 1355 1356 1357
             if (self.IsInitialized() && self.numel() > 0) {
               TransDataType(self, type, &dst);
             }
             return dst;
           })
1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370
      .def("_copy",
           [](const framework::Tensor &self, const platform::Place &place) {
             // follow fetch_op's inplementation
             framework::Tensor 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;
1371
#ifdef _WIN32
1372
           });
1373 1374
#else
           })
1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 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 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655
#ifdef PADDLE_WITH_CUDA
      .def("_share_buffer_with",
           [](framework::Tensor &self, const framework::Tensor src,
              py::tuple t) {
             auto *cuda_ipc_allocation =
                 dynamic_cast<memory::allocation::CudaIpcAllocation *>(
                     src.Holder().get());

             PADDLE_ENFORCE_NOT_NULL(
                 cuda_ipc_allocation,
                 platform::errors::PreconditionNotMet(
                     "Tensor is not Cuda IPC shared tensor. "
                     "Now only Tensor shared by cuda ipc could use this "
                     "api."));

             size_t size = t[0].cast<size_t>();
             auto dtype =
                 static_cast<paddle::experimental::DataType>(t[1].cast<int>());
             auto dims = phi::make_ddim(t[2].cast<std::vector<int>>());
             auto lod_info = t[3].cast<framework::LoD>();
             auto device_id = t[4].cast<int>();

             auto shared_reader_holder =
                 std::make_shared<memory::allocation::Allocation>(
                     cuda_ipc_allocation->ptr(),
                     cuda_ipc_allocation->base_ptr(), size,
                     platform::CUDAPlace(device_id));

             self.ResetHolderWithType(shared_reader_holder, dtype);
             self.Resize(dims);
             self.set_lod(lod_info);

             VLOG(6) << "Reconstructed tensor with buffer shared!";
           },
           R"DOC(
           Deserialize GPU Tensor for existed shared Cuda IPC tensor.

           Params:
               tensor: Shared Cuda IPC tensor.
               tuple: contrains data size, data type,
                      tensor dims, lod information, device index.

       )DOC")
      .def("_share_cuda",
           [](framework::Tensor self) {
             if (!self.IsInitialized() || self.numel() == 0)
               throw std::runtime_error(
                   "Tensor not initialized or numel is 0.  could not pass "
                   "to shared memory. ");

             auto *holder = dynamic_cast<memory::allocation::Allocation *>(
                 self.Holder().get());
             PADDLE_ENFORCE_EQ(
                 platform::is_gpu_place(holder->place()), true,
                 platform::errors::InvalidArgument(
                     "Tensor is not on GPU. share_cuda only support GPU "
                     "Tensor, share_filename is for CPU tensor."));

             void *base_ptr = holder->base_ptr();
             ptrdiff_t offset_bytes = reinterpret_cast<char *>(holder->ptr()) -
                                      reinterpret_cast<char *>(base_ptr);

             cudaIpcMemHandle_t handle;
             PADDLE_ENFORCE_GPU_SUCCESS(cudaIpcGetMemHandle(&handle, base_ptr));

             auto _handle = py::bytes(reinterpret_cast<char *>(&handle),
                                      (py::ssize_t)CUDA_IPC_HANDLE_SIZE);

             // TODO(ZHUI): use cuda event, to avoid sync.
             const auto &device_id = paddle::platform::GetCurrentDeviceId();
             auto stream =
                 paddle::platform::stream::get_current_stream(device_id);
             stream->Synchronize();

             int type_idx = static_cast<int>(self.type());
             size_t data_size =
                 self.numel() *
                 framework::SizeOfType(
                     framework::TransToProtoVarType(self.type()));

             return py::make_tuple(_handle, (py::size_t)offset_bytes, data_size,
                                   type_idx, vectorize(self.dims()), self.lod(),
                                   device_id);
           },
           R"DOC(
           Serialize GPU Tensor by cudaIpcMemHandle.

           Returns:
               tuple: contrains handle, data size, data type,
                      tensor dims, lod information, device index.

           Examples:
               .. code-block:: python

                 import paddle
                 tensor = paddle.ones([3,3])
                 metainfo = tensor.value().get_tensor()._share_cuda()

      )DOC")
      .def("_new_shared_cuda",
           [](py::tuple t) {
             if (t.size() != 7)
               throw std::runtime_error(
                   "Invalid Tensor meta info for shared cuda tensor!");

             // 1. Create a new C++ instance
             framework::Tensor tensor;

             // 2. Rebuild Allocation from handle
             const std::string &handle = t[0].cast<std::string>();
             ptrdiff_t offset_bytes = (ptrdiff_t)t[1].cast<int64_t>();
             auto device_id = t[6].cast<int>();
             auto base_ptr = memory::allocation::GetIpcBasePtr(handle);
             size_t size = t[2].cast<size_t>();
             void *dev = base_ptr.get();
             dev = reinterpret_cast<char *>(dev) + offset_bytes;

             auto shared_reader_holder =
                 std::make_shared<memory::allocation::CudaIpcAllocation>(
                     dev, size, device_id, std::move(base_ptr));

             // 3. Rebuild Tensor
             tensor.ResetHolderWithType(
                 shared_reader_holder,
                 static_cast<paddle::experimental::DataType>(t[3].cast<int>()));
             tensor.Resize(phi::make_ddim(t[4].cast<std::vector<int>>()));
             tensor.set_lod(t[5].cast<framework::LoD>());

             return tensor;
           },
           R"DOC(
           Deserialize GPU lod tensor from cudaIpcMemHandle.

           Params:
               tuple: contrains handle, data size, data type,
                      tensor dims, lod information, device index.

           Examples:
               .. code-block:: python

                 import paddle
                 tensor = paddle.ones([3,3])
                 metainfo = tensor.value().get_tensor()._share_cuda()
                 tensor_from_shared = paddle.to_tensor(paddle.fluid.core.LoDTensor._new_shared_cuda(metainfo))

        )DOC")
#endif
      .def("_share_filename",
           [](framework::Tensor &self) {
             if (!self.IsInitialized() || self.numel() == 0)
               throw std::runtime_error(
                   "Tensor not initialized or numel is 0. could not pass to "
                   "shared memory. ");

             auto holder = self.Holder();
             PADDLE_ENFORCE_EQ(
                 platform::is_cpu_place(holder->place()) ||
                     platform::is_cuda_pinned_place(holder->place()),
                 true, platform::errors::InvalidArgument(
                           "Tensor is not on CPU. share_filename only "
                           "support CPU Tensor."));

             auto *mmap_allocation = dynamic_cast<
                 memory::allocation::RefcountedMemoryMapAllocation *>(
                 holder.get());
             // If the tensor is not shared, allocate memory map allocation.
             if (mmap_allocation == nullptr) {
               void *data_ptr = self.data();
               size_t data_size =
                   self.numel() *
                   framework::SizeOfType(
                       framework::TransToProtoVarType(self.type()));

               int flags = memory::allocation::MAPPED_SHAREDMEM |
                           memory::allocation::MAPPED_EXCLUSIVE;
               std::string handle = memory::allocation::GetIPCName();
               auto shared_holder =
                   memory::allocation::AllocateRefcountedMemoryMapAllocation(
                       handle, flags, data_size);

               // copy data & reset holder
               if (platform::is_cuda_pinned_place(holder->place())) {
#ifdef PADDLE_WITH_CUDA
                 memory::Copy(platform::CPUPlace(), shared_holder->ptr(),
                              platform::CUDAPinnedPlace(), data_ptr, data_size);
#endif
               } else {
                 memory::Copy(platform::CPUPlace(), shared_holder->ptr(),
                              platform::CPUPlace(), data_ptr, data_size);
               }
               self.ResetHolder(shared_holder);
               mmap_allocation = shared_holder.get();
             }
             int type_idx = static_cast<int>(self.type());

             return py::make_tuple(mmap_allocation->ipc_name(),
                                   mmap_allocation->size(), type_idx,
                                   vectorize(self.dims()), self.lod());
           },
           R"DOC(
           Serialize CPU lod tensor in shared memory to tuple.
           If the tensor is not in shared memory, we will copy it first.

           Returns:
               tuple: contrains ipc name, data size, data type,
                      tensor dims and lod imformation.

           Examples:
               .. code-block:: python

                 import paddle
                 tensor = paddle.ones([3,3])
                 metainfo = tensor.value().get_tensor()._share_filename()

       )DOC")
      .def("_new_shared_filename",
           [](py::tuple t) {  // __setstate__
             if (t.size() != 5)
               throw std::runtime_error("Invalid Tensor meta info state!");

             framework::Tensor tensor;

             // 2. Rebuild Allocation
             const std::string &ipc_name = t[0].cast<std::string>();
             size_t size = t[1].cast<size_t>();
             int flags = memory::allocation::MAPPED_SHAREDMEM |
                         memory::allocation::MAPPED_NOCREATE;

             auto shared_holder =
                 memory::allocation::AllocateRefcountedMemoryMapAllocation(
                     ipc_name, flags, size);

             // 3. Rebuild Tensor
             tensor.ResetHolderWithType(
                 shared_holder,
                 static_cast<paddle::experimental::DataType>(t[2].cast<int>()));
             tensor.Resize(phi::make_ddim(t[3].cast<std::vector<int>>()));
             tensor.set_lod(t[4].cast<framework::LoD>());

             return tensor;
           },
           R"DOC(
           Deserialize CPU lod tensor from shared memory.

           Params:
               tuple: contrains ipc file name, data size, data type,
                      tensor dims and lod information.

           Examples:
               .. code-block:: python

                 import paddle
                 tensor = paddle.ones([3,3])
                 metainfo = tensor.value().get_tensor()._share_filename()
                 tensor_from_shared = paddle.to_tensor(paddle.fluid.core.LoDTensor._new_shared_filename(metainfo))

        )DOC")
      .def("_shared_incref",
           [](framework::Tensor &self) {
             auto *mmap_allocation = dynamic_cast<
                 memory::allocation::RefcountedMemoryMapAllocation *>(
                 self.Holder().get());
             if (mmap_allocation) {
               mmap_allocation->incref();
             }
           },
           R"DOC(
            Increase reference count of share_filename tensor.
      )DOC")
      .def("_shared_decref",
           [](framework::Tensor &self) {
             auto *mmap_allocation = dynamic_cast<
                 memory::allocation::RefcountedMemoryMapAllocation *>(
                 self.Holder().get());
             if (mmap_allocation) {
               mmap_allocation->decref();
             }
           },
           R"DOC(
            Decrease reference count of share_filename tensor.
      )DOC")
1656
      .def(py::pickle(
1657
          [](const framework::Tensor &t) {  // __getstate__
1658
            auto holder = t.Holder();
1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670
            PADDLE_ENFORCE_EQ(platform::is_cpu_place(holder->place()), true,
                              platform::errors::PreconditionNotMet(
                                  "Tensor is not on CPU."
                                  "Now only Tensor 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(
                    "Tensor is not in shared memory."
                    "Now only Tensor on shared memory can be serialized."));
1671 1672 1673
            int type_idx = static_cast<int>(t.type());

            return py::make_tuple(mmap_writer_allocation->ipc_name(),
1674 1675
                                  mmap_writer_allocation->size(), type_idx,
                                  vectorize(t.dims()), t.lod());
1676 1677 1678
          },
          [](py::tuple t) {  // __setstate__
            if (t.size() != 5)
1679
              throw std::runtime_error("Invalid Tensor state!");
1680 1681

            // 1. Create a new C++ instance
1682
            framework::Tensor tensor;
1683 1684 1685 1686 1687

            // 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 =
1688 1689
                memory::allocation::RebuildMemoryMapReaderAllocation(ipc_name,
                                                                     size);
1690 1691

            // 3. Maintain global fd set
1692
            VLOG(3) << "Tensor ipc name: " << ipc_name;
1693 1694
            memory::allocation::MemoryMapFdSet::Instance().Insert(ipc_name);

1695 1696 1697
            // 4. Rebuild Tensor
            tensor.ResetHolderWithType(
                shared_reader_holder,
1698
                static_cast<paddle::experimental::DataType>(t[2].cast<int>()));
1699
            tensor.Resize(phi::make_ddim(t[3].cast<std::vector<int>>()));
1700 1701 1702 1703 1704
            tensor.set_lod(t[4].cast<framework::LoD>());

            return tensor;
          }));
#endif
D
dangqingqing 已提交
1705

1706
  py::class_<phi::SelectedRows>(m, "SelectedRows")
Q
qijun 已提交
1707
      .def("__init__",
1708 1709
           [](phi::SelectedRows &instance) {
             new (&instance) phi::SelectedRows();
1710
           })
Q
qijun 已提交
1711
      .def("__init__",
1712 1713
           [](phi::SelectedRows &instance,
              const std::vector<int64_t> rows,
Q
qijun 已提交
1714
              const int64_t &height) {
1715
             new (&instance) phi::SelectedRows(rows, height);
Q
qijun 已提交
1716
           })
1717 1718 1719 1720
      .def(
          "get_tensor",
          [](phi::SelectedRows &self) { return self.mutable_value(); },
          py::return_value_policy::reference)
1721
      .def("numel",
1722
           [](phi::SelectedRows &self) -> int64_t {
1723 1724
             return self.value().numel();
           })
1725 1726
      .def("set_height", &phi::SelectedRows::set_height)
      .def("height", &phi::SelectedRows::height)
Q
qijun 已提交
1727
      .def("set_rows",
1728
           [](phi::SelectedRows &self, std::vector<int64_t> rows) {
1729
#if !defined(PADDLE_WITH_CUDA) && !defined(PADDLE_WITH_HIP)
Q
qijun 已提交
1730 1731 1732 1733 1734 1735
             self.set_rows(rows);
#else
        Vector<int64_t> new_rows(rows);
        self.set_rows(new_rows);
#endif
           })
1736
      .def("sync_index",
1737 1738
           [](phi::SelectedRows &instance) { instance.SyncIndex(); })
      .def("rows", [](phi::SelectedRows &self) {
1739 1740 1741 1742 1743
        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;
1744
      });
Q
qijun 已提交
1745

1746
  py::class_<Variable>(m, "Variable", R"DOC(Variable Class.
1747 1748 1749

All parameter, weight, gradient are variables in Paddle.
)DOC")
S
sneaxiy 已提交
1750
      .def(py::init<>())
1751
      .def("is_int", [](const Variable &var) { return var.IsType<int>(); })
1752
      .def("set_int",
1753 1754
           [](Variable &var, int val) -> void { *var.GetMutable<int>() = val; })
      .def("get_int", [](const Variable &var) -> int { return var.Get<int>(); })
1755 1756 1757 1758 1759 1760 1761
      .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>(); })
1762 1763 1764 1765 1766 1767
      .def(
          "get_tensor",
          [](Variable &self) -> LoDTensor * {
            return self.GetMutable<LoDTensor>();
          },
          py::return_value_policy::reference)
1768 1769 1770 1771
      .def("get_bytes",
           [](Variable &self) {
             return py::bytes(*self.GetMutable<std::string>());
           })
S
Steffy-zxf 已提交
1772 1773 1774 1775
      .def("set_string_list",
           [](Variable &self, Strings str_list) {
             *self.GetMutable<Strings>() = str_list;
           })
1776 1777 1778 1779
      .def("set_vocab",
           [](Variable &self, Vocab vocab) {
             *self.GetMutable<Vocab>() = vocab;
           })
1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805
      .def(
          "get_string_tensor",
          [](Variable &self) { return self.GetMutable<Strings>(); },
          py::return_value_policy::reference)
      .def(
          "get_map_tensor",
          [](Variable &self) { return self.GetMutable<Vocab>(); },
          py::return_value_policy::reference)
      .def(
          "get_lod_rank_table",
          [](Variable &self) { return self.GetMutable<LoDRankTable>(); },
          py::return_value_policy::reference)
      .def(
          "get_selected_rows",
          [](Variable &self) -> phi::SelectedRows * {
            return self.GetMutable<phi::SelectedRows>();
          },
          py::return_value_policy::reference)
      .def(
          "get_lod_tensor_array",
          [](Variable &self) { return self.GetMutable<LoDTensorArray>(); },
          py::return_value_policy::reference)
      .def(
          "get_fetch_list",
          [](Variable &self) { return self.GetMutable<FetchList>(); },
          py::return_value_policy::reference)
1806
#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
1807 1808 1809 1810 1811 1812
      .def(
          "get_communicator",
          [](Variable &self) -> platform::Communicator * {
            return self.GetMutable<platform::Communicator>();
          },
          py::return_value_policy::reference)
Y
Yu Yang 已提交
1813
#endif
1814 1815 1816
      .def(
          "get_reader",
          [](Variable &self) -> framework::ReaderHolder * {
1817 1818
            PADDLE_ENFORCE_EQ(self.IsType<framework::ReaderHolder>(),
                              true,
1819 1820 1821 1822 1823 1824 1825 1826 1827 1828
                              platform::errors::InvalidArgument(
                                  "The variable is not type of ReaderHolder."));
            return self.GetMutable<framework::ReaderHolder>();
          },
          py::return_value_policy::reference)
      .def(
          "get_scope",
          [](Variable &self) -> Scope * {
            auto scope_vec = self.GetMutable<std::vector<framework::Scope *>>();
            PADDLE_ENFORCE_GT(
1829 1830
                scope_vec->size(),
                0,
1831 1832 1833 1834 1835
                platform::errors::InvalidArgument(
                    "The size of scope_vec should be greater than 0"));
            return scope_vec->front();
          },
          py::return_value_policy::reference)
1836 1837 1838 1839
      .def("set_scope", [](Variable &self, Scope &scope) {
        auto scope_vec = self.GetMutable<std::vector<framework::Scope *>>();
        scope_vec->emplace_back(&scope);
      });
1840

S
sneaxiy 已提交
1841
  BindReader(&m);
Y
Refine  
Yu Yang 已提交
1842

0
0x45f 已提交
1843
  py::class_<Scope> _Scope(m, "_Scope", R"DOC(
Q
Qiao Longfei 已提交
1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856
    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

1857
          import paddle.fluid as fluid
Q
Qiao Longfei 已提交
1858 1859 1860 1861 1862
          # 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)

0
0x45f 已提交
1863 1864 1865
        )DOC");
  g_framework_scope_pytype = reinterpret_cast<PyTypeObject *>(_Scope.ptr());
  _Scope
S
sneaxiy 已提交
1866 1867
      .def("_remove_from_pool",
           [](Scope &self) { ScopePool::Instance().Remove(&self); })
1868 1869 1870 1871 1872 1873 1874
      .def(
          "var",
          [](Scope &self, const std::string &name) -> Variable * {
            return self.Var(name);
          },
          py::arg("name"),
          R"DOC(
1875
           Find or create variable named :code:`name` in the current scope.
S
sneaxiy 已提交
1876

1877
           If the variable named :code:`name` does not exist in the
S
sneaxiy 已提交
1878
           current scope, the variable would be created. Otherwise,
1879
           return the existing variable.
S
sneaxiy 已提交
1880 1881

           Args:
1882 1883
               name (str): the variable name.

S
sneaxiy 已提交
1884
           Returns:
1885
               out (core.Variable): the found or created variable.
S
sneaxiy 已提交
1886
           )DOC",
1887
          py::return_value_policy::reference)
1888 1889 1890
      .def("find_var",
           &Scope::FindVar,
           py::arg("name"),
S
sneaxiy 已提交
1891
           R"DOC(
1892
           Find variable named :code:`name` in the current scope or
1893
           its parent scope. Return None if not found. 
1894

S
sneaxiy 已提交
1895 1896
           Args:
               name (str): the variable name.
1897

S
sneaxiy 已提交
1898
           Returns:
1899
               out (core.Variable|None): the found variable or None.
S
sneaxiy 已提交
1900
           )DOC",
1901
           py::return_value_policy::reference)
1902
      .def("size", &Scope::Size)
1903 1904 1905
      .def("erase",
           &Scope::EraseVars,
           py::arg("names"),
1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916
           R"DOC(
           Find variable named :code:`name` in the current scope or
           its parent scope. Return None if not found. 

           Args:
               name (str): the variable names to be erase.

           Returns:
               None
           )DOC",
           py::return_value_policy::reference)
1917
      .def(
1918 1919
          "new_scope",
          [](Scope &self) -> Scope * { return &self.NewScope(); },
1920
          R"DOC(
S
sneaxiy 已提交
1921 1922 1923 1924 1925
           Create a new sub-scope of the current scope.

           Returns:
               out (core._Scope): the created sub-scope.
           )DOC",
1926
          py::return_value_policy::reference)
1927 1928
      .def("drop_kids",
           &Scope::DropKids,
S
sneaxiy 已提交
1929 1930
           R"DOC(
           Delete all sub-scopes of the current scope.
S
sneaxiy 已提交
1931 1932
           )DOC")
      .def("_kids", &Scope::kids);
1933

1934 1935 1936 1937 1938 1939 1940 1941
  m.def(
      "Scope",
      []() -> Scope * {
        auto *s = new Scope();
        ScopePool::Instance().Insert(std::unique_ptr<Scope>(s));
        return s;
      },
      R"DOC(
S
sneaxiy 已提交
1942
        Create a new scope.
1943

S
sneaxiy 已提交
1944 1945 1946
        Returns:
            out (core._Scope): the created scope.
        )DOC",
1947
      py::return_value_policy::reference);
S
sneaxiy 已提交
1948

Y
Yu Yang 已提交
1949 1950
  //! @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
1951 1952
  m.def("get_all_op_protos", []() -> std::vector<py::bytes> {
    std::vector<py::bytes> ret_values;
1953 1954 1955 1956
    for (auto &iter : OpInfoMap::Instance().map()) {
      auto &info = iter.second;
      if (info.HasOpProtoAndChecker()) {
        std::string str;
C
chengduo 已提交
1957
        PADDLE_ENFORCE_EQ(
1958 1959
            info.Proto().SerializeToString(&str),
            true,
1960 1961
            platform::errors::Fatal(
                "Serialize OpProto Error. This could be a bug of Paddle."));
1962 1963 1964
        ret_values.emplace_back(str);
      }
    }
Y
Yu Yang 已提交
1965 1966
    return ret_values;
  });
1967 1968 1969 1970 1971 1972 1973
  m.def("get_all_op_names", []() {
    std::vector<std::string> op_names;
    for (auto &iter : OpInfoMap::Instance().map()) {
      op_names.emplace_back(iter.first);
    }
    return op_names;
  });
1974 1975 1976 1977 1978 1979 1980 1981
  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();
1982
              res = op_checker->GetDefaultAttrsMap();
1983 1984 1985 1986
            }
          }
          return res;
        });
1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004
  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);
        });
2005 2006 2007
  m.def("has_grad_op_maker", [](const std::string op_type) {
    return framework::OpInfoMap::Instance().Get(op_type).HasGradOpMaker();
  });
2008 2009 2010 2011 2012
  m.def("has_non_empty_grad_op_maker", [](const std::string op_type) {
    return framework::OpInfoMap::Instance()
        .Get(op_type)
        .HasNonEmptyGradOpMaker();
  });
2013 2014 2015
  m.def("has_infer_inplace", [](const std::string op_type) {
    return framework::OpInfoMap::Instance().Get(op_type).HasInferInplace();
  });
2016
  m.def("infer_no_need_buffer_slots",
2017 2018
        [](const std::string op_type,
           const framework::VariableNameMap &inputs,
2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030
           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;
          }
        });
2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045
  m.def("prune",
        [](const ProgramDesc &origin,
           const std::set<std::string> &feeded_var_names,
           const std::vector<std::array<size_t, 2>> &targets) {
          ProgramDesc prog_with_targets(origin);

          for (const auto &t : targets) {
            prog_with_targets.MutableBlock(t[0])->Op(t[1])->SetIsTarget(true);
          }
          proto::ProgramDesc pruned_desc;
          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);
        });
2046 2047 2048 2049 2050 2051
  m.def(
      "prune_backward",
      [](const framework::ProgramDesc &program) {
        return PruneBackward(program);
      },
      R"DOC(
2052 2053 2054
             Prune the backward part of a program, mostly called in
             program.clone(for_test=True).
              
2055
            Args:
2056 2057 2058 2059 2060 2061 2062 2063
                   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");
2064 2065 2066 2067
  m.def("get_serialize_comile_key", [](int64_t compilation_key) {
#ifdef PADDLE_WITH_CINN
    auto compiler = framework::paddle2cinn::CinnCompiler::GetInstance();
    auto s = compiler->SerializeKey(compilation_key);
2068 2069
    VLOG(4) << s;
    return s;
2070 2071 2072 2073 2074 2075
#else
    PADDLE_THROW(
                 platform::errors::PermissionDenied(
                 "Cannot get compilation key in non-CINN version, "
                 "Please recompile or reinstall Paddle with CINN support."));
#endif
2076
  });
2077 2078 2079 2080
  m.def("empty_var_name",
        []() { return std::string(framework::kEmptyVarName); });
  m.def("grad_var_suffix",
        []() { return std::string(framework::kGradVarSuffix); });
2081 2082 2083
  m.def_submodule(
       "var_names",
       "The module will return special predefined variable name in Paddle")
Y
Yi Wang 已提交
2084 2085
      .def("empty", []() { return kEmptyVarName; })
      .def("temp", []() { return kTempVarName; });
2086

Q
qijun 已提交
2087
  // clang-format off
Y
Yu Yang 已提交
2088
  py::class_<paddle::platform::DeviceContext>(m, "DeviceContext")
Q
qijun 已提交
2089 2090
      .def_static("create",
                  [](paddle::platform::CPUPlace& place)
Q
qijun 已提交
2091
                      -> paddle::platform::DeviceContext* {
W
Wilber 已提交
2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105
    auto* context = new paddle::platform::CPUDeviceContext();
    context->SetAllocator(
      paddle::memory::allocation::AllocatorFacade::Instance()
        .GetAllocator(place)
        .get());
    context->SetHostAllocator(
      paddle::memory::allocation::AllocatorFacade::Instance()
        .GetAllocator(paddle::platform::CPUPlace())
        .get());
    context->SetZeroAllocator(
      paddle::memory::allocation::AllocatorFacade::Instance()
        .GetZeroAllocator(place)
        .get());
    return context;
Q
qijun 已提交
2106
                  })
2107 2108 2109 2110 2111 2112 2113 2114 2115
      .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
W
Wilber 已提交
2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129
      auto* context = new paddle::platform::XPUDeviceContext(place);
      context->SetAllocator(
        paddle::memory::allocation::AllocatorFacade::Instance()
          .GetAllocator(place)
          .get());
      context->SetHostAllocator(
        paddle::memory::allocation::AllocatorFacade::Instance()
          .GetAllocator(paddle::platform::CPUPlace())
          .get());
      context->SetZeroAllocator(
        paddle::memory::allocation::AllocatorFacade::Instance()
          .GetZeroAllocator(place)
          .get());
      return context;
2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141
#endif
                  })
        .def_static("create",
                  [](paddle::platform::MLUPlace& place)
                      -> paddle::platform::DeviceContext* {
#ifndef PADDLE_WITH_MLU
             PADDLE_THROW(
                 platform::errors::PermissionDenied(
                 "Cannot use MLUPlace in CPU/GPU version, "
                 "Please recompile or reinstall Paddle with MLU support."));
#else
                    return new paddle::platform::MLUDeviceContext(place);
2142 2143
#endif
                  })
2144 2145 2146 2147 2148 2149 2150 2151 2152 2153
        .def_static("create",
                    [](paddle::platform::NPUPlace& place)
                        -> paddle::platform::DeviceContext* {
#ifndef PADDLE_WITH_ASCEND_CL
             PADDLE_THROW(
                 platform::errors::PermissionDenied(
                 "Cannot use NPUPlace in CPU/GPU/XPU version, "
                 "Please recompile or reinstall Paddle with NPU support."));
#else
                return new paddle::platform::NPUDeviceContext(place);
R
ronnywang 已提交
2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166
#endif
        })
        .def_static("create",
                    [](paddle::platform::CustomPlace& place)
                        -> paddle::platform::DeviceContext* {
#ifndef PADDLE_WITH_CUSTOM_DEVICE
             PADDLE_THROW(
                 platform::errors::PermissionDenied(
                 "Cannot use CustomPlace in CPU/GPU/XPU version, "
                 "Please recompile or reinstall Paddle with "
                 "CustomDevice support."));
#else
                return new paddle::platform::CustomDeviceContext(place);
2167 2168
#endif
        })
Q
qijun 已提交
2169
      .def_static("create",
D
dzhwinter 已提交
2170
                  [](paddle::platform::CUDAPlace& place)
Q
qijun 已提交
2171
                      -> paddle::platform::DeviceContext* {
2172
#if !defined(PADDLE_WITH_CUDA) && !defined(PADDLE_WITH_HIP)
2173 2174 2175 2176
             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
2177
#else
W
Wilber 已提交
2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190
      auto* context = new paddle::platform::CUDADeviceContext(place);
      context->SetAllocator(
        paddle::memory::allocation::AllocatorFacade::Instance()
          .GetAllocator(place, context->stream())
          .get());
      context->SetHostAllocator(
        paddle::memory::allocation::AllocatorFacade::Instance()
          .GetAllocator(paddle::platform::CPUPlace())
          .get());
      context->SetZeroAllocator(
        paddle::memory::allocation::AllocatorFacade::Instance()
        .GetZeroAllocator(place)
        .get());
W
wanghuancoder 已提交
2191 2192 2193 2194
      context->SetPinnedAllocator(
        paddle::memory::allocation::AllocatorFacade::Instance()
          .GetAllocator(paddle::platform::CUDAPinnedPlace())
          .get());
W
Wilber 已提交
2195 2196
      context->PartialInitWithAllocator();
      return context;
Q
qijun 已提交
2197
#endif
C
chengduoZH 已提交
2198 2199 2200 2201
                  })
          .def_static("create",
                [](paddle::platform::CUDAPinnedPlace& place)
                        -> paddle::platform::DeviceContext* {
2202
#if !defined(PADDLE_WITH_CUDA) && !defined(PADDLE_WITH_HIP)
2203 2204 2205 2206
             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
2207 2208 2209 2210
#else
                  return new paddle::platform::CUDAPinnedDeviceContext(place);
#endif
                });;
D
Dong Zhihong 已提交
2211
// clang-format on
2212
#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
D
Dong Zhihong 已提交
2213 2214
  py::class_<platform::Communicator>(m, "Communicator").def(py::init<>());
#endif
2215 2216 2217
  m.def("get_all_device_type", []() {
    std::vector<std::string> device_types;
#ifdef PADDLE_WITH_CUSTOM_DEVICE
2218
    device_types = phi::DeviceManager::GetAllDeviceTypes();
2219
#else
R
ronnywang 已提交
2220
          VLOG(1) << string::Sprintf(
2221 2222 2223 2224
              "Cannot use get_all_device_type because you have installed"
              "CPU/GPU version PaddlePaddle.\n"
              "If you want to use get_all_device_type, please try to install"
              "CustomDevice version "
R
ronnywang 已提交
2225
              "PaddlePaddle by: pip install paddlepaddle\n");
2226 2227 2228 2229 2230 2231
#endif
    return device_types;
  });
  m.def("get_all_custom_device_type", []() {
    std::vector<std::string> device_types;
#ifdef PADDLE_WITH_CUSTOM_DEVICE
2232
    device_types = phi::DeviceManager::GetAllCustomDeviceTypes();
2233
#else
R
ronnywang 已提交
2234
          VLOG(1) << string::Sprintf(
2235 2236 2237 2238
              "Cannot use get_all_custom_device_type because you have installed"
              "CPU/GPU version PaddlePaddle.\n"
              "If you want to use get_all_custom_device_type, please try to "
              "install CustomDevice version "
R
ronnywang 已提交
2239
              "PaddlePaddle by: pip install paddlepaddle\n");
2240 2241 2242 2243 2244 2245
#endif
    return device_types;
  });
  m.def("get_available_device", [] {
    std::vector<std::string> devices;
#ifdef PADDLE_WITH_CUSTOM_DEVICE
2246
    devices = phi::DeviceManager::GetAllDeviceList();
2247
#else
R
ronnywang 已提交
2248
          VLOG(1) << string::Sprintf(
2249 2250 2251 2252
              "Cannot use get_available_device because you have installed"
              "CPU/GPU version PaddlePaddle.\n"
              "If you want to use get_available_device, please try to install"
              "CustomDevice version "
R
ronnywang 已提交
2253
              "PaddlePaddle by: pip install paddlepaddle\n");
2254 2255 2256 2257 2258 2259
#endif
    return devices;
  });
  m.def("get_available_custom_device", [] {
    std::vector<std::string> devices;
#ifdef PADDLE_WITH_CUSTOM_DEVICE
2260
    devices = phi::DeviceManager::GetAllCustomDeviceList();
2261
#else
R
ronnywang 已提交
2262
          VLOG(1) << string::Sprintf(
2263 2264 2265 2266 2267 2268
              "Cannot use get_available_custom_device because you have "
              "installed"
              "CPU/GPU version PaddlePaddle.\n"
              "If you want to use get_available_custom_device, please try to "
              "install"
              "CustomDevice version "
R
ronnywang 已提交
2269
              "PaddlePaddle by: pip install paddlepaddle\n");
2270 2271 2272
#endif
    return devices;
  });
2273 2274
  py::class_<platform::CustomPlace> customplace(m,
                                                "CustomPlace",
2275
                                                R"DOC(
2276 2277 2278 2279 2280 2281 2282 2283
    CustomPlace is a descriptor of a device.
    It represents a custom device on which a tensor will be allocated and a model will run.

    Examples:
        .. code-block:: python

          import paddle
          fake_cpu_place = paddle.CustomPlace("FakeCPU", 0)
2284 2285 2286
                                             )DOC");
  g_customplace_pytype = reinterpret_cast<PyTypeObject *>(customplace.ptr());
  customplace
2287
      .def("__init__",
2288 2289
           [](platform::CustomPlace &self,
              const std::string &device_type,
2290 2291 2292 2293 2294 2295 2296
              int dev_id) {
#ifdef PADDLE_WITH_CUSTOM_DEVICE
             if (UNLIKELY(dev_id < 0)) {
               LOG(ERROR) << string::Sprintf(
                   "Invalid CustomPlace(%s, %d), device id must be 0 "
                   "or "
                   "positive integer",
2297 2298
                   device_type,
                   dev_id);
2299 2300 2301
               std::exit(-1);
             }

2302 2303
             if (LIKELY(phi::DeviceManager::HasDeviceType(device_type) &&
                        phi::DeviceManager::IsCustom(device_type))) {
2304
               int dev_count = static_cast<int>(
2305
                   phi::DeviceManager::GetDeviceCount(device_type));
2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318
               if (UNLIKELY(dev_id >= dev_count)) {
                 if (dev_count == 0) {
                   LOG(ERROR) << "Cannot use " << device_type
                              << " because there is no " << device_type
                              << " detected on your "
                                 "machine.";
                   std::exit(-1);
                 } else {
                   LOG(ERROR) << string::Sprintf(
                       "Invalid CustomPlace(%s, %d), dev_id must "
                       "inside "
                       "[0, %d), because %s "
                       "number on your machine is %d",
2319 2320 2321 2322 2323
                       device_type,
                       dev_id,
                       dev_count,
                       device_type,
                       dev_count);
2324 2325 2326 2327 2328 2329 2330 2331 2332
                   std::exit(-1);
                 }
               }
               new (&self) platform::CustomPlace(device_type, dev_id);
             } else {
               LOG(ERROR) << string::Sprintf(
                   "Invalid CustomPlace(%s, %d), the device type is "
                   "not registered "
                   "as a custom device.",
2333 2334
                   device_type,
                   dev_id);
2335 2336 2337 2338 2339 2340 2341 2342
               std::exit(-1);
             }
#else
             LOG(ERROR) << string::Sprintf(
                 "Cannot use CustomDevice because you have installed CPU/GPU"
                 "version PaddlePaddle.\n"
                 "If you want to use CustomDevice, please try to install"
                 "CustomDevice version "
R
ronnywang 已提交
2343
                 "PaddlePaddle by: pip install paddlepaddle\n"
2344 2345 2346 2347 2348 2349
                 "If you only have CPU, please change "
                 "CustomPlace(%s, %d) to be CPUPlace().\n",
                 device_type, dev_id);
             std::exit(-1);
#endif
           })
2350
      .def("_type", &PlaceIndex<platform::CustomPlace>)
2351 2352 2353 2354 2355 2356 2357 2358
      .def("get_device_id",
           [](const platform::CustomPlace &self) { return self.GetDeviceId(); })
      .def("get_device_type",
           [](const platform::CustomPlace &self) {
             return self.GetDeviceType();
           })
      .def("__repr__", string::to_string<const platform::CustomPlace &>)
      .def("__str__", string::to_string<const platform::CustomPlace &>);
2359
  py::class_<platform::CUDAPlace> cudaplace(m, "CUDAPlace", R"DOC(
2360 2361 2362 2363 2364

    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.
2365
    The memory of CUDAPlace with different dev_id is not accessible.
2366 2367 2368 2369 2370 2371 2372 2373
    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
2374 2375 2376 2377

    Examples:
        .. code-block:: python

2378 2379 2380
          import paddle

          place = paddle.CUDAPlace(0)
L
lujun 已提交
2381

2382 2383 2384
        )DOC");
  g_cudaplace_pytype = reinterpret_cast<PyTypeObject *>(cudaplace.ptr());
  cudaplace
S
sneaxiy 已提交
2385 2386
      .def("__init__",
           [](platform::CUDAPlace &self, int dev_id) {
2387
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
2388 2389 2390 2391 2392 2393 2394 2395
             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);
             }

2396 2397
             if (UNLIKELY(dev_id >= platform::GetGPUDeviceCount())) {
               if (platform::GetGPUDeviceCount() == 0) {
2398 2399 2400 2401 2402 2403 2404 2405
                 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",
2406 2407
                     dev_id,
                     platform::GetGPUDeviceCount(),
2408
                     platform::GetGPUDeviceCount());
2409 2410 2411 2412
                 std::exit(-1);
               }
             }

S
sneaxiy 已提交
2413 2414
             new (&self) platform::CUDAPlace(dev_id);
#else
2415 2416 2417 2418 2419 2420 2421 2422 2423
             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
2424 2425
#endif
           })
2426
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
2427 2428
      .def("get_device_id",
           [](const platform::CUDAPlace &self) { return self.GetDeviceId(); })
S
sneaxiy 已提交
2429 2430 2431 2432
      .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>)
2433
      .def("_equals", &IsSamePlace<platform::CUDAPlace, platform::XPUPlace>)
2434
      .def("_equals", &IsSamePlace<platform::CUDAPlace, platform::NPUPlace>)
2435
      .def("_equals", &IsSamePlace<platform::CUDAPlace, platform::MLUPlace>)
S
sneaxiy 已提交
2436 2437
      .def("_equals",
           &IsSamePlace<platform::CUDAPlace, platform::CUDAPinnedPlace>)
2438 2439 2440
      .def("_get_device_id",
           [](platform::CUDAPlace &self) -> int { return self.GetDeviceId(); })
#endif
2441
      .def("__repr__", string::to_string<const platform::CUDAPlace &>)
D
dzhwinter 已提交
2442
      .def("__str__", string::to_string<const platform::CUDAPlace &>);
Q
qijun 已提交
2443

2444
  py::class_<platform::XPUPlace> xpuplace(m, "XPUPlace", R"DOC(
2445 2446 2447 2448 2449
    **Note**:
    Examples:
        .. code-block:: python
          import paddle.fluid as fluid
          xpu_place = fluid.XPUPlace(0)
2450 2451 2452
        )DOC");
  g_xpuplace_pytype = reinterpret_cast<PyTypeObject *>(xpuplace.ptr());
  xpuplace
2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468 2469 2470 2471 2472
      .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",
2473 2474
                     dev_id,
                     platform::GetXPUDeviceCount(),
2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491
                     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
           })
2492
#ifdef PADDLE_WITH_XPU
2493 2494 2495 2496 2497 2498 2499
      .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>)
2500 2501 2502
      .def("get_device_id",
           [](const platform::XPUPlace &self) { return self.GetDeviceId(); })
#endif
2503
      .def("__repr__", string::to_string<const platform::XPUPlace &>)
2504
      .def("__str__", string::to_string<const platform::XPUPlace &>);
2505
#ifdef PADDLE_WITH_XPU
2506 2507 2508
  py::enum_<phi::backends::xpu::XPUVersion>(m, "XPUVersion", py::arithmetic())
      .value("XPU1", phi::backends::xpu::XPUVersion::XPU1)
      .value("XPU2", phi::backends::xpu::XPUVersion::XPU2)
T
TTerror 已提交
2509
      .export_values();
2510
  m.def("get_xpu_device_count", platform::GetXPUDeviceCount);
T
TTerror 已提交
2511 2512
  m.def("get_xpu_device_version",
        [](int device_id) { return platform::get_xpu_version(device_id); });
L
Lijunhui 已提交
2513 2514 2515 2516 2517 2518
#ifdef PADDLE_WITH_XPU_KP
  m.def("get_xpu_device_op_support_types",
        [](const std::string &op_name, phi::backends::xpu::XPUVersion version) {
          return platform::get_xpu_kp_op_support_type(op_name, version);
        });
#else
2519 2520 2521 2522
  m.def("get_xpu_device_op_support_types",
        [](const std::string &op_name, phi::backends::xpu::XPUVersion version) {
          return platform::get_xpu_op_support_type(op_name, version);
        });
L
Lijunhui 已提交
2523
#endif
2524
  m.def("get_xpu_device_op_list", [](phi::backends::xpu::XPUVersion version) {
T
TTerror 已提交
2525 2526
    return platform::get_xpu_op_list(version);
  });
T
taixiurong 已提交
2527 2528
  m.def("is_float16_supported", [](const platform::XPUPlace &place) -> bool {
    // XPUs with Compute Capability > xpu2 support float16 and bfloat16
W
Wilber 已提交
2529
    return platform::get_xpu_version(place.device) >
2530
           phi::backends::xpu::XPUVersion::XPU1;
T
taixiurong 已提交
2531 2532 2533
  });
  m.def("is_bfloat16_supported", [](const platform::XPUPlace &place) -> bool {
    // XPUs with Compute Capability > xpu2 support float16 and bfloat16
W
Wilber 已提交
2534
    return platform::get_xpu_version(place.device) >
2535
           phi::backends::xpu::XPUVersion::XPU1;
T
taixiurong 已提交
2536
  });
2537
#endif
2538

2539
  py::class_<paddle::platform::CPUPlace> cpuplace(m, "CPUPlace", R"DOC(
2540
    CPUPlace is a descriptor of a device.
2541
    It represents a CPU device on which a tensor will be allocated and a model will run.
L
lujun 已提交
2542 2543 2544 2545

    Examples:
        .. code-block:: python

2546 2547
          import paddle
          cpu_place = paddle.CPUPlace()
L
lujun 已提交
2548

2549 2550 2551
        )DOC");
  g_cpuplace_pytype = reinterpret_cast<PyTypeObject *>(cpuplace.ptr());
  cpuplace.def(py::init<>())
S
sneaxiy 已提交
2552 2553
      .def("_type", &PlaceIndex<platform::CPUPlace>)
      .def("_equals", &IsSamePlace<platform::CPUPlace, platform::Place>)
2554
      .def("_equals", &IsSamePlace<platform::CPUPlace, platform::XPUPlace>)
2555
      .def("_equals", &IsSamePlace<platform::CPUPlace, platform::NPUPlace>)
S
sneaxiy 已提交
2556 2557 2558 2559
      .def("_equals", &IsSamePlace<platform::CPUPlace, platform::CUDAPlace>)
      .def("_equals", &IsSamePlace<platform::CPUPlace, platform::CPUPlace>)
      .def("_equals",
           &IsSamePlace<platform::CPUPlace, platform::CUDAPinnedPlace>)
2560
      .def("__repr__", string::to_string<const platform::CPUPlace &>)
2561
      .def("__str__", string::to_string<const platform::CPUPlace &>);
Y
Yu Yang 已提交
2562

2563 2564
  py::class_<paddle::platform::CUDAPinnedPlace> cudapinnedplace(
      m, "CUDAPinnedPlace", R"DOC(
2565 2566 2567 2568 2569 2570
    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
2571 2572 2573 2574

    Examples:
        .. code-block:: python

2575 2576
          import paddle
          place = paddle.CUDAPinnedPlace()
L
lujun 已提交
2577

2578 2579 2580 2581
        )DOC");
  g_cudapinnedplace_pytype =
      reinterpret_cast<PyTypeObject *>(cudapinnedplace.ptr());
  cudapinnedplace
S
sneaxiy 已提交
2582
      .def("__init__",
S
sneaxiy 已提交
2583
           [](platform::CUDAPinnedPlace &self) {
2584
#if !defined(PADDLE_WITH_CUDA) && !defined(PADDLE_WITH_HIP)
2585 2586 2587
             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
2588
#endif
S
sneaxiy 已提交
2589
             new (&self) platform::CUDAPinnedPlace();
S
sneaxiy 已提交
2590
           })
S
sneaxiy 已提交
2591 2592 2593 2594
      .def("_type", &PlaceIndex<platform::CUDAPinnedPlace>)
      .def("_equals", &IsSamePlace<platform::CUDAPinnedPlace, platform::Place>)
      .def("_equals",
           &IsSamePlace<platform::CUDAPinnedPlace, platform::CUDAPlace>)
2595 2596
      .def("_equals",
           &IsSamePlace<platform::CUDAPinnedPlace, platform::XPUPlace>)
2597 2598
      .def("_equals",
           &IsSamePlace<platform::CUDAPinnedPlace, platform::NPUPlace>)
S
sneaxiy 已提交
2599 2600 2601 2602
      .def("_equals",
           &IsSamePlace<platform::CUDAPinnedPlace, platform::CPUPlace>)
      .def("_equals",
           &IsSamePlace<platform::CUDAPinnedPlace, platform::CUDAPinnedPlace>)
2603
      .def("__repr__", string::to_string<const platform::CUDAPinnedPlace &>)
C
chengduoZH 已提交
2604 2605
      .def("__str__", string::to_string<const platform::CUDAPinnedPlace &>);

2606
  // NPUPlace
2607
  py::class_<platform::NPUPlace> npuplace(m, "NPUPlace", R"DOC(
2608 2609 2610 2611 2612 2613 2614 2615
    NPUPlace is a descriptor of a device.
    It represents a NPU device on which a tensor will be allocated and a model will run.

    Examples:
        .. code-block:: python
          import paddle
          npu_place = paddle.NPUPlace(0)

2616 2617 2618
        )DOC");
  g_npuplace_pytype = reinterpret_cast<PyTypeObject *>(npuplace.ptr());
  npuplace
2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638
      .def("__init__",
           [](platform::NPUPlace &self, int dev_id) {
#ifdef PADDLE_WITH_ASCEND_CL
             if (UNLIKELY(dev_id < 0)) {
               LOG(ERROR) << string::Sprintf(
                   "Invalid NPUPlace(%d), device id must be 0 or "
                   "positive integer",
                   dev_id);
               std::exit(-1);
             }
             if (UNLIKELY(dev_id >= platform::GetNPUDeviceCount())) {
               if (platform::GetNPUDeviceCount() == 0) {
                 LOG(ERROR) << "Cannot use NPU because there is no NPU "
                               "detected on your "
                               "machine.";
                 std::exit(-1);
               } else {
                 LOG(ERROR) << string::Sprintf(
                     "Invalid NPUPlace(%d), must inside [0, %d), because NPU "
                     "number on your machine is %d",
2639 2640
                     dev_id,
                     platform::GetNPUDeviceCount(),
2641 2642 2643 2644 2645 2646 2647 2648 2649 2650
                     platform::GetNPUDeviceCount());
                 std::exit(-1);
               }
             }
             new (&self) platform::NPUPlace(dev_id);
#else
             LOG(ERROR) << string::Sprintf(
                 "Cannot use NPU because you have installed CPU/GPU version "
                 "PaddlePaddle.\n"
                 "If you want to use NPU, please try to install NPU version "
2651
                 "PaddlePaddle by: pip install paddlepaddle-npu\n"
2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665
                 "If you only have CPU, please change NPUPlace(%d) to be "
                 "CPUPlace().\n",
                 dev_id);
             std::exit(-1);
#endif
           })
      .def("_type", &PlaceIndex<platform::NPUPlace>)
      .def("_equals", &IsSamePlace<platform::NPUPlace, platform::Place>)
      .def("_equals", &IsSamePlace<platform::NPUPlace, platform::CUDAPlace>)
      .def("_equals", &IsSamePlace<platform::NPUPlace, platform::CPUPlace>)
      .def("_equals", &IsSamePlace<platform::NPUPlace, platform::XPUPlace>)
      .def("_equals", &IsSamePlace<platform::NPUPlace, platform::NPUPlace>)
      .def("_equals",
           &IsSamePlace<platform::NPUPlace, platform::CUDAPinnedPlace>)
H
houj04 已提交
2666 2667
      .def("get_device_id",
           [](const platform::NPUPlace &self) { return self.GetDeviceId(); })
2668 2669
      .def("__str__", string::to_string<const platform::NPUPlace &>);

J
jianghaicheng 已提交
2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718 2719 2720 2721
  // IPUPlace
  py::class_<platform::IPUPlace>(m, "IPUPlace", R"DOC(
    IPUPlace is a descriptor of a device.
    It represents a IPU device on which a tensor will be allocated and a model will run.

    Examples:
        .. code-block:: python
          import paddle

          # required: ipu

          ipu_place = paddle.IPUPlace()

        )DOC")
      .def("__init__",
           [](platform::IPUPlace &self) {
#ifdef PADDLE_WITH_IPU
             if (platform::GetIPUDeviceCount() == 0) {
               LOG(ERROR) << "Cannot use IPU because there is no IPU "
                             "detected on your "
                             "machine.";
               std::exit(-1);
             }
             // use ipu(0) to comile, while run with the number user configure
             // in sharding and pipline.
             new (&self) platform::IPUPlace(0);
#else
             LOG(ERROR) << string::Sprintf(
                 "Cannot use IPU because you didn't install IPU version "
                 "PaddlePaddle.\n"
                 "If you want to use IPU, please try to install IPU version "
                 "PaddlePaddle by: pip install paddlepaddle*\n"
                 "If you only have CPU, please change IPUPlace to be "
                 "CPUPlace().\n");
             std::exit(-1);
#endif
           })
      .def("_type", &PlaceIndex<platform::IPUPlace>)
      .def("_equals", &IsSamePlace<platform::IPUPlace, platform::Place>)
      .def("_equals", &IsSamePlace<platform::IPUPlace, platform::CUDAPlace>)
      .def("_equals", &IsSamePlace<platform::IPUPlace, platform::CPUPlace>)
      .def("_equals", &IsSamePlace<platform::IPUPlace, platform::XPUPlace>)
      .def("_equals", &IsSamePlace<platform::IPUPlace, platform::NPUPlace>)
      .def("_equals", &IsSamePlace<platform::IPUPlace, platform::IPUPlace>)
      .def("_equals",
           &IsSamePlace<platform::IPUPlace, platform::CUDAPinnedPlace>)
#ifdef PADDLE_WITH_IPU
      .def("get_device_id",
           [](const platform::IPUPlace &self) { return self.GetDeviceId(); })
#endif
      .def("__str__", string::to_string<const platform::IPUPlace &>);

2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752 2753 2754 2755
  // MLUPlace
  py::class_<platform::MLUPlace> mluplace(m, "MLUPlace", R"DOC(
    MLUPlace is a descriptor of a device.
    It represents a MLU device on which a tensor will be allocated and a model will run.

    Examples:
        .. code-block:: python
          import paddle
          # required: mlu
          mlu_place = paddle.MLUPlace(0)

        )DOC");
  g_mluplace_pytype = reinterpret_cast<PyTypeObject *>(mluplace.ptr());
  mluplace
      .def("__init__",
           [](platform::MLUPlace &self, int dev_id) {
#ifdef PADDLE_WITH_MLU
             if (UNLIKELY(dev_id < 0)) {
               LOG(ERROR) << string::Sprintf(
                   "Invalid MLUPlace(%d), device id must be 0 or "
                   "positive integer",
                   dev_id);
               std::exit(-1);
             }
             if (UNLIKELY(dev_id >= platform::GetMLUDeviceCount())) {
               if (platform::GetMLUDeviceCount() == 0) {
                 LOG(ERROR) << "Cannot use MLU because there is no MLU "
                               "detected on your "
                               "machine.";
                 std::exit(-1);
               } else {
                 LOG(ERROR) << string::Sprintf(
                     "Invalid MLUPlace(%d), must inside [0, %d), because MLU "
                     "number on your machine is %d",
2756 2757
                     dev_id,
                     platform::GetMLUDeviceCount(),
2758 2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791
                     platform::GetMLUDeviceCount());
                 std::exit(-1);
               }
             }
             new (&self) platform::MLUPlace(dev_id);
#else
             LOG(ERROR) << string::Sprintf(
                 "Cannot use MLU because you have installed CPU/GPU/... "
                 "version "
                 "PaddlePaddle.\n"
                 "If you want to use MLU, please try to install MLU version "
                 "PaddlePaddle by: pip install paddlepaddle-mlu\n"
                 "If you only have CPU, please change MLUPlace(%d) to be "
                 "CPUPlace().\n",
                 dev_id);
             std::exit(-1);
#endif
           })
      .def("_type", &PlaceIndex<platform::MLUPlace>)
#ifdef PADDLE_WITH_MLU
      .def("_equals", &IsSamePlace<platform::MLUPlace, platform::Place>)
      .def("_equals", &IsSamePlace<platform::MLUPlace, platform::CUDAPlace>)
      .def("_equals", &IsSamePlace<platform::MLUPlace, platform::CPUPlace>)
      .def("_equals", &IsSamePlace<platform::MLUPlace, platform::XPUPlace>)
      .def("_equals", &IsSamePlace<platform::MLUPlace, platform::NPUPlace>)
      .def("_equals", &IsSamePlace<platform::MLUPlace, platform::IPUPlace>)
      .def("_equals", &IsSamePlace<platform::MLUPlace, platform::MLUPlace>)
      .def("_equals",
           &IsSamePlace<platform::MLUPlace, platform::CUDAPinnedPlace>)
      .def("get_device_id",
           [](const platform::MLUPlace &self) { return self.GetDeviceId(); })
#endif
      .def("__str__", string::to_string<const platform::MLUPlace &>);

2792 2793 2794
  py::class_<platform::Place> platformplace(m, "Place");
  g_place_pytype = reinterpret_cast<PyTypeObject *>(platformplace.ptr());
  platformplace.def(py::init<>())
S
sneaxiy 已提交
2795 2796 2797 2798
      .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>)
2799
      .def("_equals", &IsSamePlace<platform::Place, platform::XPUPlace>)
2800
      .def("_equals", &IsSamePlace<platform::Place, platform::NPUPlace>)
J
jianghaicheng 已提交
2801
      .def("_equals", &IsSamePlace<platform::Place, platform::IPUPlace>)
S
sneaxiy 已提交
2802
      .def("_equals", &IsSamePlace<platform::Place, platform::CUDAPinnedPlace>)
2803
      .def("_equals", &IsSamePlace<platform::Place, platform::MLUPlace>)
2804
      .def("_equals", &IsSamePlace<platform::Place, platform::CustomPlace>)
X
xuezhong 已提交
2805 2806
      .def("is_gpu_place",
           [](platform::Place &self) { return platform::is_gpu_place(self); })
S
sneaxiy 已提交
2807 2808
      .def("is_cpu_place",
           [](platform::Place &self) { return platform::is_cpu_place(self); })
2809 2810
      .def("is_xpu_place",
           [](platform::Place &self) { return platform::is_xpu_place(self); })
2811 2812
      .def("is_npu_place",
           [](platform::Place &self) { return platform::is_npu_place(self); })
J
jianghaicheng 已提交
2813 2814
      .def("is_ipu_place",
           [](platform::Place &self) { return platform::is_ipu_place(self); })
S
sneaxiy 已提交
2815 2816 2817 2818
      .def("is_cuda_pinned_place",
           [](platform::Place &self) {
             return platform::is_cuda_pinned_place(self);
           })
2819 2820
      .def("is_mlu_place",
           [](platform::Place &self) { return platform::is_mlu_place(self); })
2821 2822 2823
      .def(
          "is_custom_place",
          [](platform::Place &self) { return platform::is_custom_place(self); })
2824 2825 2826 2827 2828
      .def("gpu_device_id", [](platform::Place &self) { return self.device; })
      .def("xpu_device_id", [](platform::Place &self) { return self.device; })
      .def("npu_device_id", [](platform::Place &self) { return self.device; })
      .def("ipu_device_id", [](platform::Place &self) { return self.device; })
      .def("mlu_device_id", [](platform::Place &self) { return self.device; })
2829 2830
      .def("custom_device_id",
           [](platform::Place &self) { return self.device; })
2831 2832 2833 2834
      .def("set_place",
           [](platform::Place &self, const platform::Place &other) {
             self = other;
           })
Y
Yu Yang 已提交
2835 2836 2837 2838
      .def("set_place",
           [](platform::Place &self, const platform::CPUPlace &cpu_place) {
             self = cpu_place;
           })
2839 2840 2841 2842
      .def("set_place",
           [](platform::Place &self, const platform::XPUPlace &xpu_place) {
             self = xpu_place;
           })
Y
Yu Yang 已提交
2843
      .def("set_place",
D
dzhwinter 已提交
2844
           [](platform::Place &self, const platform::CUDAPlace &gpu_place) {
Y
Yu Yang 已提交
2845
             self = gpu_place;
C
chengduoZH 已提交
2846
           })
2847 2848 2849 2850 2851
      .def("set_place",
           [](platform::Place &self,
              const platform::CUDAPinnedPlace &cuda_pinned_place) {
             self = cuda_pinned_place;
           })
2852 2853 2854 2855
      .def("set_place",
           [](platform::Place &self, const platform::NPUPlace &npu_place) {
             self = npu_place;
           })
J
jianghaicheng 已提交
2856 2857 2858 2859
      .def("set_place",
           [](platform::Place &self, const platform::IPUPlace &ipu_place) {
             self = ipu_place;
           })
2860 2861 2862 2863
      .def("set_place",
           [](platform::Place &self, const platform::MLUPlace &mlu_place) {
             self = mlu_place;
           })
2864 2865 2866 2867
      .def("set_place",
           [](platform::Place &self, const platform::CustomPlace &plug_place) {
             self = plug_place;
           })
2868 2869
      .def("__repr__", string::to_string<const platform::Place &>)
      .def("__str__", string::to_string<const platform::Place &>);
Y
Yu Yang 已提交
2870

Y
Yu Yang 已提交
2871
  py::class_<OperatorBase>(m, "Operator")
2872 2873 2874 2875 2876 2877 2878
      .def_static("create",
                  [](py::bytes protobin) {
                    proto::OpDesc desc;
                    PADDLE_ENFORCE_EQ(desc.ParsePartialFromString(protobin),
                                      true,
                                      platform::errors::InvalidArgument(
                                          "Cannot parse user input to OpDesc"));
2879 2880
                    PADDLE_ENFORCE_EQ(desc.IsInitialized(),
                                      true,
2881 2882 2883 2884 2885 2886
                                      platform::errors::InvalidArgument(
                                          "The provided OpDesc is not "
                                          "initialized, the reason is: %s",
                                          desc.InitializationErrorString()));
                    return OpRegistry::CreateOp(desc);
                  })
2887
      .def("run",
2888 2889
           [](OperatorBase &self,
              const Scope &scope,
2890 2891 2892 2893
              const platform::CPUPlace &place) {
             pybind11::gil_scoped_release release;
             self.Run(scope, place);
           })
2894
      .def("run",
2895 2896
           [](OperatorBase &self,
              const Scope &scope,
2897 2898 2899 2900
              const platform::XPUPlace &place) {
             pybind11::gil_scoped_release release;
             self.Run(scope, place);
           })
2901
      .def("run",
2902 2903
           [](OperatorBase &self,
              const Scope &scope,
2904 2905 2906 2907
              const platform::NPUPlace &place) {
             pybind11::gil_scoped_release release;
             self.Run(scope, place);
           })
D
dzhwinter 已提交
2908
      .def("run",
2909 2910
           [](OperatorBase &self,
              const Scope &scope,
2911 2912 2913 2914
              const platform::CUDAPlace &place) {
             pybind11::gil_scoped_release release;
             self.Run(scope, place);
           })
C
chengduoZH 已提交
2915
      .def("run",
2916 2917
           [](OperatorBase &self,
              const Scope &scope,
C
chengduoZH 已提交
2918
              const platform::CUDAPinnedPlace &place) {
2919
             pybind11::gil_scoped_release release;
C
chengduoZH 已提交
2920 2921
             self.Run(scope, place);
           })
2922
      .def("run",
2923 2924
           [](OperatorBase &self,
              const Scope &scope,
2925 2926 2927 2928
              const platform::MLUPlace &place) {
             pybind11::gil_scoped_release release;
             self.Run(scope, place);
           })
R
ronnywang 已提交
2929
      .def("run",
2930 2931
           [](OperatorBase &self,
              const Scope &scope,
R
ronnywang 已提交
2932 2933 2934 2935
              const platform::CustomPlace &place) {
             pybind11::gil_scoped_release release;
             self.Run(scope, place);
           })
Y
Yu Yang 已提交
2936 2937 2938 2939 2940
      .def("type",
           [](const OperatorBase &op) -> std::string { return op.Type(); })
      .def("outputs",
           [](const OperatorBase &op)
               -> std::map<std::string, std::vector<std::string>> {
2941 2942
             return op.Outputs();
           })
Q
qijun 已提交
2943 2944
      .def("output_vars",
           [](const OperatorBase &op) { return op.OutputVars(true); })
Y
Yu Yang 已提交
2945
      .def("inputs", [](const OperatorBase &op) { return op.Inputs(); })
Q
qijun 已提交
2946
      .def("input_vars", [](const OperatorBase &op) { return op.InputVars(); })
Y
Yu Yang 已提交
2947 2948 2949 2950
      .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
2951

2952 2953 2954
  py::class_<framework::ExecutorPrepareContext>(m, "ExecutorPrepareContext")
      .def(py::init<const ProgramDesc &, size_t>());

2955 2956
  py::class_<framework::TrainerBase, std::shared_ptr<framework::TrainerBase>>(
      m, "TrainerBase")
2957 2958 2959 2960 2961 2962
      .def(
          "get_worker_scope",
          [](TrainerBase &self, int thread_id) -> Scope * {
            return self.GetWorkerScope(thread_id);
          },
          py::return_value_policy::reference)
2963 2964
      .def("finalize", &TrainerBase::Finalize)
      .def("ResetDataset", &TrainerBase::ResetDataset);
2965

2966 2967
  m.def("_get_eager_deletion_vars", &framework::GetEagerDeletionCleanVars);

F
fengjiayi 已提交
2968
  py::class_<framework::Executor>(m, "Executor")
D
dzhwinter 已提交
2969
      .def(py::init<const platform::Place &>())
Y
Yancey1989 已提交
2970
      .def("close", &Executor::Close)
2971 2972
      .def("run_from_dataset",
           &Executor::RunFromDataset,
2973
           py::call_guard<py::gil_scoped_release>())
2974 2975
      .def("release_trainer",
           &Executor::ReleaseTrainer,
D
Dong Daxiang 已提交
2976
           py::call_guard<py::gil_scoped_release>())
2977
      .def("init_for_dataset",
2978 2979 2980 2981
           [](Executor &self,
              const ProgramDesc &prog,
              const std::string &trainer_desc,
              Scope *scope,
2982
              Dataset *dataset) -> std::shared_ptr<TrainerBase> {
D
Dong Daxiang 已提交
2983
             pybind11::gil_scoped_release release;
2984 2985 2986 2987 2988 2989 2990
             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);
           })
2991
      .def("run_prepared_ctx",
2992 2993 2994
           [](Executor &self,
              ExecutorPrepareContext *ctx,
              Scope *scope,
2995
              std::map<std::string, const LoDTensor *> *feed_targets,
2996
              std::map<std::string, FetchType *> *fetch_targets,
2997 2998
              bool create_local_scope = true,
              bool create_vars = true,
2999 3000 3001
              const std::string &feed_holder_name = "feed",
              const std::string &fetch_holder_name = "fetch") {
             pybind11::gil_scoped_release release;
3002 3003 3004 3005 3006 3007 3008 3009
             self.RunPreparedContext(ctx,
                                     scope,
                                     feed_targets,
                                     fetch_targets,
                                     create_local_scope,
                                     create_vars,
                                     feed_holder_name,
                                     fetch_holder_name);
3010
           })
3011
      .def("run_prepared_ctx",
3012 3013 3014 3015 3016
           [](Executor &self,
              ExecutorPrepareContext *ctx,
              Scope *scope,
              bool create_local_scope = true,
              bool create_vars = true,
G
guru4elephant 已提交
3017 3018
              bool keep_kids = false) {
             pybind11::gil_scoped_release release;
3019 3020
             self.RunPreparedContext(
                 ctx, scope, create_local_scope, create_vars, keep_kids);
G
guru4elephant 已提交
3021
           })
3022
      .def("prepare",
3023 3024 3025
           [](Executor &self,
              const ProgramDesc &program,
              int block_id,
3026 3027 3028 3029
              const std::vector<std::string> &skip_ref_cnt_vars =
                  std::vector<std::string>(),
              bool force_disable_gc = false) {
             pybind11::gil_scoped_release release;
3030 3031
             return self.Prepare(
                 program, block_id, skip_ref_cnt_vars, force_disable_gc);
3032 3033
           })
      .def("create_variables", &Executor::CreateVariables)
3034 3035 3036 3037 3038 3039 3040 3041 3042 3043 3044 3045 3046 3047 3048 3049
      .def("run",
           [](Executor &self,
              const ProgramDesc &prog,
              Scope *scope,
              int block_id,
              bool create_local_scope,
              bool create_vars,
              const std::vector<std::string> &fetch_vars) {
             pybind11::gil_scoped_release release;
             self.Run(prog,
                      scope,
                      block_id,
                      create_local_scope,
                      create_vars,
                      fetch_vars);
           });
S
sneaxiy 已提交
3050

3051
  py::class_<framework::interpreter::CostInfo>(m, "CostInfo")
3052
      .def(py::init<>())
3053 3054 3055 3056 3057
      .def("total_time",
           [](interpreter::CostInfo &self) { return self.total_time; })
      .def("device_memory_bytes", [](interpreter::CostInfo &self) {
        return self.device_memory_bytes;
      });
3058

3059
  py::class_<framework::StandaloneExecutor>(m, "StandaloneExecutor")
3060 3061 3062 3063
      .def(py::init<const platform::Place &,
                    const ProgramDesc &,
                    const ProgramDesc &,
                    Scope *>())
H
hong 已提交
3064
      .def("run",
3065
           [](StandaloneExecutor &self,
3066
              Scope *scope,
H
hong 已提交
3067
              const std::unordered_map<std::string, py::array> &input_dict,
3068
              std::vector<std::string> fetch_names) {
3069
             std::vector<framework::LoDTensor> feed_tensors;
3070
             std::vector<std::string> feed_names;
H
hong 已提交
3071 3072 3073 3074 3075

             for (auto &item : input_dict) {
               framework::LoDTensor t;
               SetTensorFromPyArray<platform::CPUPlace>(
                   &t, item.second, platform::CPUPlace(), false);
3076 3077
               feed_names.push_back(item.first);
               feed_tensors.push_back(t);
H
hong 已提交
3078 3079
             }

3080 3081 3082
             paddle::framework::FetchList ret;
             {
               pybind11::gil_scoped_release release;
3083
               ret = self.Run(scope, feed_names, feed_tensors, fetch_names);
3084 3085 3086 3087 3088
             }
             return py::cast(std::move(ret));
           })
      .def("run",
           [](StandaloneExecutor &self,
3089
              Scope *scope,
3090
              const std::unordered_map<std::string, framework::LoDTensor>
3091 3092
                  &input_dict,
              std::vector<std::string> fetch_names) {
3093
             std::vector<framework::LoDTensor> feed_tensors;
3094 3095 3096 3097 3098 3099 3100
             std::vector<std::string> feed_names;

             for (auto &item : input_dict) {
               feed_names.push_back(item.first);
               feed_tensors.push_back(item.second);
             }

W
wanghuancoder 已提交
3101 3102 3103
             paddle::framework::FetchList ret;
             {
               pybind11::gil_scoped_release release;
3104
               ret = self.Run(scope, feed_names, feed_tensors, fetch_names);
H
hong 已提交
3105
             }
W
wanghuancoder 已提交
3106
             return py::cast(std::move(ret));
3107
           })
3108
      .def("run",
3109
           [](StandaloneExecutor &self,
3110
              Scope *scope,
3111
              std::vector<std::string> feed_names,
3112 3113 3114 3115
              std::vector<std::string> fetch_names) {
             paddle::framework::FetchList ret;
             {
               pybind11::gil_scoped_release release;
3116
               ret = self.Run(scope, feed_names, fetch_names);
3117 3118 3119
             }
             return py::cast(std::move(ret));
           })
3120 3121
      .def("dry_run",
           [](StandaloneExecutor &self,
3122
              Scope *scope,
3123
              const std::unordered_map<std::string, py::array> &input_dict) {
3124
             std::vector<framework::LoDTensor> feed_tensors;
3125 3126 3127 3128 3129 3130 3131 3132 3133 3134
             std::vector<std::string> feed_names;

             for (auto &item : input_dict) {
               framework::LoDTensor t;
               SetTensorFromPyArray<platform::CPUPlace>(
                   &t, item.second, platform::CPUPlace(), false);
               feed_names.push_back(item.first);
               feed_tensors.push_back(t);
             }

3135
             framework::interpreter::CostInfo cost_info;
3136 3137
             {
               pybind11::gil_scoped_release release;
3138
               cost_info = self.DryRun(scope, feed_names, feed_tensors);
3139 3140
             }
             return cost_info;
H
hong 已提交
3141 3142
           });

D
dzhwinter 已提交
3143
  m.def("init_gflags", framework::InitGflags);
Y
Yang Yu 已提交
3144
  m.def("init_glog", framework::InitGLOG);
3145 3146 3147 3148
  m.def("load_op_meta_info_and_register_op", [](const std::string dso_name) {
    egr::Controller::Instance().MergeOpMetaInfoMap(
        framework::LoadOpMetaInfoAndRegisterOp(dso_name));
  });
3149
  m.def("init_devices", []() { framework::InitDevices(); });
3150 3151
  m.def("init_default_kernel_signatures",
        []() { framework::InitDefaultKernelSignatureMap(); });
3152
  m.def("is_compiled_with_cuda", IsCompiledWithCUDA);
3153
  m.def("is_compiled_with_ascend", IsCompiledWithAscend);
3154
  m.def("is_compiled_with_rocm", IsCompiledWithROCM);
3155
  m.def("is_compiled_with_npu", IsCompiledWithNPU);
J
jianghaicheng 已提交
3156
  m.def("is_compiled_with_ipu", IsCompiledWithIPU);
3157
  m.def("is_compiled_with_xpu", IsCompiledWithXPU);
3158
  m.def("is_compiled_with_mkldnn", IsCompiledWithMKLDNN);
3159
  m.def("is_compiled_with_nccl", IsCompiledWithNCCL);
3160
  m.def("is_compiled_with_cinn", IsCompiledWithCINN);
3161
  m.def("is_compiled_with_mlu", IsCompiledWithMLU);
3162
  m.def("_is_compiled_with_heterps", IsCompiledWithHETERPS);
3163
  m.def("supports_bfloat16", SupportsBfloat16);
3164
  m.def("supports_bfloat16_fast_performance", SupportsBfloat16FastPerformance);
3165 3166
  m.def("supports_int8", SupportsInt8);
  m.def("supports_vnni", SupportsVNNI);
L
Leo Chen 已提交
3167
  m.def("op_supported_infos", imperative::OpSupportedInfos);
3168
  m.def("is_compiled_with_brpc", IsCompiledWithBrpc);
Y
update  
Yancey1989 已提交
3169
  m.def("is_compiled_with_dist", IsCompiledWithDIST);
3170 3171 3172
  m.def("_cuda_synchronize", [](const platform::CUDAPlace &place) {
    platform::DeviceContextPool::Instance().Get(place)->Wait();
  });
H
hutuxian 已提交
3173 3174 3175 3176 3177 3178 3179 3180 3181 3182 3183 3184 3185 3186 3187 3188 3189 3190 3191

  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;
  });
3192 3193 3194
  m.def("device_memory_stat_current_value",
        memory::DeviceMemoryStatCurrentValue);
  m.def("device_memory_stat_peak_value", memory::DeviceMemoryStatPeakValue);
3195 3196
  m.def(
      "run_cmd",
3197 3198
      [](const std::string &cmd,
         int time_out = -1,
3199
         int sleep_inter = -1) -> const std::string {
3200 3201
        return paddle::framework::shell_get_command_output(
            cmd, time_out, sleep_inter);
3202
      },
3203 3204 3205
      py::arg("cmd"),
      py::arg("time_out") = -1,
      py::arg("sleep_inter") = -1);
3206 3207
  m.def(
      "shell_execute_cmd",
3208 3209 3210
      [](const std::string &cmd,
         int time_out = 0,
         int sleep_inter = 0,
3211
         bool redirect_stderr = false) -> std::vector<std::string> {
3212 3213
        return paddle::framework::shell_execute_cmd(
            cmd, time_out, sleep_inter, redirect_stderr);
3214
      },
3215 3216 3217
      py::arg("cmd"),
      py::arg("time_out") = 0,
      py::arg("sleep_inter") = 0,
3218
      py::arg("redirect_stderr") = false);
G
gongweibao 已提交
3219

3220
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
3221 3222
  m.def("is_float16_supported", [](const platform::CUDAPlace &place) -> bool {
    // Only GPUs with Compute Capability >= 53 support float16
3223
    return platform::GetGPUComputeCapability(place.device) >= 53;
3224
  });
3225 3226 3227 3228
  m.def("is_bfloat16_supported", [](const platform::CUDAPlace &place) -> bool {
    // Only GPUs with Compute Capability >= 80 support bfloat16
    return platform::GetGPUComputeCapability(place.device) >= 80;
  });
3229
#endif
3230

S
Steffy-zxf 已提交
3231
  m.def("set_feed_variable",
3232 3233 3234 3235 3236
        static_cast<void (*)(  // NOLINT
            Scope *,
            const LoDTensor &,
            const std::string &,
            size_t)>(&framework::SetFeedVariable));
S
Steffy-zxf 已提交
3237
  m.def("set_feed_variable",
3238 3239 3240 3241 3242
        static_cast<void (*)(  // NOLINT
            Scope *,
            const Strings &,
            const std::string &,
            size_t)>(&framework::SetFeedVariable));
3243
  m.def("get_fetch_variable",
3244 3245
        [](const Scope &scope,
           const std::string &var_name,
3246 3247 3248
           size_t index) -> py::object {
          auto &var = framework::GetFetchVariable(scope, var_name, index);
          if (data_is_lod_tensor(var)) {
3249
            return py::cast(BOOST_GET(LoDTensor, var));
3250
          } else {
3251
            return py::cast(BOOST_GET(LoDTensorArray, var));
3252 3253
          }
        });
3254
  m.def("get_variable_tensor", framework::GetVariableTensor);
Q
qijun 已提交
3255

X
Xin Pan 已提交
3256 3257
  m.def("_is_program_version_supported", IsProgramVersionSupported);

3258 3259 3260 3261
  BindProgramDesc(&m);
  BindBlockDesc(&m);
  BindVarDsec(&m);
  BindOpDesc(&m);
H
Huihuang Zheng 已提交
3262
  BindCostModel(&m);
3263
  BindConstValue(&m);
3264
  BindGlobalValueGetterSetter(&m);
3265
  BindProcessMeshDesc(&m);
L
LiYuRio 已提交
3266
  BindFleetExecutor(&m);
3267
  BindTCPStore(&m);
Y
Yu Yang 已提交
3268

Y
Yu Yang 已提交
3269 3270 3271 3272 3273 3274 3275 3276 3277
  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;
      });

3278
  py::class_<LoDTensorArray> pylodtensorarray(m, "LoDTensorArray", R"DOC(
3279
    LoDTensorArray is array of LoDTensor, it supports operator[], len() and for-loop iteration.
Z
Zeng Jinle 已提交
3280 3281 3282

    Examples:
        .. code-block:: python
3283

Z
Zeng Jinle 已提交
3284 3285 3286
          import paddle.fluid as fluid

          arr = fluid.LoDTensorArray()
3287 3288 3289 3290
)DOC");
  g_framework_lodtensorarray_pytype =
      reinterpret_cast<PyTypeObject *>(pylodtensorarray.ptr());
  pylodtensorarray
S
sneaxiy 已提交
3291 3292
      .def("__init__",
           [](LoDTensorArray &instance) { new (&instance) LoDTensorArray(); })
3293 3294 3295 3296
      .def(
          "__getitem__",
          [](LoDTensorArray &self, size_t i) { return &self.at(i); },
          py::return_value_policy::reference)
Y
Yu Yang 已提交
3297 3298 3299
      .def("__len__", [](LoDTensorArray &self) { return self.size(); })
      .def("__setitem__",
           [](LoDTensorArray &self, size_t i, const LoDTensor &t) {
3300 3301
             PADDLE_ENFORCE_LT(i,
                               self.size(),
3302 3303 3304
                               platform::errors::InvalidArgument(
                                   "The index to set is larger than the size "
                                   "of LoDTensorArray."));
Y
Yu Yang 已提交
3305 3306 3307
             self[i].ShareDataWith(t);
             self[i].set_lod(t.lod());
           })
3308 3309 3310 3311 3312 3313 3314
      .def(
          "append",
          [](LoDTensorArray &self, const LoDTensor &t) {
            self.emplace_back();
            self.back().ShareDataWith(t);
            self.back().set_lod(t.lod());
          },
3315 3316
          py::arg("tensor"),
          R"DOC(
Z
Zeng Jinle 已提交
3317
             Append a LoDensor to LoDTensorArray.
3318 3319 3320 3321 3322 3323
              
             Args:
                   tensor (LoDTensor): The LoDTensor to be appended.

             Returns:
                   None.
Z
Zeng Jinle 已提交
3324 3325 3326 3327 3328 3329 3330 3331 3332 3333 3334

             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)
3335
           )DOC")
3336 3337 3338 3339 3340 3341 3342 3343 3344 3345 3346
      .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
3347

3348
  py::class_<FetchList>(m, "FetchList", R"DOC( FetchList is a
R
Ruibiao Chen 已提交
3349
        vector of paddle::variant<LoDTensor, LoDTensorArray>.
3350
        )DOC")
3351 3352 3353 3354 3355 3356 3357 3358 3359 3360 3361 3362 3363 3364 3365 3366 3367 3368 3369 3370 3371
      .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])) {
                auto &data = BOOST_GET(LoDTensor, self[i]);
                res[i] = py::cast(std::move(data));
              } else {
                auto &data = BOOST_GET(LoDTensorArray, self[i]);
                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)
3372

3373 3374 3375 3376 3377 3378 3379 3380 3381 3382 3383 3384 3385 3386 3387 3388 3389 3390 3391 3392 3393
      .def(
          "append",
          [](FetchList &self, const LoDTensor &t) {
            self.emplace_back();
            auto &lod_tensor = BOOST_GET(LoDTensor, self.back());
            lod_tensor.ShareDataWith(t);
            lod_tensor.set_lod(t.lod());
          },
          py::arg("var"))

      .def(
          "append",
          [](FetchList &self, const LoDTensorArray &t) {
            self.emplace_back();
            auto &lod_tensor_array = BOOST_GET(LoDTensorArray, self.back());
            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"));
3394 3395

  py::class_<FetchUnmergedList>(m, "FetchUnmergedList", R"DOC(
R
Ruibiao Chen 已提交
3396
        FetchUnmergedList is 2-D array of FetchType(paddle::variant(LoDTensor, LoDTensorArray)).
Z
Zhen Wang 已提交
3397
        )DOC")
3398 3399 3400 3401 3402 3403 3404 3405 3406 3407 3408 3409 3410 3411 3412 3413 3414 3415 3416 3417 3418 3419 3420 3421 3422 3423
      .def(
          "_move_to_list",
          [](FetchUnmergedList &self) -> py::list {
            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) {
                if (data_is_lod_tensor(self[i][j])) {
                  auto &var = BOOST_GET(LoDTensor, self[i][j]);
                  tmp[j] = py::cast(std::move(var));
                } else {
                  auto &var = BOOST_GET(LoDTensorArray, self[i][j]);
                  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);
                }
              }
              res[i] = std::move(tmp);
              self[i].clear();
            }
            self.clear();
            return res;
          },
          py::return_value_policy::take_ownership);
Z
Zhen Wang 已提交
3424

Y
Yu Yang 已提交
3425
  m.def("op_support_gpu", OpSupportGPU);
3426
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
3427
  m.def("get_cuda_device_count", platform::GetGPUDeviceCount);
3428
  m.def("get_cuda_current_device_id", &platform::GetCurrentDeviceId);
3429 3430 3431 3432 3433 3434 3435 3436
  m.def("cuda_empty_cache", [] {
    for (int dev_id : platform::GetSelectedDevices()) {
      auto *dev_ctx = platform::DeviceContextPool::Instance().GetByPlace(
          platform::CUDAPlace(dev_id));
      dev_ctx->cudnn_workspace_handle().ResetWorkspace();
    }
    platform::EmptyCache();
  });
3437 3438 3439 3440 3441 3442
  m.def(
      "get_device_properties",
      [](int id) -> const gpuDeviceProp & {
        return platform::GetDeviceProperties(id);
      },
      py::return_value_policy::copy);
3443 3444

  py::class_<gpuDeviceProp>(m, "_gpuDeviceProperties")
Y
Yanxing Shi 已提交
3445 3446 3447 3448 3449 3450 3451 3452 3453 3454 3455 3456 3457 3458 3459 3460 3461 3462 3463 3464 3465 3466 3467 3468 3469
      .def_property_readonly(
          "name", [](const gpuDeviceProp &prop) { return prop.name; })
      .def_property_readonly(
          "major", [](const gpuDeviceProp &prop) { return prop.major; })
      .def_property_readonly(
          "minor", [](const gpuDeviceProp &prop) { return prop.minor; })
      .def_property_readonly(
          "total_memory",
          [](const gpuDeviceProp &prop) { return prop.totalGlobalMem; })
      .def_property_readonly(
          "multi_processor_count",
          [](const gpuDeviceProp &prop) { return prop.multiProcessorCount; })
      .def_property_readonly(
          "is_multi_gpu_board",
          [](const gpuDeviceProp &prop) { return prop.isMultiGpuBoard; })
      .def_property_readonly(
          "is_integrated",
          [](const gpuDeviceProp &prop) { return prop.integrated; })
      .def("__repr__", [](const gpuDeviceProp &prop) {
        std::stringstream ostr;
        ostr << "_gpuDeviceProperties(name='" << prop.name
             << "', major=" << prop.major << ", minor=" << prop.minor
             << ", total_memory=" << prop.totalGlobalMem / (1024 * 1024)
             << "MB, multi_processor_count=" << prop.multiProcessorCount << ")";
        return ostr.str();
3470
      });
D
dangqingqing 已提交
3471

3472
#if !defined(PADDLE_WITH_HIP) && !defined(_WIN32)
D
dangqingqing 已提交
3473 3474 3475
  m.def("nvprof_init", platform::CudaProfilerInit);
  m.def("nvprof_start", platform::CudaProfilerStart);
  m.def("nvprof_stop", platform::CudaProfilerStop);
3476 3477 3478 3479
  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
3480
#endif
P
peizhilin 已提交
3481
#endif
Y
Yu Yang 已提交
3482

3483 3484
#ifdef PADDLE_WITH_ASCEND_CL
  m.def("get_npu_device_count", platform::GetNPUDeviceCount);
3485
  m.def("npu_finalize", []() {
3486 3487
    platform::HCCLCommContext::Instance().ReleaseHCCLComms();

3488 3489 3490
    auto &pool = platform::DeviceContextPool::Instance();
    auto devices = platform::GetSelectedNPUDevices();
    for (size_t i = 0; i < devices.size(); ++i) {
R
ronnywang 已提交
3491
      platform::NPUDeviceGuard guard(devices[i]);
3492 3493 3494 3495
      pool.Get(platform::NPUPlace(devices[i]))->Wait();
    }
    platform::AclInstance::Instance().Finalize();
  });
3496 3497 3498 3499 3500 3501 3502 3503 3504 3505 3506 3507 3508 3509 3510 3511 3512 3513 3514 3515

  py::class_<platform::NPUProfConfigWrapper>(m, "NPUProfConfigWrapper");

  m.def("npu_prof_init", platform::NPUProfilerInit);
  m.def("npu_prof_start", [](platform::NPUProfConfigWrapper c) {
    platform::NPUProfilerStart(c.ptr());
  });
  m.def("npu_prof_stop", [](platform::NPUProfConfigWrapper c) {
    platform::NPUProfilerStop(c.ptr());
  });
  m.def("npu_prof_finalize", platform::NPUProfilerFinalize);
  m.def("npu_prof_create_config", []() {
    return platform::NPUProfConfigWrapper(platform::NPUProfilerCreateConfig());
  });

  m.def("npu_prof_destropy_config", [](platform::NPUProfConfigWrapper c) {
    platform::NPUProfilerDestroyConfig(c.ptr());
  });
#endif

J
jianghaicheng 已提交
3516 3517 3518 3519
#ifdef PADDLE_WITH_IPU
  m.def("get_ipu_device_count", platform::GetIPUDeviceCount);
#endif

3520 3521 3522 3523
#ifdef PADDLE_WITH_MLU
  m.def("get_mlu_device_count", platform::GetMLUDeviceCount);
#endif

3524 3525 3526 3527 3528 3529
  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();

3530 3531 3532 3533
  py::enum_<platform::ProfilerState>(m, "ProfilerState", py::arithmetic())
      .value("kDisabled", platform::ProfilerState::kDisabled)
      .value("kCPU", platform::ProfilerState::kCPU)
      .value("kCUDA", platform::ProfilerState::kCUDA)
3534
      .value("kAll", platform::ProfilerState::kAll)
3535 3536 3537 3538 3539 3540 3541 3542 3543 3544 3545
      .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();

3546
  m.def("set_tracer_option", platform::SetTracerOption);
3547 3548
  m.def("enable_profiler", platform::EnableProfiler);
  m.def("disable_profiler", platform::DisableProfiler);
X
Xin Pan 已提交
3549
  m.def("is_profiler_enabled", platform::IsProfileEnabled);
3550
  m.def("reset_profiler", platform::ResetProfiler);
W
wuhuanzhou 已提交
3551
  m.def("register_pass", [](const std::string &pass_type, py::object callable) {
3552
    PADDLE_ENFORCE_EQ(
3553 3554
        framework::ir::PassRegistry::Instance().Has(pass_type),
        false,
3555 3556 3557
        platform::errors::AlreadyExists("Pass '%s' is registered more than "
                                        "once. Please use another name.",
                                        pass_type));
W
wuhuanzhou 已提交
3558
    callable.inc_ref();
3559 3560 3561 3562 3563 3564 3565 3566
    framework::ir::PassRegistry::Instance().Insert(
        pass_type, [pass_type, callable]() {
          py::gil_scoped_acquire guard;
          std::unique_ptr<framework::ir::Pass> pass(
              new framework::ir::GeneratePass(
                  py::cast<std::string>(callable())));
          return pass;
        });
3567
  });
3568
  m.def("get_pass", [](const std::string &pass_type) {
W
WangZhen 已提交
3569 3570 3571
    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
3572

3573
  m.def("size_of_dtype", framework::SizeOfType);
C
chenjian 已提交
3574 3575
  py::class_<paddle::platform::ProfilerResult>(m, "_ProfilerResult")
      .def(py::init<>())
3576 3577
      .def("get_data",
           &paddle::platform::ProfilerResult::GetData,
C
chenjian 已提交
3578 3579 3580 3581
           py::return_value_policy::automatic_reference)
      .def("save", &paddle::platform::ProfilerResult::Save)
      .def("get_extra_info", &paddle::platform::ProfilerResult::GetExtraInfo);

3582 3583 3584 3585 3586 3587 3588 3589 3590 3591 3592 3593 3594 3595 3596 3597 3598 3599 3600 3601
  py::class_<paddle::platform::MemPythonNode>(m, "MemPythonNode")
      .def(py::init<>())
      .def_readwrite("timestamp_ns",
                     &paddle::platform::MemPythonNode::timestamp_ns)
      .def_readwrite("addr", &paddle::platform::MemPythonNode::addr)
      .def_readwrite("type", &paddle::platform::MemPythonNode::type)
      .def_readwrite("process_id", &paddle::platform::MemPythonNode::process_id)
      .def_readwrite("thread_id", &paddle::platform::MemPythonNode::thread_id)
      .def_readwrite("increase_bytes",
                     &paddle::platform::MemPythonNode::increase_bytes)
      .def_readwrite("place", &paddle::platform::MemPythonNode::place)
      .def_readwrite("current_allocated",
                     &paddle::platform::MemPythonNode::current_allocated)
      .def_readwrite("current_reserved",
                     &paddle::platform::MemPythonNode::current_reserved)
      .def_readwrite("peak_allocated",
                     &paddle::platform::MemPythonNode::peak_allocated)
      .def_readwrite("peak_reserved",
                     &paddle::platform::MemPythonNode::peak_reserved);

C
chenjian 已提交
3602 3603 3604 3605 3606 3607 3608 3609 3610 3611 3612 3613 3614 3615 3616 3617 3618 3619 3620 3621 3622 3623
  py::class_<paddle::platform::DevicePythonNode>(m, "DevicePythonNode")
      .def(py::init<>())
      .def_readwrite("name", &paddle::platform::DevicePythonNode::name)
      .def_readwrite("type", &paddle::platform::DevicePythonNode::type)
      .def_readwrite("start_ns", &paddle::platform::DevicePythonNode::start_ns)
      .def_readwrite("end_ns", &paddle::platform::DevicePythonNode::end_ns)
      .def_readwrite("device_id",
                     &paddle::platform::DevicePythonNode::device_id)
      .def_readwrite("context_id",
                     &paddle::platform::DevicePythonNode::context_id)
      .def_readwrite("stream_id",
                     &paddle::platform::DevicePythonNode::stream_id);

  py::class_<paddle::platform::HostPythonNode>(m, "HostPythonNode")
      .def(py::init<>())
      .def_readwrite("name", &paddle::platform::HostPythonNode::name)
      .def_readwrite("type", &paddle::platform::HostPythonNode::type)
      .def_readwrite("start_ns", &paddle::platform::HostPythonNode::start_ns)
      .def_readwrite("end_ns", &paddle::platform::HostPythonNode::end_ns)
      .def_readwrite("process_id",
                     &paddle::platform::HostPythonNode::process_id)
      .def_readwrite("thread_id", &paddle::platform::HostPythonNode::thread_id)
3624 3625 3626 3627
      .def_readwrite("input_shapes",
                     &paddle::platform::HostPythonNode::input_shapes)
      .def_readwrite("dtypes", &paddle::platform::HostPythonNode::dtypes)
      .def_readwrite("callstack", &paddle::platform::HostPythonNode::callstack)
C
chenjian 已提交
3628 3629 3630 3631 3632
      .def_readwrite("children_node",
                     &paddle::platform::HostPythonNode::children_node_ptrs)
      .def_readwrite("runtime_node",
                     &paddle::platform::HostPythonNode::runtime_node_ptrs)
      .def_readwrite("device_node",
3633 3634 3635
                     &paddle::platform::HostPythonNode::device_node_ptrs)
      .def_readwrite("mem_node",
                     &paddle::platform::HostPythonNode::mem_node_ptrs);
C
chenjian 已提交
3636 3637

  py::class_<paddle::platform::Profiler>(m, "_Profiler")
3638 3639
      .def("create",
           &paddle::platform::Profiler::Create,
C
chenjian 已提交
3640
           py::return_value_policy::take_ownership)
C
chenjian 已提交
3641
      .def("is_cupti_supported", &paddle::platform::Profiler::IsCuptiSupported)
F
fwenguang 已提交
3642 3643
      .def("is_cnpapi_supported",
           &paddle::platform::Profiler::IsCnpapiSupported)
C
chenjian 已提交
3644 3645 3646 3647 3648 3649
      .def("prepare",
           [](paddle::platform::Profiler *profiler) {
             platform::EnableHostEventRecorder();
             profiler->Prepare();
           })
      .def("start", &paddle::platform::Profiler::Start)
3650 3651 3652 3653 3654 3655 3656 3657 3658 3659
      .def(
          "stop",
          [](paddle::platform::Profiler *profiler) {
            platform::DisableHostEventRecorder();
            auto result = profiler->Stop();
            framework::StaticGraphExecutorPerfStatistics(
                result->GetNodeTrees());
            return result;
          },
          py::return_value_policy::automatic_reference);
C
chenjian 已提交
3660 3661 3662 3663 3664 3665 3666 3667 3668 3669 3670 3671 3672

  py::class_<paddle::platform::ProfilerOptions>(m, "ProfilerOptions")
      .def(py::init<>())
      .def_readwrite("trace_switch",
                     &paddle::platform::ProfilerOptions::trace_switch);

  py::class_<platform::RecordEvent>(m, "_RecordEvent")
      .def(py::init([](std::string name, platform::TracerEventType type) {
        return std::make_unique<platform::RecordEvent>(
            name, type, 1, paddle::platform::EventRole::kOrdinary);
      }))
      .def("end", [](platform::RecordEvent *event) { event->End(); });

3673 3674 3675 3676 3677 3678 3679 3680
  py::enum_<paddle::platform::TracerMemEventType>(m, "TracerMemEventType")
      .value("Allocate", paddle::platform::TracerMemEventType::Allocate)
      .value("Free", paddle::platform::TracerMemEventType::Free)
      .value("ReservedAllocate",
             paddle::platform::TracerMemEventType::ReservedAllocate)
      .value("ReservedFree",
             paddle::platform::TracerMemEventType::ReservedFree);

C
chenjian 已提交
3681 3682 3683 3684 3685 3686 3687 3688 3689 3690 3691 3692 3693 3694 3695 3696 3697 3698
  py::enum_<paddle::platform::TracerEventType>(m, "TracerEventType")
      .value("Operator", paddle::platform::TracerEventType::Operator)
      .value("Dataloader", paddle::platform::TracerEventType::Dataloader)
      .value("ProfileStep", paddle::platform::TracerEventType::ProfileStep)
      .value("CudaRuntime", paddle::platform::TracerEventType::CudaRuntime)
      .value("Kernel", paddle::platform::TracerEventType::Kernel)
      .value("Memcpy", paddle::platform::TracerEventType::Memcpy)
      .value("Memset", paddle::platform::TracerEventType::Memset)
      .value("UserDefined", paddle::platform::TracerEventType::UserDefined)
      .value("OperatorInner", paddle::platform::TracerEventType::OperatorInner)
      .value("Forward", paddle::platform::TracerEventType::Forward)
      .value("Backward", paddle::platform::TracerEventType::Backward)
      .value("Optimization", paddle::platform::TracerEventType::Optimization)
      .value("Communication", paddle::platform::TracerEventType::Communication)
      .value("PythonOp", paddle::platform::TracerEventType::PythonOp)
      .value("PythonUserDefined",
             paddle::platform::TracerEventType::PythonUserDefined);
  m.def("load_profiler_result", &paddle::platform::LoadProfilerResult);
3699

3700
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
3701 3702
  m.def("set_cublas_switch", platform::SetAllowTF32Cublas);
  m.def("get_cublas_switch", platform::AllowTF32Cublas);
A
AshburnLee 已提交
3703 3704
  m.def("set_cudnn_switch", platform::SetAllowTF32Cudnn);
  m.def("get_cudnn_switch", platform::AllowTF32Cudnn);
3705
#endif  // PADDLE_WITH_CUDA
3706 3707
  m.def("clear_executor_cache",
        []() { framework::ExecutorInfoCache::Instance().Finalize(); });
3708

3709 3710 3711
  using VarQuantScale =
      std::unordered_map<std::string, std::pair<bool, LoDTensor>>;

3712 3713
  py::class_<ir::Pass, std::shared_ptr<ir::Pass>> pass(m, "Pass");
  pass.def(py::init())
W
WangZhen 已提交
3714
      .def("has", &ir::Pass::Has)
3715 3716 3717
      .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
3718
           })
3719
      .def(
3720
          "set",
3721 3722 3723
          [](ir::Pass &self, const std::string &name, const std::string &attr) {
            self.Set<std::string>(name, new std::string(attr));
          })
J
jianghaicheng 已提交
3724
      .def("set",
3725 3726 3727 3728 3729 3730 3731 3732 3733 3734
           [](ir::Pass &self, const std::string &name, bool val) {
             self.Set<bool>(name, new bool(val));
           })
      .def("set",
           [](ir::Pass &self, const std::string &name, int val) {
             self.Set<const int>(name, new int(val));
           })
      .def("set",
           [](ir::Pass &self,
              const std::string &name,
J
jianghaicheng 已提交
3735 3736 3737
              std::vector<std::string> set) {
             self.Set(name, new std::vector<std::string>(set));
           })
3738
      .def("set",
3739 3740
           [](ir::Pass &self,
              const std::string &name,
3741 3742 3743 3744
              std::unordered_set<std::string> set) {
             self.Set(name, new std::unordered_set<std::string>(set));
           })
      .def("set",
3745 3746
           [](ir::Pass &self,
              const std::string &name,
3747 3748 3749 3750 3751 3752 3753
              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
3754 3755
      .def("type", &ir::Pass::Type)
      .def("apply", [](ir::Pass &self, std::shared_ptr<ir::Graph> graph) {
3756
        self.Apply(graph.get());
F
flame 已提交
3757
      });
3758

X
fix  
Xin Pan 已提交
3759 3760
  py::class_<ir::PassBuilder, std::shared_ptr<ir::PassBuilder>> pb(
      m, "PassBuilder");
3761 3762 3763 3764 3765 3766 3767 3768 3769 3770 3771 3772 3773 3774
  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
3775
  // -- python binds for parallel executor.
Y
yuyang18 已提交
3776
  py::class_<ParallelExecutor> pe(m, "ParallelExecutor");
C
chengduo 已提交
3777 3778 3779 3780
  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.

3781 3782 3783
    Returns:
        ExecutionStrategy: An ExecutionStrategy object.

C
chengduo 已提交
3784 3785 3786
    Examples:
        .. code-block:: python

3787 3788 3789 3790 3791 3792 3793 3794 3795
          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)
3796

3797 3798
          cost = F.square_error_cost(input=y_predict, label=y)
          avg_loss = paddle.mean(cost)
3799

3800
          sgd_optimizer = paddle.optimizer.SGD(learning_rate=0.001)
3801 3802
          sgd_optimizer.minimize(avg_loss)

3803
          exec_strategy = static.ExecutionStrategy()
C
chengduo 已提交
3804 3805
          exec_strategy.num_threads = 4

3806 3807 3808
          train_exe = static.ParallelExecutor(use_cuda=False,
                                              loss_name=avg_loss.name,
                                              exec_strategy=exec_strategy)
C
chengduo 已提交
3809 3810
        )DOC");

3811 3812 3813 3814
  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);
3815

Y
yuyang18 已提交
3816
  exec_strategy.def(py::init())
Y
yuyang18 已提交
3817 3818 3819 3820 3821
      .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
3822
          },
3823 3824
          R"DOC(
            The type is INT, num_threads represents the size of thread pool that
C
chengduo 已提交
3825 3826 3827 3828 3829 3830 3831
            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
3832 3833 3834 3835 3836 3837 3838 3839 3840 3841 3842 3843 3844
            `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
3845
      .def_property(
3846 3847
          "_use_device",
          [](const ExecutionStrategy &self) { return self.use_device_; },
3848
          [](ExecutionStrategy &self, paddle::platform::DeviceType use_device) {
3849 3850 3851
            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
3852 3853 3854 3855 3856
      .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
3857 3858 3859
          },
          R"DOC(The type is BOOL, allow_op_delay represents whether to delay the
                communication operators to run, it may make the execution faster.
3860 3861
                Note that this option is invalid now, and it will be removed in
                next version. Default False.)DOC")
Y
yuyang18 已提交
3862 3863 3864 3865 3866 3867 3868
      .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
3869 3870 3871 3872
          },
          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,
3873
                because the temp variable's shape maybe the same between two iterations.
3874 3875 3876 3877 3878 3879 3880 3881 3882 3883
                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
3884

3885 3886 3887 3888 3889 3890 3891
                        import paddle
                        import paddle.static as static

                        paddle.enable_static()

                        exec_strategy = static.ExecutionStrategy()
                        exec_strategy.num_iteration_per_drop_scope = 10
3892
              )DOC")
Q
Qiao Longfei 已提交
3893 3894 3895 3896 3897 3898 3899 3900 3901
      .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
3902 3903 3904 3905 3906 3907 3908 3909 3910 3911 3912 3913
                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
3914
              )DOC")
3915 3916 3917 3918 3919 3920 3921 3922
      .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")
3923 3924 3925 3926 3927 3928
      .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
3929

Y
yuyang18 已提交
3930
  exec_strategy.def_property(
Y
yuyang18 已提交
3931 3932 3933 3934 3935 3936 3937
      "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
3938 3939
      });

C
chengduo 已提交
3940 3941 3942 3943
  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.

3944 3945 3946
    Returns:
        BuildStrategy: An BuildStrategy object.

C
chengduo 已提交
3947 3948 3949
    Examples:
        .. code-block:: python

3950
            import os
3951 3952 3953 3954
            import paddle
            import paddle.static as static

            paddle.enable_static()
3955

3956 3957
            os.environ['CPU_NUM'] = str(2)
            places = static.cpu_places()
3958

3959 3960 3961 3962
            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)
3963

3964
            build_strategy = static.BuildStrategy()
3965 3966
            build_strategy.enable_inplace = True
            build_strategy.memory_optimize = True
3967 3968
            build_strategy.reduce_strategy = static.BuildStrategy.ReduceStrategy.Reduce
            program = static.CompiledProgram(static.default_main_program())
3969
            program = program.with_data_parallel(loss_name=loss.name,
3970 3971
                                                  build_strategy=build_strategy,
                                                  places=places)
C
chengduo 已提交
3972
)DOC");
Y
yuyang18 已提交
3973 3974 3975

  py::enum_<BuildStrategy::ReduceStrategy>(build_strategy, "ReduceStrategy")
      .value("Reduce", BuildStrategy::ReduceStrategy::kReduce)
3976 3977
      .value("AllReduce", BuildStrategy::ReduceStrategy::kAllReduce)
      .value("_NoReduce", BuildStrategy::ReduceStrategy::kNoReduce);
Y
yuyang18 已提交
3978 3979 3980 3981 3982 3983 3984 3985
  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())
3986
      .def("_clear_finalized", &BuildStrategy::ClearFinalized)
Y
yuyang18 已提交
3987 3988 3989 3990
      .def_property(
          "reduce_strategy",
          [](const BuildStrategy &self) { return self.reduce_; },
          [](BuildStrategy &self, BuildStrategy::ReduceStrategy strategy) {
3991 3992
            PADDLE_ENFORCE_NE(self.IsFinalized(),
                              true,
3993 3994 3995
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
Y
yuyang18 已提交
3996
            self.reduce_ = strategy;
C
chengduo 已提交
3997
          },
3998
          R"DOC((fluid.BuildStrategy.ReduceStrategy, optional): there are two reduce
C
chengduo 已提交
3999 4000
                strategies in ParallelExecutor, AllReduce and Reduce. If you want
                that all the parameters' optimization are done on all devices independently,
4001
                you should choose AllReduce; otherwise, if you choose Reduce, all the parameters'
C
chengduo 已提交
4002 4003
                optimization will be evenly distributed to different devices, and then
                broadcast the optimized parameter to other devices.
4004
                Default is 'AllReduce'.
F
flame 已提交
4005 4006 4007 4008

                Examples:
                    .. code-block:: python

4009 4010 4011 4012 4013 4014 4015
                        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
4016
                  )DOC")
Y
yuyang18 已提交
4017 4018 4019 4020 4021
      .def_property(
          "gradient_scale_strategy",
          [](const BuildStrategy &self) { return self.gradient_scale_; },
          [](BuildStrategy &self,
             BuildStrategy::GradientScaleStrategy strategy) {
4022 4023
            PADDLE_ENFORCE_NE(self.IsFinalized(),
                              true,
4024 4025 4026
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
Y
yuyang18 已提交
4027
            self.gradient_scale_ = strategy;
C
chengduo 已提交
4028
          },
4029
          R"DOC((paddle.static.BuildStrategy.GradientScaleStrategy, optional): there are three
4030
                ways of defining :math:`loss@grad` in ParallelExecutor, that is, CoeffNumDevice,
C
chengduo 已提交
4031 4032
                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`,
4033
                you can choose Customized. Default is 'CoeffNumDevice'.
F
flame 已提交
4034 4035 4036 4037

                Examples:
                    .. code-block:: python

C
chengduo 已提交
4038 4039
                        import numpy
                        import os
4040 4041 4042 4043
                        import paddle
                        import paddle.static as static

                        paddle.enable_static()
C
chengduo 已提交
4044 4045

                        use_cuda = True
4046 4047
                        place = paddle.CUDAPlace(0) if use_cuda else paddle.CPUPlace()
                        exe = static.Executor(place)
C
chengduo 已提交
4048 4049

                        # NOTE: If you use CPU to run the program, you need
4050
                        # to specify the CPU_NUM, otherwise, paddle will use
C
chengduo 已提交
4051 4052 4053 4054 4055 4056
                        # 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)
4057
                            places = static.cpu_places()
C
chengduo 已提交
4058
                        else:
4059
                            places = static.cuda_places()
C
chengduo 已提交
4060

4061 4062 4063 4064
                        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
4065

4066
                        exe.run(static.default_startup_program())
C
chengduo 已提交
4067

4068
                        build_strategy = static.BuildStrategy()
C
chengduo 已提交
4069
                        build_strategy.gradient_scale_strategy = \
4070 4071 4072
                                  static.BuildStrategy.GradientScaleStrategy.Customized
                        compiled_prog = static.CompiledProgram(
                                  static.default_main_program()).with_data_parallel(
C
chengduo 已提交
4073
                                          loss_name=loss.name, build_strategy=build_strategy,
4074
                                          places=places)
C
chengduo 已提交
4075 4076 4077 4078 4079 4080

                        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,
4081 4082
                                              feed={"X": x, loss_grad_name : loss_grad},
                                              fetch_list=[loss.name, loss_grad_name])
F
flame 已提交
4083
                   )DOC")
Y
yuyang18 已提交
4084 4085 4086 4087
      .def_property(
          "debug_graphviz_path",
          [](const BuildStrategy &self) { return self.debug_graphviz_path_; },
          [](BuildStrategy &self, const std::string &path) {
4088 4089
            PADDLE_ENFORCE_NE(self.IsFinalized(),
                              true,
4090 4091 4092
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
Y
yuyang18 已提交
4093
            self.debug_graphviz_path_ = path;
C
chengduo 已提交
4094
          },
4095
          R"DOC((str, optional): debug_graphviz_path indicates the path that
F
flame 已提交
4096
                writing the SSA Graph to file in the form of graphviz.
4097
                It is useful for debugging. Default is empty string, that is, ""
F
flame 已提交
4098 4099 4100 4101

                Examples:
                    .. code-block:: python

4102 4103 4104 4105
                        import paddle
                        import paddle.static as static

                        paddle.enable_static()
C
chengduo 已提交
4106

4107 4108
                        build_strategy = static.BuildStrategy()
                        build_strategy.debug_graphviz_path = "./graph"
F
flame 已提交
4109
                    )DOC")
S
sneaxiy 已提交
4110 4111 4112 4113 4114 4115
      .def_property(
          "enable_sequential_execution",
          [](const BuildStrategy &self) {
            return self.enable_sequential_execution_;
          },
          [](BuildStrategy &self, bool b) {
4116 4117
            PADDLE_ENFORCE_NE(self.IsFinalized(),
                              true,
4118 4119 4120
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
S
sneaxiy 已提交
4121 4122
            self.enable_sequential_execution_ = b;
          },
4123 4124
          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
4125 4126 4127 4128

                Examples:
                    .. code-block:: python

4129 4130 4131 4132 4133 4134
                        import paddle
                        import paddle.static as static

                        paddle.enable_static()

                        build_strategy = static.BuildStrategy()
F
flame 已提交
4135 4136
                        build_strategy.enable_sequential_execution = True
          )DOC")
S
sneaxiy 已提交
4137 4138 4139 4140 4141 4142
      .def_property(
          "remove_unnecessary_lock",
          [](const BuildStrategy &self) {
            return self.remove_unnecessary_lock_;
          },
          [](BuildStrategy &self, bool b) {
4143 4144
            PADDLE_ENFORCE_NE(self.IsFinalized(),
                              true,
4145 4146 4147
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
S
sneaxiy 已提交
4148 4149
            self.remove_unnecessary_lock_ = b;
          },
4150 4151
          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
4152 4153 4154 4155

                Examples:
                    .. code-block:: python

4156 4157 4158 4159 4160 4161
                        import paddle
                        import paddle.static as static

                        paddle.enable_static()

                        build_strategy = static.BuildStrategy()
F
flame 已提交
4162 4163
                        build_strategy.remove_unnecessary_lock = True
          )DOC")
4164 4165 4166 4167
      .def_property(
          "num_trainers",
          [](const BuildStrategy &self) { return self.num_trainers_; },
          [](BuildStrategy &self, int num_trainers) {
4168
#ifdef WIN32
4169
            PADDLE_THROW(platform::errors::Unavailable(
4170
                "Distribution mode is not supported on Windows platform."));
4171
#endif
4172 4173
            self.num_trainers_ = num_trainers;
          })
4174 4175 4176 4177 4178 4179 4180
      .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;
          })
4181 4182 4183 4184 4185 4186
      .def_property(
          "trainer_id",
          [](const BuildStrategy &self) { return self.trainer_id_; },
          [](BuildStrategy &self, int trainer_id) {
            self.trainer_id_ = trainer_id;
          })
4187 4188 4189 4190 4191 4192
      .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;
          })
4193 4194 4195 4196 4197 4198
      .def_property(
          "bkcl_comm_num",
          [](const BuildStrategy &self) { return self.bkcl_comm_num_; },
          [](BuildStrategy &self, int bkcl_comm_num) {
            self.bkcl_comm_num_ = bkcl_comm_num;
          })
4199 4200 4201 4202 4203 4204 4205 4206 4207 4208 4209 4210 4211 4212 4213 4214
      .def_property(
          "use_hierarchical_allreduce",
          [](const BuildStrategy &self) {
            return self.use_hierarchical_allreduce_;
          },
          [](BuildStrategy &self, bool use) {
            self.use_hierarchical_allreduce_ = use;
          })
      .def_property(
          "hierarchical_allreduce_inter_nranks",
          [](const BuildStrategy &self) {
            return self.hierarchical_allreduce_inter_nranks_;
          },
          [](BuildStrategy &self, int nranks) {
            self.hierarchical_allreduce_inter_nranks_ = nranks;
          })
4215

C
chengduo 已提交
4216 4217 4218 4219 4220 4221
      .def_property(
          "fuse_elewise_add_act_ops",
          [](const BuildStrategy &self) {
            return self.fuse_elewise_add_act_ops_;
          },
          [](BuildStrategy &self, bool b) {
4222 4223
            PADDLE_ENFORCE_NE(self.IsFinalized(),
                              true,
4224 4225 4226
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
C
chengduo 已提交
4227 4228
            self.fuse_elewise_add_act_ops_ = b;
          },
4229
          R"DOC((bool, optional): fuse_elewise_add_act_ops indicate whether
F
flame 已提交
4230
                to fuse elementwise_add_op and activation_op,
4231
                it may make the execution faster. Default is False.
F
flame 已提交
4232 4233 4234 4235

                Examples:
                    .. code-block:: python

4236 4237 4238 4239 4240 4241
                        import paddle
                        import paddle.static as static

                        paddle.enable_static()

                        build_strategy = static.BuildStrategy()
F
flame 已提交
4242 4243
                        build_strategy.fuse_elewise_add_act_ops = True
                     )DOC")
4244 4245 4246 4247
      .def_property(
          "fuse_gemm_epilogue",
          [](const BuildStrategy &self) { return self.fuse_gemm_epilogue_; },
          [](BuildStrategy &self, bool b) {
4248 4249
            PADDLE_ENFORCE_NE(self.IsFinalized(),
                              true,
4250 4251 4252 4253 4254 4255 4256 4257 4258 4259 4260 4261 4262 4263 4264 4265 4266 4267 4268 4269
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
            self.fuse_gemm_epilogue_ = b;
          },
          R"DOC((bool, optional): fuse_gemm_epilogue indicate whether
                to fuse matmul_op, elemenewist_add_op and activation_op,
                it may make the execution faster. Default is False.

                Examples:
                    .. code-block:: python

                        import paddle
                        import paddle.static as static

                        paddle.enable_static()

                        build_strategy = static.BuildStrategy()
                        build_strategy.fuse_gemm_epilogue = True
                     )DOC")
Z
Zhen Wang 已提交
4270 4271 4272 4273
      .def_property(
          "fuse_bn_act_ops",
          [](const BuildStrategy &self) { return self.fuse_bn_act_ops_; },
          [](BuildStrategy &self, bool b) {
4274 4275
            PADDLE_ENFORCE_NE(self.IsFinalized(),
                              true,
Z
Zhen Wang 已提交
4276
                              platform::errors::PreconditionNotMet(
4277 4278
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
Z
Zhen Wang 已提交
4279 4280 4281 4282 4283 4284 4285 4286 4287
            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

4288 4289 4290 4291 4292 4293
                        import paddle
                        import paddle.static as static

                        paddle.enable_static()

                        build_strategy = static.BuildStrategy()
Z
Zhen Wang 已提交
4294 4295
                        build_strategy.fuse_bn_act_ops = True
                     )DOC")
Z
Zhang Ting 已提交
4296 4297 4298 4299
      .def_property(
          "fuse_bn_add_act_ops",
          [](const BuildStrategy &self) { return self.fuse_bn_add_act_ops_; },
          [](BuildStrategy &self, bool b) {
4300 4301
            PADDLE_ENFORCE_NE(self.IsFinalized(),
                              true,
Z
Zhang Ting 已提交
4302 4303 4304 4305 4306 4307 4308 4309 4310 4311 4312 4313 4314 4315 4316 4317 4318 4319 4320 4321
                              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")
4322 4323 4324 4325
      .def_property(
          "enable_auto_fusion",
          [](const BuildStrategy &self) { return self.enable_auto_fusion_; },
          [](BuildStrategy &self, bool b) {
4326 4327
            PADDLE_ENFORCE_NE(self.IsFinalized(),
                              true,
4328
                              platform::errors::PreconditionNotMet(
4329 4330
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
4331 4332 4333 4334 4335 4336 4337 4338 4339 4340
            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

4341 4342 4343 4344 4345 4346
                        import paddle
                        import paddle.static as static

                        paddle.enable_static()

                        build_strategy = static.BuildStrategy()
4347 4348
                        build_strategy.enable_auto_fusion = True
                    )DOC")
4349 4350 4351 4352 4353 4354
      .def_property(
          "fuse_relu_depthwise_conv",
          [](const BuildStrategy &self) {
            return self.fuse_relu_depthwise_conv_;
          },
          [](BuildStrategy &self, bool b) {
4355 4356
            PADDLE_ENFORCE_NE(self.IsFinalized(),
                              true,
4357 4358 4359
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
4360 4361
            self.fuse_relu_depthwise_conv_ = b;
          },
4362
          R"DOC((bool, optional): fuse_relu_depthwise_conv indicate whether
F
flame 已提交
4363 4364 4365
                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.
4366
                Default is False.
F
flame 已提交
4367 4368 4369 4370

                Examples:
                    .. code-block:: python

4371 4372 4373 4374 4375 4376
                        import paddle
                        import paddle.static as static

                        paddle.enable_static()

                        build_strategy = static.BuildStrategy()
F
flame 已提交
4377 4378
                        build_strategy.fuse_relu_depthwise_conv = True
          )DOC")
4379 4380 4381 4382 4383 4384 4385
      .def_property(
          "fuse_broadcast_ops",
          [](const BuildStrategy &self) {
            return self.fuse_broadcast_ops_ == true ||
                   self.fuse_broadcast_ops_ == paddle::none;
          },
          [](BuildStrategy &self, bool b) {
4386 4387
            PADDLE_ENFORCE_NE(self.IsFinalized(),
                              true,
4388 4389 4390 4391 4392 4393
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, "
                                  "cannot be configured again."));
            self.fuse_broadcast_ops_ = b;
          },
          R"DOC((bool, optional): fuse_broadcast_op indicates whether
4394 4395 4396 4397
                      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
4398 4399 4400 4401 4402
                      for NCCLReduce operations for a period of time. Default False.

                      Examples:
                          .. code-block:: python

4403 4404 4405 4406 4407 4408
                              import paddle
                              import paddle.static as static

                              paddle.enable_static()

                              build_strategy = static.BuildStrategy()
4409 4410
                              build_strategy.fuse_broadcast_ops = True
                    )DOC")
4411 4412 4413 4414 4415 4416 4417
      .def_property(
          "fuse_all_optimizer_ops",
          [](const BuildStrategy &self) {
            return self.fuse_all_optimizer_ops_ == true ||
                   self.fuse_all_optimizer_ops_ == paddle::none;
          },
          [](BuildStrategy &self, bool b) {
4418 4419
            PADDLE_ENFORCE_NE(self.IsFinalized(),
                              true,
4420 4421 4422 4423 4424
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, "
                                  "cannot be configured again."));
            self.fuse_all_optimizer_ops_ = b;
          })
Q
qingqing01 已提交
4425 4426 4427 4428
      .def_property(
          "sync_batch_norm",
          [](const BuildStrategy &self) { return self.sync_batch_norm_; },
          [](BuildStrategy &self, bool b) {
4429 4430
            PADDLE_ENFORCE_NE(self.IsFinalized(),
                              true,
4431 4432 4433
                              platform::errors::PreconditionNotMet(
                                  "BuildStrategy has been finlaized, cannot be "
                                  "configured again."));
Q
qingqing01 已提交
4434 4435
            self.sync_batch_norm_ = b;
          },
4436
          R"DOC((bool, optional): sync_batch_norm indicates whether to use
Q
qingqing01 已提交
4437 4438 4439
                synchronous batch normalization which synchronizes the mean
                and variance through multi-devices in training phase.
                Current implementation doesn't support FP16 training and CPU.
4440 4441
                And only synchronous on one machine, not all machines. 
                Default is False.
F
flame 已提交
4442 4443 4444 4445

                Examples:
                    .. code-block:: python

4446 4447 4448 4449 4450 4451
                        import paddle
                        import paddle.static as static

                        paddle.enable_static()

                        build_strategy = static.BuildStrategy()
F
flame 已提交
4452 4453
                        build_strategy.sync_batch_norm = True
                )DOC")
D
dzhwinter 已提交
4454 4455
      .def_property(
          "memory_optimize",
4456 4457 4458 4459 4460 4461 4462 4463 4464 4465
          [](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) {
4466
              self.memory_optimize_ = paddle::none;
4467 4468 4469
            } else if (PyBool_Check(py_obj)) {
              self.memory_optimize_ = (py_obj == Py_True);
            } else {
4470
              PADDLE_THROW(platform::errors::InvalidArgument(
Z
Zeng Jinle 已提交
4471 4472
                  "BuildStrategy.memory_optimize must be set to None, False "
                  "or True"));
4473 4474
            }
          },
4475
          R"DOC((bool, optional): memory opitimize aims to save total memory
4476
                consumption, set to True to enable it.
4477

4478 4479 4480
                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. 
4481 4482 4483 4484 4485 4486 4487 4488 4489 4490 4491 4492 4493 4494
                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")
4495 4496 4497
      .def_property(
          "is_distribution",
          [](const BuildStrategy &self) { return self.is_distribution_; },
4498 4499 4500
          [](BuildStrategy &self, bool b) {
#ifdef WIN32
            if (b) {
4501
              PADDLE_THROW(platform::errors::Unavailable(
4502
                  "Distribution mode is not supported on Windows platform."));
4503 4504 4505 4506 4507
            }
#else
            self.is_distribution_ = b;
#endif
          })
4508 4509 4510 4511
      .def_property(
          "async_mode",
          [](const BuildStrategy &self) { return self.async_mode_; },
          [](BuildStrategy &self, bool b) { self.async_mode_ = b; })
D
dzhwinter 已提交
4512
      .def_property(
D
dzhwinter 已提交
4513 4514 4515
          "enable_inplace",
          [](const BuildStrategy &self) { return self.enable_inplace_; },
          [](BuildStrategy &self, bool b) { self.enable_inplace_ = b; })
4516 4517 4518 4519
      .def_property(
          "enable_addto",
          [](const BuildStrategy &self) { return self.enable_addto_; },
          [](BuildStrategy &self, bool b) { self.enable_addto_ = b; })
C
chengduo 已提交
4520 4521
      .def_property(
          "fuse_all_reduce_ops",
C
chengduo 已提交
4522 4523
          [](const BuildStrategy &self) {
            return self.fuse_all_reduce_ops_ == true ||
4524
                   self.fuse_all_reduce_ops_ == paddle::none;
C
chengduo 已提交
4525
          },
C
chengduo 已提交
4526
          [](BuildStrategy &self, bool b) { self.fuse_all_reduce_ops_ = b; })
4527 4528 4529 4530 4531 4532 4533 4534
      .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;
          })
4535 4536 4537 4538
      .def_property(
          "cache_runtime_context",
          [](const BuildStrategy &self) { return self.cache_runtime_context_; },
          [](BuildStrategy &self, bool b) { self.cache_runtime_context_ = b; })
4539 4540 4541 4542 4543 4544 4545 4546 4547
      .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;
          })
Z
Zeng Jinle 已提交
4548 4549 4550 4551 4552 4553
      .def_property(
          "fix_op_run_order",
          [](const BuildStrategy &self) { return self.fix_op_run_order_; },
          [](BuildStrategy &self, bool fix_op_run_order) {
            self.fix_op_run_order_ = fix_op_run_order;
          })
4554 4555 4556 4557 4558 4559 4560 4561
      .def_property(
          "allow_cuda_graph_capture",
          [](const BuildStrategy &self) {
            return self.allow_cuda_graph_capture_;
          },
          [](BuildStrategy &self, bool allow_cuda_graph_capture) {
            self.allow_cuda_graph_capture_ = allow_cuda_graph_capture;
          })
4562 4563 4564 4565 4566 4567
      .def("_copy",
           [](const BuildStrategy &self) {
             auto new_bs = self;
             new_bs.ClearFinalized();
             return new_bs;
           })
4568 4569 4570 4571 4572 4573
      .def(
          "_finalize_strategy_and_create_passes",
          [](BuildStrategy &self) -> std::shared_ptr<ir::PassBuilder> {
            return self.CreatePassesFromStrategy(true);
          },
          R"DOC(Allow user to customized passes. Normally model-specific
4574 4575
                optimization passes should be defined in this way. BuildStrategy
                cannot be updated after being finalized.)DOC");
Y
yuyang18 已提交
4576

4577 4578 4579 4580 4581 4582
  m.def("_set_cached_executor_build_strategy",
        [](int64_t program_id, const BuildStrategy &build_strategy) {
          auto &cached_exe_info = framework::ExecutorInfoCache::Instance();
          cached_exe_info.SetBuildStrategy(program_id, build_strategy);
        });

Y
yuyang18 已提交
4583
  pe.def(py::init<const std::vector<platform::Place> &,
4584 4585 4586 4587 4588 4589 4590
                  const std::vector<std::string> &,
                  const std::string &,
                  Scope *,
                  std::vector<Scope *> &,
                  const ExecutionStrategy &,
                  const BuildStrategy &,
                  ir::Graph *>())
Y
Yu Yang 已提交
4591 4592 4593 4594
      // 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.
4595 4596 4597 4598 4599 4600
      .def(
          "local_scopes",
          [](ParallelExecutor &self) -> std::vector<Scope *> * {
            return &self.GetLocalScopes();
          },
          py::return_value_policy::reference)
4601 4602 4603
      .def("drop_local_exe_scopes", &ParallelExecutor::DropLocalExeScopes)
      .def("_need_create_local_exe_scopes",
           &ParallelExecutor::NeedCreateLocalExeScope)
Y
Yu Yang 已提交
4604 4605 4606 4607
      .def("feed_tensors_into_local_scopes",
           &ParallelExecutor::FeedTensorsIntoLocalScopes)
      .def("feed_and_split_tensor_into_local_scopes",
           &ParallelExecutor::FeedAndSplitTensorIntoLocalScopes)
4608 4609
      .def("run",
           [](ParallelExecutor &self,
Z
Zhen Wang 已提交
4610 4611 4612
              const std::vector<std::string> &fetch_tensors,
              bool return_merged) -> py::object {
             if (return_merged) {
4613 4614 4615 4616 4617 4618
               paddle::framework::FetchList ret;
               /*gil_scoped_release*/ {
                 pybind11::gil_scoped_release release;
                 ret = self.RunAndMerge(fetch_tensors);
               }
               return py::cast(std::move(ret));
Z
Zhen Wang 已提交
4619
             } else {
4620 4621 4622 4623 4624 4625
               paddle::framework::FetchUnmergedList ret;
               /*gil_scoped_release*/ {
                 pybind11::gil_scoped_release release;
                 ret = self.Run(fetch_tensors);
               }
               return py::cast(std::move(ret));
Z
Zhen Wang 已提交
4626
             }
4627 4628
           })
      .def("device_count", &ParallelExecutor::DeviceCount);
Y
Yu Yang 已提交
4629

J
jianghaicheng 已提交
4630 4631
#ifdef PADDLE_WITH_IPU
  py::class_<platform::ipu::IpuBackend,
4632 4633 4634
             std::unique_ptr<platform::ipu::IpuBackend, py::nodelete>>(
      m, "IpuBackend")
      // manage IpuBackend in C++
4635 4636 4637 4638 4639 4640 4641
      .def(
          "get_instance",
          []() {
            return std::unique_ptr<platform::ipu::IpuBackend, py::nodelete>(
                platform::ipu::IpuBackend::GetInstance());
          },
          py::return_value_policy::reference)
A
Allen Guo 已提交
4642
      .def("weights_to_host", &platform::ipu::IpuBackend::WeightsToHost)
4643 4644
      .def("detach", &platform::ipu::IpuBackend::Detach)
      .def("reset", &platform::ipu::IpuBackend::Reset)
J
jianghaicheng 已提交
4645
      .def("set_scope", &platform::ipu::IpuBackend::SetScope)
4646 4647 4648 4649 4650 4651 4652 4653 4654 4655
      .def("set_ipu_strategy", &platform::ipu::IpuBackend::SetIpuStrategy)
      .def("save_model_proto", &platform::ipu::IpuBackend::SaveModelProto);

  py::class_<platform::ipu::IpuStrategy>(m, "IpuStrategy")
      .def(py::init())
      .def("set_options",
           [](platform::ipu::IpuStrategy &self, const py::dict &opt) {
             for (auto element : opt) {
               auto option_name = element.first.cast<std::string>();
               VLOG(10) << "Set option: " << option_name;
A
Allen Guo 已提交
4656 4657 4658 4659
               if (option_name == "compilation_progress_logger") {
                 self.SetCompilationProgressLogger(
                     element.second.cast<py::function>());
               } else if (py::isinstance<py::bool_>(element.second)) {
4660 4661 4662 4663 4664 4665 4666 4667 4668 4669 4670 4671 4672 4673 4674 4675 4676 4677 4678 4679 4680 4681
                 self.AddBoolOption(option_name, element.second.cast<bool>());
               } else if (py::isinstance<py::float_>(element.second)) {
                 self.AddDoubleOption(option_name,
                                      element.second.cast<double>());
               } else if (py::isinstance<py::int_>(element.second)) {
                 self.AddUint64Option(option_name,
                                      element.second.cast<std::uint64_t>());
               } else if (py::isinstance<py::str>(element.second)) {
                 self.AddStringOption(option_name,
                                      element.second.cast<std::string>());
               } else if (py::isinstance<py::set>(element.second) ||
                          py::isinstance<py::list>(element.second)) {
                 for (auto option : element.second.cast<py::list>()) {
                   std::string option_val;
                   if (py::isinstance<py::str>(option)) {
                     option_val = option.cast<std::string>();
                   } else if (py::isinstance<py::int_>(option)) {
                     option_val = std::to_string(option.cast<std::uint64_t>());
                   } else {
                     PADDLE_THROW(platform::errors::Unimplemented(
                         "Failed to convert type: %s when set IpuStrategy "
                         "option: %s",
4682 4683
                         option.get_type(),
                         option_name));
4684 4685 4686 4687 4688 4689 4690
                   }
                   self.InsertStringOption(option_name, option_val);
                 }
               } else if (py::isinstance<py::dict>(element.second)) {
                 if (option_name.rfind("location_", 0) == 0) {
                   for (auto option : element.second.cast<py::dict>()) {
                     self.SetTensorLocation(
4691 4692
                         option_name,
                         option.first.cast<std::string>(),
4693 4694
                         option.second.cast<std::uint64_t>());
                   }
4695 4696 4697 4698 4699 4700
                 } else if (option_name == "replicated_collectives_settings") {
                   for (auto option : element.second.cast<py::dict>()) {
                     self.SetReplicatedCollectivesSettings(
                         option.first.cast<std::string>(),
                         option.second.cast<bool>());
                   }
A
Allen Guo 已提交
4701 4702 4703 4704 4705 4706 4707 4708 4709
                 } else if (option_name == "accumulate_outer_fragment") {
                   for (auto option : element.second.cast<py::dict>()) {
                     std::vector<int> values;
                     for (auto value : option.second.cast<py::list>()) {
                       values.push_back(value.cast<int>());
                     }
                     self.SetAccumulateOuterFragmentSettings(
                         option.first.cast<std::uint64_t>(), values);
                   }
4710 4711 4712 4713 4714 4715 4716 4717 4718 4719 4720 4721 4722 4723 4724 4725 4726 4727 4728 4729 4730 4731 4732 4733 4734 4735 4736 4737 4738 4739 4740 4741 4742 4743 4744 4745
                 } else if (option_name == "custom_op") {
                   std::string paddle_op;
                   std::string popart_op;
                   std::string domain;
                   int version = -1;
                   for (auto option : element.second.cast<py::dict>()) {
                     std::string option_key = option.first.cast<std::string>();
                     if (option_key == "paddle_op") {
                       paddle_op = option.second.cast<std::string>();
                     } else if (option_key == "popart_op") {
                       popart_op = option.second.cast<std::string>();
                     } else if (option_key == "domain") {
                       domain = option.second.cast<std::string>();
                     } else if (option_key == "version") {
                       version = option.second.cast<int>();
                     } else {
                       PADDLE_THROW(platform::errors::InvalidArgument(
                           "Invalid argument, key must be one of paddle_op, "
                           "popart_op, domain or version, but revecived %s",
                           option_key));
                     }
                   }
                   self.AddCustomOp(paddle_op, popart_op, domain, version);
                 } else {
                   for (auto option : element.second.cast<py::dict>()) {
                     std::string option_key = option.first.cast<std::string>();
                     std::string option_val;
                     if (py::isinstance<py::str>(option.second)) {
                       option_val = option.second.cast<std::string>();
                     } else if (py::isinstance<py::int_>(option.second)) {
                       option_val =
                           std::to_string(option.second.cast<std::uint64_t>());
                     } else {
                       PADDLE_THROW(platform::errors::Unimplemented(
                           "Failed to convert value type: %s when set "
                           "IpuStrategy option: %s",
4746 4747
                           option.second.get_type(),
                           option_key));
4748
                     }
4749 4750
                     self.InsertStringPairOption(
                         option_name, option_key, option_val);
4751 4752 4753 4754 4755 4756
                   }
                 }
               } else {
                 PADDLE_THROW(platform::errors::InvalidArgument(
                     "Invalid IpuStrategy option value type: %s, please check "
                     "input value for option: %s",
4757 4758
                     element.second.get_type(),
                     option_name));
4759 4760 4761 4762 4763 4764 4765 4766 4767 4768 4769 4770 4771 4772 4773 4774 4775 4776 4777 4778 4779 4780 4781 4782 4783 4784 4785 4786 4787 4788
               }
             }
           })
      .def("get_option",
           [](platform::ipu::IpuStrategy &self, const std::string &name) {
             py::dict res;
             auto option_type = self.GetOptionType(name);
             res["name"] = name;
             res["type"] = option_type;
             if (option_type == "vector") {
               auto value = self.GetVectorOption(name);
               res["value"] = value;
             } else if (option_type == "map") {
               auto value = self.GetMapOption(name);
               res["value"] = value;
             } else {
               auto value_s = self.GetOption(name);
               res["value_s"] = value_s;
               if (option_type == "bool") {
                 res["value"] = static_cast<bool>(std::stoi(value_s));
               } else if (option_type == "uint64") {
                 res["value"] = std::stoul(value_s);
               } else if (option_type == "double") {
                 res["value"] = std::stod(value_s);
               } else if (option_type == "string") {
                 res["value"] = value_s;
               }
             }
             return res;
           })
4789 4790
      .def("get_all_option_names",
           &platform::ipu::IpuStrategy::GetAllOptionNames)
4791 4792 4793
      .def("enable_pattern", &platform::ipu::IpuStrategy::EnablePattern)
      .def("disable_pattern", &platform::ipu::IpuStrategy::DisablePattern)
      .def("is_pattern_enabled", &platform::ipu::IpuStrategy::IsPatternEnabled);
J
jianghaicheng 已提交
4794 4795
#endif

4796 4797 4798 4799 4800 4801 4802 4803
  m.def("enable_autotune", [] {
    return phi::autotune::AutoTuneStatus::Instance().EnableAutoTune();
  });

  m.def("disable_autotune", [] {
    return phi::autotune::AutoTuneStatus::Instance().DisableAutoTune();
  });

4804
  m.def("set_autotune_range", [](int64_t start, int64_t stop) {
4805 4806 4807 4808 4809 4810 4811 4812 4813
    return phi::autotune::AutoTuneStatus::Instance().SetAutoTuneRange(start,
                                                                      stop);
  });

  m.def("update_autotune_status",
        [] { return phi::autotune::AutoTuneStatus::Instance().Update(); });

  m.def("autotune_status", [] {
    py::dict res;
4814
    phi::autotune::AutoTuneCache::Instance().UpdateStatus();
4815 4816 4817 4818 4819 4820 4821
    res["step_id"] = phi::autotune::AutoTuneStatus::Instance().StepID();
    res["cache_size"] = phi::autotune::AutoTuneCache::Instance().Size();
    res["cache_hit_rate"] =
        phi::autotune::AutoTuneCache::Instance().CacheHitRate();
    return res;
  });

4822 4823 4824 4825 4826 4827 4828 4829 4830 4831 4832 4833 4834 4835
  m.def("enable_layout_autotune", [] {
    return paddle::imperative::LayoutAutoTune::Instance()
        .EnableLayoutAutoTune();
  });

  m.def("disable_layout_autotune", [] {
    return paddle::imperative::LayoutAutoTune::Instance()
        .DisableLayoutAutoTune();
  });

  m.def("use_layout_autotune", [] {
    return paddle::imperative::LayoutAutoTune::Instance().UseLayoutAutoTune();
  });

D
dongdaxiang 已提交
4836
  BindFleetWrapper(&m);
4837
  BindIO(&m);
T
Thunderbrook 已提交
4838

T
Thunderbrook 已提交
4839
#if defined(PADDLE_WITH_PSLIB) && !defined(PADDLE_WITH_HETERPS)
T
Thunderbrook 已提交
4840
  BindHeterWrapper(&m);
4841
  BindMetrics(&m);
T
Thunderbrook 已提交
4842
#endif
T
Thunderbrook 已提交
4843
#ifdef PADDLE_WITH_HETERPS
T
Thunderbrook 已提交
4844
  BindPSGPUWrapper(&m);
T
Thunderbrook 已提交
4845 4846 4847
#ifdef PADDLE_WITH_PSLIB
  BindAfsWrapper(&m);
#endif
T
Thunderbrook 已提交
4848
#endif
4849
  BindGlooWrapper(&m);
H
hutuxian 已提交
4850
  BindBoxHelper(&m);
H
hutuxian 已提交
4851 4852 4853
#ifdef PADDLE_WITH_BOX_PS
  BindBoxWrapper(&m);
#endif
4854
#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
D
dongdaxiang 已提交
4855
  BindNCCLWrapper(&m);
4856 4857 4858
#endif
#ifdef PADDLE_WITH_GLOO
  BindGlooContext(&m);
W
wopeizl 已提交
4859
#endif
F
flame 已提交
4860 4861
  BindGraph(&m);
  BindNode(&m);
4862
  BindPass(&m);
F
flame 已提交
4863
  BindInferenceApi(&m);
4864
  BindCompatible(&m);
4865
  BindDataset(&m);
Y
yaoxuefeng 已提交
4866
  BindGenerator(&m);
4867 4868 4869
#ifndef PADDLE_ON_INFERENCE
  BindDistributed(&m);
#endif
4870 4871 4872
#ifdef PADDLE_WITH_ASCEND
  BindAscendWrapper(&m);
  BindAscendGraph(&m);
4873
  BindAscendDevice(&m);
4874
#endif
Y
Yanghello 已提交
4875 4876 4877
#ifdef PADDLE_WITH_CRYPTO
  BindCrypto(&m);
#endif
T
tangwei12 已提交
4878

T
tangwei12 已提交
4879
#if defined PADDLE_WITH_PSCORE
T
tangwei12 已提交
4880 4881
  BindDistFleetWrapper(&m);
  BindPSHost(&m);
4882
  BindCommunicatorContext(&m);
T
tangwei12 已提交
4883 4884
  BindDistCommunicator(&m);
  BindHeterClient(&m);
S
seemingwang 已提交
4885 4886 4887 4888 4889
  BindGraphPyFeatureNode(&m);
  BindGraphNode(&m);
  BindGraphPyService(&m);
  BindGraphPyServer(&m);
  BindGraphPyClient(&m);
1
123malin 已提交
4890 4891 4892 4893
  BindIndexNode(&m);
  BindTreeIndex(&m);
  BindIndexWrapper(&m);
  BindIndexSampler(&m);
4894
#ifdef PADDLE_WITH_HETERPS
4895 4896
  BindNodeQueryResult(&m);
  BindNeighborSampleQuery(&m);
4897 4898 4899
  BindNeighborSampleResult(&m);
  BindGraphGpuWrapper(&m);
#endif
4900
#endif
L
Luo Tao 已提交
4901
}
4902
}  // namespace pybind
4903
}  // namespace paddle