distributed_py.cc 36.8 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
/* Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */

#include <fcntl.h>
#ifdef _POSIX_C_SOURCE
#undef _POSIX_C_SOURCE
#endif

#ifdef _XOPEN_SOURCE
#undef _XOPEN_SOURCE
#endif

#include "paddle/fluid/distributed/collective/ProcessGroup.h"
25
#include "paddle/fluid/distributed/collective/ProcessGroupStream.h"
26
#include "paddle/fluid/distributed/collective/Types.h"
27
#include "paddle/fluid/distributed/collective/Utils.h"
28
#include "paddle/fluid/distributed/collective/reducer.h"
29 30 31 32 33 34 35
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/imperative/layer.h"
#include "paddle/fluid/pybind/distributed_py.h"
#include "paddle/fluid/pybind/eager_utils.h"
#include "paddle/phi/api/all.h"

36
#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
37 38 39
#include "paddle/fluid/distributed/collective/ProcessGroupNCCL.h"
#endif

W
wuhuachaocoding 已提交
40 41 42 43
#if defined(PADDLE_WITH_MPI)
#include "paddle/fluid/distributed/collective/ProcessGroupMPI.h"
#endif

44 45 46 47
#if defined(PADDLE_WITH_ASCEND_CL)
#include "paddle/fluid/distributed/collective/ProcessGroupHCCL.h"
#endif

48 49 50 51
#if defined(PADDLE_WITH_CUSTOM_DEVICE)
#include "paddle/fluid/distributed/collective/ProcessGroupCustom.h"
#endif

52 53 54 55 56
#if defined(PADDLE_WITH_GLOO) && defined(PADDLE_WITH_PSCORE) && \
    (defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_ASCEND_CL))
#include "paddle/fluid/distributed/collective/ProcessGroupHeter.h"
#endif

57 58 59 60 61
#if defined(PADDLE_WITH_GLOO)
#include "paddle/fluid/distributed/collective/ProcessGroupGloo.h"
#include "paddle/fluid/distributed/store/tcp_store.h"
#endif

62 63 64 65 66 67 68
namespace py = pybind11;

namespace paddle {
namespace pybind {

using Tensor = paddle::experimental::Tensor;

69 70 71 72 73
std::shared_ptr<distributed::EagerReducer> CreateEagerReducer(
    py::handle py_tensors,
    const std::vector<std::vector<size_t>> &group_indices,
    const std::vector<bool> &is_sparse_gradient,
    std::shared_ptr<distributed::ProcessGroup> process_group,
74 75
    const std::vector<size_t> &group_size_limits,
    bool find_unused_parameters) {
76
  auto params = CastPyArg2VectorOfTensor(py_tensors.ptr(), 0);
77 78 79 80 81 82
  return std::make_shared<distributed::EagerReducer>(params,
                                                     group_indices,
                                                     is_sparse_gradient,
                                                     process_group,
                                                     group_size_limits,
                                                     find_unused_parameters);
83 84
}

85 86 87 88 89 90 91 92
#if defined(PADDLE_WITH_GLOO)
using ProcessGroupGloo = paddle::distributed::ProcessGroupGloo;
using GlooStore = paddle::distributed::ProcessGroupGloo::GlooStore;
using GlooOptions = paddle::distributed::ProcessGroupGloo::GlooOptions;
#endif

static std::string GLOO_SOCKET_IFNAME_ENV = "GLOO_SOCKET_IFNAME";  // NOLINT

93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110
void BindDistributed(py::module *m) {
  py::enum_<distributed::ReduceOp>(*m, "ReduceOp")
      .value("SUM", distributed::ReduceOp::SUM)
      .value("AVG", distributed::ReduceOp::AVG)
      .value("MAX", distributed::ReduceOp::MAX)
      .value("MIN", distributed::ReduceOp::MIN)
      .value("PRODUCT", distributed::ReduceOp::PRODUCT);

  py::class_<distributed::AllreduceOptions>(*m, "AllreduceOptions")
      .def(py::init<>())
      .def_readwrite("reduce_op", &distributed::AllreduceOptions::reduce_op);

  py::class_<distributed::BroadcastOptions>(*m, "BroadcastOptions")
      .def(py::init<>())
      .def_readwrite("source_rank", &distributed::BroadcastOptions::source_rank)
      .def_readwrite("source_root",
                     &distributed::BroadcastOptions::source_root);

B
Baibaifan 已提交
111 112 113 114
  py::class_<distributed::BarrierOptions>(*m, "BarrierOptions")
      .def(py::init<>())
      .def_readwrite("place_ids", &distributed::BarrierOptions::place_ids);

115 116 117 118 119
  py::class_<distributed::ReduceOptions>(*m, "ReduceOptions")
      .def(py::init<>())
      .def_readwrite("reduce_op", &distributed::ReduceOptions::reduce_op)
      .def_readwrite("source_root", &distributed::ReduceOptions::root_rank);

120 121 122 123 124 125
  auto ProcessGroup =
      py::class_<distributed::ProcessGroup,
                 std::shared_ptr<distributed::ProcessGroup>>(*m, "ProcessGroup")
          .def("rank", &distributed::ProcessGroup::GetRank)
          .def("size", &distributed::ProcessGroup::GetSize)
          .def("name", &distributed::ProcessGroup::GetBackendName)
126 127
          .def(
              "allreduce",
128 129
              [](distributed::ProcessGroup &self,
                 py::handle py_tensor,
130 131 132 133 134 135 136 137 138
                 distributed::ReduceOp op) {
                auto tensor = CastPyArg2Tensor(py_tensor.ptr(), 0);
                distributed::AllreduceOptions opts;
                opts.reduce_op = op;
                auto dense =
                    std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl());
                std::vector<phi::DenseTensor> tensors = {*dense};
                return self.AllReduce(tensors, tensors, opts);
              },
139 140
              py::arg("tensor"),
              py::arg("op") = distributed::ReduceOp::SUM,
141 142
              py::call_guard<py::gil_scoped_release>())

143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161
          .def(
              "allreduce",
              [](distributed::ProcessGroup &self,
                 py::handle py_tensor,
                 distributed::ReduceOp op,
                 bool sync_op) {
                auto tensor = CastPyArg2Tensor(py_tensor.ptr(), 0);
                distributed::AllreduceOptions opts;
                opts.reduce_op = op;
                auto dense =
                    std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl());
                std::vector<phi::DenseTensor> tensors = {*dense};
                return self.AllReduce(tensors, tensors, opts, sync_op);
              },
              py::arg("tensor"),
              py::arg("op"),
              py::arg("sync_op"),
              py::call_guard<py::gil_scoped_release>())

162 163
          .def(
              "broadcast",
164 165
              [](distributed::ProcessGroup &self,
                 py::handle py_tensor,
166 167 168 169 170 171 172 173 174
                 int source_rank) {
                auto tensor = CastPyArg2Tensor(py_tensor.ptr(), 0);
                distributed::BroadcastOptions opts;
                opts.source_rank = source_rank;
                auto dense =
                    std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl());
                std::vector<phi::DenseTensor> tensors = {*dense};
                return self.Broadcast(tensors, tensors, opts);
              },
175 176
              py::arg("tensor"),
              py::arg("source_rank"),
177 178 179 180 181 182 183 184 185 186 187 188 189 190
              py::call_guard<py::gil_scoped_release>())

          .def(
              "barrier",
              [](distributed::ProcessGroup &self, std::vector<int> place_ids) {
                distributed::BarrierOptions opts;
                opts.place_ids = place_ids;
                return self.Barrier(opts);
              },
              py::arg("place_ids") = std::vector<int>{},
              py::call_guard<py::gil_scoped_release>())

          .def(
              "send",
191 192
              [](distributed::ProcessGroup &self,
                 py::handle py_tensor,
193 194 195 196 197 198 199
                 int dst) {
                auto tensor = CastPyArg2Tensor(py_tensor.ptr(), 0);
                auto dense =
                    std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl());
                std::vector<phi::DenseTensor> tensors = {*dense};
                return self.Send(tensors, dst);
              },
200 201
              py::arg("tensor"),
              py::arg("dst"),
202 203
              py::call_guard<py::gil_scoped_release>())

204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220
          .def(
              "send",
              [](distributed::ProcessGroup &self,
                 py::handle py_tensor,
                 int dst,
                 bool sync_op) {
                auto tensor = CastPyArg2Tensor(py_tensor.ptr(), 0);
                auto dense =
                    std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl());
                std::vector<phi::DenseTensor> tensors = {*dense};
                return self.Send(tensors, dst, sync_op);
              },
              py::arg("tensor"),
              py::arg("dst"),
              py::arg("sync_op"),
              py::call_guard<py::gil_scoped_release>())

221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241
          .def(
              "send_partial",
              [](distributed::ProcessGroup &self,
                 py::handle py_tensor,
                 int dst_rank,
                 int nranks,
                 int rank_id) {
                auto tensor = CastPyArg2Tensor(py_tensor.ptr(), 0);
                auto dense =
                    std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl());
                int numel = (*dense).numel();
                int send_numel = numel / nranks;
                int offset = send_numel * rank_id;
                return self.Send_Partial(*dense, dst_rank, offset, send_numel);
              },
              py::arg("tensor"),
              py::arg("dst"),
              py::arg("num"),
              py::arg("id"),
              py::call_guard<py::gil_scoped_release>())

242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265
          .def(
              "send_partial",
              [](distributed::ProcessGroup &self,
                 py::handle py_tensor,
                 int dst_rank,
                 int nranks,
                 int rank_id,
                 bool sync_op) {
                auto tensor = CastPyArg2Tensor(py_tensor.ptr(), 0);
                auto dense =
                    std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl());
                int numel = (*dense).numel();
                int send_numel = numel / nranks;
                int offset = send_numel * rank_id;
                return self.Send_Partial(
                    *dense, dst_rank, offset, send_numel, sync_op);
              },
              py::arg("tensor"),
              py::arg("dst"),
              py::arg("num"),
              py::arg("id"),
              py::arg("sync_op"),
              py::call_guard<py::gil_scoped_release>())

266 267
          .def(
              "recv",
268 269
              [](distributed::ProcessGroup &self,
                 py::handle py_tensor,
270 271 272 273 274 275 276
                 int src) {
                auto tensor = CastPyArg2Tensor(py_tensor.ptr(), 0);
                auto dense =
                    std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl());
                std::vector<phi::DenseTensor> tensors = {*dense};
                return self.Recv(tensors, src);
              },
277 278
              py::arg("tensor"),
              py::arg("src"),
279 280
              py::call_guard<py::gil_scoped_release>())

281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297
          .def(
              "recv",
              [](distributed::ProcessGroup &self,
                 py::handle py_tensor,
                 int src,
                 bool sync_op) {
                auto tensor = CastPyArg2Tensor(py_tensor.ptr(), 0);
                auto dense =
                    std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl());
                std::vector<phi::DenseTensor> tensors = {*dense};
                return self.Recv(tensors, src, sync_op);
              },
              py::arg("tensor"),
              py::arg("src"),
              py::arg("sync_op"),
              py::call_guard<py::gil_scoped_release>())

298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318
          .def(
              "recv_partial",
              [](distributed::ProcessGroup &self,
                 py::handle py_tensor,
                 int src_rank,
                 int nranks,
                 int rank_id) {
                auto tensor = CastPyArg2Tensor(py_tensor.ptr(), 0);
                auto dense =
                    std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl());
                int numel = (*dense).numel();
                int recv_numel = numel / nranks;
                int offset = recv_numel * rank_id;
                return self.Recv_Partial(*dense, src_rank, offset, recv_numel);
              },
              py::arg("tensor"),
              py::arg("src"),
              py::arg("num"),
              py::arg("id"),
              py::call_guard<py::gil_scoped_release>())

319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342
          .def(
              "recv_partial",
              [](distributed::ProcessGroup &self,
                 py::handle py_tensor,
                 int src_rank,
                 int nranks,
                 int rank_id,
                 bool sync_op) {
                auto tensor = CastPyArg2Tensor(py_tensor.ptr(), 0);
                auto dense =
                    std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl());
                int numel = (*dense).numel();
                int recv_numel = numel / nranks;
                int offset = recv_numel * rank_id;
                return self.Recv_Partial(
                    *dense, src_rank, offset, recv_numel, sync_op);
              },
              py::arg("tensor"),
              py::arg("src"),
              py::arg("num"),
              py::arg("id"),
              py::arg("sync_op"),
              py::call_guard<py::gil_scoped_release>())

343 344
          .def(
              "all_gather",
345 346
              [](distributed::ProcessGroup &self,
                 py::handle py_in_tensor,
347 348 349 350 351 352 353 354 355 356 357
                 py::handle py_out_tensor) {
                auto in_tensor = CastPyArg2Tensor(py_in_tensor.ptr(), 0);
                auto out_tensor = CastPyArg2Tensor(py_out_tensor.ptr(), 0);
                auto in_dense = std::dynamic_pointer_cast<phi::DenseTensor>(
                    in_tensor.impl());
                auto out_dense = std::dynamic_pointer_cast<phi::DenseTensor>(
                    out_tensor.impl());
                std::vector<phi::DenseTensor> in_tensors = {*in_dense};
                std::vector<phi::DenseTensor> out_tensors = {*out_dense};
                return self.AllGather(in_tensors, out_tensors);
              },
358 359
              py::arg("in"),
              py::arg("out"),
360 361
              py::call_guard<py::gil_scoped_release>())

362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412
          .def(
              "allgather",
              [](distributed::ProcessGroup &self,
                 py::handle py_in_tensor,
                 py::handle py_out_tensor_list,
                 bool sync_op) {
                auto in_tensor = CastPyArg2Tensor(py_in_tensor.ptr(), 0);
                auto in_dense = std::dynamic_pointer_cast<phi::DenseTensor>(
                    in_tensor.impl());
                std::vector<phi::DenseTensor> in_wrapper = {*in_dense};

                auto out_tensor_list =
                    CastPyArg2VectorOfTensor(py_out_tensor_list.ptr(), 0);
                Tensor concat_out_tensor = paddle::concat(out_tensor_list, 0);
                auto out_dense = std::dynamic_pointer_cast<phi::DenseTensor>(
                    concat_out_tensor.impl());
                std::vector<phi::DenseTensor> out_wrapper = {*out_dense};

                const auto *dev_ctx = self.GetDeviceContext(in_tensor.place());
                auto task = self.AllGather(in_wrapper, out_wrapper, sync_op);
                distributed::SplitTensor(dev_ctx, *out_dense, &out_tensor_list);
                return task;
              },
              py::arg("in"),
              py::arg("out"),
              py::arg("sync_op"),
              py::call_guard<py::gil_scoped_release>())

          .def(
              "allgather_base",
              [](distributed::ProcessGroup &self,
                 py::handle py_in_tensor,
                 py::handle py_out_tensor,
                 bool sync_op) {
                auto in_tensor = CastPyArg2Tensor(py_in_tensor.ptr(), 0);
                auto in_dense = std::dynamic_pointer_cast<phi::DenseTensor>(
                    in_tensor.impl());
                std::vector<phi::DenseTensor> in_wrapper = {*in_dense};

                auto out_tensor = CastPyArg2Tensor(py_out_tensor.ptr(), 0);
                auto out_dense = std::dynamic_pointer_cast<phi::DenseTensor>(
                    out_tensor.impl());
                std::vector<phi::DenseTensor> out_wrapper = {*out_dense};

                return self.AllGather(in_wrapper, out_wrapper, sync_op);
              },
              py::arg("in"),
              py::arg("out"),
              py::arg("sync_op"),
              py::call_guard<py::gil_scoped_release>())

413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439
          .def(
              "all_gather_partial",
              [](distributed::ProcessGroup &self,
                 py::handle py_in_tensor,
                 py::handle py_out_tensor,
                 int nranks,
                 int rank_id) {
                auto in_tensor = CastPyArg2Tensor(py_in_tensor.ptr(), 0);
                auto out_tensor = CastPyArg2Tensor(py_out_tensor.ptr(), 0);
                auto in_dense = std::dynamic_pointer_cast<phi::DenseTensor>(
                    in_tensor.impl());
                auto out_dense = std::dynamic_pointer_cast<phi::DenseTensor>(
                    out_tensor.impl());
                std::vector<phi::DenseTensor> in_tensors = {*in_dense};
                std::vector<phi::DenseTensor> out_tensors = {*out_dense};
                int numel = (*in_dense).numel();
                int send_numel = numel / nranks;
                int offset = send_numel * rank_id;
                return self.AllGather_Partial(
                    in_tensors, out_tensors, offset, send_numel);
              },
              py::arg("in"),
              py::arg("out"),
              py::arg("num"),
              py::arg("id"),
              py::call_guard<py::gil_scoped_release>())

440 441
          .def(
              "alltoall",
442 443
              [](distributed::ProcessGroup &self,
                 py::handle py_in_tensor,
444 445 446 447 448 449 450 451 452 453 454
                 py::handle py_out_tensor) {
                auto in_tensor = CastPyArg2Tensor(py_in_tensor.ptr(), 0);
                auto out_tensor = CastPyArg2Tensor(py_out_tensor.ptr(), 0);
                auto in_dense = std::dynamic_pointer_cast<phi::DenseTensor>(
                    in_tensor.impl());
                auto out_dense = std::dynamic_pointer_cast<phi::DenseTensor>(
                    out_tensor.impl());
                std::vector<phi::DenseTensor> in_tensors = {*in_dense};
                std::vector<phi::DenseTensor> out_tensors = {*out_dense};
                return self.AllToAll(in_tensors, out_tensors);
              },
455 456
              py::arg("in"),
              py::arg("out"),
457 458
              py::call_guard<py::gil_scoped_release>())

459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482
          .def(
              "alltoall_single",
              [](distributed::ProcessGroup &self,
                 py::handle py_in_tensor,
                 py::handle py_out_tensor,
                 std::vector<int64_t> in_sizes,
                 std::vector<int64_t> out_sizes) {
                auto in_tensor = CastPyArg2Tensor(py_in_tensor.ptr(), 0);
                auto out_tensor = CastPyArg2Tensor(py_out_tensor.ptr(), 0);
                auto in_dense = std::dynamic_pointer_cast<phi::DenseTensor>(
                    in_tensor.impl());
                auto out_dense = std::dynamic_pointer_cast<phi::DenseTensor>(
                    out_tensor.impl());
                std::vector<phi::DenseTensor> in_tensors = {*in_dense};
                std::vector<phi::DenseTensor> out_tensors = {*out_dense};
                return self.AllToAll_Single(
                    in_tensors, out_tensors, in_sizes, out_sizes);
              },
              py::arg("in"),
              py::arg("out"),
              py::arg("in_sizes"),
              py::arg("out_sizes"),
              py::call_guard<py::gil_scoped_release>())

483 484
          .def(
              "reduce",
485 486 487 488
              [](distributed::ProcessGroup &self,
                 py::handle py_in_tensor,
                 int dst,
                 distributed::ReduceOp op) {
489 490 491 492 493 494 495 496 497
                auto in_tensor = CastPyArg2Tensor(py_in_tensor.ptr(), 0);
                distributed::ReduceOptions opts;
                opts.reduce_op = op;
                opts.root_rank = dst;
                auto dense = std::dynamic_pointer_cast<phi::DenseTensor>(
                    in_tensor.impl());
                std::vector<phi::DenseTensor> tensors = {*dense};
                return self.Reduce(tensors, tensors, opts);
              },
498 499
              py::arg("tensor"),
              py::arg("dst"),
500 501 502 503
              py::arg("op") = distributed::ReduceOp::SUM,
              py::call_guard<py::gil_scoped_release>())
          .def(
              "scatter",
504 505 506 507
              [](distributed::ProcessGroup &self,
                 py::handle py_in_tensor,
                 py::handle py_out_tensor,
                 int src) {
508 509 510 511 512 513 514 515 516 517 518 519
                auto in_tensor = CastPyArg2Tensor(py_in_tensor.ptr(), 0);
                auto out_tensor = CastPyArg2Tensor(py_out_tensor.ptr(), 0);
                distributed::ScatterOptions opts;
                opts.root_rank = src;
                auto in_dense = std::dynamic_pointer_cast<phi::DenseTensor>(
                    in_tensor.impl());
                auto out_dense = std::dynamic_pointer_cast<phi::DenseTensor>(
                    out_tensor.impl());
                std::vector<phi::DenseTensor> in_tensors = {*in_dense};
                std::vector<phi::DenseTensor> out_tensors = {*out_dense};
                return self.Scatter(in_tensors, out_tensors, opts);
              },
520 521 522
              py::arg("in"),
              py::arg("out"),
              py::arg("src"),
523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542
              py::call_guard<py::gil_scoped_release>())
          .def(
              "_reduce_scatter_base",
              [](distributed::ProcessGroup &self,
                 py::handle py_out_tensor,
                 py::handle py_in_tensor,
                 distributed::ReduceOp op) {
                auto in_tensor = CastPyArg2Tensor(py_in_tensor.ptr(), 0);
                auto out_tensor = CastPyArg2Tensor(py_out_tensor.ptr(), 0);
                distributed::ReduceScatterOptions opts;
                opts.reduce_op = op;
                auto dense_out = std::dynamic_pointer_cast<phi::DenseTensor>(
                    out_tensor.impl());
                auto dense_in = std::dynamic_pointer_cast<phi::DenseTensor>(
                    in_tensor.impl());
                return self._ReduceScatterBase(*dense_out, *dense_in, opts);
              },
              py::arg("out_tensor"),
              py::arg("in_tensor"),
              py::arg("op") = distributed::ReduceOp::SUM,
543
              py::call_guard<py::gil_scoped_release>());
544

545 546 547 548
  auto ProcessGroupStream =
      py::class_<distributed::ProcessGroupStream,
                 std::shared_ptr<distributed::ProcessGroupStream>>(
          *m, "ProcessGroupStream", ProcessGroup)
549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602
          .def(
              "allgather_on_calc_stream",
              [](distributed::ProcessGroupStream &self,
                 py::handle py_in_tensor,
                 py::handle py_out_tensor_list) {
                auto in_tensor = CastPyArg2Tensor(py_in_tensor.ptr(), 0);
                auto in_dense = std::dynamic_pointer_cast<phi::DenseTensor>(
                    in_tensor.impl());
                std::vector<phi::DenseTensor> in_wrapper = {*in_dense};

                auto out_tensor_list =
                    CastPyArg2VectorOfTensor(py_out_tensor_list.ptr(), 0);
                Tensor concat_out_tensor = paddle::concat(out_tensor_list, 0);
                auto out_dense = std::dynamic_pointer_cast<phi::DenseTensor>(
                    concat_out_tensor.impl());
                std::vector<phi::DenseTensor> out_wrapper = {*out_dense};

                const auto *dev_ctx =
                    self.GetDeviceContext(in_tensor.place(), true);
                auto task = self.AllGather(in_wrapper,
                                           out_wrapper,
                                           /*sync_op*/ true,
                                           /*use_calc_stream*/ true);
                distributed::SplitTensor(dev_ctx, *out_dense, &out_tensor_list);
                return task;
              },
              py::arg("in"),
              py::arg("out"),
              py::call_guard<py::gil_scoped_release>())

          .def(
              "allgather_base_on_calc_stream",
              [](distributed::ProcessGroupStream &self,
                 py::handle py_in_tensor,
                 py::handle py_out_tensor) {
                auto in_tensor = CastPyArg2Tensor(py_in_tensor.ptr(), 0);
                auto in_dense = std::dynamic_pointer_cast<phi::DenseTensor>(
                    in_tensor.impl());
                std::vector<phi::DenseTensor> in_wrapper = {*in_dense};

                auto out_tensor = CastPyArg2Tensor(py_out_tensor.ptr(), 0);
                auto out_dense = std::dynamic_pointer_cast<phi::DenseTensor>(
                    out_tensor.impl());
                std::vector<phi::DenseTensor> out_wrapper = {*out_dense};

                return self.AllGather(in_wrapper,
                                      out_wrapper,
                                      /*sync_op*/ true,
                                      /*use_calc_stream*/ true);
              },
              py::arg("in"),
              py::arg("out"),
              py::call_guard<py::gil_scoped_release>())

603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621
          .def(
              "allreduce_on_calc_stream",
              [](distributed::ProcessGroupStream &self,
                 py::handle py_tensor,
                 distributed::ReduceOp op) {
                auto tensor = CastPyArg2Tensor(py_tensor.ptr(), 0);
                distributed::AllreduceOptions opts;
                opts.reduce_op = op;
                auto dense =
                    std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl());
                std::vector<phi::DenseTensor> tensors = {*dense};
                return self.AllReduce(tensors,
                                      tensors,
                                      opts,
                                      /*sync_op*/ true,
                                      /*use_calc_stream*/ true);
              },
              py::arg("tensor"),
              py::arg("op"),
622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709
              py::call_guard<py::gil_scoped_release>())

          .def(
              "send_on_calc_stream",
              [](distributed::ProcessGroupStream &self,
                 py::handle py_tensor,
                 int dst) {
                auto tensor = CastPyArg2Tensor(py_tensor.ptr(), 0);
                auto dense =
                    std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl());
                std::vector<phi::DenseTensor> tensors = {*dense};
                return self.Send(tensors,
                                 dst,
                                 /*sync_op*/ true,
                                 /*use_calc_stream*/ true);
              },
              py::arg("tensor"),
              py::arg("dst"),
              py::call_guard<py::gil_scoped_release>())

          .def(
              "send_partial_on_calc_stream",
              [](distributed::ProcessGroupStream &self,
                 py::handle py_tensor,
                 int dst_rank,
                 int nranks,
                 int rank_id) {
                auto tensor = CastPyArg2Tensor(py_tensor.ptr(), 0);
                auto dense =
                    std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl());
                int numel = (*dense).numel();
                int send_numel = numel / nranks;
                int offset = send_numel * rank_id;
                return self.Send_Partial(*dense,
                                         dst_rank,
                                         offset,
                                         send_numel,
                                         /*sync_op*/ true,
                                         /*use_calc_stream*/ true);
              },
              py::arg("tensor"),
              py::arg("dst"),
              py::arg("num"),
              py::arg("id"),
              py::call_guard<py::gil_scoped_release>())

          .def(
              "recv_on_calc_stream",
              [](distributed::ProcessGroupStream &self,
                 py::handle py_tensor,
                 int src) {
                auto tensor = CastPyArg2Tensor(py_tensor.ptr(), 0);
                auto dense =
                    std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl());
                std::vector<phi::DenseTensor> tensors = {*dense};
                return self.Recv(tensors,
                                 src,
                                 /*sync_op*/ true,
                                 /*use_calc_stream*/ true);
              },
              py::arg("tensor"),
              py::arg("src"),
              py::call_guard<py::gil_scoped_release>())

          .def(
              "recv_partial_on_calc_stream",
              [](distributed::ProcessGroupStream &self,
                 py::handle py_tensor,
                 int src_rank,
                 int nranks,
                 int rank_id) {
                auto tensor = CastPyArg2Tensor(py_tensor.ptr(), 0);
                auto dense =
                    std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl());
                int numel = (*dense).numel();
                int recv_numel = numel / nranks;
                int offset = recv_numel * rank_id;
                return self.Recv_Partial(*dense,
                                         src_rank,
                                         offset,
                                         recv_numel,
                                         /*sync_op*/ true,
                                         /*use_calc_stream*/ true);
              },
              py::arg("tensor"),
              py::arg("src"),
              py::arg("num"),
              py::arg("id"),
710 711
              py::call_guard<py::gil_scoped_release>());

712
#if defined(PADDLE_WITH_RCCL) || defined(PADDLE_WITH_NCCL)
713 714 715
  auto processGroupNCCL =
      py::class_<distributed::ProcessGroupNCCL,
                 std::shared_ptr<distributed::ProcessGroupNCCL>>(
716
          *m, "ProcessGroupNCCL", ProcessGroupStream)
717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733
          .def(py::init<const std::shared_ptr<distributed::Store> &,
                        int,
                        int,
                        const platform::CUDAPlace &,
                        int>(),
               py::arg("store"),
               py::arg("rank"),
               py::arg("world_size"),
               py::arg("place"),
               py::arg("group_id") = 0,
               py::call_guard<py::gil_scoped_release>());

  processGroupNCCL.def_static(
      "group_start", []() { distributed::ProcessGroupNCCL::GroupStart(); });
  processGroupNCCL.def_static(
      "group_end", []() { distributed::ProcessGroupNCCL::GroupEnd(); });

734
#endif
735

W
wuhuachaocoding 已提交
736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754
#if defined(PADDLE_WITH_MPI)
  py::class_<distributed::ProcessGroupMPI,
             std::shared_ptr<distributed::ProcessGroupMPI>>(
      *m, "ProcessGroupMPI", ProcessGroup)
      .def_static(
          "create",
          [](const std::vector<int> &ranks,
             int gid) -> std::shared_ptr<distributed::ProcessGroupMPI> {
            return paddle::distributed::ProcessGroupMPI::CreateProcessGroupMPI(
                ranks, gid);
          })
      .def("get_rank",
           &distributed::ProcessGroup::GetRank,
           py::call_guard<py::gil_scoped_release>())
      .def("get_world_size",
           &distributed::ProcessGroup::GetSize,
           py::call_guard<py::gil_scoped_release>());
#endif

755 756 757 758 759
#if defined(PADDLE_WITH_GLOO) && defined(PADDLE_WITH_PSCORE) && \
    (defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_ASCEND_CL))
  py::class_<distributed::ProcessGroupHeter,
             std::shared_ptr<distributed::ProcessGroupHeter>>(
      *m, "ProcessGroupHeter", ProcessGroup)
760 761 762
      .def(py::init<const std::shared_ptr<distributed::Store> &,
                    int,
                    int,
763 764 765 766 767
#if defined(PADDLE_WITH_ASCEND_CL)
                    const platform::NPUPlace &,
#else
                    const platform::CUDAPlace &,
#endif
768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790
                    int,
                    int,
                    int,
                    int,
                    int,
                    bool,
                    std::string,
                    int,
                    int>(),
           py::arg("store"),
           py::arg("rank"),
           py::arg("world_size"),
           py::arg("place"),
           py::arg("gid") = 0,
           py::arg("local_rank") = 0,
           py::arg("local_size") = 1,
           py::arg("gloo_rank") = 0,
           py::arg("gloo_size") = 1,
           py::arg("with_switch") = false,
           py::arg("switch_endpoint") = "",
           py::arg("src_rank") = "",
           py::arg("dst_rank") = "",
           py::call_guard<py::gil_scoped_release>());
791
#endif
792

793 794 795 796
#if defined(PADDLE_WITH_ASCEND_CL)
  py::class_<distributed::ProcessGroupHCCL,
             std::shared_ptr<distributed::ProcessGroupHCCL>>(
      *m, "ProcessGroupHCCL", ProcessGroup)
797 798 799 800 801 802 803 804 805 806
      .def(py::init<const std::shared_ptr<distributed::Store> &,
                    int,
                    int,
                    const platform::NPUPlace &,
                    int>(),
           py::arg("store"),
           py::arg("rank"),
           py::arg("world_size"),
           py::arg("place"),
           py::arg("group_id") = 0,
807
           py::call_guard<py::gil_scoped_release>());
808

809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826
#endif

#if defined(PADDLE_WITH_CUSTOM_DEVICE)
  py::class_<distributed::ProcessGroupCustom,
             std::shared_ptr<distributed::ProcessGroupCustom>>(
      *m, "ProcessGroupCustom", ProcessGroup)
      .def(py::init<const std::shared_ptr<distributed::Store> &,
                    int,
                    int,
                    const platform::CustomPlace &,
                    int>(),
           py::arg("store"),
           py::arg("rank"),
           py::arg("world_size"),
           py::arg("place"),
           py::arg("group_id") = 0,
           py::call_guard<py::gil_scoped_release>());

827 828
#endif

829 830 831
  py::class_<distributed::ProcessGroup::Task,
             std::shared_ptr<distributed::ProcessGroup::Task>>(*m, "task")
      .def("is_completed", &distributed::ProcessGroup::Task::IsCompleted)
832
      .def("is_sync", &distributed::ProcessGroup::Task::IsSync)
833 834
      .def("wait",
           &distributed::ProcessGroup::Task::Wait,
835 836
           py::arg("timeout") = kWaitTimeout,
           py::call_guard<py::gil_scoped_release>())
837 838
      .def("synchronize",
           &distributed::ProcessGroup::Task::Synchronize,
839 840
           py::call_guard<py::gil_scoped_release>());

841 842 843
#if defined(PADDLE_WITH_GLOO)
  py::class_<ProcessGroupGloo, std::shared_ptr<ProcessGroupGloo>>(
      *m, "ProcessGroupGloo", ProcessGroup)
844 845 846 847 848
      .def(py::init<const std::shared_ptr<paddle::distributed::Store> &,
                    int,
                    int,
                    const platform::CPUPlace &,
                    int,
849
                    std::shared_ptr<GlooOptions> &>(),
850
           py::call_guard<py::gil_scoped_release>())
851
      .def(py::init([](const std::shared_ptr<paddle::distributed::Store> &store,
852 853 854 855
                       int rank,
                       int world_size,
                       const platform::CPUPlace &place,
                       int gid) {
856 857 858 859 860 861 862 863
             auto opts = GlooOptions::create();
             char *ifname = getenv(GLOO_SOCKET_IFNAME_ENV.c_str());
             if (ifname && strlen(ifname) > 1) {
               opts->device = ProcessGroupGloo::createDeviceForInterface(
                   std::string(ifname));
             } else {
               opts->device = ProcessGroupGloo::createDefaultDevice();
             }
864 865
             return std::make_shared<ProcessGroupGloo>(
                 store, rank, world_size, place, gid, opts);
866
           }),
867 868 869 870 871
           py::arg("store"),
           py::arg("rank"),
           py::arg("world_size"),
           py::arg("place"),
           py::arg("group_id") = 0,
872
           py::call_guard<py::gil_scoped_release>())
873 874 875 876
      .def_static("create_default_device",
                  &ProcessGroupGloo::createDefaultDevice);
#endif

877 878
  m->def(
      "eager_assign_group_by_size",
879 880
      [](py::handle py_tensors,
         std::vector<bool> is_sparse_gradient,
881 882 883 884 885 886
         std::vector<size_t> group_size_limits,
         std::vector<int64_t> tensor_indices) {
        auto tensors = CastPyArg2VectorOfTensor(py_tensors.ptr(), 0);
        return distributed::Eager_AssignGroupBySize(
            tensors, is_sparse_gradient, group_size_limits, tensor_indices);
      },
887 888
      py::arg("tensors"),
      py::arg("is_sparse_gradient"),
889 890 891
      py::arg("group_size_limits") = std::vector<size_t>{25 * 1024 * 1024},
      py::arg("tensor_indices") = std::vector<int64_t>{},
      py::call_guard<py::gil_scoped_release>());
892 893

  py::class_<distributed::EagerReducer,
894 895
             std::shared_ptr<distributed::EagerReducer>>(
      *m, "EagerReducer", R"DOC()DOC")
896
      .def(py::init(&CreateEagerReducer))
897 898 899 900 901 902
      .def(
          "prepare_for_backward",
          [](distributed::EagerReducer &self, py::handle py_tensors) {
            auto params = CastPyArg2VectorOfTensor(py_tensors.ptr(), 0);
            self.PrepareForBackward(params);
          },
903 904
          py::arg("tensors"),
          py::call_guard<py::gil_scoped_release>());
905 906 907 908
}

}  // end namespace pybind
}  // namespace paddle