distributed_py.cc 56.6 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/reducer.h"
28 29 30 31 32
#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"
33
#include "paddle/fluid/pybind/process_group_utils.h"
34 35
#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_CUSTOM_DEVICE)
#include "paddle/fluid/distributed/collective/ProcessGroupCustom.h"
#endif

48 49 50 51 52
#if defined(PADDLE_WITH_GLOO)
#include "paddle/fluid/distributed/collective/ProcessGroupGloo.h"
#include "paddle/fluid/distributed/store/tcp_store.h"
#endif

J
james 已提交
53 54 55 56
#if defined(PADDLE_WITH_XPU_BKCL)
#include "paddle/fluid/distributed/collective/ProcessGroupBKCL.h"
#endif

57 58
#include "paddle/phi/kernels/sync_batch_norm_kernel.h"

59 60 61 62 63 64 65
namespace py = pybind11;

namespace paddle {
namespace pybind {

using Tensor = paddle::experimental::Tensor;

66 67 68 69 70
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,
71 72
    const std::vector<size_t> &group_size_limits,
    bool find_unused_parameters) {
73
  auto params = CastPyArg2VectorOfTensor(py_tensors.ptr(), 0);
74 75 76 77 78 79
  return std::make_shared<distributed::EagerReducer>(params,
                                                     group_indices,
                                                     is_sparse_gradient,
                                                     process_group,
                                                     group_size_limits,
                                                     find_unused_parameters);
80 81
}

82 83 84 85 86 87 88 89
#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

90 91 92
static UNUSED void *use_ccl_comm_func =
    phi::detail::GetCCLComm(phi::CPUPlace());

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

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
          .def(
L
LiYuRio 已提交
127
              "all_reduce",
128 129 130 131 132
              [](distributed::ProcessGroup &self,
                 py::handle py_tensor,
                 distributed::ReduceOp op,
                 bool sync_op) {
                auto tensor = CastPyArg2Tensor(py_tensor.ptr(), 0);
133
                auto p_dense =
134
                    std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl());
135 136 137 138
                auto *out_dense = p_dense.get();
                auto in_dense = *p_dense;
                distributed::AllreduceOptions opts{op};
                return self.AllReduce(out_dense, in_dense, opts, sync_op);
139 140 141 142 143 144
              },
              py::arg("tensor"),
              py::arg("op"),
              py::arg("sync_op"),
              py::call_guard<py::gil_scoped_release>())

145 146 147 148 149 150 151
          .def(
              "broadcast",
              [](distributed::ProcessGroup &self,
                 py::handle py_tensor,
                 int src,
                 bool sync_op) {
                auto tensor = CastPyArg2Tensor(py_tensor.ptr(), 0);
152
                auto p_dense =
153
                    std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl());
154 155 156 157
                auto *out_dense = p_dense.get();
                auto in_dense = *p_dense;
                distributed::BroadcastOptions opts{src};
                return self.Broadcast(out_dense, in_dense, opts, sync_op);
158 159 160 161 162 163
              },
              py::arg("tensor"),
              py::arg("src"),
              py::arg("sync_op"),
              py::call_guard<py::gil_scoped_release>())

164 165 166 167 168 169 170
          .def(
              "send",
              [](distributed::ProcessGroup &self,
                 py::handle py_tensor,
                 int dst,
                 bool sync_op) {
                auto tensor = CastPyArg2Tensor(py_tensor.ptr(), 0);
171
                auto p_dense =
172
                    std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl());
173
                auto *out_dense = p_dense.get();
174 175 176
                // numel == -1 indicates sending the whole tensor
                return self.Send(
                    out_dense, dst, /*offset*/ 0, /*numel*/ -1, sync_op);
177 178 179 180 181 182 183 184 185 186 187 188 189 190 191
              },
              py::arg("tensor"),
              py::arg("dst"),
              py::arg("sync_op"),
              py::call_guard<py::gil_scoped_release>())

          .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);
192
                auto p_dense =
193
                    std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl());
194 195
                auto *out_dense = p_dense.get();

196
                int64_t numel = p_dense->numel();
197 198
                int64_t send_numel = numel / nranks;
                int64_t offset = send_numel * rank_id;
199 200

                return self.Send(
201
                    out_dense, dst_rank, offset, send_numel, sync_op);
202 203 204 205 206
              },
              py::arg("tensor"),
              py::arg("dst"),
              py::arg("num"),
              py::arg("id"),
207
              py::arg("sync_op") = true,
208 209 210 211 212 213 214 215 216
              py::call_guard<py::gil_scoped_release>())

          .def(
              "recv",
              [](distributed::ProcessGroup &self,
                 py::handle py_tensor,
                 int src,
                 bool sync_op) {
                auto tensor = CastPyArg2Tensor(py_tensor.ptr(), 0);
217
                auto p_dense =
218
                    std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl());
219
                auto *in_dense = p_dense.get();
220 221 222
                // numel == -1 indicates receiving the whole tensor
                return self.Recv(
                    in_dense, src, /*offset*/ 0, /*numel*/ -1, sync_op);
223 224 225 226 227 228 229 230 231 232 233 234 235 236 237
              },
              py::arg("tensor"),
              py::arg("src"),
              py::arg("sync_op"),
              py::call_guard<py::gil_scoped_release>())

          .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);
238
                auto p_dense =
239
                    std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl());
240 241
                auto *out_dense = p_dense.get();

242
                int64_t numel = p_dense->numel();
243 244
                int64_t recv_numel = numel / nranks;
                int64_t offset = recv_numel * rank_id;
245 246

                return self.Recv(
247
                    out_dense, src_rank, offset, recv_numel, sync_op);
248 249 250 251 252
              },
              py::arg("tensor"),
              py::arg("src"),
              py::arg("num"),
              py::arg("id"),
253
              py::arg("sync_op") = true,
254 255
              py::call_guard<py::gil_scoped_release>())

256 257
          .def(
              "all_gather",
258 259
              [](distributed::ProcessGroup &self,
                 py::handle py_out_tensor_list,
260
                 py::handle py_in_tensor,
261 262 263 264
                 bool sync_op) {
                auto out_tensor_list =
                    CastPyArg2VectorOfTensor(py_out_tensor_list.ptr(), 0);
                Tensor concat_out_tensor = paddle::concat(out_tensor_list, 0);
265
                auto p_out_tensor = std::dynamic_pointer_cast<phi::DenseTensor>(
266
                    concat_out_tensor.impl());
267 268 269 270 271 272
                auto *out_dense = p_out_tensor.get();

                auto in_tensor = CastPyArg2Tensor(py_in_tensor.ptr(), 0);
                auto p_in_tensor = std::dynamic_pointer_cast<phi::DenseTensor>(
                    in_tensor.impl());
                auto in_dense = *p_in_tensor;
273

274
                const auto &dev_ctx = self.GetDeviceContext(in_tensor.place());
275 276 277 278 279
                auto task = self.AllGather(out_dense,
                                           in_dense,
                                           /*offset*/ 0,
                                           /*numel*/ -1,
                                           sync_op);
280
                SplitTensor(dev_ctx, *out_dense, &out_tensor_list);
281
                task->UpdateWaitChain(dev_ctx);
282 283 284
                return task;
              },
              py::arg("out"),
285
              py::arg("in"),
286 287 288 289
              py::arg("sync_op"),
              py::call_guard<py::gil_scoped_release>())

          .def(
L
LiYuRio 已提交
290
              "all_gather_into_tensor",
291 292
              [](distributed::ProcessGroup &self,
                 py::handle py_out_tensor,
293
                 py::handle py_in_tensor,
294 295
                 bool sync_op) {
                auto out_tensor = CastPyArg2Tensor(py_out_tensor.ptr(), 0);
296
                auto p_out_tensor = std::dynamic_pointer_cast<phi::DenseTensor>(
297
                    out_tensor.impl());
298 299 300 301 302 303
                auto *out_dense = p_out_tensor.get();

                auto in_tensor = CastPyArg2Tensor(py_in_tensor.ptr(), 0);
                auto p_in_tensor = std::dynamic_pointer_cast<phi::DenseTensor>(
                    in_tensor.impl());
                auto in_dense = *p_in_tensor;
304

305 306 307 308 309
                return self.AllGather(out_dense,
                                      in_dense,
                                      /*offset*/ 0,
                                      /*numel*/ -1,
                                      sync_op);
310 311
              },
              py::arg("out"),
312
              py::arg("in"),
313 314 315
              py::arg("sync_op"),
              py::call_guard<py::gil_scoped_release>())

316
          .def(
L
LiYuRio 已提交
317
              "all_to_all",
318 319
              [](distributed::ProcessGroup &self,
                 py::handle py_out_tensor_list,
320
                 py::handle py_in_tensor_list,
321 322 323 324
                 bool sync_op) {
                auto out_tensor_list =
                    CastPyArg2VectorOfTensor(py_out_tensor_list.ptr(), 0);
                Tensor concat_out_tensor = paddle::concat(out_tensor_list, 0);
325
                auto p_out_tensor = std::dynamic_pointer_cast<phi::DenseTensor>(
326
                    concat_out_tensor.impl());
327 328 329 330 331 332 333 334
                auto *out_dense = p_out_tensor.get();

                auto in_tensor_list =
                    CastPyArg2VectorOfTensor(py_in_tensor_list.ptr(), 0);
                Tensor concat_in_tensor = paddle::concat(in_tensor_list, 0);
                auto p_in_tensor = std::dynamic_pointer_cast<phi::DenseTensor>(
                    concat_in_tensor.impl());
                auto in_dense = *p_in_tensor;
335 336

                // in_tensor_list should not be empty
337
                const auto &dev_ctx =
338
                    self.GetDeviceContext(in_tensor_list.back().place());
339 340 341 342 343 344 345 346
                int world_size = self.GetSize();
                auto task =
                    self.AllToAll(out_dense,
                                  in_dense,
                                  GetDefaultSplitSizes(*out_dense, world_size),
                                  GetDefaultSplitSizes(in_dense, world_size),
                                  sync_op);
                SplitTensor(dev_ctx, *out_dense, &out_tensor_list);
347
                task->UpdateWaitChain(dev_ctx);
348 349 350
                return task;
              },
              py::arg("out"),
351
              py::arg("in"),
352 353 354 355
              py::arg("sync_op"),
              py::call_guard<py::gil_scoped_release>())

          .def(
L
LiYuRio 已提交
356
              "all_to_all_tensor",
357 358
              [](distributed::ProcessGroup &self,
                 py::handle py_out_tensor,
359
                 py::handle py_in_tensor,
360 361
                 bool sync_op) {
                auto out_tensor = CastPyArg2Tensor(py_out_tensor.ptr(), 0);
362
                auto p_out_tensor = std::dynamic_pointer_cast<phi::DenseTensor>(
363
                    out_tensor.impl());
364 365 366 367 368 369
                auto *out_dense = p_out_tensor.get();

                auto in_tensor = CastPyArg2Tensor(py_in_tensor.ptr(), 0);
                auto p_in_tensor = std::dynamic_pointer_cast<phi::DenseTensor>(
                    in_tensor.impl());
                auto in_dense = *p_in_tensor;
370

371 372 373 374 375 376 377
                int world_size = self.GetSize();
                return self.AllToAll(
                    out_dense,
                    in_dense,
                    GetDefaultSplitSizes(*out_dense, world_size),
                    GetDefaultSplitSizes(in_dense, world_size),
                    sync_op);
378 379
              },
              py::arg("out"),
380
              py::arg("in"),
381 382 383
              py::arg("sync_op"),
              py::call_guard<py::gil_scoped_release>())

384
          .def(
L
LiYuRio 已提交
385
              "all_to_all_single",
386 387
              [](distributed::ProcessGroup &self,
                 py::handle py_out_tensor,
388 389 390
                 py::handle py_in_tensor,
                 const std::vector<int64_t> &out_sizes,
                 const std::vector<int64_t> &in_sizes,
391 392
                 bool sync_op) {
                auto out_tensor = CastPyArg2Tensor(py_out_tensor.ptr(), 0);
393
                auto p_out_tensor = std::dynamic_pointer_cast<phi::DenseTensor>(
394
                    out_tensor.impl());
395 396 397 398 399 400
                auto *out_dense = p_out_tensor.get();

                auto in_tensor = CastPyArg2Tensor(py_in_tensor.ptr(), 0);
                auto p_in_tensor = std::dynamic_pointer_cast<phi::DenseTensor>(
                    in_tensor.impl());
                auto in_dense = *p_in_tensor;
401

402 403
                return self.AllToAll(
                    out_dense, in_dense, out_sizes, in_sizes, sync_op);
404 405
              },
              py::arg("out"),
406
              py::arg("in"),
407
              py::arg("out_sizes"),
408
              py::arg("in_sizes"),
409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541
              py::arg("sync_op"),
              py::call_guard<py::gil_scoped_release>())

          .def(
              "reduce",
              [](distributed::ProcessGroup &self,
                 py::handle py_in_tensor,
                 int dst,
                 distributed::ReduceOp op,
                 bool sync_op) {
                auto in_tensor = CastPyArg2Tensor(py_in_tensor.ptr(), 0);
                distributed::ReduceOptions opts{op, dst};
                auto dense = std::dynamic_pointer_cast<phi::DenseTensor>(
                    in_tensor.impl());
                std::vector<phi::DenseTensor> tensors = {*dense};
                return self.Reduce(tensors, tensors, opts, sync_op);
              },
              py::arg("tensor"),
              py::arg("dst"),
              py::arg("op"),
              py::arg("sync_op"),
              py::call_guard<py::gil_scoped_release>())

          .def(
              "reduce_scatter",
              [](distributed::ProcessGroup &self,
                 py::handle py_in_tensor_list,
                 py::handle py_out_tensor,
                 distributed::ReduceOp op,
                 bool sync_op) {
                auto in_tensor_list =
                    CastPyArg2VectorOfTensor(py_in_tensor_list.ptr(), 0);
                Tensor concat_in_tensor = paddle::concat(in_tensor_list, 0);
                auto in_dense = std::dynamic_pointer_cast<phi::DenseTensor>(
                    concat_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};

                distributed::ReduceScatterOptions opts{op};
                return self.ReduceScatter(
                    in_wrapper, out_wrapper, opts, sync_op);
              },
              py::arg("in"),
              py::arg("out"),
              py::arg("op"),
              py::arg("sync_op"),
              py::call_guard<py::gil_scoped_release>())

          .def(
              "reduce_scatter_tensor",
              [](distributed::ProcessGroup &self,
                 py::handle py_in_tensor,
                 py::handle py_out_tensor,
                 distributed::ReduceOp op,
                 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};

                distributed::ReduceScatterOptions opts{op};
                return self.ReduceScatter(
                    in_wrapper, out_wrapper, opts, sync_op);
              },
              py::arg("in"),
              py::arg("out"),
              py::arg("op"),
              py::arg("sync_op"),
              py::call_guard<py::gil_scoped_release>())

          .def(
              "scatter",
              [](distributed::ProcessGroup &self,
                 py::handle py_in_tensor_list,
                 py::handle py_out_tensor,
                 int src,
                 bool sync_op) {
                auto in_tensor_list =
                    CastPyArg2VectorOfTensor(py_in_tensor_list.ptr(), 0);
                Tensor concat_in_tensor = paddle::concat(in_tensor_list, 0);
                auto in_dense = std::dynamic_pointer_cast<phi::DenseTensor>(
                    concat_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};

                distributed::ScatterOptions opts{src};
                return self.Scatter(in_wrapper, out_wrapper, opts, sync_op);
              },
              py::arg("in"),
              py::arg("out"),
              py::arg("src"),
              py::arg("sync_op"),
              py::call_guard<py::gil_scoped_release>())

          .def(
              "scatter_tensor",
              [](distributed::ProcessGroup &self,
                 py::handle py_in_tensor,
                 py::handle py_out_tensor,
                 int src,
                 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};

                distributed::ScatterOptions opts{src};
                return self.Scatter(in_wrapper, out_wrapper, opts, sync_op);
              },
              py::arg("in"),
              py::arg("out"),
              py::arg("src"),
              py::arg("sync_op"),
              py::call_guard<py::gil_scoped_release>())

L
LiYuRio 已提交
542 543
          .def(
              "barrier",
544
              [](distributed::ProcessGroup &self, int8_t device_id) {
L
LiYuRio 已提交
545
                distributed::BarrierOptions opts;
546
                opts.device_id = device_id;
L
LiYuRio 已提交
547 548
                return self.Barrier(opts);
              },
549
              py::arg("device_id") = -1,
L
LiYuRio 已提交
550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639
              py::call_guard<py::gil_scoped_release>())

          // TODO(liyurui): Interface below will be removed in the future.
          .def(
              "allreduce",
              [](distributed::ProcessGroup &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);
              },
              py::arg("tensor"),
              py::arg("op") = distributed::ReduceOp::SUM,
              py::call_guard<py::gil_scoped_release>())

          .def(
              "broadcast",
              [](distributed::ProcessGroup &self,
                 py::handle py_tensor,
                 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);
              },
              py::arg("tensor"),
              py::arg("source_rank"),
              py::call_guard<py::gil_scoped_release>())

          .def(
              "send",
              [](distributed::ProcessGroup &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);
              },
              py::arg("tensor"),
              py::arg("dst"),
              py::call_guard<py::gil_scoped_release>())

          .def(
              "recv",
              [](distributed::ProcessGroup &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);
              },
              py::arg("tensor"),
              py::arg("src"),
              py::call_guard<py::gil_scoped_release>())

          .def(
              "all_gather",
              [](distributed::ProcessGroup &self,
                 py::handle py_in_tensor,
                 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);
              },
              py::arg("in"),
              py::arg("out"),
              py::call_guard<py::gil_scoped_release>())

          .def(
              "all_gather_partial",
              [](distributed::ProcessGroup &self,
                 py::handle py_out_tensor,
640
                 py::handle py_in_tensor,
L
LiYuRio 已提交
641 642 643
                 int nranks,
                 int rank_id) {
                auto out_tensor = CastPyArg2Tensor(py_out_tensor.ptr(), 0);
644
                auto p_out_tensor = std::dynamic_pointer_cast<phi::DenseTensor>(
L
LiYuRio 已提交
645
                    out_tensor.impl());
646 647 648 649 650 651 652 653
                auto *out_dense = p_out_tensor.get();

                auto in_tensor = CastPyArg2Tensor(py_in_tensor.ptr(), 0);
                auto p_in_tensor = std::dynamic_pointer_cast<phi::DenseTensor>(
                    in_tensor.impl());
                auto in_dense = *p_in_tensor;

                int64_t numel = in_dense.numel();
L
LiYuRio 已提交
654 655
                int64_t send_numel = numel / nranks;
                int64_t offset = send_numel * rank_id;
656 657
                return self.AllGather(
                    out_dense, in_dense, offset, send_numel, /*sync_op*/ true);
L
LiYuRio 已提交
658 659
              },
              py::arg("out"),
660
              py::arg("in"),
L
LiYuRio 已提交
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
              py::arg("num"),
              py::arg("id"),
              py::call_guard<py::gil_scoped_release>())

          .def(
              "alltoall",
              [](distributed::ProcessGroup &self,
                 py::handle py_in_tensor,
                 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);
              },
              py::arg("in"),
              py::arg("out"),
              py::call_guard<py::gil_scoped_release>())

          .def(
              "alltoall_single",
              [](distributed::ProcessGroup &self,
                 py::handle py_in_tensor,
                 py::handle py_out_tensor,
689 690
                 const std::vector<int64_t> in_sizes,
                 const std::vector<int64_t> out_sizes) {
L
LiYuRio 已提交
691
                auto out_tensor = CastPyArg2Tensor(py_out_tensor.ptr(), 0);
692
                auto p_out_tensor = std::dynamic_pointer_cast<phi::DenseTensor>(
L
LiYuRio 已提交
693
                    out_tensor.impl());
694 695 696 697 698 699 700 701 702
                auto *out_dense = p_out_tensor.get();

                auto in_tensor = CastPyArg2Tensor(py_in_tensor.ptr(), 0);
                auto p_in_tensor = std::dynamic_pointer_cast<phi::DenseTensor>(
                    in_tensor.impl());
                auto in_dense = *p_in_tensor;

                return self.AllToAll(
                    out_dense, in_dense, out_sizes, in_sizes, /*sync_op*/ true);
L
LiYuRio 已提交
703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750
              },
              py::arg("in"),
              py::arg("out"),
              py::arg("in_sizes"),
              py::arg("out_sizes"),
              py::call_guard<py::gil_scoped_release>())

          .def(
              "reduce",
              [](distributed::ProcessGroup &self,
                 py::handle py_in_tensor,
                 int dst,
                 distributed::ReduceOp op) {
                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);
              },
              py::arg("tensor"),
              py::arg("dst"),
              py::arg("op") = distributed::ReduceOp::SUM,
              py::call_guard<py::gil_scoped_release>())

          .def(
              "scatter",
              [](distributed::ProcessGroup &self,
                 py::handle py_in_tensor,
                 py::handle py_out_tensor,
                 int src) {
                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);
              },
              py::arg("in"),
              py::arg("out"),
              py::arg("src"),
751
              py::call_guard<py::gil_scoped_release>());
752

753 754 755 756
  auto ProcessGroupStream =
      py::class_<distributed::ProcessGroupStream,
                 std::shared_ptr<distributed::ProcessGroupStream>>(
          *m, "ProcessGroupStream", ProcessGroup)
757
          .def(
L
LiYuRio 已提交
758
              "all_gather_on_calc_stream",
759
              [](distributed::ProcessGroupStream &self,
760 761
                 py::handle py_out_tensor_list,
                 py::handle py_in_tensor) {
762 763 764
                auto out_tensor_list =
                    CastPyArg2VectorOfTensor(py_out_tensor_list.ptr(), 0);
                Tensor concat_out_tensor = paddle::concat(out_tensor_list, 0);
765
                auto p_out_tensor = std::dynamic_pointer_cast<phi::DenseTensor>(
766
                    concat_out_tensor.impl());
767 768 769 770 771 772
                auto *out_dense = p_out_tensor.get();

                auto in_tensor = CastPyArg2Tensor(py_in_tensor.ptr(), 0);
                auto p_in_tensor = std::dynamic_pointer_cast<phi::DenseTensor>(
                    in_tensor.impl());
                auto in_dense = *p_in_tensor;
773

774
                const auto &dev_ctx =
775
                    self.GetDeviceContext(in_tensor.place(), true);
776 777
                auto task = self.AllGather(out_dense,
                                           in_dense,
778 779
                                           /*offset*/ 0,
                                           /*numel*/ -1,
780 781
                                           /*sync_op*/ true,
                                           /*use_calc_stream*/ true);
782
                SplitTensor(dev_ctx, *out_dense, &out_tensor_list);
783 784 785
                return task;
              },
              py::arg("out"),
786
              py::arg("in"),
787 788 789
              py::call_guard<py::gil_scoped_release>())

          .def(
L
LiYuRio 已提交
790
              "all_gather_into_tensor_on_calc_stream",
791
              [](distributed::ProcessGroupStream &self,
792 793
                 py::handle py_out_tensor,
                 py::handle py_in_tensor) {
794
                auto out_tensor = CastPyArg2Tensor(py_out_tensor.ptr(), 0);
795
                auto p_out_tensor = std::dynamic_pointer_cast<phi::DenseTensor>(
796
                    out_tensor.impl());
797
                auto *out_dense = p_out_tensor.get();
798

799 800 801 802 803 804 805
                auto in_tensor = CastPyArg2Tensor(py_in_tensor.ptr(), 0);
                auto p_in_tensor = std::dynamic_pointer_cast<phi::DenseTensor>(
                    in_tensor.impl());
                auto in_dense = *p_in_tensor;

                return self.AllGather(out_dense,
                                      in_dense,
806 807
                                      /*offset*/ 0,
                                      /*numel*/ -1,
808 809 810 811
                                      /*sync_op*/ true,
                                      /*use_calc_stream*/ true);
              },
              py::arg("out"),
812
              py::arg("in"),
813 814
              py::call_guard<py::gil_scoped_release>())

815 816 817 818
          .def(
              "all_gather_partial_on_calc_stream",
              [](distributed::ProcessGroupStream &self,
                 py::handle py_out_tensor,
819
                 py::handle py_in_tensor,
820 821 822
                 int nranks,
                 int rank_id) {
                auto out_tensor = CastPyArg2Tensor(py_out_tensor.ptr(), 0);
823
                auto p_out_tensor = std::dynamic_pointer_cast<phi::DenseTensor>(
824
                    out_tensor.impl());
825 826 827 828 829 830 831 832
                auto *out_dense = p_out_tensor.get();

                auto in_tensor = CastPyArg2Tensor(py_in_tensor.ptr(), 0);
                auto p_in_tensor = std::dynamic_pointer_cast<phi::DenseTensor>(
                    in_tensor.impl());
                auto in_dense = *p_in_tensor;

                int64_t numel = in_dense.numel();
833 834
                int64_t send_numel = numel / nranks;
                int64_t offset = send_numel * rank_id;
835 836 837 838 839 840 841

                return self.AllGather(out_dense,
                                      in_dense,
                                      offset,
                                      send_numel,
                                      /*sync_op*/ true,
                                      /*use_calc_stream*/ true);
842 843
              },
              py::arg("out"),
844
              py::arg("in"),
845 846 847 848
              py::arg("num"),
              py::arg("id"),
              py::call_guard<py::gil_scoped_release>())

849
          .def(
L
LiYuRio 已提交
850
              "all_reduce_on_calc_stream",
851 852 853 854
              [](distributed::ProcessGroupStream &self,
                 py::handle py_tensor,
                 distributed::ReduceOp op) {
                auto tensor = CastPyArg2Tensor(py_tensor.ptr(), 0);
855
                auto p_dense =
856
                    std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl());
857 858 859 860 861
                auto in_dense = *p_dense;
                auto *out_dense = p_dense.get();
                distributed::AllreduceOptions opts{op};
                return self.AllReduce(out_dense,
                                      in_dense,
862 863 864 865 866
                                      opts,
                                      /*sync_op*/ true,
                                      /*use_calc_stream*/ true);
              },
              py::arg("tensor"),
L
LiYuRio 已提交
867
              py::arg("op") = distributed::ReduceOp::SUM,
868 869
              py::call_guard<py::gil_scoped_release>())

870
          .def(
L
LiYuRio 已提交
871
              "all_to_all_on_calc_stream",
872
              [](distributed::ProcessGroupStream &self,
873 874
                 py::handle py_out_tensor_list,
                 py::handle py_in_tensor_list) {
875 876 877
                auto out_tensor_list =
                    CastPyArg2VectorOfTensor(py_out_tensor_list.ptr(), 0);
                Tensor concat_out_tensor = paddle::concat(out_tensor_list, 0);
878
                auto p_out_tensor = std::dynamic_pointer_cast<phi::DenseTensor>(
879
                    concat_out_tensor.impl());
880 881 882 883 884 885 886 887
                auto *out_dense = p_out_tensor.get();

                auto in_tensor_list =
                    CastPyArg2VectorOfTensor(py_in_tensor_list.ptr(), 0);
                Tensor concat_in_tensor = paddle::concat(in_tensor_list, 0);
                auto p_in_tensor = std::dynamic_pointer_cast<phi::DenseTensor>(
                    concat_in_tensor.impl());
                auto in_dense = *p_in_tensor;
888

889
                // in_tensor_list should not be empty
890
                const auto &dev_ctx = self.GetDeviceContext(
891
                    in_tensor_list.back().place(), /*use_calc_stream*/ true);
892 893 894 895 896 897 898 899 900
                int world_size = self.GetSize();
                auto task =
                    self.AllToAll(out_dense,
                                  in_dense,
                                  GetDefaultSplitSizes(*out_dense, world_size),
                                  GetDefaultSplitSizes(in_dense, world_size),
                                  /*sync_op*/ true,
                                  /*use_calc_stream*/ true);
                SplitTensor(dev_ctx, *out_dense, &out_tensor_list);
901 902 903
                return task;
              },
              py::arg("out"),
904
              py::arg("in"),
905 906 907
              py::call_guard<py::gil_scoped_release>())

          .def(
L
LiYuRio 已提交
908
              "all_to_all_tensor_on_calc_stream",
909
              [](distributed::ProcessGroupStream &self,
910 911
                 py::handle py_out_tensor,
                 py::handle py_in_tensor) {
912
                auto out_tensor = CastPyArg2Tensor(py_out_tensor.ptr(), 0);
913
                auto p_out_tensor = std::dynamic_pointer_cast<phi::DenseTensor>(
914
                    out_tensor.impl());
915
                auto *out_dense = p_out_tensor.get();
916

917 918 919 920 921 922 923 924 925 926 927 928 929
                auto in_tensor = CastPyArg2Tensor(py_in_tensor.ptr(), 0);
                auto p_in_tensor = std::dynamic_pointer_cast<phi::DenseTensor>(
                    in_tensor.impl());
                auto in_dense = *p_in_tensor;

                int world_size = self.GetSize();
                return self.AllToAll(
                    out_dense,
                    in_dense,
                    GetDefaultSplitSizes(*out_dense, world_size),
                    GetDefaultSplitSizes(in_dense, world_size),
                    /*sync_op*/ true,
                    /*use_calc_stream*/ true);
930 931
              },
              py::arg("out"),
932
              py::arg("in"),
933 934 935
              py::call_guard<py::gil_scoped_release>())

          .def(
L
LiYuRio 已提交
936
              "all_to_all_single_on_calc_stream",
937 938
              [](distributed::ProcessGroupStream &self,
                 py::handle py_out_tensor,
939 940 941
                 py::handle py_in_tensor,
                 const std::vector<int64_t> &out_sizes,
                 const std::vector<int64_t> &in_sizes) {
942
                auto out_tensor = CastPyArg2Tensor(py_out_tensor.ptr(), 0);
943
                auto p_out_tensor = std::dynamic_pointer_cast<phi::DenseTensor>(
944
                    out_tensor.impl());
945
                auto *out_dense = p_out_tensor.get();
946

947 948 949 950 951 952 953 954 955 956 957
                auto in_tensor = CastPyArg2Tensor(py_in_tensor.ptr(), 0);
                auto p_in_tensor = std::dynamic_pointer_cast<phi::DenseTensor>(
                    in_tensor.impl());
                auto in_dense = *p_in_tensor;

                return self.AllToAll(out_dense,
                                     in_dense,
                                     out_sizes,
                                     in_sizes,
                                     /*sync_op*/ true,
                                     /*use_calc_stream*/ true);
958 959
              },
              py::arg("out"),
960
              py::arg("in"),
961
              py::arg("out_sizes"),
962
              py::arg("in_sizes"),
963 964 965 966 967 968 969 970
              py::call_guard<py::gil_scoped_release>())

          .def(
              "broadcast_on_calc_stream",
              [](distributed::ProcessGroupStream &self,
                 py::handle py_tensor,
                 int src) {
                auto tensor = CastPyArg2Tensor(py_tensor.ptr(), 0);
971
                auto p_dense =
972
                    std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl());
973 974 975 976 977
                auto *out_dense = p_dense.get();
                auto in_dense = *p_dense;
                distributed::BroadcastOptions opts{src};
                return self.Broadcast(out_dense,
                                      in_dense,
978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 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
                                      opts,
                                      /*sync_op*/ true,
                                      /*use_calc_stream*/ true);
              },
              py::arg("tensor"),
              py::arg("src"),
              py::call_guard<py::gil_scoped_release>())

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

          .def(
              "reduce_scatter_on_calc_stream",
              [](distributed::ProcessGroupStream &self,
                 py::handle py_in_tensor_list,
                 py::handle py_out_tensor,
                 distributed::ReduceOp op) {
                auto in_tensor_list =
                    CastPyArg2VectorOfTensor(py_in_tensor_list.ptr(), 0);
                Tensor concat_in_tensor = paddle::concat(in_tensor_list, 0);
                auto in_dense = std::dynamic_pointer_cast<phi::DenseTensor>(
                    concat_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};

                distributed::ReduceScatterOptions opts{op};
                return self.ReduceScatter(in_wrapper,
                                          out_wrapper,
                                          opts,
                                          /*sync_op*/ true,
                                          /*use_calc_stream*/ true);
              },
              py::arg("in"),
              py::arg("out"),
              py::arg("op"),
              py::call_guard<py::gil_scoped_release>())

          .def(
              "reduce_scatter_tensor_on_calc_stream",
              [](distributed::ProcessGroupStream &self,
                 py::handle py_in_tensor,
                 py::handle py_out_tensor,
                 distributed::ReduceOp op) {
                auto in_tensor = CastPyArg2Tensor(py_in_tensor.ptr(), 0);
1045 1046
                auto in_dense = std::dynamic_pointer_cast<phi::DenseTensor>(
                    in_tensor.impl());
1047 1048 1049
                std::vector<phi::DenseTensor> in_wrapper = {*in_dense};

                auto out_tensor = CastPyArg2Tensor(py_out_tensor.ptr(), 0);
1050 1051
                auto out_dense = std::dynamic_pointer_cast<phi::DenseTensor>(
                    out_tensor.impl());
1052 1053 1054 1055 1056 1057 1058 1059
                std::vector<phi::DenseTensor> out_wrapper = {*out_dense};

                distributed::ReduceScatterOptions opts{op};
                return self.ReduceScatter(in_wrapper,
                                          out_wrapper,
                                          opts,
                                          /*sync_op*/ true,
                                          /*use_calc_stream*/ true);
1060 1061 1062
              },
              py::arg("in"),
              py::arg("out"),
1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121
              py::arg("op"),
              py::call_guard<py::gil_scoped_release>())

          .def(
              "scatter_on_calc_stream",
              [](distributed::ProcessGroupStream &self,
                 py::handle py_in_tensor_list,
                 py::handle py_out_tensor,
                 int src) {
                auto in_tensor_list =
                    CastPyArg2VectorOfTensor(py_in_tensor_list.ptr(), 0);
                Tensor concat_in_tensor = paddle::concat(in_tensor_list, 0);
                auto in_dense = std::dynamic_pointer_cast<phi::DenseTensor>(
                    concat_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};

                distributed::ScatterOptions opts{src};
                return self.Scatter(in_wrapper,
                                    out_wrapper,
                                    opts,
                                    /*sync_op*/ true,
                                    /*use_calc_stream*/ true);
              },
              py::arg("in"),
              py::arg("out"),
              py::arg("src"),
              py::call_guard<py::gil_scoped_release>())

          .def(
              "scatter_tensor_on_calc_stream",
              [](distributed::ProcessGroupStream &self,
                 py::handle py_in_tensor,
                 py::handle py_out_tensor,
                 int src) {
                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};

                distributed::ScatterOptions opts{src};
                return self.Scatter(in_wrapper,
                                    out_wrapper,
                                    opts,
                                    /*sync_op*/ true,
                                    /*use_calc_stream*/ true);
              },
              py::arg("in"),
              py::arg("out"),
              py::arg("src"),
1122 1123
              py::call_guard<py::gil_scoped_release>())

1124 1125 1126 1127 1128 1129
          .def(
              "send_on_calc_stream",
              [](distributed::ProcessGroupStream &self,
                 py::handle py_tensor,
                 int dst) {
                auto tensor = CastPyArg2Tensor(py_tensor.ptr(), 0);
1130
                auto p_dense =
1131
                    std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl());
1132
                auto *out_dense = p_dense.get();
1133
                // numel == -1 indicates sending the whole tensor
1134
                return self.Send(out_dense,
1135
                                 dst,
1136 1137
                                 /*offset*/ 0,
                                 /*numel*/ -1,
1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152
                                 /*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);
1153
                auto p_dense =
1154
                    std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl());
1155 1156
                auto *out_dense = p_dense.get();

1157
                int64_t numel = p_dense->numel();
1158 1159
                int64_t send_numel = numel / nranks;
                int64_t offset = send_numel * rank_id;
1160 1161 1162 1163 1164 1165 1166

                return self.Send(out_dense,
                                 dst_rank,
                                 offset,
                                 send_numel,
                                 /*sync_op*/ true,
                                 /*use_calc_stream*/ true);
1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179
              },
              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);
1180
                auto p_dense =
1181
                    std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl());
1182
                auto *in_dense = p_dense.get();
1183
                // numel == -1 indicates receiving the whole tensor
1184
                return self.Recv(in_dense,
1185
                                 src,
1186 1187
                                 /*offset*/ 0,
                                 /*numel*/ -1,
1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202
                                 /*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);
1203
                auto p_dense =
1204
                    std::dynamic_pointer_cast<phi::DenseTensor>(tensor.impl());
1205 1206
                auto *out_dense = p_dense.get();

1207
                int64_t numel = p_dense->numel();
1208 1209
                int64_t recv_numel = numel / nranks;
                int64_t offset = recv_numel * rank_id;
1210 1211 1212 1213 1214 1215 1216

                return self.Recv(out_dense,
                                 src_rank,
                                 offset,
                                 recv_numel,
                                 /*sync_op*/ true,
                                 /*use_calc_stream*/ true);
1217 1218 1219 1220 1221
              },
              py::arg("tensor"),
              py::arg("src"),
              py::arg("num"),
              py::arg("id"),
1222 1223
              py::call_guard<py::gil_scoped_release>());

1224
#if defined(PADDLE_WITH_RCCL) || defined(PADDLE_WITH_NCCL)
1225 1226 1227
  auto processGroupNCCL =
      py::class_<distributed::ProcessGroupNCCL,
                 std::shared_ptr<distributed::ProcessGroupNCCL>>(
1228
          *m, "ProcessGroupNCCL", ProcessGroupStream)
1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243
          .def(py::init<const std::shared_ptr<distributed::Store> &,
                        int,
                        int,
                        int>(),
               py::arg("store"),
               py::arg("rank"),
               py::arg("world_size"),
               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(); });

1244
#endif
1245

W
wuhuachaocoding 已提交
1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264
#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

1265 1266 1267 1268 1269
#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> &,
1270
                    const std::string &,
1271 1272 1273 1274
                    int,
                    int,
                    int>(),
           py::arg("store"),
1275
           py::arg("device_type"),
1276 1277 1278 1279 1280
           py::arg("rank"),
           py::arg("world_size"),
           py::arg("group_id") = 0,
           py::call_guard<py::gil_scoped_release>());

1281 1282
#endif

J
james 已提交
1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298
#if defined(PADDLE_WITH_XPU_BKCL)
  auto processGroupBKCL =
      py::class_<distributed::ProcessGroupBKCL,
                 std::shared_ptr<distributed::ProcessGroupBKCL>>(
          *m, "ProcessGroupBKCL", ProcessGroup)
          .def(py::init<const std::shared_ptr<distributed::Store> &,
                        int,
                        int,
                        int>(),
               py::arg("store"),
               py::arg("rank"),
               py::arg("world_size"),
               py::arg("group_id") = 0,
               py::call_guard<py::gil_scoped_release>());
#endif

1299 1300 1301
  py::class_<distributed::ProcessGroup::Task,
             std::shared_ptr<distributed::ProcessGroup::Task>>(*m, "task")
      .def("is_completed", &distributed::ProcessGroup::Task::IsCompleted)
1302
      .def("is_sync", &distributed::ProcessGroup::Task::IsSync)
1303 1304
      .def("wait",
           &distributed::ProcessGroup::Task::Wait,
1305 1306
           py::arg("timeout") = kWaitTimeout,
           py::call_guard<py::gil_scoped_release>())
1307 1308
      .def("synchronize",
           &distributed::ProcessGroup::Task::Synchronize,
1309 1310
           py::call_guard<py::gil_scoped_release>());

1311 1312 1313
#if defined(PADDLE_WITH_GLOO)
  py::class_<ProcessGroupGloo, std::shared_ptr<ProcessGroupGloo>>(
      *m, "ProcessGroupGloo", ProcessGroup)
1314 1315 1316 1317
      .def(py::init<const std::shared_ptr<paddle::distributed::Store> &,
                    int,
                    int,
                    int,
1318
                    std::shared_ptr<GlooOptions> &>(),
1319
           py::call_guard<py::gil_scoped_release>())
1320
      .def(py::init([](const std::shared_ptr<paddle::distributed::Store> &store,
1321 1322 1323
                       int rank,
                       int world_size,
                       int gid) {
1324 1325 1326 1327 1328 1329 1330 1331
             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();
             }
1332
             return std::make_shared<ProcessGroupGloo>(
1333
                 store, rank, world_size, gid, opts);
1334
           }),
1335 1336 1337 1338
           py::arg("store"),
           py::arg("rank"),
           py::arg("world_size"),
           py::arg("group_id") = 0,
1339
           py::call_guard<py::gil_scoped_release>())
1340 1341 1342 1343
      .def_static("create_default_device",
                  &ProcessGroupGloo::createDefaultDevice);
#endif

1344 1345
  m->def(
      "eager_assign_group_by_size",
1346 1347
      [](py::handle py_tensors,
         std::vector<bool> is_sparse_gradient,
1348 1349 1350 1351 1352 1353
         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);
      },
1354 1355
      py::arg("tensors"),
      py::arg("is_sparse_gradient"),
1356 1357 1358
      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>());
1359 1360

  py::class_<distributed::EagerReducer,
1361 1362
             std::shared_ptr<distributed::EagerReducer>>(
      *m, "EagerReducer", R"DOC()DOC")
1363
      .def(py::init(&CreateEagerReducer))
1364 1365
      .def(
          "prepare_for_backward",
1366
          [](distributed::EagerReducer &self, py::handle py_tensors) {
1367
            auto params = CastPyArg2VectorOfTensor(py_tensors.ptr(), 0);
1368
            self.PrepareForBackward(params);
1369
          },
1370 1371
          py::arg("tensors"),
          py::call_guard<py::gil_scoped_release>());
1372 1373 1374 1375
}

}  // end namespace pybind
}  // namespace paddle