forked from pytorch/pytorch
-
Notifications
You must be signed in to change notification settings - Fork 9
/
init.cpp
1939 lines (1767 loc) · 76.3 KB
/
init.cpp
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
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
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
542
543
544
545
546
547
548
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
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
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
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
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
#include <torch/csrc/python_headers.h>
#include <c10/util/intrusive_ptr.h>
#include <torch/csrc/distributed/c10d/FileStore.hpp>
#include <torch/csrc/distributed/c10d/TCPStore.hpp>
#include <torch/csrc/distributed/c10d/Utils.hpp>
#ifndef _WIN32
#include <torch/csrc/distributed/c10d/HashStore.hpp>
#include <torch/csrc/distributed/c10d/ProcessGroupRoundRobin.hpp>
#endif
#include <torch/csrc/distributed/c10d/ProcessGroup.hpp>
#include <torch/csrc/distributed/c10d/PyProcessGroup.hpp>
#ifdef USE_C10D_GLOO
#include <torch/csrc/distributed/c10d/ProcessGroupGloo.hpp>
#include <torch/csrc/distributed/c10d/ProcessGroupWrapper.hpp>
#endif
#ifdef USE_C10D_NCCL
#include <torch/csrc/distributed/c10d/NCCLUtils.hpp>
#include <torch/csrc/distributed/c10d/ProcessGroupNCCL.hpp>
#endif
#ifdef USE_C10D_MPI
#include <torch/csrc/distributed/c10d/ProcessGroupMPI.hpp>
#endif
#ifdef USE_C10D_UCC
#include <torch/csrc/distributed/c10d/ProcessGroupUCC.hpp>
#endif
#include <fmt/format.h>
#include <pybind11/chrono.h>
#include <torch/csrc/distributed/c10d/PrefixStore.hpp>
#include <torch/csrc/distributed/c10d/comm.hpp>
#include <torch/csrc/distributed/c10d/debug.h>
#include <torch/csrc/distributed/c10d/logger.hpp>
#include <torch/csrc/distributed/c10d/reducer.hpp>
#include <torch/csrc/Exceptions.h>
#include <torch/csrc/distributed/c10d/Ops.hpp>
#include <torch/csrc/distributed/c10d/python_comm_hook.h>
#include <torch/csrc/jit/python/pybind_utils.h>
#include <torch/csrc/utils/object_ptr.h>
#include <torch/csrc/utils/pybind.h>
#include <torch/custom_class.h>
namespace {
// Wrapper to ensure GIL is released before destructing ProcessGroupGloo
// TODO: move this somewhere more generally useful
template <typename T>
class IntrusivePtrNoGilDestructor {
c10::intrusive_ptr<T> impl_;
public:
IntrusivePtrNoGilDestructor() = default;
IntrusivePtrNoGilDestructor(const IntrusivePtrNoGilDestructor&) = default;
IntrusivePtrNoGilDestructor(IntrusivePtrNoGilDestructor&&) = default;
IntrusivePtrNoGilDestructor& operator=(const IntrusivePtrNoGilDestructor&) =
default;
IntrusivePtrNoGilDestructor& operator=(IntrusivePtrNoGilDestructor&&) =
default;
/* implicit */ IntrusivePtrNoGilDestructor(c10::intrusive_ptr<T> impl)
: impl_(std::move(impl)) {}
// This ctor is very important; see
// https://github.com/pybind/pybind11/issues/2957
explicit IntrusivePtrNoGilDestructor(T* impl)
: impl_(c10::intrusive_ptr<T>::unsafe_steal_from_new(impl)) {}
~IntrusivePtrNoGilDestructor() {
if (impl_) {
if (PyGILState_Check()) {
pybind11::gil_scoped_release release;
impl_.reset();
} else {
impl_.reset();
}
}
}
T& operator*() const noexcept {
return *impl_;
}
T* operator->() const noexcept {
return impl_.get();
}
C10_NODISCARD T* get() const noexcept {
return impl_.get();
}
void reset() noexcept {
impl_.reset();
}
operator bool() const noexcept {
return impl_;
}
};
} // anonymous namespace
PYBIND11_DECLARE_HOLDER_TYPE(T, IntrusivePtrNoGilDestructor<T>, true);
namespace torch {
namespace distributed {
namespace c10d {
namespace {
std::vector<std::string> split(char separator, const std::string& string) {
std::vector<std::string> pieces;
std::stringstream ss(string);
std::string item;
while (std::getline(ss, item, separator)) {
pieces.push_back(std::move(item));
}
return pieces;
}
template <typename T>
using shared_ptr_class_ = py::class_<T, std::shared_ptr<T>>;
constexpr auto kDeprecationWarning =
"{} API is being deprecated, please ping "
"https://github.com/pytorch/pytorch/issues/46291 "
"if you see this warning";
template <typename T>
using intrusive_ptr_class_ = py::class_<T, c10::intrusive_ptr<T>>;
template <typename T>
using intrusive_ptr_no_gil_destructor_class_ =
py::class_<T, IntrusivePtrNoGilDestructor<T>>;
// PythonStore is a pybind11 trampoline class to allow a Python
// class to inherit from c10d.Store and implement its interface.
class PythonStore : public ::c10d::Store {
public:
using ::c10d::Store::Store;
// Note: this function manually calls the Python-side overload
// for this function instead of using the PYBIND11_OVERLOAD_XYZ
// macros. This is done so that we can call the Python-side
// function with a std::string instead of a std::vector<uint8_t>.
void set(const std::string& key, const std::vector<uint8_t>& value) override {
pybind11::gil_scoped_acquire gil;
pybind11::function fn =
pybind11::get_overload(static_cast<const ::c10d::Store*>(this), "set");
TORCH_INTERNAL_ASSERT(fn);
// Call function with a py::bytes object for the value.
fn(key,
py::bytes(reinterpret_cast<const char*>(value.data()), value.size()));
}
// Note: this function manually calls the Python-side overload
// for this function instead of using the PYBIND11_OVERLOAD_XYZ
// macros. This is done so that the Python-side function can
// return a py::bytes instead of a std::vector<uint8_t>.
std::vector<uint8_t> get(const std::string& key) override {
pybind11::gil_scoped_acquire gil;
pybind11::function fn =
pybind11::get_overload(static_cast<const ::c10d::Store*>(this), "get");
TORCH_INTERNAL_ASSERT(fn);
// Cast return value from Python to py::bytes, then implicitly
// convert that to a std::string, so that we can construct a
// std::vector<uint8_t>. There is no API for directly accessing
// the contents of the py::bytes object.
std::string str = pybind11::cast<py::bytes>(fn(key));
return std::vector<uint8_t>(str.begin(), str.end());
}
// Note: this function manually calls the Python-side overload
// for this function instead of using the PYBIND11_OVERLOAD_XYZ
// macros. This is done so that the Python-side function can
// return a py::bytes instead of a std::vector<uint8_t>.
std::vector<uint8_t> compareSet(
const std::string& key,
const std::vector<uint8_t>& expectedValue,
const std::vector<uint8_t>& desiredValue) override {
pybind11::gil_scoped_acquire gil;
pybind11::function fn = pybind11::get_overload(
static_cast<const ::c10d::Store*>(this), "compare_set");
TORCH_INTERNAL_ASSERT(fn);
// Cast return value from Python to py::bytes, then implicitly
// convert that to a std::string, so that we can construct a
// std::vector<uint8_t>. There is no API for directly accessing
// the contents of the py::bytes object.
std::string str =
pybind11::cast<py::bytes>(fn(key, expectedValue, desiredValue));
return std::vector<uint8_t>(str.begin(), str.end());
}
int64_t add(const std::string& key, int64_t value) override {
PYBIND11_OVERLOAD_PURE(int64_t, ::c10d::Store, add, key, value);
}
int64_t getNumKeys() override {
PYBIND11_OVERLOAD_PURE(int64_t, ::c10d::Store, getNumKeys);
}
bool deleteKey(const std::string& key) override {
PYBIND11_OVERLOAD_PURE(bool, ::c10d::Store, deleteKey, key);
}
bool check(const std::vector<std::string>& keys) override {
PYBIND11_OVERLOAD_PURE(bool, ::c10d::Store, check, keys);
}
void wait(const std::vector<std::string>& keys) override {
PYBIND11_OVERLOAD_PURE(void, ::c10d::Store, wait, keys);
}
void wait(
const std::vector<std::string>& keys,
const std::chrono::milliseconds& timeout) override {
PYBIND11_OVERLOAD_PURE(void, ::c10d::Store, wait, keys, timeout);
}
};
// Called from DDP's Python API to create a c10d Python comm hook object.
// The input state and callable comm_hook are Python objects. It later calls
// register_comm_hook function of the reducer input to register the hook.
void _register_comm_hook(
::c10d::Reducer& reducer,
py::object state,
py::object comm_hook) {
reducer.register_comm_hook(std::make_unique<::c10d::PythonCommHook>(
std::move(state), std::move(comm_hook)));
}
// Called from DDP's Python API to create a c10d C++ comm hook.
// The input is an enum hook type. It later calls register_builtin_comm_hook
// function of the reducer input to set the hook type.
void _register_builtin_comm_hook(
::c10d::Reducer& reducer,
::c10d::BuiltinCommHookType comm_hook_type) {
reducer.register_builtin_comm_hook(comm_hook_type);
}
PyObject* c10d_init(PyObject* _unused, PyObject* noargs) {
C10_LOG_API_USAGE_ONCE("c10d.python.import");
auto c10d_module = THPObjectPtr(PyImport_ImportModule("torch.distributed"));
if (!c10d_module) {
throw python_error();
}
auto torch_C_module = THPObjectPtr(PyImport_ImportModule("torch._C"));
if (!torch_C_module) {
throw python_error();
}
auto torch_C_m = py::handle(torch_C_module).cast<py::module>();
auto m =
torch_C_m.def_submodule("_distributed_c10d", "distributed c10d bindings");
auto module = py::handle(m).cast<py::module>();
module
.def(
"_register_comm_hook",
&_register_comm_hook,
py::arg("reducer"),
py::arg("state"),
py::arg("comm_hook"),
py::call_guard<py::gil_scoped_release>())
.def(
"_register_builtin_comm_hook",
&_register_builtin_comm_hook,
py::arg("reducer"),
py::arg("comm_hook_type"));
shared_ptr_class_<::c10d::GradBucket>(
module,
"GradBucket",
R"(
This class mainly passes a flattened gradient tensor
(returned by :meth:`~torch.distributed.GradBucket.buffer`)
to DDP communication hook.
This tensor can be further decomposed into a list of per-parameter tensors within this bucket
(returned by :meth:`~torch.distributed.GradBucket.get_per_parameter_tensors`)
to apply layer-wise operations.
)")
.def(
"index",
&::c10d::GradBucket::getIndex,
py::call_guard<py::gil_scoped_release>(),
R"(
.. warning::
Since the buckets are rebuilt after the first iteration, should not rely on the indices at the beginning of training.
Returns:
The index of a bucket that stores gradients of a few contiguous layers.
All the gradients are bucketized.
)")
.def(
"buffer",
&::c10d::GradBucket::getBuffer,
py::call_guard<py::gil_scoped_release>(),
R"(
Returns:
A flattened 1D ``torch.Tensor`` buffer,
which can be further decomposed into a list of per-parameter tensors within this bucket.
)")
.def(
"gradients",
&::c10d::GradBucket::getGradients,
py::call_guard<py::gil_scoped_release>(),
R"(
Returns:
A list of ``torch.Tensor``. Each tensor in the list corresponds to a gradient.
)")
.def(
"parameters",
&::c10d::GradBucket::getParameters,
py::call_guard<py::gil_scoped_release>(),
R"(
Returns:
A list of ``torch.Tensor``. Each tensor in the list corresponds to a model
parameter.
)")
.def(
"is_last",
&::c10d::GradBucket::isLast,
py::call_guard<py::gil_scoped_release>(),
R"(
Returns:
Whether this bucket is the last bucket to allreduce in an iteration.
This also means that this bucket corresponds to the first few layers in the forward pass.
)")
.def(
"set_buffer",
&::c10d::GradBucket::setBuffer,
py::arg("buffer"),
py::call_guard<py::gil_scoped_release>(),
R"(
Replaces the tensor in the bucket with the input tensor buffer.
)");
py::enum_<::c10d::BuiltinCommHookType>(module, "BuiltinCommHookType", R"(
An enum-like class for built-in communication hooks: ``ALLREDUCE`` and ``FP16_COMPRESS``.)")
.value("ALLREDUCE", ::c10d::BuiltinCommHookType::ALLREDUCE)
.value("FP16_COMPRESS", ::c10d::BuiltinCommHookType::FP16_COMPRESS);
shared_ptr_class_<::c10d::Reducer>(module, "Reducer")
.def(
py::init<
std::vector<at::Tensor>,
std::vector<std::vector<size_t>>,
std::vector<size_t>,
c10::intrusive_ptr<::c10d::ProcessGroup>,
std::vector<bool>,
int64_t,
bool,
bool,
std::unordered_map<size_t, std::string>,
int64_t>(),
py::arg("params"),
py::arg("bucket_indices"),
py::arg("per_bucket_size_limits"),
py::arg("process_group"),
py::arg("expect_sparse_gradients") = std::vector<bool>(),
py::arg("bucket_bytes_cap") = ::c10d::kDefaultBucketBytesCap,
py::arg("find_unused_parameters") = false,
py::arg("gradient_as_bucket_view") = false,
py::arg("param_to_name_mapping") =
std::unordered_map<size_t, std::string>(),
py::arg("first_bucket_bytes_cap") = ::c10d::kDefaultFirstBucketBytes,
py::call_guard<py::gil_scoped_release>())
.def(
"prepare_for_forward",
&::c10d::Reducer::prepare_for_forward,
py::call_guard<py::gil_scoped_release>())
.def(
"prepare_for_backward",
&::c10d::Reducer::prepare_for_backward,
py::call_guard<py::gil_scoped_release>())
.def(
"prepare_for_backward",
[](::c10d::Reducer& reducer, const at::Tensor& output) -> void {
reducer.prepare_for_backward({output});
},
py::call_guard<py::gil_scoped_release>())
.def("get_backward_stats", &::c10d::Reducer::get_backward_stats)
.def(
"_install_post_backward_futures",
[](::c10d::Reducer& reducer,
const std::vector<std::shared_ptr<jit::PythonFutureWrapper>>&
futs) {
c10::List<c10::intrusive_ptr<c10::ivalue::Future>> futures(
c10::FutureType::create(c10::TensorType::get()));
for (const auto& fut : futs) {
futures.push_back(fut->fut);
}
reducer.install_futures(std::move(futures));
},
py::call_guard<py::gil_scoped_release>())
.def(
"_rebuild_buckets",
&::c10d::Reducer::rebuild_buckets,
py::call_guard<py::gil_scoped_release>())
.def(
"_get_zeros_like_grad_buckets",
[](::c10d::Reducer& reducer) {
return reducer.get_grad_buckets(/* return_zero_tensors */ true);
},
py::call_guard<py::gil_scoped_release>())
.def(
"_push_all_rebuilt_params",
&::c10d::Reducer::push_rebuilt_params_for_all_indices,
py::call_guard<py::gil_scoped_release>())
.def(
"_set_forward_pass_work_handle",
&::c10d::Reducer::set_forward_pass_work_handle,
py::call_guard<py::gil_scoped_release>())
.def(
"_get_local_used_map", &::c10d::Reducer::get_local_used_map_on_device)
.def(
"_set_ddp_runtime_logging_sample_rate",
&::c10d::Reducer::set_ddp_runtime_logging_sample_rate,
py::arg("sample_rate"),
py::call_guard<py::gil_scoped_release>())
.def(
"_set_static_graph",
&::c10d::Reducer::set_static_graph,
py::call_guard<py::gil_scoped_release>())
.def(
"_ddp_graph_static",
&::c10d::Reducer::ddp_graph_static,
py::call_guard<py::gil_scoped_release>())
.def(
"_delay_all_reduce",
&::c10d::Reducer::delay_all_reduce,
py::call_guard<py::gil_scoped_release>())
.def(
"_run_comm_hook",
[](::c10d::Reducer& reducer, ::c10d::GradBucket& bucket)
-> std::shared_ptr<jit::PythonFutureWrapper> {
c10::intrusive_ptr<c10::ivalue::Future> fut =
reducer.run_comm_hook(bucket);
return std::make_shared<jit::PythonFutureWrapper>(fut);
},
py::call_guard<py::gil_scoped_release>())
.def(
"set_logger",
[](::c10d::Reducer& reducer,
const std::shared_ptr<::c10d::Logger> logger) {
std::weak_ptr<::c10d::Logger> logger_weakref = logger;
reducer.set_logger(logger_weakref);
});
shared_ptr_class_<::c10d::Logger>(module, "Logger")
.def(
py::init<std::shared_ptr<::c10d::Reducer>>(),
py::arg("reducer"),
py::call_guard<py::gil_scoped_release>())
.def(
"set_construction_data_and_log",
&::c10d::Logger::set_construction_data_and_log,
py::arg("module_name"),
py::arg("device_ids"),
py::arg("output_device"),
py::arg("broadcast_buffers"),
py::arg("has_sync_bn"),
py::arg("static_graph"),
py::call_guard<py::gil_scoped_release>())
.def(
"set_runtime_stats_and_log",
&::c10d::Logger::set_runtime_stats_and_log,
py::call_guard<py::gil_scoped_release>())
.def(
"set_error_and_log",
[](::c10d::Logger& logger, const std::string& error) {
logger.set_error_and_log(error);
},
py::call_guard<py::gil_scoped_release>())
.def(
"_get_ddp_logging_data",
&::c10d::Logger::get_ddp_logging_data,
py::call_guard<py::gil_scoped_release>())
.def(
"_set_comm_hook_name",
&::c10d::Logger::set_comm_hook,
py::arg("comm_hook"),
py::call_guard<py::gil_scoped_release>())
.def(
"_set_uneven_input_join",
&::c10d::Logger::set_uneven_input_join,
py::call_guard<py::gil_scoped_release>())
.def(
"_set_static_graph",
&::c10d::Logger::set_static_graph,
py::call_guard<py::gil_scoped_release>());
py::enum_<::c10d::DebugLevel>(module, "DebugLevel", R"(
An enum whose values correspond to different debug levels of the
torch.distributed package. Currently supporting OFF, INFO, and DETAIL,
which can be set via the TORCH_DISTRIBUTED_DEBUG environment variable
or via ``set_debug_level()`` function.
)")
.value("OFF", ::c10d::DebugLevel::Off)
.value("INFO", ::c10d::DebugLevel::Info)
.value("DETAIL", ::c10d::DebugLevel::Detail);
module
.def(
"get_debug_level",
::c10d::debug_level,
R"(Gets the debug level of the torch.distributed package.)")
.def(
"set_debug_level",
::c10d::setDebugLevel,
R"(Sets the debug level of the torch.distributed package.)")
.def(
"set_debug_level_from_env",
::c10d::setDebugLevelFromEnvironment,
R"(Sets the debug level of the torch.distributed package from the
``TORCH_DISTRIBUTED_DEBUG`` environment variable.)");
// TODO(crcrpar): Hardening `ReduceOp`.
// While keeping most op types as enum value,
// making `PREMUL_SUM` callable, i.e., allowing for
// `ReduceOp.PREMUL_SUM(scale)` might be better as per @wanchaol.
// https://pybind11.readthedocs.io/en/stable/classes.html#enumerations-and-internal-types
py::class_<::c10d::ReduceOp> reduce_op(module, "ReduceOp", R"(
An enum-like class for available reduction operations: ``SUM``, ``PRODUCT``,
``MIN``, ``MAX``, ``BAND``, ``BOR``, ``BXOR``, and ``PREMUL_SUM``.
``BAND``, ``BOR``, and ``BXOR`` reductions are not available when
using the ``NCCL`` backend.
``AVG`` divides values by the world size before summing across ranks.
``AVG`` is only available with the ``NCCL`` backend,
and only for NCCL versions 2.10 or later.
``PREMUL_SUM`` multiplies inputs by a given scalar locally before reduction.
``PREMUL_SUM`` is only available with the ``NCCL`` backend,
and only available for NCCL versions 2.11 or later. Users are supposed to
use ``torch.distributed._make_nccl_premul_sum``.
Additionally, ``MAX``, ``MIN`` and ``PRODUCT`` are not supported for complex tensors.
The values of this class can be accessed as attributes, e.g., ``ReduceOp.SUM``.
They are used in specifying strategies for reduction collectives, e.g.,
:func:`reduce`, :func:`all_reduce_multigpu`, etc.
This class does not support ``__members__`` property.)");
reduce_op.def(py::init<::c10d::ReduceOp::RedOpType>())
.def_readwrite("op", &::c10d::ReduceOp::op_);
// The following are for some kind of backward compatibility.
// Since c10d::ReduceOp had been an `enum class`, users can do comparison and
// take hash of `::c10d::ReduceOp`. To avoid losing these functionality, here
// I define some member methods.
reduce_op
.def(
"__eq__",
[](const ::c10d::ReduceOp& self,
const ::c10d::ReduceOp::RedOpType& other) {
return self == other;
})
.def(
"__eq__",
[](const ::c10d::ReduceOp& self, const ::c10d::ReduceOp& other) {
return self == other.op_;
})
.def("__hash__", [](const ::c10d::ReduceOp& self) {
return static_cast<uint8_t>(self.op_);
});
// note(crcrpar): Deliberately skip
// [`export_values`](https://pybind11.readthedocs.io/en/stable/classes.html#enumerations-and-internal-types)
// here and manually set values in Python side. See note "ReduceOp static
// class attributes to support `isinstance`"
py::enum_<::c10d::ReduceOp::RedOpType>(reduce_op, "RedOpType")
.value("SUM", ::c10d::ReduceOp::RedOpType::SUM)
.value("AVG", ::c10d::ReduceOp::RedOpType::AVG)
.value("PRODUCT", ::c10d::ReduceOp::RedOpType::PRODUCT)
.value("MIN", ::c10d::ReduceOp::RedOpType::MIN)
.value("MAX", ::c10d::ReduceOp::RedOpType::MAX)
.value("BAND", ::c10d::ReduceOp::RedOpType::BAND)
.value("BOR", ::c10d::ReduceOp::RedOpType::BOR)
.value("BXOR", ::c10d::ReduceOp::RedOpType::BXOR)
.value("PREMUL_SUM", ::c10d::ReduceOp::RedOpType::PREMUL_SUM);
// note(crcrpar): This could be removed because users will not pass
// `RedOpType` to reduce collective ops Ref: [Implicit
// conversions](https://pybind11.readthedocs.io/en/stable/advanced/classes.html#implicit-conversions)
// Let us skip the explicit construction of `c10d::ReduceOp` from
// `c10d::ReduceOp::RedOpType` in Python.
py::implicitly_convertible<::c10d::ReduceOp::RedOpType, ::c10d::ReduceOp>();
module
.def(
"_make_nccl_premul_sum",
&::c10d::makeNCCLPreMulSum<double>,
py::arg("factor").noconvert(),
py::return_value_policy::copy, // seems safest
py::call_guard<py::gil_scoped_release>())
.def(
"_make_nccl_premul_sum",
&::c10d::makeNCCLPreMulSum<std::vector<at::Tensor>>,
py::arg("factor").noconvert(),
py::return_value_policy::copy, // seems safest
py::call_guard<py::gil_scoped_release>());
py::class_<::c10d::BroadcastOptions>(module, "BroadcastOptions")
.def(py::init<>())
.def_readwrite("rootRank", &::c10d::BroadcastOptions::rootRank)
.def_readwrite("rootTensor", &::c10d::BroadcastOptions::rootTensor)
.def_readwrite("timeout", &::c10d::BroadcastOptions::timeout);
py::class_<::c10d::AllreduceOptions>(module, "AllreduceOptions")
.def(py::init<>())
.def_readwrite("reduceOp", &::c10d::AllreduceOptions::reduceOp)
.def_readwrite("timeout", &::c10d::AllreduceOptions::timeout);
py::class_<::c10d::AllreduceCoalescedOptions>(
module, "AllreduceCoalescedOptions")
.def(py::init<>())
.def_readwrite("reduceOp", &::c10d::AllreduceCoalescedOptions::reduceOp)
.def_readwrite("timeout", &::c10d::AllreduceCoalescedOptions::timeout);
py::class_<::c10d::ReduceOptions>(module, "ReduceOptions")
.def(py::init<>())
.def_readwrite("reduceOp", &::c10d::ReduceOptions::reduceOp)
.def_readwrite("rootRank", &::c10d::ReduceOptions::rootRank)
.def_readwrite("rootTensor", &::c10d::ReduceOptions::rootTensor)
.def_readwrite("timeout", &::c10d::ReduceOptions::timeout);
py::class_<::c10d::AllgatherOptions>(module, "AllgatherOptions")
.def(py::init<>())
.def_readwrite("timeout", &::c10d::AllgatherOptions::timeout);
py::class_<::c10d::GatherOptions>(module, "GatherOptions")
.def(py::init<>())
.def_readwrite("rootRank", &::c10d::GatherOptions::rootRank)
.def_readwrite("timeout", &::c10d::GatherOptions::timeout);
py::class_<::c10d::ScatterOptions>(module, "ScatterOptions")
.def(py::init<>())
.def_readwrite("rootRank", &::c10d::ScatterOptions::rootRank)
.def_readwrite("timeout", &::c10d::ScatterOptions::timeout);
py::class_<::c10d::ReduceScatterOptions>(module, "ReduceScatterOptions")
.def(py::init<>())
.def_readwrite("reduceOp", &::c10d::ReduceScatterOptions::reduceOp)
.def_readwrite("timeout", &::c10d::ReduceScatterOptions::timeout);
py::class_<::c10d::BarrierOptions>(module, "BarrierOptions")
.def(py::init<>())
.def_readwrite("device_ids", &::c10d::BarrierOptions::device_ids)
.def_readwrite("timeout", &::c10d::BarrierOptions::timeout);
py::class_<::c10d::AllToAllOptions>(module, "AllToAllOptions")
.def(py::init<>())
.def_readwrite("timeout", &::c10d::AllToAllOptions::timeout);
py::class_<::c10d::DistributedBackendOptions>(
module, "_DistributedBackendOptions")
.def(py::init<>())
.def_readwrite("store", &::c10d::DistributedBackendOptions::store)
.def_readwrite(
"group_rank", &::c10d::DistributedBackendOptions::group_rank)
.def_readwrite(
"group_size", &::c10d::DistributedBackendOptions::group_size)
.def_readwrite("timeout", &::c10d::DistributedBackendOptions::timeout)
.def_readwrite("group_id", &::c10d::DistributedBackendOptions::group_id)
.def_readwrite(
"global_ranks_in_group",
&::c10d::DistributedBackendOptions::global_ranks_in_group);
auto store =
py::class_<::c10d::Store, c10::intrusive_ptr<::c10d::Store>, PythonStore>(
module,
"Store",
R"(
Base class for all store implementations, such as the 3 provided by PyTorch
distributed: (:class:`~torch.distributed.TCPStore`, :class:`~torch.distributed.FileStore`,
and :class:`~torch.distributed.HashStore`).
)")
// Default constructor.
.def(py::init<>())
// Convert from std::string to std::vector<uint8>.
.def(
"set",
[](::c10d::Store& store,
const std::string& key,
const std::string& value) {
std::vector<uint8_t> value_(value.begin(), value.end());
store.set(key, value_);
},
py::call_guard<py::gil_scoped_release>(),
R"(
Inserts the key-value pair into the store based on the supplied ``key`` and
``value``. If ``key`` already exists in the store, it will overwrite the old
value with the new supplied ``value``.
Arguments:
key (str): The key to be added to the store.
value (str): The value associated with ``key`` to be added to the store.
Example::
>>> import torch.distributed as dist
>>> from datetime import timedelta
>>> store = dist.TCPStore("127.0.0.1", 0, 1, True, timedelta(seconds=30))
>>> store.set("first_key", "first_value")
>>> # Should return "first_value"
>>> store.get("first_key")
)")
.def(
"compare_set",
[](::c10d::Store& store,
const std::string& key,
const std::string& expected_value,
const std::string& desired_value) -> py::bytes {
std::vector<uint8_t> expectedValue_(
expected_value.begin(), expected_value.end());
std::vector<uint8_t> desiredValue_(
desired_value.begin(), desired_value.end());
auto value =
store.compareSet(key, expectedValue_, desiredValue_);
return py::bytes(
reinterpret_cast<char*>(value.data()), value.size());
},
py::call_guard<py::gil_scoped_release>(),
R"(
Inserts the key-value pair into the store based on the supplied ``key`` and
performs comparison between ``expected_value`` and ``desired_value`` before inserting. ``desired_value``
will only be set if ``expected_value`` for the ``key`` already exists in the store or if ``expected_value``
is an empty string.
Arguments:
key (str): The key to be checked in the store.
expected_value (str): The value associated with ``key`` to be checked before insertion.
desired_value (str): The value associated with ``key`` to be added to the store.
Example::
>>> import torch.distributed as dist
>>> from datetime import timedelta
>>> store = dist.TCPStore("127.0.0.1", 0, 1, True, timedelta(seconds=30))
>>> store.set("key", "first_value")
>>> store.compare_set("key", "first_value", "second_value")
>>> # Should return "second_value"
>>> store.get("key")
)")
// Convert from std::vector<uint8_t> to py::bytes.
// The returned value is not guaranteed to be valid UTF-8.
.def(
"get",
[](::c10d::Store& store, const std::string& key) -> py::bytes {
auto value = [&]() {
py::gil_scoped_release guard;
return store.get(key);
}();
return py::bytes(
reinterpret_cast<char*>(value.data()), value.size());
},
R"(
Retrieves the value associated with the given ``key`` in the store. If ``key`` is not
present in the store, the function will wait for ``timeout``, which is defined
when initializing the store, before throwing an exception.
Arguments:
key (str): The function will return the value associated with this key.
Returns:
Value associated with ``key`` if ``key`` is in the store.
Example::
>>> import torch.distributed as dist
>>> from datetime import timedelta
>>> store = dist.TCPStore("127.0.0.1", 0, 1, True, timedelta(seconds=30))
>>> store.set("first_key", "first_value")
>>> # Should return "first_value"
>>> store.get("first_key")
)")
.def(
"add",
&::c10d::Store::add,
py::call_guard<py::gil_scoped_release>(),
R"(
The first call to add for a given ``key`` creates a counter associated
with ``key`` in the store, initialized to ``amount``. Subsequent calls to add
with the same ``key`` increment the counter by the specified ``amount``.
Calling :meth:`~torch.distributed.store.add` with a key that has already
been set in the store by :meth:`~torch.distributed.store.set` will result
in an exception.
Arguments:
key (str): The key in the store whose counter will be incremented.
amount (int): The quantity by which the counter will be incremented.
Example::
>>> import torch.distributed as dist
>>> from datetime import timedelta
>>> # Using TCPStore as an example, other store types can also be used
>>> store = dist.TCPStore("127.0.0.1", 0, 1, True, timedelta(seconds=30))
>>> store.add("first_key", 1)
>>> store.add("first_key", 6)
>>> # Should return 7
>>> store.get("first_key")
)")
.def(
"delete_key",
&::c10d::Store::deleteKey,
py::call_guard<py::gil_scoped_release>(),
R"(
Deletes the key-value pair associated with ``key`` from the store. Returns
`true` if the key was successfully deleted, and `false` if it was not.
.. warning::
The ``delete_key`` API is only supported by the :class:`~torch.distributed.TCPStore` and :class:`~torch.distributed.HashStore`. Using this API
with the :class:`~torch.distributed.FileStore` will result in an exception.
Arguments:
key (str): The key to be deleted from the store
Returns:
`True` if ``key`` was deleted, otherwise `False`.
Example::
>>> import torch.distributed as dist
>>> from datetime import timedelta
>>> # Using TCPStore as an example, HashStore can also be used
>>> store = dist.TCPStore("127.0.0.1", 0, 1, True, timedelta(seconds=30))
>>> store.set("first_key")
>>> # This should return true
>>> store.delete_key("first_key")
>>> # This should return false
>>> store.delete_key("bad_key")
)")
.def(
"num_keys",
&::c10d::Store::getNumKeys,
py::call_guard<py::gil_scoped_release>(),
R"(
Returns the number of keys set in the store. Note that this number will typically
be one greater than the number of keys added by :meth:`~torch.distributed.store.set`
and :meth:`~torch.distributed.store.add` since one key is used to coordinate all
the workers using the store.
.. warning::
When used with the :class:`~torch.distributed.TCPStore`, ``num_keys`` returns the number of keys written to the underlying file. If the store is destructed and another store is created with the same file, the original keys will be retained.
Returns:
The number of keys present in the store.
Example::
>>> import torch.distributed as dist
>>> from datetime import timedelta
>>> # Using TCPStore as an example, other store types can also be used
>>> store = dist.TCPStore("127.0.0.1", 0, 1, True, timedelta(seconds=30))
>>> store.set("first_key", "first_value")
>>> # This should return 2
>>> store.num_keys()
)")
.def(
"set_timeout",
&::c10d::Store::setTimeout,
py::call_guard<py::gil_scoped_release>(),
R"(
Sets the store's default timeout. This timeout is used during initialization and in
:meth:`~torch.distributed.store.wait` and :meth:`~torch.distributed.store.get`.
Arguments:
timeout (timedelta): timeout to be set in the store.
Example::
>>> import torch.distributed as dist
>>> from datetime import timedelta
>>> # Using TCPStore as an example, other store types can also be used
>>> store = dist.TCPStore("127.0.0.1", 0, 1, True, timedelta(seconds=30))
>>> store.set_timeout(timedelta(seconds=10))
>>> # This will throw an exception after 10 seconds
>>> store.wait(["bad_key"])
)")
.def(
"wait",
[](::c10d::Store& store, const std::vector<std::string>& keys) {
store.wait(keys);
},
py::call_guard<py::gil_scoped_release>(),
R"(
Waits for each key in ``keys`` to be added to the store. If not all keys are
set before the ``timeout`` (set during store initialization), then ``wait``
will throw an exception.
Arguments:
keys (list): List of keys on which to wait until they are set in the store.
Example::
>>> import torch.distributed as dist
>>> from datetime import timedelta
>>> # Using TCPStore as an example, other store types can also be used
>>> store = dist.TCPStore("127.0.0.1", 0, 1, True, timedelta(seconds=30))
>>> # This will throw an exception after 30 seconds
>>> store.wait(["bad_key"])
)")
.def(
"wait",
[](::c10d::Store& store,
const std::vector<std::string>& keys,
const std::chrono::milliseconds& timeout) {
store.wait(keys, timeout);
},
py::call_guard<py::gil_scoped_release>(),
R"(
Waits for each key in ``keys`` to be added to the store, and throws an exception
if the keys have not been set by the supplied ``timeout``.
Arguments:
keys (list): List of keys on which to wait until they are set in the store.
timeout (timedelta): Time to wait for the keys to be added before throwing an exception.
Example::
>>> import torch.distributed as dist
>>> from datetime import timedelta
>>> # Using TCPStore as an example, other store types can also be used
>>> store = dist.TCPStore("127.0.0.1", 0, 1, True, timedelta(seconds=30))
>>> # This will throw an exception after 10 seconds
>>> store.wait(["bad_key"], timedelta(seconds=10))
)")
.def_property_readonly(
"timeout",
&::c10d::Store::getTimeout,
R"(Gets the timeout of the store.)");
intrusive_ptr_class_<::c10d::FileStore>(
module,
"FileStore",
store,
R"(
A store implementation that uses a file to store the underlying key-value pairs.
Arguments:
file_name (str): path of the file in which to store the key-value pairs
world_size (int, optional): The total number of processes using the store. Default is -1 (a negative value indicates a non-fixed number of store users).
Example::
>>> import torch.distributed as dist
>>> store1 = dist.FileStore("/tmp/filestore", 2)
>>> store2 = dist.FileStore("/tmp/filestore", 2)
>>> # Use any of the store methods from either the client or server after initialization
>>> store1.set("first_key", "first_value")
>>> store2.get("first_key")
)")
.def(
py::init<const std::string&, int>(),
py::arg("file_name"),
py::arg("world_size") = -1)
.def_property_readonly(
"path",
&::c10d::FileStore::getPath,
R"(Gets the path of the file used by FileStore to store key-value pairs.)");
#ifndef _WIN32
intrusive_ptr_class_<::c10d::HashStore>(
module,
"HashStore",
store,
R"(
A thread-safe store implementation based on an underlying hashmap. This store can be used
within the same process (for example, by other threads), but cannot be used across processes.
Example::
>>> import torch.distributed as dist
>>> store = dist.HashStore()
>>> # store can be used from other threads
>>> # Use any of the store methods after initialization
>>> store.set("first_key", "first_value")
)")
.def(py::init<>());
#endif
intrusive_ptr_class_<::c10d::TCPStore>(
module,
"TCPStore",
store,
R"(
A TCP-based distributed key-value store implementation. The server store holds
the data, while the client stores can connect to the server store over TCP and
perform actions such as :meth:`~torch.distributed.store.set` to insert a key-value
pair, :meth:`~torch.distributed.store.get` to retrieve a key-value pair, etc. There
should always be one server store initialized because the client store(s) will wait for
the server to establish a connection.
Arguments:
host_name (str): The hostname or IP Address the server store should run on.
port (int): The port on which the server store should listen for incoming requests.
world_size (int, optional): The total number of store users (number of clients + 1 for the server). Default is None (None indicates a non-fixed number of store users).
is_master (bool, optional): True when initializing the server store and False for client stores. Default is False.
timeout (timedelta, optional): Timeout used by the store during initialization and for methods such as :meth:`~torch.distributed.store.get` and :meth:`~torch.distributed.store.wait`. Default is timedelta(seconds=300)
wait_for_worker (bool, optional): Whether to wait for all the workers to connect with the server store. This is only applicable when world_size is a fixed value. Default is True.
Example::
>>> import torch.distributed as dist
>>> from datetime import timedelta
>>> # Run on process 1 (server)
>>> server_store = dist.TCPStore("127.0.0.1", 1234, 2, True, timedelta(seconds=30))
>>> # Run on process 2 (client)