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
|
#include <torch/csrc/distributed/c10d/Ops.hpp>
#include <ATen/core/dispatch/Dispatcher.h>
#include <torch/csrc/distributed/c10d/Types.hpp>
#include <torch/library.h>
namespace c10d {
namespace {
std::tuple<std::vector<at::Tensor>, c10::intrusive_ptr<Work>> broadcast_(
at::TensorList tensors,
const c10::intrusive_ptr<ProcessGroup>& process_group,
int64_t root_rank,
int64_t root_tensor,
int64_t timeout) {
auto tensor_vec = tensors.vec();
auto work = process_group->broadcast(
tensor_vec,
BroadcastOptions{
root_rank, root_tensor, std::chrono::milliseconds(timeout)});
return std::tuple<std::vector<at::Tensor>, c10::intrusive_ptr<Work>>(
std::move(tensor_vec), work);
}
std::tuple<std::vector<at::Tensor>, c10::intrusive_ptr<Work>> allreduce_(
at::TensorList tensors,
const c10::intrusive_ptr<ProcessGroup>& process_group,
const c10::intrusive_ptr<ReduceOp>& reduce_op,
int64_t timeout) {
auto tensor_vec = tensors.vec();
auto work = process_group->allreduce(
tensor_vec,
AllreduceOptions{*reduce_op.get(), std::chrono::milliseconds(timeout)});
// Return input tensors as output tensors to make inplace allreduce look like
// a functional API, so that make_fx can correctly build the dependencies in
// the graph later.
return std::tuple<std::vector<at::Tensor>, c10::intrusive_ptr<Work>>(
std::move(tensor_vec), work);
}
std::tuple<std::vector<std::vector<at::Tensor>>, c10::intrusive_ptr<Work>>
allgather_(
const std::vector<std::vector<at::Tensor>>& output_tensors,
const std::vector<at::Tensor>& input_tensors,
const c10::intrusive_ptr<ProcessGroup>& process_group,
int64_t timeout) {
auto work = process_group->allgather(
const_cast<std::vector<std::vector<at::Tensor>>&>(output_tensors),
const_cast<std::vector<at::Tensor>&>(input_tensors),
AllgatherOptions{std::chrono::milliseconds(timeout)});
// Copy output tensors (not storage) so that this can be used in a functional
// manner
return std::
tuple<std::vector<std::vector<at::Tensor>>, c10::intrusive_ptr<Work>>(
output_tensors, work);
}
std::tuple<std::vector<at::Tensor>, c10::intrusive_ptr<Work>> reduce_scatter_(
const std::vector<at::Tensor>& output_tensors,
const std::vector<std::vector<at::Tensor>>& input_tensors,
const c10::intrusive_ptr<ProcessGroup>& process_group,
const c10::intrusive_ptr<ReduceOp>& reduce_op,
int64_t timeout) {
auto work = process_group->reduce_scatter(
const_cast<std::vector<at::Tensor>&>(output_tensors),
const_cast<std::vector<std::vector<at::Tensor>>&>(input_tensors),
ReduceScatterOptions{
*reduce_op.get(), std::chrono::milliseconds(timeout)});
return std::tuple<std::vector<at::Tensor>, c10::intrusive_ptr<Work>>(
output_tensors, work);
}
c10::intrusive_ptr<Work> reduce_(
at::TensorList tensors,
const c10::intrusive_ptr<ProcessGroup>& process_group,
const c10::intrusive_ptr<ReduceOp>& reduce_op,
int64_t root_rank,
int64_t root_tensor,
int64_t timeout) {
auto tensor_vec = tensors.vec();
return process_group->reduce(
tensor_vec,
ReduceOptions{
*reduce_op.get(),
root_rank,
root_tensor,
std::chrono::milliseconds(timeout)});
}
c10::intrusive_ptr<Work> gather_(
const std::vector<std::vector<at::Tensor>>& output_tensors,
const std::vector<at::Tensor>& input_tensors,
const c10::intrusive_ptr<ProcessGroup>& process_group,
int64_t root_rank,
int64_t timeout) {
return process_group->gather(
const_cast<std::vector<std::vector<at::Tensor>>&>(output_tensors),
const_cast<std::vector<at::Tensor>&>(input_tensors),
GatherOptions{root_rank, std::chrono::milliseconds(timeout)});
}
std::tuple<std::vector<at::Tensor>, c10::intrusive_ptr<Work>> scatter_(
const std::vector<at::Tensor>& output_tensors,
const std::vector<std::vector<at::Tensor>>& input_tensors,
const c10::intrusive_ptr<ProcessGroup>& process_group,
int64_t root_rank,
int64_t timeout) {
auto work = process_group->scatter(
const_cast<std::vector<at::Tensor>&>(output_tensors),
const_cast<std::vector<std::vector<at::Tensor>>&>(input_tensors),
ScatterOptions{root_rank, std::chrono::milliseconds(timeout)});
return std::tuple<std::vector<at::Tensor>, c10::intrusive_ptr<Work>>(
output_tensors, work);
}
c10::intrusive_ptr<Work> alltoall_(
at::TensorList output_tensors,
at::TensorList input_tensors,
const c10::intrusive_ptr<ProcessGroup>& process_group,
int64_t timeout) {
auto output_tensors_vec = output_tensors.vec();
auto input_tensors_vec = input_tensors.vec();
return process_group->alltoall(
output_tensors_vec,
input_tensors_vec,
AllToAllOptions{std::chrono::milliseconds(timeout)});
}
c10::intrusive_ptr<Work> barrier(
const c10::intrusive_ptr<ProcessGroup>& process_group,
const std::vector<int64_t>& device_ids,
int64_t timeout) {
return process_group->barrier(
BarrierOptions{device_ids, std::chrono::milliseconds(timeout)});
}
c10::intrusive_ptr<Work> send(
at::TensorList tensors,
const c10::intrusive_ptr<ProcessGroup>& process_group,
int64_t dstRank,
int64_t tag) {
auto tensor_vec = tensors.vec();
return process_group->send(
tensor_vec, static_cast<int>(dstRank), static_cast<int>(tag));
}
c10::intrusive_ptr<Work> recv_(
at::TensorList tensors,
const c10::intrusive_ptr<ProcessGroup>& process_group,
int64_t srcRank,
int64_t tag) {
auto tensor_vec = tensors.vec();
return process_group->recv(
tensor_vec, static_cast<int>(srcRank), static_cast<int>(tag));
}
TORCH_LIBRARY(c10d, m) {
// The following ProcessGroup, Work, and ReduceOp definitions are more like
// declarations. They don't expose the details of the two classes into
// TorchScript.
m.class_<ProcessGroup>("ProcessGroup").def(torch::init<int64_t, int64_t>());
m.class_<Work>("Work")
.def(torch::init<>())
.def("wait", [](const c10::intrusive_ptr<Work>& self) { self->wait(); });
m.class_<ReduceOp>("ReduceOp").def(torch::init<>());
// It's important to register the op to the CompositeExplicitAutograd key
// instead of the CompositeImplicitAutograd key to enable
// __torch_dispatch__.
m.def(
"broadcast_",
dispatch(c10::DispatchKey::CompositeExplicitAutograd, broadcast_));
m.def(
"allreduce_",
dispatch(c10::DispatchKey::CompositeExplicitAutograd, allreduce_));
m.def(
"allgather_",
dispatch(c10::DispatchKey::CompositeExplicitAutograd, allgather_));
m.def(
"reduce_scatter_",
dispatch(c10::DispatchKey::CompositeExplicitAutograd, reduce_scatter_));
m.def(
"reduce_",
dispatch(c10::DispatchKey::CompositeExplicitAutograd, reduce_));
m.def(
"gather_",
dispatch(c10::DispatchKey::CompositeExplicitAutograd, gather_));
m.def(
"scatter_",
dispatch(c10::DispatchKey::CompositeExplicitAutograd, scatter_));
m.def(
"alltoall_",
dispatch(c10::DispatchKey::CompositeExplicitAutograd, alltoall_));
m.def(
"barrier",
dispatch(c10::DispatchKey::CompositeExplicitAutograd, barrier));
m.def("send", dispatch(c10::DispatchKey::CompositeExplicitAutograd, send));
m.def("recv_", dispatch(c10::DispatchKey::CompositeExplicitAutograd, recv_));
}
} // namespace
namespace ops {
c10::intrusive_ptr<Work> broadcast(
const c10::intrusive_ptr<ProcessGroup>& process_group,
at::TensorList tensors,
const BroadcastOptions& opts) {
static auto op =
c10::Dispatcher::singleton()
.findSchemaOrThrow("c10d::broadcast_", "")
.typed<std::tuple<std::vector<at::Tensor>, c10::intrusive_ptr<Work>>(
at::TensorList,
const c10::intrusive_ptr<::c10d::ProcessGroup>&,
int64_t,
int64_t,
int64_t)>();
// It's awakward to unbox the opts here and box them again in the custom C++
// op. But it's also complicated to make opts as a CustomClassHolder. Leave it
// as it is now.
return std::get<1>(op.call(
tensors,
process_group,
opts.rootRank,
opts.rootTensor,
opts.timeout.count()));
}
c10::intrusive_ptr<Work> allreduce(
const c10::intrusive_ptr<ProcessGroup>& process_group,
at::TensorList tensors,
const AllreduceOptions& opts) {
static auto op =
c10::Dispatcher::singleton()
.findSchemaOrThrow("c10d::allreduce_", "")
.typed<std::tuple<std::vector<at::Tensor>, c10::intrusive_ptr<Work>>(
at::TensorList,
const c10::intrusive_ptr<::c10d::ProcessGroup>&,
const c10::intrusive_ptr<::c10d::ReduceOp>&,
int64_t)>();
return std::get<1>(op.call(
tensors,
process_group,
c10::make_intrusive<ReduceOp>(opts.reduceOp),
opts.timeout.count()));
}
c10::intrusive_ptr<Work> allgather(
const c10::intrusive_ptr<ProcessGroup>& process_group,
const std::vector<std::vector<at::Tensor>>& output_tensors,
const std::vector<at::Tensor>& input_tensors,
const AllgatherOptions& opts) {
static auto op = c10::Dispatcher::singleton()
.findSchemaOrThrow("c10d::allgather_", "")
.typed<std::tuple<
std::vector<std::vector<at::Tensor>>,
c10::intrusive_ptr<Work>>(
const std::vector<std::vector<at::Tensor>>&,
const std::vector<at::Tensor>&,
const c10::intrusive_ptr<::c10d::ProcessGroup>&,
int64_t)>();
return std::get<1>(op.call(
output_tensors, input_tensors, process_group, opts.timeout.count()));
}
c10::intrusive_ptr<Work> reduce_scatter(
const c10::intrusive_ptr<ProcessGroup>& process_group,
const std::vector<at::Tensor>& output_tensors,
const std::vector<std::vector<at::Tensor>>& input_tensors,
const ReduceScatterOptions& opts) {
static auto op =
c10::Dispatcher::singleton()
.findSchemaOrThrow("c10d::reduce_scatter_", "")
.typed<std::tuple<std::vector<at::Tensor>, c10::intrusive_ptr<Work>>(
const std::vector<at::Tensor>&,
const std::vector<std::vector<at::Tensor>>&,
const c10::intrusive_ptr<::c10d::ProcessGroup>&,
const c10::intrusive_ptr<::c10d::ReduceOp>&,
int64_t)>();
return std::get<1>(op.call(
output_tensors,
input_tensors,
process_group,
c10::make_intrusive<::c10d::ReduceOp>(opts.reduceOp),
opts.timeout.count()));
}
c10::intrusive_ptr<Work> reduce(
const c10::intrusive_ptr<ProcessGroup>& process_group,
at::TensorList tensors,
const ReduceOptions& opts) {
static auto op = c10::Dispatcher::singleton()
.findSchemaOrThrow("c10d::reduce_", "")
.typed<c10::intrusive_ptr<::c10d::Work>(
at::TensorList,
const c10::intrusive_ptr<::c10d::ProcessGroup>&,
const c10::intrusive_ptr<::c10d::ReduceOp>&,
int64_t,
int64_t,
int64_t)>();
return op.call(
tensors,
process_group,
c10::make_intrusive<ReduceOp>(opts.reduceOp),
opts.rootRank,
opts.rootTensor,
opts.timeout.count());
}
c10::intrusive_ptr<Work> gather(
const c10::intrusive_ptr<ProcessGroup>& process_group,
const std::vector<std::vector<at::Tensor>>& output_tensors,
const std::vector<at::Tensor>& input_tensors,
const GatherOptions& opts) {
static auto op = c10::Dispatcher::singleton()
.findSchemaOrThrow("c10d::gather_", "")
.typed<c10::intrusive_ptr<::c10d::Work>(
const std::vector<std::vector<at::Tensor>>&,
const std::vector<at::Tensor>&,
const c10::intrusive_ptr<::c10d::ProcessGroup>&,
int64_t,
int64_t)>();
return op.call(
output_tensors,
input_tensors,
process_group,
opts.rootRank,
opts.timeout.count());
}
c10::intrusive_ptr<Work> scatter(
const c10::intrusive_ptr<ProcessGroup>& process_group,
const std::vector<at::Tensor>& output_tensors,
const std::vector<std::vector<at::Tensor>>& input_tensors,
const ScatterOptions& opts) {
static auto op =
c10::Dispatcher::singleton()
.findSchemaOrThrow("c10d::scatter_", "")
.typed<std::tuple<std::vector<at::Tensor>, c10::intrusive_ptr<Work>>(
const std::vector<at::Tensor>&,
const std::vector<std::vector<at::Tensor>>&,
const c10::intrusive_ptr<::c10d::ProcessGroup>&,
int64_t,
int64_t)>();
return std::get<1>(op.call(
output_tensors,
input_tensors,
process_group,
opts.rootRank,
opts.timeout.count()));
}
c10::intrusive_ptr<Work> alltoall(
const c10::intrusive_ptr<ProcessGroup>& process_group,
at::TensorList output_tensors,
at::TensorList input_tensors,
const AllToAllOptions& opts) {
static auto op = c10::Dispatcher::singleton()
.findSchemaOrThrow("c10d::alltoall_", "")
.typed<c10::intrusive_ptr<::c10d::Work>(
at::TensorList,
at::TensorList,
const c10::intrusive_ptr<::c10d::ProcessGroup>&,
int64_t)>();
return op.call(
output_tensors, input_tensors, process_group, opts.timeout.count());
}
c10::intrusive_ptr<Work> barrier(
const c10::intrusive_ptr<ProcessGroup>& process_group,
const BarrierOptions& opts) {
static auto op = c10::Dispatcher::singleton()
.findSchemaOrThrow("c10d::barrier", "")
.typed<c10::intrusive_ptr<::c10d::Work>(
const c10::intrusive_ptr<::c10d::ProcessGroup>&,
const std::vector<int64_t>&,
int64_t)>();
return op.call(process_group, opts.device_ids, opts.timeout.count());
}
c10::intrusive_ptr<Work> send(
const c10::intrusive_ptr<ProcessGroup>& process_group,
at::TensorList tensors,
int64_t dstRank,
int64_t tag) {
static auto op = c10::Dispatcher::singleton()
.findSchemaOrThrow("c10d::send", "")
.typed<c10::intrusive_ptr<::c10d::Work>(
at::TensorList,
const c10::intrusive_ptr<::c10d::ProcessGroup>&,
int64_t,
int64_t)>();
return op.call(tensors, process_group, dstRank, tag);
}
c10::intrusive_ptr<Work> recv(
const c10::intrusive_ptr<ProcessGroup>& process_group,
at::TensorList tensors,
int64_t srcRank,
int64_t tag) {
static auto op = c10::Dispatcher::singleton()
.findSchemaOrThrow("c10d::recv_", "")
.typed<c10::intrusive_ptr<::c10d::Work>(
at::TensorList,
const c10::intrusive_ptr<::c10d::ProcessGroup>&,
int64_t,
int64_t)>();
return op.call(tensors, process_group, srcRank, tag);
}
} // namespace ops
} // namespace c10d
|
