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
|
#include <c10/util/intrusive_ptr.h>
#include <torch/csrc/distributed/c10d/ProcessGroup.hpp>
#include <torch/csrc/distributed/c10d/Types.hpp>
#include <torch/library.h>
namespace c10d {
namespace ops {
// Below are ProcessGroup's corresponding ops for each backend. Ops are but
// routed through the dispatcher to be dispatched to the appropriate backend.
// Currently a no-op as the process group does not have a list of backends.
c10::intrusive_ptr<Work> send_cpu(
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> send_cuda(
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_cpu_(
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));
}
c10::intrusive_ptr<Work> recv_cuda_(
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));
}
std::tuple<std::vector<at::Tensor>, c10::intrusive_ptr<Work>> broadcast_cpu_(
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>> broadcast_cuda_(
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_cpu_(
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<at::Tensor>, c10::intrusive_ptr<Work>> allreduce_cuda_(
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);
}
// register functions to dispatcher
namespace {
TORCH_LIBRARY_IMPL(c10d, CPU, m) {
m.impl("send", send_cpu);
}
TORCH_LIBRARY_IMPL(c10d, CUDA, m) {
m.impl("send", send_cuda);
}
TORCH_LIBRARY_IMPL(c10d, CPU, m) {
m.impl("recv_", recv_cpu_);
}
TORCH_LIBRARY_IMPL(c10d, CUDA, m) {
m.impl("recv_", recv_cuda_);
}
TORCH_LIBRARY_IMPL(c10d, CPU, m) {
m.impl("broadcast_", broadcast_cpu_);
}
TORCH_LIBRARY_IMPL(c10d, CUDA, m) {
m.impl("broadcast_", broadcast_cuda_);
}
TORCH_LIBRARY_IMPL(c10d, CPU, m) {
m.impl("allreduce_", allreduce_cpu_);
}
// TODO: The SparseCPU/SparseCUDA dispatched methods are only used to support
// sparse all_reduce in the Gloo backend
TORCH_LIBRARY_IMPL(c10d, SparseCPU, m) {
m.impl("allreduce_", allreduce_cpu_);
}
TORCH_LIBRARY_IMPL(c10d, SparseCUDA, m) {
m.impl("allreduce_", allreduce_cuda_);
}
TORCH_LIBRARY_IMPL(c10d, CUDA, m) {
m.impl("allreduce_", allreduce_cuda_);
}
} // namespace
} // namespace ops
} // namespace c10d
|
