File: allgather_ops.h

package info (click to toggle)
pytorch 1.13.1%2Bdfsg-4
  • links: PTS, VCS
  • area: main
  • in suites: bookworm
  • size: 139,252 kB
  • sloc: cpp: 1,100,274; python: 706,454; ansic: 83,052; asm: 7,618; java: 3,273; sh: 2,841; javascript: 612; makefile: 323; xml: 269; ruby: 185; yacc: 144; objc: 68; lex: 44
file content (130 lines) | stat: -rw-r--r-- 3,833 bytes parent folder | download
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
/**
 * Copyright (c) 2017-present, Facebook, Inc.
 *
 * 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.
 */

#pragma once

#include <algorithm>

#include "caffe2/contrib/gloo/common.h"
#include "caffe2/core/operator.h"
#include "caffe2/core/types.h"

#include <gloo/algorithm.h>
#include <gloo/common/error.h>
#include <gloo/context.h>

namespace caffe2 {
namespace gloo {

template <class Context>
class AllgatherOp final : public Operator<Context> {
 public:
  USE_OPERATOR_CONTEXT_FUNCTIONS;

  AllgatherOp(const OperatorDef& operator_def, Workspace* ws)
      : Operator<Context>(operator_def, ws),
        ws_(ws),
        status_blob_(
            OperatorBase::GetSingleArgument<std::string>("status_blob", "")) {
    if (status_blob_ != "") {
      ws_->CreateBlob(status_blob_);
    }
  }

  virtual ~AllgatherOp() {}

  bool RunOnDevice() override {
    std::call_once(once_, [&] { initialize(); });

    // If any parameter has changed in between runs, the initialized
    // algorithm is invalid and cannot be used.
    update(current_);
    CAFFE_ENFORCE(current_ == init_, "Inputs/outputs have changed");

    try {
      algorithm_->run();
    } catch (::gloo::IoException& ioe) {
      LOG(ERROR) << "Caught gloo IO exception: " << ioe.what();
      if (status_blob_ != "") {
        signalFailure(ws_->GetBlob(status_blob_), ioe);
        return false;
      } else {
        throw;
      }
    }
    return true;
  }

 protected:
  void initialize() {
    // Allocate output tensor
    CAFFE_ENFORCE_EQ(OutputSize(), 1);
    auto comm_size =
        OperatorBase::Input<std::shared_ptr<::gloo::Context>>(0)->size;
    const auto dims = std::vector<int64_t>(
        1, (InputSize() - 1) * Input(1).numel() * comm_size);
    Output(0)->Resize(dims);

    // Store which inputs/outputs this instance initialized with
    update(init_);

    CAFFE_ENFORCE_EQ(init_.outputs.size(), 1);

    // Verify tensors all have same size
    size_t size = Input(1).numel();
    for (const auto i : c10::irange(2, InputSize())) {
      CAFFE_ENFORCE_EQ(Input(i).numel(), size);
    }

    // Verify tensors all have same type
    TypeMeta meta = Input(1).dtype();
    for (const auto i : c10::irange(2, InputSize())) {
      CAFFE_ENFORCE(Input(i).dtype() == meta);
    }

    // Finally initialize the algorithm
    initializeAlgorithm();
  }

  void initializeAlgorithm();

  std::once_flag once_;
  std::unique_ptr<::gloo::Algorithm> algorithm_;

  // Captures the parameters passed to Gloo when first initialized.
  // An instance is updated every time this op runs and is compared
  // to the reference instance for equality. If any parameter has
  // changed from run to run, the initialized algorithm is invalid.
  void update(GlooParameters& params) {
    params.context = OperatorBase::Input<std::shared_ptr<::gloo::Context>>(0);
    params.inputs.resize(InputSize() - 1);
    params.size = Input(1).numel();
    params.meta = Input(1).dtype();
    for (const auto i : c10::irange(params.inputs.size())) {
      params.inputs[i] = Input(i + 1).raw_data();
    }
    params.outputs.resize(OutputSize());
    params.outputs[0] = Output(0)->raw_mutable_data(params.meta);
  }

  GlooParameters init_;
  GlooParameters current_;
  Workspace* ws_;
  std::string status_blob_;
};

} // namespace gloo
} // namespace caffe2