File: PyProcessGroup.hpp

package info (click to toggle)
pytorch-cuda 2.6.0%2Bdfsg-7
  • links: PTS, VCS
  • area: contrib
  • in suites: forky, sid, trixie
  • size: 161,620 kB
  • sloc: python: 1,278,832; cpp: 900,322; ansic: 82,710; asm: 7,754; java: 3,363; sh: 2,811; javascript: 2,443; makefile: 597; ruby: 195; xml: 84; objc: 68
file content (299 lines) | stat: -rw-r--r-- 9,297 bytes parent folder | download | duplicates (3)
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
#pragma once

#include <torch/csrc/distributed/c10d/ProcessGroup.hpp>
#include <torch/csrc/jit/python/pybind_utils.h>
#include <torch/csrc/utils/pybind.h>

namespace c10d {

// PyProcessGroup is a pybind11 trampoline class to allow a Python
// class to inherit from torch.distributed.ProcessGroup
class PyProcessGroup : public ProcessGroup {
 public:
  // PyWork is a pybind11 trampoline class to allow a Python
  // class to inherit from torch.distributed.Work
  class TORCH_PYTHON_API PyWork : public Work {
   public:
    PyWork() = default;

    bool wait(std::chrono::milliseconds timeout = kNoTimeout) override {
      PYBIND11_OVERRIDE(
          bool, /* Return type */
          Work, /* Parent class */
          wait, /* Name of function in C++ */
          timeout);
    }

    c10::intrusive_ptr<c10::ivalue::Future> getFuture() override {
      // We cannot use PYBIND11_OVERRIDE because:
      // 1. We have to >MANUALLY< unwrap the PyFutureWrapper and
      // 2. The python name is get_future
      pybind11::gil_scoped_acquire gil;
      auto override =
          pybind11::get_override(static_cast<const Work*>(this), "get_future");

      if (override) {
        py::object o = override();
        auto futWrapper =
            o.cast<std::shared_ptr<torch::jit::PythonFutureWrapper>>();
        return futWrapper->fut;
      }

      return Work::getFuture();
    }

    // Take a reference of the corresponding py::object.
    // With functional collectives, ownership of work objects is generally
    // transferred to C++. For pure C++ work objects, it is sufficient to
    // transfer the ownership of work object. For user-defined work objects in
    // Python, it is necessary to keep the corresponding py::object alive in
    // addition to ensure that the user-defined methods can be executed.
    void ref_py_object() {
      py_obj_ = py::cast(this);
    }

   private:
    py::object py_obj_;
  };

  using ProcessGroup::ProcessGroup;

  const std::string getBackendName() const override {
    PYBIND11_OVERRIDE(
        std::string, /* Return type */
        ProcessGroup, /* Parent class */
        getBackendName, /* Name of function in C++ */
    );
  }

  int getRank() const override {
    PYBIND11_OVERRIDE(
        int, /* Return type */
        ProcessGroup, /* Parent class */
        getRank, /* Name of function in C++ */
    );
  }

  int getSize() const override {
    PYBIND11_OVERRIDE(
        int, /* Return type */
        ProcessGroup, /* Parent class */
        getSize, /* Name of function in C++ */
    );
  }

  const std::string& getGroupName() const override {
    PYBIND11_OVERRIDE(
        const std::string&, /* Return type */
        ProcessGroup, /* Parent class */
        getGroupName, /* Name of function in C++ */
    );
  }

  void setGroupName(const std::string& group_name) override {
    PYBIND11_OVERRIDE(
        void, /* Return type */
        ProcessGroup, /* Parent class */
        setGroupName, /* Name of function in C++ */
        group_name);
  }

  const std::string& getGroupDesc() const override {
    PYBIND11_OVERRIDE(
        const std::string&, /* Return type */
        ProcessGroup, /* Parent class */
        getGroupDesc, /* Name of function in C++ */
    );
  }

  void setGroupDesc(const std::string& group_desc) override {
    PYBIND11_OVERRIDE(
        void, /* Return type */
        ProcessGroup, /* Parent class */
        setGroupDesc, /* Name of function in C++ */
        group_desc);
  }

  c10::intrusive_ptr<Work> allgather(
      std::vector<std::vector<at::Tensor>>& outputTensors,
      std::vector<at::Tensor>& inputTensors,
      const AllgatherOptions& opts = AllgatherOptions()) override {
    PYBIND11_OVERRIDE(
        c10::intrusive_ptr<Work>, /* Return type */
        ProcessGroup, /* Parent class */
        allgather, /* Name of function in C++ */
        outputTensors,
        inputTensors,
        opts);
  }

  c10::intrusive_ptr<Work> allgather_into_tensor_coalesced(
      std::vector<at::Tensor>& outputTensors,
      std::vector<at::Tensor>& inputTensors,
      const AllgatherOptions& opts = AllgatherOptions()) override {
    PYBIND11_OVERRIDE(
        c10::intrusive_ptr<Work>, /* Return type */
        ProcessGroup, /* Parent class */
        allgather_into_tensor_coalesced, /* Name of function in C++ */
        outputTensors,
        inputTensors,
        opts);
  }

  c10::intrusive_ptr<Work> allreduce(
      std::vector<at::Tensor>& tensors,
      const AllreduceOptions& opts = AllreduceOptions()) override {
    PYBIND11_OVERRIDE(
        c10::intrusive_ptr<Work>, /* Return type */
        ProcessGroup, /* Parent class */
        allreduce, /* Name of function in C++ */
        tensors,
        opts);
  }

  c10::intrusive_ptr<Work> allreduce_coalesced(
      std::vector<at::Tensor>& tensors,
      const AllreduceCoalescedOptions& opts =
          AllreduceCoalescedOptions()) override {
    PYBIND11_OVERRIDE(
        c10::intrusive_ptr<Work>, /* Return type */
        ProcessGroup, /* Parent class */
        allreduce_coalesced, /* Name of function in C++ */
        tensors,
        opts);
  }

  c10::intrusive_ptr<Work> alltoall_base(
      at::Tensor& outputBuffer,
      at::Tensor& inputBuffer,
      std::vector<int64_t>& outputSplitSizes,
      std::vector<int64_t>& inputSplitSizes,
      const AllToAllOptions& opts = AllToAllOptions()) override {
    PYBIND11_OVERRIDE(
        c10::intrusive_ptr<Work>, /* Return type */
        ProcessGroup, /* Parent class */
        alltoall_base, /* Name of function in C++ */
        outputBuffer,
        inputBuffer,
        outputSplitSizes,
        inputSplitSizes,
        opts);
  }

  c10::intrusive_ptr<Work> barrier(
      const BarrierOptions& opts = BarrierOptions()) override {
    PYBIND11_OVERRIDE(
        c10::intrusive_ptr<Work>, /* Return type */
        ProcessGroup, /* Parent class */
        barrier, /* Name of function in C++ */
        opts);
  }

  c10::intrusive_ptr<Work> broadcast(
      std::vector<at::Tensor>& tensors,
      const BroadcastOptions& opts = BroadcastOptions()) override {
    PYBIND11_OVERRIDE(
        c10::intrusive_ptr<Work>, /* Return type */
        ProcessGroup, /* Parent class */
        broadcast, /* Name of function in C++ */
        tensors,
        opts);
  }

  c10::intrusive_ptr<Work> reduce_scatter(
      std::vector<at::Tensor>& outputTensors,
      std::vector<std::vector<at::Tensor>>& inputTensors,
      const ReduceScatterOptions& opts = ReduceScatterOptions()) override {
    PYBIND11_OVERRIDE(
        c10::intrusive_ptr<Work>, /* Return type */
        ProcessGroup, /* Parent class */
        reduce_scatter, /* Name of function in C++ */
        outputTensors,
        inputTensors,
        opts);
  }

  c10::intrusive_ptr<Work> reduce_scatter_tensor_coalesced(
      std::vector<at::Tensor>& outputTensors,
      std::vector<at::Tensor>& inputTensors,
      const ReduceScatterOptions& opts = ReduceScatterOptions()) override {
    PYBIND11_OVERRIDE(
        c10::intrusive_ptr<Work>, /* Return type */
        ProcessGroup, /* Parent class */
        reduce_scatter_tensor_coalesced, /* Name of function in C++ */
        outputTensors,
        inputTensors,
        opts);
  }

  c10::intrusive_ptr<Work> send(
      std::vector<at::Tensor>& tensors,
      int dstRank,
      int tag) override {
    PYBIND11_OVERRIDE(
        c10::intrusive_ptr<Work>, /* Return type */
        ProcessGroup, /* Parent class */
        send, /* Name of function in C++ */
        tensors,
        dstRank,
        tag);
  }

  c10::intrusive_ptr<Work> recv(
      std::vector<at::Tensor>& tensors,
      int srcRank,
      int tag) override {
    PYBIND11_OVERRIDE(
        c10::intrusive_ptr<Work>, /* Return type */
        ProcessGroup, /* Parent class */
        recv, /* Name of function in C++ */
        tensors,
        srcRank,
        tag);
  }
};

class TORCH_PYTHON_API PythonOnCompletionHook {
 public:
  // Wraps a py::object hook and acquires Python GIL in dtor before
  // destructing the hook object.
  PythonOnCompletionHook(py::object hook) : hook_(std::move(hook)) {}
  PythonOnCompletionHook(const PythonOnCompletionHook&) = default;

  // NOLINTNEXTLINE(bugprone-exception-escape)
  ~PythonOnCompletionHook() {
    py::gil_scoped_acquire ag;
    hook_.dec_ref();
    // Explicitly set hook_ to nullptr to prevent py::object's dtor
    // to decref on the PyObject again.
    // See Note [Destructing py::object] in python_ivalue.h
    hook_.ptr() = nullptr;
  }

  void operator()(const std::shared_ptr<WorkInfo>& workInfo) const {
    std::exception_ptr eptr;
    {
      py::gil_scoped_acquire acquire;
      try {
        hook_(workInfo);
      } catch (py::error_already_set& e) {
        // py::error_already_set requires GIL to destruct, take
        // special care.
        eptr = std::make_exception_ptr(std::runtime_error(e.what()));
        e.restore();
        PyErr_Clear();
      } catch (std::exception& e) {
        eptr = std::current_exception();
      }
    }
    // No more Python-related stuff at this point, i.e., this
    // exception can be captured and handled by PG backend.
    if (eptr)
      std::rethrow_exception(eptr);
  }

 private:
  py::object hook_;
};

} // namespace c10d