Repository navigation
Expand file tree
/
Copy pathpython_function.cpp
More file actions
1971 lines (1791 loc) · 64.9 KB
/
Copy pathpython_function.cpp
File metadata and controls
1971 lines (1791 loc) · 64.9 KB
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/autograd/python_function.h>
#include <atomic>
#include <ATen/ATen.h>
#include <ATen/SequenceNumber.h>
#include <c10/util/irange.h>
#include <pybind11/pybind11.h>
#include <structmember.h>
#include <torch/csrc/PyInterpreter.h>
#include <torch/csrc/python_headers.h>
#include <torch/csrc/utils/pybind.h>
#include <ATen/FuncTorchTLS.h>
#include <ATen/functorch/DynamicLayer.h>
#include <torch/csrc/DynamicTypes.h>
#include <torch/csrc/Exceptions.h>
#include <torch/csrc/THP.h>
#include <torch/csrc/autograd/functions/accumulate_grad.h>
#include <torch/csrc/autograd/functions/basic_ops.h>
#include <torch/csrc/autograd/functions/utils.h>
#include <torch/csrc/autograd/grad_mode.h>
#include <torch/csrc/autograd/graph_task.h>
#include <torch/csrc/autograd/python_anomaly_mode.h>
#include <torch/csrc/autograd/python_cpp_function.h>
#include <torch/csrc/autograd/python_hook.h>
#include <torch/csrc/autograd/saved_variable.h>
#include <torch/csrc/autograd/utils/wrap_outputs.h>
#include <torch/csrc/dynamo/compiled_autograd.h>
#include <torch/csrc/jit/frontend/tracer.h>
#include <torch/csrc/jit/ir/ir.h>
#include <torch/csrc/jit/python/pybind_utils.h>
#include <torch/csrc/jit/python/python_tracer.h>
#include <torch/csrc/profiler/api.h>
#include <torch/csrc/utils/python_numbers.h>
#include <torch/csrc/utils/python_strings.h>
#include <torch/csrc/utils/tensor_dtypes.h>
#include <torch/csrc/autograd/function.h>
#include <functional>
#include <memory>
#include <stdexcept>
#include <string>
#include <unordered_map>
#include <unordered_set>
#include <utility>
#include <vector>
using namespace torch;
using namespace torch::autograd;
using at::Tensor;
PyObject* THPFunctionClass = nullptr;
PyObject* THPGradientEdgeClass = nullptr;
#define THPFunction_assert(condition, ...) \
if (!(condition)) { \
THPUtils_setError(__VA_ARGS__); \
throw python_error(); \
}
// Anonymous namespace for helpful functions used in this file
namespace {
inline void check_legacy_fn_attr_access(
const std::shared_ptr<torch::autograd::Node>& cdata,
const char* attr) {
TORCH_CHECK(
cdata,
"Attribute '",
attr,
"' is invalid for this instance of _C._FunctionBase. "
"Accessing this attribute directly on an instance of autograd.Function "
"is a legacy access pattern that is no longer supported. For examples "
"on how to use new‑style autograd functions, see "
"https://pytorch.org/docs/stable/autograd.html#torch.autograd.Function ");
}
// TODO: We shouldn't need to call this function because the engine
// can already persist the errors for us. This still seems to be
// needed for the DistEngine however.
//
// python test/distributed/rpc/test_tensorpipe_agent.py -k
// test_backward_autograd_engine_error
//
// See Note [ Persisting PyErr state across autograd engine threads ]
void throw_python_error() {
python_error err;
err.persist();
throw std::move(err);
}
static PyObject* unpack_saved_variables(
THPFunction* self,
const std::function<PyObject*(const Variable&)>& unpack_fn) {
HANDLE_TH_ERRORS
TORCH_CHECK(!self->has_freed_buffers, ERR_BACKWARD_TWICE);
auto& saved_variables = self->saved_variables;
if (saved_variables.empty())
return PyTuple_New(0);
auto num_saved = saved_variables.size();
THPObjectPtr saved(PyTuple_New(static_cast<Py_ssize_t>(num_saved)));
if (!saved)
return nullptr;
auto saved_for = self->cdata.lock();
// This is really a true assert, because we've already tested for the
// self->has_freed_buffers case at the beginning of this function:
// buffers are freed when PyNode dies; if the buffers are not freed,
// PyNode must be live. (Note that the buffers could be freed
// even though the PyNode is live, but that doesn't matter here
// because we will never hit this line of code if the buffers are freed--
// and in any case saved_for will be non-NULL.)
TORCH_INTERNAL_ASSERT(saved_for);
for (const auto i : c10::irange(num_saved)) {
auto unpacked_var = saved_variables[i].unpack(saved_for);
THPObjectPtr value;
if (!unpacked_var.defined()) {
Py_INCREF(Py_None);
value = Py_None;
} else {
value = unpack_fn(unpacked_var);
}
PyTuple_SET_ITEM(saved.get(), i, value.release());
}
return saved.release();
END_HANDLE_TH_ERRORS
}
PyObject* to_py_size(const std::vector<c10::SymInt>& size) {
c10::SymIntArrayRef sym_sizes(size);
auto ret = THPObjectPtr(THPSizeType.tp_alloc(
&THPSizeType, static_cast<Py_ssize_t>(sym_sizes.size())));
if (!ret)
throw python_error();
for (auto i : c10::irange(sym_sizes.size())) {
auto symint = sym_sizes[i];
if (auto maybe_int = symint.maybe_as_int(); maybe_int.has_value()) {
PyTuple_SET_ITEM(ret.get(), i, THPUtils_packInt64(*maybe_int));
} else {
auto py_symint = py::cast(symint).release().ptr();
PyTuple_SET_ITEM(ret.get(), i, py_symint);
}
}
return ret.release();
}
} // namespace
namespace torch::autograd {
// NOTE: this function is written in a way that assumes it's only called for
// backward; it's used by engine.cpp. This is responsible for forwarding a call
// from C++'s Node::apply to a Python method "apply".
// NOLINTNEXTLINE(*-rvalue-reference*)
auto PyNode::apply(variable_list&& inputs) -> variable_list {
pybind11::gil_scoped_acquire gil;
at::OptionalDeviceGuard _device_guard;
THPFunction* py_fn = (THPFunction*)obj;
// Massage a C++ variable_list into a Python arguments tuple
THPObjectPtr pyInputs(to_py_args(inputs, &_device_guard));
inputs.clear();
THPObjectPtr r;
if (py_fn->boxed_grads_call) {
// Move grad tensors from the immutable args tuple into a plain list
// and call apply_boxed instead of apply. This lets backward pop/clear
// individual grads to free memory mid-execution, because the mutable
// list (not the C++ tuple) is the only container holding grad refs.
auto num_inputs = PyTuple_GET_SIZE(pyInputs.get());
THPObjectPtr gradsList(PyList_New(num_inputs));
if (!gradsList)
throw_python_error();
for (Py_ssize_t i = 0; i < num_inputs; i++) {
PyObject* item = PyTuple_GET_ITEM(pyInputs.get(), i);
Py_INCREF(item);
PyList_SET_ITEM(gradsList.get(), i, item);
}
// Release the tuple so its refs to individual grads are dropped
pyInputs = nullptr;
THPObjectPtr boxedArgs(PyTuple_New(1));
if (!boxedArgs)
throw_python_error();
PyTuple_SET_ITEM(boxedArgs.get(), 0, gradsList.release());
THPObjectPtr apply_fn(PyObject_GetAttrString(obj, "apply_boxed"));
if (!apply_fn)
throw_python_error();
r = THPObjectPtr(PyObject_CallObject(apply_fn, boxedArgs.get()));
} else {
THPObjectPtr apply_fn(PyObject_GetAttrString(obj, "apply"));
if (!apply_fn)
throw_python_error();
r = THPObjectPtr(PyObject_CallObject(apply_fn, pyInputs.get()));
}
pyInputs = nullptr;
if (!r)
throw_python_error();
ensure_tuple(r);
auto& is_variable_input = py_fn->is_variable_input;
auto num_outputs = PyTuple_GET_SIZE(r.get());
auto num_forward_inputs = static_cast<Py_ssize_t>(is_variable_input.size());
// Returning too many results is ok, but only as long as they're all None.
// Truncate the result tuple in that case.
if (num_outputs > num_forward_inputs) {
bool all_none = true;
for (const auto i : c10::irange(num_forward_inputs, num_outputs)) {
all_none &= Py_IsNone(PyTuple_GET_ITEM(r.get(), i));
}
if (all_none) {
num_outputs = num_forward_inputs;
r = PyTuple_GetSlice(r.get(), 0, num_forward_inputs);
if (!r)
throw_python_error();
}
}
// Now the number of gradients should match
TORCH_CHECK(
num_outputs == num_forward_inputs,
"function ",
name(),
" returned an incorrect number of gradients (expected ",
num_forward_inputs,
", got ",
num_outputs,
")");
// Massage the Python results tuple back into a C++ variable_list
return to_variable_list(r.get(), is_variable_input);
}
auto PyNode::apply_with_saved_impl(
const variable_list& inputs,
const SwapSavedVariables& saved) -> variable_list {
pybind11::gil_scoped_acquire gil;
at::OptionalDeviceGuard _device_guard;
THPFunction* py_fn = (THPFunction*)obj;
// Massage a C++ variable_list into a Python arguments tuple
THPObjectPtr pyInputs(to_py_args(inputs, &_device_guard));
const auto& is_variable_input = py_fn->is_variable_input;
const auto& input_infos = py_fn->input_info;
// input_info only contains info from variable inputs and should be a subset
TORCH_INTERNAL_ASSERT(is_variable_input.size() >= input_infos.size());
// The gradients returned in the backwards need to match the number of inputs
// to the forward, and their metadata, so we pass the fwdInputs
THPObjectPtr fwdInputMetadatas(
PyTuple_New(static_cast<Py_ssize_t>(is_variable_input.size())));
if (!fwdInputMetadatas)
throw python_error();
int offset = 0;
for (const auto i : c10::irange(is_variable_input.size())) {
if (!is_variable_input[i]) {
// input at i is not a variable, skip index
PyTuple_SET_ITEM(fwdInputMetadatas.get(), i, Py_None);
offset++;
continue;
}
const auto& input_info = input_infos[i - offset];
PyObject* device(THPDevice_New(input_info.device));
if (!device)
throw_python_error();
// Metadata is a tuple of 4 elements: (layout, device, dtype, size)
PyObject* fwdInputMetadata = PyTuple_Pack(
4,
autograd::utils::wrap(input_info.layout),
device,
autograd::utils::wrap(input_info.scalar_type),
to_py_size(input_info.size));
if (!fwdInputMetadata)
throw python_error();
PyTuple_SET_ITEM(fwdInputMetadatas.get(), i, fwdInputMetadata);
}
THPObjectPtr saved_tensors(unpack_saved_variables(
py_fn, [](const Variable& var) { return THPVariable_Wrap(var); }));
auto [bwd_idx, maybe_bwd_state_idx, opaque_obj_indices] =
saved.retrieve_pynode_objs(this);
PyObject* backward_state_idx = Py_None;
if (maybe_bwd_state_idx.has_value()) {
backward_state_idx = THPUtils_packUInt64(maybe_bwd_state_idx.value());
// this might be simplifiable now that we no longer inline
Py_CLEAR(py_fn->compiled_autograd_backward_state);
}
THPObjectPtr opaque_indices_list(
PyList_New(static_cast<Py_ssize_t>(opaque_obj_indices.size())));
if (!opaque_indices_list) {
throw_python_error();
}
for (size_t i = 0; i < opaque_obj_indices.size(); i += 1) {
PyList_SET_ITEM(
opaque_indices_list.get(),
i,
THPUtils_packUInt64(opaque_obj_indices[i]));
}
THPObjectPtr r(PyObject_CallMethod(
saved.get_py_compiler(),
"proxy_call_backward",
"OOOiOOO",
pyInputs.get(),
fwdInputMetadatas.get(),
saved_tensors.get(),
bwd_idx,
obj,
backward_state_idx,
opaque_indices_list.get()));
if (!r)
throw_python_error();
ensure_tuple(r);
// Massage the Python results tuple back into a C++ variable_list
return to_variable_list(r.get(), is_variable_input);
}
auto PyNode::is_traceable() -> bool {
pybind11::gil_scoped_acquire gil;
THPObjectPtr forward_class{PyObject_GetAttrString(obj, "_forward_cls")};
if (!forward_class)
throw_python_error();
THPObjectPtr traceable_py_bool{
PyObject_GetAttrString(forward_class, "is_traceable")};
if (!traceable_py_bool)
throw_python_error();
return Py_IsTrue(traceable_py_bool);
}
auto PyNode::release_variables() -> void {
// This function is called as part of the Node destructor!
// Since this object might be kept alive by C++, it is possible
// that the python interpreter is already dead here. In that case
// we just leak the saved objects.
if (Py_IsInitialized()) {
pybind11::gil_scoped_acquire gil;
auto f = (THPFunction*)obj;
f->saved_variables.clear();
f->has_freed_buffers = 1;
}
}
auto PyNode::name() const -> std::string {
pybind11::gil_scoped_acquire gil;
auto f = (THPFunction*)obj;
auto name = std::string(Py_TYPE(f)->tp_name);
return name;
}
bool PyNode::is_aot_backward() const {
py::handle handle(obj);
return py::hasattr(py::getattr(handle, "_forward_cls"), "_aot_id");
}
void PyNode::compiled_args(CompiledNodeArgs& args) const {
static PyObject* method_name =
PyUnicode_InternFromString("_compiled_autograd_key");
THPObjectPtr pykey(PyObject_CallMethodObjArgs(obj, method_name, nullptr));
if (!pykey)
throw_python_error();
TORCH_CHECK(
PyTuple_CheckExact(pykey.get()),
"_compiled_autograd_key should return tuple of ints");
auto size = PyTuple_GET_SIZE(pykey.get());
TORCH_INTERNAL_ASSERT(size > 0);
// first value is unique id managed by AUTOGRAD_FUNCTION_COUNTER
auto key = PyLong_AsSsize_t(PyTuple_GET_ITEM(pykey.get(), 0));
if (C10_UNLIKELY(key < 0)) {
TORCH_CHECK(PyErr_Occurred(), "key must be positive");
throw_python_error();
}
args.collect_size(static_cast<size_t>(key));
args.collect_size(static_cast<size_t>(size));
auto f = (THPFunction*)obj;
f->compiled_autograd_symints.clear();
f->compiled_autograd_symints.reserve(size - 1);
for (const auto i : c10::irange(1, size)) {
auto val = PyLong_AsSsize_t(PyTuple_GET_ITEM(pykey.get(), i));
if (C10_UNLIKELY(val == -1 && PyErr_Occurred()))
throw_python_error();
f->compiled_autograd_symints.emplace_back(val);
}
// AotAutograd symints are all dynamic
auto prior =
args.set_default_dyn_type(torch::dynamo::autograd::SizeInput::DYNAMIC);
args.collect(f->compiled_autograd_symints);
args.set_default_dyn_type(prior);
args.collect(f->saved_variables, true); // always unpacked as output in eager
args.collect(f->materialize_grads);
args.collect(f->is_variable_input);
args.collect(f->needs_input_grad);
args.collect(f->materialize_non_diff_grads);
args.collect(f->output_info);
args.collect(f->input_info);
Py_INCREF(obj);
c10::SafePyObject backward_obj(obj, getPyInterpreter());
std::optional<c10::SafePyObject> backward_state_obj;
PyObject* bw_state = f->compiled_autograd_backward_state;
if (args.cond(bw_state != nullptr)) {
Py_INCREF(bw_state);
backward_state_obj = c10::SafePyObject(bw_state, getPyInterpreter());
}
std::vector<c10::SafePyObject> opaque_objs;
if (THPObjectPtr opaque_objs_ptr =
THPObjectPtr(PyObject_GetAttrString(obj, "opaque_objects"))) {
Py_ssize_t size = PySequence_Size(opaque_objs_ptr.get());
if (size > 0) {
opaque_objs.reserve(static_cast<size_t>(size));
for (Py_ssize_t i = 0; i < size; i += 1) {
opaque_objs.emplace_back(c10::SafePyObject(
PySequence_GetItem(opaque_objs_ptr.get(), i), getPyInterpreter()));
}
} else if (size < 0) {
PyErr_Clear();
}
} else {
PyErr_Clear();
}
args.collect_pynode_objs(
this,
std::move(backward_obj),
std::move(backward_state_obj),
std::move(opaque_objs));
}
variable_list PyNode::apply_with_saved(
const variable_list& inputs,
SwapSavedVariables& saved) {
auto f = (THPFunction*)obj;
saved.before(f->compiled_autograd_symints);
saved.before(f->saved_variables);
saved.before(f->needs_input_grad);
saved.before(f->materialize_non_diff_grads);
saved.before(f->output_info);
saved.before(f->input_info);
variable_list result = apply_with_saved_impl(variable_list(inputs), saved);
saved.after(f->compiled_autograd_symints);
saved.after(f->saved_variables);
saved.after(f->needs_input_grad);
saved.after(f->materialize_non_diff_grads);
saved.after(f->output_info);
saved.after(f->input_info);
return result;
}
PyObject* PyNode::to_py_args(
const variable_list& inputs,
at::OptionalDeviceGuard* device_guard) {
THPFunction* py_fn = (THPFunction*)obj;
auto zeros_without_gil = [](const VariableInfo& variable,
at::OptionalDeviceGuard& dg) {
pybind11::gil_scoped_release gil;
return variable.zeros(dg);
};
auto num_inputs = inputs.size();
PyObject* pyInputs = PyTuple_New(static_cast<Py_ssize_t>(num_inputs));
if (!pyInputs)
throw_python_error();
auto& output_info = py_fn->output_info;
for (const auto i : c10::irange(num_inputs)) {
PyObject* input = nullptr;
if (inputs[i].defined() || !py_fn->materialize_grads ||
(input_metadata(i).was_default_constructed() &&
!py_fn->materialize_non_diff_grads)) {
input = THPVariable_Wrap(inputs[i]);
} else {
input =
THPVariable_Wrap(zeros_without_gil(output_info[i], *device_guard));
}
if (!input)
throw_python_error();
PyTuple_SET_ITEM(pyInputs, i, input);
}
return pyInputs;
}
variable_list PyNode::to_variable_list(
const PyObject* outputs,
const std::vector<bool>& is_variable_input) {
auto num_outputs = PyTuple_GET_SIZE(outputs);
variable_list results;
results.reserve(num_outputs);
for (int i = 0; i != num_outputs; ++i) {
PyObject* output = PyTuple_GET_ITEM(outputs, i);
bool was_variable = is_variable_input[i];
if (!was_variable) {
TORCH_CHECK(
Py_IsNone(output),
"function ",
name(),
" returned a gradient different than None at position ",
i + 1,
", but the corresponding forward input was not a Variable");
continue;
}
if (Py_IsNone(output)) {
results.emplace_back();
} else {
TORCH_CHECK(
THPVariable_Check(output),
"expected Variable or None (got ",
THPUtils_typename(output),
")");
results.emplace_back(THPVariable_Unpack(output));
}
}
return results;
}
} // namespace torch::autograd
// Traverse and clear are required for supporting Python's GC cycle handling.
static int THPFunction_traverse(THPFunction* self, visitproc visit, void* arg) {
// NB: We should not traverse PyObbject stored on PyNode, since we only hold
// as weak reference to the PyNode.
Py_VISIT(self->to_save);
Py_VISIT(self->non_differentiable);
Py_VISIT(self->dirty_tensors);
Py_VISIT(self->compiled_autograd_backward_state);
Py_VISIT(self->saved_for_forward);
return 0;
}
static int THPFunction_clear(THPFunction* self) {
// Note that the cdata might not be expired yet in the case where this
// object is part of a cycle and the GC happens to tp_clear this PyObject
// before the other ones that trigger the de-allocation of the cdata
Py_CLEAR(self->needs_input_grad);
Py_CLEAR(self->to_save);
Py_CLEAR(self->non_differentiable);
Py_CLEAR(self->dirty_tensors);
Py_CLEAR(self->compiled_autograd_backward_state);
Py_CLEAR(self->saved_for_forward);
self->output_info.clear();
self->input_info.clear();
self->saved_variables.clear();
self->is_variable_input.clear();
return 0;
}
static void THPFunction_dealloc(THPFunction* self) {
// Why is this guaranteed to be true? Suppose that self->cdata is non-null
// (otherwise the condition is trivially true). Then there is a PyNode
// which contains an owning reference to this object. But we are only
// allowed to clear if all owning references are gone! Contradiction.
//
// However, note that THPFunction_clear is typically called in the shared_ptr
// destructor of PyNode; in that case, per
// https://cplusplus.github.io/LWG/lwg-active.html#2751 it's not currently
// specified in the standard that this is guaranteed. If you see this
// assert triggering in the wild, feel free to comment it out. They're
// likely to standardize that you ARE guaranteed to see the weak pointers
// as expired in the destructor in the future, so we'll keep this for now.
TORCH_INTERNAL_ASSERT(self->cdata.expired());
PyObject_GC_UnTrack(self);
THPFunction_clear(self);
self->cdata.~weak_ptr<PyNode>();
self->output_info.~vector();
self->input_info.~vector();
self->saved_variables.~vector();
self->is_variable_input.~vector();
Py_TYPE(self)->tp_free((PyObject*)self);
}
static PyObject* THPFunction_new(
PyTypeObject* type,
PyObject* args,
PyObject* kwargs) {
PyObject* obj = type->tp_alloc(type, 0);
if (!obj)
return nullptr;
// Python zero-initializes the object memory, so there's no need to initialize
// most fields
THPFunction* self = (THPFunction*)obj;
// Setup the PyNode later; we can't keep it live here
new (&self->cdata) std::weak_ptr<PyNode>();
new (&self->output_info) std::vector<VariableInfo>();
new (&self->input_info) std::vector<VariableInfo>();
new (&self->saved_variables) std::vector<SavedVariable>();
new (&self->is_variable_input) std::vector<bool>();
self->materialize_grads = true;
self->pure_view = false;
self->materialize_non_diff_grads = true;
self->clear_saved_tensors_on_access = false;
self->saved_tensors_accessed_and_cleared = false;
return obj;
}
////////////////////////////////////////////////////////////////////////////////
// Forward
////////////////////////////////////////////////////////////////////////////////
// Bump the counters of all recorded dirty input tensors, adding each of them
// into dirty_inputs. Also does some sanity checking.
static std::unordered_set<at::TensorImpl*> _mark_dirty(THPFunction* self) {
// Increase versions of modified tensors
std::unordered_set<at::TensorImpl*> dirty_inputs;
if (!self->dirty_tensors)
return dirty_inputs;
THPFunction_assert(
PyTuple_Check(self->dirty_tensors),
"autograd "
"internal error: dirty_tensors attribute is expected to be a tuple "
"but is ",
THPUtils_typename(self->dirty_tensors));
Py_ssize_t num_dirty = PyTuple_GET_SIZE(self->dirty_tensors);
dirty_inputs.reserve(num_dirty);
for (const auto i : c10::irange(num_dirty)) {
PyObject* obj = PyTuple_GET_ITEM(self->dirty_tensors, i);
THPFunction_assert(
THPVariable_Check(obj),
"mark_dirty can "
"only accept variables, but argument ",
i,
" is of type ",
THPUtils_typename(obj));
const auto& tensor = THPVariable_Unpack(obj);
dirty_inputs.insert(tensor.unsafeGetTensorImpl());
torch::autograd::impl::bump_version(tensor);
}
// We're not going to ever need this so let's remove references now
Py_CLEAR(self->dirty_tensors);
return dirty_inputs;
}
static std::unordered_set<at::TensorImpl*> _parse_non_differentiable(
THPFunction* self);
// Given a Python tuple of raw output tensors (raw_output), set each of
// the corresponding entries in a different Python tuple (outputs) with
// these tensors wrapped with variables. We save the gradient function (self)
// to the variable if the output requires grad.
//
// There is a considerable amount of complexity to handle if the operation
// that produced these output tensors is inplace. A mapping of *input*
// tensors to variables (t2var) is used to test if this occurred, and
// the set of dirty tensors (dirty_inputs) is used to figure out what to
// do in this case. After this method is run, t2var is extended with
// mappings for output tensors as well.
static void _wrap_outputs(
const std::shared_ptr<PyNode>& cdata,
THPFunction* self,
const variable_list& input_vars,
PyObject* raw_output,
PyObject* outputs,
bool is_executable,
const std::unordered_set<at::TensorImpl*>& to_save_if_setup_context) {
auto cdata_if_executable = is_executable ? cdata : nullptr;
Py_ssize_t num_outputs = PyTuple_GET_SIZE(raw_output);
if (is_executable) {
self->output_info.clear();
self->output_info.reserve(num_outputs);
}
auto non_differentiable = _parse_non_differentiable(self);
auto dirty_inputs = _mark_dirty(self);
std::vector<std::optional<Variable>> raw_output_vars;
raw_output_vars.reserve(num_outputs);
for (const auto i : c10::irange(num_outputs)) {
PyObject* obj = PyTuple_GET_ITEM(raw_output, i);
// Only process tensors as outputs for autograd purposes.
if (THPVariable_Check(obj)) {
raw_output_vars.emplace_back(THPVariable_Unpack(obj));
} else {
raw_output_vars.emplace_back();
}
}
_jvp_fn_t jvp_user_function = [self](
variable_list inputs,
variable_list grad_inputs) {
pybind11::gil_scoped_acquire gil;
// Massage a C++ variable_list into a Python arguments tuple
// Making sure to introduce the proper None for non-Tensor inputs
auto num_inputs = self->is_variable_input.size();
THPObjectPtr pyInputs(PyTuple_New(static_cast<Py_ssize_t>(num_inputs)));
if (!pyInputs)
throw_python_error();
int64_t variable_idx = 0;
for (const auto i : c10::irange(num_inputs)) {
PyObject* input = nullptr;
if (self->is_variable_input[i]) {
if (grad_inputs[variable_idx].defined() || !self->materialize_grads ||
!isDifferentiableType(inputs[variable_idx].scalar_type())) {
input = THPVariable_Wrap(grad_inputs[variable_idx]);
} else {
input = THPVariable_Wrap(at::zeros_like(inputs[variable_idx]));
}
if (!input) {
throw_python_error();
}
variable_idx++;
} else {
Py_INCREF(Py_None);
input = Py_None;
}
PyTuple_SET_ITEM(pyInputs.get(), i, input);
}
THPObjectPtr apply_jvp_fn(
PyObject_GetAttrString((PyObject*)self, "apply_jvp"));
if (!apply_jvp_fn)
throw_python_error();
THPObjectPtr r(PyObject_CallObject(apply_jvp_fn, pyInputs.get()));
if (!r)
throw_python_error();
ensure_tuple(r);
// Massage the Python results tuple back into a C++ variable_list
// Don't do any check on the number of results here as
// it is handled by the caller
const int num_outputs = PyTuple_GET_SIZE(r.get());
variable_list results;
results.reserve(num_outputs);
for (const auto i : c10::irange(num_outputs)) {
PyObject* output = PyTuple_GET_ITEM(r.get(), i);
if (Py_IsNone(output)) {
results.emplace_back();
} else {
TORCH_CHECK(
THPVariable_Check(output),
"expected Variable or None (got ",
THPUtils_typename(output),
") for grad output ",
i,
".")
results.emplace_back(THPVariable_Unpack(output));
}
}
return results;
};
auto view_as_self_fn = [](const at::Tensor& x) -> at::Tensor {
pybind11::gil_scoped_acquire gil;
THPObjectPtr py_x(THPVariable_Wrap(x));
THPObjectPtr py_view_as_method(PyObject_GetAttrString(py_x, "view_as"));
if (!py_view_as_method)
throw python_error();
THPObjectPtr args(PyTuple_Pack(1, py_x.get()));
if (!args)
throw python_error();
THPObjectPtr result(PyObject_CallObject(py_view_as_method, args));
if (!result)
throw python_error();
return THPVariable_Unpack(result);
};
// Wrap only the tensor outputs.
auto wrapped_outputs = _wrap_outputs(
input_vars,
non_differentiable,
dirty_inputs,
raw_output_vars,
cdata_if_executable,
jvp_user_function,
to_save_if_setup_context,
view_as_self_fn,
self->pure_view);
for (const auto i : c10::irange(num_outputs)) {
PyObject* obj = PyTuple_GetItem(raw_output, i);
const auto& wrapped_output = wrapped_outputs[i];
// Keep the non-tensor outputs as is.
if (!THPVariable_Check(obj) || !wrapped_output.has_value()) {
if (is_executable) {
self->output_info.emplace_back();
}
Py_INCREF(obj);
PyTuple_SetItem(outputs, i, obj);
} else {
if (is_executable) {
// If one of the grad outputs is undefined, a correctly-shaped zeros
// should be used instead. To construct these for NJT, zeros_like() must
// be used until we have factory function support.
bool is_differentiable =
(non_differentiable.count(wrapped_output->unsafeGetTensorImpl()) ==
0 &&
isDifferentiableType(wrapped_output->scalar_type()));
bool use_zeros_like =
is_differentiable && num_outputs > 1 && wrapped_output->is_nested();
self->output_info.emplace_back(wrapped_output.value(), use_zeros_like);
}
PyTuple_SetItem(outputs, i, THPVariable_Wrap(wrapped_output.value()));
}
}
}
static void _get_tensors_to_save(
THPFunction* self,
std::unordered_set<at::TensorImpl*>& to_save_if_setup_context,
std::vector<std::optional<at::Tensor>>& tensors_to_save,
bool overridden_setup_context,
bool is_executable) {
if (self->saved_for_forward && overridden_setup_context) {
// We look at saved_for_forward here purely for the purpose of populating
// to_save_if_setup_context, the actual saving is not done here.
THPFunction_assert(
PyTuple_Check(self->saved_for_forward),
"autograd internal "
"error: saved_for_forward attribute is expected to be a tuple but is ",
THPUtils_typename(self->saved_for_forward));
Py_ssize_t num_saved_for_forward =
PyTuple_GET_SIZE(self->saved_for_forward);
for (const auto i : c10::irange(num_saved_for_forward)) {
PyObject* obj = PyTuple_GET_ITEM(self->saved_for_forward, i);
if (THPVariable_Check(obj)) {
const auto& tensor = THPVariable_Unpack(obj);
to_save_if_setup_context.insert(tensor.unsafeGetTensorImpl());
}
}
}
if (self->to_save) {
THPFunction_assert(
PyTuple_Check(self->to_save),
"autograd internal "
"error: to_save attribute is expected to be a tuple but is ",
THPUtils_typename(self->to_save));
Py_ssize_t num_saved = PyTuple_GET_SIZE(self->to_save);
for (const auto i : c10::irange(num_saved)) {
PyObject* obj = PyTuple_GET_ITEM(self->to_save, i);
if (Py_IsNone(obj)) {
tensors_to_save.emplace_back(std::nullopt);
continue;
} else if (THPVariable_Check(obj)) {
const auto& tensor = THPVariable_Unpack(obj);
if (overridden_setup_context) {
to_save_if_setup_context.insert(tensor.unsafeGetTensorImpl());
}
if (is_executable) {
tensors_to_save.emplace_back(tensor);
}
} else {
if (is_executable) {
// TODO: We should really just ALWAYS throw an error here, but
// doing so will break some internal tests. We should fix those.
TORCH_CHECK_TYPE(
false,
fmt::format(
"save_for_backward can only save variables, but argument {} is of "
"type {}",
i,
Py_TYPE(obj)->tp_name));
}
}
}
Py_CLEAR(self->to_save);
}
}
// Save any variables that requested by to_save
static void _save_variables(
const std::vector<std::optional<at::Tensor>>& tensors_to_save,
const std::shared_ptr<PyNode>& cdata_ptr,
THPFunction* self,
PyObject* outputs,
int64_t num_outputs) {
if (tensors_to_save.empty())
return;
size_t num_saved = tensors_to_save.size();
self->saved_variables.clear();
self->saved_variables.reserve(num_saved);
std::unordered_set<at::TensorImpl*> output_impls{};
output_impls.reserve(num_outputs);
for (const auto i : c10::irange(num_outputs)) {
PyObject* obj = PyTuple_GET_ITEM(outputs, i);
if (THPVariable_Check(obj)) {
const auto& tensor = THPVariable_Unpack(obj);
output_impls.insert(tensor.unsafeGetTensorImpl());
}
}
for (const auto& opt_tensor : tensors_to_save) {
if (!opt_tensor.has_value()) {
self->saved_variables.emplace_back();
} else {
bool is_output =
output_impls.count(opt_tensor.value().unsafeGetTensorImpl()) > 0;
self->saved_variables.emplace_back(opt_tensor.value(), is_output);
}
}
}
// Mark requires_grad = 0 on non-differentiable variables (as per
// non_differentiable)
static std::unordered_set<at::TensorImpl*> _parse_non_differentiable(
THPFunction* self) {
std::unordered_set<at::TensorImpl*> set;
if (!self->non_differentiable)
return set;
THPFunction_assert(
PyTuple_Check(self->non_differentiable),
"autograd "
"internal error: non_differentiable attribute is expected to be a "
"tuple but is ",
THPUtils_typename(self->non_differentiable));
Py_ssize_t num_nondiff = PyTuple_GET_SIZE(self->non_differentiable);
set.reserve(num_nondiff);
for (const auto i : c10::irange(num_nondiff)) {
PyObject* t = PyTuple_GET_ITEM(self->non_differentiable, i);
THPFunction_assert(
THPVariable_Check(t),
"mark_non_differentiable "
"only accepts variable arguments, but got ",
THPUtils_typename(t));
set.insert(THPVariable_Unpack(t).unsafeGetTensorImpl());
}
Py_CLEAR(self->non_differentiable);
return set;
}
struct UnpackedInput {
THPObjectPtr input_tuple;
variable_list input_vars;
// record_function_inputs is for RECORD_FUNCTION only
std::vector<c10::IValue> record_function_inputs;
};
struct InputFlags {
bool is_executable = false;
edge_list next_edges;
THPObjectPtr needs_input_grad;
std::vector<bool> is_variable_input;
};
namespace {
template <bool enforce_variables>
std::pair<UnpackedInput, InputFlags> unpack_input(PyObject* args) {
UnpackedInput unpacked;
InputFlags flags;
auto num_args = PyTuple_GET_SIZE(args);
unpacked.input_tuple = PyTuple_New(num_args);
flags.needs_input_grad = PyTuple_New(num_args);
bool profiler_need_input = torch::autograd::profiler::profilerEnabled() &&
torch::autograd::profiler::getProfilerConfig().report_input_shapes;
for (const auto i : c10::irange(num_args)) {
PyObject* arg = PyTuple_GET_ITEM(args, i);
bool is_variable = THPVariable_Check(arg);
flags.is_variable_input.push_back(is_variable);
if (!is_variable) {
// TODO: remove this code path once Variable and Tensor are merged in
// Python
if (enforce_variables) {
THPUtils_setError(
"expected a Tensor argument, but got ", THPUtils_typename(arg));
throw python_error();
}
Py_INCREF(Py_False);
PyTuple_SET_ITEM(flags.needs_input_grad.get(), i, Py_False);
if (profiler_need_input) {
// The following conversion from PyObject to IValue is expensive
// Only do it if profiler is enabled and needs input shapes
auto match = torch::jit::tryToInferPrimitiveType(arg);
if (match.success()) {
unpacked.record_function_inputs.push_back(
torch::jit::toIValue(arg, match.type()));
}
}
} else {
const auto& tensor = THPVariable_Unpack(arg);
unpacked.input_vars.push_back(tensor);