Repository navigation
Expand file tree
/
Copy pathModule.cpp
More file actions
3043 lines (2788 loc) · 94 KB
/
Copy pathModule.cpp
File metadata and controls
3043 lines (2788 loc) · 94 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 <ATen/DeviceAccelerator.h>
#include <fmt/core.h>
#include <sys/types.h>
#include <torch/csrc/python_headers.h>
#include <csignal>
#include <optional>
#ifndef _MSC_VER
#include <sys/socket.h>
#endif
#include <ATen/ATen.h>
#include <ATen/BlasBackend.h>
#include <ATen/CachedTensorUtils.h>
#include <ATen/DLConvertor.h>
#include <ATen/ExpandUtils.h>
#include <ATen/LegacyVmapMode.h>
#include <ATen/LinalgBackend.h>
#include <ATen/Parallel.h>
#include <ATen/Utils.h>
#include <ATen/core/Vitals.h>
#include <ATen/dlpack.h>
#include <ATen/native/ConvUtils.h>
#include <ATen/native/ForeachUtils.h>
#include <ATen/native/Normalization.h>
#include <c10/core/Device.h>
#include <c10/core/DispatchKeySet.h>
#include <c10/core/impl/DeviceGuardImplInterface.h>
#include <c10/util/AbortHandler.h>
#include <c10/util/Backtrace.h>
#include <c10/util/Logging.h>
#include <c10/util/irange.h>
#include <c10/util/thread_name.h>
#include <libshm.h>
#include <pybind11/pybind11.h>
#include <pybind11/stl.h>
#include <torch/csrc/THConcat.h>
#include <torch/csrc/utils/pybind.h>
#include <cstdlib>
#include <iostream>
#include <unordered_map>
#include <ATen/ThreadLocalPythonObjects.h>
#include <torch/csrc/DataLoader.h>
#include <torch/csrc/Device.h>
#include <torch/csrc/DeviceAccelerator.h>
#include <torch/csrc/Dtype.h>
#include <torch/csrc/DynamicTypes.h>
#include <torch/csrc/Event.h>
#include <torch/csrc/Export.h>
#include <torch/csrc/Generator.h>
#include <torch/csrc/Layout.h>
#include <torch/csrc/MemoryFormat.h>
#include <torch/csrc/QScheme.h>
#include <torch/csrc/Stream.h>
#include <torch/csrc/THP.h>
#include <torch/csrc/TypeInfo.h>
#include <torch/csrc/acc/Module.h>
#include <torch/csrc/api/include/torch/python/init.h>
#include <torch/csrc/autograd/generated/python_return_types.h>
#include <torch/csrc/autograd/python_cpp_function.h>
#include <torch/csrc/autograd/python_enum_tag.h>
#include <torch/csrc/autograd/python_fft_functions.h>
#include <torch/csrc/autograd/python_function.h>
#include <torch/csrc/autograd/python_legacy_variable.h>
#include <torch/csrc/autograd/python_linalg_functions.h>
#include <torch/csrc/autograd/python_nested_functions.h>
#include <torch/csrc/autograd/python_nn_functions.h>
#include <torch/csrc/autograd/python_sparse_functions.h>
#include <torch/csrc/autograd/python_special_functions.h>
#include <torch/csrc/autograd/python_variable.h>
#include <torch/csrc/cpu/Module.h>
#include <torch/csrc/distributed/python_placement.h>
#include <torch/csrc/dynamo/init.h>
#include <torch/csrc/export/pybind.h>
#include <torch/csrc/functionalization/Module.h>
#include <torch/csrc/functorch/init.h>
#include <torch/csrc/fx/node.h>
#include <torch/csrc/inductor/aoti_package/pybind.h>
#include <torch/csrc/inductor/aoti_runner/pybind.h>
#include <torch/csrc/instruction_counter/Module.h>
#include <torch/csrc/jit/python/init.h>
#include <torch/csrc/jit/python/python_ir.h>
#include <torch/csrc/jit/python/python_tracer.h>
#include <torch/csrc/jit/serialization/pickler.h>
#include <torch/csrc/lazy/python/init.h>
#include <torch/csrc/monitor/python_init.h>
#include <torch/csrc/mps/Module.h>
#include <torch/csrc/mtia/Module.h>
#include <torch/csrc/multiprocessing/init.h>
#include <torch/csrc/onnx/init.h>
#include <torch/csrc/profiler/python/init.h>
#include <torch/csrc/tensor/python_tensor.h>
#include <torch/csrc/utils/disable_torch_function.h>
#include <torch/csrc/utils/init.h>
#include <torch/csrc/utils/pycfunction_helpers.h>
#include <torch/csrc/utils/python_arg_parser.h>
#include <torch/csrc/utils/python_compat.h>
#include <torch/csrc/utils/python_dispatch.h>
#include <torch/csrc/utils/python_strings.h>
#include <torch/csrc/utils/tensor_dtypes.h>
#include <torch/csrc/utils/tensor_layouts.h>
#include <torch/csrc/utils/tensor_memoryformats.h>
#include <torch/csrc/utils/tensor_new.h>
#include <torch/csrc/utils/tensor_numpy.h>
#include <torch/csrc/utils/tensor_qschemes.h>
#include <torch/csrc/utils/verbose.h>
#include <ATen/native/transformers/sdp_utils_cpp.h>
#include <torch/csrc/profiler/combined_traceback.h>
#include <torch/csrc/profiler/kineto_client_interface.h>
#include <sstream>
#ifdef USE_CUDA
#include <ATen/ROCmFABackend.h>
#include <ATen/cuda/CUDABlas.h>
#include <ATen/cuda/CUDAConfig.h>
#include <ATen/native/transformers/cuda/sdp_utils.h>
#include <torch/csrc/inductor/static_cuda_launcher.h>
#ifdef __HIP_PLATFORM_AMD__
#include <ATen/native/cudnn/hip/BatchNorm.h>
#else
#include <ATen/native/cudnn/BatchNorm.h>
#endif
#endif
#ifdef USE_XPU
#include <ATen/native/transformers/xpu/sdp_utils.h>
#endif
#ifdef USE_DISTRIBUTED
#ifdef USE_C10D
#include <torch/csrc/distributed/autograd/python_autograd.h>
#include <torch/csrc/distributed/c10d/c10d.h>
#include <torch/csrc/distributed/rpc/rpc.h>
#include <torch/csrc/distributed/rpc/testing/testing.h>
#endif
#endif
#if defined(USE_VALGRIND)
#include <callgrind.h>
#endif
#ifdef USE_ITT
#include <torch/csrc/itt.h>
#endif
#include <torch/nativert/python/Bindings.h>
namespace py = pybind11;
static PyObject* module;
static THPGenerator* THPDefaultCPUGenerator = nullptr;
////////////////////////////////////////////////////////////////////////////////
////////////////////////////////////////////////////////////////////////////////
static PyObject* THPModule_initNames(PyObject* self, PyObject* arg) {
HANDLE_TH_ERRORS
static std::vector<std::string> names;
THPObjectPtr types(PySequence_Fast(arg, "expected a sequence"));
if (!types)
return nullptr;
// NOLINTNEXTLINE(bugprone-branch-clone)
auto num_classes = PySequence_Fast_GET_SIZE(types.get());
names.reserve(names.size() + num_classes);
for (Py_ssize_t i = 0; i < num_classes; i++) {
PyObject* obj = PySequence_Fast_GET_ITEM(types.get(), i);
TORCH_CHECK(PyType_Check(obj), "expected a PyTypeObject");
PyTypeObject* type = reinterpret_cast<PyTypeObject*>(obj);
THPObjectPtr module_name(PyObject_GetAttrString(obj, "__module__"));
if (!module_name)
return nullptr;
TORCH_CHECK(
THPUtils_checkString(module_name.get()),
"expected __module__ to be a string");
std::string name = THPUtils_unpackString(module_name.get());
names.emplace_back(name + "." + type->tp_name);
type->tp_name = names.back().c_str();
}
Py_RETURN_NONE;
END_HANDLE_TH_ERRORS
}
//
// Callback for python part. Used for additional initialization of python
// classes
static PyObject* THPModule_initExtension(
PyObject* _unused,
PyObject* shm_manager_path) {
HANDLE_TH_ERRORS
#if !defined(FBCODE_CAFFE2) && !defined(__aarch64__)
if (torch::get_cpp_stacktraces_enabled()) {
c10::SetStackTraceFetcher([]() -> std::string {
auto tb = torch::CapturedTraceback::gather(false, false, true);
if (torch::get_symbolize_mode() == torch::unwind::Mode::addr2line) {
LOG(WARNING)
<< "symbolizing C++ stack trace for exception; if this hangs, rerun with TORCH_DISABLE_ADDR2LINE=1..."
<< '\n';
}
auto s_tbs = torch::symbolize({tb.get()});
std::stringstream oss;
oss << "C++ CapturedTraceback:" << '\n';
const auto& s_tb = s_tbs.tracebacks.at(0);
for (auto idx : c10::irange(s_tb.size())) {
// Skip the first few frames:
// #1 torch::CapturedTraceback::gather(bool, bool, bool)
// #2 THPModule_initExtension
// #3 THPModule_initExtension(_object*, _object*)::{lambda()#1}
if (idx <= 3) {
continue;
}
auto frame_id = s_tb[idx];
const auto& frame = s_tbs.all_frames.at(frame_id);
oss << '#' << idx << ' ' << frame.funcname << " from " << frame.filename
<< ':' << frame.lineno << '\n';
}
return oss.str();
});
}
#endif
if (!THPUtils_checkString(shm_manager_path)) {
THPUtils_setError(
"initialization error - expected bytes/string object as shm_manager_path!");
return nullptr;
}
torch::utils::initializeLayouts();
torch::utils::initializeMemoryFormats();
torch::utils::initializeQSchemes();
torch::utils::initializeDtypes();
torch::tensors::initialize_python_bindings();
std::string path = THPUtils_unpackString(shm_manager_path);
libshm_init(path.c_str());
auto module = THPObjectPtr(PyImport_ImportModule("torch"));
if (!module)
throw python_error();
THPStorage_postInit(module);
THPAutograd_initFunctions();
Py_RETURN_NONE;
END_HANDLE_TH_ERRORS
}
// The idea behind these functions is to make it easy to test if we are
// built with ASAN: they're designed not to crash if ASAN is not enabled, but
// to trigger ASAN if it is enabled. This lets us run a "canary" tests which
// checks if our build environment is misconfigured.
static PyObject* THPModule_crashIfCsrcASAN(PyObject* module, PyObject* arg) {
HANDLE_TH_ERRORS
TORCH_CHECK(
THPUtils_checkLong(arg),
"crash_if_csrc_asan expects an int, but got ",
THPUtils_typename(arg));
// NOLINTNEXTLINE(cppcoreguidelines-avoid-c-arrays, modernize-avoid-c-arrays)
volatile char x[3];
x[THPUtils_unpackInt(arg)] = 0;
// NOLINTNEXTLINE(clang-analyzer-core.CallAndMessage)
return THPUtils_packInt32(x[0]);
END_HANDLE_TH_ERRORS
}
static PyObject* THPModule_crashIfCsrcUBSAN(PyObject* module, PyObject* arg) {
HANDLE_TH_ERRORS
TORCH_CHECK(
THPUtils_checkLong(arg),
"crash_if_csrc_ubsan expects an int, but got ",
THPUtils_typename(arg));
int32_t x = THPUtils_unpackInt(arg);
double y = 1.0 / x;
return THPUtils_packInt32(static_cast<int>(y));
END_HANDLE_TH_ERRORS
}
static PyObject* THPModule_crashIfvptrUBSAN(PyObject* module, PyObject* noarg) {
// This code should work perfectly fine, as vtables are identical for Foo and
// Baz unless rtti and ubsan are enabled
struct Foo {
virtual int bar() = 0;
virtual ~Foo() = default;
};
struct Baz {
virtual int bar() {
return 17;
}
virtual ~Baz() = default;
};
Baz x{};
// Purposely cast through `void*` so there's no fixups applied.
// NOLINTNEXTLINE(bugprone-casting-through-void,-warnings-as-errors)
auto y = static_cast<Foo*>(static_cast<void*>(&x));
auto rc = y->bar();
return THPUtils_packInt32(rc);
}
static PyObject* THPModule_crashIfATenASAN(PyObject* module, PyObject* arg) {
HANDLE_TH_ERRORS
TORCH_CHECK(
THPUtils_checkLong(arg),
"crash_if_aten_asan expects an int, "
"but got ",
THPUtils_typename(arg));
return THPUtils_packInt32(at::_crash_if_asan(THPUtils_unpackInt(arg)));
END_HANDLE_TH_ERRORS
}
static PyObject* THPModule_abort(PyObject* module, PyObject* noargs) {
std::terminate();
Py_RETURN_NONE;
}
static PyObject* THPModule_crashIfDebugAssertsFail(
PyObject* module,
PyObject* arg) {
HANDLE_TH_ERRORS
TORCH_CHECK(
THPUtils_checkLong(arg),
"crash_if_debug_asserts_fail expects an int, but got ",
THPUtils_typename(arg));
TORCH_INTERNAL_ASSERT_DEBUG_ONLY(
THPUtils_unpackInt(arg) != 424242,
"Expect anything but 424242 as an input for debug builds");
return THPUtils_packInt32(0);
END_HANDLE_TH_ERRORS
}
static PyObject* THPModule_getNumThreads(PyObject* module, PyObject* noargs) {
return THPUtils_packInt32(at::get_num_threads());
}
static PyObject* THPModule_setNumThreads(PyObject* module, PyObject* arg) {
HANDLE_TH_ERRORS
TORCH_CHECK(
THPUtils_checkLong(arg),
"set_num_threads expects an int, but got ",
THPUtils_typename(arg));
int nthreads = THPUtils_unpackInt(arg);
TORCH_CHECK(nthreads > 0, "set_num_threads expects a positive integer");
at::set_num_threads(nthreads);
Py_RETURN_NONE;
END_HANDLE_TH_ERRORS
}
static PyObject* THPModule_getNumInteropThreads(
PyObject* module,
PyObject* noargs) {
return THPUtils_packUInt64(at::get_num_interop_threads());
}
static PyObject* THPModule_setNumInteropThreads(
PyObject* module,
PyObject* arg) {
HANDLE_TH_ERRORS
TORCH_CHECK(
THPUtils_checkLong(arg),
"set_num_interop_threads expects an int, "
"but got ",
THPUtils_typename(arg));
int nthreads = THPUtils_unpackInt(arg);
TORCH_CHECK(
nthreads > 0, "set_num_interop_threads expects a positive integer");
at::set_num_interop_threads(nthreads);
Py_RETURN_NONE;
END_HANDLE_TH_ERRORS
}
static PyObject* THPModule_setDefaultTensorType(
PyObject* _unused,
PyObject* type) {
HANDLE_TH_ERRORS
torch::tensors::py_set_default_tensor_type(type);
Py_RETURN_NONE;
END_HANDLE_TH_ERRORS
}
static PyObject* THPModule_setDefaultDtype(PyObject* _unused, PyObject* dtype) {
HANDLE_TH_ERRORS
torch::tensors::py_set_default_dtype(dtype);
Py_RETURN_NONE;
END_HANDLE_TH_ERRORS
}
static PyObject* THPModule_swap_tensor_impl(PyObject* _unused, PyObject* args) {
HANDLE_TH_ERRORS
PyObject* a_ = nullptr;
PyObject* b_ = nullptr;
if (!PyArg_ParseTuple(args, "OO", &a_, &b_)) {
return nullptr;
}
// Ensure we have Tensors
TORCH_CHECK(THPVariable_Check(a_));
TORCH_CHECK(THPVariable_Check(b_));
THPVariable* a = reinterpret_cast<THPVariable*>(a_);
THPVariable* b = reinterpret_cast<THPVariable*>(b_);
// weak_use_count() adds 1 if use_count is non-zero
TORCH_CHECK(
a->cdata.weak_use_count() == 1,
"Expected no weakrefs to t1's Tensor object but got ",
a->cdata.weak_use_count() - 1);
TORCH_CHECK(
b->cdata.weak_use_count() == 1,
"Expected no weakrefs to t2's Tensor object but got ",
b->cdata.weak_use_count() - 1);
// NB: Creating local copies of *both* Tensors here ensures that they each
// hold a strong reference to their PyObject. This avoids having to fix up
// reference counts when we swap the PyObject slots below.
at::Tensor tmp_a = a->cdata;
at::Tensor tmp_b = b->cdata;
// Swap the Tensor Impl
a->cdata = tmp_b;
b->cdata = tmp_a;
// Fix up the PyObjects associated with each TensorImpl
a->cdata.unsafeGetTensorImpl()->pyobj_slot()->store_pyobj(a_);
b->cdata.unsafeGetTensorImpl()->pyobj_slot()->store_pyobj(b_);
Py_RETURN_NONE;
END_HANDLE_TH_ERRORS
}
static PyObject* THPModule_addDocStr(PyObject* _unused, PyObject* args) {
// adds a __doc__ string to a function, similar to numpy's arr_add_docstring
static std::vector<std::string> all_docs;
PyObject* obj = nullptr;
PyObject* doc_obj = nullptr;
if (!PyArg_ParseTuple(args, "OO", &obj, &doc_obj)) {
return nullptr;
}
const char* doc_str = "<invalid string>";
if (THPUtils_checkString(doc_obj)) {
all_docs.push_back(THPUtils_unpackString(doc_obj));
doc_str = all_docs.back().c_str();
}
if (Py_TYPE(obj) == &PyCFunction_Type) {
PyCFunctionObject* f = reinterpret_cast<PyCFunctionObject*>(obj);
if (f->m_ml->ml_doc) {
return PyErr_Format(
PyExc_RuntimeError,
"function '%s' already has a docstring",
f->m_ml->ml_name);
}
f->m_ml->ml_doc = doc_str;
} else if (strcmp(Py_TYPE(obj)->tp_name, "method_descriptor") == 0) {
PyMethodDescrObject* m = reinterpret_cast<PyMethodDescrObject*>(obj);
if (m->d_method->ml_doc) {
return PyErr_Format(
PyExc_RuntimeError,
"method '%s' already has a docstring",
m->d_method->ml_name);
}
m->d_method->ml_doc = doc_str;
} else if (strcmp(Py_TYPE(obj)->tp_name, "getset_descriptor") == 0) {
PyGetSetDescrObject* m = reinterpret_cast<PyGetSetDescrObject*>(obj);
if (m->d_getset->doc) {
return PyErr_Format(
PyExc_RuntimeError,
"attribute '%s' already has a docstring",
m->d_getset->name);
}
m->d_getset->doc = doc_str;
} else if (Py_TYPE(obj) == &PyType_Type) {
PyTypeObject* t = reinterpret_cast<PyTypeObject*>(obj);
if (t->tp_doc) {
return PyErr_Format(
PyExc_RuntimeError, "Type '%s' already has a docstring", t->tp_name);
}
t->tp_doc = doc_str;
} else {
return PyErr_Format(
PyExc_TypeError,
"don't know how to add docstring to type '%s'",
Py_TYPE(obj)->tp_name);
}
Py_INCREF(obj);
return obj;
}
static PyObject* THPModule_inferSize(PyObject* _unused, PyObject* args) {
HANDLE_TH_ERRORS
Py_ssize_t num_args = args ? PyTuple_Size(args) : 0;
TORCH_CHECK(num_args == 2, "expected exactly 2 arguments");
PyObject* arg1 = PyTuple_GET_ITEM(args, 0);
TORCH_CHECK(THPSize_Check(arg1), "expected a torch.Size as argument 1");
PyObject* arg2 = PyTuple_GET_ITEM(args, 1);
TORCH_CHECK(THPSize_Check(arg2), "expected a torch.Size as argument 2");
auto size1 = THPUtils_unpackLongs(arg1);
auto size2 = THPUtils_unpackLongs(arg2);
auto sizes = at::infer_size(size1, size2);
return THPSize_NewFromSizes(static_cast<int64_t>(sizes.size()), sizes.data());
END_HANDLE_TH_ERRORS
}
static PyObject* THPModule_setBackcompatBroadcastWarn(
PyObject* module,
PyObject* arg) {
HANDLE_TH_ERRORS
TORCH_CHECK(
PyBool_Check(arg),
"set_backcompat_broadcast_warn expects a bool, "
"but got ",
THPUtils_typename(arg));
setBackCompatBroadcastWarn(arg == Py_True);
Py_RETURN_NONE;
END_HANDLE_TH_ERRORS
}
static PyObject* THPModule_getBackcompatBroadcastWarn(
PyObject* module,
PyObject* noargs) {
if (getBackCompatBroadcastWarn())
Py_RETURN_TRUE;
else
Py_RETURN_FALSE;
}
static PyObject* THPModule_setBackcompatKeepdimWarn(
PyObject* module,
PyObject* arg) {
HANDLE_TH_ERRORS
TORCH_CHECK(
PyBool_Check(arg),
"set_backcompat_keepdim_warn expects a bool, "
"but got ",
THPUtils_typename(arg));
setBackCompatKeepdimWarn(arg == Py_True);
Py_RETURN_NONE;
END_HANDLE_TH_ERRORS
}
static PyObject* THPModule_getBackcompatKeepdimWarn(
PyObject* module,
PyObject* noargs) {
if (getBackCompatKeepdimWarn())
Py_RETURN_TRUE;
else
Py_RETURN_FALSE;
}
static PyObject* THPModule_hasDistributed(PyObject* _unused, PyObject* noargs) {
#ifdef USE_DISTRIBUTED
Py_RETURN_TRUE;
#else
Py_RETURN_FALSE;
#endif
}
static PyObject* THPModule_showConfig(PyObject* module, PyObject* noargs) {
HANDLE_TH_ERRORS
return THPUtils_packString(at::show_config());
END_HANDLE_TH_ERRORS
}
static PyObject* THPModule_cxxFlags(PyObject* module, PyObject* noargs) {
HANDLE_TH_ERRORS
return THPUtils_packString(at::get_cxx_flags());
END_HANDLE_TH_ERRORS
}
static PyObject* THPModule_parallelInfo(PyObject* module, PyObject* noargs) {
HANDLE_TH_ERRORS
return THPUtils_packString(at::get_parallel_info());
END_HANDLE_TH_ERRORS
}
static PyObject* THPModule_getCpuCapability(
PyObject* module,
PyObject* noargs) {
HANDLE_TH_ERRORS
return THPUtils_packString(at::get_cpu_capability());
END_HANDLE_TH_ERRORS
}
namespace {
template <class T>
void DLPack_Capsule_Destructor(PyObject* data) {
if (C10_LIKELY(!PyCapsule_IsValid(data, at::DLPackTraits<T>::capsule))) {
// early out, see DLPack spec: if a consuming library sets the capsule
// name to something else, they own it and we don't need to do anything
return;
}
HANDLE_TH_ERRORS
// Causes overheads for validity checks again, but this case is rare
// since consuming libraries should rename the capsule according to spec.
// Note that this cannot set a python error (we checked validity above),
// so we don't need to handle python error state here.
T* tensor = (T*)PyCapsule_GetPointer(data, at::DLPackTraits<T>::capsule);
// the dlMTensor has not been consumed, call deleter ourselves.
// DLPack spec mentions that deleter may be NULL, but deleter from
// `at::toDLPack` is never NULL, so no need for an additional check here.
tensor->deleter(tensor);
END_HANDLE_TH_ERRORS_RET()
}
template <class T>
PyObject* THPModule_toDLPackImpl(
PyObject* self,
PyObject* args,
PyObject* kwargs) {
HANDLE_TH_ERRORS
static torch::PythonArgParser parser(
{"_to_dlpack(Tensor data, *, IntArrayRef? dl_device=None, bool? copy=None)"});
torch::ParsedArgs<3> parsed_args{};
auto r = parser.parse(args, kwargs, parsed_args);
TORCH_INTERNAL_ASSERT(r.idx == 0);
auto data = r.tensor(0);
auto dl_device = r.intlist(1);
auto copy = r.toBoolOptional(2);
// Parse the int list into a tuple.
std::optional<DLDevice> optional_dl_device;
if (!dl_device.empty()) {
TORCH_CHECK(
dl_device.size() == 2,
"dl_device must be either None or a tuple of ints");
optional_dl_device = DLDevice{
static_cast<DLDeviceType>(dl_device[0]),
static_cast<int32_t>(dl_device[1])};
}
auto tensor = at::DLPackTraits<T>::toDLPack(
at::maybeCopyTensor(data, optional_dl_device, copy));
return PyCapsule_New(
tensor, at::DLPackTraits<T>::capsule, DLPack_Capsule_Destructor<T>);
END_HANDLE_TH_ERRORS
}
} // namespace
static PyObject* THPModule_toDLPack(
PyObject* self,
PyObject* args,
PyObject* kwargs) {
return THPModule_toDLPackImpl<DLManagedTensor>(self, args, kwargs);
}
static PyObject* THPModule_toDLPackVersioned(
PyObject* self,
PyObject* args,
PyObject* kwargs) {
return THPModule_toDLPackImpl<DLManagedTensorVersioned>(self, args, kwargs);
}
static PyObject* THPModule_fromDLPack(PyObject* _unused, PyObject* data) {
using namespace torch::autograd;
HANDLE_TH_ERRORS
auto tensor = torch::utils::tensor_fromDLPack(data);
return THPVariable_Wrap(tensor);
END_HANDLE_TH_ERRORS
}
static PyObject* THPModule_torchDeviceToDLDevice(
PyObject* _unused,
PyObject* data) {
HANDLE_TH_ERRORS
TORCH_CHECK(
THPDevice_Check(data),
"torchDeviceToDLDevice: expected torch.device argument.");
auto device = reinterpret_cast<THPDevice*>(data)->device;
auto dl_device = at::torchDeviceToDLDevice(device);
auto tuple = PyTuple_New(2);
if (!tuple) {
throw python_error();
}
PyTuple_SET_ITEM(tuple, 0, THPUtils_packInt64(dl_device.device_type));
PyTuple_SET_ITEM(tuple, 1, THPUtils_packInt64(dl_device.device_id));
return tuple;
END_HANDLE_TH_ERRORS
}
struct TorchDLPackExchangeAPI : public DLPackExchangeAPI {
TorchDLPackExchangeAPI() {
header.version.major = DLPACK_MAJOR_VERSION;
header.version.minor = DLPACK_MINOR_VERSION;
header.prev_api = nullptr;
managed_tensor_allocator = ManagedTensorAllocator;
managed_tensor_from_py_object_no_sync = ManagedTensorFromPyObjectNoSync;
managed_tensor_to_py_object_no_sync = ManagedTensorToPyObjectNoSync;
dltensor_from_py_object_no_sync = DLTensorFromPyObjectNoSync;
current_work_stream = CurrentWorkStream;
}
static const DLPackExchangeAPI* Global() {
static TorchDLPackExchangeAPI inst;
return &inst;
}
private:
// Fast non-owning PyObject→DLTensor conversion
static int DLTensorFromPyObjectNoSync(void* py_obj, DLTensor* out) {
try {
// Use handle (non-owning) to avoid unnecessary refcount operations
py::handle handle(static_cast<PyObject*>(py_obj));
at::Tensor tensor = handle.cast<at::Tensor>();
at::toDLPackNonOwning(tensor, out);
return 0;
} catch (const std::exception& e) {
PyErr_SetString(PyExc_RuntimeError, e.what());
return -1;
}
}
// PyObject→DLManagedTensorVersioned conversion
static int ManagedTensorFromPyObjectNoSync(
void* py_obj,
DLManagedTensorVersioned** out) {
try {
py::handle handle(static_cast<PyObject*>(py_obj));
at::Tensor tensor = handle.cast<at::Tensor>();
*out = at::toDLPackVersioned(tensor);
return 0;
} catch (const std::exception& e) {
PyErr_SetString(PyExc_RuntimeError, e.what());
return -1;
}
}
// DLManagedTensorVersioned→PyObject conversion
static int ManagedTensorToPyObjectNoSync(
DLManagedTensorVersioned* src,
void** py_obj_out) {
try {
at::Tensor tensor = at::fromDLPackVersioned(src, nullptr);
*py_obj_out = THPVariable_Wrap(tensor);
return 0;
} catch (const std::exception& e) {
PyErr_SetString(PyExc_RuntimeError, e.what());
return -1;
}
}
// Allocate new tensor from prototype
static int ManagedTensorAllocator(
DLTensor* prototype,
DLManagedTensorVersioned** out,
void* error_ctx,
void (
*SetError)(void* error_ctx, const char* kind, const char* message)) {
try {
at::IntArrayRef shape(
prototype->shape, prototype->shape + prototype->ndim);
at::TensorOptions options =
at::TensorOptions()
.dtype(at::toScalarType(prototype->dtype))
.device(at::dlDeviceToTorchDevice(
prototype->device.device_type, prototype->device.device_id));
at::Tensor tensor = at::empty(shape, options);
*out = at::toDLPackVersioned(tensor);
return 0;
} catch (const std::exception& e) {
SetError(error_ctx, "MemoryError", e.what());
return -1;
}
}
// Get current CUDA/ROCm work stream
static int CurrentWorkStream(
DLDeviceType device_type,
int32_t device_id,
void** out_stream) {
try {
#ifdef USE_CUDA
if (device_type == kDLCUDA || device_type == kDLROCM) {
*out_stream = at::cuda::getCurrentCUDAStream(device_id).stream();
return 0;
}
#endif
// For CPU and other devices, return NULL (no stream concept)
*out_stream = nullptr;
return 0;
} catch (const std::exception& e) {
PyErr_SetString(PyExc_RuntimeError, e.what());
return -1;
}
}
};
static PyObject* THPModule_DLPackExchangeAPI(
PyObject* _unused,
PyObject* noargs) {
HANDLE_TH_ERRORS
return PyCapsule_New(
const_cast<DLPackExchangeAPI*>(TorchDLPackExchangeAPI::Global()),
"dlpack_exchange_api",
nullptr);
END_HANDLE_TH_ERRORS
}
static PyObject* THModule_getCppBacktrace(PyObject* _unused, PyObject* args) {
HANDLE_TH_ERRORS
size_t frames_to_skip = 0;
size_t maximum_number_of_frames = 0;
if (!PyArg_ParseTuple(
args, "LL", &frames_to_skip, &maximum_number_of_frames)) {
return nullptr;
}
return THPUtils_packString(
c10::get_backtrace(frames_to_skip, maximum_number_of_frames, true));
END_HANDLE_TH_ERRORS
}
static PyObject* THModule_rename_privateuse1_backend(
PyObject* _unused,
PyObject* arg) {
HANDLE_TH_ERRORS
TORCH_CHECK(
THPUtils_checkString(arg),
"_rename_privateuse1_backend expects a str, but got ",
THPUtils_typename(arg));
const std::string backend_name = THPUtils_unpackString(arg);
c10::register_privateuse1_backend(backend_name);
Py_RETURN_NONE;
END_HANDLE_TH_ERRORS
}
static PyObject* THModule_get_privateuse1_backend_name(
PyObject* _unused,
PyObject* arg) {
HANDLE_TH_ERRORS
return THPUtils_packString(c10::get_privateuse1_backend());
END_HANDLE_TH_ERRORS
}
static PyObject* THPModule_setAllowTF32CuDNN(PyObject* _unused, PyObject* arg) {
HANDLE_TH_ERRORS
TORCH_CHECK(
PyBool_Check(arg),
"set_allow_tf32_cublas expects a bool, "
"but got ",
THPUtils_typename(arg));
at::globalContext().setAllowTF32CuDNN(arg == Py_True);
Py_RETURN_NONE;
END_HANDLE_TH_ERRORS
}
static PyObject* THPModule_allowTF32CuDNN(PyObject* _unused, PyObject* noargs) {
HANDLE_TH_ERRORS
if (at::globalContext().allowTF32CuDNN())
Py_RETURN_TRUE;
else
Py_RETURN_FALSE;
END_HANDLE_TH_ERRORS
}
static PyObject* THPModule_setFloat32MatmulPrecision(
PyObject* _unused,
PyObject* arg) {
HANDLE_TH_ERRORS
TORCH_CHECK(
THPUtils_checkString(arg),
"set_float32_matmul_precision expects a str, "
"but got ",
THPUtils_typename(arg));
std::string s = THPUtils_unpackString(arg);
at::globalContext().setFloat32MatmulPrecision(s);
Py_RETURN_NONE;
END_HANDLE_TH_ERRORS
}
static PyObject* THPModule_float32MatmulPrecision(
PyObject* _unused,
PyObject* noargs) {
HANDLE_TH_ERRORS
std::string s = "highest";
auto p = at::globalContext().float32MatmulPrecision();
if (p == at::Float32MatmulPrecision::HIGH) {
s = "high";
} else if (p == at::Float32MatmulPrecision::MEDIUM) {
s = "medium";
}
return THPUtils_packString(s);
END_HANDLE_TH_ERRORS
}
static PyObject* THPModule_setSDPPriorityOrder(
PyObject* _unused,
PyObject* arg) {
HANDLE_TH_ERRORS
auto priority_order = THPUtils_unpackLongs(arg);
at::globalContext().setSDPPriorityOrder(priority_order);
Py_RETURN_NONE;
END_HANDLE_TH_ERRORS
}
static PyObject* THPModule_sDPPriorityOrder(
PyObject* _unused,
PyObject* noargs) {
auto ordervec = at::globalContext().sDPPriorityOrder();
auto order =
THPObjectPtr(PyList_New(static_cast<Py_ssize_t>(ordervec.size())));
for (const auto i : c10::irange(ordervec.size())) {
PyObject* i64 = THPUtils_packInt64(static_cast<int64_t>(ordervec[i]));
if (!i64)
return nullptr;
PyList_SET_ITEM(order.get(), i, i64);
}
return order.release();
}
static PyObject* THPModule_setSDPUseFlash(PyObject* _unused, PyObject* arg) {
HANDLE_TH_ERRORS
TORCH_CHECK(
PyBool_Check(arg),
"set_sdp_use_math expects a bool, "
"but got ",
THPUtils_typename(arg));
at::globalContext().setSDPUseFlash(arg == Py_True);
Py_RETURN_NONE;
END_HANDLE_TH_ERRORS
}
static PyObject* THPModule_userEnabledFlashSDP(
PyObject* _unused,
PyObject* noargs) {
if (at::globalContext().userEnabledFlashSDP())
Py_RETURN_TRUE;
else
Py_RETURN_FALSE;
}
static PyObject* THPModule_setSDPUseMemEfficient(
PyObject* _unused,
PyObject* arg) {
HANDLE_TH_ERRORS
TORCH_CHECK(
PyBool_Check(arg),
"set_sdp_use_math expects a bool, "
"but got ",
THPUtils_typename(arg));
at::globalContext().setSDPUseMemEfficient(arg == Py_True);
Py_RETURN_NONE;
END_HANDLE_TH_ERRORS
}
static PyObject* userEnabledMemEfficientSDP(
PyObject* _unused,
PyObject* noargs) {
if (at::globalContext().userEnabledMemEfficientSDP())
Py_RETURN_TRUE;
else
Py_RETURN_FALSE;
}
static PyObject* THPModule_setSDPUseMath(PyObject* _unused, PyObject* arg) {
HANDLE_TH_ERRORS
TORCH_CHECK(
PyBool_Check(arg),
"set_sdp_use_math expects a bool, "
"but got ",
THPUtils_typename(arg));
at::globalContext().setSDPUseMath(arg == Py_True);
Py_RETURN_NONE;
END_HANDLE_TH_ERRORS
}
static PyObject* THPModule_userEnabledMathSDP(
PyObject* _unused,
PyObject* noargs) {
if (at::globalContext().userEnabledMathSDP())
Py_RETURN_TRUE;
else
Py_RETURN_FALSE;
}
static PyObject* THPModule_setAllowFP16BF16ReductionMathSDP(
PyObject* _unused,
PyObject* arg) {
HANDLE_TH_ERRORS
TORCH_CHECK(
PyBool_Check(arg),
"set_sdp_use_math expects a bool, "
"but got ",
THPUtils_typename(arg));
at::globalContext().setAllowFP16BF16ReductionMathSDP(arg == Py_True);
Py_RETURN_NONE;
END_HANDLE_TH_ERRORS
}
static PyObject* THPModule_allowFP16BF16ReductionMathSDP(
PyObject* _unused,
PyObject* noargs) {
if (at::globalContext().allowFP16BF16ReductionMathSDP())
Py_RETURN_TRUE;
else
Py_RETURN_FALSE;
}
static PyObject* THPModule_setSDPUseOverrideable(
PyObject* _unused,
PyObject* arg) {