forked from pytorch/pytorch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
setup.py
1460 lines (1322 loc) · 50.4 KB
/
setup.py
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
# Welcome to the PyTorch setup.py.
#
# Environment variables you are probably interested in:
#
# DEBUG
# build with -O0 and -g (debug symbols)
#
# REL_WITH_DEB_INFO
# build with optimizations and -g (debug symbols)
#
# MAX_JOBS
# maximum number of compile jobs we should use to compile your code
#
# USE_CUDA=0
# disables CUDA build
#
# CFLAGS
# flags to apply to both C and C++ files to be compiled (a quirk of setup.py
# which we have faithfully adhered to in our build system is that CFLAGS
# also applies to C++ files (unless CXXFLAGS is set), in contrast to the
# default behavior of autogoo and cmake build systems.)
#
# CC
# the C/C++ compiler to use
#
# Environment variables for feature toggles:
#
# DEBUG_CUDA=1
# if used in conjunction with DEBUG or REL_WITH_DEB_INFO, will also
# build CUDA kernels with -lineinfo --source-in-ptx. Note that
# on CUDA 12 this may cause nvcc to OOM, so this is disabled by default.
# USE_CUDNN=0
# disables the cuDNN build
#
# USE_CUSPARSELT=0
# disables the cuSPARSELt build
#
# USE_FBGEMM=0
# disables the FBGEMM build
#
# USE_KINETO=0
# disables usage of libkineto library for profiling
#
# USE_NUMPY=0
# disables the NumPy build
#
# BUILD_TEST=0
# disables the test build
#
# USE_MKLDNN=0
# disables use of MKLDNN
#
# USE_MKLDNN_ACL
# enables use of Compute Library backend for MKLDNN on Arm;
# USE_MKLDNN must be explicitly enabled.
#
# MKLDNN_CPU_RUNTIME
# MKL-DNN threading mode: TBB or OMP (default)
#
# USE_STATIC_MKL
# Prefer to link with MKL statically - Unix only
# USE_ITT=0
# disable use of Intel(R) VTune Profiler's ITT functionality
#
# USE_NNPACK=0
# disables NNPACK build
#
# USE_QNNPACK=0
# disables QNNPACK build (quantized 8-bit operators)
#
# USE_DISTRIBUTED=0
# disables distributed (c10d, gloo, mpi, etc.) build
#
# USE_TENSORPIPE=0
# disables distributed Tensorpipe backend build
#
# USE_GLOO=0
# disables distributed gloo backend build
#
# USE_MPI=0
# disables distributed MPI backend build
#
# USE_SYSTEM_NCCL=0
# disables use of system-wide nccl (we will use our submoduled
# copy in third_party/nccl)
#
# BUILD_CAFFE2_OPS=0
# disable Caffe2 operators build
#
# BUILD_CAFFE2=0
# disable Caffe2 build
#
# USE_IBVERBS
# toggle features related to distributed support
#
# USE_OPENCV
# enables use of OpenCV for additional operators
#
# USE_OPENMP=0
# disables use of OpenMP for parallelization
#
# USE_FFMPEG
# enables use of ffmpeg for additional operators
#
# USE_FLASH_ATTENTION=0
# disables building flash attention for scaled dot product attention
#
# USE_MEM_EFF_ATTENTION=0
# disables building memory efficient attention for scaled dot product attention
#
# USE_LEVELDB
# enables use of LevelDB for storage
#
# USE_LMDB
# enables use of LMDB for storage
#
# BUILD_BINARY
# enables the additional binaries/ build
#
# ATEN_AVX512_256=TRUE
# ATen AVX2 kernels can use 32 ymm registers, instead of the default 16.
# This option can be used if AVX512 doesn't perform well on a machine.
# The FBGEMM library also uses AVX512_256 kernels on Xeon D processors,
# but it also has some (optimized) assembly code.
#
# PYTORCH_BUILD_VERSION
# PYTORCH_BUILD_NUMBER
# specify the version of PyTorch, rather than the hard-coded version
# in this file; used when we're building binaries for distribution
#
# TORCH_CUDA_ARCH_LIST
# specify which CUDA architectures to build for.
# ie `TORCH_CUDA_ARCH_LIST="6.0;7.0"`
# These are not CUDA versions, instead, they specify what
# classes of NVIDIA hardware we should generate PTX for.
#
# PYTORCH_ROCM_ARCH
# specify which AMD GPU targets to build for.
# ie `PYTORCH_ROCM_ARCH="gfx900;gfx906"`
#
# ONNX_NAMESPACE
# specify a namespace for ONNX built here rather than the hard-coded
# one in this file; needed to build with other frameworks that share ONNX.
#
# BLAS
# BLAS to be used by Caffe2. Can be MKL, Eigen, ATLAS, FlexiBLAS, or OpenBLAS. If set
# then the build will fail if the requested BLAS is not found, otherwise
# the BLAS will be chosen based on what is found on your system.
#
# MKL_THREADING
# MKL threading mode: SEQ, TBB or OMP (default)
#
# USE_REDIS
# Whether to use Redis for distributed workflows (Linux only)
#
# USE_ZSTD
# Enables use of ZSTD, if the libraries are found
#
# Environment variables we respect (these environment variables are
# conventional and are often understood/set by other software.)
#
# CUDA_HOME (Linux/OS X)
# CUDA_PATH (Windows)
# specify where CUDA is installed; usually /usr/local/cuda or
# /usr/local/cuda-x.y
# CUDAHOSTCXX
# specify a different compiler than the system one to use as the CUDA
# host compiler for nvcc.
#
# CUDA_NVCC_EXECUTABLE
# Specify a NVCC to use. This is used in our CI to point to a cached nvcc
#
# CUDNN_LIB_DIR
# CUDNN_INCLUDE_DIR
# CUDNN_LIBRARY
# specify where cuDNN is installed
#
# MIOPEN_LIB_DIR
# MIOPEN_INCLUDE_DIR
# MIOPEN_LIBRARY
# specify where MIOpen is installed
#
# NCCL_ROOT
# NCCL_LIB_DIR
# NCCL_INCLUDE_DIR
# specify where nccl is installed
#
# NVFUSER_SOURCE_DIR
# specify nvfuser root directory
#
# NVTOOLSEXT_PATH (Windows only)
# specify where nvtoolsext is installed
#
# ACL_ROOT_DIR
# specify where Compute Library is installed
#
# LIBRARY_PATH
# LD_LIBRARY_PATH
# we will search for libraries in these paths
#
# ATEN_THREADING
# ATen parallel backend to use for intra- and inter-op parallelism
# possible values:
# OMP - use OpenMP for intra-op and native backend for inter-op tasks
# NATIVE - use native thread pool for both intra- and inter-op tasks
# TBB - using TBB for intra- and native thread pool for inter-op parallelism
#
# USE_TBB
# enable TBB support
#
# USE_SYSTEM_TBB
# Use system-provided Intel TBB.
#
# USE_SYSTEM_LIBS (work in progress)
# Use system-provided libraries to satisfy the build dependencies.
# When turned on, the following cmake variables will be toggled as well:
# USE_SYSTEM_CPUINFO=ON USE_SYSTEM_SLEEF=ON BUILD_CUSTOM_PROTOBUF=OFF
#
# USE_MIMALLOC
# Static link mimalloc into C10, and use mimalloc in alloc_cpu & alloc_free.
# By default, It is only enabled on Windows.
import sys
if sys.platform == "win32" and sys.maxsize.bit_length() == 31:
print(
"32-bit Windows Python runtime is not supported. Please switch to 64-bit Python."
)
sys.exit(-1)
import platform
python_min_version = (3, 8, 0)
python_min_version_str = ".".join(map(str, python_min_version))
if sys.version_info < python_min_version:
print(
f"You are using Python {platform.python_version()}. Python >={python_min_version_str} is required."
)
sys.exit(-1)
import filecmp
import glob
import importlib
import json
import os
import shutil
import subprocess
import sysconfig
import time
from collections import defaultdict
import setuptools.command.build_ext
import setuptools.command.install
import setuptools.command.sdist
from setuptools import Extension, find_packages, setup
from setuptools.dist import Distribution
from tools.build_pytorch_libs import build_caffe2
from tools.generate_torch_version import get_torch_version
from tools.setup_helpers.cmake import CMake
from tools.setup_helpers.env import build_type, IS_DARWIN, IS_LINUX, IS_WINDOWS
################################################################################
# Parameters parsed from environment
################################################################################
VERBOSE_SCRIPT = True
RUN_BUILD_DEPS = True
# see if the user passed a quiet flag to setup.py arguments and respect
# that in our parts of the build
EMIT_BUILD_WARNING = False
RERUN_CMAKE = False
CMAKE_ONLY = False
filtered_args = []
for i, arg in enumerate(sys.argv):
if arg == "--cmake":
RERUN_CMAKE = True
continue
if arg == "--cmake-only":
# Stop once cmake terminates. Leave users a chance to adjust build
# options.
CMAKE_ONLY = True
continue
if arg == "rebuild" or arg == "build":
arg = "build" # rebuild is gone, make it build
EMIT_BUILD_WARNING = True
if arg == "--":
filtered_args += sys.argv[i:]
break
if arg == "-q" or arg == "--quiet":
VERBOSE_SCRIPT = False
if arg in ["clean", "egg_info", "sdist"]:
RUN_BUILD_DEPS = False
filtered_args.append(arg)
sys.argv = filtered_args
if VERBOSE_SCRIPT:
def report(*args):
print(*args)
else:
def report(*args):
pass
# Make distutils respect --quiet too
setuptools.distutils.log.warn = report
# Constant known variables used throughout this file
cwd = os.path.dirname(os.path.abspath(__file__))
lib_path = os.path.join(cwd, "torch", "lib")
third_party_path = os.path.join(cwd, "third_party")
caffe2_build_dir = os.path.join(cwd, "build")
# CMAKE: full path to python library
if IS_WINDOWS:
cmake_python_library = "{}/libs/python{}.lib".format(
sysconfig.get_config_var("prefix"), sysconfig.get_config_var("VERSION")
)
# Fix virtualenv builds
if not os.path.exists(cmake_python_library):
cmake_python_library = "{}/libs/python{}.lib".format(
sys.base_prefix, sysconfig.get_config_var("VERSION")
)
else:
cmake_python_library = "{}/{}".format(
sysconfig.get_config_var("LIBDIR"), sysconfig.get_config_var("INSTSONAME")
)
cmake_python_include_dir = sysconfig.get_path("include")
################################################################################
# Version, create_version_file, and package_name
################################################################################
package_name = os.getenv("TORCH_PACKAGE_NAME", "torch")
package_type = os.getenv("PACKAGE_TYPE", "wheel")
version = get_torch_version()
report(f"Building wheel {package_name}-{version}")
cmake = CMake()
def get_submodule_folders():
git_modules_path = os.path.join(cwd, ".gitmodules")
default_modules_path = [
os.path.join(third_party_path, name)
for name in [
"gloo",
"cpuinfo",
"tbb",
"onnx",
"foxi",
"QNNPACK",
"fbgemm",
"cutlass",
]
]
if not os.path.exists(git_modules_path):
return default_modules_path
with open(git_modules_path) as f:
return [
os.path.join(cwd, line.split("=", 1)[1].strip())
for line in f.readlines()
if line.strip().startswith("path")
]
def check_submodules():
def check_for_files(folder, files):
if not any(os.path.exists(os.path.join(folder, f)) for f in files):
report("Could not find any of {} in {}".format(", ".join(files), folder))
report("Did you run 'git submodule update --init --recursive'?")
sys.exit(1)
def not_exists_or_empty(folder):
return not os.path.exists(folder) or (
os.path.isdir(folder) and len(os.listdir(folder)) == 0
)
if bool(os.getenv("USE_SYSTEM_LIBS", False)):
return
folders = get_submodule_folders()
# If none of the submodule folders exists, try to initialize them
if all(not_exists_or_empty(folder) for folder in folders):
try:
print(" --- Trying to initialize submodules")
start = time.time()
subprocess.check_call(
["git", "submodule", "update", "--init", "--recursive"], cwd=cwd
)
end = time.time()
print(f" --- Submodule initialization took {end - start:.2f} sec")
except Exception:
print(" --- Submodule initalization failed")
print("Please run:\n\tgit submodule update --init --recursive")
sys.exit(1)
for folder in folders:
check_for_files(
folder,
[
"CMakeLists.txt",
"Makefile",
"setup.py",
"LICENSE",
"LICENSE.md",
"LICENSE.txt",
],
)
check_for_files(
os.path.join(third_party_path, "fbgemm", "third_party", "asmjit"),
["CMakeLists.txt"],
)
check_for_files(
os.path.join(third_party_path, "onnx", "third_party", "benchmark"),
["CMakeLists.txt"],
)
# Windows has very bad support for symbolic links.
# Instead of using symlinks, we're going to copy files over
def mirror_files_into_torchgen():
# (new_path, orig_path)
# Directories are OK and are recursively mirrored.
paths = [
(
"torchgen/packaged/ATen/native/native_functions.yaml",
"aten/src/ATen/native/native_functions.yaml",
),
("torchgen/packaged/ATen/native/tags.yaml", "aten/src/ATen/native/tags.yaml"),
("torchgen/packaged/ATen/templates", "aten/src/ATen/templates"),
("torchgen/packaged/autograd", "tools/autograd"),
("torchgen/packaged/autograd/templates", "tools/autograd/templates"),
]
for new_path, orig_path in paths:
# Create the dirs involved in new_path if they don't exist
if not os.path.exists(new_path):
os.makedirs(os.path.dirname(new_path), exist_ok=True)
# Copy the files from the orig location to the new location
if os.path.isfile(orig_path):
shutil.copyfile(orig_path, new_path)
continue
if os.path.isdir(orig_path):
if os.path.exists(new_path):
# copytree fails if the tree exists already, so remove it.
shutil.rmtree(new_path)
shutil.copytree(orig_path, new_path)
continue
raise RuntimeError("Check the file paths in `mirror_files_into_torchgen()`")
# all the work we need to do _before_ setup runs
def build_deps():
report("-- Building version " + version)
check_submodules()
check_pydep("yaml", "pyyaml")
build_caffe2(
version=version,
cmake_python_library=cmake_python_library,
build_python=True,
rerun_cmake=RERUN_CMAKE,
cmake_only=CMAKE_ONLY,
cmake=cmake,
)
if CMAKE_ONLY:
report(
'Finished running cmake. Run "ccmake build" or '
'"cmake-gui build" to adjust build options and '
'"python setup.py install" to build.'
)
sys.exit()
# Use copies instead of symbolic files.
# Windows has very poor support for them.
sym_files = [
"tools/shared/_utils_internal.py",
"torch/utils/benchmark/utils/valgrind_wrapper/callgrind.h",
"torch/utils/benchmark/utils/valgrind_wrapper/valgrind.h",
]
orig_files = [
"torch/_utils_internal.py",
"third_party/valgrind-headers/callgrind.h",
"third_party/valgrind-headers/valgrind.h",
]
for sym_file, orig_file in zip(sym_files, orig_files):
same = False
if os.path.exists(sym_file):
if filecmp.cmp(sym_file, orig_file):
same = True
else:
os.remove(sym_file)
if not same:
shutil.copyfile(orig_file, sym_file)
################################################################################
# Building dependent libraries
################################################################################
missing_pydep = """
Missing build dependency: Unable to `import {importname}`.
Please install it via `conda install {module}` or `pip install {module}`
""".strip()
def check_pydep(importname, module):
try:
importlib.import_module(importname)
except ImportError as e:
raise RuntimeError(
missing_pydep.format(importname=importname, module=module)
) from e
class build_ext(setuptools.command.build_ext.build_ext):
# Copy libiomp5.dylib inside the wheel package on OS X
def _embed_libiomp(self):
lib_dir = os.path.join(self.build_lib, "torch", "lib")
libtorch_cpu_path = os.path.join(lib_dir, "libtorch_cpu.dylib")
if not os.path.exists(libtorch_cpu_path):
return
# Parse libtorch_cpu load commands
otool_cmds = (
subprocess.check_output(["otool", "-l", libtorch_cpu_path])
.decode("utf-8")
.split("\n")
)
rpaths, libs = [], []
for idx, line in enumerate(otool_cmds):
if line.strip() == "cmd LC_LOAD_DYLIB":
lib_name = otool_cmds[idx + 2].strip()
assert lib_name.startswith("name ")
libs.append(lib_name.split(" ", 1)[1].rsplit("(", 1)[0][:-1])
if line.strip() == "cmd LC_RPATH":
rpath = otool_cmds[idx + 2].strip()
assert rpath.startswith("path ")
rpaths.append(rpath.split(" ", 1)[1].rsplit("(", 1)[0][:-1])
omp_lib_name = "libiomp5.dylib"
if os.path.join("@rpath", omp_lib_name) not in libs:
return
# Copy libiomp5 from rpath locations
for rpath in rpaths:
source_lib = os.path.join(rpath, omp_lib_name)
if not os.path.exists(source_lib):
continue
target_lib = os.path.join(self.build_lib, "torch", "lib", omp_lib_name)
self.copy_file(source_lib, target_lib)
break
def run(self):
# Report build options. This is run after the build completes so # `CMakeCache.txt` exists and we can get an
# accurate report on what is used and what is not.
cmake_cache_vars = defaultdict(lambda: False, cmake.get_cmake_cache_variables())
if cmake_cache_vars["USE_NUMPY"]:
report("-- Building with NumPy bindings")
else:
report("-- NumPy not found")
if cmake_cache_vars["USE_CUDNN"]:
report(
"-- Detected cuDNN at "
+ cmake_cache_vars["CUDNN_LIBRARY"]
+ ", "
+ cmake_cache_vars["CUDNN_INCLUDE_DIR"]
)
else:
report("-- Not using cuDNN")
if cmake_cache_vars["USE_CUDA"]:
report("-- Detected CUDA at " + cmake_cache_vars["CUDA_TOOLKIT_ROOT_DIR"])
else:
report("-- Not using CUDA")
if cmake_cache_vars["USE_MKLDNN"]:
report("-- Using MKLDNN")
if cmake_cache_vars["USE_MKLDNN_ACL"]:
report("-- Using Compute Library for the Arm architecture with MKLDNN")
else:
report(
"-- Not using Compute Library for the Arm architecture with MKLDNN"
)
if cmake_cache_vars["USE_MKLDNN_CBLAS"]:
report("-- Using CBLAS in MKLDNN")
else:
report("-- Not using CBLAS in MKLDNN")
else:
report("-- Not using MKLDNN")
if cmake_cache_vars["USE_NCCL"] and cmake_cache_vars["USE_SYSTEM_NCCL"]:
report(
"-- Using system provided NCCL library at {}, {}".format(
cmake_cache_vars["NCCL_LIBRARIES"],
cmake_cache_vars["NCCL_INCLUDE_DIRS"],
)
)
elif cmake_cache_vars["USE_NCCL"]:
report("-- Building NCCL library")
else:
report("-- Not using NCCL")
if cmake_cache_vars["USE_DISTRIBUTED"]:
if IS_WINDOWS:
report("-- Building without distributed package")
else:
report("-- Building with distributed package: ")
report(
" -- USE_TENSORPIPE={}".format(cmake_cache_vars["USE_TENSORPIPE"])
)
report(" -- USE_GLOO={}".format(cmake_cache_vars["USE_GLOO"]))
report(" -- USE_MPI={}".format(cmake_cache_vars["USE_OPENMPI"]))
else:
report("-- Building without distributed package")
if cmake_cache_vars["STATIC_DISPATCH_BACKEND"]:
report(
"-- Using static dispatch with backend {}".format(
cmake_cache_vars["STATIC_DISPATCH_BACKEND"]
)
)
if cmake_cache_vars["USE_LIGHTWEIGHT_DISPATCH"]:
report("-- Using lightweight dispatch")
if cmake_cache_vars["BUILD_EXECUTORCH"]:
report("-- Building Executorch")
if cmake_cache_vars["USE_ITT"]:
report("-- Using ITT")
else:
report("-- Not using ITT")
if cmake_cache_vars["BUILD_NVFUSER"]:
report("-- Building nvfuser")
else:
report("-- Not Building nvfuser")
# Do not use clang to compile extensions if `-fstack-clash-protection` is defined
# in system CFLAGS
c_flags = str(os.getenv("CFLAGS", ""))
if (
IS_LINUX
and "-fstack-clash-protection" in c_flags
and "clang" in os.environ.get("CC", "")
):
os.environ["CC"] = str(os.environ["CC"])
# It's an old-style class in Python 2.7...
setuptools.command.build_ext.build_ext.run(self)
if IS_DARWIN and package_type != "conda":
self._embed_libiomp()
# Copy the essential export library to compile C++ extensions.
if IS_WINDOWS:
build_temp = self.build_temp
ext_filename = self.get_ext_filename("_C")
lib_filename = ".".join(ext_filename.split(".")[:-1]) + ".lib"
export_lib = os.path.join(
build_temp, "torch", "csrc", lib_filename
).replace("\\", "/")
build_lib = self.build_lib
target_lib = os.path.join(build_lib, "torch", "lib", "_C.lib").replace(
"\\", "/"
)
# Create "torch/lib" directory if not exists.
# (It is not created yet in "develop" mode.)
target_dir = os.path.dirname(target_lib)
if not os.path.exists(target_dir):
os.makedirs(target_dir)
self.copy_file(export_lib, target_lib)
def build_extensions(self):
self.create_compile_commands()
# The caffe2 extensions are created in
# tmp_install/lib/pythonM.m/site-packages/caffe2/python/
# and need to be copied to build/lib.linux.... , which will be a
# platform dependent build folder created by the "build" command of
# setuptools. Only the contents of this folder are installed in the
# "install" command by default.
# We only make this copy for Caffe2's pybind extensions
caffe2_pybind_exts = [
"caffe2.python.caffe2_pybind11_state",
"caffe2.python.caffe2_pybind11_state_gpu",
"caffe2.python.caffe2_pybind11_state_hip",
]
i = 0
while i < len(self.extensions):
ext = self.extensions[i]
if ext.name not in caffe2_pybind_exts:
i += 1
continue
fullname = self.get_ext_fullname(ext.name)
filename = self.get_ext_filename(fullname)
report(f"\nCopying extension {ext.name}")
relative_site_packages = (
sysconfig.get_path("purelib")
.replace(sysconfig.get_path("data"), "")
.lstrip(os.path.sep)
)
src = os.path.join("torch", relative_site_packages, filename)
if not os.path.exists(src):
report(f"{src} does not exist")
del self.extensions[i]
else:
dst = os.path.join(os.path.realpath(self.build_lib), filename)
report(f"Copying {ext.name} from {src} to {dst}")
dst_dir = os.path.dirname(dst)
if not os.path.exists(dst_dir):
os.makedirs(dst_dir)
self.copy_file(src, dst)
i += 1
# Copy functorch extension
for i, ext in enumerate(self.extensions):
if ext.name != "functorch._C":
continue
fullname = self.get_ext_fullname(ext.name)
filename = self.get_ext_filename(fullname)
fileext = os.path.splitext(filename)[1]
src = os.path.join(os.path.dirname(filename), "functorch" + fileext)
dst = os.path.join(os.path.realpath(self.build_lib), filename)
if os.path.exists(src):
report(f"Copying {ext.name} from {src} to {dst}")
dst_dir = os.path.dirname(dst)
if not os.path.exists(dst_dir):
os.makedirs(dst_dir)
self.copy_file(src, dst)
# Copy nvfuser extension
for i, ext in enumerate(self.extensions):
if ext.name != "nvfuser._C":
continue
fullname = self.get_ext_fullname(ext.name)
filename = self.get_ext_filename(fullname)
fileext = os.path.splitext(filename)[1]
src = os.path.join(os.path.dirname(filename), "nvfuser" + fileext)
dst = os.path.join(os.path.realpath(self.build_lib), filename)
if os.path.exists(src):
report(f"Copying {ext.name} from {src} to {dst}")
dst_dir = os.path.dirname(dst)
if not os.path.exists(dst_dir):
os.makedirs(dst_dir)
self.copy_file(src, dst)
setuptools.command.build_ext.build_ext.build_extensions(self)
def get_outputs(self):
outputs = setuptools.command.build_ext.build_ext.get_outputs(self)
outputs.append(os.path.join(self.build_lib, "caffe2"))
report(f"setup.py::get_outputs returning {outputs}")
return outputs
def create_compile_commands(self):
def load(filename):
with open(filename) as f:
return json.load(f)
ninja_files = glob.glob("build/*compile_commands.json")
cmake_files = glob.glob("torch/lib/build/*/compile_commands.json")
all_commands = [entry for f in ninja_files + cmake_files for entry in load(f)]
# cquery does not like c++ compiles that start with gcc.
# It forgets to include the c++ header directories.
# We can work around this by replacing the gcc calls that python
# setup.py generates with g++ calls instead
for command in all_commands:
if command["command"].startswith("gcc "):
command["command"] = "g++ " + command["command"][4:]
new_contents = json.dumps(all_commands, indent=2)
contents = ""
if os.path.exists("compile_commands.json"):
with open("compile_commands.json") as f:
contents = f.read()
if contents != new_contents:
with open("compile_commands.json", "w") as f:
f.write(new_contents)
class concat_license_files:
"""Merge LICENSE and LICENSES_BUNDLED.txt as a context manager
LICENSE is the main PyTorch license, LICENSES_BUNDLED.txt is auto-generated
from all the licenses found in ./third_party/. We concatenate them so there
is a single license file in the sdist and wheels with all of the necessary
licensing info.
"""
def __init__(self, include_files=False):
self.f1 = "LICENSE"
self.f2 = "third_party/LICENSES_BUNDLED.txt"
self.include_files = include_files
def __enter__(self):
"""Concatenate files"""
old_path = sys.path
sys.path.append(third_party_path)
try:
from build_bundled import create_bundled
finally:
sys.path = old_path
with open(self.f1) as f1:
self.bsd_text = f1.read()
with open(self.f1, "a") as f1:
f1.write("\n\n")
create_bundled(
os.path.relpath(third_party_path), f1, include_files=self.include_files
)
def __exit__(self, exception_type, exception_value, traceback):
"""Restore content of f1"""
with open(self.f1, "w") as f:
f.write(self.bsd_text)
try:
from wheel.bdist_wheel import bdist_wheel
except ImportError:
# This is useful when wheel is not installed and bdist_wheel is not
# specified on the command line. If it _is_ specified, parsing the command
# line will fail before wheel_concatenate is needed
wheel_concatenate = None
else:
# Need to create the proper LICENSE.txt for the wheel
class wheel_concatenate(bdist_wheel):
"""check submodules on sdist to prevent incomplete tarballs"""
def run(self):
with concat_license_files(include_files=True):
super().run()
class install(setuptools.command.install.install):
def run(self):
super().run()
class clean(setuptools.Command):
user_options = []
def initialize_options(self):
pass
def finalize_options(self):
pass
def run(self):
import glob
import re
with open(".gitignore") as f:
ignores = f.read()
pat = re.compile(r"^#( BEGIN NOT-CLEAN-FILES )?")
for wildcard in filter(None, ignores.split("\n")):
match = pat.match(wildcard)
if match:
if match.group(1):
# Marker is found and stop reading .gitignore.
break
# Ignore lines which begin with '#'.
else:
# Don't remove absolute paths from the system
wildcard = wildcard.lstrip("./")
for filename in glob.glob(wildcard):
try:
os.remove(filename)
except OSError:
shutil.rmtree(filename, ignore_errors=True)
class sdist(setuptools.command.sdist.sdist):
def run(self):
with concat_license_files():
super().run()
def get_cmake_cache_vars():
try:
return defaultdict(lambda: False, cmake.get_cmake_cache_variables())
except FileNotFoundError:
# CMakeCache.txt does not exist. Probably running "python setup.py clean" over a clean directory.
return defaultdict(lambda: False)
def configure_extension_build():
r"""Configures extension build options according to system environment and user's choice.
Returns:
The input to parameters ext_modules, cmdclass, packages, and entry_points as required in setuptools.setup.
"""
cmake_cache_vars = get_cmake_cache_vars()
################################################################################
# Configure compile flags
################################################################################
library_dirs = []
extra_install_requires = []
if IS_WINDOWS:
# /NODEFAULTLIB makes sure we only link to DLL runtime
# and matches the flags set for protobuf and ONNX
extra_link_args = ["/NODEFAULTLIB:LIBCMT.LIB"]
# /MD links against DLL runtime
# and matches the flags set for protobuf and ONNX
# /EHsc is about standard C++ exception handling
extra_compile_args = ["/MD", "/FS", "/EHsc"]
else:
extra_link_args = []
extra_compile_args = [
"-Wall",
"-Wextra",
"-Wno-strict-overflow",
"-Wno-unused-parameter",
"-Wno-missing-field-initializers",
"-Wno-unknown-pragmas",
# This is required for Python 2 declarations that are deprecated in 3.
"-Wno-deprecated-declarations",
# Python 2.6 requires -fno-strict-aliasing, see
# http://legacy.python.org/dev/peps/pep-3123/
# We also depend on it in our code (even Python 3).
"-fno-strict-aliasing",
# Clang has an unfixed bug leading to spurious missing
# braces warnings, see
# https://bugs.llvm.org/show_bug.cgi?id=21629
"-Wno-missing-braces",
]
library_dirs.append(lib_path)
main_compile_args = []
main_libraries = ["torch_python"]
main_link_args = []
main_sources = ["torch/csrc/stub.c"]
if cmake_cache_vars["USE_CUDA"]:
library_dirs.append(os.path.dirname(cmake_cache_vars["CUDA_CUDA_LIB"]))
if build_type.is_debug():
if IS_WINDOWS:
extra_compile_args.append("/Z7")
extra_link_args.append("/DEBUG:FULL")
else:
extra_compile_args += ["-O0", "-g"]
extra_link_args += ["-O0", "-g"]
if build_type.is_rel_with_deb_info():
if IS_WINDOWS:
extra_compile_args.append("/Z7")
extra_link_args.append("/DEBUG:FULL")
else:
extra_compile_args += ["-g"]
extra_link_args += ["-g"]
# special CUDA 11.7 package that requires installation of cuda runtime, cudnn and cublas
pytorch_extra_install_requirements = os.getenv(
"PYTORCH_EXTRA_INSTALL_REQUIREMENTS", ""
)
if pytorch_extra_install_requirements:
report(
f"pytorch_extra_install_requirements: {pytorch_extra_install_requirements}"
)
extra_install_requires += pytorch_extra_install_requirements.split("|")
# Cross-compile for M1
if IS_DARWIN:
macos_target_arch = os.getenv("CMAKE_OSX_ARCHITECTURES", "")
if macos_target_arch in ["arm64", "x86_64"]:
macos_sysroot_path = os.getenv("CMAKE_OSX_SYSROOT")
if macos_sysroot_path is None:
macos_sysroot_path = (
subprocess.check_output(
["xcrun", "--show-sdk-path", "--sdk", "macosx"]
)
.decode("utf-8")
.strip()
)
extra_compile_args += [
"-arch",
macos_target_arch,
"-isysroot",
macos_sysroot_path,
]
extra_link_args += ["-arch", macos_target_arch]
def make_relative_rpath_args(path):
if IS_DARWIN:
return ["-Wl,-rpath,@loader_path/" + path]