forked from NVIDIA/apex
-
Notifications
You must be signed in to change notification settings - Fork 0
/
setup.py
276 lines (247 loc) · 15.8 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
import torch
from setuptools import setup, find_packages
import subprocess
import sys
import warnings
import os
if not torch.cuda.is_available():
# https://github.com/NVIDIA/apex/issues/486
# Extension builds after https://github.com/pytorch/pytorch/pull/23408 attempt to query torch.cuda.get_device_capability(),
# which will fail if you are compiling in an environment without visible GPUs (e.g. during an nvidia-docker build command).
print('\nWarning: Torch did not find available GPUs on this system.\n',
'If your intention is to cross-compile, this is not an error.\n'
'By default, Apex will cross-compile for Pascal (compute capabilities 6.0, 6.1, 6.2),\n'
'Volta (compute capability 7.0), and Turing (compute capability 7.5).\n'
'If you wish to cross-compile for a single specific architecture,\n'
'export TORCH_CUDA_ARCH_LIST="compute capability" before running setup.py.\n')
if os.environ.get("TORCH_CUDA_ARCH_LIST", None) is None:
os.environ["TORCH_CUDA_ARCH_LIST"] = "6.0;6.1;6.2;7.0;7.5"
print("torch.__version__ = ", torch.__version__)
TORCH_MAJOR = int(torch.__version__.split('.')[0])
TORCH_MINOR = int(torch.__version__.split('.')[1])
if TORCH_MAJOR == 0 and TORCH_MINOR < 4:
raise RuntimeError("Apex requires Pytorch 0.4 or newer.\n" +
"The latest stable release can be obtained from https://pytorch.org/")
cmdclass = {}
ext_modules = []
extras = {}
if "--pyprof" in sys.argv:
with open('requirements.txt') as f:
required_packages = f.read().splitlines()
extras['pyprof'] = required_packages
try:
sys.argv.remove("--pyprof")
except:
pass
else:
warnings.warn("Option --pyprof not specified. Not installing PyProf dependencies!")
if "--cpp_ext" in sys.argv or "--cuda_ext" in sys.argv:
if TORCH_MAJOR == 0:
raise RuntimeError("--cpp_ext requires Pytorch 1.0 or later, "
"found torch.__version__ = {}".format(torch.__version__))
from torch.utils.cpp_extension import BuildExtension
cmdclass['build_ext'] = BuildExtension
if "--cpp_ext" in sys.argv:
from torch.utils.cpp_extension import CppExtension
sys.argv.remove("--cpp_ext")
ext_modules.append(
CppExtension('apex_C',
['csrc/flatten_unflatten.cpp',]))
def check_cuda_torch_binary_vs_bare_metal(cuda_dir):
raw_output = subprocess.check_output([cuda_dir + "/bin/nvcc", "-V"], universal_newlines=True)
output = raw_output.split()
release_idx = output.index("release") + 1
release = output[release_idx].split(".")
bare_metal_major = release[0]
bare_metal_minor = release[1][0]
torch_binary_major = torch.version.cuda.split(".")[0]
torch_binary_minor = torch.version.cuda.split(".")[1]
print("\nCompiling cuda extensions with")
print(raw_output + "from " + cuda_dir + "/bin\n")
if (bare_metal_major != torch_binary_major) or (bare_metal_minor != torch_binary_minor):
raise RuntimeError("Cuda extensions are being compiled with a version of Cuda that does " +
"not match the version used to compile Pytorch binaries. " +
"Pytorch binaries were compiled with Cuda {}.\n".format(torch.version.cuda) +
"In some cases, a minor-version mismatch will not cause later errors: " +
"https://github.com/NVIDIA/apex/pull/323#discussion_r287021798. "
"You can try commenting out this check (at your own risk).")
# Set up macros for forward/backward compatibility hack around
# https://github.com/pytorch/pytorch/commit/4404762d7dd955383acee92e6f06b48144a0742e
# and
# https://github.com/NVIDIA/apex/issues/456
# https://github.com/pytorch/pytorch/commit/eb7b39e02f7d75c26d8a795ea8c7fd911334da7e#diff-4632522f237f1e4e728cb824300403ac
version_ge_1_1 = []
if (TORCH_MAJOR > 1) or (TORCH_MAJOR == 1 and TORCH_MINOR > 0):
version_ge_1_1 = ['-DVERSION_GE_1_1']
version_ge_1_3 = []
if (TORCH_MAJOR > 1) or (TORCH_MAJOR == 1 and TORCH_MINOR > 2):
version_ge_1_3 = ['-DVERSION_GE_1_3']
version_dependent_macros = version_ge_1_1 + version_ge_1_3
if "--cuda_ext" in sys.argv:
from torch.utils.cpp_extension import CUDAExtension
sys.argv.remove("--cuda_ext")
if torch.utils.cpp_extension.CUDA_HOME is None:
raise RuntimeError("--cuda_ext was requested, but nvcc was not found. Are you sure your environment has nvcc available? If you're installing within a container from https://hub.docker.com/r/pytorch/pytorch, only images whose names contain 'devel' will provide nvcc.")
else:
check_cuda_torch_binary_vs_bare_metal(torch.utils.cpp_extension.CUDA_HOME)
ext_modules.append(
CUDAExtension(name='amp_C',
sources=['csrc/amp_C_frontend.cpp',
'csrc/multi_tensor_sgd_kernel.cu',
'csrc/multi_tensor_scale_kernel.cu',
'csrc/multi_tensor_axpby_kernel.cu',
'csrc/multi_tensor_l2norm_kernel.cu',
'csrc/multi_tensor_lamb_stage_1.cu',
'csrc/multi_tensor_lamb_stage_2.cu',
'csrc/multi_tensor_adam.cu',
'csrc/multi_tensor_novograd.cu',
'csrc/multi_tensor_lamb.cu'],
extra_compile_args={'cxx': ['-O3'] + version_dependent_macros,
'nvcc':['-lineinfo',
'-O3',
# '--resource-usage',
'--use_fast_math'] + version_dependent_macros}))
ext_modules.append(
CUDAExtension(name='syncbn',
sources=['csrc/syncbn.cpp',
'csrc/welford.cu'],
extra_compile_args={'cxx': ['-O3'] + version_dependent_macros,
'nvcc':['-O3'] + version_dependent_macros}))
ext_modules.append(
CUDAExtension(name='fused_layer_norm_cuda',
sources=['csrc/layer_norm_cuda.cpp',
'csrc/layer_norm_cuda_kernel.cu'],
extra_compile_args={'cxx': ['-O3'] + version_dependent_macros,
'nvcc':['-maxrregcount=50',
'-O3',
'--use_fast_math'] + version_dependent_macros}))
if "--bnp" in sys.argv:
from torch.utils.cpp_extension import CUDAExtension
sys.argv.remove("--bnp")
from torch.utils.cpp_extension import BuildExtension
cmdclass['build_ext'] = BuildExtension
if torch.utils.cpp_extension.CUDA_HOME is None:
raise RuntimeError("--bnp was requested, but nvcc was not found. Are you sure your environment has nvcc available? If you're installing within a container from https://hub.docker.com/r/pytorch/pytorch, only images whose names contain 'devel' will provide nvcc.")
else:
ext_modules.append(
CUDAExtension(name='bnp',
sources=['apex/contrib/csrc/groupbn/batch_norm.cu',
'apex/contrib/csrc/groupbn/ipc.cu',
'apex/contrib/csrc/groupbn/interface.cpp',
'apex/contrib/csrc/groupbn/batch_norm_add_relu.cu'],
include_dirs=['csrc'],
extra_compile_args={'cxx': [] + version_dependent_macros,
'nvcc':['-DCUDA_HAS_FP16=1',
'-D__CUDA_NO_HALF_OPERATORS__',
'-D__CUDA_NO_HALF_CONVERSIONS__',
'-D__CUDA_NO_HALF2_OPERATORS__'] + version_dependent_macros}))
if "--xentropy" in sys.argv:
from torch.utils.cpp_extension import CUDAExtension
sys.argv.remove("--xentropy")
from torch.utils.cpp_extension import BuildExtension
cmdclass['build_ext'] = BuildExtension
if torch.utils.cpp_extension.CUDA_HOME is None:
raise RuntimeError("--xentropy was requested, but nvcc was not found. Are you sure your environment has nvcc available? If you're installing within a container from https://hub.docker.com/r/pytorch/pytorch, only images whose names contain 'devel' will provide nvcc.")
else:
ext_modules.append(
CUDAExtension(name='xentropy_cuda',
sources=['apex/contrib/csrc/xentropy/interface.cpp',
'apex/contrib/csrc/xentropy/xentropy_kernel.cu'],
include_dirs=['csrc'],
extra_compile_args={'cxx': ['-O3'] + version_dependent_macros,
'nvcc':['-O3'] + version_dependent_macros}))
if "--deprecated_fused_adam" in sys.argv:
from torch.utils.cpp_extension import CUDAExtension
sys.argv.remove("--deprecated_fused_adam")
from torch.utils.cpp_extension import BuildExtension
cmdclass['build_ext'] = BuildExtension
if torch.utils.cpp_extension.CUDA_HOME is None:
raise RuntimeError("--deprecated_fused_adam was requested, but nvcc was not found. Are you sure your environment has nvcc available? If you're installing within a container from https://hub.docker.com/r/pytorch/pytorch, only images whose names contain 'devel' will provide nvcc.")
else:
ext_modules.append(
CUDAExtension(name='fused_adam_cuda',
sources=['apex/contrib/csrc/optimizers/fused_adam_cuda.cpp',
'apex/contrib/csrc/optimizers/fused_adam_cuda_kernel.cu'],
include_dirs=['csrc'],
extra_compile_args={'cxx': ['-O3',] + version_dependent_macros,
'nvcc':['-O3',
'--use_fast_math'] + version_dependent_macros}))
if "--fast_multihead_attn" in sys.argv:
from torch.utils.cpp_extension import CUDAExtension
sys.argv.remove("--fast_multihead_attn")
from torch.utils.cpp_extension import BuildExtension
cmdclass['build_ext'] = BuildExtension
if torch.utils.cpp_extension.CUDA_HOME is None:
raise RuntimeError("--fast_multihead_attn was requested, but nvcc was not found. Are you sure your environment has nvcc available? If you're installing within a container from https://hub.docker.com/r/pytorch/pytorch, only images whose names contain 'devel' will provide nvcc.")
else:
import subprocess
subprocess.run(["git", "submodule", "update", "--init", "apex/contrib/csrc/multihead_attn/cutlass"])
ext_modules.append(
CUDAExtension(name='fast_self_multihead_attn',
sources=['apex/contrib/csrc/multihead_attn/self_multihead_attn.cpp',
'apex/contrib/csrc/multihead_attn/self_multihead_attn_cuda.cu'],
extra_compile_args={'cxx': ['-O3',] + version_dependent_macros,
'nvcc':['-O3',
'-gencode', 'arch=compute_70,code=sm_70',
'-I./apex/contrib/csrc/multihead_attn/cutlass/',
'-U__CUDA_NO_HALF_OPERATORS__',
'-U__CUDA_NO_HALF_CONVERSIONS__',
'--expt-relaxed-constexpr',
'--expt-extended-lambda',
'--use_fast_math'] + version_dependent_macros}))
ext_modules.append(
CUDAExtension(name='fast_self_multihead_attn_norm_add',
sources=['apex/contrib/csrc/multihead_attn/self_multihead_attn_norm_add.cpp',
'apex/contrib/csrc/multihead_attn/self_multihead_attn_norm_add_cuda.cu'],
extra_compile_args={'cxx': ['-O3',] + version_dependent_macros,
'nvcc':['-O3',
'-gencode', 'arch=compute_70,code=sm_70',
'-I./apex/contrib/csrc/multihead_attn/cutlass/',
'-U__CUDA_NO_HALF_OPERATORS__',
'-U__CUDA_NO_HALF_CONVERSIONS__',
'--expt-relaxed-constexpr',
'--expt-extended-lambda',
'--use_fast_math'] + version_dependent_macros}))
ext_modules.append(
CUDAExtension(name='fast_encdec_multihead_attn',
sources=['apex/contrib/csrc/multihead_attn/encdec_multihead_attn.cpp',
'apex/contrib/csrc/multihead_attn/encdec_multihead_attn_cuda.cu'],
extra_compile_args={'cxx': ['-O3',] + version_dependent_macros,
'nvcc':['-O3',
'-gencode', 'arch=compute_70,code=sm_70',
'-I./apex/contrib/csrc/multihead_attn/cutlass/',
'-U__CUDA_NO_HALF_OPERATORS__',
'-U__CUDA_NO_HALF_CONVERSIONS__',
'--expt-relaxed-constexpr',
'--expt-extended-lambda',
'--use_fast_math'] + version_dependent_macros}))
ext_modules.append(
CUDAExtension(name='fast_encdec_multihead_attn_norm_add',
sources=['apex/contrib/csrc/multihead_attn/encdec_multihead_attn_norm_add.cpp',
'apex/contrib/csrc/multihead_attn/encdec_multihead_attn_norm_add_cuda.cu'],
extra_compile_args={'cxx': ['-O3',] + version_dependent_macros,
'nvcc':['-O3',
'-gencode', 'arch=compute_70,code=sm_70',
'-I./apex/contrib/csrc/multihead_attn/cutlass/',
'-U__CUDA_NO_HALF_OPERATORS__',
'-U__CUDA_NO_HALF_CONVERSIONS__',
'--expt-relaxed-constexpr',
'--expt-extended-lambda',
'--use_fast_math'] + version_dependent_macros}))
setup(
name='apex',
version='0.1',
packages=find_packages(exclude=('build',
'csrc',
'include',
'tests',
'dist',
'docs',
'tests',
'examples',
'apex.egg-info',)),
description='PyTorch Extensions written by NVIDIA',
ext_modules=ext_modules,
cmdclass=cmdclass,
extras_require=extras,
)