-
Notifications
You must be signed in to change notification settings - Fork 2
/
setup.py
236 lines (203 loc) · 12.5 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
import torch
from torch.utils import cpp_extension
from setuptools import setup, find_packages
import subprocess
import sys
import warnings
import os
# ninja build does not work unless include_dirs are abs path
this_dir = os.path.dirname(os.path.abspath(__file__))
def get_cuda_bare_metal_version(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]
return raw_output, bare_metal_major, bare_metal_minor
if not torch.cuda.is_available():
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, it will cross-compile for Volta (compute capability 7.0), Turing (compute capability 7.5),\n'
'and, if the CUDA version is >= 11.0, Ampere (compute capability 8.0).\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:
_, bare_metal_major, _ = get_cuda_bare_metal_version(cpp_extension.CUDA_HOME)
if int(bare_metal_major) == 11:
os.environ["TORCH_CUDA_ARCH_LIST"] = "7.0;7.5;8.0"
else:
os.environ["TORCH_CUDA_ARCH_LIST"] = "7.0;7.5"
print("\n\ntorch.__version__ = {}\n\n".format(torch.__version__))
TORCH_MAJOR = int(torch.__version__.split('.')[0])
TORCH_MINOR = int(torch.__version__.split('.')[1])
if not (TORCH_MAJOR >= 1 and TORCH_MINOR >= 4):
raise RuntimeError("Requires Pytorch 1.4 or newer.\n" +
"The latest stable release can be obtained from https://pytorch.org/")
cmdclass = {}
ext_modules = []
extras = {}
def get_cuda_bare_metal_version(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]
return raw_output, bare_metal_major, bare_metal_minor
def check_cuda_torch_binary_vs_bare_metal(cuda_dir):
raw_output, bare_metal_major, bare_metal_minor = get_cuda_bare_metal_version(cuda_dir)
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))
from torch.utils.cpp_extension import CUDAExtension
from torch.utils.cpp_extension import BuildExtension
cmdclass['build_ext'] = BuildExtension
if torch.utils.cpp_extension.CUDA_HOME is None:
raise RuntimeError("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.")
check_cuda_torch_binary_vs_bare_metal(torch.utils.cpp_extension.CUDA_HOME)
generator_flag = []
torch_dir = torch.__path__[0]
if os.path.exists(os.path.join(torch_dir, 'include', 'ATen', 'CUDAGenerator.h')):
generator_flag = ['-DOLD_GENERATOR']
ext_modules.append(
CUDAExtension(name='fused_rounding',
sources=['csrc/rounding/interface.cpp',
'csrc/rounding/fp32_to_bf16.cu'],
include_dirs=[os.path.join(this_dir, 'csrc')],
extra_compile_args={'cxx': ['-O3',] + generator_flag,
'nvcc':['-O3', '--use_fast_math',
'-gencode', 'arch=compute_70,code=sm_70',
'-gencode', 'arch=compute_80,code=sm_80',
'-U__CUDA_NO_HALF_OPERATORS__',
'-U__CUDA_NO_BFLOAT16_OPERATORS__',
'-U__CUDA_NO_HALF_CONVERSIONS__',
'-U__CUDA_NO_BFLOAT16_CONVERSIONS__',
'--expt-relaxed-constexpr',
'--expt-extended-lambda'] + generator_flag}))
ext_modules.append(
CUDAExtension(name='multi_tensor',
sources=['csrc/multi_tensor/interface.cpp',
'csrc/multi_tensor/multi_tensor_l2norm_kernel.cu'],
include_dirs=[os.path.join(this_dir, 'csrc')],
extra_compile_args={'cxx': ['-O3'],
'nvcc':['-O3', '--use_fast_math',
'-gencode', 'arch=compute_70,code=sm_70',
'-gencode', 'arch=compute_80,code=sm_80',
'-U__CUDA_NO_HALF_OPERATORS__',
'-U__CUDA_NO_BFLOAT16_OPERATORS__',
'-U__CUDA_NO_HALF_CONVERSIONS__',
'-U__CUDA_NO_BFLOAT16_CONVERSIONS__',
'--expt-relaxed-constexpr',
'--expt-extended-lambda']
}))
ext_modules.append(
CUDAExtension(name='fused_xentropy_cuda',
sources=['csrc/xentropy/interface.cpp',
'csrc/xentropy/xentropy_kernel.cu'],
include_dirs=[os.path.join(this_dir, 'csrc')],
extra_compile_args={'cxx': ['-O3'],
'nvcc':['-O3']}))
ext_modules.append(
CUDAExtension(name='fused_adam_cuda_v2',
sources=['csrc/adam/interface.cpp',
'csrc/adam/adam_kernel.cu'],
include_dirs=[os.path.join(this_dir, 'csrc')],
extra_compile_args={'cxx': ['-O3'],
'nvcc':['-O3', '--use_fast_math']}))
ext_modules.append(
CUDAExtension(name='fused_softmax_dropout_fast_cuda',
sources=['csrc/softmax_dropout/interface.cpp',
'csrc/softmax_dropout/softmax_dropout_kernel.cu'],
include_dirs=[os.path.join(this_dir, 'csrc')],
extra_compile_args={'cxx': ['-O3',] + generator_flag,
'nvcc':['-O3', '--use_fast_math',
'-gencode', 'arch=compute_70,code=sm_70',
'-gencode', 'arch=compute_80,code=sm_80',
'-U__CUDA_NO_HALF_OPERATORS__',
'-U__CUDA_NO_BFLOAT16_OPERATORS__',
'-U__CUDA_NO_HALF_CONVERSIONS__',
'-U__CUDA_NO_BFLOAT16_CONVERSIONS__',
'--expt-relaxed-constexpr',
'--expt-extended-lambda'] + generator_flag}))
ext_modules.append(
CUDAExtension(name='fused_rmsnorm_fast_cuda',
sources=['csrc/rmsnorm/interface.cpp',
'csrc/rmsnorm/rmsnorm.cu'],
include_dirs=[os.path.join(this_dir, 'csrc')],
extra_compile_args={'cxx': ['-O3',] + generator_flag,
'nvcc':['-O3', '--use_fast_math',
'-gencode', 'arch=compute_70,code=sm_70',
'-gencode', 'arch=compute_80,code=sm_80',
'-U__CUDA_NO_HALF_OPERATORS__',
'-U__CUDA_NO_BFLOAT16_OPERATORS__',
'-U__CUDA_NO_HALF_CONVERSIONS__',
'-U__CUDA_NO_BFLOAT16_CONVERSIONS__',
'--expt-relaxed-constexpr',
'--expt-extended-lambda'] + generator_flag}))
ext_modules.append(
CUDAExtension(name='fused_rmsnorm_backward_gamma_cuda',
sources=['csrc/rmsnorm/interface_gamma.cpp',
'csrc/rmsnorm/rmsnorm_backward.cu'],
include_dirs=[os.path.join(this_dir, 'csrc')],
extra_compile_args={'cxx': ['-O3',] + generator_flag,
'nvcc':['-O3', '--use_fast_math', '-maxrregcount=50',
'-gencode', 'arch=compute_70,code=sm_70',
'-gencode', 'arch=compute_80,code=sm_80',
'-U__CUDA_NO_HALF_OPERATORS__',
'-U__CUDA_NO_BFLOAT16_OPERATORS__',
'-U__CUDA_NO_HALF_CONVERSIONS__',
'-U__CUDA_NO_BFLOAT16_CONVERSIONS__',
'--expt-relaxed-constexpr',
'--expt-extended-lambda'] + generator_flag}))
ext_modules.append(
CUDAExtension(name='fused_layernorm_fast_cuda',
sources=['csrc/layernorm/interface.cpp',
'csrc/layernorm/layernorm.cu'],
include_dirs=[os.path.join(this_dir, 'csrc')],
extra_compile_args={'cxx': ['-O3',] + generator_flag,
'nvcc':['-O3', '--use_fast_math',
'-gencode', 'arch=compute_70,code=sm_70',
'-gencode', 'arch=compute_80,code=sm_80',
'-U__CUDA_NO_HALF_OPERATORS__',
'-U__CUDA_NO_BFLOAT16_OPERATORS__',
'-U__CUDA_NO_HALF_CONVERSIONS__',
'-U__CUDA_NO_BFLOAT16_CONVERSIONS__',
'--expt-relaxed-constexpr',
'--expt-extended-lambda'] + generator_flag}))
ext_modules.append(
CUDAExtension(name='fused_layernorm_backward_gamma_beta_cuda',
sources=['csrc/layernorm/interface_gamma_beta.cpp',
'csrc/layernorm/layernorm_backward.cu'],
include_dirs=[os.path.join(this_dir, 'csrc')],
extra_compile_args={'cxx': ['-O3',] + generator_flag,
'nvcc':['-O3', '--use_fast_math', '-maxrregcount=50',
'-gencode', 'arch=compute_70,code=sm_70',
'-gencode', 'arch=compute_80,code=sm_80',
'-U__CUDA_NO_HALF_OPERATORS__',
'-U__CUDA_NO_BFLOAT16_OPERATORS__',
'-U__CUDA_NO_HALF_CONVERSIONS__',
'-U__CUDA_NO_BFLOAT16_CONVERSIONS__',
'--expt-relaxed-constexpr',
'--expt-extended-lambda'] + generator_flag}))
setup(
name='fused_ops',
version='0.1',
packages=find_packages(exclude=('build',
'csrc',
'include',
'tests',
'dist',
'docs',
'tests',
'examples',)),
description='Fused ops',
ext_modules=ext_modules,
cmdclass=cmdclass,
extras_require=extras,
)