setup.py 30.7 KB
Newer Older
1
import torch
ptrblck's avatar
ptrblck committed
2
from torch.utils import cpp_extension
3
from setuptools import setup, find_packages
mcarilli's avatar
mcarilli committed
4
import subprocess
5

jjsjann123's avatar
jjsjann123 committed
6
import sys
Marek Kolodziej's avatar
Marek Kolodziej committed
7
import warnings
mcarilli's avatar
mcarilli committed
8
import os
jjsjann123's avatar
jjsjann123 committed
9

10
11
12
# ninja build does not work unless include_dirs are abs path
this_dir = os.path.dirname(os.path.abspath(__file__))

ptrblck's avatar
ptrblck committed
13
14
15
16
17
18
19
20
21
22
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

Jithun Nair's avatar
Jithun Nair committed
23
24
25
26
print("\n\ntorch.__version__  = {}\n\n".format(torch.__version__))
TORCH_MAJOR = int(torch.__version__.split('.')[0])
TORCH_MINOR = int(torch.__version__.split('.')[1])

27
28
def check_if_rocm_pytorch():
    is_rocm_pytorch = False
Jithun Nair's avatar
Jithun Nair committed
29
    if TORCH_MAJOR > 1 or (TORCH_MAJOR == 1 and TORCH_MINOR >= 5):
30
31
32
33
34
35
36
        from torch.utils.cpp_extension import ROCM_HOME
        is_rocm_pytorch = True if ((torch.version.hip is not None) and (ROCM_HOME is not None)) else False

    return is_rocm_pytorch

IS_ROCM_PYTORCH = check_if_rocm_pytorch()

37
38
39
40
41
42
if IS_ROCM_PYTORCH:
    rocm_include_dirs = ["/opt/rocm/include/hiprand", "/opt/rocm/include/rocrand"]
else:
    rocm_include_dirs = []

include_dirs=[os.path.join(this_dir, 'csrc')] + rocm_include_dirs
43
if not torch.cuda.is_available() and not IS_ROCM_PYTORCH:
mcarilli's avatar
mcarilli committed
44
45
46
47
48
49
    # 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'
ptrblck's avatar
ptrblck committed
50
51
          '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'
mcarilli's avatar
mcarilli committed
52
53
54
          '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:
ptrblck's avatar
ptrblck committed
55
56
57
58
59
        _, bare_metal_major, _ = get_cuda_bare_metal_version(cpp_extension.CUDA_HOME)
        if int(bare_metal_major) == 11:
            os.environ["TORCH_CUDA_ARCH_LIST"] = "6.0;6.1;6.2;7.0;7.5;8.0"
        else:
            os.environ["TORCH_CUDA_ARCH_LIST"] = "6.0;6.1;6.2;7.0;7.5"
60
61
62
63
64
elif not torch.cuda.is_available() and IS_ROCM_PYTORCH:
    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 the same gfx targets\n'
          'used by default in ROCm PyTorch\n')
65

66
if TORCH_MAJOR == 0 and TORCH_MINOR < 4:
Michael Carilli's avatar
Michael Carilli committed
67
      raise RuntimeError("Apex requires Pytorch 0.4 or newer.\n" +
68
69
                         "The latest stable release can be obtained from https://pytorch.org/")

jjsjann123's avatar
jjsjann123 committed
70
71
72
cmdclass = {}
ext_modules = []

ptrblck's avatar
ptrblck committed
73
extras = {}
Marek Kolodziej's avatar
Marek Kolodziej committed
74
if "--pyprof" in sys.argv:
75
76
77
78
79
    string = "\n\nPyprof has been moved to its own dedicated repository and will " + \
             "soon be removed from Apex.  Please visit\n" + \
             "https://github.com/NVIDIA/PyProf\n" + \
             "for the latest version."
    warnings.warn(string, DeprecationWarning)
Marek Kolodziej's avatar
Marek Kolodziej committed
80
81
    with open('requirements.txt') as f:
        required_packages = f.read().splitlines()
ptrblck's avatar
ptrblck committed
82
        extras['pyprof'] = required_packages
Marek Kolodziej's avatar
Marek Kolodziej committed
83
84
85
86
87
88
89
    try:
        sys.argv.remove("--pyprof")
    except:
        pass
else:
    warnings.warn("Option --pyprof not specified. Not installing PyProf dependencies!")

90
if "--cpp_ext" in sys.argv or "--cuda_ext" in sys.argv:
Michael Carilli's avatar
Michael Carilli committed
91
92
    if TORCH_MAJOR == 0:
        raise RuntimeError("--cpp_ext requires Pytorch 1.0 or later, "
93
                           "found torch.__version__ = {}".format(torch.__version__))
94
95
96
97
98
99
100
101
102
103
    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',]))

ptrblck's avatar
ptrblck committed
104
def get_cuda_bare_metal_version(cuda_dir):
mcarilli's avatar
mcarilli committed
105
106
107
108
109
110
    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]
ptrblck's avatar
ptrblck committed
111
112
113
114
115

    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)
mcarilli's avatar
mcarilli committed
116
117
118
119
120
121
122
    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):
Michael Carilli's avatar
Michael Carilli committed
123
124
125
126
127
128
        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).")
mcarilli's avatar
mcarilli committed
129

mcarilli's avatar
mcarilli committed
130
131
132
133
134
135
136
137
138
139
140
# 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']
141
142
143
144
version_ge_1_5 = []
if (TORCH_MAJOR > 1) or (TORCH_MAJOR == 1 and TORCH_MINOR > 4):
    version_ge_1_5 = ['-DVERSION_GE_1_5']
version_dependent_macros = version_ge_1_1 + version_ge_1_3 + version_ge_1_5
mcarilli's avatar
mcarilli committed
145

146
147
148
149
150
151
152
if "--distributed_adam" in sys.argv:
    from torch.utils.cpp_extension import CUDAExtension
    sys.argv.remove("--distributed_adam")

    from torch.utils.cpp_extension import BuildExtension
    cmdclass['build_ext'] = BuildExtension

153
    if torch.utils.cpp_extension.CUDA_HOME is None and not IS_ROCM_PYTORCH:
154
155
        raise RuntimeError("--distributed_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:
156
157
        nvcc_args_adam = ['-O3', '--use_fast_math'] + version_dependent_macros
        hipcc_args_adam = ['-O3'] + version_dependent_macros
158
159
        ext_modules.append(
            CUDAExtension(name='distributed_adam_cuda',
160
161
162
                          sources=['./apex/contrib/csrc/optimizers/multi_tensor_distopt_adam.cpp',
                                   './apex/contrib/csrc/optimizers/multi_tensor_distopt_adam_kernel.cu'],
                          include_dirs=include_dirs + [this_dir + '/apex/contrib/csrc/optimizers/'],
163
                          extra_compile_args={'cxx': ['-O3',] + version_dependent_macros,
164
                                              'nvcc':nvcc_args_adam if not IS_ROCM_PYTORCH else hipcc_args_adam}))
165

166
167
168
169
170
171
172
if "--distributed_lamb" in sys.argv:
    from torch.utils.cpp_extension import CUDAExtension
    sys.argv.remove("--distributed_lamb")

    from torch.utils.cpp_extension import BuildExtension
    cmdclass['build_ext'] = BuildExtension

173
    if torch.utils.cpp_extension.CUDA_HOME is None and not IS_ROCM_PYTORCH:
174
175
        raise RuntimeError("--distributed_lamb 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:
176
177
178
        print ("INFO: Building the distributed_lamb extension.")
        nvcc_args_distributed_lamb = ['-O3', '--use_fast_math'] + version_dependent_macros
        hipcc_args_distributed_lamb = ['-O3'] + version_dependent_macros
179
180
181
182
183
184
        ext_modules.append(
            CUDAExtension(name='distributed_lamb_cuda',
                          sources=['apex/contrib/csrc/optimizers/multi_tensor_distopt_lamb.cpp',
                                   'apex/contrib/csrc/optimizers/multi_tensor_distopt_lamb_kernel.cu'],
                          include_dirs=[os.path.join(this_dir, 'csrc')],
                          extra_compile_args={'cxx': ['-O3',] + version_dependent_macros,
185
                                              'nvcc': nvcc_args_distributed_lamb if not IS_ROCM_PYTORCH else hipcc_args_distributed_lamb}))
186

jjsjann123's avatar
jjsjann123 committed
187
if "--cuda_ext" in sys.argv:
188
    from torch.utils.cpp_extension import CUDAExtension
jjsjann123's avatar
jjsjann123 committed
189
    sys.argv.remove("--cuda_ext")
190

191
    if torch.utils.cpp_extension.CUDA_HOME is None and not IS_ROCM_PYTORCH:
Michael Carilli's avatar
Michael Carilli committed
192
        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.")
193
    else:
194
        if not IS_ROCM_PYTORCH:
195
196
            check_cuda_torch_binary_vs_bare_metal(torch.utils.cpp_extension.CUDA_HOME)

197
198
199
200
        print ("INFO: Building the multi-tensor apply extension.")
        nvcc_args_multi_tensor = ['-lineinfo', '-O3', '--use_fast_math'] + version_dependent_macros
        hipcc_args_multi_tensor = ['-O3'] + version_dependent_macros
        ext_modules.append(
201
202
203
204
205
206
207
208
209
210
211
212
            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_adagrad.cu',
                                   'csrc/multi_tensor_novograd.cu',
                                   'csrc/multi_tensor_lamb.cu'],
213
                          include_dirs=[os.path.join(this_dir, 'csrc')],
214
215
                          extra_compile_args={'cxx': ['-O3'] + version_dependent_macros,
                                              'nvcc': nvcc_args_multi_tensor if not IS_ROCM_PYTORCH else hipcc_args_multi_tensor}))
216

lcskrishna's avatar
lcskrishna committed
217
        print ("INFO: Building syncbn extension.")
218
        ext_modules.append(
219
220
221
            CUDAExtension(name='syncbn',
                          sources=['csrc/syncbn.cpp',
                                   'csrc/welford.cu'],
222
                          include_dirs=[os.path.join(this_dir, 'csrc')],
223
224
                          extra_compile_args={'cxx': ['-O3'] + version_dependent_macros,
                                              'nvcc':['-O3'] + version_dependent_macros}))
225

226
        nvcc_args_layer_norm = ['-maxrregcount=50', '-O3', '--use_fast_math'] + version_dependent_macros
227
228
229
        hipcc_args_layer_norm = ['-O3'] + version_dependent_macros
        print ("INFO: Building fused layernorm extension.")
        ext_modules.append(
230
231
232
            CUDAExtension(name='fused_layer_norm_cuda',
                          sources=['csrc/layer_norm_cuda.cpp',
                                   'csrc/layer_norm_cuda_kernel.cu'],
233
                          include_dirs=[os.path.join(this_dir, 'csrc')],
234
235
                          extra_compile_args={'cxx': ['-O3'] + version_dependent_macros,
                                              'nvcc': nvcc_args_layer_norm if not IS_ROCM_PYTORCH else hipcc_args_layer_norm}))
236

237
238
        print ("INFO: Building the MLP Extension.")
        ext_modules.append(
239
240
241
            CUDAExtension(name='mlp_cuda',
                          sources=['csrc/mlp.cpp',
                                   'csrc/mlp_cuda.cu'],
242
                          include_dirs=[os.path.join(this_dir, 'csrc')],
243
244
                          extra_compile_args={'cxx': ['-O3'] + version_dependent_macros,
                                              'nvcc':['-O3'] + version_dependent_macros}))
245

jjsjann123's avatar
jjsjann123 committed
246
247
248
249
250
251
252
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

253
    if torch.utils.cpp_extension.CUDA_HOME is None and not IS_ROCM_PYTORCH:
254
        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.")
jjsjann123's avatar
jjsjann123 committed
255
256
257
258
259
260
261
    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'],
262
263
                          include_dirs=[os.path.join(this_dir, 'csrc'),
                                        os.path.join(this_dir, 'apex/contrib/csrc/groupbn')],
mcarilli's avatar
mcarilli committed
264
                          extra_compile_args={'cxx': [] + version_dependent_macros,
jjsjann123's avatar
jjsjann123 committed
265
266
267
                                              'nvcc':['-DCUDA_HAS_FP16=1',
                                                      '-D__CUDA_NO_HALF_OPERATORS__',
                                                      '-D__CUDA_NO_HALF_CONVERSIONS__',
268
                                                      '-D__CUDA_NO_HALF2_OPERATORS__'] + version_dependent_macros}))
jjsjann123's avatar
jjsjann123 committed
269

270
271
272
273
274
275
276
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

277
    if torch.utils.cpp_extension.CUDA_HOME is None and not IS_ROCM_PYTORCH:
278
279
        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:
280
281
        print ("INFO: Building the xentropy extension.")
        ext_modules.append(
282
            CUDAExtension(name='xentropy_cuda',
283
284
285
                          sources=['./apex/contrib/csrc/xentropy/interface.cpp',
                                   './apex/contrib/csrc/xentropy/xentropy_kernel.cu'],
                          include_dirs=include_dirs + [this_dir + '/apex/contrib/csrc/xentropy/'],
286
287
                          extra_compile_args={'cxx': ['-O3'] + version_dependent_macros,
                                              'nvcc':['-O3'] + version_dependent_macros}))
288

289

290
291
292
293
294
295
296
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

297
    if torch.utils.cpp_extension.CUDA_HOME is None and not IS_ROCM_PYTORCH:
298
299
        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:
300
301
302
303
        print ("INFO: Building deprecated fused adam extension.")
        nvcc_args_fused_adam = ['-O3', '--use_fast_math'] + version_dependent_macros
        hipcc_args_fused_adam = ['-O3'] + version_dependent_macros
        ext_modules.append(
304
            CUDAExtension(name='fused_adam_cuda',
305
306
307
                          sources=['./apex/contrib/csrc/optimizers/fused_adam_cuda.cpp',
                                   './apex/contrib/csrc/optimizers/fused_adam_cuda_kernel.cu'],
                          include_dirs=include_dirs + [this_dir + '/apex/contrib/csrc/optimizers/'],
308
309
310
                          extra_compile_args={'cxx': ['-O3'] + version_dependent_macros,
                                              'nvcc' : nvcc_args_fused_adam if not IS_ROCM_PYTORCH else hipcc_args_fused_adam}))

311
312
313
314
315
316
317
if "--deprecated_fused_lamb" in sys.argv:
    from torch.utils.cpp_extension import CUDAExtension
    sys.argv.remove("--deprecated_fused_lamb")

    from torch.utils.cpp_extension import BuildExtension
    cmdclass['build_ext'] = BuildExtension

318
    if torch.utils.cpp_extension.CUDA_HOME is None and not IS_ROCM_PYTORCH:
319
320
        raise RuntimeError("--deprecated_fused_lamb 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:
321
322
323
324
        print ("INFO: Building deprecated fused lamb extension.")
        nvcc_args_fused_lamb = ['-O3', '--use_fast_math'] + version_dependent_macros
        hipcc_args_fused_lamb = ['-O3'] + version_dependent_macros
        ext_modules.append(
325
326
327
328
329
330
            CUDAExtension(name='fused_lamb_cuda',
                          sources=['apex/contrib/csrc/optimizers/fused_lamb_cuda.cpp',
                                   'apex/contrib/csrc/optimizers/fused_lamb_cuda_kernel.cu',
                                   'csrc/multi_tensor_l2norm_kernel.cu'],
                          include_dirs=[os.path.join(this_dir, 'csrc')],
                          extra_compile_args = nvcc_args_fused_lamb if not IS_ROCM_PYTORCH else hipcc_args_fused_lamb))
331

332
# Check, if ATen/CUDAGenerator.h is found, otherwise use the new ATen/CUDAGeneratorImpl.h, due to breaking change in https://github.com/pytorch/pytorch/pull/36026
ptrblck's avatar
ptrblck committed
333
334
335
336
337
generator_flag = []
torch_dir = torch.__path__[0]
if os.path.exists(os.path.join(torch_dir, 'include', 'ATen', 'CUDAGenerator.h')):
    generator_flag = ['-DOLD_GENERATOR']

yjk21's avatar
yjk21 committed
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
if "--fast_layer_norm" in sys.argv:
    from torch.utils.cpp_extension import CUDAExtension
    sys.argv.remove("--fast_layer_norm")

    from torch.utils.cpp_extension import BuildExtension
    cmdclass['build_ext'] = BuildExtension.with_options(use_ninja=False)

    if torch.utils.cpp_extension.CUDA_HOME is None:
        raise RuntimeError("--fast_layer_norm 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, if CUDA11 is installed for compute capability 8.0
        cc_flag = []
        _, bare_metal_major, _ = get_cuda_bare_metal_version(cpp_extension.CUDA_HOME)
        if int(bare_metal_major) >= 11:
            cc_flag.append('-gencode')
            cc_flag.append('arch=compute_80,code=sm_80')

        ext_modules.append(
            CUDAExtension(name='fast_layer_norm',
                          sources=['apex/contrib/csrc/layer_norm/ln_api.cpp',
                                   'apex/contrib/csrc/layer_norm/ln_fwd_cuda_kernel.cu',
                                   'apex/contrib/csrc/layer_norm/ln_bwd_semi_cuda_kernel.cu',
                                   ],
                          extra_compile_args={'cxx': ['-O3',] + version_dependent_macros + generator_flag,
                                              'nvcc':['-O3',
                                                      '-gencode', 'arch=compute_70,code=sm_70',
                                                      '-U__CUDA_NO_HALF_OPERATORS__',
                                                      '-U__CUDA_NO_HALF_CONVERSIONS__',
                                                      '-I./apex/contrib/csrc/layer_norm/',
                                                      '--expt-relaxed-constexpr',
                                                      '--expt-extended-lambda',
                                                      '--use_fast_math'] + version_dependent_macros + generator_flag + cc_flag}))
ptrblck's avatar
ptrblck committed
370

371
372
373
374
375
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
376
    cmdclass['build_ext'] = BuildExtension.with_options(use_ninja=False)
377

378
    if torch.utils.cpp_extension.CUDA_HOME is None and not IS_ROCM_PYTORCH: 
379
380
        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:
ptrblck's avatar
ptrblck committed
381
382
        # Check, if CUDA11 is installed for compute capability 8.0
        cc_flag = []
383
384
385
386
387
        if not IS_ROCM_PYTORCH:
            _, bare_metal_major, _ = get_cuda_bare_metal_version(cpp_extension.CUDA_HOME)
            if int(bare_metal_major) >= 11:
                cc_flag.append('-gencode')
                cc_flag.append('arch=compute_80,code=sm_80')
ptrblck's avatar
ptrblck committed
388

389
        subprocess.run(["git", "submodule", "update", "--init", "apex/contrib/csrc/multihead_attn/cutlass"])
390
391
392
        nvcc_args_mha = ['-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 + generator_flag + cc_flag
        hipcc_args_mha = ['-O3', '-I./apex/contrib/csrc/multihead_attn/cutlass/', '-U__CUDA_NO_HALF_OPERATORS__', '-U__CUDA_NO_HALF_CONVERSIONS__'] + version_dependent_macros + generator_flag

393
394
395
396
397
398
399
400
401
402
403
404
        ext_modules.append(
            CUDAExtension(name='fast_additive_mask_softmax_dropout',
                          sources=['apex/contrib/csrc/multihead_attn/additive_masked_softmax_dropout.cpp',
                                   'apex/contrib/csrc/multihead_attn/additive_masked_softmax_dropout_cuda.cu'],
                          extra_compile_args={'cxx': ['-O3',] + version_dependent_macros + generator_flag,
                                              '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',
ptrblck's avatar
ptrblck committed
405
                                                      '--use_fast_math'] + version_dependent_macros + generator_flag + cc_flag}))
406
407
408
409
410
411
412
413
414
415
416
417
        ext_modules.append(
            CUDAExtension(name='fast_mask_softmax_dropout',
                          sources=['apex/contrib/csrc/multihead_attn/masked_softmax_dropout.cpp',
                                   'apex/contrib/csrc/multihead_attn/masked_softmax_dropout_cuda.cu'],
                          extra_compile_args={'cxx': ['-O3',] + version_dependent_macros + generator_flag,
                                              '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',
ptrblck's avatar
ptrblck committed
418
                                                      '--use_fast_math'] + version_dependent_macros + generator_flag + cc_flag}))
419
420
421
422
423
424
425
426
427
428
429
430
        ext_modules.append(
            CUDAExtension(name='fast_self_multihead_attn_bias_additive_mask',
                          sources=['apex/contrib/csrc/multihead_attn/self_multihead_attn_bias_additive_mask.cpp',
                                   'apex/contrib/csrc/multihead_attn/self_multihead_attn_bias_additive_mask_cuda.cu'],
                          extra_compile_args={'cxx': ['-O3',] + version_dependent_macros + generator_flag,
                                              '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',
ptrblck's avatar
ptrblck committed
431
                                                      '--use_fast_math'] + version_dependent_macros + generator_flag + cc_flag}))
432
433
434
435
436
437
438
439
440
441
442
443
        ext_modules.append(
            CUDAExtension(name='fast_self_multihead_attn_bias',
                          sources=['apex/contrib/csrc/multihead_attn/self_multihead_attn_bias.cpp',
                                   'apex/contrib/csrc/multihead_attn/self_multihead_attn_bias_cuda.cu'],
                          extra_compile_args={'cxx': ['-O3',] + version_dependent_macros + generator_flag,
                                              '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',
ptrblck's avatar
ptrblck committed
444
                                                      '--use_fast_math'] + version_dependent_macros + generator_flag + cc_flag}))
445
446
        ext_modules.append(
            CUDAExtension(name='fast_self_multihead_attn',
447
                          sources=['apex/contrib/csrc/multihead_attn/self_multihead_attn.cpp',
448
                                   'apex/contrib/csrc/multihead_attn/self_multihead_attn_cuda.cu'],
ptrblck's avatar
ptrblck committed
449
                          extra_compile_args={'cxx': ['-O3',] + version_dependent_macros + generator_flag,
450
                                              'nvcc':['-O3',
451
                                                      '-gencode', 'arch=compute_70,code=sm_70',
452
453
454
455
456
                                                      '-I./apex/contrib/csrc/multihead_attn/cutlass/',
                                                      '-U__CUDA_NO_HALF_OPERATORS__',
                                                      '-U__CUDA_NO_HALF_CONVERSIONS__',
                                                      '--expt-relaxed-constexpr',
                                                      '--expt-extended-lambda',
ptrblck's avatar
ptrblck committed
457
                                                      '--use_fast_math'] + version_dependent_macros + generator_flag + cc_flag}))
458
459
        ext_modules.append(
            CUDAExtension(name='fast_self_multihead_attn_norm_add',
460
461
462
                          sources=['./apex/contrib/csrc/multihead_attn/self_multihead_attn_norm_add.cpp',
                                   './apex/contrib/csrc/multihead_attn/self_multihead_attn_norm_add_cuda.cu'],
                          include_dirs=include_dirs + [this_dir + '/apex/contrib/csrc/multihead_attn/'],
ptrblck's avatar
ptrblck committed
463
                          extra_compile_args={'cxx': ['-O3',] + version_dependent_macros + generator_flag,
464
                                              'nvcc':nvcc_args_mha if not IS_ROCM_PYTORCH else hipcc_args_mha}))
465
466
        ext_modules.append(
            CUDAExtension(name='fast_encdec_multihead_attn',
467
                          sources=['apex/contrib/csrc/multihead_attn/encdec_multihead_attn.cpp',
468
                                   'apex/contrib/csrc/multihead_attn/encdec_multihead_attn_cuda.cu'],
ptrblck's avatar
ptrblck committed
469
                          extra_compile_args={'cxx': ['-O3',] + version_dependent_macros + generator_flag,
470
                                              'nvcc':['-O3',
471
                                                      '-gencode', 'arch=compute_70,code=sm_70',
472
473
474
475
476
                                                      '-I./apex/contrib/csrc/multihead_attn/cutlass/',
                                                      '-U__CUDA_NO_HALF_OPERATORS__',
                                                      '-U__CUDA_NO_HALF_CONVERSIONS__',
                                                      '--expt-relaxed-constexpr',
                                                      '--expt-extended-lambda',
ptrblck's avatar
ptrblck committed
477
                                                      '--use_fast_math'] + version_dependent_macros + generator_flag + cc_flag}))
478
479
        ext_modules.append(
            CUDAExtension(name='fast_encdec_multihead_attn_norm_add',
480
481
482
                          sources=['./apex/contrib/csrc/multihead_attn/encdec_multihead_attn_norm_add.cpp',
                                   './apex/contrib/csrc/multihead_attn/encdec_multihead_attn_norm_add_cuda.cu'],
                          include_dirs=include_dirs + [this_dir + '/apex/contrib/csrc/multihead_attn/'],
ptrblck's avatar
ptrblck committed
483
                          extra_compile_args={'cxx': ['-O3',] + version_dependent_macros + generator_flag,
484
                                              'nvcc':nvcc_args_mha if not IS_ROCM_PYTORCH else hipcc_args_mha}))
485

Christian Sarofeen's avatar
Christian Sarofeen committed
486
setup(
487
488
    name='apex',
    version='0.1',
489
490
491
492
    packages=find_packages(exclude=('build',
                                    'csrc',
                                    'include',
                                    'tests',
493
494
495
496
497
                                    'dist',
                                    'docs',
                                    'tests',
                                    'examples',
                                    'apex.egg-info',)),
Christian Sarofeen's avatar
Christian Sarofeen committed
498
    description='PyTorch Extensions written by NVIDIA',
jjsjann123's avatar
jjsjann123 committed
499
500
    ext_modules=ext_modules,
    cmdclass=cmdclass,
ptrblck's avatar
ptrblck committed
501
    extras_require=extras,
Christian Sarofeen's avatar
Christian Sarofeen committed
502
)