setup.py 29.1 KB
Newer Older
1
import torch
2
from setuptools import setup, find_packages
mcarilli's avatar
mcarilli committed
3
import subprocess
4

jjsjann123's avatar
jjsjann123 committed
5
import sys
Marek Kolodziej's avatar
Marek Kolodziej committed
6
import warnings
mcarilli's avatar
mcarilli committed
7
import os
jjsjann123's avatar
jjsjann123 committed
8

9
10
from torch.utils.hipify import hipify_python

11
12
13
# ninja build does not work unless include_dirs are abs path
this_dir = os.path.dirname(os.path.abspath(__file__))

14
if not torch.cuda.is_available():
mcarilli's avatar
mcarilli committed
15
16
17
18
19
20
21
22
23
24
25
    # 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"
26

27
print("\n\ntorch.__version__  = {}\n\n".format(torch.__version__))
28
29
30
31
TORCH_MAJOR = int(torch.__version__.split('.')[0])
TORCH_MINOR = int(torch.__version__.split('.')[1])

if TORCH_MAJOR == 0 and TORCH_MINOR < 4:
Michael Carilli's avatar
Michael Carilli committed
32
      raise RuntimeError("Apex requires Pytorch 0.4 or newer.\n" +
33
34
                         "The latest stable release can be obtained from https://pytorch.org/")

jjsjann123's avatar
jjsjann123 committed
35
36
37
cmdclass = {}
ext_modules = []

ptrblck's avatar
ptrblck committed
38
extras = {}
Marek Kolodziej's avatar
Marek Kolodziej committed
39
if "--pyprof" in sys.argv:
40
41
42
43
44
    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
45
46
    with open('requirements.txt') as f:
        required_packages = f.read().splitlines()
ptrblck's avatar
ptrblck committed
47
        extras['pyprof'] = required_packages
Marek Kolodziej's avatar
Marek Kolodziej committed
48
49
50
51
52
53
54
    try:
        sys.argv.remove("--pyprof")
    except:
        pass
else:
    warnings.warn("Option --pyprof not specified. Not installing PyProf dependencies!")

55
if "--cpp_ext" in sys.argv or "--cuda_ext" in sys.argv:
Michael Carilli's avatar
Michael Carilli committed
56
57
    if TORCH_MAJOR == 0:
        raise RuntimeError("--cpp_ext requires Pytorch 1.0 or later, "
58
                           "found torch.__version__ = {}".format(torch.__version__))
59
60
61
62
63
64
65
66
67
68
    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',]))

mcarilli's avatar
mcarilli committed
69
70
71
72
73
74
75
76
77
78
79
80
81
82
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):
Michael Carilli's avatar
Michael Carilli committed
83
84
85
86
87
88
        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
89

90
91
92
93
94
95
96
97
def check_if_rocm_pytorch():
    is_rocm_pytorch = False
    if torch.__version__ >= '1.5':
        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

mcarilli's avatar
mcarilli committed
98
99
100
101
102
103
104
105
106
107
108
# 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']
109
110
111
112
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
113

114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
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

    if torch.utils.cpp_extension.CUDA_HOME is None:
        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:
        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,
                                              'nvcc':['-O3',
                                                      '--use_fast_math'] + version_dependent_macros}))

jjsjann123's avatar
jjsjann123 committed
133
if "--cuda_ext" in sys.argv:
134
    from torch.utils.cpp_extension import CUDAExtension
jjsjann123's avatar
jjsjann123 committed
135
    sys.argv.remove("--cuda_ext")
136

137
138
139
140
    is_rocm_pytorch = False
    if torch.__version__ >= '1.5':
        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
141

142
    if torch.utils.cpp_extension.CUDA_HOME is None and (not is_rocm_pytorch):
Michael Carilli's avatar
Michael Carilli committed
143
        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.")
144
    else:
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
        if not is_rocm_pytorch:
            check_cuda_torch_binary_vs_bare_metal(torch.utils.cpp_extension.CUDA_HOME)

        if is_rocm_pytorch:
            import shutil
            with hipify_python.GeneratedFileCleaner(keep_intermediates=True) as clean_ctx:
                hipify_python.hipify(project_directory=this_dir, output_directory=this_dir, includes="csrc/*",
                                        show_detailed=True, is_pytorch_extension=True, clean_ctx=clean_ctx)
            shutil.copy("csrc/compat.h", "csrc/hip/compat.h")
            shutil.copy("csrc/type_shim.h", "csrc/hip/type_shim.h")

        if not is_rocm_pytorch:
            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',
167
                                       'csrc/multi_tensor_adagrad.cu',
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
                                       '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}))
        else:
            print ("INFO: Building Multitensor apply extension")
            ext_modules.append(
                CUDAExtension(name='amp_C',
                              sources=['csrc/amp_C_frontend.cpp',
                                       'csrc/hip/multi_tensor_sgd_kernel.hip',
                                       'csrc/hip/multi_tensor_scale_kernel.hip',
                                       'csrc/hip/multi_tensor_axpby_kernel.hip',
                                       'csrc/hip/multi_tensor_l2norm_kernel.hip',
                                       'csrc/hip/multi_tensor_lamb_stage_1.hip',
                                       'csrc/hip/multi_tensor_lamb_stage_2.hip',
                                       'csrc/hip/multi_tensor_adam.hip',
187
                                       'csrc/hip/multi_tensor_adagrad.hip',
188
189
                                       'csrc/hip/multi_tensor_novograd.hip',
                                       'csrc/hip/multi_tensor_lamb.hip'],
190
                              extra_compile_args=['-O3'] + version_dependent_macros))
191
192
193
194
195
196
197
198
199

        if not is_rocm_pytorch:
            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}))
        else:
200
201
202
203
204
205
            print ("INFO: Building syncbn extension.")
            ext_modules.append(
                CUDAExtension(name='syncbn',
                              sources=['csrc/syncbn.cpp',
                                       'csrc/hip/welford.hip'],
                              extra_compile_args=['-O3'] + version_dependent_macros))
206

207

208
209
210
211
212
213
214
215
216
217
        if not is_rocm_pytorch:
            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}))
        else:
218
219
220
221
222
223
224
            print ("INFO: Building FusedLayerNorm extension.")
            ext_modules.append(
                CUDAExtension(name='fused_layer_norm_cuda',
                              sources=['csrc/layer_norm_cuda.cpp',
                                       'csrc/hip/layer_norm_hip_kernel.hip'],
                              extra_compile_args={'cxx' : ['-O3'] + version_dependent_macros,
                                                  'nvcc' : []}))
225
226
227
228
229
230
231
232
233

        if not is_rocm_pytorch:
            ext_modules.append(
                CUDAExtension(name='mlp_cuda',
                              sources=['csrc/mlp.cpp',
                                       'csrc/mlp_cuda.cu'],
                              extra_compile_args={'cxx': ['-O3'] + version_dependent_macros,
                                                  'nvcc':['-O3'] + version_dependent_macros}))
        else:
234
235
236
237
238
239
240
            print ("INFO: Building MLP extension")
            ext_modules.append(
                CUDAExtension(name='mlp_cuda',
                              sources=['csrc/mlp.cpp',
                                       'csrc/hip/mlp_hip.hip'],
                              extra_compile_args={'cxx' : ['-O3'] + version_dependent_macros,
                                                  'nvcc' : []}))
241

jjsjann123's avatar
jjsjann123 committed
242
243
244
245
246
247
248
249
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:
250
        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
251
252
253
254
255
256
257
    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'],
258
                          include_dirs=[os.path.join(this_dir, 'csrc')],
mcarilli's avatar
mcarilli committed
259
                          extra_compile_args={'cxx': [] + version_dependent_macros,
jjsjann123's avatar
jjsjann123 committed
260
261
262
                                              'nvcc':['-DCUDA_HAS_FP16=1',
                                                      '-D__CUDA_NO_HALF_OPERATORS__',
                                                      '-D__CUDA_NO_HALF_CONVERSIONS__',
263
                                                      '-D__CUDA_NO_HALF2_OPERATORS__'] + version_dependent_macros}))
jjsjann123's avatar
jjsjann123 committed
264

265
266
267
268
269
270
271
272
273
274
275
276
277
278
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'],
279
                          include_dirs=[os.path.join(this_dir, 'csrc')],
mcarilli's avatar
mcarilli committed
280
281
                          extra_compile_args={'cxx': ['-O3'] + version_dependent_macros,
                                              'nvcc':['-O3'] + version_dependent_macros}))
282

283
284
285
286
287
288
289
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

290
291
292
    is_rocm_pytorch = check_if_rocm_pytorch()

    if torch.utils.cpp_extension.CUDA_HOME is None and (not is_rocm_pytorch):
293
294
        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:
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
        if not is_rocm_pytorch:
            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=[os.path.join(this_dir, 'csrc')],
                              extra_compile_args={'cxx': ['-O3',] + version_dependent_macros,
                                                  'nvcc':['-O3',
                                                          '--use_fast_math'] + version_dependent_macros}))
        else:
            print ("INFO: Building deprecated fused adam.")
            ext_modules.append(
                CUDAExtension(name='fused_adam_cuda',
                              sources=['apex/contrib/csrc/optimizers/fused_adam_cuda.cpp',
                                       'apex/contrib/csrc/optimizers/hip/fused_adam_hip_kernel.hip'],
                              include_dirs=[os.path.join(this_dir, 'csrc/hip')],
                              extra_compile_args=['-O3'] + version_dependent_macros))
312

313
314
315
316
317
318
319
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

320
321
322
    is_rocm_pytorch = check_if_rocm_pytorch()

    if torch.utils.cpp_extension.CUDA_HOME is None and (not is_rocm_pytorch):
323
324
        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:
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
        if not is_rocm_pytorch:
            ext_modules.append(
                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={'cxx': ['-O3',] + version_dependent_macros,
                                                  'nvcc':['-O3',
                                                          '--use_fast_math'] + version_dependent_macros}))
        else:
            print ("INFO: Building deprecated fused lamb.")
            ext_modules.append(
                CUDAExtension(name='fused_lamb_cuda',
                              sources=['apex/contrib/csrc/optimizers/fused_lamb_cuda.cpp',
                                       'apex/contrib/csrc/optimizers/hip/fused_lamb_hip_kernel.hip',
                                       'csrc/hip/multi_tensor_l2norm_kernel.hip'],
                              include_dirs=[os.path.join(this_dir, 'csrc/hip')],
                              extra_compile_args=['-O3'] + version_dependent_macros))
344

345
# 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
346
347
348
349
350
generator_flag = []
torch_dir = torch.__path__[0]
if os.path.exists(os.path.join(torch_dir, 'include', 'ATen', 'CUDAGenerator.h')):
    generator_flag = ['-DOLD_GENERATOR']

351
352
353
354
355
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
356
    cmdclass['build_ext'] = BuildExtension.with_options(use_ninja=False)
357
358
359
360
361

    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:
        subprocess.run(["git", "submodule", "update", "--init", "apex/contrib/csrc/multihead_attn/cutlass"])
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
        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',
                                                      '--use_fast_math'] + version_dependent_macros + generator_flag}))
        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',
                                                      '--use_fast_math'] + version_dependent_macros + generator_flag}))
        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',
                                                      '--use_fast_math'] + version_dependent_macros + generator_flag}))
        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',
                                                      '--use_fast_math'] + version_dependent_macros + generator_flag}))
414
415
        ext_modules.append(
            CUDAExtension(name='fast_self_multihead_attn',
416
                          sources=['apex/contrib/csrc/multihead_attn/self_multihead_attn.cpp',
417
                                   'apex/contrib/csrc/multihead_attn/self_multihead_attn_cuda.cu'],
ptrblck's avatar
ptrblck committed
418
                          extra_compile_args={'cxx': ['-O3',] + version_dependent_macros + generator_flag,
419
                                              'nvcc':['-O3',
420
                                                      '-gencode', 'arch=compute_70,code=sm_70',
421
422
423
424
425
                                                      '-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
426
                                                      '--use_fast_math'] + version_dependent_macros + generator_flag}))
427
428
        ext_modules.append(
            CUDAExtension(name='fast_self_multihead_attn_norm_add',
429
                          sources=['apex/contrib/csrc/multihead_attn/self_multihead_attn_norm_add.cpp',
430
                                   'apex/contrib/csrc/multihead_attn/self_multihead_attn_norm_add_cuda.cu'],
ptrblck's avatar
ptrblck committed
431
                          extra_compile_args={'cxx': ['-O3',] + version_dependent_macros + generator_flag,
432
                                              'nvcc':['-O3',
433
                                                      '-gencode', 'arch=compute_70,code=sm_70',
434
435
436
437
438
                                                      '-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
439
                                                      '--use_fast_math'] + version_dependent_macros + generator_flag}))
440
441
        ext_modules.append(
            CUDAExtension(name='fast_encdec_multihead_attn',
442
                          sources=['apex/contrib/csrc/multihead_attn/encdec_multihead_attn.cpp',
443
                                   'apex/contrib/csrc/multihead_attn/encdec_multihead_attn_cuda.cu'],
ptrblck's avatar
ptrblck committed
444
                          extra_compile_args={'cxx': ['-O3',] + version_dependent_macros + generator_flag,
445
                                              'nvcc':['-O3',
446
                                                      '-gencode', 'arch=compute_70,code=sm_70',
447
448
449
450
451
                                                      '-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
452
                                                      '--use_fast_math'] + version_dependent_macros + generator_flag}))
453
454
        ext_modules.append(
            CUDAExtension(name='fast_encdec_multihead_attn_norm_add',
455
                          sources=['apex/contrib/csrc/multihead_attn/encdec_multihead_attn_norm_add.cpp',
456
                                   'apex/contrib/csrc/multihead_attn/encdec_multihead_attn_norm_add_cuda.cu'],
ptrblck's avatar
ptrblck committed
457
                          extra_compile_args={'cxx': ['-O3',] + version_dependent_macros + generator_flag,
458
                                              'nvcc':['-O3',
459
                                                      '-gencode', 'arch=compute_70,code=sm_70',
460
461
462
463
464
                                                      '-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
465
                                                      '--use_fast_math'] + version_dependent_macros + generator_flag}))
466

Christian Sarofeen's avatar
Christian Sarofeen committed
467
setup(
468
469
    name='apex',
    version='0.1',
470
471
472
473
    packages=find_packages(exclude=('build',
                                    'csrc',
                                    'include',
                                    'tests',
474
475
476
477
478
                                    'dist',
                                    'docs',
                                    'tests',
                                    'examples',
                                    'apex.egg-info',)),
Christian Sarofeen's avatar
Christian Sarofeen committed
479
    description='PyTorch Extensions written by NVIDIA',
jjsjann123's avatar
jjsjann123 committed
480
481
    ext_modules=ext_modules,
    cmdclass=cmdclass,
ptrblck's avatar
ptrblck committed
482
    extras_require=extras,
Christian Sarofeen's avatar
Christian Sarofeen committed
483
)