setup.py 34.1 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

23
if not torch.cuda.is_available():
mcarilli's avatar
mcarilli committed
24
25
26
27
28
29
    # 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
30
31
          '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
32
33
34
          '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
35
36
37
38
39
        _, 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"
40

41
print("\n\ntorch.__version__  = {}\n\n".format(torch.__version__))
42
43
44
45
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
46
      raise RuntimeError("Apex requires Pytorch 0.4 or newer.\n" +
47
48
                         "The latest stable release can be obtained from https://pytorch.org/")

jjsjann123's avatar
jjsjann123 committed
49
50
51
cmdclass = {}
ext_modules = []

ptrblck's avatar
ptrblck committed
52
extras = {}
Marek Kolodziej's avatar
Marek Kolodziej committed
53
if "--pyprof" in sys.argv:
54
55
56
57
58
    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
59
60
    with open('requirements.txt') as f:
        required_packages = f.read().splitlines()
ptrblck's avatar
ptrblck committed
61
        extras['pyprof'] = required_packages
Marek Kolodziej's avatar
Marek Kolodziej committed
62
63
64
65
66
67
68
    try:
        sys.argv.remove("--pyprof")
    except:
        pass
else:
    warnings.warn("Option --pyprof not specified. Not installing PyProf dependencies!")

69
if "--cpp_ext" in sys.argv or "--cuda_ext" in sys.argv:
Michael Carilli's avatar
Michael Carilli committed
70
71
    if TORCH_MAJOR == 0:
        raise RuntimeError("--cpp_ext requires Pytorch 1.0 or later, "
72
                           "found torch.__version__ = {}".format(torch.__version__))
73
74
75
76
77
78
79
80
81
82
    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
83
def get_cuda_bare_metal_version(cuda_dir):
mcarilli's avatar
mcarilli committed
84
85
86
87
88
89
    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
90
91
92
93
94

    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
95
96
97
98
99
100
101
    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
102
103
104
105
106
107
        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
108

ptrblck's avatar
ptrblck committed
109

mcarilli's avatar
mcarilli committed
110
111
112
113
114
115
116
117
118
119
120
# 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']
121
122
123
124
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
125

126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
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

    if torch.utils.cpp_extension.CUDA_HOME is None:
        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:
        ext_modules.append(
            CUDAExtension(name='distributed_adam_cuda',
                          sources=['apex/contrib/csrc/optimizers/multi_tensor_distopt_adam.cpp',
                                   'apex/contrib/csrc/optimizers/multi_tensor_distopt_adam_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}))

145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
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
164
if "--cuda_ext" in sys.argv:
165
    from torch.utils.cpp_extension import CUDAExtension
jjsjann123's avatar
jjsjann123 committed
166
    sys.argv.remove("--cuda_ext")
167
168

    if torch.utils.cpp_extension.CUDA_HOME is None:
Michael Carilli's avatar
Michael Carilli committed
169
        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.")
170
    else:
mcarilli's avatar
mcarilli committed
171
172
        check_cuda_torch_binary_vs_bare_metal(torch.utils.cpp_extension.CUDA_HOME)

173
174
        ext_modules.append(
            CUDAExtension(name='amp_C',
175
                          sources=['csrc/amp_C_frontend.cpp',
Michael Carilli's avatar
Michael Carilli committed
176
                                   'csrc/multi_tensor_sgd_kernel.cu',
177
                                   'csrc/multi_tensor_scale_kernel.cu',
178
                                   'csrc/multi_tensor_axpby_kernel.cu',
179
                                   'csrc/multi_tensor_l2norm_kernel.cu',
180
181
                                   'csrc/multi_tensor_lamb_stage_1.cu',
                                   'csrc/multi_tensor_lamb_stage_2.cu',
182
                                   'csrc/multi_tensor_adam.cu',
Andrew Tulloch's avatar
Andrew Tulloch committed
183
                                   'csrc/multi_tensor_adagrad.cu',
184
185
                                   'csrc/multi_tensor_novograd.cu',
                                   'csrc/multi_tensor_lamb.cu'],
mcarilli's avatar
mcarilli committed
186
                          extra_compile_args={'cxx': ['-O3'] + version_dependent_macros,
Michael Carilli's avatar
Michael Carilli committed
187
                                              'nvcc':['-lineinfo',
188
                                                      '-O3',
189
                                                      # '--resource-usage',
mcarilli's avatar
mcarilli committed
190
                                                      '--use_fast_math'] + version_dependent_macros}))
191
192
193
        ext_modules.append(
            CUDAExtension(name='syncbn',
                          sources=['csrc/syncbn.cpp',
mcarilli's avatar
mcarilli committed
194
195
196
197
                                   'csrc/welford.cu'],
                          extra_compile_args={'cxx': ['-O3'] + version_dependent_macros,
                                              'nvcc':['-O3'] + version_dependent_macros}))

198
199
        ext_modules.append(
            CUDAExtension(name='fused_layer_norm_cuda',
200
201
                          sources=['csrc/layer_norm_cuda.cpp',
                                   'csrc/layer_norm_cuda_kernel.cu'],
mcarilli's avatar
mcarilli committed
202
                          extra_compile_args={'cxx': ['-O3'] + version_dependent_macros,
203
                                              'nvcc':['-maxrregcount=50',
mcarilli's avatar
mcarilli committed
204
                                                      '-O3',
mcarilli's avatar
mcarilli committed
205
                                                      '--use_fast_math'] + version_dependent_macros}))
jjsjann123's avatar
jjsjann123 committed
206

207
208
209
210
211
212
213
        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}))

jjsjann123's avatar
jjsjann123 committed
214
215
216
217
218
219
220
221
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:
222
        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
223
224
225
226
227
228
229
    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'],
230
                          include_dirs=[os.path.join(this_dir, 'csrc')],
mcarilli's avatar
mcarilli committed
231
                          extra_compile_args={'cxx': [] + version_dependent_macros,
jjsjann123's avatar
jjsjann123 committed
232
233
234
                                              'nvcc':['-DCUDA_HAS_FP16=1',
                                                      '-D__CUDA_NO_HALF_OPERATORS__',
                                                      '-D__CUDA_NO_HALF_CONVERSIONS__',
235
                                                      '-D__CUDA_NO_HALF2_OPERATORS__'] + version_dependent_macros}))
jjsjann123's avatar
jjsjann123 committed
236

237
238
239
240
241
242
243
244
245
246
247
248
249
250
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'],
251
                          include_dirs=[os.path.join(this_dir, 'csrc')],
mcarilli's avatar
mcarilli committed
252
253
                          extra_compile_args={'cxx': ['-O3'] + version_dependent_macros,
                                              'nvcc':['-O3'] + version_dependent_macros}))
254

255
256
257
258
259
260
261
262
263
264
265
266
267
268
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'],
269
                          include_dirs=[os.path.join(this_dir, 'csrc')],
270
271
272
273
                          extra_compile_args={'cxx': ['-O3',] + version_dependent_macros,
                                              'nvcc':['-O3',
                                                      '--use_fast_math'] + version_dependent_macros}))

274
275
276
277
278
279
280
281
282
283
284
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

    if torch.utils.cpp_extension.CUDA_HOME is None:
        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:
        ext_modules.append(
Kexin Yu's avatar
Kexin Yu committed
285
            CUDAExtension(name='fused_lamb_cuda',
286
                          sources=['apex/contrib/csrc/optimizers/fused_lamb_cuda.cpp',
Kexin Yu's avatar
Kexin Yu committed
287
288
                                   'apex/contrib/csrc/optimizers/fused_lamb_cuda_kernel.cu',
                                   'csrc/multi_tensor_l2norm_kernel.cu'],
289
290
291
292
293
                          include_dirs=[os.path.join(this_dir, 'csrc')],
                          extra_compile_args={'cxx': ['-O3',] + version_dependent_macros,
                                              'nvcc':['-O3',
                                                      '--use_fast_math'] + version_dependent_macros}))

294
# 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
295
296
297
298
299
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
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
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}))
yjk21's avatar
yjk21 committed
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
if "--fmha" in sys.argv:
    from torch.utils.cpp_extension import CUDAExtension
    sys.argv.remove("--fmha")

    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("--fmha 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:
            raise RuntimeError("--fmha only supported on SM80")

        ext_modules.append(
            CUDAExtension(name='fmhalib',
                          sources=[
                                   'apex/contrib/csrc/fmha/fmha_api.cpp',
                                   'apex/contrib/csrc/fmha/src/fmha_fprop_fp16_128_64_kernel.sm80.cu',
                                   'apex/contrib/csrc/fmha/src/fmha_fprop_fp16_256_64_kernel.sm80.cu',
                                   'apex/contrib/csrc/fmha/src/fmha_fprop_fp16_384_64_kernel.sm80.cu',
                                   'apex/contrib/csrc/fmha/src/fmha_fprop_fp16_512_64_kernel.sm80.cu',
                                   'apex/contrib/csrc/fmha/src/fmha_dgrad_fp16_128_64_kernel.sm80.cu',
                                   'apex/contrib/csrc/fmha/src/fmha_dgrad_fp16_256_64_kernel.sm80.cu',
                                   'apex/contrib/csrc/fmha/src/fmha_dgrad_fp16_384_64_kernel.sm80.cu',
                                   'apex/contrib/csrc/fmha/src/fmha_dgrad_fp16_512_64_kernel.sm80.cu',
                                   ],
                          extra_compile_args={'cxx': ['-O3',
                                                      '-I./apex/contrib/csrc/fmha/src',
                                                      ] + version_dependent_macros + generator_flag,
                                              'nvcc':['-O3',
                                                      '-gencode', 'arch=compute_80,code=sm_80',
                                                      '-U__CUDA_NO_HALF_OPERATORS__',
                                                      '-U__CUDA_NO_HALF_CONVERSIONS__',
                                                      '-I./apex/contrib/csrc/',
                                                      '-I./apex/contrib/csrc/fmha/src',
                                                      '--expt-relaxed-constexpr',
                                                      '--expt-extended-lambda',
                                                      '--use_fast_math'] + version_dependent_macros + generator_flag + cc_flag}))

ptrblck's avatar
ptrblck committed
374

375
376
377
378
379
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
380
    cmdclass['build_ext'] = BuildExtension.with_options(use_ninja=False)
381
382
383
384

    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:
ptrblck's avatar
ptrblck committed
385
386
387
388
389
390
391
        # 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')

392
        subprocess.run(["git", "submodule", "update", "--init", "apex/contrib/csrc/multihead_attn/cutlass"])
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
                          sources=['apex/contrib/csrc/multihead_attn/self_multihead_attn_norm_add.cpp',
461
                                   'apex/contrib/csrc/multihead_attn/self_multihead_attn_norm_add_cuda.cu'],
ptrblck's avatar
ptrblck committed
462
                          extra_compile_args={'cxx': ['-O3',] + version_dependent_macros + generator_flag,
463
                                              'nvcc':['-O3',
464
                                                      '-gencode', 'arch=compute_70,code=sm_70',
465
466
467
468
469
                                                      '-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
470
                                                      '--use_fast_math'] + version_dependent_macros + generator_flag + cc_flag}))
471
472
        ext_modules.append(
            CUDAExtension(name='fast_encdec_multihead_attn',
473
                          sources=['apex/contrib/csrc/multihead_attn/encdec_multihead_attn.cpp',
474
                                   'apex/contrib/csrc/multihead_attn/encdec_multihead_attn_cuda.cu'],
ptrblck's avatar
ptrblck committed
475
                          extra_compile_args={'cxx': ['-O3',] + version_dependent_macros + generator_flag,
476
                                              'nvcc':['-O3',
477
                                                      '-gencode', 'arch=compute_70,code=sm_70',
478
479
480
481
482
                                                      '-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
483
                                                      '--use_fast_math'] + version_dependent_macros + generator_flag + cc_flag}))
484
485
        ext_modules.append(
            CUDAExtension(name='fast_encdec_multihead_attn_norm_add',
486
                          sources=['apex/contrib/csrc/multihead_attn/encdec_multihead_attn_norm_add.cpp',
487
                                   'apex/contrib/csrc/multihead_attn/encdec_multihead_attn_norm_add_cuda.cu'],
ptrblck's avatar
ptrblck committed
488
                          extra_compile_args={'cxx': ['-O3',] + version_dependent_macros + generator_flag,
489
                                              'nvcc':['-O3',
490
                                                      '-gencode', 'arch=compute_70,code=sm_70',
491
492
493
494
495
                                                      '-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
496
                                                      '--use_fast_math'] + version_dependent_macros + generator_flag + cc_flag}))
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
if "--transducer" in sys.argv:
    from torch.utils.cpp_extension import CUDAExtension
    sys.argv.remove("--transducer")

    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("--transducer 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='transducer_joint_cuda',
                          sources=['apex/contrib/csrc/transducer/transducer_joint.cpp',
                                   'apex/contrib/csrc/transducer/transducer_joint_kernel.cu'],
                          include_dirs=[os.path.join(this_dir, 'csrc')],
                          extra_compile_args={'cxx': ['-O3'] + version_dependent_macros,
                                              'nvcc':['-O3'] + version_dependent_macros}))
        ext_modules.append(
            CUDAExtension(name='transducer_loss_cuda',
                          sources=['apex/contrib/csrc/transducer/transducer_loss.cpp',
                                   'apex/contrib/csrc/transducer/transducer_loss_kernel.cu'],
                          include_dirs=[os.path.join(this_dir, 'csrc')],
                          extra_compile_args={'cxx': ['-O3'] + version_dependent_macros,
                                              'nvcc':['-O3'] + version_dependent_macros}))

523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
if "--fast_bottleneck" in sys.argv:
    from torch.utils.cpp_extension import CUDAExtension
    sys.argv.remove("--fast_bottleneck")

    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_bottleneck 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/cudnn-frontend/"])
        ext_modules.append(
            CUDAExtension(name='fast_bottleneck',
                          sources=['apex/contrib/csrc/bottleneck/bottleneck.cpp'],
                          include_dirs=['apex/contrib/csrc/cudnn-frontend/include'],
                          extra_compile_args={'cxx': ['-O3',] + version_dependent_macros + generator_flag}))

Christian Sarofeen's avatar
Christian Sarofeen committed
540
setup(
541
542
    name='apex',
    version='0.1',
543
544
545
546
    packages=find_packages(exclude=('build',
                                    'csrc',
                                    'include',
                                    'tests',
547
548
549
550
551
                                    'dist',
                                    'docs',
                                    'tests',
                                    'examples',
                                    'apex.egg-info',)),
Christian Sarofeen's avatar
Christian Sarofeen committed
552
    description='PyTorch Extensions written by NVIDIA',
jjsjann123's avatar
jjsjann123 committed
553
554
    ext_modules=ext_modules,
    cmdclass=cmdclass,
ptrblck's avatar
ptrblck committed
555
    extras_require=extras,
Christian Sarofeen's avatar
Christian Sarofeen committed
556
)