1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
|
# Copyright 2013-2019 Lawrence Livermore National Security, LLC and other
# Spack Project Developers. See the top-level COPYRIGHT file for details.
#
# SPDX-License-Identifier: (Apache-2.0 OR MIT)
class PyHorovod(PythonPackage):
"""Horovod is a distributed deep learning training framework for
TensorFlow, Keras, PyTorch, and Apache MXNet."""
homepage = "https://github.com/horovod"
git = "https://github.com/horovod/horovod.git"
maintainers = ['adamjstewart']
version('master', branch='master', submodules=True)
version('0.19.1', tag='v0.19.1', submodules=True)
version('0.19.0', tag='v0.19.0', submodules=True)
version('0.18.2', tag='v0.18.2', submodules=True)
version('0.18.1', tag='v0.18.1', submodules=True)
version('0.18.0', tag='v0.18.0', submodules=True)
version('0.17.1', tag='v0.17.1', submodules=True)
version('0.17.0', tag='v0.17.0', submodules=True)
version('0.16.4', tag='v0.16.4', submodules=True)
version('0.16.3', tag='v0.16.3', submodules=True)
version('0.16.2', tag='v0.16.2', submodules=True)
# Deep learning frameworks
variant('pytorch', default=True, description='Enables PyTorch')
variant('tensorflow', default=False, description='Enables TensorFlow')
variant('mxnet', default=False, description='Enables Apache MXNet')
# Distributed support
variant('gloo', default=False, description='Enables features related to distributed support')
variant('mpi', default=True, description='Enables MPI build')
# GPU support
variant('cuda', default=True, description='Enables CUDA build')
variant('gpu_allreduce', default='mpi',
description='Backend to use for GPU_ALLREDUCE',
values=('mpi', 'nccl'), multi=False) # DDL support is deprecated
variant('gpu_allgather', default='mpi',
description='Backend to use for GPU_ALLGATHER',
values=('mpi',), multi=False)
variant('gpu_broadcast', default='mpi',
description='Backend to use for GPU_BROADCAST',
values=('mpi', 'nccl'), multi=False)
# Required dependencies
depends_on('py-setuptools', type='build')
depends_on('py-cloudpickle', type=('build', 'run'))
depends_on('py-psutil', type=('build', 'run'))
depends_on('py-pyyaml', type=('build', 'run'))
depends_on('py-six', type=('build', 'run'))
# Deep learning frameworks
depends_on('py-torch@0.4.0:', type=('build', 'run'), when='+pytorch')
depends_on('py-torch+cuda', type=('build', 'run'), when='+pytorch+cuda')
depends_on('py-cffi@1.4.0:', type=('build', 'run'), when='+pytorch')
depends_on('py-tensorflow@1.1.0:', type=('build', 'link', 'run'), when='+tensorflow')
depends_on('mxnet@1.4.0:+python', type=('build', 'link', 'run'), when='+mxnet')
depends_on('mxnet+cuda', type=('build', 'link', 'run'), when='+mxnet+cuda')
# Distributed support
# There does not appear to be a way to use an external Gloo installation
depends_on('cmake', type='build', when='+gloo')
depends_on('mpi', when='+mpi')
depends_on('mpi', when='gpu_allreduce=mpi')
depends_on('mpi', when='gpu_allgather=mpi')
depends_on('mpi', when='gpu_broadcast=mpi')
# GPU support
depends_on('cuda', when='+cuda')
depends_on('nccl@2.0:', when='gpu_allreduce=nccl')
depends_on('nccl@2.0:', when='gpu_broadcast=nccl')
# Test dependencies
depends_on('py-mock', type='test')
depends_on('py-pytest', type='test')
depends_on('py-pytest-forked', type='test')
conflicts('+gloo', when='platform=darwin', msg='Gloo cannot be compiled on MacOS')
conflicts('~gloo~mpi', msg='One of Gloo or MPI are required for Horovod to run')
conflicts('~pytorch~tensorflow~mxnet', msg='At least one deep learning backend is required')
def setup_build_environment(self, env):
# Deep learning frameworks
if '~pytorch' in self.spec:
env.set('HOROVOD_WITHOUT_PYTORCH', 1)
if '~tensorflow' in self.spec:
env.set('HOROVOD_WITHOUT_TENSORFLOW', 1)
if '~mxnet' in self.spec:
env.set('HOROVOD_WITHOUT_MXNET', 1)
# Distributed support
if '~gloo' in self.spec:
env.set('HOROVOD_WITHOUT_GLOO', 1)
if '+mpi' in self.spec:
env.set('HOROVOD_WITH_MPI', 1)
else:
env.set('HOROVOD_WITHOUT_MPI', 1)
# GPU support
if '+cuda' in self.spec:
env.set('HOROVOD_CUDA_HOME', self.spec['cuda'].prefix)
env.set('HOROVOD_CUDA_INCLUDE',
self.spec['cuda'].headers.directories[0])
env.set('HOROVOD_CUDA_LIB', self.spec['cuda'].libs.directories[0])
if '^nccl' in self.spec:
env.set('HOROVOD_NCCL_HOME', self.spec['nccl'].prefix)
env.set('HOROVOD_NCCL_INCLUDE',
self.spec['nccl'].headers.directories[0])
env.set('HOROVOD_NCCL_LIB', self.spec['nccl'].libs.directories[0])
env.set('HOROVOD_GPU_ALLREDUCE',
self.spec.variants['gpu_allreduce'].value.upper())
env.set('HOROVOD_GPU_ALLGATHER',
self.spec.variants['gpu_allgather'].value.upper())
env.set('HOROVOD_GPU_BROADCAST',
self.spec.variants['gpu_broadcast'].value.upper())
env.set('HOROVOD_ALLOW_MIXED_GPU_IMPL', 1)
|
