forked from horovod/horovod
-
Notifications
You must be signed in to change notification settings - Fork 0
/
docker-compose.test.yml
283 lines (275 loc) · 11.5 KB
/
docker-compose.test.yml
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
version: '2.3'
services:
test-cpu-base:
build:
context: .
dockerfile: Dockerfile.test.cpu
args:
UBUNTU_VERSION: 20.04
GPP_VERSION: 7
MPI_KIND: None
PYTHON_VERSION: 3.8
TENSORFLOW_PACKAGE: tensorflow-cpu==2.12.0
KERAS_PACKAGE: keras==2.12.0
PYTORCH_PACKAGE: torch==2.0.0+cpu
PYTORCH_LIGHTNING_PACKAGE: pytorch-lightning==1.5.9
TORCHVISION_PACKAGE: torchvision==0.15.1+cpu
MXNET_PACKAGE: mxnet==1.9.1
PYSPARK_PACKAGE: pyspark==3.4.0
SPARK_PACKAGE: spark-3.4.0/spark-3.4.0-bin-hadoop3.tgz
HOROVOD_BUILD_FLAGS: HOROVOD_WITH_GLOO=1
privileged: true
shm_size: 8gb
# our baseline first
test-cpu-gloo-py3_8-tf2_12_0-keras2_12_0-torch2_0_0-mxnet1_9_1-pyspark3_4_0:
extends: test-cpu-base
# permute MPI kinds
test-cpu-mpich-py3_8-tf2_12_0-keras2_12_0-torch2_0_0-mxnet1_9_1-pyspark3_4_0:
extends: test-cpu-base
build:
args:
MPI_KIND: MPICH
HOROVOD_BUILD_FLAGS: HOROVOD_WITHOUT_GLOO=1
test-cpu-oneccl-py3_8-tf2_12_0-keras2_12_0-torch2_0_0-mxnet1_9_1-pyspark3_4_0:
extends: test-cpu-base
build:
args:
MPI_KIND: ONECCL
HOROVOD_BUILD_FLAGS: HOROVOD_WITHOUT_GLOO=1
test-cpu-openmpi-py3_8-tf2_12_0-keras2_12_0-torch2_0_0-mxnet1_9_1-pyspark3_4_0:
extends: test-cpu-base
build:
args:
MPI_KIND: OpenMPI
HOROVOD_BUILD_FLAGS: HOROVOD_WITHOUT_GLOO=1
test-cpu-openmpi-gloo-py3_8-tf2_12_0-keras2_12_0-torch2_0_0-mxnet1_9_1-pyspark3_4_0:
extends: test-cpu-base
build:
args:
MPI_KIND: OpenMPI
# run_gloo_integration expects tf1 to have Gloo mpi kind to run 'Elastic Spark * Tests'
# Tensorflow 1.15.5 is only available for Python 3.7
# Python 3.7 is only available on Ubuntu 18.04
# torch==1.8.1 is the latest we can test in this setup
# there is no mxnet-1.6.0.post0 and mxnet-1.6.0 does not work with horovod
# https://github.com/apache/incubator-mxnet/issues/16193
# so we test with mxnet 1.5.1
test-cpu-gloo-py3_7-tf1_15_5-keras2_2_4-torch1_8_1-mxnet1_5_1_p0-pyspark3_4_0:
extends: test-cpu-base
build:
args:
# On Ubuntu 18.04 our setup.py will pull in a recent CMake and use that only to build Horovod
UBUNTU_VERSION: 18.04
PYTHON_VERSION: 3.7
# there is no tensorflow-cpu>1.15.0, so we use tensorflow==1.15.5
TENSORFLOW_PACKAGE: tensorflow==1.15.5
KERAS_PACKAGE: keras==2.2.4
PYTORCH_PACKAGE: torch==1.8.1+cpu
TORCHVISION_PACKAGE: torchvision==0.9.1+cpu
MXNET_PACKAGE: mxnet==1.5.1.post0
test-cpu-gloo-py3_8-tf2_10_1-keras2_10_0-torch1_12_1-mxnet1_7_0_p2-pyspark3_4_0:
extends: test-cpu-base
build:
args:
TENSORFLOW_PACKAGE: tensorflow-cpu==2.10.1
KERAS_PACKAGE: keras==2.10.0
PYTORCH_PACKAGE: torch==1.12.1+cpu
TORCHVISION_PACKAGE: torchvision==0.13.1+cpu
MXNET_PACKAGE: mxnet==1.7.0.post2
test-cpu-gloo-py3_8-tf2_11_1-keras2_11_0-torch1_13_1-mxnet1_8_0_p0-pyspark3_4_0:
extends: test-cpu-base
build:
args:
TENSORFLOW_PACKAGE: tensorflow-cpu==2.11.1
KERAS_PACKAGE: keras==2.11.0
PYTORCH_PACKAGE: torch==1.13.1+cpu
TORCHVISION_PACKAGE: torchvision==0.14.1+cpu
MXNET_PACKAGE: mxnet==1.8.0.post0
# then our baseline again, omitted ...
test-cpu-openmpi-gloo-py3_8-tfhead-keras_none-torchhead-mxnethead-pyspark3_4_0:
extends: test-cpu-base
build:
args:
MPI_KIND: OpenMPI
TENSORFLOW_PACKAGE: tf-nightly
KERAS_PACKAGE: None
PYTORCH_PACKAGE: torch-nightly
TORCHVISION_PACKAGE: torchvision
PYTORCH_LIGHTNING_PACKAGE: pytorch-lightning==1.5.9
MXNET_PACKAGE: mxnet-nightly
# these are the lowest framework versions that Horovod compiles with, but they are not tested
test-cpu-openmpi-gloo-py3_7-tfmin-kerasmin-torchmin-mxnetmin-pysparkmin:
extends: test-cpu-base
build:
args:
UBUNTU_VERSION: 18.04
PYTHON_VERSION: 3.7
MPI_KIND: OpenMPI
TENSORFLOW_PACKAGE: tensorflow-cpu==1.15.0
KERAS_PACKAGE: keras==2.2.4
PYTORCH_PACKAGE: torch==1.5.0+cpu
PYTORCH_LIGHTNING_PACKAGE: pytorch-lightning==0.7.3
TORCHVISION_PACKAGE: torchvision==0.6.0+cpu
MXNET_PACKAGE: mxnet==1.4.1
PYSPARK_PACKAGE: pyspark==2.4.0
SPARK_PACKAGE: spark-2.4.0/spark-2.4.0-bin-hadoop2.7.tgz
# we deviate from baseline here because PySpark 2.4 requires Python 3.7 and
# Tensorflow 2.11.0 is the last version that supports that Python
# Torch 1.13.1 is the last version that supports that Python
test-cpu-gloo-py3_7-tf2_11_0-keras2_11_0-torch1_13_1-mxnet1_9_1-pyspark2_4_8:
extends: test-cpu-base
build:
args:
# PySpark 2.4.8 is only available for Python 3.7
# Python 3.7 is only available on Ubuntu 18.04
# Tensorflow 2.11.0 is the last version supporting that Python
# Torch 1.13.1 is the last version supporting that Python
UBUNTU_VERSION: 18.04
PYTHON_VERSION: 3.7
TENSORFLOW_PACKAGE: tensorflow-cpu==2.11.0
KERAS_PACKAGE: keras==2.11.0
PYTORCH_PACKAGE: torch==1.13.1+cpu
TORCHVISION_PACKAGE: torchvision==0.14.1+cpu
PYSPARK_PACKAGE: pyspark==2.4.8
SPARK_PACKAGE: spark-2.4.8/spark-2.4.8-bin-hadoop2.7.tgz
test-cpu-gloo-py3_8-tf2_12_0-keras2_12_0-torch2_0_0-mxnet1_9_1-pyspark3_3_2:
extends: test-cpu-base
build:
args:
PYTHON_VERSION: 3.8
PYSPARK_PACKAGE: pyspark==3.3.2
SPARK_PACKAGE: spark-3.3.2/spark-3.3.2-bin-hadoop2.tgz
# then our baseline again, omitted ...
test-gpu-base:
build:
context: .
dockerfile: Dockerfile.test.gpu
args:
GPP_VERSION: 7
MPI_KIND: None
PYTHON_VERSION: 3.8
PYSPARK_PACKAGE: pyspark==3.4.0
SPARK_PACKAGE: spark-3.4.0/spark-3.4.0-bin-hadoop3.tgz
HOROVOD_BUILD_FLAGS: HOROVOD_GPU_OPERATIONS=NCCL
HOROVOD_MIXED_INSTALL: 0
runtime: nvidia
# We plumb CUDA_VISIBLE_DEVICES instead of NVIDIA_VISIBLE_DEVICES because
# the latter does not work in privileged mode that we use in the containers.
environment:
- CUDA_VISIBLE_DEVICES
privileged: true
shm_size: 8gb
# available versions for CUDNN_VERSION and NCCL_VERSION_OVERRIDE can be found at
# https://developer.download.nvidia.com/compute/cuda/repos/{OS}/x86_64/
# Mainline tensorflow-gpu==1.15.5 is compiled against and linked to CUDA 10.0, but appropriate containers aren't
# available anymore. Hence, we use the updated Python 3.8 wheel provided by Nvidia, see
# https://github.com/NVIDIA/tensorflow. For this reason versions of torch and mxnet also deviate from the CPU path.
test-gpu-gloo-py3_8-tf1_15_5-keras2_2_4-torch1_12_1-mxnet1_8_0_p0-pyspark3_4_0:
extends: test-gpu-base
build:
args:
CUDA_DOCKER_VERSION: 11.6.2-devel-ubuntu20.04
CUDNN_VERSION: 8.4.1.50-1+cuda11.6
NCCL_VERSION_OVERRIDE: 2.11.4-1+cuda11.6
PYTHON_VERSION: 3.8
TENSORFLOW_PACKAGE: nvidia-tensorflow==1.15.5+nv22.4
KERAS_PACKAGE: keras==2.2.4
PYTORCH_PACKAGE: torch==1.12.1+cu116
PYTORCH_LIGHTNING_PACKAGE: pytorch-lightning==1.5.9
TORCHVISION_PACKAGE: torchvision==0.13.1+cu116
MXNET_PACKAGE: mxnet-cu112==1.8.0.post0
# The container isn't provided for CUDA 10 anymore. The lowest version of mxnet available for cu112 is 1.8.0.post0.
test-gpu-gloo-py3_8-tf2_10_1-keras2_10_0-torch1_12_1-mxnet1_8_0_p0-pyspark3_4_0:
extends: test-gpu-base
build:
args:
CUDA_DOCKER_VERSION: 11.6.2-devel-ubuntu20.04
CUDNN_VERSION: 8.4.1.50-1+cuda11.6
NCCL_VERSION_OVERRIDE: 2.11.4-1+cuda11.6
TENSORFLOW_PACKAGE: tensorflow-gpu==2.10.1
KERAS_PACKAGE: keras==2.10.0
PYTORCH_PACKAGE: torch==1.12.1+cu116
PYTORCH_LIGHTNING_PACKAGE: pytorch-lightning==1.5.9
TORCHVISION_PACKAGE: torchvision==0.13.1+cu116
MXNET_PACKAGE: mxnet-cu112==1.8.0.post0
test-gpu-gloo-py3_8-tf2_11_1-keras2_11_0-torch1_13_1-mxnet1_8_0_p0-pyspark3_4_0:
extends: test-gpu-base
build:
args:
CUDA_DOCKER_VERSION: 11.6.2-devel-ubuntu20.04
CUDNN_VERSION: 8.4.1.50-1+cuda11.6
NCCL_VERSION_OVERRIDE: 2.11.4-1+cuda11.6
# tensorflow package supports GPU from 2.11.1 and 2.12.0 on
TENSORFLOW_PACKAGE: tensorflow==2.11.1
KERAS_PACKAGE: keras==2.11.0
PYTORCH_PACKAGE: torch==1.13.1+cu116
PYTORCH_LIGHTNING_PACKAGE: pytorch-lightning==1.5.9
TORCHVISION_PACKAGE: torchvision==0.14.1+cu116
MXNET_PACKAGE: mxnet-cu112==1.8.0.post0
test-gpu-openmpi-gloo-py3_8-tf2_12_0-keras2_12_0-torch2_0_0-mxnet1_9_1-pyspark3_4_0:
extends: test-gpu-base
build:
args:
CUDA_DOCKER_VERSION: 11.8.0-devel-ubuntu20.04
CUDNN_VERSION: 8.6.0.163-1+cuda11.8
NCCL_VERSION_OVERRIDE: 2.16.5-1+cuda11.8
MPI_KIND: OpenMPI
# tensorflow package supports GPU from 2.11.1 and 2.12.0 on
TENSORFLOW_PACKAGE: tensorflow==2.12.0
KERAS_PACKAGE: keras==2.12.0
PYTORCH_PACKAGE: torch==2.0.0+cu118
PYTORCH_LIGHTNING_PACKAGE: pytorch-lightning==1.5.9
TORCHVISION_PACKAGE: torchvision==0.15.1+cu118
MXNET_PACKAGE: mxnet-cu112==1.9.1
test-gpu-openmpi-gloo-py3_8-tfhead-keras_none-torchhead-mxnethead-pyspark3_4_0:
extends: test-gpu-base
build:
args:
CUDA_DOCKER_VERSION: 11.8.0-devel-ubuntu20.04
CUDNN_VERSION: 8.6.0.163-1+cuda11.8
NCCL_VERSION_OVERRIDE: 2.16.5-1+cuda11.8
MPI_KIND: OpenMPI
TENSORFLOW_PACKAGE: tf-nightly
KERAS_PACKAGE: None
PYTORCH_PACKAGE: torch-nightly-cu118
PYTORCH_LIGHTNING_PACKAGE: pytorch-lightning==1.5.9
TORCHVISION_PACKAGE: torchvision
MXNET_PACKAGE: mxnet-nightly-cu112
# These are the lowest framework versions that Horovod compiles with on the CUDA 11.x container, but they are not tested.
# Versions of python, mxnet, and pyspark differ from the CPU build with minimum versions.
test-gpu-openmpi-gloo-py3_8-tfmin-kerasmin-torchmin-mxnetmin-pysparkmin:
extends: test-gpu-base
build:
args:
CUDA_DOCKER_VERSION: 11.6.2-devel-ubuntu20.04
CUDNN_VERSION: 8.4.1.50-1+cuda11.6
NCCL_VERSION_OVERRIDE: 2.11.4-1+cuda11.6
MPI_KIND: OpenMPI
PYTHON_VERSION: 3.8
TENSORFLOW_PACKAGE: nvidia-tensorflow==1.15.5+nv22.4
KERAS_PACKAGE: keras==2.2.4
# torch ships its own CUDA libraries
PYTORCH_PACKAGE: torch==1.5.0+cu101
PYTORCH_LIGHTNING_PACKAGE: pytorch-lightning==0.7.3
TORCHVISION_PACKAGE: torchvision==0.6.0+cu101
MXNET_PACKAGE: mxnet-cu112==1.8.0.post0
# On Python 3.8 Spark 3.0.0 is the lowest supported version
PYSPARK_PACKAGE: pyspark==3.0.0
SPARK_PACKAGE: spark-3.0.0/spark-3.0.0-bin-hadoop2.7.tgz
test-mixed-openmpi-gloo-py3_8-tf2_12_0-keras2_12_0-torch2_0_0-mxnet1_9_1-pyspark3_4_0:
extends: test-gpu-base
build:
args:
CUDA_DOCKER_VERSION: 11.8.0-devel-ubuntu20.04
CUDNN_VERSION: 8.6.0.163-1+cuda11.8
NCCL_VERSION_OVERRIDE: 2.16.5-1+cuda11.8
MPI_KIND: OpenMPI
# tensorflow package supports GPU from 2.11.1 and 2.12.0 on
TENSORFLOW_PACKAGE: tensorflow==2.12.0
KERAS_PACKAGE: keras==2.12.0
PYTORCH_PACKAGE: torch==2.0.0+cu118
PYTORCH_LIGHTNING_PACKAGE: pytorch-lightning==1.5.9
TORCHVISION_PACKAGE: torchvision==0.15.1+cu118
MXNET_PACKAGE: mxnet-cu112==1.9.1
HOROVOD_BUILD_FLAGS: ""
HOROVOD_MIXED_INSTALL: 1