forked from NVIDIA/Megatron-LM
-
Notifications
You must be signed in to change notification settings - Fork 0
/
.gitlab-ci.yml
357 lines (321 loc) · 13 KB
/
.gitlab-ci.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
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
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
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
workflow:
rules:
- if: $CI_PIPELINE_SOURCE == "schedule"
variables:
FUNCTIONAL_TEST: "yes"
- if: $CI_PIPELINE_SOURCE == "web"
- if: $CI_COMMIT_BRANCH == $CI_DEFAULT_BRANCH
variables:
FUNCTIONAL_TEST: "no"
- if: $CI_COMMIT_BRANCH =~ /^core_r/
variables:
FUNCTIONAL_TEST: "no"
- if: $CI_MERGE_REQUEST_LABELS =~ /Run tests and nightly/
variables:
FUNCTIONAL_TEST: "yes"
SLURM_CLUSTER: dgxa100_dracooci
SCOPE: mr-and-nightly
- if: $CI_MERGE_REQUEST_LABELS =~ /Run tests/
variables:
FUNCTIONAL_TEST: "yes"
SLURM_CLUSTER: dgxa100_dracooci
SCOPE: mr
- if: $CI_PIPELINE_SOURCE == "merge_request_event"
variables:
FUNCTIONAL_TEST: "no"
- when: never
auto_cancel:
on_new_commit: interruptible
stages:
- build
- unit_tests
- functional_tests
default:
interruptible: true
variables:
FUNCTIONAL_TEST: "yes"
SCOPE:
value: "mr"
options:
- "mr"
- "nightly"
- "mr-and-nightly"
- "weekly"
- "release"
description: "Testsuite to run"
SLURM_CLUSTER:
value: "dgxa100_dracooci"
options:
- "dgxa100_dracooci"
- "dgxh100_eos"
description: '"dgxa100_dracooci" for OCI-IAD, "dgxh100_eos" for EOS'
# CI wide variables
CI_MCORE_IMAGE: gitlab-master.nvidia.com:5005/adlr/megatron-lm/mcore_ci
CI_NEMO_IMAGE: gitlab-master.nvidia.com:5005/adlr/megatron-lm/nemo_ci
LINTING_IMAGE: gitlab-master.nvidia.com:5005/adlr/megatron-lm/mcore_linting
metadata:
image: python:3.10
stage: .pre
tags:
- os/linux
script:
- set -x
- env
- JET_CUSTOM_FILTER="type == 'basic'"
- |
if [[ $SLURM_CLUSTER == dgxh100_eos ]]; then
JET_CI_BRANCH=mcore/eos
JET_CUSTOM_FILTER="$JET_CUSTOM_FILTER and 'dgx_h100' in spec.platforms"
elif [[ $SLURM_CLUSTER == dgxa100_dracooci ]]; then
JET_CI_BRANCH=mcore/draco-oci
JET_CUSTOM_FILTER="$JET_CUSTOM_FILTER and 'dgx_a100' in spec.platforms"
fi
- |
if [[ $SCOPE == mr ]]; then
JET_CUSTOM_FILTER="$JET_CUSTOM_FILTER and 'mr' in spec.scope"
elif [[ $SCOPE == nightly ]]; then
JET_CUSTOM_FILTER="$JET_CUSTOM_FILTER and 'nightly' in spec.scope"
elif [[ $SCOPE == mr-and-nightly ]]; then
JET_CUSTOM_FILTER="$JET_CUSTOM_FILTER and ('mr' in spec.scope or 'nightly' in spec.scope)"
elif [[ $SCOPE == weekly ]]; then
JET_CUSTOM_FILTER="$JET_CUSTOM_FILTER and 'weekly' in spec.scope"
elif [[ $SCOPE == release ]]; then
JET_CUSTOM_FILTER="$JET_CUSTOM_FILTER and 'release' in spec.scope"
fi
- |
if [[ "$JET_CUSTOM_FILTER" == "type == 'basic'" ]]; then
JET_CUSTOM_FILTER="False"
fi
- echo "JET_CI_BRANCH=$JET_CI_BRANCH" | tee -a build.env
- echo "JET_CUSTOM_FILTER=$JET_CUSTOM_FILTER" | tee -a build.env
artifacts:
reports:
dotenv: build.env
rules:
- if: '$FUNCTIONAL_TEST == "yes"'
ppp_capacity_statistics:
tags: [mcore-ssh-agent]
stage: .pre
script:
- |
set -x
ALL_USER=$(sshare -aP | grep coreai_dlalgo_mcore | tail -n +2 | awk -F '|' '{print $2}' | tr '\n' ',')
# Get the current year, month, and day
YEAR=$(date +%Y)
MONTH=$(date +%m)
DAY=$([[ $(date +%-d) -le 15 ]] && echo "01" || echo "15")
TIMESTAMP="${YEAR}-${MONTH}-${DAY}T00:00:01"
CLUSTER_ID=$(curl "${RESOURCE_ENDPOINT}/api/v1/clusters" \
-H "accept: application/json, text/plain, */*" \
-H "accept-language: en-US,en;q=0.9" \
-H "authorization: Bearer $CSRG_API_KEY" | jq '.[] | select(.name == "draco-oci-iad") | .id' | tr -d '"')
INITIATIVE_ITEM_ID=$(curl "${RESOURCE_ENDPOINT}/api/v1/initiative-items" \
-H "accept: application/json, text/plain, */*" \
-H "accept-language: en-US,en;q=0.9" \
-H "authorization: Bearer $CSRG_API_KEY" | jq '.[] | select(.name == "coreai_dlalgo_mcore") | .id' | tr -d '"')
QUOTA=$(curl "${RESOURCE_ENDPOINT}/api/v1/capacity-requests" \
-H "accept: application/json, text/plain, */*" \
-H "accept-language: en-US,en;q=0.9" \
-H "authorization: Bearer $CSRG_API_KEY" | jq --arg CLUSTER_ID $CLUSTER_ID --arg INITIATIVE_ITEM_ID $INITIATIVE_ITEM_ID '[.[] | select(.clusterId == $CLUSTER_ID and .initiativeItemId == $INITIATIVE_ITEM_ID)] | to_entries | [last] | .[0].value.quantity')
USED_CAPA=$(sacct \
-u ${ALL_USER} \
--partition batch_block1,batch_block3,batch_block4 \
--truncate \
-A coreai_dlalgo_mcore \
-S ${TIMESTAMP} \
-X \
--format JobID,JobName%20,Partition,AllocNodes,ElapsedRaw \
-p \
-n \
| awk -F "|" '{{sum+=$4*$5}} END {{print sum*8/3600}}')
TOTAL_CAPA=$(( $QUOTA*24*30 ))
USAGE=$(echo "$USED_CAPA $TOTAL_CAPA" | awk '{print (1 - $1/$2)*100}')%
echo "Usage left: $USAGE"
echo "Disclaimer: Please be careful with this number. Usage does not imply
what we are guaranteed to get a slot, SLURM scheduling is more complicated
than that. The number is rather a proxy to the FairShare that determines
our job-scheduling-priority.
Most important take-away of this number is to get a sense how much much
we are eating up our budget such that we can discuss this with capacity planning.
"
build_image:
tags:
- mcore-docker-node
image: docker:26.1.4-dind
needs: [] # May start ASAP
stage: build
timeout: 45m
parallel:
matrix:
- IMAGE: CI_MCORE_IMAGE
FILE: Dockerfile.ci
BASE_IMAGE: nvcr.io/nvidia/pytorch:24.01-py3
- IMAGE: CI_NEMO_IMAGE
FILE: Dockerfile.ci
BASE_IMAGE: nvcr.io/nvidian/nemo:nightly
- IMAGE: LINTING_IMAGE
FILE: Dockerfile.linting
BASE_IMAGE: python:3.10
before_script:
- echo "$NGC_API_KEY" | docker login nvcr.io -u '$oauthtoken' --password-stdin
- echo "$CI_REGISTRY_PASSWORD" | docker login $CI_REGISTRY -u $CI_REGISTRY_USER --password-stdin
script:
- |
set -x
eval "IMAGE=\$$IMAGE"
OLD_IMAGES=$(docker image ls --format "{{.ID}} {{.Repository}}:{{.Tag}}" \
| grep -v 'nvcr.io/nvidia/pytorch:24.01-py3' \
| grep -v 'gitlab-master.nvidia.com:5005/adlr/megatron-lm/mcore_ci:buildcache' \
| grep -v 'gitlab-master.nvidia.com:5005/adlr/megatron-lm/mcore_nemo:buildcache' \
| grep -v 'gitlab-master.nvidia.com:5005/adlr/megatron-lm/mcore_linting:buildcache' \
| grep -v 'nvcr.io/nvidian/nemo:nightly' \
| grep -v 'python:3.10' | awk '{ print $1 }'
)
docker rmi $OLD_IMAGES || true
docker builder prune -a --filter "until=24h" -f
if [[ "$CI_COMMIT_BRANCH" == "$CI_DEFAULT_BRANCH" ]]; then
ADDITIONAL_PARAMS="--pull"
fi
docker build \
-f $FILE \
-t ${IMAGE}:${CI_PIPELINE_ID} \
--cache-to type=inline \
--cache-from type=registry,ref=${IMAGE}:buildcache \
--build-arg FROM_IMAGE_NAME=$BASE_IMAGE \
${ADDITIONAL_PARAMS} .
docker push ${IMAGE}:${CI_PIPELINE_ID}
if [[ "$CI_COMMIT_BRANCH" == "$CI_DEFAULT_BRANCH" ]]; then
docker tag ${IMAGE}:${CI_PIPELINE_ID} ${IMAGE}:buildcache
docker push ${IMAGE}:buildcache
fi
if [[ $CI_COMMIT_BRANCH == core_r* ]]; then
docker tag ${IMAGE}:${CI_PIPELINE_ID} ${IMAGE}:v${CI_COMMIT_BRANCH#core_r}-${CI_PIPELINE_ID}
docker push ${IMAGE}:v${CI_COMMIT_BRANCH#core_r}-${CI_PIPELINE_ID}
fi
.unit_test_common:
image: ${CI_MCORE_IMAGE}:${CI_PIPELINE_ID}
stage: unit_tests
needs: [build_image]
tags:
- 8xL40S
variables:
MOE_GROUPED_GEMM: 0 # Set to 1 to enable grouped gemm for MoE
retry:
max: 2
when: job_execution_timeout
unit_tests:
extends: [.unit_test_common]
script:
- torchrun --nproc_per_node=8 -m pytest -x -v -s --cov-report=term --cov-report=html --cov=megatron/core --no-cov-on-fail tests/unit_tests
coverage: '/(?i)total.*? (100(?:\.0+)?\%|[1-9]?\d(?:\.\d+)?\%)$/'
artifacts:
paths:
- coverage
expire_in: 30 days
rules:
- if: '$FUNCTIONAL_TEST == "yes" && $CI_PIPELINE_SOURCE == "merge_request_event" && ($CI_MERGE_REQUEST_TARGET_BRANCH_NAME != $CI_DEFAULT_BRANCH && $CI_MERGE_REQUEST_TARGET_BRANCH_NAME !~ /^core_r/)'
allow_failure: true
- if: '$FUNCTIONAL_TEST == "yes"'
unit_tests-data:
extends: [.unit_test_common]
script:
- torchrun --nproc_per_node=8 -m pytest -x -v -s tests/unit_tests/data
rules:
- if: '$FUNCTIONAL_TEST == "no" && $CI_PIPELINE_SOURCE == "merge_request_event" && ($CI_MERGE_REQUEST_TARGET_BRANCH_NAME != $CI_DEFAULT_BRANCH && $CI_MERGE_REQUEST_TARGET_BRANCH_NAME !~ /^core_r/)'
allow_failure: true
- if: '$FUNCTIONAL_TEST == "no"'
unit_tests-dist-checkpointing:
extends: [.unit_test_common]
script:
- torchrun --nproc_per_node=8 -m pytest -x -v -s tests/unit_tests/dist_checkpointing
rules:
- if: '$FUNCTIONAL_TEST == "no" && $CI_PIPELINE_SOURCE == "merge_request_event" && ($CI_MERGE_REQUEST_TARGET_BRANCH_NAME != $CI_DEFAULT_BRANCH && $CI_MERGE_REQUEST_TARGET_BRANCH_NAME !~ /^core_r/)'
allow_failure: true
- if: '$FUNCTIONAL_TEST == "no"'
unit_tests-fusions:
extends: [.unit_test_common]
script:
- torchrun --nproc_per_node=8 -m pytest -x -v -s tests/unit_tests/fusions
rules:
- if: '$FUNCTIONAL_TEST == "no" && $CI_PIPELINE_SOURCE == "merge_request_event" && ($CI_MERGE_REQUEST_TARGET_BRANCH_NAME != $CI_DEFAULT_BRANCH && $CI_MERGE_REQUEST_TARGET_BRANCH_NAME !~ /^core_r/)'
allow_failure: true
- if: '$FUNCTIONAL_TEST == "no"'
unit_tests-inference:
extends: [.unit_test_common]
script:
- torchrun --nproc_per_node=8 -m pytest -x -v -s tests/unit_tests/inference
rules:
- if: '$FUNCTIONAL_TEST == "no" && $CI_PIPELINE_SOURCE == "merge_request_event" && ($CI_MERGE_REQUEST_TARGET_BRANCH_NAME != $CI_DEFAULT_BRANCH && $CI_MERGE_REQUEST_TARGET_BRANCH_NAME !~ /^core_r/)'
allow_failure: true
- if: '$FUNCTIONAL_TEST == "no"'
unit_tests-models:
extends: [.unit_test_common]
script:
- torchrun --nproc_per_node=8 -m pytest -x -v -s tests/unit_tests/models
rules:
- if: '$FUNCTIONAL_TEST == "no" && $CI_PIPELINE_SOURCE == "merge_request_event" && ($CI_MERGE_REQUEST_TARGET_BRANCH_NAME != $CI_DEFAULT_BRANCH && $CI_MERGE_REQUEST_TARGET_BRANCH_NAME !~ /^core_r/)'
allow_failure: true
- if: '$FUNCTIONAL_TEST == "no"'
unit_tests-pipeline-parallel:
extends: [.unit_test_common]
script:
- torchrun --nproc_per_node=8 -m pytest -x -v -s tests/unit_tests/pipeline_parallel
rules:
- if: '$FUNCTIONAL_TEST == "no" && $CI_PIPELINE_SOURCE == "merge_request_event" && ($CI_MERGE_REQUEST_TARGET_BRANCH_NAME != $CI_DEFAULT_BRANCH && $CI_MERGE_REQUEST_TARGET_BRANCH_NAME !~ /^core_r/)'
allow_failure: true
- if: '$FUNCTIONAL_TEST == "no"'
unit_tests-tensor-parallel:
extends: [.unit_test_common]
script:
- torchrun --nproc_per_node=8 -m pytest -x -v -s tests/unit_tests/tensor_parallel
rules:
- if: '$FUNCTIONAL_TEST == "no" && $CI_PIPELINE_SOURCE == "merge_request_event" && ($CI_MERGE_REQUEST_TARGET_BRANCH_NAME != $CI_DEFAULT_BRANCH && $CI_MERGE_REQUEST_TARGET_BRANCH_NAME !~ /^core_r/)'
allow_failure: true
- if: '$FUNCTIONAL_TEST == "no"'
unit_tests-transformer:
extends: [.unit_test_common]
script:
- torchrun --nproc_per_node=8 -m pytest -x -v -s tests/unit_tests/transformer
rules:
- if: '$FUNCTIONAL_TEST == "no" && $CI_PIPELINE_SOURCE == "merge_request_event" && ($CI_MERGE_REQUEST_TARGET_BRANCH_NAME != $CI_DEFAULT_BRANCH && $CI_MERGE_REQUEST_TARGET_BRANCH_NAME !~ /^core_r/)'
allow_failure: true
- if: '$FUNCTIONAL_TEST == "no"'
unit_tests-top-py:
extends: [.unit_test_common]
script:
- torchrun --nproc_per_node=8 -m pytest -x -v -s tests/unit_tests/*.py
rules:
- if: '$FUNCTIONAL_TEST == "no" && $CI_PIPELINE_SOURCE == "merge_request_event" && ($CI_MERGE_REQUEST_TARGET_BRANCH_NAME != $CI_DEFAULT_BRANCH && $CI_MERGE_REQUEST_TARGET_BRANCH_NAME !~ /^core_r/)'
allow_failure: true
- if: '$FUNCTIONAL_TEST == "no"'
docs_build_test:
image: gitlab-master.nvidia.com:5005/adlr/megatron-lm/python-format:0.0.1
stage: unit_tests
tags:
- os/linux
script:
- cd ..
- rm -rf documentation && git clone https://gitlab-ci-token:${CI_JOB_TOKEN}@gitlab-master.nvidia.com/nemo-megatron-core-tme/documentation.git
- mv megatron-lm/ documentation/
- cd documentation/
- ./repo docs
allow_failure: true
except:
- main
interruptible: true
formatting:
image: ${LINTING_IMAGE}:${CI_PIPELINE_ID}
tags:
- os/linux
stage: unit_tests
before_script:
- git fetch origin main
script:
- CHECK_ONLY=true bash tools/autoformat.sh
rules:
- if: '$CI_PIPELINE_SOURCE == "merge_request_event" && ($CI_MERGE_REQUEST_TARGET_BRANCH_NAME != $CI_DEFAULT_BRANCH && $CI_MERGE_REQUEST_TARGET_BRANCH_NAME !~ /^core_r/)'
allow_failure: true
- when: always
interruptible: true
include:
- jet-tests.yml