@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / e3c664bfcb
Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 16, 19:30 UTC

[Build] Add initial conditional testing spec (#6841)

Simon Mo committed Aug 5, 2024 at 17:39 UTC e3c664bfcb14a41e43ddb6078ed1464ae9b7852f
1 file changed +234 -156
.buildkite/test-pipeline.yaml
+234 -156
@@ -5,11 +5,47 @@
5 # https://github.com/vllm-project/buildkite-ci/blob/main/scripts/test-template-aws.j2
6 # to generate the final pipeline yaml file.
7
8 +# Documentation
9 +# label(str): the name of the test. emoji allowed.
10 +# fast_check(bool): whether to run this on each commit on fastcheck pipeline.
11 +# fast_check_only(bool): run this test on fastcheck pipeline only
12 +# command(str): the single command to run for tests. incompatible with commands.
13 +# commands(list): the list of commands to run for test. incompatbile with command.
14 +# mirror_hardwares(list): the list of hardwares to run the test on as well. currently only supports [amd]
15 +# gpu(str): override the GPU selection for the test. default is on L4 GPUs. currently only supports a100
16 +# num_gpus(int): override the number of GPUs for the test. default to 1 GPU. currently support 2,4.
17 +# num_nodes(int): whether to simulate multi-node setup by launch multiple containers on one host,
18 +# in this case, commands must be specified. the first command runs on first host, the second
19 +# command runs on the second host.
20 +# working_dir(str): specify the place where command should execute, default to /vllm-workspace/tests
21 +# source_file_dependencies(list): the list of prefix to opt-in the test for, if empty, the test will always run.
22 +
23 +# When adding a test
24 +# - If the test belong to an existing group, add it there
25 +# - If the test is short, add to any existing step
26 +# - If the test takes more than 10min, then it is okay to create a new step.
27 +# Note that all steps execute in parallel.
28
29 steps:
10 -- label: Async Engine, Inputs, Utils, Worker Test
30 +##### fast check tests #####
31 +
32 +- label: Documentation Build # 2min
33 + working_dir: "/vllm-workspace/test_docs/docs"
34 fast_check: true
12 - fast_check_only: true
35 + no_gpu: True
36 + commands:
37 + - pip install -r requirements-docs.txt
38 + - SPHINXOPTS=\"-W\" make html
39 +
40 +- label: Async Engine, Inputs, Utils, Worker Test # 15min
41 + fast_check: true
42 + source_file_dependencies:
43 + - vllm/
44 + - tests/async_engine
45 + - tests/test_inputs
46 + - tests/multimodal
47 + - tests/test_utils
48 + - tests/worker
49 commands:
50 - pytest -v -s async_engine # Async Engine
51 - pytest -v -s test_inputs.py
@@ -17,31 +53,12 @@ steps:
53 - pytest -v -s test_utils.py # Utils
54 - pytest -v -s worker # Worker
55
20 -- label: Metrics, Tracing Test
21 - fast_check: true
22 - fast_check_only: true
23 - commands:
24 - - pytest -v -s metrics # Metrics
25 - - "pip install \
26 - opentelemetry-sdk \
27 - opentelemetry-api \
28 - opentelemetry-exporter-otlp \
29 - opentelemetry-semantic-conventions-ai" # Tracing
30 - - pytest -v -s tracing
31 -
32 -- label: Regression Test
33 - mirror_hardwares: [amd]
34 - fast_check: true
35 - command: pytest -v -s test_regression.py
36 - working_dir: "/vllm-workspace/tests" # optional
37 -
38 -- label: AsyncEngine Test
39 - #mirror_hardwares: [amd]
40 - command: pytest -v -s async_engine
41 -
42 -- label: Basic Correctness Test
56 +- label: Basic Correctness Test # 30min
57 mirror_hardwares: [amd]
58 fast_check: true
59 + source_file_dependencies:
60 + - vllm/
61 + - tests/basic_correctness
62 commands:
63 # This flashinfer installation will fail on AMD ROCm, so it is set as optional.
64 - pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.1.2/flashinfer-0.1.2+cu121torch2.4-cp310-cp310-linux_x86_64.whl || true
@@ -50,215 +67,264 @@ steps:
67 - VLLM_ATTENTION_BACKEND=XFORMERS pytest -v -s basic_correctness/test_chunked_prefill.py
68 - VLLM_ATTENTION_BACKEND=FLASH_ATTN pytest -v -s basic_correctness/test_chunked_prefill.py
69 - VLLM_TEST_ENABLE_ARTIFICIAL_PREEMPT=1 pytest -v -s basic_correctness/test_preemption.py
53 -
54 -- label: Core Test
70 +
71 +- label: Core Test # 10min
72 mirror_hardwares: [amd]
73 fast_check: true
74 + source_file_dependencies:
75 + - vllm/core
76 + - vllm/distributed
77 + - tests/core
78 commands:
79 - pytest -v -s core
80
60 -- label: Distributed Comm Ops Test
61 - #mirror_hardwares: [amd]
62 - working_dir: "/vllm-workspace/tests"
63 - num_gpus: 2
64 - commands:
65 - - pytest -v -s distributed/test_comm_ops.py
66 - - pytest -v -s distributed/test_shm_broadcast.py
67 -
68 -- label: 2 Node Tests (4 GPUs in total)
69 - working_dir: "/vllm-workspace/tests"
70 - num_gpus: 2
71 - num_nodes: 2
72 - commands:
73 - - # the following commands are for the first node, with ip 192.168.10.10 (ray environment already set up)
74 - - VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py
75 - - VLLM_MULTI_NODE=1 pytest -v -s distributed/test_pipeline_parallel.py
76 - - # the following commands are for the second node, with ip 192.168.10.11 (ray environment already set up)
77 - - VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py
78 -
79 -- label: Distributed Tests (2 GPUs)
81 +- label: Entrypoints Test # 20min
82 + fast_check: true
83 mirror_hardwares: [amd]
81 - working_dir: "/vllm-workspace/tests"
82 - num_gpus: 2
84 + source_file_dependencies:
85 + - vllm/entrypoints
86 + - tests/entrypoints
87 commands:
84 - - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py
85 - - TARGET_TEST_SUITE=L4 pytest -v -s distributed/test_basic_distributed_correctness.py
86 - - pytest -v -s distributed/test_chunked_prefill_distributed.py
87 - - pytest -v -s distributed/test_multimodal_broadcast.py
88 - - pytest -v -s spec_decode/e2e/test_integration_dist_tp2.py
89 - - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s test_sharded_state_loader.py
90 - - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s distributed/test_utils.py
88 + - pytest -v -s entrypoints/llm
89 + - pytest -v -s entrypoints/openai
90
92 -- label: Distributed Tests (4 GPUs)
93 - #mirror_hardwares: [amd]
91 +- label: Distributed Tests (4 GPUs) # 10min
92 working_dir: "/vllm-workspace/tests"
93 num_gpus: 4
94 fast_check: true
95 + source_file_dependencies:
96 + - vllm/
97 + - tests/distributed
98 + - tests/spec_decode/e2e/test_integration_dist_tp4
99 commands:
100 - pytest -v -s distributed/test_pynccl.py
101 - pytest -v -s spec_decode/e2e/test_integration_dist_tp4.py
102
101 -- label: Pipeline Parallelism Test
102 - working_dir: "/vllm-workspace/tests"
103 - num_gpus: 4
103 +##### fast check tests #####
104 +##### 1 GPU test #####
105 +
106 +- label: Metrics, Tracing Test # 10min
107 + source_file_dependencies:
108 + - vllm/
109 + - tests/metrics
110 + - tests/tracing
111 commands:
105 - - pytest -v -s distributed/test_pipeline_parallel.py
112 + - pytest -v -s metrics
113 + - "pip install \
114 + opentelemetry-sdk \
115 + opentelemetry-api \
116 + opentelemetry-exporter-otlp \
117 + opentelemetry-semantic-conventions-ai"
118 + - pytest -v -s tracing
119
107 -- label: Engine Test
120 +- label: Regression Test # 5min
121 mirror_hardwares: [amd]
122 + source_file_dependencies:
123 + - vllm/
124 + - tests/test_regression
125 + command: pytest -v -s test_regression.py
126 + working_dir: "/vllm-workspace/tests" # optional
127 +
128 +- label: Engine Test # 10min
129 + mirror_hardwares: [amd]
130 + source_file_dependencies:
131 + - vllm/
132 + - tests/engine
133 + - tests/tokenization
134 commands:
135 - pytest -v -s engine test_sequence.py test_config.py test_logger.py
136 # OOM in the CI unless we run this separately
137 - pytest -v -s tokenization
138
114 -- label: Entrypoints Test
115 - fast_check: true
116 - mirror_hardwares: [amd]
117 -
118 - commands:
119 - - pytest -v -s entrypoints/llm
120 - - pytest -v -s entrypoints/openai
121 -
122 -- label: Examples Test
139 +- label: Examples Test # 12min
140 working_dir: "/vllm-workspace/examples"
141 mirror_hardwares: [amd]
142 + source_file_dependencies:
143 + - vllm/entrypoints
144 + - examples/
145 commands:
126 - # install tensorizer for tensorize_vllm_model.py
127 - - pip install awscli tensorizer
146 + - pip install awscli tensorizer # for llava example and tensorizer test
147 - python3 offline_inference.py
148 - python3 cpu_offload.py
149 - python3 offline_inference_with_prefix.py
150 - python3 llm_engine_example.py
132 - - python3 offline_inference_vision_language.py
151 + - python3 llava_example.py
152 - python3 tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 tensorize_vllm_model.py --model facebook/opt-125m deserialize --path-to-tensors /tmp/vllm/facebook/opt-125m/v1/model.tensors
153
135 -- label: Inputs Test
136 - #mirror_hardwares: [amd]
137 - commands:
138 - - pytest -v -s test_inputs.py
139 - - pytest -v -s multimodal
140 -
141 -# - label: Kernels Test %N
142 -# #mirror_hardwares: [amd]
143 -# commands:
144 -# - pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.0.8/flashinfer-0.0.8+cu121torch2.3-cp310-cp310-linux_x86_64.whl
145 -# - pytest -v -s kernels --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
146 -# parallelism: 4
147 -
148 -- label: Models Test
149 - #mirror_hardwares: [amd]
154 +- label: Models Test # 1hr10min
155 + source_file_dependencies:
156 + - vllm/
157 + - tests/models
158 commands:
159 - pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.1.2/flashinfer-0.1.2+cu121torch2.4-cp310-cp310-linux_x86_64.whl
160 - pytest -v -s models -m \"not vlm\"
161
154 -- label: Vision Language Models Test
162 +- label: Vision Language Models Test # 42min
163 mirror_hardwares: [amd]
164 + source_file_dependencies:
165 + - vllm/
166 commands:
167 - pytest -v -s models -m vlm
168
159 -- label: Prefix Caching Test
169 +- label: Prefix Caching Test # 7min
170 mirror_hardwares: [amd]
171 + source_file_dependencies:
172 + - vllm/
173 + - tests/prefix_caching
174 commands:
175 - pytest -v -s prefix_caching
176
164 -- label: Samplers Test
165 - #mirror_hardwares: [amd]
177 +- label: Samplers Test # 18min
178 + source_file_dependencies:
179 + - vllm/model_executor/layers
180 + - vllm/sampling_metadata.py
181 + - tests/samplers
182 command: pytest -v -s samplers
183
168 -- label: LogitsProcessor Test
184 +- label: LogitsProcessor Test # 5min
185 mirror_hardwares: [amd]
186 + source_file_dependencies:
187 + - vllm/model_executor/layers
188 + - tests/test_logits_processor
189 command: pytest -v -s test_logits_processor.py
190
172 -- label: Utils Test
173 - commands:
174 - - pytest -v -s test_utils.py
175 - - pytest -v -s test_embedded_commit.py
176 -
177 -- label: Worker Test
178 - mirror_hardwares: [amd]
179 - command: pytest -v -s worker
180 -
181 -- label: Speculative decoding tests
182 - #mirror_hardwares: [amd]
191 +- label: Speculative decoding tests # 22min
192 + source_file_dependencies:
193 + - vllm/spec_decode
194 + - tests/spec_decode
195 commands:
196 # See https://github.com/vllm-project/vllm/issues/5152
197 - export VLLM_ATTENTION_BACKEND=XFORMERS
198 - pytest -v -s spec_decode
199
188 -# - label: LoRA Test %N
189 -# #mirror_hardwares: [amd]
190 -# command: pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_long_context.py
191 -# parallelism: 4
192 -
193 -# - label: LoRA Long Context (Distributed)
194 -# #mirror_hardwares: [amd]
195 -# num_gpus: 4
196 -# # This test runs llama 13B, so it is required to run on 4 GPUs.
197 -# commands:
198 -# # FIXIT: find out which code initialize cuda before running the test
199 -# # before the fix, we need to use spawn to test it
200 -# - export VLLM_WORKER_MULTIPROC_METHOD=spawn
201 -# - pytest -v -s -x lora/test_long_context.py
202 -
203 -- label: Tensorizer Test
204 - #mirror_hardwares: [amd]
205 - fast_check: true
200 +- label: LoRA Test %N # 30min each
201 + source_file_dependencies:
202 + - vllm/lora
203 + - csrc/punica
204 + - tests/lora
205 + command: pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_long_context.py
206 + parallelism: 4
207 +
208 +- label: Kernels Test %N # 30min each
209 + source_file_dependencies:
210 + - csrc/
211 + - vllm/attention
212 + - tests/kernels
213 + commands:
214 + - pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.1.2/flashinfer-0.1.2+cu121torch2.4-cp310-cp310-linux_x86_64.whl
215 + - pytest -v -s kernels --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
216 + parallelism: 4
217 +
218 +- label: Tensorizer Test # 11min
219 + soft_fail: true
220 + source_file_dependencies:
221 + - vllm/model_executor/model_loader
222 + - tests/tensorizer_loader
223 commands:
224 - apt-get install -y curl libsodium23
225 - export VLLM_WORKER_MULTIPROC_METHOD=spawn
226 - pytest -v -s tensorizer_loader
227
211 -- label: Metrics Test
212 - mirror_hardwares: [amd]
213 - command: pytest -v -s metrics
214 -
215 -- label: Quantization Test
216 - #mirror_hardwares: [amd]
217 - command: pytest -v -s quantization
218 -
219 -- label: Tracing Test
220 - commands:
221 - - "pip install \
222 - opentelemetry-sdk \
223 - opentelemetry-api \
224 - opentelemetry-exporter-otlp \
225 - opentelemetry-semantic-conventions-ai"
226 - - pytest -v -s tracing
227 -
228 -- label: Benchmarks
228 +- label: Benchmarks # 9min
229 working_dir: "/vllm-workspace/.buildkite"
230 mirror_hardwares: [amd]
231 + source_file_dependencies:
232 + - benchmarks/
233 commands:
234 - pip install aiohttp
235 - bash run-benchmarks.sh
236
235 -- label: LM Eval Small Models
237 +- label: Quantization Test # 15min
238 + source_file_dependencies:
239 + - csrc/
240 + - vllm/model_executor/layers/quantization
241 + - tests/quantization
242 + command: pytest -v -s quantization
243 +
244 +- label: LM Eval Small Models # 53min
245 working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
246 + source_file_dependencies:
247 + - csrc/
248 + - vllm/model_executor/layers/quantization
249 commands:
250 - pip install lm-eval
251 - export VLLM_WORKER_MULTIPROC_METHOD=spawn
252 - bash ./run-tests.sh -c configs/models-small.txt -t 1
253
242 -- label: LM Eval Large Models
243 - gpu: a100
254 +
255 +##### 1 GPU test #####
256 +##### multi gpus test #####
257 +
258 +- label: Distributed Comm Ops Test # 7min
259 + working_dir: "/vllm-workspace/tests"
260 + num_gpus: 2
261 + source_file_dependencies:
262 + - vllm/distributed
263 + - tests/distributed
264 + commands:
265 + - pytest -v -s distributed/test_comm_ops.py
266 + - pytest -v -s distributed/test_shm_broadcast.py
267 +
268 +- label: 2 Node Tests (4 GPUs in total) # 16min
269 + working_dir: "/vllm-workspace/tests"
270 + num_gpus: 2
271 + num_nodes: 2
272 + source_file_dependencies:
273 + - vllm/
274 + - tests/distributed/test_same_node
275 + commands:
276 + - # the following commands are for the first node, with ip 192.168.10.10 (ray environment already set up)
277 + - VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py
278 + - VLLM_MULTI_NODE=1 pytest -v -s distributed/test_pipeline_parallel.py
279 + - # the following commands are for the second node, with ip 192.168.10.11 (ray environment already set up)
280 + - VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py
281 +
282 +- label: Distributed Tests (2 GPUs) # 28min
283 + mirror_hardwares: [amd]
284 + working_dir: "/vllm-workspace/tests"
285 + num_gpus: 2
286 + source_file_dependencies:
287 + - vllm/
288 + - tests/distributed
289 + commands:
290 + - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py
291 + - TARGET_TEST_SUITE=L4 pytest -v -s distributed/test_basic_distributed_correctness.py
292 + - pytest -v -s distributed/test_chunked_prefill_distributed.py
293 + - pytest -v -s distributed/test_multimodal_broadcast.py
294 + - pytest -v -s spec_decode/e2e/test_integration_dist_tp2.py
295 + - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s test_sharded_state_loader.py
296 + - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s distributed/test_utils.py
297 +
298 +- label: Pipeline Parallelism Test # 23min
299 + working_dir: "/vllm-workspace/tests"
300 num_gpus: 4
245 - working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
301 + source_file_dependencies:
302 + - vllm/
303 + - tests/distributed/test_pipeline_parallel
304 commands:
247 - - pip install lm-eval
248 - - export VLLM_WORKER_MULTIPROC_METHOD=spawn
249 - - bash ./run-tests.sh -c configs/models-large.txt -t 4
305 + - pytest -v -s distributed/test_pipeline_parallel.py
306
251 -- label: Documentation Build
252 - working_dir: "/vllm-workspace/test_docs/docs"
253 - fast_check: true
254 - no_gpu: True
307 +- label: LoRA Long Context (Distributed) # 11min
308 + # This test runs llama 13B, so it is required to run on 4 GPUs.
309 + num_gpus: 4
310 + source_file_dependencies:
311 + - vllm/lora
312 + - csrc/punica
313 + - tests/lora/test_long_context
314 commands:
256 - - pip install -r requirements-docs.txt
257 - - SPHINXOPTS=\"-W\" make html
315 + # FIXIT: find out which code initialize cuda before running the test
316 + # before the fix, we need to use spawn to test it
317 + - export VLLM_WORKER_MULTIPROC_METHOD=spawn
318 + - pytest -v -s -x lora/test_long_context.py
319 +
320 +##### multi gpus test #####
321 +##### A100 test #####
322
259 -- label: Distributed Tests (A100)
323 +- label: Distributed Tests (A100) # optional
324 gpu: a100
325 num_gpus: 4
326 + source_file_dependencies:
327 + - vllm/
328 commands:
329 # NOTE: don't test llama model here, it seems hf implementation is buggy
330 # see https://github.com/vllm-project/vllm/pull/5689 for details
@@ -266,3 +332,15 @@ steps:
332 - pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.1.2/flashinfer-0.1.2+cu121torch2.4-cp310-cp310-linux_x86_64.whl
333 - TARGET_TEST_SUITE=A100 pytest -v -s distributed/test_basic_distributed_correctness.py
334 - pytest -v -s -x lora/test_mixtral.py
335 +
336 +- label: LM Eval Large Models # optional
337 + gpu: a100
338 + num_gpus: 4
339 + working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
340 + source_file_dependencies:
341 + - csrc/
342 + - vllm/model_executor/layers/quantization
343 + commands:
344 + - pip install lm-eval
345 + - export VLLM_WORKER_MULTIPROC_METHOD=spawn
346 + - bash ./run-tests.sh -c configs/models-large.txt -t 4