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 14, 05:30 UTC
[CI/Build] Split up Distributed Tests (#25572)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Cyrus Leung committed
Sep 26, 2025 at 20:49 UTC
bc9d7b5595887d4a1358926579b638b42368efd7
2 files changed
+28
-18
.buildkite/test-pipeline.yaml
+27
-16
@@ -870,25 +870,27 @@ steps:
870
- NUM_NODES=2 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_node_count.py | grep 'Node count test passed'
871
- python3 ../examples/offline_inference/data_parallel.py --dp-size=2 --tp-size=1 --node-size=2 --node-rank=1 --master-addr=192.168.10.10 --master-port=12345 --enforce-eager --trust-remote-code
872
873
-- label: Distributed Tests (2 GPUs) # 110min
874
- timeout_in_minutes: 150
873
+- label: Distributed Tests (2 GPUs) # 68min
874
+ timeout_in_minutes: 90
875
mirror_hardwares: [amdexperimental]
876
working_dir: "/vllm-workspace/tests"
877
num_gpus: 2
878
source_file_dependencies:
879
+ - vllm/compilation/
880
- vllm/distributed/
881
- vllm/engine/
882
- vllm/executor/
882
- - vllm/model_executor/models/
883
- - tests/distributed/
884
- - vllm/compilation
883
- vllm/worker/worker_base.py
886
- - entrypoints/llm/test_collective_rpc.py
884
+ - vllm/v1/engine/
885
+ - vllm/v1/worker/
886
+ - tests/compile/test_basic_correctness.py
887
+ - tests/compile/test_wrapper.py
888
+ - tests/distributed/
889
+ - tests/entrypoints/llm/test_collective_rpc.py
890
- tests/v1/test_async_llm_dp.py
891
- tests/v1/test_external_lb_dp.py
892
- tests/v1/entrypoints/openai/test_multi_api_servers.py
890
- - vllm/v1/engine/
891
- - vllm/v1/worker/
893
+ - tests/v1/shutdown
894
- tests/v1/worker/test_worker_memory_snapshot.py
895
commands:
896
- TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/test_async_llm_dp.py
@@ -898,20 +900,29 @@ steps:
900
- pytest -v -s ./compile/test_basic_correctness.py
901
- pytest -v -s ./compile/test_wrapper.py
902
- VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
903
+ - pytest -v -s distributed/test_sequence_parallel.py
904
+ - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s v1/shutdown
905
+ - pytest -v -s v1/worker/test_worker_memory_snapshot.py
906
+
907
+- label: Distributed Model Tests (2 GPUs) # 37min
908
+ timeout_in_minutes: 50
909
+ mirror_hardwares: [amdexperimental]
910
+ working_dir: "/vllm-workspace/tests"
911
+ num_gpus: 2
912
+ source_file_dependencies:
913
+ - vllm/model_executor/model_loader/sharded_state_loader.py
914
+ - vllm/model_executor/models/
915
+ - tests/basic_correctness/
916
+ - tests/model_executor/model_loader/test_sharded_state_loader.py
917
+ - tests/models/
918
+ commands:
919
- TARGET_TEST_SUITE=L4 pytest basic_correctness/ -v -s -m 'distributed(num_gpus=2)'
920
+ - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s model_executor/model_loader/test_sharded_state_loader.py
921
# Avoid importing model tests that cause CUDA reinitialization error
922
- pytest models/test_transformers.py -v -s -m 'distributed(num_gpus=2)'
923
- pytest models/language -v -s -m 'distributed(num_gpus=2)'
924
- pytest models/multimodal -v -s -m 'distributed(num_gpus=2)' --ignore models/multimodal/generation/test_whisper.py
925
- VLLM_WORKER_MULTIPROC_METHOD=spawn pytest models/multimodal/generation/test_whisper.py -v -s -m 'distributed(num_gpus=2)'
907
- # test sequence parallel
908
- - pytest -v -s distributed/test_sequence_parallel.py
909
- # this test fails consistently.
910
- # TODO: investigate and fix
911
- - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s test_sharded_state_loader.py
912
- - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s v1/shutdown
913
- - pytest -v -s models/multimodal/generation/test_maverick.py
914
- - pytest -v -s v1/worker/test_worker_memory_snapshot.py
926
927
- label: Plugin Tests (2 GPUs) # 40min
928
timeout_in_minutes: 60
tests/model_executor/model_loader/test_sharded_state_loader.py
renamed
+1
-2
@@ -91,8 +91,7 @@ def _run_generate(input_dir, queue: mp.Queue, **kwargs):
91
@pytest.mark.parametrize("enable_lora", [False, True])
92
@pytest.mark.parametrize("tp_size", [1, 2])
93
def test_sharded_state_loader(enable_lora, tp_size, num_gpus_available,
94
- llama_3p2_1b_files,
95
- monkeypatch: pytest.MonkeyPatch):
94
+ llama_3p2_1b_files):
95
if num_gpus_available < tp_size:
96
pytest.skip(f"Not enough GPUs for tensor parallelism {tp_size}")
97