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 6, 13:15 UTC
[CI] Split Distributed Tests (4 GPUs) and Kernel MoE tests (#37100)
Signed-off-by: Avinash Singh <avinashsingh.rcoem@gmail.com> Signed-off-by: Avinash Singh <107198269+avinashsingh77@users.noreply.github.com> Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> Co-authored-by: Kevin H. Luu <khluu000@gmail.com>
Avinash Singh committed
Mar 18, 2026 at 00:14 UTC
c5030c439db3944f2cdbdfbc1283b431e863f73f
2 files changed
+41
-11
.buildkite/test_areas/distributed.yaml
+39
-9
@@ -15,26 +15,19 @@ steps:
15
- pytest -v -s distributed/test_shm_buffer.py
16
- pytest -v -s distributed/test_shm_storage.py
17
18
-- label: Distributed (2 GPUs)
19
- timeout_in_minutes: 60
18
+- label: Distributed DP Tests (2 GPUs)
19
+ timeout_in_minutes: 20
20
working_dir: "/vllm-workspace/tests"
21
num_devices: 2
22
source_file_dependencies:
23
- - vllm/compilation/
23
- vllm/distributed/
24
- vllm/engine/
25
- vllm/executor/
26
- vllm/worker/worker_base.py
27
- vllm/v1/engine/
28
- vllm/v1/worker/
30
- - tests/compile/fullgraph/test_basic_correctness.py
31
- - tests/compile/test_wrapper.py
32
- - tests/distributed/
33
- - tests/entrypoints/llm/test_collective_rpc.py
29
- tests/v1/distributed
30
- tests/v1/entrypoints/openai/test_multi_api_servers.py
36
- - tests/v1/shutdown
37
- - tests/v1/worker/test_worker_memory_snapshot.py
31
commands:
32
# https://github.com/NVIDIA/nccl/issues/1838
33
- export NCCL_CUMEM_HOST_ENABLE=0
@@ -42,9 +35,46 @@ steps:
35
- TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_eagle_dp.py
36
- TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_external_lb_dp.py
37
- DP_SIZE=2 pytest -v -s v1/entrypoints/openai/test_multi_api_servers.py
38
+
39
+- label: Distributed Compile + RPC Tests (2 GPUs)
40
+ timeout_in_minutes: 20
41
+ working_dir: "/vllm-workspace/tests"
42
+ num_devices: 2
43
+ source_file_dependencies:
44
+ - vllm/compilation/
45
+ - vllm/distributed/
46
+ - vllm/engine/
47
+ - vllm/executor/
48
+ - vllm/worker/worker_base.py
49
+ - vllm/v1/engine/
50
+ - vllm/v1/worker/
51
+ - tests/compile/fullgraph/test_basic_correctness.py
52
+ - tests/compile/test_wrapper.py
53
+ - tests/entrypoints/llm/test_collective_rpc.py
54
+ commands:
55
+ # https://github.com/NVIDIA/nccl/issues/1838
56
+ - export NCCL_CUMEM_HOST_ENABLE=0
57
- pytest -v -s entrypoints/llm/test_collective_rpc.py
58
- pytest -v -s ./compile/fullgraph/test_basic_correctness.py
59
- pytest -v -s ./compile/test_wrapper.py
60
+
61
+- label: Distributed Torchrun + Shutdown Tests (2 GPUs)
62
+ timeout_in_minutes: 20
63
+ working_dir: "/vllm-workspace/tests"
64
+ num_devices: 2
65
+ source_file_dependencies:
66
+ - vllm/distributed/
67
+ - vllm/engine/
68
+ - vllm/executor/
69
+ - vllm/worker/worker_base.py
70
+ - vllm/v1/engine/
71
+ - vllm/v1/worker/
72
+ - tests/distributed/
73
+ - tests/v1/shutdown
74
+ - tests/v1/worker/test_worker_memory_snapshot.py
75
+ commands:
76
+ # https://github.com/NVIDIA/nccl/issues/1838
77
+ - export NCCL_CUMEM_HOST_ENABLE=0
78
- VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
79
- VLLM_TEST_SAME_HOST=1 VLLM_TEST_WITH_DEFAULT_DEVICE_SET=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
80
- CUDA_VISIBLE_DEVICES=0,1 pytest -v -s v1/shutdown
.buildkite/test_areas/kernels.yaml
+2
-2
@@ -35,7 +35,7 @@ steps:
35
parallelism: 2
36
37
- label: Kernels MoE Test %N
38
- timeout_in_minutes: 60
38
+ timeout_in_minutes: 25
39
source_file_dependencies:
40
- csrc/quantization/cutlass_w8a8/moe/
41
- csrc/moe/
@@ -47,7 +47,7 @@ steps:
47
commands:
48
- pytest -v -s kernels/moe --ignore=kernels/moe/test_modular_oai_triton_moe.py --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
49
- pytest -v -s kernels/moe/test_modular_oai_triton_moe.py --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
50
- parallelism: 2
50
+ parallelism: 5
51
52
- label: Kernels Mamba Test
53
timeout_in_minutes: 45