[CI] Remove MRV2-specific tests (#54823)
Signed-off-by: Nick Hill <nickhill123@gmail.com>
Nick Hill committed
Sep 1, 2026 at 14:15 UTC
a56e74afd6238914e263a61bcb33f89fd179b842
3 files changed
+11
-172
.buildkite/test_areas/distributed.yaml
+9
-1
@@ -36,7 +36,7 @@ steps:
36
37
- label: ":nvidia: (L4) Distributed DP Basic"
38
key: distributed-dp-tests-2-gpus
39
- timeout_in_minutes: 35
39
+ timeout_in_minutes: 45
40
working_dir: "/vllm-workspace/tests"
41
device: l4
42
num_devices: 2
@@ -50,6 +50,9 @@ steps:
50
- vllm/v1/worker/
51
- tests/v1/distributed
52
- tests/entrypoints/launchers/api_server/test_multi_api_servers.py
53
+ - tests/v1/e2e/general/test_sharded_sampling.py
54
+ - tests/v1/e2e/spec_decode/test_sharded_sampling.py
55
+ - tests/v1/e2e/spec_decode/utils.py
56
commands:
57
# https://github.com/NVIDIA/nccl/issues/1838
58
- export NCCL_CUMEM_HOST_ENABLE=0
@@ -57,6 +60,9 @@ steps:
60
- TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_eagle_dp.py
61
- TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_external_lb_dp.py
62
- DP_SIZE=2 pytest -v -s entrypoints/launchers/api_server/test_multi_api_servers.py
63
+ # Batch-sharded sampling requires TP=2
64
+ - pytest -v -s v1/e2e/general/test_sharded_sampling.py
65
+ - pytest -v -s v1/e2e/spec_decode/test_sharded_sampling.py
66
mirror:
67
amd:
68
label: ":amd: (MI300) Distributed DP Basic"
@@ -324,9 +330,11 @@ steps:
330
- vllm/executor/
331
- vllm/model_executor/models/
332
- tests/distributed/
333
+ - tests/v1/distributed/test_pp_dp_v2.py
334
commands:
335
- pytest -v -s distributed/test_pp_cudagraph.py
336
- pytest -v -s distributed/test_pipeline_parallel.py
337
+ - pytest -v -s v1/distributed/test_pp_dp_v2.py
338
339
- label: ":nvidia: (L4) RayExecutorV2"
340
key: rayexecutorv2-4-gpus
.buildkite/test_areas/model_runner_v2.yaml
deleted
-169
@@ -1,169 +0,0 @@
1
-group: Model Runner V2
2
-depends_on:
3
- - image-build
4
-steps:
5
-- label: ":nvidia: (H200 MIG 35GB) Model Runner V2 Core"
6
- device: h200_35gb
7
- key: model-runner-v2-core-tests
8
- timeout_in_minutes: 35
9
- source_file_dependencies:
10
- - vllm/v1/worker/gpu/
11
- - vllm/v1/worker/gpu_worker.py
12
- - vllm/v1/core/sched/
13
- - vllm/v1/attention/
14
- - tests/v1/engine/test_llm_engine.py
15
- - tests/v1/e2e/
16
- - tests/entrypoints/llm/test_struct_output_generate.py
17
- commands:
18
- - set -x
19
- - export VLLM_USE_V2_MODEL_RUNNER=1
20
- - pytest -v -s v1/engine/test_llm_engine.py -k "not test_engine_metrics"
21
- - pytest -v -s v1/e2e/general/test_async_scheduling.py -k "not ngram"
22
- - pytest -v -s v1/e2e/general/test_context_length.py
23
- - pytest -v -s v1/e2e/general/test_min_tokens.py
24
- # Temporary hack filter to exclude ngram spec decoding based tests.
25
- - pytest -v -s entrypoints/llm/test_struct_output_generate.py -k "xgrammar and not speculative_config6 and not speculative_config7 and not speculative_config8 and not speculative_config0"
26
-
27
-- label: ":nvidia: (H200 MIG 35GB) Model Runner V2 Examples"
28
- device: h200_35gb
29
- key: model-runner-v2-examples
30
- timeout_in_minutes: 35
31
- working_dir: "/vllm-workspace/examples"
32
- source_file_dependencies:
33
- - vllm/v1/worker/gpu/
34
- - vllm/v1/core/sched/
35
- - vllm/v1/worker/gpu_worker.py
36
- - examples/basic/offline_inference/
37
- - examples/generate/multimodal/
38
- - examples/features/
39
- - examples/pooling/embed/vision_embedding_offline.py
40
- - examples/features/tensorize_vllm_model.py
41
- commands:
42
- - set -x
43
- - export VLLM_USE_V2_MODEL_RUNNER=1
44
- - pip install --no-deps tensorizer # for tensorizer test
45
- - python3 basic/offline_inference/chat.py # for basic
46
- - python3 basic/offline_inference/generate.py --model facebook/opt-125m
47
- #- python3 basic/offline_inference/generate.py --model meta-llama/Llama-2-13b-chat-hf --cpu-offload-gb 10 # TODO
48
- #- python3 basic/offline_inference/embed.py # TODO
49
- # for multi-modal models
50
- - python3 generate/multimodal/audio_language_offline.py --seed 0
51
- - python3 generate/multimodal/vision_language_offline.py --seed 0
52
- - python3 generate/multimodal/vision_language_multi_image_offline.py --seed 0
53
- - python3 generate/multimodal/encoder_decoder_multimodal_offline.py --model-type whisper --seed 0
54
- # for pooling models
55
- - python3 pooling/embed/vision_embedding_offline.py --seed 0
56
- # for features demo
57
- - python3 features/automatic_prefix_caching/prefix_caching_offline.py
58
- - python3 deployment/llm_engine_example.py
59
- - python3 features/tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 features/tensorize_vllm_model.py --model facebook/opt-125m deserialize --path-to-tensors /tmp/vllm/facebook/opt-125m/v1/model.tensors
60
- - python3 features/speculative_decoding/spec_decode_offline.py --test --method eagle --num_spec_tokens 3 --dataset-name hf --dataset-path philschmid/mt-bench --num-prompts 80 --temp 0 --top-p 1.0 --top-k -1 --tp 1 --enable-chunked-prefill --max-model-len 2048
61
- # https://github.com/vllm-project/vllm/pull/26682 uses slightly more memory in PyTorch 2.9+ causing this test to OOM in 1xL4 GPU
62
- - python3 features/speculative_decoding/spec_decode_offline.py --test --method eagle3 --num_spec_tokens 3 --dataset-name hf --dataset-path philschmid/mt-bench --num-prompts 80 --temp 0 --top-p 1.0 --top-k -1 --tp 1 --enable-chunked-prefill --max-model-len 1536
63
-
64
-- label: ":nvidia: (L4) Model Runner V2 Distributed"
65
- key: model-runner-v2-distributed-2-gpus
66
- timeout_in_minutes: 40
67
- working_dir: "/vllm-workspace/tests"
68
- device: l4
69
- num_devices: 2
70
- source_file_dependencies:
71
- - vllm/v1/worker/gpu/
72
- - vllm/v1/worker/gpu_worker.py
73
- - tests/basic_correctness/test_basic_correctness.py
74
- - tests/v1/distributed/test_async_llm_dp.py
75
- - tests/v1/distributed/test_eagle_dp.py
76
- - tests/v1/e2e/general/test_sharded_sampling.py
77
- - tests/v1/e2e/spec_decode/test_sharded_sampling.py
78
- commands:
79
- - set -x
80
- - export VLLM_USE_V2_MODEL_RUNNER=1
81
- # The "and not True" here is a hacky way to exclude the prompt_embeds cases which aren't yet supported.
82
- - TARGET_TEST_SUITE=L4 pytest -v -s basic_correctness/test_basic_correctness.py -m 'distributed(num_gpus=2)' -k "not ray and not True"
83
- # https://github.com/NVIDIA/nccl/issues/1838
84
- - export NCCL_CUMEM_HOST_ENABLE=0
85
- - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_async_llm_dp.py -k "not ray"
86
- - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_eagle_dp.py
87
- - pytest -v -s v1/e2e/general/test_sharded_sampling.py
88
- - pytest -v -s v1/e2e/spec_decode/test_sharded_sampling.py
89
-
90
- mirror:
91
- amd:
92
- label: ":amd: (MI300) Model Runner V2 Distributed"
93
- dind: false
94
- device: mi300_2
95
- timeout_in_minutes: 45
96
- depends_on:
97
- - image-build-amd
98
- source_file_dependencies:
99
- - vllm/v1/worker/gpu/
100
- - vllm/v1/worker/gpu_worker.py
101
- - tests/basic_correctness/test_basic_correctness.py
102
- - tests/v1/distributed/test_async_llm_dp.py
103
- - tests/v1/distributed/test_eagle_dp.py
104
- - vllm/platforms/rocm.py
105
- commands:
106
- - set -x
107
- - export VLLM_USE_V2_MODEL_RUNNER=1
108
- - TARGET_TEST_SUITE=MI300 pytest -v -s basic_correctness/test_basic_correctness.py -m 'distributed(num_gpus=2)' -k "not ray and not True"
109
- - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_async_llm_dp.py -k "not ray"
110
- - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_eagle_dp.py
111
-
112
-- label: ":nvidia: (L4) Model Runner V2 Pipeline Parallelism"
113
- key: model-runner-v2-pipeline-parallelism-4-gpus
114
- timeout_in_minutes: 50
115
- working_dir: "/vllm-workspace/tests"
116
- device: l4
117
- num_devices: 4
118
- source_file_dependencies:
119
- - vllm/v1/worker/gpu/
120
- - vllm/v1/worker/gpu_worker.py
121
- - tests/distributed/test_pipeline_parallel.py
122
- - tests/distributed/test_pp_cudagraph.py
123
- - tests/v1/distributed/test_pp_dp_v2.py
124
- commands:
125
- - set -x
126
- - export VLLM_USE_V2_MODEL_RUNNER=1
127
- - pytest -v -s distributed/test_pipeline_parallel.py -k "not ray"
128
- - pytest -v -s distributed/test_pp_cudagraph.py -k "not ray"
129
- - pytest -v -s v1/distributed/test_pp_dp_v2.py
130
-
131
-- label: ":nvidia: (H200 MIG 35GB) Model Runner V2 Spec Decode"
132
- device: h200_35gb
133
- key: model-runner-v2-spec-decode
134
- timeout_in_minutes: 50
135
- working_dir: "/vllm-workspace/tests"
136
- source_file_dependencies:
137
- - vllm/v1/worker/gpu/
138
- - vllm/v1/worker/gpu_worker.py
139
- - tests/v1/spec_decode/test_max_len.py
140
- - tests/v1/spec_decode/test_rejection_sampler_utils.py
141
- - tests/v1/spec_decode/test_synthetic_rejection_sampler_utils.py
142
- - tests/v1/e2e/spec_decode/
143
- commands:
144
- - set -x
145
- - export VLLM_USE_V2_MODEL_RUNNER=1
146
- - pytest -v -s v1/spec_decode/test_max_len.py -k "eagle or mtp"
147
- - pytest -v -s v1/spec_decode/test_rejection_sampler_utils.py
148
- - pytest -v -s v1/spec_decode/test_synthetic_rejection_sampler_utils.py
149
- - pytest -v -s v1/e2e/spec_decode/eagle/
150
- - pytest -v -s v1/e2e/spec_decode/speculators/
151
- - >-
152
- pytest -v -s
153
- 'v1/e2e/spec_decode/mtp/test_mtp.py::test_mtp_correctness[qwen3_5-hybrid]'
154
- mirror:
155
- amd:
156
- label: ":amd: (MI355) Model Runner V2 Spec Decode"
157
- dind: false
158
- device: mi355_1
159
- timeout_in_minutes: 75
160
- depends_on:
161
- - image-build-amd
162
- source_file_dependencies:
163
- - vllm/platforms/rocm.py
164
- - vllm/v1/worker/gpu/
165
- - vllm/v1/worker/gpu_worker.py
166
- - tests/v1/spec_decode/test_max_len.py
167
- - tests/v1/spec_decode/test_rejection_sampler_utils.py
168
- - tests/v1/spec_decode/test_synthetic_rejection_sampler_utils.py
169
- - tests/v1/e2e/spec_decode/
tests/v1/distributed/test_pp_dp_v2.py
+2
-2
@@ -37,8 +37,8 @@ _GPU_SKIP = _gpu_skip_reason()
37
38
pytestmark = [
39
pytest.mark.skipif(
40
- os.environ.get("VLLM_USE_V2_MODEL_RUNNER", "0") != "1",
41
- reason="VLLM_USE_V2_MODEL_RUNNER=1 required",
40
+ os.environ.get("VLLM_USE_V2_MODEL_RUNNER", "1") == "0",
41
+ reason="V2 model runner disabled via VLLM_USE_V2_MODEL_RUNNER=0",
42
),
43
pytest.mark.skipif(_GPU_SKIP is not None, reason=_GPU_SKIP or ""),
44
]