[CI] Move smaller H200 workloads to 18GB MIG slices (#56695)
Signed-off-by: Kevin Luu <51931015+khluu@users.noreply.github.com> Co-authored-by: OpenAI Codex <codex@openai.com>
Kevin H. Luu committed
Sep 13, 2026 at 21:30 UTC
214248c7d896a6d9430ffe6870ee78c6f0ffb556
4 files changed
+40
-11
.buildkite/test_areas/cuda.yaml
+2
-2
@@ -30,8 +30,8 @@ steps:
30
- vllm/
31
- tests/cuda
32
33
-- label: ":nvidia: (H200 MIG 35GB) CUDAGraph"
34
- device: h200_35gb
33
+- label: ":nvidia: (H200 MIG 18GB) CUDAGraph"
34
+ device: h200_18gb
35
key: cudagraph
36
timeout_in_minutes: 30
37
source_file_dependencies:
.buildkite/test_areas/engine.yaml
+17
-3
@@ -74,18 +74,20 @@ steps:
74
depends_on:
75
- image-build-amd
76
77
-- label: ":nvidia: (H200 MIG 35GB) E2E Core"
78
- device: h200_35gb
77
+- label: ":nvidia: (H200 MIG 18GB) E2E Core"
78
+ device: h200_18gb
79
key: e2e-core-1-gpu
80
timeout_in_minutes: 50
81
source_file_dependencies:
82
- vllm/v1/
83
- tests/v1/e2e/general/
84
commands:
85
- - pytest -v -s v1/e2e/general --ignore v1/e2e/general/test_async_scheduling.py
85
+ - pytest -v -s v1/e2e/general --ignore v1/e2e/general/test_async_scheduling.py --ignore v1/e2e/general/test_kv_sharing_fast_prefill.py -k "not test_mamba_prefix_cache_mrv2"
86
mirror:
87
amd:
88
label: ":amd: (MI250) E2E Core"
89
+ commands:
90
+ - pytest -v -s v1/e2e/general --ignore v1/e2e/general/test_async_scheduling.py
91
device: mi250_1
92
timeout_in_minutes: 50
93
depends_on:
@@ -95,6 +97,18 @@ steps:
97
- tests/v1/e2e/general/
98
- vllm/platforms/rocm.py
99
100
+# Keep the large Gemma-3n and Mamba MRV2 warmup allocations on 35GB slices.
101
+- label: ":nvidia: (H200 MIG 35GB) E2E Core Large Memory"
102
+ device: h200_35gb
103
+ key: e2e-core-large-memory
104
+ timeout_in_minutes: 50
105
+ source_file_dependencies:
106
+ - vllm/v1/
107
+ - tests/v1/e2e/general/
108
+ commands:
109
+ - pytest -v -s v1/e2e/general/test_kv_sharing_fast_prefill.py
110
+ - pytest -v -s v1/e2e/general/test_mamba_prefix_cache.py::test_mamba_prefix_cache_mrv2 v1/e2e/general/test_mamba_prefix_cache.py::test_mamba_prefix_cache_mrv2_async
111
+
112
- label: ":nvidia: (L4) V1 E2E"
113
key: v1-e2e-2-gpus
114
timeout_in_minutes: 25 # TODO: Fix timeout after we have more confidence in the test stability
.buildkite/test_areas/misc.yaml
+2
-2
@@ -103,8 +103,8 @@ steps:
103
depends_on:
104
- image-build-amd
105
106
-- label: ":nvidia: (H200 MIG 35GB) V1 Executor + Worker"
107
- device: h200_35gb
106
+- label: ":nvidia: (H200 MIG 18GB) V1 Executor + Worker"
107
+ device: h200_18gb
108
key: v1-executor-worker
109
timeout_in_minutes: 45
110
source_file_dependencies: *v1-core-kv-metrics-deps
.buildkite/test_areas/samplers.yaml
+19
-4
@@ -2,8 +2,8 @@ group: Samplers
2
depends_on:
3
- image-build
4
steps:
5
-- label: ":nvidia: (H200 MIG 35GB) Samplers"
6
- device: h200_35gb
5
+- label: ":nvidia: (H200 MIG 18GB) Samplers"
6
+ device: h200_18gb
7
key: samplers-test
8
timeout_in_minutes: 60
9
source_file_dependencies:
@@ -15,8 +15,8 @@ steps:
15
commands:
16
# VLLM_USE_FLASHINFER_SAMPLER defaults to 1 now, so we need to pin both
17
# values explicitly to still cover the PyTorch-native (Triton) path.
18
- - VLLM_USE_FLASHINFER_SAMPLER=0 pytest -v -s samplers
19
- - VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers
18
+ - VLLM_USE_FLASHINFER_SAMPLER=0 pytest -v -s samplers -k "not test_beam_search_passes_multimodal_data"
19
+ - VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers -k "not test_beam_search_passes_multimodal_data"
20
mirror:
21
amd:
22
label: ":amd: (MI250) Samplers"
@@ -35,3 +35,18 @@ steps:
35
- vllm/platforms/rocm.py
36
commands:
37
- pytest -v -s samplers
38
+
39
+# Qwen2-Audio-7B leaves no room for KV cache on a 16 GiB slice.
40
+- label: ":nvidia: (H200 MIG 35GB) Samplers Multimodal Beam Search"
41
+ device: h200_35gb
42
+ key: samplers-multimodal-beam-search
43
+ timeout_in_minutes: 30
44
+ source_file_dependencies:
45
+ - vllm/model_executor/layers
46
+ - vllm/sampling_metadata.py
47
+ - tests/samplers
48
+ - tests/conftest.py
49
+ - vllm/entrypoints/generate/beam_search
50
+ commands:
51
+ - VLLM_USE_FLASHINFER_SAMPLER=0 pytest -v -s samplers/test_beam_search.py::test_beam_search_passes_multimodal_data
52
+ - VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers/test_beam_search.py::test_beam_search_passes_multimodal_data