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 5, 12:15 UTC
[ci] Switch some CI jobs to H200 MIG slices (#38956)
Kevin H. Luu committed
Apr 5, 2026 at 13:26 UTC
56de443db1d50dce2624a21765c11a79dffa8740
14 files changed
+25
.buildkite/test_areas/basic_correctness.yaml
+1
@@ -4,6 +4,7 @@ depends_on:
4
steps:
5
- label: Basic Correctness
6
timeout_in_minutes: 30
7
+ device: h200_18gb
8
source_file_dependencies:
9
- vllm/
10
- tests/basic_correctness/test_basic_correctness
.buildkite/test_areas/benchmarks.yaml
+1
@@ -4,6 +4,7 @@ depends_on:
4
steps:
5
- label: Benchmarks CLI Test
6
timeout_in_minutes: 20
7
+ device: h200_18gb
8
source_file_dependencies:
9
- vllm/
10
- tests/benchmarks/
.buildkite/test_areas/cuda.yaml
+1
@@ -4,6 +4,7 @@ depends_on:
4
steps:
5
- label: Platform Tests (CUDA)
6
timeout_in_minutes: 15
7
+ device: h200_18gb
8
source_file_dependencies:
9
- vllm/
10
- tests/cuda
.buildkite/test_areas/engine.yaml
+2
@@ -4,6 +4,7 @@ depends_on:
4
steps:
5
- label: Engine
6
timeout_in_minutes: 15
7
+ device: h200_18gb
8
source_file_dependencies:
9
- vllm/
10
- tests/engine
@@ -25,6 +26,7 @@ steps:
26
27
- label: e2e Scheduling (1 GPU)
28
timeout_in_minutes: 30
29
+ device: h200_18gb
30
source_file_dependencies:
31
- vllm/v1/
32
- tests/v1/e2e/general/
.buildkite/test_areas/entrypoints.yaml
+2
@@ -61,6 +61,7 @@ steps:
61
62
- label: Entrypoints Integration (API Server openai - Part 3)
63
timeout_in_minutes: 50
64
+ device: h200_18gb
65
working_dir: "/vllm-workspace/tests"
66
source_file_dependencies:
67
- vllm/
@@ -105,6 +106,7 @@ steps:
106
107
- label: OpenAI API Correctness
108
timeout_in_minutes: 30
109
+ device: h200_18gb
110
source_file_dependencies:
111
- csrc/
112
- vllm/entrypoints/openai/
.buildkite/test_areas/expert_parallelism.yaml
+1
@@ -4,6 +4,7 @@ depends_on:
4
steps:
5
- label: EPLB Algorithm
6
timeout_in_minutes: 15
7
+ device: h200_18gb
8
working_dir: "/vllm-workspace/tests"
9
source_file_dependencies:
10
- vllm/distributed/eplb
.buildkite/test_areas/kernels.yaml
+1
@@ -4,6 +4,7 @@ depends_on:
4
steps:
5
- label: vLLM IR Tests
6
timeout_in_minutes: 10
7
+ device: h200_18gb
8
working_dir: "/vllm-workspace/"
9
source_file_dependencies:
10
- vllm/ir
.buildkite/test_areas/misc.yaml
+2
@@ -19,6 +19,7 @@ steps:
19
20
- label: V1 Sample + Logits
21
timeout_in_minutes: 30
22
+ device: h200_18gb
23
source_file_dependencies:
24
- vllm/
25
- tests/v1/sample
@@ -86,6 +87,7 @@ steps:
87
88
- label: Regression
89
timeout_in_minutes: 20
90
+ device: h200_18gb
91
source_file_dependencies:
92
- vllm/
93
- tests/test_regression
.buildkite/test_areas/models_basic.yaml
+1
@@ -4,6 +4,7 @@ depends_on:
4
steps:
5
- label: Basic Models Tests (Initialization)
6
timeout_in_minutes: 45
7
+ device: h200_18gb
8
torch_nightly: true
9
source_file_dependencies:
10
- vllm/
.buildkite/test_areas/models_language.yaml
+2
@@ -67,6 +67,7 @@ steps:
67
68
- label: Language Models Test (PPL)
69
timeout_in_minutes: 110
70
+ device: h200_18gb
71
optional: true
72
source_file_dependencies:
73
- vllm/
@@ -90,6 +91,7 @@ steps:
91
92
- label: Language Models Test (MTEB)
93
timeout_in_minutes: 110
94
+ device: h200_18gb
95
optional: true
96
source_file_dependencies:
97
- vllm/
.buildkite/test_areas/models_multimodal.yaml
+4
@@ -4,6 +4,7 @@ depends_on:
4
steps:
5
- label: "Multi-Modal Models (Standard) 1: qwen2"
6
timeout_in_minutes: 45
7
+ device: h200_18gb
8
source_file_dependencies:
9
- vllm/
10
- tests/models/multimodal
@@ -19,6 +20,7 @@ steps:
20
21
- label: "Multi-Modal Models (Standard) 2: qwen3 + gemma"
22
timeout_in_minutes: 45
23
+ device: h200_18gb
24
source_file_dependencies:
25
- vllm/
26
- tests/models/multimodal
@@ -77,6 +79,7 @@ steps:
79
80
- label: Multi-Modal Processor # 44min
81
timeout_in_minutes: 60
82
+ device: h200_18gb
83
source_file_dependencies:
84
- vllm/
85
- tests/models/multimodal
@@ -131,6 +134,7 @@ steps:
134
135
- label: Multi-Modal Models (Extended Pooling)
136
optional: true
137
+ device: h200_18gb
138
source_file_dependencies:
139
- vllm/
140
- tests/models/multimodal/pooling
.buildkite/test_areas/pytorch.yaml
+2
@@ -49,6 +49,7 @@ steps:
49
50
- label: PyTorch Fullgraph
51
timeout_in_minutes: 30
52
+ device: h200_18gb
53
source_file_dependencies:
54
- vllm/
55
- tests/compile
@@ -60,6 +61,7 @@ steps:
61
# if this test fails, it means the nightly torch version is not compatible with some
62
# of the dependencies. Please check the error message and add the package to whitelist
63
# in /vllm/tools/pre_commit/generate_nightly_torch_test.py
64
+ device: h200_18gb
65
soft_fail: true
66
source_file_dependencies:
67
- requirements/nightly_torch_test.txt
.buildkite/test_areas/ray_compat.yaml
+1
@@ -7,6 +7,7 @@ steps:
7
# If this fails, it means the PR introduces a dependency that
8
# conflicts with Ray's dependency constraints.
9
# See https://github.com/vllm-project/vllm/issues/33599
10
+ device: h200_18gb
11
soft_fail: true
12
timeout_in_minutes: 10
13
source_file_dependencies:
.buildkite/test_areas/spec_decode.yaml
+4
@@ -4,6 +4,7 @@ depends_on:
4
steps:
5
- label: Spec Decode Eagle
6
timeout_in_minutes: 30
7
+ device: h200_18gb
8
source_file_dependencies:
9
- vllm/v1/spec_decode/
10
- vllm/v1/worker/gpu/spec_decode/
@@ -13,6 +14,7 @@ steps:
14
15
- label: Spec Decode Speculators + MTP
16
timeout_in_minutes: 30
17
+ device: h200_18gb
18
source_file_dependencies:
19
- vllm/v1/spec_decode/
20
- vllm/v1/worker/gpu/spec_decode/
@@ -23,6 +25,7 @@ steps:
25
26
- label: Spec Decode Ngram + Suffix
27
timeout_in_minutes: 30
28
+ device: h200_18gb
29
source_file_dependencies:
30
- vllm/v1/spec_decode/
31
- vllm/v1/worker/gpu/spec_decode/
@@ -32,6 +35,7 @@ steps:
35
36
- label: Spec Decode Draft Model
37
timeout_in_minutes: 30
38
+ device: h200_18gb
39
source_file_dependencies:
40
- vllm/v1/spec_decode/
41
- vllm/v1/worker/gpu/spec_decode/