@tduncan / Nvidia-Optimization-Stack-VLLM
Code Commits Issues Pulls Cloud Agent CI/CD
main
main
yaml 243 lines 10.4 KB
Raw
1 group: Models - Language
2 depends_on:
3 - image-build
4 steps:
5 - label: ":nvidia: (H200 MIG 18GB) Language Models (Standard)"
6 key: language-models-tests-standard
7 timeout_in_minutes: 30
8 device: h200_18gb
9 source_file_dependencies:
10 - vllm/
11 - "!vllm/distributed/kv_transfer/"
12 - tests/models/language
13 commands:
14 # Test standard language models, excluding a subset of slow tests
15 - pip freeze | grep -E 'torch'
16 - pytest -v -s models/language -m 'core_model and (not slow_test)'
17 mirror:
18 amd:
19 label: ":amd: (MI355 DPX) Language Models (Standard)"
20 dind: false
21 device: mi355_dpx
22 timeout_in_minutes: 60
23 depends_on:
24 - image-build-amd
25
26 - label: ":nvidia: (H200 MIG 35GB) Language Models (Extra Standard) Shard %N"
27 device: h200_35gb
28 key: language-models-tests-extra-standard
29 timeout_in_minutes: 40
30 source_file_dependencies:
31 - vllm/model_executor/models/
32 - tests/models/language/pooling/test_embedding.py
33 - tests/models/language/generation/test_common.py
34 - tests/models/language/pooling/test_classification.py
35 commands:
36 # Shard slow subset of standard language models tests. Only run when model
37 # source is modified, or when specified test files are modified
38 - pip freeze | grep -E 'torch'
39 - pytest -v -s models/language -m 'core_model and slow_test' --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
40 parallelism: 2
41 mirror:
42 amd:
43 label: ":amd: (MI355 DPX) Language Models (Extra Standard) Shard %N"
44 dind: false
45 device: mi355_dpx
46 timeout_in_minutes: 40
47 depends_on:
48 - image-build-amd
49 source_file_dependencies:
50 - vllm/model_executor/models/
51 - vllm/model_executor/model_loader/
52 - vllm/model_executor/layers/
53 - vllm/v1/attention/backends/
54 - vllm/v1/attention/selector.py
55 - tests/models/language/pooling/test_embedding.py
56 - tests/models/language/generation/test_common.py
57 - tests/models/language/pooling/test_classification.py
58 - vllm/_aiter_ops.py
59 - vllm/platforms/rocm.py
60 - label: ":nvidia: (H200 MIG 35GB) Language Models (Hybrid) Shard %N"
61 device: h200_35gb
62 key: language-models-tests-hybrid
63 timeout_in_minutes: 65
64 env:
65 # GitHub intermittently rejects HTTP/2 upload-pack requests from the H200
66 # CI image. Force Git's smart protocol to use HTTP/1.1 for source installs.
67 GIT_CONFIG_COUNT: "1"
68 GIT_CONFIG_KEY_0: http.version
69 GIT_CONFIG_VALUE_0: HTTP/1.1
70 source_file_dependencies:
71 - vllm/
72 - "!vllm/distributed/kv_transfer/"
73 - tests/models/language/generation
74 commands:
75 # Install fast path packages for testing against transformers
76 # torch>=2.14 nightly requires C++20 for ATen headers (pytorch/pytorch#178150);
77 # mamba@v2.3.0 pins -std=c++17, so build from a patched checkout instead of the git URL.
78 - rm -rf /tmp/mamba-src && git clone --depth 1 --branch v2.3.0 https://github.com/state-spaces/mamba /tmp/mamba-src && sed -i 's/-std=c++17/-std=c++20/g' /tmp/mamba-src/setup.py && MAMBA_FORCE_BUILD=TRUE uv pip install --system --no-build-isolation /tmp/mamba-src
79 - CAUSAL_CONV1D_FORCE_BUILD=TRUE uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
80 # Shard the hybrid language model tests that are numerically stable on Hopper.
81 - pytest -v -s models/language/generation -m hybrid_model -k 'not granite-4.0-tiny-preview' --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
82 parallelism: 2
83 mirror:
84 amd:
85 label: ":amd: (MI355 DPX) Language Models (Hybrid) Shard %N"
86 dind: false
87 device: mi355_dpx
88 timeout_in_minutes: 65
89 depends_on:
90 - image-build-amd
91 commands:
92 - MAMBA_FORCE_BUILD=TRUE uv pip install --system --no-build-isolation 'git+https://github.com/AndreasKaratzas/mamba@fix-rocm-7.0-warp-size-constexpr'
93 - CAUSAL_CONV1D_FORCE_BUILD=TRUE uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
94 - pytest -v -s models/language/generation -m hybrid_model -k 'not granite-4.0-tiny-preview' --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
95
96 # Granite 4 hybrid generation is sensitive to hardware-specific Triton SSD
97 # autotuning (https://github.com/vllm-project/vllm/issues/25194). Keep this one
98 # correctness test on L4 until its H200 output matches the Transformers reference.
99 - label: ":nvidia: (L4) Granite Language Model Compatibility"
100 key: language-models-tests-granite-l4-compatibility
101 device: l4
102 num_devices: 1
103 timeout_in_minutes: 65
104 source_file_dependencies:
105 - vllm/
106 - "!vllm/distributed/kv_transfer/"
107 - tests/models/language/generation
108 commands:
109 # torch>=2.14 nightly requires C++20 for ATen headers (pytorch/pytorch#178150);
110 # mamba@v2.3.0 pins -std=c++17, so build from a patched checkout instead of the git URL.
111 - rm -rf /tmp/mamba-src && git clone --depth 1 --branch v2.3.0 https://github.com/state-spaces/mamba /tmp/mamba-src && sed -i 's/-std=c++17/-std=c++20/g' /tmp/mamba-src/setup.py && MAMBA_FORCE_BUILD=TRUE uv pip install --system --no-build-isolation /tmp/mamba-src
112 - CAUSAL_CONV1D_FORCE_BUILD=TRUE uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
113 - pytest -v -s models/language/generation -m hybrid_model -k 'granite-4.0-tiny-preview'
114 mirror:
115 amd:
116 label: ":amd: (MI355 DPX) Granite Language Model Compatibility"
117 dind: false
118 device: mi355_dpx
119 num_devices: 1
120 timeout_in_minutes: 80
121 working_dir: /vllm-workspace/tests
122 depends_on:
123 - image-build-amd
124 source_file_dependencies:
125 - vllm/
126 - "!vllm/distributed/kv_transfer/"
127 - tests/models/language/generation
128 commands:
129 - export GIT_CONFIG_COUNT=1
130 - export GIT_CONFIG_KEY_0=http.version
131 - export GIT_CONFIG_VALUE_0=HTTP/1.1
132 - MAMBA_FORCE_BUILD=TRUE uv pip install --system --no-build-isolation 'git+https://github.com/AndreasKaratzas/mamba@fix-rocm-7.0-warp-size-constexpr'
133 - CAUSAL_CONV1D_FORCE_BUILD=TRUE uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
134 - pytest -v -s models/language/generation -m hybrid_model -k 'granite-4.0-tiny-preview'
135
136 - label: ":nvidia: (H200 MIG 35GB) Language Models (Extended Generation)"
137 device: h200_35gb
138 key: language-models-test-extended-generation
139 timeout_in_minutes: 80
140 optional: true
141 source_file_dependencies:
142 - vllm/
143 - "!vllm/distributed/kv_transfer/"
144 - tests/models/language/generation
145 commands:
146 # Install fast path packages for testing against transformers
147 # torch>=2.14 nightly requires C++20 for ATen headers (pytorch/pytorch#178150);
148 # mamba@v2.3.0 pins -std=c++17, so build from a patched checkout instead of the git URL.
149 - rm -rf /tmp/mamba-src && git clone --depth 1 --branch v2.3.0 https://github.com/state-spaces/mamba /tmp/mamba-src && sed -i 's/-std=c++17/-std=c++20/g' /tmp/mamba-src/setup.py && MAMBA_FORCE_BUILD=TRUE uv pip install --system --no-build-isolation /tmp/mamba-src
150 - CAUSAL_CONV1D_FORCE_BUILD=TRUE uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
151 - pytest -v -s models/language/generation -m '(not core_model) and (not hybrid_model)'
152 mirror:
153 amd:
154 label: ":amd: (MI355 DPX) Language Models (Extended Generation)"
155 dind: false
156 device: mi355_dpx
157 timeout_in_minutes: 95
158 depends_on:
159 - image-build-amd
160 commands:
161 - MAMBA_FORCE_BUILD=TRUE uv pip install --system --no-build-isolation 'git+https://github.com/AndreasKaratzas/mamba@fix-rocm-7.0-warp-size-constexpr'
162 - CAUSAL_CONV1D_FORCE_BUILD=TRUE uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
163 - pytest -v -s models/language/generation -m '(not core_model) and (not hybrid_model)'
164
165 - label: ":nvidia: (H200 MIG 18GB) Language Models (PPL)"
166 key: language-models-test-ppl
167 timeout_in_minutes: 30
168 device: h200_18gb
169 optional: true
170 source_file_dependencies:
171 - vllm/
172 - "!vllm/distributed/kv_transfer/"
173 - tests/models/language/generation_ppl_test
174 commands:
175 - pytest -v -s models/language/generation_ppl_test
176 mirror:
177 amd:
178 label: ":amd: (MI250) Language Models (PPL)"
179 dind: false
180 device: mi250_1
181 timeout_in_minutes: 65
182 depends_on:
183 - image-build-amd
184 source_file_dependencies:
185 - vllm/model_executor/models/qwen3_5.py
186 - vllm/model_executor/models/qwen3_5_mtp.py
187 - vllm/transformers_utils/configs/qwen3_5.py
188 - vllm/transformers_utils/configs/qwen3_5_moe.py
189 - vllm/model_executor/models/qwen2.py
190 - vllm/model_executor/models/qwen3.py
191 - vllm/model_executor/models/qwen3_next.py
192 - vllm/model_executor/models/qwen3_next_mtp.py
193 - vllm/model_executor/layers/fla/ops/
194 - vllm/_aiter_ops.py
195 - vllm/v1/attention/backends/triton_attn.py
196 - vllm/v1/attention/backends/rocm_attn.py
197 - vllm/v1/attention/backends/rocm_aiter_unified_attn.py
198 - vllm/v1/attention/backends/rocm_aiter_fa.py
199 - vllm/v1/attention/backends/flex_attention.py
200 - vllm/v1/attention/ops/
201 - vllm/platforms/rocm.py
202 - tests/models/language/generation_ppl_test
203
204 - label: ":nvidia: (H200 MIG 35GB) Language Models (Extended Pooling) Shard %N"
205 device: h200_35gb
206 key: language-models-test-extended-pooling
207 timeout_in_minutes: 120
208 parallelism: 4
209 optional: true
210 source_file_dependencies:
211 - vllm/
212 - "!vllm/distributed/kv_transfer/"
213 - tests/models/language/pooling
214 commands:
215 - pytest -v -s models/language/pooling -m 'not core_model' --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
216 mirror:
217 amd:
218 label: ":amd: (MI300) Language Models (Extended Pooling) Shard %N"
219 dind: false
220 device: mi300_1
221 timeout_in_minutes: 95
222 depends_on:
223 - image-build-amd
224
225 - label: ":nvidia: (H200 MIG 18GB) Language Models (MTEB)"
226 key: language-models-test-mteb
227 timeout_in_minutes: 68
228 device: h200_18gb
229 optional: true
230 source_file_dependencies:
231 - vllm/
232 - "!vllm/distributed/kv_transfer/"
233 - tests/models/language/pooling_mteb_test
234 commands:
235 - pytest -v -s models/language/pooling_mteb_test
236 mirror:
237 amd:
238 label: ":amd: (MI250) Language Models (MTEB)"
239 dind: false
240 device: mi250_1
241 timeout_in_minutes: 60
242 depends_on:
243 - image-build-amd