@tduncan / Nvidia-Optimization-Stack-VLLM
Code Commits Issues Pulls Cloud Agent CI/CD
main
main
yaml 284 lines 7.98 KB
Raw
1 group: PyTorch
2 depends_on:
3 - image-build
4 steps:
5 - label: ":nvidia: (H200 MIG 35GB) PyTorch Compilation"
6 device: h200_35gb
7 key: pytorch-compilation-unit-tests
8 timeout_in_minutes: 30
9 source_file_dependencies: &pytorch-compilation-deps
10 - vllm/__init__.py
11 - vllm/_aiter_ops.py
12 - vllm/_custom_ops.py
13 - vllm/compilation/
14 - vllm/config/
15 - vllm/distributed/
16 - "!vllm/distributed/kv_transfer/"
17 - vllm/engine/
18 - vllm/env_override.py
19 - vllm/envs.py
20 - vllm/forward_context.py
21 - vllm/inputs/
22 - vllm/ir/
23 - vllm/kernels/
24 - vllm/logger.py
25 - vllm/model_executor/
26 - vllm/multimodal/
27 - vllm/platforms/
28 - vllm/plugins/
29 - vllm/sampling_params.py
30 - vllm/sequence.py
31 - vllm/transformers_utils/
32 - vllm/triton_utils/
33 - vllm/utils/
34 - vllm/v1/
35 - tests/compile
36 commands:
37 # Run unit tests defined directly under compile/,
38 # not including subdirectories, which are usually heavier
39 # tests covered elsewhere.
40 # Use `find` to launch multiple instances of pytest so that
41 # they do not suffer from https://github.com/vllm-project/vllm/issues/28965
42 # However, find does not normally propagate error codes, so we combine it with xargs
43 # (using -0 for proper path handling)
44 - "find compile/ -maxdepth 1 -name 'test_*.py' -print0 | xargs -0 -n1 -I{} pytest -s -v '{}'"
45 mirror:
46 amd:
47 label: ":amd: (MI300) PyTorch Compilation"
48 dind: false
49 device: mi300_1
50 timeout_in_minutes: 45
51 depends_on:
52 - image-build-amd
53 source_file_dependencies:
54 - csrc/
55 - vllm/platforms/rocm.py
56
57 - label: ":nvidia: (H200 MIG 35GB) PyTorch Compilation / Dynamic Shapes"
58 device: h200_35gb
59 key: pytorch-compilation-dynamic-shapes
60 timeout_in_minutes: 30
61 source_file_dependencies: *pytorch-compilation-deps
62 commands:
63 - pytest -s -v compile/dynamic_shapes/
64 mirror:
65 amd:
66 label: ":amd: (MI300) PyTorch Compilation / Dynamic Shapes"
67 dind: false
68 device: mi300_1
69 timeout_in_minutes: 45
70 depends_on:
71 - image-build-amd
72 source_file_dependencies:
73 - csrc/
74 - vllm/platforms/rocm.py
75
76 - label: ":nvidia: (H200 MIG 18GB) PyTorch Compilation H100 Cases"
77 key: pytorch-compilation-unit-tests-h100
78 timeout_in_minutes: 30
79 device: h200_18gb
80 source_file_dependencies:
81 - vllm/__init__.py
82 - vllm/_aiter_ops.py
83 - vllm/_custom_ops.py
84 - vllm/compilation/
85 - vllm/config/
86 - vllm/distributed/
87 - "!vllm/distributed/kv_transfer/"
88 - vllm/engine/
89 - vllm/env_override.py
90 - vllm/envs.py
91 - vllm/forward_context.py
92 - vllm/inputs/
93 - vllm/ir/
94 - vllm/kernels/
95 - vllm/logger.py
96 - vllm/model_executor/
97 - vllm/multimodal/
98 - vllm/platforms/
99 - vllm/plugins/
100 - vllm/sampling_params.py
101 - vllm/sequence.py
102 - vllm/transformers_utils/
103 - vllm/triton_utils/
104 - vllm/utils/
105 - vllm/v1/
106 - tests/compile/h100/
107 commands:
108 - "find compile/h100/ -name 'test_*.py' -print0 | xargs -0 -n1 -I{} pytest -s -v '{}'"
109 mirror:
110 amd:
111 label: ":amd: (MI355 DPX) PyTorch Compilation H100 Cases"
112 dind: false
113 device: mi355_dpx
114 timeout_in_minutes: 45
115 working_dir: "/vllm-workspace/tests"
116 depends_on:
117 - image-build-amd
118 source_file_dependencies:
119 - tests/utils.py
120 - csrc/
121 commands:
122 - pytest -v -s compile/h100/test_startup.py
123
124 - label: ":nvidia: (L4) PyTorch Compilation Passes"
125 key: pytorch-compilation-passes-unit-tests
126 device: l4
127 num_devices: 1
128 timeout_in_minutes: 45
129 source_file_dependencies:
130 - vllm/__init__.py
131 - vllm/_aiter_ops.py
132 - vllm/_custom_ops.py
133 - vllm/compilation/
134 - vllm/config/
135 - vllm/distributed/
136 - "!vllm/distributed/kv_transfer/"
137 - vllm/engine/
138 - vllm/env_override.py
139 - vllm/envs.py
140 - vllm/forward_context.py
141 - vllm/inputs/
142 - vllm/ir/
143 - vllm/kernels/
144 - vllm/logger.py
145 - vllm/model_executor/
146 - vllm/multimodal/
147 - vllm/platforms/
148 - vllm/plugins/
149 - vllm/sampling_params.py
150 - vllm/sequence.py
151 - vllm/transformers_utils/
152 - vllm/triton_utils/
153 - vllm/utils/
154 - vllm/v1/
155 - tests/compile/passes
156 commands:
157 - pytest -s -v compile/passes --ignore compile/passes/distributed
158 mirror:
159 amd:
160 label: ':amd: (MI300) PyTorch Compilation Passes'
161 dind: false
162 device: mi300_1
163 timeout_in_minutes: 80
164 working_dir: /vllm-workspace/tests
165 depends_on:
166 - image-build-amd
167 source_file_dependencies:
168 - vllm/
169 - tests/compile/passes
170 - vllm/platforms/rocm.py
171 commands:
172 - pytest -s -v compile/passes --ignore compile/passes/distributed
173
174 - label: ":nvidia: (H200 MIG 35GB) PyTorch Fullgraph"
175 device: h200_35gb
176 key: pytorch-fullgraph-test
177 timeout_in_minutes: 90
178 source_file_dependencies:
179 - vllm/__init__.py
180 - vllm/_aiter_ops.py
181 - vllm/_custom_ops.py
182 - vllm/compilation/
183 - vllm/config/
184 - vllm/distributed/
185 - "!vllm/distributed/kv_transfer/"
186 - vllm/engine/
187 - vllm/env_override.py
188 - vllm/envs.py
189 - vllm/forward_context.py
190 - vllm/inputs/
191 - vllm/ir/
192 - vllm/kernels/
193 - vllm/logger.py
194 - vllm/model_executor/
195 - vllm/multimodal/
196 - vllm/platforms/
197 - vllm/plugins/
198 - vllm/sampling_params.py
199 - vllm/sequence.py
200 - vllm/transformers_utils/
201 - vllm/triton_utils/
202 - vllm/utils/
203 - vllm/v1/
204 - tests/compile
205 - csrc/
206 commands:
207 # Run smoke tests under fullgraph directory, except test_full_cudagraph.py
208 # as it is a heavy test that is covered in other steps.
209 # Use `find` to launch multiple instances of pytest so that
210 # they do not suffer from https://github.com/vllm-project/vllm/issues/28965
211 - "find compile/fullgraph/ -name 'test_*.py' -not -name 'test_full_cudagraph.py' -print0 | xargs -0 -n1 -I{} pytest -s -v '{}'"
212 mirror:
213 amd:
214 label: ":amd: (MI355 DPX) PyTorch Fullgraph"
215 dind: false
216 device: mi355_dpx
217 timeout_in_minutes: 90
218 # Remove once the ROCm Triton pin includes triton-lang/triton#11567.
219 # Avoid overlapping LLVM codegen/LLD by disabling background precompilation.
220 env:
221 TORCHINDUCTOR_PRECOMPILATION_TIMEOUT_SECONDS: "0"
222 depends_on:
223 - image-build-amd
224
225 # Hopper-only DeepSeek-V2-Lite cases in this file require two 29.3-GiB model
226 # instances and cannot fit a 35GB MIG slice. L4 retains the original coverage:
227 # those SM90 cases skip while the architecture-compatible cases still run.
228 - label: ":nvidia: (L4) PyTorch Fullgraph CUDAGraph Compatibility"
229 key: pytorch-fullgraph-cudagraph-l4-compatibility
230 device: l4
231 num_devices: 1
232 timeout_in_minutes: 60
233 source_file_dependencies:
234 - vllm/__init__.py
235 - vllm/_aiter_ops.py
236 - vllm/_custom_ops.py
237 - vllm/compilation/
238 - vllm/config/
239 - vllm/distributed/
240 - "!vllm/distributed/kv_transfer/"
241 - vllm/engine/
242 - vllm/env_override.py
243 - vllm/envs.py
244 - vllm/forward_context.py
245 - vllm/inputs/
246 - vllm/ir/
247 - vllm/kernels/
248 - vllm/logger.py
249 - vllm/model_executor/
250 - vllm/multimodal/
251 - vllm/platforms/
252 - vllm/plugins/
253 - vllm/sampling_params.py
254 - vllm/sequence.py
255 - vllm/transformers_utils/
256 - vllm/triton_utils/
257 - vllm/utils/
258 - vllm/v1/
259 - tests/compile
260 commands:
261 - pytest -s -v compile/fullgraph/test_full_cudagraph.py
262
263 - label: ":nvidia: (H200 MIG 18GB) PyTorch Nightly Dependency Override Check"
264 key: pytorch-nightly-dependency-override-check
265 # if this test fails, it means the nightly torch version is not compatible with some
266 # of the dependencies. Please check the error message and add the package to whitelist
267 # in /vllm/tools/pre_commit/generate_nightly_torch_test.py
268 device: h200_18gb
269 soft_fail: true
270 source_file_dependencies:
271 - requirements/test/nightly-torch.txt
272 commands:
273 - bash standalone_tests/pytorch_nightly_dependency.sh
274 mirror:
275 amd:
276 label: ":amd: (MI300) PyTorch Nightly Dependency Override Check"
277 dind: false
278 device: mi300_1
279 timeout_in_minutes: 30
280 depends_on:
281 - image-build-amd
282 source_file_dependencies:
283 - requirements/test/nightly-torch.txt
284 - vllm/platforms/rocm.py