@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / a943839e9a
Fork 0

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 4, 22:30 UTC

[ROCm][CI] Introducing new MI300 nodes (#39531)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>

Andreas Karatzas committed Apr 20, 2026 at 03:09 UTC a943839e9a7c9ee00e57fcfabbbf0a453393cc97
2 files changed +1966 -2319
.buildkite/test-amd.yaml
+1964 -2317
@@ -1,8 +1,8 @@
1 # In this file, you can add more tests to run either by adding a new step or
2 # adding a new command to an existing step. See different options here for examples.
3
4 -# This script will be feed into Jinja template in `test-template-aws.j2` at
5 -# https://github.com/vllm-project/buildkite-ci/blob/main/scripts/test-template-aws.j2
4 +# This script will be feed into Jinja template in `test-template-amd.j2` at
5 +# https://github.com/vllm-project/buildkite-ci/blob/main/scripts/test-template-amd.j2
6 # to generate the final pipeline yaml file.
7
8 # Documentation
@@ -39,8 +39,8 @@
39 #####################################################################################################################################
40 # #
41 # IMPORTANT: #
42 -# * Currently AMD CI has MI250 agents, MI325 agents, and MI355 agents. All upcoming feature improvements are tracked in: #
43 -# https://github.com/vllm-project/vllm/issues/34994 #
42 +# * Currently AMD CI has MI250 agents, MI300 agents, MI325 agents, and MI355 agents. All upcoming feature improvements are #
43 +# tracked in: https://github.com/vllm-project/vllm/issues/34994 #
44 # #
45 #-----------------------------------------------------------------------------------------------------------------------------------#
46 # #
@@ -104,193 +104,195 @@
104 #####################################################################################################################################
105
106
107 -
108 -
107 steps:
108
109 +#########################################################################################################################################
110 +# #
111 +# MI250 (gfx90a) tests #
112 +# #
113 +#########################################################################################################################################
114
112 -#####################################################################################################################################
113 -# #
114 -# MI250 test definitions ( currently the test set is completely mirrored // TBD which tests are to be routed there ultimately) #
115 -# #
116 -#####################################################################################################################################
115 +#----------------------------------------------------- mi250 · basic_correctness -----------------------------------------------------#
116
118 -- label: Pytorch Nightly Dependency Override Check # TBD
117 +- label: Distributed Model Tests (2 GPUs) # TBD
118 timeout_in_minutes: 180
119 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
121 - agent_pool: mi250_1
122 - optional: true
123 - soft_fail: true
120 + agent_pool: mi250_2
121 + num_gpus: 2
122 working_dir: "/vllm-workspace/tests"
123 source_file_dependencies:
126 - - requirements/test/nightly-torch.txt
124 + - vllm/model_executor/model_loader/sharded_state_loader.py
125 + - vllm/model_executor/models/
126 + - vllm/model_executor/layers/
127 + - vllm/v1/attention/backends/
128 + - vllm/v1/attention/selector.py
129 + - tests/basic_correctness/
130 + - tests/model_executor/model_loader/test_sharded_state_loader.py
131 + - tests/models/
132 + - vllm/_aiter_ops.py
133 - vllm/platforms/rocm.py
134 commands:
129 - - bash standalone_tests/pytorch_nightly_dependency.sh
135 + - TARGET_TEST_SUITE=L4 pytest basic_correctness/ -v -s -m 'distributed(num_gpus=2)'
136 + - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s model_executor/model_loader/test_sharded_state_loader.py -m '(not slow_test)'
137 + - pytest models/test_transformers.py -v -s -m 'distributed(num_gpus=2)'
138 + - pytest models/language -v -s -m 'distributed(num_gpus=2)'
139 + - pytest models/multimodal -v -s -m 'distributed(num_gpus=2)' --ignore models/multimodal/generation/test_whisper.py
140 + - VLLM_WORKER_MULTIPROC_METHOD=spawn pytest models/multimodal/generation/test_whisper.py -v -s -m 'distributed(num_gpus=2)'
141
142 +#-------------------------------------------------------- mi250 · benchmarks ---------------------------------------------------------#
143
132 -- label: Async Engine, Inputs, Utils, Worker # TBD
144 +- label: Benchmarks # TBD
145 timeout_in_minutes: 180
146 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
147 agent_pool: mi250_1
136 - working_dir: "/vllm-workspace/tests"
148 + working_dir: "/vllm-workspace/.buildkite"
149 source_file_dependencies:
138 - - vllm/
139 - - tests/detokenizer
140 - - tests/multimodal
141 - - tests/utils_
150 + - benchmarks/
151 + - vllm/platforms/rocm.py
152 commands:
143 - - pytest -v -s detokenizer
144 - - pytest -v -s -m 'not cpu_test' multimodal
145 - - pytest -v -s utils_
153 + - bash scripts/run-benchmarks.sh
154
155 +#---------------------------------------------------------- mi250 · compile ----------------------------------------------------------#
156
148 -- label: Async Engine, Inputs, Utils, Worker, Config (CPU) # TBD
157 +- label: PyTorch Compilation Unit Tests # TBD
158 timeout_in_minutes: 180
159 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
160 agent_pool: mi250_1
161 + torch_nightly: true
162 optional: true
153 - no_gpu: true
163 working_dir: "/vllm-workspace/tests"
164 source_file_dependencies:
156 - - vllm/
157 - - tests/test_inputs.py
158 - - tests/test_outputs.py
159 - - tests/test_pooling_params.py
160 - - tests/test_ray_env.py
161 - - tests/multimodal
162 - - tests/renderers
163 - - tests/standalone_tests/lazy_imports.py
164 - - tests/tokenizers_
165 - - tests/tool_parsers
166 - - tests/transformers_utils
167 - - tests/config
165 + - vllm/compilation/
166 + - vllm/model_executor/layers/
167 + - vllm/v1/worker/
168 + - vllm/v1/attention/
169 + - vllm/v1/cudagraph_dispatcher.py
170 + - vllm/config/compilation.py
171 + - csrc/
172 + - tests/compile
173 + - vllm/platforms/rocm.py
174 commands:
169 - - python3 standalone_tests/lazy_imports.py
170 - - pytest -v -s test_inputs.py
171 - - pytest -v -s test_outputs.py
172 - - pytest -v -s test_pooling_params.py
173 - - pytest -v -s test_ray_env.py
174 - - pytest -v -s -m 'cpu_test' multimodal
175 - - pytest -v -s renderers
176 - - pytest -v -s tokenizers_
177 - - pytest -v -s tool_parsers
178 - - pytest -v -s transformers_utils
179 - - pytest -v -s config
180 -
175 + - "find compile/ -maxdepth 1 -name 'test_*.py' -print0 | xargs -0 -n1 -I{} pytest -s -v '{}'"
176
182 -- label: Python-only Installation # TBD
177 +- label: PyTorch Fullgraph # TBD
178 timeout_in_minutes: 180
179 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
180 agent_pool: mi250_1
181 optional: true
182 + torch_nightly: true
183 working_dir: "/vllm-workspace/tests"
184 source_file_dependencies:
189 - - tests/standalone_tests/python_only_compile.sh
190 - - setup.py
185 + - vllm/compilation/
186 + - vllm/model_executor/
187 + - vllm/v1/attention/
188 + - vllm/config/compilation.py
189 + - csrc/
190 + - tests/compile
191 - vllm/platforms/rocm.py
192 commands:
193 - - bash standalone_tests/python_only_compile.sh
194 -
193 + - pytest -v -s compile/fullgraph/test_full_graph.py -k 'not test_fp8_kv_scale_compile'
194
196 -- label: Basic Correctness # TBD
195 +- label: PyTorch Fullgraph Smoke Test # TBD
196 timeout_in_minutes: 180
197 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
198 agent_pool: mi250_1
200 - fast_check: true
199 + optional: true
200 torch_nightly: true
201 working_dir: "/vllm-workspace/tests"
202 source_file_dependencies:
204 - - vllm/
205 - - tests/basic_correctness/test_basic_correctness
206 - - tests/basic_correctness/test_cpu_offload
207 - - tests/basic_correctness/test_cumem.py
208 - commands:
209 - - export VLLM_WORKER_MULTIPROC_METHOD=spawn
210 - - pytest -v -s basic_correctness/test_cumem.py
211 - - pytest -v -s basic_correctness/test_basic_correctness.py
212 - - pytest -v -s basic_correctness/test_cpu_offload.py
213 -
214 -
215 -- label: Entrypoints Unit Tests # TBD
216 - timeout_in_minutes: 180
217 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
218 - agent_pool: mi250_1
219 - fast_check: true
220 - working_dir: "/vllm-workspace/tests"
221 - source_file_dependencies:
222 - - vllm/entrypoints
223 - - tests/entrypoints/
203 + - vllm/compilation/
204 + - vllm/model_executor/
205 + - vllm/v1/attention/
206 + - vllm/config/compilation.py
207 + - csrc/
208 + - tests/compile
209 - vllm/platforms/rocm.py
210 commands:
226 - - pytest -v -s entrypoints/openai/tool_parsers
227 - - pytest -v -s entrypoints/ --ignore=entrypoints/llm --ignore=entrypoints/rpc --ignore=entrypoints/sleep --ignore=entrypoints/serve/instrumentator --ignore=entrypoints/openai --ignore=entrypoints/offline_mode --ignore=entrypoints/test_chat_utils.py --ignore=entrypoints/pooling
228 -
211 + - "find compile/fullgraph/ -name 'test_*.py' -not -name 'test_full_graph.py' -exec pytest -s -v {} \\\\;"
212
230 -- label: Entrypoints Integration (LLM) # TBD
213 +- label: Distributed Compile + RPC Tests (2 GPUs) # TBD
214 timeout_in_minutes: 180
215 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
233 - agent_pool: mi250_1
234 - fast_check: true
235 - torch_nightly: true
216 + agent_pool: mi250_2
217 + num_gpus: 2
218 + optional: true
219 working_dir: "/vllm-workspace/tests"
220 source_file_dependencies:
238 - - vllm/
239 - - tests/entrypoints/llm
240 - - tests/entrypoints/offline_mode
221 + - vllm/compilation/
222 + - vllm/distributed/
223 + - vllm/engine/
224 + - vllm/executor/
225 + - vllm/worker/worker_base.py
226 + - vllm/v1/engine/
227 + - vllm/v1/worker/
228 + - tests/compile/fullgraph/test_basic_correctness.py
229 + - tests/compile/test_wrapper.py
230 + - tests/entrypoints/llm/test_collective_rpc.py
231 + - vllm/platforms/rocm.py
232 commands:
242 - - export VLLM_WORKER_MULTIPROC_METHOD=spawn
243 - - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_collective_rpc.py
244 - - pytest -v -s entrypoints/llm/test_generate.py
245 - - pytest -v -s entrypoints/offline_mode
233 + - export TORCH_NCCL_BLOCKING_WAIT=1
234 + - pytest -v -s entrypoints/llm/test_collective_rpc.py
235 + - pytest -v -s ./compile/fullgraph/test_basic_correctness.py
236 + - pytest -v -s ./compile/test_wrapper.py
237
238 +#-------------------------------------------------------- mi250 · distributed --------------------------------------------------------#
239
248 -- label: Entrypoints Integration (API Server 2) # TBD
240 +- label: Distributed Comm Ops # TBD
241 timeout_in_minutes: 180
242 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
251 - agent_pool: mi250_1
252 - fast_check: true
253 - torch_nightly: true
243 + agent_pool: mi250_2
244 + num_gpus: 2
245 working_dir: "/vllm-workspace/tests"
246 source_file_dependencies:
256 - - vllm/
257 - - tests/entrypoints/rpc
258 - - tests/entrypoints/serve/instrumentator
259 - - tests/tool_use
247 + - vllm/distributed
248 + - tests/distributed
249 + - vllm/platforms/rocm.py
250 commands:
261 - - export VLLM_WORKER_MULTIPROC_METHOD=spawn
262 - - pytest -v -s entrypoints/serve/instrumentator
263 - - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/rpc
264 - - pytest -v -s tool_use
265 -
251 + - pytest -v -s distributed/test_comm_ops.py
252 + - pytest -v -s distributed/test_shm_broadcast.py
253 + - pytest -v -s distributed/test_shm_buffer.py
254 + - pytest -v -s distributed/test_shm_storage.py
255
267 -- label: Entrypoints Integration (Responses API) # TBD
256 +- label: Distributed Torchrun + Shutdown Tests (2 GPUs) # TBD
257 timeout_in_minutes: 180
258 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
270 - agent_pool: mi250_1
271 - fast_check: true
272 - torch_nightly: true
259 + agent_pool: mi250_2
260 + num_gpus: 2
261 + optional: true
262 working_dir: "/vllm-workspace/tests"
263 source_file_dependencies:
275 - - vllm/
276 - - tests/entrypoints/openai/responses
264 + - vllm/distributed/
265 + - vllm/engine/
266 + - vllm/executor/
267 + - vllm/worker/worker_base.py
268 + - vllm/v1/engine/
269 + - vllm/v1/worker/
270 + - tests/distributed/
271 + - tests/v1/shutdown
272 + - tests/v1/worker/test_worker_memory_snapshot.py
273 + - vllm/platforms/rocm.py
274 commands:
278 - - export VLLM_WORKER_MULTIPROC_METHOD=spawn
279 - - pytest -v -s entrypoints/openai/responses
280 -
275 + - export TORCH_NCCL_BLOCKING_WAIT=1
276 + - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
277 + - VLLM_TEST_SAME_HOST=1 VLLM_TEST_WITH_DEFAULT_DEVICE_SET=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
278 + - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s v1/shutdown
279 + - pytest -v -s v1/worker/test_worker_memory_snapshot.py
280
282 -- label: EPLB Algorithm # TBD
281 +- label: Elastic EP Scaling Test # TBD
282 timeout_in_minutes: 180
283 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
285 - agent_pool: mi250_1
284 + agent_pool: mi250_4
285 + num_gpus: 4
286 working_dir: "/vllm-workspace/tests"
287 source_file_dependencies:
288 - - vllm/distributed/eplb
289 - - tests/distributed/test_eplb_algo.py
288 + - vllm/distributed/
289 + - vllm/engine/
290 + - vllm/executor/
291 + - vllm/compilation/
292 + - tests/distributed/
293 - vllm/platforms/rocm.py
294 commands:
292 - - pytest -v -s distributed/test_eplb_algo.py
293 -
295 + - pytest -v -s distributed/test_elastic_ep.py
296
297 - label: EPLB Execution # TBD
298 timeout_in_minutes: 180
@@ -307,8 +309,7 @@ steps:
309 - pytest -v -s distributed/test_eplb_execute.py
310 - pytest -v -s distributed/test_eplb_spec_decode.py
311
310 -
311 -- label: Elastic EP Scaling Test # TBD
312 +- label: Pipeline + Context Parallelism (4 GPUs) # TBD
313 timeout_in_minutes: 180
314 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
315 agent_pool: mi250_4
@@ -318,1620 +319,1317 @@ steps:
319 - vllm/distributed/
320 - vllm/engine/
321 - vllm/executor/
321 - - vllm/compilation/
322 + - vllm/model_executor/models/
323 + - vllm/model_executor/layers/
324 + - vllm/v1/attention/backends/
325 + - vllm/v1/attention/selector.py
326 - tests/distributed/
327 + - vllm/_aiter_ops.py
328 - vllm/platforms/rocm.py
329 commands:
325 - - pytest -v -s distributed/test_elastic_ep.py
330 + - pytest -v -s distributed/test_pp_cudagraph.py
331 + - pytest -v -s distributed/test_pipeline_parallel.py
332
333 +#---------------------------------------------------------- mi250 · engine -----------------------------------------------------------#
334
328 -- label: Metrics, Tracing (2 GPUs) # TBD
335 +- label: Engine # TBD
336 timeout_in_minutes: 180
337 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
338 agent_pool: mi250_1
332 - num_gpus: 2
339 + optional: true
340 working_dir: "/vllm-workspace/tests"
341 source_file_dependencies:
342 - vllm/
336 - - tests/v1/tracing
343 + - tests/engine
344 + - tests/test_sequence
345 + - tests/test_config
346 + - tests/test_logger
347 + - tests/test_vllm_port
348 commands:
338 - - "pip install \
339 - 'opentelemetry-sdk>=1.26.0' \
340 - 'opentelemetry-api>=1.26.0' \
341 - 'opentelemetry-exporter-otlp>=1.26.0' \
342 - 'opentelemetry-semantic-conventions-ai>=0.4.1'"
343 - - pytest -v -s v1/tracing
349 + - pytest -v -s engine test_sequence.py test_config.py test_logger.py test_vllm_port.py
350
351 +#----------------------------------------------------------- mi250 · evals -----------------------------------------------------------#
352
346 -- label: Regression # TBD
353 +- label: Multi-Modal Accuracy Eval (Small Models) # TBD
354 timeout_in_minutes: 180
355 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
356 agent_pool: mi250_1
350 - working_dir: "/vllm-workspace/tests"
357 + optional: true
358 + working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
359 source_file_dependencies:
352 - - vllm/
353 - - tests/test_regression
360 + - vllm/multimodal/
361 + - vllm/inputs/
362 + - vllm/v1/core/
363 + - vllm/platforms/rocm.py
364 + - vllm/model_executor/model_loader/
365 commands:
355 - - pip install modelscope
356 - - pytest -v -s test_regression.py
366 + - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-mm-small.txt --tp-size=1
367
368 +#--------------------------------------------------------- mi250 · examples ----------------------------------------------------------#
369
359 -- label: Engine # TBD
370 +- label: Examples # TBD
371 timeout_in_minutes: 180
372 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
373 agent_pool: mi250_1
374 optional: true
364 - working_dir: "/vllm-workspace/tests"
375 + working_dir: "/vllm-workspace/examples"
376 source_file_dependencies:
366 - - vllm/
367 - - tests/engine
368 - - tests/test_sequence
369 - - tests/test_config
370 - - tests/test_logger
371 - - tests/test_vllm_port
377 + - vllm/entrypoints
378 + - vllm/multimodal
379 + - examples/
380 + - vllm/platforms/rocm.py
381 commands:
373 - - pytest -v -s engine test_sequence.py test_config.py test_logger.py test_vllm_port.py
382 + - pip install tensorizer
383 + # Basic
384 + - python3 basic/offline_inference/chat.py --attention-backend TRITON_ATTN
385 + - python3 basic/offline_inference/generate.py --model facebook/opt-125m
386 + - python3 basic/offline_inference/generate.py --model meta-llama/Llama-2-13b-chat-hf --cpu-offload-gb 10
387 + - python3 basic/offline_inference/classify.py
388 + - python3 basic/offline_inference/embed.py
389 + - python3 basic/offline_inference/score.py
390 + # Multi-modal models
391 + - python3 offline_inference/audio_language.py --seed 0
392 + - python3 offline_inference/vision_language.py --seed 0
393 + - python3 offline_inference/vision_language_multi_image.py --seed 0
394 + - python3 offline_inference/encoder_decoder_multimodal.py --model-type whisper --seed 0
395 + # Pooling models
396 + - python3 pooling/embed/vision_embedding_offline.py --seed 0
397 + # Features demo
398 + - python3 offline_inference/prefix_caching.py
399 + - python3 offline_inference/llm_engine_example.py
400 + - python3 others/tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 others/tensorize_vllm_model.py --model facebook/opt-125m deserialize --path-to-tensors /tmp/vllm/facebook/opt-125m/v1/model.tensors
401 + - python3 offline_inference/spec_decode.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
402 + - python3 offline_inference/spec_decode.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
403
404 +#---------------------------------------------------------- mi250 · kernels ----------------------------------------------------------#
405
376 -- label: Engine (1 GPU) # TBD
406 +- label: Kernels Core Operation Test # TBD
407 timeout_in_minutes: 180
408 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
409 agent_pool: mi250_1
410 optional: true
411 working_dir: "/vllm-workspace/tests"
412 source_file_dependencies:
383 - - vllm/v1/
384 - - tests/v1/engine/
413 + - csrc/
414 + - tests/kernels/core
415 + - tests/kernels/test_top_k_per_row.py
416 + - tests/kernels/test_concat_mla_q.py
417 + - vllm/model_executor/layers/rotary_embedding/
418 + - vllm/_aiter_ops.py
419 - vllm/platforms/rocm.py
420 commands:
387 - - pytest -v -s v1/engine/test_preprocess_error_handling.py
388 - - pytest -v -s v1/engine --ignore v1/engine/test_preprocess_error_handling.py
421 + - pytest -v -s kernels/core --ignore=kernels/core/test_minimax_reduce_rms.py kernels/test_concat_mla_q.py kernels/test_top_k_per_row.py
422
390 -
391 -- label: e2e Scheduling (1 GPU) # TBD
423 +- label: Kernels Helion Test # TBD
424 timeout_in_minutes: 180
425 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
426 agent_pool: mi250_1
427 optional: true
428 working_dir: "/vllm-workspace/tests"
429 source_file_dependencies:
398 - - vllm/v1/
399 - - tests/v1/e2e/general/
430 + - vllm/utils/import_utils.py
431 + - tests/kernels/helion/
432 - vllm/platforms/rocm.py
433 commands:
402 - - pytest -v -s v1/e2e/general/test_async_scheduling.py
403 -
434 + - pip install helion==1.0.0
435 + - pytest -v -s kernels/helion/
436
405 -- label: e2e Core (1 GPU) # TBD
437 +- label: Kernels Mamba Test # TBD
438 timeout_in_minutes: 180
439 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
440 agent_pool: mi250_1
441 optional: true
442 working_dir: "/vllm-workspace/tests"
443 source_file_dependencies:
412 - - vllm/v1/
413 - - tests/v1/e2e/general/
444 + - csrc/mamba/
445 + - tests/kernels/mamba
446 + - vllm/model_executor/layers/mamba/ops
447 - vllm/platforms/rocm.py
448 commands:
416 - - pytest -v -s v1/e2e/general --ignore v1/e2e/general/test_async_scheduling.py
449 + - pytest -v -s kernels/mamba
450
451 +#----------------------------------------------------------- mi250 · lora ------------------------------------------------------------#
452
419 -- label: Spec Decode Speculators + MTP # TBD
453 +- label: LoRA %N # TBD
454 timeout_in_minutes: 180
455 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
456 agent_pool: mi250_1
457 + parallelism: 4
458 optional: true
459 working_dir: "/vllm-workspace/tests"
460 source_file_dependencies:
426 - - vllm/v1/spec_decode/
427 - - vllm/v1/worker/gpu/spec_decode/
428 - - vllm/model_executor/model_loader/
429 - - vllm/v1/sample/
430 - - vllm/model_executor/layers/
431 - - vllm/transformers_utils/configs/speculators/
432 - - tests/v1/e2e/spec_decode/
461 + - vllm/lora
462 + - tests/lora
463 - vllm/platforms/rocm.py
464 commands:
435 - - pytest -v -s v1/e2e/spec_decode -k "speculators or mtp_correctness"
465 + - pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py --ignore=lora/test_llm_with_multi_loras.py --ignore=lora/test_olmoe_tp.py --ignore=lora/test_deepseekv2_tp.py --ignore=lora/test_gptoss_tp.py --ignore=lora/test_qwen3moe_tp.py --ignore=lora/test_qwen35_densemodel_lora.py
466
467 +#------------------------------------------------------ mi250 · model_executor -------------------------------------------------------#
468
438 -- label: Spec Decode Ngram + Suffix # TBD
469 +- label: Model Executor # TBD
470 timeout_in_minutes: 180
471 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
472 agent_pool: mi250_1
473 optional: true
474 + torch_nightly: true
475 working_dir: "/vllm-workspace/tests"
476 source_file_dependencies:
445 - - vllm/v1/spec_decode/
446 - - vllm/v1/worker/gpu/spec_decode/
447 - - vllm/model_executor/model_loader/
448 - - vllm/v1/sample/
449 - - vllm/model_executor/layers/
450 - - tests/v1/e2e/spec_decode/
477 + - vllm/engine/arg_utils.py
478 + - vllm/config/model.py
479 + - vllm/model_executor
480 + - tests/model_executor
481 + - tests/entrypoints/openai/completion/test_tensorizer_entrypoint.py
482 + - vllm/_aiter_ops.py
483 - vllm/platforms/rocm.py
484 commands:
453 - - pytest -v -s v1/e2e/spec_decode -k "ngram or suffix"
485 + - apt-get update && apt-get install -y curl libsodium23
486 + - export VLLM_WORKER_MULTIPROC_METHOD=spawn
487 + - pytest -v -s model_executor -m '(not slow_test)'
488 + - pytest -v -s entrypoints/openai/completion/test_tensorizer_entrypoint.py
489
490 +#---------------------------------------------------------- mi250 · models -----------------------------------------------------------#
491
456 -- label: Spec Decode Draft Model # TBD
492 +- label: Basic Models Test (Other CPU) # TBD
493 timeout_in_minutes: 180
494 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
495 agent_pool: mi250_1
496 + no_gpu: true
497 optional: true
498 + torch_nightly: true
499 working_dir: "/vllm-workspace/tests"
500 source_file_dependencies:
463 - - vllm/v1/spec_decode/
464 - - vllm/v1/worker/gpu/spec_decode/
465 - - vllm/model_executor/model_loader/
466 - - vllm/v1/sample/
501 + - vllm/
502 + - tests/models/test_utils.py
503 + - tests/models/test_vision.py
504 + commands:
505 + - pytest -v -s models/test_utils.py models/test_vision.py
506 +
507 +- label: Basic Models Tests (Extra Initialization) %N # TBD
508 + timeout_in_minutes: 180
509 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
510 + agent_pool: mi250_1
511 + torch_nightly: true
512 + parallelism: 2
513 + working_dir: "/vllm-workspace/tests"
514 + source_file_dependencies:
515 + - vllm/model_executor/models/
516 - vllm/model_executor/layers/
468 - - tests/v1/e2e/spec_decode/
517 + - tests/models/test_initialization.py
518 + - tests/models/registry.py
519 + - vllm/_aiter_ops.py
520 - vllm/platforms/rocm.py
521 commands:
471 - - pytest -v -s v1/e2e/spec_decode -k "draft_model or no_sync or batch_inference"
472 -
522 + - pytest -v -s models/test_initialization.py -k 'not test_can_initialize_small_subset' --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
523
474 -- label: V1 e2e (2 GPUs) # TBD
524 +- label: Basic Models Tests (Initialization) # TBD
525 timeout_in_minutes: 180
526 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
527 agent_pool: mi250_1
478 - optional: true
528 + torch_nightly: true
529 working_dir: "/vllm-workspace/tests"
530 source_file_dependencies:
531 - vllm/
482 - - tests/v1/e2e
532 + - tests/models/test_initialization.py
533 + - tests/models/registry.py
534 commands:
484 - - pytest -v -s v1/e2e/spec_decode/test_spec_decode.py -k "tensor_parallelism"
485 -
535 + - pytest -v -s models/test_initialization.py::test_can_initialize_small_subset
536
487 -- label: V1 Sample + Logits # TBD
488 - timeout_in_minutes: 60
537 +- label: Basic Models Tests (Other) # TBD
538 + timeout_in_minutes: 180
539 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
540 agent_pool: mi250_1
491 - optional: true
541 + torch_nightly: true
542 working_dir: "/vllm-workspace/tests"
543 source_file_dependencies:
544 - vllm/
495 - - tests/v1/sample
496 - - tests/v1/logits_processors
497 - - tests/v1/test_oracle.py
498 - - tests/v1/test_request.py
499 - - tests/v1/test_outputs.py
545 + - tests/models/test_terratorch.py
546 + - tests/models/test_transformers.py
547 + - tests/models/test_registry.py
548 commands:
501 - - pytest -v -s v1/sample
502 - - pytest -v -s v1/logits_processors
503 - - pytest -v -s v1/test_oracle.py
504 - - pytest -v -s v1/test_request.py
505 - - pytest -v -s v1/test_outputs.py
506 -
549 + - pytest -v -s models/test_terratorch.py models/test_transformers.py models/test_registry.py
550
508 -- label: V1 Core + KV + Metrics # TBD
509 - timeout_in_minutes: 60
551 +- label: Language Models Test (MTEB) # TBD
552 + timeout_in_minutes: 180
553 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
554 agent_pool: mi250_1
512 - optional: true
555 working_dir: "/vllm-workspace/tests"
556 source_file_dependencies:
557 - vllm/
516 - - tests/v1/core
517 - - tests/v1/executor
518 - - tests/v1/kv_offload
519 - - tests/v1/worker
520 - - tests/v1/kv_connector/unit
521 - - tests/v1/metrics
522 - - tests/entrypoints/openai/correctness/test_lmeval.py
558 + - tests/models/language/pooling_mteb_test
559 commands:
524 - - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
525 - - pytest -v -s -m 'not cpu_test' v1/core
526 - - pytest -v -s v1/executor
527 - - pytest -v -s v1/kv_offload
528 - - pytest -v -s v1/worker
529 - - pytest -v -s -m 'not cpu_test' v1/kv_connector/unit
530 - - pytest -v -s -m 'not cpu_test' v1/metrics
531 - - pip install -U git+https://github.com/robertgshaw2-redhat/lm-evaluation-harness.git@streaming-api
532 - - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
560 + - pytest -v -s models/language/pooling_mteb_test
561
562 +- label: Language Models Test (PPL) # TBD
563 + timeout_in_minutes: 180
564 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
565 + agent_pool: mi250_1
566 + working_dir: "/vllm-workspace/tests"
567 + source_file_dependencies:
568 + - vllm/
569 + - tests/models/language/generation_ppl_test
570 + commands:
571 + - pytest -v -s models/language/generation_ppl_test
572
535 -- label: V1 attention (H100-MI250) # TBD
573 +- label: Language Models Tests (Extra Standard) %N # TBD
574 timeout_in_minutes: 180
575 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
576 agent_pool: mi250_1
577 + torch_nightly: true
578 + parallelism: 2
579 working_dir: "/vllm-workspace/tests"
580 source_file_dependencies:
541 - - vllm/config/attention.py
542 - - vllm/model_executor/layers/attention
543 - - vllm/v1/attention
544 - - tests/v1/attention
581 + - vllm/model_executor/models/
582 + - vllm/model_executor/model_loader/
583 + - vllm/model_executor/layers/
584 + - vllm/v1/attention/backends/
585 + - vllm/v1/attention/selector.py
586 + - tests/models/language/pooling/test_embedding.py
587 + - tests/models/language/generation/test_common.py
588 + - tests/models/language/pooling/test_classification.py
589 - vllm/_aiter_ops.py
546 - - vllm/envs.py
590 - vllm/platforms/rocm.py
591 commands:
549 - - pytest -v -s v1/attention
550 -
592 + - pip freeze | grep -E 'torch'
593 + - export TORCH_NCCL_BLOCKING_WAIT=1
594 + - pytest -v -s models/language -m 'core_model and slow_test' --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
595
552 -- label: V1 others (CPU) # TBD
596 +- label: Multi-Modal Models (Extended Generation 2) # TBD
597 timeout_in_minutes: 180
598 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
599 agent_pool: mi250_1
556 - no_gpu: true
557 - optional: true
600 working_dir: "/vllm-workspace/tests"
601 source_file_dependencies:
602 - vllm/
561 - - tests/v1
603 + - tests/models/multimodal/generation
604 commands:
563 - - pytest -v -s -m 'cpu_test' v1/core
564 - - pytest -v -s v1/structured_output
565 - - pytest -v -s v1/test_serial_utils.py
566 - - pytest -v -s -m 'cpu_test' v1/kv_connector/unit
567 - - pytest -v -s -m 'cpu_test' v1/metrics
568 -
605 + - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
606 + - pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=0) and not core_model'
607
570 -- label: Examples # TBD
608 +- label: Multi-Modal Models (Extended Pooling) # TBD
609 timeout_in_minutes: 180
610 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
611 agent_pool: mi250_1
574 - optional: true
575 - working_dir: "/vllm-workspace/examples"
612 + working_dir: "/vllm-workspace/tests"
613 source_file_dependencies:
577 - - vllm/entrypoints
578 - - vllm/multimodal
579 - - examples/
580 - - vllm/platforms/rocm.py
614 + - vllm/
615 + - tests/models/multimodal/pooling
616 commands:
582 - - pip install tensorizer
583 - # Basic
584 - - python3 basic/offline_inference/chat.py --attention-backend TRITON_ATTN
585 - - python3 basic/offline_inference/generate.py --model facebook/opt-125m
586 - - python3 basic/offline_inference/generate.py --model meta-llama/Llama-2-13b-chat-hf --cpu-offload-gb 10
587 - - python3 basic/offline_inference/classify.py
588 - - python3 basic/offline_inference/embed.py
589 - - python3 basic/offline_inference/score.py
590 - # Multi-modal models
591 - - python3 offline_inference/audio_language.py --seed 0
592 - - python3 offline_inference/vision_language.py --seed 0
593 - - python3 offline_inference/vision_language_multi_image.py --seed 0
594 - - python3 offline_inference/encoder_decoder_multimodal.py --model-type whisper --seed 0
595 - # Pooling models
596 - - python3 pooling/embed/vision_embedding_offline.py --seed 0
597 - # Features demo
598 - - python3 offline_inference/prefix_caching.py
599 - - python3 offline_inference/llm_engine_example.py
600 - - python3 others/tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 others/tensorize_vllm_model.py --model facebook/opt-125m deserialize --path-to-tensors /tmp/vllm/facebook/opt-125m/v1/model.tensors
601 - - python3 offline_inference/spec_decode.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
602 - - python3 offline_inference/spec_decode.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
603 -
617 + - pytest -v -s models/multimodal/pooling -m 'not core_model'
618
605 -- label: Platform Tests (CUDA) # TBD
619 +- label: "Multi-Modal Models (Standard) 3: llava + qwen2_vl" # TBD
620 timeout_in_minutes: 180
621 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
622 agent_pool: mi250_1
623 + torch_nightly: true
624 working_dir: "/vllm-workspace/tests"
625 source_file_dependencies:
626 - vllm/
612 - - tests/cuda
627 + - tests/models/multimodal
628 commands:
614 - - pytest -v -s cuda/test_cuda_context.py
615 - - pytest -v -s cuda/test_platform_no_cuda_init.py
629 + - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
630 + - pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "not qwen2 and not qwen3 and not gemma"
631 + - pytest -v -s models/multimodal/generation/test_qwen2_vl.py -m core_model
632
633 +#---------------------------------------------------------- mi250 · plugins ----------------------------------------------------------#
634
618 -- label: Samplers Test # TBD
635 +- label: Plugin Tests (2 GPUs) # TBD
636 timeout_in_minutes: 180
637 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
621 - agent_pool: mi250_1
638 + agent_pool: mi250_2
639 + num_gpus: 2
640 optional: true
641 working_dir: "/vllm-workspace/tests"
642 source_file_dependencies:
625 - - vllm/model_executor/layers
626 - - vllm/sampling_metadata.py
627 - - vllm/v1/sample/
628 - - vllm/beam_search.py
629 - - tests/samplers
630 - - tests/conftest.py
631 - - vllm/_aiter_ops.py
643 + - vllm/plugins/
644 + - tests/plugins/
645 - vllm/platforms/rocm.py
646 commands:
634 - - pytest -v -s samplers
647 + # BEGIN: platform plugin and general plugin tests, all the code in-between runs on dummy platform
648 + - pip install -e ./plugins/vllm_add_dummy_platform
649 + - pytest -v -s plugins_tests/test_platform_plugins.py
650 + - pip uninstall vllm_add_dummy_platform -y
651 + # END: platform plugin tests
652 + # BEGIN: `io_processor` plugins test, all the code in between uses the `prithvi_io_processor` plugin
653 + - pip install -e ./plugins/prithvi_io_processor_plugin
654 + - pytest -v -s plugins_tests/test_io_processor_plugins.py
655 + - pytest -v -s plugins_tests/test_terratorch_io_processor_plugins.py
656 + - pip uninstall prithvi_io_processor_plugin -y
657 + # END: `io_processor` plugins test
658 + # BEGIN: `bge_m3_sparse io_processor` test
659 + - pip install -e ./plugins/bge_m3_sparse_plugin
660 + - pytest -v -s plugins_tests/test_bge_m3_sparse_io_processor_plugins.py
661 + - pip uninstall bge_m3_sparse_plugin -y
662 + # END: `bge_m3_sparse io_processor` test
663 + # BEGIN: `stat_logger` plugins test
664 + - pip install -e ./plugins/vllm_add_dummy_stat_logger
665 + - pytest -v -s plugins_tests/test_stats_logger_plugins.py
666 + - pip uninstall dummy_stat_logger -y
667 + # END: `stat_logger` plugins test
668 + # BEGIN: other tests
669 + - pytest -v -s plugins_tests/test_scheduler_plugins.py
670 + - pip install -e ./plugins/vllm_add_dummy_model
671 + - pytest -v -s distributed/test_distributed_oot.py
672 + - pytest -v -s entrypoints/openai/chat_completion/test_oot_registration.py
673 + - pytest -v -s models/test_oot_registration.py
674 + - pytest -v -s plugins/lora_resolvers
675
676 +#------------------------------------------------------------ mi250 · v1 -------------------------------------------------------------#
677
637 -- label: LoRA %N # TBD
678 +- label: Batch Invariance (H100-MI250) # TBD
679 timeout_in_minutes: 180
680 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
681 agent_pool: mi250_1
641 - parallelism: 4
682 optional: true
683 working_dir: "/vllm-workspace/tests"
684 source_file_dependencies:
645 - - vllm/lora
646 - - tests/lora
685 + - vllm/v1/attention
686 + - vllm/model_executor/layers
687 + - tests/v1/determinism/
688 + - vllm/_aiter_ops.py
689 - vllm/platforms/rocm.py
690 commands:
649 - - pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py --ignore=lora/test_llm_with_multi_loras.py --ignore=lora/test_olmoe_tp.py --ignore=lora/test_deepseekv2_tp.py --ignore=lora/test_gptoss_tp.py --ignore=lora/test_qwen3moe_tp.py
650 -
691 + - export VLLM_WORKER_MULTIPROC_METHOD=spawn
692 + - pip install pytest-timeout pytest-forked
693 + - pytest -v -s v1/determinism/test_batch_invariance.py
694 + - pytest -v -s v1/determinism/test_rms_norm_batch_invariant.py
695
652 -- label: PyTorch Compilation Unit Tests # TBD
696 +- label: Cudagraph # TBD
697 timeout_in_minutes: 180
698 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
699 agent_pool: mi250_1
656 - torch_nightly: true
700 optional: true
701 working_dir: "/vllm-workspace/tests"
702 source_file_dependencies:
660 - - vllm/compilation/
661 - - vllm/model_executor/layers/
662 - - vllm/v1/worker/
663 - - vllm/v1/attention/
703 + - tests/v1/cudagraph
704 - vllm/v1/cudagraph_dispatcher.py
705 - vllm/config/compilation.py
666 - - csrc/
667 - - tests/compile
706 + - vllm/compilation
707 - vllm/platforms/rocm.py
708 commands:
670 - - "find compile/ -maxdepth 1 -name 'test_*.py' -print0 | xargs -0 -n1 -I{} pytest -s -v '{}'"
671 -
709 + - pytest -v -s v1/cudagraph/test_cudagraph_dispatch.py
710 + - pytest -v -s v1/cudagraph/test_cudagraph_mode.py
711
673 -- label: PyTorch Fullgraph Smoke Test # TBD
712 +- label: e2e Core (1 GPU) # TBD
713 timeout_in_minutes: 180
714 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
715 agent_pool: mi250_1
716 optional: true
678 - torch_nightly: true
717 working_dir: "/vllm-workspace/tests"
718 source_file_dependencies:
681 - - vllm/compilation/
682 - - vllm/model_executor/
683 - - vllm/v1/attention/
684 - - vllm/config/compilation.py
685 - - csrc/
686 - - tests/compile
719 + - vllm/v1/
720 + - tests/v1/e2e/general/
721 - vllm/platforms/rocm.py
722 commands:
689 - - "find compile/fullgraph/ -name 'test_*.py' -not -name 'test_full_graph.py' -exec pytest -s -v {} \\\\;"
690 -
723 + - pytest -v -s v1/e2e/general --ignore v1/e2e/general/test_async_scheduling.py
724
692 -- label: PyTorch Fullgraph # TBD
725 +- label: e2e Scheduling (1 GPU) # TBD
726 timeout_in_minutes: 180
727 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
728 agent_pool: mi250_1
729 optional: true
697 - torch_nightly: true
730 working_dir: "/vllm-workspace/tests"
731 source_file_dependencies:
700 - - vllm/compilation/
701 - - vllm/model_executor/
702 - - vllm/v1/attention/
703 - - vllm/config/compilation.py
704 - - csrc/
705 - - tests/compile
732 + - vllm/v1/
733 + - tests/v1/e2e/general/
734 - vllm/platforms/rocm.py
735 commands:
708 - - pytest -v -s compile/fullgraph/test_full_graph.py -k 'not test_fp8_kv_scale_compile'
709 -
736 + - pytest -v -s v1/e2e/general/test_async_scheduling.py
737
711 -- label: Cudagraph # TBD
738 +- label: Engine (1 GPU) # TBD
739 timeout_in_minutes: 180
740 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
741 agent_pool: mi250_1
742 optional: true
743 working_dir: "/vllm-workspace/tests"
744 source_file_dependencies:
718 - - tests/v1/cudagraph
719 - - vllm/v1/cudagraph_dispatcher.py
720 - - vllm/config/compilation.py
721 - - vllm/compilation
745 + - vllm/v1/
746 + - tests/v1/engine/
747 - vllm/platforms/rocm.py
748 commands:
724 - - pytest -v -s v1/cudagraph/test_cudagraph_dispatch.py
725 - - pytest -v -s v1/cudagraph/test_cudagraph_mode.py
726 -
749 + - pytest -v -s v1/engine/test_preprocess_error_handling.py
750 + - pytest -v -s v1/engine --ignore v1/engine/test_preprocess_error_handling.py
751
728 -- label: Kernels Core Operation Test # TBD
752 +- label: Spec Decode Draft Model # TBD
753 timeout_in_minutes: 180
754 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
755 agent_pool: mi250_1
756 optional: true
757 working_dir: "/vllm-workspace/tests"
758 source_file_dependencies:
735 - - csrc/
736 - - tests/kernels/core
737 - - tests/kernels/test_top_k_per_row.py
738 - - tests/kernels/test_concat_mla_q.py
739 - - vllm/model_executor/layers/rotary_embedding/
740 - - vllm/_aiter_ops.py
759 + - vllm/v1/spec_decode/
760 + - vllm/v1/worker/gpu/spec_decode/
761 + - vllm/model_executor/model_loader/
762 + - vllm/v1/sample/
763 + - vllm/model_executor/layers/
764 + - tests/v1/e2e/spec_decode/
765 - vllm/platforms/rocm.py
766 commands:
743 - - pytest -v -s kernels/core kernels/test_top_k_per_row.py
744 -
767 + - pytest -v -s v1/e2e/spec_decode -k "draft_model or no_sync or batch_inference"
768
746 -- label: Kernels Mamba Test # TBD
769 +- label: Spec Decode Speculators + MTP # TBD
770 timeout_in_minutes: 180
771 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
772 agent_pool: mi250_1
773 optional: true
774 working_dir: "/vllm-workspace/tests"
775 source_file_dependencies:
753 - - csrc/mamba/
754 - - tests/kernels/mamba
755 - - vllm/model_executor/layers/mamba/ops
776 + - vllm/v1/spec_decode/
777 + - vllm/v1/worker/gpu/spec_decode/
778 + - vllm/model_executor/model_loader/
779 + - vllm/v1/sample/
780 + - vllm/model_executor/layers/
781 + - vllm/transformers_utils/configs/speculators/
782 + - tests/v1/e2e/spec_decode/
783 - vllm/platforms/rocm.py
784 commands:
758 - - pytest -v -s kernels/mamba
759 -
785 + - pytest -v -s v1/e2e/spec_decode -k "speculators or mtp_correctness"
786
761 -- label: Kernels Helion Test # TBD
787 +- label: V1 attention (H100-MI250) # TBD
788 timeout_in_minutes: 180
789 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
790 agent_pool: mi250_1
765 - optional: true
791 working_dir: "/vllm-workspace/tests"
792 source_file_dependencies:
768 - - vllm/utils/import_utils.py
769 - - tests/kernels/helion/
793 + - vllm/config/attention.py
794 + - vllm/model_executor/layers/attention
795 + - vllm/v1/attention
796 + - tests/v1/attention
797 + - vllm/_aiter_ops.py
798 + - vllm/envs.py
799 - vllm/platforms/rocm.py
800 commands:
772 - - pip install helion==1.0.0
773 - - pytest -v -s kernels/helion/
774 -
801 + - pytest -v -s v1/attention
802
776 -- label: Model Executor # TBD
777 - timeout_in_minutes: 180
803 +- label: V1 Core + KV + Metrics # TBD
804 + timeout_in_minutes: 60
805 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
806 agent_pool: mi250_1
807 optional: true
781 - torch_nightly: true
808 working_dir: "/vllm-workspace/tests"
809 source_file_dependencies:
784 - - vllm/engine/arg_utils.py
785 - - vllm/config/model.py
786 - - vllm/model_executor
787 - - tests/model_executor
788 - - tests/entrypoints/openai/completion/test_tensorizer_entrypoint.py
789 - - vllm/_aiter_ops.py
790 - - vllm/platforms/rocm.py
791 - commands:
792 - - apt-get update && apt-get install -y curl libsodium23
793 - - export VLLM_WORKER_MULTIPROC_METHOD=spawn
794 - - pytest -v -s model_executor -m '(not slow_test)'
795 - - pytest -v -s entrypoints/openai/completion/test_tensorizer_entrypoint.py
796 -
797 -
798 -- label: Benchmarks # TBD
799 - timeout_in_minutes: 180
800 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
801 - agent_pool: mi250_1
802 - working_dir: "/vllm-workspace/.buildkite"
803 - source_file_dependencies:
804 - - benchmarks/
805 - - vllm/platforms/rocm.py
810 + - vllm/
811 + - tests/v1/core
812 + - tests/v1/executor
813 + - tests/v1/kv_offload
814 + - tests/v1/worker
815 + - tests/v1/kv_connector/unit
816 + - tests/v1/metrics
817 + - tests/entrypoints/openai/correctness/test_lmeval.py
818 commands:
807 - - bash scripts/run-benchmarks.sh
808 -
819 + - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
820 + - pytest -v -s -m 'not cpu_test' v1/core
821 + - pytest -v -s v1/executor
822 + - pytest -v -s v1/kv_offload
823 + - pytest -v -s v1/worker
824 + - pytest -v -s -m 'not cpu_test' v1/kv_connector/unit
825 + - pytest -v -s -m 'not cpu_test' v1/metrics
826 + - pip install -U git+https://github.com/robertgshaw2-redhat/lm-evaluation-harness.git@streaming-api
827 + - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
828
810 -- label: Benchmarks CLI Test # TBD
811 - timeout_in_minutes: 180
829 +- label: V1 Sample + Logits # TBD
830 + timeout_in_minutes: 60
831 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
832 agent_pool: mi250_1
833 + optional: true
834 working_dir: "/vllm-workspace/tests"
835 source_file_dependencies:
836 - vllm/
817 - - tests/benchmarks/
837 + - tests/v1/sample
838 + - tests/v1/logits_processors
839 + - tests/v1/test_oracle.py
840 + - tests/v1/test_request.py
841 + - tests/v1/test_outputs.py
842 commands:
819 - - pytest -v -s benchmarks/
820 -
843 + - pytest -v -s v1/sample
844 + - pytest -v -s v1/logits_processors
845 + - pytest -v -s v1/test_oracle.py
846 + - pytest -v -s v1/test_request.py
847 + - pytest -v -s v1/test_outputs.py
848
822 -- label: OpenAI API correctness # TBD
849 +- label: Distributed DP Tests (2 GPUs) # TBD
850 timeout_in_minutes: 180
851 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
825 - agent_pool: mi250_1
852 + agent_pool: mi250_2
853 + num_gpus: 2
854 + optional: true
855 working_dir: "/vllm-workspace/tests"
856 source_file_dependencies:
828 - - csrc/
829 - - vllm/entrypoints/openai/
830 - - vllm/model_executor/models/whisper.py
831 - - vllm/model_executor/layers/
832 - - vllm/v1/attention/backends/
833 - - vllm/v1/attention/selector.py
834 - - vllm/_aiter_ops.py
857 + - vllm/distributed/
858 + - vllm/engine/
859 + - vllm/executor/
860 + - vllm/worker/worker_base.py
861 + - vllm/v1/engine/
862 + - vllm/v1/worker/
863 + - tests/v1/distributed
864 + - tests/entrypoints/openai/test_multi_api_servers.py
865 - vllm/platforms/rocm.py
836 - - vllm/model_executor/model_loader/
866 commands:
838 - - bash ../tools/install_torchcodec_rocm.sh || exit 1
839 - - pytest -s entrypoints/openai/correctness/
867 + - export TORCH_NCCL_BLOCKING_WAIT=1
868 + - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_async_llm_dp.py
869 + - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_eagle_dp.py
870 + - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_external_lb_dp.py
871 + - DP_SIZE=2 pytest -v -s entrypoints/openai/test_multi_api_servers.py
872
873 +- label: NixlConnector PD + Spec Decode acceptance (2 GPUs) # TBD
874 + timeout_in_minutes: 180
875 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
876 + agent_pool: mi250_2
877 + num_gpus: 2
878 + working_dir: "/vllm-workspace/tests"
879 + source_file_dependencies:
880 + - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
881 + - vllm/v1/worker/kv_connector_model_runner_mixin.py
882 + - tests/v1/kv_connector/nixl_integration/
883 + - vllm/platforms/rocm.py
884 + commands:
885 + - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
886 + - ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/spec_decode_acceptance_test.sh
887
842 -- label: Basic Models Tests (Initialization) # TBD
888 +- label: V1 e2e (2 GPUs) # TBD
889 timeout_in_minutes: 180
890 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
845 - agent_pool: mi250_1
846 - torch_nightly: true
891 + agent_pool: mi250_2
892 + optional: true
893 working_dir: "/vllm-workspace/tests"
894 source_file_dependencies:
895 - vllm/
850 - - tests/models/test_initialization.py
851 - - tests/models/registry.py
896 + - tests/v1/e2e
897 commands:
853 - - pytest -v -s models/test_initialization.py::test_can_initialize_small_subset
854 -
898 + - pytest -v -s v1/e2e/spec_decode/test_spec_decode.py -k "tensor_parallelism"
899
856 -- label: Basic Models Tests (Extra Initialization) %N # TBD
900 +- label: Distributed NixlConnector PD accuracy (4 GPUs) # TBD
901 timeout_in_minutes: 180
902 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
859 - agent_pool: mi250_1
860 - torch_nightly: true
861 - parallelism: 2
903 + agent_pool: mi250_4
904 + num_gpus: 4
905 working_dir: "/vllm-workspace/tests"
906 source_file_dependencies:
864 - - vllm/model_executor/models/
865 - - vllm/model_executor/layers/
866 - - tests/models/test_initialization.py
867 - - tests/models/registry.py
868 - - vllm/_aiter_ops.py
907 + - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
908 + - tests/v1/kv_connector/nixl_integration/
909 - vllm/platforms/rocm.py
910 commands:
871 - - pytest -v -s models/test_initialization.py -k 'not test_can_initialize_small_subset' --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
911 + - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
912 + - ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
913
914 +#------------------------------------------------------------- mi250 · misc ------------------------------------------------------------#
915
874 -- label: Basic Models Tests (Other) # TBD
916 +- label: Async Engine, Inputs, Utils, Worker, Config (CPU) # TBD
917 timeout_in_minutes: 180
918 mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
919 agent_pool: mi250_1
878 - torch_nightly: true
920 + optional: true
921 + no_gpu: true
922 working_dir: "/vllm-workspace/tests"
923 source_file_dependencies:
924 - vllm/
882 - - tests/models/test_terratorch.py
883 - - tests/models/test_transformers.py
884 - - tests/models/test_registry.py
925 + - tests/test_inputs.py
926 + - tests/test_outputs.py
927 + - tests/test_pooling_params.py
928 + - tests/test_ray_env.py
929 + - tests/multimodal
930 + - tests/renderers
931 + - tests/standalone_tests/lazy_imports.py
932 + - tests/tokenizers_
933 + - tests/tool_parsers
934 + - tests/transformers_utils
935 + - tests/config
936 commands:
886 - - pytest -v -s models/test_terratorch.py models/test_transformers.py models/test_registry.py
937 + - python3 standalone_tests/lazy_imports.py
938 + - pytest -v -s test_inputs.py
939 + - pytest -v -s test_outputs.py
940 + - pytest -v -s test_pooling_params.py
941 + - pytest -v -s test_ray_env.py
942 + - pytest -v -s -m 'cpu_test' multimodal
943 + - pytest -v -s renderers
944 + - pytest -v -s tokenizers_
945 + - pytest -v -s reasoning --ignore=reasoning/test_seedoss_reasoning_parser.py --ignore=reasoning/test_glm4_moe_reasoning_parser.py --ignore=reasoning/test_gemma4_reasoning_parser.py
946 + - pytest -v -s tool_parsers
947 + - pytest -v -s transformers_utils
948 + - pytest -v -s config
949
950 +#########################################################################################################################################
951 +# #
952 +# MI300 (gfx942) tests #
953 +# #
954 +#########################################################################################################################################
955
889 -- label: Basic Models Test (Other CPU) # TBD
956 +#----------------------------------------------------- mi300 · basic_correctness -----------------------------------------------------#
957 +
958 +- label: Basic Correctness # TBD
959 timeout_in_minutes: 180
891 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
892 - agent_pool: mi250_1
893 - no_gpu: true
894 - optional: true
960 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
961 + agent_pool: mi300_1
962 + fast_check: true
963 torch_nightly: true
964 working_dir: "/vllm-workspace/tests"
965 source_file_dependencies:
966 - vllm/
899 - - tests/models/test_utils.py
900 - - tests/models/test_vision.py
967 + - tests/basic_correctness/test_basic_correctness
968 + - tests/basic_correctness/test_cpu_offload
969 + - tests/basic_correctness/test_cumem.py
970 commands:
902 - - pytest -v -s models/test_utils.py models/test_vision.py
903 -
971 + - export VLLM_WORKER_MULTIPROC_METHOD=spawn
972 + - pytest -v -s basic_correctness/test_cumem.py
973 + - pytest -v -s basic_correctness/test_basic_correctness.py
974 + - pytest -v -s basic_correctness/test_cpu_offload.py
975
905 -- label: Language Models Tests (Extra Standard) %N # TBD
976 +- label: Distributed Model Tests (2 GPUs) # TBD
977 timeout_in_minutes: 180
907 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
908 - agent_pool: mi250_1
909 - torch_nightly: true
910 - parallelism: 2
978 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
979 + agent_pool: mi300_2
980 + num_gpus: 2
981 + optional: true
982 working_dir: "/vllm-workspace/tests"
983 source_file_dependencies:
984 + - vllm/model_executor/model_loader/sharded_state_loader.py
985 - vllm/model_executor/models/
914 - - vllm/model_executor/model_loader/
986 - vllm/model_executor/layers/
987 - vllm/v1/attention/backends/
988 - vllm/v1/attention/selector.py
918 - - tests/models/language/pooling/test_embedding.py
919 - - tests/models/language/generation/test_common.py
920 - - tests/models/language/pooling/test_classification.py
989 - vllm/_aiter_ops.py
990 - vllm/platforms/rocm.py
991 + - tests/basic_correctness/
992 + - tests/model_executor/model_loader/test_sharded_state_loader.py
993 + - tests/models/
994 commands:
924 - - pip freeze | grep -E 'torch'
925 - - export TORCH_NCCL_BLOCKING_WAIT=1
926 - - pytest -v -s models/language -m 'core_model and slow_test' --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
995 + - TARGET_TEST_SUITE=L4 pytest basic_correctness/ -v -s -m 'distributed(num_gpus=2)'
996 + - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s model_executor/model_loader/test_sharded_state_loader.py -m '(not slow_test)'
997 + - pytest models/test_transformers.py -v -s -m 'distributed(num_gpus=2)'
998 + - pytest models/language -v -s -m 'distributed(num_gpus=2)'
999 + - pytest models/multimodal -v -s -m 'distributed(num_gpus=2)' --ignore models/multimodal/generation/test_whisper.py --ignore models/multimodal/generation/test_phi4siglip.py
1000 + - pytest models/multimodal/generation/test_phi4siglip.py -v -s -m 'distributed(num_gpus=2)'
1001 + - VLLM_WORKER_MULTIPROC_METHOD=spawn pytest models/multimodal/generation/test_whisper.py -v -s -m 'distributed(num_gpus=2)'
1002
1003 +#-------------------------------------------------------- mi300 · benchmarks ---------------------------------------------------------#
1004
929 -- label: Language Models Test (PPL) # TBD
1005 +- label: Benchmarks # TBD
1006 timeout_in_minutes: 180
931 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
932 - agent_pool: mi250_1
1007 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1008 + agent_pool: mi300_1
1009 + optional: true
1010 + working_dir: "/vllm-workspace/.buildkite"
1011 + source_file_dependencies:
1012 + - benchmarks/
1013 + - vllm/platforms/rocm.py
1014 + commands:
1015 + - bash scripts/run-benchmarks.sh
1016 +
1017 +- label: Benchmarks CLI Test # TBD
1018 + timeout_in_minutes: 180
1019 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1020 + agent_pool: mi300_1
1021 working_dir: "/vllm-workspace/tests"
1022 source_file_dependencies:
1023 - vllm/
936 - - tests/models/language/generation_ppl_test
1024 + - tests/benchmarks/
1025 commands:
938 - - pytest -v -s models/language/generation_ppl_test
1026 + - pytest -v -s benchmarks/
1027
1028 +#---------------------------------------------------------- mi300 · compile ----------------------------------------------------------#
1029
941 -- label: Language Models Test (Extended Pooling) # TBD
1030 +- label: Fusion E2E Config Sweep (H100-MI300) # TBD
1031 timeout_in_minutes: 180
943 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
944 - agent_pool: mi250_1
945 - working_dir: "/vllm-workspace/tests"
1032 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1033 + agent_pool: mi300_1
1034 + num_gpus: 1
1035 + working_dir: "/vllm-workspace/"
1036 source_file_dependencies:
947 - - vllm/
948 - - tests/models/language/pooling
1037 + - csrc/quantization/
1038 + - vllm/compilation/
1039 + - vllm/model_executor/layers/layernorm.py
1040 + - vllm/model_executor/layers/activation.py
1041 + - vllm/model_executor/layers/attention/attention.py
1042 + - vllm/model_executor/layers/quantization/input_quant_fp8.py
1043 + - tests/compile/fusions_e2e/
1044 + - vllm/_aiter_ops.py
1045 + - vllm/platforms/rocm.py
1046 commands:
950 - - pytest -v -s models/language/pooling -m 'not core_model'
1047 + - rocm-smi
1048 + - pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k "llama-3"
1049
1050 +- label: Fusion E2E Quick (H100-MI300) # TBD
1051 + timeout_in_minutes: 180
1052 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1053 + agent_pool: mi300_1
1054 + num_gpus: 1
1055 + working_dir: "/vllm-workspace/"
1056 + source_file_dependencies:
1057 + - csrc/quantization/
1058 + - vllm/model_executor/
1059 + - vllm/v1/attention/
1060 + - vllm/compilation/
1061 + - tests/compile/fusions_e2e/
1062 + - vllm/_aiter_ops.py
1063 + - vllm/platforms/rocm.py
1064 + commands:
1065 + - rocm-smi
1066 + # Run all models and attn backends but only Inductor partition and native custom ops
1067 + - "pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k 'inductor_partition and not +rms_norm and not +quant_fp8'"
1068 + # Different from CUDA, Qwen requires +rms_norm and +quant_fp8 as rms+quant fusion is only supported on AITER
1069 + - "pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k 'inductor_partition and +rms_norm and +quant_fp8 and qwen3'"
1070
953 -- label: Language Models Test (MTEB) # TBD
1071 +- label: PyTorch Compilation Passes Unit Tests # TBD
1072 timeout_in_minutes: 180
955 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
956 - agent_pool: mi250_1
1073 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1074 + agent_pool: mi300_1
1075 working_dir: "/vllm-workspace/tests"
1076 source_file_dependencies:
1077 - vllm/
960 - - tests/models/language/pooling_mteb_test
1078 + - tests/compile/passes
1079 commands:
962 - - pytest -v -s models/language/pooling_mteb_test
963 -
1080 + - pytest -s -v compile/passes --ignore compile/passes/distributed
1081
965 -- label: Multi-Modal Processor (CPU) # TBD
1082 +- label: Pytorch Nightly Dependency Override Check # TBD
1083 timeout_in_minutes: 180
967 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
968 - agent_pool: mi250_1
969 - no_gpu: true
1084 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1085 + agent_pool: mi300_1
1086 optional: true
1087 + soft_fail: true
1088 working_dir: "/vllm-workspace/tests"
1089 source_file_dependencies:
973 - - vllm/
974 - - tests/models/multimodal
975 - - tests/models/registry.py
1090 + - requirements/test/nightly-torch.txt
1091 + - vllm/platforms/rocm.py
1092 commands:
977 - - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
978 - - pytest -v -s models/multimodal/processing --ignore models/multimodal/processing/test_tensor_schema.py
979 -
1093 + - bash standalone_tests/pytorch_nightly_dependency.sh
1094
981 -- label: Multi-Modal Accuracy Eval (Small Models) # TBD
1095 +- label: Distributed Compile Unit Tests (2xH100-2xMI300) # TBD
1096 timeout_in_minutes: 180
983 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
984 - agent_pool: mi250_1
985 - optional: true
986 - working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
1097 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1098 + agent_pool: mi300_2
1099 + num_gpus: 2
1100 + working_dir: "/vllm-workspace/"
1101 source_file_dependencies:
988 - - vllm/multimodal/
989 - - vllm/inputs/
990 - - vllm/v1/core/
1102 + - vllm/compilation/
1103 + - vllm/model_executor/layers
1104 + - tests/compile/passes/distributed/
1105 + - vllm/_aiter_ops.py
1106 - vllm/platforms/rocm.py
992 - - vllm/model_executor/model_loader/
1107 commands:
994 - - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-mm-small.txt --tp-size=1
1108 + - export VLLM_TEST_CLEAN_GPU_MEMORY=1
1109 + - VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/compile/passes/distributed/test_async_tp.py
1110 + - pytest -v -s tests/compile/passes/distributed/test_sequence_parallelism.py
1111
1112 +#----------------------------------------------------------- mi300 · cuda ------------------------------------------------------------#
1113
997 -- label: "Multi-Modal Models (Standard) 1: qwen2" # TBD
1114 +- label: Platform Tests (CUDA) # TBD
1115 timeout_in_minutes: 180
999 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
1000 - agent_pool: mi250_1
1001 - torch_nightly: true
1116 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1117 + agent_pool: mi300_1
1118 working_dir: "/vllm-workspace/tests"
1119 source_file_dependencies:
1120 - vllm/
1005 - - tests/models/multimodal
1121 + - tests/cuda
1122 commands:
1007 - - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
1008 - - pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "qwen2"
1009 - - pytest -v -s models/multimodal/generation/test_ultravox.py -m core_model
1123 + - pytest -v -s cuda/test_cuda_context.py
1124 + - pytest -v -s cuda/test_platform_no_cuda_init.py
1125
1126 +#-------------------------------------------------------- mi300 · detokenizer --------------------------------------------------------#
1127
1012 -- label: "Multi-Modal Models (Standard) 2: qwen3 + gemma" # TBD
1128 +- label: Async Engine, Inputs, Utils, Worker # TBD
1129 timeout_in_minutes: 180
1014 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
1015 - agent_pool: mi250_1
1016 - torch_nightly: true
1017 - working_dir: "/vllm-workspace/tests"
1018 - source_file_dependencies:
1019 - - vllm/
1020 - - tests/models/multimodal
1021 - commands:
1022 - - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
1023 - - pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "qwen3 or gemma"
1024 - - pytest -v -s models/multimodal/generation/test_qwen2_5_vl.py -m core_model
1025 -
1026 -
1027 -- label: "Multi-Modal Models (Standard) 3: llava + qwen2_vl" # TBD
1028 - timeout_in_minutes: 180
1029 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
1030 - agent_pool: mi250_1
1031 - torch_nightly: true
1032 - working_dir: "/vllm-workspace/tests"
1033 - source_file_dependencies:
1034 - - vllm/
1035 - - tests/models/multimodal
1036 - commands:
1037 - - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
1038 - - pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "not qwen2 and not qwen3 and not gemma"
1039 - - pytest -v -s models/multimodal/generation/test_qwen2_vl.py -m core_model
1040 -
1041 -
1042 -- label: "Multi-Modal Models (Standard) 4: other + whisper" # TBD
1043 - timeout_in_minutes: 180
1044 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
1045 - agent_pool: mi250_1
1046 - torch_nightly: true
1047 - working_dir: "/vllm-workspace/tests"
1048 - source_file_dependencies:
1049 - - vllm/
1050 - - tests/models/multimodal/generation
1051 - - tests/models/multimodal/test_mapping.py
1052 - commands:
1053 - - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
1054 - - pytest -v -s models/multimodal -m core_model --ignore models/multimodal/generation/test_common.py --ignore models/multimodal/generation/test_ultravox.py --ignore models/multimodal/generation/test_qwen2_5_vl.py --ignore models/multimodal/generation/test_qwen2_vl.py --ignore models/multimodal/generation/test_whisper.py --ignore models/multimodal/generation/test_memory_leak.py --ignore models/multimodal/processing
1055 - - pytest -v -s models/multimodal/generation/test_memory_leak.py -m core_model
1056 - - cd .. && VLLM_WORKER_MULTIPROC_METHOD=spawn pytest -v -s tests/models/multimodal/generation/test_whisper.py -m core_model
1057 -
1058 -
1059 -- label: Multi-Modal Models (Extended Generation 1) # TBD
1060 - timeout_in_minutes: 180
1061 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
1062 - agent_pool: mi250_1
1130 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1131 + agent_pool: mi300_1
1132 working_dir: "/vllm-workspace/tests"
1133 source_file_dependencies:
1134 - vllm/
1066 - - tests/models/multimodal/generation
1067 - - tests/models/multimodal/test_mapping.py
1135 + - tests/detokenizer
1136 + - tests/multimodal
1137 + - tests/utils_
1138 commands:
1069 - - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
1070 - - pytest -v -s models/multimodal/generation -m 'not core_model' --ignore models/multimodal/generation/test_common.py
1071 - - pytest -v -s models/multimodal/test_mapping.py
1139 + - pytest -v -s detokenizer
1140 + - pytest -v -s -m 'not cpu_test' multimodal
1141 + - pytest -v -s utils_
1142
1143 +#-------------------------------------------------------- mi300 · distributed --------------------------------------------------------#
1144
1074 -- label: Multi-Modal Models (Extended Generation 2) # TBD
1145 +- label: EPLB Algorithm # TBD
1146 timeout_in_minutes: 180
1076 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
1077 - agent_pool: mi250_1
1147 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1148 + agent_pool: mi300_1
1149 working_dir: "/vllm-workspace/tests"
1150 source_file_dependencies:
1080 - - vllm/
1081 - - tests/models/multimodal/generation
1151 + - vllm/distributed/eplb
1152 + - tests/distributed/test_eplb_algo.py
1153 + - vllm/platforms/rocm.py
1154 commands:
1083 - - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
1084 - - pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=0) and not core_model'
1085 -
1155 + - pytest -v -s distributed/test_eplb_algo.py
1156 + - pytest -v -s distributed/test_eplb_utils.py
1157
1087 -- label: Multi-Modal Models (Extended Generation 3) # TBD
1158 +- label: Distributed Tests (2xH100-2xMI250) # TBD
1159 timeout_in_minutes: 180
1089 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
1090 - agent_pool: mi250_1
1091 - working_dir: "/vllm-workspace/tests"
1160 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1161 + agent_pool: mi300_2
1162 + num_gpus: 2
1163 + working_dir: "/vllm-workspace/"
1164 source_file_dependencies:
1093 - - vllm/
1094 - - tests/models/multimodal/generation
1165 + - vllm/distributed/
1166 + - vllm/v1/distributed/
1167 + - vllm/model_executor/layers/fused_moe/
1168 + - vllm/v1/attention/backends/
1169 + - vllm/v1/attention/selector.py
1170 + - tests/distributed/test_context_parallel.py
1171 + - examples/offline_inference/data_parallel.py
1172 + - vllm/_aiter_ops.py
1173 + - vllm/platforms/rocm.py
1174 commands:
1096 - - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
1097 - - pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=1) and not core_model'
1098 -
1175 + - export TORCH_NCCL_BLOCKING_WAIT=1
1176 + - pytest -v -s tests/distributed/test_context_parallel.py
1177 + - VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model=Qwen/Qwen1.5-MoE-A2.7B -tp=1 -dp=2 --max-model-len=2048 --all2all-backend=allgather_reducescatter --disable-nccl-for-dp-synchronization
1178
1100 -- label: Multi-Modal Models (Extended Pooling) # TBD
1179 +- label: Distributed Tests (4xA100-4xMI300) # TBD
1180 timeout_in_minutes: 180
1102 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
1103 - agent_pool: mi250_1
1181 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1182 + agent_pool: mi300_4
1183 + num_gpus: 4
1184 + optional: true
1185 working_dir: "/vllm-workspace/tests"
1186 source_file_dependencies:
1187 - vllm/
1107 - - tests/models/multimodal/pooling
1188 commands:
1109 - - pytest -v -s models/multimodal/pooling -m 'not core_model'
1110 -
1189 + - export TORCH_NCCL_BLOCKING_WAIT=1
1190 + - pytest -v -s distributed/test_custom_all_reduce.py
1191 + - torchrun --nproc_per_node=2 distributed/test_ca_buffer_sharing.py
1192 + - TARGET_TEST_SUITE=A100 pytest basic_correctness/ -v -s -m 'distributed(num_gpus=2)'
1193 + - pytest -v -s -x lora/test_mixtral.py
1194
1112 -- label: Distributed Comm Ops # TBD
1195 +- label: Distributed Torchrun + Examples (4 GPUs) # TBD
1196 timeout_in_minutes: 180
1114 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
1115 - agent_pool: mi250_1
1116 - num_gpus: 2
1197 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1198 + agent_pool: mi300_4
1199 + num_gpus: 4
1200 working_dir: "/vllm-workspace/tests"
1201 source_file_dependencies:
1119 - - vllm/distributed
1120 - - tests/distributed
1202 + - vllm/distributed/
1203 + - tests/distributed/test_torchrun_example.py
1204 + - tests/distributed/test_torchrun_example_moe.py
1205 + - examples/rl/
1206 + - tests/examples/offline_inference/data_parallel.py
1207 - vllm/platforms/rocm.py
1208 commands:
1123 - - pytest -v -s distributed/test_comm_ops.py
1124 - - pytest -v -s distributed/test_shm_broadcast.py
1125 - - pytest -v -s distributed/test_shm_buffer.py
1126 - - pytest -v -s distributed/test_shm_storage.py
1127 -
1209 + - export TORCH_NCCL_BLOCKING_WAIT=1
1210 + - torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
1211 + - PP_SIZE=2 torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
1212 + - TP_SIZE=4 torchrun --nproc-per-node=4 distributed/test_torchrun_example_moe.py
1213 + - PP_SIZE=2 TP_SIZE=2 torchrun --nproc-per-node=4 distributed/test_torchrun_example_moe.py
1214 + - DP_SIZE=4 ENABLE_EP=1 torchrun --nproc-per-node=4 distributed/test_torchrun_example_moe.py
1215 + - TP_SIZE=2 DP_SIZE=2 ENABLE_EP=1 torchrun --nproc-per-node=4 distributed/test_torchrun_example_moe.py
1216 + - python3 ../examples/offline_inference/data_parallel.py --enforce-eager
1217 + # rlhf examples
1218 + - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 ../examples/rl/rlhf_nccl.py
1219 + - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 ../examples/rl/rlhf_ipc.py
1220
1129 -- label: Distributed DP Tests (2 GPUs) # TBD
1221 +- label: Elastic EP Scaling Test # TBD
1222 timeout_in_minutes: 180
1131 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
1132 - agent_pool: mi250_2
1133 - num_gpus: 2
1223 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1224 + agent_pool: mi300_4
1225 + num_gpus: 4
1226 optional: true
1227 working_dir: "/vllm-workspace/tests"
1228 source_file_dependencies:
1229 - vllm/distributed/
1230 - vllm/engine/
1231 - vllm/executor/
1140 - - vllm/worker/worker_base.py
1141 - - vllm/v1/engine/
1142 - - vllm/v1/worker/
1143 - - tests/v1/distributed
1144 - - tests/entrypoints/openai/test_multi_api_servers.py
1232 + - vllm/compilation/
1233 + - tests/distributed/
1234 - vllm/platforms/rocm.py
1235 commands:
1147 - - export TORCH_NCCL_BLOCKING_WAIT=1
1148 - - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_async_llm_dp.py
1149 - - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_eagle_dp.py
1150 - - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_external_lb_dp.py
1151 - - DP_SIZE=2 pytest -v -s entrypoints/openai/test_multi_api_servers.py
1152 -
1236 + - pytest -v -s distributed/test_elastic_ep.py
1237
1154 -- label: Distributed Compile + RPC Tests (2 GPUs) # TBD
1238 +- label: RayExecutorV2 (4 GPUs) # TBD
1239 timeout_in_minutes: 180
1156 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
1157 - agent_pool: mi250_2
1158 - num_gpus: 2
1159 - optional: true
1240 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1241 + agent_pool: mi300_4
1242 + num_gpus: 4
1243 working_dir: "/vllm-workspace/tests"
1244 source_file_dependencies:
1162 - - vllm/compilation/
1163 - - vllm/distributed/
1164 - - vllm/engine/
1165 - - vllm/executor/
1166 - - vllm/worker/worker_base.py
1167 - - vllm/v1/engine/
1168 - - vllm/v1/worker/
1169 - - tests/compile/fullgraph/test_basic_correctness.py
1170 - - tests/compile/test_wrapper.py
1171 - - tests/entrypoints/llm/test_collective_rpc.py
1245 + - vllm/v1/executor/ray_executor_v2.py
1246 + - vllm/v1/executor/abstract.py
1247 + - vllm/v1/executor/multiproc_executor.py
1248 + - tests/distributed/test_ray_v2_executor.py
1249 + - tests/distributed/test_ray_v2_executor_e2e.py
1250 + - tests/distributed/test_pipeline_parallel.py
1251 + - tests/basic_correctness/test_basic_correctness.py
1252 - vllm/platforms/rocm.py
1253 commands:
1254 + - export VLLM_USE_RAY_V2_EXECUTOR_BACKEND=1
1255 - export TORCH_NCCL_BLOCKING_WAIT=1
1175 - - pytest -v -s entrypoints/llm/test_collective_rpc.py
1176 - - pytest -v -s ./compile/fullgraph/test_basic_correctness.py
1177 - - pytest -v -s ./compile/test_wrapper.py
1178 -
1256 + - pytest -v -s distributed/test_ray_v2_executor.py
1257 + - pytest -v -s distributed/test_ray_v2_executor_e2e.py
1258 + - pytest -v -s distributed/test_pipeline_parallel.py -k "ray"
1259 + - TARGET_TEST_SUITE=L4 pytest -v -s basic_correctness/test_basic_correctness.py -k "ray"
1260
1180 -- label: Distributed Torchrun + Shutdown Tests (2 GPUs) # TBD
1261 +- label: Distributed Tests (8xH100-8xMI300) # TBD
1262 timeout_in_minutes: 180
1182 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
1183 - agent_pool: mi250_2
1184 - num_gpus: 2
1263 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1264 + agent_pool: mi300_8
1265 + num_gpus: 8
1266 optional: true
1267 working_dir: "/vllm-workspace/tests"
1268 source_file_dependencies:
1269 + - examples/offline_inference/torchrun_dp_example.py
1270 + - vllm/config/parallel.py
1271 - vllm/distributed/
1189 - - vllm/engine/
1190 - - vllm/executor/
1191 - - vllm/worker/worker_base.py
1192 - - vllm/v1/engine/
1193 - - vllm/v1/worker/
1194 - - tests/distributed/
1195 - - tests/v1/shutdown
1196 - - tests/v1/worker/test_worker_memory_snapshot.py
1272 + - vllm/v1/engine/llm_engine.py
1273 + - vllm/v1/executor/uniproc_executor.py
1274 + - vllm/v1/worker/gpu_worker.py
1275 - vllm/platforms/rocm.py
1276 commands:
1277 - export TORCH_NCCL_BLOCKING_WAIT=1
1200 - - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
1201 - - VLLM_TEST_SAME_HOST=1 VLLM_TEST_WITH_DEFAULT_DEVICE_SET=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
1202 - - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s v1/shutdown
1203 - - pytest -v -s v1/worker/test_worker_memory_snapshot.py
1278 + - torchrun --nproc-per-node=8 ../examples/offline_inference/torchrun_dp_example.py --tp-size=2 --pp-size=1 --dp-size=4 --enable-ep
1279
1280 +#-------------------------------------------------------- mi300 · entrypoints --------------------------------------------------------#
1281
1206 -- label: Distributed Model Tests (2 GPUs) # TBD
1282 +- label: Entrypoints Integration (API Server 2) # TBD
1283 timeout_in_minutes: 180
1208 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
1209 - agent_pool: mi250_2
1210 - num_gpus: 2
1284 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1285 + agent_pool: mi300_1
1286 + fast_check: true
1287 + torch_nightly: true
1288 working_dir: "/vllm-workspace/tests"
1289 source_file_dependencies:
1213 - - vllm/model_executor/model_loader/sharded_state_loader.py
1214 - - vllm/model_executor/models/
1215 - - vllm/model_executor/layers/
1216 - - vllm/v1/attention/backends/
1217 - - vllm/v1/attention/selector.py
1218 - - tests/basic_correctness/
1219 - - tests/model_executor/model_loader/test_sharded_state_loader.py
1220 - - tests/models/
1221 - - vllm/_aiter_ops.py
1222 - - vllm/platforms/rocm.py
1290 + - vllm/
1291 + - tests/entrypoints/rpc
1292 + - tests/entrypoints/serve/instrumentator
1293 + - tests/tool_use
1294 commands:
1224 - - TARGET_TEST_SUITE=L4 pytest basic_correctness/ -v -s -m 'distributed(num_gpus=2)'
1225 - - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s model_executor/model_loader/test_sharded_state_loader.py -m '(not slow_test)'
1226 - - pytest models/test_transformers.py -v -s -m 'distributed(num_gpus=2)'
1227 - - pytest models/language -v -s -m 'distributed(num_gpus=2)'
1228 - - pytest models/multimodal -v -s -m 'distributed(num_gpus=2)' --ignore models/multimodal/generation/test_whisper.py
1229 - - VLLM_WORKER_MULTIPROC_METHOD=spawn pytest models/multimodal/generation/test_whisper.py -v -s -m 'distributed(num_gpus=2)'
1230 -
1295 + - export VLLM_WORKER_MULTIPROC_METHOD=spawn
1296 + - pytest -v -s entrypoints/serve/instrumentator
1297 + - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/rpc
1298 + - pytest -v -s tool_use
1299
1232 -- label: Plugin Tests (2 GPUs) # TBD
1300 +- label: Entrypoints Integration (API Server openai - Part 1) # TBD
1301 timeout_in_minutes: 180
1234 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
1235 - agent_pool: mi250_2
1236 - num_gpus: 2
1237 - optional: true
1302 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1303 + agent_pool: mi300_1
1304 + fast_check: true
1305 + torch_nightly: true
1306 working_dir: "/vllm-workspace/tests"
1307 source_file_dependencies:
1240 - - vllm/plugins/
1241 - - tests/plugins/
1242 - - vllm/platforms/rocm.py
1308 + - vllm/
1309 + - tests/entrypoints/openai
1310 + - tests/entrypoints/test_chat_utils
1311 commands:
1244 - # BEGIN: platform plugin and general plugin tests, all the code in-between runs on dummy platform
1245 - - pip install -e ./plugins/vllm_add_dummy_platform
1246 - - pytest -v -s plugins_tests/test_platform_plugins.py
1247 - - pip uninstall vllm_add_dummy_platform -y
1248 - # END: platform plugin tests
1249 - # BEGIN: `io_processor` plugins test, all the code in between uses the `prithvi_io_processor` plugin
1250 - - pip install -e ./plugins/prithvi_io_processor_plugin
1251 - - pytest -v -s plugins_tests/test_io_processor_plugins.py
1252 - - pip uninstall prithvi_io_processor_plugin -y
1253 - # END: `io_processor` plugins test
1254 - # BEGIN: `bge_m3_sparse io_processor` test
1255 - - pip install -e ./plugins/bge_m3_sparse_plugin
1256 - - pytest -v -s plugins_tests/test_bge_m3_sparse_io_processor_plugins.py
1257 - - pip uninstall bge_m3_sparse_plugin -y
1258 - # END: `bge_m3_sparse io_processor` test
1259 - # BEGIN: `stat_logger` plugins test
1260 - - pip install -e ./plugins/vllm_add_dummy_stat_logger
1261 - - pytest -v -s plugins_tests/test_stats_logger_plugins.py
1262 - - pip uninstall dummy_stat_logger -y
1263 - # END: `stat_logger` plugins test
1264 - # BEGIN: other tests
1265 - - pytest -v -s plugins_tests/test_scheduler_plugins.py
1266 - - pip install -e ./plugins/vllm_add_dummy_model
1267 - - pytest -v -s distributed/test_distributed_oot.py
1268 - - pytest -v -s entrypoints/openai/chat_completion/test_oot_registration.py
1269 - - pytest -v -s models/test_oot_registration.py
1270 - - pytest -v -s plugins/lora_resolvers
1271 - # END: other tests
1272 -
1312 + - export VLLM_WORKER_MULTIPROC_METHOD=spawn
1313 + - pytest -v -s entrypoints/openai/chat_completion --ignore=entrypoints/openai/chat_completion/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/chat_completion/test_oot_registration.py
1314
1274 -- label: Pipeline + Context Parallelism (4 GPUs) # TBD
1315 +- label: Entrypoints Integration (API Server openai - Part 2) # TBD
1316 timeout_in_minutes: 180
1276 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
1277 - agent_pool: mi250_4
1278 - num_gpus: 4
1317 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1318 + agent_pool: mi300_1
1319 + fast_check: true
1320 + torch_nightly: true
1321 working_dir: "/vllm-workspace/tests"
1322 source_file_dependencies:
1281 - - vllm/distributed/
1282 - - vllm/engine/
1283 - - vllm/executor/
1284 - - vllm/model_executor/models/
1285 - - vllm/model_executor/layers/
1286 - - vllm/v1/attention/backends/
1287 - - vllm/v1/attention/selector.py
1288 - - tests/distributed/
1289 - - vllm/_aiter_ops.py
1290 - - vllm/platforms/rocm.py
1323 + - vllm/
1324 + - tests/entrypoints/openai
1325 + - tests/entrypoints/test_chat_utils
1326 commands:
1292 - - pytest -v -s distributed/test_pp_cudagraph.py
1293 - - pytest -v -s distributed/test_pipeline_parallel.py
1294 -
1327 + - export VLLM_WORKER_MULTIPROC_METHOD=spawn
1328 + - pytest -v -s entrypoints/openai/completion --ignore=entrypoints/openai/completion/test_tensorizer_entrypoint.py
1329 + - pytest -v -s entrypoints/openai/speech_to_text/
1330 + - pytest -v -s entrypoints/test_chat_utils.py
1331
1296 -- label: Ray Dependency Compatibility Check # TBD
1332 +- label: Entrypoints Integration (API Server openai - Part 3) # TBD
1333 timeout_in_minutes: 180
1298 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
1299 - agent_pool: mi250_1
1300 - working_dir: "/"
1334 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1335 + agent_pool: mi300_1
1336 + fast_check: true
1337 + torch_nightly: true
1338 + working_dir: "/vllm-workspace/tests"
1339 source_file_dependencies:
1302 - - requirements/
1303 - - setup.py
1304 - - vllm/platforms/rocm.py
1340 + - vllm/
1341 + - tests/entrypoints/openai
1342 + - tests/entrypoints/test_chat_utils
1343 commands:
1306 - - bash /vllm-workspace/.buildkite/scripts/check-ray-compatibility.sh
1307 -
1344 + - export VLLM_WORKER_MULTIPROC_METHOD=spawn
1345 + - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/chat_completion --ignore=entrypoints/openai/completion --ignore=entrypoints/openai/speech_to_text/ --ignore=entrypoints/openai/correctness/ --ignore=entrypoints/openai/tool_parsers/ --ignore=entrypoints/openai/responses --ignore=entrypoints/openai/test_multi_api_servers.py
1346
1309 -- label: Distributed NixlConnector PD accuracy (4 GPUs) # TBD
1347 +- label: Entrypoints Integration (LLM) # TBD
1348 timeout_in_minutes: 180
1311 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
1312 - agent_pool: mi250_4
1313 - num_gpus: 4
1349 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1350 + agent_pool: mi300_1
1351 + optional: true
1352 + fast_check: true
1353 + torch_nightly: true
1354 working_dir: "/vllm-workspace/tests"
1355 source_file_dependencies:
1316 - - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
1317 - - tests/v1/kv_connector/nixl_integration/
1318 - - vllm/platforms/rocm.py
1356 + - vllm/
1357 + - tests/entrypoints/llm
1358 + - tests/entrypoints/offline_mode
1359 commands:
1320 - - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
1321 - - ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
1322 -
1360 + - export VLLM_WORKER_MULTIPROC_METHOD=spawn
1361 + - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_collective_rpc.py
1362 + - pytest -v -s entrypoints/llm/test_generate.py
1363 + - pytest -v -s entrypoints/offline_mode
1364
1324 -- label: DP EP Distributed NixlConnector PD accuracy tests (4 GPUs) # TBD
1365 +- label: Entrypoints Integration (Pooling) # TBD
1366 timeout_in_minutes: 180
1326 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
1327 - agent_pool: mi250_4
1328 - num_gpus: 4
1367 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1368 + agent_pool: mi300_1
1369 + fast_check: true
1370 + torch_nightly: true
1371 working_dir: "/vllm-workspace/tests"
1372 source_file_dependencies:
1331 - - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
1332 - - tests/v1/kv_connector/nixl_integration/
1333 - - vllm/platforms/rocm.py
1373 + - vllm/
1374 + - tests/entrypoints/pooling
1375 commands:
1335 - - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
1336 - - DP_EP=1 ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
1337 -
1376 + - export VLLM_WORKER_MULTIPROC_METHOD=spawn
1377 + - pytest -v -s entrypoints/pooling
1378
1339 -- label: NixlConnector PD + Spec Decode acceptance (2 GPUs) # TBD
1379 +- label: Entrypoints Integration (Responses API) # TBD
1380 timeout_in_minutes: 180
1341 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
1342 - agent_pool: mi250_2
1343 - num_gpus: 2
1381 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1382 + agent_pool: mi300_1
1383 + fast_check: true
1384 + torch_nightly: true
1385 working_dir: "/vllm-workspace/tests"
1386 source_file_dependencies:
1346 - - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
1347 - - vllm/v1/worker/kv_connector_model_runner_mixin.py
1348 - - tests/v1/kv_connector/nixl_integration/
1349 - - vllm/platforms/rocm.py
1387 + - vllm/
1388 + - tests/entrypoints/openai/responses
1389 commands:
1351 - - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
1352 - - ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/spec_decode_acceptance_test.sh
1353 -
1390 + - export VLLM_WORKER_MULTIPROC_METHOD=spawn
1391 + - pytest -v -s entrypoints/openai/responses
1392
1355 -- label: CrossLayer KV layout Distributed NixlConnector PD accuracy tests (4 GPUs) # TBD
1393 +- label: Entrypoints Unit Tests # TBD
1394 timeout_in_minutes: 180
1357 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
1358 - agent_pool: mi250_4
1359 - num_gpus: 4
1395 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1396 + agent_pool: mi300_1
1397 + fast_check: true
1398 working_dir: "/vllm-workspace/tests"
1399 source_file_dependencies:
1362 - - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
1363 - - tests/v1/kv_connector/nixl_integration/
1400 + - vllm/entrypoints
1401 + - tests/entrypoints/
1402 - vllm/platforms/rocm.py
1403 commands:
1366 - - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
1367 - - CROSS_LAYERS_BLOCKS=True ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
1368 -
1404 + - pytest -v -s entrypoints/openai/tool_parsers
1405 + - pytest -v -s entrypoints/ --ignore=entrypoints/llm --ignore=entrypoints/rpc --ignore=entrypoints/sleep --ignore=entrypoints/serve/instrumentator --ignore=entrypoints/openai --ignore=entrypoints/offline_mode --ignore=entrypoints/test_chat_utils.py --ignore=entrypoints/pooling
1406
1370 -- label: Hyrbid SSM NixlConnector PD accuracy tests (4 GPUs) # TBD
1407 +- label: OpenAI API correctness # TBD
1408 timeout_in_minutes: 180
1372 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
1373 - agent_pool: mi250_4
1374 - num_gpus: 4
1409 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1410 + agent_pool: mi300_1
1411 working_dir: "/vllm-workspace/tests"
1412 source_file_dependencies:
1377 - - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
1378 - - tests/v1/kv_connector/nixl_integration/
1413 + - csrc/
1414 + - vllm/entrypoints/openai/
1415 + - vllm/model_executor/models/whisper.py
1416 + - vllm/model_executor/layers/
1417 + - vllm/v1/attention/backends/
1418 + - vllm/v1/attention/selector.py
1419 + - vllm/_aiter_ops.py
1420 - vllm/platforms/rocm.py
1421 + - vllm/model_executor/model_loader/
1422 commands:
1381 - - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
1382 - - HYBRID_SSM=1 ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
1423 + - bash ../tools/install_torchcodec_rocm.sh || exit 1
1424 + - pytest -s entrypoints/openai/correctness/
1425
1426 +#----------------------------------------------------------- mi300 · evals -----------------------------------------------------------#
1427
1385 -- label: Distributed Tests (2 GPUs)(H100-MI250) # TBD
1428 +- label: DeepSeek V2-Lite Prefetch Offload Accuracy (H100-MI300) # TBD
1429 timeout_in_minutes: 180
1387 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
1388 - agent_pool: mi325_2
1389 - num_gpus: 2
1390 - working_dir: "/vllm-workspace/"
1430 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1431 + agent_pool: mi300_1
1432 + num_gpus: 1
1433 + working_dir: "/vllm-workspace"
1434 source_file_dependencies:
1392 - - vllm/distributed/
1393 - - vllm/v1/distributed/
1435 + - vllm/model_executor/models/
1436 + - vllm/model_executor/model_loader/
1437 - vllm/model_executor/layers/fused_moe/
1438 + - vllm/model_executor/layers/quantization/
1439 - vllm/v1/attention/backends/
1440 + - vllm/v1/attention/backends/mla/
1441 - vllm/v1/attention/selector.py
1397 - - tests/distributed/test_context_parallel.py
1398 - - examples/offline_inference/data_parallel.py
1442 + - .buildkite/scripts/scheduled_integration_test/
1443 - vllm/_aiter_ops.py
1444 - vllm/platforms/rocm.py
1445 commands:
1402 - - export TORCH_NCCL_BLOCKING_WAIT=1
1403 - - pytest -v -s tests/distributed/test_context_parallel.py
1404 - - VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model=Qwen/Qwen1.5-MoE-A2.7B -tp=1 -dp=2 --max-model-len=2048 --all2all-backend=allgather_reducescatter --disable-nccl-for-dp-synchronization
1405 -
1406 -
1407 -#####################################################################################################################################
1408 -# #
1409 -# gfx942 #
1410 -# #
1411 -#####################################################################################################################################
1412 -
1446 + - bash .buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_prefetch_offload.sh 0.25 200 8030
1447
1414 -- label: Entrypoints Integration (LLM) # 13.1m
1415 - timeout_in_minutes: 22
1416 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1417 - agent_pool: mi325_1
1418 - optional: true
1419 - fast_check: true
1420 - torch_nightly: true
1448 +- label: LM Eval Small Models # TBD
1449 + timeout_in_minutes: 180
1450 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1451 + agent_pool: mi300_1
1452 working_dir: "/vllm-workspace/tests"
1453 source_file_dependencies:
1423 - - vllm/
1424 - - tests/entrypoints/llm
1425 - - tests/entrypoints/offline_mode
1454 + - csrc/
1455 + - vllm/model_executor/layers/quantization
1456 + - vllm/model_executor/models/
1457 + - vllm/model_executor/model_loader/
1458 + - vllm/v1/attention/backends/
1459 + - vllm/v1/attention/selector.py
1460 + - vllm/_aiter_ops.py
1461 + - vllm/platforms/rocm.py
1462 commands:
1427 - - export VLLM_WORKER_MULTIPROC_METHOD=spawn
1428 - - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_collective_rpc.py
1429 - - pytest -v -s entrypoints/llm/test_generate.py
1430 - - pytest -v -s entrypoints/offline_mode
1431 -
1463 + - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small.txt
1464
1433 -- label: Entrypoints Integration (API Server openai - Part 1) # TBD
1465 +- label: LM Eval Small Models (MI300) # TBD
1466 timeout_in_minutes: 180
1435 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1436 - agent_pool: mi325_1
1437 - fast_check: true
1438 - torch_nightly: true
1439 - working_dir: "/vllm-workspace/tests"
1467 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1468 + agent_pool: mi300_1
1469 + working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
1470 source_file_dependencies:
1441 - - vllm/
1442 - - tests/entrypoints/openai
1443 - - tests/entrypoints/test_chat_utils
1471 + - csrc/
1472 + - vllm/model_executor/layers/quantization
1473 + - vllm/model_executor/models/
1474 + - vllm/model_executor/model_loader/
1475 + - vllm/v1/attention/backends/
1476 + - vllm/v1/attention/selector.py
1477 + - vllm/_aiter_ops.py
1478 + - vllm/platforms/rocm.py
1479 commands:
1445 - - export VLLM_WORKER_MULTIPROC_METHOD=spawn
1446 - - pytest -v -s entrypoints/openai/chat_completion --ignore=entrypoints/openai/chat_completion/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/chat_completion/test_oot_registration.py
1447 -
1480 + - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-small-rocm.txt
1481
1449 -- label: Entrypoints Integration (API Server openai - Part 2) # TBD
1482 +- label: GPQA Eval (GPT-OSS) (2xH100-2xMI300) # TBD
1483 timeout_in_minutes: 180
1451 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1452 - agent_pool: mi325_1
1453 - fast_check: true
1454 - torch_nightly: true
1484 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1485 + agent_pool: mi300_2
1486 + num_gpus: 2
1487 + optional: true
1488 working_dir: "/vllm-workspace/tests"
1489 source_file_dependencies:
1457 - - vllm/
1458 - - tests/entrypoints/openai
1459 - - tests/entrypoints/test_chat_utils
1490 + - csrc/
1491 + - vllm/model_executor/layers/quantization
1492 + - vllm/model_executor/models/
1493 + - vllm/model_executor/model_loader/
1494 + - vllm/v1/attention/backends/
1495 + - vllm/v1/attention/selector.py
1496 + - vllm/model_executor/layers/fused_moe/
1497 + - vllm/_aiter_ops.py
1498 + - vllm/platforms/rocm.py
1499 + - tests/evals/gpt_oss/
1500 commands:
1461 - - export VLLM_WORKER_MULTIPROC_METHOD=spawn
1462 - - pytest -v -s entrypoints/openai/completion --ignore=entrypoints/openai/completion/test_tensorizer_entrypoint.py
1463 - - pytest -v -s entrypoints/openai/speech_to_text/
1464 - - pytest -v -s entrypoints/test_chat_utils.py
1465 -
1501 + - uv pip install --system 'gpt-oss[eval]==0.0.5'
1502 + - pytest -s -v evals/gpt_oss/test_gpqa_correctness.py --config-list-file=configs/models-gfx942.txt
1503
1467 -- label: Entrypoints Integration (API Server openai - Part 3) # TBD
1504 +- label: LM Eval Small Models (2xB200-2xMI300) # TBD
1505 timeout_in_minutes: 180
1469 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1470 - agent_pool: mi325_1
1471 - fast_check: true
1472 - torch_nightly: true
1473 - working_dir: "/vllm-workspace/tests"
1474 - source_file_dependencies:
1475 - - vllm/
1476 - - tests/entrypoints/openai
1477 - - tests/entrypoints/test_chat_utils
1478 - commands:
1479 - - export VLLM_WORKER_MULTIPROC_METHOD=spawn
1480 - - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/chat_completion --ignore=entrypoints/openai/completion --ignore=entrypoints/openai/speech_to_text/ --ignore=entrypoints/openai/correctness/ --ignore=entrypoints/openai/tool_parsers/ --ignore=entrypoints/openai/responses
1481 -
1482 -
1483 -- label: Entrypoints Integration (API Server 2) #26.9m
1484 - timeout_in_minutes: 45
1485 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1486 - agent_pool: mi325_1
1487 - optional: true
1488 - fast_check: true
1489 - torch_nightly: true
1490 - working_dir: "/vllm-workspace/tests"
1491 - source_file_dependencies:
1492 - - vllm/
1493 - - tests/entrypoints/rpc
1494 - - tests/entrypoints/serve/instrumentator
1495 - - tests/tool_use
1496 - commands:
1497 - - export VLLM_WORKER_MULTIPROC_METHOD=spawn
1498 - - pytest -v -s entrypoints/serve/instrumentator
1499 - - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/rpc
1500 - - pytest -v -s tool_use
1501 -
1502 -
1503 -- label: Entrypoints Integration (Pooling) # 22.8m
1504 - timeout_in_minutes: 48
1505 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1506 - agent_pool: mi325_1
1507 - fast_check: true
1508 - torch_nightly: true
1509 - working_dir: "/vllm-workspace/tests"
1510 - source_file_dependencies:
1511 - - vllm/
1512 - - tests/entrypoints/pooling
1513 - commands:
1514 - - export VLLM_WORKER_MULTIPROC_METHOD=spawn
1515 - - pytest -v -s entrypoints/pooling
1516 -
1517 -
1518 -- label: Distributed Torchrun + Examples (4 GPUs) # TBD
1519 - timeout_in_minutes: 80
1520 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1521 - agent_pool: mi325_4
1522 - num_gpus: 4
1523 - working_dir: "/vllm-workspace/tests"
1524 - source_file_dependencies:
1525 - - vllm/distributed/
1526 - - tests/distributed/test_torchrun_example.py
1527 - - tests/distributed/test_torchrun_example_moe.py
1528 - - examples/rl/
1529 - - tests/examples/offline_inference/data_parallel.py
1530 - - vllm/platforms/rocm.py
1531 - commands:
1532 - - export TORCH_NCCL_BLOCKING_WAIT=1
1533 - - torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
1534 - - PP_SIZE=2 torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
1535 - - TP_SIZE=4 torchrun --nproc-per-node=4 distributed/test_torchrun_example_moe.py
1536 - - PP_SIZE=2 TP_SIZE=2 torchrun --nproc-per-node=4 distributed/test_torchrun_example_moe.py
1537 - - DP_SIZE=4 ENABLE_EP=1 torchrun --nproc-per-node=4 distributed/test_torchrun_example_moe.py
1538 - - TP_SIZE=2 DP_SIZE=2 ENABLE_EP=1 torchrun --nproc-per-node=4 distributed/test_torchrun_example_moe.py
1539 - - python3 ../examples/offline_inference/data_parallel.py --enforce-eager
1540 - # rlhf examples
1541 - - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 ../examples/rl/rlhf_nccl.py
1542 - - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 ../examples/rl/rlhf_ipc.py
1543 -
1544 -
1545 -- label: Distributed DP Tests (4 GPUs) # TBD
1546 - timeout_in_minutes: 60
1547 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1548 - agent_pool: mi325_4
1549 - num_gpus: 4
1550 - working_dir: "/vllm-workspace/tests"
1551 - source_file_dependencies:
1552 - - vllm/distributed/
1553 - - tests/v1/distributed
1554 - - tests/v1/engine/test_engine_core_client.py
1555 - - tests/distributed/test_utils
1556 - - vllm/platforms/rocm.py
1557 - commands:
1558 - - export TORCH_NCCL_BLOCKING_WAIT=1
1559 - - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/distributed/test_async_llm_dp.py
1560 - - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/distributed/test_eagle_dp.py
1561 - - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/distributed/test_external_lb_dp.py
1562 - - TP_SIZE=1 DP_SIZE=4 pytest -v -s v1/distributed/test_internal_lb_dp.py
1563 - - TP_SIZE=1 DP_SIZE=4 pytest -v -s v1/distributed/test_hybrid_lb_dp.py
1564 - - pytest -v -s v1/engine/test_engine_core_client.py::test_kv_cache_events_dp
1565 - - pytest -v -s distributed/test_utils.py
1566 -
1567 -
1568 -- label: Distributed Compile + Comm (4 GPUs) # TBD
1569 - timeout_in_minutes: 40
1570 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1571 - agent_pool: mi325_4
1572 - num_gpus: 4
1573 - working_dir: "/vllm-workspace/tests"
1574 - source_file_dependencies:
1575 - - vllm/distributed/
1576 - - tests/distributed/test_pynccl
1577 - - tests/distributed/test_events
1578 - - tests/compile/fullgraph/test_basic_correctness.py
1579 - - tests/distributed/test_symm_mem_allreduce.py
1580 - - tests/distributed/test_multiproc_executor.py
1581 - - vllm/platforms/rocm.py
1582 - commands:
1583 - - export TORCH_NCCL_BLOCKING_WAIT=1
1584 - - pytest -v -s compile/fullgraph/test_basic_correctness.py
1585 - - pytest -v -s distributed/test_pynccl.py
1586 - - pytest -v -s distributed/test_events.py
1587 - - pytest -v -s distributed/test_symm_mem_allreduce.py
1588 - - pytest -v -s distributed/test_multiproc_executor.py::test_multiproc_executor_multi_node
1589 -
1590 -
1591 -- label: Distributed Tests (8 GPUs)(H100-MI325) # 6.4m
1592 - timeout_in_minutes: 10
1593 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1594 - agent_pool: mi325_8
1595 - num_gpus: 8
1596 - optional: true
1506 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1507 + agent_pool: mi300_2
1508 working_dir: "/vllm-workspace/tests"
1509 source_file_dependencies:
1599 - - examples/offline_inference/torchrun_dp_example.py
1600 - - vllm/config/parallel.py
1601 - - vllm/distributed/
1602 - - vllm/v1/engine/llm_engine.py
1603 - - vllm/v1/executor/uniproc_executor.py
1604 - - vllm/v1/worker/gpu_worker.py
1510 + - csrc/
1511 + - vllm/model_executor/layers/quantization
1512 + - vllm/model_executor/models/
1513 + - vllm/model_executor/model_loader/
1514 + - vllm/v1/attention/backends/
1515 + - vllm/v1/attention/selector.py
1516 + - vllm/_aiter_ops.py
1517 - vllm/platforms/rocm.py
1518 commands:
1607 - - export TORCH_NCCL_BLOCKING_WAIT=1
1608 - - torchrun --nproc-per-node=8 ../examples/offline_inference/torchrun_dp_example.py --tp-size=2 --pp-size=1 --dp-size=4 --enable-ep
1609 -
1519 + - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-mi3xx-fp8-and-mixed.txt
1520
1611 -- label: Elastic EP Scaling Test # TBD
1521 +- label: DeepSeek V2-Lite Accuracy (4xH100-4xMI300) # TBD
1522 timeout_in_minutes: 180
1613 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1614 - agent_pool: mi325_4
1523 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1524 + agent_pool: mi300_4
1525 num_gpus: 4
1526 optional: true
1617 - working_dir: "/vllm-workspace/tests"
1618 - source_file_dependencies:
1619 - - vllm/distributed/
1620 - - vllm/engine/
1621 - - vllm/executor/
1622 - - vllm/compilation/
1623 - - tests/distributed/
1624 - - vllm/platforms/rocm.py
1625 - commands:
1626 - - pytest -v -s distributed/test_elastic_ep.py
1627 -
1628 -
1629 -- label: Engine # 11.3m
1630 - timeout_in_minutes: 35
1631 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1632 - agent_pool: mi325_1
1633 - working_dir: "/vllm-workspace/tests"
1634 - source_file_dependencies:
1635 - - vllm/
1636 - - tests/engine
1637 - - tests/test_sequence
1638 - - tests/test_config
1639 - - tests/test_logger
1640 - - tests/test_vllm_port
1641 - commands:
1642 - - pytest -v -s engine test_sequence.py test_config.py test_logger.py test_vllm_port.py
1643 -
1644 -
1645 -- label: Engine (1 GPU) # TBD
1646 - timeout_in_minutes: 180
1647 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1648 - agent_pool: mi325_1
1649 - optional: true
1650 - working_dir: "/vllm-workspace/tests"
1651 - source_file_dependencies:
1652 - - vllm/v1/engine/
1653 - - tests/v1/engine/
1654 - - vllm/platforms/rocm.py
1655 - commands:
1656 - - pytest -v -s v1/engine/test_preprocess_error_handling.py
1657 - - pytest -v -s v1/engine --ignore v1/engine/test_preprocess_error_handling.py
1658 -
1659 -
1660 -- label: e2e Scheduling (1 GPU) # TBD
1661 - timeout_in_minutes: 180
1662 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1663 - agent_pool: mi325_1
1664 - optional: true
1665 - working_dir: "/vllm-workspace/tests"
1666 - source_file_dependencies:
1667 - - vllm/v1/
1668 - - tests/v1/e2e/general/
1669 - - vllm/platforms/rocm.py
1670 - commands:
1671 - - pytest -v -s v1/e2e/general/test_async_scheduling.py
1672 -
1673 -
1674 -- label: e2e Core (1 GPU) # TBD
1675 - timeout_in_minutes: 180
1676 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1677 - agent_pool: mi325_1
1678 - optional: true
1679 - working_dir: "/vllm-workspace/tests"
1680 - source_file_dependencies:
1681 - - vllm/v1/
1682 - - tests/v1/e2e/
1683 - - vllm/platforms/rocm.py
1684 - commands:
1685 - - pytest -v -s v1/e2e/general --ignore v1/e2e/general/test_async_scheduling.py
1686 -
1687 -
1688 -- label: Spec Decode Eagle # TBD
1689 - timeout_in_minutes: 180
1690 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1691 - agent_pool: mi325_1
1692 - working_dir: "/vllm-workspace/tests"
1527 + working_dir: "/vllm-workspace"
1528 source_file_dependencies:
1694 - - vllm/v1/spec_decode/
1695 - - vllm/v1/worker/gpu/spec_decode/
1529 + - vllm/model_executor/models/
1530 - vllm/model_executor/model_loader/
1697 - - vllm/v1/sample/
1698 - - vllm/model_executor/layers/
1699 - - tests/v1/e2e/spec_decode/
1531 + - vllm/distributed/eplb
1532 + - vllm/model_executor/layers/fused_moe/
1533 + - vllm/model_executor/layers/quantization/
1534 + - vllm/v1/attention/backends/
1535 + - vllm/v1/attention/backends/mla/
1536 + - vllm/v1/attention/selector.py
1537 + - .buildkite/scripts/scheduled_integration_test/
1538 + - vllm/_aiter_ops.py
1539 - vllm/platforms/rocm.py
1540 commands:
1702 - - pytest -v -s v1/e2e/spec_decode -k "eagle_correctness"
1703 -
1541 + - bash .buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_ep_eplb.sh 0.25 200 8010
1542
1705 -- label: Spec Decode Speculators + MTP # TBD
1543 +- label: LM Eval Large Models (4xA100-4xMI300) # TBD
1544 timeout_in_minutes: 180
1707 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1708 - agent_pool: mi325_1
1545 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1546 + agent_pool: mi300_4
1547 + num_gpus: 4
1548 optional: true
1710 - working_dir: "/vllm-workspace/tests"
1549 + working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
1550 source_file_dependencies:
1712 - - vllm/v1/spec_decode/
1713 - - vllm/v1/worker/gpu/spec_decode/
1551 + - csrc/
1552 + - vllm/model_executor/layers/quantization
1553 + - vllm/model_executor/models/
1554 - vllm/model_executor/model_loader/
1715 - - vllm/v1/sample/
1716 - - vllm/model_executor/layers/
1717 - - vllm/transformers_utils/configs/speculators/
1718 - - tests/v1/e2e/spec_decode/
1555 + - vllm/v1/attention/backends/
1556 + - vllm/v1/attention/selector.py
1557 + - vllm/_aiter_ops.py
1558 - vllm/platforms/rocm.py
1559 commands:
1721 - - pytest -v -s v1/e2e/spec_decode -k "speculators or mtp_correctness"
1722 -
1560 + - export VLLM_WORKER_MULTIPROC_METHOD=spawn
1561 + - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large.txt --tp-size=4
1562
1724 -- label: Spec Decode Ngram + Suffix # TBD
1563 +- label: Qwen3-30B-A3B-FP8-block Accuracy (4xH100-4xMI300) # TBD
1564 timeout_in_minutes: 180
1726 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1727 - agent_pool: mi325_1
1565 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1566 + agent_pool: mi300_4
1567 optional: true
1729 - working_dir: "/vllm-workspace/tests"
1568 + working_dir: "/vllm-workspace"
1569 source_file_dependencies:
1731 - - vllm/v1/spec_decode/
1732 - - vllm/v1/worker/gpu/spec_decode/
1570 + - vllm/model_executor/models/
1571 - vllm/model_executor/model_loader/
1734 - - vllm/v1/sample/
1735 - - vllm/model_executor/layers/
1736 - - tests/v1/e2e/spec_decode/
1572 + - vllm/model_executor/layers/quantization/
1573 + - vllm/distributed/eplb
1574 + - vllm/model_executor/layers/fused_moe/
1575 + - vllm/v1/attention/backends/
1576 + - vllm/v1/attention/selector.py
1577 + - .buildkite/scripts/scheduled_integration_test/
1578 + - vllm/_aiter_ops.py
1579 - vllm/platforms/rocm.py
1580 commands:
1739 - - pytest -v -s v1/e2e/spec_decode -k "ngram or suffix"
1740 -
1581 + - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh 0.8 200 8020
1582
1742 -- label: Spec Decode Draft Model # TBD
1583 +- label: Qwen3-Next-80B-A3B-Instruct MTP Async EPLB Accuracy # TBD
1584 timeout_in_minutes: 180
1744 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1745 - agent_pool: mi325_1
1585 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1586 + agent_pool: mi300_4
1587 + num_gpus: 4
1588 optional: true
1747 - working_dir: "/vllm-workspace/tests"
1589 + working_dir: "/vllm-workspace"
1590 source_file_dependencies:
1749 - - vllm/v1/spec_decode/
1750 - - vllm/v1/worker/gpu/spec_decode/
1591 + - vllm/model_executor/models/
1592 - vllm/model_executor/model_loader/
1752 - - vllm/v1/sample/
1753 - - vllm/model_executor/layers/
1754 - - tests/v1/e2e/spec_decode/
1755 - - vllm/platforms/rocm.py
1756 - commands:
1757 - - pytest -v -s v1/e2e/spec_decode -k "draft_model or no_sync or batch_inference"
1758 -
1759 -
1760 -- label: V1 e2e (2 GPUs) # 7.1m
1761 - timeout_in_minutes: 12
1762 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1763 - agent_pool: mi325_2
1764 - optional: true
1765 - working_dir: "/vllm-workspace/tests"
1766 - source_file_dependencies:
1767 - - vllm/
1768 - - tests/v1/e2e
1769 - commands:
1770 - - pytest -v -s v1/e2e/spec_decode/test_spec_decode.py -k "tensor_parallelism"
1771 -
1772 -
1773 -- label: V1 e2e (4 GPUs) # 52.6m
1774 - timeout_in_minutes: 106
1775 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1776 - agent_pool: mi325_4
1777 - optional: true
1778 - working_dir: "/vllm-workspace/tests"
1779 - source_file_dependencies:
1780 - - vllm/
1781 - - tests/v1/e2e
1782 - commands:
1783 - - pytest -v -s v1/e2e/spec_decode/test_spec_decode.py -k "eagle_correctness_heavy"
1784 -
1785 -
1786 -- label: V1 e2e (4xH100-4xMI325) # TBD
1787 - timeout_in_minutes: 180
1788 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1789 - agent_pool: mi325_4
1790 - optional: true
1791 - source_file_dependencies:
1792 - - vllm/v1/attention/backends/utils.py
1793 - - vllm/v1/worker/gpu_model_runner.py
1794 - - tests/v1/e2e/test_hybrid_chunked_prefill.py
1795 - commands:
1796 - - pytest -v -s v1/e2e/test_hybrid_chunked_prefill.py
1797 -
1798 -
1799 -- label: V1 Spec Decode # TBD
1800 - timeout_in_minutes: 40
1801 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1802 - agent_pool: mi325_1
1803 - working_dir: "/vllm-workspace/tests"
1804 - source_file_dependencies:
1805 - - vllm/
1806 - - tests/v1/spec_decode
1807 - commands:
1808 - - pytest -v -s -m 'not slow_test' v1/spec_decode
1809 -
1810 -
1811 -- label: V1 Sample + Logits # TBD
1812 - timeout_in_minutes: 40
1813 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1814 - agent_pool: mi325_1
1815 - optional: true
1816 - working_dir: "/vllm-workspace/tests"
1817 - source_file_dependencies:
1818 - - vllm/
1819 - - tests/v1/sample
1820 - - tests/v1/logits_processors
1821 - - tests/v1/test_oracle.py
1822 - - tests/v1/test_request.py
1823 - - tests/v1/test_outputs.py
1824 - commands:
1825 - - pytest -v -s v1/sample
1826 - - pytest -v -s v1/logits_processors
1827 - - pytest -v -s v1/test_oracle.py
1828 - - pytest -v -s v1/test_request.py
1829 - - pytest -v -s v1/test_outputs.py
1830 -
1831 -
1832 -- label: V1 Core + KV + Metrics # TBD
1833 - timeout_in_minutes: 40
1834 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1835 - agent_pool: mi325_1
1836 - optional: true
1837 - working_dir: "/vllm-workspace/tests"
1838 - source_file_dependencies:
1839 - - vllm/
1840 - - tests/v1/core
1841 - - tests/v1/executor
1842 - - tests/v1/kv_offload
1843 - - tests/v1/worker
1844 - - tests/v1/kv_connector/unit
1845 - - tests/v1/metrics
1846 - - tests/entrypoints/openai/correctness/test_lmeval.py
1847 - commands:
1848 - - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
1849 - - pytest -v -s -m 'not cpu_test' v1/core
1850 - - pytest -v -s v1/executor
1851 - - pytest -v -s v1/kv_offload
1852 - - pytest -v -s v1/worker
1853 - - pytest -v -s -m 'not cpu_test' v1/kv_connector/unit
1854 - - pytest -v -s -m 'not cpu_test' v1/metrics
1855 - - pip install -U git+https://github.com/robertgshaw2-redhat/lm-evaluation-harness.git@streaming-api
1856 - # - export HSA_NO_SCRATCH_RECLAIM=1
1857 - - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
1858 -
1859 -
1860 -
1861 -- label: Acceptance Length Test (Large Models) # TBD
1862 - timeout_in_minutes: 180
1863 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1864 - agent_pool: mi325_1
1865 - optional: true
1866 - working_dir: "/vllm-workspace/tests"
1867 - source_file_dependencies:
1593 - vllm/v1/spec_decode/
1869 - - vllm/model_executor/models/mlp_speculator.py
1870 - - tests/v1/spec_decode/test_acceptance_length.py
1871 - - vllm/platforms/rocm.py
1872 - commands:
1873 - - export VLLM_ALLOW_INSECURE_SERIALIZATION=1
1874 - - pytest -v -s v1/spec_decode/test_acceptance_length.py
1875 -
1876 -
1877 -- label: V1 attention (H100-MI325) # 14.5m
1878 - timeout_in_minutes: 40
1879 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1880 - agent_pool: mi325_1
1881 - working_dir: "/vllm-workspace/tests"
1882 - source_file_dependencies:
1883 - - vllm/config/attention.py
1884 - - vllm/model_executor/layers/attention
1885 - - vllm/v1/attention
1886 - - tests/v1/attention
1887 - - vllm/_aiter_ops.py
1888 - - vllm/envs.py
1889 - - vllm/platforms/rocm.py
1890 - commands:
1891 - - pytest -v -s v1/attention
1892 -
1893 -
1894 -- label: Batch Invariance (H100-MI325) # 5.2m
1895 - timeout_in_minutes: 12
1896 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1897 - agent_pool: mi325_1
1898 - optional: true
1899 - working_dir: "/vllm-workspace/tests"
1900 - source_file_dependencies:
1901 - - vllm/v1/attention
1902 - - vllm/model_executor/layers
1903 - - tests/v1/determinism/
1594 + - vllm/distributed/eplb
1595 + - vllm/model_executor/layers/fused_moe/
1596 + - vllm/model_executor/layers/quantization/
1597 + - vllm/v1/attention/backends/
1598 + - vllm/v1/attention/selector.py
1599 + - .buildkite/scripts/scheduled_integration_test/
1600 - vllm/_aiter_ops.py
1601 - vllm/platforms/rocm.py
1602 commands:
1907 - - export VLLM_WORKER_MULTIPROC_METHOD=spawn
1908 - - pip install pytest-timeout pytest-forked
1909 - - pytest -v -s v1/determinism/test_batch_invariance.py
1910 - - pytest -v -s v1/determinism/test_rms_norm_batch_invariant.py
1911 -
1603 + - bash .buildkite/scripts/scheduled_integration_test/qwen3_next_mtp_async_eplb.sh 0.8 1319 8040
1604
1913 -- label: V1 others (CPU) # 10.4m
1914 - timeout_in_minutes: 28
1915 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1916 - agent_pool: mi325_1
1917 - no_gpu: true
1605 +- label: LM Eval Large Models (8xH200-8xMI300) # TBD
1606 + timeout_in_minutes: 180
1607 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1608 + agent_pool: mi300_8
1609 optional: true
1610 + num_gpus: 8
1611 working_dir: "/vllm-workspace/tests"
1612 source_file_dependencies:
1921 - - vllm/
1922 - - tests/v1
1613 + - vllm/model_executor/models/
1614 + - vllm/model_executor/model_loader/
1615 + - vllm/model_executor/layers/quantization/
1616 + - vllm/v1/attention/backends/
1617 + - vllm/v1/attention/selector.py
1618 + - vllm/model_executor/layers/layernorm.py
1619 + - csrc/
1620 + - vllm/_aiter_ops.py
1621 + - vllm/platforms/rocm.py
1622 + - tests/evals/
1623 commands:
1924 - - pytest -v -s -m 'cpu_test' v1/core
1925 - - pytest -v -s v1/structured_output
1926 - - pytest -v -s v1/test_serial_utils.py
1927 - - pytest -v -s -m 'cpu_test' v1/kv_connector/unit
1928 - - pytest -v -s -m 'cpu_test' v1/metrics
1624 + - export VLLM_WORKER_MULTIPROC_METHOD=spawn
1625 + - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-mi3xx.txt
1626
1627 +#--------------------------------------------------------- mi300 · examples ----------------------------------------------------------#
1628
1931 -- label: Examples # 24.5m
1932 - timeout_in_minutes: 55
1933 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1934 - agent_pool: mi325_1
1629 +- label: Examples # TBD
1630 + timeout_in_minutes: 180
1631 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1632 + agent_pool: mi300_1
1633 optional: true
1634 working_dir: "/vllm-workspace/examples"
1635 source_file_dependencies:
@@ -1962,54 +1660,12 @@ steps:
1660 - python3 offline_inference/spec_decode.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
1661 - python3 offline_inference/spec_decode.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
1662
1663 +#---------------------------------------------------------- mi300 · kernels ----------------------------------------------------------#
1664
1966 -- label: Platform Tests (CUDA) # 5.0m
1967 - timeout_in_minutes: 9
1968 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1969 - agent_pool: mi325_1
1970 - optional: true
1971 - working_dir: "/vllm-workspace/tests"
1972 - source_file_dependencies:
1973 - - vllm/
1974 - - tests/cuda
1975 - commands:
1976 - - pytest -v -s cuda/test_cuda_context.py
1977 - - pytest -v -s cuda/test_platform_no_cuda_init.py
1978 -
1979 -
1980 -- label: PyTorch Compilation Passes Unit Tests # TBD
1665 +- label: Kernels Attention Test %N # TBD
1666 timeout_in_minutes: 180
1982 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1983 - agent_pool: mi325_1
1984 - working_dir: "/vllm-workspace/tests"
1985 - source_file_dependencies:
1986 - - vllm/
1987 - - tests/compile/passes
1988 - commands:
1989 - - pytest -s -v compile/passes --ignore compile/passes/distributed
1990 -
1991 -
1992 -- label: Kernels Core Operation Test # 26.8m
1993 - timeout_in_minutes: 38
1994 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1995 - agent_pool: mi325_1
1996 - working_dir: "/vllm-workspace/tests"
1997 - source_file_dependencies:
1998 - - csrc/
1999 - - tests/kernels/core
2000 - - tests/kernels/test_top_k_per_row.py
2001 - - tests/kernels/test_concat_mla_q.py
2002 - - vllm/model_executor/layers/rotary_embedding/
2003 - - vllm/_aiter_ops.py
2004 - - vllm/platforms/rocm.py
2005 - commands:
2006 - - pytest -v -s kernels/core kernels/test_top_k_per_row.py
2007 -
2008 -
2009 -- label: Kernels Attention Test %N # 17.7m
2010 - timeout_in_minutes: 28
2011 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2012 - agent_pool: mi325_1
1667 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1668 + agent_pool: mi300_1
1669 parallelism: 2
1670 working_dir: "/vllm-workspace/tests"
1671 source_file_dependencies:
@@ -2023,29 +1679,26 @@ steps:
1679 commands:
1680 - pytest -v -s kernels/attention --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
1681
2026 -
2027 -- label: Kernels Quantization Test %N # 15.2m
2028 - timeout_in_minutes: 24
2029 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2030 - agent_pool: mi325_1
2031 - parallelism: 2
1682 +- label: Kernels Core Operation Test # TBD
1683 + timeout_in_minutes: 180
1684 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1685 + agent_pool: mi300_1
1686 working_dir: "/vllm-workspace/tests"
1687 source_file_dependencies:
2034 - - csrc/quantization/
2035 - - vllm/model_executor/layers/quantization
2036 - - tests/kernels/quantization
2037 - - tests/kernels/quantization/test_rocm_skinny_gemms.py
1688 + - csrc/
1689 + - tests/kernels/core
1690 + - tests/kernels/test_top_k_per_row.py
1691 + - tests/kernels/test_concat_mla_q.py
1692 + - vllm/model_executor/layers/rotary_embedding/
1693 - vllm/_aiter_ops.py
1694 - vllm/platforms/rocm.py
2040 - - vllm/model_executor/kernels/
1695 commands:
2042 - - pytest -v -s kernels/quantization --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
2043 -
1696 + - pytest -v -s kernels/core --ignore=kernels/core/test_minimax_reduce_rms.py kernels/test_concat_mla_q.py kernels/test_top_k_per_row.py
1697
1698 - label: Kernels MoE Test %N # TBD
2046 - timeout_in_minutes: 19
2047 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2048 - agent_pool: mi325_1
1699 + timeout_in_minutes: 180
1700 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1701 + agent_pool: mi300_1
1702 parallelism: 4
1703 working_dir: "/vllm-workspace/tests"
1704 source_file_dependencies:
@@ -2062,11 +1715,27 @@ steps:
1715 - pytest -v -s kernels/moe --ignore=kernels/moe/test_modular_oai_triton_moe.py --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
1716 - pytest -v -s kernels/moe/test_modular_oai_triton_moe.py --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
1717
1718 +- label: Kernels Quantization Test %N # TBD
1719 + timeout_in_minutes: 180
1720 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1721 + agent_pool: mi300_1
1722 + parallelism: 2
1723 + working_dir: "/vllm-workspace/tests"
1724 + source_file_dependencies:
1725 + - csrc/quantization/
1726 + - vllm/model_executor/layers/quantization
1727 + - tests/kernels/quantization
1728 + - tests/kernels/quantization/test_rocm_skinny_gemms.py
1729 + - vllm/_aiter_ops.py
1730 + - vllm/platforms/rocm.py
1731 + - vllm/model_executor/kernels/
1732 + commands:
1733 + - pytest -v -s kernels/quantization --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
1734
2066 -- label: Kernels FP8 MoE Test # TBD
1735 +- label: Kernels FP8 MoE Test (2xH100-2xMI300) # TBD
1736 timeout_in_minutes: 180
2068 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2069 - agent_pool: mi325_2
1737 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1738 + agent_pool: mi300_2
1739 optional: true
1740 working_dir: "/vllm-workspace/tests"
1741 source_file_dependencies:
@@ -2080,40 +1749,217 @@ steps:
1749 commands:
1750 - pytest -v -s kernels/moe/test_deepep_moe.py
1751
1752 +#----------------------------------------------------------- mi300 · lora ------------------------------------------------------------#
1753
2084 -- label: ROCm AITER Ops Test # TBD
1754 +- label: LoRA TP (Distributed) # TBD
1755 timeout_in_minutes: 180
2086 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2087 - agent_pool: mi325_1
1756 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1757 + agent_pool: mi300_4
1758 + num_gpus: 4
1759 working_dir: "/vllm-workspace/tests"
1760 source_file_dependencies:
2090 - - vllm/_aiter_ops.py
2091 - - vllm/envs.py
1761 + - vllm/lora
1762 + - tests/lora
1763 - vllm/platforms/rocm.py
2093 - - tests/rocm/aiter/
2094 - - vllm/v1/attention/backends/mla/rocm_aiter_mla.py
2095 - - vllm/v1/attention/selector.py
1764 commands:
2097 - - pytest -v -s rocm/aiter/
1765 + - export VLLM_WORKER_MULTIPROC_METHOD=spawn
1766 + - export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
1767 + - pytest -v -s -x lora/test_chatglm3_tp.py
1768 + - pytest -v -s -x lora/test_llama_tp.py
1769 + - pytest -v -s -x lora/test_llm_with_multi_loras.py
1770 + - pytest -v -s -x lora/test_olmoe_tp.py
1771 + - pytest -v -s -x lora/test_gptoss_tp.py
1772 + - pytest -v -s -x lora/test_qwen35_densemodel_lora.py
1773
1774 +#---------------------------------------------------------- mi300 · models -----------------------------------------------------------#
1775
2100 -- label: Benchmarks # 8.2m
2101 - timeout_in_minutes: 20
2102 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2103 - agent_pool: mi325_1
1776 +- label: Language Models Test (Extended Pooling) # TBD
1777 + timeout_in_minutes: 180
1778 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1779 + agent_pool: mi300_1
1780 + working_dir: "/vllm-workspace/tests"
1781 + source_file_dependencies:
1782 + - vllm/
1783 + - tests/models/language/pooling
1784 + commands:
1785 + - pytest -v -s models/language/pooling -m 'not core_model'
1786 +
1787 +- label: Language Models Tests (Standard) # TBD
1788 + timeout_in_minutes: 180
1789 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1790 + agent_pool: mi300_1
1791 optional: true
2105 - working_dir: "/vllm-workspace/.buildkite"
1792 + torch_nightly: true
1793 + working_dir: "/vllm-workspace/tests"
1794 source_file_dependencies:
2107 - - benchmarks/
1795 + - vllm/
1796 + - tests/models/language
1797 + commands:
1798 + - pip freeze | grep -E 'torch'
1799 + - pytest -v -s models/language -m 'core_model and (not slow_test)'
1800 +
1801 +- label: Multi-Modal Models (Extended Generation 1) # TBD
1802 + timeout_in_minutes: 180
1803 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1804 + agent_pool: mi300_1
1805 + working_dir: "/vllm-workspace/tests"
1806 + source_file_dependencies:
1807 + - vllm/
1808 + - tests/models/multimodal/generation
1809 + - tests/models/multimodal/test_mapping.py
1810 + commands:
1811 + - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
1812 + - pytest -v -s models/multimodal/generation -m 'not core_model' --ignore models/multimodal/generation/test_common.py
1813 + - pytest -v -s models/multimodal/test_mapping.py
1814 +
1815 +- label: Multi-Modal Models (Extended Generation 2) # TBD
1816 + timeout_in_minutes: 180
1817 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1818 + agent_pool: mi300_1
1819 + optional: true
1820 + working_dir: "/vllm-workspace/tests"
1821 + source_file_dependencies:
1822 + - vllm/
1823 + - tests/models/multimodal/generation
1824 + commands:
1825 + - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
1826 + - pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=0) and not core_model'
1827 +
1828 +- label: Multi-Modal Models (Extended Generation 3) # TBD
1829 + timeout_in_minutes: 180
1830 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1831 + agent_pool: mi300_1
1832 + optional: true
1833 + working_dir: "/vllm-workspace/tests"
1834 + source_file_dependencies:
1835 + - vllm/
1836 + - tests/models/multimodal/generation
1837 + commands:
1838 + - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
1839 + - pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=1) and not core_model'
1840 +
1841 +- label: "Multi-Modal Models (Standard) 1: qwen2" # TBD
1842 + timeout_in_minutes: 180
1843 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1844 + agent_pool: mi300_1
1845 + torch_nightly: true
1846 + working_dir: "/vllm-workspace/tests"
1847 + source_file_dependencies:
1848 + - vllm/
1849 + - tests/models/multimodal
1850 + commands:
1851 + - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
1852 + - pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "qwen2"
1853 + - pytest -v -s models/multimodal/generation/test_ultravox.py -m core_model
1854 +
1855 +- label: "Multi-Modal Models (Standard) 3: llava + qwen2_vl" # TBD
1856 + timeout_in_minutes: 180
1857 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1858 + agent_pool: mi300_1
1859 + torch_nightly: true
1860 + optional: true
1861 + working_dir: "/vllm-workspace/tests"
1862 + source_file_dependencies:
1863 + - vllm/
1864 + - tests/models/multimodal/generation
1865 + - tests/models/multimodal/test_mapping.py
1866 + commands:
1867 + - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
1868 + - pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "not qwen2 and not qwen3 and not gemma"
1869 + - pytest -v -s models/multimodal/generation/test_qwen2_vl.py -m core_model
1870 +
1871 +- label: "Multi-Modal Models (Standard) 4: other + whisper" # TBD
1872 + timeout_in_minutes: 180
1873 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1874 + agent_pool: mi300_1
1875 + torch_nightly: true
1876 + working_dir: "/vllm-workspace/tests"
1877 + source_file_dependencies:
1878 + - vllm/
1879 + - tests/models/multimodal/generation
1880 + - tests/models/multimodal/test_mapping.py
1881 + commands:
1882 + - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
1883 + - pytest -v -s models/multimodal -m core_model --ignore models/multimodal/generation/test_common.py --ignore models/multimodal/generation/test_ultravox.py --ignore models/multimodal/generation/test_qwen2_5_vl.py --ignore models/multimodal/generation/test_qwen2_vl.py --ignore models/multimodal/generation/test_whisper.py --ignore models/multimodal/generation/test_memory_leak.py --ignore models/multimodal/processing
1884 + - pytest -v -s models/multimodal/generation/test_memory_leak.py -m core_model
1885 + - cd .. && VLLM_WORKER_MULTIPROC_METHOD=spawn pytest -v -s tests/models/multimodal/generation/test_whisper.py -m core_model
1886 +
1887 +- label: Multi-Modal Processor # 1h 42m
1888 + timeout_in_minutes: 138
1889 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1890 + agent_pool: mi300_1
1891 + optional: true
1892 + working_dir: "/vllm-workspace/tests"
1893 + source_file_dependencies:
1894 + - vllm/
1895 + - tests/models/multimodal
1896 + - tests/models/registry.py
1897 + commands:
1898 + - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
1899 + - pytest -v -s models/multimodal/processing/test_tensor_schema.py
1900 +
1901 +- label: Multi-Modal Processor (CPU) # TBD
1902 + timeout_in_minutes: 180
1903 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1904 + agent_pool: mi300_1
1905 + no_gpu: true
1906 + optional: true
1907 + working_dir: "/vllm-workspace/tests"
1908 + source_file_dependencies:
1909 + - vllm/
1910 + - tests/models/multimodal
1911 + - tests/models/registry.py
1912 + commands:
1913 + - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
1914 + - pytest -v -s models/multimodal/processing --ignore models/multimodal/processing/test_tensor_schema.py
1915 +
1916 +- label: Quantized Models Test # TBD
1917 + timeout_in_minutes: 180
1918 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1919 + agent_pool: mi300_1
1920 + working_dir: "/vllm-workspace/tests"
1921 + source_file_dependencies:
1922 + - vllm/model_executor/layers/quantization
1923 + - vllm/_aiter_ops.py
1924 + - vllm/platforms/rocm.py
1925 + - tests/models/quantization
1926 + - vllm/model_executor/model_loader/
1927 + commands:
1928 + - pytest -v -s models/quantization
1929 +
1930 +- label: Transformers Nightly Models # TBD
1931 + timeout_in_minutes: 180
1932 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1933 + agent_pool: mi300_1
1934 + optional: true
1935 + working_dir: "/vllm-workspace/"
1936 + source_file_dependencies:
1937 + - vllm/model_executor/models/
1938 + - vllm/model_executor/model_loader/
1939 + - vllm/multimodal/
1940 + - vllm/model_executor/layers/
1941 + - vllm/v1/attention/backends/
1942 + - vllm/v1/attention/selector.py
1943 + - vllm/_aiter_ops.py
1944 - vllm/platforms/rocm.py
1945 + - tests/models/
1946 + - examples/
1947 commands:
2110 - - bash scripts/run-benchmarks.sh
1948 + - pip install --upgrade git+https://github.com/huggingface/transformers
1949 + - pytest -v -s tests/models/test_initialization.py
1950 + - pytest -v -s tests/models/test_transformers.py
1951 + - pytest -v -s tests/models/multimodal/processing/
1952 + - pytest -v -s tests/models/multimodal/test_mapping.py
1953 + - python3 examples/basic/offline_inference/chat.py
1954 + - python3 examples/offline_inference/vision_language.py --model-type qwen2_5_vl
1955 + - VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/offline_inference/audio_language.py --model-type whisper
1956
1957 +#------------------------------------------------------- mi300 · quantization --------------------------------------------------------#
1958
2113 -- label: Quantization # 36.1m
2114 - timeout_in_minutes: 60
2115 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2116 - agent_pool: mi325_1
1959 +- label: Quantization # TBD
1960 + timeout_in_minutes: 180
1961 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1962 + agent_pool: mi300_1
1963 working_dir: "/vllm-workspace/tests"
1964 source_file_dependencies:
1965 - csrc/
@@ -2134,262 +1980,293 @@ steps:
1980 - uv pip install --system conch-triton-kernels
1981 - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py
1982
1983 +#----------------------------------------------------------- mi300 · rocm ------------------------------------------------------------#
1984
2138 -- label: Language Models Tests (Standard) # 22.8m
2139 - timeout_in_minutes: 38
2140 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2141 - agent_pool: mi325_1
2142 - optional: true
2143 - torch_nightly: true
1985 +- label: ROCm AITER Ops Test # TBD
1986 + timeout_in_minutes: 180
1987 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1988 + agent_pool: mi300_1
1989 working_dir: "/vllm-workspace/tests"
1990 source_file_dependencies:
2146 - - vllm/
2147 - - tests/models/language
1991 + - vllm/_aiter_ops.py
1992 + - vllm/envs.py
1993 + - vllm/platforms/rocm.py
1994 + - tests/rocm/aiter/
1995 + - vllm/v1/attention/backends/mla/rocm_aiter_mla.py
1996 + - vllm/v1/attention/selector.py
1997 commands:
2149 - - pip freeze | grep -E 'torch'
2150 - - pytest -v -s models/language -m 'core_model and (not slow_test)'
1998 + - pytest -v -s rocm/aiter/
1999
2000 +#--------------------------------------------------------- mi300 · samplers ----------------------------------------------------------#
2001
2153 -- label: Language Models Tests (Hybrid) %N # 34.9m
2154 - timeout_in_minutes: 55
2155 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2156 - agent_pool: mi325_1
2157 - torch_nightly: true
2158 - parallelism: 2
2002 +- label: Samplers Test # TBD
2003 + timeout_in_minutes: 180
2004 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
2005 + agent_pool: mi300_1
2006 + optional: true
2007 working_dir: "/vllm-workspace/tests"
2008 source_file_dependencies:
2161 - - vllm/
2162 - - tests/models/language/generation
2009 + - vllm/model_executor/layers
2010 + - vllm/sampling_metadata.py
2011 + - vllm/v1/sample/
2012 + - vllm/beam_search.py
2013 + - tests/samplers
2014 + - tests/conftest.py
2015 + - vllm/_aiter_ops.py
2016 + - vllm/platforms/rocm.py
2017 commands:
2164 - - uv pip install --system --no-build-isolation 'git+https://github.com/AndreasKaratzas/mamba@fix-rocm-7.0-warp-size-constexpr'
2165 - - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.2'
2166 - - pytest -v -s models/language/generation -m hybrid_model --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
2018 + - pytest -v -s samplers
2019
2020 +#------------------------------------------------------------ mi300 · misc ------------------------------------------------------------#
2021
2169 -- label: Language Models Test (Extended Generation) # 32.2m
2170 - timeout_in_minutes: 55
2171 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2172 - agent_pool: mi325_1
2022 +- label: Python-only Installation # TBD
2023 + timeout_in_minutes: 180
2024 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
2025 + agent_pool: mi300_1
2026 + optional: true
2027 working_dir: "/vllm-workspace/tests"
2028 source_file_dependencies:
2175 - - vllm/
2176 - - tests/models/language/generation
2029 + - tests/standalone_tests/python_only_compile.sh
2030 + - setup.py
2031 + - vllm/platforms/rocm.py
2032 commands:
2178 - - uv pip install --system --no-build-isolation 'git+https://github.com/AndreasKaratzas/mamba@fix-rocm-7.0-warp-size-constexpr'
2179 - - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.2'
2180 - - pytest -v -s models/language/generation -m '(not core_model) and (not hybrid_model)'
2181 -
2033 + - bash standalone_tests/python_only_compile.sh
2034
2183 -- label: Multi-Modal Processor # 1h 42m
2184 - timeout_in_minutes: 138
2185 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2186 - agent_pool: mi325_1
2187 - optional: true
2035 +- label: Regression # TBD
2036 + timeout_in_minutes: 180
2037 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
2038 + agent_pool: mi300_1
2039 working_dir: "/vllm-workspace/tests"
2040 source_file_dependencies:
2041 - vllm/
2191 - - tests/models/multimodal
2192 - - tests/models/registry.py
2042 + - tests/test_regression
2043 commands:
2194 - - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
2195 - - pytest -v -s models/multimodal/processing/test_tensor_schema.py
2044 + - pip install modelscope
2045 + - pytest -v -s test_regression.py
2046
2047 +#--------------------------------------------------------- mi300 · ray_compat ---------------------------------------------------------#
2048
2198 -- label: "Multi-Modal Models (Standard) 1: qwen2" # TBD
2049 +- label: Ray Dependency Compatibility Check # TBD
2050 timeout_in_minutes: 180
2200 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2201 - agent_pool: mi325_1
2202 - torch_nightly: true
2203 - optional: true
2204 - working_dir: "/vllm-workspace/tests"
2051 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
2052 + agent_pool: mi300_1
2053 + working_dir: "/"
2054 source_file_dependencies:
2206 - - vllm/
2207 - - tests/models/multimodal
2055 + - requirements/
2056 + - setup.py
2057 + - vllm/platforms/rocm.py
2058 commands:
2209 - - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
2210 - - pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "qwen2"
2211 - - pytest -v -s models/multimodal/generation/test_ultravox.py -m core_model
2059 + - bash /vllm-workspace/.buildkite/scripts/check-ray-compatibility.sh
2060
2061 +#------------------------------------------------------------ mi300 · v1 -------------------------------------------------------------#
2062
2214 -- label: "Multi-Modal Models (Standard) 2: qwen3 + gemma" # TBD
2063 +- label: Acceptance Length Test (Large Models) # TBD
2064 timeout_in_minutes: 180
2216 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2217 - agent_pool: mi325_1
2218 - torch_nightly: true
2065 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
2066 + agent_pool: mi300_1
2067 + optional: true
2068 working_dir: "/vllm-workspace/tests"
2069 source_file_dependencies:
2221 - - vllm/
2222 - - tests/models/multimodal
2070 + - vllm/v1/spec_decode/
2071 + - vllm/model_executor/models/mlp_speculator.py
2072 + - tests/v1/spec_decode/test_acceptance_length.py
2073 + - vllm/platforms/rocm.py
2074 commands:
2224 - - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
2225 - - pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "qwen3 or gemma"
2226 - - pytest -v -s models/multimodal/generation/test_qwen2_5_vl.py -m core_model
2075 + - export VLLM_ALLOW_INSECURE_SERIALIZATION=1
2076 + - pytest -v -s v1/spec_decode/test_acceptance_length.py -m slow_test
2077
2078 +- label: e2e Core (1 GPU) # TBD
2079 + timeout_in_minutes: 180
2080 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
2081 + agent_pool: mi300_1
2082 + optional: true
2083 + working_dir: "/vllm-workspace/tests"
2084 + source_file_dependencies:
2085 + - vllm/v1/
2086 + - tests/v1/e2e/
2087 + - vllm/platforms/rocm.py
2088 + commands:
2089 + - pytest -v -s v1/e2e/general --ignore v1/e2e/general/test_async_scheduling.py
2090
2229 -- label: "Multi-Modal Models (Standard) 3: llava + qwen2_vl" # TBD
2091 +- label: e2e Scheduling (1 GPU) # TBD
2092 timeout_in_minutes: 180
2231 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2232 - agent_pool: mi325_1
2233 - torch_nightly: true
2093 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
2094 + agent_pool: mi300_1
2095 optional: true
2096 working_dir: "/vllm-workspace/tests"
2097 source_file_dependencies:
2237 - - vllm/
2238 - - tests/models/multimodal/generation
2239 - - tests/models/multimodal/test_mapping.py
2098 + - vllm/v1/
2099 + - tests/v1/e2e/general/
2100 + - vllm/platforms/rocm.py
2101 commands:
2241 - - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
2242 - - pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "not qwen2 and not qwen3 and not gemma"
2243 - - pytest -v -s models/multimodal/generation/test_qwen2_vl.py -m core_model
2102 + - pytest -v -s v1/e2e/general/test_async_scheduling.py
2103
2104 +- label: Engine (1 GPU) # TBD
2105 + timeout_in_minutes: 180
2106 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
2107 + agent_pool: mi300_1
2108 + optional: true
2109 + working_dir: "/vllm-workspace/tests"
2110 + source_file_dependencies:
2111 + - vllm/v1/engine/
2112 + - tests/v1/engine/
2113 + - vllm/platforms/rocm.py
2114 + commands:
2115 + - pytest -v -s v1/engine/test_preprocess_error_handling.py
2116 + - pytest -v -s v1/engine --ignore v1/engine/test_preprocess_error_handling.py
2117
2246 -- label: "Multi-Modal Models (Standard) 4: other + whisper" # TBD
2118 +- label: Spec Decode Draft Model # TBD
2119 timeout_in_minutes: 180
2248 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2249 - agent_pool: mi325_1
2250 - torch_nightly: true
2120 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
2121 + agent_pool: mi300_1
2122 optional: true
2123 working_dir: "/vllm-workspace/tests"
2124 source_file_dependencies:
2254 - - vllm/
2255 - - tests/models/multimodal/generation
2125 + - vllm/v1/spec_decode/
2126 + - vllm/v1/worker/gpu/spec_decode/
2127 + - vllm/model_executor/model_loader/
2128 + - vllm/v1/sample/
2129 + - vllm/model_executor/layers/
2130 + - tests/v1/e2e/spec_decode/
2131 + - vllm/platforms/rocm.py
2132 commands:
2257 - - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
2258 - - pytest -v -s models/multimodal -m core_model --ignore models/multimodal/generation/test_common.py --ignore models/multimodal/generation/test_ultravox.py --ignore models/multimodal/generation/test_qwen2_5_vl.py --ignore models/multimodal/generation/test_qwen2_vl.py --ignore models/multimodal/generation/test_whisper.py --ignore models/multimodal/generation/test_memory_leak.py --ignore models/multimodal/processing
2259 - - pytest -v -s models/multimodal/generation/test_memory_leak.py -m core_model
2260 - - cd .. && VLLM_WORKER_MULTIPROC_METHOD=spawn pytest -v -s tests/models/multimodal/generation/test_whisper.py -m core_model
2133 + - pytest -v -s v1/e2e/spec_decode -k "draft_model or no_sync or batch_inference"
2134
2135 +- label: Spec Decode Eagle # TBD
2136 + timeout_in_minutes: 180
2137 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
2138 + agent_pool: mi300_1
2139 + working_dir: "/vllm-workspace/tests"
2140 + source_file_dependencies:
2141 + - vllm/v1/spec_decode/
2142 + - vllm/v1/worker/gpu/spec_decode/
2143 + - vllm/model_executor/model_loader/
2144 + - vllm/v1/sample/
2145 + - vllm/model_executor/layers/
2146 + - tests/v1/e2e/spec_decode/
2147 + - vllm/platforms/rocm.py
2148 + commands:
2149 + - pytest -v -s v1/e2e/spec_decode -k "eagle_correctness"
2150
2263 -- label: Multi-Modal Models (Extended Generation 1) # 1h 2m
2264 - timeout_in_minutes: 106
2265 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2266 - agent_pool: mi325_1
2151 +- label: Spec Decode Ngram + Suffix # TBD
2152 + timeout_in_minutes: 180
2153 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
2154 + agent_pool: mi300_1
2155 optional: true
2156 working_dir: "/vllm-workspace/tests"
2157 source_file_dependencies:
2270 - - vllm/
2271 - - tests/models/multimodal/generation
2272 - - tests/models/multimodal/test_mapping.py
2158 + - vllm/v1/spec_decode/
2159 + - vllm/v1/worker/gpu/spec_decode/
2160 + - vllm/model_executor/model_loader/
2161 + - vllm/v1/sample/
2162 + - vllm/model_executor/layers/
2163 + - tests/v1/e2e/spec_decode/
2164 + - vllm/platforms/rocm.py
2165 commands:
2274 - - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
2275 - - pytest -v -s models/multimodal/generation -m 'not core_model' --ignore models/multimodal/generation/test_common.py
2276 - - pytest -v -s models/multimodal/test_mapping.py
2277 -
2166 + - pytest -v -s v1/e2e/spec_decode -k "ngram or suffix"
2167
2279 -- label: Multi-Modal Models (Extended Generation 2) # TBD
2168 +- label: Spec Decode Speculators + MTP # TBD
2169 timeout_in_minutes: 180
2281 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2282 - agent_pool: mi325_1
2170 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
2171 + agent_pool: mi300_1
2172 optional: true
2173 working_dir: "/vllm-workspace/tests"
2174 source_file_dependencies:
2286 - - vllm/
2287 - - tests/models/multimodal/generation
2175 + - vllm/v1/spec_decode/
2176 + - vllm/v1/worker/gpu/spec_decode/
2177 + - vllm/model_executor/model_loader/
2178 + - vllm/v1/sample/
2179 + - vllm/model_executor/layers/
2180 + - vllm/transformers_utils/configs/speculators/
2181 + - tests/v1/e2e/spec_decode/
2182 + - vllm/platforms/rocm.py
2183 commands:
2289 - - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
2290 - - pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=0) and not core_model'
2184 + - pytest -v -s v1/e2e/spec_decode -k "speculators or mtp_correctness"
2185
2186 +- label: V1 attention (H100-MI300) # TBD
2187 + timeout_in_minutes: 180
2188 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
2189 + agent_pool: mi300_1
2190 + working_dir: "/vllm-workspace/tests"
2191 + source_file_dependencies:
2192 + - vllm/config/attention.py
2193 + - vllm/model_executor/layers/attention
2194 + - vllm/v1/attention
2195 + - tests/v1/attention
2196 + - vllm/_aiter_ops.py
2197 + - vllm/envs.py
2198 + - vllm/platforms/rocm.py
2199 + commands:
2200 + - pytest -v -s v1/attention
2201
2293 -- label: Multi-Modal Models (Extended Generation 3) # TBD
2202 +- label: V1 Core + KV + Metrics # TBD
2203 timeout_in_minutes: 180
2295 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2296 - agent_pool: mi325_1
2204 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
2205 + agent_pool: mi300_1
2206 optional: true
2207 working_dir: "/vllm-workspace/tests"
2208 source_file_dependencies:
2209 - vllm/
2301 - - tests/models/multimodal/generation
2210 + - tests/v1/core
2211 + - tests/v1/executor
2212 + - tests/v1/kv_offload
2213 + - tests/v1/worker
2214 + - tests/v1/kv_connector/unit
2215 + - tests/v1/metrics
2216 + - tests/entrypoints/openai/correctness/test_lmeval.py
2217 commands:
2303 - - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
2304 - - pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=1) and not core_model'
2305 -
2218 + - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
2219 + - pytest -v -s -m 'not cpu_test' v1/core
2220 + - pytest -v -s v1/executor
2221 + - pytest -v -s v1/kv_offload
2222 + - pytest -v -s v1/worker
2223 + - pytest -v -s -m 'not cpu_test' v1/kv_connector/unit
2224 + - pytest -v -s -m 'not cpu_test' v1/metrics
2225 + - pip install -U git+https://github.com/robertgshaw2-redhat/lm-evaluation-harness.git@streaming-api
2226 + # - export HSA_NO_SCRATCH_RECLAIM=1
2227 + - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
2228
2307 -- label: Multi-Modal Models (Extended Pooling) # TBD
2229 +- label: V1 others (CPU) # TBD
2230 timeout_in_minutes: 180
2309 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2310 - agent_pool: mi325_1
2231 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
2232 + agent_pool: mi300_1
2233 + no_gpu: true
2234 + optional: true
2235 working_dir: "/vllm-workspace/tests"
2236 source_file_dependencies:
2237 - vllm/
2314 - - tests/models/multimodal/pooling
2315 - commands:
2316 - - pytest -v -s models/multimodal/pooling -m 'not core_model'
2317 -
2318 -
2319 -- label: Quantized Models Test # 21.4m
2320 - timeout_in_minutes: 38
2321 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2322 - agent_pool: mi325_1
2323 - working_dir: "/vllm-workspace/tests"
2324 - source_file_dependencies:
2325 - - vllm/model_executor/layers/quantization
2326 - - vllm/_aiter_ops.py
2327 - - vllm/platforms/rocm.py
2328 - - tests/models/quantization
2329 - - vllm/model_executor/model_loader/
2238 + - tests/v1
2239 commands:
2331 - - pytest -v -s models/quantization
2332 -
2240 + - pytest -v -s -m 'cpu_test' v1/core
2241 + - pytest -v -s v1/structured_output
2242 + - pytest -v -s v1/test_serial_utils.py
2243 + - pytest -v -s -m 'cpu_test' v1/kv_connector/unit
2244 + - pytest -v -s -m 'cpu_test' v1/metrics
2245
2334 -- label: Transformers Nightly Models # 50.9m
2335 - timeout_in_minutes: 102
2336 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2337 - agent_pool: mi325_1
2246 +- label: V1 Sample + Logits # TBD
2247 + timeout_in_minutes: 180
2248 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
2249 + agent_pool: mi300_1
2250 optional: true
2339 - working_dir: "/vllm-workspace/"
2251 + working_dir: "/vllm-workspace/tests"
2252 source_file_dependencies:
2341 - - vllm/model_executor/models/
2342 - - vllm/model_executor/model_loader/
2343 - - vllm/multimodal/
2344 - - vllm/model_executor/layers/
2345 - - vllm/v1/attention/backends/
2346 - - vllm/v1/attention/selector.py
2347 - - vllm/_aiter_ops.py
2348 - - vllm/platforms/rocm.py
2349 - - tests/models/
2350 - - examples/
2253 + - vllm/
2254 + - tests/v1/sample
2255 + - tests/v1/logits_processors
2256 + - tests/v1/test_oracle.py
2257 + - tests/v1/test_request.py
2258 + - tests/v1/test_outputs.py
2259 commands:
2352 - - pip install --upgrade git+https://github.com/huggingface/transformers
2353 - - pytest -v -s tests/models/test_initialization.py
2354 - - pytest -v -s tests/models/test_transformers.py
2355 - - pytest -v -s tests/models/multimodal/processing/
2356 - - pytest -v -s tests/models/multimodal/test_mapping.py
2357 - - python3 examples/basic/offline_inference/chat.py
2358 - - python3 examples/offline_inference/vision_language.py --model-type qwen2_5_vl
2359 - - VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/offline_inference/audio_language.py --model-type whisper
2360 -
2260 + - pytest -v -s v1/sample
2261 + - pytest -v -s v1/logits_processors
2262 + - pytest -v -s v1/test_oracle.py
2263 + - pytest -v -s v1/test_request.py
2264 + - pytest -v -s v1/test_outputs.py
2265
2362 -- label: Quantized MoE Test (B200-MI325) # TBD
2266 +- label: Distributed DP Tests (2 GPUs) # TBD
2267 timeout_in_minutes: 180
2364 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2365 - agent_pool: mi325_1
2366 - working_dir: "/vllm-workspace/"
2367 - source_file_dependencies:
2368 - - tests/quantization/test_gfx3xx_moe.py
2369 - - vllm/model_executor/models/deepseek_v2.py
2370 - - vllm/model_executor/models/gpt_oss.py
2371 - - vllm/model_executor/models/llama4.py
2372 - - vllm/model_executor/layers/fused_moe
2373 - - vllm/model_executor/layers/quantization/compressed_tensors
2374 - - vllm/model_executor/layers/quantization/modelopt.py
2375 - - vllm/model_executor/layers/quantization/mxfp4.py
2376 - - vllm/v1/attention/backends/triton_attn.py
2377 - - vllm/v1/attention/backends/rocm_attn.py
2378 - - vllm/v1/attention/backends/rocm_aiter_fa.py
2379 - - vllm/v1/attention/backends/mla/
2380 - - vllm/v1/attention/selector.py
2381 - - vllm/model_executor/layers/layernorm.py
2382 - - vllm/_aiter_ops.py
2383 - - vllm/platforms/rocm.py
2384 - - vllm/model_executor/model_loader/
2385 - commands:
2386 - - pytest -s -v tests/quantization/test_gfx3xx_moe.py
2387 -
2388 -
2389 -- label: Distributed DP Tests (2 GPUs) # 56.1m
2390 - timeout_in_minutes: 102
2391 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2392 - agent_pool: mi325_2
2268 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
2269 + agent_pool: mi300_2
2270 num_gpus: 2
2271 working_dir: "/vllm-workspace/tests"
2272 source_file_dependencies:
@@ -2409,315 +2286,287 @@ steps:
2286 - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_external_lb_dp.py
2287 - DP_SIZE=2 pytest -v -s entrypoints/openai/test_multi_api_servers.py
2288
2412 -
2413 -- label: Distributed Compile + RPC Tests (2 GPUs) # 56.1m
2414 - timeout_in_minutes: 102
2415 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2416 - agent_pool: mi325_2
2289 +- label: Distributed Tests (2xH100-2xMI300) # TBD
2290 + timeout_in_minutes: 180
2291 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
2292 + agent_pool: mi300_2
2293 num_gpus: 2
2418 - working_dir: "/vllm-workspace/tests"
2294 + working_dir: "/vllm-workspace/"
2295 source_file_dependencies:
2420 - - vllm/compilation/
2296 - vllm/distributed/
2422 - - vllm/engine/
2423 - - vllm/executor/
2424 - - vllm/worker/worker_base.py
2425 - - vllm/v1/engine/
2426 - - vllm/v1/worker/
2427 - - tests/compile/fullgraph/test_basic_correctness.py
2428 - - tests/compile/test_wrapper.py
2429 - - tests/entrypoints/llm/test_collective_rpc.py
2297 + - vllm/v1/distributed/
2298 + - vllm/model_executor/layers/fused_moe/
2299 + - tests/v1/distributed/test_dbo.py
2300 + - vllm/_aiter_ops.py
2301 - vllm/platforms/rocm.py
2302 commands:
2303 - export TORCH_NCCL_BLOCKING_WAIT=1
2433 - - pytest -v -s entrypoints/llm/test_collective_rpc.py
2434 - - pytest -v -s ./compile/fullgraph/test_basic_correctness.py
2435 - - pytest -v -s ./compile/test_wrapper.py
2436 -
2304 + - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 examples/rl/rlhf_async_new_apis.py
2305 + - VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model=Qwen/Qwen1.5-MoE-A2.7B -tp=1 -dp=2 --max-model-len=2048 --all2all-backend=deepep_high_throughput
2306 + - pytest -v -s tests/v1/distributed/test_dbo.py
2307 + - VLLM_ALLOW_INSECURE_SERIALIZATION=1 pytest -v -s tests/distributed/test_weight_transfer.py
2308 + - pytest -v -s tests/distributed/test_packed_tensor.py
2309
2438 -- label: Distributed Torchrun + Shutdown Tests (2 GPUs) # 56.1m
2439 - timeout_in_minutes: 102
2440 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2441 - agent_pool: mi325_2
2310 +- label: Metrics, Tracing (2 GPUs) # TBD
2311 + timeout_in_minutes: 180
2312 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
2313 + agent_pool: mi300_2
2314 num_gpus: 2
2315 working_dir: "/vllm-workspace/tests"
2316 source_file_dependencies:
2445 - - vllm/distributed/
2446 - - vllm/engine/
2447 - - vllm/executor/
2448 - - vllm/worker/worker_base.py
2449 - - vllm/v1/engine/
2450 - - vllm/v1/worker/
2451 - - tests/distributed/
2452 - - tests/v1/shutdown
2453 - - tests/v1/worker/test_worker_memory_snapshot.py
2454 - - vllm/platforms/rocm.py
2317 + - vllm/
2318 + - tests/v1/tracing
2319 commands:
2456 - - export TORCH_NCCL_BLOCKING_WAIT=1
2457 - - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
2458 - - VLLM_TEST_SAME_HOST=1 VLLM_TEST_WITH_DEFAULT_DEVICE_SET=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
2459 - - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s v1/shutdown
2460 - - pytest -v -s v1/worker/test_worker_memory_snapshot.py
2461 -
2320 + - "pip install \
2321 + 'opentelemetry-sdk>=1.26.0' \
2322 + 'opentelemetry-api>=1.26.0' \
2323 + 'opentelemetry-exporter-otlp>=1.26.0' \
2324 + 'opentelemetry-semantic-conventions-ai>=0.4.1'"
2325 + - pytest -v -s v1/tracing
2326
2463 -- label: Distributed Model Tests (2 GPUs) # 19.3m
2464 - timeout_in_minutes: 38
2465 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2466 - agent_pool: mi325_2
2467 - num_gpus: 2
2327 +- label: V1 e2e (2 GPUs) # TBD
2328 + timeout_in_minutes: 180
2329 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
2330 + agent_pool: mi300_2
2331 optional: true
2332 working_dir: "/vllm-workspace/tests"
2333 source_file_dependencies:
2471 - - vllm/model_executor/model_loader/sharded_state_loader.py
2472 - - vllm/model_executor/models/
2473 - - vllm/model_executor/layers/
2474 - - vllm/v1/attention/backends/
2475 - - vllm/v1/attention/selector.py
2476 - - vllm/_aiter_ops.py
2477 - - vllm/platforms/rocm.py
2478 - - tests/basic_correctness/
2479 - - tests/model_executor/model_loader/test_sharded_state_loader.py
2480 - - tests/models/
2334 + - vllm/
2335 + - tests/v1/e2e
2336 commands:
2482 - - TARGET_TEST_SUITE=L4 pytest basic_correctness/ -v -s -m 'distributed(num_gpus=2)'
2483 - - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s model_executor/model_loader/test_sharded_state_loader.py -m '(not slow_test)'
2484 - - pytest models/test_transformers.py -v -s -m 'distributed(num_gpus=2)'
2485 - - pytest models/language -v -s -m 'distributed(num_gpus=2)'
2486 - - pytest models/multimodal -v -s -m 'distributed(num_gpus=2)' --ignore models/multimodal/generation/test_whisper.py
2487 - - VLLM_WORKER_MULTIPROC_METHOD=spawn pytest models/multimodal/generation/test_whisper.py -v -s -m 'distributed(num_gpus=2)'
2488 -
2337 + - pytest -v -s v1/e2e/spec_decode/test_spec_decode.py -k "tensor_parallelism"
2338
2490 -- label: LoRA TP (Distributed) # 9.8m
2491 - timeout_in_minutes: 18
2492 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2493 - agent_pool: mi325_4
2339 +- label: CrossLayer KV layout Distributed NixlConnector PD accuracy tests (4 GPUs) # TBD
2340 + timeout_in_minutes: 180
2341 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
2342 + agent_pool: mi300_4
2343 num_gpus: 4
2344 working_dir: "/vllm-workspace/tests"
2345 source_file_dependencies:
2497 - - vllm/lora
2498 - - tests/lora
2346 + - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
2347 + - tests/v1/kv_connector/nixl_integration/
2348 - vllm/platforms/rocm.py
2349 commands:
2501 - - export VLLM_WORKER_MULTIPROC_METHOD=spawn
2502 - - export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
2503 - - pytest -v -s -x lora/test_chatglm3_tp.py
2504 - - pytest -v -s -x lora/test_llama_tp.py
2505 - - pytest -v -s -x lora/test_llm_with_multi_loras.py
2506 - - pytest -v -s -x lora/test_olmoe_tp.py
2507 - - pytest -v -s -x lora/test_gptoss_tp.py
2508 - - pytest -v -s -x lora/test_qwen35_densemodel_lora.py
2509 -
2350 + - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
2351 + - CROSS_LAYERS_BLOCKS=True ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
2352
2511 -- label: Weight Loading Multiple GPU # 7.5m
2512 - timeout_in_minutes: 14
2513 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2514 - agent_pool: mi325_2
2515 - num_gpus: 2
2353 +- label: Distributed DP Tests (4 GPUs) # TBD
2354 + timeout_in_minutes: 180
2355 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
2356 + agent_pool: mi300_4
2357 + num_gpus: 4
2358 working_dir: "/vllm-workspace/tests"
2359 source_file_dependencies:
2518 - - vllm/
2519 - - tests/weight_loading
2360 + - vllm/distributed/
2361 + - tests/v1/distributed
2362 + - tests/v1/engine/test_engine_core_client.py
2363 + - tests/distributed/test_utils
2364 + - vllm/platforms/rocm.py
2365 commands:
2521 - - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models-amd.txt
2522 -
2366 + - export TORCH_NCCL_BLOCKING_WAIT=1
2367 + - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/distributed/test_async_llm_dp.py
2368 + - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/distributed/test_eagle_dp.py
2369 + - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/distributed/test_external_lb_dp.py
2370 + - TP_SIZE=1 DP_SIZE=4 pytest -v -s v1/distributed/test_internal_lb_dp.py
2371 + - TP_SIZE=1 DP_SIZE=4 pytest -v -s v1/distributed/test_hybrid_lb_dp.py
2372 + - pytest -v -s v1/engine/test_engine_core_client.py::test_kv_cache_events_dp
2373 + - pytest -v -s distributed/test_utils.py
2374
2524 -- label: Weight Loading Multiple GPU - Large Models # 12.6m
2525 - timeout_in_minutes: 26
2526 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2527 - agent_pool: mi325_2
2528 - num_gpus: 2
2375 +- label: Distributed NixlConnector PD accuracy (4 GPUs) # TBD
2376 + timeout_in_minutes: 180
2377 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
2378 + agent_pool: mi300_4
2379 + num_gpus: 4
2380 optional: true
2381 working_dir: "/vllm-workspace/tests"
2382 source_file_dependencies:
2532 - - vllm/
2533 - - tests/weight_loading
2383 + - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
2384 + - tests/v1/kv_connector/nixl_integration/
2385 + - vllm/platforms/rocm.py
2386 commands:
2535 - - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models-large-amd.txt
2536 -
2387 + - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
2388 + - ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
2389
2538 -- label: Ray Dependency Compatibility Check # TBD
2390 +- label: DP EP Distributed NixlConnector PD accuracy tests (4 GPUs) # TBD
2391 timeout_in_minutes: 180
2540 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2541 - agent_pool: mi325_1
2542 - optional: true
2543 - working_dir: "/"
2392 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
2393 + agent_pool: mi300_4
2394 + num_gpus: 4
2395 + working_dir: "/vllm-workspace/tests"
2396 source_file_dependencies:
2545 - - requirements/
2546 - - setup.py
2397 + - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
2398 + - tests/v1/kv_connector/nixl_integration/
2399 - vllm/platforms/rocm.py
2400 commands:
2549 - - bash /vllm-workspace/.buildkite/scripts/check-ray-compatibility.sh
2550 -
2401 + - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
2402 + - DP_EP=1 ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
2403
2552 -- label: Distributed NixlConnector PD accuracy (4 GPUs) # 27.4m
2553 - timeout_in_minutes: 44
2554 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2555 - agent_pool: mi325_4
2404 +- label: Hyrbid SSM NixlConnector PD accuracy tests (4 GPUs) # TBD
2405 + timeout_in_minutes: 180
2406 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
2407 + agent_pool: mi300_4
2408 num_gpus: 4
2557 - optional: true
2409 working_dir: "/vllm-workspace/tests"
2410 source_file_dependencies:
2411 - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
2412 - tests/v1/kv_connector/nixl_integration/
2413 - vllm/platforms/rocm.py
2414 commands:
2564 - - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
2565 - - ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
2415 + - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
2416 + - HYBRID_SSM=1 ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
2417 +
2418 +- label: V1 e2e (4 GPUs) # TBD
2419 + timeout_in_minutes: 180
2420 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
2421 + agent_pool: mi300_4
2422 + optional: true
2423 + working_dir: "/vllm-workspace/tests"
2424 + source_file_dependencies:
2425 + - vllm/
2426 + - tests/v1/e2e
2427 + commands:
2428 + - pytest -v -s v1/e2e/spec_decode/test_spec_decode.py -k "eagle_correctness_heavy"
2429 +
2430 +- label: V1 e2e (4xH100-4xMI300) # TBD
2431 + timeout_in_minutes: 180
2432 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
2433 + agent_pool: mi300_4
2434 + optional: true
2435 + source_file_dependencies:
2436 + - vllm/v1/attention/backends/utils.py
2437 + - vllm/v1/worker/gpu_model_runner.py
2438 + - tests/v1/e2e/test_hybrid_chunked_prefill.py
2439 + commands:
2440 + - pytest -v -s v1/e2e/test_hybrid_chunked_prefill.py
2441
2442 +#------------------------------------------------------ mi300 · weight_loading -------------------------------------------------------#
2443
2568 -- label: CrossLayer KV layout Distributed NixlConnector PD accuracy tests (4 GPUs) # TBD
2444 +- label: Weight Loading Multiple GPU # TBD
2445 timeout_in_minutes: 180
2570 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2571 - agent_pool: mi325_4
2572 - num_gpus: 4
2573 - optional: true
2446 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
2447 + agent_pool: mi300_2
2448 + num_gpus: 2
2449 working_dir: "/vllm-workspace/tests"
2450 source_file_dependencies:
2576 - - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
2577 - - tests/v1/kv_connector/nixl_integration/
2578 - - vllm/platforms/rocm.py
2451 + - vllm/
2452 + - tests/weight_loading
2453 commands:
2580 - - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
2581 - - CROSS_LAYERS_BLOCKS=True ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
2582 -
2454 + - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models-amd.txt
2455
2584 -- label: Distributed Tests (4 GPUs)(A100-MI325) # 20.9m
2585 - timeout_in_minutes: 37
2586 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2587 - agent_pool: mi325_4
2588 - num_gpus: 4
2456 +- label: Weight Loading Multiple GPU - Large Models # TBD
2457 + timeout_in_minutes: 180
2458 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
2459 + agent_pool: mi300_2
2460 + num_gpus: 2
2461 optional: true
2462 working_dir: "/vllm-workspace/tests"
2463 source_file_dependencies:
2464 - vllm/
2465 + - tests/weight_loading
2466 commands:
2594 - - export TORCH_NCCL_BLOCKING_WAIT=1
2595 - - pytest -v -s distributed/test_custom_all_reduce.py
2596 - - torchrun --nproc_per_node=2 distributed/test_ca_buffer_sharing.py
2597 - - TARGET_TEST_SUITE=A100 pytest basic_correctness/ -v -s -m 'distributed(num_gpus=2)'
2598 - - pytest -v -s -x lora/test_mixtral.py
2467 + - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models-large-amd.txt
2468
2469 +#########################################################################################################################################
2470 +# #
2471 +# MI325 (gfx942) tests #
2472 +# #
2473 +#########################################################################################################################################
2474
2601 -- label: Distributed Tests (2 GPUs)(H100-MI325) # TBD
2475 +#---------------------------------------------------------- mi325 · compile ----------------------------------------------------------#
2476 +
2477 +- label: Distributed Compile + RPC Tests (2 GPUs) # TBD
2478 timeout_in_minutes: 180
2479 mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2480 agent_pool: mi325_2
2481 num_gpus: 2
2606 - working_dir: "/vllm-workspace/"
2482 + working_dir: "/vllm-workspace/tests"
2483 source_file_dependencies:
2484 + - vllm/compilation/
2485 - vllm/distributed/
2609 - - vllm/v1/distributed/
2610 - - vllm/model_executor/layers/fused_moe/
2611 - - tests/v1/distributed/test_dbo.py
2612 - - vllm/_aiter_ops.py
2486 + - vllm/engine/
2487 + - vllm/executor/
2488 + - vllm/worker/worker_base.py
2489 + - vllm/v1/engine/
2490 + - vllm/v1/worker/
2491 + - tests/compile/fullgraph/test_basic_correctness.py
2492 + - tests/compile/test_wrapper.py
2493 + - tests/entrypoints/llm/test_collective_rpc.py
2494 - vllm/platforms/rocm.py
2495 commands:
2496 - export TORCH_NCCL_BLOCKING_WAIT=1
2616 - - VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model=Qwen/Qwen1.5-MoE-A2.7B -tp=1 -dp=2 --max-model-len=2048 --all2all-backend=deepep_high_throughput
2617 - - pytest -v -s tests/v1/distributed/test_dbo.py
2497 + - pytest -v -s entrypoints/llm/test_collective_rpc.py
2498 + - pytest -v -s ./compile/fullgraph/test_basic_correctness.py
2499 + - pytest -v -s ./compile/test_wrapper.py
2500
2501 +#-------------------------------------------------------- mi325 · distributed --------------------------------------------------------#
2502
2620 -- label: Distributed Compile Unit Tests (2xH100-2xMI325) # 14.3m
2621 - timeout_in_minutes: 32
2503 +- label: Distributed Torchrun + Shutdown Tests (2 GPUs) # TBD
2504 + timeout_in_minutes: 180
2505 mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2506 agent_pool: mi325_2
2507 num_gpus: 2
2625 - working_dir: "/vllm-workspace/"
2626 - source_file_dependencies:
2627 - - vllm/compilation/
2628 - - vllm/model_executor/layers
2629 - - tests/compile/passes/distributed/
2630 - - vllm/_aiter_ops.py
2631 - - vllm/platforms/rocm.py
2632 - commands:
2633 - - export VLLM_TEST_CLEAN_GPU_MEMORY=1
2634 - - VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/compile/passes/distributed/test_async_tp.py
2635 - - pytest -v -s tests/compile/passes/distributed/test_sequence_parallelism.py
2636 - # TODO: this test is not supported on ROCm, there are aiter kernels for this.
2637 - # - pytest -v -s tests/compile/passes/distributed/test_fusion_all_reduce.py
2638 - # - pytest -v -s tests/compile/distributed/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm
2639 - # - "VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/compile/distributed/test_fusions_e2e.py -k 'not Llama-4'"
2640 -
2641 -
2642 -- label: LM Eval Small Models # 13.3m
2643 - timeout_in_minutes: 23
2644 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2645 - agent_pool: mi325_1
2508 working_dir: "/vllm-workspace/tests"
2509 source_file_dependencies:
2648 - - csrc/
2649 - - vllm/model_executor/layers/quantization
2650 - - vllm/model_executor/models/
2651 - - vllm/model_executor/model_loader/
2652 - - vllm/v1/attention/backends/
2653 - - vllm/v1/attention/selector.py
2654 - - vllm/_aiter_ops.py
2655 - - vllm/platforms/rocm.py
2656 - commands:
2657 - - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small.txt
2658 -
2659 -
2660 -- label: LM Eval Small Models (MI325) # TBD
2661 - timeout_in_minutes: 180
2662 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2663 - agent_pool: mi325_1
2664 - working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
2665 - source_file_dependencies:
2666 - - csrc/
2667 - - vllm/model_executor/layers/quantization
2668 - - vllm/model_executor/models/
2669 - - vllm/model_executor/model_loader/
2670 - - vllm/v1/attention/backends/
2671 - - vllm/v1/attention/selector.py
2672 - - vllm/_aiter_ops.py
2510 + - vllm/distributed/
2511 + - vllm/engine/
2512 + - vllm/executor/
2513 + - vllm/worker/worker_base.py
2514 + - vllm/v1/engine/
2515 + - vllm/v1/worker/
2516 + - tests/distributed/
2517 + - tests/v1/shutdown
2518 + - tests/v1/worker/test_worker_memory_snapshot.py
2519 - vllm/platforms/rocm.py
2520 commands:
2675 - - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-small-rocm.txt
2676 -
2521 + - export TORCH_NCCL_BLOCKING_WAIT=1
2522 + - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
2523 + - VLLM_TEST_SAME_HOST=1 VLLM_TEST_WITH_DEFAULT_DEVICE_SET=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
2524 + - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s v1/shutdown
2525 + - pytest -v -s v1/worker/test_worker_memory_snapshot.py
2526
2678 -- label: LM Eval Small Models (B200-MI325) # TBD
2527 +- label: Distributed Compile + Comm (4 GPUs) # TBD
2528 timeout_in_minutes: 180
2529 mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2681 - agent_pool: mi325_2
2530 + agent_pool: mi325_4
2531 + num_gpus: 4
2532 working_dir: "/vllm-workspace/tests"
2533 source_file_dependencies:
2684 - - csrc/
2685 - - vllm/model_executor/layers/quantization
2686 - - vllm/model_executor/models/
2687 - - vllm/model_executor/model_loader/
2688 - - vllm/v1/attention/backends/
2689 - - vllm/v1/attention/selector.py
2690 - - vllm/_aiter_ops.py
2534 + - vllm/distributed/
2535 + - tests/distributed/test_pynccl
2536 + - tests/distributed/test_events
2537 + - tests/compile/fullgraph/test_basic_correctness.py
2538 + - tests/distributed/test_symm_mem_allreduce.py
2539 + - tests/distributed/test_multiproc_executor.py
2540 - vllm/platforms/rocm.py
2541 commands:
2693 - - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-mi3xx-fp8-and-mixed.txt
2542 + - export TORCH_NCCL_BLOCKING_WAIT=1
2543 + - pytest -v -s compile/fullgraph/test_basic_correctness.py
2544 + - pytest -v -s distributed/test_pynccl.py
2545 + - pytest -v -s distributed/test_events.py
2546 + - pytest -v -s distributed/test_symm_mem_allreduce.py
2547 + - pytest -v -s distributed/test_multiproc_executor.py::test_multiproc_executor_multi_node
2548
2549 +#---------------------------------------------------------- mi325 · engine -----------------------------------------------------------#
2550
2696 -- label: LM Eval Large Models (H200-MI325) # TBD
2551 +- label: Engine # TBD
2552 timeout_in_minutes: 180
2553 mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2699 - agent_pool: mi325_8
2700 - optional: true
2701 - num_gpus: 8
2554 + agent_pool: mi325_1
2555 working_dir: "/vllm-workspace/tests"
2556 source_file_dependencies:
2704 - - vllm/model_executor/models/
2705 - - vllm/model_executor/model_loader/
2706 - - vllm/model_executor/layers/quantization/
2707 - - vllm/v1/attention/backends/
2708 - - vllm/v1/attention/selector.py
2709 - - vllm/model_executor/layers/layernorm.py
2710 - - csrc/
2711 - - vllm/_aiter_ops.py
2712 - - vllm/platforms/rocm.py
2713 - - tests/evals/
2557 + - vllm/
2558 + - tests/engine
2559 + - tests/test_sequence
2560 + - tests/test_config
2561 + - tests/test_logger
2562 + - tests/test_vllm_port
2563 commands:
2715 - - export VLLM_WORKER_MULTIPROC_METHOD=spawn
2716 - - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-mi3xx.txt
2564 + - pytest -v -s engine test_sequence.py test_config.py test_logger.py test_vllm_port.py
2565
2566 +#----------------------------------------------------------- mi325 · evals -----------------------------------------------------------#
2567
2719 -- label: LM Eval Large Models (4 GPUs)(FP8) # 24.8m
2720 - timeout_in_minutes: 42
2568 +- label: LM Eval Large Models (4xH100-4xMI325) # TBD
2569 + timeout_in_minutes: 180
2570 mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2571 agent_pool: mi325_4
2572 num_gpus: 4
@@ -2736,29 +2585,7 @@ steps:
2585 - export VLLM_USE_DEEP_GEMM=0
2586 - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large-rocm-fp8.txt --tp-size=4
2587
2739 -
2740 -- label: LM Eval Large Models (4 GPUs)(A100-MI325) # 17.3m
2741 - timeout_in_minutes: 27
2742 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2743 - agent_pool: mi325_4
2744 - num_gpus: 4
2745 - optional: true
2746 - working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
2747 - source_file_dependencies:
2748 - - csrc/
2749 - - vllm/model_executor/layers/quantization
2750 - - vllm/model_executor/models/
2751 - - vllm/model_executor/model_loader/
2752 - - vllm/v1/attention/backends/
2753 - - vllm/v1/attention/selector.py
2754 - - vllm/_aiter_ops.py
2755 - - vllm/platforms/rocm.py
2756 - commands:
2757 - - export VLLM_WORKER_MULTIPROC_METHOD=spawn
2758 - - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large.txt --tp-size=4
2759 -
2760 -
2761 -- label: ROCm LM Eval Large Models (8 Card) # TBD
2588 +- label: ROCm LM Eval Large Models (8 GPUs) # TBD
2589 timeout_in_minutes: 180
2590 mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2591 agent_pool: mi325_8
@@ -2779,232 +2606,143 @@ steps:
2606 - export VLLM_WORKER_MULTIPROC_METHOD=spawn
2607 - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large-rocm.txt --tp-size=8
2608
2609 +#---------------------------------------------------------- mi325 · models -----------------------------------------------------------#
2610
2783 -- label: GPQA Eval (GPT-OSS) (H100-MI325) # TBD
2611 +- label: Language Models Test (Extended Generation) # TBD
2612 timeout_in_minutes: 180
2613 mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2786 - agent_pool: mi325_2
2787 - num_gpus: 2
2788 - optional: true
2614 + agent_pool: mi325_1
2615 working_dir: "/vllm-workspace/tests"
2616 source_file_dependencies:
2791 - - csrc/
2792 - - vllm/model_executor/layers/quantization
2793 - - vllm/model_executor/models/
2794 - - vllm/model_executor/model_loader/
2795 - - vllm/v1/attention/backends/
2796 - - vllm/v1/attention/selector.py
2797 - - vllm/model_executor/layers/fused_moe/
2798 - - vllm/_aiter_ops.py
2799 - - vllm/platforms/rocm.py
2800 - - tests/evals/gpt_oss/
2617 + - vllm/
2618 + - tests/models/language/generation
2619 commands:
2802 - - uv pip install --system 'gpt-oss[eval]==0.0.5'
2803 - - pytest -s -v evals/gpt_oss/test_gpqa_correctness.py --config-list-file=configs/models-gfx942.txt
2804 -
2620 + - uv pip install --system --no-build-isolation 'git+https://github.com/AndreasKaratzas/mamba@fix-rocm-7.0-warp-size-constexpr'
2621 + - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
2622 + - pytest -v -s models/language/generation -m '(not core_model) and (not hybrid_model)'
2623
2806 -- label: DeepSeek V2-Lite Accuracy # 6.7m
2807 - timeout_in_minutes: 12
2624 +- label: Language Models Tests (Hybrid) %N # TBD
2625 + timeout_in_minutes: 180
2626 mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2809 - agent_pool: mi325_4
2810 - num_gpus: 4
2811 - optional: true
2812 - working_dir: "/vllm-workspace"
2627 + agent_pool: mi325_1
2628 + torch_nightly: true
2629 + parallelism: 2
2630 + working_dir: "/vllm-workspace/tests"
2631 source_file_dependencies:
2814 - - vllm/model_executor/models/
2815 - - vllm/model_executor/model_loader/
2816 - - vllm/distributed/eplb
2817 - - vllm/model_executor/layers/fused_moe/
2818 - - vllm/model_executor/layers/quantization/
2819 - - vllm/v1/attention/backends/
2820 - - vllm/v1/attention/backends/mla/
2821 - - vllm/v1/attention/selector.py
2822 - - .buildkite/scripts/scheduled_integration_test/
2823 - - vllm/_aiter_ops.py
2824 - - vllm/platforms/rocm.py
2632 + - vllm/
2633 + - tests/models/language/generation
2634 commands:
2826 - - bash .buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_ep_eplb.sh 0.25 200 8010
2827 -
2635 + - uv pip install --system --no-build-isolation 'git+https://github.com/AndreasKaratzas/mamba@fix-rocm-7.0-warp-size-constexpr'
2636 + - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
2637 + - pytest -v -s models/language/generation -m hybrid_model --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
2638
2829 -- label: DeepSeek V2-Lite Prefetch Offload Accuracy (H100-MI325) # TBD
2639 +- label: Multi-Modal Models (Extended Pooling) # TBD
2640 timeout_in_minutes: 180
2641 mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2642 agent_pool: mi325_1
2833 - num_gpus: 1
2834 - working_dir: "/vllm-workspace"
2643 + working_dir: "/vllm-workspace/tests"
2644 source_file_dependencies:
2836 - - vllm/model_executor/models/
2837 - - vllm/model_executor/model_loader/
2838 - - vllm/model_executor/layers/fused_moe/
2839 - - vllm/model_executor/layers/quantization/
2840 - - vllm/v1/attention/backends/
2841 - - vllm/v1/attention/backends/mla/
2842 - - vllm/v1/attention/selector.py
2843 - - .buildkite/scripts/scheduled_integration_test/
2844 - - vllm/_aiter_ops.py
2845 - - vllm/platforms/rocm.py
2645 + - vllm/
2646 + - tests/models/multimodal/pooling
2647 commands:
2847 - - bash .buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_prefetch_offload.sh 0.25 200 8030
2848 -
2648 + - pytest -v -s models/multimodal/pooling -m 'not core_model'
2649
2850 -- label: Qwen3-30B-A3B-FP8-block Accuracy # 6.4m
2851 - timeout_in_minutes: 11
2650 +- label: "Multi-Modal Models (Standard) 2: qwen3 + gemma" # TBD
2651 + timeout_in_minutes: 180
2652 mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2853 - agent_pool: mi325_4
2854 - optional: true
2855 - working_dir: "/vllm-workspace"
2653 + agent_pool: mi325_1
2654 + torch_nightly: true
2655 + working_dir: "/vllm-workspace/tests"
2656 source_file_dependencies:
2857 - - vllm/model_executor/models/
2858 - - vllm/model_executor/model_loader/
2859 - - vllm/model_executor/layers/quantization/
2860 - - vllm/distributed/eplb
2861 - - vllm/model_executor/layers/fused_moe/
2862 - - vllm/v1/attention/backends/
2863 - - vllm/v1/attention/selector.py
2864 - - .buildkite/scripts/scheduled_integration_test/
2865 - - vllm/_aiter_ops.py
2866 - - vllm/platforms/rocm.py
2657 + - vllm/
2658 + - tests/models/multimodal
2659 commands:
2868 - - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh 0.8 200 8020
2660 + - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
2661 + - pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "qwen3 or gemma"
2662 + - pytest -v -s models/multimodal/generation/test_qwen2_5_vl.py -m core_model
2663
2664 +#------------------------------------------------------------ mi325 · v1 -------------------------------------------------------------#
2665
2871 -- label: Qwen3-Next-80B-A3B-Instruct MTP Async EPLB Accuracy # 10.9m
2872 - timeout_in_minutes: 22
2666 +- label: V1 Spec Decode # TBD
2667 + timeout_in_minutes: 180
2668 mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2874 - agent_pool: mi325_4
2875 - num_gpus: 4
2876 - optional: true
2877 - working_dir: "/vllm-workspace"
2669 + agent_pool: mi325_1
2670 + working_dir: "/vllm-workspace/tests"
2671 source_file_dependencies:
2879 - - vllm/model_executor/models/
2880 - - vllm/model_executor/model_loader/
2881 - - vllm/v1/spec_decode/
2882 - - vllm/distributed/eplb
2883 - - vllm/model_executor/layers/fused_moe/
2884 - - vllm/model_executor/layers/quantization/
2885 - - vllm/v1/attention/backends/
2886 - - vllm/v1/attention/selector.py
2887 - - .buildkite/scripts/scheduled_integration_test/
2888 - - vllm/_aiter_ops.py
2889 - - vllm/platforms/rocm.py
2672 + - vllm/
2673 + - tests/v1/spec_decode
2674 commands:
2891 - - bash .buildkite/scripts/scheduled_integration_test/qwen3_next_mtp_async_eplb.sh 0.8 1319 8040
2892 -
2893 -##### .buildkite/test_areas/compile.yaml #####
2894 -# Slowly setting up the tests so that it is also easier for the
2895 -# CI team to review and upstream to the pipelinev2.
2896 -# The following tests are important for vLLM IR Ops refactoring,
2897 -# which affects fusion passes on ROCm. So we have to
2898 -# enable them as as soon as possible.
2675 + - pytest -v -s -m 'not slow_test' v1/spec_decode
2676
2900 -## TODO: Enable the test in this group
2901 -# # corresponds to .buildkite/test_areas/compile.yaml
2902 -# - label: Fusion and Compile Unit Tests (2xB200-2xMI325) # TBD
2903 -# timeout_in_minutes: 180
2904 -# mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325, tj]
2905 -# agent_pool: mi325_1 # changed to 1 GPU until the fusion all reduce is enabled then only revert back to 2 GPUs
2906 -# num_gpus: 1
2907 -# working_dir: "/vllm-workspace/"
2908 -# source_file_dependencies:
2909 -# - csrc/quantization/fp4/
2910 -# - vllm/model_executor/layers/quantization/
2911 -# - vllm/model_executor/layers/layernorm.py
2912 -# - vllm/model_executor/layers/activation.py
2913 -# - vllm/model_executor/layers/attention/attention.py
2914 -# - vllm/v1/attention/backends/flashinfer.py
2915 -# - vllm/compilation/ # TODO(luka) limit to vllm/compilation/passes
2916 -# - tests/compile/test_fusion_attn.py
2917 -# - tests/compile/test_silu_mul_quant_fusion.py
2918 -# - tests/compile/distributed/test_fusion_all_reduce.py
2919 -# - tests/compile/fullgraph/test_full_graph.py
2920 -# commands:
2921 -# - rocm-smi
2922 -# # we run all backend tests on ROCm
2923 -# # These two tests are covered in "PyTorch Compilation Passes Unit Tests"
2924 -# # - "pytest -v -s tests/compile/passes/test_fusion_attn.py"
2925 -# # - "pytest -v -s tests/compile/passes/test_silu_mul_quant_fusion.py"
2926 -# # TODO: this test is not supported on ROCm, there are aiter kernels for this.
2927 -# # - pytest -v -s tests/compile/passes/distributed/test_fusion_all_reduce.py
2928 -# # TODO: find out more details
2929 -# # - pytest -v -s tests/compile/fullgraph/test_full_graph.py::test_fp8_kv_scale_compile
2677 +#########################################################################################################################################
2678 +# #
2679 +# MI355 (gfx950) tests #
2680 +# #
2681 +#########################################################################################################################################
2682
2683 +#-------------------------------------------------------- mi355 · benchmarks ---------------------------------------------------------#
2684
2932 -- label: Fusion E2E Quick (H100-MI325) # TBD
2685 +- label: Attention Benchmarks Smoke Test (B200-MI355) # TBD
2686 timeout_in_minutes: 180
2934 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2935 - agent_pool: mi325_1
2936 - num_gpus: 1
2687 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
2688 + agent_pool: mi355_2
2689 + num_gpus: 2
2690 working_dir: "/vllm-workspace/"
2691 source_file_dependencies:
2939 - - csrc/quantization/
2940 - - vllm/model_executor/
2692 + - benchmarks/attention_benchmarks/
2693 - vllm/v1/attention/
2942 - - vllm/compilation/
2943 - - tests/compile/fusions_e2e/
2694 - vllm/_aiter_ops.py
2695 - vllm/platforms/rocm.py
2696 commands:
2947 - - rocm-smi
2948 - # Run all models and attn backends but only Inductor partition and native custom ops
2949 - - "pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k 'inductor_partition and not +rms_norm and not +quant_fp8'"
2950 - # Different from CUDA, Qwen requires +rms_norm and +quant_fp8 as rms+quant fusion is only supported on AITER
2951 - - "pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k 'inductor_partition and +rms_norm and +quant_fp8 and qwen3'"
2697 + - python3 benchmarks/attention_benchmarks/benchmark.py --backends ROCM_ATTN ROCM_AITER_FA ROCM_AITER_UNIFIED_ATTN --batch-specs "8q1s1k" --repeats 1 --warmup-iters 1
2698
2699 +#-------------------------------------------------------- mi355 · distributed --------------------------------------------------------#
2700
2954 -- label: Fusion E2E Config Sweep (H100-MI325) # TBD
2701 +- label: Distributed Tests (2xH100-2xMI355) # TBD
2702 timeout_in_minutes: 180
2956 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
2957 - agent_pool: mi325_1
2958 - num_gpus: 1
2703 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
2704 + agent_pool: mi355_2
2705 + num_gpus: 2
2706 + optional: true
2707 working_dir: "/vllm-workspace/"
2708 source_file_dependencies:
2961 - - csrc/quantization/
2962 - - vllm/compilation/
2963 - - vllm/model_executor/layers/layernorm.py
2964 - - vllm/model_executor/layers/activation.py
2965 - - vllm/model_executor/layers/attention/attention.py
2966 - - vllm/model_executor/layers/quantization/input_quant_fp8.py
2967 - - tests/compile/fusions_e2e/
2709 + - vllm/distributed/
2710 + - vllm/v1/distributed/
2711 + - vllm/model_executor/layers/fused_moe/
2712 + - vllm/v1/attention/backends/
2713 + - vllm/v1/attention/selector.py
2714 + - tests/distributed/test_context_parallel.py
2715 + - tests/v1/distributed/test_dbo.py
2716 + - examples/offline_inference/data_parallel.py
2717 - vllm/_aiter_ops.py
2718 - vllm/platforms/rocm.py
2719 commands:
2971 - - rocm-smi
2972 - - pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k "llama-3"
2973 -
2974 -## There are no ops on ROCm for these tests.
2975 -## The test still passes but the logs are not useful.
2976 -## fused ops just call torch.ops.symm_mem which
2977 -## exists in ROCm even though they don't work
2978 -# - label: AsyncTP Correctness Tests (2xH100-2xMI325)
2979 -# - label: Fusion E2E TP2 Quick (H100-MI325)
2980 -# - label: Fusion E2E TP2 AsyncTP Config Sweep (H100-MI325)
2981 -# - label: Fusion E2E TP2 (B200-MI325)
2982 -# - label: Sequence Parallel Correctness Tests (2xH100-2xMI325)
2983 -
2720 + - export TORCH_NCCL_BLOCKING_WAIT=1
2721 + - pytest -v -s tests/distributed/test_context_parallel.py
2722 + - pytest -v -s tests/v1/distributed/test_dbo.py
2723
2985 -#####################################################################################################################################
2986 -# #
2987 -# gfx950 #
2988 -# #
2989 -#####################################################################################################################################
2724 +#-------------------------------------------------------- mi355 · entrypoints --------------------------------------------------------#
2725
2991 -- label: Entrypoints Integration (API Server openai - Part 1) # TBD
2726 +- label: Entrypoints Integration (API Server 2) # TBD
2727 timeout_in_minutes: 180
2728 mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
2729 agent_pool: mi355_1
2730 + optional: true
2731 fast_check: true
2732 torch_nightly: true
2733 working_dir: "/vllm-workspace/tests"
2734 source_file_dependencies:
2735 - vllm/
3000 - - tests/entrypoints/openai
3001 - - tests/entrypoints/test_chat_utils
2736 + - tests/entrypoints/rpc
2737 + - tests/entrypoints/serve/instrumentator
2738 + - tests/tool_use
2739 commands:
2740 - export VLLM_WORKER_MULTIPROC_METHOD=spawn
3004 - - pytest -v -s entrypoints/openai/chat_completion --ignore=entrypoints/openai/chat_completion/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/chat_completion/test_oot_registration.py
3005 -
2741 + - pytest -v -s entrypoints/serve/instrumentator
2742 + - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/rpc
2743 + - pytest -v -s tool_use
2744
3007 -- label: Entrypoints Integration (API Server openai - Part 2) # TBD
2745 +- label: Entrypoints Integration (API Server openai - Part 1) # TBD
2746 timeout_in_minutes: 180
2747 mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
2748 agent_pool: mi355_1
@@ -3017,12 +2755,9 @@ steps:
2755 - tests/entrypoints/test_chat_utils
2756 commands:
2757 - export VLLM_WORKER_MULTIPROC_METHOD=spawn
3020 - - pytest -v -s entrypoints/openai/completion --ignore=entrypoints/openai/completion/test_tensorizer_entrypoint.py
3021 - - pytest -v -s entrypoints/openai/speech_to_text/
3022 - - pytest -v -s entrypoints/test_chat_utils.py
3023 -
2758 + - pytest -v -s entrypoints/openai/chat_completion --ignore=entrypoints/openai/chat_completion/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/chat_completion/test_oot_registration.py
2759
3025 -- label: Entrypoints Integration (API Server openai - Part 3) # TBD
2760 +- label: Entrypoints Integration (API Server openai - Part 2) # TBD
2761 timeout_in_minutes: 180
2762 mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
2763 agent_pool: mi355_1
@@ -3035,28 +2770,24 @@ steps:
2770 - tests/entrypoints/test_chat_utils
2771 commands:
2772 - export VLLM_WORKER_MULTIPROC_METHOD=spawn
3038 - - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/chat_completion --ignore=entrypoints/openai/completion --ignore=entrypoints/openai/speech_to_text/ --ignore=entrypoints/openai/correctness/ --ignore=entrypoints/openai/tool_parsers/ --ignore=entrypoints/openai/responses
3039 -
2773 + - pytest -v -s entrypoints/openai/completion --ignore=entrypoints/openai/completion/test_tensorizer_entrypoint.py
2774 + - pytest -v -s entrypoints/openai/speech_to_text/
2775 + - pytest -v -s entrypoints/test_chat_utils.py
2776
3041 -- label: Entrypoints Integration (API Server 2) # TBD
2777 +- label: Entrypoints Integration (API Server openai - Part 3) # TBD
2778 timeout_in_minutes: 180
2779 mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
2780 agent_pool: mi355_1
3045 - optional: true
2781 fast_check: true
2782 torch_nightly: true
2783 working_dir: "/vllm-workspace/tests"
2784 source_file_dependencies:
2785 - vllm/
3051 - - tests/entrypoints/rpc
3052 - - tests/entrypoints/serve/instrumentator
3053 - - tests/tool_use
2786 + - tests/entrypoints/openai
2787 + - tests/entrypoints/test_chat_utils
2788 commands:
2789 - export VLLM_WORKER_MULTIPROC_METHOD=spawn
3056 - - pytest -v -s entrypoints/serve/instrumentator
3057 - - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/rpc
3058 - - pytest -v -s tool_use
3059 -
2790 + - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/chat_completion --ignore=entrypoints/openai/completion --ignore=entrypoints/openai/speech_to_text/ --ignore=entrypoints/openai/correctness/ --ignore=entrypoints/openai/tool_parsers/ --ignore=entrypoints/openai/responses --ignore=entrypoints/openai/test_multi_api_servers.py
2791
2792 - label: Entrypoints Integration (Pooling) # TBD
2793 timeout_in_minutes: 180
@@ -3072,95 +2803,112 @@ steps:
2803 - export VLLM_WORKER_MULTIPROC_METHOD=spawn
2804 - pytest -v -s entrypoints/pooling
2805
2806 +#----------------------------------------------------------- mi355 · evals -----------------------------------------------------------#
2807
3076 -- label: Regression # TBD
2808 +- label: GPQA Eval (GPT-OSS) (2xB200-2xMI355) # TBD
2809 timeout_in_minutes: 180
2810 mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
3079 - agent_pool: mi355_1
2811 + agent_pool: mi355_2
2812 + num_gpus: 2
2813 optional: true
2814 working_dir: "/vllm-workspace/tests"
2815 source_file_dependencies:
3083 - - vllm/
3084 - - tests/test_regression
2816 + - csrc/
2817 + - vllm/model_executor/layers/quantization
2818 + - vllm/model_executor/models/
2819 + - vllm/model_executor/model_loader/
2820 + - vllm/v1/attention/backends/
2821 + - vllm/v1/attention/selector.py
2822 + - vllm/model_executor/layers/fused_moe/
2823 + - tests/evals/gpt_oss/
2824 + - vllm/_aiter_ops.py
2825 + - vllm/platforms/rocm.py
2826 commands:
3086 - - pip install modelscope
3087 - - pytest -v -s test_regression.py
3088 -
2827 + - uv pip install --system 'gpt-oss[eval]==0.0.5'
2828 + - pytest -s -v evals/gpt_oss/test_gpqa_correctness.py --config-list-file=configs/models-gfx950.txt
2829
3090 -- label: V1 Spec Decode # TBD
3091 - timeout_in_minutes: 60
2830 +- label: LM Eval Qwen3-5 Models (B200-MI355) # TBD
2831 + timeout_in_minutes: 120
2832 mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
3093 - agent_pool: mi355_1
2833 + agent_pool: mi355_2
2834 + num_gpus: 2
2835 + optional: true
2836 working_dir: "/vllm-workspace/tests"
2837 source_file_dependencies:
3096 - - vllm/
3097 - - tests/v1/spec_decode
2838 + - vllm/model_executor/models/qwen3_5.py
2839 + - vllm/model_executor/models/qwen3_5_mtp.py
2840 + - vllm/transformers_utils/configs/qwen3_5.py
2841 + - vllm/transformers_utils/configs/qwen3_5_moe.py
2842 + - vllm/model_executor/models/qwen.py
2843 + - vllm/model_executor/models/qwen2.py
2844 + - vllm/model_executor/models/qwen3.py
2845 + - vllm/model_executor/models/qwen3_next.py
2846 + - vllm/model_executor/models/qwen3_next_mtp.py
2847 + - vllm/model_executor/layers/fla/ops/
2848 + - vllm/_aiter_ops.py
2849 + - vllm/platforms/rocm.py
2850 commands:
3099 - - pytest -v -s -m 'not slow_test' v1/spec_decode
3100 -
2851 + - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-qwen35-mi355.txt
2852
3102 -- label: V1 Sample + Logits # TBD
3103 - timeout_in_minutes: 60
2853 +- label: LM Eval Small Models (2xB200-2xMI355) # TBD
2854 + timeout_in_minutes: 180
2855 mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
3105 - agent_pool: mi355_1
2856 + agent_pool: mi355_2
2857 + optional: true
2858 working_dir: "/vllm-workspace/tests"
2859 source_file_dependencies:
3108 - - vllm/
3109 - - tests/v1/sample
3110 - - tests/v1/logits_processors
3111 - - tests/v1/test_oracle.py
3112 - - tests/v1/test_request.py
3113 - - tests/v1/test_outputs.py
2860 + - csrc/
2861 + - vllm/model_executor/layers/quantization
2862 + - vllm/model_executor/models/
2863 + - vllm/model_executor/model_loader/
2864 + - vllm/v1/attention/backends/
2865 + - vllm/v1/attention/selector.py
2866 + - vllm/_aiter_ops.py
2867 + - vllm/platforms/rocm.py
2868 commands:
3115 - - pytest -v -s v1/sample
3116 - - pytest -v -s v1/logits_processors
3117 - - pytest -v -s v1/test_oracle.py
3118 - - pytest -v -s v1/test_request.py
3119 - - pytest -v -s v1/test_outputs.py
3120 -
2869 + - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-mi3xx-fp8-and-mixed.txt
2870
3122 -- label: V1 Core + KV + Metrics # TBD
3123 - timeout_in_minutes: 60
2871 +- label: Qwen3-30B-A3B-FP8-block Accuracy (B200-MI355) # TBD
2872 + timeout_in_minutes: 180
2873 mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
3125 - agent_pool: mi355_1
3126 - working_dir: "/vllm-workspace/tests"
2874 + agent_pool: mi355_2
2875 + num_gpus: 2
2876 + working_dir: "/vllm-workspace"
2877 source_file_dependencies:
3128 - - vllm/
3129 - - tests/v1/core
3130 - - tests/v1/executor
3131 - - tests/v1/kv_offload
3132 - - tests/v1/worker
3133 - - tests/v1/kv_connector/unit
3134 - - tests/v1/metrics
3135 - - tests/entrypoints/openai/correctness/test_lmeval.py
2878 + - vllm/model_executor/models/
2879 + - vllm/model_executor/model_loader/
2880 + - vllm/model_executor/layers/quantization/
2881 + - vllm/model_executor/layers/fused_moe/
2882 + - vllm/distributed/eplb
2883 + - vllm/v1/attention/backends/
2884 + - vllm/v1/attention/selector.py
2885 + - .buildkite/scripts/scheduled_integration_test/
2886 + - vllm/_aiter_ops.py
2887 + - vllm/platforms/rocm.py
2888 commands:
3137 - - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
3138 - - pytest -v -s -m 'not cpu_test' v1/core
3139 - - pytest -v -s v1/executor
3140 - - pytest -v -s v1/kv_offload
3141 - - pytest -v -s v1/worker
3142 - - pytest -v -s -m 'not cpu_test' v1/kv_connector/unit
3143 - - pytest -v -s -m 'not cpu_test' v1/metrics
3144 - - pip install -U git+https://github.com/robertgshaw2-redhat/lm-evaluation-harness.git@streaming-api
3145 - - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
3146 -
2889 + - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh 0.8 200 8020 2 1
2890
3148 -- label: V1 attention (B200-MI355) # TBD
2891 +- label: LM Eval Large Models (4xH100-4xMI355) # TBD
2892 timeout_in_minutes: 180
2893 mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
3151 - agent_pool: mi355_1
3152 - working_dir: "/vllm-workspace/tests"
3153 - source_file_dependencies:
3154 - - vllm/config/attention.py
3155 - - vllm/model_executor/layers/attention
3156 - - vllm/v1/attention
3157 - - tests/v1/attention
2894 + agent_pool: mi355_4
2895 + num_gpus: 4
2896 + optional: true
2897 + working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
2898 + source_file_dependencies:
2899 + - csrc/
2900 + - vllm/model_executor/layers/quantization
2901 + - vllm/model_executor/models/
2902 + - vllm/model_executor/model_loader/
2903 + - vllm/v1/attention/backends/
2904 + - vllm/v1/attention/selector.py
2905 - vllm/_aiter_ops.py
3159 - - vllm/envs.py
2906 - vllm/platforms/rocm.py
2907 commands:
3162 - - pytest -v -s v1/attention
2908 + - export VLLM_USE_DEEP_GEMM=0
2909 + - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large-rocm-fp8.txt --tp-size=4
2910
2911 +#--------------------------------------------------------- mi355 · examples ----------------------------------------------------------#
2912
2913 - label: Examples # TBD
2914 timeout_in_minutes: 180
@@ -3195,6 +2943,31 @@ steps:
2943 - python3 offline_inference/spec_decode.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
2944 - python3 offline_inference/spec_decode.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
2945
2946 +#---------------------------------------------------------- mi355 · kernels ----------------------------------------------------------#
2947 +
2948 +- label: Kernels (B200-MI355) # TBD
2949 + timeout_in_minutes: 180
2950 + mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
2951 + agent_pool: mi355_1
2952 + working_dir: "/vllm-workspace/"
2953 + source_file_dependencies:
2954 + - csrc/quantization/fp4/
2955 + - csrc/attention/mla/
2956 + - csrc/quantization/cutlass_w8a8/moe/
2957 + - vllm/model_executor/layers/fused_moe/cutlass_moe.py
2958 + - vllm/v1/attention/backends/triton_attn.py
2959 + - vllm/v1/attention/backends/rocm_attn.py
2960 + - vllm/v1/attention/backends/rocm_aiter_fa.py
2961 + - vllm/v1/attention/backends/rocm_aiter_unified_attn.py
2962 + - vllm/v1/attention/backends/mla/aiter_triton_mla.py
2963 + - vllm/v1/attention/backends/mla/rocm_aiter_mla.py
2964 + - vllm/v1/attention/selector.py
2965 + - vllm/platforms/rocm.py
2966 + - vllm/_aiter_ops.py
2967 + commands:
2968 + - rocm-smi
2969 + - python3 examples/basic/offline_inference/chat.py
2970 + - pytest -v -s tests/kernels/attention/test_attention_selector.py
2971
2972 - label: Kernels Attention Test %N # TBD
2973 timeout_in_minutes: 180
@@ -3214,25 +2987,6 @@ steps:
2987 commands:
2988 - pytest -v -s kernels/attention --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
2989
3217 -
3218 -- label: Kernels Quantization Test %N # TBD
3219 - timeout_in_minutes: 180
3220 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
3221 - agent_pool: mi355_1
3222 - parallelism: 2
3223 - working_dir: "/vllm-workspace/tests"
3224 - source_file_dependencies:
3225 - - csrc/quantization/
3226 - - vllm/model_executor/layers/quantization
3227 - - tests/kernels/quantization
3228 - - tests/kernels/quantization/test_rocm_skinny_gemms.py
3229 - - vllm/_aiter_ops.py
3230 - - vllm/platforms/rocm.py
3231 - - vllm/model_executor/kernels/
3232 - commands:
3233 - - pytest -v -s kernels/quantization --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
3234 -
3235 -
2990 - label: Kernels MoE Test %N # TBD
2991 timeout_in_minutes: 180
2992 mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
@@ -3253,54 +3007,40 @@ steps:
3007 - pytest -v -s kernels/moe --ignore=kernels/moe/test_modular_oai_triton_moe.py --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
3008 - pytest -v -s kernels/moe/test_modular_oai_triton_moe.py --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
3009
3256 -
3257 -- label: Kernels FP8 MoE Test # TBD
3258 - timeout_in_minutes: 180
3259 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
3260 - agent_pool: mi355_2
3261 - working_dir: "/vllm-workspace/tests"
3262 - source_file_dependencies:
3263 - - csrc/moe/
3264 - - csrc/quantization/w8a8/cutlass/moe/
3265 - - vllm/model_executor/layers/fused_moe/
3266 - - tests/kernels/moe/test_deepep_moe.py
3267 - - vllm/_aiter_ops.py
3268 - - vllm/platforms/rocm.py
3269 - - vllm/envs.py
3270 - commands:
3271 - - pytest -v -s kernels/moe/test_deepep_moe.py
3272 -
3273 -
3274 -- label: Quantization # TBD
3010 +- label: Kernels Quantization Test %N # TBD
3011 timeout_in_minutes: 180
3012 mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
3013 agent_pool: mi355_1
3014 + parallelism: 2
3015 working_dir: "/vllm-workspace/tests"
3016 source_file_dependencies:
3280 - - csrc/
3017 + - csrc/quantization/
3018 - vllm/model_executor/layers/quantization
3282 - - tests/quantization
3019 + - tests/kernels/quantization
3020 + - tests/kernels/quantization/test_rocm_skinny_gemms.py
3021 - vllm/_aiter_ops.py
3022 - vllm/platforms/rocm.py
3023 + - vllm/model_executor/kernels/
3024 commands:
3286 - - uv pip install --system torchao==0.17.0
3287 - - uv pip install --system conch-triton-kernels
3288 - - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py
3289 -
3025 + - pytest -v -s kernels/quantization --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
3026
3291 -- label: Language Models Tests (Standard) # TBD
3027 +- label: Kernels FP8 MoE Test (2xH100-2xMI355) # TBD
3028 timeout_in_minutes: 180
3029 mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
3294 - agent_pool: mi355_1
3295 - torch_nightly: true
3030 + agent_pool: mi355_2
3031 working_dir: "/vllm-workspace/tests"
3032 source_file_dependencies:
3298 - - vllm/
3299 - - tests/models/language
3033 + - csrc/moe/
3034 + - csrc/quantization/w8a8/cutlass/moe/
3035 + - vllm/model_executor/layers/fused_moe/
3036 + - tests/kernels/moe/test_deepep_moe.py
3037 + - vllm/_aiter_ops.py
3038 + - vllm/platforms/rocm.py
3039 + - vllm/envs.py
3040 commands:
3301 - - pip freeze | grep -E 'torch'
3302 - - pytest -v -s models/language -m 'core_model and (not slow_test)'
3041 + - pytest -v -s kernels/moe/test_deepep_moe.py
3042
3043 +#---------------------------------------------------------- mi355 · models -----------------------------------------------------------#
3044
3045 - label: Language Models Test (Extended Generation) # TBD
3046 timeout_in_minutes: 180
@@ -3312,10 +3052,9 @@ steps:
3052 - tests/models/language/generation
3053 commands:
3054 - uv pip install --system --no-build-isolation 'git+https://github.com/AndreasKaratzas/mamba@fix-rocm-7.0-warp-size-constexpr'
3315 - - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.2'
3055 + - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
3056 - pytest -v -s models/language/generation -m '(not core_model) and (not hybrid_model)'
3057
3318 -
3058 - label: Language Models Test (Extended Pooling) # TBD
3059 timeout_in_minutes: 180
3060 mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
@@ -3327,72 +3066,46 @@ steps:
3066 commands:
3067 - pytest -v -s models/language/pooling -m 'not core_model'
3068
3330 -
3331 -- label: "Multi-Modal Models (Standard) 1: qwen2" # TBD
3332 - timeout_in_minutes: 180
3333 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
3334 - agent_pool: mi355_1
3335 - torch_nightly: true
3336 - optional: true
3337 - working_dir: "/vllm-workspace/tests"
3338 - source_file_dependencies:
3339 - - vllm/
3340 - - tests/models/multimodal
3341 - commands:
3342 - - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
3343 - - pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "qwen2"
3344 - - pytest -v -s models/multimodal/generation/test_ultravox.py -m core_model
3345 -
3346 -
3347 -- label: "Multi-Modal Models (Standard) 2: qwen3 + gemma" # TBD
3348 - timeout_in_minutes: 180
3349 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
3350 - agent_pool: mi355_1
3351 - torch_nightly: true
3352 - optional: true
3353 - working_dir: "/vllm-workspace/tests"
3354 - source_file_dependencies:
3355 - - vllm/
3356 - - tests/models/multimodal
3357 - commands:
3358 - - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
3359 - - pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "qwen3 or gemma"
3360 - - pytest -v -s models/multimodal/generation/test_qwen2_5_vl.py -m core_model
3361 -
3362 -
3363 -- label: "Multi-Modal Models (Standard) 3: llava + qwen2_vl" # TBD
3069 +- label: Language Models Test (PPL) # TBD
3070 timeout_in_minutes: 180
3071 mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
3072 agent_pool: mi355_1
3367 - torch_nightly: true
3368 - optional: true
3073 working_dir: "/vllm-workspace/tests"
3074 source_file_dependencies:
3371 - - vllm/
3372 - - tests/models/multimodal/generation
3373 - - tests/models/multimodal/test_mapping.py
3075 + - vllm/model_executor/models/qwen3_5.py
3076 + - vllm/model_executor/models/qwen3_5_mtp.py
3077 + - vllm/transformers_utils/configs/qwen3_5.py
3078 + - vllm/transformers_utils/configs/qwen3_5_moe.py
3079 + - vllm/model_executor/models/qwen.py
3080 + - vllm/model_executor/models/qwen2.py
3081 + - vllm/model_executor/models/qwen3.py
3082 + - vllm/model_executor/models/qwen3_next.py
3083 + - vllm/model_executor/models/qwen3_next_mtp.py
3084 + - vllm/model_executor/layers/fla/ops/
3085 + - vllm/_aiter_ops.py
3086 + - vllm/v1/attention/backends/triton_attn.py
3087 + - vllm/v1/attention/backends/rocm_attn.py
3088 + - vllm/v1/attention/backends/rocm_aiter_unified_attn.py
3089 + - vllm/v1/attention/backends/rocm_aiter_fa.py
3090 + - vllm/v1/attention/backends/flex_attention.py
3091 + - vllm/v1/attention/ops/
3092 + - vllm/platforms/rocm.py
3093 + - tests/models/language/generation_ppl_test
3094 commands:
3375 - - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
3376 - - pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "not qwen2 and not qwen3 and not gemma"
3377 - - pytest -v -s models/multimodal/generation/test_qwen2_vl.py -m core_model
3378 -
3095 + - pytest -v -s models/language/generation_ppl_test
3096
3380 -- label: "Multi-Modal Models (Standard) 4: other + whisper" # TBD
3097 +- label: Language Models Tests (Standard) # TBD
3098 timeout_in_minutes: 180
3099 mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
3100 agent_pool: mi355_1
3101 torch_nightly: true
3385 - optional: true
3102 working_dir: "/vllm-workspace/tests"
3103 source_file_dependencies:
3104 - vllm/
3389 - - tests/models/multimodal/generation
3105 + - tests/models/language
3106 commands:
3391 - - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
3392 - - pytest -v -s models/multimodal -m core_model --ignore models/multimodal/generation/test_common.py --ignore models/multimodal/generation/test_ultravox.py --ignore models/multimodal/generation/test_qwen2_5_vl.py --ignore models/multimodal/generation/test_qwen2_vl.py --ignore models/multimodal/generation/test_whisper.py --ignore models/multimodal/generation/test_memory_leak.py --ignore models/multimodal/processing
3393 - - pytest -v -s models/multimodal/generation/test_memory_leak.py -m core_model
3394 - - cd .. && VLLM_WORKER_MULTIPROC_METHOD=spawn pytest -v -s tests/models/multimodal/generation/test_whisper.py -m core_model
3395 -
3107 + - pip freeze | grep -E 'torch'
3108 + - pytest -v -s models/language -m 'core_model and (not slow_test)'
3109
3110 - label: Multi-Modal Models (Extended Generation 1) # TBD
3111 timeout_in_minutes: 180
@@ -3409,21 +3122,6 @@ steps:
3122 - pytest -v -s models/multimodal/generation -m 'not core_model' --ignore models/multimodal/generation/test_common.py
3123 - pytest -v -s models/multimodal/test_mapping.py
3124
3412 -
3413 -- label: Multi-Modal Models (Extended Generation 2) # TBD
3414 - timeout_in_minutes: 180
3415 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
3416 - agent_pool: mi355_1
3417 - optional: true
3418 - working_dir: "/vllm-workspace/tests"
3419 - source_file_dependencies:
3420 - - vllm/
3421 - - tests/models/multimodal/generation
3422 - commands:
3423 - - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
3424 - - pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=0) and not core_model'
3425 -
3426 -
3125 - label: Multi-Modal Models (Extended Generation 3) # TBD
3126 timeout_in_minutes: 180
3127 mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
@@ -3437,7 +3135,6 @@ steps:
3135 - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
3136 - pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=1) and not core_model'
3137
3440 -
3138 - label: Multi-Modal Models (Extended Pooling) # TBD
3139 timeout_in_minutes: 180
3140 mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
@@ -3450,302 +3147,252 @@ steps:
3147 commands:
3148 - pytest -v -s models/multimodal/pooling -m 'not core_model'
3149
3453 -
3454 -- label: Quantized Models Test # TBD
3455 - timeout_in_minutes: 180
3456 - mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
3457 - agent_pool: mi355_1
3458 - working_dir: "/vllm-workspace/tests"
3459 - source_file_dependencies:
3460 - - vllm/model_executor/layers/quantization
3461 - - tests/models/quantization
3462 - - vllm/_aiter_ops.py
3463 - - vllm/platforms/rocm.py
3464 - - vllm/model_executor/model_loader/
3465 - commands:
3466 - - pytest -v -s models/quantization
3467 -
3468 -
3469 -- label: Kernels (B200-MI355) # TBD
3150 +- label: "Multi-Modal Models (Standard) 1: qwen2" # TBD
3151 timeout_in_minutes: 180
3152 mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
3153 agent_pool: mi355_1

This file is too large to show in full.

tests/quantization/test_gfx950_moe.py renamed
+2 -2
@@ -2,5 +2,5 @@
2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
4
5 -def test_mi3xx_moe():
6 - print("TODO: add tests for Mi3xx MoE quantization")
5 +def test_mi355_moe():
6 + print("TODO: add tests for Mi355 MoE quantization")