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 12, 21:00 UTC
Updating the mirror of test-amd.yaml as of 2025-11-18 (#29016)
Signed-off-by: Alexei V. Ivanov <alexei.ivanov@amd.com>
Alexei-V-Ivanov-AMD committed
Nov 20, 2025 at 11:07 UTC
22924383e14a7a37ee86cf6e15f39e13efc86f7c
1 file changed
+27
-14
.buildkite/test-amd.yaml
+27
-14
@@ -61,7 +61,7 @@ steps:
61
- pytest -v -s -m 'not cpu_test' multimodal
62
- pytest -v -s utils_
63
64
-- label: Async Engine, Inputs, Utils, Worker Test (CPU) # 4 mins
64
+- label: Async Engine, Inputs, Utils, Worker, Config Test (CPU) # 4 mins
65
timeout_in_minutes: 10
66
mirror_hardwares: [amdexperimental, amdproduction]
67
agent_pool: mi325_1
@@ -73,6 +73,7 @@ steps:
73
- tests/multimodal
74
- tests/standalone_tests/lazy_imports.py
75
- tests/transformers_utils
76
+ - tests/config
77
no_gpu: true
78
commands:
79
- python3 standalone_tests/lazy_imports.py
@@ -80,6 +81,7 @@ steps:
81
- pytest -v -s test_outputs.py
82
- pytest -v -s -m 'cpu_test' multimodal
83
- pytest -v -s transformers_utils
84
+ - pytest -v -s config
85
86
- label: Python-only Installation Test # 10min
87
timeout_in_minutes: 20
@@ -390,6 +392,15 @@ steps:
392
commands:
393
- pytest -v -s v1/attention
394
395
+- label: V1 Test attention (B200) # 10min
396
+ timeout_in_minutes: 30
397
+ gpu: b200
398
+ source_file_dependencies:
399
+ - vllm/v1/attention
400
+ - tests/v1/attention
401
+ commands:
402
+ - VLLM_DISABLE_FLASHINFER_PREFILL=1 pytest -v -s v1/attention # TODO: FI prefill is bugged and causes incorrectness, fix this
403
+
404
- label: V1 Test others (CPU) # 5 mins
405
mirror_hardwares: [amdexperimental, amdproduction]
406
agent_pool: mi325_1
@@ -529,7 +540,7 @@ steps:
540
- pytest -v -s compile/fullgraph/test_full_graph.py -k 'not test_fp8_kv_scale_compile'
541
# Limit to no custom ops to reduce running time
542
# Wrap with quotes to escape yaml and avoid starting -k string with a -
532
- - "pytest -v -s compile/distributed/test_fusions_e2e.py -k 'TRITON and -quant_fp8'"
543
+ - "pytest -v -s compile/distributed/test_fusions_e2e.py -k 'TRITON and not +quant_fp8 and not Llama-4'"
544
545
- label: Cudagraph test
546
timeout_in_minutes: 20
@@ -694,7 +705,7 @@ steps:
705
- vllm/model_executor/models/whisper.py
706
commands: # LMEval
707
# Transcription WER check is skipped because encoder-decoder models are not supported on ROCm, see https://github.com/vllm-project/vllm/issues/27442
697
- - pytest -s entrypoints/openai/correctness/ --ignore entrypoints/openai/correctness/test_transcription_api_correctness.py
708
+ - pytest -s entrypoints/openai/correctness/
709
710
- label: OpenAI-Compatible Tool Use # 23 min
711
timeout_in_minutes: 35
@@ -995,12 +1006,12 @@ steps:
1006
optional: true
1007
commands:
1008
- pip install --upgrade git+https://github.com/huggingface/transformers
998
- - pytest -v -s tests/models/test_initialization.py
1009
+ - pytest -v -s tests/models/test_initialization.py -k 'not (Gemma3 or ModernBert or Qwen2_5_VL or Qwen2_5vl or Qwen2VL or TransformersMultiModalEmbeddingModel or TransformersMultiModalForSequenceClassification or Ultravox or Phi4Multimodal or LlavaNextVideo or MiniCPMO or Lfm2Moe or PaliGemma or RobertaForSequenceClassification or Ovis2_5 or Fuyu or DeepseekOCR or KimiVL)'
1010
- pytest -v -s tests/models/test_transformers.py
1000
- - pytest -v -s tests/models/multimodal/processing/
1001
- - pytest -v -s tests/models/multimodal/test_mapping.py
1011
+ # - pytest -v -s tests/models/multimodal/processing/
1012
+ - pytest -v -s tests/models/multimodal/test_mapping.py -k 'not (Gemma3 or Qwen2VL or Qwen2_5_VL)'
1013
- python3 examples/offline_inference/basic/chat.py
1003
- - python3 examples/offline_inference/vision_language.py --model-type qwen2_5_vl
1014
+ # - python3 examples/offline_inference/vision_language.py --model-type qwen2_5_vl
1015
# Whisper needs spawn method to avoid deadlock
1016
- VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/offline_inference/audio_language.py --model-type whisper
1017
@@ -1045,7 +1056,7 @@ steps:
1056
- pytest -v -s tests/kernels/moe/test_ocp_mx_moe.py
1057
- pytest -v -s tests/kernels/moe/test_flashinfer.py
1058
1048
-- label: Blackwell Fusion Tests # 30 min
1059
+- label: Blackwell Fusion and Compile Tests # 30 min
1060
timeout_in_minutes: 40
1061
working_dir: "/vllm-workspace/"
1062
gpu: b200
@@ -1066,7 +1077,9 @@ steps:
1077
- pytest -v -s tests/compile/distributed/test_fusion_all_reduce.py
1078
# Limit to Inductor partition, no custom ops, and allreduce & attn fusion to reduce running time
1079
# Wrap with quotes to escape yaml
1069
- - "pytest -v -s tests/compile/distributed/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm -k 'True and Llama-3.1 and -quant_fp8 and -rms_norm'"
1080
+ - "pytest -v -s tests/compile/distributed/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm -k 'True and not +quant_fp8 and not +rms_norm'"
1081
+ # test_fp8_kv_scale_compile requires FlashAttention (not supported on default L4/L40)
1082
+ - pytest -v -s tests/compile/distributed/test_full_graph.py::test_fp8_kv_scale_compile
1083
1084
- label: Blackwell Fusion E2E Tests # 30 min
1085
timeout_in_minutes: 40
@@ -1088,15 +1101,13 @@ steps:
1101
commands:
1102
- nvidia-smi
1103
# Run all e2e fusion tests
1091
- - pytest -v -s tests/compile/distributed/test_fusions_e2e.py
1092
- # test_fp8_kv_scale_compile requires FlashAttention (not supported on default L4/L40)
1093
- - pytest -v -s tests/compile/fullgraph/test_full_graph.py::test_fp8_kv_scale_compile
1104
+ - pytest -v -s tests/compile/test_fusions_e2e.py
1105
1106
- label: ROCm GPT-OSS Eval
1107
timeout_in_minutes: 60
1108
working_dir: "/vllm-workspace/"
1109
agent_pool: mi325_1
1099
- mirror_hardwares: [amdproduction]
1110
+ mirror_hardwares: [amdexperimental, amdproduction]
1111
optional: true # run on nightlies
1112
source_file_dependencies:
1113
- tests/evals/gpt_oss
@@ -1416,7 +1427,9 @@ steps:
1427
- pytest -v -s tests/compile/distributed/test_async_tp.py
1428
- pytest -v -s tests/compile/distributed/test_sequence_parallelism.py
1429
- pytest -v -s tests/compile/distributed/test_fusion_all_reduce.py
1419
- - pytest -v -s tests/compile/distributed/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm
1430
+ #- pytest -v -s tests/compile/distributed/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm
1431
+ - "pytest -v -s tests/compile/distributed/test_fusions_e2e.py -k 'not Llama-4'"
1432
+ - pytest -v -s tests/compile/distributed/test_sequence_parallel.py
1433
- pytest -v -s tests/distributed/test_context_parallel.py
1434
- CUDA_VISIBLE_DEVICES=1,2 VLLM_ALL2ALL_BACKEND=deepep_high_throughput VLLM_USE_DEEP_GEMM=1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model Qwen/Qwen1.5-MoE-A2.7B --tp-size=1 --dp-size=2 --max-model-len 2048
1435
- pytest -v -s tests/v1/distributed/test_dbo.py