@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 49c00fe304

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 14, 05:30 UTC

Mirroring changes in test-pipeline.yaml into test-amd.yaml (#27242)

Signed-off-by: Alexei V. Ivanov <alexei.ivanov@amd.com>

Alexei-V-Ivanov-AMD committed Oct 22, 2025 at 08:59 UTC 49c00fe304d4890f87dbcd99aad9f46222c22111

1 file changed +65 -14

.buildkite/test-amd.yaml

+65 -14

@@ -454,8 +454,8 @@ steps:

454 - pytest -v -s compile/test_fusion_attn.py

455 - pytest -v -s compile/test_functionalization.py

456 - pytest -v -s compile/test_silu_mul_quant_fusion.py

457 - - pytest -v -s compile/test_sequence_parallelism.py

458 - - pytest -v -s compile/test_async_tp.py

457 + # - pytest -v -s compile/test_sequence_parallelism.py

458 + # - pytest -v -s compile/test_async_tp.py

459 - pytest -v -s compile/test_fusion_all_reduce.py

460 - pytest -v -s compile/test_decorator.py

461 - pytest -v -s compile/test_noop_elimination.py

@@ -474,8 +474,8 @@ steps:

474 - pytest -v -s compile/test_basic_correctness.py

475 - pytest -v -s compile/piecewise/

476

477 -- label: PyTorch Fullgraph Test # 20min

478 - timeout_in_minutes: 30

477 +- label: PyTorch Fullgraph Test # 22min

478 + timeout_in_minutes: 35

479 mirror_hardwares: [amdexperimental, amdproduction]

480 agent_pool: mi325_1

481 # grade: Blocking

@@ -485,6 +485,7 @@ steps:

485 - tests/compile

486 commands:

487 - pytest -v -s compile/test_full_graph.py

488 + - pytest -v -s compile/test_fusions_e2e.py

489

490 - label: Kernels Core Operation Test # 48min

491 timeout_in_minutes: 75

@@ -494,6 +495,7 @@ steps:

495 source_file_dependencies:

496 - csrc/

497 - tests/kernels/core

498 + - tests/kernels/test_top_k_per_row.py

499 commands:

500 - pytest -v -s kernels/core kernels/test_top_k_per_row.py

501

@@ -606,7 +608,7 @@ steps:

608 # we can only upgrade after this is resolved

609 # TODO(jerryzh168): resolve the above comment

610 - uv pip install --system torchao==0.13.0

609 - - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/

611 + - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py

612

613 - label: LM Eval Small Models # 53min

614 timeout_in_minutes: 75

@@ -848,6 +850,18 @@ steps:

850 - pytest -v -s models/multimodal -m core_model --ignore models/multimodal/generation/test_whisper.py --ignore models/multimodal/processing

851 - cd .. && VLLM_WORKER_MULTIPROC_METHOD=spawn pytest -v -s tests/models/multimodal/generation/test_whisper.py -m core_model # Otherwise, mp_method="spawn" doesn't work

852

853 +- label: Multi-Modal Accuracy Eval (Small Models) # 50min

854 + mirror_hardwares: [amdexperimental]

855 + agent_pool: mi325_1

856 + timeout_in_minutes: 70

857 + working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"

858 + source_file_dependencies:

859 + - vllm/multimodal/

860 + - vllm/inputs/

861 + - vllm/v1/core/

862 + commands:

863 + - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-mm-small.txt --tp-size=1

864 +

865 - label: Multi-Modal Models Test (Extended) 1

866 mirror_hardwares: [amdexperimental]

867 agent_pool: mi325_1

@@ -923,8 +937,8 @@ steps:

937 # Whisper needs spawn method to avoid deadlock

938 - VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/offline_inference/audio_language.py --model-type whisper

939

926 -- label: Blackwell Test # 38 min

927 - timeout_in_minutes: 60

940 +- label: Blackwell Test # 21 min

941 + timeout_in_minutes: 30

942 working_dir: "/vllm-workspace/"

943 gpu: b200

944 # optional: true

@@ -937,8 +951,6 @@ steps:

951 - vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py

952 - vllm/model_executor/layers/quantization/utils/flashinfer_utils.py

953 - vllm/v1/attention/backends/flashinfer.py

940 - - vllm/compilation/fusion.py

941 - - vllm/compilation/fusion_attn.py

954 commands:

955 - nvidia-smi

956 - python3 examples/offline_inference/basic/chat.py

@@ -955,13 +967,32 @@ steps:

967 - pytest -v -s tests/kernels/quantization/test_nvfp4_scaled_mm.py

968 - pytest -v -s tests/kernels/quantization/test_flashinfer_scaled_mm.py

969 - pytest -v -s tests/kernels/quantization/test_flashinfer_nvfp4_scaled_mm.py

970 + - pytest -v -s tests/kernels/quantization/test_nvfp4_qutlass.py

971 + - pytest -v -s tests/kernels/quantization/test_mxfp4_qutlass.py

972 - pytest -v -s tests/kernels/moe/test_nvfp4_moe.py

973 - pytest -v -s tests/kernels/moe/test_ocp_mx_moe.py

960 - # Fusion

961 - - pytest -v -s tests/compile/test_fusion_all_reduce.py

962 - - pytest -v -s tests/compile/test_fusion_attn.py::test_attention_quant_pattern

974 - pytest -v -s tests/kernels/moe/test_flashinfer.py

975 +

976 +- label: Blackwell Fusion Tests # 30 min

977 + timeout_in_minutes: 40

978 + working_dir: "/vllm-workspace/"

979 + gpu: b200

980 + source_file_dependencies:

981 + - csrc/quantization/fp4/

982 + - vllm/model_executor/layers/quantization/utils/flashinfer_utils.py

983 + - vllm/v1/attention/backends/flashinfer.py

984 + - vllm/compilation/

985 + # can affect pattern matching

986 + - vllm/model_executor/layers/layernorm.py

987 + - vllm/model_executor/layers/activation.py

988 + - vllm/model_executor/layers/quantization/input_quant_fp8.py

989 + commands:

990 + - nvidia-smi

991 + - pytest -v -s tests/compile/test_fusion_attn.py

992 - pytest -v -s tests/compile/test_silu_mul_quant_fusion.py

993 + # this runner has 2 GPUs available even though num_gpus=2 is not set

994 + - pytest -v -s tests/compile/test_fusion_all_reduce.py

995 + - pytest -v -s tests/compile/test_fusions_e2e.py

996

997 - label: Blackwell GPT-OSS Eval

998 timeout_in_minutes: 60

@@ -1129,6 +1160,11 @@ steps:

1160 - pytest -v -s plugins_tests/test_io_processor_plugins.py

1161 - pip uninstall prithvi_io_processor_plugin -y

1162 # end io_processor plugins test

1163 + # begin stat_logger plugins test

1164 + - pip install -e ./plugins/vllm_add_dummy_stat_logger

1165 + - pytest -v -s plugins_tests/test_stats_logger_plugins.py

1166 + - pip uninstall dummy_stat_logger -y

1167 + # end stat_logger plugins test

1168 # other tests continue here:

1169 - pytest -v -s plugins_tests/test_scheduler_plugins.py

1170 - pip install -e ./plugins/vllm_add_dummy_model

@@ -1173,7 +1209,6 @@ steps:

1209 - pytest -v -s -x lora/test_llama_tp.py

1210 - pytest -v -s -x lora/test_llm_with_multi_loras.py

1211

1176 -

1212 - label: Weight Loading Multiple GPU Test # 33min

1213 timeout_in_minutes: 45

1214 mirror_hardwares: [amdexperimental]

@@ -1202,6 +1237,18 @@ steps:

1237 commands:

1238 - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models-large.txt

1239

1240 +- label: NixlConnector PD accuracy tests (Distributed) # 30min

1241 + mirror_hardwares: [amdexperimental]

1242 + agent_pool: mi325_4

1243 + timeout_in_minutes: 30

1244 + working_dir: "/vllm-workspace/tests"

1245 + num_gpus: 4

1246 + source_file_dependencies:

1247 + - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py

1248 + - tests/v1/kv_connector/nixl_integration/

1249 + commands:

1250 + - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt

1251 + - bash v1/kv_connector/nixl_integration/tp_config_sweep_accuracy_test.sh

1252

1253 ##### multi gpus test #####

1254 ##### A100 test #####

@@ -1233,12 +1280,16 @@ steps:

1280 - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large.txt --tp-size=4

1281

1282 ##### H200 test #####

1236 -- label: Distrubted Tests (H200) # optional

1283 +- label: Distributed Tests (H200) # optional

1284 gpu: h200

1285 optional: true

1286 working_dir: "/vllm-workspace/"

1287 num_gpus: 2

1288 commands:

1289 + - pytest -v -s tests/compile/test_async_tp.py

1290 + - pytest -v -s tests/compile/test_sequence_parallelism.py

1291 + - pytest -v -s tests/compile/test_fusion_all_reduce.py

1292 + - pytest -v -s tests/compile/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm

1293 - pytest -v -s tests/distributed/test_context_parallel.py

1294 - 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

1295