| 1 | group: Samplers |
| 2 | depends_on: |
| 3 | - image-build |
| 4 | steps: |
| 5 | - label: ":nvidia: (H200 MIG 18GB) Samplers" |
| 6 | device: h200_18gb |
| 7 | key: samplers-test |
| 8 | timeout_in_minutes: 60 |
| 9 | source_file_dependencies: |
| 10 | - vllm/model_executor/layers |
| 11 | - vllm/sampling_metadata.py |
| 12 | - tests/samplers |
| 13 | - tests/conftest.py |
| 14 | - vllm/entrypoints/generate/beam_search |
| 15 | commands: |
| 16 | # VLLM_USE_FLASHINFER_SAMPLER defaults to 1 now, so we need to pin both |
| 17 | # values explicitly to still cover the PyTorch-native (Triton) path. |
| 18 | - VLLM_USE_FLASHINFER_SAMPLER=0 pytest -v -s samplers -k "not test_beam_search_passes_multimodal_data" |
| 19 | - VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers -k "not test_beam_search_passes_multimodal_data" |
| 20 | mirror: |
| 21 | amd: |
| 22 | label: ":amd: (MI250) Samplers" |
| 23 | dind: false |
| 24 | device: mi250_1 |
| 25 | timeout_in_minutes: 40 |
| 26 | depends_on: |
| 27 | - image-build-amd |
| 28 | source_file_dependencies: |
| 29 | - vllm/model_executor/layers |
| 30 | - vllm/sampling_metadata.py |
| 31 | - vllm/v1/sample/ |
| 32 | - vllm/entrypoints/generate/beam_search/ |
| 33 | - tests/samplers |
| 34 | - tests/conftest.py |
| 35 | - vllm/_aiter_ops.py |
| 36 | - vllm/platforms/rocm.py |
| 37 | commands: |
| 38 | - pytest -v -s samplers -k "not test_beam_search_passes_multimodal_data" |
| 39 | |
| 40 | # Qwen2-Audio-7B leaves no room for KV cache on a 16 GiB slice. |
| 41 | - label: ":nvidia: (H200 MIG 35GB) Samplers Multimodal Beam Search" |
| 42 | device: h200_35gb |
| 43 | key: samplers-multimodal-beam-search |
| 44 | timeout_in_minutes: 30 |
| 45 | source_file_dependencies: |
| 46 | - vllm/model_executor/layers |
| 47 | - vllm/sampling_metadata.py |
| 48 | - tests/samplers |
| 49 | - tests/conftest.py |
| 50 | - vllm/entrypoints/generate/beam_search |
| 51 | commands: |
| 52 | - VLLM_USE_FLASHINFER_SAMPLER=0 pytest -v -s samplers/test_beam_search.py::test_beam_search_passes_multimodal_data |
| 53 | - VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers/test_beam_search.py::test_beam_search_passes_multimodal_data |
| 54 | mirror: |
| 55 | amd: |
| 56 | label: ":amd: (MI355 DPX) Samplers Multimodal Beam Search" |
| 57 | dind: false |
| 58 | device: mi355_dpx |
| 59 | num_devices: 1 |
| 60 | timeout_in_minutes: 45 |
| 61 | working_dir: /vllm-workspace/tests |
| 62 | depends_on: |
| 63 | - image-build-amd |
| 64 | source_file_dependencies: |
| 65 | - vllm/model_executor/layers |
| 66 | - vllm/sampling_metadata.py |
| 67 | - vllm/v1/sample/ |
| 68 | - vllm/entrypoints/generate/beam_search/ |
| 69 | - tests/samplers |
| 70 | - tests/conftest.py |
| 71 | - vllm/_aiter_ops.py |
| 72 | - vllm/platforms/rocm.py |
| 73 | commands: |
| 74 | - pytest -v -s samplers/test_beam_search.py::test_beam_search_passes_multimodal_data |