| 1 | group: Fault Tolerance |
| 2 | depends_on: |
| 3 | - image-build |
| 4 | steps: |
| 5 | - label: ":nvidia: (H100) Fault Tolerance E2E" |
| 6 | key: fault-tolerance-e2e-2xh100 |
| 7 | timeout_in_minutes: 35 |
| 8 | device: h100 |
| 9 | num_devices: 2 |
| 10 | working_dir: "/vllm-workspace/tests" |
| 11 | source_file_dependencies: |
| 12 | - vllm/v1/fault_tolerance/ |
| 13 | - vllm/v1/worker/sentinel/ |
| 14 | - vllm/entrypoints/serve/fault_tolerance/ |
| 15 | - vllm/distributed/elastic_ep/ |
| 16 | - vllm/distributed/device_communicators/ |
| 17 | - vllm/v1/engine/ |
| 18 | - vllm/v1/worker/ |
| 19 | - tests/v1/fault_tolerance/ |
| 20 | - tests/v1/distributed/test_external_lb_dp.py |
| 21 | commands: |
| 22 | # Base image has no nixl; install it or has_nixl_ep() skips the tests. |
| 23 | - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh |
| 24 | # https://github.com/NVIDIA/nccl/issues/1838 |
| 25 | - export NCCL_CUMEM_HOST_ENABLE=0 |
| 26 | - pytest -v -s v1/fault_tolerance/test_fault_tolerance_e2e.py |
| 27 | mirror: |
| 28 | amd: |
| 29 | label: ":amd: (MI355) MoRI EP Healthy Baseline" |
| 30 | dind: false |
| 31 | device: mi355_2 |
| 32 | num_devices: 2 |
| 33 | timeout_in_minutes: 10 |
| 34 | working_dir: "/vllm-workspace/tests" |
| 35 | depends_on: |
| 36 | - image-build-amd |
| 37 | source_file_dependencies: |
| 38 | - csrc/moe/ |
| 39 | - csrc/rocm/ |
| 40 | - tests/kernels/moe/ |
| 41 | - vllm/model_executor/layers/fused_moe/ |
| 42 | - vllm/model_executor/layers/quantization/ |
| 43 | - vllm/distributed/ |
| 44 | - vllm/config/ |
| 45 | - vllm/forward_context.py |
| 46 | - vllm/v1/worker/workspace.py |
| 47 | - vllm/utils/import_utils.py |
| 48 | - vllm/utils/math_utils.py |
| 49 | - vllm/utils/torch_utils.py |
| 50 | - vllm/platforms/ |
| 51 | - vllm/_aiter_ops.py |
| 52 | # TODO: Add fault injection, detection and recovery once MoRI supports them. |
| 53 | # This mirror currently checks healthy EP execution only. |
| 54 | commands: |
| 55 | - VLLM_WORKER_MULTIPROC_METHOD=spawn pytest -v -s -rs kernels/moe/test_moe_layer.py::test_moe_layer_mori_graph |
| 56 | - VLLM_TEST_ENABLE_MORI_MOE_LAYER=1 VLLM_WORKER_MULTIPROC_METHOD=spawn pytest -v -s -rs 'kernels/moe/test_moe_layer.py::test_moe_layer[False-mori_high_throughput-2-1-True]' --subtests='[32-512-512-8-2-bfloat16-None-False-False-False-False-mori_high_throughput-2-2-1]' |