defaults:
  rocprofv3:
    # keys can be used with --filter-rocprofv3 option
    full: [--hsa-trace, --hip-trace, --kernel-trace, --memory-copy-trace, --rccl-trace, --memory-allocation-trace]
    kernel: [--kernel-trace]
    hip: [--hip-trace]
    hsa: [--hsa-trace]
    memcpy: [--memory-copy-trace]
    malloc: [--memory-allocation-trace]
    rccl: [--rccl-trace]
  benchmark:
    - baseline
    - disabled-sdk-contexts
    - sdk-buffer-overhead
    - sdk-callback-overhead
    - tool-runtime-overhead
  group:
    - vllm

jobs:
  # vLLM throughput benchmark with Qwen3-30B-A3B model
  - name: vllm-qwen3-30b
    group: [vllm, inference, llm]
    command: [vllm, bench, throughput,
              --model, /model/Qwen/Qwen3-30B-A3B,
              --num-prompts=100,
              --input-len, "256",
              --output-len, "256",
              --tensor-parallel-size, "4",
              --dtype, float16]
    rocprofv3:
      - [--hsa-trace, --hip-trace, --kernel-trace, --memory-copy-trace, --rccl-trace, --memory-allocation-trace]
    benchmark:
      - baseline
      - disabled-sdk-contexts
      - tool-runtime-overhead