# Run this exact file on both the local GPU and one H100; see README.md.
experiment:
  name: report-showcase
  output_dir: /var/tmp/vllm-optimizer-demo/runs
  seed: 42
server:
  model: /var/tmp/vllm-optimizer-demo/models/opt-125m
  host: 127.0.0.1
  port: 8766
  dtype: float16
  seed: 42
  tensor-parallel-size: 1
  gpu-memory-utilization: 0.4
  max-model-len: 512
  enforce-eager: true
  enable-chunked-prefill: true
  max-num-seqs: 1
  max-num-batched-tokens: 256
env:
  CUDA_VISIBLE_DEVICES: "0"
  VLLM_USE_V2_MODEL_RUNNER: "0"
  VLLM_USE_FLASHINFER_SAMPLER: "0"
tune:
  max-num-seqs:
    # Zero deliberately exercises real server validation failure diagnostics.
    values: [1, 4, 16, 0]
  max-num-batched-tokens:
    values: [256, 512]
benchmark:
  engine: vllm
  warmup_repeats: 1
  repeats: 4
  min_repeats: 4
  max_failure_percentage: 0
  runs:
    - name: short-chat
      args:
        dataset-name: random
        random-input-len: 64
        random-output-len: 32
        num-prompts: 16
        max-concurrency: 8
        request-rate: inf
        seed: 42
        temperature: 0
        ignore-eos: true
        percentile-metrics: ttft,tpot,itl,e2el
        metric-percentiles: 50,95,99
    - name: longer-prompt
      args:
        dataset-name: random
        random-input-len: 256
        random-output-len: 64
        num-prompts: 16
        max-concurrency: 8
        request-rate: inf
        seed: 42
        temperature: 0
        ignore-eos: true
        percentile-metrics: ttft,tpot,itl,e2el
        metric-percentiles: 50,95,99
baseline:
  enabled: true
optimization:
  maximize: output_tokens_per_second
  sampler: grid
analysis:
  # Demonstration policy: any nonzero drift triggers finalist validation.
  drift_threshold: 0
timeouts:
  startup: 180s
  benchmark: 120s
execution:
  retry:
    max_attempts: 2
logging:
  level: INFO