# Complete vLLM Optimizer configuration reference.
# Values are examples; comments show mutually exclusive forms and defaults.
schema_version: 1

experiment:
  name: complete-example
  output_dir: runs
  seed: 42

# Use model for inline vLLM settings, or replace it with config.
server:
  model: /models/Qwen3-32B
  # config: ./vllm-config.yaml
  host: 0.0.0.0
  port: 8000
  tensor-parallel-size: 4
  dtype: bfloat16
  gpu-memory-utilization: 0.90
  max-model-len: 32768
  enforce-eager: false
  enable-prefix-caching: true
  # Any other vLLM serve argument is accepted here as a pass-through field.

# Every entry is either categorical (`values`) or an inclusive range.
tune:
  attention-backend:
    values: [FLASH_ATTN, FLASHINFER]
  max-num-seqs:
    values: [64, 128, 256]
  enforce-eager:
    values: [true, false]
  max-num-batched-tokens:
    min: 4096
    max: 16384
    step: 4096
  gpu-memory-utilization:
    min: 0.85
    max: 0.95
    step: 0.05

env:
  CUDA_VISIBLE_DEVICES: "0,1,2,3"
  VLLM_LOG_STATS_INTERVAL: "5"

tune_env:
  VLLM_USE_FLASHINFER_SAMPLER:
    values: ["0", "1"]
  WORKER_COUNT:
    min: 1
    max: 4
    step: 1

benchmark:
  engine: guidellm
  repeats: 4
  warmup_repeats: 1
  min_repeats: 2
  max_failure_percentage: 2
  accept_any_request_failures: false
  adaptive_repeats:
    minimum_relative_score: 0.7
  runs:
    - name: concurrent-chat
      request_format: /v1/chat/completions
      profile:
        kind: concurrent
        streams: [4, 16, 32]
        max_concurrency: 32
      constraints:
        - kind: max_requests
          count: 500
        - kind: max_duration
          seconds: 2m
      data:
        - kind: synthetic_text
          prompt_tokens: 512
          prompt_tokens_stdev: 32
          prompt_tokens_min: 128
          prompt_tokens_max: 1024
          output_tokens: 128
          output_tokens_stdev: 16
          output_tokens_min: 32
          output_tokens_max: 256
          turns: 1

  # To use vLLM Bench Serve, replace engine and runs with:
  # engine: vllm
  # runs:
  #   - name: random-throughput
  #     args:
  #       dataset-name: random
  #       random-input-len: 512
  #       random-output-len: 128
  #       random-prefix-len: 0
  #       num-prompts: 1000
  #       request-rate: inf
  #       max-concurrency: 32
  #       percentile-metrics: ttft,tpot,itl,e2el
  #       metric-percentiles: 50,90,95,99
  #       ignore-eos: true

baseline:
  enabled: true

optimization:
  maximize: output_tokens_per_second
  sampler: tpe
  trials: 20
  # Grid alternative: sampler: grid, with trials omitted.
  # Random alternative: sampler: random, with trials retained.

analysis:
  drift_threshold: 0.05
  finalist_validation:
    top_k: 2
    repeats: 6
  llm_summary:
    base_url: https://api.example.com/v1
    model: your-model
    api_key_env: VLLM_OPTIMIZER_LLM_API_KEY
    timeout: 30

timeouts:
  startup: 15m
  benchmark: 20m

logging:
  level: INFO

execution:
  mode: sequential
  host: 127.0.0.1
  health_path: /health
  shutdown_grace: 15
  drain_grace: 15
  retry:
    max_attempts: 2

  # For local parallel execution, remove server.port and
  # env.CUDA_VISIBLE_DEVICES, then replace mode and uncomment these fields:
  # mode: local_parallel
  # max_parallel_trials: 2
  # gpu_allocation:
  #   strategy: explicit
  #   allow_sharing: false
  #   workers:
  #     - name: worker-0
  #       devices: [0, 1]
  #     - name: worker-1
  #       devices: [2, 3]
  # ports:
  #   min: 8100
  #   max: 8199