---
# Audio Transcription Latency Test
# Focus on per-request latency characteristics

test_scenario:
  name: "transcription-latency"
  type: "latency"
  description: "Measure latency characteristics for audio transcription requests"

  # Test configuration
  backend: "openai-audio"
  endpoint: "/v1/audio/transcriptions"
  request_type: "audio_transcriptions"

  # Dataset configuration
  dataset:
    name: "openslr/librispeech_asr"
    config: "clean"
    split: "test"
    audio_column: "audio"
    num_samples: 50

  # Audio preprocessing
  audio_config:
    format: "wav"
    sample_rate: 16000
    mono: true

  # vLLM server configuration
  server:
    dtype: "float16"
    env_vars:
      VLLM_CPU_KVCACHE_SPACE: "2GiB"

  # Test stages - Different load levels
  stages:
    # Stage 1: Single user (no contention)
    - name: "single-user-latency"
      description: "Measure baseline latency with no concurrency"
      profile: "synchronous"
      max_requests: 50
      result_filename_suffix: "single-user"
      metrics_focus:
        - "Mean latency (E2E)"
        - "P50/P95/P99 latency"
        - "Consistency (latency variance)"

    # Stage 2: Light load (2 concurrent users)
    - name: "light-load-latency"
      description: "Measure latency under light concurrent load"
      profile: "concurrent"
      rate: 2
      max_requests: 50
      result_filename_suffix: "light-load"

    # Stage 3: Medium load (5 concurrent users)
    - name: "medium-load-latency"
      description: "Measure latency under medium concurrent load"
      profile: "concurrent"
      rate: 5
      max_requests: 50
      result_filename_suffix: "medium-load"

    # Stage 4: Heavy load (10 concurrent users)
    - name: "heavy-load-latency"
      description: "Measure latency under heavy concurrent load"
      profile: "concurrent"
      rate: 10
      max_requests: 50
      result_filename_suffix: "heavy-load"

  # Key metrics
  metrics:
    primary:
      - "Mean E2E latency (ms)"
      - "Median E2E latency (ms)"
      - "P95 E2E latency (ms)"
      - "P99 E2E latency (ms)"
      - "Latency standard deviation"
      - "Time to First Token (TTFT) - if available"

    derived:
      - "Latency degradation vs single-user (%)"
      - "Queueing time estimate (latency increase under load)"

  # Success criteria
  success_criteria:
    - "P99 latency < 2000ms under light load"
    - "P99 latency < 5000ms under heavy load"
    - "Latency degradation < 3x from single-user to heavy load"

  results:
    format: "json"
    location: "results/audio-models/{model}/transcription-latency/"
    graphs:
      - type: "latency_percentiles"
        title: "Latency Percentiles by Load Level"
        x_axis: "Load Level"
        y_axis: "Latency (ms)"
        series:
          - "P50"
          - "P95"
          - "P99"

      - type: "latency_vs_audio_duration"
        title: "Latency vs Audio Duration"
        x_axis: "Audio Duration (seconds)"
        y_axis: "E2E Latency (ms)"