---
# Audio Duration Scaling Test
# How does performance scale with audio duration?

test_scenario:
  name: "audio-duration-scaling"
  type: "scaling"
  description: "Measure how transcription performance scales with audio clip duration"

  # 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: 100

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

  # Test stages - Different audio duration buckets
  #
  # NOTE: GuideLLM's encode_media preprocessor only supports max_duration
  # (truncation), not min_duration or exclusive range filtering.  Stages
  # therefore use truncation-based buckets: a clip shorter than the
  # previous stage's max_duration will appear at its natural length in
  # later stages as well.  For mutually exclusive duration analysis,
  # post-process per-request audio_seconds from benchmarks.json.
  stages:
    # Stage 1: Short audio clips (truncated to ≤5 seconds)
    - name: "short-clips"
      description: "Audio truncated to ≤5s"
      profile: "synchronous"
      max_requests: 100
      audio_config:
        format: "mp3"
        bitrate: "64k"
        sample_rate: 16000
        mono: true
        max_duration: 5.0
      result_filename_suffix: "short-5sec"
      expected_behavior: "Lowest latency, highest files/sec"

    # Stage 2: Medium audio clips (truncated to ≤15 seconds)
    - name: "medium-clips"
      description: "Audio truncated to ≤15s"
      profile: "synchronous"
      max_requests: 100
      audio_config:
        format: "mp3"
        bitrate: "64k"
        sample_rate: 16000
        mono: true
        max_duration: 15.0
      result_filename_suffix: "medium-15sec"
      expected_behavior: "Moderate latency"

    # Stage 3: Long audio clips (truncated to ≤30 seconds)
    - name: "long-clips"
      description: "Audio truncated to ≤30s"
      profile: "synchronous"
      max_requests: 100
      audio_config:
        format: "mp3"
        bitrate: "64k"
        sample_rate: 16000
        mono: true
        max_duration: 30.0
      result_filename_suffix: "long-30sec"
      expected_behavior: "Higher latency, lower files/sec"

    # Stage 4: Full-length clips (no truncation)
    - name: "full-length"
      description: "Full-length audio, no truncation"
      profile: "synchronous"
      max_requests: 100
      audio_config:
        format: "mp3"
        bitrate: "64k"
        sample_rate: 16000
        mono: true
        max_duration: null
      result_filename_suffix: "full-length"
      expected_behavior: "Highest latency, assess real audio distribution"

  # Key metrics
  metrics:
    primary:
      # Per-file metrics
      - "Mean latency per file (ms)"
      - "Throughput (files/sec)"

      # Per-audio-second metrics (normalized)
      - "Latency per audio second (ms/audio_sec)"
      - "Processing rate (audio_seconds/wall_clock_sec)"

      # Audio characteristics
      - "Mean audio duration (seconds)"
      - "Total audio seconds processed"

    derived:
      - "Real-time factor = processing_time / audio_duration"
      - "Scalability factor = latency_ratio / duration_ratio"
        # < 1.0 = sub-linear scaling (good)
        # = 1.0 = linear scaling
        # > 1.0 = super-linear scaling (bad)

  # Analysis
  analysis:
    questions:
      - "Is processing time linear with audio duration?"
      - "What is the per-second processing cost?"
      - "Does performance degrade for longer clips?"
      - "What is the optimal audio chunk size for throughput?"

  # Success criteria
  success_criteria:
    - "Linear or sub-linear scaling with audio duration"
    - "Real-time factor < 1.0 for all durations"
    - "Consistent per-audio-second processing rate"

  results:
    format: "json"
    location: "results/audio-models/{model}/audio-duration-scaling/"
    graphs:
      - type: "latency_vs_duration"
        title: "Latency vs Audio Duration"
        x_axis: "Audio Duration (seconds)"
        y_axis: "E2E Latency (ms)"
        fit_line: "linear regression"

      - type: "throughput_vs_duration"
        title: "Throughput by Audio Duration Bucket"
        x_axis: "Duration Bucket"
        y_axis: "Audio Seconds per Second"

      - type: "realtime_factor"
        title: "Real-time Factor by Duration"
        x_axis: "Audio Duration (seconds)"
        y_axis: "Real-time Factor"
        reference_line: 1.0  # Real-time processing line

# GuideLLM command examples
guidellm_commands:
  short_clips:
    cmd: |
      guidellm benchmark \
        --target "http://localhost:8000" \
        --request-type audio_transcriptions \
        --profile synchronous \
        --max-requests 100 \
        --data openslr/librispeech_asr \
        --data-args '{"name": "clean", "split": "test"}' \
        --data-column-mapper '{"audio_column": "audio"}' \
        --request-formatter-kwargs '{"encode_kwargs": {"max_duration": 5.0, "audio_format": "mp3"}}'

  long_clips:
    cmd: |
      guidellm benchmark \
        --target "http://localhost:8000" \
        --request-type audio_transcriptions \
        --profile synchronous \
        --max-requests 100 \
        --data openslr/librispeech_asr \
        --data-args '{"name": "clean", "split": "test"}' \
        --data-column-mapper '{"audio_column": "audio"}' \
        --request-formatter-kwargs '{"encode_kwargs": {"max_duration": 30.0, "audio_format": "mp3"}}'