# Compatibility alias for breeze_tts_2_throughput.yaml; keep both profiles synchronized.
# Breeze-TTS-2 non-streaming throughput profile, tested on one H100 80 GB.
# Both streamed and complete HTTP responses use the stateful codec pipeline.
async_chunk: true
connectors:
  shm:
    name: SharedMemoryConnector
    extra:
      codec_streaming: true
      codec_chunk_frames: 16
      initial_codec_chunk_frames: 16
      codec_chunk_ramp: [16]
      codec_left_context_frames: 0
      decode_cudagraph_batch_sizes: [1, 2, 4, 8, 16, 32]
      connector_get_sleep_s: 0.001
      connector_get_max_wait_first_chunk: 1000
      connector_get_max_wait: 300
stages:
  - stage_id: 0
    devices: "0"
    max_num_seqs: 32
    max_model_len: 2048
    max_num_batched_tokens: 4096
    gpu_memory_utilization: 0.65
    # Reserve enough KV for thirty-two full-length requests.
    kv_cache_memory_bytes: 8589934592
    enforce_eager: false
    compilation_config:
      mode: 3
      cudagraph_mode: FULL_AND_PIECEWISE
      cudagraph_capture_sizes: [1, 2, 4, 8, 32, 64, 128, 256]
    enable_prefix_caching: false
    enable_chunked_prefill: false
    async_scheduling: false
    engine_extras:
      additional_config:
        cfg_prefill_delay_tokens: 0
    output_connectors:
      to_stage_1: shm
    default_sampling_params:
      temperature: 0.9
      top_p: 1.0
      top_k: 50
      repetition_penalty: 1.1
      max_tokens: 1500
      seed: 42
  - stage_id: 1
    devices: "0"
    max_num_seqs: 32
    max_model_len: 32768
    max_num_batched_tokens: 32768
    gpu_memory_utilization: 0.15
    enforce_eager: false
    enable_prefix_caching: false
    async_scheduling: true
    input_connectors:
      from_stage_0: shm
    default_sampling_params:
      temperature: 0.0
      top_p: 1.0
      top_k: -1
      repetition_penalty: 1.0
      max_tokens: 32768