# MOSS-TTS (MossTTSDelayModel, 8B, n_vq=32, 24 kHz)
# HF: OpenMOSS-Team/MOSS-TTS
# General-purpose TTS with voice cloning. Production-recommended variant.
async_chunk: true
trust_remote_code: true  # OpenMOSS-Team/MOSS-TTS ships custom modeling code

connectors:
  shm:
    name: SharedMemoryConnector
    extra:
      codec_streaming: true
      connector_get_sleep_s: 0.01
      connector_get_max_wait_first_chunk: 3000
      connector_get_max_wait: 300
      codec_chunk_frames: 25
      codec_left_context_frames: 0
      decode_cudagraph_capture_sizes: [4, 8, 16, 25, 32, 50, 64, 100, 128, 200, 256]

stages:
  - stage_id: 0
    max_num_seqs: 4
    gpu_memory_utilization: 0.85
    enable_prefix_caching: false
    async_scheduling: true
    max_num_batched_tokens: 4096
    max_model_len: 4096
    devices: "0"
    output_connectors:
      to_stage_1: shm
    default_sampling_params:
      temperature: 1.7
      top_p: 0.8
      top_k: 25
      max_tokens: 4096
      seed: 42
      repetition_penalty: 1.0
      stop_token_ids: [151645]  # <|im_end|> — terminates the talker AR loop

  - stage_id: 1
    max_num_seqs: 1
    gpu_memory_utilization: 0.12
    enforce_eager: false
    enable_prefix_caching: false
    async_scheduling: true
    # Talker can emit up to max_tokens=4096 frames × n_vq=32 = 131072 codec
    # tokens per request; cap codec stage above that so a single long output
    # doesn't trigger "could not broadcast (>65536,) into (65536,)".
    max_num_batched_tokens: 196608
    max_model_len: 196608
    devices: "0"
    input_connectors:
      from_stage_0: shm
    default_sampling_params:
      temperature: 0.0
      top_p: 1.0
      top_k: -1
      max_tokens: 65536
      seed: 42
      repetition_penalty: 1.0