# MOSS-VoiceGenerator (MossTTSDelayModel, 1.7B, n_vq=16, 24 kHz)
# HF: OpenMOSS-Team/MOSS-VoiceGenerator
# Zero-shot voice design — create a new voice from a text description.
async_chunk: true
# OpenMOSS-Team/MOSS-VoiceGenerator ships custom modeling code; the newer vLLM
# ModelConfig hard-rejects it at validation without this (was lenient before).
trust_remote_code: true

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

stages:
  - stage_id: 0
    max_num_seqs: 8
    gpu_memory_utilization: 0.60
    # The MOSS talker (custom audio heads + per-request GPU buffers) is not
    # CUDA-graph-safe; FULL-decode graph capture crashes the worker. Run the
    # LLM stage eager, matching the other TTS talker configs (ming_tts,
    # moss_tts_nano).
    enforce_eager: true
    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: 1.0
      top_p: 0.95
      top_k: 50
      max_tokens: 4096
      seed: 42
      repetition_penalty: 1.0

  - stage_id: 1
    max_num_seqs: 1
    gpu_memory_utilization: 0.12
    enforce_eager: true
    enable_prefix_caching: false
    async_scheduling: true
    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