# Breeze-TTS-2: talker -> stateful Qwen3-TTS codec decoder.
# Stage 0 batches codec frames and sends incremental codebook-major chunks.
async_chunk: true
trust_remote_code: true

connectors:
  connector_of_shared_memory:
    name: SharedMemoryConnector
    extra:
      codec_streaming: true
      connector_get_sleep_s: 0.01
      connector_get_max_wait_first_chunk: 3000
      connector_get_max_wait: 300
      breeze_codec_chunk_frames: 8

stages:
  - stage_id: 0
    max_num_seqs: 8
    gpu_memory_utilization: 0.80
    enable_prefix_caching: false
    # The talker derives one codec frame from the last hidden row of every
    # scheduled request. A partially scheduled prompt would yield a frame from
    # a mid-prompt hidden state, so keep each prefill in a single step.
    enable_chunked_prefill: false
    async_scheduling: true
    max_num_batched_tokens: 4096
    max_model_len: 4096
    devices: "0"
    output_connectors:
      to_stage_1: connector_of_shared_memory
    default_sampling_params:
      temperature: 0.0
      top_p: 1.0
      top_k: -1
      # Match upstream's penalty on generated codebook-0 history.
      repetition_penalty: 1.1
      max_tokens: 2048
      # min_tokens also masks the tokenizer's EOS 1 despite ignore_eos.
      # Keep valid codec token 1 and an immediate backbone EOS available.
      min_tokens: 0

  - stage_id: 1
    max_num_seqs: 8
    gpu_memory_utilization: 0.15
    enforce_eager: true
    enable_prefix_caching: false
    async_scheduling: true
    max_num_batched_tokens: 65536
    max_model_len: 65536
    devices: "0"
    default_sampling_params:
      temperature: 0.0
      top_p: 1.0
      top_k: -1
      max_tokens: 65536