# Single-GPU JoyAI-VL-Interaction -> Qwen3-TTS all-sync profile.
pipeline: joyai_vl_interaction
async_chunk: false
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
      codec_chunk_frames: 25
      codec_left_context_frames: 72
      initial_codec_chunk_frames: 1
      # Qwen3-TTS Code2Wav buckets: 25, 72+1, 72+25, 72+72+25, and 300+25 frames.
      decode_cudagraph_capture_sizes: [25, 73, 97, 169, 325]

stages:
  # GPU 0: JoyAI action model.
  - stage_id: 0
    model: jdopensource/JoyAI-VL-Interaction-Preview
    devices: "0"
    gpu_memory_utilization: 0.5
    max_num_seqs: 4
    # Some OmniInteract video prompts contain more than 12K input tokens.
    max_model_len: 16384
    max_num_batched_tokens: 16384
    limit_mm_per_prompt: {image: 256, video: 1}
    enable_prefix_caching: true
    default_sampling_params:
      temperature: 0.8
      top_p: 0.9
      top_k: 40
      max_tokens: 256
      repetition_penalty: 1.1
      presence_penalty: 1.5

  # GPU 0: Qwen3-TTS Talker.
  - stage_id: 1
    model: Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice
    devices: "0"
    gpu_memory_utilization: 0.2
    max_num_seqs: 8
    max_model_len: 4096
    max_num_batched_tokens: 512
    enable_prefix_caching: false
    output_connectors:
      to_stage_2: connector_of_shared_memory
    default_sampling_params:
      temperature: 0.9
      top_k: 50
      min_tokens: 2
      max_tokens: 4096
      repetition_penalty: 1.05
    subtalker_sampling_params:
      do_sample: true
      temperature: 0.9
      top_k: 50
      top_p: 1.0

  # GPU 0: Qwen3-TTS Code2Wav.
  - stage_id: 2
    model: Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice
    devices: "0"
    gpu_memory_utilization: 0.2
    max_num_seqs: 8
    max_model_len: 65536
    max_num_batched_tokens: 65536
    enable_prefix_caching: false
    input_connectors:
      from_stage_1: connector_of_shared_memory
    default_sampling_params:
      temperature: 0.0
      top_p: 1.0
      top_k: -1
      max_tokens: 65536
      repetition_penalty: 1.0