# PersonaPlex deploy: talker -> code2wav (Mimi) via shared-memory chunk streaming.
#
# PersonaPlex is a Moshi finetune (full-duplex S2S), served offline as a 2-stage
# audio->audio pipeline that reuses the Qwen3-TTS staged deploy layout. Fields
# omitted from a stage fall back to StageDeployConfig defaults (see
# vllm_omni/config/stage_config.py).
#
# Notable choices for this model:
#   * Stage 0 (talker) keeps a small max_num_batched_tokens for async-chunk
#     first-chunk latency (not correctness).
#   * Stage 1 (code2wav) wraps the external Mimi codec. Each codec frame is
#     8 active codebooks (cb 0..7) -> 1920 PCM samples @ 24 kHz / 12.5 Hz.
#   * Stage 0 runs cudagraph by default; Stage 1 (Mimi) is eager because the
#     external decoder performs host-to-device copies that cannot be captured.
pipeline: personaplex
session_mode: duplex
duplex_session:
  max_sessions: 2

async_chunk: true

connectors:
  connector_of_shared_memory:
    name: SharedMemoryConnector
    extra:
      shm_threshold_bytes: 65536
      codec_streaming: true
      connector_get_sleep_s: 0.01
      connector_get_max_wait_first_chunk: 3000
      connector_get_max_wait: 300
      # Mimi runs at 12.5 Hz (80 ms / frame). Stream a few frames per chunk.
      codec_chunk_frames: 5
      codec_left_context_frames: 10
      # Emit only the first audio chunk early, then return to codec_chunk_frames.
      initial_codec_chunk_frames: 1

stages:
  - stage_id: 0
    max_num_seqs: 10
    gpu_memory_utilization: 0.4
    trust_remote_code: true
    skip_tokenizer_init: true
    # Prompt IDs are placeholders; voice/persona/audio semantics live in
    # request-specific embeddings, so token-ID prefix caching is invalid.
    enable_prefix_caching: false
    async_scheduling: true
    max_num_batched_tokens: 512
    # Temporal transformer sliding_window is 3000 frames; keep model len within.
    max_model_len: 3000
    devices: "0"
    output_connectors:
      to_stage_1: connector_of_shared_memory
    default_sampling_params:
      # Greedy talker decoding: the HeliumModel temporal matches Moshi within cos
      # 0.999 (not bit-exact), and chaotic sampling amplifies that into divergence.
      # Greedy is reliably coherent; raise temperature only once the temporal is
      # bit-exact.
      temperature: 0.0
      top_k: 1
      max_tokens: 3000
      seed: 42
      repetition_penalty: 1.0

  - stage_id: 1
    max_num_seqs: 10
    gpu_memory_utilization: 0.3
    enforce_eager: true
    trust_remote_code: true
    skip_tokenizer_init: true
    enable_prefix_caching: false
    async_scheduling: true
    # Codec prefill length is num_codebooks * num_frames; keep ample headroom.
    max_num_batched_tokens: 65536
    max_model_len: 65536
    devices: "0"
    input_connectors:
      from_stage_0: connector_of_shared_memory
    default_sampling_params:
      temperature: 0.0
      top_p: 1.0
      top_k: -1
      max_tokens: 65536
      seed: 42
      repetition_penalty: 1.0

platforms:
  npu:
    stages:
      - stage_id: 0
        enforce_eager: false