# Ming-omni-tts-16.8B-A3B (MoE) deploy: stage-0 bailing_moe LLM+flow -> stage-1 audio VAE.
# Same two-stage pipeline as the dense ming_tts.yaml; the Stage-0 backbone is the
# 16.8B-A3B bailing_moe MoE (~31GB bf16).
#
# pipeline: HF config reports model_type="bailingmm", which is not a registry key,
# already declared by MING_TTS_MOE_PIPELINE for auto-detection.
#
# NOTE: ~31GB bf16 + the tiny audio VAE fit one 80GB+ GPU, so both stages colocate
# on GPU0 (Stage-0 0.60 util ≈ 58GB + Stage-1 0.25 util ≈ 24GB ≈ 0.85, fits 97GB
# H20). For smaller cards raise Stage-0 tensor_parallel_size or split the stages
# onto separate devices. Multi-request batching is not yet validated (max_num_seqs: 1).
pipeline: ming_tts_moe
async_chunk: true
# The MoE ships a custom tokenizer (tokenization_bailing.BailingTokenizer) that
# must be executed to load. The model config still resolves to the registered
# MingMoeConfig (vLLM _CONFIG_REGISTRY precedence over the repo auto_map).
trust_remote_code: true
dtype: bfloat16
enable_prefix_caching: false

connectors:
  connector_of_shared_memory:
    name: SharedMemoryConnector
    extra:
      latent_chunk_size: 25
      initial_latent_chunk_size: 4
      latent_left_context: 0

stages:
  - stage_id: 0
    max_num_seqs: 1
    tensor_parallel_size: 1
    gpu_memory_utilization: 0.60
    enforce_eager: false
    compilation_config:
      cudagraph_mode: PIECEWISE
    enable_prefix_caching: false
    async_scheduling: false
    max_num_batched_tokens: 8192
    max_model_len: 8192
    # Pin the unquantized MoE kernel (wer diff between vllm 0.22 vs 0.23)
    moe_backend: triton
    devices: "0"
    output_connectors:
      to_stage_1: connector_of_shared_memory
    default_sampling_params:
      temperature: 0.0
      top_p: 1.0
      top_k: -1
      max_tokens: 512
      detokenize: true

  - stage_id: 1
    max_num_seqs: 1
    gpu_memory_utilization: 0.25
    enforce_eager: true
    enable_prefix_caching: false
    async_scheduling: false
    max_num_batched_tokens: 8192
    max_model_len: 8192
    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: 1
      detokenize: false