vllm-omni/vllm_omni/deploy/minicpmo_4_5.yaml

# MiniCPM-o 4.5 three-stage deploy on one large-memory GPU.
pipeline: minicpmo_4_5
async_chunk: true
# Use minicpmo_4_5_turn.yaml to select the ordinary turn-based serving engine.
session_mode: duplex
# Keep the active window aligned with the session and stage capacities below.
# A smaller window serializes first-packet generation for concurrent chat turns.
duplex_session:
  # Matches Stage 0/1 max_num_seqs below.
  max_sessions: 16

trust_remote_code: true
enable_prefix_caching: false

connectors:
  connector_of_shared_memory:
    name: SharedMemoryConnector
    extra:
      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: 3
      # HiFT capture buckets are derived from codec_chunk_frames and
      # codec_left_context_frames. (#5869)
      # A future initial_codec_chunk_frames shape will use lazy capture unless
      # HiFTGraphWrapper.derive_capture_bucket_size explicitly includes it.
      enable_hift_graph: true
      enable_cfm_graph: true
      # Whole-Euler CUDA graphs exist for batch 1, 2, 4, ... up to
      # micro_batch_size, which defaults to Stage 2 max_num_seqs: all graphs
      # share one attention-cache arena of that many requests (~0.6 GiB each
      # in fp32, 10 GiB at 16). Larger batches replay micro-batches;
      # B > max_graph_batch runs eager.
      max_graph_batch: 16
      cfm_max_graphs: 32

stages:
  - stage_id: 0
    max_num_seqs: 16
    # Leave headroom on the shared GPU for Talker KV + concurrent Code2Wav.
    gpu_memory_utilization: 0.55
    enforce_eager: false
    tensor_parallel_size: 1
    max_num_batched_tokens: 16384
    devices: "0"
    # Required for MiniCPM interleaved image/audio Daily-Omni packs (~78% recipe).
    # Keep these in YAML so they are applied through --deploy-config.
    interleave_mm_strings: true
    media_io_kwargs:
      video:
        fps: 1
        num_frames: 128
    limit_mm_per_prompt:
      # Daily-Omni: up to 64 interleaved image/audio pairs (1fps).
      image: 64
      audio: 64
      video: 1
    default_sampling_params:
      temperature: 0.0
      top_p: 1.0
      top_k: -1
      max_tokens: 2048
      seed: 42
      detokenize: true
      # 1.0 for MCQ accuracy; 1.1 can bias short A–D answers.
      repetition_penalty: 1.0

  - stage_id: 1
    max_num_seqs: 16
    # Preserve the checkpoint's full-attention Talker prompt while it fits;
    # native duplex rolls over to a one-chunk window at context capacity.
    # Leave device-level headroom for the colocated HiFi-GAN cuDNN workspace.
    gpu_memory_utilization: 0.15
    enforce_eager: false
    max_num_batched_tokens: 8192
    # Talker context is 4096 (tts_config.max_position_embeddings).
    max_model_len: 4096
    devices: "0"
    output_connectors:
      to_stage_2: connector_of_shared_memory
    # Talker is a single-vocab codec LM. These knobs go to vLLM's Sampler
    # (head_code logits). EOS is tts_config.eos_token_id (num_audio_tokens-1).
    # Values track upstream utils.TTSSamplingParams (min_p/win_size/tau_r are
    # declared there but never read, so they stay out). repetition_penalty is
    # consumed by the Talker, which rescores it over upstream's 16-frame window
    # instead of the Sampler's whole-stream presence penalty.
    # min_tokens is upstream's min_new_token.
    default_sampling_params:
      temperature: 0.8
      top_p: 0.85
      top_k: 25
      repetition_penalty: 1.05
      min_tokens: 50
      max_tokens: 4096
      seed: 42
      detokenize: false

  - stage_id: 2
    max_num_seqs: 16
    gpu_memory_utilization: 0.18
    enforce_eager: true
    enable_chunked_prefill: false
    max_num_batched_tokens: 65536
    max_model_len: 65536
    devices: "0"
    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
      detokenize: true

platforms:
  cuda:
    stages:
      - stage_id: 1
        # Full-attention occupancy is bounded by max_num_seqs * max_model_len
        # (16 * 4096 = 65536 tokens). Keep the reported CUDA KV token capacity
        # above that bound because physical-preemption replay is unsupported
        # (2 GiB holds ~35k tokens). The explicit cap leaves device headroom
        # for concurrent Stage 2.
        kv_cache_memory_bytes: 4294967296  # 4 GiB

  npu:
    stages:
      - stage_id: 0
        max_num_batched_tokens: 8192
        compilation_config:
          cudagraph_mode: PIECEWISE
      - stage_id: 1
        max_num_batched_tokens: 8192
        compilation_config:
          cudagraph_mode: PIECEWISE
      - stage_id: 2
        # Stage 2 stays outer-eager and captures only exact-shape CFM DiT
        # estimator kernels through the NPU platform patch.
        additional_config:
          code2wav_enable_npu_graph: true
          code2wav_max_npu_graphs: 32