vllm-omni/vllm_omni/deploy/minicpmo_4_5_8x4090.yaml

# MiniCPM-o 4.5 deploy: 8x4090 layout (24GB cards).
#
# Stage 0 (thinker): GPU 0,1,2,3 — 4-way TP, ~85% mem (≈20.4 GB / card,
#                                  thinker weights take ~17.5 GB per card
#                                  on 4090, leaving ~3 GB for KV cache /
#                                  activations).
# Stage 1 (talker) : GPU 4       — MiniCPMTTS codec generation, ~90% mem.
# Stage 2 (code2wav): GPU 5      — codec-to-waveform generation, ~90% mem.
#
# max_model_len is pinned to 4096 because the upstream config.json
# advertises max_position_embeddings=40960 (tuned for H20 141GB); on a
# 24 GB 4090 a 40k-token KV cache request OOMs the worker process
# before serving the first request. 8192 also OOMs in practice with
# this thinker model + vision/audio encoders + Whisper preallocation,
# so 4096 is the empirically-verified safe ceiling.
pipeline: minicpmo_4_5
async_chunk: true
# session_mode: duplex makes vllm-omni serve run this model duplex-only
# (/v1/realtime?duplex=1). The default duplex_session caps sessions at 1,
# matching Stage 0/1 max_num_seqs below.
session_mode: duplex

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
      cfm_max_graphs: 32

stages:
  - stage_id: 0
    max_num_seqs: 1
    gpu_memory_utilization: 0.85
    enforce_eager: true
    tensor_parallel_size: 4
    max_model_len: 4096
    max_num_batched_tokens: 4096
    devices: "0,1,2,3"
    interleave_mm_strings: true
    media_io_kwargs:
      video:
        fps: 1
        num_frames: 128
    limit_mm_per_prompt:
      video: 1
    default_sampling_params:
      temperature: 0.0
      top_p: 1.0
      top_k: -1
      max_tokens: 2048
      seed: 42
      detokenize: true
      repetition_penalty: 1.1

  - stage_id: 1
    max_num_seqs: 1
    gpu_memory_utilization: 0.9
    enforce_eager: true
    max_num_batched_tokens: 8192
    devices: "4"
    output_connectors:
      to_stage_2: connector_of_shared_memory
    default_sampling_params:
      temperature: 0.8
      top_p: 0.85
      top_k: 25
      # See minicpmo_4_5.yaml: the Talker rescores this over a 16-frame window.
      repetition_penalty: 1.05
      min_tokens: 50
      max_tokens: 4096
      seed: 42
      detokenize: false

  - stage_id: 2
    max_num_seqs: 1
    gpu_memory_utilization: 0.9
    enforce_eager: true
    enable_chunked_prefill: false
    max_num_batched_tokens: 65536
    max_model_len: 65536
    devices: "5"
    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: {}