# Ming-flash-omni-2.0 deploy: thinker + talker.
#
# Stage 0: Thinker: multimodal understanding -> text generation.
#          TP=4 (or PP=4) across GPUs 0-3.
# Stage 1: Talker: text -> audio waveform via CFM + AudioVAE.
#          Generation worker, runs eager, colocates on the last GPU.
#
# Fields omitted from a stage fall back to StageDeployConfig dataclass
# defaults (see vllm_omni/config/stage_config.py).
#   * Stage 0 (thinker) owns GPUs 0-3 with gpu_memory_utilization=0.74,
#     as we tend to fit talker in one of the devices. (For now,
#     gpu mem util on different devices are not supported to be set
#     independently for a single stage.)
#   * Stage 1 (talker) colocates on GPU 3 with gpu_memory_utilization=0.18.
#     This custom settings is verified and required on devices with 80GB memory.
pipeline: ming_flash_omni
async_chunk: false

trust_remote_code: true
distributed_executor_backend: mp
enable_prefix_caching: false

stages:
  - stage_id: 0
    max_num_seqs: 1
    gpu_memory_utilization: 0.74
    enforce_eager: false
    tensor_parallel_size: 4
    max_num_batched_tokens: 32768
    devices: "0,1,2,3"
    # disable fused allreduce+rms for flashinfer version mismatch for now
    compilation_config:
      pass_config:
        fuse_allreduce_rms: false
    default_sampling_params:
      temperature: 0.4
      top_p: 0.9
      max_tokens: 2048
      repetition_penalty: 1.05
      detokenize: true
      ignore_eos: false

  - stage_id: 1
    max_num_seqs: 1
    gpu_memory_utilization: 0.18
    enforce_eager: true
    max_num_batched_tokens: 1000000
    devices: "3"
    default_sampling_params:
      temperature: 0.0
      max_tokens: 1

# Ming-flash-omni-2.0 has only been verified on CUDA devices,
# left empty intentionally. Add per-platform overrides here
# when other backends (ROCm/XPU/NPU) are verified end-to-end.
platforms: {}