# VoxCPM two-stage (latent → VAE) without async_chunk: one-shot latent then decode.
stage_args:
  - stage_id: 0
    stage_type: llm
    is_comprehension: true
    runtime:
      devices: "0"
      max_batch_size: 1
    engine_args:
      dtype: bfloat16
      model_stage: latent_generator
      model_arch: VoxCPMForConditionalGeneration
      # Optional persistent HF-compatible config dir for native VoxCPM models.
      hf_config_path: ${oc.env:VLLM_OMNI_VOXCPM_HF_CONFIG_PATH,}
      worker_type: ar
      scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
      enforce_eager: true
      trust_remote_code: true
      async_scheduling: false
      enable_prefix_caching: false
      engine_output_type: latent
      gpu_memory_utilization: 0.7
      distributed_executor_backend: "mp"
      max_num_batched_tokens: 4096
      max_model_len: 4096
    default_sampling_params:
      temperature: 0.0
      top_p: 1.0
      top_k: -1
      max_tokens: 4096
      stop_token_ids: [2]
      seed: 42
      detokenize: false
      repetition_penalty: 1.0
    final_output: false

  - stage_id: 1
    stage_type: llm
    runtime:
      devices: "0"
      max_batch_size: 1
    engine_args:
      dtype: float32
      model_stage: vae
      model_arch: VoxCPMForConditionalGeneration
      hf_config_path: ${oc.env:VLLM_OMNI_VOXCPM_HF_CONFIG_PATH,}
      worker_type: generation
      scheduler_cls: vllm_omni.core.sched.omni_generation_scheduler.OmniGenerationScheduler
      enforce_eager: true
      trust_remote_code: true
      async_scheduling: false
      enable_prefix_caching: false
      engine_output_type: audio
      gpu_memory_utilization: 0.15
      distributed_executor_backend: "mp"
      max_num_batched_tokens: 8192
      max_model_len: 4096
    engine_input_source: [0]
    custom_process_input_func: vllm_omni.model_executor.stage_input_processors.voxcpm.latent2vae
    final_output: true
    final_output_type: audio
    default_sampling_params:
      temperature: 0.0
      top_p: 1.0
      top_k: -1
      max_tokens: 1
      seed: 42
      detokenize: true
      repetition_penalty: 1.0