Skip to content

vllm_omni.config.pipeline_registry

Pipeline registry and factory for vllm-omni.

OMNI_PIPELINES maps each model_type to either a PipelineConfig instance or a resolver callable that accepts an optional HF config and returns a PipelineConfig.

To add a new pipeline
  1. Define the PipelineConfig instance as a module-level variable in vllm_omni/.../pipeline.py.
  2. If the model needs to support several configurations, e.g., because some stages are optional, implement a resolver that consumes the HF config and returns a PipelineConfig.
  3. Update the registry to map the key to the new config object (in the case of new keys) or to the resolver func.

Out of tree pipeline configs or resolvers can also be registered with register_pipeline.

NOTE: Single-stage diffusion models continue to use the _create_default_diffusion_stage_cfg fallback in async_omni_engine.py; for now we do not add them to registry.

OMNI_PIPELINES module-attribute 

OMNI_PIPELINES: dict[
    str, PipelineConfig | PipelineResolverFunc
] = {
    "aura_omni": AURA_OMNI_PIPELINE,
    "qwen2_5_omni": QWEN2_5_OMNI_PIPELINE,
    "qwen2_5_omni_thinker_only": QWEN2_5_OMNI_THINKER_ONLY_PIPELINE,
    "qwen3_omni_moe": resolve_qwen3_omni_pipeline,
    "qwen3_tts": QWEN3_TTS_PIPELINE,
    "covo_audio": COVO_AUDIO_PIPELINE,
    "bagel": BAGEL_PIPELINE,
    "bagel_think": BAGEL_THINK_PIPELINE,
    "bagel_single_stage": BAGEL_SINGLE_STAGE_PIPELINE,
    "lance": LANCE_PIPELINE,
    "dreamzero": DREAMZERO_PIPELINE,
    "Gr00tN1d7": GR00T_N1D7_PIPELINE,
    "glm_image": GLM_IMAGE_PIPELINE,
    "hunyuan_image_3_moe": HUNYUAN_IMAGE3_PIPELINE,
    "hunyuan_image3_ar": HUNYUAN_IMAGE3_AR_PIPELINE,
    "hunyuan_image3_dit": HUNYUAN_IMAGE3_DIT_PIPELINE,
    "voxcpm2": VOXCPM2_PIPELINE,
    "cosyvoice3": COSYVOICE3_PIPELINE,
    "mimo_audio": MIMO_AUDIO_PIPELINE,
    "ming_tts": MING_TTS_PIPELINE,
    "ming_tts_moe": MING_TTS_MOE_PIPELINE,
    "voxtral_tts": VOXTRAL_TTS_PIPELINE,
    "glm_tts": GLM_TTS_PIPELINE,
    "fish_qwen3_omni": FISH_SPEECH_PIPELINE,
    "ming_flash_omni": MING_FLASH_OMNI_PIPELINE,
    "ming_flash_omni_tts": MING_FLASH_OMNI_TTS_PIPELINE,
    "ming_flash_omni_thinker_only": MING_FLASH_OMNI_THINKER_ONLY_PIPELINE,
    "ming_flash_omni_image": MING_FLASH_OMNI_IMAGE_PIPELINE,
    "moss_tts_nano": MOSS_TTS_NANO_PIPELINE,
    "omnivoice": OMNIVOICE_PIPELINE,
    "mammoth_moda2": MAMMOTH_MODA2_PIPELINE,
    "mammoth_moda2_ar": MAMMOTH_MODA2_AR_PIPELINE,
    "moss_tts_delay": MOSS_TTS_PIPELINE,
    "moss_tts_realtime": MOSS_TTS_REALTIME_PIPELINE,
    "moss_tts_local": MOSS_TTS_LOCAL_PIPELINE,
    "minicpmo_4_5": MINICPMO_4_5_PIPELINE,
    "higgs_audio_v2": HIGGS_AUDIO_V2_PIPELINE,
    "higgs_multimodal_qwen3": HIGGS_AUDIO_V3_PIPELINE,
    "dynin_omni": DYNIN_OMNI_PIPELINE,
    "indextts2": INDEXTTS2_PIPELINE,
}

PipelineResolverFunc module-attribute 

PipelineResolverFunc: TypeAlias = Callable[
    [PretrainedConfig | None], PipelineConfig | None
]

logger module-attribute 

logger = init_logger(__name__)

register_pipeline 

register_pipeline(
    pipeline: PipelineConfig | PipelineResolverFunc,
    model_type: str | None = None,
)

Register an out of tree pipeline or PipelineResolverFunc to a model_type key. If a PipelineConfig is provided, model_type is optional, and pipeline.model_type will be used by default. If a callable is provided, model_type must be provided, since resolvers can return multiple different PipelineConfigs depending on the consumed config.