Skip to content

vllm_omni.model_executor.models.covo_audio.pipeline

Covo-Audio pipeline topology (frozen).

fused_thinker_talker — multimodal understanding + interleaved text/audio

token generation (AR).

Stage 1: code2wav — audio codes → waveform (generation).

COVO_AUDIO_PIPELINE module-attribute 

COVO_AUDIO_PIPELINE = PipelineConfig(
    model_type="covo_audio",
    model_arch="CovoAudioForConditionalGeneration",
    stages=(
        StagePipelineConfig(
            stage_id=0,
            model_stage="fused_thinker_talker",
            execution_type=LLM_AR,
            input_sources=(),
            final_output=True,
            final_output_type="text",
            owns_tokenizer=True,
            requires_multimodal_data=True,
            engine_output_type="latent",
            custom_process_next_stage_input_func=f"{_PROC}.llm2code2wav_full_payload",
            sampling_constraints={
                "detokenize": True,
                "stop_token_ids": [151645],
                "ignore_eos": True,
            },
        ),
        StagePipelineConfig(
            stage_id=1,
            model_stage="code2wav",
            execution_type=LLM_GENERATION,
            input_sources=(0,),
            final_output=True,
            final_output_type="audio",
            engine_output_type="audio",
            custom_process_input_func=f"{_PROC}.llm2code2wav",
            sync_process_input_func=f"{_PROC}.llm2code2wav_token_only",
            sampling_constraints={"detokenize": False},
        ),
    ),
)