Skip to content

vllm_omni.model_executor.models.voxcpm2.pipeline ¶

VoxCPM2 pipeline topology (frozen).

Single-stage AR TTS: text → speech waveform in one pass. Uses the native MiniCPM4 base_lm with a per-request StaticKVCache that the talker restores into the paged attention layer at step boundaries.

VOXCPM2_PIPELINE `module-attribute` ¶

VOXCPM2_PIPELINE = PipelineConfig(
    model_type="voxcpm2",
    model_arch="VoxCPM2TalkerForConditionalGeneration",
    stages=(
        StagePipelineConfig(
            stage_id=0,
            model_stage="latent_generator",
            execution_type=LLM_AR,
            input_sources=(),
            final_output=True,
            final_output_type="audio",
            owns_tokenizer=True,
            engine_output_type="audio",
            sampling_constraints={
                "detokenize": False,
                "stop_token_ids": [1],
            },
        ),
    ),
)