vllm_omni.model_executor.common.duplex.data_plane ¶

Data plane for models whose final stage streams cumulative audio and text.

The stage runner re-sends the whole waveform (and transcript) it has produced so far for one request; the session needs only what is new. This data plane keeps one cursor per request, slices the delta, encodes it for the wire and reports it as one internal response.output_audio.delta-shaped result. Model-specific constants (stage role, runtime metadata, default rate) are class attributes for a subclass to set.

EncodeAudio = Callable[
    [object, int, str, float | None], str | None
]

default_sample_rate_hz: int = 24000

owned_runtime: bool = False

runner_kv_backed: bool = True

runtime_impl: str = 'scheduler_data_plane'

stage_role: str = 'tts'

uses_model_runner_scheduler: bool = True

begin_request(request_id: str) -> None

close_session(
    session_id: str, *, active_request_id: str | None = None
) -> None

close_stream(request_id: str) -> None

is_terminal(request_id: str | None) -> bool

mark_terminal(request_id: str) -> None

project(
    result: object, *, context: object | None = None
) -> Iterator[dict[str, object]]