vllm_omni.protocol.realtime.session ¶

Reading the Realtime session object.

session.update carries one JSON object that a client may spell in either of two shapes: the GA nest (session.audio.input.format, session.audio.output.sample_rate_hz) or the older flat beta keys (input_audio_format, sample_rate_hz). Everything downstream --- the append decoder, the projection that renders session.created back out, a model's runtime configuration --- has to agree on what a given session object means, so the readers live here once.

:class:RealtimeInputDefaults is the resolved answer for the input path: the format, sample rate and silence threshold an input_audio_buffer.append inherits when it omits them. A connection keeps one of these and re-derives it on every session.update.

overlap_policy / playback_commit_policy are vLLM-Omni extension fields on the session object rather than GA OpenAI fields. They are read here because this module owns what the wire may say; what a server does with them is the consumer's business.

input_audio_format: str = 'pcm16'

input_sample_rate_hz: int = 16000

output_audio_format: str = 'pcm16'

output_sample_rate_hz: int | None = None

overlap_silence_rms: float = 0.003

with_session_payload(
    session_payload: Mapping[str, object],
) -> RealtimeInputDefaults

apply_realtime_session_defaults(
    defaults: RealtimeInputDefaults,
    session_payload: Mapping[str, object],
) -> RealtimeInputDefaults

input_audio_transcription_config(
    session_payload: Mapping[str, object],
) -> dict[str, object] | None

json_safe_realtime_payload(
    payload: Mapping[str, object],
) -> dict[str, object]

realtime_max_output_tokens(value: object) -> int | None

realtime_overlap_fields(
    session_payload: Mapping[str, object],
) -> dict[str, object]