vllm_omni.model_executor.common.duplex.pcm_buffer ¶

Transactional framing of client PCM into fixed model units.

A frame-locked model (PersonaPlex: 1920 samples every 80 ms at 24 kHz) takes exactly one frame per append. Clients send arbitrary chunks, so the session runner buffers them here and takes one whole frame out at a time as a reservation: committed once the stage accepted the append, rolled back (the bytes go back to the front of the buffer) when it did not.

chunk_period_ms = int(chunk_period_ms)

frame_bytes: int

frame_samples = int(frame_samples)

model = model

pending_byte_count: int

sample_rate_hz = int(sample_rate_hz)

clear() -> None

clear_force_listen() -> None

flush(*, chunk_period_ms: int) -> dict[str, object] | None

has_pending() -> bool

has_reserved() -> bool

prepare_append(
    payload: dict[str, object],
    *,
    operation_id: str,
    chunk_period_ms: int,
    allow_emit: bool,
) -> FixedFramePcmAppendReservation | None

prepare_commit(
    *, operation_id: str, chunk_period_ms: int
) -> FixedFramePcmAppendReservation

active: bool

byte_count: int

operation_id = operation_id

payload = payload

commit() -> None

rollback() -> None