Wire protocols vLLM-Omni speaks, independent of who serves or runs them.
A module under vllm_omni.protocol describes what is on the wire: how a client payload is parsed and validated, and what shape a server payload has. It never decides how a request is executed, which model answers it, or where the session state lives. That keeps one protocol implementation usable by more than one runtime (see vllm_omni.protocol.realtime).
Dependency rule, asserted by tests/protocol/realtime/test_protocol_import_boundary.py: nothing here may import vllm_omni.engine, vllm_omni.entrypoints, vllm_omni.model_executor, vllm_omni.worker or vllm_omni.clients. The arrows point inwards --- those packages import the protocol, never the other way round.
Modules:
| Name | Description |
|---|---|
duplex | vLLM-Omni's full-duplex extension of the OpenAI Realtime wire protocol. |
realtime | The OpenAI Realtime wire codec, shared by every vLLM-Omni Realtime surface. |