[Metrics] Add audio SLOs + cross-stage transfer families + per-(stage, replica) wrap for upstream vllm:* (#3576)

Signed-off-by: vraiti <vraiti@redhat.com>
Signed-off-by: LHXuuu <xulianhao.xlh@antgroup.com>
Co-authored-by: vraiti <vraiti@redhat.com>

vllm_omni/entrypoints/client_request_state.py

@@ -17,3 +17,23 @@ class ClientRequestState:
         self.stage_id: int | None = None
         self.queue = queue if queue is not None else asyncio.Queue()
         self.metrics: OrchestratorAggregator | None = None
+        # Wall-clock time at which the user's request arrived in the engine
+        # entrypoint. Set in async_omni.generate() before the orchestrator
+        # accepts the request. Used as the t0 anchor for audio_ttfp.
+        self.request_arrival_ts: float = 0.0
+        # Wall-clock time at which the first audio packet was observed for
+        # this request. None means the streaming hook hasn't fired yet.
+        # Used as the once-per-request guard for audio_ttfp_s emit.
+        self.first_audio_ts: float | None = None
+        # Per-chunk timeline (seconds since request_arrival_ts) and PCM byte
+        # counts for the audio streaming response. Populated by the streaming
+        # endpoint on every audio.chunk emit; consumed at request finalize to
+        # compute audio_underrun_s and audio_continuity_ok_total.
+        self.audio_chunk_arrivals_s: list[float] = []
+        self.audio_chunk_bytes: list[int] = []
+        self.audio_sample_rate: int | None = None
+        # Stage / replica that produced the audio packets — captured at the
+        # first-packet hook so the finalize-time emit can label correctly
+        # without re-querying stage_pools.
+        self.audio_emit_stage_id: int | None = None
+        self.audio_emit_replica_id: int | None = None