[CI/Build] Fix Qwen3-Omni Talker benchmark workload (#8057)

Signed-off-by: psv666 <2693925048@qq.com>

vllm_omni/benchmarks/patch/patch.py

@@ -1203,6 +1203,27 @@ def _update_output_stage_metrics_from_payload(
         output.stage_metrics.update(stage_snapshot)
 
 
+# Per-request stage fields persisted in benchmark results. Full snapshots carry
+# per-token latency lists, which would inflate every chat-omni result file.
+_REQUEST_STAGE_METRIC_FIELDS = (
+    defs.NUM_TOKENS_OUT,
+    "finish_reason",
+    defs.AUDIO_FRAMES,
+    f"{defs.AUDIO_DURATION}_s",
+)
+
+
+def _compact_request_stage_metrics(snapshot: object) -> dict[str, dict] | None:
+    """Keep the stage fields used for workload checks; empty snapshots become None."""
+    if not isinstance(snapshot, dict) or not snapshot:
+        return None
+    return {
+        stage: {field: metrics[field] for field in _REQUEST_STAGE_METRIC_FIELDS if field in metrics}
+        for stage, metrics in snapshot.items()
+        if isinstance(metrics, dict)
+    }
+
+
 def _apply_chat_stage0_token_timings(output: MixRequestFuncOutput) -> bool:
     """Apply native Stage 0 timings from a chat response snapshot."""
     stage_metrics = output.stage_metrics
@@ -3635,6 +3656,14 @@ async def benchmark(
             "input_lens": [output.prompt_len for output in outputs],
             "errors": [output.error for output in outputs],
         }
+    # Preserve request order, including missing snapshots, so CI can verify
+    # fixed stage workloads without parsing logs or storing audio payloads.
+    request_stage_metrics = [
+        _compact_request_stage_metrics(getattr(output, "stage_metrics", None)) for output in outputs
+    ]
+    if any(request_stage_metrics):
+        result["request_stage_metrics"] = request_stage_metrics
+
     # Plain-vLLM backends (e.g. the vLLM-text perf config) return upstream
     # RequestFuncOutput objects without the Mix duplex fields; read them
     # tolerantly or the whole benchmark result is discarded ("fallback to