[Core] Add span metrics for model_forward, scheduler and sampler time (#7089)

vllm/config.py

@@ -1656,11 +1656,26 @@ class ObservabilityConfig:
     """Configuration for observability."""
     otlp_traces_endpoint: Optional[str] = None
 
+    # Collecting detailed timing information for each request can be expensive.
+
+    # If set, collects the model forward time for the request.
+    collect_model_forward_time: bool = False
+
+    # If set, collects the model execute time for the request.
+    collect_model_execute_time: bool = False
+
     def __post_init__(self):
         if not is_otel_installed() and self.otlp_traces_endpoint is not None:
             raise ValueError("OpenTelemetry packages must be installed before "
                              "configuring 'otlp_traces_endpoint'")
 
+        if ((self.collect_model_forward_time
+             or self.collect_model_execute_time)
+                and self.otlp_traces_endpoint is None):
+            raise ValueError(
+                "collect_model_forward_time or collect_model_execute_time "
+                "requires --otlp-traces-endpoint to be set.")
+
 
 @dataclass(frozen=True)
 class EngineConfig: