[observability] Fix missing e2e/decode/inference latency span attributes (#37789)

python/sglang/srt/managers/tokenizer_manager.py

@@ -2494,11 +2494,12 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
                 state.time_stats.set_first_token_time()
 
             if state.finished:
-                if state.time_stats.trace_ctx.tracing_enable:
-                    state.time_stats.trace_ctx.trace_set_root_attrs(
-                        self.convert_to_span_attrs(state, recv_obj, i)
-                    )
-                state.time_stats.set_finished_time()
+                span_attrs = (
+                    self.convert_to_span_attrs(state, recv_obj, i)
+                    if state.time_stats.trace_ctx.tracing_enable
+                    else None
+                )
+                state.time_stats.set_finished_time(span_attrs=span_attrs)
                 meta_info["e2e_latency"] = state.time_stats.get_e2e_latency()
 
                 if get_spec().speculative_algorithm:
@@ -3658,8 +3659,8 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
                 [finish_reason]
             )
 
-        # Latency attributes
-        span_attrs.update(state.time_stats.convert_to_gen_ai_span_attrs())
+        # Latency attributes are added by set_finished_time(), which stamps
+        # finished_time before deriving them.
 
         return span_attrs
 

python/sglang/srt/observability/req_time_stats.py

@@ -415,12 +415,18 @@ class APIServerReqTimeStats(ReqTimeStatsBase):
                 convert_time_to_realtime_ns(ts),
             )
 
-    def set_finished_time(self, ts=None):
+    def set_finished_time(self, ts=None, span_attrs=None):
         ts = ts or time.perf_counter()
         self.finished_time = ts
 
         if self.trace_ctx.tracing_enable:
-            self.trace_ctx.trace_req_finish(convert_time_to_realtime_ns(ts))
+            # The latency attrs are derived from finished_time and the root span is
+            # closed below, so they must be merged in here rather than by the caller.
+            attrs = dict(span_attrs) if span_attrs else {}
+            attrs.update(self.convert_to_gen_ai_span_attrs())
+            self.trace_ctx.trace_req_finish(
+                convert_time_to_realtime_ns(ts), attrs=attrs
+            )
 
     def set_first_token_time(self, ts=None):
         ts = ts or time.perf_counter()