Add request-level timestamp for when prefill finishes (#14860)

python/sglang/srt/metrics/collector.py

@@ -64,6 +64,13 @@ class TimeStats:
     prefill_start_time_host: float = 0.0
     prefill_end_time_host: float = 0.0
 
+    # Timestamp when prefill phase finishes, obtained from `time.time()`.
+    # Note that this differs from the other `_time` fields tracked by the
+    # `TimeStats` class, which are obtained from `time.perf_counter()`.
+    # We use `time.time()` instead of `time.perf_counter()` here in order to
+    # maintain unit consistency with other timestamp fields tracked by the `ReqState` class.
+    prefill_finished_ts: float = 0.0
+
     def get_queueing_time(self) -> float:
         return self.forward_entry_time - self.wait_queue_entry_time
 
@@ -77,6 +84,11 @@ class TimeStats:
             return self.prefill_end_time_host - self.prefill_start_time_host
         return None
 
+    def get_prefill_finished_ts(self) -> Optional[float]:
+        if self.prefill_finished_ts > 0.0:
+            return self.prefill_finished_ts
+        return None
+
     def convert_to_duration(self) -> str:
         if self.disagg_mode == DisaggregationMode.NULL:
             queue_duration = self.forward_entry_time - self.wait_queue_entry_time