diff --git a/python/sglang/srt/disaggregation/base/conn.py b/python/sglang/srt/disaggregation/base/conn.py
index a7bf9904a..fbd9186b5 100644
--- a/python/sglang/srt/disaggregation/base/conn.py
+++ b/python/sglang/srt/disaggregation/base/conn.py
@@ -24,6 +24,8 @@ class StateType(str, enum.Enum):
 class KVTransferMetric:
     # Backends that cannot isolate transfer latency can leave this as None.
     transfer_latency_s: Optional[float] = None
+    # Backends that cannot isolate allocation wait latency can leave this as None.
+    alloc_latency_s: Optional[float] = None
     transfer_total_bytes: Optional[int] = None
 
 
diff --git a/python/sglang/srt/observability/req_time_stats.py b/python/sglang/srt/observability/req_time_stats.py
index 2de10730c..f672045ad 100644
--- a/python/sglang/srt/observability/req_time_stats.py
+++ b/python/sglang/srt/observability/req_time_stats.py
@@ -903,7 +903,13 @@ class SchedulerReqTimeStats(ReqTimeStatsBase):
             bootstrap_ms = (
                 self.bootstrap_done_time - self.prefill_bootstrap_queue_entry_time
             ) * 1000
-            alloc_ms = (self.wait_queue_entry_time - self.bootstrap_done_time) * 1000
+            if transfer_metric.alloc_latency_s is not None:
+                alloc_ms = transfer_metric.alloc_latency_s * 1000
+            else:
+                alloc_ms = (
+                    max(0.0, self.wait_queue_entry_time - self.bootstrap_done_time)
+                    * 1000
+                )
 
             result["bootstrap_ms"] = bootstrap_ms
             result["alloc_ms"] = alloc_ms