[Feat] Add received_time in serving_base

python/sglang/srt/entrypoints/openai/serving_base.py

@@ -88,6 +88,9 @@ async def handle_request(
         """Handle the specific request type with common pattern
         If you want to override this method, you should be careful to record the validation time.
         """
+        received_time = time.time()
+        received_time_perf = time.perf_counter()
+
         try:
             # Validate request
             validation_start = time.perf_counter()
@@ -103,6 +106,12 @@ async def handle_request(
             if hasattr(adapted_request, "validation_time"):
                 adapted_request.validation_time = validation_time
 
+            if hasattr(adapted_request, "received_time"):
+                adapted_request.received_time = received_time
+
+            if hasattr(adapted_request, "received_time_perf"):
+                adapted_request.received_time_perf = received_time_perf
+
             # Note(Xinyuan): raw_request below is only used for detecting the connection of the client
             if hasattr(request, "stream") and request.stream:
                 return await self._handle_streaming_request(

python/sglang/srt/managers/io_struct.py

@@ -204,6 +204,12 @@ class GenerateReqInput(BaseReq):
     # Validation step duration
     validation_time: Optional[float] = None
 
+    # For metrics
+    received_time: Optional[float] = None
+
+    # Perf_counter equivalents for accurate time calculations
+    received_time_perf: Optional[float] = None
+
     # For data parallel rank routing
     data_parallel_rank: Optional[int] = None
 

python/sglang/srt/managers/tokenizer_manager.py

@@ -431,7 +431,7 @@ async def generate_request(
         obj: Union[GenerateReqInput, EmbeddingReqInput],
         request: Optional[fastapi.Request] = None,
     ):
-        created_time = time.time()
+        created_time = obj.received_time if obj.received_time else time.time()
         self.auto_create_handle_loop()
         obj.normalize_batch_and_arguments()