[Bugfix] Fix ITL recording in serving benchmark (vllm-project#7372)

charlifu · Aug 9, 2024 · 70d268a · 70d268a
1 parent 249b882
commit 70d268a
Showing 1 changed file with 3 additions and 2 deletions.
diff --git a/benchmarks/backend_request_func.py b/benchmarks/backend_request_func.py
@@ -276,8 +276,9 @@ async def async_request_openai_completions(
                                     output.ttft = ttft
 
                                 # Decoding phase
-                                output.itl.append(timestamp -
-                                                  most_recent_timestamp)
+                                else:
+                                    output.itl.append(timestamp -
+                                                      most_recent_timestamp)
 
                                 most_recent_timestamp = timestamp
                                 generated_text += data["choices"][0]["text"]