[BugFix] Fix inference_start_time (#4922) (#4930)

* fix inference_start_time

* fix inference_start_time

Author: kxz2002

fastdeploy/entrypoints/openai/serving_chat.py

@@ -192,6 +192,7 @@ async def chat_completion_stream_generator(
         num_cached_tokens = 0
         num_image_tokens = [0] * num_choices
         tool_called = [False] * num_choices
+        inference_start_time = [0] * num_choices
         max_streaming_response_tokens = (
             request.max_streaming_response_tokens
             if request.max_streaming_response_tokens is not None
@@ -268,9 +269,9 @@ async def chat_completion_stream_generator(
 
                     if res["metrics"]["first_token_time"] is not None:
                         arrival_time = res["metrics"]["first_token_time"]
-                        inference_start_time = res["metrics"]["inference_start_time"]
+                        inference_start_time[idx] = res["metrics"]["inference_start_time"]
                     else:
-                        arrival_time = res["metrics"]["arrival_time"] - inference_start_time
+                        arrival_time = res["metrics"]["arrival_time"] - inference_start_time[idx]
                     if first_iteration:
                         num_prompt_tokens = len(prompt_token_ids)
                         num_cached_tokens = res.get("num_cached_tokens", 0)