Fixes the misuse/mixuse of time.time()/time.monotonic() (#3220)

Signed-off-by: Tao He <sighingnow@gmail.com>
Co-authored-by: simon-mo <simon.mo@hey.com>

vllm/core/scheduler.py

@@ -160,7 +160,7 @@ class Scheduler:
         blocks_to_copy: Dict[int, List[int]] = {}
 
         # Fix the current time.
-        now = time.monotonic()
+        now = time.time()
 
         # Join waiting sequences if possible.
         if not self.swapped:

vllm/engine/async_llm_engine.py

@@ -604,8 +604,7 @@ class AsyncLLMEngine:
             >>> ...
         """
         # Preprocess the request.
-        # This should not be used for logging, as it is monotonic time.
+        arrival_time = time.time()
-        arrival_time = time.monotonic()
 
         try:
             stream = await self.add_request(

vllm/engine/llm_engine.py

@@ -244,7 +244,7 @@ class LLMEngine:
             raise ValueError(f"Cannot request more than "
                              f"{max_logprobs} logprobs.")
         if arrival_time is None:
-            arrival_time = time.monotonic()
+            arrival_time = time.time()
         prompt_token_ids = self.encode_request(
             request_id=request_id,
             prompt=prompt,
@@ -628,7 +628,7 @@ class LLMEngine:
     def _get_stats(self,
                    scheduler_outputs: Optional[SchedulerOutputs]) -> Stats:
         """Get Stats to be Logged to Prometheus."""
-        now = time.monotonic()
+        now = time.time()
 
         # KV Cache Usage in %.
         num_total_gpu = self.cache_config.num_gpu_blocks

vllm/entrypoints/openai/serving_chat.py

@@ -103,7 +103,7 @@ class OpenAIServingChat(OpenAIServing):
     ) -> Union[ErrorResponse, AsyncGenerator[str, None]]:
 
         model_name = request.model
-        created_time = int(time.monotonic())
+        created_time = int(time.time())
         chunk_object_type = "chat.completion.chunk"
         first_iteration = True
 
@@ -244,7 +244,7 @@ class OpenAIServingChat(OpenAIServing):
             request_id: str) -> Union[ErrorResponse, ChatCompletionResponse]:
 
         model_name = request.model
-        created_time = int(time.monotonic())
+        created_time = int(time.time())
         final_res: RequestOutput = None
 
         async for res in result_generator:

vllm/entrypoints/openai/serving_completion.py

@@ -118,7 +118,7 @@ class OpenAIServingCompletion(OpenAIServing):
 
         model_name = request.model
         request_id = f"cmpl-{random_uuid()}"
-        created_time = int(time.monotonic())
+        created_time = int(time.time())
 
         # Schedule the request and get the result generator.
         generators = []