Use monotonic time where appropriate (#1249)

vllm/core/scheduler.py

@@ -121,7 +121,7 @@ class Scheduler:
         blocks_to_copy: Dict[int, List[int]] = {}
 
         # Fix the current time.
-        now = time.time()
+        now = time.monotonic()
 
         # Join waiting sequences if possible.
         if not self.swapped:

vllm/engine/async_llm_engine.py

@@ -417,7 +417,8 @@ class AsyncLLMEngine:
             request.
         """
         # Preprocess the request.
-        arrival_time = time.time()
+        # This should not be used for logging, as it is monotonic time.
+        arrival_time = time.monotonic()
 
         try:
             stream = await self.add_request(request_id,

vllm/engine/llm_engine.py

@@ -256,10 +256,10 @@ class LLMEngine:
             prompt_token_ids: The token IDs of the prompt. If None, we
                 use the tokenizer to convert the prompts to token IDs.
             arrival_time: The arrival time of the request. If None, we use
-                the current time.
+                the current monotonic time.
         """
         if arrival_time is None:
-            arrival_time = time.time()
+            arrival_time = time.monotonic()
         if prompt_token_ids is None:
             assert prompt is not None
             prompt_token_ids = self.tokenizer.encode(prompt)
@@ -568,7 +568,7 @@ class LLMEngine:
         prompt_run: bool,
         num_batched_tokens: int,
     ) -> None:
-        now = time.time()
+        now = time.monotonic()
         # Log the number of batched input tokens.
         if prompt_run:
             self.num_prompt_tokens.append((now, num_batched_tokens))

vllm/entrypoints/openai/api_server.py

@@ -210,7 +210,7 @@ async def create_chat_completion(request: ChatCompletionRequest,
 
     model_name = request.model
     request_id = f"cmpl-{random_uuid()}"
-    created_time = int(time.time())
+    created_time = int(time.monotonic())
     try:
         sampling_params = SamplingParams(
             n=request.n,
@@ -411,7 +411,7 @@ async def create_completion(request: CompletionRequest, raw_request: Request):
     if error_check_ret is not None:
         return error_check_ret
 
-    created_time = int(time.time())
+    created_time = int(time.monotonic())
     try:
         sampling_params = SamplingParams(
             n=request.n,