A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 15, 19:45 UTC
[V1][Minor] Minor enhancements on scheduler (#14732)
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
Woosuk Kwon committed
Mar 13, 2025 at 08:53 UTC
01b3fd0af7d487f1a3b5aa11807b78d1f330ab95
1 file changed
+3
-5
vllm/v1/core/scheduler.py
+3
-5
@@ -587,9 +587,6 @@ class Scheduler:
587
if spec_token_ids is not None:
588
request.spec_token_ids = spec_token_ids[req_index]
589
590
- # Get prompt logprobs for this request.
591
- prompt_logprobs_tensors = prompt_logprobs_dict.get(req_id)
592
-
590
stopped = False
591
new_logprobs = None
592
new_token_ids: list[int] = []
@@ -622,6 +619,8 @@ class Scheduler:
619
new_token_ids,
620
)
621
622
+ # Get prompt logprobs for this request.
623
+ prompt_logprobs_tensors = prompt_logprobs_dict.get(req_id)
624
# Transmit partial if chunked prefill & prompt logprobs is enabled
625
if new_token_ids or prompt_logprobs_tensors is not None:
626
# Add EngineCoreOutput for this Request.
@@ -693,8 +692,7 @@ class Scheduler:
692
693
if request.status == RequestStatus.RUNNING:
694
self.running.remove(request)
696
- if request.request_id in self.scheduled_req_ids:
697
- self.scheduled_req_ids.remove(request.request_id)
695
+ self.scheduled_req_ids.discard(request.request_id)
696
else:
697
self.waiting.remove(request)
698
request.status = finished_status