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 16, 09:30 UTC
[Core] Refactor: Clean up unused argument in Scheduler._preempt (#9696)
Signed-off-by: André Jonasson <andre.jonasson@gmail.com>
André Jonasson committed
Nov 1, 2024 at 19:41 UTC
4581d2cc02f655e76233f9cb129f07c6b65d39f4
1 file changed
+3
-8
vllm/core/scheduler.py
+3
-8
@@ -828,8 +828,7 @@ class Scheduler:
828
num_running_seqs)
829
830
#Preempt out the victim sequence group
831
- self._preempt(vseq_group, blocks_to_swap_out,
832
- PreemptionMode.RECOMPUTE)
831
+ self._preempt(vseq_group, blocks_to_swap_out)
832
waiting_queue.appendleft(vseq_group)
833
force_preemption_count += 1
834
#Put the sequence back into the waiting queue
@@ -1451,12 +1450,8 @@ class Scheduler:
1450
if len(cows) > 0:
1451
blocks_to_copy.extend(cows)
1452
1454
- def _preempt(
1455
- self,
1456
- seq_group: SequenceGroup,
1457
- blocks_to_swap_out: List[Tuple[int, int]],
1458
- preemption_mode: Optional[PreemptionMode] = None,
1459
- ) -> PreemptionMode:
1453
+ def _preempt(self, seq_group: SequenceGroup,
1454
+ blocks_to_swap_out: List[Tuple[int, int]]) -> PreemptionMode:
1455
# If preemption mode is not specified, we determine the mode as follows:
1456
# We use recomputation by default since it incurs lower overhead than
1457
# swapping. However, when the sequence group has multiple sequences