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 13, 02:00 UTC
[Misc] Fix wrong comment in scheduler (#28880)
Signed-off-by: Zhuohan Li <zhuohan123@gmail.com>
Zhuohan Li committed
Nov 17, 2025 at 15:32 UTC
552cac95b5da283844a9994b94d4b1308a0a0565
1 file changed
+4
-4
vllm/v1/core/sched/scheduler.py
+4
-4
@@ -472,9 +472,9 @@ class Scheduler(SchedulerInterface):
472
num_computed_tokens = (
473
num_new_local_computed_tokens + num_external_computed_tokens
474
)
475
- # KVTransfer: WAITING reqs have num_computed_tokens > 0
476
- # after async KV recvs are completed.
475
else:
476
+ # KVTransfer: WAITING reqs have num_computed_tokens > 0
477
+ # after async KV recvs are completed.
478
new_computed_blocks = self.kv_cache_manager.empty_kv_cache_blocks
479
num_new_local_computed_tokens = 0
480
num_computed_tokens = request.num_computed_tokens
@@ -483,12 +483,12 @@ class Scheduler(SchedulerInterface):
483
external_load_encoder_input = []
484
new_encoder_compute_budget = encoder_compute_budget
485
486
- # KVTransfer: loading remote KV, do not allocate for new work.
486
if load_kv_async:
487
+ # KVTransfer: loading remote KV, do not allocate for new work.
488
assert num_external_computed_tokens > 0
489
num_new_tokens = 0
490
- # Number of tokens to be scheduled.
490
else:
491
+ # Number of tokens to be scheduled.
492
# We use `request.num_tokens` instead of
493
# `request.num_prompt_tokens` to consider the resumed
494
# requests, which have output tokens.