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 14, 05:30 UTC
[Bugfix] fixed inconsistent finish_reason handling between V0 and V1 engines (#27555)
Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>
Chauncey committed
Oct 28, 2025 at 10:18 UTC
61fbfe52742bb34ce8a04d3fb734582268bdcd10
1 file changed
+6
-7
vllm/v1/core/sched/utils.py
+6
-7
@@ -42,13 +42,6 @@ def remove_all(lst: list, items_to_remove: set) -> list:
42
def check_stop(
43
request: Request, max_model_len: int, pooler_output: torch.Tensor | None = None
44
) -> bool:
45
- if (
46
- request.num_tokens >= max_model_len
47
- or request.num_output_tokens >= request.max_tokens
48
- ):
49
- request.status = RequestStatus.FINISHED_LENGTH_CAPPED
50
- return True
51
-
45
if request.pooling_params:
46
if pooler_output is not None:
47
request.status = RequestStatus.FINISHED_STOPPED
@@ -70,4 +63,10 @@ def check_stop(
63
request.status = RequestStatus.FINISHED_STOPPED
64
request.stop_reason = last_token_id
65
return True
66
+ if (
67
+ request.num_tokens >= max_model_len
68
+ or request.num_output_tokens >= request.max_tokens
69
+ ):
70
+ request.status = RequestStatus.FINISHED_LENGTH_CAPPED
71
+ return True
72
return False