[Structured Output] Let terminal grammars stop under min_tokens (II) (#54218)

Signed-off-by: Artem Perevedentsev <aperevedents@nvidia.com>
This commit is contained in:
Artem Perevedentsev
2026-08-31 20:39:27 +04:00
committed by GitHub
parent e9dd6d4834
commit 39e276eaeb
7 changed files with 501 additions and 60 deletions
+9 -3
View File
@@ -97,9 +97,6 @@ def check_stop(request: Request, max_model_len: int) -> bool:
sampling_params = request.sampling_params
assert sampling_params is not None
if request.num_output_tokens < sampling_params.min_tokens:
return False
last_token_id = request.output_token_ids[-1]
if last_token_id == sampling_params.eos_token_id:
request.status = RequestStatus.FINISHED_STOPPED
@@ -109,6 +106,7 @@ def check_stop(request: Request, max_model_len: int) -> bool:
request.status = RequestStatus.FINISHED_STOPPED
request.stop_reason = last_token_id
return True
if (
request.num_tokens >= max_model_len
or request.num_output_tokens >= request.max_tokens
@@ -116,6 +114,14 @@ def check_stop(request: Request, max_model_len: int) -> bool:
request.status = RequestStatus.FINISHED_LENGTH_CAPPED
return True
# Note(arpera):
# Order of checks is important for min_tokens
# If you decide to change the order, first look at these PRs please:
# 47489, 49521, and 51299
# These PRs show some reasoning about the order of checks
if request.num_output_tokens < sampling_params.min_tokens:
return False
repetition_detection = sampling_params.repetition_detection
if repetition_detection is not None and (
check_sequence_repetition(