[Frontend] Skip stop in reasoning content (#14550)

Signed-off-by: Ce Gao <cegao@tensorchord.ai>
Co-authored-by: Chauncey <chaunceyjiang@gmail.com>
This commit is contained in:
Ce Gao
2025-09-15 06:04:55 +00:00
committed by GitHub
co-authored by Chauncey
parent 3f3313981c
commit f4a948f33f
3 changed files with 256 additions and 4 deletions
+15 -2
View File
@@ -40,6 +40,7 @@ from vllm.multimodal.cache import processor_only_cache_from_config
from vllm.multimodal.processing import EncDecMultiModalProcessor
from vllm.outputs import (PoolingRequestOutput, RequestOutput,
RequestOutputFactory)
from vllm.reasoning import ReasoningParser, ReasoningParserManager
from vllm.sampling_params import RequestOutputKind, SamplingParams
from vllm.sequence import (ExecuteModelRequest, ParallelSampleSequenceGroup,
Sequence, SequenceGroup, SequenceGroupBase,
@@ -372,6 +373,14 @@ class LLMEngine:
"vllm.llm_engine",
self.observability_config.otlp_traces_endpoint)
# Initialize reasoning parser if reasoning backend is set.
if self.decoding_config.reasoning_backend and \
self.tokenizer:
reasoner_class = ReasoningParserManager.get_reasoning_parser(
self.decoding_config.reasoning_backend)
self.reasoner: ReasoningParser = reasoner_class(
self.tokenizer.get_lora_tokenizer())
# Create sequence output processor, e.g. for beam search or
# speculative decoding.
self.output_processor = (
@@ -381,8 +390,12 @@ class LLMEngine:
self.scheduler,
self.seq_counter,
get_tokenizer_for_seq,
stop_checker=StopChecker(self.scheduler_config.max_model_len,
get_tokenizer_for_seq),
stop_checker=StopChecker(
self.scheduler_config.max_model_len,
get_tokenizer_for_seq,
self.reasoner if self.decoding_config.reasoning_backend
and self.tokenizer else None,
),
))
self.seq_id_to_seq_group: Dict[str, SequenceGroupBase] = {}
+13 -2
View File
@@ -4,6 +4,7 @@
from typing import Callable, List, Optional, Tuple
from vllm.lora.request import LoRARequest
from vllm.reasoning import ReasoningParser
from vllm.sampling_params import SamplingParams
from vllm.sequence import Sequence, SequenceStatus
from vllm.transformers_utils.tokenizer import AnyTokenizer
@@ -16,11 +17,16 @@ class StopChecker:
emitted, or if we have exceeded the max model len.
"""
def __init__(self, max_model_len: int,
get_tokenizer_for_seq: Callable[[Sequence], AnyTokenizer]):
def __init__(
self,
max_model_len: int,
get_tokenizer_for_seq: Callable[[Sequence], AnyTokenizer],
reasoner: Optional[ReasoningParser] = None,
):
# Do not use it directly, but use `self._get_max_model_len`.
self._max_model_len = max_model_len
self.get_tokenizer_for_seq = get_tokenizer_for_seq
self.reasoner = reasoner
def _get_max_model_len(self, lora_req: Optional[LoRARequest]):
if lora_req and lora_req.long_lora_max_len:
@@ -57,6 +63,11 @@ class StopChecker:
seq.status = SequenceStatus.FINISHED_STOPPED
return
# Skip stop string/token checks if in reasoning content generation
if self.reasoner is not None and \
not self.reasoner.is_reasoning_end(seq.get_token_ids()):
return
# Check if a stop token was encountered.
# This assumes a single token produced per step.
last_token_id = seq.get_last_token_id()