EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

fix: kimi k2 xgrammar crash (#8367)

Browse Source 
Co-authored-by: cicirori <32845984+cicirori@users.noreply.github.com>
Co-authored-by: gongwei-130 <56567052+gongwei-130@users.noreply.github.com>
This commit is contained in:
Yineng Zhang
2025-07-25 15:44:01 -07:00
committed by GitHub
parent d8ee15643b
commit f8ca2368b2
2 changed files with 15 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
python/sglang/srt/managers/schedule_batch.py
View File

@@ -431,6 +431,7 @@ class Req:
bootstrap_port: Optional[int] = None,
bootstrap_room: Optional[int] = None,
data_parallel_rank: Optional[int] = None,
vocab_size: Optional[int] = None,
):
# Input and output info
self.rid = rid
@@ -480,6 +481,7 @@ class Req:
self.to_abort_message: str = None
self.stream = stream
self.eos_token_ids = eos_token_ids
self.vocab_size = vocab_size
# For incremental decoding
# ----- | --------- read_ids -------|
@@ -713,6 +715,14 @@ class Req:
self.finished_reason = FINISH_MATCHED_TOKEN(matched=last_token_id)
return
if last_token_id > self.vocab_size or last_token_id < 0:
if self.sampling_params.stop_token_ids:
self.output_ids[-1] = next(iter(self.sampling_params.stop_token_ids))
if self.eos_token_ids:
self.output_ids[-1] = next(iter(self.eos_token_ids))
self.finished_reason = FINISH_MATCHED_STR(matched="NaN happened")
return
# Check stop strings
if len(self.sampling_params.stop_strs) > 0:
tail_str = self.tokenizer.decode(

7
python/sglang/srt/managers/scheduler.py
View File

@@ -1129,6 +1129,7 @@ class Scheduler(
bootstrap_port=recv_req.bootstrap_port,
bootstrap_room=recv_req.bootstrap_room,
data_parallel_rank=recv_req.data_parallel_rank,
vocab_size=self.model_config.vocab_size,
)
req.tokenizer = self.tokenizer
@@ -1395,8 +1396,10 @@ class Scheduler(
logger.info(f)
if self.enable_metrics:
cache_hit_rate = adder.log_hit_tokens / (
adder.log_input_tokens + adder.log_hit_tokens
total_tokens = adder.log_input_tokens + adder.log_hit_tokens
cache_hit_rate = (
adder.log_hit_tokens / total_tokens if total_tokens > 0 else 0.0
)
self.stats.num_running_reqs = running_bs
self.stats.num_used_tokens = num_used