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 7, 16:30 UTC
[Bug] Fix missing <think> tag after tool call in MiniMax 2.1 (#35352)
Signed-off-by: 冬马 <chenxinke@cai-inc.com> Co-authored-by: 冬马 <chenxinke@cai-inc.com>
stingoChen committed
Feb 26, 2026 at 22:11 UTC
7fea7250a46c88c1ba9684d7774d2c4ac17c4b90
1 file changed
+6
-1
vllm/reasoning/minimax_m2_reasoning_parser.py
+6
-1
@@ -87,10 +87,15 @@ class MiniMaxM2AppendThinkReasoningParser(ReasoningParser):
87
def __init__(self, tokenizer: TokenizerLike, *args, **kwargs):
88
super().__init__(tokenizer, *args, **kwargs)
89
self.end_token_id = self.vocab.get("</think>")
90
+ self.start_token_id = self.vocab.get("<think>")
91
92
def is_reasoning_end(self, input_ids: Sequence[int]) -> bool:
93
end_token_id = self.end_token_id
93
- return any(input_id == end_token_id for input_id in reversed(input_ids))
94
+ start_token_id = self.start_token_id
95
+ for input_id in reversed(input_ids):
96
+ if input_id in (end_token_id, start_token_id):
97
+ return input_id == end_token_id
98
+ return False
99
100
def extract_content_ids(self, input_ids: list[int]) -> list[int]:
101
return input_ids