minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[perf] fix kimi tokenizer to improve ttft (#25265)

Browse Source
This commit is contained in:
Qiaolin Yu
2026-05-14 19:11:25 -07:00
committed by GitHub
parent dca9ba6321
commit 7af4320d67
1 changed files with 10 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/managers/tokenizer_manager.py
+10 -3
View File
@@ -689,9 +689,16 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
) )
else: else:
logger.debug(f"Using regular tokenizer for {len(tokenizer_input)} inputs") logger.debug(f"Using regular tokenizer for {len(tokenizer_input)} inputs")
encoded = self.tokenizer(tokenizer_input, **tokenizer_kwargs)
input_ids = encoded["input_ids"] if not is_cross_encoder and (not getattr(self.tokenizer, "is_fast", False)):
token_type_ids = encoded.get("token_type_ids") if is_cross_encoder else None input_ids = [self.tokenizer.encode(t) for t in tokenizer_input]
token_type_ids = None
else:
encoded = self.tokenizer(tokenizer_input, **tokenizer_kwargs)
input_ids = encoded["input_ids"]
token_type_ids = (
encoded.get("token_type_ids") if is_cross_encoder else None
)
# Step 4: Extract results based on input format # Step 4: Extract results based on input format
return self._extract_tokenizer_results( return self._extract_tokenizer_results(