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

python/sglang/srt/managers/tokenizer_manager.py

@@ -689,9 +689,16 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
             )
         else:
             logger.debug(f"Using regular tokenizer for {len(tokenizer_input)} inputs")
-            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
+
+            if not is_cross_encoder and (not getattr(self.tokenizer, "is_fast", False)):
+                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
         return self._extract_tokenizer_results(