fast regex decode

Auto-detect constant str path in regex FSM, then extend instead.

python/sglang/srt/constrained/fsm_cache.py

@@ -1,6 +1,8 @@
 from sglang.srt.constrained.fsm import RegexFSM
 from sglang.srt.constrained.tokenizer import TransformerTokenizer
 
+_enable_memory_cache = True
+
 
 class FSMCache:
     def __init__(self, tokenizer_path, tokenizer_args_dict):
@@ -10,8 +12,10 @@ class FSMCache:
         )
 
     def init_fsm(self, regex):
-        if regex not in self.cache:
-            fsm = RegexFSM(regex, self.outlines_tokenizer)
-            self.cache[regex] = fsm
+        if _enable_memory_cache:
+            if regex not in self.cache:
+                fsm = RegexFSM(regex, self.outlines_tokenizer)
+                self.cache[regex] = fsm
+            return self.cache[regex]
 
-        return self.cache[regex]
+        return RegexFSM(regex, self.outlines_tokenizer)