[Bugfix] use diskcache in outlines _get_guide vllm-project#5436 (vllm…

…-project#6203)

vllm/model_executor/guided_decoding/outlines_logits_processors.py

@@ -21,6 +21,7 @@
 from typing import Callable, DefaultDict, Dict, List, Union
 
 import torch
+from outlines.caching import cache
 from outlines.fsm.guide import CFGGuide, Generate, Guide, RegexGuide, Write
 from outlines.fsm.json_schema import build_regex_from_schema
 from pydantic import BaseModel
@@ -67,7 +68,7 @@ def __call__(self, input_ids: List[int],
 class RegexLogitsProcessor(BaseLogitsProcessor):
 
     @classmethod
-    @lru_cache(maxsize=32)
+    @cache()
     def _get_guide(cls, regex_string: str,
                    tokenizer: PreTrainedTokenizerBase) -> Guide:
         tokenizer = _adapt_tokenizer(tokenizer)
@@ -126,7 +127,7 @@ def __init__(self, schema: Union[str, Dict, BaseModel],
 class CFGLogitsProcessor(BaseLogitsProcessor):
 
     @classmethod
-    @lru_cache(maxsize=32)
+    @cache()
     def _get_guide(cls, cfg: str, tokenizer: PreTrainedTokenizerBase) -> Guide:
         tokenizer = _adapt_tokenizer(tokenizer)
         return CFGGuide(cfg, tokenizer)