update RegexGuide to conform with outlines-core

Author: lapp0

Diff for: benchmarks/bench_json_schema.py

@@ -77,4 +77,4 @@ def time_json_schema_to_regex(self, schema_name):
     @cache_disabled()
     def time_json_schema_to_fsm(self, schema_name):
         regex = build_regex_from_schema(self.schema)
-        RegexGuide(regex, self.tokenizer)
+        RegexGuide.from_regex(regex, self.tokenizer)

Diff for: benchmarks/bench_regex_guide.py

@@ -25,7 +25,7 @@ def setup(self, pattern_name):
 
     @cache_disabled()
     def time_regex_to_guide(self, pattern_name):
-        RegexGuide(self.pattern, self.tokenizer)
+        RegexGuide.from_regex(self.pattern, self.tokenizer)
 
 
 class MemoryRegexGuideBenchmark:
@@ -37,4 +37,4 @@ def setup(self, pattern_name):
 
     @cache_disabled()
     def peakmem_regex_to_guide(self, pattern_name):
-        RegexGuide(self.pattern, self.tokenizer)
+        RegexGuide.from_regex(self.pattern, self.tokenizer)

Diff for: outlines/fsm/guide.py

@@ -74,8 +74,8 @@ def copy(self):
 
 
 @cache()
-def create_states_mapping(regex_string, tokenizer):
-    return uncached_create_states_mapping(regex_string, tokenizer)
+def cached_create_states_mapping(regex_string, tokenizer, *args, **kwargs):
+    return uncached_create_states_mapping(regex_string, tokenizer, *args, **kwargs)
 
 
 class RegexGuide(CoreRegexGuide):
@@ -84,15 +84,20 @@ class RegexGuide(CoreRegexGuide):
     CoreRegexGuide with outlines cache
     """
 
-    def __init__(self, regex_string: str, tokenizer: "Tokenizer"):
-        (
-            self.states_to_token_maps,
-            self.empty_token_ids,
-            fsm_finals,
-        ) = create_states_mapping(regex_string, tokenizer)
-        self.eos_token_id = tokenizer.eos_token_id
-        self.final_states = fsm_finals | {-1}
-        self._cache_state_to_token_tensor()
+    @classmethod
+    def from_regex(
+        cls,
+        regex_string: str,
+        tokenizer,
+        _create_states_mapping=cached_create_states_mapping,
+        **kwargs,
+    ):
+        return super().from_regex(
+            regex_string,
+            tokenizer,
+            _create_states_mapping=_create_states_mapping,
+            **kwargs,
+        )
 
 
 CFGState = collections.namedtuple("CFGState", ["parser_state", "prev_token"])

Diff for: outlines/processors/structured.py

@@ -149,7 +149,7 @@ def __init__(self, regex_string: str, tokenizer: "Tokenizer"):
         tokenizer
             An Outlines tokenizer
         """
-        guide = RegexGuide(regex_string, tokenizer)
+        guide = RegexGuide.from_regex(regex_string, tokenizer)
         super().__init__(tokenizer=tokenizer, guide=guide)
 
 

Diff for: tests/fsm/test_guide.py

@@ -43,7 +43,7 @@ def convert_token_to_string(self, token):
     regex_str = "[1-9]"
 
     with pytest.raises(ValueError, match="The vocabulary"):
-        RegexGuide(regex_str, MockTokenizer())
+        RegexGuide.from_regex(regex_str, MockTokenizer())
 
 
 def test_regex():
@@ -57,7 +57,7 @@ def convert_token_to_string(self, token):
 
     regex_str = "[1-9]"
     tokenizer = MockTokenizer()
-    fsm = RegexGuide(regex_str, tokenizer)
+    fsm = RegexGuide.from_regex(regex_str, tokenizer)
 
     assert fsm.states_to_token_maps == {0: {1: 1}}
 
@@ -98,7 +98,7 @@ def convert_token_to_string(self, token):
 
     regex_str = "[😁-😎]"
     tokenizer = MockTokenizer()
-    fsm = RegexGuide(regex_str, tokenizer)
+    fsm = RegexGuide.from_regex(regex_str, tokenizer)
 
     assert fsm.states_to_token_maps == {
         0: {5: 1, 4: 2},
@@ -145,7 +145,7 @@ def convert_token_to_string(self, token):
 
     regex_str = " [😁-😎]"
     tokenizer = MockTokenizer()
-    fsm = RegexGuide(regex_str, tokenizer)
+    fsm = RegexGuide.from_regex(regex_str, tokenizer)
 
     assert fsm.states_to_token_maps == {
         0: {5: 1, 10: 2},
@@ -180,7 +180,7 @@ def convert_token_to_string(self, token):
 
     regex_str = r"`\n(\.\n)?`\n"
     tokenizer = MockTokenizer()
-    fsm = RegexGuide(regex_str, tokenizer)
+    fsm = RegexGuide.from_regex(regex_str, tokenizer)
 
     state = fsm.get_next_state(state=4, token_id=103)
     assert state == 5