Use a persistent Tokenizer hash and add tokenizer to cache signature

outlines/fsm/guide.py

@@ -112,7 +112,9 @@ class RegexGuide(Guide):
 
     def __init__(self, regex_string: str, tokenizer):
         @cache()
-        def create_states_mapping(regex_string: str) -> Tuple[dict, set, set]:
+        def create_states_mapping(
+            regex_string: str, tokenizer: "Tokenizer"
+        ) -> Tuple[dict, set, set]:
             """Create the variables related to the mapping between states and tokens
             The parameters of the function are used for caching purpose
             """
@@ -142,7 +144,7 @@ def create_states_mapping(regex_string: str) -> Tuple[dict, set, set]:
             self.states_to_token_maps,
             self.empty_token_ids,
             fsm_finals,
-        ) = create_states_mapping(regex_string)
+        ) = create_states_mapping(regex_string, tokenizer)
         self.eos_token_id = tokenizer.eos_token_id
         self.final_states = fsm_finals | {-1}
 

outlines/models/transformers.py

@@ -1,5 +1,7 @@
 from typing import TYPE_CHECKING, List, Optional, Tuple, Union
 
+from datasets.fingerprint import Hasher
+
 from outlines.models.tokenizer import Tokenizer
 
 if TYPE_CHECKING:
@@ -109,9 +111,7 @@ def __eq__(self, other):
         return NotImplemented
 
     def __hash__(self):
-        from datasets.fingerprint import Hasher
-
-        return hash(Hasher.hash(self.tokenizer))
+        return int(Hasher.hash(self.tokenizer), 16)
 
 
 class Transformers:

pyproject.toml

@@ -35,6 +35,7 @@ dependencies = [
    "referencing",
    "jsonschema",
    "requests",
+   "datasets",
 ]
 dynamic = ["version"]
 
@@ -49,7 +50,6 @@ test = [
     "diff-cover",
     "accelerate",
     "beartype<0.16.0",
-    "datasets",
     "responses",
     "llama-cpp-python",
     "huggingface_hub",

tests/models/test_transformers.py

@@ -107,6 +107,15 @@ def test_tokenizer_eq_hash():
     tokenizer_hf = AutoTokenizer.from_pretrained("gpt2")
 
     tokenizer = TransformerTokenizer(tokenizer_hf)
-    tokenizer2 = TransformerTokenizer(tokenizer_hf)
-    assert tokenizer == tokenizer2
-    assert hash(tokenizer) == hash(tokenizer2)
+    tokenizer_2 = TransformerTokenizer(tokenizer_hf)
+
+    assert tokenizer == tokenizer_2
+    assert hash(tokenizer) == 1485020377002298312
+    assert hash(tokenizer) == hash(tokenizer_2)
+
+    tokenizer_hf_2 = AutoTokenizer.from_pretrained("gpt2")
+    tokenizer_hf_2.add_tokens(["test_token"])
+
+    tokenizer_3 = TransformerTokenizer(tokenizer_hf_2)
+    assert tokenizer != tokenizer_3
+    assert hash(tokenizer) != hash(tokenizer_3)