vllm 0.3.0 compatibility

outlines/serve/vllm.py

@@ -8,6 +8,8 @@
 from outlines.fsm.fsm import FSMState, RegexFSM
 from outlines.fsm.json_schema import build_regex_from_object
 
+from vllm.transformers_utils.tokenizer import TokenizerGroup
+
 
 def _patched_apply_logits_processors(
     logits,
@@ -38,14 +40,16 @@ def _patched_apply_logits_processors(
     return logits
 
 
-def adapt_tokenizer(tokenizer):
+def adapt_tokenizer(tokenizer_group):
     """Adapt vLLM's tokenizer to use to compile the FSM.
 
     The API of Outlines tokenizers is slightly different to that of
     `transformers`. In addition, we need to handle the missing spaces to
     Llama's tokenizer to be able to compile FSMs for this model.
 
     """
+    assert isinstance(tokenizer_group, TokenizerGroup)
+    tokenizer = tokenizer_group.tokenizer
     tokenizer.vocabulary = tokenizer.get_vocab()
     tokenizer.special_tokens = set(tokenizer.all_special_tokens)
 

pyproject.toml

@@ -38,7 +38,7 @@ dependencies = [
    "referencing",
    "jsonschema",
    "requests",
-   "transformers==4.36.2",  # TODO: unpin when HF fixes release (4.37.0 is broken)
+   "transformers",
 ]
 dynamic = ["version"]