Fix vLLM integration (#711)

When integrating Outlines with vLLM I faced the following issues, which
are fixed in this PR:

1. When calling `vllm.LLM.generate` then within the internals of vLLM a
`copy.deepcopy` of the vLLM `SamplingParams` is made, which includes the
logits processor from Outlines (`RegexLogitsProcessor`, say). This
requires everything to be pickleable, and the
`RegexLogitsProcessor.fsm.vocabulary` is a `dict_values` object, which
doesn't satisfy that. The fix is easy: just convert it to a list. This
doesn't affect how this `vocabulary` variable is being used in the code.
2. The `RegexLogitsProcessor` takes an `llm` argument, which the
docstring states should be a `vllm.LLM` object, but then attempts to
extract the underlying tokenizer via `llm.tokenizer.tokenizer`. The
tokenizer of `vllm.LLM` currently lies in the
`vllm.LLM.llm_engine.tokenizer.tokenizer` attribute, but this is a big
mess and isn't backwards compatible with previous vLLM versions.
Instead, they have a convenience method, `vllm.LLM.get_tokenizer`, which
fetches the tokenizer. To remain backwards compatibility, in case people
have supplied `vllm.LLM.llm_engine` directly into
`RegexLogitsProcessor`, it falls back to a `tokenizer` or
`tokenizer.tokenizer` attribute.

I also updated the vLLM example script, as that was outdated as well
(used the previous `_patched_apply_logits_processors`).

Closes #704

.gitignore

@@ -5,3 +5,4 @@ docs/build
 .coverage
 .idea/
 *.gguf
+.venv

examples/vllm_integration.py

@@ -1,20 +1,16 @@
 import vllm
-import vllm.model_executor.layers.sampler as sampler
 from pydantic import BaseModel
 
-from outlines.serve.vllm import JSONLogitsProcessor, _patched_apply_logits_processors
-
-# Patch the _apply_logits_processors so it is compatible with `JSONLogitsProcessor`
-sampler._apply_logits_processors = _patched_apply_logits_processors
+from outlines.serve.vllm import JSONLogitsProcessor
 
 
 class User(BaseModel):
     id: int
     name: str
 
 
-llm = vllm.LLM(model="gpt2")
-logits_processor = JSONLogitsProcessor(User, llm)
+llm = vllm.LLM(model="openai-community/gpt2")
+logits_processor = JSONLogitsProcessor(schema=User, llm=llm)
 result = llm.generate(
     ["A prompt", "Another prompt"],
     sampling_params=vllm.SamplingParams(

outlines/fsm/fsm.py

@@ -121,7 +121,7 @@ def create_states_mapping(
         self.states_to_token_maps, self.empty_token_ids = create_states_mapping(
             regex_string, tuple(sorted(tokenizer.vocabulary.items()))
         )
-        self.vocabulary = tokenizer.vocabulary.values()
+        self.vocabulary = list(tokenizer.vocabulary.values())
         self.eos_token_id = tokenizer.eos_token_id
 
     def allowed_token_ids(self, state: FSMState) -> List[int]:
@@ -218,7 +218,7 @@ def create_states_mapping_from_interegular_fsm(
         ) = create_states_mapping_from_interegular_fsm(
             interegular_fsm, tuple(sorted(tokenizer.vocabulary.items()))
         )
-        from_interegular_instance.vocabulary = tokenizer.vocabulary.values()
+        from_interegular_instance.vocabulary = list(tokenizer.vocabulary.values())
         from_interegular_instance.eos_token_id = tokenizer.eos_token_id
         return from_interegular_instance
 

outlines/serve/vllm.py

@@ -47,7 +47,19 @@ def __init__(self, regex_string, llm):
             An instance of `vllm.LLM`
 
         """
-        tokenizer = self.adapt_tokenizer(llm.tokenizer.tokenizer)
+        if hasattr(llm, "get_tokenizer"):
+            tokenizer = llm.get_tokenizer()
+        elif hasattr(llm, "tokenizer"):
+            if hasattr(llm.tokenizer, "tokenizer"):
+                tokenizer = llm.tokenizer.tokenizer
+            else:
+                tokenizer = llm.tokenizer
+        else:
+            raise ValueError(
+                "The provided LLM instance in `RegexLogitsProcessor` neither has a "
+                "`tokenizer` attribute or a `get_tokenizer` method."
+            )
+        tokenizer = self.adapt_tokenizer(tokenizer=tokenizer)
 
         fsm = RegexFSM(regex_string, tokenizer)
         self.fsm = fsm