Add example of integration with vLLM

examples/vllm_integration.py

@@ -0,0 +1,70 @@
+import math
+from typing import List
+
+import torch
+import vllm
+from pydantic import BaseModel
+
+from outlines.fsm.fsm import RegexFSM
+from outlines.fsm.json_schema import build_regex_from_object
+
+
+class User(BaseModel):
+    id: int
+    name: str
+
+
+class JSONLogitsProcessor:
+    def __init__(self, pydantic_model, llm):
+        schema = pydantic_model.schema_json()
+        regex_str = build_regex_from_object(schema)
+
+        # Adapt tokenizer to Outlines's interface
+        # vocabulary and special tokens can probably be modified on Outlines' side
+        # `convert_token_to_string` might however be necessary
+        tokenizer = llm.get_tokenizer()
+        tokenizer.vocabulary = tokenizer.get_vocab()
+        tokenizer.special_tokens = set(tokenizer.all_special_tokens)
+
+        def convert_token_to_string(token: str) -> str:
+            from transformers.file_utils import SPIECE_UNDERLINE
+
+            string = tokenizer.convert_tokens_to_string([token])
+
+            # A hack to handle missing spaces to HF's Llama tokenizers
+            if token.startswith(SPIECE_UNDERLINE) or token == "<0x20>":
+                return " " + string
+
+            return string
+
+        tokenizer.convert_token_to_string = convert_token_to_string
+
+        fsm = RegexFSM(regex_str, tokenizer)
+        self.fsm = fsm
+        self.fsm_state = 0
+
+    def __call__(self, input_ids: List[int], scores: torch.Tensor) -> torch.Tensor:
+        if len(input_ids) != 0:
+            last_token = input_ids[-1]
+            self.fsm_state = self.fsm.next_state(self.fsm_state, last_token)
+
+        allowed_tokens = self.fsm.allowed_token_ids(self.fsm_state)
+
+        mask = torch.full((scores.shape[-1],), -math.inf, device=scores.device)
+        mask[allowed_tokens] = 0
+        biased_scores = scores + mask
+
+        return biased_scores
+
+
+model_name = "gpt2"
+llm = vllm.LLM(model=model_name)
+logits_processor = JSONLogitsProcessor(User, llm)
+result = llm.generate(
+    "A prompt",
+    sampling_params=vllm.SamplingParams(
+        max_tokens=100, logits_processors=[logits_processor]
+    ),
+)
+
+print(result)