[Frontend][Core] Update Outlines Integration from FSM to Guide

requirements-common.txt

@@ -12,6 +12,6 @@ pydantic >= 2.0  # Required for OpenAI server.
 prometheus_client >= 0.18.0
 tiktoken == 0.6.0  # Required for DBRX tokenizer
 lm-format-enforcer == 0.9.3
-outlines == 0.0.34 # Requires torch >= 2.1.0
+outlines >= 0.0.37
 typing_extensions
 filelock >= 3.10.4 # filelock starts to support `mode` argument from 3.10.4

tests/entrypoints/test_openai_server.py

@@ -82,7 +82,8 @@ def __init__(self, args):
         env = os.environ.copy()
         env["PYTHONUNBUFFERED"] = "1"
         self.proc = subprocess.Popen(
-            ["python3", "-m", "vllm.entrypoints.openai.api_server"] + args,
+            [sys.executable, "-m", "vllm.entrypoints.openai.api_server"] +
+            args,
             env=env,
             stdout=sys.stdout,
             stderr=sys.stderr,

vllm/model_executor/guided_decoding/outlines_logits_processors.py

@@ -21,17 +21,16 @@
 from typing import Callable, DefaultDict, Dict, List, Optional, Union
 
 import torch
-from outlines.fsm.fsm import CFGFSM, FSM, RegexFSM
+from outlines.fsm.guide import CFGGuide, Generate, Guide, RegexGuide, Write
 from outlines.fsm.json_schema import build_regex_from_schema
 from pydantic import BaseModel
 from transformers import PreTrainedTokenizerBase
 
 
 class BaseLogitsProcessor:
 
-    def __init__(self):
-        # Child class should use initialize in their init.
-        self.fsm: FSM
+    def __init__(self, guide: Guide):
+        self._guide: Guide = guide
 
     def init_state(self):
         """Initialize the FSM states."""
@@ -47,10 +46,17 @@ def __call__(self, input_ids: List[int],
         else:
             last_token = input_ids[-1]
             last_seq_id = hash(tuple(input_ids[:-1]))
-            self.fsm_state[seq_id] = self.fsm.next_state(
-                self.fsm_state[last_seq_id], last_token)
+            self.fsm_state[seq_id] = self._guide.get_next_state(
+                state=self.fsm_state[last_seq_id], token_id=last_token)
 
-        allowed_tokens = self.fsm.allowed_token_ids(self.fsm_state[seq_id])
+        instruction = self._guide.get_next_instruction(
+            state=self.fsm_state[seq_id])
+
+        if type(instruction) == Generate:
+            allowed_tokens = instruction.tokens
+        elif type(instruction) == Write:
+            # TODO: support fast forward tokens
+            allowed_tokens = [instruction.tokens[0]]
 
         mask = torch.full((scores.shape[-1], ),
                           -math.inf,
@@ -74,8 +80,7 @@ def __init__(self, regex_string: str, tokenizer: PreTrainedTokenizerBase):
 
         """
         tokenizer = _adapt_tokenizer(tokenizer)
-        fsm = RegexFSM(regex_string, tokenizer)
-        self.fsm = fsm
+        super().__init__(RegexGuide(regex_string, tokenizer))
 
 
 class JSONLogitsProcessor(RegexLogitsProcessor):
@@ -128,8 +133,7 @@ def __init__(self, cfg: str, tokenizer: PreTrainedTokenizerBase):
 
         """
         tokenizer = _adapt_tokenizer(tokenizer)
-        fsm = CFGFSM(cfg, tokenizer)
-        self.fsm = fsm
+        super().__init__(CFGGuide(cfg, tokenizer))
 
 
 @lru_cache