Implement token alignment for RegexFSM and StopAtFSM

Author: RobinPicard

outlines/fsm/fsm.py

@@ -1,7 +1,6 @@
 from copy import deepcopy
-from typing import TYPE_CHECKING, List, NewType, Protocol
+from typing import TYPE_CHECKING, Dict, List, NewType, Protocol, Tuple
 
-import cloudpickle
 import interegular
 from lark import Lark
 
@@ -15,6 +14,9 @@
 
 
 class FSM(Protocol):
+    def align_prompt_tokens(self, prompt: str) -> str:
+        ...
+
     def allowed_token_ids(self, state: FSMState) -> List[int]:
         ...
 
@@ -39,8 +41,23 @@ class StopAtTokenFSM(FSM):
 
     def __init__(self, tokenizer: "Tokenizer", stop_token_id: int):
         self.stop_token_id = stop_token_id
-        self.vocabulary = tokenizer.vocabulary.values()
-        self.final_states = {1}
+        self.tokenizer = tokenizer
+        self.vocabulary = tokenizer.vocabulary
+        self.final_states = {2}
+        self.valid_alignment_tokens: List[int] = []
+
+    def align_prompt_tokens(self, prompt: str) -> str:
+        """Remove the last token from the prompt and set the value of self.valid_alignment_tokens"""
+        token_ids, _ = self.tokenizer.encode(prompt)
+        last_token_id = int(token_ids[0][-1])
+        last_token_text = self.tokenizer.decode([last_token_id])[0]
+        # select the tokens that start with the text removed from the prompt
+        self.valid_alignment_tokens = [
+            token
+            for text, token in self.vocabulary.items()
+            if text.startswith(last_token_text)
+        ]
+        return prompt[: -len(last_token_text)]
 
     def allowed_token_ids(self, state: FSMState) -> List[int]:
         """Generate a list of allowed tokens for the next step.
@@ -59,7 +76,9 @@ def allowed_token_ids(self, state: FSMState) -> List[int]:
 
         """
         if state == 0:
-            return list(self.vocabulary)
+            return self.valid_alignment_tokens
+        elif state == 1:
+            return list(self.vocabulary.values())
         else:
             return [self.stop_token_id]
 
@@ -83,17 +102,17 @@ def next_state(self, state: FSMState, token_id: int) -> FSMState:
 
         """
         if token_id == self.stop_token_id:
-            return FSMState(1)
+            return FSMState(2)
 
-        return FSMState(0)
+        return FSMState(1)
 
     def is_final_state(self, state: FSMState) -> bool:
         """Determine whether the current state of the FSM is a final state."""
         return state in self.final_states
 
     def copy(self) -> "StopAtTokenFSM":
         """Create a copy of the FSM."""
-        return self
+        return deepcopy(self)
 
 
 class RegexFSM(FSM):
@@ -122,41 +141,61 @@ def __init__(self, regex_string: str, tokenizer: "Tokenizer"):
             -1
         }  # Include the EOS token in final states
         self.tokenizer = tokenizer
-        self.vocabulary = tokenizer.vocabulary.values()
+        self.vocabulary = tokenizer.vocabulary
         self.end_token_id = tokenizer.eos_token_id
 
     def align_prompt_tokens(self, prompt: str) -> str:
         """Remove the last token from the prompt and update the states_to_token_maps accordingly"""
         token_ids, _ = self.tokenizer.encode(prompt)
         last_token_id = int(token_ids[0][-1])
         last_token_text = self.tokenizer.decode([last_token_id])[0]
-        vocabulary = {
-            self.tokenizer.decode([token_id])[0]: token_id
-            for token_id in range(len(self.vocabulary))
-        }
-        starting_state_tokens = {
-            self.tokenizer.decode([token_id])[0]: self.states_to_token_maps[0][token_id]
-            for token_id in self.states_to_token_maps[0]
-        }
-        # select the tokens that start with the text removed from the prompt and whose text after the
-        # initial prompt corresponds to that of one of the allowed tokens of the starting state
-        possible_tokens = {
-            vocabulary[token_text]: starting_state_tokens[token_text[len(last_token_text):]]
-            for token_text in vocabulary
-            if (
-               token_text.startswith(last_token_text)
-               and starting_state_tokens.get(token_text[len(last_token_text):])
-            )
+        last_token_length = len(last_token_text)
+        # select the tokens that start with the text removed from the prompt
+        crossing_tokens = {
+            token: text
+            for text, token in self.vocabulary.items()
+            if text.startswith(last_token_text)
         }
+        # keep only the tokens whose text after the boundary matches the fsm
+        valid_tokens_states = self.find_valid_crossing_tokens(
+            crossing_tokens, last_token_length
+        )
         # update the states_to_token_maps in the following manner:
         # the value of the starting state is assigned to a new state, the starting state is now the
-        # possible_tokens found above + the last_token we removed (that leads to the new state)
-        additional_state_id = max(list(self.states_to_token_maps.keys()) + list(self.final_states)) + 1
+        # valid_tokens_states found above
+        additional_state_id = (
+            max(list(self.states_to_token_maps.keys()) + list(self.final_states)) + 1
+        )
         self.states_to_token_maps[additional_state_id] = self.states_to_token_maps[0]
-        self.states_to_token_maps[0] = {**possible_tokens, last_token_id: additional_state_id}
-
-        return prompt[:-len(last_token_text)]
-
+        self.states_to_token_maps[0] = {}
+        for token, state in valid_tokens_states:
+            if state == 0:
+                self.states_to_token_maps[0][token] = additional_state_id
+            else:
+                self.states_to_token_maps[0][token] = state
+        return prompt[: -len(last_token_text)]
+
+    def find_valid_crossing_tokens(
+        self, crossing_tokens: Dict[int, str], last_token_length: int
+    ) -> List[Tuple[int, int]]:
+        """For each crossing token, check that the characters after the boundary match the FSM
+        and find the state it would lead to. Return the valid tokens with the associated state
+        """
+        valid_tokens = []
+        for token, text in crossing_tokens.items():
+            is_valid = True
+            crossing_text = text[last_token_length:]
+            state = 0
+            for char in crossing_text:
+                char_token = self.vocabulary.get(char)
+                try:
+                    state = self.states_to_token_maps[state][char_token]  # type: ignore
+                except KeyError:
+                    is_valid = False
+                    break
+            if is_valid:
+                valid_tokens.append((token, state))
+        return valid_tokens
 
     def allowed_token_ids(self, state: FSMState) -> List[int]:
         """Generate a list of allowed tokens for the next step.
@@ -222,12 +261,7 @@ def is_final_state(self, state: FSMState) -> bool:
 
     def copy(self) -> "RegexFSM":
         """Create a copy of the FSM."""
-        # temporary solution to the problem of unpickleable dict_values
-        self.vocabulary = cloudpickle.dumps(self.vocabulary)
-        copy = deepcopy(self)
-        self.vocabulary = cloudpickle.loads(self.vocabulary)
-        copy.vocabulary = cloudpickle.loads(copy.vocabulary)
-        return copy
+        return deepcopy(self)
 
 
 class CFGFSM(FSM):
@@ -257,6 +291,10 @@ def __init__(self, cfg_string: str, tokenizer: "Tokenizer"):
         self.done = False
         self.regex_fsm: RegexFSM
 
+    def align_prompt_tokens(self, prompt: str) -> str:
+        """Not implemented for CFGFSM"""
+        return prompt
+
     def _set_next_regex_fsm(self) -> None:
         """Use the CFG incremental parser to set the next regex FSM.
 
@@ -278,7 +316,6 @@ def _set_next_regex_fsm(self) -> None:
             self.allow_eos = True
             options.add("")
             assert len(options) > 1
-
         regex_string = r"(" + r"|".join([r"(" + x + r")" for x in options]) + r")"
         self.regex_fsm = RegexFSM(regex_string, self.tokenizer)
         self.reset_state = True

outlines/generate/api.py

@@ -1,5 +1,6 @@
 import json as pyjson
 import warnings
+from copy import deepcopy
 from typing import Callable, Iterator, List, Optional, Tuple, Union
 
 import torch
@@ -77,12 +78,15 @@ def get_generated_token_ids(
         return token_ids
 
     def get_generated_sequences(
-        self, generated_token_ids: List[torch.Tensor], initial_prompts: List[str], prompts: List[str]
+        self,
+        generated_token_ids: List[torch.Tensor],
+        initial_prompts: List[str],
+        prompts: List[str],
     ) -> List[str]:
         """Give the text sequences generated based on the tokens generated and the initial prompts"""
         generated_tokens_text = self.tokenizer.decode(generated_token_ids)
         return [
-            generated_tokens_text[i][len(initial_prompts[i]) - len(prompts[i]):]
+            generated_tokens_text[i][len(initial_prompts[i]) - len(prompts[i]) :]
             for i in range(len(generated_tokens_text))
         ]
 
@@ -196,7 +200,7 @@ def __call__(
 
         if isinstance(prompts, str):
             prompts = [prompts]
-        initial_prompts = copy.deepcopy(prompts)
+        initial_prompts = deepcopy(prompts)
 
         if isinstance(stop_at, str):
             stop_at = [stop_at]
@@ -205,7 +209,9 @@ def __call__(
         max_tokens = max_tokens or self.max_tokens
         num_sequences = len(prompts)
         fsms = [self.fsm.copy() for _ in prompts]
-        prompts = [fsm.align_prompt_tokens(prompt) for fsm, prompt in zip(fsms, prompts)]
+        prompts = [
+            fsm.align_prompt_tokens(prompt) for fsm, prompt in zip(fsms, prompts)
+        ]
 
         if rng is None:
             rng = torch.Generator(device=self.device)
@@ -239,7 +245,9 @@ def __call__(
         generated_token_ids = self.get_generated_token_ids(
             init_state, initial_prompts, last_state
         )
-        generated = self.get_generated_sequences(generated_token_ids, initial_prompts, prompts)
+        generated = self.get_generated_sequences(
+            generated_token_ids, initial_prompts, prompts
+        )
         stripped = [
             self.strip_stop_sequences(sequence, stop_sequences)
             for sequence in generated