TinyLlama-v1.1-GAD-Demo / transformers_gad /token_grammar_recognizer.py
kjcjohnson's picture
Add GAD libraries
901bbd9
import copy
import logging
from abc import ABC
from functools import lru_cache
from typing import List
import torch
from transformers_gad.recognizer import StringRecognizer, AcceptState
from transformers_gad.parser import parse_ebnf
from transformers_gad.trie import ByteTrie
from transformers_gad.utf8_utils import PartialUTF8
from .vocab_struct import LEAF, TokenTrie
from transformers_gad.mapping import get_mapping
logger = logging.getLogger(__name__)
class AbsTokenRecognizer(ABC):
def __init__(self, grammar_str, tokenizer, start_rule_name="root", unicode=False):
parsed_grammar = parse_ebnf(grammar_str)
grammar_encoding = parsed_grammar.grammar_encoding
self.start_rule_id = parsed_grammar.symbol_table.get(start_rule_name)
self.byte_encoding = unicode
if unicode and not tokenizer.__class__.__name__.lower().startswith(
"gpt2"
): # gpt2tokenizer or gpt2tokenizerfast
raise ValueError(
"Constrained decoding with unicode is only supported for GPT2 model. Support for other models is coming soon."
"Or you can use the constraints with only ascii characters."
)
self.eos_token_id = tokenizer.eos_token_id
self.token_trie = TokenTrie(tokenizer)
self.tokenizer = tokenizer
self.string_recognizer = StringRecognizer(grammar_encoding, self.start_rule_id)
self.unicode_trie = ByteTrie.from_tokenizer(tokenizer, unicode=unicode)
self.mapping = get_mapping(tokenizer, unicode=unicode)
assert len(self.mapping) == len(
self.token_trie
), f"{len(self.mapping)}, {len(self.token_trie)}"
def _consume_token_id(
self, token_id: int, accept_state: AcceptState
) -> AcceptState:
if self.string_recognizer._must_stop(accept_state.stacks):
if token_id == self.eos_token_id:
return self.string_recognizer.get_termination_accept_state()
else:
raise ValueError(
f"All stacks are empty, so the only token accepted is EOS({self.eos_token_id}), but got {token_id}"
)
if token_id == self.eos_token_id:
if self.string_recognizer._can_stop(accept_state.stacks):
# if at least one of the stack is empty, we can stop
# we clear all the stacks, meaning that we don't accept any token after EOS
return self.string_recognizer.get_termination_accept_state()
else:
raise ValueError(
f"At least one of the stack should be empty when EOS is reached. However, "
f"the stacks are {accept_state.stacks}"
)
bytes_or_codepoints = self.mapping.map(token_id)
accept_state = self.string_recognizer._consume_bytes(
bytes_or_codepoints, accept_state
)
return accept_state
def probe_token_id(self, token_id: int, accept_state: AcceptState) -> bool:
stacks = accept_state.stacks
if self.string_recognizer._must_stop(stacks):
if token_id == self.eos_token_id:
return True
else:
return False
if token_id == self.eos_token_id:
if self.string_recognizer._can_stop(stacks):
# if at least one of the stack is empty, we can stop
# we clear all the stacks, meaning that we don't accept any token after EOS
return True
else:
return False
# for code_point in self.mapping.map(token_id):
# stacks = self.grammar._consume_char_code_point(code_point, stacks)
bytes_or_codepoints = self.mapping.map(token_id, verbose=False)
new_acc_state = self.string_recognizer._consume_bytes(
bytes_or_codepoints, accept_state, verbose=False
)
return len(new_acc_state.stacks) > 0
def advance_token_ids(self, *args, **kwargs):
"""Process a list of tokens according to the grammar rules."""
raise NotImplementedError
def batch_filter_vocab(self, batch_accept_states, device) -> torch.Tensor:
batch_acceptance = []
for accept_state in batch_accept_states:
batch_acceptance.append(self.filter_vocab(accept_state, device))
return torch.stack(batch_acceptance)
def filter_vocab(self, accept_state, device) -> torch.Tensor:
if not accept_state.stacks: # Check if stacks is empty
# Handle the empty case: for example, return a tensor of False
# The size of the tensor should match the size of your vocabulary
vocab_size = len(self.mapping)
logger.debug(f"Empty stack, sum of acceptance: {0}")
# size of the vocab
accepts = [False] * vocab_size
accepts[self.eos_token_id] = True
return torch.tensor(accepts, dtype=torch.bool, device=device)
acceptance = self.get_token_acceptance(accept_state, device)
return acceptance
def get_token_acceptance(self, accept_state, device) -> torch.Tensor:
acceptance_matrix = torch.cat(
[
self.get_token_acceptance_array_for_stack(
tuple(stack), accept_state.partial_utf8, device
)
for stack in accept_state.stacks
]
)
# Merge stacks: any True => True
acceptance = acceptance_matrix.reshape(len(accept_state.stacks), -1).any(dim=0)
return acceptance
@lru_cache(maxsize=32768)
def get_token_acceptance_array_for_stack(self, stack, partial_utf8, device):
# stack = list(stack) # needs to come in as a tuple for lru_cache
assert isinstance(stack, tuple)
stack = list(stack)
if self.byte_encoding:
accept_f = lambda x: self.string_recognizer._probe_bytes(
x, [stack], partial_utf8=partial_utf8
)
token_acceptance = self.unicode_trie.get_token_acceptance(
accept=accept_f, accept_eos=False, eos_token_id=self.eos_token_id
)
else:
accepts = [False] * len(self.mapping)
token_acceptance = check_token_acceptance_in_trie(
self.token_trie.trie,
[stack],
self.string_recognizer,
self.eos_token_id,
accepts,
)
x = torch.tensor(token_acceptance, dtype=torch.bool, device=device)
x_eos = self.validate_and_set_eos_acceptance(x)
return x_eos
def validate_and_set_eos_acceptance(self, acceptance: torch.Tensor) -> torch.Tensor:
if torch.any(acceptance) == 0:
acceptance[self.eos_token_id] = True
else:
if acceptance[self.eos_token_id]:
raise ValueError()
acceptance[self.eos_token_id] = False
return acceptance
class IncrementalTokenRecognizer(AbsTokenRecognizer):
def __init__(self, grammar_str, start_rule_name, tokenizer, unicode=False):
super().__init__(grammar_str, tokenizer, start_rule_name, unicode)
self.last_size = None
self.is_incremental = True
# if self.last_size is not set (which would be the case when processing the first token).
# In this case, do nothing.
def advance_token_ids(self, input_ids, batch_accept_states, parse_start_index=None):
if self.last_size is None:
prefix_to_parse = [
single_input_ids[parse_start_index:]
if parse_start_index is not None
else []
for single_input_ids in input_ids
]
# self.grammar_acceptor.accept_token_ids(prefix_to_parse, self.stacks)
batch_accept_states = [
self._consume_token_ids(prefix, accept_state)
for prefix, accept_state in zip(prefix_to_parse, batch_accept_states)
]
# if the length of the current input IDs (input_ids[0]) is exactly one more than self.last_size.
# This is expected in a scenario where inputs are processed incrementally, one token at a time.
elif len(input_ids[0]) == self.last_size + 1:
batch_accept_states = [
self._consume_token_id(single_input_ids[-1], accept_state)
for single_input_ids, accept_state in zip(
input_ids, batch_accept_states
)
]
# ensure that the input size is consistent with the expected incremental processing
# (i.e., one token at a time).
else:
# here we check if the input_ids are one token longer than the last time we processed
# but we don't check if input_ids are actually valid.
# Imagine a scenario where we generate 10 tokens, then we replace the 10 generated tokens with 10 new tokens.
# In this case, the input_ids will be consistent with the last_size, but the input_ids are not valid.
# However, should we really check if the input_ids are valid here?
# If we do, then we need to reparse the whole input_ids at each call, which is not efficient.
# Maybe we should just trust the user to provide valid input_ids?
# The conclusion is that, we assume the input_ids are valid, and our generation will be correct.
# If the input_ids are not valid, then the generation result will be wrong and we don't take responsibility for that.
raise RuntimeError(
"Input ID's length is inconsistent with the current state of "
"the GrammarConstrainedLogitsProcessor. If you want to process "
"another input sequence, please instantiate a new "
"GrammarConstrainedLogitsProcessor "
"or call reset_parser method of GrammarAlignedOracleLogitsProcessor"
)
self.last_size = len(input_ids[0])
return batch_accept_states
def _consume_token_ids(
self, token_ids: List[int], accept_state: AcceptState = None, as_string=True
):
if accept_state is None:
accept_state = self.string_recognizer.get_initial_accept_state()
if as_string:
string = self.tokenizer.decode(token_ids)
accept_state = self.string_recognizer._consume_string(string, accept_state)
else:
for i, token_id in enumerate(token_ids):
accept_state = self._consume_token_id(token_id, accept_state)
if len(accept_state.stacks) > 0:
cur_token_ids = token_ids[: i + 1]
logging.debug(f"{cur_token_ids} is accepted")
decoded_string = self.tokenizer.decode(cur_token_ids)
logging.debug(f"The decoded string is {decoded_string}")
return accept_state
def reset(self):
self.last_size = None
def check_token_acceptance_in_trie(trie, stacks, grammar, eos_token_id, accepts):
for byte, next_trie in trie.items():
if byte == LEAF:
token_id = next_trie
if token_id != eos_token_id:
# if the stacks is not empty, it means we can still continue to parse
# so we should accept the token
accepts[token_id] = bool(stacks)
continue
new_stacks = []
for stk in stacks:
if not stk:
continue
next_element_offset = stk[-1]
num_chars = grammar.grammar_encoding[next_element_offset]
if not grammar.char_acceptance_at_element(next_element_offset).get(
byte, False
):
# if the current byte is not accepted by the current rule, we need to try next rule
continue
next_element_offset += num_chars + 1
new_stack = stk[:-1]
if grammar.grammar_encoding[next_element_offset]:
new_stack.append(next_element_offset)
new_stacks.extend(grammar.advance_stack(tuple(new_stack)))
if new_stacks:
check_token_acceptance_in_trie(
next_trie, new_stacks, grammar, eos_token_id, accepts
)
return accepts
if __name__ == "__main__":
from transformers import AutoTokenizer
with open("examples/grammars/japanese.ebnf", "r") as file:
input_text = file.read()
parsed_grammar = parse_ebnf(input_text)
parsed_grammar.print()
tokenizer = AutoTokenizer.from_pretrained("gpt2")
tokenRecognizer = IncrementalTokenRecognizer(
grammar_str=input_text, start_rule_name="root", tokenizer=tokenizer
)
japanese = "γƒˆγƒͺγƒΌγƒ " # "こんにけは"
token_ids = tokenizer.encode(japanese)
# 13298, 12675, 12045, 254
stacks = tokenRecognizer._consume_token_ids(
token_ids, tokenRecognizer.string_recognizer.stacks, as_string=False
)
if stacks:
print("The Japanese input is accepted")
else:
print("The Japanese input is not accepted")
korean = "μ•ˆλ…•ν•˜μ„Έμš”"
token_ids = tokenizer.encode(korean)
try:
stacks = tokenRecognizer._consume_token_ids(
token_ids, tokenRecognizer.string_recognizer.stacks, as_string=False
)
if stacks:
print("The Korean input is accepted")
else:
print("The Korean input is not accepted")
except ValueError as e:
print("The Korean input is not accepted")