mmarch/text-generation-webui
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

Better HF grammar implementation (#4953)

Browse source
This commit is contained in:
oobabooga 2023-12-17 02:01:23 -03:00 • committed by GitHub
parent aa200f8723
commit 12690d3ffc
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
19 changed files with 830 additions and 116 deletions
Download patch file Download diff file Expand all files Collapse all files

687
modules/grammar/grammar_utils.py Normal file
View file

@ -0,0 +1,687 @@
'''
This file has been 100% copied from this PR to the Transformers library:
https://github.com/huggingface/transformers/pull/27557
Author: Saibo-creator
Author GitHub: https://github.com/Saibo-creator
All credits go to the author.
'''
import logging
import re
import time
from abc import ABC
from functools import lru_cache
from typing import Dict, List
import torch
from modules import shared
logger = logging.getLogger(__name__)
########################
# EBNF Grammar Parsing #
########################
END_OF_ALTERNATE_MARKER = 0
END_OF_RULE_MARKER = 0
TO_BE_FILLED_MARKER = 0
REF_RULE_MARKER = 1
LITERAL_MARKER = 2
class ParseState:
def __init__(self):
self.symbol_ids = {}
self.grammar_encoding = [] # old name: out_grammar
def get_symbol_id(state, src):
if src not in state.symbol_ids:
state.symbol_ids[src] = len(state.symbol_ids)
return state.symbol_ids[src]
def generate_symbol_id(state, base_name):
next_id = len(state.symbol_ids)
state.symbol_ids[base_name + "_" + str(next_id)] = next_id
return next_id
def is_word_char(c):
return c.isalnum() or c == "-" or c == "_"
def hex_to_int(c):
if c.isdigit():
return int(c)
elif "a" <= c.lower() <= "f":
return ord(c.lower()) - ord("a") + 10
return -1
def remove_leading_white_space(src, newline_ok):
"""
Skips over whitespace and comments in the input string.
This function processes the input string, skipping over any spaces, tabs,
and content following a '#' character, which denotes a comment. The parsing
of a comment continues until the end of the line (denoted by newline characters
'\r' or '\n'). If the 'newline_ok' parameter is set to False, the function
will stop processing and return the remaining string upon encountering a
newline character, otherwise it will skip over newline characters as well.
Parameters:
src (str): The input string to be processed.
newline_ok (bool): A flag indicating whether encountering a newline character
should stop the parsing (False) or if it should be skipped (True).
Returns:
str: The remaining portion of the input string after skipping whitespace and comments.
"""
pos = 0
while pos < len(src) and (src[pos].isspace() or src[pos] == "#"):
if src[pos] == "#":
while pos < len(src) and src[pos] not in ("\r", "\n"):
pos += 1
else:
if not newline_ok and src[pos] in ("\r", "\n"):
break
pos += 1
return src[pos:]
def parse_name(src):
pos = 0
while pos < len(src) and is_word_char(src[pos]):
pos += 1
if pos == 0:
raise RuntimeError("expecting name at " + src)
return src[:pos], src[pos:]
def parse_char(src):
"""
parse the leading char from the input string
:param src:
:return: char, remaining_src
"""
# if we have a backslash, it's maybe an escape
if src[0] == "\\":
esc = src[1]
if esc == "x":
first = hex_to_int(src[2])
if first > -1:
second = hex_to_int(src[3])
if second > -1:
return (first << 4) + second, src[4:]
raise RuntimeError("expecting \\xNN at " + src)
elif esc in ('"', "[", "]"):
return esc, src[2:]
elif esc == "r":
return "\r", src[2:]
elif esc == "n":
return "\n", src[2:]
elif esc == "t":
return "\t", src[2:]
raise RuntimeError("unknown escape at " + src)
elif src:
return src[0], src[1:]
raise RuntimeError("unexpected end of input")
def parse_sequence(state, src, rule_name, outbuf, is_nested):
out_start_pos = len(outbuf)
# sequence size, will be replaced at end when known
outbuf.append(TO_BE_FILLED_MARKER)
last_sym_start = len(outbuf)
remaining_src = src
while remaining_src:
if remaining_src[0] == '"': # literal string
remaining_src = remaining_src[1:]
last_sym_start = len(outbuf)
while remaining_src[0] != '"':
char, remaining_src = parse_char(remaining_src)
# each char of a literal is encoded as a "range" of char - char
outbuf.append(LITERAL_MARKER)
outbuf.append(ord(char))
outbuf.append(ord(char))
remaining_src = remove_leading_white_space(remaining_src[1:], is_nested)
elif remaining_src[0] == "[": # char range(s)
remaining_src = remaining_src[1:]
last_sym_start = len(outbuf)
# num chars in range - replaced at end of loop
outbuf.append(TO_BE_FILLED_MARKER)
while remaining_src[0] != "]":
char, remaining_src = parse_char(remaining_src)
outbuf.append(ord(char))
if remaining_src[0] == "-" and remaining_src[1] != "]":
endchar_pair, remaining_src = parse_char(remaining_src[1:])
outbuf.append(ord(endchar_pair))
else:
# chars that aren't part of a c1-c2 range are just doubled (i.e., c-c)
outbuf.append(ord(char))
# replace num chars with actual
outbuf[last_sym_start] = len(outbuf) - last_sym_start - 1
remaining_src = remove_leading_white_space(remaining_src[1:], is_nested)
elif is_word_char(remaining_src[0]): # rule reference
name, remaining_src = parse_name(remaining_src)
ref_rule_id = get_symbol_id(state, name)
remaining_src = remove_leading_white_space(remaining_src, is_nested)
last_sym_start = len(outbuf)
outbuf.append(REF_RULE_MARKER)
outbuf.append(ref_rule_id)
elif remaining_src[0] == "(": # grouping
# parse nested alternates into synthesized rule
remaining_src = remove_leading_white_space(remaining_src[1:], True)
sub_rule_id = generate_symbol_id(state, rule_name)
remaining_src = parse_alternates(state, remaining_src, rule_name, sub_rule_id, True)
last_sym_start = len(outbuf)
# output reference to synthesized rule
outbuf.append(REF_RULE_MARKER)
outbuf.append(sub_rule_id)
if remaining_src[0] != ")":
raise RuntimeError("expecting ')' at " + remaining_src)
remaining_src = remove_leading_white_space(remaining_src[1:], is_nested)
elif remaining_src[0] in ("*", "+", "?"): # repetition operator
if len(outbuf) - out_start_pos - 1 == 0:
raise RuntimeError("expecting preceeding item to */+/? at " + remaining_src)
out_grammar = state.grammar_encoding
# apply transformation to previous symbol (last_sym_start -
# end) according to rewrite rules:
# S* --> S' ::= S S' |
# S+ --> S' ::= S S' | S
# S? --> S' ::= S |
sub_rule_id = generate_symbol_id(state, rule_name)
out_grammar.append(sub_rule_id)
sub_rule_start = len(out_grammar)
# placeholder for size of 1st alternate
out_grammar.append(TO_BE_FILLED_MARKER)
# add preceding symbol to generated rule
out_grammar.extend(outbuf[last_sym_start:])
if remaining_src[0] in ("*", "+"):
# cause generated rule to recurse
out_grammar.append(REF_RULE_MARKER)
out_grammar.append(sub_rule_id)
# apply actual size
out_grammar[sub_rule_start] = len(out_grammar) - sub_rule_start
# mark end of 1st alternate
out_grammar.append(END_OF_ALTERNATE_MARKER)
sub_rule_start = len(out_grammar)
# placeholder for size of 2nd alternate
out_grammar.append(TO_BE_FILLED_MARKER)
if remaining_src[0] == "+":
# add preceding symbol as alternate only for '+'
out_grammar.extend(outbuf[last_sym_start:])
# apply actual size of 2nd alternate
out_grammar[sub_rule_start] = len(out_grammar) - sub_rule_start
# mark end of 2nd alternate, then end of rule
out_grammar.append(END_OF_ALTERNATE_MARKER)
out_grammar.append(END_OF_RULE_MARKER)
# in original rule, replace previous symbol with reference to generated rule
outbuf[last_sym_start:] = [1, sub_rule_id]
remaining_src = remove_leading_white_space(remaining_src[1:], is_nested)
else:
break
# apply actual size of this alternate sequence
outbuf[out_start_pos] = len(outbuf) - out_start_pos
# mark end of alternate
outbuf.append(END_OF_ALTERNATE_MARKER)
return remaining_src
def parse_alternates(state, src, rule_name, rule_id, is_nested):
outbuf = []
remaining_src = parse_sequence(state, src, rule_name, outbuf, is_nested)
while remaining_src and remaining_src[0] == "|":
remaining_src = remove_leading_white_space(remaining_src[1:], True)
remaining_src = parse_sequence(state, remaining_src, rule_name, outbuf, is_nested)
state.grammar_encoding.append(rule_id)
state.grammar_encoding.extend(outbuf)
state.grammar_encoding.append(0)
return remaining_src
def parse_rule(state, src):
name, remaining_src = parse_name(src)
remaining_src = remove_leading_white_space(remaining_src, False)
rule_id = get_symbol_id(state, name)
if remaining_src[:3] != "::=":
raise RuntimeError("expecting ::= at " + remaining_src)
remaining_src = remove_leading_white_space(remaining_src[3:], True)
remaining_src = parse_alternates(state, remaining_src, name, rule_id, False)
if remaining_src and remaining_src[0] == "\r":
remaining_src = remaining_src[2:] if remaining_src[1] == "\n" else remaining_src[1:]
elif remaining_src and remaining_src[0] == "\n":
remaining_src = remaining_src[1:]
elif remaining_src:
raise RuntimeError("expecting newline or end at " + remaining_src)
return remove_leading_white_space(remaining_src, True)
def parse_ebnf(src):
try:
state = ParseState()
grammar_repr = remove_leading_white_space(src, True)
last_grammar_repr = ""
while grammar_repr:
if last_grammar_repr:
last_parsed_rule_len = len(last_grammar_repr) - len(grammar_repr)
logger.debug(f"last_parsed_rule: {last_grammar_repr[:last_parsed_rule_len]}")
last_grammar_repr = grammar_repr
grammar_repr = parse_rule(state, grammar_repr)
state.grammar_encoding.append(0xFFFF)
return state
except RuntimeError as err:
logger.warning("error parsing grammar:", err)
return ParseState()
def print_rule(file, grammar_encoding, index, symbol_id_names):
rule_id = grammar_encoding[index]
print(f"<{index}>{symbol_id_names[rule_id]} ::=", end=" ", file=file)
pos = index + 1
while grammar_encoding[pos]:
if pos - 1 > index:
print("|", end=" ", file=file)
pos += 1 # sequence size, not needed here
while grammar_encoding[pos]:
if grammar_encoding[pos] == REF_RULE_MARKER:
ref_rule_id = grammar_encoding[pos + 1]
print(
f"<{pos}>{symbol_id_names[ref_rule_id]}",
end=" ",
file=file,
)
pos += 2
else:
print("<{}>[".format(pos), end="", file=file)
num_chars = grammar_encoding[pos]
pos += 1
for i in range(0, num_chars, 2):
print("{}-".format(chr(grammar_encoding[pos + i])), end="", file=file)
if i + 1 < num_chars:
print("{}".format(chr(grammar_encoding[pos + i + 1])), end="", file=file)
print("]", end=" ", file=file)
pos += num_chars
pos += 1
print(file=file)
return pos + 1
def print_grammar(file, state):
pos = 0
symbol_id_names = {v: k for k, v in state.symbol_ids.items()}
print("Grammar Rules:", file=file)
while state.grammar_encoding[pos] != 0xFFFF:
pos = print_rule(file, state.grammar_encoding, pos, symbol_id_names)
pos = 0
print("\nBinary representation:", file=file)
while state.grammar_encoding[pos] != 0xFFFF:
print(f"{state.grammar_encoding[pos]:04x}", end=" ", file=file)
pos += 1
print("ffff\n")
###################################
# EBNF Grammar Parsing ends here #
###################################
class GrammarConstraint(ABC):
def __init__(self, grammar_str, start_rule_name, tokenizer):
self.tt = 0
self.nt = 0
state = parse_ebnf(grammar_str)
grammar_encoding = state.grammar_encoding
self.start_rule_id = state.symbol_ids.get(start_rule_name)
self.eos_token_id = tokenizer.eos_token_id
self.token_trie = TokenTrie(tokenizer)
self.tokenizer = tokenizer
self.grammar_encoding = grammar_encoding
pos = 0
rules: Dict[int, int] = {}
while grammar_encoding[pos] != 0xFFFF:
rule_id = grammar_encoding[pos]
# Store the current position in the 'rules' list at the index corresponding to rule_id.
# This effectively maps each rule_id to its position in the grammar encoding.
rules[rule_id] = pos
pos += 1
# Continue to the next rule in the encoding.
# The loop advances by the size indicated at the current position (grammar_encoding[pos])
# plus one for the size field itself.
while grammar_encoding[pos]:
pos += 1 + grammar_encoding[pos]
# Now we're at the end of the rule,
# so advance to the next rule by skipping the 0, which means 'end of rule'.
pos += 1
self.start_rule_pos = rules[self.start_rule_id]
self.rules_pos_dict: Dict[int, int] = rules
def init_stacks(self):
# suppose the start rule position is 0, then grammar_encoding[0] = rule_id
# grammar_encoding[1] = rule_size
# grammar_encoding[2] = rule_type
# this is why we need to add 2 to the start rule position
stack = [self.start_rule_pos + 2]
# convert to tuple for caching(immutable)
return self.advance_stack(tuple(stack))
# For each stack, resolve rules to find the actual characters that are
# accepted by this stack (not the set of sub-rules).
# This is where the parsing happens.
# The parsing is a top-down, left-to-right, depth-first traversal of the
# grammar.
@lru_cache(maxsize=32768)
def advance_stack(self, stack):
stack = list(stack)
# If the stack is empty, we're done. Because no more tokens should be accepted.
if len(stack) == 0:
return [stack]
# Get the top of the stack.
pos = stack[-1]
# If the stack head is a terminal(literal), we can resolve it immediately.
# literal is marked with 2 in the grammar encoding.
if self.grammar_encoding[pos] > 1:
return [stack]
# The stack head is a nonterminal (a rule reference, 1 in the grammar encoding).
# Resolving this rule gives a set of one or more possible positions
# (e.g. two in `a ::= b | c`)
# We pop the current rule off the stack and, for each option, push:
# - the symbol following this symbol in the current rule; then
# - the first symbol of the resolved rule.
referenced_rule_id = self.grammar_encoding[pos + 1]
# subpos should points to the size of the subrule
subpos = self.rules_pos_dict[referenced_rule_id] + 1
stacks: List[List[int]] = []
# do depth-first search to find all possible rules and check the next terminal
# When this value is non-zero, it indicates that subpos is not yet at the end of the rule, so we can continue.
# here subpos is a pointer, and the value in the rule encoding can never be 0 except for the end of the rule.
while self.grammar_encoding[subpos]:
new_stack = stack[:-1]
if self.grammar_encoding[pos + 2]:
# check if there is a next symbol in the current rule, e.g. `a ::= b c | d`
# if yes, push the pos to rule_size to the stack
new_stack.append(pos + 2)
# if the type of the next symbol is not "empty", push the first symbol of the resolved rule to the stack
if self.grammar_encoding[subpos + 1]:
new_stack.append(subpos + 1)
stacks.extend(self.advance_stack(tuple(new_stack)))
# The increment subpos += self.grammar_encoding[subpos] + 1
# moves subpos forward in the grammar encoding array to the next alternative in the current rule.
subpos += self.grammar_encoding[subpos] + 1
return stacks
def accept_char(self, *args, **kwargs):
"""Process a byte according to the grammar rules."""
raise NotImplementedError
def accept_token_id(self, *args, **kwargs):
"""Process a token according to the grammar rules."""
raise NotImplementedError
def filter_vocab(self, *args, **kwargs):
raise NotImplementedError
class IncrementalGrammarConstraint(GrammarConstraint):
def __init__(self, grammar_str, start_rule_name, tokenizer):
super().__init__(grammar_str, start_rule_name, tokenizer)
def accept_char(self, byte, stacks):
new_stacks = []
for stack in stacks:
# stack is empty
if not stack:
continue
pos = stack[-1]
num_chars = self.grammar_encoding[pos]
# to make pos point to the size of the char range rule
pos += 1
found = False
for i in range(0, num_chars, 2):
if self.grammar_encoding[pos + i] <= byte and byte <= self.grammar_encoding[pos + i + 1]:
found = True
break
if not found:
continue
pos += num_chars
new_stack = stack[:-1]
if self.grammar_encoding[pos]:
new_stack.append(pos)
new_stacks.extend(self.advance_stack(tuple(new_stack)))
return new_stacks
def accept_string(self, string: str, stacks: List[List[int]]):
_bytes = bytes(string, "utf-8")
for byte in _bytes:
stacks = self.accept_char(byte, stacks)
return stacks
def accept_token_id(self, token_id: int, stacks: List[List[int]]):
if token_id == self.eos_token_id:
if stacks and all(len(stack) != 0 for stack in stacks):
raise Exception(
f"At least one of the stack should be empty when EOS is reached. However, "
f"the stacks are {stacks}"
)
return []
for byte in self.token_trie.id2str(token_id):
stacks = self.accept_char(byte, stacks)
# check updated stacks
# TODO, I commented this out because it will fail when the stack is empty
# empty stack means the end of the grammar
# assert stacks != []
return stacks
def accept_token_ids(self, token_ids: List[int], stacks: List[List[int]], as_string=True):
if as_string:
string = self.tokenizer.decode(token_ids)
stacks = self.accept_string(string, stacks)
else:
for token_id in token_ids:
stacks = self.accept_token_id(token_id, stacks)
return stacks
def batch_filter_vocab(self, batch_stacks, device):
batch_acceptance = []
for stacks in batch_stacks:
batch_acceptance.append(self.filter_vocab(stacks, device))
return torch.stack(batch_acceptance)
def filter_vocab(self, stacks, device):
if not 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.token_trie)
logger.debug(f"sum of acceptance: {0}")
return torch.zeros(vocab_size, dtype=torch.bool, device=device)
acceptance_matrix = torch.cat([self.token_acceptance_for_stack(tuple(stack), device) for stack in stacks])
# Merge stacks: any True => True
acceptance = acceptance_matrix.reshape(len(stacks), -1).any(dim=0)
logger.debug(f"sum of acceptance: {acceptance.sum()}")
return acceptance
# For each sub-rule in the grammar, cache whether each byte is accepted.
@lru_cache(maxsize=None)
def pos_char_acceptance(self, pos):
acceptance = [False] * 256
num_chars = self.grammar_encoding[pos]
pos += 1
for i in range(0, num_chars, 2):
start = self.grammar_encoding[pos + i]
end = self.grammar_encoding[pos + i + 1]
for j in range(start, end + 1):
acceptance[j] = True
return acceptance
# Probably this should be configurable. If the grammar has an exceedingly
# large number of states, the correct setting is a tradeoff between GPU
# RAM usage and recomputation time.
#
# The main variable that pushes usage up here is number of states in the
# grammar.
@lru_cache(maxsize=32768)
def token_acceptance_for_stack(self, stack, device):
st = time.time()
stack = list(stack) # needs to come in as a tuple for lru_cache
accepts = [False] * len(self.token_trie)
accepts[self.eos_token_id] = len(stack) == 0
if len(stack) == 0:
logger.debug("empty stack")
def traverse_trie(trie, stacks):
for byte, next_trie in trie.items():
if byte == LEAF:
token_id = next_trie
if token_id != self.eos_token_id:
accepts[token_id] = bool(stacks)
continue
new_stacks = []
for stk in stacks:
if not stk:
continue
pos = stk[-1]
num_chars = self.grammar_encoding[pos]
if not self.pos_char_acceptance(pos)[byte]:
continue
pos += num_chars + 1
new_stack = stk[:-1]
if self.grammar_encoding[pos]:
new_stack.append(pos)
new_stacks.extend(self.advance_stack(tuple(new_stack)))
if new_stacks:
traverse_trie(next_trie, new_stacks)
traverse_trie(self.token_trie.trie, [stack])
et = time.time() - st
x = torch.tensor(accepts, dtype=torch.bool, device=device)
self.tt += et
self.nt += 1
return x
class StaticGrammarConstraint(GrammarConstraint):
def __init__(self, grammar_str, start_rule_name, tokenizer):
super().__init__(grammar_str, start_rule_name, tokenizer)
def accept_char(self):
raise NotImplementedError
#################
# DATA STRUCTURES
#################
LEAF = -1
class TokenTrie:
def __init__(self, tokenizer):
self.eos_token_id = tokenizer.eos_token_id
self.tokens = []
self.trie = {}
self.load_tokens(tokenizer)
def id2str(self, token_id):
return self.tokens[token_id]
def __len__(self):
return len(self.tokens)
def load_tokens(self, tokenizer):
def replace_hex(match):
hex_value = match.group(1)
return chr(int(hex_value, 16))
if "gpt2" in tokenizer.__class__.__name__.lower():
special = tokenizer.additional_special_tokens_ids
# Here, the decoder does a string replace on a bunch of sequences
# like ' .' for '.'. This interferes with our assumptions, where a
# token should always have exactly one representation.
# Fortunately(?) text-generation-inference doesn't seem to run this
# cleanup, so we get extraneous spaces. So, in order to generate
# the right token set for TGI, we have to skip the space trimming.
# See:
# https://github.com/huggingface/transformers/blob/main/src/transformers/tokenization_utils_base.py#L3588-L3600
def fmt_token(id):
if id in special:
return None
return bytes(tokenizer.decode([id], clean_up_tokenization_spaces=False), "utf-8")
elif "llama" in tokenizer.__class__.__name__.lower():
def fmt_token(id):
token = tokenizer.convert_ids_to_tokens(id)
token = re.sub(r"<0x([0-9a-fA-F]{2})>", replace_hex, token)
token = token.replace("▁", " ")
return bytes(token, "utf-8")
else:
print("Warning: unrecognized tokenizer: using default token formatting")
def fmt_token(id):
token = tokenizer.convert_ids_to_tokens(id)
return bytes(token, "utf-8")
# note: vocab_size doesn't work here because there are also
# get_added_vocab() tokens
self.tokens = [fmt_token(i) for i in range(len(tokenizer.get_vocab()))]
for token_id, token_bytes in enumerate(self.tokens):
if token_bytes is not None:
self.insert_into_trie(self.trie, token_bytes, token_id)
def insert_into_trie(self, trie, token_bytes, token_id):
current = trie
for byte in token_bytes:
if byte not in current:
current[byte] = {}
current = current[byte]
current[LEAF] = token_id
@lru_cache(maxsize=5)
def initialize_grammar(grammar_string):
return IncrementalGrammarConstraint(grammar_string.strip(), start_rule_name="root", tokenizer=shared.tokenizer)