vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Fix when params.n_vocab < tokenizer vocab size

Browse source
This commit is contained in:
wonjun Jang 2023-11-18 12:55:14 +09:00 • committed by GitHub
parent cc1f3fcfad
commit 026eb7cd01
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
1 changed files with 19 additions and 12 deletions
Download patch file Download diff file Expand all files Collapse all files

31
convert.py
View file

@ -298,7 +298,7 @@ class Params:
class VocabLoader: class VocabLoader:
def __init__(self, fname_tokenizer: Path) -> None: def __init__(self, params: Params, fname_tokenizer: Path) -> None:
try: try:
from transformers import AutoTokenizer from transformers import AutoTokenizer
except ImportError as e: except ImportError as e:
@ -309,10 +309,18 @@ class VocabLoader:
self.tokenizer = AutoTokenizer.from_pretrained(str(fname_tokenizer)) self.tokenizer = AutoTokenizer.from_pretrained(str(fname_tokenizer))
vocab_set = {encoded_tok for encoded_tok, id in self.tokenizer.vocab.items()} vocab_set = {encoded_tok for encoded_tok, id in self.tokenizer.vocab.items()}
self.added_tokens_list = [tok for tok in self.tokenizer.get_added_vocab()] self.added_tokens_list = []
self.added_tokens_dict = dict(self.tokenizer.get_added_vocab()) self.added_tokens_dict = dict()
self.added_tokens_ids = set(self.tokenizer.get_added_vocab().values()) self.added_tokens_ids = set()
for tok, tokidx in self.tokenizer.get_added_vocab().items():
if tokidx >= params.n_vocab or toksize < self.tokenizer.vocab_size:
continue
self.added_tokens_list.append(tok)
self.added_tokens_dict[tok] = tokidx
self.added_tokens_ids.add(tokidx)
self.unk_token_id = self.tokenizer.unk_token_id self.unk_token_id = self.tokenizer.unk_token_id
self.specials = { self.specials = {
@ -321,8 +329,8 @@ class VocabLoader:
} }
print(self.specials) print(self.specials)
self.special_ids = set(self.tokenizer.all_special_ids) self.special_ids = set(self.tokenizer.all_special_ids)
self.vocab_size_base: int = len(vocab_set) self.vocab_size_base: int = self.tokenizer.vocab_size
self.vocab_size: int = len(vocab_set) self.vocab_size: int = self.vocab_size_base + len(self.added_tokens_list)
self.fname_tokenizer = fname_tokenizer self.fname_tokenizer = fname_tokenizer
vocab_file = "tokenizer.model" vocab_file = "tokenizer.model"
@ -374,7 +382,6 @@ class VocabLoader:
def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
for text in self.added_tokens_list: for text in self.added_tokens_list:
if text in self.specials: if text in self.specials:
toktype = self.get_token_type(self.specials[text]) toktype = self.get_token_type(self.specials[text])
@ -1095,14 +1102,14 @@ def vocab_check_and_append_path(path: Path, vocab_file: str) -> bool:
return path return path
def load_vocab(path: Path) -> Vocab: def load_vocab(params: Params, path: Path) -> Vocab:
# Be extra-friendly and accept either a file or a directory. Also, if it's # Be extra-friendly and accept either a file or a directory. Also, if it's
# a directory, it might be the model directory, and tokenizer.model might # a directory, it might be the model directory, and tokenizer.model might
# be in the parent of that. # be in the parent of that.
print(f"Loading vocab file '{path}'") print(f"Loading vocab file '{path}'")
return VocabLoader(path) return VocabLoader(params, path)
def default_outfile(model_paths: list[Path], file_type: GGMLFileType) -> Path: def default_outfile(model_paths: list[Path], file_type: GGMLFileType) -> Path:
@ -1183,7 +1190,7 @@ def main(args_in: list[str] | None = None) -> None:
if not args.outfile: if not args.outfile:
raise ValueError("need --outfile if using --vocab-only") raise ValueError("need --outfile if using --vocab-only")
# FIXME: Try to respect vocab_dir somehow? # FIXME: Try to respect vocab_dir somehow?
vocab = load_vocab(args.vocab_dir or args.model) vocab = load_vocab(params, args.vocab_dir or args.model)
special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent, special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent,
load_merges = True, load_merges = True,
n_vocab = vocab.vocab_size) n_vocab = vocab.vocab_size)
@ -1197,7 +1204,7 @@ def main(args_in: list[str] | None = None) -> None:
vocab = model_plus.vocab vocab = model_plus.vocab
else: else:
vocab_dir = args.vocab_dir if args.vocab_dir else model_plus.paths[0].parent vocab_dir = args.vocab_dir if args.vocab_dir else model_plus.paths[0].parent
vocab = load_vocab(vocab_dir) vocab = load_vocab(params, vocab_dir)
# FIXME: Try to respect vocab_dir somehow? # FIXME: Try to respect vocab_dir somehow?
print(f"Vocab info: {vocab}") print(f"Vocab info: {vocab}")