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

convert : fix tokenizer conversion

Browse source 
ref: https://github.com/ggerganov/llama.cpp/pull/6852
This commit is contained in:
Georgi Gerganov 2024-04-23 22:11:09 +03:00
parent 1732737232
commit 5dcccb3a7d
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
1 changed files with 15 additions and 16 deletions
Download patch file Download diff file Expand all files Collapse all files

31
convert-hf-to-gguf.py
View file

@ -338,7 +338,12 @@ class Model(ABC):
tokenizer = SentencePieceProcessor(str(tokenizer_path))
vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
scores: list[float] = [-10000.0] * vocab_size
toktypes: list[int] = [SentencePieceTokenTypes.UNKNOWN] * vocab_size
for token_id in range(tokenizer.vocab_size()):
piece = tokenizer.id_to_piece(token_id)
text = piece.encode("utf-8")
score = tokenizer.get_score(token_id)
@ -353,9 +358,9 @@ class Model(ABC):
elif tokenizer.is_byte(token_id):
toktype = SentencePieceTokenTypes.BYTE
tokens.append(text)
scores.append(score)
toktypes.append(toktype)
tokens[token_id] = text
scores[token_id] = score
toktypes[token_id] = toktype
added_tokens_file = self.dir_model / 'added_tokens.json'
if added_tokens_file.is_file():
@ -363,20 +368,14 @@ class Model(ABC):
added_tokens_json = json.load(f)
for key in added_tokens_json:
key = key.encode("utf-8")
if key not in tokens:
tokens.append(key)
scores.append(-1000.0)
toktypes.append(SentencePieceTokenTypes.USER_DEFINED)
token_id = added_tokens_json[key]
if (token_id >= vocab_size):
print(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}')
continue
# pad remaining tokens
for i in range(vocab_size - len(tokens)):
print(f"gguf: padding token {i}")
tokens.append(f"[PAD{i}]")
scores.append(-1000.0)
toktypes.append(SentencePieceTokenTypes.USER_DEFINED)
assert len(tokens) == vocab_size
tokens[token_id] = key.encode("utf-8")
scores[token_id] = -1000.0
toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
self.gguf_writer.add_tokenizer_model("llama")
self.gguf_writer.add_token_list(tokens)