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

convert : fix set_vocab_sentencepiece

Browse source
This commit is contained in:
Georgi Gerganov 2024-04-24 10:19:38 +03:00
parent 3fec68be4e
commit f2588b0b70
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
1 changed files with 15 additions and 19 deletions
Download patch file Download diff file Expand all files Collapse all files

34
convert-hf-to-gguf.py
View file

@ -332,7 +332,12 @@ class Model(ABC):
tokenizer = SentencePieceProcessor(str(tokenizer_path)) tokenizer = SentencePieceProcessor(str(tokenizer_path))
vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size()) 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()): for token_id in range(tokenizer.vocab_size()):
piece = tokenizer.id_to_piece(token_id) piece = tokenizer.id_to_piece(token_id)
text = piece.encode("utf-8") text = piece.encode("utf-8")
score = tokenizer.get_score(token_id) score = tokenizer.get_score(token_id)
@ -347,9 +352,9 @@ class Model(ABC):
elif tokenizer.is_byte(token_id): elif tokenizer.is_byte(token_id):
toktype = SentencePieceTokenTypes.BYTE toktype = SentencePieceTokenTypes.BYTE
tokens.append(text) tokens[token_id] = text
scores.append(score) scores[token_id] = score
toktypes.append(toktype) toktypes[token_id] = toktype
added_tokens_file = self.dir_model / 'added_tokens.json' added_tokens_file = self.dir_model / 'added_tokens.json'
if added_tokens_file.is_file(): if added_tokens_file.is_file():
@ -357,23 +362,14 @@ class Model(ABC):
added_tokens_json = json.load(f) added_tokens_json = json.load(f)
for key in added_tokens_json: for key in added_tokens_json:
key = key.encode("utf-8") token_id = added_tokens_json[key]
if key not in tokens: if (token_id >= vocab_size):
tokens.append(key) print(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}')
scores.append(-1000.0) continue
toktypes.append(SentencePieceTokenTypes.USER_DEFINED)
if vocab_size > len(tokens): tokens[token_id] = key.encode("utf-8")
pad_count = vocab_size - len(tokens) scores[token_id] = -1000.0
print( toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]"
)
for i in range(1, pad_count + 1):
tokens.append(f"[PAD{i}]")
scores.append(-1000.0)
toktypes.append(SentencePieceTokenTypes.UNUSED)
assert len(tokens) == vocab_size
self.gguf_writer.add_tokenizer_model("llama") self.gguf_writer.add_tokenizer_model("llama")
self.gguf_writer.add_token_list(tokens) self.gguf_writer.add_token_list(tokens)