Archives/gpt4all
2
0
Fork 0
mirror of https://github.com/nomic-ai/gpt4all synced 2024-11-02 09:40:42 +00:00
Code Issues Packages Projects Releases Wiki Activity

convert_mpt_hf_to_gguf.py: better tokenizer decoding

Browse Source
This commit is contained in:
Cebtenzzre 2023-09-29 10:02:04 -04:00 committed by Adam Treat
parent 25297786db
commit cca9e6ce81
1 changed files with 15 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
gpt4all-backend/scripts/convert_mpt_hf_to_gguf.py
View File

@ -101,17 +101,27 @@ tokenizer = AutoTokenizer.from_pretrained(model_name)
special_ids = tokenizer.all_special_ids special_ids = tokenizer.all_special_ids
reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
added_tokens = tokenizer.get_added_vocab().values()
byte_encoder = bytes_to_unicode()
byte_decoder = {v: k for k, v in byte_encoder.items()}
tokens: list[bytearray] = [] tokens: list[bytearray] = []
toktypes: list[gguf.TokenType] = [] toktypes: list[gguf.TokenType] = []
# TODO(cebtenzzre): this is probably wrong, but I don't know what else to put here
dot_token = tokenizer.encode('.')[0]
# The number of tokens in tokenizer.json can differ from the expected vocab size. # The number of tokens in tokenizer.json can differ from the expected vocab size.
# This causes downstream issues with mismatched tensor sizes when running the inference # This causes downstream issues with mismatched tensor sizes when running the inference
for i in range(config.vocab_size): for i in range(config.vocab_size):
text = tokenizer.decode([dot_token, i]).encode('utf-8') if i not in reverse_vocab:
text = text[1:] # remove the first byte (it's always '.') print(f"Key {i} not in tokenizer vocabulary. Padding with an arbitrary token.")
pad_token = f"[PAD{i}]".encode("utf8")
text = bytearray(pad_token)
elif i in added_tokens:
# these tokens are not encoded, for some reason
text = bytearray(reverse_vocab[i].encode('utf-8'))
else:
text = bytearray([byte_decoder[c] for c in reverse_vocab[i]])
tokens.append(text) tokens.append(text)
# TODO(cebtenzzre): is there a better way to do this? # TODO(cebtenzzre): is there a better way to do this?