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

Fix handling pre-normalized tokens

Browse source
This commit is contained in:
Billel Mokeddem 2024-12-22 20:12:46 +00:00
parent 92e41ec4b9
commit a1f146dba1
1 changed files with 5 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

9
convert_hf_to_gguf.py
View file

@ -525,10 +525,9 @@ class Model:
else: else:
token: str = reverse_vocab[i] token: str = reverse_vocab[i]
if token in added_vocab: if token in added_vocab:
# We need to manually encode and decode the added tokens in case special characters # The tokenizer in llama.cpp assumes the CONTROL and USER_DEFINED tokens are pre-normalized.
# used for `\n` / `\t` have been manually added in the added tokens # To avoid unexpected issues - we make sure to normalize non-normalized tokens
# To avoid unexpected issues - we make sure to encode single-char tokens if not tokenizer.added_tokens_decoder[i].normalized:
if len(token) == 1:
previous_token = token previous_token = token
token = tokenizer.decode(tokenizer.encode(token, add_special_tokens=False)) token = tokenizer.decode(tokenizer.encode(token, add_special_tokens=False))
if previous_token != token: if previous_token != token:
@ -537,6 +536,8 @@ class Model:
if tokenizer.added_tokens_decoder[i].special or self.does_token_look_special(token): if tokenizer.added_tokens_decoder[i].special or self.does_token_look_special(token):
toktypes.append(gguf.TokenType.CONTROL) toktypes.append(gguf.TokenType.CONTROL)
else: else:
# NOTE: this was added for Gemma.
# Encoding and decoding the tokens above isn't sufficient for this case.
token = token.replace(b"\xe2\x96\x81".decode("utf-8"), " ") # pre-normalize user-defined spaces token = token.replace(b"\xe2\x96\x81".decode("utf-8"), " ") # pre-normalize user-defined spaces
toktypes.append(gguf.TokenType.USER_DEFINED) toktypes.append(gguf.TokenType.USER_DEFINED)
else: else: