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

Respect tokenizer.ggml.add_bos_token value when tokenizing (#4040)

Browse source 
* gguf-py: gguf-dump: Respect --no-tensor flag in JSON mode.

* Respect add_bos_token GGUF metadata value

* gguf-py: Try to fix SpecialVocab giving up too easily for the Nth time
This commit is contained in:
Kerfuffle 2023-11-16 19:14:37 -07:00 • committed by GitHub
parent 8da46278e1
commit 91f6499393
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
12 changed files with 85 additions and 29 deletions
Download patch file Download diff file Expand all files Collapse all files

25
gguf-py/gguf/vocab.py
View file

@ -117,17 +117,18 @@ class SpecialVocab:
def _try_load_from_tokenizer_json(self, path: Path) -> bool:
tokenizer_file = path / 'tokenizer.json'
if not tokenizer_file.is_file():
return False
with open(tokenizer_file, encoding = 'utf-8') as f:
tokenizer = json.load(f)
if self.load_merges:
merges = tokenizer.get('model', {}).get('merges')
if isinstance(merges, list) and merges and isinstance(merges[0], str):
self.merges = merges
if tokenizer_file.is_file():
with open(tokenizer_file, encoding = 'utf-8') as f:
tokenizer = json.load(f)
if self.load_merges:
merges = tokenizer.get('model', {}).get('merges')
if isinstance(merges, list) and merges and isinstance(merges[0], str):
self.merges = merges
added_tokens = tokenizer.get('added_tokens', {})
else:
added_tokens = {}
tokenizer_config_file = path / 'tokenizer_config.json'
added_tokens = tokenizer.get('added_tokens')
if added_tokens is None or not tokenizer_config_file.is_file():
if not tokenizer_config_file.is_file():
return True
with open(tokenizer_config_file, encoding = 'utf-8') as f:
tokenizer_config = json.load(f)
@ -135,6 +136,10 @@ class SpecialVocab:
add_entry = tokenizer_config.get(f'add_{typ}_token')
if isinstance(add_entry, bool):
self.add_special_token[typ] = add_entry
if not added_tokens:
# We will need this to get the content for the token, so if it's empty
# may as well just give up.
continue
entry = tokenizer_config.get(f'{typ}_token')
if isinstance(entry, str):
tc_content = entry