fix(tokenizer): handle fast tokenizer properly for bos/eos (#914)

This commit is contained in:
NanoCode012
2023-12-08 11:31:13 +09:00
committed by GitHub
parent 06ae39200b
commit fde091cb12

View File

@@ -92,6 +92,7 @@ def load_tokenizer(cfg):
"LlamaTokenizer",
"LlamaTokenizerFast",
"CodeLlamaTokenizer",
"CodeLlamaTokenizerFast",
]
and hasattr(tokenizer, "pad_token")
and not tokenizer.pad_token
@@ -124,6 +125,23 @@ def load_tokenizer(cfg):
tokenizer.add_special_tokens(
{k: AddedToken(val, rstrip=False, lstrip=False, normalized=False)}
)
# If we add bos_token and eos_token, we need to update the post processor to
# handle them correctly.
# https://github.com/huggingface/transformers/pull/24132
bos_or_eos_in_special_tokens = (
"bos_token" in cfg.special_tokens and "eos_token" in cfg.special_tokens
)
if (
tokenizer.__class__.__name__
in (
"LlamaTokenizerFast",
"CodeLlamaTokenizerFast",
)
and bos_or_eos_in_special_tokens
):
tokenizer.update_post_processor()
if cfg.tokens:
tokenizer.add_tokens(
[