Skip to content

[Bugfix] Fix bad words for Mistral models #17753

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
May 8, 2025
Merged

[Bugfix] Fix bad words for Mistral models #17753

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
82af104
Remove special tokens for Mistral when using bad words
qionghuang6 May 6, 2025
601f6ec
Set add_special_tokens=True in for Mistral in logits_process as well
qionghuang6 May 7, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
17 changes: 7 additions & 10 deletions vllm/logits_process.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,11 +4,12 @@

import torch

from vllm.transformers_utils.tokenizer import AnyTokenizer, MistralTokenizer
from vllm.transformers_utils.tokenizer import AnyTokenizer

LogitsProcessor = Union[Callable[[list[int], torch.Tensor], torch.Tensor],
Callable[[list[int], list[int], torch.Tensor],
torch.Tensor]]
LogitsProcessor = Union[
Callable[[list[int], torch.Tensor], torch.Tensor],
Callable[[list[int], list[int], torch.Tensor], torch.Tensor],
]
"""LogitsProcessor is a function that takes a list
of previously generated tokens, the logits tensor
for the next token and, optionally, prompt tokens as a
Expand All @@ -29,12 +30,8 @@ def get_bad_words_logits_processors(
prefix = " " if add_prefix_space else ""
prompt = prefix + bad_word.lstrip()

if isinstance(tokenizer, MistralTokenizer):
# Mistral tokenizers should not add special tokens
prompt_token_ids = tokenizer.encode(text=prompt)
else:
prompt_token_ids = tokenizer.encode(text=prompt,
add_special_tokens=False)
prompt_token_ids = tokenizer.encode(text=prompt,
add_special_tokens=False)

# If no space at the beginning
# or if prefix space produces a new word token
Expand Down
10 changes: 2 additions & 8 deletions vllm/sampling_params.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,7 +13,6 @@
from vllm.logger import init_logger
from vllm.logits_process import LogitsProcessor
from vllm.transformers_utils.tokenizer import AnyTokenizer
from vllm.transformers_utils.tokenizers.mistral import MistralTokenizer

logger = init_logger(__name__)

Expand Down Expand Up @@ -491,13 +490,8 @@ def update_from_tokenizer(self, tokenizer: AnyTokenizer) -> None:
for add_prefix_space in [False, True]:
prefix = " " if add_prefix_space else ""
prompt = prefix + bad_word.lstrip()

if isinstance(tokenizer, MistralTokenizer):
# Mistral tokenizers should not add special tokens
prompt_token_ids = tokenizer.encode(text=prompt)
Comment on lines -495 to -497
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Looks like the issue was caused by the BOS token so this makes sense

vllm/vllm/transformers_utils/tokenizers/mistral.py

Lines 369 to 374 in 98c89e1

if add_special_tokens is not None:
return self.tokenizer.encode(text,
bos=add_special_tokens,
eos=add_special_tokens)
else:
return self.tokenizer.encode(text, bos=True, eos=False)

else:
prompt_token_ids = tokenizer.encode(
text=prompt, add_special_tokens=False)
prompt_token_ids = tokenizer.encode(text=prompt,
add_special_tokens=False)

# If no space at the beginning
# or if prefix space produces a new word token
Expand Down