Skip to content

Added option to prevent tokens from being penalized for repetition #426

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

Closed
wants to merge 1 commit into from
Closed

Added option to prevent tokens from being penalized for repetition #426

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
9 changes: 7 additions & 2 deletions exllamav2/generator/sampler.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,6 +19,7 @@ class Settings:

token_frequency_penalty: float = 0.0
token_presence_penalty: float = 0.0
non_rep_penalized_tokens: list[int] = field(default_factory = list)

temperature: float = 0.8
smoothing_factor: float = 0.0
Expand Down Expand Up @@ -159,14 +160,18 @@ def sample(logits: torch.tensor,
if settings.token_repetition_penalty != 1.0 or \
settings.token_frequency_penalty != 0.0 or \
settings.token_presence_penalty != 0.0:

ext_c.apply_rep_penalty(sequence_ids[:, :],

hold_back_rep_penalty = len(settings.non_rep_penalized_tokens) > 0 # Revert the logits for certain tokens
if hold_back_rep_penalty: original_logits = torch.clone(logits[:, :]) # Copy the original logits
ext_c.apply_rep_penalty(sequence_ids[:, :], # Apply the repetition penalty from cpp side
settings.token_repetition_penalty,
settings.token_repetition_range,
settings.token_repetition_decay,
settings.token_frequency_penalty,
settings.token_presence_penalty,
logits)
if hold_back_rep_penalty: # Restore the original logits for non-penalized tokens
for i in settings.non_rep_penalized_tokens: logits[:, i] = original_logits[:, i]

# Token bias

Expand Down