Skip to content
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

FIX [PromptTuning] Simple fix for transformers >= 4.38 #1484

Merged
merged 2 commits into from
Feb 19, 2024
Merged

FIX [PromptTuning] Simple fix for transformers >= 4.38 #1484

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
d683bb4
fix for transformers >= 4.38
younesbelkada Feb 19, 2024
1c457ea
style
younesbelkada Feb 19, 2024
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
6 changes: 6 additions & 0 deletions src/peft/peft_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1205,6 +1205,12 @@ def prepare_inputs_for_generation(self, *args, task_ids: torch.Tensor = None, **
model_kwargs["inputs_embeds"] = torch.cat((prompts, inputs_embeds), dim=1)
model_kwargs["input_ids"] = None

# For transformers>=4.38.0 - for some architectures such as Llama, `cache_position` is
# passed in the forward pass to keep track of the position ids of the cache. We have to
# pop that from `model_kwargs` as `cache_position` is properly created by the model, using the passed
# `inputs_embeds`: https://github.com/huggingface/transformers/blob/593230f0a1150ea9c0477b9d859f25daf73c8c33/src/transformers/models/llama/modeling_llama.py#L956
_ = model_kwargs.pop("cache_position", None)

return model_kwargs


Expand Down
Loading