Skip to content

Conversation

@ChiGod0502
Copy link

1, Remove BOS and EOS tokens during both training and testing; otherwise, one of the vision embedding tokens will be replaced by the BOS token.

2, Ensure that only a single checkpoint is saved for each saving_step when gradient accumulation is enabled.

Only save once at saving step when accumulation_step != 1
otherwise the action_hidden_states will contains a text token, and the task_latten_states will contais a bos token
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant