Skip to content

fix ppo callbacks #9872

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

Open
leuitong wants to merge 1 commit into
base: main
Choose a base branch
from
+7 −1
Open

fix ppo callbacks #9872

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
8 changes: 7 additions & 1 deletion swift/rlhf_trainers/ppo_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,6 +10,7 @@

from swift.trainers import SwiftMixin
from swift.utils import patch_getattr
from swift.callbacks import callbacks_map

if version.parse(trl.__version__) >= version.parse('0.26.0'):
from trl.experimental.ppo import PPOTrainer as HFPPOTrainer
Expand Down Expand Up @@ -49,7 +50,6 @@ def __init__(self, model: PreTrainedModel, ref_model: PreTrainedModel, *_args, *
'reward_model',
'value_model',
'eval_dataset',
'callbacks',
]
}
parameters = inspect.signature(ppo_trainer_init).parameters
Expand All @@ -62,6 +62,12 @@ def __init__(self, model: PreTrainedModel, ref_model: PreTrainedModel, *_args, *
else:
new_kwargs['tokenizer'] = self.tokenizer
ppo_trainer_init(self, model=model, ref_model=ref_model, **new_kwargs)
# TRL's `PPOTrainer.__init__` rebuilds `callback_handler`, discarding the user callbacks that
# `SwiftMixin` registered on the HF Trainer. Re-register them onto the new handler here. We
# call `add_callback` directly (rather than `SwiftMixin._add_callbacks`/`_get_callbacks`,
# whose names/signatures differ between swift branches) so this works across versions.
for callback in self.args.callbacks:
self.add_callback(callbacks_map[callback](self.args, self))
unwrap_model = self.accelerator.unwrap_model(self.model)
patch_getattr(unwrap_model.__class__, 'policy')

Expand Down