Skip to content

fix lora_target_modules in PPO #2274

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
Oct 18, 2024
Merged

fix lora_target_modules in PPO #2274

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
9c4ea11
fix
hjh0119 Oct 17, 2024
248edca
fix
hjh0119 Oct 17, 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
15 changes: 9 additions & 6 deletions swift/llm/rlhf.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,26 +18,29 @@ def llm_rlhf(args: RLHFArguments) -> Dict[str, Any]:
if is_generation:
logger.warning(f"Please check if args.template_type: '{args.template_type}' is correct.")

msg = {}
model, ref_model, template, callbacks = prepare_model_template_train(args)
kwargs = {}
if args.rlhf_type == 'ppo':
from copy import deepcopy
reward_model_args = deepcopy(args)
reward_model_args, value_model_args = deepcopy(args), deepcopy(args)
args_to_modified = ['model_id_or_path', 'model_type', 'model_revision']
for arg in args_to_modified:
setattr(reward_model_args, arg, getattr(args, f'reward_{arg}'))
for model_args in [reward_model_args, value_model_args]:
for arg in args_to_modified:
setattr(model_args, arg, getattr(args, f'reward_{arg}'))
reward_model_args.ref_model_free = True # avoid to create ref model
value_model_args.ref_model_free = True
reward_model, _, _, _ = prepare_model_template_train(reward_model_args)
reward_model.requires_grad_(False).eval()

reward_model = get_model_with_value_head(reward_model) # add and load value head
# hack here to customize the value model
value_model, _, _, _ = prepare_model_template_train(reward_model_args)
value_model, _, _, _ = prepare_model_template_train(value_model_args)
value_model = get_model_with_value_head(value_model)
kwargs['reward_model'] = reward_model
kwargs['value_model'] = value_model

msg = {}
model, ref_model, template, callbacks = prepare_model_template_train(args)

with TrainerFactory.patch_template(args, template):
train_dataset, val_dataset = prepare_dataset(args, template, msg)

Expand Down
Loading