[Multi-Adapter PPO] Fix and Refactor reward model adapter #982
[Multi-Adapter PPO] Fix and Refactor reward model adapter #982younesbelkada merged 4 commits intohuggingface:mainfrom
Conversation
more flexible, clearer args
unwrap model since it is DDP downside, with reward adapter it seems we need to use find_unused_parameters=True
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. |
younesbelkada
left a comment
younesbelkada
left a comment
There was a problem hiding this comment.
Great clean up! Thanks a ton for working on this @mnoukhov !
I left only one question, what do you think?
| """ | ||
| transformers_parent_class = None | ||
| supported_args = None | ||
| supported_args = ("score_module", "supports_rm_adapter", "rm_adapter_name") |
There was a problem hiding this comment.
Do you know if these are used? It seems we always overwrite them: https://github.com/huggingface/trl/blob/main/trl/models/modeling_value_head.py#L90
There was a problem hiding this comment.
I made this change so it would be a bit more obvious the args being passed in. I'm happy to change it back if you'd like
I actually don't think we use supported_args anywhere. I assumed it would be a future feature for some sort of argparsing. To be more in line with transformers we would probably want a config class instead of supported args anyways so it isn't a big deal either way.
There was a problem hiding this comment.
OK I see, thanks!
Yes it is not a big deal, technically I have set it to None so that the PreTrainedModelWrapper can't be used as it is (I should have set that class as an abstract class). Yes it would be great if you can change this back then we can merge I think
| """ | ||
| transformers_parent_class = None | ||
| supported_args = None | ||
| supported_args = ("score_module", "supports_rm_adapter", "rm_adapter_name") |
There was a problem hiding this comment.
| supported_args = ("score_module", "supports_rm_adapter", "rm_adapter_name") | |
| supported_args = None |
…e#982) * reward adapter loaded as part of init more flexible, clearer args * fixed script for multi gpu unwrap model since it is DDP downside, with reward adapter it seems we need to use find_unused_parameters=True * remove gradient from reward score calculation * change supported_args back to None
…e#982) * reward adapter loaded as part of init more flexible, clearer args * fixed script for multi gpu unwrap model since it is DDP downside, with reward adapter it seems we need to use find_unused_parameters=True * remove gradient from reward score calculation * change supported_args back to None
3 participants
A simpler and cleaner version of #472 focused on the main issues
examples/multi_adapter_ppo.pyfails with multi-gpumodel.compute_reward_scoreifmodelis wrapped inDDPunwrap_modelRuntimeError: Expected to have finished reduction in the prior iteration before starting a new one....torch.no_grad()and make sure that the reward adapter'sscorehasrequires_grad=Falsein
modeling_base.py, create the reward adapter before initializing the model and pass score intoinitPeftModel.load_adapterwhich includes a lot of useful logic