[core] Push v_head when using AutoModelForCausalLMWithValueHead
#86
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
younesbelkada
changed the title
v_head when using AutoModelForCausalLMWithValueHead
|
The documentation is not available anymore as the PR was closed or merged. |
lvwerra
reviewed
Jan 19, 2023
trl/models/modeling_base.py
Outdated
| # state_dict is removed from the model after loading it. | ||
| # since this is only done in edge cases we force the user to do it only | ||
| # if they explicitly set resume_training to True | ||
| if resume_training: |
There was a problem hiding this comment.
If someone is using the AutoModelForCausalLMWithValueHead to load a model I think we can assume that they want to use the Value head so I we should always load it if it's available. What do you think?
There was a problem hiding this comment.
OK for me!
We should advice users to use AutoModelForCausalLM when using the model outside trl (i.e. not for training with PPO)
There was a problem hiding this comment.
Added more description in 853d52d
- update based on comments - add more tests - update docs
younesbelkada
commented
Jan 19, 2023
lvwerra
approved these changes
Jan 19, 2023
lvwerra
approved these changes
Jan 19, 2023
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What does this PR do?
This PR addresses saving
v_headweights when callingsave_pretrained. Before this PR if a user wanted to share intermediate-trainedv_headweights it was not possible. The workaround that I have found needs to load the state dict twice, once by thefrom_pretrainedmethod that is called by thepretrained_modelattribute and once afterwards to manually load thev_head.This PR introduces a new kwarg on the method
from_pretrainedtermed asresume_trainingas I think this is quite an edge case.Added also some tests
TODOs:
push_to_hubcc @lvwerra @lewtun