Reduce memory consumption when training with PPO #2571
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
qgallouedec
reviewed
Jan 17, 2025
| @@ -176,7 +176,8 @@ def forward( | |||
| **kwargs, | |||
| ) | |||
|
|
|||
| last_hidden_state = base_model_output.hidden_states[-1] | |||
| # last_hidden_state = base_model_output.hidden_states[-1] | |||
There was a problem hiding this comment.
Suggested change
| # last_hidden_state = base_model_output.hidden_states[-1] |
qgallouedec
reviewed
Jan 17, 2025
| @@ -485,6 +485,7 @@ def generate( | |||
| if generate_ref_response: | |||
| ref_model = self.model if self.is_peft_model else self.ref_model | |||
| if isinstance(query_tensor, List): | |||
| self.model.eval() | |||
There was a problem hiding this comment.
why do you need it in eval model here?
qgallouedec
reviewed
Jan 17, 2025
| full_kl_penalty = self.config.kl_penalty == "full" | ||
|
|
||
| # torch.cuda.memory._record_memory_history() |
There was a problem hiding this comment.
Suggested change
| # torch.cuda.memory._record_memory_history() |
qgallouedec
reviewed
Jan 17, 2025
| @@ -164,7 +164,7 @@ def forward( | |||
| kwargs (`dict`, `optional`): | |||
| Additional keyword arguments, that are passed to the wrapped model. | |||
| """ | |||
| kwargs["output_hidden_states"] = True # this had already been set in the LORA / PEFT examples | |||
| kwargs["output_hidden_states"] = False # this had already been set in the LORA / PEFT examples | |||
There was a problem hiding this comment.
Isn't it the default value? maybe we can simply remove it
qgallouedec
reviewed
Jan 17, 2025
| @@ -417,7 +417,7 @@ def prepare_dataloader(self, dataset: Union[torch.utils.data.Dataset, Dataset], | |||
| dataset, | |||
| batch_size=self.config.batch_size, | |||
| collate_fn=data_collator, | |||
| shuffle=True, | |||
| shuffle=False, | |||
qgallouedec
reviewed
Jan 17, 2025
| @@ -744,6 +745,7 @@ def step( | |||
| model_inputs, | |||
| response_masks=response_masks, | |||
| return_logits=full_kl_penalty, | |||
| use_cache=False | |||
There was a problem hiding this comment.
Why do you disable the cache?
qgallouedec
reviewed
Jan 17, 2025
| ) | ||
|
|
||
| # torch.cuda.memory._dump_snapshot(f"ppo_batched_forward_pass-{time.time()}.pickle") |
There was a problem hiding this comment.
Suggested change
| # torch.cuda.memory._dump_snapshot(f"ppo_batched_forward_pass-{time.time()}.pickle") |
qgallouedec
reviewed
Jan 17, 2025
| @@ -813,13 +816,14 @@ def step( | |||
| mini_batch_dict[k] = batch_dict[k][mini_batch_inds] | |||
| with self.accelerator.accumulate(self.model): | |||
| model_inputs = {k: mini_batch_dict[k] for k in model_inputs_names} | |||
|
|
|||
| # torch.cuda.memory._record_memory_history() | |||
There was a problem hiding this comment.
Suggested change
| # torch.cuda.memory._record_memory_history() |
|
Thanks for the PR. It seems to be still a draft right? |
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?
Fixes # (issue)
Before submitting
Pull Request section?
to it if that's the case.
documentation guidelines.
Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.