[Enhancement] Support ZeRO-3 when using BAdam #4352
Merged
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
hiyouga
reviewed
Jun 18, 2024
| @@ -371,6 +371,12 @@ def _create_badam_optimizer( | |||
| dict(params=decay_params, weight_decay=training_args.weight_decay), | |||
| ] | |||
|
|
|||
| ds_zero3_enabled = False | |||
| if hasattr(training_args, "deepspeed_plugin") and training_args.deepspeed_plugin is not None: | |||
There was a problem hiding this comment.
Why not use from transformers.integrations import is_deepspeed_zero3_enabled
There was a problem hiding this comment.
Thanks for the suggestion, I have changed it to use is_deepspeed_zero3_enabled for cleaner expressions.
hiyouga
added
solved
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Outdated suggestions cannot be applied.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 enables BAdam algorithm to use model parallelism, based on the implementation of deepspeed ZeRO-3.
A sample script is provided in "examples/extras/badam/train_zero3.sh". The running command generated by the current webUI works correctly as well.
When training Llama 3-8B on "alpaca_en_demo" dataset with batch size 1, the maximum per device allocated memory is about 13/10/8 GB when training with 2/3/4 RTX3090 GPUs, respectively. I suppose it would be feasible to finetune a Llama 3-70B model given 8 RTX-3090 / 3 A100-80G using BAdam, though I haven't conduct a comprehensive test due to the limited computation resources.
The main change in code is to add the a BAdam's callback during Trainer's initialization, when
use_badamand ZeRO-3 mode are detected.Thanks for review!
Before submitting