[RFC] Proposal to Update PPO Test to Add LR Scheduler #2423
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/torchtune/2423
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 17cb9ed with merge base 952078e ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
|
Hi @Seoley! Thank you for your pull request and welcome to our community. Action RequiredIn order to merge any pull request (code, docs, etc.), we require contributors to sign our Contributor License Agreement, and we don't seem to have one on file for you. ProcessIn order for us to review and merge your suggested changes, please sign at https://code.facebook.com/cla. If you are contributing on behalf of someone else (eg your employer), the individual CLA may not be sufficient and your employer may need to sign the corporate CLA. Once the CLA is signed, our tooling will perform checks and validations. Afterwards, the pull request will be tagged with If you have received this in error or have any questions, please contact us at cla@meta.com. Thanks! |
facebook-github-bot
added
the
CLA Signed
|
Thank you for signing our Contributor License Agreement. We can now accept your code for this (and any) Meta Open Source project. Thanks! |
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #2423 +/- ##
==========================================
+ Coverage 63.87% 65.30% +1.43%
==========================================
Files 368 374 +6
Lines 21873 22163 +290
==========================================
+ Hits 13971 14474 +503
+ Misses 7902 7689 -213 ☔ View full report in Codecov by Sentry. |
| self._steps_per_epoch = ( | ||
| len(self._dataloader) // self._gradient_accumulation_steps | ||
| ) | ||
| self.global_step = self._epochs_run * self._steps_per_epoch |
There was a problem hiding this comment.
global_step is already defined here
| @@ -257,6 +258,18 @@ def setup(self, cfg: DictConfig) -> None: | |||
| * (self.batch_size // self._ppo_batch_size) | |||
| ) | |||
|
|
|||
| self._steps_per_epoch = ( | |||
There was a problem hiding this comment.
I think we should name this lr_steps and the correct value would be the total number of optimizer steps being taken, which should be self._total_steps * self._ppo_epochs * (self.batch_size // self._ppo_batch_size), right?
|
Hi @Seoley! Thanks so much for taking this on. Sorry for the delay in getting round to a review here. I think this generally looks good - I think the only comments I have are how we calculate the number of steps for the scheduler, which I've left inline. Let me know what you think : ) |
Context
Changelog
Related Issue: #2064 (comment)
I refered to #2263 and add an lr_scheduler option to the Mistral PPO recipe. Then, run the recipe using the following command:
tune run ppo_full_finetune_single_device --config mistral/7B_full_ppo_low_memoryHere are the loss curves for reference:
And then, I need to update the test code in
recipes/test_ppo_full_finetune_single_device.pyThis commit includes changes where I added the lr_scheduler option to recipes/test_ppo_full_finetune_single_device.py, but the test has failed. I suspect the failure might be due to using LLaMA.
How should I resolve this issue? I would like to ask for some advice.
Test plan
Please make sure to do each of the following if applicable to your PR. If you're unsure about any one of these just ask and we will happily help. We also have a contributing page for some guidance on contributing.
pre-commit install)pytest testspytest tests -m integration_testUX
If your function changed a public API, please add a dummy example of what the user experience will look like when calling it.
Here is a docstring example
and a tutorial example