Add profiler to full finetune recipes

[gau-nernst]

Context

What is the purpose of this PR? Is it to

• add a new feature
• fix a bug
• update tests and/or documentation
• other (please add here)

Please link to any issues this PR addresses. #1279

Changelog

What are the changes made in this PR?

Add profiler to full_finetune_distributed.py and full_finetune_single_device.py. Basically copied over from LoRA recipes.

Test plan

Please make sure to do each of the following if applicable to your PR. (If you're not sure about any one of these just ask and we will happily help. We also have a contributing page for some guidance on contributing.)

• run pre-commit hooks and linters (make sure you've first installed via pre-commit install)
• add unit tests for any new functionality
• update docstrings for any new or updated methods or classes
• run unit tests via pytest tests
• run recipe tests via pytest tests -m integration_test
• manually run any new or modified recipes with sufficient proof of correctness
• include relevant commands and any other artifacts in this summary (pastes of loss curves, eval results, etc.)

Just like the LoRA recipe, to profile, run:

tune run full_finetune_single_device \
  --config llama2/7B_full_low_memory \
  profiler.enabled=True \
  profiler.output_dir=./profile-test

Then we can open the trace under profile-test/iteration_12 with something like https://ui.perfetto.dev

And it reveals that, in my setup, most of the time is spent transferring data back and forth between GPU and CPU for paged Adam.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/torchtune/1288

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit 56f4adc with merge base 150a011 ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[SalmanMohammadi]

nit:

Suggested change

	self, cfg_profiler: Optional[DictConfig]
	self, cfg_profiler: Optional[DictConfig] = None

[SalmanMohammadi]

nit:

Suggested change

	cfg_profiler (DictConfig): `profiler` section of the top-level `cfg` (the main config passed to `recipe.main`)
	cfg_profiler (Optional[DictConfig]): ``profiler`` section of the top-level ``cfg`` (the main config passed to ``recipe.main``). Default None.

[SalmanMohammadi]

I think this is the case in all the recipes. Sorry about our annoying linter : )

[SalmanMohammadi]

Minor nit to appease our linter but LGTM. Thanks for getting this out!!

[SalmanMohammadi]

cc @ebsmothers @joecummings to double check whether a quick example profile would be good to see

[ebsmothers]

Thanks @gau-nernst, this looks great! Re @SalmanMohammadi's comment: can you include a screenshot of one of the traces you got in the PR summary just to explicitly show that it works via an example? Also regarding the failing CI signals, can you check the linter one? (You can ignore the build signals, there is an unrelated breakage we are sorting out right now.) Let me know if you need any pointers on the linter and I'll be happy to provide them. After that I think this is good to go!

[gau-nernst]

Fix the linter issue. Sorry I didn't install pre-commit before making the first commit, so it slips through subsequent checks. Manually run pylintdoc to double check and no issue now. Also add the full finetune low memory trace to the PR description.

[ebsmothers]

Looks great, thank you for adding this!