When profile_memory, also export and save the snapshot.pickle for lora_finetune_single_device.py #1382

janeyx99 · 2024-08-20T23:08:59Z

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.

Changelog

What are the changes made in this PR?

enable the snapshot.pickle to be exported when profile_memory is True.
the above required adding attributes to self for recipes/lora_finetune_single_device.py
needing to export the snapshot.pickle does mean that it takes a little longer to export.

Previously, you only had the html from the profiler, which is not interactive:

Now, you get that in addition to:

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.)

UX

If your function changed a public API, please add a dummy example of what the user experience will look like when calling it.
Example of docstring:

torchtune/torchtune/modules/vision_transformer.py

Line 285 in 6a7951f

Examples:

Example in our docs: https://pytorch.org/torchtune/main/tutorials/qat_finetune.html#applying-qat-to-llama3-models

I did not change any public API;
I have added an example to docs or docstrings;

pytorch-bot · 2024-08-20T23:09:02Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/torchtune/1382

📄 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 e797f3b with merge base 3c580fc ():
💚 Looks good so far! There are no failures yet. 💚

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

ebsmothers

Thanks for adding this! Can you fix the linter complaints? After that this should be good to go

felipemello1

@ebsmothers, wdyt? We were missing the memory snapshot. My only concern is hardcoding torch.cuda if we plan to support xpu

Edit: you already approved it :P

felipemello1 · 2024-08-20T23:15:07Z

recipes/lora_finetune_single_device.py

@@ -579,6 +584,9 @@ def train(self) -> None:
 ):
 break



Suggested change

# Memory profiling

if curr_epoch == 0 and self.profiler_profile_memory and idx == (self.profiler_wait_steps + self.profiler_warmup_steps):

torch.cuda.memory._record_memory_history()

felipemello1 · 2024-08-20T23:16:08Z

recipes/lora_finetune_single_device.py

@@ -579,6 +584,9 @@ def train(self) -> None:
 ):
 break

+ if curr_epoch == 0 and self.profiler_profile_memory and idx == self.profiler_wait_steps + self.profiler_warmup_steps:
+ torch.cuda.memory._record_memory_history()


The issue is if in the future we support other backend, like intel. It makes me think that this should be a function, instead of being hardcoded as "torch.cuda"

cc: @ebsmothers

The issue is if in the future we support other backend, like intel. It makes me think that this should be a function, instead of being hardcoded as "torch.cuda"

cc: @ebsmothers

I think it's OK, idk that memory snapshot would be supported on XPU anyways. We can do a check on device + profile_memory combo just to be safe though

felipemello1 · 2024-08-20T23:17:02Z

recipes/lora_finetune_single_device.py

@@ -626,6 +634,9 @@ def train(self) -> None:
 num_tokens = 0
 t0 = time.perf_counter()

+ if curr_epoch == 0 and self.profiler_profile_memory and idx == self.profiler_wait_steps + self.profiler_warmup_steps + self.profiler_active_steps:


Suggested change

if curr_epoch == 0 and self.profiler_profile_memory and idx == self.profiler_wait_steps + self.profiler_warmup_steps + self.profiler_active_steps:

# Stop memory profiling

if curr_epoch == 0 and self.profiler_profile_memory and idx == self.profiler_wait_steps + self.profiler_warmup_steps + self.profiler_active_steps:

felipemello1 · 2024-08-20T23:19:15Z

should we just delete the HTML to save time, since it takes a while to save? Or is it useful?

ebsmothers · 2024-08-20T23:30:23Z

recipes/lora_finetune_single_device.py

@@ -328,6 +328,11 @@ def _setup_profiler(

 log.info(f" Profiler config after instantiation: {profiler_cfg}")

+ self.profiler_wait_steps = profiler_cfg["wait_steps"]


This seems to be throwing an error in our recipe tests, can you just quickly inspect profiler_cfg after running e.g.

pytest tests/recipes/test_lora_finetune_single_device.py -m integration_test -k 'test_loss'

I thought these fields would be defined based on this, but seems like something weird is happening here.

Ah, it's because profiling isn't enabled for tests, so it returns here

torchtune/torchtune/utils/_profiler.py

Line 262 in 2f8ed7a

return DummyProfiler(), DictConfig({"enabled": False})

janeyx99 · 2024-08-21T14:49:30Z

should we just delete the HTML to save time, since it takes a while to save? Or is it useful?

Writing the html is very light so it doesn't make a tangible difference, I'm not sure if people are using it today so it'd be bad to just remove it if there were actual users.

codecov-commenter · 2024-08-21T15:11:01Z

Codecov Report

Attention: Patch coverage is 0% with 10 lines in your changes missing coverage. Please review.

Project coverage is 70.56%. Comparing base (3c580fc) to head (e797f3b).
Report is 5 commits behind head on main.

Files	Patch %	Lines
recipes/lora_finetune_single_device.py	0.00%	9 Missing ⚠️
torchtune/utils/_profiler.py	0.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1382      +/-   ##
==========================================
- Coverage   72.57%   70.56%   -2.01%     
==========================================
  Files         271      272       +1     
  Lines       12736    12895     +159     
==========================================
- Hits         9243     9100     -143     
- Misses       3493     3795     +302

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

facebook-github-bot · 2024-08-21T15:39:04Z

@janeyx99 has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

When profile_memory, also export and save the snapshot.pickle

2c6fc92

facebook-github-bot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Aug 20, 2024

janeyx99 requested review from ebsmothers and felipemello1 August 20, 2024 23:09

ebsmothers approved these changes Aug 20, 2024

View reviewed changes

felipemello1 reviewed Aug 20, 2024

View reviewed changes

ebsmothers reviewed Aug 20, 2024

View reviewed changes

Fix lint and fix tests

e797f3b

janeyx99 merged commit 9e65fa9 into pytorch:main Aug 21, 2024
20 checks passed

janeyx99 deleted the snapshot-pickle branch August 21, 2024 16:04

janeyx99 mentioned this pull request Aug 21, 2024

Allow snapshot.pickle to be saved for all recipes with profiler support #1384

Merged

13 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

When profile_memory, also export and save the snapshot.pickle for lora_finetune_single_device.py #1382

When profile_memory, also export and save the snapshot.pickle for lora_finetune_single_device.py #1382

janeyx99 commented Aug 20, 2024

pytorch-bot bot commented Aug 20, 2024 •

edited

Loading

ebsmothers left a comment

felipemello1 left a comment •

edited

Loading

felipemello1 Aug 20, 2024

felipemello1 Aug 20, 2024

ebsmothers Aug 20, 2024

felipemello1 Aug 20, 2024

felipemello1 commented Aug 20, 2024

ebsmothers Aug 20, 2024

janeyx99 Aug 21, 2024

janeyx99 commented Aug 21, 2024

codecov-commenter commented Aug 21, 2024

facebook-github-bot commented Aug 21, 2024

+ # Memory profiling
+ if curr_epoch == 0 and self.profiler_profile_memory and idx == (self.profiler_wait_steps + self.profiler_warmup_steps):
+ torch.cuda.memory._record_memory_history()

	if curr_epoch == 0 and self.profiler_profile_memory and idx == self.profiler_wait_steps + self.profiler_warmup_steps + self.profiler_active_steps:
	# Stop memory profiling
	if curr_epoch == 0 and self.profiler_profile_memory and idx == self.profiler_wait_steps + self.profiler_warmup_steps + self.profiler_active_steps:

		@@ -328,6 +328,11 @@ def _setup_profiler(

		log.info(f" Profiler config after instantiation: {profiler_cfg}")

		self.profiler_wait_steps = profiler_cfg["wait_steps"]

When profile_memory, also export and save the snapshot.pickle for lora_finetune_single_device.py #1382

When profile_memory, also export and save the snapshot.pickle for lora_finetune_single_device.py #1382

Conversation

janeyx99 commented Aug 20, 2024

Context

Changelog

Test plan

UX

pytorch-bot bot commented Aug 20, 2024 • edited Loading

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/torchtune/1382

✅ No Failures

ebsmothers left a comment

Choose a reason for hiding this comment

felipemello1 left a comment • edited Loading

Choose a reason for hiding this comment

felipemello1 Aug 20, 2024

Choose a reason for hiding this comment

felipemello1 Aug 20, 2024

Choose a reason for hiding this comment

ebsmothers Aug 20, 2024

Choose a reason for hiding this comment

felipemello1 Aug 20, 2024

Choose a reason for hiding this comment

felipemello1 commented Aug 20, 2024

ebsmothers Aug 20, 2024

Choose a reason for hiding this comment

janeyx99 Aug 21, 2024

Choose a reason for hiding this comment

janeyx99 commented Aug 21, 2024

codecov-commenter commented Aug 21, 2024

Codecov Report

facebook-github-bot commented Aug 21, 2024

pytorch-bot bot commented Aug 20, 2024 •

edited

Loading

felipemello1 left a comment •

edited

Loading