enables default data step in megatron parallel to operate on a wider variety of tensors by jomitchellnv · Pull Request #9641 · NVIDIA-NeMo/NeMo

jomitchellnv · 2024-07-08T18:35:20Z

…variety of tensors coming out of the dataloader

Add more flexibility to the default data step inside megatron parallel.

Add a one line overview of what this PR aims to accomplish.

Collection: [Note which collection this PR will affect]

Changelog

Megatron parallel's default data step can operate on wrapped tuples and unwrapped items coming directly out of the dataloader

Usage

You can potentially add a usage example below

# Add a code snippet demonstrating how to use this

GitHub Actions CI

The Jenkins CI system has been replaced by GitHub Actions self-hosted runners.

The GitHub Actions CI will run automatically when the "Run CICD" label is added to the PR.
To re-run CI remove and add the label again.
To run CI on an untrusted fork, a NeMo user with write access must first click "Approve and run".

Before your PR is "Ready for review"

Pre checks:

Make sure you read and followed Contributor guidelines
Did you write any new necessary tests?
Did you add or update any necessary documentation?
Does the PR affect components that are optional to install? (Ex: Numba, Pynini, Apex etc)
- Reviewer: Does the PR have correct import guards for all optional libraries?

PR Type:

New Feature
Bugfix
Documentation

If you haven't finished some of the above items you can still open "Draft" PR.

Who can review?

Anyone in the community is free to review the PR once the checks have passed.
Contributor guidelines contains specific people who can review PRs to various areas.

Additional Information

Related to # (issue)

…variety of tensors coming out of the dataloader

nemo/lightning/megatron_parallel.py

Signed-off-by: jomitchellnv <jomitchellnv@users.noreply.github.com>

nemo/lightning/megatron_parallel.py

+        case _:
+            batch = None
+
+    if batch is not None:


jstjohn · 2024-07-08T19:11:57Z

nemo/lightning/megatron_parallel.py

+        # Canonical case.
+        case batch:
+            pass
+        # If the dataloader_iter is empty, return None.


This comment is not what happens anymore. Now in your else you say raise ValueError(...):. Are we sure we don't want to pass through a None on the iter? Otherwise we should prob change line 79 or 84 to return None and drop the ValueError piece.

Actually what I hope would happen is that the match(dataloader_iter) would trigger an StopIteration or whatever it is exception rather than returning None. The StopIteration exception ideally would raise through this function (what happens now on the match) and be handled elsewhere. So I think your case _ should never get hit...

Maybe a better error message on line 84 would be "None batch returned at the end of iteration rather than a StopIteration exception being raised, why?"

So I only added this to be past the linter. I updated the comment so that its more accurate. My only concern would be how it plays out with drop_last=True since maybe that would dump some None's?

than just dictionaries Signed-off-by: Jonathan Mitchell <jomitchell@nvidia.com>

…variety of tensors (#9641) * enables default data step in megatron parallel to operate on a wider variety of tensors coming out of the dataloader * handles the case where a batch is empty * Apply isort and black reformatting Signed-off-by: jomitchellnv <jomitchellnv@users.noreply.github.com> * Allows the default data step to operate on more types than just dictionaries Signed-off-by: Jonathan Mitchell <jomitchell@nvidia.com> --------- Signed-off-by: jomitchellnv <jomitchellnv@users.noreply.github.com> Signed-off-by: Jonathan Mitchell <jomitchell@nvidia.com> Co-authored-by: jomitchellnv <jomitchellnv@users.noreply.github.com> Co-authored-by: Marc Romeyn <mromeijn@nvidia.com>

…variety of tensors (#9641) * enables default data step in megatron parallel to operate on a wider variety of tensors coming out of the dataloader * handles the case where a batch is empty * Apply isort and black reformatting Signed-off-by: jomitchellnv <jomitchellnv@users.noreply.github.com> * Allows the default data step to operate on more types than just dictionaries Signed-off-by: Jonathan Mitchell <jomitchell@nvidia.com> --------- Signed-off-by: jomitchellnv <jomitchellnv@users.noreply.github.com> Signed-off-by: Jonathan Mitchell <jomitchell@nvidia.com> Co-authored-by: jomitchellnv <jomitchellnv@users.noreply.github.com> Co-authored-by: Marc Romeyn <mromeijn@nvidia.com> Signed-off-by: Tugrul Konuk <ertkonuk@gmail.com>

…variety of tensors (NVIDIA-NeMo#9641) * enables default data step in megatron parallel to operate on a wider variety of tensors coming out of the dataloader * handles the case where a batch is empty * Apply isort and black reformatting Signed-off-by: jomitchellnv <jomitchellnv@users.noreply.github.com> * Allows the default data step to operate on more types than just dictionaries Signed-off-by: Jonathan Mitchell <jomitchell@nvidia.com> --------- Signed-off-by: jomitchellnv <jomitchellnv@users.noreply.github.com> Signed-off-by: Jonathan Mitchell <jomitchell@nvidia.com> Co-authored-by: jomitchellnv <jomitchellnv@users.noreply.github.com> Co-authored-by: Marc Romeyn <mromeijn@nvidia.com> Signed-off-by: tonyjie <jl4257@cornell.edu>

…variety of tensors (#9641) * enables default data step in megatron parallel to operate on a wider variety of tensors coming out of the dataloader * handles the case where a batch is empty * Apply isort and black reformatting Signed-off-by: jomitchellnv <jomitchellnv@users.noreply.github.com> * Allows the default data step to operate on more types than just dictionaries Signed-off-by: Jonathan Mitchell <jomitchell@nvidia.com> --------- Signed-off-by: jomitchellnv <jomitchellnv@users.noreply.github.com> Signed-off-by: Jonathan Mitchell <jomitchell@nvidia.com> Co-authored-by: jomitchellnv <jomitchellnv@users.noreply.github.com> Co-authored-by: Marc Romeyn <mromeijn@nvidia.com>

…variety of tensors (NVIDIA-NeMo#9641) * enables default data step in megatron parallel to operate on a wider variety of tensors coming out of the dataloader * handles the case where a batch is empty * Apply isort and black reformatting Signed-off-by: jomitchellnv <jomitchellnv@users.noreply.github.com> * Allows the default data step to operate on more types than just dictionaries Signed-off-by: Jonathan Mitchell <jomitchell@nvidia.com> --------- Signed-off-by: jomitchellnv <jomitchellnv@users.noreply.github.com> Signed-off-by: Jonathan Mitchell <jomitchell@nvidia.com> Co-authored-by: jomitchellnv <jomitchellnv@users.noreply.github.com> Co-authored-by: Marc Romeyn <mromeijn@nvidia.com> Signed-off-by: Hainan Xu <hainanx@nvidia.com>

…variety of tensors (NVIDIA-NeMo#9641) * enables default data step in megatron parallel to operate on a wider variety of tensors coming out of the dataloader * handles the case where a batch is empty * Apply isort and black reformatting Signed-off-by: jomitchellnv <jomitchellnv@users.noreply.github.com> * Allows the default data step to operate on more types than just dictionaries Signed-off-by: Jonathan Mitchell <jomitchell@nvidia.com> --------- Signed-off-by: jomitchellnv <jomitchellnv@users.noreply.github.com> Signed-off-by: Jonathan Mitchell <jomitchell@nvidia.com> Co-authored-by: jomitchellnv <jomitchellnv@users.noreply.github.com> Co-authored-by: Marc Romeyn <mromeijn@nvidia.com>

enables default data step in megatron parallel to operate on a wider …

4dfbe36

…variety of tensors coming out of the dataloader

github-advanced-security bot found potential problems Jul 8, 2024

View reviewed changes

nemo/lightning/megatron_parallel.py Fixed Show fixed Hide fixed

handles the case where a batch is empty

70cd4c4

jomitchellnv force-pushed the jm/default-data-step branch from 3f53682 to 70cd4c4 Compare July 8, 2024 18:57

Apply isort and black reformatting

1778c4f

Signed-off-by: jomitchellnv <jomitchellnv@users.noreply.github.com>

github-advanced-security bot found potential problems Jul 8, 2024

View reviewed changes

nemo/lightning/megatron_parallel.py

case _:

batch = None

if batch is not None:

Check failure

Code scanning / CodeQL

Potentially uninitialized local variable

Local variable 'batch' may be used before it is initialized.

jstjohn reviewed Jul 8, 2024

View reviewed changes

ericharper requested a review from marcromeyn July 8, 2024 20:36

marcromeyn previously approved these changes Jul 9, 2024

View reviewed changes

Merge branch 'main' into jm/default-data-step

339b0b6

marcromeyn added the Run CICD label Jul 9, 2024

jomitchellnv dismissed marcromeyn’s stale review via 0e08f92 July 9, 2024 19:03

Allows the default data step to operate on more types

5971ff5

than just dictionaries Signed-off-by: Jonathan Mitchell <jomitchell@nvidia.com>

jomitchellnv force-pushed the jm/default-data-step branch from 14c5733 to 5971ff5 Compare July 9, 2024 22:57

Merge branch 'main' into jm/default-data-step

1240a7f

marcromeyn added Run CICD and removed Run CICD labels Jul 10, 2024

marcromeyn self-requested a review July 10, 2024 08:19

marcromeyn approved these changes Jul 10, 2024

View reviewed changes

marcromeyn merged commit f5d5221 into NVIDIA-NeMo:main Jul 10, 2024

marcromeyn mentioned this pull request Jul 10, 2024

Revert "enables default data step in megatron parallel to operate on a wider variety of tensors" #9666

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

enables default data step in megatron parallel to operate on a wider variety of tensors#9641

enables default data step in megatron parallel to operate on a wider variety of tensors#9641
marcromeyn merged 6 commits intoNVIDIA-NeMo:mainfrom
jomitchellnv:jm/default-data-step

jomitchellnv commented Jul 8, 2024

Uh oh!

Uh oh!

Check failure

jstjohn Jul 8, 2024

Uh oh!

jstjohn Jul 8, 2024 •

edited

Loading

Uh oh!

jstjohn Jul 8, 2024 •

edited

Loading

Uh oh!

jomitchellnv Jul 9, 2024

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

jomitchellnv commented Jul 8, 2024

Changelog

Usage

GitHub Actions CI

Before your PR is "Ready for review"

Who can review?

Additional Information

Uh oh!

Uh oh!

Check failure

jstjohn Jul 8, 2024

Choose a reason for hiding this comment

Uh oh!

jstjohn Jul 8, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

jstjohn Jul 8, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

jomitchellnv Jul 9, 2024

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

jstjohn Jul 8, 2024 •

edited

Loading

jstjohn Jul 8, 2024 •

edited

Loading