Add MemoryProfileCallback

[ShriyaPalsamudram]

What does this PR do ?

This callback enables recording a timeline of memory allocations during training.

Collection: [Note which collection this PR will affect]

Changelog

• Add specific line by line info of high level changes in this PR.

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The Jenkins CI system has been replaced by GitHub Actions self-hosted runners.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

[maanug-nv]

Currently, this will dump the snapshot for all ranks. Should we add an option to only save a snapshot from rank 0? or a specifiable rank?

[ShriyaPalsamudram]

Currently, this will dump the snapshot for all ranks. Should we add an option to only save a snapshot from rank 0? or a specifiable rank?

Addressed in latest commit

[maanug-nv]

Callback LGTM. I have no concerns about memory profiling, since this is almost identical to the one callback we've been using. For memory cycles, we should test by introducing an intentional cycle. Also curious if this produces warnings about any cycles not related to our code (eg in PTL).

[maanug-nv]

also, can we allow disabling the profiling, so users can only run the cycle detector?

[hemildesai]

LGTM, just one minor comment.

[hemildesai]

Can we move this import inside the callback to be safe?

[ShriyaPalsamudram]

Isn't it better to fail early if there is an issue with this import than fail inside the callback?

[hemildesai]

Yeah I was just being cautious to avoid side effects of the import but I don't think this import should have any, so it's probably fine. Get Outlook for iOS<https://aka.ms/o0ukef>

________________________________ From: Shriya Rishab ***@***.***> Sent: Friday, August 23, 2024 10:07:29 AM To: NVIDIA/NeMo ***@***.***> Cc: Hemil Desai ***@***.***>; Comment ***@***.***> Subject: Re: [NVIDIA/NeMo] Add MemoryProfileCallback (PR #10166) @ShriyaPalsamudram commented on this pull request.

________________________________ In nemo/lightning/pytorch/callbacks/memory_profiler.py<#10166 (comment)>:

@@ -0,0 +1,78 @@

+import os + +import torch +from pytorch_lightning.callbacks.callback import Callback +from torch.utils.viz._cycles import warn_tensor_cycles Isn't it better to fail early if there is an issue with this import than fail inside the callback? — Reply to this email directly, view it on GitHub<#10166 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AB6Q25H6ZIEEJZOY5WCDAWDZS5F3DAVCNFSM6AAAAABMS6X6TCVHI2DSMVQWIX3LMV43YUDVNRWFEZLROVSXG5CSMV3GSZLXHMZDENJXGQ3TONZVGY>. You are receiving this because you commented.Message ID: ***@***.***>