Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

First initialize dist with gloo #1133

Merged
merged 16 commits into from
Apr 24, 2024
Merged

First initialize dist with gloo #1133

merged 16 commits into from
Apr 24, 2024

Conversation

dakinggg
Copy link
Collaborator

@dakinggg dakinggg commented Apr 24, 2024

There seems to be an issue with the first nccl barrier, where the timeout does not properly apply. It does apply correctly to a gloo barrier though. So we add a gloo barrier at the start to make sure all ranks are alive, and then proceed with the normal process group intialization.

Run with this that runs successfully: gloo-baseline-1-cAcvEB
Run with this change and a rank sleeping in the gloo section (does timeout): gloo-messed-up-no-gloo-2-CZw9ru (on a7616ea)
Run with this change and a rank sleeping in the nccl section (does not timeout): gloo-messed-up-5-HB6gOH (on ac258b9)

@dakinggg dakinggg marked this pull request as ready for review April 24, 2024 17:47
scripts/train/train.py Outdated Show resolved Hide resolved
scripts/train/train.py Outdated Show resolved Hide resolved
scripts/train/train.py Outdated Show resolved Hide resolved
Copy link
Contributor

@irenedea irenedea left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

minor comments, thank you!

scripts/train/train.py Outdated Show resolved Hide resolved
scripts/train/train.py Outdated Show resolved Hide resolved
@mvpatel2000
Copy link
Collaborator

This might break on >300 GPUs per @sandeep-krishnamurthy-db's point on GLOO's limitations. Can you doubel check max size for GLOO?

dakinggg and others added 5 commits April 24, 2024 12:10
Co-authored-by: Mihir Patel <mihir.v.patel7@gmail.com>
Copy link
Collaborator

@mvpatel2000 mvpatel2000 left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

CC: @bigning this is roughly what im thinking for control plane except we wouldn't destroy it and would name it something else

@dakinggg dakinggg merged commit 76f74b6 into mosaicml:main Apr 24, 2024
9 checks passed
dakinggg added a commit to dakinggg/llm-foundry that referenced this pull request Apr 25, 2024
dakinggg added a commit that referenced this pull request Apr 25, 2024
dakinggg added a commit to dakinggg/llm-foundry that referenced this pull request Apr 25, 2024
@dakinggg dakinggg deleted the gloo branch June 22, 2024 20:46
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

None yet

3 participants