Skip to content

[JAX] Make MoEBlock aware of Dense TP axes - #3462

Draft
jberchtold-nvidia wants to merge 3 commits into
NVIDIA:mainfrom
jberchtold-nvidia:jberchtold/fix-tp-etp1-awareness-in-moeblock
Draft

[JAX] Make MoEBlock aware of Dense TP axes#3462
jberchtold-nvidia wants to merge 3 commits into
NVIDIA:mainfrom
jberchtold-nvidia:jberchtold/fix-tp-etp1-awareness-in-moeblock

Conversation

@jberchtold-nvidia

@jberchtold-nvidia jberchtold-nvidia commented Sep 2, 2026

Copy link
Copy Markdown
Collaborator

Description

Before this PR, the TE MoEBlock supported EP+FSDP but had some constraints that prevented using TP in other parts of the model. Importantly, this PR does not add TP support to the MoEBlock's grouped GEMMs. This PR simply makes the TE MoEBlock aware of TP axes used in Dense layers of the model so those GPUs can participate as further EP GPUs when inside the context of the MoEBlock if desired.

Type of change

  • Documentation change (change only to the documentation, either a fix or a new content)
  • Bug fix (non-breaking change which fixes an issue)
  • New feature (non-breaking change which adds functionality)
  • Breaking change (fix or feature that would cause existing functionality to not work as expected)
  • Infra/Build change
  • Code refactoring

Changes

Please list the changes introduced in this PR:

  • Change A
  • Change B

Checklist:

  • I have read and followed the contributing guidelines
  • The functionality is complete
  • I have commented my code, particularly in hard-to-understand areas
  • I have made corresponding changes to the documentation
  • My changes generate no new warnings
  • I have added tests that prove my fix is effective or that my feature works
  • New and existing unit tests pass locally with my changes

Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>
@jberchtold-nvidia
jberchtold-nvidia marked this pull request as draft September 2, 2026 16:41
@jberchtold-nvidia
jberchtold-nvidia force-pushed the jberchtold/fix-tp-etp1-awareness-in-moeblock branch from ff57020 to 4dfb6b7 Compare September 2, 2026 16:43
@jberchtold-nvidia jberchtold-nvidia changed the title [JAX] Make MoEBlock aware of TP/ETP axes that can participate as EP domain in context of MoEBlock [JAX] Make MoEBlock aware of Dense TP axes Sep 2, 2026
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.

1 participant