Skip to content

Support direct hybrid MTP layer specs - #6298

Draft
Phlip79 wants to merge 1 commit into
NVIDIA:pull-request/6297from
Phlip79:philip/direct-hybrid-mtp
Draft

Support direct hybrid MTP layer specs#6298
Phlip79 wants to merge 1 commit into
NVIDIA:pull-request/6297from
Phlip79:philip/direct-hybrid-mtp

Conversation

@Phlip79

@Phlip79 Phlip79 commented Aug 5, 2026

Copy link
Copy Markdown
Member

Summary

  • Build and repeat direct MTP templates on the final logical PP/VPP chunk.
  • Preserve checkpoint keys and assign direct MoE metrics per prediction depth.
  • Leave legacy MTP construction, repeated-layer behavior, and metrics unchanged.

Stacked on #6297.

@copy-pr-bot

copy-pr-bot Bot commented Aug 5, 2026

Copy link
Copy Markdown

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

Phlip79 commented Aug 5, 2026

Copy link
Copy Markdown
Member Author

/ok to test 597a182

@Phlip79

Phlip79 commented Aug 5, 2026

Copy link
Copy Markdown
Member Author

/ok to test 748c5d1

@Phlip79

Phlip79 commented Aug 5, 2026

Copy link
Copy Markdown
Member Author

/ok to test e2734a3

Signed-off-by: Philip Petrakian <ppetrakian@nvidia.com>
@Phlip79
Phlip79 force-pushed the philip/direct-hybrid-mtp branch from e2734a3 to dff1c46 Compare August 5, 2026 23:18
@copy-pr-bot
copy-pr-bot Bot force-pushed the pull-request/6297 branch from 89d2874 to e344ee2 Compare August 5, 2026 23:19
@Phlip79

Phlip79 commented Aug 5, 2026

Copy link
Copy Markdown
Member Author

/ok to test dff1c46

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