Skip to content

[Fix] Improve MTP compatibility and mitigate MoE load imbalance by skipping pad tokens#1952

Open
tina-wen wants to merge 2 commits into
InternLM:mainfrom
tina-wen:mem-opt-bugfix
Open

[Fix] Improve MTP compatibility and mitigate MoE load imbalance by skipping pad tokens#1952
tina-wen wants to merge 2 commits into
InternLM:mainfrom
tina-wen:mem-opt-bugfix

Conversation

@tina-wen

Copy link
Copy Markdown
Contributor

This PR includes two targeted fixes:

  • Improve MTP compatibility in the MTP block and MoE path to ensure correct behavior in related training flows.
  • Mitigate MoE load imbalance by skipping pad tokens in the MoE decoder layer, reducing unnecessary expert dispatch.

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.

2 participants