Add MoE to Gemma4 TP plan#45219
Conversation
Signed-off-by: Wang, Yi <[email protected]>
Rocketknight1
left a comment
There was a problem hiding this comment.
Happy to approve this one and people can yell at me later if there's any problem!
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
|
@Rocketknight1 thanks for the approval, I check the failure in ci, nothing to do with the PR |
Cyrilvallez
left a comment
There was a problem hiding this comment.
Thanks, indeed the MoE part was forgotten! Sorry @Rocketknight1, I dismissed your review as I saw it was in the merging queue and I panicked thinking the mlp part should be removed, but it should stay as well haha
|
[For maintainers] Suggested jobs to run (before merge) run-slow: gemma4 |
reduce memory for gemma4 moe model in tp Signed-off-by: Wang, Yi <[email protected]> Co-authored-by: Cyril Vallez <[email protected]>
reduce memory for gemma4 moe model in tp Signed-off-by: Wang, Yi <[email protected]> Co-authored-by: Cyril Vallez <[email protected]>
reduce memory for gemma4 moe model in tp Signed-off-by: Wang, Yi <[email protected]> Co-authored-by: Cyril Vallez <[email protected]>
What does this PR do?
google/gemma-4-26B-A4B-it
tp 2, memory is 46G per rank wo the change, drop to about 25G w per rank with the change