deepseek-ai/DeepSeek-MoE
1
0
Fork 0
mirror of https://github.com/deepseek-ai/DeepSeek-MoE.git synced 2026-11-05 00:57:43 +00:00
Code Issues 24 Projects Packages Wiki Activity

[GH-ISSUE #29] About expert capacities: Is there token-dropping during training? #26

New issue
Closed
opened 2026-10-26 21:01:49 +00:00 by bsakaguchi · 3 comments
bsakaguchi commented 2026-10-26 21:01:49 +00:00
Owner
Copy link

Originally created by @Spico197 on GitHub (Mar 4, 2024).
Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-MoE/issues/29

Hi there, thanks for open-sourcing such an amazing LM~

I'm wondering if you limit the maximum expert capacity by dropping some tokens during training like Switch Transformer.
Since training large MoE models is pretty costly, maybe it's necessary to add a hard constraint to ensure faster expert parallelism?

Thanks again and looking forward to hearing from you.

image

Originally created by @Spico197 on GitHub (Mar 4, 2024). Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-MoE/issues/29 Hi there, thanks for open-sourcing such an amazing LM~ I'm wondering if you limit the maximum expert capacity by dropping some tokens during training like Switch Transformer. Since training large MoE models is pretty costly, maybe it's necessary to add a hard constraint to ensure faster expert parallelism? Thanks again and looking forward to hearing from you. ![image](https://github.com/deepseek-ai/DeepSeek-MoE/assets/22840952/6a2c7216-1dfb-476a-8cdf-65a14dfb3b8f)
bsakaguchi commented 2026-10-26 21:01:50 +00:00
Author
Owner
Copy link

@zwd003 commented on GitHub (Apr 3, 2024):

@DeepSeekDDM

<!-- gh-comment-id:2034703272 --> @zwd003 commented on GitHub (Apr 3, 2024): @DeepSeekDDM
bsakaguchi commented 2026-10-26 21:01:50 +00:00
Author
Owner
Copy link

@DeepSeekDDM commented on GitHub (Apr 3, 2024):

DeepSeekMoE 16B does not have a dropping strategy since it does not use expert parallelism. If using expert parallelism, token dropping will be applied to ensure training efficiency.

<!-- gh-comment-id:2035053064 --> @DeepSeekDDM commented on GitHub (Apr 3, 2024): DeepSeekMoE 16B does not have a dropping strategy since it does not use expert parallelism. If using expert parallelism, token dropping will be applied to ensure training efficiency.
bsakaguchi commented 2026-10-26 21:01:51 +00:00
Author
Owner
Copy link

@Spico197 commented on GitHub (Apr 7, 2024):

Thanks for your valuable comments~ I found the relevant contents from the technical report. Thanks again for your response~

image

<!-- gh-comment-id:2041311811 --> @Spico197 commented on GitHub (Apr 7, 2024): Thanks for your valuable comments~ I found the relevant contents from the technical report. Thanks again for your response~ ![image](https://github.com/deepseek-ai/DeepSeek-MoE/assets/22840952/cdfa8af4-b08a-4386-80e9-04e57ca95966)
Sign in to join this conversation.
No Branch/Tag specified
Branches Tags
main
No results found.
Labels
Clear labels   
pull-request
Mirrored from GitHub Pull Request

  
WIP
No labels
pull-request
WIP
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
deepseek-ai/DeepSeek-MoE#26
No description provided.