Jump to content
Main menu
Main menu
move to sidebar
hide
Navigation
Main page
Recent changes
Random page
freem
Search
Search
Appearance
Create account
Log in
Personal tools
Create account
Log in
Pages for logged out editors
learn more
Contributions
Talk
Editing
Openai/692aba85-5b10-8006-817a-d5a83b020f8e
(section)
Add languages
Page
Discussion
English
Read
Edit
Edit source
View history
Tools
Tools
move to sidebar
hide
Actions
Read
Edit
Edit source
View history
General
What links here
Related changes
Special pages
Page information
Appearance
move to sidebar
hide
Warning:
You are not logged in. Your IP address will be publicly visible if you make any edits. If you
log in
or
create an account
, your edits will be attributed to your username, along with other benefits.
Anti-spam check. Do
not
fill this in!
====== 2.1.8 Mixture-of-Experts の効率化 ====== MoE 層には torchtitan の実装を使い、grouped GEMM カーネルで専門家をまとめて実行。 * 実験の結果、 - シーケンス長と hidden dim が大きい今回の設定では、 既に grouped GEMM が十分に飽和しており、 - Expert Parallelism(EP)を入れてもスループットがむしろ下がることが分かった(scatter/gather オーバーヘッドのみ増える)。INTELLECT_3_Technical_Report * 図5(p9)のベンチマークでは、 シーケンス長 32k〜65k では 128 experts まではTFLOPSが飽和領域にとどまるため、EPは不要という判断。 推論側では HF の MoE 実装を使う必要があるため、トレーナ側の state dict をHF形式にオンザフライ変換して vLLM に渡す工夫もしています。
Summary:
Please note that all contributions to freem are considered to be released under the Creative Commons Attribution-ShareAlike 4.0 (see
Freem:Copyrights
for details). If you do not want your writing to be edited mercilessly and redistributed at will, then do not submit it here.
You are also promising us that you wrote this yourself, or copied it from a public domain or similar free resource.
Do not submit copyrighted work without permission!
Cancel
Editing help
(opens in new window)