⚠ Switch to EXCALIDRAW VIEW in the MORE OPTIONS menu of this document. ⚠ You can decompress Drawing data with the command palette: ‘Decompress current Excalidraw file’. For more info check in plugin settings under ‘Saving’

Excalidraw Data

Text Elements

Compute Stream

Communication Stream (A2A)

Stream

BASELINE(No Overlap)

F/Attention

F/Dispatch a2a x 2

F/MLP

F/Combine a2a x 1

B/Combine a2a x 1

B/MLP

B/Attention

Overlap waith W/D Split mlp

F/Attention

F/Dispatch ag + a2a

F/MLP

F/Combine a2a x 1

B/Combine a2a x 1

B/ DgradMLP

B/ WgradMLP

B/Dispatch a2a

B/Dispatch a2a x 1

B/Attention

Speed Up

permute rank-expert major

permute rank-expert major

permute rank-expert major

permute rank-expert major

Compute Stream

Communication Stream (A2A)

F/Attention

F/Dispatch ag + a2a

F/MLP

F/Combine a2a

permute rank-expert major

Overlap waith W/D Split mlp and 2 micro batch

AMB(Another Micro Batch)

B/Combine a2a

AMB

B/ DgradMLP

B/ WgradMLP

permute rank-expert major

B/Dispatch a2a

AMB

AMB

AMB

AMB

B/Attention

AMB

Speed Up

当前的 comb 来看的话 如果先调 forward 会导致 AMB 的 backward 的 mlp 放不到 forward mlp 前面