USP: Unified (a.k.a. Hybrid, 2D) Sequence Parallel Attention for Long Context Transformers Model Training and Inference
313
stars
19
forks
source link
convert input tensor layout inside stripe and zigzag ring attention #46
Closed
feifeibear closed 5 months ago
Now, we have to prepare a correct input layout for zigzag and stripe before execute forward().