Repository navigation

Fix DFlash attention masking - #905

Closed

cih9088 wants to merge 7 commits into

sgl-project:mainfrom

cih9088:fix/dflash-swa

Commits on Sep 28, 2026
- feat: respect is_causal attribute for dflash family models
  cih9088
  committed
  e0ba80d
- feat: build attention mask for inference
  cih9088
  committed
  ae70e5d
- refactor: use existing flex attention compile friendly functions
  cih9088
  committed
  1862823
- test: cover causality and swa masking
  cih9088
  committed
  351a994
- test: mask and cached context
  cih9088
  committed
  9d33d51
- style: make isort happy
  cih9088
  committed
  3e06b44
- test: fix
  cih9088
  committed
  7ccc036