Repository navigation

Support fdma for models with attention bias - #41882

Closed

LoserCheems wants to merge 11 commits into

huggingface:mainfrom

LoserCheems:support-fdma

Closed

Support fdma for models with attention bias#41882
LoserCheems wants to merge 11 commits into
huggingface:mainfrom
LoserCheems:support-fdma

Commits on Oct 27, 2025
- Adds flash_dmattn availability check
  LoserCheems
  committed
  83df813
- Adds flash dmattn availability check export
  LoserCheems
  committed
  ac44a55
- Add test gate for Flash Dynamic Mask Attention
  LoserCheems
  committed
  ef0d943
- Adds Flash Dynamic Mask Attention integration
  LoserCheems
  committed
  d100255
- Adds flash dynamic mask attention utilities
  LoserCheems
  committed
  a5ba48c
- Add Flash Dynamic Mask Attention support
  LoserCheems
  committed
  0cca929
- Adds flash_dmattn to attention backends docs
  LoserCheems
  committed
  4cd7e56
- Removes unused import and simplifies causal check
  LoserCheems
  committed
  5d8cba2
- Reformat Flash Dynamic Mask Attention availability error messages for readability
  LoserCheems
  committed
  4e9d978
- Reformat require_flash_dmattn skipUnless call for readability
  LoserCheems
  committed
  50e3b1d
- Rename flash_dmattn to flash_dynamic_mask_attention in attention_interface.md
  LoserCheems
  committed
  ed0faf6