[FEATURE SUPPORT] Variable-Length Attention with Padding-Free Execution#188

Merged

LoserCheems merged 13 commits intomainfrom

Oct 11, 2025

Commits on Oct 10, 2025

Fix spelling of "optimisation" to "optimization" in issue and pull request templates
algo-home
committed

Commits on Oct 11, 2025

Rename 'scale' parameter to 'softmax_scale' in dynamic mask attention functions for clarity
algo-home
committed
Rename 'scale' parameter to 'softmax_scale' in API documentation and integration guide for consistency
algo-home
committed
Add integration guide for Flash Dynamic Mask Attention with detailed architecture and implementation details
algo-home
committed
Rename 'scale' parameter to 'softmax_scale' in example code for clarity and consistency
algo-home
committed
Rename 'scale' parameter to 'softmax_scale' in flex and triton attention functions for consistency
algo-home
committed
Removes mask/bias from varlen; defaults nondet
algo-home
committed
Add padding/unpadding helpers for QKV attention
algo-home
committed
Adds varlen FDMA and padding-free attention
algo-home
committed
Enable per-call override of causal flag and window
algo-home
committed
Add varlen MHA fwd/bwd; enable paged KV (no mask)
algo-home
committed
Update flash_dmattn/integrations/modeling_flash_dynamic_mask_attention_utils.py

LoserCheems
and
Copilot
authored
Update flash_dmattn/integrations/modeling_flash_dynamic_mask_attention_utils.py

LoserCheems
and
Copilot
authored