Implement variable-length attention with mask and bias support#185

Merged

LoserCheems merged 15 commits intomainfrom

Oct 9, 2025

Commits on Oct 5, 2025

Implement varlen MHA fwd with mask/bias support
algo-home
committed

Commits on Oct 6, 2025

Commits on Oct 7, 2025

Add Chinese API reference documentation for Flash Dynamic Mask Attention
algo-home
committed
Updates API docs with quick start and backend guide
algo-home
committed

Commits on Oct 9, 2025

Rename keep_window_size to window_size in DogeConfig and DogeAttention classes for consistency
algo-home
committed
Rename keep_window_size to window_size in flash_dynamic_mask_attention_forward for consistency
algo-home
committed
Refactor mask and bias parameters to be optional in flash_dmattn functions
algo-home
committed
Rename scale parameter to softmax_scale in flash_dmattn_func for clarity and consistency
algo-home
committed
Adds varlen Flash-DM attention (CUDA) + autograd
algo-home
committed
Import flash_dmattn_varlen_func alongside flash_dmattn_func for backend availability
algo-home
committed
Rename keep_window_size to window_size for consistency in README files
algo-home
committed
Rename keep_window_size to window_size for consistency in API documentation
algo-home
committed
Refactor fdma_peft_integration_check and _flash_dynamic_mask_attention_forward for clarity and consistency; rename keep_window_size to window_size and enhance FlashDynamicMaskAttentionKwargs docume…
algo-home
committed