Dao-AILab/flash-attention fa4-v4.0.0.beta9

on GitHub

What's Changed

• feat(cute): implement softcap backward pass, correct math formula, and resolve JIT cache bug by @CaesarG in #2402
• [Cute,Sm100,Fwd] add MLA 64/512 with topk sparsity for MQA 128 heads by @jayhshah in #2441
• Handle linter for flash mla file by @jayhshah in #2459

New Contributors

• @CaesarG made their first contribution in #2402

Full Changelog: fa4-v4.0.0.beta8...fa4-v4.0.0.beta9