Dao-AILab/flash-attention fa4-v4.0.0.beta30

on GitHub

What's Changed

• [CuTe, SM80/SM120] Guard invalid varlen forward tiles by @eamonn-zh in #2763
• [CuTe] Include callable defaults in JIT cache hashes by @guoriyue in #2773
• [CuTe,Bwd,Sm100] Support head_dim not a multiple of 32 in backward by @KaijingOfficial in #2698
• [CuTe] Speed up SM100 hdim-64 bwd: dedicated P/dS TMEM slots + per-warp signaling by @jack-carlisle-2025 in #2804
• cute: don't widen intentionally-empty offset windows to full attention by @cora-codes in #2490
• Fix seqlen_k_loaded treating a window bound of 0 as unbounded (sliding window) by @lollinng in #2624
• Fix SM90 bwd crash for head_dim in (128, 192] by @cora-codes in #2482

New Contributors

• @guoriyue made their first contribution in #2773
• @KaijingOfficial made their first contribution in #2698
• @jack-carlisle-2025 made their first contribution in #2804
• @cora-codes made their first contribution in #2490
• @lollinng made their first contribution in #2624

Full Changelog: fa4-v4.0.0.beta29...fa4-v4.0.0.beta30