Revision History
Flash Attention · 12 revisions
Sizes are character counts of the article source. The signed number is the change from the previous revision.
Recent edit summaries
Detailed summaries recorded by editors. Generic maintenance summaries are omitted here; every recorded revision remains below. Dates describe the edit, not necessarily the event it covers.
- Correction: FlashAttention-3 pingpong scheduling described as overlapping softmax of one consumer warp group with another's matmuls (not producer/consumer stages), per the NeurIPS 2024 paper; Ninja build-time wording fixed. Update: FlashAttention-4 MLSys 2026 publication, release status to October 2026, PyTorch 2.10-2.15 and Transformers FA4 integration, ROCm backend details, loop order, version history.
Version 14 · Oct 31, 2026, 03:43 AM
- Improve internal links (2026-10-13)
Version 13 · Oct 13, 2026, 10:07 AM
- Correction: fact-checked FlashAttention algorithms, versions, support, and scoped benchmarks
Version 12 · Jul 31, 2026, 01:03 PM
- Added 6 contextual internal links
Version 11 · Jul 23, 2026, 03:03 PM
- TEST: large deletion for testing alerts
Version 7 · Apr 12, 2026, 06:33 PM