Revision History
Flash Attention 3 · 8 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.
- Improve internal links (2026-10-13)
Version 8 · Oct 13, 2026, 10:20 AM
- Correction: remove three uncited claims (an engine-spread attribution absent from the cited post, an sm_121 prediction written as fact, and a sweeping claim that three unrelated lines of work build on the FA3 codebase)
Version 7 · Oct 1, 2026, 06:47 PM
- Correction: the FP8 ablation table had its two rows reversed and the conclusion drawn from them was backwards; register reallocation is dynamic; several figures re-sourced from the NeurIPS camera-ready rather than the PyTorch blog; FA4 and FlexAttention claims corrected; the FlashMLA inheritance claim reduced to what its README says
Version 6 · Oct 1, 2026, 06:47 PM
- Correction: 660 TFLOPS is the April 2025 update, not a later one (day one was 580); TFLOPS unit
Version 5 · Oct 1, 2026, 10:04 AM
- Added 6 contextual internal links
Version 4 · Jul 23, 2026, 03:29 PM