https://localfirstai.eu/tags/flash-attention/Recent content in Flash-Attention on Local First AIHugoen-usSun, 07 Jun 2026 00:00:00 +0000https://localfirstai.eu/posts/2026-06-07-the-cliff-that-wasnt/Sun, 07 Jun 2026 00:00:00 +0000https://localfirstai.eu/posts/2026-06-07-the-cliff-that-wasnt/Six months of cascade architecture rested on a 20K token prefill cliff. The cliff was real. The cause was a flag that should have been off. Two experiments to find it, one daemon restart to fix it.