https://blog.ans20xx.com/tags/flashattention/Recent content in FlashAttention on 安橙的博客Hugo -- 0.163.3zhSat, 20 Jun 2026 00:00:00 +0000https://blog.ans20xx.com/posts/ai/day14/Sat, 20 Jun 2026 00:00:00 +0000https://blog.ans20xx.com/posts/ai/day14/Phase 1 收官:复盘 PyTorch 核心抽象、Autograd、算子后端、显存管理、AMP 与 torch.compile,再用 FlashAttention 拆开算子融合与 IO-aware kernel 的本质。https://blog.ans20xx.com/posts/ai/day26/Sat, 20 Jun 2026 00:00:00 +0000https://blog.ans20xx.com/posts/ai/day26/进入 Transformer 算子层加速:理解 FlashAttention v2/v3、PyTorch SDPA、xFormers 与 Apex Fused Kernels 的适用边界,动手替换 attention,并用 TFLOPS / MFU 判断优化是否真的生效。