CANN/cannbot-skills Flash Attention优化笔记

Deep Note:agent/example/kernels/a2/flash_attn_full_pj_hif8_commonub.py

【免费下载链接】cannbot-skillsCANNBot 是面向 CANN 开发的用于提升开发效率的系列智能体，本仓库为其提供可复用的 Skills 模块。项目地址: https://gitcode.com/cann/cannbot-skills

Open this file only after the short catalog entry confirmed the kernel is relevant.

What this kernel is really for

• comparing againstflash_attn_full_pj_hif8.pyafter the math contract is already understood
• studying how a shared vec-side slot buffer changes queueing structure without changing the visible formula

Decisions worth copying

• move vec scratch from two plainTensorviews onto one sharedDBufffamily:ub_score_pv + score_pv_cnt
• keepstage1_cntandstage2_cntseparate even though the shared scratch family exists
• treat the gain as a same-side vecubinqueueing improvement, not as a new cross-side ownership model
• do not expect UB-footprint reduction here; the point is cleaner overlap between the next preload and current vec compute

Prefer another kernel when

• you are still deriving the math contract and want the simpler readable baseline
• you are debugging row-max / row-sum correctness and do not want shared vec scratch lineage in the picture yet

【免费下载链接】cannbot-skillsCANNBot 是面向 CANN 开发的用于提升开发效率的系列智能体，本仓库为其提供可复用的 Skills 模块。项目地址: https://gitcode.com/cann/cannbot-skills