自注意力
- softmax.py safe/online
- fa.py np/torch/block sparse
Basic idea behind flash attention (V1) - en
From Online Softmax to FlashAttention. Zihao Ye. 2023https://courses.cs.washington.edu/courses/cse599m/23sp/notes/flashattn.pdf
【手撕LLM-Flash Attention】从softmax说起,保姆级超长文!!小冬瓜AIGC
【手撕LLM-FlashAttention2】只因For循环优化的太美 小冬瓜AIGC