https://github.com/HazyResearch/flash-attention#memory https://github.com/HazyResearch/flash-attention#memory "standard attention has memory quadratic in seque

https://github.com/HazyResearch/flash-attention#memory https://github.com/HazyResearch/flash-attention#memory "standard attention has memory quadratic in sequence length, whereas FlashAttention has memory linear in sequence length." I guess you have just reported how many times the layer will need to access the memory, not how much memory usage scales with sequence length.