"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.