#llm-serving

Attention Is a Memory Allocator Problem: FlashAttention-2 and PagedAttention Under the Microscope

Modern transformer performance is limited less by math and more by how precisely we move and allocate memory.