Attention is seven questions wearing one name
Change the mask, share the key-value cache, then assemble a mechanism from seven design choices. Causal attention, GQA, FlashAttention and RoPE each answer a different one.
Tools / models / mechanisms
Small interactive tools for seeing how models work inside: change one design choice and watch what it changes.
02 tools
Change the mask, share the key-value cache, then assemble a mechanism from seven design choices. Causal attention, GQA, FlashAttention and RoPE each answer a different one.
Follow a sentence from token lookup, through attention that reads both ways, to one pooled vector you can compare with cosine similarity.