graphs
beyond softmax: next-gen attention
beyond softmax: next-gen attention
softmax is the bottleneck
swap to sparse top-k kernel
dry-dawn-9050
linear-attn-v2
gqa-headcount sweep
calm-sun-5567
ship the sparse variant
sparsity beats width
KV-cache is the real bottleneck
KINDSrootempiricalhypothesisinsightdecisioncontrol
STATUSdonerunningfailedstaged
click to focus · ⇧ click to select · esc to clear
GRAPHSNODES