Repository navigation

cuda window_attention: head dim 512 over a linear cache, for global attention - #1124

Closed

jschmied wants to merge 2 commits into

ashhart:mainfrom

jschmied:cuda-attn-global

Closed

cuda window_attention: head dim 512 over a linear cache, for global attention#1124
jschmied wants to merge 2 commits into
ashhart:mainfrom
jschmied:cuda-attn-global

Commits on Oct 19, 2026
- cuda window_attention: head dim 256, any GQA, a row's last key and fp32 probabilities
  jschmied
  committed
  a2bed92
Commits on Oct 20, 2026
- cuda window_attention: head dim 512 over a linear cache, for global attention
  jschmied
  committed
  7424800