Repository navigation

cuda: implement shared FLASH_ATTN_BACK with direct quantized cache reads - #63

Draft

joelteply wants to merge 3 commits into

fable/p1-3b-run-accumulatorfrom

codex/cuda-flash-attn-back

Draft

cuda: implement shared FLASH_ATTN_BACK with direct quantized cache reads#63
joelteply wants to merge 3 commits into
fable/p1-3b-run-accumulatorfrom
codex/cuda-flash-attn-back

Commits on Oct 14, 2026
- cuda: implement shared flash attention backward
  joelteply
  committed
  34dd2c1
- cuda: use compensated MMA in flash attention backward
  joelteply
  committed
  a0ea747
- cuda: share online attention backward reductions
  joelteply
  committed
  8798492