fix(kernel): make Triton top-k/top-p sampling exact - #329

Merged

jason-fxz merged 4 commits into

FlashML-org:mainfrom

taking-lying-flat:fix/triton-topk-correctness

Sep 3, 2026

Commits on Sep 1, 2026

fix(sampling): handle Triton top-k candidate limits
taking-lying-flat
committed
9be573d

Commits on Sep 2, 2026

fix(kernel): exact single-launch triton top-k/top-p sampling

jason-fxz
authored and
taking-lying-flat
committed
de5b6cb
fix(sampling): make fused filtering exact and portable
taking-lying-flat
committed
b51c580

Commits on Sep 3, 2026

fix(sampling): preserve boundary ties
taking-lying-flat
committed
65fedd4