Make loss calculation deterministic and sync-free - #103

Open

ndryden wants to merge 3 commits into

triton-kernelsfrom

loss-determinism

Commits on Aug 6, 2026

Sum the cross-entropy numerator outside the loss kernel
ndryden
committed
b4027d7
Build the CE normalizer without reading device memory
ndryden
committed
61b3918

Commits on Aug 7, 2026

Correct two comments the block-list emptying invalidated
ndryden
committed
785dcf4