NVIDIA NeMo · Automodel
Remote
- Found masked cross-entropy losses overwriting the caller's labels in place, which zeroed the loss on a reused batch (#3917), and opened the fix.
- Made ChunkedCrossEntropy reductions independent of chunk length, and returned zero instead of NaN for batches with no supervised tokens.
- Flagged CI going red on main when a PR passes against a stale base (#4170).