Proud of my @RutgersCS undergrad @Chen94751623484 for this work! 🎓🔥 No NaNs ≠ correct gradients. Even after fixing the visible training instability, large errors can hide in the attention backward. The culprit: BF16 rounding breaks a zero-sum conservation law. A small rounding error can overwhelm the true gradient as keys grow and attention sharpens. 🛠️ GProj restores that law after rounding. In our experiments: query/key gradient accuracy and final loss on par with FP32 attention, with just ~4.7% extra step time. 🔍 Training looks healthy? Check the backward pass anyway.