May 5, 2026Fused Linear Cross-Entropy : Why fusing the LM head projection with cross-entropy is the single biggest memory win for training LLMs at long context.