2026년 5월 5일Fused Linear Cross-Entropy : Why fusing the LM head projection with cross-entropy is the single biggest memory win for training LLMs at long context.#training#kernels#memory