#sparsity 1 item 8 сен Don't Drop Dropout: Optimizing Layer Sparsity for Efficient LLM Training and Inference Cerebras research