lookahead-sparse-attention

Proposes Lookahead Sparse Attention with a Neural Memory Indexer on DeepSeek-V4, reducing GPU memory usage to ~13.5% of full-context baseline while maintaining or slightly improving accuracy.