FlashMemory Cuts DeepSeek-V4's KV Cache to 13.5%: Lookahead Sparse Attention
Researchers have developed a new technique called Lookahead Sparse Attention (LSA) that significantly reduces the memory footprint of large language models when processing long contexts. By training a lightweight Neural Memory Indexer, LSA predicts and loads only the essential parts of the KV cache, cutting the memory usage to 13.5% of the full cache size. This method was demonstrated on the DeepSeek-V4 model, showing a reduction in KV cache size and a slight improvement in accuracy. AI
IMPACT Reduces memory costs for long-context LLMs, potentially making them more accessible and efficient for deployment.