@Akashi203 on Backlist

DeepSeek sparse attention’s hidden fp32 tensor problem

DeepSeek’s compressed sparse attention can still materialize a massive fp32 indexing tensor before top-k selection, making long context memory behavior nontrivial