Technical overview of DeepSeek V4 Compressed Sparse Attention (CSA) and Hierarchical Context Attention (HCA) for million-token context and KV cache memory optimization.