DeepSeek Cuts KV Cache Footprint Fourfold
A causal encoder-decoder MoE pairs cross-layer sparse-attention reuse with FP4 caching to reduce global context memory to 890 bytes per token.
Sep 19, 20264 min2609.19969
2 articles on SOTA Papers
A causal encoder-decoder MoE pairs cross-layer sparse-attention reuse with FP4 caching to reduce global context memory to 890 bytes per token.
ROCS separates request-side and candidate-side computation, reaching up to 3× higher retrieval QPS and 50% higher ranking QPS in production.
