Unverified50% confidenceFactExact time
The KV Cache size of sliding window attention is constant, achieving O(1) memory usage rather than O(n).
1
Sources
50%
Confidence
Long-term
Relevance
8/31/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe computational complexity of liquid convolution scales linearly with sequence length at O(n), rather than the quadratic O(n²) of self-attention, resulting in approximately 1/1000 the computation of attention layers when processing 128K tokens.68% similarUnverifiedTechniques such as Sliding Window Attention and RoPE positional encoding extension have expanded the context window from the early 2K tokens to 128K or even the million-token level.61% similarUnverifiedThe context window of modern large models has expanded from the early 4K tokens to 128K or even longer, and the computational complexity of the attention mechanism grows quadratically with context length.60% similarUnverified上下文越长，模型需要缓存的 KV Cache 越大，KV Cache 使模型生成新 Token 时无需重新计算所有历史 Token 的注意力权重59% similarUnverifiedvLLM compresses the KV Cache memory fragmentation rate from above 60% to near zero through PagedAttention technology, and can increase inference throughput by several times in high-concurrency scenarios.59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/827470API
curl https://kongchang.com/api/v1/knowledge/claims/827470MCP
get_claim(id=827470)