[KongchangAI]
Verified65% confidenceFactExact time

MLA 通过低秩分解将 KV Cache 压缩为低维潜在向量,其 d_latent 约为 d_model 的 1/8,可将 KV Cache 显存占用降低至传统多头注意力的 5%-13%

3
Sources
65%
Confidence
Long-term
Relevance
7/16/2026
First Seen

Sources

Related Claims

Cite This Claim

Stable URI
https://kongchang.com/claim/533862
API
curl https://kongchang.com/api/v1/knowledge/claims/533862
MCP
get_claim(id=533862)