Unverified50% confidenceFactExact time
The 70B version of LLaMA 2 uses 8-group GQA, reducing the KV cache to 1/4 of MHA while barely affecting model quality.
1
Sources
50%
Confidence
Long-term
Relevance
8/31/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedGQA groups attention heads to share KV projections, which can compress the KV cache to 1/4 or less of the original size with almost no loss in model quality.64% similarUnverifiedDeepSeek-V2 论文显示 MLA 可将 KV Cache 显存降低至标准 MHA 的约5%至13%，同时维持相近性能64% similarUnverifiedMLA theoretically reduces the GPU memory footprint of KV Cache to 5% to 13% of standard MHA by compressing KV matrices into low-dimensional latent vector caches.63% similarUnverifiedMLA（Multi-head Latent Attention）通过低秩压缩将 KV 缓存的显存占用降低至传统 MHA 的约 1/863% similarUnverifiedLlama 2开放了7B到70B参数规模的模型权重62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/827283API
curl https://kongchang.com/api/v1/knowledge/claims/827283MCP
get_claim(id=827283)