Unverified50% confidenceFactExact time
模型中注意力层的QKV投影权重和输出投影通常比MLP层的中间权重更敏感，模型首尾层比中间层更脆弱
1
Sources
50%
Confidence
Long-term
Relevance
8/24/2026
First Seen
Sources
Related Entities
Related Claims
Unverified通过Hessian矩阵或Fisher信息矩阵进行逐层敏感性分析可识别模型中的量化瓶颈层，注意力层Q/K投影通常为高敏感层，FFN层大部分权重可安全压缩至FP8甚至INT873% similarUnverified对于图像生成Transformer骨干，注意力层的Q/K投影矩阵通常被识别为高敏感层，而FFN层大部分权重可安全压缩至FP8甚至INT870% similarUnverified研究表明Transformer架构中注意力层（Q/K/V投影和输出投影）对精度损失更敏感，而FFN线性层对低精度表示鲁棒性更强，嵌入层和LayerNorm层通常保留高精度70% similarUnverifiedThe Qwen series introduces the GQA (Grouped Query Attention) mechanism at the architectural level, reducing KV cache memory consumption compared to standard multi-head attention.68% similarUnverifiedThe Multi-Query Attention technique allows multiple attention heads to share a single set of Key and Value projection matrices, compressing KV cache memory requirements to 1/H of the original, at the cost of a slight reduction in model expressiveness.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/795636API
curl https://kongchang.com/api/v1/knowledge/claims/795636MCP
get_claim(id=795636)