Unverified50% confidenceSolutionExact time
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INT8
1
Sources
50%
Confidence
Long-term
Relevance
7/11/2026
First Seen
Sources
Krea2 Tested: Does the FP8 vs BF16 Quality Gap Still Exist?
redditr/StableDiffusion7/9/2026
Related Claims
Unverified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INT883% similarUnverifiedIn the standard attention mechanism, the Query matrix and Key matrix perform matrix multiplication to obtain an n×n attention score matrix, with a computational complexity of O(n²·d).67% similarUnverified选择性SSM本质上等价于一种特殊的半可分矩阵（semiseparable matrix）形式的注意力66% 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66% similarUnverifiedSage Attention reduces bandwidth requirements and computational load by quantizing the Key matrix to INT8 and the Value matrix to FP8.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/475414API
curl https://kongchang.com/api/v1/knowledge/claims/475414MCP
get_claim(id=475414)