Unverified50% confidenceFactExact time
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
1
Sources
50%
Confidence
Long-term
Relevance
7/12/2026
First Seen
Sources
LingBot-Video 1.3B Selective FP8 Quantization Benchmark: 22% Faster Sampling
redditr/comfyui7/11/2026
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INT881% similarUnverifiedThe DiT architecture replaces the convolutional and attention layers in U-Net with Transformer blocks, and generation quality improves more significantly as the number of model parameters increases.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/494670API
curl https://kongchang.com/api/v1/knowledge/claims/494670MCP
get_claim(id=494670)