Unverified50% confidenceFactExact time

52层中有23层是Mamba、只有6层是注意力，Mamba基于状态空间模型的递归特性天然具有从左到右的方向性偏置

1

Sources

50%

Confidence

Long-term

Relevance

7/20/2026

First Seen

Sources

NVIDIA's Dual-Tower Diffusion Language Model Explained: How Freezing Half the Network Achieves 2.4× Speedup

Related Claims

Unverified针对长序列的架构改进方案包括线性注意力近似的Performer、基于状态空间模型的Mamba、以及结合滑动窗口注意力的Longformer60% similar UnverifiedLoRA的秩r通常取8至64之间，target_modules通常注入注意力层的Q、K、V、O矩阵56% similar Unverified区域提示技术在多个角色LoRA同时加载时，因不同LoRA的低秩矩阵方向冲突同时修改同一批注意力层权重，导致角色面部融合或特征错位55% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/565270

API

curl https://kongchang.com/api/v1/knowledge/claims/565270

MCP

get_claim(id=565270)