背景
Rotary Position Embedding (RoPE) typically uses uniform coverage across Transformer layers despite heterogeneous positional requirements.
干预观察
Phase-off interventions reveal non-uniform, model-specific rotary-phase dependence, while coverage interventions relate it to coverage-upgrade benefits.
方法贡献
We propose profiling-guided layer-wise RoPE allocation, assigning high/low coverage under a fixed budget using early coverage-benefit profiles.
框架能力
The framework supports fixed-width query/key representations and physical rotary-dimension allocation in Multi-head Latent Attention (MLA).
映射与迁移
Normalized-depth mapping transfers profiles within model families, enabling small-model measurements to guide larger-model allocations before pretraining.
分析
A brief analysis relates selection and transfer to score calibration and inter-layer interactions.
实验结果
From-scratch experiments demonstrate validation-loss improvements over available budget-matched baselines, while analytical resource accounting identifies favorable quality–storage–compute trade-offs for physical MLA.
结论
These results connect positional demand to transferable, efficient architecture design.