Rank-Aware Spectral Bounds on Attention Logits for Stable Low-Precision Training
This addresses overflow issues in low-precision training for large transformer models, offering a principled solution that is incremental but improves stability without sacrificing performance.
The paper tackles the problem of overflow risk in low-precision training of transformers by deriving a rank-aware concentration inequality for attention scores, which yields 8-28x tighter bounds than previous methods, and applies this to develop geometry-aware scale factors that eliminate overflows in models like GPT-2 XL to Llama-2-70B while maintaining comparable accuracy.
Attention scores in transformers are bilinear forms $S_{ij} = x_i^\top M x_j / \sqrt{d_h}$ whose maximum magnitude governs overflow risk in low-precision training. We derive a \emph{rank-aware concentration inequality}: when the interaction matrix $M = W^Q W^{K\top}$ has rank $r \ll d$, tail probabilities for $\max_{i,j}|S_{ij}|$ decay as $\exp(-d^{2}α^{2}/(γr))$ rather than $\exp(-dα^{2})$, where $γ> 1$ is a typicality parameter. For transformer attention where $r = d_h$, this yields $8$--$28\times$ tighter concentration than rank-agnostic bounds in modern architectures. We apply this result to FP8 training, deriving \emph{geometry-aware scale factors} that provide principled overflow guarantees without observing activations. The method computes per-layer scales from the spectral norm $\|W^Q W^{K\top}\|_2$ via implicit power iteration, includes a grouped query attention formulation that avoids key expansion, and remains compatible with fused attention kernels. Across GPT-2 XL to Llama-2-70B, geometry-aware scaling eliminates overflows in transient scenarios where delayed scaling fails, while achieving comparable downstream MMLU accuracy.