Gradient Boosting within a Single Attention Layer

BF16 is now standard in large-scale pretraining, including in fused attention kernels such as FlashAttention, and these kernels are widely trusted. When we used FlashAttention-3 to pretrain a 450M-parameter transformer on 50B tokens, however, we ran into a problem: training was healthy for 25B tokens, then the gradient norm grew a thousandfold and the loss ended 0.2 nats above FP32 attention, without a single NaN. Recomputing the attention backward of just two layers in FP32 removes almost all of the excess gradient. Part of the cause is known: a fused multiply-add in the forward softmax, so far treated as an extreme-input NaN case and never fixed in FlashAttention-3. Repairing it stops the blow-up, but the query gradient is still wrong by more than its own size, and training still drives attention logits to thousands of times their size under accurate gradients. The remaining error comes from a broken conservation law. The softmax score gradient sums to zero along every row, which makes the query gradient blind to where the keys sit as a group; rounding it to BF16 leaves a small nonzero sum that leaks the mean key into the gradient, and the leak grows exactly as late training makes keys large and attention sharp. We introduce GProj (gauge projection), which restores the zero sum after the cast with two rank-one corrections per row. It cuts the remaining median query/key gradient errors from 219%/13% to 0.34%/0.37%, on par with FP32 attention, for 4.7% more time per training step. In matched from-scratch runs it trains to the same loss as FP32 attention, while FlashAttention-3 and key smoothing both destabilize.

Recent algebraic analysis shows that in decoder-only and encoder-only transformers, the Query projection $W_Q$ may be set to identity without noticeable performance deterioration. This is possible because attention depends on $X$ only through the products $XW_Q, XW_K, XW_V$, allowing basis transformations to be absorbed by adjacent layers and propagated through the network. We replace $W_Q \in \R^{d \times d}$ with a nonlinear residual of the form $Q(X) = X + f_θ(X)$, where $f_θ$ is a bottleneck MLP with $d^2 + O(d)$ parameters. The identity term anchors the nonlinearity to a known-good prior. Experiments on GPT-3 small style models show consistent improvement over the baseline ($2.40\%$ lower validation log-loss, $6.81\%$ lower perplexity), comfortably outperforming a model with 12.5% more non-embedding parameters. These results motivate investigation at larger scales and across modalities.

We derive an exact gradient-step representation of the RoPE-softmax forward pass. For every deterministic RoPE-softmax attention head with arbitrary affine projection weights, we construct a query-dependent effective matrix $ΔM_i$ satisfying $y_i = μ_i + u_i^\top ΔM_i$, where $μ_i$ is the uniform mean of the attended values and $u_i$ is the augmented query input. The construction applies the classical exponential divided difference $ρ= φ_1$ to retain the softmax exactly. Its positive coefficients give a unit gradient-step representation on a query-conditioned quadratic objective. The same function connects the RoPE generator to exact positional finite differences. We derive a tokenwise formula for the error of reusing one query's matrix and prove that a nonconstant finite-cache head cannot admit a globally exact affine query readout. Reconstruction checks and frozen-reuse calibration on one pretrained Qwen2.5-0.5B layer verify the representation and quantify the correction required when one query's matrix is reused.