Fast Gauss Sums via Flash Attention

We present Flash-GMM, a fused Triton kernel for efficient computation of Gaussian Mixture Models (GMMs) over large-scale data in a single GPU pass. By eliminating the need to materialize the full responsibility matrix in GPU memory, Flash-GMM achieves a \textbf{20$\times$} speedup over existing implementations and enables training on datasets more than \textbf{100$\times$} larger than previously feasible on one device. To demonstrate its impact, we integrate Flash-GMM into the IVF coarse quantizer for approximate nearest-neighbor (ANN) search. We show that soft GMM clustering is now a viable drop-in replacement for $k$-means, and that GMM responsibilities can be leveraged to assign border vectors to multiple clusters. Our approach reaches fixed recall targets with up to $1.7\times$ fewer distance computations, or equivalently, yields $+2$--$12$ recall@10 at matched computational cost. We release the kernel as an open-source project.

Self-attention in Transformers is typically implemented as $\mathrm{softmax}(QK^\top/\sqrt{d})V$, where $Q=XW_Q$, $K=XW_K$, and $V=XW_V$ are learned linear projections of the input $X$. We ask whether these learned projections are necessary, or whether they can be replaced by a simpler similarity-based diffusion operator. We introduce Gaussian Kernel Attention (GKA), a drop-in replacement for dot-product attention that computes token affinities directly using a Gaussian radial basis function (RBF) kernel applied to per-head token features. Each head learns only a bandwidth parameter $σ_h$, while a single output projection $W_O$ preserves compatibility with the standard Transformer interface. GKA can be interpreted as normalized kernel regression over tokens, linking modern Transformer architectures to classical non-local filtering and kernel smoothing methods. We evaluate GKA in both vision and language modeling settings. For autoregressive language modeling within the nanochat framework, we implement causal masking and sliding-window constraints by masking and renormalizing the Gaussian kernel. At depth 20, a GKA model with $0.42\times$ the parameters and $0.49\times$ the total training FLOPs of a standard attention baseline trains stably, exhibits a near-zero train-validation gap, and demonstrates competitive behavior on standard benchmarks, albeit with higher bits-per-byte (BPB) at this compute scale. Overall, GKA provides a minimal, interpretable attention mechanism with an explicit locality scale, offering a dimension in the accuracy-efficiency trade-off for Transformer design.

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.