Transformer Inference

Transformers with self-attention modules as their core components have become an integral architecture in modern large language and foundation models. In this paper, we study the evolution of tokens in deep encoder-only transformers at inference time which is described in the large-token limit by a mean-field continuity equation. Leveraging ideas from the convergence analysis of interacting multi-particle systems, with particles corresponding to tokens, we prove that the token distribution rapidly concentrates onto the push-forward of the initial distribution under a projection map induced by the key, query, and value matrices, and remains metastable for moderate times. Specifically, we show that the Wasserstein distance of the two distributions scales like $\sqrt{{\log(β+1)}/β}\exp(Ct)+\exp(-ct)$ in terms of the temperature parameter $β^{-1}\to 0$ and inference time $t\geq 0$. For the proof, we establish Lyapunov-type estimates for the zero-temperature equation, identify its limit as $t\to\infty$, and employ a stability estimate in Wasserstein space together with a quantitative Laplace principle to couple the two equations. Our result implies that for time scales of order $\logβ$ the token distribution concentrates at the identified limiting distribution. Numerical experiments confirm this and, beyond that, complement our theory by showing that for finite $β$ and large $t$ the dynamics enter a different terminal phase, dominated by the spectrum of the value matrix.

We present Key-Value Means ("KVM"), a novel block-recurrence for attention that can accommodate either fixed-size or growing state. Equipping a strong transformer baseline with fixed-size KVM attention layers yields a strong $O(N)$ chunked RNN, while adding only an insignificant number of new parameters. We train a transformer with a growable KVM cache and show it performs competitively on long-context tests with only subquadratic prefill time and sublinear state growth. KVM is implementable with standard operations and without custom kernels, and supports chunk-wise parallelizable training and prefill. It provides many of the benefits of both traditional transformers (expandable context memory, chunk-wise parallelizable training and prefill) and linear RNNs in a single unified package. It can be used on every layer, saving KV-cache memory, and allowing a continuous range of choices of prefill time complexity between $O(N)$ and $O(N^2)$. It can also be implemented in a hybrid solution in tandem with LRNN layers in place of traditional attention, to supplement the LRNN with improved sublinear memory growth context length usage and long context decoding. We release our code at https://github.com/featherless-ai/KVM-paper and trained models at https://huggingface.co/collections/featherless-ai/kvm-paper under the Apache 2.0 license.

For Transformer models, cryptographically secure inference ensures that the client learns only the final output, while the server learns nothing about the client's input. However, securely computing nonlinear layers remains a major efficiency bottleneck due to the substantial communication rounds and data transmission required. To address this issue, prior works reveal intermediate activations to the client, allowing nonlinear operations to be computed in plaintext. Although this approach significantly improves efficiency, exposing activations enables adversaries to extract model weights. To mitigate this risk, existing works employ a shuffling defense that reveals only randomly permuted activations to the client. In this work, we show that the shuffling defense is not as robust as previously claimed. We propose an attack that aligns differently shuffled activations to a common permutation and subsequently exploits them to extract model weights. Experiments on Pythia-70m and GPT-2 demonstrate that the proposed attack can align shuffled activations with mean squared errors ranging from $10^{-9}$ to $10^{-6}$. With a query cost of approximately $1, the adversary can recover model weights with L1-norm differences ranging from $10^{-4}$ to $10^{-2}$ compared to the oracle weights.

A method is presented for accelerating inference in transformer language models by exploiting the low effective rank of the token activation manifold at each layer. The method decomposes each activation vector into a subspace component and a residual, computes the linear-layer output on the subspace component via a cached low-rank weight image at reduced memory bandwidth, and applies a per-token gate that determines whether the residual correction is computed or skipped. The gate ensures that the output distribution is preserved to within a controllable tolerance. Validation on three model families (GPT-2 124M, GPT-J 6B, OPT 6.7B) on AMD MI300X demonstrates effective speedups of 3.0x to 10.5x on linear-layer weight reads with perplexity ratios below 1.00 and top-1 token agreement above 98%. The method requires no retraining, no architectural modification, and no approximation of the attention mechanism. At the operating point (k = 256, ε = 0.05) on GPT-J 14 6B, the accelerated model produces character-for-character identical output to the baseline.

We present tensor and sequence parallelism (TSP), a parallel execution strategy that folds tensor parallelism and sequence parallelism onto a single device axis. In conventional multi-dimensional parallelism layouts, tensor parallelism (TP) shards model weights while sequence parallelism (SP) shards tokens, reducing per-device parameter or activation memory, respectively. Traditionally, each scheme is assigned its own mesh dimension. TSP instead assigns each rank both a weight shard and a sequence shard, reducing both parameter and activation memory along the same device axis. We implement this design with two runtime schedules. For attention, ranks iterate over broadcast parameter shards and reconstruct context through a sequence-wise key/value exchange. For gated MLPs, weight shards circulate in a ring while partial outputs accumulate locally. By sharding both weights and activations across the same devices, TSP trades additional communication volume for reduced memory overhead. We provide a theoretical communication and memory analysis, describe our implementation of TSP attention and gated MLP blocks, and benchmark TSP against TP, SP, and TP+SP. These results position TSP as a hardware-aware alternative for long-context and memory-constrained model training, and as a viable axis of parallelism in concert with existing parallelism schemes such as pipeline and expert parallelism for dense and mixture-of-expert models.

The key-value (KV) cache is the dominant memory bottleneck during Transformer inference, yet little is known theoretically about how aggressively it can be compressed before multi-step reasoning degrades. We study this through $k$-hop pointer chasing on $n$ tokens under a shared KV cache of size $s$, attention dimension $m$, $H$ heads, $p$-bit precision, and a locality-respecting cache controller (satisfied by all standard KV-compression methods). We give three results. (1) Product depth lower bound (conjectured). We conjecture that any such Transformer ($n \geq 4k$, $s \leq \sqrt{n}/4$) requires depth $L = Ω(\lceil k/s \rceil \cdot \lceil \log_2 n/(Hmp) \rceil)$, and isolate the sole remaining gap as a probabilistic step on the joint distribution of cache trace and pointer chain. Unconditionally, we prove a matching upper bound $L = O(\min(k, \lceil k/s \rceil \log s) \cdot \log n/(mp))$ via windowed pointer doubling, and a max-bound $L = Ω(\max(\lceil k/s \rceil, \log n/(Hmp)))$. Closing the conjecture amounts to upgrading max to product. (2) Bandwidth barrier. The product bound binds only when $Hmp \lesssim \log n$. Any lower bound provable via per-window distinguishability counting -- including reachability, bandwidth, and combinations -- cannot exceed $\lceil k/s \rceil$ once $Hmp \geq \log_2 n$. Breaking this requires lifting unconditional communication-complexity bounds for pointer chasing to Cache-Transformer depth. (3) Adaptive vs oblivious error scaling. Under random cache over $T = \lceil \log_2 k \rceil$ doubling stages, oblivious caches give $\Pr[\mathcal{E}] \leq (s/(n-T))^T + 2T^3/n$ (exponential in $T$), while adaptive locality-respecting caches achieve $\Pr[\mathcal{E}] = s/n$ exactly, independent of $T$. The $Ω((n/s)^{T-1})$ separation explains why heavy-hitter eviction empirically dominates random eviction for multi-hop reasoning.

Gravitational-wave data analysis relies on accurate and efficient methods to extract physical information from noisy detector signals, yet the increasing rate and complexity of observations represent a growing challenge. Deep learning provides a powerful alternative to traditional inference, but existing neural models typically lack the flexibility to handle variations in data analysis settings. Such variations accommodate imperfect observations or are required for specialized tests, and could include changes in detector configurations, overall frequency ranges, or localized cuts. We introduce a flexible transformer-based architecture paired with a training strategy that enables adaptation to diverse analysis settings at inference time. Applied to parameter estimation, we demonstrate that a single flexible model, called Dingo-T1, can (i) analyze 48 gravitational-wave events from the third LIGO-Virgo-KAGRA Observing Run under a wide range of analysis configurations, (ii) enable systematic studies of how detector and frequency configurations impact inferred posteriors, and (iii) perform inspiral-merger-ringdown consistency tests probing general relativity. Dingo-T1 also improves median sample efficiency on real events from a baseline of 1.4% to 4.2%. Our approach thus demonstrates flexible and scalable inference with a principled framework for handling missing or incomplete data, key capabilities for current and next-generation observatories.

Sparse attention offers a promising strategy to extend long-context capabilities in Transformer LLMs, yet its efficiency-accuracy trade-offs remain unclear due to the lack of comprehensive evaluation. We address this gap with the largest-scale empirical analysis to date of training-free sparse attention, evaluating six methods across multiple model families and sizes, sequences up to 128K tokens, and sparsity levels up to 0.95 (i.e., $1/20$ attention budget) on nine diverse tasks. We first organise the rapidly evolving landscape of sparse attention methods into a taxonomy along four design axes. Our analysis then yields actionable insights: 1) sparse attention is effective: larger sparse models outperform smaller dense ones at equivalent cost, improving the Pareto frontier; 2) for the training-free methods we study, fine-grained per-query importance estimation during prefilling remains impractical-due to both the cost of estimation and the lack of sparse kernels that translate fine-grained sparsity into wall-clock gains-forcing a task-dependent choice between global-to-token and block-to-block selection. Instead, during decoding, token-to-page selection becomes feasible, enabling better generalisation and higher sparsity tolerance; 3) longer sequences tolerate higher sparsity, suggesting that fixed-budget methods in production are suboptimal. Together, these findings provide practical guidance for deploying sparse attention and methodological recommendations for future evaluations. Our code is available at https://github.com/PiotrNawrot/sparse-frontier.

Transformers are the state-of-the-art architecture for large language models, and a key to their scalability is the strategic usage of low-precision arithmetic. We develop a mixed-precision analysis of transformer inference, deriving bounds for the condition numbers and forward error of the architecture's constituent parts. Notably, we compare the numerical stability of LayerNorm and RMSNorm in the massive-outlier regime, tighten the error bound of softmax in the presence of attention sinks, and quantify the impact of its shifted evaluation on the sensitivity to perturbations. Furthermore, we derive novel sequence-length-independent bounds on the local Lipschitz constant of self-attention. Our worst-case error bound for transformer inference suggests that its numerical stability is determined by the interplay between weight magnitude and the growth of the residual stream. Crucially, and as validated by experiments with GPT-2, our analysis establishes that the scaling of residual-projection weights preserves the propagation of the relative rounding error unless it forces a qualitative transition in the dynamics of the residual stream.