Efficient Transformer Inference

Decision Transformer (DT) formulates offline reinforcement learning as autoregressive sequence modeling, achieving promising results by predicting actions from a sequence of Return-to-Go (RTG), state, and action tokens. However, RTG is a scalar that summarizes future rewards, containing far less information than typical state or action vectors, yet it consumes the same computational budget per token. Worse, the self-attention cost of Transformers grows quadratically with sequence length, so including RTG as a separate token adds unnecessary overhead. We propose SlimDT, which removes RTG from the autoregressive sequence. Instead, we inject RTG information into the state representations before the sequential modeling step, allowing the Transformer to process only a compact (state, action) sequence. This reduces the sequence length by one-third, directly improving inference efficiency. On the D4RL benchmark, SlimDT surpasses standard DT across various tasks and achieves performance comparable to existing state-of-the-art methods. Decoupling a sparse conditioning signal from an information-rich sequence thus yields both computational gains and higher task performance.

Transformers usually expose one inference cost per trained model, while deployed systems often need multiple cost-quality operating points. We study Budgeted Attention Allocation, a monotone head-gating mechanism conditioned on a requested attention budget. Dense warm-starting is important for stability: on a robust synthetic sequence task, one budgeted model reaches 99.7% accuracy at 0.303 estimated attention cost and 100.0% accuracy at 0.504 cost. On held-out AG News with a custom word-level transformer, hard-gate adaptation turns soft cost control into measured single-thread CPU speed, reaching 82.1% accuracy with 1.28x speedup at budget 0.50. In pretrained BERT-Mini AG News, budgeted structural pruning reaches 87.6% accuracy with 1.20x speedup at budget 0.50; a validation-ranked zero-shot dense post-hoc structural baseline reaches 86.1%, and one recovery epoch raises that per-budget specialist to 87.9%. On DBpedia14, BERT-Mini budgeted gates reach 97.4% at exact budget 0.50 versus 96.6% for dense full attention. Static fixed-budget gates and recovered dense specialists remain strong. The contribution is therefore not universal dominance, but a reproducible feasibility study of one controllable checkpoint across budgets that can trade attention cost for accuracy and be converted into measured structural speedups on small CPU benchmarks.

While linear-complexity attention mechanisms offer a promising alternative to Softmax attention for overcoming the quadratic bottleneck, training such models from scratch remains prohibitively expensive. Inheriting weights from pretrained Transformers provides an appealing shortcut, yet the fundamental representational gap between Softmax and linear attention prevents effective weight transfer. In this work, we address this conversion challenge from two perspectives: architectural alignment and representational alignment. We identify Test-Time Training (TTT) as a linear-complexity architecture whose two-layer dynamic formulation is structurally aligned with Softmax attention, enabling direct inheritance of pretrained attention weights. To further align representational properties, including key shift-invariance and locality, we introduce key instance normalization and a lightweight locality enhancement module. We validate our approach by linearizing Stable Diffusion 3.5 and introduce SD3.5-T$^5$ (Transformer To Test Time Training). With only 1 hour of fine-tuning on 4$\times$H20 GPUs, SD3.5-T$^5$ achieves comparable text-to-image quality to the fine-tuned Softmax model, while accelerating inference by 1.32$\times$ and 1.47$\times$ at 1K and 2K resolutions. Code is available at https://github.com/LeapLabTHU/Transformer-to-TTT.

Layer-aligned distillation and convergence-based early exit represent two predominant computational efficiency paradigms for transformer inference; yet we establish that they exhibit systematic incompatibility under standard deployment conditions for convergence-based early exit. Distillation objectives that align intermediate student layers to teacher representations suppress the representational convergence that early-exit mechanisms exploit, rendering such mechanisms ineffective on distilled models. We introduce LEAP (Layer-wise Exit-Aware Pretraining), an auxiliary training objective that reconciles this incompatibility. LEAP requires no architectural modifications; it augments standard distillation with a single constraint ensuring intermediate layers approximate final-layer representations. LEAP-MiniLM achieves 1.61$\times$ measured wall-clock speedup (batch=1, NVIDIA L4) at $θ$=0.95, with 91.9% of samples exiting by layer 7 and 1.80$\times$ theoretical layer reduction, where standard distilled models achieve zero effective speedup. We validate across sentence similarity (STS-B: 0.760 $\pm$ 0.006) and retrieval benchmarks (BEIR), providing operational guidance including latency measurements, decision thresholds, and deployment criteria.

Existing attention accelerators often trade exact softmax semantics, depend on fused Tensor Core kernels, or incur sequential depth that limits FP32 throughput on long sequences. We present \textbf{ELSA}, an algorithmic reformulation of online softmax attention that (i)~preserves exact softmax semantics in real arithmetic with a \emph{provable} $\mathcal{O}(u\log n)$ FP32 relative error bound; (ii)~casts the online softmax update as a prefix scan over an associative monoid $(m,S,W)$, yielding $O(n)$ extra memory and $O(\log n)$ parallel depth; and (iii)~is Tensor-Core independent, implemented in Triton and CUDA C++, and deployable as a \emph{drop-in replacement} requiring no retraining or weight modification. Unlike FlashAttention-2/3, which rely on HMMA/GMMA Tensor Core instructions and provide no compatible FP32 path, ELSA operates identically on A100s and resource-constrained edge devices such as Jetson TX2 -- making it the only hardware-agnostic exact-attention kernel that reduces parallel depth to $O(\log n)$ at full precision. On A100 FP32 benchmarks (1K--16K tokens), ELSA delivers $1.3$--$3.5\times$ speedup over memory-efficient SDPA and $1.97$--$2.27\times$ on BERT; on Jetson TX2, ELSA achieves $1.5$--$1.6\times$ over Math (64--900 tokens), with $17.8$--$20.2\%$ throughput gains under LLaMA-13B offloading at $\ge$32K. In FP16, ELSA approaches hardware-fused baselines at long sequences while retaining full FP32 capability, offering a unified kernel for high-precision inference across platforms. Our code and implementation are available at https://github.com/ming053l/ELSA.

NVIDIA's CUDA Tile (CuTile) introduces a Python-based, tile-centric abstraction for GPU kernel development that aims to simplify programming while retaining Tensor Core and Tensor Memory Accelerator (TMA) efficiency on modern GPUs. We present the first independent, cross-architecture evaluation of CuTile against established approaches such as cuBLAS, Triton, WMMA, and raw SIMT on three NVIDIA GPUs spanning Hopper and Blackwell: H100 NVL, B200, and RTX PRO 6000 Blackwell Server Edition. We benchmark representative AI workloads, including GEMM, fused multi-head attention, and end-to-end LLM inference in BF16/FP16 precision, to assess both performance and portability. Our results show that CuTile effectiveness is strongly workload- and architecture-dependent. On datacenter-class Blackwell (B200), CuTile achieves up to 1007 TFLOP/s for fused attention, outperforming FlashAttention-2 by 2.5x while requiring only 60 lines of Python kernel code. For GEMM, CuTile reaches 52-79% of cuBLAS performance in 22 lines of code (versus 123 for WMMA), making it a practical replacement for hand-written CUDA kernels but not yet for vendor-optimized libraries. However, the same CuTile attention kernel achieves only 53% of FlashAttention-2 throughput on RTX PRO 6000 (sm_120), exposing significant cross-architecture optimization gaps. In contrast, Triton sustains 62-101% of cuBLAS performance across all tested platforms without architecture-specific tuning, demonstrating substantially stronger portability.

Transformers process tokens in parallel but are temporally shallow: at position $t$, each layer attends to key-value pairs computed based on the previous layer, yielding a depth capped by the number of layers. Recurrent models offer unbounded temporal depth but suffer from optimization instability and historically underutilize modern accelerators. We introduce the Recurrent Transformer, a simple architectural change where each layer attends to key-value pairs computed off its own activations, yielding layerwise recurrent memory while preserving standard autoregressive decoding cost. We show that the architecture can emulate both (i) a conventional Transformer and (ii) token-to-token recurrent updates under mild assumptions, while avoiding optimization instability. Naively, prefill/training appears bandwidth-bound with effective arithmetic intensity near $1$ because keys and values are revealed sequentially; we give an exact tiling-based algorithm that preserves the mathematical computation while reducing HBM traffic from $Θ(N^2)$ to $Θ(N\log N)$, increasing effective arithmetic intensity to $Θ(N/\log N)$ for sequence length $N$. On 150M and 300M parameter C4 pretraining, Recurrent Transformers improve cross-entropy over a parameter-matched Transformer baseline and achieve the improvement with fewer layers (fixed parameters), suggesting that recurrence can trade depth for width, thus reducing KV cache memory footprint and inference latency.

Named Entity Recognition (NER) is a key component in industrial information extraction pipelines, where systems must satisfy strict latency and throughput constraints in addition to strong accuracy. State-of-the-art NER accuracy is often achieved by span-based frameworks, which construct span representations from token encodings and classify candidate spans. However, many span-based methods enumerate large numbers of candidates and process each candidate with marker-augmented inputs, substantially increasing inference cost and limiting scalability in large-scale deployments. In this work, we propose SpanDec, an efficient span-based NER framework that targets this bottleneck. Our main insight is that span representation interactions can be computed effectively at the final transformer stage, avoiding redundant computation in earlier layers via a lightweight decoder dedicated to span representations. We further introduce a span filtering mechanism during enumeration to prune unlikely candidates before expensive processing. Across multiple benchmarks, SpanDec matches competitive span-based baselines while improving throughput and reducing computational cost, yielding a better accuracy-efficiency trade-off suitable for high-volume serving and on-device applications.

Face Image Quality Assessment is crucial for reliable face recognition systems, yet existing Vision Transformer-based approaches rely exclusively on final-layer representations, ignoring quality-relevant information captured at intermediate network depths. This paper presents the first comprehensive investigation of how intermediate representations within ViTs contribute to face quality assessment through early exit mechanisms and score fusion strategies. We systematically analyze all twelve transformer blocks of ViT-FIQA architectures, demonstrating that different depths capture distinct and complementary quality-relevant information, as evidenced by varying attention patterns and performance characteristics across network layers. We propose a score fusion framework that combines quality predictions from multiple transformer blocks without architectural modifications or additional training. Our early exit analysis reveals optimal performance-efficiency trade-offs, enabling significant computational savings while maintaining competitive performance. Through extensive evaluation across eight benchmark datasets using four FR models, we demonstrate that our fusion strategy improves upon single-exit approaches. Our proposed quality fusion approach employs depth-weighted averaging that assigns progressively higher importance to deeper transformer blocks, achieving the best quality assessment performance by effectively leveraging the hierarchical nature of feature learning in ViTs. Our work challenges the conventional wisdom that only deep features matter for face analysis, revealing that intermediate representations contain valuable information for quality assessment. The proposed framework offers practical benefits for real-world biometric systems by enabling adaptive computation based on resource constraints while maintaining competitive quality assessment capabilities.

Visual Autoregressive (VAR) modeling inefficiently applies a fixed computational depth to each position when generating high-resolution images. While existing methods accelerate inference by pruning tokens using frequency maps, their binary hard-pruning approach is fundamentally limited and fails to improve quality even with better frequency estimation. Observing that VAR models possess significant depth redundancy, we propose a paradigm shift from pruning entire tokens to adaptively allocating per-token computational depth. To this end, we introduce DepthVAR, a training-free framework that dynamically allocates computation. It integrates an adaptive depth scheduler, which assigns computational depth via a cyclic rotated schedule for balanced, non-static refinement, with a dynamic inference process that translates these depths into layer-major masks, selectively applies transformer blocks, and blends the resulting codes to ensure each token's influence is proportional to its processing depth. Extensive experiments show that DepthVAR achieves 2.3$\times$-3.1$\times$ acceleration with minimal quality loss, offering a competitive compute-performance trade-off compared to existing hard-pruning approaches. Code is available at https://github.com/STOVAGtz/DepthVAR

Standard transformer architectures apply the same number of layers to every token regardless of contextual difficulty. We present Token-Selective Attention (TSA), a learned per-token gate on residual updates between consecutive transformer blocks. Each gate is a lightweight two-layer multi-layer perceptron (MLP) that produces a continuous halting probability, making the mechanism end-to-end differentiable with 1.7% parameter overhead and no changes to the base architecture. Notably, TSA learns difficulty-proportional routing without any explicit depth pressure: even at $λ=0$ (no depth regularisation), the task-loss gradient alone drives the router to skip 20% of token-layer operations. On character-level language modeling, TSA saved 14-23% of token-layer operations (TLOps) across Tiny-Shakespeare and enwik8 at <0.5% quality loss. At matched efficiency, TSA achieved 0.7% lower validation loss than early exit, and the learned routing transfers directly to inference-time sparse execution for real wall-clock speedup.

Cryptographically secure neural network inference typically relies on secure computing techniques such as Secure Multi-Party Computation (MPC), enabling cloud servers to process client inputs without decrypting them. Although prior privacy-preserving inference systems co-design network optimizations with MPC, they remain slow and costly, limiting real-world deployment. A major bottleneck is their use of a single, fixed transformer model for all encrypted inputs, ignoring that different inputs require different model sizes to balance efficiency and accuracy. We present SecureRouter, an end-to-end encrypted routing and inference framework that accelerates secure transformer inference through input-adaptive model selection under encryption. SecureRouter establishes a unified encrypted pipeline that integrates a secure router with an MPC-optimized model pool, enabling coordinated routing, inference, and protocol execution while preserving full data and model confidentiality. The framework includes training-phase and inference-phase components: an MPC-cost-aware secure router that predicts per-model utility and cost from encrypted features, and an MPC-optimized model pool whose architectures and quantization schemes are co-trained to minimize MPC communication and computation overhead. Compared to prior work, SecureRouter achieves a latency reduction by 1.95x with negligible accuracy loss, offering a practical path toward scalable and efficient secure AI inference. Our open-source implementation is available at: https://github.com/UCF-ML-Research/SecureRouter

Large Language Model (LLM) deployment is increasingly shifting to cost-efficient accelerators like Google's Tensor Processing Units (TPUs), prioritizing both performance and total cost of ownership (TCO). However, existing LLM inference kernels and serving systems remain largely GPU-centric, and there is no well-established approach for efficiently mapping LLM workloads onto TPU architectures--particularly under the dynamic and ragged execution patterns common in modern serving. In this paper, we present Ragged Paged Attention (RPA), a high-performance and flexible attention kernel for TPUs, implemented using Pallas and Mosaic. RPA addresses these challenges through three key techniques: (1) fine-grained tiling to enable efficient dynamic slicing over ragged memory, (2) a custom software pipeline that fuses KV cache updates with attention computation, and (3) a distribution-aware compilation strategy that generates specialized kernels for decode, prefill, and mixed workloads. Evaluated on Llama 3 8B on TPU7x, RPA achieves up to 86% memory bandwidth utilization (MBU) in decode and 73% model FLOPs utilization (MFU) in prefill. Integrated as the primary TPU backend in vLLM and SGLang, RPA provides a production-grade foundation for efficient TPU inference and offers practical insights into kernel design.

Token pruning methods for Vision Transformers (ViTs) promise quadratic reductions in attention FLOPs by dropping uninformative patches. Yet standard variable-length attention APIs -- including FlashAttention-2's varlen and PyTorch's NestedTensor SDPA -- fail to translate these savings into proportional wall-clock gains at the short post-pruning sequence lengths typical of ViTs ($\leq$197 tokens). We identify a dispatch-overhead bottleneck: at these lengths, host-side kernel dispatch consumes ${\sim}$50,$μ$s regardless of workload, exceeding the actual GPU compute time at moderate-to-high pruning rates. We present a lightweight bidirectional Triton attention kernel whose dispatch floor is ${\sim}$24,$μ$s -- roughly 2.17$\times$ lower than FlashAttention-2 varlen -- allowing pruning savings to become visible in wall-clock time. Integrated into a complete pack-attend-unpack pipeline and evaluated on an NVIDIA RTX 4000 Ada Generation GPU, our system achieves 1.88$\times$ end-to-end throughput over padded PyTorch SDPA at standard 224$\times$224 inputs, scaling to 2.51$\times$ at 384$\times$384. Against FlashAttention-2 varlen -- the strongest baseline -- our kernel delivers 9-12% higher throughput at serving batch sizes (BS=1-4), and 2.17$\times$ lower kernel latency at 80% token pruning. Numerical correctness is verified with max absolute logit difference $<$0.004 and bit-exact top-1 predictions.

Standard Transformers have a fixed computational depth, limiting their ability to generalize to tasks that require variable-depth reasoning. The usual remedy, Chain-of-Thought (CoT), spends tokens to reason, inflating the key--value cache and making latency grow with the step count, so memory becomes the limiting cost when reasoning is served over large query batches. We study a depth-recurrent Transformer that decouples computational depth from parameter count by iterating a shared-weight block, so that each added reasoning step costs flat memory and linear latency, with no token generation. Three ingredients keep the recurrence stable for 20+ thinking steps: a silent thinking objective that supervises only the final output, LayerScale initialization, and an identity-biased gate that opens a gradient highway across steps. We characterize it on three compositional domains with decreasing structural bias: graph reachability (adjacency masking), nested boolean logic (relative positioning), and unstructured relational text (no positional cue). We find a \emph{computational frontier}: accuracy climbs once the thinking-step count meets the task's complexity, reaching near-perfect performance on the two structured tasks and a lower plateau on unstructured text. How it climbs depends on the structural bias---abruptly from chance on the graph task, gradually on the other two. Depth recurrence extrapolates beyond the training range: it succeeds on the graph task where fixed-depth models barely extrapolate, and on the two sequence tasks comes within two points of fixed-depth Transformers that use $4$--$6.4\times$ more parameters. On the graph task, whose adjacency mask makes propagation depth verifiable, intermediate per-step supervision---a standard recipe for deep iterative models---consistently \emph{harms} this extrapolation. We release the code for reproducibility.

The architectural shape of dense Transformers has remained remarkably stable: narrow-wide-narrow feed-forward networks (FFNs) consume most non-embedding parameters. Motivated by theoretical and empirical evidences that residual wide-narrow-wide (hourglass) MLPs remain expressive despite bottlenecks, we revisit whether this architectural convention is necessary for dense language models. We study Hourglass Transformers, which replace the conventional FFN with residual stacks of hourglass sub-MLPs and use hourglass attention to decouple residual-stream width from attention width. This exposes a practical depth-width trade-off: compressing the FFN intermediate dimension allows wider hidden states and fewer layers at matched parameter budgets. Across model scales from 113M to 8B parameters, Hourglass Transformers achieve language-modeling and downstream performance comparable to conventional Transformers, while improving training compute efficiency by $8.7\%$ at matched average downstream accuracy across the 906M, 3B, and 8B scales. After long-context extension, the 8B Hourglass model also outperforms its matched conventional baseline across 4k-64k context lengths. At 64k context, the reduced attention layer count lowers both computation and KV-cache requirements, yielding up to $1.93\times$ faster token decoding and $50\%$ lower KV-cache memory at the 1B scale. These results identify hourglass structures as a practical architecture-efficiency alternative for compute- and latency-conscious Transformer design.

Native 4K (2176$\times$3840) video generation remains a critical challenge due to the quadratic computational explosion of full-attention as spatiotemporal resolution increases, making it difficult for models to strike a balance between efficiency and quality. This paper proposes a novel Transformer retrofit strategy termed T3 ($\textbf{T}$ransform $\textbf{T}$rained $\textbf{T}$ransformer) that, without altering the core architecture of full-attention pretrained models, significantly reduces compute requirements by optimizing their forward logic. Specifically, $\textbf{T3-Video}$ introduces a multi-scale weight-sharing window attention mechanism and, via hierarchical blocking together with an axis-preserving full-attention design, can effect an "attention pattern" transformation of a pretrained model using only modest compute and data. Results on $\textbf{4K-VBench}$ show that $\textbf{T3-Video}$ substantially outperforms existing approaches: while delivering performance improvements (+4.29$\uparrow$ VQA and +0.08$\uparrow$ VTC), it accelerates native 4K video generation by more than 10$\times$. Project page at https://zhangzjn.github.io/projects/T3-Video

Existing methods for expressive music performance rendering, a conditional generation task that aims to generate a human-like performance from a symbolic score, rely on supervised learning over small labeled datasets, which limits scaling of both data volume and model size, despite the availability of vast unlabeled music, as in vision and language. To address this gap, we introduce Pianist Transformer, with three key contributions: 1) introducing large-scale self-supervised learning into expressive piano performance rendering through a unified Musical Instrument Digital Interface (MIDI) representation, enabling pre-training on 10B tokens of unlabeled MIDI data; 2) an efficient asymmetric Transformer with note-level compression, substantially improving training efficiency, memory usage, and inference speed for long-context music modeling; 3) a state-of-the-art rendering model with an editable workflow, achieving strong objective and subjective results and enabling integration into real-world music production workflows. Overall, Pianist Transformer outlines a scalable path toward human-like performance synthesis in the music domain. Code, audio samples, and model checkpoints are available on our project page: https://yhj137.github.io/pianist-transformer-demo/.

The substantial inference costs of attention in transformers motivated the development of efficient sequence mixers: namely sparse and sliding window attention, convolutions and linear attention. Although these approaches result in impressive reductions in inference costs, they often trade-off with quality, specifically in-context recall. Apriori fixing this quality-cost tradeoff at training time means being suboptimal from the get-go: some downstream applications might fundamentally require more memory for in-context recall, while other tasks may require lower latency and memory. We propose a conceptually simple meta-sequence mixer with inference-cost controllability: the Compress & Attend Transformer (CAT). CAT decodes chunks of tokens by attending to compressed chunks of the sequence so far. Both compression and decoding can use any existing sequence mixer. Decoding from the compressed sequence yields compute and memory savings, with chunk size setting the operating point on the quality-cost trade-off. Importantly, training CAT across multiple chunk sizes at once unlocks test-time control of this trade-off without any retraining, all in a single model. Instantiated with the most basic choice, dense attention as the mixer, CAT surprisingly suffices to match 10 popular and diverse efficient models (linear, hybrids, sparse) on real-world long-context recall at comparable inference costs, all from a single trained model. CAT further performs competitively on long-context understanding benchmarks while providing 1.4-3.7x higher generation throughput than a dense transformer. Code is at: https://github.com/rajesh-lab/cat-transformer