Daily Papers

Reasoning-focused LLMs improve answer quality by generating longer reasoning traces, but the additional tokens dramatically increase serving cost, motivating inference optimization. We extend and apply Puzzle, a post-training neural architecture search (NAS) framework, to gpt-oss-120B to produce gpt-oss-puzzle-88B, a deployment-optimized derivative. Our approach combines heterogeneous MoE expert pruning, selective replacement of full-context attention with window attention, FP8 KV-cache quantization with calibrated scales, and post-training reinforcement learning to recover accuracy, while maintaining low generation length. In terms of per-token speeds, on an 8XH100 node we achieve 1.63X and 1.22X throughput speedups in long-context and short-context settings, respectively. gpt-oss-puzzle-88B also delivers throughput speedups of 2.82X on a single NVIDIA H100 GPU. However, because token counts can change with reasoning effort and model variants, per-token throughput (tok/s) and latency (ms/token) do not necessarily lead to end-to-end speedups: a 2X throughput gain is erased if traces grow 2X. Conversely, throughput gains can be spent on more reasoning tokens to improve accuracy; we therefore advocate request-level efficiency metrics that normalize throughput by tokens generated and trace an accuracy--speed frontier across reasoning efforts. We show that gpt-oss-puzzle-88B improves over gpt-oss-120B along the entire frontier, delivering up to 1.29X higher request-level efficiency. Across various benchmarks, gpt-oss-puzzle-88B matches or slightly exceeds the parent on suite-average accuracy across reasoning efforts, with retention ranging from 100.8% (high) to 108.2% (low), showing that post-training architecture search can substantially reduce inference costs without sacrificing quality.

Diffusion models have emerged as a powerful class of generative models across various modalities, including image, video, and audio synthesis. However, their deployment is often limited by significant inference latency, primarily due to the inherently sequential nature of the denoising process. While existing parallelization strategies attempt to accelerate inference by distributing computation across multiple devices, they typically incur high communication overhead, hindering deployment on commercial hardware. To address this challenge, we propose ParaStep, a novel parallelization method based on a reuse-then-predict mechanism that parallelizes diffusion inference by exploiting similarity between adjacent denoising steps. Unlike prior approaches that rely on layer-wise or stage-wise communication, ParaStep employs lightweight, step-wise communication, substantially reducing overhead. ParaStep achieves end-to-end speedups of up to 3.88times on SVD, 2.43times on CogVideoX-2b, and 6.56times on AudioLDM2-large, while maintaining generation quality. These results highlight ParaStep as a scalable and communication-efficient solution for accelerating diffusion inference, particularly in bandwidth-constrained environments.

Autoregressive video diffusion models enable streaming generation, opening the door to long-form synthesis, video world models, and interactive neural game engines. However, their core attention layers become a major bottleneck at inference time: as generation progresses, the KV cache grows, causing both increasing latency and escalating GPU memory, which in turn restricts usable temporal context and harms long-range consistency. In this work, we study redundancy in autoregressive video diffusion and identify three persistent sources: near-duplicate cached keys across frames, slowly evolving (largely semantic) queries/keys that make many attention computations redundant, and cross-attention over long prompts where only a small subset of tokens matters per frame. Building on these observations, we propose a unified, training-free attention framework for autoregressive diffusion: TempCache compresses the KV cache via temporal correspondence to bound cache growth; AnnCA accelerates cross-attention by selecting frame-relevant prompt tokens using fast approximate nearest neighbor (ANN) matching; and AnnSA sparsifies self-attention by restricting each query to semantically matched keys, also using a lightweight ANN. Together, these modules reduce attention, compute, and memory and are compatible with existing autoregressive diffusion backbones and world models. Experiments demonstrate up to x5--x10 end-to-end speedups while preserving near-identical visual quality and, crucially, maintaining stable throughput and nearly constant peak GPU memory usage over long rollouts, where prior methods progressively slow down and suffer from increasing memory usage.

Mixture-of-Experts (MoE) model architecture has emerged as a promising solution for scaling transformer models efficiently, offering sparse activation that reduces computational costs while increasing model capacity. However, as MoE models scale, they need to be distributed across GPU devices, thus face critical performance bottlenecks due to their large memory footprint. Expert parallelism distributes experts across GPUs, however, faces key challenges including an unbalanced token routing and expert activation, resulting in communication tail latency and processing inefficiencies. While existing solutions address some of these issues, they fail to resolve the dual challenges of load imbalance and communication skew. The imbalance in token processing load across experts causes uneven processing times on different GPUs, while communication skew between GPUs leads to unbalanced inter-GPU data transfers. These factors degrade the performance of MoE models by increasing tail latency and reducing overall throughput. To address these limitations, we propose an Integer Linear Programming (ILP) formulation to optimize expert placement by jointly considering token load, communication, and computation costs. We exploit the property that there is a token routing dependency across layers, where tokens routed to a specific expert in one layer are likely to be routed to a limited set of experts in the subsequent layer. Our solution, MoETuner, offers an optimal expert-to-GPU assignment that minimizes inter-GPU token routing costs and balances token processing across devices, thereby reducing tail latency and end-to-end execution time. Experimental results demonstrate 9.3% and 17.5% of end-to-end speedups for single-node and multi-node inference respectively, showcasing the potential of our ILP-based optimization for offering expert parallel solutions for next-generation MoEs.

Large Language Models (LLMs) are increasingly being deployed on edge devices for long-context settings, creating a growing need for fast and efficient long-context inference. In these scenarios, the Key-Value (KV) cache is the primary bottleneck in terms of both GPU memory and latency, as the full KV cache must be loaded for each decoding step. While speculative decoding is a widely accepted technique to accelerate autoregressive decoding, existing methods often struggle to achieve significant speedups due to inefficient KV cache optimization strategies and result in low acceptance rates. To address these challenges, we propose a novel self-speculative decoding framework, QuantSpec, where the draft model shares the architecture of the target model but employs a hierarchical 4-bit quantized KV cache and 4-bit quantized weights for acceleration. QuantSpec maintains high acceptance rates (>90%) and reliably provides consistent end-to-end speedups upto sim2.5times, outperforming other self-speculative decoding methods that use sparse KV cache for long-context LLM inference. QuantSpec also reduces the memory requirements by sim 1.3times compared to these alternatives.

Diffusion-based world models have shown strong potential for unified world simulation, but the iterative denoising remains too costly for interactive use and long-horizon rollouts. While feature caching can accelerate inference without training, we find that policies designed for single-modal diffusion transfer poorly to world models due to two world-model-specific obstacles: token heterogeneity from multi-modal coupling and spatial variation, and non-uniform temporal dynamics where a small set of hard tokens drives error growth, making uniform skipping either unstable or overly conservative. We propose WorldCache, a caching framework tailored to diffusion world models. We introduce Curvature-guided Heterogeneous Token Prediction, which uses a physics-grounded curvature score to estimate token predictability and applies a Hermite-guided damped predictor for chaotic tokens with abrupt direction changes. We also design Chaotic-prioritized Adaptive Skipping, which accumulates a curvature-normalized, dimensionless drift signal and recomputes only when bottleneck tokens begin to drift. Experiments on diffusion world models show that WorldCache delivers up to 3.7times end-to-end speedups while maintaining 98\% rollout quality, demonstrating the vast advantages and practicality of WorldCache in resource-constrained scenarios. Our code is released in https://github.com/FofGofx/WorldCache.

The computational and memory demands of vanilla attention scale quadratically with the sequence length N, posing significant challenges for processing long sequences in Transformer models. FlashAttention alleviates these challenges by eliminating the O(N^2) memory dependency and reducing attention latency through IO-aware memory optimizations. However, its native support for certain attention mask types is limited, and it does not inherently accommodate more complex masking requirements. Previous approaches resort to using dense masks with O(N^2) memory complexity, leading to inefficiencies. In this paper, we propose FlashMask, an extension of FlashAttention that introduces a column-wise sparse representation of attention masks. This approach efficiently represents a wide range of mask types and facilitates the development of optimized kernel implementations. By adopting this novel representation, FlashMask achieves linear memory complexity O(N), suitable for modeling long-context sequences. Moreover, this representation enables kernel optimizations that eliminate unnecessary computations by leveraging sparsity in the attention mask, without sacrificing computational accuracy, resulting in higher computational efficiency. We evaluate FlashMask's performance in fine-tuning and alignment training of LLMs such as SFT, LoRA, DPO, and RM. FlashMask achieves significant throughput improvements, with end-to-end speedups ranging from 1.65x to 3.22x compared to existing FlashAttention dense method. Additionally, our kernel-level comparisons demonstrate that FlashMask surpasses the latest counterpart, FlexAttention, by 12.1% to 60.7% in terms of kernel TFLOPs/s, achieving 37.8% to 62.3% of the theoretical maximum FLOPs/s on the A100 GPU. The code is open-sourced on PaddlePaddle and integrated into PaddleNLP, supporting models with over 100 billion parameters for contexts up to 128K tokens.

Speculative decoding (SD) accelerates large language model inference by leveraging a draft-then-verify paradigm. To maximize the acceptance rate, recent methods construct expansive draft trees, which unfortunately incur severe VRAM bandwidth and computational overheads that bottleneck end-to-end speedups. While dynamic-depth pruning can reduce this latency by removing marginal branches, it also discards potentially valid candidates, preventing the acceptance rate from reaching the upper bound of dense trees. In this paper, we identify a critical opportunity in resource allocation: the transition from dense to pruned drafting frees up significant computational budget. To break this Pareto tradeoff, we introduce Graft, a compensation framework that couples pruning and retrieval as mutually reinforcing operations. Pruning supplies sufficient budget for retrieval, while retrieval compensates for pruning-induced coverage loss and recovers accepted length. By employing a sequential `prune-then-graft' mechanism, Graft attaches highly predictive retrieved tokens into positions opened by pruning, filling the topological gaps with near-zero overhead. Graft is entirely training-free and lossless. Comprehensive evaluations show that Graft establishes a new Pareto frontier across practical deployment settings, including short-context generation, long-context generation, and large-scale models. On short-context benchmarks, it achieves up to 5.41times speedup and improves average speedup over EAGLE-3 by up to 21.8% on the large-scale Qwen3-235B. We also provide a preliminary exploration of applying Graft to the DFlash-style block drafting paradigm, offering initial evidence and insights for extending grafting beyond autoregressive draft trees.

With the growing deployment of large language models (LLMs), LLM inference cost has become a key challenge. Pruning techniques that introduce sparsity into weight matrices can accelerate inference. However, maintaining model quality typically limits pruning to moderate unstructured sparsity (around 50\%). At these sparsity levels, none of the existing GPU kernels for sparse matrix multiplication (SpMM) can outperform their dense counterparts. This paper proposes an efficient GPU inference method for LLMs with moderate sparsity. We propose a three-layer matrix storage format comprising: (i) a Sparse-TC layer enabling sparse tensor cores to accelerate SpMM; (ii) a Slot-Filling layer using parallel differential distance for matrix compression while supporting low-cost on-chip decoding; (iii) a lightweight Residual Layer ensuring correct SpMM computation. Building on this format, we design a SpMM kernel that jointly utilizes sparse tensor cores and CUDA cores. This design enables an efficient execution pipeline and overlaps on-chip computation with memory access. Evaluations show that our work is the first to outperform dense matrix multiplication on modern GPUs equipped with high-bandwidth memory (HBM). It achieves up to 1.64x kernel-level speedup over SpInfer (EuroSys'25, Best paper) and up to 1.41x end-to-end speedups over FlashLLM (VLDB'24). Our source code: https://github.com/moui0/cudac.

Large language model (LLM) merging has become a key technique in modern LLM development pipelines, enabling the integration of multiple task- or domain-specific expert models without retraining. However, as the number of experts grows, existing merging implementations treat model parameters as unstructured files and execute merges in a stateless, one-shot manner, leading to excessive disk I/O, redundant parameter scans, and poor scalability. In this paper, we present MergePipe, a parameter management system for scalable LLM merging. MergePipe is the first system that treats LLM merging as a data management and execution problem, and introduces a catalog-driven abstraction over model parameters, merge plans, and execution lineage. At its core, MergePipe employs a cost-aware planner that explicitly models expert parameter I/O and enforces user-specified I/O budgets, followed by a streaming execution engine that materializes merged models under transactional guarantees. Our key insight is that while base model reads and output writes are unavoidable, expert parameter reads dominate merge cost and constitute the primary optimization target. By making expert access budget-aware throughout planning and execution, MergePipe mitigates the O(K) I/O growth of naive pipelines and achieves predictable scaling behavior. Experiments show that MergePipe reduces total I/O by up to an order of magnitude and delivers up to 11times end-to-end speedups (up to 90\% wall-time reduction) over state-of-the-art LLM merging pipelines.

Accelerating large language model (LLM) inference is critical for real-world deployments requiring high throughput and low latency. Contextual sparsity, where each token dynamically activates only a small subset of the model parameters, shows promise but does not scale to large batch sizes due to union of active neurons quickly approaching dense computation. We introduce Polar Sparsity, highlighting a key shift in sparsity importance from MLP to Attention layers as we scale batch size and sequence length. While MLP layers become more compute-efficient under batching, their sparsity vanishes. In contrast, attention becomes increasingly more expensive at scale, while their head sparsity remains stable and batch-invariant. We develop hardware-efficient, sparsity-aware GPU kernels for selective MLP and Attention computations, delivering up to \(2.2\times\) end-to-end speedups for models like OPT, LLaMA-2 \& 3, across various batch sizes and sequence lengths without compromising accuracy. To our knowledge, this is the first work to demonstrate that contextual sparsity can scale effectively to large batch sizes, delivering substantial inference acceleration with minimal changes, making Polar Sparsity practical for large-scale, high-throughput LLM deployment systems. Our code is available at: https://github.com/susavlsh10/Polar-Sparsity.

Speculative decoding accelerates LLM inference, but SOTA hidden-state-based drafters suffer from long-range decay: draft accuracy degrades as the speculative step increases. Existing work attributes this decay to train-inference mismatch and proposes test-time training (TTT) as a remedy, yet we observe that long-range decay persists even in TTT-trained drafters. We revisit long-range decay from the perspective of context information preservation. In hidden-state reuse, we argue the target hidden state acts as a biased context compression: it aggregates historical token information according to the attention query at the current position, yielding a compact representation optimized for immediate next-token prediction. This compression can suppress information less relevant to the current query but important for later speculative steps. In contrast, the target model's KV cache serves as an explicit context, retaining the complete set of token-wise KV representations. We therefore posit the KV-Reuse Hypothesis: allowing the draft model to reuse the target KV cache can provide richer signals for long-horizon drafting. To test this hypothesis, we introduce KVShot, a diagnostic framework that compares three reuse paradigms: hidden-only, KV-only, and hybrid. Extensive evaluations on Qwen3-8B show that KV-Reuse improves long-range acceptance, although end-to-end speedups remain marginal under current training pipelines. Our analysis identifies two key structural bottlenecks: shallow drafters struggle to estimate target queries accurately, and draft-side KV projections receive sparse gradient signals. These findings suggest that realizing the full potential of KV-aware decoding requires moving beyond TTT toward block-wise training paradigms. By exposing these bottlenecks, KVShot provides a foundational diagnostic testbed and a clear roadmap for designing next-generation inference architectures.

Speculative decoding is a powerful way to accelerate autoregressive large language models (LLMs), but directly porting it to vision-language models (VLMs) faces unique systems constraints: the prefill stage is dominated by visual tokens whose count scales with image resolution and video length, inflating both compute and memory, especially the key-value (KV) cache. We study speculative decoding for VLMs and introduce SpecVLM, a practical system that (1) establishes a strong EAGLE-2-style baseline, EagleVLM, delivering 1.5--2.3x end-to-end speedups over full autoregressive inference, and (2) further accelerates VLM inference with an elastic visual compressor that adaptively selects among pruning, pooling, convolution, and resampler primitives to balance FLOPs/parameters and accuracy per input. To avoid costly offline distillation corpora, we propose an online-logit distillation protocol that trains the draft model with on-the-fly teacher logits and penultimate features using a combined cross-entropy and Smooth L1 objective, eliminating storage and preprocessing while remaining compute-efficient. This protocol reveals a training-time scaling effect: longer online training monotonically increases the draft model's average accepted length, improving speculative efficiency. Empirically, SpecVLM achieves additional acceleration, culminating in 2.5--2.9x end-to-end speedups within 5 epochs across LLaVA and MMMU, consistently over resolutions and task difficulties, while preserving the target model's output distribution (lossless decoding). Our code is available at https://github.com/haiduo/SpecVLM.

The increasing demand to process long and high-resolution videos significantly burdens Large Vision-Language Models (LVLMs) due to the enormous number of visual tokens. Existing token reduction methods primarily focus on importance-based token pruning, which overlooks the redundancy caused by frame resemblance and repetitive visual elements. In this paper, we analyze the high vision token similarities in LVLMs. We reveal that token similarity distribution condenses as layers deepen while maintaining ranking consistency. Leveraging the unique properties of similarity over importance, we introduce FrameFusion, a novel approach that combines similarity-based merging with importance-based pruning for better token reduction in LVLMs. FrameFusion identifies and merges similar tokens before pruning, opening up a new perspective for token reduction. We evaluate FrameFusion on diverse LVLMs, including Llava-Video-{7B,32B,72B}, and MiniCPM-V-8B, on video understanding, question-answering, and retrieval benchmarks. Experiments show that FrameFusion reduces vision tokens by 70%, achieving 3.4-4.4x LLM speedups and 1.6-1.9x end-to-end speedups, with an average performance impact of less than 3%. Our code is available at https://github.com/thu-nics/FrameFusion.

We consider the problem of accurate sparse finetuning of large language models (LLMs), that is, finetuning pretrained LLMs on specialized tasks, while inducing sparsity in their weights. On the accuracy side, we observe that standard loss-based finetuning may fail to recover accuracy, especially at high sparsities. To address this, we perform a detailed study of distillation-type losses, determining an L2-based distillation approach we term SquareHead which enables accurate recovery even at higher sparsities, across all model types. On the practical efficiency side, we show that sparse LLMs can be executed with speedups by taking advantage of sparsity, for both CPU and GPU runtimes. While the standard approach is to leverage sparsity for computational reduction, we observe that in the case of memory-bound LLMs sparsity can also be leveraged for reducing memory bandwidth. We exhibit end-to-end results showing speedups due to sparsity, while recovering accuracy, on T5 (language translation), Whisper (speech translation), and open GPT-type (MPT for text generation). For MPT text generation, we show for the first time that sparse finetuning can reach 75% sparsity without accuracy drops, provide notable end-to-end speedups for both CPU and GPU inference, and highlight that sparsity is also compatible with quantization approaches. Models and software for reproducing our results are provided in Section 6.

Large language models (LLMs) now support context windows of hundreds of thousands to millions of tokens, enabling applications such as long-document summarization, large-scale code synthesis, multi-document question answering and persistent multi-turn dialogue. However, such extended contexts exacerbate the quadratic cost of self-attention, leading to severe latency in autoregressive decoding. Existing sparse attention methods alleviate these costs but rely on heuristic patterns that struggle to recall critical key-value (KV) pairs for each query, resulting in accuracy degradation. We introduce Adamas, a lightweight yet highly accurate sparse attention mechanism designed for long-context inference. Adamas applies the Hadamard transform, bucketization and 2-bit compression to produce compact representations, and leverages Manhattan-distance estimation for efficient top-k selections. Experiments show that Adamas matches the accuracy of full attention with only a 64-token budget, achieves near-lossless performance at 128, and supports up to 8x higher sparsity than prior state-of-the-art (SOTA) methods while delivering up to 4.4x self-attention and 1.5x end-to-end speedups on 32K-length sequences. Remarkably, Adamas attains comparable or even lower perplexity than full attention, underscoring its effectiveness in maintaining accuracy under aggressive sparsity.

Communication-reduction techniques are a popular way to improve scalability in data-parallel training of deep neural networks (DNNs). The recent emergence of large language models such as GPT has created the need for new approaches to exploit data-parallelism. Among these, fully-sharded data parallel (FSDP) training is highly popular, yet it still encounters scalability bottlenecks. One reason is that applying compression techniques to FSDP is challenging: as the vast majority of the communication involves the model's weights, direct compression alters convergence and leads to accuracy loss. We present QSDP, a variant of FSDP which supports both gradient and weight quantization with theoretical guarantees, is simple to implement and has essentially no overheads. To derive QSDP we prove that a natural modification of SGD achieves convergence even when we only maintain quantized weights, and thus the domain over which we train consists of quantized points and is, therefore, highly non-convex. We validate this approach by training GPT-family models with up to 1.3 billion parameters on a multi-node cluster. Experiments show that QSDP preserves model accuracy, while completely removing the communication bottlenecks of FSDP, providing end-to-end speedups of up to 2.2x.

Recent research has shown that training low-rank neural networks can effectively reduce the total number of trainable parameters without sacrificing predictive accuracy, resulting in end-to-end speedups. However, low-rank model training necessitates adjusting several additional factorization hyperparameters, such as the rank of the factorization at each layer. In this paper, we tackle this challenge by introducing Cuttlefish, an automated low-rank training approach that eliminates the need for tuning factorization hyperparameters. Cuttlefish leverages the observation that after a few epochs of full-rank training, the stable rank (i.e., an approximation of the true rank) of each layer stabilizes at a constant value. Cuttlefish switches from full-rank to low-rank training once the stable ranks of all layers have converged, setting the dimension of each factorization to its corresponding stable rank. Our results show that Cuttlefish generates models up to 5.6 times smaller than full-rank models, and attains up to a 1.2 times faster end-to-end training process while preserving comparable accuracy. Moreover, Cuttlefish outperforms state-of-the-art low-rank model training methods and other prominent baselines. The source code for our implementation can be found at: https://github.com/hwang595/Cuttlefish.

Attention scales quadratically with sequence length, fundamentally limiting long-context inference. Existing block-granularity sparsification can reduce latency, but coarse blocks impose an intrinsic sparsity ceiling, making further improvements difficult even with carefully engineered designs. We present S2O, which performs early stopping for sparse attention via online permutation. Inspired by virtual-to-physical address mapping in memory systems, S2O revisits and factorizes FlashAttention execution, enabling inference to load non-contiguous tokens rather than a contiguous span in the original order. Motivated by fine-grained structures in attention heatmaps, we transform explicit permutation into an online, index-guided, discrete loading policy; with extremely lightweight preprocessing and index-remapping overhead, it concentrates importance on a small set of high-priority blocks. Building on this importance-guided online permutation for loading, S2O further introduces an early-stopping rule: computation proceeds from high to low importance; once the current block score falls below a threshold, S2O terminates early and skips the remaining low-contribution blocks, thereby increasing effective sparsity and reducing computation under a controlled error budget. As a result, S2O substantially raises the practical sparsity ceiling. On Llama-3.1-8B under a 128K context, S2O reduces single-operator MSE by 3.82times at matched sparsity, and reduces prefill compute density by 3.31times at matched MSE; meanwhile, it preserves end-to-end accuracy and achieves 7.51times attention and 3.81times end-to-end speedups.

Vision Transformers (ViTs) have shown impressive performance and have become a unified backbone for multiple vision tasks. But both attention and multi-layer perceptions (MLPs) in ViTs are not efficient enough due to dense multiplications, resulting in costly training and inference. To this end, we propose to reparameterize the pre-trained ViT with a mixture of multiplication primitives, e.g., bitwise shifts and additions, towards a new type of multiplication-reduced model, dubbed ShiftAddViT, which aims for end-to-end inference speedups on GPUs without the need of training from scratch. Specifically, all MatMuls among queries, keys, and values are reparameterized by additive kernels, after mapping queries and keys to binary codes in Hamming space. The remaining MLPs or linear layers are then reparameterized by shift kernels. We utilize TVM to implement and optimize those customized kernels for practical hardware deployment on GPUs. We find that such a reparameterization on (quadratic or linear) attention maintains model accuracy, while inevitably leading to accuracy drops when being applied to MLPs. To marry the best of both worlds, we further propose a new mixture of experts (MoE) framework to reparameterize MLPs by taking multiplication or its primitives as experts, e.g., multiplication and shift, and designing a new latency-aware load-balancing loss. Such a loss helps to train a generic router for assigning a dynamic amount of input tokens to different experts according to their latency. In principle, the faster experts run, the larger amount of input tokens are assigned. Extensive experiments consistently validate the effectiveness of our proposed ShiftAddViT, achieving up to 5.18\times$ latency reductions on GPUs and 42.9%$ energy savings, while maintaining comparable accuracy as original or efficient ViTs.

Generative Pre-trained Transformer models, known as GPT or OPT, set themselves apart through breakthrough performance across complex language modelling tasks, but also by their extremely high computational and storage costs. Specifically, due to their massive size, even inference for large, highly-accurate GPT models may require multiple performant GPUs, which limits the usability of such models. While there is emerging work on relieving this pressure via model compression, the applicability and performance of existing compression techniques is limited by the scale and complexity of GPT models. In this paper, we address this challenge, and propose GPTQ, a new one-shot weight quantization method based on approximate second-order information, that is both highly-accurate and highly-efficient. Specifically, GPTQ can quantize GPT models with 175 billion parameters in approximately four GPU hours, reducing the bitwidth down to 3 or 4 bits per weight, with negligible accuracy degradation relative to the uncompressed baseline. Our method more than doubles the compression gains relative to previously-proposed one-shot quantization methods, preserving accuracy, allowing us for the first time to execute an 175 billion-parameter model inside a single GPU for generative inference. Moreover, we also show that our method can still provide reasonable accuracy in the extreme quantization regime, in which weights are quantized to 2-bit or even ternary quantization levels. We show experimentally that these improvements can be leveraged for end-to-end inference speedups over FP16, of around 3.25x when using high-end GPUs (NVIDIA A100) and 4.5x when using more cost-effective ones (NVIDIA A6000). The implementation is available at https://github.com/IST-DASLab/gptq.

We present MegaBlocks, a system for efficient Mixture-of-Experts (MoE) training on GPUs. Our system is motivated by the limitations of current frameworks, which restrict the dynamic routing in MoE layers to satisfy the constraints of existing software and hardware. These formulations force a tradeoff between model quality and hardware efficiency, as users must choose between dropping tokens from the computation or wasting computation and memory on padding. To address these limitations, we reformulate MoE computation in terms of block-sparse operations and develop new block-sparse GPU kernels that efficiently handle the dynamism present in MoEs. Our approach never drops tokens and maps efficiently to modern hardware, enabling end-to-end training speedups of up to 40% over MoEs trained with the state-of-the-art Tutel library and 2.4x over DNNs trained with the highly-optimized Megatron-LM framework.

We introduce Soft Anisotropic Diagrams (SAD), an explicit and differentiable image representation parameterized by a set of adaptive sites in the image plane. In SAD, each site specifies an anisotropic metric and an additively weighted distance score, and we compute pixel colors as a softmax blend over a small per-pixel top-K subset of sites. We induce a soft anisotropic additively weighted Voronoi partition (i.e., an Apollonius diagram) with learnable per-site temperatures, preserving informative gradients while allowing clear, content-aligned boundaries and explicit ownership. Such a formulation enables efficient rendering by maintaining a per-query top-K map that approximates nearest neighbors under the same shading score, allowing GPU-friendly, fixed-size local computation. We update this list using our top-K propagation scheme inspired by jump flooding, augmented with stochastic injection to provide probabilistic global coverage. Training follows a GPU-first pipeline with gradient-weighted initialization, Adam optimization, and adaptive budget control through densification and pruning. Across standard benchmarks, SAD consistently outperforms Image-GS and Instant-NGP at matched bitrate. On Kodak, SAD reaches 46.0 dB PSNR with 2.2 s encoding time (vs. 28 s for Image-GS), and delivers 4-19 times end-to-end training speedups over state-of-the-art baselines. We demonstrate the effectiveness of SAD by showcasing the seamless integration with differentiable pipelines for forward and inverse problems, efficiency of fast random access, and compact storage.

Communication has emerged as a critical bottleneck in the distributed training of large language models (LLMs). While numerous approaches have been proposed to reduce communication overhead, the potential of lossless compression has remained largely underexplored since compression and decompression typically consume larger overheads than the benefits of reduced communication traffic. We observe that the communication data, including activations, gradients and parameters, during training often follows a near-Gaussian distribution, which is a key feature for data compression. Thus, we introduce ZipCCL, a lossless compressed communication library of collectives for LLM training. ZipCCL is equipped with our novel techniques: (1) theoretically grounded exponent coding that exploits the Gaussian distribution of LLM tensors to accelerate compression without expensive online statistics, (2) GPU-optimized compression and decompression kernels that carefully design memory access patterns and pipeline using communication-aware data layout, and (3) adaptive communication strategies that dynamically switch collective operations based on workload patterns and system characteristics. Evaluated on a 64-GPU cluster using both mixture-of-experts and dense transformer models, ZipCCL reduces communication time by up to 1.35times and achieves end-to-end training speedups of up to 1.18times without any impact on model quality.

As inference on Large Language Models (LLMs) emerges as an important workload in machine learning applications, weight quantization has become a standard technique for efficient GPU deployment. Quantization not only reduces model size, but has also been shown to yield substantial speedups for single-user inference, due to reduced memory movement, with low accuracy impact. Yet, it remains open whether speedups are achievable also in batched settings with multiple parallel clients, which are highly relevant for practical serving. It is unclear whether GPU kernels can be designed to remain practically memory-bound, while supporting the substantially increased compute requirements of batched workloads. This paper resolves this question positively by describing the design of Mixed-precision Auto-Regressive LINear kernels, called MARLIN. Concretely, given a model whose weights are compressed via quantization to, e.g., 4 bits per element, MARLIN shows that batchsizes up to 16-32 can be supported with close to maximum (4times) quantization speedup, and larger batchsizes up to 64-128 with gradually decreasing, but still significant, acceleration. MARLIN accomplishes this via a combination of techniques, such as asynchronous memory access, complex task scheduling and pipelining, and bespoke quantization support. Our experiments show that MARLIN's near-optimal performance on individual LLM layers across different scenarios can also lead to end-to-end LLM inference speedups (of up to 2.8times) when integrated with the popular vLLM serving engine. Finally, MARLIN is extensible to further compression techniques, like NVIDIA 2:4 sparsity, leading to additional speedups.

We provide a new efficient version of the backpropagation algorithm, specialized to the case where the weights of the neural network being trained are sparse. Our algorithm is general, as it applies to arbitrary (unstructured) sparsity and common layer types (e.g., convolutional or linear). We provide a fast vectorized implementation on commodity CPUs, and show that it can yield speedups in end-to-end runtime experiments, both in transfer learning using already-sparsified networks, and in training sparse networks from scratch. Thus, our results provide the first support for sparse training on commodity hardware.

The deployment of Large Language Models is constrained by the memory and bandwidth demands of static weights and dynamic Key-Value cache. SVD-based compression provides a hardware-friendly solution to reduce these costs. However, existing methods suffer from two key limitations: some are suboptimal in reconstruction error, while others are theoretically optimal but practically inefficient. In this paper, we propose Swift-SVD, an activation-aware, closed-form compression framework that simultaneously guarantees theoretical optimum, practical efficiency and numerical stability. Swift-SVD incrementally aggregates covariance of output activations given a batch of inputs and performs a single eigenvalue decomposition after aggregation, enabling training-free, fast, and optimal layer-wise low-rank approximation. We employ effective rank to analyze local layer-wise compressibility and design a dynamic rank allocation strategy that jointly accounts for local reconstruction loss and end-to-end layer importance. Extensive experiments across six LLMs and eight datasets demonstrate that Swift-SVD outperforms state-of-the-art baselines, achieving optimal compression accuracy while delivering 3-70X speedups in end-to-end compression time. Our code will be released upon acceptance.

We present Galactic, a large-scale simulation and reinforcement-learning (RL) framework for robotic mobile manipulation in indoor environments. Specifically, a Fetch robot (equipped with a mobile base, 7DoF arm, RGBD camera, egomotion, and onboard sensing) is spawned in a home environment and asked to rearrange objects - by navigating to an object, picking it up, navigating to a target location, and then placing the object at the target location. Galactic is fast. In terms of simulation speed (rendering + physics), Galactic achieves over 421,000 steps-per-second (SPS) on an 8-GPU node, which is 54x faster than Habitat 2.0 (7699 SPS). More importantly, Galactic was designed to optimize the entire rendering + physics + RL interplay since any bottleneck in the interplay slows down training. In terms of simulation+RL speed (rendering + physics + inference + learning), Galactic achieves over 108,000 SPS, which 88x faster than Habitat 2.0 (1243 SPS). These massive speed-ups not only drastically cut the wall-clock training time of existing experiments, but also unlock an unprecedented scale of new experiments. First, Galactic can train a mobile pick skill to >80% accuracy in under 16 minutes, a 100x speedup compared to the over 24 hours it takes to train the same skill in Habitat 2.0. Second, we use Galactic to perform the largest-scale experiment to date for rearrangement using 5B steps of experience in 46 hours, which is equivalent to 20 years of robot experience. This scaling results in a single neural network composed of task-agnostic components achieving 85% success in GeometricGoal rearrangement, compared to 0% success reported in Habitat 2.0 for the same approach. The code is available at github.com/facebookresearch/galactic.

We show that the majority of the inference computations for large generative models such as LLaMA and OPT can be performed with both weights and activations being cast to 4 bits, in a way that leads to practical speedups while at the same time maintaining good accuracy. We achieve this via a hybrid quantization strategy called QUIK, which compresses most of the weights and activations to 4-bit, while keeping some outlier weights and activations in higher-precision. Crucially, our scheme is designed with computational efficiency in mind: we provide GPU kernels with highly-efficient layer-wise runtimes, which lead to practical end-to-end throughput improvements of up to 3.1x relative to FP16 execution. Code and models are provided at https://github.com/IST-DASLab/QUIK.

Speculative decoding accelerates LLM inference by using a draft model to look ahead, but gains are capped by the cost of autoregressive draft generation: increasing draft size elevates acceptance rates but introduces additional latency overhead exacerbating the speed-accuracy tradeoff. Prior methods (Medusa, Hydra, EAGLE) partially reduce draft cost but either degrade acceptance or introduce overheads that limit scaling. We present Mirror Speculative Decoding (Mirror-SD), an inference algorithm that breaks the latency-acceptance tradeoff. Mirror-SD launches branch-complete rollouts from early-exit signals in parallel with the target model's suffix and explicitly maps computation across heterogeneous accelerators (GPU and NPU) to exploit cross-device parallelism. The draft speculates forward continuations for the target to verify, while the target simultaneously speculates correction paths for the draft, converting speculation into two complementary execution pipelines. To further cut draft latency without weakening acceptance semantics, we add speculative streaming so the draft emits multiple tokens per step. This dual strategy of parallel heterogeneous execution plus multi-token speculative streaming pushes speculative decoding toward its ideal regime of high acceptance with low overhead. On SpecBench with server-scale models from 14B to 66B parameters, Mirror-SD delivers consistent end-to-end gains, achieving 2.8x-5.8x wall-time speedups across diverse tasks and a 30% average relative improvement over the strongest baseline, EAGLE3.

Long-context inference enhances the reasoning capability of Large Language Models (LLMs), but incurs significant computational overhead. Token-oriented methods, such as pruning and skipping, have shown great promise in reducing inference latency, yet still suffer from inherently insufficient structure optimization, outdated selection criteria, and redundancy interference, resulting in suboptimal speed-accuracy trade-off. To address these issues, we propose a novel training-free framework dubbed Self-Predictive Token Skipping (SPTS), for efficient long-context LLM inference. Specifically, motivated by probing the influence of target layers prior to skipping, we design two selective token skipping strategies for typical structures, including Partial Attention Probing (PAP) for multi-head attention and Low-rank Transformation Probing (LTP) for feed forward network. The former selects informative tokens via partial forward attention computation, while the latter constructs a low-rank proxy network to predict token transformations. In addition, a Multi-Stage Delayed Pruning (MSDP) strategy reallocates skipping budgets and progressively removes redundant tokens across layers. Extensive experiments display the effectiveness of our method, achieving up to 2.46times and 2.29times speedups for prefilling and end-to-end generation, respectively, while maintaining state-of-the-art accuracy. We will release the source code upon acceptance.

Speculative decoding accelerates LLM inference by drafting multiple tokens and verifying them in parallel with the target model. However, its practical speedup is constrained by the trade-off between draft quality and drafting cost: autoregressive drafters model causal dependencies among draft tokens but incur sequential overhead, while parallel drafters reduce drafting cost but weaken intra-block dependency modeling. In this paper, we propose Domino, a speculative decoding framework that decouples causal dependency modeling from expensive autoregressive draft execution. Domino first uses a parallel draft backbone to produce preliminary draft distributions for the entire block, and then applies a lightweight Domino head to refine them with prefix-dependent causal information. To stabilize teacher-forced causal encoding, we further introduce a base-anchored training curriculum that first strengthens the parallel backbone and then gradually shifts optimization toward the causally corrected final distribution. Experiments on Qwen3 models show that Domino achieves up to \(5.49\times\) end-to-end speedup under the Transformers backend and up to \(5.8\times\) throughput speedup under SGLang serving.

Long-context inference in LLMs faces the dual challenges of quadratic attention complexity and prohibitive KV cache memory. While token-level sparse attention offers superior accuracy, its indexing overhead is costly; block-level methods improve efficiency but sacrifice precision. We propose AsyncTLS, a hierarchical sparse attention system that combines coarse-grained block filtering with fine-grained token selection to balance accuracy and efficiency, coupled with an asynchronous offloading engine that overlaps KV cache transfers with computation via temporal locality exploitation. Evaluated on Qwen3 and GLM-4.7-Flash across GQA, and MLA architectures, AsyncTLS achieves accuracy comparable to full attention while delivering 1.2x - 10.0x operator speedups and 1.3x - 4.7x end-to-end throughput improvements on 48k - 96k contexts.

Long-context decoding in LLMs is IO-bound: each token re-reads an ever-growing KV cache. Prior accelerations cut bytes via compression, which lowers fidelity, or selection/eviction, which restricts what remains accessible, and both can degrade delayed recall and long-form generation. We introduce MAC-Attention, a fidelity- and access-preserving alternative that accelerates decoding by reusing prior attention computations for semantically similar recent queries. It starts with a match stage that performs pre-RoPE L2 matching over a short local window; an amend stage rectifies the reused attention by recomputing a small band near the match boundary; and a complete stage fuses the rectified results with fresh attention computed on the KV tail through a numerically stable merge. On a match hit, the compute and bandwidth complexity is constant regardless of context length. The method is model-agnostic and composes with IO-aware kernels, paged-KV managers, and MQA/GQA. Across LongBench v2 (120K), RULER (120K), and LongGenBench (16K continuous generation), compared to the latest FlashInfer library, MAC-Attention reduces KV accesses by up to 99%, cuts token generation latency by over 60% at 128K, and achieves over 14.3x attention-phase speedups, up to 2.6x end-to-end, while maintaining full-attention quality. By reusing computation, MAC-Attention delivers long-context inference that is both fast and faithful. Code is available here: https://github.com/YJHMITWEB/MAC-Attention.git

FlashAttention series has been widely applied in the inference of large language models (LLMs). However, FlashAttention series only supports the high-level GPU architectures, e.g., Ampere and Hopper. At present, FlashAttention series is not easily transferrable to NPUs and low-resource GPUs. Moreover, FlashAttention series is inefficient for multi- NPUs or GPUs inference scenarios. In this work, we propose FastAttention which pioneers the adaptation of FlashAttention series for NPUs and low-resource GPUs to boost LLM inference efficiency. Specifically, we take Ascend NPUs and Volta-based GPUs as representatives for designing our FastAttention. We migrate FlashAttention series to Ascend NPUs by proposing a novel two-level tiling strategy for runtime speedup, tiling-mask strategy for memory saving and the tiling-AllReduce strategy for reducing communication overhead, respectively. Besides, we adapt FlashAttention for Volta-based GPUs by redesigning the operands layout in shared memory and introducing a simple yet effective CPU-GPU cooperative strategy for efficient memory utilization. On Ascend NPUs, our FastAttention can achieve a 10.7times speedup compared to the standard attention implementation. Llama-7B within FastAttention reaches up to 5.16times higher throughput than within the standard attention. On Volta architecture GPUs, FastAttention yields 1.43times speedup compared to its equivalents in xformers. Pangu-38B within FastAttention brings 1.46times end-to-end speedup using FasterTransformer. Coupled with the propose CPU-GPU cooperative strategy, FastAttention supports a maximal input length of 256K on 8 V100 GPUs. All the codes will be made available soon.

Block diffusion LLMs are emerging as a promising next paradigm for language generation, but their use of KV caching makes memory access a dominant bottleneck in long-context settings. While dynamic sparse attention has been actively explored, existing methods designed for autoregressive LLMs rely on approximate importance estimation and perform poorly when adapted to block diffusion. This work identifies a key opportunity unique to block diffusion: attention at the first All-[MASK] denoising step reliably predicts important KV entries and budget requirements, enabling MAGE to perform a single exact attention pass per block and reuse it for training-free sparse denoising. Across long-context benchmarks including LongBench and Needle-in-a-Haystack, MAGE achieves near-lossless accuracy with a fraction of the KV budget while delivering up to 3-4x end-to-end speedup, consistently outperforming AR-oriented sparse attention baselines. A lightweight fine-tuning strategy further strengthens [MASK]-guided patterns with minimal cost, requiring only a few hours of training on a single NVIDIA H100 GPU for both 1.5B and 7B models.

Video diffusion transformers have achieved remarkable progress in high-quality video generation, but remain computationally expensive due to the quadratic complexity of attention over high-dimensional video sequences. Recent acceleration methods enhance the efficiency by exploiting the local sparsity of attention scores; yet they often struggle with accelerating the long-range computation. To address this problem, we propose VORTA, an acceleration framework with two novel components: 1) a sparse attention mechanism that efficiently captures long-range dependencies, and 2) a routing strategy that adaptively replaces full 3D attention with specialized sparse attention variants. VORTA achieves an end-to-end speedup 1.76times without loss of quality on VBench. Furthermore, it can seamlessly integrate with various other acceleration methods, such as model caching and step distillation, reaching up to speedup 14.41times with negligible performance degradation. VORTA demonstrates its efficiency and enhances the practicality of video diffusion transformers in real-world settings. Codes and weights are available at https://github.com/wenhao728/VORTA.

With the wide adoption of language models for IR -- and specifically RAG systems -- the latency of the underlying LLM becomes a crucial bottleneck, since the long contexts of retrieved passages lead large prompts and therefore, compute increase. Prompt compression, which reduces the size of input prompts while aiming to preserve performance on downstream tasks, has established itself as a cost-effective and low-latency method for accelerating inference in large language models. However, its usefulness depends on whether the additional preprocessing time during generation is offset by faster decoding. We present the first systematic, large-scale study of this trade-off, with thousands of runs and 30,000 queries across several open-source LLMs and three GPU classes. Our evaluation separates compression overhead from decoding latency while tracking output quality and memory usage. LLMLingua achieves up to 18% end-to-end speed-ups, when prompt length, compression ratio, and hardware capacity are well matched, with response quality remaining statistically unchanged across summarization, code generation, and question answering tasks. Outside this operating window, however, the compression step dominates and cancels out the gains. We also show that effective compression can reduce memory usage enough to offload workloads from data center GPUs to commodity cards, with only a 0.3s increase in latency. Our open-source profiler predicts the latency break-even point for each model-hardware setup, providing practical guidance on when prompt compression delivers real-world benefits.

While Diffusion Transformers (DiTs) have revolutionized high-fidelity video generation, their reliance on 3D full attention creates a quadratic computational bottleneck. Existing sparse methods face a dilemma: dynamic pruning suffers from prohibitive runtime overhead and memory fragmentation, while static heuristics fail to capture fine-grained dependencies. In this work, we propose ScalingAttention, a training-free framework grounded in a key inductive bias: while individual activations are input-dependent, the high-mass attention regions for each head rapidly converge to a stable, prompt-agnostic Intrinsic Sparse Topology. This topology is weight-encoded, scale-invariant, and efficient to extract. ScalingAttention decouples topology discovery from sparsity control via: (1) WEST (Weight-Encoded Sparse Topology), which extracts a robust block-sparse prior mask offline to eliminate runtime search; (2) FAST (Fidelity-Aware Sensitivity Tuning), which adaptively tunes head-wise sparsity based on diffusion fidelity requirements. To ensure practical acceleration, we co-design a hardware-aligned bit-wise block-sparse kernel. Experiments on Wan2.1 show up to 1.90X end-to-end speedup with superior fidelity, establishing a new Pareto frontier over state-of-the-art baselines.

LLM-based function calling enables intelligent agents to interact with external tools and environments, yet autoregressive decoding imposes a fundamental latency bottleneck that limits real-time applications such as embodied intelligence, game AI, and interactive avatars (e.g., 10 Hz control frequency). We observe that function calling differs fundamentally from free-form text generation: structured outputs exhibit substantial token redundancy (delimiters, parameter names), and arguments exhibit weak causal dependencies. Crucially, these two properties must be exploited jointly to achieve real-time performance. We present SimpleTool, which introduces special tokens that serve a dual role: compressing low-entropy tokens (4-6x reduction) while acting as mode selectors that enable independent parallel generation of function name and arguments. This synergistic design achieves 3-6x end-to-end speedup (up to 9.6x) with only +8.2% parallelization overhead. Experiments on five benchmarks across Qwen-series models (0.5B-14B) demonstrate substantial speedup while maintaining competitive or improved accuracy. On Mobile Actions, ST-Qwen-0.5B outperforms Google's FunctionGemma in both accuracy and latency consistency. With quantization on consumer-grade GPU, SimpleTool achieves 61.2ms P50 latency, enabling 16 Hz real-time control at 4B model scale, bridging the gap between LLM function calling and latency-critical real-world deployment.

Reinforcement learning (RL) post-training has proven effective at unlocking reasoning, self-reflection, and tool-use capabilities in large language models. As models extend to omni-modal inputs and agentic multi-turn workflows, RL training systems face three interdependent challenges: heterogeneous data flows, operational robustness at scale, and the staleness -- throughput tradeoff. We present Relax (Reinforcement Engine Leveraging Agentic X-modality), an open-source RL training engine that addresses these challenges through three co-designed architectural layers. First, an omni-native architecture builds multimodal support into the full stack -- from data preprocessing and modality-aware parallelism to inference generation -- rather than retrofitting it onto a text-centric pipeline. Second, each RL role runs as an independent, fault-isolated service that can be scaled, recovered, and upgraded without global coordination. Third, service-level decoupling enables asynchronous training via the TransferQueue data bus, where a single staleness parameter smoothly interpolates among on-policy, near-on-policy, and fully asynchronous execution. Relax achieves a 1.20times end-to-end speedup over veRL on Qwen3-4B on-policy training. Its fully async mode delivers a 1.76times speedup over colocate on Qwen3-4B and a 2.00times speedup on Qwen3-Omni-30B, while all modes converge to the same reward level. Relax supports R3 (Rollout Routing Replay)~ma2025r3 for MoE models with only 1.9\% overhead, compared to 32\% degradation in veRL under the same configuration. It further demonstrates stable omni-modal RL convergence on Qwen3-Omni across image, text, and audio, sustaining over 2{,}000 steps on video without degradation. Relax is available at https://github.com/rednote-ai/Relax.

With the rapid advancement of artificial intelligence technologies such as ChatGPT, AI agents, and video generation, contemporary mobile systems have begun integrating these AI capabilities on local devices to enhance privacy and reduce response latency. To meet the computational demands of AI tasks, current mobile SoCs are equipped with diverse AI accelerators, including GPUs and Neural Processing Units (NPUs). However, there has not been a comprehensive characterization of these heterogeneous processors, and existing designs typically only leverage a single AI accelerator for LLM inference, leading to suboptimal use of computational resources and memory bandwidth. In this paper, we first summarize key performance characteristics of heterogeneous processors, SoC memory bandwidth, etc. Drawing on these observations, we propose different heterogeneous parallel mechanisms to fully exploit both GPU and NPU computational power and memory bandwidth. We further design a fast synchronization mechanism between heterogeneous processors that leverages the unified memory architecture. By employing these techniques, we present HeteroInfer, the fastest LLM inference engine in mobile devices which supports GPU-NPU heterogeneous execution. Evaluation shows that HeteroInfer delivers a 1.34x to 6.02x end-to-end speedup over state-of-the-art GPU-only and NPU-only LLM engines, while maintaining negligible interference with other applications.

Low-Rank Adaptation (LoRA) has become the leading Parameter-Efficient Fine-Tuning (PEFT) method for Large Language Models (LLMs), as it significantly reduces GPU memory usage while maintaining competitive fine-tuned model quality on downstream tasks. Despite these benefits, we identify two key inefficiencies in existing LoRA fine-tuning systems. First, they incur substantial runtime overhead due to redundant memory accesses on large activation tensors. Second, they miss the opportunity to concurrently fine-tune multiple independent LoRA adapters that share the same base model on the same set of GPUs. This leads to missed performance gains such as reduced pipeline bubbles, better communication overlap, and improved GPU load balance. To address these issues, we introduce LoRAFusion, an efficient LoRA fine-tuning system for LLMs. At the kernel level, we propose a graph-splitting method that fuses memory-bound operations. This design eliminates unnecessary memory accesses and preserves the performance of compute-bound GEMMs without incurring the cost of recomputation or synchronization. At the scheduling level, LoRAFusion introduces an adaptive batching algorithm for multi-job fine-tuning. It first splits LoRA adapters into groups to intentionally stagger batch execution across jobs, and then solves a bin-packing problem within each group to generate balanced, dependency-aware microbatches. LoRAFusion achieves up to 1.96times (1.47times on average) end-to-end speedup compared to Megatron-LM, and up to 1.46times (1.29times on average) improvement over mLoRA, the state-of-the-art multi-LoRA fine-tuning system. Our fused kernel achieves up to 1.39times (1.27times on average) kernel performance improvement and can directly serve as a plug-and-play replacement in existing LoRA systems. We open-source LoRAFusion at https://github.com/CentML/lorafusion.

Diffusion models deliver high-fidelity synthesis but remain slow due to iterative sampling. We empirically observe there exists feature invariance in deterministic sampling, and present InvarDiff, a training-free acceleration method that exploits the relative temporal invariance across timestep-scale and layer-scale. From a few deterministic runs, we compute a per-timestep, per-layer, per-module binary cache plan matrix and use a re-sampling correction to avoid drift when consecutive caches occur. Using quantile-based change metrics, this matrix specifies which module at which step is reused rather than recomputed. The same invariance criterion is applied at the step scale to enable cross-timestep caching, deciding whether an entire step can reuse cached results. During inference, InvarDiff performs step-first and layer-wise caching guided by this matrix. When applied to DiT and FLUX, our approach reduces redundant compute while preserving fidelity. Experiments show that InvarDiff achieves 2-3times end-to-end speed-ups with minimal impact on standard quality metrics. Qualitatively, we observe almost no degradation in visual quality compared with full computations.

Large language models (LLMs) have demonstrated impressive abilities in various domains while the inference cost is expensive. The state-of-the-art methods use 2-bit quantization for mainstream LLMs. However, challenges still exist: (1) Nonnegligible accuracy loss for 2-bit quantization. Weights are quantized by groups, while the ranges of weights are large in some groups, resulting in large quantization errors and nonnegligible accuracy loss (e.g. >3% for Llama2-7b with 2-bit quantization in GPTQ and Greenbit). (2) Limited accuracy improvement by adding 4-bit weights. Increasing 10% extra average bit more 4-bit weights only leads to <0.5% accuracy improvement on a quantized Llama2-7b. (3) Time-consuming dequantization operations on GPUs. The dequantization operations lead to >50% execution time, hindering the potential of reducing LLM inference cost. To tackle these challenges, we propose the following techniques: (1) We only quantize a small fraction of groups with the larger range using 4-bit with memory alignment consideration on GPUs.(2) We design the asynchronous dequantization on GPUs, leading to up to 3.92X speedup. We conduct extensive experiments on different model sizes. We achieve 2.85-bit for each weight and the end-to-end speedup for Llama2-7b is 1.74X over the original model, and we reduce both runtime cost and hardware cost by up to 2.70X and 2.81X with less GPU requirements.

SAM3D enables scalable, open-world 3D reconstruction from complex scenes, yet its deployment is hindered by prohibitive inference latency. In this work, we conduct the first systematic investigation into its inference dynamics, revealing that generic acceleration strategies are brittle in this context. We demonstrate that these failures stem from neglecting the pipeline's inherent multi-level heterogeneity: the kinematic distinctiveness between shape and layout, the intrinsic sparsity of texture refinement, and the spectral variance across geometries. To address this, we present Fast-SAM3D, a training-free framework that dynamically aligns computation with instantaneous generation complexity. Our approach integrates three heterogeneity-aware mechanisms: (1) Modality-Aware Step Caching to decouple structural evolution from sensitive layout updates; (2) Joint Spatiotemporal Token Carving to concentrate refinement on high-entropy regions; and (3) Spectral-Aware Token Aggregation to adapt decoding resolution. Extensive experiments demonstrate that Fast-SAM3D delivers up to 2.67times end-to-end speedup with negligible fidelity loss, establishing a new Pareto frontier for efficient single-view 3D generation. Our code is released in https://github.com/wlfeng0509/Fast-SAM3D.

Speculative decoding (SD) accelerates autoregressive Large Language Models (LLMs) by drafting multiple tokens and verifying them in parallel, but it faces a scaling limitation: increasing the draft budget improves speed only when acceptance remains high and drafting overhead stays low. This ceiling has been difficult to break because prior head-based SD methods face a causality-efficiency dilemma. Autoregressive drafters produce path-conditioned candidates that are effective for tree speculative decoding with higher acceptance length, but their drafting cost grows with tree depth. Bidirectional block-diffusion drafters generate all positions in one pass, but their branch-agnostic marginals can form individually plausible yet mutually inconsistent trees, wasting budget and reducing acceptance. We propose JetSpec, a head-based SD framework that combines one-forward drafting efficiency with branch-wise causal conditioning. JetSpec trains a causal parallel draft head over fused hidden states from the frozen target model, producing candidate trees whose scores align with the target model's autoregressive factorization. This enables JetSpec to convert larger draft budgets into longer accepted prefixes and higher end-to-end speedup. Across math, coding, and chat benchmarks on dense and MoE Qwen3 models, JetSpec consistently outperforms bidirectional-head and tree-based SD baselines. On H100 GPUs, JetSpec achieves up to 9.64x speedup on MATH-500 and 4.58x on open-ended conversational workloads, with further latency gains demonstrated through vLLM integration under realistic serving loads. Our code and models are available at https://github.com/hao-ai-lab/JetSpec.

k-means has historically been positioned primarily as an offline processing primitive, typically used for dataset organization or embedding preprocessing rather than as a first-class component in online systems. In this work, we revisit this classical algorithm under the lens of modern AI system design and enable k-means as an online primitive. We point out that existing GPU implementations of k-means remain fundamentally bottlenecked by low-level system constraints rather than theoretical algorithmic complexity. Specifically, the assignment stage suffers from a severe IO bottleneck due to the massive explicit materialization of the N times K distance matrix in High Bandwidth Memory (HBM). Simultaneously, the centroid update stage is heavily penalized by hardware-level atomic write contention caused by irregular, scatter-style token aggregations. To bridge this performance gap, we propose flash-kmeans, an IO-aware and contention-free k-means implementation for modern GPU workloads. Flash-kmeans introduces two core kernel-level innovations: (1) FlashAssign, which fuses distance computation with an online argmin to completely bypass intermediate memory materialization; (2) sort-inverse update, which explicitly constructs an inverse mapping to transform high-contention atomic scatters into high-bandwidth, segment-level localized reductions. Furthermore, we integrate algorithm-system co-designs, including chunked-stream overlap and cache-aware compile heuristics, to ensure practical deployability. Extensive evaluations on NVIDIA H200 GPUs demonstrate that flash-kmeans achieves up to 17.9times end-to-end speedup over best baselines, while outperforming industry-standard libraries like cuML and FAISS by 33times and over 200times, respectively.

Efficient long-sequence generation is a critical challenge for Large Language Models. While recent sparse decoding methods improve efficiency, they suffer from KV cache misalignment, where approximation errors accumulate and degrade generation quality. In this work, we propose Rectified Sparse Attention (ReSA), a simple yet effective method that combines block-sparse attention with periodic dense rectification. By refreshing the KV cache at fixed intervals using a dense forward pass, ReSA bounds error accumulation and preserves alignment with the pretraining distribution. Experiments across math reasoning, language modeling, and retrieval tasks demonstrate that ReSA achieves near-lossless generation quality with significantly improved efficiency. Notably, ReSA delivers up to 2.42times end-to-end speedup under decoding at 256K sequence length, making it a practical solution for scalable long-context inference. Code is available at https://aka.ms/ReSA-LM.

Large language model-based deep research agents have been increasingly popular for addressing long-horizon information-seeking tasks, but they often incur high end-to-end latency due to extensive reasoning and frequent tool use. Speculation frameworks aim to reduce latency by overlapping action execution with reasoning; however, existing approaches typically rely on uniform speculation strategies and strict action matching, which limits inference speedups and robustness. In this work, we revisit the speculate-verify paradigm for deep research agents through the lens of action heterogeneity. We show that Search and Visit actions exhibit fundamentally different reasoning and model capacity requirements: entropy-based analysis reveals that Search decisions have higher uncertainty and benefit significantly from explicit reasoning, whereas Visit decisions have lower entropy and depend primarily on model capacity. Motivated by this dual-process characteristic, we propose DualSpec, a heterogeneous speculation framework equipped with a lightweight, confidence-based semantic verifier. Experiments across multiple models and benchmarks demonstrate that DualSpec achieves up to 3.28times end-to-end speedup while maintaining accuracy comparable to fully reasoning agents.

Using a diffusion model for parallel drafting is a promising approach for speculative decoding. By predicting tokens at multiple future positions in a single forward pass, diffusion drafters substantially reduce drafting latency. However, this shifts the bottleneck to verification: verifying a single sequence limits acceptance length, while verifying large draft trees incurs excessive target-model latency. We identify a key mismatch in existing draft-tree methods: existing diffusion-tree methods rank nodes by the marginal probability, ignoring that verification is prefix-conditioned. As a result, they may verify unreachable descendants of rejected prefixes, increasing latency with limited acceptance gains. To address this, we propose TAPS, a target-aware prefix selection method that turns diffusion marginals into path-conditioned acceptance estimates. TAPS then selects a compact prefix-closed subtree under a fixed verification budget, improving the acceptance-cost tradeoff rather than simply expanding the draft tree. Experiments across diverse datasets and model families demonstrate that TAPS achieves up to 7.9x lossless end-to-end speedup over vanilla autoregressive decoding, outperforming state-of-the-art DFlash and DDTree by 1.36x and 1.74x respectively. Our work is available at https://anonymous.4open.science/r/TAPS-EMNLP2026-53DD

Scaling Diffusion Transformers to generate high-resolution, long videos is constrained by the quadratic cost of self-attention, and existing sparse attention methods degrade under high sparsity. We show empirically that generation quality is determined not by the sparsity ratio itself, but by how well the sparse mask aligns with the tile-wise geometry of full attention. Based on this insight, we propose Veda, a distilled sparse attention framework that formulates tile selection as an explicit reconstruction problem from full attention. Veda integrates statistics-aware tile scoring with head-aware tiling to reduce estimation error and structural mismatch, enabling aggressive sparsity. A hardware-efficient tile-skipping kernel converts theoretical sparsity into practical wall-clock speedups. Experiments on large video diffusion models, including Waver and Wan2.1, demonstrate substantial acceleration with no noticeable degradation in generation quality. To generate 720P 10-second videos on Waver-T2V-12B, Veda achieves a 5.1times end-to-end speedup and a 10.5times self-attention speedup, reducing attention overhead from 92% to 50%. Notably, the gains increase with sequence length, indicating that Veda scales favorably with spatiotemporal resolution across models.

Tree-based speculative decoding accelerates autoregressive generation by verifying a branching tree of draft tokens in a single target-model forward pass. However, existing methods prioritize maximizing token-level likelihood or the number of accepted tokens while ignoring a critical ``efficiency paradox'': the computational overhead of drafting and verifying big trees can grow super-linearly, particularly at scale. This often leads to negative wall-clock speedup when batch sizes increase or hardware saturation limits are reached. To address this, we propose SMART, a system-aware marginal analysis framework for runtime tree construction. SMART reformulates tree expansion as a hardware-aware optimization problem that directly maximizes end-to-end speedup. By applying a principled marginal benefit--cost rule at inference time, SMART expands a node only when its marginal benefit--cost ratio exceeds the tree-level speedup. SMART is training-free and serves as a plug-and-play controller for existing frameworks like MSD and EAGLE. Extensive evaluations across three MLLMs (e.g., LLaVA, Qwen2-VL) and four LLMs (e.g., Llama-3.1, DeepSeek-R1) demonstrate that SMART consistently outperforms state-of-the-art baselines. It delivers an average additional speedup of 20.0\% for MLLMs and 15.4\% for LLMs across compute-bound batching regimes and diverse GPU architectures without performance loss.

Quantized training of Large Language Models (LLMs) remains an open challenge, as maintaining accuracy while performing all matrix multiplications in low precision has proven difficult. This is particularly the case when fine-tuning pre-trained models, which often already have large weight and activation outlier values that render quantized optimization difficult. We present HALO, a novel quantization-aware training approach for Transformers that enables accurate and efficient low-precision training by combining 1) strategic placement of Hadamard rotations in both forward and backward passes, to mitigate outliers during the low-precision computation, 2) FSDP integration for low-precision communication, and 3) high-performance kernel support. Our approach ensures that all large matrix multiplications during the forward and backward passes are executed in lower precision. Applied to LLAMA-family models, HALO achieves near-full-precision-equivalent results during fine-tuning on various tasks, while delivering up to 1.31x end-to-end speedup for full fine-tuning on RTX 4090 GPUs. Our method supports both standard and parameter-efficient fine-tuning (PEFT) methods, both backed by efficient kernel implementations. Our results demonstrate the first practical approach to fully quantized LLM fine-tuning that maintains accuracy in FP8 precision, while delivering performance benefits.

We present Ovis2.5, a successor to Ovis2 designed for native-resolution visual perception and strong multimodal reasoning. Ovis2.5 integrates a native-resolution vision transformer that processes images at their native, variable resolutions, avoiding the degradation from fixed-resolution tiling and preserving both fine detail and global layout -- crucial for visually dense content like complex charts. To strengthen reasoning, we train the model to move beyond linear chain-of-thought and perform reflection -- including self-checking and revision. This advanced capability is exposed as an optional "thinking mode" at inference time, allowing users to trade latency for enhanced accuracy on difficult inputs. The model is trained via a comprehensive five-phase curriculum that progressively builds its skills. The process begins with foundational visual and multimodal pretraining, advances through large-scale instruction tuning, and culminates in alignment and reasoning enhancement using DPO and GRPO. To scale these upgrades efficiently, we employ multimodal data packing and hybrid parallelism, yielding a significant end-to-end speedup. We release two open-source models: Ovis2.5-9B and Ovis2.5-2B. The latter continues the "small model, big performance" philosophy of Ovis2, making it ideal for resource-constrained, on-device scenarios. On the OpenCompass multimodal leaderboard, Ovis2.5-9B averages 78.3, marking a substantial improvement over its predecessor, Ovis2-8B, and achieving state-of-the-art results among open-source MLLMs in the sub-40B parameter range; Ovis2.5-2B scores 73.9, establishing SOTA for its size. Beyond aggregate scores, Ovis2.5 achieves leading results on STEM benchmarks, exhibits strong capabilities on grounding and video tasks, and achieves open-source SOTA at its scale for complex chart analysis.

Reinforcement learning (RL) is essential for enhancing the complex reasoning capabilities of large language models (LLMs). However, existing RL training pipelines are computationally inefficient and resource-intensive, with the rollout phase accounting for over 70% of total training time. Quantized RL training, particularly using FP8 precision, offers a promising approach to mitigating this bottleneck. A commonly adopted strategy applies FP8 precision during rollout while retaining BF16 precision for training. In this work, we present the first comprehensive study of FP8 RL training and demonstrate that the widely used BF16-training + FP8-rollout strategy suffers from severe training instability and catastrophic accuracy collapse under long-horizon rollouts and challenging tasks. Our analysis shows that these failures stem from the off-policy nature of the approach, which introduces substantial numerical mismatch between training and inference. Motivated by these observations, we propose Jet-RL, an FP8 RL training framework that enables robust and stable RL optimization. The key idea is to adopt a unified FP8 precision flow for both training and rollout, thereby minimizing numerical discrepancies and eliminating the need for inefficient inter-step calibration. Extensive experiments validate the effectiveness of Jet-RL: our method achieves up to 33% speedup in the rollout phase, up to 41% speedup in the training phase, and a 16% end-to-end speedup over BF16 training, while maintaining stable convergence across all settings and incurring negligible accuracy degradation.

Transformer-based Large Language Models rely critically on the KV cache to efficiently handle extended contexts during the decode phase. Yet, the size of the KV cache grows proportionally with the input length, burdening both memory bandwidth and capacity as decoding progresses. To address this challenge, we present RocketKV, a training-free KV cache compression strategy containing two consecutive stages. In the first stage, it performs coarse-grain permanent KV cache eviction on the input sequence tokens. In the second stage, it adopts a hybrid sparse attention method to conduct fine-grain top-k sparse attention, approximating the attention scores by leveraging both head and sequence dimensionality reductions. We show that RocketKV provides a compression ratio of up to 400times, end-to-end speedup of up to 3.7times as well as peak memory reduction of up to 32.6% in the decode phase on an NVIDIA A100 GPU compared to the full KV cache baseline, while achieving negligible accuracy loss on a variety of long-context tasks. We also propose a variant of RocketKV for multi-turn scenarios, which consistently outperforms other existing methods and achieves accuracy nearly on par with an oracle top-k attention scheme. The source code is available here: https://github.com/NVlabs/RocketKV.

Large reasoning models (LRMs) achieve strong performance on complex reasoning tasks by generating long, multi-step reasoning trajectories, but inference-time scaling incurs substantial deployment cost. A key challenge is that generation difficulty varies within a single output, whereas existing efficiency-oriented approaches either ignore this intra-generation variation or rely on supervised token-level routing with high system complexity. We present RelayGen, a training-free, segment-level runtime model switching framework that exploits difficulty variation in long-form reasoning. Through offline analysis of generation uncertainty using token probability margins, we show that coarse-grained segment-level control is sufficient to capture difficulty transitions within a reasoning trajectory. RelayGen identifies model-specific switch cues that signal transitions to lower-difficulty segments and dynamically delegates their continuation to a smaller model, while preserving high-difficulty reasoning on the large model. Across multiple reasoning benchmarks, RelayGen substantially reduces inference latency while preserving most of the accuracy of large models. When combined with speculative decoding, RelayGen achieves up to 2.2times end-to-end speedup with less than 2\% accuracy degradation, without requiring additional training or learned routing components.

Scaling the context length of large language models (LLMs) offers significant benefits but is computationally expensive. This expense stems primarily from the self-attention mechanism, whose O(N^2) complexity with respect to sequence length presents a major bottleneck for both memory and latency. Fortunately, the attention matrix is often sparse, particularly for long sequences, suggesting an opportunity for optimization. Block-sparse attention has emerged as a promising solution that partitions sequences into blocks and skips computation for a subset of these blocks. However, the effectiveness of this method is highly dependent on the underlying attention patterns, which can lead to sub-optimal block-level sparsity. For instance, important key tokens for queries within a single block may be scattered across numerous other blocks, leading to computational redundancy. In this work, we propose Permuted Block-Sparse Attention (PBS-Attn), a plug-and-play method that leverages the permutation properties of attention to increase block-level sparsity and enhance the computational efficiency of LLM prefilling. We conduct comprehensive experiments on challenging real-world long-context datasets, demonstrating that PBS-Attn consistently outperforms existing block-sparse attention methods in model accuracy and closely matches the full attention baseline. Powered by our custom permuted-FlashAttention kernels, PBS-Attn achieves an end-to-end speedup of up to 2.75times in long-context prefilling, confirming its practical viability. Code available at https://github.com/xinghaow99/pbs-attn

KV-Cache compression methods generally sample a KV-Cache of effectual tokens or quantize it into lower bits. However, these methods cannot exploit the redundancy of the hidden dimension of KV tensors. This paper investigates a unique hidden dimension approach called Palu, a novel KV-Cache compression framework that utilizes low-rank projection. Palu decomposes the linear layers into low-rank matrices, caches the smaller intermediate states, and reconstructs the full keys and values on the fly. To improve accuracy, compression rate, and efficiency, Palu further encompasses (1) a medium-grained low-rank decomposition scheme, (2) an efficient rank search algorithm, (3) a low-rank-aware quantization algorithm, and (4) matrix fusion with optimized GPU kernels. Our extensive experiments with popular LLMs show that Palu can compress KV-Cache by more than 91.25% while maintaining a significantly better accuracy (up to 1.19 lower perplexity) than state-of-the-art KV-Cache quantization methods at a similar or even higher memory usage. When compressing KV-Cache for 50%, Palu delivers up to 1.61x end-to-end speedup for the attention module. Our code is publicly available at https://github.com/shadowpa0327/Palu.

Diffusion transformer-based video generation models (DiTs) have recently attracted widespread attention for their excellent generation quality. However, their computational cost remains a major bottleneck-attention alone accounts for over 80% of total latency, and generating just 8 seconds of 720p video takes tens of minutes-posing serious challenges to practical application and scalability. To address this, we propose the DraftAttention, a training-free framework for the acceleration of video diffusion transformers with dynamic sparse attention on GPUs. We apply down-sampling to each feature map across frames in the compressed latent space, enabling a higher-level receptive field over the latent composed of hundreds of thousands of tokens. The low-resolution draft attention map, derived from draft query and key, exposes redundancy both spatially within each feature map and temporally across frames. We reorder the query, key, and value based on the draft attention map to guide the sparse attention computation in full resolution, and subsequently restore their original order after the attention computation. This reordering enables structured sparsity that aligns with hardware-optimized execution. Our theoretical analysis demonstrates that the low-resolution draft attention closely approximates the full attention, providing reliable guidance for constructing accurate sparse attention. Experimental results show that our method outperforms existing sparse attention approaches in video generation quality and achieves up to 1.75x end-to-end speedup on GPUs. Code: https://github.com/shawnricecake/draft-attention

W4A4 quantization promises full utilization of INT4 Tensor Cores, yet group dequantization overhead on CUDA Cores has driven existing systems to mixed-precision fallbacks. We present the first systematic study of how intra-SM compute balance governs this bottleneck. Through controlled benchmarks across four GPUs from Ampere and Ada architectures, we identify the Tensor Cores to CUDA Cores throughput ratio (ρ) as the primary hardware indicator: the W4A4-g128 kernel yields 2.0--2.5times speedup on RTX~3090 (ρ=16) yet degrades to 0.43--0.47times on A100 (ρ=64) in compute-bond scenarios, establishing W4A4 viability as platform-dependent rather than universally infeasible. Guided by this finding, we build APEX4, which co-designs pure INT4 GEMM kernels with ρ-aware granularity adaptation to mitigate the CUDA Cores dequantization bottleneck. APEX4 achieves perplexity within 0.63 of FP16 on LLaMA-2-70B and outperforms W4Ax Atom-g128 by 4.0\%--4.4\% in zero-shot accuracy. Deployed as a drop-in replacement in unmodified vLLM, it delivers up to 1.66times end-to-end speedup on L40S (ρ=8), and 1.78times on RTX~3090 (ρ=16), 2.09times on A40 (ρ=16), while recovering A100 (ρ=64) to 1.20--1.40times via the mixed-granularity mode. Our code is available at https://github.com/APEX4-W4A4/APEX4-W4A4.

Deep learning on point clouds has received increased attention thanks to its wide applications in AR/VR and autonomous driving. These applications require low latency and high accuracy to provide real-time user experience and ensure user safety. Unlike conventional dense workloads, the sparse and irregular nature of point clouds poses severe challenges to running sparse CNNs efficiently on the general-purpose hardware. Furthermore, existing sparse acceleration techniques for 2D images do not translate to 3D point clouds. In this paper, we introduce TorchSparse, a high-performance point cloud inference engine that accelerates the sparse convolution computation on GPUs. TorchSparse directly optimizes the two bottlenecks of sparse convolution: irregular computation and data movement. It applies adaptive matrix multiplication grouping to trade computation for better regularity, achieving 1.4-1.5x speedup for matrix multiplication. It also optimizes the data movement by adopting vectorized, quantized and fused locality-aware memory access, reducing the memory movement cost by 2.7x. Evaluated on seven representative models across three benchmark datasets, TorchSparse achieves 1.6x and 1.5x measured end-to-end speedup over the state-of-the-art MinkowskiEngine and SpConv, respectively.

Semi-structured 2:4 sparsity is widely supported by modern accelerators, providing up to a 2x theoretical speedup. However, its strict 50% sparsity constraint often causes non-negligible accuracy degradation under post-training pruning. Meanwhile, existing relaxed sparsity formats either require specialized compiler support or introduce runtime overheads that limit end-to-end speedup. We propose Spense, a practical hybrid sparse-dense format that splits each weight matrix into a 2:4 sparse region and a dense region. This design relaxes the effective sparsity constraint while remaining compatible with existing high-performance sparse and dense GEMM libraries, avoiding both custom compiler support and input activation expansion. Building on this format, we introduce SpenseGPT, a one-shot post-training pruning method that produces sparse and dense regions. Notably, we show that selecting the right dense regions is important, and we devise two different strategies to choose them. Experiments on Qwen3-32B and Seed-OSS-36B demonstrate that our method achieves up to 1.2x end-to-end decoding speedup on B200 GPUs with FP8 precision, while preserving accuracy. To the best of our knowledge, this is the first one-shot pruning demonstration of real-world end-to-end LLM decoding speedup from semi-structured sparse tensor cores on recent GPUs such as B200s, while maintaining model quality.

Diffusion Transformers (DiTs) dominate video generation but their high computational cost severely limits real-world applicability, usually requiring tens of minutes to generate a few seconds of video even on high-performance GPUs. This inefficiency primarily arises from the quadratic computational complexity of 3D Full Attention with respect to the context length. In this paper, we propose a training-free framework termed Sparse VideoGen (SVG) that leverages the inherent sparsity in 3D Full Attention to boost inference efficiency. We reveal that the attention heads can be dynamically classified into two groups depending on distinct sparse patterns: (1) Spatial Head, where only spatially-related tokens within each frame dominate the attention output, and (2) Temporal Head, where only temporally-related tokens across different frames dominate. Based on this insight, SVG proposes an online profiling strategy to capture the dynamic sparse patterns and predicts the type of attention head. Combined with a novel hardware-efficient tensor layout transformation and customized kernel implementations, SVG achieves up to 2.28x and 2.33x end-to-end speedup on CogVideoX-v1.5 and HunyuanVideo, respectively, while preserving generation quality.

Increasing the context length of large language models (LLMs) unlocks fundamentally new capabilities, but also significantly increases the memory footprints of training. Previous model-parallel systems such as Megatron-LM partition and compute different attention heads in parallel, resulting in large communication volumes, so they cannot scale beyond the number of attention heads, thereby hindering its adoption. In this paper, we introduce a new approach, LightSeq, for long-context LLMs training. LightSeq has many notable advantages. First, LightSeq partitions over the sequence dimension, hence is agnostic to model architectures and readily applicable for models with varying numbers of attention heads, such as Multi-Head, Multi-Query and Grouped-Query attention. Second, LightSeq not only requires up to 4.7x less communication than Megatron-LM on popular LLMs but also overlaps the communication with computation. To further reduce the training time, LightSeq features a novel gradient checkpointing scheme to bypass an forward computation for memory-efficient attention. We evaluate LightSeq on Llama-7B and its variants with sequence lengths from 32K to 512K. Through comprehensive experiments on single and cross-node training, we show that LightSeq achieves up to 1.24-2.01x end-to-end speedup, and a 2-8x longer sequence length on models with fewer heads, compared to Megatron-LM. Codes will be available at https://github.com/RulinShao/LightSeq.

Vision Language Models (VLMs) have rapidly advanced in integrating visual and textual reasoning, powering applications across high-resolution image understanding, long-video analysis, and multi-turn conversation. However, their scalability remains limited by the growing number of visual tokens that dominate inference latency. We present SparseVILA, a new paradigm for efficient VLM inference that decouples visual sparsity across the prefilling and decoding stages. SparseVILA distributes sparsity across stages by pruning redundant visual tokens during prefill and retrieving only query-relevant tokens during decoding. This decoupled design matches leading prefill pruning methods while preserving multi-turn fidelity by retaining most of the visual cache so that query-aware tokens can be retrieved at each conversation round. Built on an AWQ-optimized inference pipeline, SparseVILA achieves up to 4.0 times faster prefilling, 2.5 times faster decoding, and an overall 2.6 times end-to-end speedup on long-context video tasks -- while improving accuracy on document-understanding and reasoning tasks. By decoupling query-agnostic pruning and query-aware retrieval, SparseVILA establishes a new direction for efficient multimodal inference, offering a training-free, architecture-agnostic framework for accelerating large VLMs without sacrificing capability.

Diffusion language models offer parallel token generation and inherent bidirectionality, promising more efficient and powerful sequence modeling compared to autoregressive approaches. However, state-of-the-art diffusion models (e.g., Dream 7B, LLaDA 8B) suffer from slow inference. While they match the quality of similarly sized Autoregressive (AR) Models (e.g., Qwen2.5 7B, Llama3 8B), their iterative denoising requires multiple full-sequence forward passes, resulting in high computational costs and latency, particularly for long input prompts and long-context scenarios. Furthermore, parallel token generation introduces token incoherence problems, and current sampling heuristics suffer from significant quality drops with decreasing denoising steps. We address these limitations with two training-free techniques. First, we propose FreeCache, a Key-Value (KV) approximation caching technique that reuses stable KV projections across denoising steps, effectively reducing the computational cost of DLM inference. Second, we introduce Guided Diffusion, a training-free method that uses a lightweight pretrained autoregressive model to supervise token unmasking, dramatically reducing the total number of denoising iterations without sacrificing quality. We conduct extensive evaluations on open-source reasoning benchmarks, and our combined methods deliver up to a 34x end-to-end speedup without compromising accuracy. For the first time, diffusion language models achieve a comparable and even faster latency as the widely adopted autoregressive models. Our work successfully paved the way for scaling up the diffusion language model to a broader scope of applications across different domains.

Diffusion transformers have achieved remarkable success in high-quality video generation, yet their reliance on spatiotemporal 3D full attention incurs prohibitive computational cost due to the quadratic complexity of attention. Block sparse attention is a common approach to mitigate this by focusing computation on important regions. However, attention maps in DiTs exhibit inherently dynamic and fine-grained sparsity, which causes existing block sparse attention methods to degrade significantly in quality, especially at high sparsity ratios. In this paper, we revisit block sparse attention and derive a theoretical lower bound on attention recall to characterize the key factors governing its effectiveness. Guided by these insights, we propose DFSAttn, a training-free sparse attention framework that enables dynamic, fine-grained sparsification efficiently. DFSAttn incorporates three core designs: Hilbert curve-based token reordering to achieve fine-grained sparsity while preserving efficient GPU execution, hierarchical block scoring for accurate block importance estimation, and sparse mask caching with adaptive ratios to balance accuracy and efficiency. Experimental results demonstrate that DFSAttn consistently outperforms prior methods under high sparsity, achieving up to 2.1times end-to-end speedup while maintaining high generation quality. Our code is open-sourced and available at https://github.com/jessica-hujie/DFSAttn.

In Diffusion Transformer (DiT) models, particularly for video generation, attention latency is a major bottleneck due to the long sequence length and the quadratic complexity. We find that attention weights can be separated into two parts: a small fraction of large weights with high rank and the remaining weights with very low rank. This naturally suggests applying sparse acceleration to the first part and low-rank acceleration to the second. Based on this finding, we propose SLA (Sparse-Linear Attention), a trainable attention method that fuses sparse and linear attention to accelerate diffusion models. SLA classifies attention weights into critical, marginal, and negligible categories, applying O(N^2) attention to critical weights, O(N) attention to marginal weights, and skipping negligible ones. SLA combines these computations into a single GPU kernel and supports both forward and backward passes. With only a few fine-tuning steps using SLA, DiT models achieve a 20x reduction in attention computation, resulting in significant acceleration without loss of generation quality. Experiments show that SLA reduces attention computation by 95% without degrading end-to-end generation quality, outperforming baseline methods. In addition, we implement an efficient GPU kernel for SLA, which yields a 13.7x speedup in attention computation and a 2.2x end-to-end speedup in video generation on Wan2.1-1.3B.

Speculative decoding speeds up autoregressive generation in Large Language Models (LLMs) through a two-step procedure, where a lightweight draft model proposes tokens which the target model then verifies in a single forward pass. Although the drafter network is small in modern architectures, its LM-head still performs projection to a large vocabulary, becoming one of the major computational bottlenecks. In prior work this issue has been predominantly addressed via static or dynamic vocabulary truncation. Yet mitigating the bottleneck, these methods bring in extra complexity, such as special vocabulary curation, sophisticated inference-time logic or modifications of the training setup. In this paper, we propose SlimSpec, a low-rank parameterization of the drafter's LM-head that compresses the inner representation rather than the output, preserving full vocabulary support. We evaluate our method with EAGLE-3 drafter across three target models and diverse benchmarks in both latency- and throughput-bound inference regimes. SlimSpec achieves 4-5times acceleration over the standard LM-head architecture while maintaining a competitive acceptance length, surpassing existing methods by up to 8-9% of the end-to-end speedup. Our method requires minimal adjustments of training and inference pipelines. Combined with the aforementioned speedup improvements, it makes SlimSpec a strong alternative across wide variety of draft LM-head architectures.

Advanced autoregressive (AR) video generation models have improved visual fidelity and interactivity, but the quadratic complexity of attention remains a primary bottleneck for efficient deployment. While existing sparse attention solutions have shown promise on bidirectional models, we identify that applying these solutions to AR models leads to considerable performance degradation for two reasons: isolated consideration of chunk generation and insufficient utilization of past informative context. Motivated by these observations, we propose Light Forcing, the first sparse attention solution tailored for AR video generation models. It incorporates a Chunk-Aware Growth mechanism to quantitatively estimate the contribution of each chunk, which determines their sparsity allocation. This progressive sparsity increase strategy enables the current chunk to inherit prior knowledge in earlier chunks during generation. Additionally, we introduce a Hierarchical Sparse Attention to capture informative historical and local context in a coarse-to-fine manner. Such two-level mask selection strategy (\ie, frame and block level) can adaptively handle diverse attention patterns. Extensive experiments demonstrate that our method outperforms existing sparse attention in quality (\eg, 84.5 on VBench) and efficiency (\eg, 1.2{sim}1.3times end-to-end speedup). Combined with FP8 quantization and LightVAE, Light Forcing further achieves a 2.3times speedup and 19.7\,FPS on an RTX~5090 GPU. Code will be released at https://github.com/chengtao-lv/LightForcing{https://github.com/chengtao-lv/LightForcing}.

Long-context inference in modern LLMs is increasingly constrained by decoding efficiency, especially in reasoning-heavy settings where models generate long intermediate chains of thought. Existing sparse attention methods often face a practical efficiency-quality trade-off. Structured block sparse methods typically provide stronger acceleration but incur noticeable quality loss, while token sparse methods are usually more accurate yet deliver limited end-to-end speedup because top-k routing over the full cache remains expensive. In this work, we propose cross-layer sparse attention (CLSA), which is built on top of KV-sharing architectures such as YOCO. The core idea is to share not only the KV cache across cross-decoder layers, but also the routing index. A single indexer computes token-level top-k selection once and reuses the resulting index across layers, thereby preserving the fine-grained selectivity of token sparse attention while amortizing the routing overhead. The resulting architecture improves all major inference bottlenecks jointly, including pre-filling, KV-cache storage, and long-context decoding. Experiments across short-context and long-context benchmarks show that CLSA is both accurate and efficient, achieving up to 7.6x decoding speedup and 17.1x overall throughput improvement at 128K context. These results suggest a more complete architectural solution for long-context LLMs that jointly advances model quality and inference efficiency.

Large language models (LLMs) deliver impressive performance but incur prohibitive memory and compute costs at deployment. Model pruning is an effective way to reduce these overheads, yet existing approaches face challenges: unstructured sparsity, where nonzeros can appear anywhere, preserves accuracy but yields irregular access patterns that prevent GPU acceleration, while semi-structured 2:4 sparsity is hardware-friendly but enforces a rigid 50% pattern that degrades model quality. To bridge this gap, we introduce PATCH, a hybrid sparsity framework that enables a continuous sparsity ratio between 0% and 50%. PATCH partitions weight matrices into tiles, assigning each tile to be either dense or 2:4 sparse via a learnable mask selection mechanism. This design provides fine-grained control over accuracy-acceleration tradeoffs and supports non-uniform sparsity across layers, leading to superior overall quality. Across models from 0.5B to 8B parameters, PATCH consistently narrows the gap to dense accuracy while delivering practical speedups. For instance, on LLaMA-2 7B with an A6000 GPU, PATCH achieves 1.18x-1.38x end-to-end speedup over dense baselines while improving accuracy by 0.37%-2.96% compared to the state-of-the-art 2:4 pruning method, MaskLLM.

Diffusion generative models have become the standard for producing high-quality, coherent video content, yet their slow inference speeds and high computational demands hinder practical deployment. Although both quantization and sparsity can independently accelerate inference while maintaining generation quality, naively combining these techniques in existing training-free approaches leads to significant performance degradation due to the lack of joint optimization. We introduce FPSAttention, a novel training-aware co-design of FP8 quantization and sparsity for video generation, with a focus on the 3D bi-directional attention mechanism. Our approach features three key innovations: 1) A unified 3D tile-wise granularity that simultaneously supports both quantization and sparsity; 2) A denoising step-aware strategy that adapts to the noise schedule, addressing the strong correlation between quantization/sparsity errors and denoising steps; 3) A native, hardware-friendly kernel that leverages FlashAttention and is implemented with optimized Hopper architecture features for highly efficient execution. Trained on Wan2.1's 1.3B and 14B models and evaluated on the VBench benchmark, FPSAttention achieves a 7.09x kernel speedup for attention operations and a 4.96x end-to-end speedup for video generation compared to the BF16 baseline at 720p resolution-without sacrificing generation quality.

Many sparse attention mechanisms such as Neighborhood Attention have typically failed to consistently deliver speedup over the self attention baseline. This is largely due to the level of complexity in attention infrastructure, and the rapid evolution of AI hardware architecture. At the same time, many state-of-the-art foundational models, particularly in computer vision, are heavily bound by attention, and need reliable sparsity to escape the O(n^2) complexity. In this paper, we study a class of promising sparse attention mechanisms that focus on locality, and aim to develop a better analytical model of their performance improvements. We first introduce Generalized Neighborhood Attention (GNA), which can describe sliding window, strided sliding window, and blocked attention. We then consider possible design choices in implementing these approaches, and create a simulator that can provide much more realistic speedup upper bounds for any given setting. Finally, we implement GNA on top of a state-of-the-art fused multi-headed attention (FMHA) kernel designed for the NVIDIA Blackwell architecture in CUTLASS. Our implementation can fully realize the maximum speedup theoretically possible in many perfectly block-sparse cases, and achieves an effective utilization of 1.3 petaFLOPs/second in FP16. In addition, we plug various GNA configurations into off-the-shelf generative models, such as Cosmos-7B, HunyuanVideo, and FLUX, and show that it can deliver 28% to 46% end-to-end speedup on B200 without any fine-tuning. We will open source our simulator and Blackwell kernels directly through the NATTEN project.

Denoising generative models deliver high-fidelity generation but remain bottlenecked by inference latency due to the many iterative denoiser calls required during sampling. Training-free acceleration methods reduce latency by either sparsifying the model architecture or shortening the sampling trajectory. Current training-free acceleration methods are more complex than necessary: higher-order predictors amplify error under aggressive speedups, and architectural modifications hinder deployment. Beyond 2x acceleration, step skipping creates structural scarcity -- at most one fresh evaluation per local window -- leaving the computed output and its backward difference as the only causally grounded information. Based on this, we propose ZEUS, an acceleration method that predicts reduced denoiser evaluations using a second-order predictor, and stabilizes aggressive consecutive skipping with an interleaved scheme that avoids back-to-back extrapolations. ZEUS adds essentially zero overhead, no feature caches, and no architectural modifications, and it is compatible with different backbones, prediction objectives, and solver choices. Across image and video generation, ZEUS consistently improves the speed-fidelity performance over recent training-free baselines, achieving up to 3.2x end-to-end speedup while maintaining perceptual quality. Our code is available at: https://github.com/Ting-Justin-Jiang/ZEUS.

Large Language Models (LLMs) demonstrate exceptional capability across diverse tasks. However, their deployment in long-context scenarios is hindered by two challenges: computational inefficiency and redundant information. We propose RAM (Read As HuMan), a context compression framework that adopts an adaptive hybrid reading strategy, to address these challenges. Inspired by human reading behavior (i.e., close reading important content while skimming less relevant content), RAM partitions the context into segments and encodes them with the input query in parallel. High-relevance segments are fully retained (close reading), while low-relevance ones are query-guided compressed into compact summary vectors (skimming). Both explicit textual segments and implicit summary vectors are concatenated and fed into decoder to achieve both superior performance and natural language format interpretability. To refine the decision boundary between close reading and skimming, we further introduce a contrastive learning objective based on positive and negative query-segment pairs. Experiments demonstrate that RAM outperforms existing baselines on multiple question answering and summarization benchmarks across two backbones, while delivering up to a 12x end-to-end speedup on long inputs (average length 16K; maximum length 32K).

Large Language Models (LLMs), constrained by their auto-regressive nature, suffer from slow decoding. Speculative decoding methods have emerged as a promising solution to accelerate LLM decoding, attracting attention from both systems and AI research communities. Recently, the pursuit of better draft quality has driven a trend toward parametrically larger draft models, which inevitably introduces substantial computational overhead. While existing work attempts to balance the trade-off between prediction accuracy and compute latency, we address this fundamental dilemma through architectural innovation. We propose PRISM, which disaggregates the computation of each predictive step across different parameter sets, refactoring the computational pathways of draft models to successfully decouple model capacity from inference cost. Through extensive experiments, we demonstrate that PRISM outperforms all existing draft architectures, achieving exceptional acceptance lengths while maintaining minimal draft latency for superior end-to-end speedup. We also re-examine scaling laws with PRISM, revealing that PRISM scales more effectively with expanding data volumes than other draft architectures. Through rigorous and fair comparison, we show that PRISM boosts the decoding throughput of an already highly optimized inference engine by more than 2.6x.

Large reasoning models (LRMs) achieve state-of-the-art performance on challenging benchmarks by generating long chains of intermediate steps, but their inference cost is dominated by decoding, where each new token must attend to the entire growing sequence. Existing sparse attention methods reduce computation by pruning the key-value (KV) cache, yet they suffer from severe accuracy degradation on reasoning tasks due to cumulative selection errors and the dynamic importance of tokens over long derivations. We present DELTA, a training-free sparse attention mechanism that achieves computational efficiency without sacrificing model accuracy. DELTA partitions transformer layers into three groups: initial layers that use full attention, a small set of selection layers that identify salient tokens via aggregated head-level attention scores, and subsequent sparse-attention layers that attend only to the selected subset. This design preserves the full KV cache in GPU memory for accuracy, while avoiding expensive full-attention computation over many layers. On reasoning benchmarks such as AIME and GPQA-Diamond, DELTA matches or surpasses full attention in accuracy, while reducing the number of attended tokens by up to 5times and delivering 1.5times end-to-end speedup. Our results show that selective reuse of intermediate attention maps offers a robust path toward efficient long-context reasoning.

Text-to-image generation models, especially Multimodal Diffusion Transformers (MMDiT), have shown remarkable progress in generating high-quality images. However, these models often face significant computational bottlenecks, particularly in attention mechanisms, which hinder their scalability and efficiency. In this paper, we introduce DiTFastAttnV2, a post-training compression method designed to accelerate attention in MMDiT. Through an in-depth analysis of MMDiT's attention patterns, we identify key differences from prior DiT-based methods and propose head-wise arrow attention and caching mechanisms to dynamically adjust attention heads, effectively bridging this gap. We also design an Efficient Fused Kernel for further acceleration. By leveraging local metric methods and optimization techniques, our approach significantly reduces the search time for optimal compression schemes to just minutes while maintaining generation quality. Furthermore, with the customized kernel, DiTFastAttnV2 achieves a 68% reduction in attention FLOPs and 1.5x end-to-end speedup on 2K image generation without compromising visual fidelity.

Given the success of deep learning in classical machine learning, quantum algorithms for traditional neural network architectures may provide one of the most promising settings for quantum machine learning. Considering a fully-connected feedforward neural network, we show that conditions amenable to classical trainability via gradient descent coincide with those necessary for efficiently solving quantum linear systems. We propose a quantum algorithm to approximately train a wide and deep neural network up to O(1/n) error for a training set of size n by performing sparse matrix inversion in O(log n) time. To achieve an end-to-end exponential speedup over gradient descent, the data distribution must permit efficient state preparation and readout. We numerically demonstrate that the MNIST image dataset satisfies such conditions; moreover, the quantum algorithm matches the accuracy of the fully-connected network. Beyond the proven architecture, we provide empirical evidence for O(log n) training of a convolutional neural network with pooling.

FP8 training has emerged as a promising method for improving training efficiency. Existing frameworks accelerate training by applying FP8 computation to linear layers while leaving optimizer states and activations in higher precision, which fails to fully optimize memory usage. This paper introduces COAT (Compressing Optimizer States and Activations for FP8 Training), a novel FP8 training framework designed to significantly reduce memory footprint when training large models. COAT addresses current limitations through two key innovations: (1) Dynamic Range Expansion, which aligns optimizer state distributions more closely with the FP8 representation range, thereby reducing quantization error, and (2) Mixed-Granularity Activation Quantization, which optimizes activation memory using a combination of per-tensor and per-group quantization strategies. Experiments demonstrate that COAT effectively reduces end-to-end training memory footprint by 1.54x compared to BF16 while achieving nearly lossless performance across various tasks, such as Large Language Model pretraining and fine-tuning and Vision Language Model training. COAT also achieves a 1.43x end-to-end training speedup compared to BF16, performing on par with or surpassing TransformerEngine's speedup. COAT enables efficient full-parameter training of large models on fewer GPUs, and facilitates doubling the batch size in distributed training settings, providing a practical solution for scaling large-scale model training. The code is available at https://github.com/NVlabs/COAT.

The quadratic complexity of the attention mechanism and the substantial memory footprint of the Key-Value (KV) cache present severe computational and memory challenges for Large Language Models (LLMs) processing long contexts. Existing retrieval-based methods often compromise semantic integrity through fixed-size chunking and suffer from inefficient linear scanning. In this paper, we propose LycheeCluster, a novel method for efficient KV cache management. LycheeCluster preserves local semantic coherence via boundary-aware chunking and constructs a recursive hierarchical index rooted in the triangle inequality. This design transforms cache retrieval from a linear scan into a theoretically bounded, logarithmic-time pruning process, while a lazy update strategy supports efficient streaming generation. Experiments demonstrate that LycheeCluster achieves up to a 3.6x end-to-end inference speedup with negligible degradation in model performance, outperforming state-of-the-art KV cache management methods (e.g., Quest, ClusterKV). We will release our code and kernels after publication.

Transformer models have achieved remarkable success across various AI applications but face significant training costs. Low-bit training, such as INT8 training, can leverage computational units with higher throughput, and has already demonstrated its effectiveness on GPT2 models with block-level quantization. However, it struggles with modern Transformer variants incorporating GLU units. This is because those variants demonstrate complex distributions of activation outliers. To address the challenge, we propose Fallback Quantization, implementing mixed-precision GEMM that dynamically falls back 8-bit to 16-bit for activation blocks containing outliers. Experiments show that our approach is robustly competent in both fine-tuning and pretraining settings. Moreover, our method achieves a 1.57x end-to-end training speedup on RTX4090 GPUs.

Recent progress in sparse attention mechanisms has demonstrated strong potential for reducing the computational cost of long-context training and inference in large language models (LLMs). Native Sparse Attention (NSA), a state-of-the-art approach, introduces natively trainable, hardware-aligned sparse attention that delivers substantial system-level performance gains while maintaining accuracy comparable to full attention. However, the kernel implementation of NSA relies on a query-grouping strategy that is efficient only with large Grouped Query Attention (GQA) sizes, whereas modern LLMs typically adopt much smaller GQA groups, which limits the applicability of this sparse algorithmic advance. In this work, we propose Flash Sparse Attention (FSA), which includes an alternative kernel design that enables efficient NSA computation across a wide range of popular LLMs with varied smaller GQA group sizes on modern GPUs. Compared to vanilla NSA kernel implementation, our empirical evaluation demonstrates that FSA achieves (i) up to 3.5times and on average 1.6times kernel-level latency reduction, (ii) up to 1.25times and 1.09times on average end-to-end training speedup on state-of-the-art LLMs, and (iii) up to 1.36times and 1.11times on average end-to-end prefill speedup on state-of-the-art LLMs. The source code is open-sourced and publicly available at https://github.com/Relaxed-System-Lab/Flash-Sparse-Attention.

Vision-Language-Action (VLA) models provide a unified paradigm for robotic manipulation, yet their real-world deployment is often bottlenecked by execution efficiency. While existing efforts predominantly focus on compute-centric efficiency to reduce per-step inference latency, the intrinsic policy efficiency of these models remains largely unexplored. Policy efficiency is fundamentally affected by two factors, namely the effective executable length of predicted action chunks and the total physical steps required to complete a task. These two factors jointly determine the total number of forward inference calls during execution. We observe that current VLA policies struggle with planning unreliability and action redundancy, suffering from severe prediction degradation at the tail of action chunks and tending to generate unnecessarily redundant physical steps. To address this, we propose PolicyTrim, a reinforcement learning-based post-training framework that extends the reliable action chunk length and reduces redundant physical steps. For reliable chunk extension, we employ a dynamic exploration strategy that explicitly rewards the successful completion of longer executable lengths, progressively pushing the trustworthy prediction horizon to its empirical limit. For step efficiency, we design a redundancy-aware reward that directly favors successful task completions with fewer steps while penalizing unreproducible shortcuts, effectively eliminating redundant physical actions. Extensive experiments across three benchmarks and three VLA models demonstrate that PolicyTrim improves action chunk utilization by 3times and reduces physical execution steps by 51.4\%. Ultimately, our framework delivers up to a 5.83times end-to-end deployment speedup without compromising task success rates.

Diffusion Large Language Models (dLLMs) have recently emerged as a promising alternative to autoregressive models, offering competitive performance while naturally supporting parallel decoding. However, as dLLMs are increasingly integrated with Mixture-of-Experts (MoE) architectures to scale model capacity, a fundamental mismatch arises between block parallel decoding and token-level expert selection. Specifically, each dLLM forward pass processes multiple tokens with bidirectional dependencies, whereas conventional MoE layers route each token independently. This mismatch substantially increases the number of uniquely activated experts, making inference increasingly memory-bound. To address this, we propose dMoE, a simple yet effective block-level MoE framework. The central idea of dMoE is to aggregate token-level expert distributions within each block into a unified block-level expert distribution, which is then used to guide expert routing in a more coherent manner. In this way, dMoE substantially reduces the number of uniquely activated experts during inference without sacrificing performance, thereby mitigating the memory-bound bottleneck. Extensive experiments across a variety of benchmarks demonstrate the effectiveness of dMoE. On average, dMoE reduces the number of uniquely activated experts from 69.5 to 14.6 while retaining 99.11% of the original performance. Meanwhile, it reduces memory usage by 76.64% to 79.84% and achieves 1.14times to 1.66times end-to-end latency speedup. Code is available at: https://github.com/fscdc/dMoE

Vision-language models (VLMs) predominantly rely on autoregressive decoding, which generates tokens one at a time and fundamentally limits inference throughput. This limitation is especially acute in physical AI scenarios such as robotics and autonomous driving, where VLMs are deployed on edge devices at batch size one, making AR decoding memory-bandwidth-bound and leaving hardware parallelism underutilized. While block-wise discrete diffusion has shown promise for parallel text generation, extending it to VLMs remains challenging due to the need to jointly handle continuous visual representations and discrete text tokens while preserving pretrained multimodal capabilities. We present Fast-dVLM, a block-diffusion-based VLM that enables KV-cache-compatible parallel decoding and speculative block decoding for inference acceleration. We systematically compare two AR-to-diffusion conversion strategies: a two-stage approach that first adapts the LLM backbone with text-only diffusion fine-tuning before multimodal training, and a direct approach that converts the full AR VLM in one stage. Under comparable training budgets, direct conversion proves substantially more efficient by leveraging the already multimodally aligned VLM; we therefore adopt it as our recommended recipe. We introduce a suite of multimodal diffusion adaptations, block size annealing, causal context attention, auto-truncation masking, and vision efficient concatenation, that collectively enable effective block diffusion in the VLM setting. Extensive experiments across 11 multimodal benchmarks show Fast-dVLM matches its autoregressive counterpart in generation quality. With SGLang integration and FP8 quantization, Fast-dVLM achieves over 6x end-to-end inference speedup over the AR baseline.

Modern RL post-training methods such as GRPO and DAPO train on N response sequences of R tokens sampled from a shared prompt of P tokens, but standard FlashAttention replicates all P prompt tokens N times across both forward and backward passes -- duplicating compute and memory on identical hidden states. In large-rollout, long-context RL training (N{geq}16, P{geq}8K), this redundancy dominates the policy update cost. We observe that in decoder-only models, causal masking makes prompt representations invariant across sequences at every layer, so all per-token operations (norms, projections, MLP) and attention can process the prompt once -- a property not yet exploited at the kernel level for training. We propose DualKV, the first FlashAttention kernel variant that eliminates shared-prompt replication during RL training, via (1)~fused CUDA forward and backward kernels that iterate over two disjoint KV regions -- shared context and per-sequence response -- in a single kernel launch, and (2)~a data-pipeline redesign in veRL that repacks N(P{+}R) tokens into P{+}NR tokens per micro-batch, extending the token reduction from attention to the entire model by a factor ρ= N(P{+}R)/(P{+}NR). DualKV is mathematically equivalent to standard attention and introduces no approximation. On Qwen3-8B GRPO training with 8timesH100 GPUs (N{=}32, 8K-context), DualKV achieves 1.63--2.09times policy-update speedup, enables 2times larger micro-batches, and raises MFU from 36% to 76%. Similar gains hold for DAPO (2.47times speedup, 77% MFU). At 30B MoE scale on 16timesH100, DualKV achieves 3.82times policy-update and 3.38times end-to-end step speedup over FlashAttention (which requires 4-way Ulysses sequence parallelism to avoid OOM).

Speculative decoding can significantly accelerate LLM serving, yet most deployments today disentangle speculator training from serving, treating speculator training as a standalone offline modeling problem. We show that this decoupled formulation introduces substantial deployment and adaptation lag: (1) high time-to-serve, since a speculator must be trained offline for a considerable period before deployment; (2) delayed utility feedback, since the true end-to-end decoding speedup is only known after training and cannot be inferred reliably from acceptance rate alone due to model-architecture and system-level overheads; and (3) domain-drift degradation, as the target model is repurposed to new domains and the speculator becomes stale and less effective. To address these issues, we present Aurora, a unified training-serving system that closes the loop by continuously learning a speculator directly from live inference traces. Aurora reframes online speculator learning as an asynchronous reinforcement-learning problem: accepted tokens provide positive feedback, while rejected speculator proposals provide implicit negative feedback that we exploit to improve sample efficiency. Our design integrates an SGLang-based inference server with an asynchronous training server, enabling hot-swapped speculator updates without service interruption. Crucially, Aurora supports day-0 deployment: a speculator can be served immediately and rapidly adapted to live traffic, improving system performance while providing immediate utility feedback. Across experiments, Aurora achieves a 1.5x day-0 speedup on recently released frontier models (e.g., MiniMax M2.1 229B and Qwen3-Coder-Next 80B). Aurora also adapts effectively to distribution shifts in user traffic, delivering an additional 1.25x speedup over a well-trained but static speculator on widely used models (e.g., Qwen3 and Llama3).

Large reasoning models (LRMs) improve complex problem-solving by generating long intermediate reasoning traces, but this substantially increases inference costs. NVFP4 inference offers a promising approach to reduce both computational and memory costs through hardware-supported low-precision execution. However, directly applying NVFP4 to LRMs introduces two practical limitations: reasoning accuracy degrades under quantization, and existing NVFP4 kernels do not fully realize latency benefits in small-batch autoregressive decoding. In this work, we analyze the effect of NVFP4 quantization on token-level uncertainty during reasoning. We show that quantization increases incorrect sampling at low-entropy symbolic tokens, while causing over-concentration on a small set of tokens in high-uncertainty reasoning steps. Based on this observation, we propose ReSET, a reasoning-step entropy-based temperature-scaling method that estimates step-level uncertainty online and adapts the decoding temperature using both token-level and step-level entropy signals. To address the latency gap, we further design a CUDA-core small-M NVFP4 kernel for latency-critical autoregressive decoding. Across reasoning benchmarks and model scales, ReSET improves NVFP4 reasoning accuracy by up to sim!2 points over the NVFP4 baseline. Our CUDA-core small-M kernel further improves latency-critical decoding, delivering up to 2.5!times kernel-level speedup over NVFP4 vLLM and approximately 2!times end-to-end decoding speedup over BF16. Code is available at https://github.com/aiha-lab/ReSET.

Recursive or looped language models have recently emerged as a new scaling axis by iteratively refining the same model computation over latent states to deepen reasoning. We extend such scaling principle from a single model to multi-agent systems, and ask: Can agent collaboration itself be scaled through recursion? To this end, we introduce RecursiveMAS, a recursive multi-agent framework that casts the entire system as a unified latent-space recursive computation. RecursiveMAS connects heterogeneous agents as a collaboration loop through the lightweight RecursiveLink module, enabling in-distribution latent thoughts generation and cross-agent latent state transfer. To optimize our framework, we develop an inner-outer loop learning algorithm for iterative whole-system co-optimization through shared gradient-based credit assignment across recursion rounds. Theoretical analyses of runtime complexity and learning dynamics establish that RecursiveMAS is more efficient than standard text-based MAS and maintains stable gradients during recursive training. Empirically, we instantiate RecursiveMAS under 4 representative agent collaboration patterns and evaluate across 9 benchmarks spanning mathematics, science, medicine, search, and code generation. In comparison with advanced single/multi-agent and recursive computation baselines, RecursiveMAS consistently delivers an average accuracy improvement of 8.3%, together with 1.2times-2.4times end-to-end inference speedup, and 34.6%-75.6% token usage reduction. Code and Data are provided in https://recursivemas.github.io.

Mixture-of-Experts (MoE) scales language models efficiently through sparse expert activation, and its dynamic variant further reduces computation by adjusting the activated experts in an input-dependent manner. Existing dynamic MoE methods usually rely on pre-training from scratch or task-specific adaptation, leaving the practical conversion of fully trained MoE underexplored. Enabling such adaptation would directly alleviate the inference costs by allowing easy tokens to bypass unnecessary expert during serving. This paper introduces Zero-Expert Self-Distillation Adaptation (ZEDA), a low-cost framework that transforms post-trained static MoE models into efficient dynamic ones. To stabilize this architectural conversion, ZEDA injects parameter-free zero-output experts into each MoE layer and adapts the augmented model through two-stage self-distillation, utilizing the original MoE as a frozen teacher and applying a group-level balancing loss. On Qwen3-30B-A3B and GLM-4.7-Flash across 11 benchmarks spanning math, code, and instruction following, ZEDA eliminates over 50% of expert FLOPs at marginal accuracy loss. It outperforms the strongest dynamic MoE baseline by 6.1 and 4.0 points on the two models, and delivers ~1.20times end-to-end inference speedup.

RL post-training of frontier language models is increasingly bottlenecked by autoregressive rollout generation, making rollout acceleration a central systems challenge. Many existing efficiency methods improve throughput by changing the rollout or optimization regime, for example, through off-policy execution, replay, or lower-precision generation. We study speculative decoding as a lossless acceleration primitive for RL rollouts that preserves the target model's output distribution. We implement speculative decoding in NeMo-RL with a vLLM backend, supporting both synchronous and asynchronous pipelines and enabling speculation during RL rollouts. This benefit is realizable across speculation mechanisms, such as pretrained MTP heads, small external draft models or even techniques such as Eagle3, which are traditionally applied after RL phase. This yields a deployment path for state-of-the-art speculative decoding inside RL training. In a reasoning post-training workload at 8B scale under synchronous RL, speculative decoding improves rollout throughput by 1.8x. Using a high-fidelity performance simulator, we project that combining speculative decoding with asynchronous RL yields up to 2.5x end-to-end training speedup at 235B scale.

Large language models incur high inference latency due to sequential autoregressive decoding. Speculative decoding alleviates this bottleneck by using a lightweight draft model to propose multiple tokens for batched verification. However, its adoption has been limited by the lack of high-quality draft models and scalable training infrastructure. We introduce SpecForge, an open-source, production-oriented framework for training speculative decoding models with full support for EAGLE-3. SpecForge incorporates target-draft decoupling, hybrid parallelism, optimized training kernels, and integration with production-grade inference engines, enabling up to 9.9x faster EAGLE-3 training for Qwen3-235B-A22B. In addition, we release SpecBundle, a suite of production-grade EAGLE-3 draft models trained with SpecForge for mainstream open-source LLMs. Through a systematic study of speculative decoding training recipes, SpecBundle addresses the scarcity of high-quality drafts in the community, and our draft models achieve up to 4.48x end-to-end inference speedup on SGLang, establishing SpecForge as a practical foundation for real-world speculative decoding deployment.

As long-context inference becomes central to large language models (LLMs), attention over growing key-value caches emerges as a dominant decoding bottleneck, motivating sparse attention for scalable inference. Fixed-budget top-k sparse attention cannot adapt to heterogeneous attention distributions across heads and layers, whereas top-p sparse attention directly preserves attention mass and provides stronger accuracy guarantees. Existing top-p methods, however, fail to jointly optimize top-p accuracy, selection overhead, and sparse attention cost, which limits their overall efficiency. We present Double-P, a hierarchical sparse attention framework that optimizes all three stages. Double-P first performs coarse-grained top-p estimation at the cluster level using size-weighted centroids, then adaptively refines computation through a second top-p stage that allocates token-level attention only when needed. Across long-context benchmarks, Double-P consistently achieves near-zero accuracy drop, reducing attention computation overhead by up to 1.8x and delivers up to 1.3x end-to-end decoding speedup over state-of-the-art fixed-budget sparse attention methods.

Pretraining transformers are generally time-consuming. Fully quantized training (FQT) is a promising approach to speed up pretraining. However, most FQT methods adopt a quantize-compute-dequantize procedure, which often leads to suboptimal speedup and significant performance degradation when used in transformers due to the high memory access overheads and low-precision computations. In this work, we propose Jetfire, an efficient and accurate INT8 training method specific to transformers. Our method features an INT8 data flow to optimize memory access and a per-block quantization method to maintain the accuracy of pretrained transformers. Extensive experiments demonstrate that our INT8 FQT method achieves comparable accuracy to the FP16 training baseline and outperforms the existing INT8 training works for transformers. Moreover, for a standard transformer block, our method offers an end-to-end training speedup of 1.42x and a 1.49x memory reduction compared to the FP16 baseline.

We present Mach-Mind-4-Flash, a 35B-parameter Mixture-of-Experts (MoE) agentic model with 3B activated parameters. Through post-training optimization alone without scaling pre-training compute, the model achieves performance on par with or surpassing that of 100B-parameter-class models. By introducing scalable agentic interaction environments for large-scale reinforcement learning, the model attains significant performance gains on real-world application tasks. Our pipeline comprises three stages: (1) a unified RL/OPD training infrastructure with dynamic multi-teacher scheduling and operator-level acceleration, delivering 17\% end-to-end training speedup; (2) multiple domain-specific RL experts trained in parallel across Reasoning, General, and Agent tracks, then fused into a single generalist via Multi-Teacher On-Policy Distillation (MOPD) -- a routed reverse-KL objective that eliminates the see-saw degradation of mixed-reward RL; (3) Hybrid Median-length Policy Optimization (HMPO), a single-stage token-efficiency method that compresses reasoning chains by 19--46\% with le0.7 percentage-point accuracy loss. Mach-Mind-4-Flash scores 92.70 on AIME'26, 82.82 on IFBench, 80.74 on Behavioral-SafetyBench, 75.80 on BFCL-v4, 72.31 on BrowseComp-zh, and 84.20 on ClawBench -- leading or matching models with 10--30times its activated size at a fraction of the inference cost.

In visual generation, the quadratic complexity of attention mechanisms results in high memory and computational costs, especially for longer token sequences required in high-resolution image or multi-frame video generation. To address this, prior research has explored techniques such as sparsification and quantization. However, these techniques face significant challenges under low density and reduced bitwidths. Through systematic analysis, we identify that the core difficulty stems from the dispersed and irregular characteristics of visual attention patterns. Therefore, instead of introducing specialized sparsification and quantization design to accommodate such patterns, we propose an alternative strategy: *reorganizing* the attention pattern to alleviate the challenges. Inspired by the local aggregation nature of visual feature extraction, we design a novel **Pattern-Aware token ReOrdering (PARO)** technique, which unifies the diverse attention patterns into a hardware-friendly block-wise pattern. This unification substantially simplifies and enhances both sparsification and quantization. We evaluate the performance-efficiency trade-offs of various design choices and finalize a methodology tailored for the unified pattern. Our approach, **PAROAttention**, achieves video and image generation with lossless metrics, and nearly identical results from full-precision (FP) baselines, while operating at notably lower density (~20%-30%) and bitwidth (**INT8/INT4**), achieving a **1.9x** to **2.7x** end-to-end latency speedup.

Reinforcement learning (RL) has demonstrated immense potential in advancing artificial general intelligence, agentic intelligence, and embodied intelligence. However, the inherent heterogeneity and dynamicity of RL workflows often lead to low hardware utilization and slow training on existing systems. In this paper, we present RLinf, a high-performance RL training system based on our key observation that the major roadblock to efficient RL training lies in system flexibility. To maximize flexibility and efficiency, RLinf is built atop a novel RL system design paradigm called macro-to-micro flow transformation (M2Flow), which automatically breaks down high-level, easy-to-compose RL workflows at both the temporal and spatial dimensions, and recomposes them into optimized execution flows. Supported by RLinf worker's adaptive communication capability, we devise context switching and elastic pipelining to realize M2Flow transformation, and a profiling-guided scheduling policy to generate optimal execution plans. Extensive evaluations on both reasoning RL and embodied RL tasks demonstrate that RLinf consistently outperforms state-of-the-art systems, achieving 1.1x-2.13x speedup in end-to-end training throughput.

The emergence of Large Language Models (LLMs) with strong reasoning capabilities marks a significant milestone, unlocking new frontiers in complex problem-solving. However, training these reasoning models, typically using Reinforcement Learning (RL), encounters critical efficiency bottlenecks: response generation during RL training exhibits a persistent long-tail distribution, where a few very long responses dominate execution time, wasting resources and inflating costs. To address this, we propose TLT, a system that accelerates reasoning RL training losslessly by integrating adaptive speculative decoding. Applying speculative decoding in RL is challenging due to the dynamic workloads, evolving target model, and draft model training overhead. TLT overcomes these obstacles with two synergistic components: (1) Adaptive Drafter, a lightweight draft model trained continuously on idle GPUs during long-tail generation to maintain alignment with the target model at no extra cost; and (2) Adaptive Rollout Engine, which maintains a memory-efficient pool of pre-captured CUDAGraphs and adaptively select suitable SD strategies for each input batch. Evaluations demonstrate that TLT achieves over 1.7x end-to-end RL training speedup over state-of-the-art systems, preserves the model accuracy, and yields a high-quality draft model as a free byproduct suitable for efficient deployment. Code is released at https://github.com/mit-han-lab/fastrl.

An efficient attention implementation is essential for large models due to its quadratic time complexity. Fortunately, attention commonly exhibits sparsity, i.e., many values in the attention map are near zero, allowing for the omission of corresponding computations. Many studies have utilized the sparse pattern to accelerate attention. However, most existing works focus on optimizing attention within specific models by exploiting certain sparse patterns of the attention map. A universal sparse attention that guarantees both the speedup and end-to-end performance of diverse models remains elusive. In this paper, we propose SpargeAttn, a universal sparse and quantized attention for any model. Our method uses a two-stage online filter: in the first stage, we rapidly and accurately predict the attention map, enabling the skip of some matrix multiplications in attention. In the second stage, we design an online softmax-aware filter that incurs no extra overhead and further skips some matrix multiplications. Experiments show that our method significantly accelerates diverse models, including language, image, and video generation, without sacrificing end-to-end metrics. The codes are available at https://github.com/thu-ml/SpargeAttn.

Transformers are slow and memory-hungry on long sequences, since the time and memory complexity of self-attention are quadratic in sequence length. Approximate attention methods have attempted to address this problem by trading off model quality to reduce the compute complexity, but often do not achieve wall-clock speedup. We argue that a missing principle is making attention algorithms IO-aware -- accounting for reads and writes between levels of GPU memory. We propose FlashAttention, an IO-aware exact attention algorithm that uses tiling to reduce the number of memory reads/writes between GPU high bandwidth memory (HBM) and GPU on-chip SRAM. We analyze the IO complexity of FlashAttention, showing that it requires fewer HBM accesses than standard attention, and is optimal for a range of SRAM sizes. We also extend FlashAttention to block-sparse attention, yielding an approximate attention algorithm that is faster than any existing approximate attention method. FlashAttention trains Transformers faster than existing baselines: 15% end-to-end wall-clock speedup on BERT-large (seq. length 512) compared to the MLPerf 1.1 training speed record, 3times speedup on GPT-2 (seq. length 1K), and 2.4times speedup on long-range arena (seq. length 1K-4K). FlashAttention and block-sparse FlashAttention enable longer context in Transformers, yielding higher quality models (0.7 better perplexity on GPT-2 and 6.4 points of lift on long-document classification) and entirely new capabilities: the first Transformers to achieve better-than-chance performance on the Path-X challenge (seq. length 16K, 61.4% accuracy) and Path-256 (seq. length 64K, 63.1% accuracy).

Context-augmented generation (CAG) techniques, including RAG and ICL, require the efficient combination of multiple contexts to generate responses to user queries. Directly inputting these contexts as a sequence introduces a considerable computational burden by re-encoding the combined selection of contexts for every request. To address this, we explore the promising potential of parallel encoding to independently pre-compute and cache each context's KV states. This approach enables the direct loading of cached states during inference while accommodating more contexts through position reuse across contexts. However, due to misalignments in attention distribution, directly applying parallel encoding results in a significant performance drop. To enable effective and efficient CAG, we propose Adaptive Parallel Encoding (APE), which brings shared prefix, attention temperature, and scaling factor to align the distribution of parallel encoding with sequential encoding. Results on RAG and ICL tasks demonstrate that APE can preserve 98% and 93% sequential encoding performance using the same inputs while outperforming parallel encoding by 3.6% and 7.9%, respectively. It also scales to many-shot CAG, effectively encoding hundreds of contexts in parallel. Efficiency evaluation shows that APE can achieve an end-to-end 4.5times speedup by reducing 28times prefilling time for a 128K-length context.

Convolution models with long filters have demonstrated state-of-the-art reasoning abilities in many long-sequence tasks but lag behind the most optimized Transformers in wall-clock time. A major bottleneck is the Fast Fourier Transform (FFT)--which allows long convolutions to run in O(N logN) time in sequence length N but has poor hardware utilization. In this paper, we study how to optimize the FFT convolution. We find two key bottlenecks: the FFT does not effectively use specialized matrix multiply units, and it incurs expensive I/O between layers of the memory hierarchy. In response, we propose FlashFFTConv. FlashFFTConv uses a matrix decomposition that computes the FFT using matrix multiply units and enables kernel fusion for long sequences, reducing I/O. We also present two sparse convolution algorithms--1) partial convolutions and 2) frequency-sparse convolutions--which can be implemented simply by skipping blocks in the matrix decomposition, enabling further opportunities for memory and compute savings. FlashFFTConv speeds up exact FFT convolutions by up to 7.93times over PyTorch and achieves up to 4.4times speedup end-to-end. Given the same compute budget, FlashFFTConv allows Hyena-GPT-s to achieve 2.3 points better perplexity on the PILE and M2-BERT-base to achieve 3.3 points higher GLUE score--matching models with twice the parameter count. FlashFFTConv also achieves 96.1% accuracy on Path-512, a high-resolution vision task where no model had previously achieved better than 50%. Furthermore, partial convolutions enable longer-sequence models--yielding the first DNA model that can process the longest human genes (2.3M base pairs)--and frequency-sparse convolutions speed up pretrained models while maintaining or improving model quality.

Despite their strong performance on reasoning tasks, large reasoning models (LRMs) often suffer from overthinking, producing unnecessarily long outputs and incurring high end-to-end latency, a significant limitation to their real-world deployment. To address overthinking, early-exit mechanisms have been proposed to terminate reasoning before typical completion, showing that this approach can effectively shorten generation length with minimal impact on accuracy. However, their reliance on probing mechanisms introduces a detection overhead that limits their end-to-end latency gains and compromises their generalizability across diverse problems. Inspired by the use of hidden states in speculative decoding, we propose SpecExit, a novel framework that predicts both future tokens and an early-exit signal directly from a lightweight draft model without probing overhead. Our method offers significant improvements, reducing average generation length by 66\% and achieving a 2.5x speedup in end-to-end latency compared to the speculative decoding baseline, without compromising accuracy. Our method leverages the inherent signals from hidden states to provide effective early-exit signals, suggesting broader use of hidden states for efficient reasoning. Our code is available at https://github.com/Tencent/AngelSlim.

The scaling of Large Language Models (LLMs) drives interest in matrix-based optimizers (e.g., Shampoo, Muon, SOAP) for their convergence efficiency; yet their requirement for holistic updates conflicts with the tensor fragmentation in distributed frameworks like Megatron. Existing solutions are suboptimal: synchronous approaches suffer from computational redundancy, while layer-wise partitioning fails to reconcile this conflict without violating the geometric constraints of efficient communication primitives. To bridge this gap, we propose Canzona, a Unified, Asynchronous, and Load-Balanced framework that decouples logical optimizer assignment from physical parameter distribution. For Data Parallelism, we introduce an alpha-Balanced Static Partitioning strategy that respects atomicity while neutralizing the load imbalance. For Tensor Parallelism, we design an Asynchronous Compute pipeline utilizing Micro-Group Scheduling to batch fragmented updates and hide reconstruction overhead. Extensive evaluations on the Qwen3 model family (up to 32B parameters) on 256 GPUs demonstrate that our approach preserves the efficiency of established parallel architectures, achieving a 1.57x speedup in end-to-end iteration time and reducing optimizer step latency by 5.8x compared to the baseline.

Large language models (LLMs) have achieved impressive performance in a variety of natural language processing (NLP) tasks. However, when applied to long-context scenarios, they face two challenges, i.e., low computational efficiency and much redundant information. This paper introduces GMSA, a context compression framework based on the encoder-decoder architecture, which addresses these challenges by reducing input sequence length and redundant information. Structurally, GMSA has two key components: Group Merging and Layer Semantic Alignment (LSA). Group merging is used to effectively and efficiently extract summary vectors from the original context. Layer semantic alignment, on the other hand, aligns the high-level summary vectors with the low-level primary input semantics, thus bridging the semantic gap between different layers. In the training process, GMSA first learns soft tokens that contain complete semantics through autoencoder training. To furtherly adapt GMSA to downstream tasks, we propose Knowledge Extraction Fine-tuning (KEFT) to extract knowledge from the soft tokens for downstream tasks. We train GMSA by randomly sampling the compression rate for each sample in the dataset. Under this condition, GMSA not only significantly outperforms the traditional compression paradigm in context restoration but also achieves stable and significantly faster convergence with only a few encoder layers. In downstream question-answering (QA) tasks, GMSA can achieve approximately a 2x speedup in end-to-end inference while outperforming both the original input prompts and various state-of-the-art (SOTA) methods by a large margin.

Large language models (LLMs) have proven to be highly effective across various natural language processing tasks. However, their large number of parameters poses significant challenges for practical deployment. Pruning, a technique aimed at reducing the size and complexity of LLMs, offers a potential solution by removing redundant components from the network. Despite the promise of pruning, existing methods often struggle to achieve substantial end-to-end LLM inference speedup. In this paper, we introduce SLEB, a novel approach designed to streamline LLMs by eliminating redundant transformer blocks. We choose the transformer block as the fundamental unit for pruning, because LLMs exhibit block-level redundancy with high similarity between the outputs of neighboring blocks. This choice allows us to effectively enhance the processing speed of LLMs. Our experimental results demonstrate that SLEB successfully accelerates LLM inference without compromising the linguistic capabilities of these models, making it a promising technique for optimizing the efficiency of LLMs. The code is available at: https://github.com/leapingjagg-dev/SLEB

Quantization has been widely used to compress and accelerate inference of large language models (LLMs). Existing methods focus on exploring the per-token dynamic calibration to ensure both inference acceleration and model accuracy under 4-bit quantization. However, in autoregressive generation inference of long sequences, the overhead of repeated dynamic quantization and dequantization steps becomes considerably expensive. In this work, we propose MergeQuant, an accurate and efficient per-channel static quantization framework. MergeQuant integrates the per-channel quantization steps with the corresponding scalings and linear mappings through a Quantization Step Migration (QSM) method, thereby eliminating the quantization overheads before and after matrix multiplication. Furthermore, in view of the significant differences between the different channel ranges, we propose dimensional reconstruction and adaptive clipping to address the non-uniformity of quantization scale factors and redistribute the channel variations to the subsequent modules to balance the parameter distribution under QSM. Within the static quantization setting of W4A4, MergeQuant reduces the accuracy gap on zero-shot tasks compared to FP16 baseline to 1.3 points on Llama-2-70B model. On Llama-2-7B model, MergeQuant achieves up to 1.77x speedup in decoding, and up to 2.06x speedup in end-to-end compared to FP16 baseline.

Long-context reasoning has significantly empowered large language models (LLMs) to tackle complex tasks, yet it introduces severe efficiency bottlenecks due to the computational complexity. Existing efficient approaches often rely on complex additional training or external models for compression, which limits scalability and discards critical fine-grained information. In this paper, we propose VTC-R1, a new efficient reasoning paradigm that integrates vision-text compression into the reasoning process. Instead of processing lengthy textual traces, VTC-R1 renders intermediate reasoning segments into compact images, which are iteratively fed back into vision-language models as "optical memory." We construct a training dataset based on OpenR1-Math-220K achieving 3.4x token compression and fine-tune representative VLMs-Glyph and Qwen3-VL. Extensive experiments on benchmarks such as MATH500, AIME25, AMC23 and GPQA-D demonstrate that VTC-R1 consistently outperforms standard long-context reasoning. Furthermore, our approach significantly improves inference efficiency, achieving 2.7x speedup in end-to-end latency, highlighting its potential as a scalable solution for reasoning-intensive applications. Our code is available at https://github.com/w-yibo/VTC-R1.

Vision-language-action (VLA) models unify perception, language, and control for embodied agents but face significant challenges in practical deployment due to rapidly increasing compute and memory demands, especially as models scale to longer horizons and larger backbones. To address these bottlenecks, we introduce QuantVLA, a training-free post-training quantization (PTQ) framework that, to our knowledge, is the first PTQ approach for VLA systems and the first to successfully quantize a diffusion transformer (DiT) action head. QuantVLA incorporates three scale-calibrated components: (1) a selective quantization layout that integerizes all linear layers in both the language backbone and the DiT while keeping attention projections in floating point to preserve the original operator schedule; (2) attention temperature matching, a lightweight per-head scaling mechanism that stabilizes attention logits and is folded into the dequantization scales at inference; and (3) output head balancing, a per-layer residual interface calibration that mitigates post-projection energy drift. The framework requires no additional training, uses only a small unlabeled calibration buffer, and supports integer kernels for low-bit weights and activations while leaving the architecture unchanged. Across representative VLA models on LIBERO, QuantVLA exceeds the task success rates of full-precision baselines, achieves about 70% relative memory savings on the quantized components, and delivers a 1.22x speedup in end-to-end inference latency, providing a practical pathway toward scalable low-bit embodied intelligence under strict compute, memory, and power constraints.

Large language models (LLMs) excel at complex reasoning when they include intermediate steps, known as "chains of thought" (CoTs). However, these rationales are often overly verbose, even for simple problems, leading to wasted context, increased latency, and higher energy consumption. We observe that verbose, English-heavy CoTs and concise, math-centric CoTs occupy distinct regions in the model's residual-stream activation space. By extracting and injecting a "steering vector" to transition between these modes, we can reliably shift generation toward more concise reasoning, effectively compressing CoTs without retraining. We formalize this approach as Activation-Steered Compression (ASC), an inference-time technique that shortens reasoning traces by directly modifying hidden representations. In addition, we provide a theoretical analysis of the impact of ASC on the output distribution, derived from a closed-form KL-divergence-bounded constraint to regulate steering strength. Using only 100 paired verbose and concise examples, ASC achieves up to 67.43% reduction in CoT length on MATH500 and GSM8K datasets, while maintaining accuracy across 7B, 8B, and 32B parameter models. As a training-free method, ASC introduces negligible runtime overhead and, on MATH500, delivers an average 2.73x speedup in end-to-end reasoning wall-clock time on an 8B model. This makes ASC a practical and efficient tool for streamlining the deployment of reasoning-capable LLMs in latency- or cost-sensitive settings. The code is available at: https://github.com/ArminAzizi98/ASC

Training vision-language web agents with multi-step RL is compute-intensive, with two dominant forms of inefficiency: idle GPUs in synchronous RL, and trajectories that use more steps and tokens than necessary. We present AsyncWebRL, which addresses both. On the system side, an asynchronous design overlaps rollout, gradient update, and policy refresh across iterations, paired with two web-agent-specific adaptations, namely an everlasting rollout pool and lightweight screenshot handling, that together deliver up to a 2.9times end-to-end training-throughput speedup over the previously fastest open synchronous pipeline (WebGym). On the algorithmic side, we identify the per-trajectory normalizer 1/|τ_i| in multi-step GRPO as the root cause of trajectory-level and token-level inefficiency: because failures are systematically longer than successes, it down-weights the negative gradient on failed tokens, so the policy keeps producing verbose memory schemas. Replacing 1/|τ_i| with a constant 1/k breaks this coupling, contracting trajectories while preserving aggregate success. Together, these contributions set a new open-source state of the art on the WebGym out-of-distribution test split (+5.8% relative over the 42.9% prior best), with the largest gains on the harder slices (+42% relative on Medium, +48% relative on Hard).

Vision-Language-Action (VLA) models based on flow matching -- such as pi0, pi0.5, and SmolVLA -- achieve state-of-the-art generalist robotic manipulation, yet their iterative denoising, typically 10 ODE steps, introduces substantial latency: on a modern GPU, denoising alone accounts for 80% of end-to-end inference time. Naively reducing the step count is unreliable, degrading success on most tasks due to the velocity field being uncalibrated for single-step jumps. We present SnapFlow, a plug-and-play self-distillation method that compresses multi-step denoising into a single forward pass (1-NFE) for flow-matching VLAs. SnapFlow mixes standard flow-matching samples with consistency samples whose targets are two-step Euler shortcut velocities computed from the model's own marginal velocity predictions, avoiding the trajectory drift caused by conditional velocities, as we analyze theoretically. A zero-initialized target-time embedding lets the network switch between local velocity estimation and global one-step generation within a single architecture. SnapFlow requires no external teacher, no architecture changes, and trains in ~12h on a single GPU. We validate on two VLA architectures spanning a 6x parameter range, with identical hyperparameters: on pi0.5 (3B) across four LIBERO suites (40 tasks, 400 episodes), SnapFlow achieves 98.75% average success -- matching the 10-step teacher at 97.75% and slightly exceeding it -- with 9.6x denoising speedup and end-to-end latency reduced from 274ms to 83ms; on SmolVLA (500M), it reduces MSE by 8.3% with 3.56x end-to-end acceleration. An action-step sweep on long-horizon tasks reveals that SnapFlow maintains its advantage across execution horizons, achieving 93% at n_act=5 where the baseline reaches only 90%. SnapFlow is orthogonal to layer-distillation and token-pruning approaches, enabling compositional speedups.

Autoregressive decoding is the only part of sequence-to-sequence models that prevents them from massive parallelization at inference time. Non-autoregressive models enable the decoder to generate all output symbols independently in parallel. We present a novel non-autoregressive architecture based on connectionist temporal classification and evaluate it on the task of neural machine translation. Unlike other non-autoregressive methods which operate in several steps, our model can be trained end-to-end. We conduct experiments on the WMT English-Romanian and English-German datasets. Our models achieve a significant speedup over the autoregressive models, keeping the translation quality comparable to other non-autoregressive models.

Monocular 4D human-object interaction (HOI) reconstruction - recovering a moving human and a manipulated object from a single RGB video - remains challenging due to depth ambiguity and frequent occlusions. Existing methods often rely on multi-stage pipelines or iterative optimization, leading to high inference latency, failing to meet real-time requirements, and susceptibility to error accumulation. To address these limitations, we propose THO, an end-to-end Spatial-Temporal Transformer that predicts human motion and coordinated object motion in a forward fashion from the given video and 3D template. THO achieves this by leveraging spatial-temporal HOI tuple priors. Spatial priors exploit contact-region proximity to infer occluded object features from human cues, while temporal priors capture cross-frame kinematic correlations to refine object representations and enforce physical coherence. Extensive experiments demonstrate that THO operates at an inference speed of 31.5 FPS on a single RTX 4090 GPU, achieving a >600x speedup over prior optimization-based methods while simultaneously improving reconstruction accuracy and temporal consistency. The project page is available at: https://nianheng.github.io/THO-project/

Crop row detection is essential for enabling autonomous navigation in GPS-denied environments, such as under-canopy agricultural settings. Traditional methods often struggle with occlusions, variable lighting conditions, and the structural variability of crop rows. To address these challenges, RowDetr, a novel end-to-end neural network architecture, is introduced for robust and efficient row detection. A new dataset of approximately 6,900 images is curated, capturing a diverse range of real-world agricultural conditions, including occluded rows, uneven terrain, and varying crop densities. Unlike previous approaches, RowDetr leverages smooth polynomial functions to precisely delineate crop boundaries in the image space, ensuring a more structured and interpretable representation of row geometry. A key innovation of this approach is PolyOptLoss, a novel energy-based loss function designed to enhance learning robustness, even in the presence of noisy or imperfect labels. This loss function significantly improves model stability and generalization by optimizing polynomial curve fitting directly in image space. Extensive experiments demonstrate that RowDetr significantly outperforms existing frameworks, including Agronav and RowColAttention, across key performance metrics. Additionally, RowDetr achieves a sixfold speedup over Agronav, making it highly suitable for real-time deployment on resource-constrained edge devices. To facilitate better comparisons across future studies, lane detection metrics from autonomous driving research are adapted, providing a more standardized and meaningful evaluation framework for crop row detection. This work establishes a new benchmark in under-canopy

Large language models (LLMs) are increasingly deployed on customer devices. To support them, current devices are adopting SoCs (System on Chip) with NPUs (Neural Processing Unit) installed. Although high performance is expected, LLM inference on NPUs is slower than its CPU counterpart. The reason is that NPUs have poor performance on computations other than GEMM, like dequantization. Current works either disaggregate prefill on the NPUs and decoding on the CPUs, or put both on the NPUs but with an accuracy loss. To solve this issue, based on the insight that low-bit can enable target computation encoded within an acceptably sized table, we propose table lookup to subsume hardware operations otherwise unsupported. To realize this, we overcome the conflicting hardware behavior of prefill and decoding to design a unified table layout and tiling through (1) fused two-level table-based dequantization and (2) concurrency-hierarchy-guided tiling. Based on that, we implement the prefill phase by three-stage pipeline and map the table-lookup-based decoding to NPU's vector units. Results show 1.4x and 3.1x speedup for prefill and decoding respectively, and 84% energy savings compared to the baseline NPU methods. The code is available at https://github.com/microsoft/T-MAC/tree/main/t-man.