Pathology foundation models (PFMs) have emerged as powerful tools for analyzing whole slide images (WSIs). However, adapting these pretrained PFMs for specific clinical tasks presents considerable challenges, primarily due to the availability of only weak (WSI-level) labels for gigapixel images, necessitating multiple instance learning (MIL) paradigm for effective WSI analysis. This paper proposes a novel approach for single-GPU \textbf{T}ask \textbf{A}daptation of \textbf{PFM}s (TAPFM) that uses vision transformer (\vit) attention for MIL aggregation while optimizing both for feature representations and attention weights. The proposed approach maintains separate computational graphs for MIL aggregator and the PFM to create stable training dynamics that align with downstream task objectives during end-to-end adaptation. Evaluated on mutation prediction tasks for bladder cancer and lung adenocarcinoma across institutional and TCGA cohorts, TAPFM consistently outperforms conventional approaches, with H-Optimus-0 (TAPFM) outperforming the benchmarks. TAPFM effectively handles multi-label classification of actionable mutations as well. Thus, TAPFM makes adaptation of powerful pre-trained PFMs practical on standard hardware for various clinical applications.
Video diffusion transformers (vDiTs) have made impressive progress in text-to-video generation, but their high computational demands present major challenges for practical deployment. While existing acceleration methods reduce workload at various granularities, they often rely on heuristics, limiting their applicability. We introduce ASTRAEA, an automatic framework that searches for near-optimal configurations for vDiT-based video generation. At its core, ASTRAEA proposes a lightweight token selection mechanism and a memory-efficient, GPU-parallel sparse attention strategy, enabling linear reductions in execution time with minimal impact on generation quality. To determine optimal token reduction for different timesteps, we further design a search framework that leverages a classic evolutionary algorithm to automatically determine the distribution of the token budget effectively. Together, ASTRAEA achieves up to 2.4x inference speedup on a single GPU with great scalability (up to 13.2x speedup on 8 GPUs) while retaining better video quality compared to the state-of-the-art methods (<0.5% loss on the VBench score compared to the baseline vDiT models).
Existing LiDAR 3D object detection methods predominantely rely on sparse convolutions and/or transformers, which can be challenging to run on resource-constrained edge devices, due to irregular memory access patterns and high computational costs. In this paper, we propose FALO, a hardware-friendly approach to LiDAR 3D detection, which offers both state-of-the-art (SOTA) detection accuracy and fast inference speed. More specifically, given the 3D point cloud and after voxelization, FALO first arranges sparse 3D voxels into a 1D sequence based on their coordinates and proximity. The sequence is then processed by our proposed ConvDotMix blocks, consisting of large-kernel convolutions, Hadamard products, and linear layers. ConvDotMix provides sufficient mixing capability in both spatial and embedding dimensions, and introduces higher-order nonlinear interaction among spatial features. Furthermore, when going through the ConvDotMix layers, we introduce implicit grouping, which balances the tensor dimensions for more efficient inference and takes into account the growing receptive field. All these operations are friendly to run on resource-constrained platforms and proposed FALO can readily deploy on compact, embedded devices. Our extensive evaluation on LiDAR 3D detection benchmarks such as nuScenes and Waymo shows that FALO achieves competitive performance. Meanwhile, FALO is 1.6~9.8x faster than the latest SOTA on mobile Graphics Processing Unit (GPU) and mobile Neural Processing Unit (NPU).
Transducer models have emerged as a promising choice for end-to-end ASR systems, offering a balanced trade-off between recognition accuracy, streaming capabilities, and inference speed in greedy decoding. However, beam search significantly slows down Transducers due to repeated evaluations of key network components, limiting practical applications. This paper introduces a universal method to accelerate beam search for Transducers, enabling the implementation of two optimized algorithms: ALSD++ and AES++. The proposed method utilizes batch operations, a tree-based hypothesis structure, novel blank scoring for enhanced shallow fusion, and CUDA graph execution for efficient GPU inference. This narrows the speed gap between beam and greedy modes to only 10-20% for the whole system, achieves 14-30% relative improvement in WER compared to greedy decoding, and improves shallow fusion for low-resource up to 11% compared to existing implementations. All the algorithms are open sourced.
Vision-language models (VLMs) pretrained on large-scale multimodal datasets encode rich visual and linguistic knowledge, making them a strong foundation for robotics. Rather than training robotic policies from scratch, recent approaches adapt VLMs into vision-language-action (VLA) models that enable natural language-driven perception and control. However, existing VLAs are typically massive--often with billions of parameters--leading to high training costs and limited real-world deployability. Moreover, they rely on academic and industrial datasets, overlooking the growing availability of community-collected data from affordable robotic platforms. In this work, we present SmolVLA, a small, efficient, and community-driven VLA that drastically reduces both training and inference costs, while retaining competitive performance. SmolVLA is designed to be trained on a single GPU and deployed on consumer-grade GPUs or even CPUs. To further improve responsiveness, we introduce an asynchronous inference stack decoupling perception and action prediction from action execution, allowing higher control rates with chunked action generation. Despite its compact size, SmolVLA achieves performance comparable to VLAs that are 10x larger. We evaluate SmolVLA on a range of both simulated as well as real-world robotic benchmarks and release all code, pretrained models, and training data.
We present eACGM, a full-stack AI/ML system monitoring framework based on eBPF. eACGM collects real-time performance data from key hardware components, including the GPU and network communication layer, as well as from key software stacks such as CUDA, Python, and PyTorch, all without requiring any code instrumentation or modifications. Additionally, it leverages libnvml to gather process-level GPU resource usage information. By applying a Gaussian Mixture Model (GMM) to the collected multidimensional performance metrics for statistical modeling and clustering analysis, eACGM effectively identifies complex failure modes, such as latency anomalies, hardware failures, and communication inefficiencies, enabling rapid diagnosis of system bottlenecks and abnormal behaviors. To evaluate eACGM's effectiveness and practicality, we conducted extensive empirical studies and case analyses in multi-node distributed training scenarios. The results demonstrate that eACGM, while maintaining a non-intrusive and low-overhead profile, successfully captures critical performance anomalies during model training and inference. Its stable anomaly detection performance and comprehensive monitoring capabilities validate its applicability and scalability in real-world production environments, providing strong support for performance optimization and fault diagnosis in large-scale AI/ML systems.
Efficient inference of Multi-Head Latent Attention (MLA) is challenged by deploying the DeepSeek-R1 671B model on a single Multi-GPU server. This paper introduces FlashMLA-ETAP, a novel framework that enhances MLA inference for the single-instance deployment scenario on NVIDIA H20 GPUs. We propose the Efficient Transpose Attention Pipeline (ETAP), which reconfigures attention computation through transposition to align the KV context length with the \(M\)-dimension in WGMMA operations, significantly reducing redundant computations. FlashMLA-ETAP achieves a 2.78x speedup over FlashMLA at 64K sequence length (batch size 16), with 5.24x and 4.94x improvements over FlashAttention-3 and FlashInfer, respectively, while maintaining numerical stability with a 15.2x lower RMSE (\(1.25 \times 10^{-5}\)) than FlashAttention-3. Furthermore, ETAP's design enables seamless integration into frameworks like FlashAttention-3 and FlashInfer, supported by a detailed theoretical analysis. Our work addresses a critical gap in resource-constrained inference, offering a scalable solution for mid-tier GPUs and paving the way for broader adoption in hardware-aware optimization. Code is available at https://github.com/pengcuo/FlashMLA-ETAP.
The computational sparsity of Mixture-of-Experts (MoE) models enables sub-linear growth in compute cost as model size increases, thus offering a scalable path to training massive neural networks. However, existing implementations suffer from \emph{low GPU utilization}, \emph{significant latency overhead}, and a fundamental \emph{inability to leverage task locality}, primarily due to CPU-managed scheduling, host-initiated communication, and frequent kernel launches. To overcome these limitations, we develop FlashDMoE, a fully GPU-resident MoE operator that fuses expert computation and inter-GPU communication into a \emph{single persistent GPU kernel}. FlashDMoE enables fine-grained pipelining of dispatch, compute, and combine phases, eliminating launch overheads and reducing idle gaps. Unlike existing work, FlashDMoE obviates bulk-synchronous collectives for one-sided, device-initiated, inter-GPU (R)DMA transfers, thus unlocking \emph{payload efficiency}, where we eliminate bloated or redundant network payloads in sparsely activated layers. When evaluated on a single 8-H100 GPU node with MoE models having up to 128 experts and 16K token sequences, FlashDMoE achieves up to \textbf{9}$\times$ higher GPU utilization, \textbf{6}$\times$ lower latency, \textbf{5.7}$\times$ higher throughput, and \textbf{4}$\times$ better overlap efficiency compared to state-of-the-art baselines, despite using FP32 while baselines use FP16. FlashDMoE demonstrates that principled GPU kernel-hardware co-design is key to unlocking the performance ceiling of large-scale distributed ML workloads.
The meteoric rise and proliferation of autonomous Large Language Model (LLM) agents promise significant capabilities across various domains. However, their deployment is increasingly constrained by substantial computational demands, specifically for Graphics Processing Unit (GPU) resources. This paper addresses the critical problem of optimizing resource utilization in LLM agent systems. We introduce COALESCE (Cost-Optimized and Secure Agent Labour Exchange via Skill-based Competence Estimation), a novel framework designed to enable autonomous LLM agents to dynamically outsource specific subtasks to specialized, cost-effective third-party LLM agents. The framework integrates mechanisms for hybrid skill representation, dynamic skill discovery, automated task decomposition, a unified cost model comparing internal execution costs against external outsourcing prices, simplified market-based decision-making algorithms, and a standardized communication protocol between LLM agents. Comprehensive validation through 239 theoretical simulations demonstrates 41.8\% cost reduction potential, while large-scale empirical validation across 240 real LLM tasks confirms 20.3\% cost reduction with proper epsilon-greedy exploration, establishing both theoretical viability and practical effectiveness. The emergence of proposed open standards like Google's Agent2Agent (A2A) protocol further underscores the need for frameworks like COALESCE that can leverage such standards for efficient agent interaction. By facilitating a dynamic market for agent capabilities, potentially utilizing protocols like A2A for communication, COALESCE aims to significantly reduce operational costs, enhance system scalability, and foster the emergence of specialized agent economies, making complex LLM agent functionalities more accessible and economically viable.
The rapid growth of large-language models (LLMs) is driving a new wave of specialized hardware for inference. This paper presents the first workload-centric, cross-architectural performance study of commercial AI accelerators, spanning GPU-based chips, hybrid packages, and wafer-scale engines. We compare memory hierarchies, compute fabrics, and on-chip interconnects, and observe up to 3.7x performance variation across architectures as batch size and sequence length change. Four scaling techniques for trillion-parameter models are examined; expert parallelism offers an 8.4x parameter-to-compute advantage but incurs 2.1x higher latency variance than tensor parallelism. These findings provide quantitative guidance for matching workloads to accelerators and reveal architectural gaps that next-generation designs must address.
LiDAR semantic segmentation plays a vital role in autonomous driving. Existing voxel-based methods for LiDAR semantic segmentation apply uniform partition to the 3D LiDAR point cloud to form a structured representation based on cartesian/cylindrical coordinates. Although these methods show impressive performance, the drawback of existing voxel-based methods remains in two aspects: (1) it requires a large enough input voxel resolution, which brings a large amount of computation cost and memory consumption. (2) it does not well handle the unbalanced point distribution of LiDAR point cloud. In this paper, we propose a non-uniform cylindrical partition network named NUC-Net to tackle the above challenges. Specifically, we propose the Arithmetic Progression of Interval (API) method to non-uniformly partition the radial axis and generate the voxel representation which is representative and efficient. Moreover, we propose a non-uniform multi-scale aggregation method to improve contextual information. Our method achieves state-of-the-art performance on SemanticKITTI and nuScenes datasets with much faster speed and much less training time. And our method can be a general component for LiDAR semantic segmentation, which significantly improves both the accuracy and efficiency of the uniform counterpart by $4 \times$ training faster and $2 \times$ GPU memory reduction and $3 \times$ inference speedup. We further provide theoretical analysis towards understanding why NUC is effective and how point distribution affects performance. Code is available at \href{https://github.com/alanWXZ/NUC-Net}{https://github.com/alanWXZ/NUC-Net}.
Reinforcement learning (RL) has become a dominant paradigm for training large language models (LLMs), particularly for reasoning tasks. Effective RL for LLMs requires massive parallelization and poses an urgent need for efficient training systems. Most existing large-scale RL systems for LLMs are synchronous, alternating generation and training in a batch setting where rollouts in each training batch are generated by the same model. This approach stabilizes RL training but suffers from severe system-level inefficiency: generation must wait until the longest output in the batch is completed before model updates, resulting in GPU underutilization. We present AReaL, a fully asynchronous RL system that completely decouples generation from training. Rollout workers in AReaL continuously generate new outputs without waiting, while training workers update the model whenever a batch of data is collected. AReaL also incorporates a collection of system-level optimizations, leading to substantially higher GPU utilization. To stabilize RL training, AReaL balances the workload of rollout and training workers to control data staleness, and adopts a staleness-enhanced PPO variant to better handle outdated training samples. Extensive experiments on math and code reasoning benchmarks show that AReaL achieves up to 2.77$\times$ training speedup compared to synchronous systems with the same number of GPUs and matched or improved final performance. The code of AReaL is available at https://github.com/inclusionAI/AReaL/.
Network pruning reduces the computational requirements of large neural networks, with N:M sparsity -- retaining only N out of every M consecutive weights -- offering a compelling balance between compressed model quality and hardware acceleration. However, N:M sparsity only accelerates forward-pass computations, as N:M patterns are not preserved during matrix transposition, limiting efficiency during training where both passes are computationally intensive. While transposable N:M sparsity has been proposed to address this limitation, existing methods for finding transposable N:M sparse masks either fail to scale to large models or are restricted to M=4 which results in suboptimal compression-accuracy trade-off. We introduce an efficient solver for transposable N:M masks that scales to billion-parameter models. We formulate mask generation as optimal transport problems and solve through entropy regularization and Dykstra's algorithm, followed by a rounding procedure. Our tensor-based implementation exploits GPU parallelism, achieving up to 100x speedup with only 1-10% error compared to existing methods. Our approach can be integrated with layer-wise N:M pruning frameworks including Wanda, SparseGPT and ALPS to produce transposable N:M sparse models with arbitrary N:M values. Experiments show that LLaMA3.2-8B with transposable 16:32 sparsity maintains performance close to its standard N:M counterpart and outperforms standard 2:4 sparse model, showing the practical value of our approach.
Reinforcement Learning (RL) has become the most effective post-training approach for improving the capabilities of Large Language Models (LLMs). In practice, because of the high demands on latency and memory, it is particularly challenging to develop an efficient RL framework that reliably manages policy models with hundreds to thousands of billions of parameters. In this paper, we present LlamaRL, a fully distributed, asynchronous RL framework optimized for efficient training of large-scale LLMs with various model sizes (8B, 70B, and 405B parameters) on GPU clusters ranging from a handful to thousands of devices. LlamaRL introduces a streamlined, single-controller architecture built entirely on native PyTorch, enabling modularity, ease of use, and seamless scalability to thousands of GPUs. We also provide a theoretical analysis of LlamaRL's efficiency, including a formal proof that its asynchronous design leads to strict RL speed-up. Empirically, by leveraging best practices such as colocated model offloading, asynchronous off-policy training, and distributed direct memory access for weight synchronization, LlamaRL achieves significant efficiency gains -- up to 10.7x speed-up compared to DeepSpeed-Chat-like systems on a 405B-parameter policy model. Furthermore, the efficiency advantage continues to grow with increasing model scale, demonstrating the framework's suitability for future large-scale RL training.
Feature matching dominats the time costs in structure from motion (SfM). The primary contribution of this study is a GPU data schedule algorithm for efficient feature matching of Unmanned aerial vehicle (UAV) images. The core idea is to divide the whole dataset into blocks based on the matrix band reduction (MBR) and achieve efficient feature matching via GPU-accelerated cascade hashing. First, match pairs are selected by using an image retrieval technique, which converts images into global descriptors and searches high-dimension nearest neighbors with graph indexing. Second, compact image blocks are iteratively generated from a MBR-based data schedule strategy, which exploits image connections to avoid redundant data IO (input/output) burden and increases the usage of GPU computing power. Third, guided by the generated image blocks, feature matching is executed sequentially within the framework of GPU-accelerated cascade hashing, and initial candidate matches are refined by combining a local geometric constraint and RANSAC-based global verification. For further performance improvement, these two seps are designed to execute parallelly in GPU and CPU. Finally, the performance of the proposed solution is evaluated by using large-scale UAV datasets. The results demonstrate that it increases the efficiency of feature matching with speedup ratios ranging from 77.0 to 100.0 compared with KD-Tree based matching methods, and achieves comparable accuracy in relative and absolute bundle adjustment (BA). The proposed algorithm is an efficient solution for feature matching of UAV images.
Transformer-based large language models suffer from quadratic complexity at inference on long sequences. Linear attention methods are efficient alternatives, however, they fail to provide an accurate approximation of softmax attention. By additionally incorporating sliding window attention into each linear attention head, this gap can be closed for short context-length tasks. Unfortunately, these approaches cannot recall important information from long contexts due to "memory collisions". In this paper , we propose LoLA: Low-rank Linear Attention with sparse caching. LoLA separately stores additional key-value pairs that would otherwise interfere with past associative memories. Moreover, LoLA further closes the gap between linear attention models and transformers by distributing past key-value pairs into three forms of memory: (i) recent pairs in a local sliding window; (ii) difficult-to-memorize pairs in a sparse, global cache; and (iii) generic pairs in the recurrent hidden state of linear attention. As an inference-only strategy, LoLA enables pass-key retrieval on up to 8K context lengths on needle-in-a-haystack tasks from RULER. It boosts the accuracy of the base subquadratic model from 0.6% to 97.4% at 4K context lengths, with a 4.6x smaller cache than that of Llama-3.1 8B. LoLA demonstrates strong performance on zero-shot commonsense reasoning tasks among 1B and 8B parameter subquadratic models. Finally, LoLA is an extremely lightweight approach: Nearly all of our results can be reproduced on a single consumer GPU.
Speculative decoding and quantization effectively accelerate memory-bound inference of large language models. Speculative decoding mitigates the memory bandwidth bottleneck by verifying multiple tokens within a single forward pass, which increases computational effort. Quantization achieves this optimization by compressing weights and activations into lower bit-widths and also reduces computations via low-bit matrix multiplications. To further leverage their strengths, we investigate the integration of these two techniques. Surprisingly, experiments applying the advanced speculative decoding method EAGLE-2 to various quantized models reveal that the memory benefits from 4-bit weight quantization are diminished by the computational load from speculative decoding. Specifically, verifying a tree-style draft incurs significantly more time overhead than a single-token forward pass on 4-bit weight quantized models. This finding led to our new speculative decoding design: a hierarchical framework that employs a small model as an intermediate stage to turn tree-style drafts into sequence drafts, leveraging the memory access benefits of the target quantized model. Experimental results show that our hierarchical approach achieves a 2.78$\times$ speedup across various tasks for the 4-bit weight Llama-3-70B model on an A100 GPU, outperforming EAGLE-2 by 1.31$\times$. Code available at https://github.com/AI9Stars/SpecMQuant.
3D occupancy prediction has attracted much attention in the field of autonomous driving due to its powerful geometric perception and object recognition capabilities. However, existing methods have not explored the most essential distribution patterns of voxels, resulting in unsatisfactory results. This paper first explores the inter-class distribution and geometric distribution of voxels, thereby solving the long-tail problem caused by the inter-class distribution and the poor performance caused by the geometric distribution. Specifically, this paper proposes SHTOcc (Sparse Head-Tail Occupancy), which uses sparse head-tail voxel construction to accurately identify and balance key voxels in the head and tail classes, while using decoupled learning to reduce the model's bias towards the dominant (head) category and enhance the focus on the tail class. Experiments show that significant improvements have been made on multiple baselines: SHTOcc reduces GPU memory usage by 42.2%, increases inference speed by 58.6%, and improves accuracy by about 7%, verifying its effectiveness and efficiency. The code is available at https://github.com/ge95net/SHTOcc
Diffusion Transformer (DiT)-based video diffusion models generate high-quality videos at scale but incur prohibitive processing latency and memory costs for long videos. To address this, we propose a novel distributed inference strategy, termed DualParal. The core idea is that, instead of generating an entire video on a single GPU, we parallelize both temporal frames and model layers across GPUs. However, a naive implementation of this division faces a key limitation: since diffusion models require synchronized noise levels across frames, this implementation leads to the serialization of original parallelisms. We leverage a block-wise denoising scheme to handle this. Namely, we process a sequence of frame blocks through the pipeline with progressively decreasing noise levels. Each GPU handles a specific block and layer subset while passing previous results to the next GPU, enabling asynchronous computation and communication. To further optimize performance, we incorporate two key enhancements. Firstly, a feature cache is implemented on each GPU to store and reuse features from the prior block as context, minimizing inter-GPU communication and redundant computation. Secondly, we employ a coordinated noise initialization strategy, ensuring globally consistent temporal dynamics by sharing initial noise patterns across GPUs without extra resource costs. Together, these enable fast, artifact-free, and infinitely long video generation. Applied to the latest diffusion transformer video generator, our method efficiently produces 1,025-frame videos with up to 6.54$\times$ lower latency and 1.48$\times$ lower memory cost on 8$\times$RTX 4090 GPUs.
Recent advances in 3D Gaussian Splatting have shown remarkable potential for novel view synthesis. However, most existing large-scale scene reconstruction methods rely on the divide-and-conquer paradigm, which often leads to the loss of global scene information and requires complex parameter tuning due to scene partitioning and local optimization. To address these limitations, we propose MixGS, a novel holistic optimization framework for large-scale 3D scene reconstruction. MixGS models the entire scene holistically by integrating camera pose and Gaussian attributes into a view-aware representation, which is decoded into fine-detailed Gaussians. Furthermore, a novel mixing operation combines decoded and original Gaussians to jointly preserve global coherence and local fidelity. Extensive experiments on large-scale scenes demonstrate that MixGS achieves state-of-the-art rendering quality and competitive speed, while significantly reducing computational requirements, enabling large-scale scene reconstruction training on a single 24GB VRAM GPU. The code will be released at https://github.com/azhuantou/MixGS.