The ever-increasing sizes of large language models necessitate distributed solutions for fast inference that exploit multi-dimensional parallelism, where computational loads are split across various accelerators such as GPU clusters. However, this approach often introduces significant communication overhead, especially on devices with limited bandwidth. In this paper, we introduce Flash Communication, a novel low-bit compression technique designed to alleviate the tensor-parallelism communication bottleneck during inference. Our method substantially boosts intra-node communication speed by more than 3x and reduces the time-to-first-token by 2x, with nearly no sacrifice in model accuracy. Extensive experiments on various up-to-date LLMs demonstrate the effectiveness of our approach.
Efficiently deploying large language models (LLMs) in real-world scenarios remains a critical challenge, primarily due to hardware heterogeneity, inference framework limitations, and workload complexities.Efficiently deploying large language models (LLMs) in real-world scenarios remains a critical challenge, primarily due to hardware heterogeneity, inference framework limitations, and workload complexities. These challenges often lead to inefficiencies in memory utilization, latency, and throughput, hindering the effective deployment of LLMs, especially for non-experts. Through extensive experiments, we identify key performance bottlenecks, including sudden drops in memory utilization, latency fluctuations with varying batch sizes, and inefficiencies in multi-GPU configurations. These insights reveal a vast optimization space shaped by the intricate interplay of hardware, frameworks, and workload parameters. This underscores the need for a systematic approach to optimize LLM inference, motivating the design of our framework, GUIDE. GUIDE leverages dynamic modeling and simulation-based optimization to address these issues, achieving prediction errors between 9.9% and 42.3% for key metrics such as batch latency, TTFT, and decode throughput. By effectively bridging the gap between theoretical performance and practical deployment, our framework empowers practitioners, particularly non-specialists, to make data-driven decisions and unlock the full potential of LLMs in heterogeneous environments cheaply.
Large language models (LLMs) are notoriously memory-intensive during training, particularly with the popular AdamW optimizer. This memory burden necessitates using more or higher-end GPUs or reducing batch sizes, limiting training scalability and throughput. To address this, various memory-efficient optimizers have been proposed to reduce optimizer memory usage. However, they face critical challenges: (i) reliance on costly SVD operations; (ii) significant performance trade-offs compared to AdamW; and (iii) still substantial optimizer memory overhead to maintain competitive performance. In this work, we identify that AdamW's learning rate adaptation rule can be effectively coarsened as a structured learning rate update. Based on this insight, we propose Approximated Gradient Scaling for Memory-Efficient LLM Optimization (APOLLO), which approximates learning rate scaling using an auxiliary low-rank optimizer state based on pure random projection. This structured learning rate update rule makes APOLLO highly tolerant to further memory reductions while delivering comparable pre-training performance. Even its rank-1 variant, APOLLO-Mini, achieves superior pre-training performance compared to AdamW with SGD-level memory costs. Extensive experiments demonstrate that the APOLLO series performs on-par with or better than AdamW, while achieving greater memory savings by nearly eliminating the optimization states of AdamW. These savings provide significant system-level benefits: (1) Enhanced Throughput: 3x throughput on an 8xA100-80GB setup compared to AdamW by supporting 4x larger batch sizes. (2) Improved Model Scalability: Pre-training LLaMA-13B with naive DDP on A100-80GB GPUs without system-level optimizations. (3) Low-End GPU Friendly Pre-training: Pre-training LLaMA-7B on a single GPU using less than 12 GB of memory with weight quantization.
Large Language Models (LLMs) exhibit impressive performance across various tasks, but deploying them for inference poses challenges. Their high resource demands often necessitate complex, costly multi-GPU pipelines, or the use of smaller, less capable models. While quantization offers a promising solution utilizing lower precision for model storage, existing methods frequently experience significant performance drops at lower precision levels. Additionally, they typically provide only a limited set of solutions at specific bit levels, many of which are extensively manually tuned. To address these challenges, we propose a new method called SKIM: Scaled K-means clustering wIth Mixed precision. Our approach introduces two novel techniques: 1. A greedy algorithm to solve approximately optimal bit allocation across weight channels, and 2. A trainable scaling vector for non-differentiable K-means clustering. These techniques substantially improve performance and can be adapted to any given bit. Notably, in terms of model perplexity, our method narrows the gap between 3-bit quantized LLaMA models and their full precision counterparts by 16.3% on average.
Large language models (LLMs) increasingly play an important role in a wide range of information processing and management tasks. Many of these tasks are performed in large batches or even offline, and the performance indictor for which is throughput. These tasks usually show the characteristic of prefix sharing, where different prompt input can partially show the common prefix. However, the existing LLM inference engines tend to optimize the streaming requests and show limitations of supporting the large batched tasks with the prefix sharing characteristic. The existing solutions use the LRU-based cache to reuse the KV context of common prefix between requests. The KV context that are about to be reused may prematurely evicted with the implicit cache management. Besides, the streaming oriented systems do not leverage the request-batch information and can not mix the decoding tokens with the prefill chunks to the best for the batched scenarios, and thus fails to saturate the GPU. We propose BatchLLM to address the above problems. BatchLLM explicitly identifies the common prefixes globally. The requests sharing the same prefix will be scheduled together to reuse the KV context the best. BatchLLM reorders the requests and schedules the requests with larger ratio of decoding first to better mix the decoding tokens with the latter prefill chunks, and applies memory-centric token batching to enlarge the token-batch sizes, which helps to increase the GPU utilization. Finally, BatchLLM optimizes the prefix-shared Attention kernel with horizontal fusion to reduce tail effect and kernel launch overhead. Extensive evaluation shows that BatchLLM outperforms vLLM and SGLang by 1.3$\times$ to 10.8$\times$ on a set of microbenchmarks and a typical industry workload under different hardware environments.
Despite the remarkable performance of multimodal large language models (MLLMs) across diverse tasks, the substantial training and inference costs impede their advancement. The majority of computation stems from the overwhelming volume of vision tokens processed by the transformer decoder. In this paper, we propose to build efficient MLLMs by leveraging the Mixture-of-Depths (MoD) mechanism, where each transformer decoder layer selects essential vision tokens to process while skipping redundant ones. However, integrating MoD into MLLMs is non-trivial. To address the challenges of training and inference stability as well as limited training data, we adapt the MoD module with two novel designs: tanh-gated weight normalization (TanhNorm) and symmetric token reweighting (STRing). Moreover, we observe that vision tokens exhibit higher redundancy in deeper layer and thus design a progressive ratio decay (PRD) strategy, which gradually reduces the token retention ratio layer by layer, employing a shifted cosine schedule. This crucial design fully unleashes the potential of MoD, significantly boosting the efficiency and performance of our models. To validate the effectiveness of our approach, we conduct extensive experiments with two baseline models across 14 benchmarks. Our model, p-MoD, matches or even surpasses the performance of the baseline models, with only 55.6% TFLOPs and 53.8% KV cache storage during inference, and 77.7% GPU hours during training.
We introduce DEIM, an innovative and efficient training framework designed to accelerate convergence in real-time object detection with Transformer-based architectures (DETR). To mitigate the sparse supervision inherent in one-to-one (O2O) matching in DETR models, DEIM employs a Dense O2O matching strategy. This approach increases the number of positive samples per image by incorporating additional targets, using standard data augmentation techniques. While Dense O2O matching speeds up convergence, it also introduces numerous low-quality matches that could affect performance. To address this, we propose the Matchability-Aware Loss (MAL), a novel loss function that optimizes matches across various quality levels, enhancing the effectiveness of Dense O2O. Extensive experiments on the COCO dataset validate the efficacy of DEIM. When integrated with RT-DETR and D-FINE, it consistently boosts performance while reducing training time by 50%. Notably, paired with RT-DETRv2, DEIM achieves 53.2% AP in a single day of training on an NVIDIA 4090 GPU. Additionally, DEIM-trained real-time models outperform leading real-time object detectors, with DEIM-D-FINE-L and DEIM-D-FINE-X achieving 54.7% and 56.5% AP at 124 and 78 FPS on an NVIDIA T4 GPU, respectively, without the need for additional data. We believe DEIM sets a new baseline for advancements in real-time object detection. Our code and pre-trained models are available at https://github.com/ShihuaHuang95/DEIM.
Optimizing deep learning algorithms currently requires slow, manual derivation, potentially leaving much performance untapped. Methods like FlashAttention have achieved a x6 performance improvement over native PyTorch by avoiding unnecessary data transfers, but required three iterations over three years to be developed. Automated compiled methods have consistently lagged behind. This paper extends Neural Circuit Diagrams for deep learning models to consider resource usage and the distribution of tasks across a GPU hierarchy. We show how diagrams can use simple relabellings to derive high-level streaming and tiling optimization strategies along with performance models. We show how this high-level performance model allows the effects of quantization and multi-level GPU hierarchies to be readily considered. We develop a methodology for representing intermediate-level pseudocode with diagrams, allowing hardware-aware algorithms to be derived step-by-step. Finally, we show how our methodology can be used to better understand existing techniques like FlashAttention. This work uses a theoretical framework to link assumptions about GPU behaviour to claims about performance. We aim to lay the groundwork for a scientific approach to GPU optimization where experiments can address clear hypotheses rather than post-hoc rationalizations.
Large language models (LLMs) exhibit exceptional performance but incur significant serving costs due to their substantial memory requirements, with the key-value (KV) cache being a primary bottleneck. Existing KV cache compression techniques, such as quantization and pruning, apply uniform treatment to both keys and values, and discard unimportant tokens entirely, overlooking the fine-grained differences in significance of various components within the KV cache. To address these limitations, we introduce LeanKV, a framework that advances KV cache compression by exploiting three levels of differentiation in the KV cache: (1) the differing impact of keys and values on attention computation, (2) the varying importance of tokens, and (3) the diverse dynamic sparsity patterns across attention heads. At the core of LeanKV is an on-GPU memory manager that compacts fragmented free memory list into contiguous regions in parallel, effectively translating sparsity in the KV cache into performance gains. We evaluate LeanKV on several mainstream models, including the recent "thinking model". LeanKV is able to compress the KV cache by $2.7\times$ to $5.7\times$ with near-lossless accuracy on complex workloads requiring sophisticated reasoning and long-generation capabilities, and enhances throughput by $1.9\times$ to $5.4\times$.
In this paper, we present DiffusionVLA, a novel framework that seamlessly combines the autoregression model with the diffusion model for learning visuomotor policy. Central to our approach is a next-token prediction objective, enabling the model to reason effectively over the user's query in the context of current observations. Subsequently, a diffusion model is attached to generate robust action outputs. To enhance policy learning through self-reasoning, we introduce a novel reasoning injection module that integrates reasoning phrases directly into the policy learning process. The whole framework is simple and flexible, making it easy to deploy and upgrade. We conduct extensive experiments using multiple real robots to validate the effectiveness of DiffusionVLA. Our tests include a challenging factory sorting task, where DiffusionVLA successfully categorizes objects, including those not seen during training. We observe that the reasoning module makes the model interpretable. It allows observers to understand the model thought process and identify potential causes of policy failures. Additionally, we test DiffusionVLA on a zero-shot bin-picking task, achieving 63.7\% accuracy on 102 previously unseen objects. Our method demonstrates robustness to visual changes, such as distractors and new backgrounds, and easily adapts to new embodiments. Furthermore, DiffusionVLA can follow novel instructions and retain conversational ability. Notably, DiffusionVLA is data-efficient and fast at inference; our smallest DiffusionVLA-2B runs 82Hz on a single A6000 GPU and can train from scratch on less than 50 demonstrations for a complex task. Finally, we scale the model from 2B to 72B parameters, showcasing improved generalization capabilities with increased model size.
Masked Image Modeling (MIM) has emerged as a promising approach for Self-Supervised Learning (SSL) of visual representations. However, the out-of-the-box performance of MIMs is typically inferior to competing approaches. Most users cannot afford fine-tuning due to the need for large amounts of data, high GPU consumption, and specialized user knowledge. Therefore, the practical use of MIM representations is limited. In this paper we ask what is the reason for the poor out-of-the-box performance of MIMs. Is it due to weaker features produced by MIM models, or is it due to suboptimal usage? Through detailed analysis, we show that attention in MIMs is spread almost uniformly over many patches, leading to ineffective aggregation by the [cls] token. Based on this insight, we propose Selective Aggregation to better capture the rich semantic information retained in patch tokens, which significantly improves the out-of-the-box performance of MIM.
Unsupervised 3D representation learning via masked-and-reconstruction with differentiable rendering is promising to reduce the labeling burden for fusion 3D perception. However, previous literature conduct pre-training for different modalities separately because of the hight GPU memory consumption. Consequently, the interaction between the two modalities (images and point clouds) is neglected during pre-training. In this paper, we explore joint unsupervised pre-training for fusion 3D perception via differentiable rendering and propose CLAP, short for Curvature sampLing and swApping Prototype assignment prediction. The contributions are three-fold. 1) To overcome the GPU memory consumption problem, we propose Curvature Sampling to sample the more informative points/pixels for pre-training. 2) We propose to use learnable prototypes to represent parts of the scenes in a common feature space and bring the idea of swapping prototype assignment prediction to learn the interaction between the two modalities. 3) To further optimize learnable prototypes, we propose an Expectation-Maximization training scheme to maximize the similarity between embeddings and prototypes, followed by a Gram Matrix Regularization Loss to avoid collapse. Experiment results on NuScenes show that CLAP achieves 300% more performance gain as compared to previous SOTA 3D pre-training method via differentiable rendering. Codes and models will be released.
Constructing datasets representative of the target domain is essential for training effective machine learning models. Active learning (AL) is a promising method that iteratively extends training data to enhance model performance while minimizing data acquisition costs. However, current AL workflows often require human intervention and lack parallelism, leading to inefficiencies and underutilization of modern computational resources. In this work, we introduce PAL, an automated, modular, and parallel active learning library that integrates AL tasks and manages their execution and communication on shared- and distributed-memory systems using the Message Passing Interface (MPI). PAL provides users with the flexibility to design and customize all components of their active learning scenarios, including machine learning models with uncertainty estimation, oracles for ground truth labeling, and strategies for exploring the target space. We demonstrate that PAL significantly reduces computational overhead and improves scalability, achieving substantial speed-ups through asynchronous parallelization on CPU and GPU hardware. Applications of PAL to several real-world scenarios - including ground-state reactions in biomolecular systems, excited-state dynamics of molecules, simulations of inorganic clusters, and thermo-fluid dynamics - illustrate its effectiveness in accelerating the development of machine learning models. Our results show that PAL enables efficient utilization of high-performance computing resources in active learning workflows, fostering advancements in scientific research and engineering applications.
The creation of a metric-semantic map, which encodes human-prior knowledge, represents a high-level abstraction of environments. However, constructing such a map poses challenges related to the fusion of multi-modal sensor data, the attainment of real-time mapping performance, and the preservation of structural and semantic information consistency. In this paper, we introduce an online metric-semantic mapping system that utilizes LiDAR-Visual-Inertial sensing to generate a global metric-semantic mesh map of large-scale outdoor environments. Leveraging GPU acceleration, our mapping process achieves exceptional speed, with frame processing taking less than 7ms, regardless of scenario scale. Furthermore, we seamlessly integrate the resultant map into a real-world navigation system, enabling metric-semantic-based terrain assessment and autonomous point-to-point navigation within a campus environment. Through extensive experiments conducted on both publicly available and self-collected datasets comprising 24 sequences, we demonstrate the effectiveness of our mapping and navigation methodologies. Code has been publicly released: https://github.com/gogojjh/cobra
In the field of data-driven 3D shape analysis and generation, the estimation of global topological features from localized representations such as point clouds, voxels, and neural implicit fields is a longstanding challenge. This paper introduces a novel, differentiable algorithm tailored to accurately estimate the global topology of 3D shapes, overcoming the limitations of traditional methods rooted in mesh reconstruction and topological data analysis. The proposed method ensures high accuracy, efficiency, and instant computation with GPU compatibility. It begins with an efficient calculation of the self-adjoint Weingarten map for point clouds and its adaptations for other modalities. The curvatures are then extracted, and their integration over tangent differentiable Voronoi elements is utilized to estimate key topological invariants, including the Euler number and Genus. Additionally, an auto-optimization mechanism is implemented to refine the local moving frames and area elements based on the integrity of topological invariants. Experimental results demonstrate the method's superior performance across various datasets. The robustness and differentiability of the algorithm ensure its seamless integration into deep learning frameworks, offering vast potential for downstream tasks in 3D shape analysis.
Adopting advances in recommendation systems is often challenging in industrial settings due to unique constraints. This paper aims to highlight these constraints through the lens of feature interactions. Feature interactions are critical for accurately predicting user behavior in recommendation systems and online advertising. Despite numerous novel techniques showing superior performance on benchmark datasets like Criteo, their direct application in industrial settings is hindered by constraints such as model latency, GPU memory limitations and model reproducibility. In this paper, we share our learnings from improving feature interactions in Pinterest's Homefeed ranking model under such constraints. We provide details about the specific challenges encountered, the strategies employed to address them, and the trade-offs made to balance performance with practical limitations. Additionally, we present a set of learning experiments that help guide the feature interaction architecture selection. We believe these insights will be useful for engineers who are interested in improving their model through better feature interaction learning.
We introduce Orthus, an autoregressive (AR) transformer that excels in generating images given textual prompts, answering questions based on visual inputs, and even crafting lengthy image-text interleaved contents. Unlike prior arts on unified multimodal modeling, Orthus simultaneously copes with discrete text tokens and continuous image features under the AR modeling principle. The continuous treatment of visual signals minimizes the information loss for both image understanding and generation while the fully AR formulation renders the characterization of the correlation between modalities straightforward. The key mechanism enabling Orthus to leverage these advantages lies in its modality-specific heads -- one regular language modeling (LM) head predicts discrete text tokens and one diffusion head generates continuous image features conditioning on the output of the backbone. We devise an efficient strategy for building Orthus -- by substituting the Vector Quantization (VQ) operation in the existing unified AR model with a soft alternative, introducing a diffusion head, and tuning the added modules to reconstruct images, we can create an Orthus-base model effortlessly (e.g., within mere 72 A100 GPU hours). Orthus-base can further embrace post-training to better model interleaved images and texts. Empirically, Orthus surpasses competing baselines including Show-o and Chameleon across standard benchmarks, achieving a GenEval score of 0.58 and an MME-P score of 1265.8 using 7B parameters. Orthus also shows exceptional mixed-modality generation capabilities, reflecting the potential for handling intricate practical generation tasks.
Deep neural network (DNN)-based policy models like vision-language-action (VLA) models are transformative in automating complex decision-making across applications by interpreting multi-modal data. However, scaling these models greatly increases computational costs, which presents challenges in fields like robot manipulation and autonomous driving that require quick, accurate responses. To address the need for deployment on resource-limited hardware, we propose a new quantization framework for IL-based policy models that fine-tunes parameters to enhance robustness against low-bit precision errors during training, thereby maintaining efficiency and reliability under constrained conditions. Our evaluations with representative robot manipulation for 4-bit weight-quantization on a real edge GPU demonstrate that our framework achieves up to 2.5x speedup and 2.5x energy savings while preserving accuracy. For 4-bit weight and activation quantized self-driving models, the framework achieves up to 3.7x speedup and 3.1x energy saving on a low-end GPU. These results highlight the practical potential of deploying IL-based policy models on resource-constrained devices.
Real-time 2D keypoint detection plays an essential role in computer vision. Although CNN-based and Transformer-based methods have achieved breakthrough progress, they often fail to deliver superior performance and real-time speed. This paper introduces MamKPD, the first efficient yet effective mamba-based pose estimation framework for 2D keypoint detection. The conventional Mamba module exhibits limited information interaction between patches. To address this, we propose a lightweight contextual modeling module (CMM) that uses depth-wise convolutions to model inter-patch dependencies and linear layers to distill the pose cues within each patch. Subsequently, by combining Mamba for global modeling across all patches, MamKPD effectively extracts instances' pose information. We conduct extensive experiments on human and animal pose estimation datasets to validate the effectiveness of MamKPD. Our MamKPD-L achieves 77.3% AP on the COCO dataset with 1492 FPS on an NVIDIA GTX 4090 GPU. Moreover, MamKPD achieves state-of-the-art results on the MPII dataset and competitive results on the AP-10K dataset while saving 85% of the parameters compared to ViTPose. Our project page is available at https://mamkpd.github.io/.
Despite significant advances in vision-language understanding, implementing image segmentation within multimodal architectures remains a fundamental challenge in modern artificial intelligence systems. Existing vision-language models, which primarily rely on backbone architectures or CLIP-based embedding learning, demonstrate inherent limitations in fine-grained spatial localization and operational capabilities. This paper introduces SJTU: Spatial Judgments in Multimodal Models - Towards Unified Segmentation through Coordinate Detection, a framework that leverages spatial coordinate understanding to bridge vision-language interaction and precise segmentation, enabling accurate target identification through natural language instructions. The framework presents an approach for integrating segmentation techniques with vision-language models through spatial inference in multimodal space. By utilizing normalized coordinate detection for bounding boxes and transforming them into actionable segmentation outputs, we establish a connection between spatial and language representations in multimodal architectures. Experimental results demonstrate superior performance across benchmark datasets, achieving IoU scores of 0.5958 on COCO 2017 and 0.6758 on Pascal VOC. Testing on a single NVIDIA RTX 3090 GPU with 512x512 resolution images yields an average inference time of 7 seconds per image, demonstrating the framework's effectiveness in both accuracy and practical deployability. The project code is available at https://github.com/jw-chae/SJTU