Squeezed Attention: Accelerating Long Context Length LLM Inference

Emerging Large Language Model (LLM) applications require long input prompts to perform complex downstream tasks like document analysis and code generation. For these long context length applications, the length of the input prompt poses a significant challenge in terms of inference efficiency since the inference costs increase linearly with sequence length. However, for many of these applications, much of the context in the prompt is fixed across different user inputs, thereby providing the opportunity to perform offline optimizations to process user inputs quickly, as they are received. In this work, we propose Squeezed Attention as a mechanism to accelerate LLM applications where a large portion of the input prompt is fixed. We first leverage K-means clustering offline to group the keys for the fixed context based on semantic similarity and represent each cluster with a single centroid value. During inference, we compare query tokens from the user input with the centroids to predict which of the keys from the fixed context are semantically relevant and need to be loaded during inference. We then compute exact attention using only these important keys from the fixed context, thereby reducing bandwidth and computational costs. We also extend our method to use a hierarchical centroid lookup to identify important keys, which can reduce the complexity of attention from linear to logarithmic with respect to the context length. We implement optimized Triton kernels for centroid comparison and sparse FlashAttention with important keys, achieving more than 4x speedups during both the prefill and generation phases for long-context inference. Furthermore, we have extensively evaluated our method on various long-context benchmarks including LongBench, where it achieves a 3x reduction in KV cache budget without accuracy loss and up to an 8x reduction with <0.5 point accuracy gap for various models.

Stochastic Communication Avoidance for Recommendation Systems

One of the major bottlenecks for efficient deployment of neural network based recommendation systems is the memory footprint of their embedding tables. Although many neural network based recommendation systems could benefit from the faster on-chip memory access and increased computational power of hardware accelerators, the large embedding tables in these models often cannot fit on the constrained memory of accelerators. Despite the pervasiveness of these models, prior methods in memory optimization and parallelism fail to address the memory and communication costs of large embedding tables on accelerators. As a result, the majority of models are trained on CPUs, while current implementations of accelerators are hindered by issues such as bottlenecks in inter-device communication and main memory lookups. In this paper, we propose a theoretical framework that analyses the communication costs of arbitrary distributed systems that use lookup tables. We use this framework to propose algorithms that maximize throughput subject to memory, computation, and communication constraints. Furthermore, we demonstrate that our method achieves strong theoretical performance across dataset distributions and memory constraints, applicable to a wide range of use cases from mobile federated learning to warehouse-scale computation. We implement our framework and algorithms in PyTorch and achieve up to 6x increases in training throughput on GPU systems over baselines, on the Criteo Terabytes dataset.

DQRM: Deep Quantized Recommendation Models

Large-scale recommendation models are currently the dominant workload for many large Internet companies. These recommenders are characterized by massive embedding tables that are sparsely accessed by the index for user and item features. The size of these 1TB+ tables imposes a severe memory bottleneck for the training and inference of recommendation models. In this work, we propose a novel recommendation framework that is small, powerful, and efficient to run and train, based on the state-of-the-art Deep Learning Recommendation Model (DLRM). The proposed framework makes inference more efficient on the cloud servers, explores the possibility of deploying powerful recommenders on smaller edge devices, and optimizes the workload of the communication overhead in distributed training under the data parallelism settings. Specifically, we show that quantization-aware training (QAT) can impose a strong regularization effect to mitigate the severe overfitting issues suffered by DLRMs. Consequently, we achieved INT4 quantization of DLRM models without any accuracy drop. We further propose two techniques that improve and accelerate the conventional QAT workload specifically for the embedding tables in the recommendation models. Furthermore, to achieve efficient training, we quantize the gradients of the embedding tables into INT8 on top of the well-supported specified sparsification. We show that combining gradient sparsification and quantization together significantly reduces the amount of communication. Briefly, DQRM models with INT4 can achieve 79.07% accuracy on Kaggle with 0.27 GB model size, and 81.21% accuracy on the Terabyte dataset with 1.57 GB, which even outperform FP32 DLRMs that have much larger model sizes (2.16 GB on Kaggle and 12.58 on Terabyte).

COAT: Compressing Optimizer states and Activation for Memory-Efficient FP8 Training

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.

PixelGaussian: Generalizable 3D Gaussian Reconstruction from Arbitrary Views

We propose PixelGaussian, an efficient feed-forward framework for learning generalizable 3D Gaussian reconstruction from arbitrary views. Most existing methods rely on uniform pixel-wise Gaussian representations, which learn a fixed number of 3D Gaussians for each view and cannot generalize well to more input views. Differently, our PixelGaussian dynamically adapts both the Gaussian distribution and quantity based on geometric complexity, leading to more efficient representations and significant improvements in reconstruction quality. Specifically, we introduce a Cascade Gaussian Adapter to adjust Gaussian distribution according to local geometry complexity identified by a keypoint scorer. CGA leverages deformable attention in context-aware hypernetworks to guide Gaussian pruning and splitting, ensuring accurate representation in complex regions while reducing redundancy. Furthermore, we design a transformer-based Iterative Gaussian Refiner module that refines Gaussian representations through direct image-Gaussian interactions. Our PixelGaussian can effectively reduce Gaussian redundancy as input views increase. We conduct extensive experiments on the large-scale ACID and RealEstate10K datasets, where our method achieves state-of-the-art performance with good generalization to various numbers of views. Code: https://github.com/Barrybarry-Smith/PixelGaussian.

UniDrive: Towards Universal Driving Perception Across Camera Configurations

Vision-centric autonomous driving has demonstrated excellent performance with economical sensors. As the fundamental step, 3D perception aims to infer 3D information from 2D images based on 3D-2D projection. This makes driving perception models susceptible to sensor configuration (e.g., camera intrinsics and extrinsics) variations. However, generalizing across camera configurations is important for deploying autonomous driving models on different car models. In this paper, we present UniDrive, a novel framework for vision-centric autonomous driving to achieve universal perception across camera configurations. We deploy a set of unified virtual cameras and propose a ground-aware projection method to effectively transform the original images into these unified virtual views. We further propose a virtual configuration optimization method by minimizing the expected projection error between original cameras and virtual cameras. The proposed virtual camera projection can be applied to existing 3D perception methods as a plug-and-play module to mitigate the challenges posed by camera parameter variability, resulting in more adaptable and reliable driving perception models. To evaluate the effectiveness of our framework, we collect a dataset on Carla by driving the same routes while only modifying the camera configurations. Experimental results demonstrate that our method trained on one specific camera configuration can generalize to varying configurations with minor performance degradation.

SparseVLM: Visual Token Sparsification for Efficient Vision-Language Model Inference

In vision-language models (VLMs), visual tokens usually consume a significant amount of computational overhead, despite their sparser information density compared to text tokens. To address this, most existing methods learn a network to prune redundant visual tokens and require additional training data. Differently, we propose an efficient training-free token optimization mechanism dubbed SparseVLM without extra parameters or fine-tuning costs. Concretely, given that visual tokens complement text tokens in VLMs for linguistic reasoning, we select visual-relevant text tokens to rate the significance of vision tokens within the self-attention matrix extracted from the VLMs. Then we progressively prune irrelevant tokens. To maximize sparsity while retaining essential information, we introduce a rank-based strategy to adaptively determine the sparsification ratio for each layer, alongside a token recycling method that compresses pruned tokens into more compact representations. Experimental results show that our SparseVLM improves the efficiency of various VLMs across a range of image and video understanding tasks. In particular, LLaVA equipped with SparseVLM reduces 61% to 67% FLOPs with a compression ratio of 78% while maintaining 93% of the accuracy. Our code is available at https://github.com/Gumpest/SparseVLMs.

One Model is All You Need: ByT5-Sanskrit, a Unified Model for Sanskrit NLP Tasks

Morphologically rich languages are notoriously challenging to process for downstream NLP applications. This paper presents a new pretrained language model, ByT5-Sanskrit, designed for NLP applications involving the morphologically rich language Sanskrit. We evaluate ByT5-Sanskrit on established Sanskrit word segmentation tasks, where it outperforms previous data-driven approaches by a considerable margin and matches the performance of the current best lexicon-based model. It is easier to deploy and more robust to data not covered by external linguistic resources. It also achieves new state-of-the-art results in Vedic Sanskrit dependency parsing and OCR post-correction tasks. Additionally, based on the Digital Corpus of Sanskrit, we introduce a novel multitask dataset for the joint training of Sanskrit word segmentation, lemmatization, and morphosyntactic tagging tasks. We fine-tune ByT5-Sanskrit on this dataset, creating a versatile multitask model for various downstream Sanskrit applications. We have used this model in Sanskrit linguistic annotation projects, in information retrieval setups, and as a preprocessing step in a Sanskrit machine translation pipeline. We also show that our approach yields new best scores for lemmatization and dependency parsing of other morphologically rich languages. We thus demonstrate that byte-level pretrained language models can achieve excellent performance for morphologically rich languages, outperforming tokenizer-based models and presenting an important vector of exploration when constructing NLP pipelines for such languages.

RoVi-Aug: Robot and Viewpoint Augmentation for Cross-Embodiment Robot Learning

Scaling up robot learning requires large and diverse datasets, and how to efficiently reuse collected data and transfer policies to new embodiments remains an open question. Emerging research such as the Open-X Embodiment (OXE) project has shown promise in leveraging skills by combining datasets including different robots. However, imbalances in the distribution of robot types and camera angles in many datasets make policies prone to overfit. To mitigate this issue, we propose RoVi-Aug, which leverages state-of-the-art image-to-image generative models to augment robot data by synthesizing demonstrations with different robots and camera views. Through extensive physical experiments, we show that, by training on robot- and viewpoint-augmented data, RoVi-Aug can zero-shot deploy on an unseen robot with significantly different camera angles. Compared to test-time adaptation algorithms such as Mirage, RoVi-Aug requires no extra processing at test time, does not assume known camera angles, and allows policy fine-tuning. Moreover, by co-training on both the original and augmented robot datasets, RoVi-Aug can learn multi-robot and multi-task policies, enabling more efficient transfer between robots and skills and improving success rates by up to 30%.

Efficient and Scalable Estimation of Tool Representations in Vector Space

Recent advancements in function calling and tool use have significantly enhanced the capabilities of large language models (LLMs) by enabling them to interact with external information sources and execute complex tasks. However, the limited context window of LLMs presents challenges when a large number of tools are available, necessitating efficient methods to manage prompt length and maintain accuracy. Existing approaches, such as fine-tuning LLMs or leveraging their reasoning capabilities, either require frequent retraining or incur significant latency overhead. A more efficient solution involves training smaller models to retrieve the most relevant tools for a given query, although this requires high quality, domain-specific data. To address those challenges, we present a novel framework for generating synthetic data for tool retrieval applications and an efficient data-driven tool retrieval strategy using small encoder models. Empowered by LLMs, we create ToolBank, a new tool retrieval dataset that reflects real human user usages. For tool retrieval methodologies, we propose novel approaches: (1) Tool2Vec: usage-driven tool embedding generation for tool retrieval, (2) ToolRefiner: a staged retrieval method that iteratively improves the quality of retrieved tools, and (3) MLC: framing tool retrieval as a multi-label classification problem. With these new methods, we achieve improvements of up to 27.28 in Recall@K on the ToolBench dataset and 30.5 in Recall@K on ToolBank. Additionally, we present further experimental results to rigorously validate our methods. Our code is available at \url{https://github.com/SqueezeAILab/Tool2Vec}