Dobi-SVD: Differentiable SVD for LLM Compression and Some New Perspectives

We provide a new LLM-compression solution via SVD, unlocking new possibilities for LLM compression beyond quantization and pruning. We point out that the optimal use of SVD lies in truncating activations, rather than merely using activations as an optimization distance. Building on this principle, we address three critical challenges in SVD-based LLM compression: including (1) How can we determine the optimal activation truncation position for each weight matrix in LLMs? (2) How can we efficiently reconstruct the weight matrices based on truncated activations? (3) How can we address the inherent "injection" nature that results in the information loss of the SVD? We propose Dobi-SVD, which establishes a new, principled approach to SVD-based LLM compression.

Sparse VideoGen: Accelerating Video Diffusion Transformers with Spatial-Temporal Sparsity

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.

DrivingRecon: Large 4D Gaussian Reconstruction Model For Autonomous Driving

Photorealistic 4D reconstruction of street scenes is essential for developing real-world simulators in autonomous driving. However, most existing methods perform this task offline and rely on time-consuming iterative processes, limiting their practical applications. To this end, we introduce the Large 4D Gaussian Reconstruction Model (DrivingRecon), a generalizable driving scene reconstruction model, which directly predicts 4D Gaussian from surround view videos. To better integrate the surround-view images, the Prune and Dilate Block (PD-Block) is proposed to eliminate overlapping Gaussian points between adjacent views and remove redundant background points. To enhance cross-temporal information, dynamic and static decoupling is tailored to better learn geometry and motion features. Experimental results demonstrate that DrivingRecon significantly improves scene reconstruction quality and novel view synthesis compared to existing methods. Furthermore, we explore applications of DrivingRecon in model pre-training, vehicle adaptation, and scene editing. Our code is available at https://github.com/EnVision-Research/DriveRecon.

Driv3R: Learning Dense 4D Reconstruction for Autonomous Driving

Realtime 4D reconstruction for dynamic scenes remains a crucial challenge for autonomous driving perception. Most existing methods rely on depth estimation through self-supervision or multi-modality sensor fusion. In this paper, we propose Driv3R, a DUSt3R-based framework that directly regresses per-frame point maps from multi-view image sequences. To achieve streaming dense reconstruction, we maintain a memory pool to reason both spatial relationships across sensors and dynamic temporal contexts to enhance multi-view 3D consistency and temporal integration. Furthermore, we employ a 4D flow predictor to identify moving objects within the scene to direct our network focus more on reconstructing these dynamic regions. Finally, we align all per-frame pointmaps consistently to the world coordinate system in an optimization-free manner. We conduct extensive experiments on the large-scale nuScenes dataset to evaluate the effectiveness of our method. Driv3R outperforms previous frameworks in 4D dynamic scene reconstruction, achieving 15x faster inference speed compared to methods requiring global alignment. Code: https://github.com/Barrybarry-Smith/Driv3R.

A Lesson in Splats: Teacher-Guided Diffusion for 3D Gaussian Splats Generation with 2D Supervision

We introduce a diffusion model for Gaussian Splats, SplatDiffusion, to enable generation of three-dimensional structures from single images, addressing the ill-posed nature of lifting 2D inputs to 3D. Existing methods rely on deterministic, feed-forward predictions, which limit their ability to handle the inherent ambiguity of 3D inference from 2D data. Diffusion models have recently shown promise as powerful generative models for 3D data, including Gaussian splats; however, standard diffusion frameworks typically require the target signal and denoised signal to be in the same modality, which is challenging given the scarcity of 3D data. To overcome this, we propose a novel training strategy that decouples the denoised modality from the supervision modality. By using a deterministic model as a noisy teacher to create the noised signal and transitioning from single-step to multi-step denoising supervised by an image rendering loss, our approach significantly enhances performance compared to the deterministic teacher. Additionally, our method is flexible, as it can learn from various 3D Gaussian Splat (3DGS) teachers with minimal adaptation; we demonstrate this by surpassing the performance of two different deterministic models as teachers, highlighting the potential generalizability of our framework. Our approach further incorporates a guidance mechanism to aggregate information from multiple views, enhancing reconstruction quality when more than one view is available. Experimental results on object-level and scene-level datasets demonstrate the effectiveness of our framework.

DeSiRe-GS: 4D Street Gaussians for Static-Dynamic Decomposition and Surface Reconstruction for Urban Driving Scenes

We present DeSiRe-GS, a self-supervised gaussian splatting representation, enabling effective static-dynamic decomposition and high-fidelity surface reconstruction in complex driving scenarios. Our approach employs a two-stage optimization pipeline of dynamic street Gaussians. In the first stage, we extract 2D motion masks based on the observation that 3D Gaussian Splatting inherently can reconstruct only the static regions in dynamic environments. These extracted 2D motion priors are then mapped into the Gaussian space in a differentiable manner, leveraging an efficient formulation of dynamic Gaussians in the second stage. Combined with the introduced geometric regularizations, our method are able to address the over-fitting issues caused by data sparsity in autonomous driving, reconstructing physically plausible Gaussians that align with object surfaces rather than floating in air. Furthermore, we introduce temporal cross-view consistency to ensure coherence across time and viewpoints, resulting in high-quality surface reconstruction. Comprehensive experiments demonstrate the efficiency and effectiveness of DeSiRe-GS, surpassing prior self-supervised arts and achieving accuracy comparable to methods relying on external 3D bounding box annotations. Code is available at \url{https://github.com/chengweialan/DeSiRe-GS}

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.