Gradient sparsification (GS) is an effective method for reducing communication overhead in distributed training. For the first time, we introduce the concept of Multi-dimensional information into GS and propose a new gradient sparsification method named Multi-dimensional information-aware Gradient sparsification (MultiGS), which achieves high compression ratio with negligible accuracy loss and is applicable to mainstream network architectures. MultiGS reconstructs the layer-wise gradient by combining the high-frequency components of the local gradient and the low-frequency components of the sparsified global gradient that effectively addresses the issue of stale gradients and alleviates model bifurcation. Through the convergence proof of MultiGS for smooth non-convex problems and comparison with momentum SGD in convergence speed, we show that such new perspective approach is theoretically reasonable and practically effective. As validated with several mainstream model families (i.e., ResNets, VGGNet, LSTM, Vision Transformer, and Large Language Models), our MultiGS shows better accuracy over previous GS methods. Moreover, empirical results show that when a sufficient number of training nodes are available, MultiGS accelerates the distributed training by more than 3×, which is better than existing sparsification method.

Vision-Language-Action (VLA) systems are crucial for autonomous decision-making in embodied intelligence. While current systems have advanced the instruction-following capabilities, their limited spatial perception often leads to suboptimal performance for mobile manipulation tasks in unstructured environments. To address this challenge, we propose Uni-Sight, an end-to-end VLA system for robust mobile manipulation. Uni-Sight unifies decision-making, perception, and control through joint training, enabling synchronized cross-component optimization. Within the system, we introduce Latent Feature Aligner (LFA) that ensures accurate target localization by aligning multi-view data. Specifically, we develop Domain Transfer Policy (DTP), a hierarchical policy constrained by LiDAR-guided spatial priors, which ensures 3D spatial understanding with limited visual coverage. Extensive experiments on 20 real-world mobile manipulation tasks demonstrate the high task success rate and robust execution performance of Uni-Sight. Our Uni-Sight achieves a 3.04× the success rate of existing methods, and exhibits superior generalization in both long-horizon and zero-shot scenes. Code and dataset are publicly available at https://github.com/trantor2nd/Uni-Sight.

Neural network training is a memory- and compute-intensive task. Quantization, which enables low-bitwidth formats in training, can significantly mitigate the workload. To reduce quantization error, recent methods have developed new data formats and additional pre-processing operations on quantizers. However, it remains quite challenging to achieve high accuracy and efficiency simultaneously. In this paper, we explore sub-8-bit integer training from its essence of gradient descent optimization. Our integer training framework includes two components: ShiftQuant to realize accurate gradient estimation, and L1 normalization to smoothen the loss landscape. ShiftQuant attains performance that approaches the theoretical upper bound of group quantization. Furthermore, it liberates group quantization from inefficient memory rearrangement. The L1 normalization facilitates the implementation of fully quantized normalization layers with impressive convergence accuracy. Our method frees sub-8-bit integer training from pre-processing and supports general devices. This framework achieves negligible accuracy loss across various neural networks and tasks (0.92% on 4-bit ResNets, 0.61% on 6-bit Transformers). The prototypical implementation of ShiftQuant achieves more than 1.85×/15.3% performance improvement on CPU/GPU compared to its FP16 counterparts, and 33.9% resource consumption reduction on FPGA than the FP16 counterparts. The proposed fully-quantized L1 normalization layers achieve more than 35.54% improvement in throughout on CPU compared to traditional L2 normalization layers. Moreover, theoretical analysis verifies the advancement of our method.

Gradient sparsification and quantization offer a promising prospect to alleviate the communication overhead problem in distributed learning. However, direct combination of the two results in suboptimal solutions, due to the fact that sparsification and quantization haven't been learned together. In this paper, we propose Joint Sparsification-Quantization (JointSQ) inspired by the discovery that sparsification can be treated as 0-bit quantization, regardless of architectures. Specifically, we mathematically formu-late JointSQ as a mixed-precision quantization problem, expanding the solution space. It can be solved by the designed MCKP-Greedy algorithm. Theoretical analysis demon-strates the minimal compression noise of JointSQ, and ex-tensive experiments on various network architectures, including CNN, RNN, and Transformer, also validate this point. Under the introduction of computation overhead consistent with or even lower than previous methods, JointSQ achieves a compression ratio of 1000× on different models while maintaining near-lossless accuracy and brings 1.4× to 2.9× speedup over existing methods.