Existing methods for integerized training speed up deep learning by using low-bitwidth integerized weights, activations, gradients, and optimizer buffers. However, they overlook the issue of full-precision latent weights, which consume excessive memory to accumulate gradient-based updates for optimizing the integerized weights. In this paper, we propose the first latent weight quantization schema for general integerized training, which minimizes quantization perturbation to training process via residual quantization with optimized dual quantizer. We leverage residual quantization to eliminate the correlation between latent weight and integerized weight for suppressing quantization noise. We further propose dual quantizer with optimal nonuniform codebook to avoid frozen weight and ensure statistically unbiased training trajectory as full-precision latent weight. The codebook is optimized to minimize the disturbance on weight update under importance guidance and achieved with a three-segment polyline approximation for hardware-friendly implementation. Extensive experiments show that the proposed schema allows integerized training with lowest 4-bit latent weight for various architectures including ResNets, MobileNetV2, and Transformers, and yields negligible performance loss in image classification and text generation. Furthermore, we successfully fine-tune Large Language Models with up to 13 billion parameters on one single GPU using the proposed schema.

KV cache stores key and value states from previous tokens to avoid re-computation, yet it demands substantial storage space, especially for long sequences. Adaptive KV cache compression seeks to discern the saliency of tokens, preserving vital information while aggressively compressing those of less importance. However, previous methods of this approach exhibit significant performance degradation at high compression ratios due to inaccuracies in identifying salient tokens. Additionally, the compression process introduces excessive overhead, substantially increasing memory burdens and the generation latency. In this paper, we present ZipCache, an accurate and efficient KV cache quantization method for large language models (LLMs). First, we construct a strong baseline for quantizing KV cache. Through the proposed channel-separable tokenwise quantization scheme, the memory overhead of quantization parameters are substantially reduced compared to fine-grained groupwise quantization. To enhance the compression ratio, we propose normalized attention score as an effective metric for identifying salient tokens by considering the lower triangle characteristics of the attention matrix. The quantization bit-width for each token is then adaptively assigned based on their saliency. Moreover, we develop an efficient approximation method that decouples the saliency metric from full attention scores, enabling compatibility with fast attention implementations like FlashAttention. Extensive experiments demonstrate that ZipCache achieves superior compression ratios, fast generation speed and minimal performance losses compared with previous KV cache compression methods. For instance, when evaluating Mistral-7B model on GSM8k dataset, ZipCache is capable of compressing the KV cache by $4.98\times$, with only a 0.38% drop in accuracy. In terms of efficiency, ZipCache also showcases a 37.3% reduction in prefill-phase latency, a 56.9% reduction in decoding-phase latency, and a 19.8% reduction in GPU memory usage when evaluating LLaMA3-8B model with a input length of 4096. Code is available at https://github.com/ThisisBillhe/ZipCache/.

Diverse human motion prediction (HMP) is a fundamental application in computer vision that has recently attracted considerable interest. Prior methods primarily focus on the stochastic nature of human motion, while neglecting the specific impact of external environment, leading to the pronounced artifacts in prediction when applied to real-world scenarios. To fill this gap, this work introduces a novel task: predicting diverse human motion within real-world 3D scenes. In contrast to prior works, it requires harmonizing the deterministic constraints imposed by the surrounding 3D scenes with the stochastic aspect of human motion. For this purpose, we propose DiMoP3D, a diverse motion prediction framework with 3D scene awareness, which leverages the 3D point cloud and observed sequence to generate diverse and high-fidelity predictions. DiMoP3D is able to comprehend the 3D scene, and determines the probable target objects and their desired interactive pose based on the historical motion. Then, it plans the obstacle-free trajectory towards these interested objects, and generates diverse and physically-consistent future motions. On top of that, DiMoP3D identifies deterministic factors in the scene and integrates them into the stochastic modeling, making the diverse HMP in realistic scenes become a controllable stochastic generation process. On two real-captured benchmarks, DiMoP3D has demonstrated significant improvements over state-of-the-art methods, showcasing its effectiveness in generating diverse and physically-consistent motion predictions within real-world 3D environments.

Binary neural networks leverage Sign function to binarize weights and activations, which require gradient estimators to overcome its non-differentiability and will inevitably bring gradient errors during backpropagation. Although many hand-designed soft functions have been proposed as gradient estimators to better approximate gradients, their mechanism is not clear and there are still huge performance gaps between binary models and their full-precision counterparts. To address these issues and reduce gradient error, we propose to tackle network binarization as a binary classification problem and use a multi-layer perceptron (MLP) as the classifier in the forward pass and gradient estimator in the backward pass. Benefiting from the MLP’s theoretical capability to fit any continuous function, it can be adaptively learned to binarize networks and backpropagate gradients without any prior knowledge of soft functions. From this perspective, we further empirically justify that even a simple linear function can outperform previous complex soft functions. Extensive experiments demonstrate that the proposed method yields surprising performance both in image classification and human pose estimation tasks. Specifically, we achieve 65.7% top-1 accuracy of ResNet-34 on ImageNet dataset, with an absolute improvement of 2.6%. Moreover, we take binarization as a lightweighting approach for pose estimation models and propose well-designed binary pose estimation networks SBPN and BHRNet. When evaluating on the challenging Microsoft COCO keypoint dataset, the proposed method enables binary networks to achieve a mAP of up to 60.6 for the first time. Experiments conducted on real platforms demonstrate that BNN achieves a better balance between performance and computational complexity, especially when computational resources are extremely low.

Model binarization can significantly compress model size, reduce energy consumption, and accelerate inference through efficient bit-wise operations. Although binarizing convolutional neural networks have been extensively studied, there is little work on exploring binarization of vision Transformers which underpin most recent breakthroughs in visual recognition. To this end, we propose to solve two fundamental challenges to push the horizon of Binary Vision Transformers (BiViT). First, the traditional binary method does not take the long-tailed distribution of softmax attention into consideration, bringing large binarization errors in the attention module. To solve this, we propose Softmax-aware Binarization, which dynamically adapts to the data distribution and reduces the error caused by binarization. Second, to better preserve the information of the pretrained model and restore accuracy, we propose a Cross-layer Binarization scheme that decouples the binarization of self-attention and multi-layer perceptrons (MLPs), and Parameterized Weight Scales which introduce learnable scaling factors for weight binarization. Overall, our method performs favorably against state-of-the-arts by 19.8% on the TinyImageNet dataset. On ImageNet, our BiViT achieves a competitive 75.6% Top-1 accuracy over Swin-S model. Additionally, on COCO object detection, our method achieves an mAP of 40.8 with a Swin-T backbone over Cascade Mask R-CNN framework.