Mixture-of-Experts large language models (MoE-LLMs) marks a significant step forward of language models, however, they encounter two critical challenges in practice: 1) expert parameters lead to considerable memory consumption and loading latency; and 2) the current activated experts are redundant, as many tokens may only require a single expert. Motivated by these issues, we investigate the MoE-LLMs and make two key observations: a) different experts exhibit varying behaviors on activation reconstruction error, routing scores, and activated frequencies, highlighting their differing importance, and b) not all tokens are equally important-only a small subset is critical. Building on these insights, we propose MC-MoE, a training-free Mixture-Compressor for MoE-LLMs, which leverages the significance of both experts and tokens to achieve an extreme compression. First, to mitigate storage and loading overheads, we introduce Pre-Loading Mixed-Precision Quantization (PMQ), which formulates the adaptive bit-width allocation as a Linear Programming (LP) problem, where the objective function balances multi-factors reflecting the importance of each expert. Additionally, we develop Online Dynamic Pruning (ODP), which identifies important tokens to retain and dynamically select activated experts for other tokens during inference to optimize efficiency while maintaining performance. Our MC-MoE integrates static quantization and dynamic pruning to collaboratively achieve extreme compression for MoE-LLMs with less accuracy loss, ensuring an optimal trade-off between performance and efficiency. Extensive experiments confirm the effectiveness of our approach. During dynamic inference, we further reduce activated parameters by 15%, with a performance drop of less than 0.6%. Remarkably, MC-MoE even surpasses floating-point 13b dense LLMs with significantly smaller parameter sizes, suggesting that mixture compression in MoE-LLMs has the potential to outperform both comparable and larger dense LLMs.

Abstract--Although considerable progress has been obtained in neural network quantization for efficient inference, existing methods are not scalable to heterogeneous devices as one dedicated model needs to be trained, transmitted, and stored for one specific hardware setting, incurring considerable costs in model training and maintenance. In this paper, we study a new vertical-layered representation of neural network weights for encapsulating all quantized models into a single one. It represents weights as a group of bits (i.e., vertical layers) organized from the most significant bit (also called the basic layer) to less significant bits (i.e., enhance layers). Hence, a neural network with an arbitrary quantization precision can be obtained by adding corresponding enhance layers to the basic layer. With this representation, we can theoretically achieve any precision network for on-demand service while only needing to train and maintain one model. However, we empirically find that models obtained with existing quantization methods suffer severe performance degradation if adapted to vertical-layered weight representation. A significant issue is that once a quantization model is well-trained for a certain precision, it is difficult to reuse the weights and transfer them to another precision since essential information lying in the full precision source model is discarded after dedicated training. To this end, we propose a simple once quantization-aware training (QAT) scheme for obtaining high-performance vertical-layered models. Our design incorporates a cascade downsampling mechanism which allows us to obtain multiple quantized networks from one full precision source model by progressively mapping the higher precision weights to their adjacent lower precision counterparts.

Modern deep networks can be better generalized when trained with noisy samples and regularization techniques. Mixup and CutMix have been proven to be effective for data augmentation to help avoid overfitting. Previous Mixup-based methods linearly combine images and labels to generate additional training data. However, this is problematic if the object does not occupy the whole image as we demonstrate in Figure 1. Correctly assigning the label weights is hard even for human beings and there is no clear criterion to measure it. To tackle this problem, in this paper, we propose LUMix, which models such uncertainty by adding label perturbation during training. LUMix is simple as it can be implemented in just a few lines of code and can be universally applied to any deep networks \eg CNNs and Vision Transformers, with minimal computational cost. Extensive experiments show that our LUMix can consistently boost the performance for networks with a wide range of diversity and capacity on ImageNet, \eg $+0.7\%$ for a small model DeiT-S and $+0.6\%$ for a large variant XCiT-L. We also demonstrate that LUMix can lead to better robustness when evaluated on ImageNet-O and ImageNet-A. The source code can be found \href{https://github.com/kevin-ssy/LUMix}{here}