Mixture-of-Experts (MoE) Transformer, the backbone architecture The profound impact of the Transformer architecture in the domain of multiple phenomenal language models, leverages sparsity of machine learning is undeniable, for the fields including by activating only a fraction of model parameters for each input natural language processing [3, 14, 18, 45, 48, 61] and computer token. The sparse structure, while allowing constant time costs, vision [17, 39, 64], to name a few. To further improve the capabilities results in space inefficiency: we still need to load all the model of pre-trained large language models (LLMs), one general parameters during inference. We introduce ResMoE, an innovative strategy is to scale up their parameters. Mixture-of-Experts (MoE) MoE approximation framework that utilizes Wasserstein barycenter [52] extends the traditional feedforward neural network (FFN) layer to extract a common expert (barycenter expert) and approximate by replacing a single multilayer perceptron (MLP) with multiple the residuals between this barycenter expert and the original ones. MLPs, referred to as "experts". While enhancing the performance, ResMoE enhances the space efficiency for inference of large-scale sparse MoE keeps computing costs (FLOPs) comparable to the original MoE Transformers in a one-shot and data-agnostic manner without dense model, as only a few selected experts will be activated retraining while maintaining minimal accuracy loss, thereby each time. The framework of an MoE layer is demonstrated in paving the way for broader accessibility to large language models.

Transformer-based Large Language Models rely critically on KV cache to efficiently handle extended contexts during the decode phase. Yet, the size of the KV cache grows proportionally with the input length, burdening both memory bandwidth and capacity as decoding progresses. To address this challenge, we present RocketKV, a training-free KV cache compression strategy designed specifically to reduce both memory bandwidth and capacity demand of KV cache during the decode phase. RocketKV contains two consecutive stages. In the first stage, it performs coarse-grain KV cache eviction on the input sequence tokens with SnapKV++, a method improved upon SnapKV by introducing adaptive pooling size and full compatibility with grouped-query attention. In the second stage, it adopts a hybrid attention method to conduct fine-grain top-k sparse attention, approximating the attention scores by leveraging both head and sequence dimensional reductions. Combining these two stages, RocketKV achieves significant KV cache fetching bandwidth and storage savings while maintaining comparable accuracy to full KV cache attention. We show that RocketKV provides end-to-end speedup by up to 3$\times$ as well as peak memory reduction by up to 31% in the decode phase on an NVIDIA H100 GPU compared to the full KV cache baseline, while achieving negligible accuracy loss on a variety of long-context tasks.

Quantizing deep neural networks ,reducing the precision (bit-width) of their computations, can remarkably decrease memory usage and accelerate processing, making these models more suitable for large-scale medical imaging applications with limited computational resources. However, many existing methods studied "fake quantization", which simulates lower precision operations during inference, but does not actually reduce model size or improve real-world inference speed. Moreover, the potential of deploying real 3D low-bit quantization on modern GPUs is still unexplored. In this study, we introduce a real post-training quantization (PTQ) framework that successfully implements true 8-bit quantization on state-of-the-art (SOTA) 3D medical segmentation models, i.e., U-Net, SegResNet, SwinUNETR, nnU-Net, UNesT, TransUNet, ST-UNet,and VISTA3D. Our approach involves two main steps. First, we use TensorRT to perform fake quantization for both weights and activations with unlabeled calibration dataset. Second, we convert this fake quantization into real quantization via TensorRT engine on real GPUs, resulting in real-world reductions in model size and inference latency. Extensive experiments demonstrate that our framework effectively performs 8-bit quantization on GPUs without sacrificing model performance. This advancement enables the deployment of efficient deep learning models in medical imaging applications where computational resources are constrained. The code and models have been released, including U-Net, TransUNet pretrained on the BTCV dataset for abdominal (13-label) segmentation, UNesT pretrained on the Whole Brain Dataset for whole brain (133-label) segmentation, and nnU-Net, SegResNet, SwinUNETR and VISTA3D pretrained on TotalSegmentator V2 for full body (104-label) segmentation. https://github.com/hrlblab/PTQ.

Narrow bit-width data formats are key to reducing the computational and storage costs of modern deep learning applications. This paper evaluates Microscaling (MX) data formats that combine a per-block scaling factor with narrow floating-point and integer types for individual elements. MX formats balance the competing needs of hardware efficiency, model accuracy, and user friction. Empirical results on over two dozen benchmarks demonstrate practicality of MX data formats as a drop-in replacement for baseline FP32 for AI inference and training with low user friction. We also show the first instance of training generative language models at sub-8-bit weights, activations, and gradients with minimal accuracy loss and no modifications to the training recipe.

This paper introduces Block Data Representations (BDR), a framework for exploring and evaluating a wide spectrum of narrow-precision formats for deep learning. It enables comparison of popular quantization standards, and through BDR, new formats based on shared microexponents (MX) are identified, which outperform other state-of-the-art quantization approaches, including narrow-precision floating-point and block floating-point. MX utilizes multiple levels of quantization scaling with ultra-fine scaling factors based on shared microexponents in the hardware. The effectiveness of MX is demonstrated on real-world models including large-scale generative pretraining and inferencing, and production-scale recommendation systems.

Outliers in weights and activations pose a key challenge for fixed-point quantization of neural networks. While outliers can be addressed by fine-tuning, this is not practical for machine learning (ML) service providers (e.g., Google, Microsoft) who often receive customers' models without the training data. Specialized hardware for handling outliers can enable low-precision DNNs, but incurs nontrivial area overhead. In this paper, we propose overwrite quantization (OverQ), a novel hardware technique which opportunistically increases bitwidth for outliers by letting them overwrite adjacent values. An FPGA prototype shows OverQ can significantly improve ResNet-18 accuracy at 4 bits while incurring relatively little increase in resource utilization.

Quantization can improve the execution latency and energy efficiency of neural networks on both commodity GPUs and specialized accelerators. The majority of existing literature focuses on training quantized DNNs, while this work examines the less-studied topic of quantizing a floating-point model without (re)training. DNN weights and activations follow a bell-shaped distribution post-training, while practical hardware uses a linear quantization grid. This leads to challenges in dealing with outliers in the distribution. Prior work has addressed this by clipping the outliers or using specialized hardware. In this work, we propose outlier channel splitting (OCS), which duplicates channels containing outliers, then halves the channel values. The network remains functionally identical, but affected outliers are moved toward the center of the distribution. OCS requires no additional training and works on commodity hardware. Experimental evaluation on ImageNet classification and language modeling shows that OCS can outperform state-of-the-art clipping techniques with only minor overhead.

We propose unitary group convolutions (UGConvs), a building block for CNNs which compose a group convolution with unitary transforms in feature space to learn a richer set of representations than group convolution alone. UGConvs generalize two disparate ideas in CNN architecture, channel shuffling (i.e. ShuffleNet) and block-circulant networks (i.e. CirCNN), and provide unifying insights that lead to a deeper understanding of each technique. We experimentally demonstrate that dense unitary transforms can outperform channel shuffling in DNN accuracy. On the other hand, different dense transforms exhibit comparable accuracy performance. Based on these observations we propose HadaNet, a UGConv network using Hadamard transforms. HadaNets achieve similar accuracy to circulant networks with lower computation complexity, and better accuracy than ShuffleNets with the same number of parameters and floating-point multiplies.