We propose LLM-FP4 for quantizing both weights and activations in large language models (LLMs) down to 4-bit floating-point values, in a post-training manner. Existing post-training quantization (PTQ) solutions are primarily integer-based and struggle with bit widths below 8 bits. Compared to integer quantization, floating-point (FP) quantization is more flexible and can better handle long-tail or bell-shaped distributions, and it has emerged as a default choice in many hardware platforms. One characteristic of FP quantization is that its performance largely depends on the choice of exponent bits and clipping range. In this regard, we construct a strong FP-PTQ baseline by searching for the optimal quantization parameters. Furthermore, we observe a high inter-channel variance and low intra-channel variance pattern in activation distributions, which adds activation quantization difficulty. We recognize this pattern to be consistent across a spectrum of transformer models designed for diverse tasks, such as LLMs, BERT, and Vision Transformer models. To tackle this, we propose per-channel activation quantization and show that these additional scaling factors can be reparameterized as exponential biases of weights, incurring a negligible cost. Our method, for the first time, can quantize both weights and activations in the LLaMA-13B to only 4-bit and achieves an average score of 63.1 on the common sense zero-shot reasoning tasks, which is only 5.8 lower than the full-precision model, significantly outperforming the previous state-of-the-art by 12.7 points. Code is available at: https://github.com/nbasyl/LLM-FP4.

The expanding model size and computation of deep neural networks (DNNs) have increased the demand for efficient model deployment methods. Quantization-aware training (QAT) is a representative model compression method to leverage redundancy in weights and activations. However, most existing QAT methods require end-to-end training on the entire dataset, which suffers from long training time and high energy costs. Coreset selection, aiming to improve data efficiency utilizing the redundancy of training data, has also been widely used for efficient training. In this work, we propose a new angle through the coreset selection to improve the training efficiency of quantization-aware training. Based on the characteristics of QAT, we propose two metrics: error vector score and disagreement score, to quantify the importance of each sample during training. Guided by these two metrics of importance, we proposed a quantization-aware adaptive coreset selection (ACS) method to select the data for the current training epoch. We evaluate our method on various networks (ResNet-18, MobileNetV2), datasets(CIFAR-100, ImageNet-1K), and under different quantization settings. Compared with previous coreset selection methods, our method significantly improves QAT performance with different dataset fractions. Our method can achieve an accuracy of 68.39% of 4-bit quantized ResNet-18 on the ImageNet-1K dataset with only a 10% subset, which has an absolute gain of 4.24% compared to the random baseline.

Weight-sharing supernet has become a vital component for performance estimation in the state-of-the-art (SOTA) neural architecture search (NAS) frameworks. Although supernet can directly generate different subnetworks without retraining, there is no guarantee for the quality of these subnetworks because of weight sharing. In NLP tasks such as machine translation and pre-trained language modeling, we observe that given the same model architecture, there is a large performance gap between supernet and training from scratch. Hence, supernet cannot be directly used and retraining is necessary after finding the optimal architectures. In this work, we propose mixture-of-supernets, a generalized supernet formulation where mixture-of-experts (MoE) is adopted to enhance the expressive power of the supernet model, with negligible training overhead. In this way, different subnetworks do not share the model weights directly, but through an architecture-based routing mechanism. As a result, model weights of different subnetworks are customized towards their specific architectures and the weight generation is learned by gradient descent. Compared to existing weight-sharing supernet for NLP, our method can minimize the retraining time, greatly improving training efficiency. In addition, the proposed method achieves the SOTA performance in NAS for building fast machine translation models, yielding better latency-BLEU tradeoff compared to HAT, state-of-the-art NAS for MT. We also achieve the SOTA performance in NAS for building memory-efficient task-agnostic BERT models, outperforming NAS-BERT and AutoDistil in various model sizes.

Weight oscillation is an undesirable side effect of quantization-aware training, in which quantized weights frequently jump between two quantized levels, resulting in training instability and a sub-optimal final model. We discover that the learnable scaling factor, a widely-used $\textit{de facto}$ setting in quantization aggravates weight oscillation. In this study, we investigate the connection between the learnable scaling factor and quantized weight oscillation and use ViT as a case driver to illustrate the findings and remedies. In addition, we also found that the interdependence between quantized weights in $\textit{query}$ and $\textit{key}$ of a self-attention layer makes ViT vulnerable to oscillation. We, therefore, propose three techniques accordingly: statistical weight quantization ($\rm StatsQ$) to improve quantization robustness compared to the prevalent learnable-scale-based method; confidence-guided annealing ($\rm CGA$) that freezes the weights with $\textit{high confidence}$ and calms the oscillating weights; and $\textit{query}$-$\textit{key}$ reparameterization ($\rm QKR$) to resolve the query-key intertwined oscillation and mitigate the resulting gradient misestimation. Extensive experiments demonstrate that these proposed techniques successfully abate weight oscillation and consistently achieve substantial accuracy improvement on ImageNet. Specifically, our 2-bit DeiT-T/DeiT-S algorithms outperform the previous state-of-the-art by 9.8% and 7.7%, respectively. Code and models are available at: https://github.com/nbasyl/OFQ.

Ternary and binary neural networks enable multiplication-free computation and promise multiple orders of magnitude efficiency gains over full-precision networks if implemented on specialized hardware. However, since both the parameter and the output space are highly discretized, such networks have proven very difficult to optimize. The difficulties are compounded for the class of transformer text generation models due to the sensitivity of the attention operation to quantization and the noise-compounding effects of autoregressive decoding in the high-cardinality output space. We approach the problem with a mix of statistics-based quantization for the weights and elastic quantization of the activations and demonstrate the first ternary and binary transformer models on the downstream tasks of summarization and machine translation. Our ternary BART base achieves an R1 score of 41 on the CNN/DailyMail benchmark, which is merely 3.9 points behind the full model while being 16x more efficient. Our binary model, while less accurate, achieves a highly non-trivial score of 35.6. For machine translation, we achieved BLEU scores of 21.7 and 17.6 on the WMT16 En-Ro benchmark, compared with a full precision mBART model score of 26.8. We also compare our approach in the 8-bit activation setting, where our ternary and even binary weight models can match or outperform the best existing 8-bit weight models in the literature. Our code and models are available at: https://github.com/facebookresearch/Ternary_Binary_Transformer

Several post-training quantization methods have been applied to large language models (LLMs), and have been shown to perform well down to 8-bits. We find that these methods break down at lower bit precision, and investigate quantization aware training for LLMs (LLM-QAT) to push quantization levels even further. We propose a data-free distillation method that leverages generations produced by the pre-trained model, which better preserves the original output distribution and allows quantizing any generative model independent of its training data, similar to post-training quantization methods. In addition to quantizing weights and activations, we also quantize the KV cache, which is critical for increasing throughput and support long sequence dependencies at current model sizes. We experiment with LLaMA models of sizes 7B, 13B, and 30B, at quantization levels down to 4-bits. We observe large improvements over training-free methods, especially in the low-bit settings.

Modern pre-trained transformers have rapidly advanced the state-of-the-art in machine learning, but have also grown in parameters and computational complexity, making them increasingly difficult to deploy in resource-constrained environments. Binarization of the weights and activations of the network can significantly alleviate these issues, however, is technically challenging from an optimization perspective. In this work, we identify a series of improvements that enables binary transformers at a much higher accuracy than what was possible previously. These include a two-set binarization scheme, a novel elastic binary activation function with learned parameters, and a method to quantize a network to its limit by successively distilling higher precision models into lower precision students. These approaches allow for the first time, fully binarized transformer models that are at a practical level of accuracy, approaching a full-precision BERT baseline on the GLUE language understanding benchmark within as little as 5.9%.

This paper explores the feasibility of finding an optimal sub-model from a vision transformer and introduces a pure vision transformer slimming (ViT-Slim) framework that can search such a sub-structure from the original model end-to-end across multiple dimensions, including the input tokens, MHSA and MLP modules with state-of-the-art performance. Our method is based on a learnable and unified l1 sparsity constraint with pre-defined factors to reflect the global importance in the continuous searching space of different dimensions. The searching process is highly efficient through a single-shot training scheme. For instance, on DeiT-S, ViT-Slim only takes ~43 GPU hours for searching process, and the searched structure is flexible with diverse dimensionalities in different modules. Then, a budget threshold is employed according to the requirements of accuracy-FLOPs trade-off on running devices, and a re-training process is performed to obtain the final models. The extensive experiments show that our ViT-Slim can compress up to 40% of parameters and 40% FLOPs on various vision transformers while increasing the accuracy by ~0.6% on ImageNet. We also demonstrate the advantage of our searched models on several downstream datasets. Our source code will be publicly available.

This paper aims to explore the feasibility of neural architecture search (NAS) given only a pre-trained model without using any original training data. This is an important circumstance for privacy protection, bias avoidance, etc., in real-world scenarios. To achieve this, we start by synthesizing usable data through recovering the knowledge from a pre-trained deep neural network. Then we use the synthesized data and their predicted soft-labels to guide neural architecture search. We identify that the NAS task requires the synthesized data (we target at image domain here) with enough semantics, diversity, and a minimal domain gap from the natural images. For semantics, we propose recursive label calibration to produce more informative outputs. For diversity, we propose a regional update strategy to generate more diverse and semantically-enriched synthetic data. For minimal domain gap, we use input and feature-level regularization to mimic the original data distribution in latent space. We instantiate our proposed framework with three popular NAS algorithms: DARTS, ProxylessNAS and SPOS. Surprisingly, our results demonstrate that the architectures discovered by searching with our synthetic data achieve accuracy that is comparable to, or even higher than, architectures discovered by searching from the original ones, for the first time, deriving the conclusion that NAS can be done effectively with no need of access to the original or called natural data if the synthesis method is well designed. Our code will be publicly available.

The nonuniform quantization strategy for compressing neural networks usually achieves better performance than its counterpart, i.e., uniform strategy, due to its superior representational capacity. However, many nonuniform quantization methods overlook the complicated projection process in implementing the nonuniformly quantized weights/activations, which incurs non-negligible time and space overhead in hardware deployment. In this study, we propose Nonuniform-to-Uniform Quantization (N2UQ), a method that can maintain the strong representation ability of nonuniform methods while being hardware-friendly and efficient as the uniform quantization for model inference. We achieve this through learning the flexible in-equidistant input thresholds to better fit the underlying distribution while quantizing these real-valued inputs into equidistant output levels. To train the quantized network with learnable input thresholds, we introduce a generalized straight-through estimator (G-STE) for intractable backward derivative calculation w.r.t. threshold parameters. Additionally, we consider entropy preserving regularization to further reduce information loss in weight quantization. Even under this adverse constraint of imposing uniformly quantized weights and activations, our N2UQ outperforms state-of-the-art nonuniform quantization methods by 0.7~1.8% on ImageNet, demonstrating the contribution of N2UQ design. Code will be made publicly available.