Quantizable Transformers: Removing Outliers by Helping Attention Heads Do Nothing

• URL: http://arxiv.org/abs/2306.12929v2
• Date: Thu, 9 Nov 2023 14:05:51 GMT
• Title: Quantizable Transformers: Removing Outliers by Helping Attention Heads Do Nothing
• Authors: Yelysei Bondarenko, Markus Nagel, Tijmen Blankevoort
• Abstract summary: Modern transformer models tend to learn strong outliers in their activations, making them difficult to quantize. We show that strong outliers are related to very specific behavior of attention heads that try to learn a "no-op" or just a partial update of the residual. We propose two simple (independent) modifications to the attention mechanism - clipped softmax and gated attention.
• Score: 18.673619610942197
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Transformer models have been widely adopted in various domains over the last years, and especially large language models have advanced the field of AI significantly. Due to their size, the capability of these networks has increased tremendously, but this has come at the cost of a significant increase in necessary compute. Quantization is one of the most effective ways to reduce the computational time and memory consumption of neural networks. Many studies have shown, however, that modern transformer models tend to learn strong outliers in their activations, making them difficult to quantize. To retain acceptable performance, the existence of these outliers requires activations to be in higher bitwidth or the use of different numeric formats, extra fine-tuning, or other workarounds. We show that strong outliers are related to very specific behavior of attention heads that try to learn a "no-op" or just a partial update of the residual. To achieve the exact zeros needed in the attention matrix for a no-update, the input to the softmax is pushed to be larger and larger during training, causing outliers in other parts of the network. Based on these observations, we propose two simple (independent) modifications to the attention mechanism - clipped softmax and gated attention. We empirically show that models pre-trained using our methods learn significantly smaller outliers while maintaining and sometimes even improving the floating-point task performance. This enables us to quantize transformers to full INT8 quantization of the activations without any additional effort. We demonstrate the effectiveness of our methods on both language models (BERT, OPT) and vision transformers.

Related papers

• RecurFormer: Not All Transformer Heads Need Self-Attention [14.331807060659902]
  Transformer-based large language models (LLMs) excel in modeling complex language patterns but face significant computational costs during inference. We propose RecurFormer, a novel architecture that replaces certain attention heads with linear recurrent neural networks.
  arXiv  Detail & Related papers  (2024-10-10T15:24:12Z)
• QuantAttack: Exploiting Dynamic Quantization to Attack Vision Transformers [29.957089564635083]
  We present QuantAttack, a novel attack that targets the availability of quantized models. We show that carefully crafted adversarial examples, which are designed to exhaust the resources of the operating system, can trigger worst-case performance.
  arXiv  Detail & Related papers  (2023-12-03T18:31:19Z)
• Learning to Grow Pretrained Models for Efficient Transformer Training [72.20676008625641]
  We learn to grow pretrained transformers, where we learn to linearly map the parameters of the smaller model to initialize the larger model. Experiments across both language and vision transformers demonstrate that our learned Linear Growth Operator (LiGO) can save up to 50% computational cost of training from scratch.
  arXiv  Detail & Related papers  (2023-03-02T05:21:18Z)
• Robust representations of oil wells' intervals via sparse attention mechanism [2.604557228169423]
  We introduce the class of efficient Transformers named Regularized Transformers (Reguformers) The focus in our experiments is on oil&gas data, namely, well logs. To evaluate our models for such problems, we work with an industry-scale open dataset consisting of well logs of more than 20 wells.
  arXiv  Detail & Related papers  (2022-12-29T09:56:33Z)
• Outlier Suppression: Pushing the Limit of Low-bit Transformer Language Models [57.933500846742234]
  Recent work recognizes that structured outliers are the critical bottleneck for quantization performance. We propose an outlier suppression framework including two components: Gamma Migration and Token-Wise Clipping. This framework effectively suppresses the outliers and can be used in a plug-and-play mode.
  arXiv  Detail & Related papers  (2022-09-27T12:05:59Z)
• Confident Adaptive Language Modeling [95.45272377648773]
  CALM is a framework for dynamically allocating different amounts of compute per input and generation timestep. We demonstrate the efficacy of our framework in reducing compute -- potential speedup of up to $times 3$ -- while provably maintaining high performance.
  arXiv  Detail & Related papers  (2022-07-14T17:00:19Z)
• Continual Learning with Transformers for Image Classification [12.028617058465333]
  In computer vision, neural network models struggle to continually learn new concepts without forgetting what has been learnt in the past. We develop a solution called Adaptive Distillation of Adapters (ADA), which is developed to perform continual learning. We empirically demonstrate on different classification tasks that this method maintains a good predictive performance without retraining the model.
  arXiv  Detail & Related papers  (2022-06-28T15:30:10Z)
• Mesa: A Memory-saving Training Framework for Transformers [58.78933015299703]
  We present Mesa, a memory-saving training framework for Transformers. Mesa uses exact activations during forward pass while storing a low-precision version of activations to reduce memory consumption during training. Experiments on ImageNet, CIFAR-100 and ADE20K demonstrate that Mesa can reduce half of the memory footprints during training.
  arXiv  Detail & Related papers  (2021-11-22T11:23:01Z)
• When Vision Transformers Outperform ResNets without Pretraining or Strong Data Augmentations [111.44860506703307]
  Vision Transformers (ViTs) and existing VisionNets signal efforts on replacing hand-wired features or inductive throughputs with general-purpose neural architectures. This paper investigates ViTs and Res-Mixers from the lens of loss geometry, intending to improve the models' data efficiency at training and inference. We show that the improved robustness attributes to sparser active neurons in the first few layers. The resultant ViTs outperform Nets of similar size and smoothness when trained from scratch on ImageNet without large-scale pretraining or strong data augmentations.
  arXiv  Detail & Related papers  (2021-06-03T02:08:03Z)
• Train Large, Then Compress: Rethinking Model Size for Efficient Training and Inference of Transformers [94.43313684188819]
  We study the impact of model size in this setting, focusing on Transformer models for NLP tasks that are limited by compute. We first show that even though smaller Transformer models execute faster per iteration, wider and deeper models converge in significantly fewer steps. This leads to an apparent trade-off between the training efficiency of large Transformer models and the inference efficiency of small Transformer models.
  arXiv  Detail & Related papers  (2020-02-26T21:17:13Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.