Understanding and Overcoming the Challenges of Efficient Transformer Quantization

Abstract

Transformer-based architectures have become the de-facto standard models for a wide range of Natural Language Processing tasks. However, their memory footprint and high latency are prohibitive for efficient deployment and inference on resource-limited devices. In this work, we explore quantization for transformers. We show that transformers have unique quantization challenges -namely, high dynamic activation ranges that are difficult to represent with a low bit fixed-point format. We establish that these activations contain structured outliers in the residual connections that encourage specific attention patterns, such as attending to the special separator token.To combat these challenges, we present three solutions based on post-training quantization and quantization-aware training, each with a different set of compromises for accuracy, model size, and ease of use. In particular, we introduce a novel quantization scheme -per-embedding-group quantization. We demonstrate the effectiveness of our methods on the GLUE benchmark using BERT, establishing state-of-the-art results for post-training quantization. Finally, we show that transformer weights and embeddings can be quantized to ultra-low bit-widths, leading to significant memory savings with a minimum accuracy loss.

WuDaoCorpora: A super large-scale Chinese corpora for pre-training language models

“…This can be attributed to the larger range differences among different channels in large networks compared to smaller networks. As depicted in Table 5, we compared RPTQ with PEG [3]. Due to the original paper was only tested on small models, we applied its method to the OPT model and used the same group settings for PEG as in RPTQ.…”

“…However, PTQ-SL mainly focuses on the quantization of weights in convolutional networks, and does not address the quantization issues of activations. PGQ [3] employs a range-based permutation of the embedding dimensions and share quantization parameters among elements in the same group to address the problem of activation quantization. Nonetheless, it only consider for the dynamic range and utilizes uniformly divided groups, rendering it less efficacious for LLMs.…”

“…We will evaluate our proposed reorder-based post-training quantization (RPTQ) on large language models. As our work focus on processing the problem in quantizing activations, we use GPTQ [12] to quantize the weights in LLMs 3 . We apply static quantization to all the weights and input activations.…”

Greener yet Powerful: Taming Large Code Generation Models with Quantization

“…3) Quantization. This technique uses fewer bits to represent the weights of parameterized functions [17,27].…”

BiT: Robustly Binarized Multi-distilled Transformer

“…Transformer quantization Compared to the CNNs, transformers with attention layers are naturally more challenging to quantize (Bondarenko et al, 2021). Previous research mainly focused on 8-bit quantization (Zafrir et al, 2019; or 4-bit quantization (Shen et al, 2020;Zadeh et al, 2020).…”