MKQ-BERT: Quantized BERT with 4-bits Weights and Activations

Abstract

Recently, pre-trained Transformer based language models, such as BERT, have shown great superiority over the traditional methods in many Natural Language Processing (NLP) tasks. However, the computational cost for deploying these models is prohibitive on resource-restricted devices. One method to alleviate this computation overhead is to quantize the original model into fewer bits representation, and previous work has proved that we can at most quantize both weights and activations of BERT into 8-bits, without degrading its performance. In this work, we propose MKQ-BERT, which further improves the compression level and uses 4-bits for quantization. In MKQ-BERT, we propose a novel way for computing the gradient of the quantization scale, combined with an advanced distillation strategy. On the one hand, we prove that MKQ-BERT outperforms the existing BERT quantization methods for achieving a higher accuracy under the same compression level. On the other hand, we are the first work that successfully deploys the 4-bits BERT and achieves an end-toend speedup for inference. Our results suggest that we could achieve 5.3x of bits reduction without degrading the model accuracy, and the inference speed of one int4 layer is 15x faster than a float32 layer in Transformer based model.

HadSkip: Homotopic and Adaptive Layer Skipping of Pre-trained Language Models for Efficient Inference

“…Weight quantization involves mapping model weights to low-precision integers and floating-point numbers, making them more hardware-friendly for computation. Specifically, (Xiao et al, 2022) proposed 8-bit quantization for BERT; (Tang et al, 2022) explored how to use 4 bits to quantize BERT; (Bai et al, 2020;Tian et al, 2023; studied how to quantize BERT into 1-bit or 2-bits. Pruning, on the other hand, focuses on setting redundant parameters to zero to create a sparse network, enabling accelerated sparse matrix operations on specific hardware platforms.…”

Understanding INT4 Quantization for Transformer Models: Latency Speedup, Composability, and Failure Cases

“…Wu et al (2022) prove that even the binary network can result in only a small degradation if applying QAT with knowledge distillation (KD) (Hinton et al, 2014) and longer training, but the activations are quantized to INT8 (using INT8 computation,not INT4). Tang et al (2022) are the first to claim to apply W4A4 to BERT for inference with QAT and KD. However, their quantization method fails to enable W4A4 for all but only the last two layers in a four-layer TinyBERT model (otherwise causing drastic accuracy drops).…”

“…2: End-to-end inference time (ms) for running one layer in BERT-base model with different batch size and sequence length on NVIDIA T4 GPUs. Column 2 to 4 are numbers taken fromTang et al (2022). FasterTransformer(FT) requires sequence length to be multiple of 32, thus the inputs in the parenthesis are used to run FasterTransformer.…”