Integer Scale: A Free Lunch for Faster Fine-grained Quantization of LLMs

We introduce Integer Scale, a novel post-training quantization scheme for large language models that effectively resolves the inference bottleneck in current fine-grained quantization approaches while maintaining similar accuracies. Integer Scale is a free lunch as it requires no extra calibration or fine-tuning which will otherwise incur additional costs. It can be used plug-and-play for most fine-grained quantization methods. Its integration results in at most 1.85x end-to-end speed boost over the original counterpart with comparable accuracy. Additionally, due to the orchestration of the proposed Integer Scale and fine-grained quantization, we resolved the quantization difficulty for Mixtral-8x7B and LLaMA-3 models with negligible performance degradation, and it comes with an end-to-end speed boost of 2.13x, and 2.31x compared with their FP16 versions respectively.

A Speed Odyssey for Deployable Quantization of LLMs

The large language model era urges faster and less costly inference. Prior model compression works on LLMs tend to undertake a software-centric approach primarily focused on the simulated quantization performance. By neglecting the feasibility of deployment, these approaches are typically disabled in real practice. They used to drastically push down the quantization bit range for a reduced computation which might not be supported by the mainstream hardware, or involve sophisticated algorithms that introduce extra computation or memory access overhead. We argue that pursuing a hardware-centric approach in the construction of quantization algorithms is crucial. In this regard, we are driven to build our compression method on top of hardware awareness, eliminating impractical algorithm choices while maximizing the benefit of hardware acceleration. Our method, OdysseyLLM, comes with a novel W4A8 kernel implementation called FastGEMM and a combined recipe of quantization strategies. Extensive experiments manifest the superiority of our W4A8 method which brings the actual speed boosting up to \textbf{4$\times$} compared to Hugging Face FP16 inference and \textbf{2.23$\times$} vs. the state-of-the-art inference engine TensorRT-LLM in FP16, and \textbf{1.45$\times$} vs. TensorRT-LLM in INT8, yet without substantially harming the performance.

ESimCSE Unsupervised Contrastive Learning Jointly with UDA Semi-Supervised Learning for Large Label System Text Classification Mode

The challenges faced by text classification with large tag systems in natural language processing tasks include multiple tag systems, uneven data distribution, and high noise. To address these problems, the ESimCSE unsupervised comparative learning and UDA semi-supervised comparative learning models are combined through the use of joint training techniques in the models.The ESimCSE model efficiently learns text vector representations using unlabeled data to achieve better classification results, while UDA is trained using unlabeled data through semi-supervised learning methods to improve the prediction performance of the models and stability, and further improve the generalization ability of the model. In addition, adversarial training techniques FGM and PGD are used in the model training process to improve the robustness and reliability of the model. The experimental results show that there is an 8% and 10% accuracy improvement relative to Baseline on the public dataset Ruesters as well as on the operational dataset, respectively, and a 15% improvement in manual validation accuracy can be achieved on the operational dataset, indicating that the method is effective.