BiSup: Bidirectional Quantization Error Suppression for Large Language Models

As the size and context length of Large Language Models (LLMs) grow, weight-activation quantization has emerged as a crucial technique for efficient deployment of LLMs. Compared to weight-only quantization, weight-activation quantization presents greater challenges due to the presence of outliers in activations. Existing methods have made significant progress by exploring mixed-precision quantization and outlier suppression. However, these methods primarily focus on optimizing the results of single matrix multiplication, neglecting the bidirectional propagation of quantization errors in LLMs. Specifically, errors accumulate vertically within the same token through layers, and diffuse horizontally across different tokens due to self-attention mechanisms. To address this issue, we introduce BiSup, a Bidirectional quantization error Suppression method. By constructing appropriate optimizable parameter spaces, BiSup utilizes a small amount of data for quantization-aware parameter-efficient fine-tuning to suppress the error vertical accumulation. Besides, BiSup employs prompt mixed-precision quantization strategy, which preserves high precision for the key-value cache of system prompts, to mitigate the error horizontal diffusion. Extensive experiments on Llama and Qwen families demonstrate that BiSup can improve performance over two state-of-the-art methods (the average WikiText2 perplexity decreases from 13.26 to 9.41 for Atom and from 14.33 to 7.85 for QuaRot under the W3A3-g128 configuration), further facilitating the practical applications of low-bit weight-activation quantization.

Attribute Simulation for Item Embedding Enhancement in Multi-interest Recommendation

Although multi-interest recommenders have achieved significant progress in the matching stage, our research reveals that existing models tend to exhibit an under-clustered item embedding space, which leads to a low discernibility between items and hampers item retrieval. This highlights the necessity for item embedding enhancement. However, item attributes, which serve as effective and straightforward side information for enhancement, are either unavailable or incomplete in many public datasets due to the labor-intensive nature of manual annotation tasks. This dilemma raises two meaningful questions: 1. Can we bypass manual annotation and directly simulate complete attribute information from the interaction data? And 2. If feasible, how to simulate attributes with high accuracy and low complexity in the matching stage? In this paper, we first establish an inspiring theoretical feasibility that the item-attribute correlation matrix can be approximated through elementary transformations on the item co-occurrence matrix. Then based on formula derivation, we propose a simple yet effective module, SimEmb (Item Embedding Enhancement via Simulated Attribute), in the multi-interest recommendation of the matching stage to implement our findings. By simulating attributes with the co-occurrence matrix, SimEmb discards the item ID-based embedding and employs the attribute-weighted summation for item embedding enhancement. Comprehensive experiments on four benchmark datasets demonstrate that our approach notably enhances the clustering of item embedding and significantly outperforms SOTA models with an average improvement of 25.59% on Recall@20.