Transformer-Lite: High-efficiency Deployment of Large Language Models on Mobile Phone GPUs

The Large Language Model (LLM) is widely employed for tasks such as intelligent assistants, text summarization, translation, and multi-modality on mobile phones. However, the current methods for on-device LLM deployment maintain slow inference speed, which causes poor user experience. To facilitate high-efficiency LLM deployment on device GPUs, we propose four optimization techniques: (a) a symbolic expression-based approach to support dynamic shape model inference; (b) operator optimizations and execution priority setting to enhance inference speed and reduce phone lagging; (c) an FP4 quantization method termed M0E4 to reduce dequantization overhead; (d) a sub-tensor-based technique to eliminate the need for copying KV cache after LLM inference. Furthermore, we implement these methods in our mobile inference engine, Transformer-Lite, which is compatible with both Qualcomm and MTK processors. We evaluated Transformer-Lite's performance using LLMs with varied architectures and parameters ranging from 2B to 14B. Specifically, we achieved prefill and decoding speeds of 121 token/s and 14 token/s for ChatGLM2 6B, and 330 token/s and 30 token/s for smaller Gemma 2B, respectively. Compared with CPU-based FastLLM and GPU-based MLC-LLM, our engine attains over 10x speedup for the prefill speed and 2~3x speedup for the decoding speed.

GBG++: A Fast and Stable Granular Ball Generation Method for Classification

Granular ball computing (GBC), as an efficient, robust, and scalable learning method, has become a popular research topic of granular computing. GBC includes two stages: granular ball generation (GBG) and multi-granularity learning based on the granular ball (GB). However, the stability and efficiency of existing GBG methods need to be further improved due to their strong dependence on $k$-means or $k$-division. In addition, GB-based classifiers only unilaterally consider the GB's geometric characteristics to construct classification rules, but the GB's quality is ignored. Therefore, in this paper, based on the attention mechanism, a fast and stable GBG (GBG++) method is proposed first. Specifically, the proposed GBG++ method only needs to calculate the distances from the data-driven center to the undivided samples when splitting each GB, instead of randomly selecting the center and calculating the distances between it to all samples. Moreover, an outlier detection method is introduced to identify local outliers. Consequently, the GBG++ method can significantly improve effectiveness, robustness, and efficiency while being absolutely stable. Second, considering the influence of the sample size within the GB on the GB's quality, based on the GBG++ method, a $k$-nearest neighbors algorithm (GB$k$NN++) which can reduce misclassification at class boundary to some extent is presented. Finally, the experimental results indicate that the proposed method outperforms several existing GB-based classifiers and classical machine learning classifiers on $20$ public benchmark data sets.