HeadInfer: Memory-Efficient LLM Inference by Head-wise Offloading

Transformer-based large language models (LLMs) demonstrate impressive performance in long context generation. Extending the context length has disproportionately shifted the memory footprint of LLMs during inference to the key-value cache (KV cache). In this paper, we propose HEADINFER, which offloads the KV cache to CPU RAM while avoiding the need to fully store the KV cache for any transformer layer on the GPU. HEADINFER employs a fine-grained, head-wise offloading strategy, maintaining only selective attention heads KV cache on the GPU while computing attention output dynamically. Through roofline analysis, we demonstrate that HEADINFER maintains computational efficiency while significantly reducing memory footprint. We evaluate HEADINFER on the Llama-3-8B model with a 1-million-token sequence, reducing the GPU memory footprint of the KV cache from 128 GB to 1 GB and the total GPU memory usage from 207 GB to 17 GB, achieving a 92% reduction compared to BF16 baseline inference. Notably, HEADINFER enables 4-million-token inference with an 8B model on a single consumer GPU with 24GB memory (e.g., NVIDIA RTX 4090) without approximation methods.

Survey and Improvement Strategies for Gene Prioritization with Large Language Models

Rare diseases are challenging to diagnose due to limited patient data and genetic diversity. Despite advances in variant prioritization, many cases remain undiagnosed. While large language models (LLMs) have performed well in medical exams, their effectiveness in diagnosing rare genetic diseases has not been assessed. To identify causal genes, we benchmarked various LLMs for gene prioritization. Using multi-agent and Human Phenotype Ontology (HPO) classification, we categorized patients based on phenotypes and solvability levels. As gene set size increased, LLM performance deteriorated, so we used a divide-and-conquer strategy to break the task into smaller subsets. At baseline, GPT-4 outperformed other LLMs, achieving near 30% accuracy in ranking causal genes correctly. The multi-agent and HPO approaches helped distinguish confidently solved cases from challenging ones, highlighting the importance of known gene-phenotype associations and phenotype specificity. We found that cases with specific phenotypes or clear associations were more accurately solved. However, we observed biases toward well-studied genes and input order sensitivity, which hindered gene prioritization. Our divide-and-conquer strategy improved accuracy by overcoming these biases. By utilizing HPO classification, novel multi-agent techniques, and our LLM strategy, we improved causal gene identification accuracy compared to our baseline evaluation. This approach streamlines rare disease diagnosis, facilitates reanalysis of unsolved cases, and accelerates gene discovery, supporting the development of targeted diagnostics and therapies.

PLPP: Prompt Learning with Perplexity Is Self-Distillation for Vision-Language Models

Pre-trained Vision-Language (VL) models such as CLIP have demonstrated their excellent performance across numerous downstream tasks. A recent method, Context Optimization (CoOp), further improves the performance of VL models on downstream tasks by introducing prompt learning. CoOp optimizes a set of learnable vectors, aka prompt, and freezes the whole CLIP model. However, relying solely on CLIP loss to fine-tune prompts can lead to models that are prone to overfitting on downstream task. To address this issue, we propose a plug-in prompt-regularization method called PLPP (Prompt Learning with PerPlexity), which use perplexity loss to regularize prompt learning. PLPP designs a two-step operation to compute the perplexity for prompts: (a) calculating cosine similarity between the weight of the embedding layer and prompts to get labels, (b) introducing a language model (LM) head that requires no training behind text encoder to output word probability distribution. Meanwhile, we unveil that the essence of PLPP is inherently a form of self-distillation. To further prevent overfitting as well as to reduce the additional computation introduced by PLPP, we turn the hard label to soft label and choose top-$k$ values for calculating the perplexity loss. For accelerating model convergence, we introduce mutual self-distillation learning, that is perplexity and inverted perplexity loss. The experiments conducted on four classification tasks indicate that PLPP exhibits superior performance compared to existing methods.

COAP: Memory-Efficient Training with Correlation-Aware Gradient Projection

Training large-scale neural networks in vision, and multimodal domains demands substantial memory resources, primarily due to the storage of optimizer states. While LoRA, a popular parameter-efficient method, reduces memory usage, it often suffers from suboptimal performance due to the constraints of low-rank updates. Low-rank gradient projection methods (e.g., GaLore, Flora) reduce optimizer memory by projecting gradients and moment estimates into low-rank spaces via singular value decomposition or random projection. However, they fail to account for inter-projection correlation, causing performance degradation, and their projection strategies often incur high computational costs. In this paper, we present COAP (Correlation-Aware Gradient Projection), a memory-efficient method that minimizes computational overhead while maintaining training performance. Evaluated across various vision, language, and multimodal tasks, COAP outperforms existing methods in both training speed and model performance. For LLaMA-1B, it reduces optimizer memory by 61% with only 2% additional time cost, achieving the same PPL as AdamW. With 8-bit quantization, COAP cuts optimizer memory by 81% and achieves 4x speedup over GaLore for LLaVA-v1.5-7B fine-tuning, while delivering higher accuracy.

Reconciling Semantic Controllability and Diversity for Remote Sensing Image Synthesis with Hybrid Semantic Embedding

Significant advancements have been made in semantic image synthesis in remote sensing. However, existing methods still face formidable challenges in balancing semantic controllability and diversity. In this paper, we present a Hybrid Semantic Embedding Guided Generative Adversarial Network (HySEGGAN) for controllable and efficient remote sensing image synthesis. Specifically, HySEGGAN leverages hierarchical information from a single source. Motivated by feature description, we propose a hybrid semantic Embedding method, that coordinates fine-grained local semantic layouts to characterize the geometric structure of remote sensing objects without extra information. Besides, a Semantic Refinement Network (SRN) is introduced, incorporating a novel loss function to ensure fine-grained semantic feedback. The proposed approach mitigates semantic confusion and prevents geometric pattern collapse. Experimental results indicate that the method strikes an excellent balance between semantic controllability and diversity. Furthermore, HySEGGAN significantly improves the quality of synthesized images and achieves state-of-the-art performance as a data augmentation technique across multiple datasets for downstream tasks.

Exploring Accuracy-Fairness Trade-off in Large Language Models

Large Language Models (LLMs) have made significant strides in the field of artificial intelligence, showcasing their ability to interact with humans and influence human cognition through information dissemination. However, recent studies have brought to light instances of bias inherent within these LLMs, presenting a critical issue that demands attention. In our research, we delve deeper into the intricate challenge of harmonising accuracy and fairness in the enhancement of LLMs. While improving accuracy can indeed enhance overall LLM performance, it often occurs at the expense of fairness. Overemphasising optimisation of one metric invariably leads to a significant degradation of the other. This underscores the necessity of taking into account multiple considerations during the design and optimisation phases of LLMs. Therefore, we advocate for reformulating the LLM training process as a multi-objective learning task. Our investigation reveals that multi-objective evolutionary learning (MOEL) methodologies offer promising avenues for tackling this challenge. Our MOEL framework enables the simultaneous optimisation of both accuracy and fairness metrics, resulting in a Pareto-optimal set of LLMs. In summary, our study sheds valuable lights on the delicate equilibrium between accuracy and fairness within LLMs, which is increasingly significant for their real-world applications. By harnessing MOEL, we present a promising pathway towards fairer and more efficacious AI technologies.

An Exploration of Higher Education Course Evaluation by Large Language Models

Course evaluation is a critical component in higher education pedagogy. It not only serves to identify limitations in existing course designs and provide a basis for curricular innovation, but also to offer quantitative insights for university administrative decision-making. Traditional evaluation methods, primarily comprising student surveys, instructor self-assessments, and expert reviews, often encounter challenges, including inherent subjectivity, feedback delays, inefficiencies, and limitations in addressing innovative teaching approaches. Recent advancements in large language models (LLMs) within artificial intelligence (AI) present promising new avenues for enhancing course evaluation processes. This study explores the application of LLMs in automated course evaluation from multiple perspectives and conducts rigorous experiments across 100 courses at a major university in China. The findings indicate that: (1) LLMs can be an effective tool for course evaluation; (2) their effectiveness is contingent upon appropriate fine-tuning and prompt engineering; and (3) LLM-generated evaluation results demonstrate a notable level of rationality and interpretability.

MoE-I$^2$: Compressing Mixture of Experts Models through Inter-Expert Pruning and Intra-Expert Low-Rank Decomposition

The emergence of Mixture of Experts (MoE) LLMs has significantly advanced the development of language models. Compared to traditional LLMs, MoE LLMs outperform traditional LLMs by achieving higher performance with considerably fewer activated parameters. Despite this efficiency, their enormous parameter size still leads to high deployment costs. In this paper, we introduce a two-stage compression method tailored for MoE to reduce the model size and decrease the computational cost. First, in the inter-expert pruning stage, we analyze the importance of each layer and propose the Layer-wise Genetic Search and Block-wise KT-Reception Field with the non-uniform pruning ratio to prune the individual expert. Second, in the intra-expert decomposition stage, we apply the low-rank decomposition to further compress the parameters within the remaining experts. Extensive experiments on Qwen1.5-MoE-A2.7B, DeepSeek-V2-Lite, and Mixtral-8$\times$7B demonstrate that our proposed methods can both reduce the model size and enhance inference efficiency while maintaining performance in various zero-shot tasks. The code will be available at \url{https://github.com/xiaochengsky/MoEI-2.git}

RescueADI: Adaptive Disaster Interpretation in Remote Sensing Images with Autonomous Agents

Current methods for disaster scene interpretation in remote sensing images (RSIs) mostly focus on isolated tasks such as segmentation, detection, or visual question-answering (VQA). However, current interpretation methods often fail at tasks that require the combination of multiple perception methods and specialized tools. To fill this gap, this paper introduces Adaptive Disaster Interpretation (ADI), a novel task designed to solve requests by planning and executing multiple sequentially correlative interpretation tasks to provide a comprehensive analysis of disaster scenes. To facilitate research and application in this area, we present a new dataset named RescueADI, which contains high-resolution RSIs with annotations for three connected aspects: planning, perception, and recognition. The dataset includes 4,044 RSIs, 16,949 semantic masks, 14,483 object bounding boxes, and 13,424 interpretation requests across nine challenging request types. Moreover, we propose a new disaster interpretation method employing autonomous agents driven by large language models (LLMs) for task planning and execution, proving its efficacy in handling complex disaster interpretations. The proposed agent-based method solves various complex interpretation requests such as counting, area calculation, and path-finding without human intervention, which traditional single-task approaches cannot handle effectively. Experimental results on RescueADI demonstrate the feasibility of the proposed task and show that our method achieves an accuracy 9% higher than existing VQA methods, highlighting its advantages over conventional disaster interpretation approaches. The dataset will be publicly available.

An Efficient Real-Time Object Detection Framework on Resource-Constricted Hardware Devices via Software and Hardware Co-design

The fast development of object detection techniques has attracted attention to developing efficient Deep Neural Networks (DNNs). However, the current state-of-the-art DNN models can not provide a balanced solution among accuracy, speed, and model size. This paper proposes an efficient real-time object detection framework on resource-constrained hardware devices through hardware and software co-design. The Tensor Train (TT) decomposition is proposed for compressing the YOLOv5 model. By unitizing the unique characteristics given by the TT decomposition, we develop an efficient hardware accelerator based on FPGA devices. Experimental results show that the proposed method can significantly reduce the model size and improve the execution time.