Router-Tuning: A Simple and Effective Approach for Enabling Dynamic-Depth in Transformers

Traditional transformer models often allocate a fixed amount of computational resources to every input token, leading to inefficient and unnecessary computation. To address this, the Mixture of Depths (MoD) was introduced to dynamically adjust the computational depth by skipping less important layers. Despite its promise, current MoD approaches remain under-explored and face two main challenges: (1) \textit{high training costs due to the need to train the entire model along with the routers that determine which layers to skip}, and (2) \textit{the risk of performance degradation when important layers are bypassed}. In response to the first issue, we propose Router-Tuning, a method that fine-tunes only the router on a small dataset, drastically reducing the computational overhead associated with full model training. For the second challenge, we propose MindSkip, which deploys \textit{Attention with Dynamic Depths}. This method preserves the model's performance while significantly enhancing computational and memory efficiency. Extensive experiments demonstrate that our approach delivers competitive results while dramatically improving the computation efficiency, e.g., 21\% speedup and only a 0.2\% performance drop. The code is released at \url{https://github.com/CASE-Lab-UMD/Router-Tuning}.

What Matters in Transformers? Not All Attention is Needed

Scaling Transformer-based large language models (LLMs) has demonstrated promising performance across various tasks. However, this scaling also introduces redundant structures, posing challenges for real-world deployment. Despite some recognition of redundancy in LLMs, the variability of redundancy across different structures, such as MLP and Attention layers, is under-explored. In this work, we investigate the varying redundancy across different modules within Transformers, including Blocks, MLP, and Attention layers, using a similarity-based metric. This metric operates on the premise that redundant structures produce outputs highly similar to their inputs. Surprisingly, while attention layers are essential for transformers and distinguish them from other mainstream architectures, we found that a large proportion of attention layers exhibit excessively high similarity and can be safely pruned without degrading performance, leading to reduced memory and computation costs. Additionally, we further propose a method that jointly drops Attention and MLP layers, achieving improved performance and dropping ratios. Extensive experiments demonstrate the effectiveness of our methods, e.g., Llama-3-70B maintains comparable performance even after pruning half of the attention layers. Our findings provide valuable insights for future network architecture design. The code will be released at: \url{https://github.com/Shwai-He/LLM-Drop}.

Demystifying the Compression of Mixture-of-Experts Through a Unified Framework

Scaling large language models has revolutionized the performance across diverse domains, yet the continual growth in model size poses significant challenges for real-world deployment. The Mixture of Experts (MoE) approach addresses this by dynamically selecting and activating only a subset of experts, significantly reducing computational costs while maintaining high performance. However, MoE introduces potential redundancy (e.g., parameters) and extra costs (e.g., communication overhead). Despite numerous compression techniques developed for mitigating the redundancy in dense models, the compression of MoE remains under-explored. We first bridge this gap with a cutting-edge unified framework that not only seamlessly integrates mainstream compression methods but also helps systematically understand MoE compression. This framework approaches compression from two perspectives: Expert Slimming which compresses individual experts and Expert Trimming which removes structured modules. Within this framework, we explore the optimization space unexplored by existing methods,and further introduce aggressive Expert Trimming techniques, i.e., Layer Drop and Block Drop, to eliminate redundancy at larger scales. Based on these insights,we present a comprehensive recipe to guide practitioners in compressing MoE effectively. Extensive experimental results demonstrate the effectiveness of the compression methods under our framework and the proposed recipe, achieving a 6.05x speedup and only 20.0GB memory usage while maintaining over 92% of performance on Mixtral-8x7B.

Loki: Low-Rank Keys for Efficient Sparse Attention

Inference on large language models can be expensive in terms of the compute and memory costs involved, especially when long sequence lengths are used. In particular, the self-attention mechanism used in such models contributes significantly to these costs, which has resulted in several recent works that propose sparse attention approximations for inference. In this work, we propose to approximate the self-attention computation by focusing on the dimensionality of key vectors computed in the attention block. Our analysis reveals that the key vectors lie in a significantly lower-dimensional space, consistently across several datasets and models. Exploiting this observation, we propose Loki, a novel sparse attention method that ranks and selects tokens in the KV-cache based on attention scores computed in low-dimensional space. Our evaluations show that Loki is able to maintain the efficacy of the models better than other popular approximation methods, while speeding up the attention computation due to reduced data movement (load/store) and compute costs.

RESSA: Repair Sparse Vision-Language Models via Sparse Cross-Modality Adaptation

Vision-Language Models (VLMs), integrating diverse information from multiple modalities, have shown remarkable success across various tasks. However, deploying VLMs, comprising large-scale vision and language models poses challenges in resource-constrained scenarios. While pruning followed by finetuning offers a potential solution to maintain performance with smaller model sizes, its application to VLMs remains relatively unexplored, presenting two main questions: how to distribute sparsity across different modality-specific models, and how to repair the performance of pruned sparse VLMs. To answer the first question, we conducted preliminary studies on VLM pruning and found that pruning vision models and language models with the same sparsity ratios contribute to nearly optimal performance. For the second question, unlike finetuning unimodal sparse models, sparse VLMs involve cross-modality interactions, requiring specialized techniques for post-pruning performance repair. Moreover, while parameter-efficient LoRA finetuning has been proposed to repair the performance of sparse models, a significant challenge of weights merging arises due to the incompatibility of dense LoRA modules with sparse models that destroy the sparsity of pruned models. To tackle these challenges, we propose to Repair Sparse Vision-Language Models via Sparse Cross-modality Adaptation (RESSA). RESSA utilizes cross-modality finetuning to enhance task-specific performance and facilitate knowledge distillation from original dense models. Additionally, we introduce SparseLoRA, which applies sparsity directly to LoRA weights, enabling seamless integration with sparse models. Our experimental results validate the effectiveness of RESSA, showcasing significant enhancements, such as an 11.3\% improvement under 2:4 sparsity and a remarkable 47.6\% enhancement under unstructured 70\% sparsity.

Reformatted Alignment

The quality of finetuning data is crucial for aligning large language models (LLMs) with human values. Current methods to improve data quality are either labor-intensive or prone to factual errors caused by LLM hallucinations. This paper explores elevating the quality of existing instruction data to better align with human values, introducing a simple and effective approach named ReAlign, which reformats the responses of instruction data into a format that better aligns with pre-established criteria and the collated evidence. This approach minimizes human annotation, hallucination, and the difficulty in scaling, remaining orthogonal to existing alignment techniques. Experimentally, ReAlign significantly boosts the general alignment ability, math reasoning, factuality, and readability of the LLMs. Encouragingly, without introducing any additional data or advanced training techniques, and merely by reformatting the response, LLaMA-2-13B's mathematical reasoning ability on GSM8K can be improved from 46.77% to 56.63% in accuracy. Additionally, a mere 5% of ReAlign data yields a 67% boost in general alignment ability measured by the Alpaca dataset. This work highlights the need for further research into the science and mechanistic interpretability of LLMs. We have made the associated code and data publicly accessible to support future studies at https://github.com/GAIR-NLP/ReAlign.

Selective Reflection-Tuning: Student-Selected Data Recycling for LLM Instruction-Tuning

Instruction tuning is critical to large language models (LLMs) for achieving better instruction following and task adaptation capabilities but its success heavily relies on the training data quality. Many recent methods focus on improving the data quality but often overlook the compatibility of the data with the student model being finetuned. This paper introduces Selective Reflection-Tuning, a novel paradigm that synergizes a teacher LLM's reflection and introspection for improving existing data quality with the data selection capability of the student LLM, to automatically refine existing instruction-tuning data. This teacher-student collaboration produces high-quality and student-compatible instruction-response pairs, resulting in sample-efficient instruction tuning and LLMs of superior performance. Selective Reflection-Tuning is a data augmentation and synthesis that generally improves LLM finetuning and self-improvement without collecting brand-new data. We apply our method to Alpaca and WizardLM data and achieve much stronger and top-tier 7B and 13B LLMs. Our codes, models, and data will be released at https://github.com/tianyi-lab/Reflection_Tuning.

Superfiltering: Weak-to-Strong Data Filtering for Fast Instruction-Tuning

Instruction tuning is critical to improve LLMs but usually suffers from low-quality and redundant data. Data filtering for instruction tuning has proved important in improving both the efficiency and performance of the tuning process. But it also leads to extra cost and computation due to the involvement of LLMs in this process. To reduce the filtering cost, we study Superfiltering: Can we use a smaller and weaker model to select data for finetuning a larger and stronger model? Despite the performance gap between weak and strong language models, we find their highly consistent capability to perceive instruction difficulty and data selection results. This enables us to use a much smaller and more efficient model to filter the instruction data used to train a larger language model. Not only does it largely speed up the data filtering, but the filtered-data-finetuned LLM achieves even better performance on standard benchmarks. Extensive experiments validate the efficacy and efficiency of our approach.

Merging Experts into One: Improving Computational Efficiency of Mixture of Experts

Scaling the size of language models usually leads to remarkable advancements in NLP tasks. But it often comes with a price of growing computational cost. Although a sparse Mixture of Experts (MoE) can reduce the cost by activating a small subset of parameters (e.g., one expert) for each input, its computation escalates significantly if increasing the number of activated experts, limiting its practical utility. Can we retain the advantages of adding more experts without substantially increasing the computational costs? In this paper, we first demonstrate the superiority of selecting multiple experts and then propose a computation-efficient approach called \textbf{\texttt{Merging Experts into One}} (MEO), which reduces the computation cost to that of a single expert. Extensive experiments show that MEO significantly improves computational efficiency, e.g., FLOPS drops from 72.0G of vanilla MoE to 28.6G (MEO). Moreover, we propose a token-level attention block that further enhances the efficiency and performance of token-level MEO, e.g., 83.3\% (MEO) vs. 82.6\% (vanilla MoE) average score on the GLUE benchmark. Our code will be released upon acceptance. Code will be released at: \url{https://github.com/Shwai-He/MEO}.

Reflection-Tuning: Data Recycling Improves LLM Instruction-Tuning

Recent advancements in Large Language Models (LLMs) have expanded the horizons of natural language understanding and generation. Notably, the output control and alignment with the input of LLMs can be refined through instruction tuning. However, as highlighted in several studies, low-quality data in the training set are usually detrimental to instruction tuning, resulting in inconsistent or even misleading LLM outputs. We propose a novel method, termed "reflection-tuning," which addresses the problem by self-improvement and judging capabilities of LLMs. This approach utilizes an oracle LLM to recycle the original training data by introspecting and enhancing the quality of instructions and responses in the data. Extensive experiments on widely used evaluation benchmarks show that LLMs trained with our recycled data outperform those trained with existing datasets in various benchmarks.