As the field of Large Language Models (LLMs) continues to evolve, the context length in inference is steadily growing. Key-Value Cache (KVCache), a crucial component in LLM inference, has now become the primary memory bottleneck due to limited GPU memory. Current methods selectively determine suitable keys and values for self-attention computation in LLMs to address the issue. However, they either fall short in maintaining model quality or result in high serving latency. Drawing inspiration from advanced embedding retrieval techniques used in the database community, we consider the storage and searching of KVCache as a typical embedding retrieval problem. We propose PQCache, which employs Product Quantization (PQ) to manage KVCache, maintaining model quality while ensuring low serving latency. During the prefilling phase, we apply PQ to tokens' keys for each LLM layer and head. During the autoregressive decoding phase, for each newly generated token, we first identify important tokens through Maximum Inner-Product Search (MIPS) using PQ codes and centroids, then fetch the corresponding key-value pairs for self-attention computation. Through meticulous design of overlapping and caching, we minimize any additional computation and communication overhead during both phases. Extensive experiments show that PQCache achieves both effectiveness and efficiency. It maintains model quality even when only 1/5 of the tokens are involved in attention, while attaining acceptable system latency.

The advent of large language models (LLMs) like GPT-4 has catalyzed the exploration of multi-task learning (MTL), in which a single model demonstrates proficiency across diverse tasks. Task arithmetic has emerged as a cost-effective approach for MTL. It enables performance enhancement across multiple tasks by adding their corresponding task vectors to a pre-trained model. However, the current lack of a method that can simultaneously achieve optimal performance, computational efficiency, and data privacy limits their application to LLMs. In this paper, we propose \textbf{M}odel \textbf{E}xclusive \textbf{T}ask \textbf{A}rithmetic for merging \textbf{GPT}-scale models, which formalizes the objective of model merging into a multi-task learning framework, aiming to minimize the average loss difference between the merged model and each individual task model. Since data privacy limits the use of multi-task training data, we leverage LLMs' local linearity and task vectors' orthogonality to separate the data term and scaling coefficients term and derive a model-exclusive task arithmetic method. Our proposed MetaGPT is data-agnostic and bypasses the heavy search process, making it cost-effective and easy to implement for LLMs.Extensive experiments demonstrate that MetaGPT leads to improvements in task arithmetic and achieves state-of-the-art performance on multiple tasks.

Retrieval-augmented generation (RAG) effectively addresses issues of static knowledge and hallucination in large language models. Existing studies mostly focus on question scenarios with clear user intents and concise answers. However, it is prevalent that users issue broad, open-ended queries with diverse sub-intents, for which they desire rich and long-form answers covering multiple relevant aspects. To tackle this important yet underexplored problem, we propose a novel RAG framework, namely RichRAG. It includes a sub-aspect explorer to identify potential sub-aspects of input questions, a multi-faceted retriever to build a candidate pool of diverse external documents related to these sub-aspects, and a generative list-wise ranker, which is a key module to provide the top-k most valuable documents for the final generator. These ranked documents sufficiently cover various query aspects and are aware of the generator's preferences, hence incentivizing it to produce rich and comprehensive responses for users. The training of our ranker involves a supervised fine-tuning stage to ensure the basic coverage of documents, and a reinforcement learning stage to align downstream LLM's preferences to the ranking of documents. Experimental results on two publicly available datasets prove that our framework effectively and efficiently provides comprehensive and satisfying responses to users.

Synthetic data has been proposed as a solution to address the issue of high-quality data scarcity in the training of large language models (LLMs). Studies have shown that synthetic data can effectively improve the performance of LLMs on downstream benchmarks. However, despite its potential benefits, our analysis suggests that there may be inherent flaws in synthetic data. The uniform format of synthetic data can lead to pattern overfitting and cause significant shifts in the output distribution, thereby reducing the model's instruction-following capabilities. Our work delves into these specific flaws associated with question-answer (Q-A) pairs, a prevalent type of synthetic data, and presents a method based on unlearning techniques to mitigate these flaws. The empirical results demonstrate the effectiveness of our approach, which can reverse the instruction-following issues caused by pattern overfitting without compromising performance on benchmarks at relatively low cost. Our work has yielded key insights into the effective use of synthetic data, aiming to promote more robust and efficient LLM training.

Deep Neural Networks (DNNs) excel in various domains but face challenges in providing accurate uncertainty estimates, which are crucial for high-stakes applications. Large Language Models (LLMs) have recently emerged as powerful tools, demonstrating exceptional performance in language tasks. However, traditional calibration metrics such as Expected Calibration Error (ECE) and classwise-ECE (cw-ECE) are inadequate for LLMs due to their vast vocabularies, data complexity, and distributional focus. To address this, we propose a novel calibration concept called full calibration and introduce its corresponding metric, Full-ECE. Full-ECE evaluates the entire predicted probability distribution, offering a more accurate and robust measure of calibration for LLMs.

Transformer-based large language models (LLMs) typically have a limited context window, resulting in significant performance degradation when processing text beyond the length of the context window. Extensive studies have been proposed to extend the context window and achieve length extrapolation of LLMs, but there is still a lack of in-depth interpretation of these approaches. In this study, we explore the positional information within and beyond the context window for deciphering the underlying mechanism of LLMs. By using a mean-based decomposition method, we disentangle positional vectors from hidden states of LLMs and analyze their formation and effect on attention. Furthermore, when texts exceed the context window, we analyze the change of positional vectors in two settings, i.e., direct extrapolation and context window extension. Based on our findings, we design two training-free context window extension methods, positional vector replacement and attention window extension. Experimental results show that our methods can effectively extend the context window length.

Position embedding is a core component of current Large Language Models (LLMs). Rotary position embedding (RoPE), a technique that encodes the position information with a rotation matrix, has been the de facto choice for position embedding in many LLMs, such as the Llama series. RoPE has been further utilized to extend long context capability, which is roughly based on adjusting the \textit{base} parameter of RoPE to mitigate out-of-distribution (OOD) problems in position embedding. However, in this paper, we find that LLMs may obtain a superficial long-context ability based on the OOD theory. We revisit the role of RoPE in LLMs and propose a novel property of long-term decay, we derive that the \textit{base of RoPE bounds context length}: there is an absolute lower bound for the base value to obtain certain context length capability. Our work reveals the relationship between context length and RoPE base both theoretically and empirically, which may shed light on future long context training.

Large language models (LLMs) suffer from low efficiency as the mismatch between the requirement of auto-regressive decoding and the design of most contemporary GPUs. Specifically, billions to trillions of parameters must be loaded to the GPU cache through its limited memory bandwidth for computation, but only a small batch of tokens is actually computed. Consequently, the GPU spends most of its time on memory transfer instead of computation. Recently, parallel decoding, a type of speculative decoding algorithms, is becoming more popular and has demonstrated impressive efficiency improvement in generation. It introduces extra decoding heads to large models, enabling them to predict multiple subsequent tokens simultaneously and verify these candidate continuations in a single decoding step. However, this approach deviates from the training objective of next token prediction used during pre-training, resulting in a low hit rate for candidate tokens. In this paper, we propose a new speculative decoding algorithm, Clover, which integrates sequential knowledge into the parallel decoding process. This enhancement improves the hit rate of speculators and thus boosts the overall efficiency. Clover transmits the sequential knowledge from pre-speculated tokens via the Regressive Connection, then employs an Attention Decoder to integrate these speculated tokens. Additionally, Clover incorporates an Augmenting Block that modifies the hidden states to better align with the purpose of speculative generation rather than next token prediction. The experiment results demonstrate that Clover outperforms the baseline by up to 91% on Baichuan-Small and 146% on Baichuan-Large, respectively, and exceeds the performance of the previously top-performing method, Medusa, by up to 37% on Baichuan-Small and 57% on Baichuan-Large, respectively.

The rapid proliferation of large language models (LLMs) such as GPT-4 and Gemini underscores the intense demand for resources during their training processes, posing significant challenges due to substantial computational and environmental costs. To alleviate this issue, we propose checkpoint merging in pretraining LLM. This method utilizes LLM checkpoints with shared training trajectories, and is rooted in an extensive search space exploration for the best merging weight via Bayesian optimization. Through various experiments, we demonstrate that: (1) Our proposed methodology exhibits the capacity to augment pretraining, presenting an opportunity akin to obtaining substantial benefits at minimal cost; (2) Our proposed methodology, despite requiring a given held-out dataset, still demonstrates robust generalization capabilities across diverse domains, a pivotal aspect in pretraining.

As Large Language Models (LLMs) continue to advance in performance, their size has escalated significantly, with current LLMs containing billions or even trillions of parameters. However, in this study, we discovered that many layers of LLMs exhibit high similarity, and some layers play a negligible role in network functionality. Based on this observation, we define a metric called Block Influence (BI) to gauge the significance of each layer in LLMs. We then propose a straightforward pruning approach: layer removal, in which we directly delete the redundant layers in LLMs based on their BI scores. Experiments demonstrate that our method, which we call ShortGPT, significantly outperforms previous state-of-the-art (SOTA) methods in model pruning. Moreover, ShortGPT is orthogonal to quantization-like methods, enabling further reduction in parameters and computation. The ability to achieve better results through simple layer removal, as opposed to more complex pruning techniques, suggests a high degree of redundancy in the model architecture.