The inference process for large language models is slow and memory-intensive, with one of the most critical bottlenecks being excessive Key-Value (KV) cache accesses. This paper introduces "Double Sparsity," a novel post-training sparse attention technique designed to alleviate this bottleneck by reducing KV cache access. Double Sparsity combines token sparsity, which focuses on utilizing only the important tokens for computing self-attention, with channel sparsity, an approach that uses important feature channels for identifying important tokens. Our key insight is that the pattern of channel sparsity is relatively static, allowing us to use offline calibration to make it efficient at runtime, thereby enabling accurate and efficient identification of important tokens. Moreover, this method can be combined with offloading to achieve significant memory usage reduction. Experimental results demonstrate that Double Sparsity can achieve $\frac{1}{16}$ token and channel sparsity with minimal impact on accuracy across various tasks, including wiki-2 perplexity, key-value retrieval, and long context benchmarks with models including Llama-2-7B, Llama-2-70B, and Mixtral-8x7B. It brings up to a 14.1$\times$ acceleration in attention operations and a 1.9$\times$ improvement in end-to-end inference on GPUs. With offloading, it achieves a decoding speed acceleration of 16.3$\times$ compared to state-of-the-art solutions at a sequence length of 256K. Our code is publicly available at //github.com/andy-yang-1/DoubleSparse.
相關內容
Batch prompting is a common technique in large language models (LLMs) used to process multiple inputs simultaneously, aiming to improve computational efficiency. However, as batch sizes increase, performance degradation often occurs due to the model's difficulty in handling lengthy context inputs. Existing methods that attempt to mitigate these issues rely solely on batch data arrangement and majority voting rather than improving the design of the batch prompt itself. In this paper, we address these limitations by proposing "Auto-Demo Prompting," a novel approach that leverages the question-output pairs from earlier questions within a batch as demonstrations for subsequent answer inference. We provide a formal theoretical analysis of how Auto-Demo Prompting functions within the autoregressive generation process of LLMs, illustrating how it utilizes prior outputs to optimize the model's internal representations. Our method effectively bridges the gap between batch prompting and few-shot prompting, enhancing performance with only a slight compromise in token usage. Experimental results across five NLP tasks demonstrate its effectiveness in mitigating performance degradation and occasionally outperforming single prompts. Furthermore, it opens new avenues for applying few-shot learning techniques, such as demonstration selection, within batch prompting, making it a robust solution for real-world applications.
Large language models (LLMs) have achieved remarkable success across diverse tasks, yet their inference processes are hindered by substantial time and energy demands due to single-token generation at each decoding step. While previous methods such as speculative decoding mitigate these inefficiencies by producing multiple tokens per step, each token is still generated by its single-token distribution, thereby enhancing speed without improving effectiveness. In contrast, our work simultaneously enhances inference speed and improves the output effectiveness. We consider multi-token joint decoding (MTJD), which generates multiple tokens from their joint distribution at each iteration, theoretically reducing perplexity and enhancing task performance. However, MTJD suffers from the high cost of sampling from the joint distribution of multiple tokens. Inspired by speculative decoding, we introduce multi-token assisted decoding (MTAD), a novel framework designed to accelerate MTJD. MTAD leverages a smaller auxiliary model to approximate the joint distribution of a larger model, incorporating a verification mechanism that not only ensures the accuracy of this approximation, but also improves the decoding efficiency over conventional speculative decoding. Theoretically, we demonstrate that MTAD closely approximates exact MTJD with bounded error. Empirical evaluations using Llama-2 and OPT models ranging from 13B to 70B parameters across various tasks reveal that MTAD reduces perplexity by 21.2% and improves downstream performance compared to standard single-token sampling. Furthermore, MTAD achieves a 1.42x speed-up and consumes 1.54x less energy than conventional speculative decoding methods. These results highlight MTAD's ability to make multi-token joint decoding both effective and efficient, promoting more sustainable and high-performance deployment of LLMs.
Multimodal language modeling constitutes a recent breakthrough which leverages advances in large language models to pretrain capable multimodal models. The integration of natural language during pretraining has been shown to significantly improve learned representations, particularly in computer vision. However, the efficacy of multimodal language modeling in the realm of functional brain data, specifically for advancing pathology detection, remains unexplored. This study pioneers EEG-language models trained on clinical reports and 15000 EEGs. We extend methods for multimodal alignment to this novel domain and investigate which textual information in reports is useful for training EEG-language models. Our results indicate that models learn richer representations from being exposed to a variety of report segments, including the patient's clinical history, description of the EEG, and the physician's interpretation. Compared to models exposed to narrower clinical text information, we find such models to retrieve EEGs based on clinical reports (and vice versa) with substantially higher accuracy. Yet, this is only observed when using a contrastive learning approach. Particularly in regimes with few annotations, we observe that representations of EEG-language models can significantly improve pathology detection compared to those of EEG-only models, as demonstrated by both zero-shot classification and linear probes. In sum, these results highlight the potential of integrating brain activity data with clinical text, suggesting that EEG-language models represent significant progress for clinical applications.
With the rapid advancement of large language models (LLMs), the diversity of multi-LLM tasks and the variability in their pricing structures have become increasingly important, as costs can vary greatly between different LLMs. To tackle these challenges, we introduce the \textit{C2MAB-V}, a \underline{C}ost-effective \underline{C}ombinatorial \underline{M}ulti-armed \underline{B}andit with \underline{V}ersatile reward models for optimal LLM selection and usage. This online model differs from traditional static approaches or those reliant on a single LLM without cost consideration. With multiple LLMs deployed on a scheduling cloud and a local server dedicated to handling user queries, \textit{C2MAB-V} facilitates the selection of multiple LLMs over a combinatorial search space, specifically tailored for various collaborative task types with different reward models. Based on our designed online feedback mechanism and confidence bound technique, \textit{C2MAB-V} can effectively address the multi-LLM selection challenge by managing the exploration-exploitation trade-off across different models, while also balancing cost and reward for diverse tasks. The NP-hard integer linear programming problem for selecting multiple LLMs with trade-off dilemmas is addressed by: i) decomposing the integer problem into a relaxed form by the local server, ii) utilizing a discretization rounding scheme that provides optimal LLM combinations by the scheduling cloud, and iii) continual online updates based on feedback. Theoretically, we prove that \textit{C2MAB-V} offers strict guarantees over versatile reward models, matching state-of-the-art results for regret and violations in some degenerate cases. Empirically, we show that \textit{C2MAB-V} effectively balances performance and cost-efficiency with nine LLMs for three application scenarios.
Knowledge distillation (KD) is known as a promising solution to compress large language models (LLMs) via transferring their knowledge to smaller models. During this process, white-box KD methods usually minimize the distance between the output distributions of the two models so that more knowledge can be transferred. However, in the current white-box KD framework, the output distributions are from the respective output spaces of the two models, using their own prediction heads. We argue that the space discrepancy will lead to low similarity between the teacher model and the student model on both representation and distribution levels. Furthermore, this discrepancy also hinders the KD process between models with different vocabularies, which is common for current LLMs. To address these issues, we propose a dual-space knowledge distillation (DSKD) framework that unifies the output spaces of the two models for KD. On the basis of DSKD, we further develop a cross-model attention mechanism, which can automatically align the representations of the two models with different vocabularies. Thus, our framework is not only compatible with various distance functions for KD (e.g., KL divergence) like the current framework, but also supports KD between any two LLMs regardless of their vocabularies. Experiments on task-agnostic instruction-following benchmarks show that DSKD significantly outperforms the current white-box KD framework with various distance functions, and also surpasses existing KD methods for LLMs with different vocabularies.
Matrix-variate distributions are a recent addition to the model-based clustering field, thereby making it possible to analyze data in matrix form with complex structure such as images and time series. Due to its recent appearance, there is limited literature on matrix-variate data, with even less on dealing with outliers in these models. An approach for clustering matrix-variate normal data with outliers is discussed. The approach, which uses the distribution of subset log-likelihoods, extends the OCLUST algorithm to matrix-variate normal data and uses an iterative approach to detect and trim outliers.
Despite the recent progress in deep learning, most approaches still go for a silo-like solution, focusing on learning each task in isolation: training a separate neural network for each individual task. Many real-world problems, however, call for a multi-modal approach and, therefore, for multi-tasking models. Multi-task learning (MTL) aims to leverage useful information across tasks to improve the generalization capability of a model. This thesis is concerned with multi-task learning in the context of computer vision. First, we review existing approaches for MTL. Next, we propose several methods that tackle important aspects of multi-task learning. The proposed methods are evaluated on various benchmarks. The results show several advances in the state-of-the-art of multi-task learning. Finally, we discuss several possibilities for future work.
Federated Learning (FL) is a decentralized machine-learning paradigm, in which a global server iteratively averages the model parameters of local users without accessing their data. User heterogeneity has imposed significant challenges to FL, which can incur drifted global models that are slow to converge. Knowledge Distillation has recently emerged to tackle this issue, by refining the server model using aggregated knowledge from heterogeneous users, other than directly averaging their model parameters. This approach, however, depends on a proxy dataset, making it impractical unless such a prerequisite is satisfied. Moreover, the ensemble knowledge is not fully utilized to guide local model learning, which may in turn affect the quality of the aggregated model. Inspired by the prior art, we propose a data-free knowledge distillation} approach to address heterogeneous FL, where the server learns a lightweight generator to ensemble user information in a data-free manner, which is then broadcasted to users, regulating local training using the learned knowledge as an inductive bias. Empirical studies powered by theoretical implications show that, our approach facilitates FL with better generalization performance using fewer communication rounds, compared with the state-of-the-art.
Domain shift is a fundamental problem in visual recognition which typically arises when the source and target data follow different distributions. The existing domain adaptation approaches which tackle this problem work in the closed-set setting with the assumption that the source and the target data share exactly the same classes of objects. In this paper, we tackle a more realistic problem of open-set domain shift where the target data contains additional classes that are not present in the source data. More specifically, we introduce an end-to-end Progressive Graph Learning (PGL) framework where a graph neural network with episodic training is integrated to suppress underlying conditional shift and adversarial learning is adopted to close the gap between the source and target distributions. Compared to the existing open-set adaptation approaches, our approach guarantees to achieve a tighter upper bound of the target error. Extensive experiments on three standard open-set benchmarks evidence that our approach significantly outperforms the state-of-the-arts in open-set domain adaptation.
While existing machine learning models have achieved great success for sentiment classification, they typically do not explicitly capture sentiment-oriented word interaction, which can lead to poor results for fine-grained analysis at the snippet level (a phrase or sentence). Factorization Machine provides a possible approach to learning element-wise interaction for recommender systems, but they are not directly applicable to our task due to the inability to model contexts and word sequences. In this work, we develop two Position-aware Factorization Machines which consider word interaction, context and position information. Such information is jointly encoded in a set of sentiment-oriented word interaction vectors. Compared to traditional word embeddings, SWI vectors explicitly capture sentiment-oriented word interaction and simplify the parameter learning. Experimental results show that while they have comparable performance with state-of-the-art methods for document-level classification, they benefit the snippet/sentence-level sentiment analysis.
小貼士
登錄享
相關主題
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191