Large language models (LLMs) power a new generation of interactive AI applications exemplified by ChatGPT. The interactive nature of these applications demand low job completion time (JCT) for model inference. Existing LLM serving systems use run-to-completion processing for inference jobs, which suffers from head-of-line blocking and long JCT. We present FastServe, a distributed inference serving system for LLMs. FastServe exploits the autoregressive pattern of LLM inference to enable preemption at the granularity of each output token. FastServe uses preemptive scheduling to minimize JCT with a novel skip-join Multi-Level Feedback Queue scheduler. Based on the new semi information-agnostic setting of LLM inference, the scheduler leverages the input length information to assign an appropriate initial queue for each arrival job to join. The higher priority queues than the joined queue are skipped to reduce demotions. We design an efficient GPU memory management mechanism that proactively offloads and uploads intermediate states between GPU memory and host memory for LLM inference. We build a system prototype of FastServe based on NVIDIA FasterTransformer. Experimental results show that compared to the state-of-the-art solution Orca, FastServe improves the average and tail JCT by up to 5.1$\times$ and 6.4$\times$, respectively.
相關內容
Power analysis poses a significant threat to the security of cryptographic algorithms, as it can be leveraged to recover secret keys. While various software-based countermeasures exist to mitigate this non-invasive attack, they often involve a trade-off between time and space constraints. Techniques such as masking and shuffling, while effective, can noticeably impact execution speed and rely heavily on run-time random number generators. On the contrary, internally encoded implementations of block ciphers offer an alternative approach that does not rely on run-time random sources, but it comes with the drawback of requiring substantial memory space to accommodate lookup tables. Internal encoding, commonly employed in white-box cryptography, suffers from a security limitation as it does not effectively protect the secret key against statistical analysis. To overcome this weakness, this paper introduces a secure internal encoding method for an AES implementation. By addressing the root cause of vulnerabilities found in previous encoding methods, we propose a balanced encoding technique that aims to minimize the problematic correlation with key-dependent intermediate values. We analyze the potential weaknesses associated with the balanced encoding and present a method that utilizes complementary sets of lookup tables. In this approach, the size of the lookup tables is approximately 512KB, and the number of table lookups is 1,024. This is comparable to the table size of non-protected white-box AES-128 implementations, while requiring only half the number of lookups. By adopting this method, our aim is to introduce a non-masking technique that mitigates the vulnerability to statistical analysis present in current internally-encoded AES implementations.
Text-to-image generation models represent the next step of evolution in image synthesis, offering a natural way to achieve flexible yet fine-grained control over the result. One emerging area of research is the fast adaptation of large text-to-image models to smaller datasets or new visual concepts. However, many efficient methods of adaptation have a long training time, which limits their practical applications, slows down research experiments, and spends excessive GPU resources. In this work, we study the training dynamics of popular text-to-image personalization methods (such as Textual Inversion or DreamBooth), aiming to speed them up. We observe that most concepts are learned at early stages and do not improve in quality later, but standard model convergence metrics fail to indicate that. Instead, we propose a simple drop-in early stopping criterion that only requires computing the regular training objective on a fixed set of inputs for all training iterations. Our experiments on Stable Diffusion for a range of concepts and for three personalization methods demonstrate the competitive performance of our approach, making adaptation up to 8 times faster with no significant drops in quality.
Fault diagnosis is essential in industrial processes for monitoring the conditions of important machines. With the ever-increasing complexity of working conditions and demand for safety during production and operation, different diagnosis methods are required, and more importantly, an integrated fault diagnosis system that can cope with multiple tasks is highly desired. However, the diagnosis subtasks are often studied separately, and the currently available methods still need improvement for such a generalized system. To address this issue, we propose the Generalized Out-of-distribution Fault Diagnosis (GOOFD) framework to integrate diagnosis subtasks, such as fault detection, fault classification, and novel fault diagnosis. Additionally, a unified fault diagnosis method based on internal contrastive learning is put forward to underpin the proposed generalized framework. The method extracts features utilizing the internal contrastive learning technique and then recognizes the outliers based on the Mahalanobis distance. Experiments are conducted on a simulated benchmark dataset as well as two practical process datasets to evaluate the proposed framework. As demonstrated in the experiments, the proposed method achieves better performance compared with several existing techniques and thus verifies the effectiveness of the proposed framework.
Grey-box fuzzing is the lightweight approach of choice for finding bugs in sequential programs. It provides a balance between efficiency and effectiveness by conducting a biased random search over the domain of program inputs using a feedback function from observed test executions. For distributed system testing, however, the state-of-practice is represented today by only black-box tools that do not attempt to infer and exploit any knowledge of the system's past behaviours to guide the search for bugs. In this work, we present Mallory: the first framework for grey-box fuzz-testing of distributed systems. Unlike popular black-box distributed system fuzzers, such as Jepsen, that search for bugs by randomly injecting network partitions and node faults or by following human-defined schedules, Mallory is adaptive. It exercises a novel metric to learn how to maximize the number of observed system behaviors by choosing different sequences of faults, thus increasing the likelihood of finding new bugs. The key enablers for our approach are the new ideas of timeline-driven testing and timeline abstraction that provide the feedback function guiding a biased random search for failures. Mallory dynamically constructs Lamport timelines of the system behaviour, abstracts these timelines into happens-before summaries, and introduces faults guided by its real-time observation of the summaries. We have evaluated Mallory on a diverse set of widely-used industrial distributed systems. Compared to the start-of-the-art black-box fuzzer Jepsen, Mallory explores more behaviours and takes less time to find bugs. Mallory discovered 22 zero-day bugs (of which 18 were confirmed by developers), including 10 new vulnerabilities, in rigorously-tested distributed systems such as Braft, Dqlite, and Redis. 6 new CVEs have been assigned.
Using large language models (LLMs) for source code has recently gained attention. LLMs, such as Transformer-based models like Codex and ChatGPT, have been shown to be highly capable of solving a wide range of programming problems. However, the extent to which LLMs understand problem descriptions and generate programs accordingly or just retrieve source code from the most relevant problem in training data based on superficial cues has not been discovered yet. To explore this research question, we conduct experiments to understand the robustness of several popular LLMs, CodeGen and GPT-3.5 series models, capable of tackling code generation tasks in introductory programming problems. Our experimental results show that CodeGen and Codex are sensitive to the superficial modifications of problem descriptions and significantly impact code generation performance. Furthermore, we observe that Codex relies on variable names, as randomized variables decrease the solved rate significantly. However, the state-of-the-art (SOTA) models, such as InstructGPT and ChatGPT, show higher robustness to superficial modifications and have an outstanding capability for solving programming problems. This highlights the fact that slight modifications to the prompts given to the LLMs can greatly affect code generation performance, and careful formatting of prompts is essential for high-quality code generation, while the SOTA models are becoming more robust to perturbations.
Researchers have proposed a wide range of ransomware detection and analysis schemes. However, most of these efforts have focused on older families targeting Windows 7/8 systems. Hence there is a critical need to develop efficient solutions to tackle the latest threats, many of which may have relatively fewer samples to analyze. This paper presents a machine learning (ML) framework for early ransomware detection and attribution. The solution pursues a data-centric approach which uses a minimalist ransomware dataset and implements static analysis using portable executable (PE) files. Results for several ML classifiers confirm strong performance in terms of accuracy and zero-day threat detection.
After just a few hundred training updates, a standard probabilistic model for language generation has likely not yet learnt many semantic or syntactic rules of natural language, making it difficult to estimate the probability distribution over next tokens. Yet around this point, these models have identified a simple, loss-minimising behaviour: to output the unigram distribution of the target training corpus. The use of such a heuristic raises the question: Can we initialise our models with this behaviour and save precious compute resources and model capacity? Here we show that we can effectively endow standard neural language generation models with a separate module that reflects unigram frequency statistics as prior knowledge, simply by initialising the bias term in a model's final linear layer with the log-unigram distribution. We use neural machine translation as a test bed for this simple technique and observe that it: (i) improves learning efficiency; (ii) achieves better overall performance; and perhaps most importantly (iii) appears to disentangle strong frequency effects by encouraging the model to specialise in non-frequency-related aspects of language.
As Large Language Models (LLMs) become increasingly integrated into our everyday lives, understanding their ability to comprehend human mental states becomes critical for ensuring effective interactions. However, despite the recent attempts to assess the Theory-of-Mind (ToM) reasoning capabilities of LLMs, the degree to which these models can align with human ToM remains a nuanced topic of exploration. This is primarily due to two distinct challenges: (1) the presence of inconsistent results from previous evaluations, and (2) concerns surrounding the validity of existing evaluation methodologies. To address these challenges, we present a novel framework for procedurally generating evaluations with LLMs by populating causal templates. Using our framework, we create a new social reasoning benchmark (BigToM) for LLMs which consists of 25 controls and 5,000 model-written evaluations. We find that human participants rate the quality of our benchmark higher than previous crowd-sourced evaluations and comparable to expert-written evaluations. Using BigToM, we evaluate the social reasoning capabilities of a variety of LLMs and compare model performances with human performance. Our results suggest that GPT4 has ToM capabilities that mirror human inference patterns, though less reliable, while other LLMs struggle.
The dominating NLP paradigm of training a strong neural predictor to perform one task on a specific dataset has led to state-of-the-art performance in a variety of applications (eg. sentiment classification, span-prediction based question answering or machine translation). However, it builds upon the assumption that the data distribution is stationary, ie. that the data is sampled from a fixed distribution both at training and test time. This way of training is inconsistent with how we as humans are able to learn from and operate within a constantly changing stream of information. Moreover, it is ill-adapted to real-world use cases where the data distribution is expected to shift over the course of a model's lifetime. The first goal of this thesis is to characterize the different forms this shift can take in the context of natural language processing, and propose benchmarks and evaluation metrics to measure its effect on current deep learning architectures. We then proceed to take steps to mitigate the effect of distributional shift on NLP models. To this end, we develop methods based on parametric reformulations of the distributionally robust optimization framework. Empirically, we demonstrate that these approaches yield more robust models as demonstrated on a selection of realistic problems. In the third and final part of this thesis, we explore ways of efficiently adapting existing models to new domains or tasks. Our contribution to this topic takes inspiration from information geometry to derive a new gradient update rule which alleviate catastrophic forgetting issues during adaptation.
As soon as abstract mathematical computations were adapted to computation on digital computers, the problem of efficient representation, manipulation, and communication of the numerical values in those computations arose. Strongly related to the problem of numerical representation is the problem of quantization: in what manner should a set of continuous real-valued numbers be distributed over a fixed discrete set of numbers to minimize the number of bits required and also to maximize the accuracy of the attendant computations? This perennial problem of quantization is particularly relevant whenever memory and/or computational resources are severely restricted, and it has come to the forefront in recent years due to the remarkable performance of Neural Network models in computer vision, natural language processing, and related areas. Moving from floating-point representations to low-precision fixed integer values represented in four bits or less holds the potential to reduce the memory footprint and latency by a factor of 16x; and, in fact, reductions of 4x to 8x are often realized in practice in these applications. Thus, it is not surprising that quantization has emerged recently as an important and very active sub-area of research in the efficient implementation of computations associated with Neural Networks. In this article, we survey approaches to the problem of quantizing the numerical values in deep Neural Network computations, covering the advantages/disadvantages of current methods. With this survey and its organization, we hope to have presented a useful snapshot of the current research in quantization for Neural Networks and to have given an intelligent organization to ease the evaluation of future research in this area.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191