Ziming Huang

Biao Sun, Ziming Huang, Hanyu Zhao et al.

Inference serving for large language models (LLMs) is the key to unleashing their potential in people's daily lives. However, efficient LLM serving remains challenging today because the requests are inherently heterogeneous and unpredictable in terms of resource and latency requirements, as a result of the diverse applications and the dynamic execution nature of LLMs. Existing systems are fundamentally limited in handling these characteristics and cause problems such as severe queuing delays, poor tail latencies, and SLO violations. We introduce Llumnix, an LLM serving system that reacts to such heterogeneous and unpredictable requests by runtime rescheduling across multiple model instances. Similar to context switching across CPU cores in modern operating systems, Llumnix reschedules requests to improve load balancing and isolation, mitigate resource fragmentation, and differentiate request priorities and SLOs. Llumnix implements the rescheduling with an efficient and scalable live migration mechanism for requests and their in-memory states, and exploits it in a dynamic scheduling policy that unifies the multiple rescheduling scenarios elegantly. Our evaluations show that Llumnix improves tail latencies by an order of magnitude, accelerates high-priority requests by up to 1.5x, and delivers up to 36% cost savings while achieving similar tail latencies, compared against state-of-the-art LLM serving systems. Llumnix is publicly available at https://github.com/AlibabaPAI/llumnix.

Zixuan Yuan, Yada Zhu, Wei Zhang et al.

Earnings call (EC), as a periodic teleconference of a publicly-traded company, has been extensively studied as an essential market indicator because of its high analytical value in corporate fundamentals. The recent emergence of deep learning techniques has shown great promise in creating automated pipelines to benefit the EC-supported financial applications. However, these methods presume all included contents to be informative without refining valuable semantics from long-text transcript and suffer from EC scarcity issue. Meanwhile, these black-box methods possess inherent difficulties in providing human-understandable explanations. To this end, in this paper, we propose a Multi-Domain Transformer-Based Counterfactual Augmentation, named MTCA, to address the above problems. Specifically, we first propose a transformer-based EC encoder to attentively quantify the task-inspired significance of critical EC content for market inference. Then, a multi-domain counterfactual learning framework is developed to evaluate the gradient-based variations after we perturb limited EC informative texts with plentiful cross-domain documents, enabling MTCA to perform unsupervised data augmentation. As a bonus, we discover a way to use non-training data as instance-based explanations for which we show the result with case studies. Extensive experiments on the real-world financial datasets demonstrate the effectiveness of interpretable MTCA for improving the volatility evaluation ability of the state-of-the-art by 14.2\% in accuracy.

Wei Zhang, Ziming Huang, Yada Zhu et al.

In the recent advances of natural language processing, the scale of the state-of-the-art models and datasets is usually extensive, which challenges the application of sample-based explanation methods in many aspects, such as explanation interpretability, efficiency, and faithfulness. In this work, for the first time, we can improve the interpretability of explanations by allowing arbitrary text sequences as the explanation unit. On top of this, we implement a hessian-free method with a model faithfulness guarantee. Finally, to compare our method with the others, we propose a semantic-based evaluation metric that can better align with humans' judgment of explanations than the widely adopted diagnostic or re-training measures. The empirical results on multiple real data sets demonstrate the proposed method's superior performance to popular explanation techniques such as Influence Function or TracIn on semantic evaluation.

Zhuoxuan Jiang, Ziming Huang, Dong Sheng Li et al.

In end-to-end dialogue modeling and agent learning, it is important to (1) effectively learn knowledge from data, and (2) fully utilize heterogeneous information, e.g., dialogue act flow and utterances. However, the majority of existing methods cannot simultaneously satisfy the two conditions. For example, rule definition and data labeling during system design take too much manual work, and sequence-to-sequence methods only model one-side utterance information. In this paper, we propose a novel joint end-to-end model by multi-task representation learning, which can capture the knowledge from heterogeneous information through automatically learning knowledgeable low-dimensional embeddings from data, named with DialogAct2Vec. The model requires little manual work for intervention in system design and we find that the multi-task learning can greatly improve the effectiveness of representation learning. Extensive experiments on a public dataset for restaurant reservation show that the proposed method leads to significant improvements against the state-of-the-art baselines on both the act prediction task and utterance prediction task.

Zhuoxuan Jiang, Xian-Ling Mao, Ziming Huang et al.

Learning an efficient manager of dialogue agent from data with little manual intervention is important, especially for goal-oriented dialogues. However, existing methods either take too many manual efforts (e.g. reinforcement learning methods) or cannot guarantee the dialogue efficiency (e.g. sequence-to-sequence methods). In this paper, we address this problem by proposing a novel end-to-end learning model to train a dialogue agent that can look ahead for several future turns and generate an optimal response to make the dialogue efficient. Our method is data-driven and does not require too much manual work for intervention during system design. We evaluate our method on two datasets of different scenarios and the experimental results demonstrate the efficiency of our model.