Taming the Titans: A Survey of Efficient LLM Inference Serving

Large Language Models (LLMs) for Generative AI have achieved remarkable progress, evolving into sophisticated and versatile tools widely adopted across various domains and applications. However, the substantial memory overhead caused by their vast number of parameters, combined with the high computational demands of the attention mechanism, poses significant challenges in achieving low latency and high throughput for LLM inference services. Recent advancements, driven by groundbreaking research, have significantly accelerated progress in this field. This paper provides a comprehensive survey of these methods, covering fundamental instance-level approaches, in-depth cluster-level strategies, emerging scenario directions, and other miscellaneous but important areas. At the instance level, we review model placement, request scheduling, decoding length prediction, storage management, and the disaggregation paradigm. At the cluster level, we explore GPU cluster deployment, multi-instance load balancing, and cloud service solutions. For emerging scenarios, we organize the discussion around specific tasks, modules, and auxiliary methods. To ensure a holistic overview, we also highlight several niche yet critical areas. Finally, we outline potential research directions to further advance the field of LLM inference serving.

Kimi-Audio Technical Report

We present Kimi-Audio, an open-source audio foundation model that excels in audio understanding, generation, and conversation. We detail the practices in building Kimi-Audio, including model architecture, data curation, training recipe, inference deployment, and evaluation. Specifically, we leverage a 12.5Hz audio tokenizer, design a novel LLM-based architecture with continuous features as input and discrete tokens as output, and develop a chunk-wise streaming detokenizer based on flow matching. We curate a pre-training dataset that consists of more than 13 million hours of audio data covering a wide range of modalities including speech, sound, and music, and build a pipeline to construct high-quality and diverse post-training data. Initialized from a pre-trained LLM, Kimi-Audio is continual pre-trained on both audio and text data with several carefully designed tasks, and then fine-tuned to support a diverse of audio-related tasks. Extensive evaluation shows that Kimi-Audio achieves state-of-the-art performance on a range of audio benchmarks including speech recognition, audio understanding, audio question answering, and speech conversation. We release the codes, model checkpoints, as well as the evaluation toolkits in https://github.com/MoonshotAI/Kimi-Audio.

MoE Parallel Folding: Heterogeneous Parallelism Mappings for Efficient Large-Scale MoE Model Training with Megatron Core

Mixture of Experts (MoE) models enhance neural network scalability by dynamically selecting relevant experts per input token, enabling larger model sizes while maintaining manageable computation costs. However, efficient training of large-scale MoE models across thousands of GPUs presents significant challenges due to limitations in existing parallelism strategies. We introduce an end-to-end training framework for large-scale MoE models that utilizes five-dimensional hybrid parallelism: Tensor Parallelism, Expert Parallelism, Context Parallelism, Data Parallelism, and Pipeline Parallelism. Central to our approach is MoE Parallel Folding, a novel strategy that decouples the parallelization of attention and MoE layers in Transformer models, allowing each layer type to adopt optimal parallel configurations. Additionally, we develop a flexible token-level dispatcher that supports both token-dropping and token-dropless MoE training across all five dimensions of parallelism. This dispatcher accommodates dynamic tensor shapes and coordinates different parallelism schemes for Attention and MoE layers, facilitating complex parallelism implementations. Our experiments demonstrate significant improvements in training efficiency and scalability. We achieve up to 49.3% Model Flops Utilization (MFU) for the Mixtral 8x22B model and 39.0% MFU for the Qwen2-57B-A14B model on H100 GPUs, outperforming existing methods. The framework scales efficiently up to 1,024 GPUs and maintains high performance with sequence lengths up to 128K tokens, validating its effectiveness for large-scale MoE model training. The code is available in Megatron-Core.

Align in Depth: Defending Jailbreak Attacks via Progressive Answer Detoxification

Large Language Models (LLMs) are vulnerable to jailbreak attacks, which use crafted prompts to elicit toxic responses. These attacks exploit LLMs' difficulty in dynamically detecting harmful intents during the generation process. Traditional safety alignment methods, often relying on the initial few generation steps, are ineffective due to limited computational budget. This paper proposes DEEPALIGN, a robust defense framework that fine-tunes LLMs to progressively detoxify generated content, significantly improving both the computational budget and effectiveness of mitigating harmful generation. Our approach uses a hybrid loss function operating on hidden states to directly improve LLMs' inherent awareness of toxity during generation. Furthermore, we redefine safe responses by generating semantically relevant answers to harmful queries, thereby increasing robustness against representation-mutation attacks. Evaluations across multiple LLMs demonstrate state-of-the-art defense performance against six different attack types, reducing Attack Success Rates by up to two orders of magnitude compared to previous state-of-the-art defense while preserving utility. This work advances LLM safety by addressing limitations of conventional alignment through dynamic, context-aware mitigation.

AVD2: Accident Video Diffusion for Accident Video Description

Traffic accidents present complex challenges for autonomous driving, often featuring unpredictable scenarios that hinder accurate system interpretation and responses. Nonetheless, prevailing methodologies fall short in elucidating the causes of accidents and proposing preventive measures due to the paucity of training data specific to accident scenarios. In this work, we introduce AVD2 (Accident Video Diffusion for Accident Video Description), a novel framework that enhances accident scene understanding by generating accident videos that aligned with detailed natural language descriptions and reasoning, resulting in the contributed EMM-AU (Enhanced Multi-Modal Accident Video Understanding) dataset. Empirical results reveal that the integration of the EMM-AU dataset establishes state-of-the-art performance across both automated metrics and human evaluations, markedly advancing the domains of accident analysis and prevention. Project resources are available at https://an-answer-tree.github.io

Memory Helps, but Confabulation Misleads: Understanding Streaming Events in Videos with MLLMs

Multimodal large language models (MLLMs) have demonstrated strong performance in understanding videos holistically, yet their ability to process streaming videos-videos are treated as a sequence of visual events-remains underexplored. Intuitively, leveraging past events as memory can enrich contextual and temporal understanding of the current event. In this paper, we show that leveraging memories as contexts helps MLLMs better understand video events. However, because such memories rely on predictions of preceding events, they may contain misinformation, leading to confabulation and degraded performance. To address this, we propose a confabulation-aware memory modification method that mitigates confabulated memory for memory-enhanced event understanding.

Multi-Class Traffic Assignment using Multi-View Heterogeneous Graph Attention Networks

Solving traffic assignment problem for large networks is computationally challenging when conventional optimization-based methods are used. In our research, we develop an innovative surrogate model for a traffic assignment when multi-class vehicles are involved. We do so by employing heterogeneous graph neural networks which use a multiple-view graph attention mechanism tailored to different vehicle classes, along with additional links connecting origin-destination pairs. We also integrate the node-based flow conservation law into the loss function. As a result, our model adheres to flow conservation while delivering highly accurate predictions for link flows and utilization ratios. Through numerical experiments conducted on urban transportation networks, we demonstrate that our model surpasses traditional neural network approaches in convergence speed and predictive accuracy in both user equilibrium and system optimal versions of traffic assignment.

FocalPO: Enhancing Preference Optimizing by Focusing on Correct Preference Rankings

Efficient preference optimization algorithms such as Direct Preference Optimization (DPO) have become a popular approach in aligning large language models (LLMs) with human preferences. These algorithms implicitly treat the LLM as a reward model, and focus on training it to correct misranked preference pairs. However, recent work~\citep{chen2024preference} empirically finds that DPO training \textit{rarely improves these misranked preference pairs}, despite its gradient emphasizing on these cases. We introduce FocalPO, a DPO variant that instead \textit{down-weighs} misranked preference pairs and prioritizes enhancing the model's understanding of pairs that it can already rank correctly. Inspired by Focal Loss used in vision tasks, FocalPO achieves this by adding a modulating factor to dynamically scale DPO loss. Our experiment demonstrates that FocalPO surpasses DPO and its variants on popular benchmarks like Alpaca Eval 2.0 using Mistral-Base-7B and Llama-3-Instruct-8B. Additionally, we empirically reveals how FocalPO affects training on correct and incorrect sample groups, further underscoring its effectiveness.

All-domain Moveline Evolution Network for Click-Through Rate Prediction

E-commerce app users exhibit behaviors that are inherently logically consistent. A series of multi-scenario user behaviors interconnect to form the scene-level all-domain user moveline, which ultimately reveals the user's true intention. Traditional CTR prediction methods typically focus on the item-level interaction between the target item and the historically interacted items. However, the scene-level interaction between the target item and the user moveline remains underexplored. There are two challenges when modeling the interaction with preceding all-domain user moveline: (i) Heterogeneity between items and scenes: Unlike traditional user behavior sequences that utilize items as carriers, the user moveline utilizes scenes as carriers. The heterogeneity between items and scenes complicates the process of aligning interactions within a unified representation space. (ii) Temporal misalignment of linked scene-level and item-level behaviors: In the preceding user moveline with a fixed sampling length, certain critical scene-level behaviors are closely linked to subsequent item-level behaviors. However, it is impossible to establish a complete temporal alignment that clearly identifies which specific scene-level behaviors correspond to which item-level behaviors. To address these challenges and pioneer modeling user intent from the perspective of the all-domain moveline, we propose All-domain Moveline Evolution Network (AMEN). AMEN not only transfers interactions between items and scenes to homogeneous representation spaces, but also introduces a Temporal Sequential Pairwise (TSP) mechanism to understand the nuanced associations between scene-level and item-level behaviors, ensuring that the all-domain user moveline differentially influences CTR predictions for user's favored and unfavored items. Online A/B testing demonstrates that our method achieves a +11.6% increase in CTCVR.

Collaborative Contrastive Network for Click-Through Rate Prediction

E-commerce platforms provide entrances for customers to enter mini-apps to meet their specific shopping needs. At the entrance of a mini-app, a trigger item recommended based on customers' historical preferences, is displayed to attract customers to enter the mini-app. Existing Click-Through Rate (CTR) prediction approaches have two significant weaknesses: (i) A portion of customer entries is driven by their interest in the mini-app itself rather than the trigger item. In such cases, approaches highly hinging on the trigger item tend to recommend similar items, thus misunderstanding the customers' real intention; (ii) Approaches that consider customers' intention toward mini-apps, require the regular existence of mini-apps for customers to cultivate routine shopping habits, making such approaches less robust for mini-apps that are available for only short periods (1 or 3 days) in Explosive Promotional Scenarios (EPS), such as the Black Friday and China's Double 11 Shopping Carnival. To address the above-mentioned issues, we introduce a more general and robust CTR prediction approach, dubbed Collaborative Contrastive Network (CCN). Given a user, CCN learns to identify two item clusters that can represent the user's interests and disinterests, via leveraging the collaborative relationship of co-click/co-non-click or the non-collaborative relationship of mono-click as the supervision signal for contrastive learning. This paradigm does not need to explicitly estimate user's binary entry intention and avoids amplifying the impact of the trigger item. Online A/B testing on large-scale real-world data demonstrates that CCN sets a new state-of-the-art performance on Taobao, boosting CTR by 12.3% and order volume by 12.7%.