Search

In this technical report, we tackle the challenges of training large-scale Mixture of Experts (MoE) models, focusing on overcoming cost inefficiency and resource limitations prevalent in such systems. To address these issues, we present two differently sized MoE large language models (LLMs), namely Ling-Lite and Ling-Plus (referred to as "Bailing" in Chinese, spelled B\v{a}il\'ing in Pinyin). Ling-Lite contains 16.8 billion parameters with 2.75 billion activated parameters, while Ling-Plus boasts 290 billion parameters with 28.8 billion activated parameters. Both models exhibit comparable performance to leading industry benchmarks. This report offers actionable insights to improve the efficiency and accessibility of AI development in resource-constrained settings, promoting more scalable and sustainable technologies. Specifically, to reduce training costs for large-scale MoE models, we propose innovative methods for (1) optimization of model architecture and training processes, (2) refinement of training anomaly handling, and (3) enhancement of model evaluation efficiency. Additionally, leveraging high-quality data generated from knowledge graphs, our models demonstrate superior capabilities in tool use compared to other models. Ultimately, our experimental findings demonstrate that a 300B MoE LLM can be effectively trained on lower-performance devices while achieving comparable performance to models of a similar scale, including dense and MoE models. Compared to high-performance devices, utilizing a lower-specification hardware system during the pre-training phase demonstrates significant cost savings, reducing computing costs by approximately 20%. The models can be accessed at https://huggingface.co/inclusionAI., Comment: 34 pages

Image personalization has garnered attention for its ability to customize Text-to-Image generation using only a few reference images. However, a key challenge in image personalization is the issue of conceptual coupling, where the limited number of reference images leads the model to form unwanted associations between the personalization target and other concepts. Current methods attempt to tackle this issue indirectly, leading to a suboptimal balance between text control and personalization fidelity. In this paper, we take a direct approach to the concept coupling problem through statistical analysis, revealing that it stems from two distinct sources of dependence discrepancies. We therefore propose two complementary plug-and-play loss functions: Denoising Decouple Loss and Prior Decouple loss, each designed to minimize one type of dependence discrepancy. Extensive experiments demonstrate that our approach achieves a superior trade-off between text control and personalization fidelity.

Large language models (LLMs) have shown impressive abilities in answering questions across various domains, but they often encounter hallucination issues on questions that require professional and up-to-date knowledge. To address this limitation, retrieval-augmented generation (RAG) techniques have been proposed, which retrieve relevant information from external sources to inform their responses. However, existing RAG methods typically focus on a single type of external data, such as vectorized text database or knowledge graphs, and cannot well handle real-world questions on semi-structured data containing both text and relational information. To bridge this gap, we introduce PASemiQA, a novel approach that jointly leverages text and relational information in semi-structured data to answer questions. PASemiQA first generates a plan to identify relevant text and relational information to answer the question in semi-structured data, and then uses an LLM agent to traverse the semi-structured data and extract necessary information. Our empirical results demonstrate the effectiveness of PASemiQA across different semi-structured datasets from various domains, showcasing its potential to improve the accuracy and reliability of question answering systems on semi-structured data.

Commit messages concisely describe code changes in natural language and are important for software maintenance. Several approaches have been proposed to automatically generate commit messages, but they still suffer from critical limitations, such as time-consuming training and poor generalization ability. To tackle these limitations, we propose to borrow the weapon of large language models (LLMs) and in-context learning (ICL). Our intuition is based on the fact that the training corpora of LLMs contain extensive code changes and their pairwise commit messages, which makes LLMs capture the knowledge about commits, while ICL can exploit the knowledge hidden in the LLMs and enable them to perform downstream tasks without model tuning. However, it remains unclear how well LLMs perform on commit message generation via ICL. In this paper, we conduct an empirical study to investigate the capability of LLMs to generate commit messages via ICL. Specifically, we first explore the impact of different settings on the performance of ICL-based commit message generation. We then compare ICL-based commit message generation with state-of-the-art approaches on a popular multilingual dataset and a new dataset we created to mitigate potential data leakage. The results show that ICL-based commit message generation significantly outperforms state-of-the-art approaches on subjective evaluation and achieves better generalization ability. We further analyze the root causes for LLM's underperformance and propose several implications, which shed light on future research directions for using LLMs to generate commit messages., Comment: Accepted by the 47th IEEE/ACM International Conference on Software Engineering (ICSE'25)

Multi-task learning (MTL) benefits the fine-tuning of large language models (LLMs) by providing a single model with improved performance and generalization ability across tasks, presenting a resource-efficient alternative to developing separate models for each task. Yet, existing MTL strategies for LLMs often fall short by either being computationally intensive or failing to ensure simultaneous task convergence. This paper presents CoBa, a new MTL approach designed to effectively manage task convergence balance with minimal computational overhead. Utilizing Relative Convergence Scores (RCS), Absolute Convergence Scores (ACS), and a Divergence Factor (DF), CoBa dynamically adjusts task weights during the training process, ensuring that the validation loss of all tasks progress towards convergence at an even pace while mitigating the issue of individual task divergence. The results of our experiments involving three disparate datasets underscore that this approach not only fosters equilibrium in task convergence but enhances the LLMs' performance by up to 13% relative to the second-best baselines. Code is open-sourced at https://github.com/codefuse-ai/MFTCoder., Comment: 15 pages, main conference of EMNLP 2024

Recent advancements in Transformer-based large language models (LLMs) have set new standards in natural language processing. However, the classical softmax attention incurs significant computational costs, leading to a $O(T)$ complexity for per-token generation, where $T$ represents the context length. This work explores reducing LLMs' complexity while maintaining performance by introducing Rodimus and its enhanced version, Rodimus$+$. Rodimus employs an innovative data-dependent tempered selection (DDTS) mechanism within a linear attention-based, purely recurrent framework, achieving significant accuracy while drastically reducing the memory usage typically associated with recurrent models. This method exemplifies semantic compression by maintaining essential input information with fixed-size hidden states. Building on this, Rodimus$+$ combines Rodimus with the innovative Sliding Window Shared-Key Attention (SW-SKA) in a hybrid approach, effectively leveraging the complementary semantic, token, and head compression techniques. Our experiments demonstrate that Rodimus$+$-1.6B, trained on 1 trillion tokens, achieves superior downstream performance against models trained on more tokens, including Qwen2-1.5B and RWKV6-1.6B, underscoring its potential to redefine the accuracy-efficiency balance in LLMs. Model code and pre-trained checkpoints will be available soon.

In the realm of Large Language Models (LLMs), the ability to process long contexts is increasingly crucial for tasks such as multi-round dialogues, code generation, and document summarization. This paper addresses the challenges of enhancing the long-context performance, reducing computational complexity, and leveraging pretrained models collectively termed the "impossible triangle." We introduce E2LLM (Encoder Elongated Large Language Models), a novel approach that effectively navigates this paradox. The method involves splitting long contexts into chunks, compressing each into embedding vectors via a pretrained text encoder, and utilizing an adapter to align these representations with a decoder-only LLM. Two training objectives, focusing on reconstruction of the encoder output and long-context instruction fine-tuning, are employed to facilitate the understanding of soft prompts by the LLM. Experimental results demonstrate that E2LLM achieves superior performance in long-context scenarios while balancing efficiency, performance, and compatibility with pretrained models. Our framework thus represents a significant advancement in the field, contributing to effective long-text modeling., Comment: 12 pages, 4 figures

Programming languages possess rich semantic information such as data flow that is represented by graphs and not available from the surface form of source code. Recent code language models have scaled to billions of parameters, but model source code solely as text tokens while ignoring any other structural information. Conversely, models that do encode structural information of code make modifications to the Transformer architecture, limiting their scale and compatibility with pretrained LLMs. In this work, we take the best of both worlds with GALLa - Graph Aligned Large Language Model. GALLa utilizes graph neural networks and cross-modal alignment technologies to inject the structural information of code into LLMs as an auxiliary task during finetuning. This framework is both model-agnostic and task-agnostic, as it can be applied to any code LLM for any code downstream task, and requires the structural graph data only at training time from a corpus unrelated to the finetuning data, while incurring no cost at inference time over the baseline LLM. Experiments on five code tasks with four different baseline LLMs ranging in size from 350M to 8B validate the effectiveness of GALLa, demonstrating consistent improvement over the baseline, even for powerful models such as LLaMA3.

As large language models (LLMs) are gaining increasing popularity across a wide range of web applications, it is of great importance to optimize service-level objectives (SLOs) for LLM inference services to enhance user satisfaction and improve the competitiveness of cloud vendors. In this paper, we observe that adjusting the parameters of LLM inference engines can improve service performance, and the optimal parameter configurations of different services are different. Therefore, we propose SCOOT, an automatic performance tuning system to optimize SLOs for each LLM inference service by tuning the parameters of the inference engine. SCOOT jointly exploits single-objective and multiple-objective Bayesian optimization (BO) techniques to handle various optimization objectives via exploration and exploitation. Moreover, SCOOT prunes the search space with known constraints and adopts a random forest to learn hidden constraints during the tuning process to mitigate invalid exploration. To improve the tuning efficiency, SCOOT utilizes the parallel suggestion to accelerate the tuning process. Extensive experiments demonstrate that SCOOT considerably outperforms existing tuning techniques in SLO optimization while greatly improving the tuning efficiency. Moreover, SCOOT is universally applicable to various LLM inference engines including vLLM and TensorRT-LLM. Currently, SCOOT has already been implemented in the production environment at Ant Group., Comment: Accepted by The ACM Web Conference 2025 (WWW'25)

Text-to-SQL conversion is a critical innovation, simplifying the transition from complex SQL to intuitive natural language queries, especially significant given SQL's prevalence in the job market across various roles. The rise of Large Language Models (LLMs) like GPT-3.5 and GPT-4 has greatly advanced this field, offering improved natural language understanding and the ability to generate nuanced SQL statements. However, the potential of open-source LLMs in Text-to-SQL applications remains underexplored, with many frameworks failing to leverage their full capabilities, particularly in handling complex database queries and incorporating feedback for iterative refinement. Addressing these limitations, this paper introduces SQLfuse, a robust system integrating open-source LLMs with a suite of tools to enhance Text-to-SQL translation's accuracy and usability. SQLfuse features four modules: schema mining, schema linking, SQL generation, and a SQL critic module, to not only generate but also continuously enhance SQL query quality. Demonstrated by its leading performance on the Spider Leaderboard and deployment by Ant Group, SQLfuse showcases the practical merits of open-source LLMs in diverse business contexts.

The key challenge in semantic search is to create models that are both accurate and efficient in pinpointing relevant sentences for queries. While BERT-style bi-encoders excel in efficiency with pre-computed embeddings, they often miss subtle nuances in search tasks. Conversely, GPT-style LLMs with cross-encoder designs capture these nuances but are computationally intensive, hindering real-time applications. In this paper, we present D2LLMs-Decomposed and Distilled LLMs for semantic search-that combines the best of both worlds. We decompose a cross-encoder into an efficient bi-encoder integrated with Pooling by Multihead Attention and an Interaction Emulation Module, achieving nuanced understanding and pre-computability. Knowledge from the LLM is distilled into this model using contrastive, rank, and feature imitation techniques. Our experiments show that D2LLM surpasses five leading baselines in terms of all metrics across three tasks, particularly improving NLI task performance by at least 6.45%. The source code is available at https://github.com/codefuse-ai/D2LLM.

Large language models (LLMs) iteratively generate text token by token, with memory usage increasing with the length of generated token sequences. The unpredictability of generation lengths makes it difficult to estimate the time and memory needed to process requests, posing a challenge for effective request scheduling. Conventional sequence-level scheduling (SLS) serves requests in a first-come first-served (FCFS) manner with static batching where requests with short generation lengths are delayed until those with long ones have finished generation, which hurts computational efficiency. Besides, to avoid out-of-memory (OOM) errors, SLS batches requests with a small batch size, which limits throughput. Recently proposed iteration-level scheduling (ILS) enhances computational efficiency with continuous batching to return completed requests timely and dynamically add new requests for processing. However, many ILS schedulers limit the number of parallel-processing requests to avoid OOM errors while achieving a fast inference speed, which compromises throughput. Moreover, existing SLS and ILS schedulers fail to balance the workload across multiple deployed LLM instances. To tackle these challenges, we propose slice-level scheduling (SCLS). By splitting the predefined maximal generation length limit into slices and serving batches slice by slice, it provides a precise range of serving time and memory usage for batched requests, laying the foundation for effective scheduling. Experiments confirm that compared with SLS and ILS schedulers, SCLS can improve throughput by up to 315.8% and greatly mitigate load imbalance with proposed batching and offloading algorithms., Comment: 13 pages, 22 figures

Current directed graph embedding methods build upon undirected techniques but often inadequately capture directed edge information, leading to challenges such as: (1) Suboptimal representations for nodes with low in/out-degrees, due to the insufficient neighbor interactions; (2) Limited inductive ability for representing new nodes post-training; (3) Narrow generalizability, as training is overly coupled with specific tasks. In response, we propose DUPLEX, an inductive framework for complex embeddings of directed graphs. It (1) leverages Hermitian adjacency matrix decomposition for comprehensive neighbor integration, (2) employs a dual GAT encoder for directional neighbor modeling, and (3) features two parameter-free decoders to decouple training from particular tasks. DUPLEX outperforms state-of-the-art models, especially for nodes with sparse connectivity, and demonstrates robust inductive capability and adaptability across various tasks. The code is available at https://github.com/alipay/DUPLEX.

Nowadays, large language models (LLMs) are published as a service and can be accessed by various applications via APIs, also known as language-model-as-a-service (LMaaS). Without knowing the generation length of requests, existing serving systems serve requests in a first-come, first-served (FCFS) manner with a fixed batch size, which leads to two problems that affect batch serving efficiency. First, the generation lengths of requests in a batch vary, and requests with short generation lengths must wait for requests with long generation lengths to finish during the batch serving procedure. Second, requests with longer generation lengths consume more memory during serving. Without knowing the generation lengths of batched requests, the batch size is always set small to avoid the out-of-memory (OOM) error, thus preventing the GPU from being fully utilized. In this paper, we find that a significant number of popular applications in the LMaaS scenario have a positive correlation between the generation length and the length of raw user input. Based on this observation, we propose Magnus, which can accurately predict the request generation length with the user input length, application-level, and user-level semantic features. Accordingly, Magnus can achieve high request throughput by batching requests of similar generation lengths together with adaptive batch sizes. Besides, Magnus can also schedule batches with the highest response ratio next (HRRN) policy to reduce request response time. Experiments conducted on our testbed show that Magnus improves request throughput by up to 234\% and reduces response time by up to 89.7\% compared to baselines., Comment: 12 pages, 14 figures

Vector data management systems (VDMSs) have become an indispensable cornerstone in large-scale information retrieval and machine learning systems like large language models. To enhance the efficiency and flexibility of similarity search, VDMS exposes many tunable index parameters and system parameters for users to specify. However, due to the inherent characteristics of VDMS, automatic performance tuning for VDMS faces several critical challenges, which cannot be well addressed by the existing auto-tuning methods. In this paper, we introduce VDTuner, a learning-based automatic performance tuning framework for VDMS, leveraging multi-objective Bayesian optimization. VDTuner overcomes the challenges associated with VDMS by efficiently exploring a complex multi-dimensional parameter space without requiring any prior knowledge. Moreover, it is able to achieve a good balance between search speed and recall rate, delivering an optimal configuration. Extensive evaluations demonstrate that VDTuner can markedly improve VDMS performance (14.12% in search speed and 186.38% in recall rate) compared with default setting, and is more efficient compared with state-of-the-art baselines (up to 3.57 times faster in terms of tuning time). In addition, VDTuner is scalable to specific user preference and cost-aware optimization objective. VDTuner is available online at https://github.com/tiannuo-yang/VDTuner., Comment: Accepted by ICDE 2024

Action Quality Assessment (AQA) evaluates diverse skills but models struggle with non-stationary data. We propose Continual AQA (CAQA) to refine models using sparse new data. Feature replay preserves memory without storing raw inputs. However, the misalignment between static old features and the dynamically changing feature manifold causes severe catastrophic forgetting. To address this novel problem, we propose Manifold-Aligned Graph Regularization (MAGR), which first aligns deviated old features to the current feature manifold, ensuring representation consistency. It then constructs a graph jointly arranging old and new features aligned with quality scores. Experiments show MAGR outperforms recent strong baselines with up to 6.56%, 5.66%, 15.64%, and 9.05% correlation gains on the MTL-AQA, FineDiving, UNLV-Dive, and JDM-MSA split datasets, respectively. This validates MAGR for continual assessment challenges arising from non-stationary skill variations. Code is available at https://github.com/ZhouKanglei/MAGR_CAQA}{https://github.com/ZhouKanglei/MAGR_CAQA., Comment: Accepted by ECCV 2024 as an oral paper

Due to the sheer size of software logs, developers rely on automated log analysis. Log parsing, which parses semi-structured logs into a structured format, is a prerequisite of automated log analysis. However, existing log parsers are unsatisfactory when applied in practice because: 1) they ignore categories of variables, and 2) have poor generalization ability. To address the limitations of existing approaches, we propose LogPTR, the first end-to-end variable-aware log parser that can extract the static and dynamic parts in logs, and further identify the categories of variables. The key of LogPTR is using pointer network to copy words from the log message. We have performed extensive experiments on 16 public log datasets and the results show that LogPTR outperforms state-of-the-art log parsers both on general log parsing that extracts the log template and variable-aware log parsing that further identifies the category of variables.

In this work we systematically review the recent advancements in software engineering with language models, covering 70+ models, 40+ evaluation tasks, 180+ datasets, and 900 related works. Unlike previous works, we integrate software engineering (SE) with natural language processing (NLP) by discussing the perspectives of both sides: SE applies language models for development automation, while NLP adopts SE tasks for language model evaluation. We break down code processing models into general language models represented by the GPT family and specialized models that are specifically pretrained on code, often with tailored objectives. We discuss the relations and differences between these models, and highlight the historical transition of code modeling from statistical models and RNNs to pretrained Transformers and LLMs, which is exactly the same course that had been taken by NLP. We also go beyond programming and review LLMs' application in other software engineering activities including requirement engineering, testing, deployment, and operations in an endeavor to provide a global view of NLP in SE, and identify key challenges and potential future directions in this domain. We keep the survey open and updated on GitHub at https://github.com/codefuse-ai/Awesome-Code-LLM., Comment: Repo: https://github.com/codefuse-ai/Awesome-Code-LLM. 9 figures, 18 tables, and 902 references. Under review

Code LLMs have emerged as a specialized research field, with remarkable studies dedicated to enhancing model's coding capabilities through fine-tuning on pre-trained models. Previous fine-tuning approaches were typically tailored to specific downstream tasks or scenarios, which meant separate fine-tuning for each task, requiring extensive training resources and posing challenges in terms of deployment and maintenance. Furthermore, these approaches failed to leverage the inherent interconnectedness among different code-related tasks. To overcome these limitations, we present a multi-task fine-tuning framework, MFTcoder, that enables simultaneous and parallel fine-tuning on multiple tasks. By incorporating various loss functions, we effectively address common challenges in multi-task learning, such as data imbalance, varying difficulty levels, and inconsistent convergence speeds. Extensive experiments have conclusively demonstrated that our multi-task fine-tuning approach outperforms both individual fine-tuning on single tasks and fine-tuning on a mixed ensemble of tasks. Moreover, MFTcoder offers efficient training capabilities, including efficient data tokenization modes and PEFT fine-tuning, resulting in significantly improved speed compared to traditional fine-tuning methods. MFTcoder seamlessly integrates with several mainstream open-source LLMs, such as CodeLLama and Qwen. Leveraging the CodeLLama foundation, our MFTcoder fine-tuned model, \textsc{CodeFuse-CodeLLama-34B}, achieves an impressive pass@1 score of 74.4\% on the HumaneEval benchmark, surpassing GPT-4 performance (67\%, zero-shot). MFTCoder is open-sourced at \url{https://github.com/codefuse-ai/MFTCOder}

Bases have become an integral part of modern deep learning-based models for time series forecasting due to their ability to act as feature extractors or future references. To be effective, a basis must be tailored to the specific set of time series data and exhibit distinct correlation with each time series within the set. However, current state-of-the-art methods are limited in their ability to satisfy both of these requirements simultaneously. To address this challenge, we propose BasisFormer, an end-to-end time series forecasting architecture that leverages learnable and interpretable bases. This architecture comprises three components: First, we acquire bases through adaptive self-supervised learning, which treats the historical and future sections of the time series as two distinct views and employs contrastive learning. Next, we design a Coef module that calculates the similarity coefficients between the time series and bases in the historical view via bidirectional cross-attention. Finally, we present a Forecast module that selects and consolidates the bases in the future view based on the similarity coefficients, resulting in accurate future predictions. Through extensive experiments on six datasets, we demonstrate that BasisFormer outperforms previous state-of-the-art methods by 11.04\% and 15.78\% respectively for univariate and multivariate forecasting tasks. Code is available at: \url{https://github.com/nzl5116190/Basisformer}, Comment: NeurIPS 2023(poster)

Code Large Language Models (Code LLMs) have gained significant attention in the industry due to their wide applications in the full lifecycle of software engineering. However, the effectiveness of existing models in understanding non-English inputs for multi-lingual code-related tasks is still far from well studied. This paper introduces CodeFuse-13B, an open-sourced pre-trained code LLM. It is specifically designed for code-related tasks with both English and Chinese prompts and supports over 40 programming languages. CodeFuse achieves its effectiveness by utilizing a high quality pre-training dataset that is carefully filtered by program analyzers and optimized during the training process. Extensive experiments are conducted using real-world usage scenarios, the industry-standard benchmark HumanEval-x, and the specially designed CodeFuseEval for Chinese prompts. To assess the effectiveness of CodeFuse, we actively collected valuable human feedback from the AntGroup's software development process where CodeFuse has been successfully deployed. The results demonstrate that CodeFuse-13B achieves a HumanEval pass@1 score of 37.10%, positioning it as one of the top multi-lingual code LLMs with similar parameter sizes. In practical scenarios, such as code generation, code translation, code comments, and testcase generation, CodeFuse performs better than other models when confronted with Chinese prompts., Comment: Accepted by ICSE-SEIP 2024

Microservice architecture has sprung up over recent years for managing enterprise applications, due to its ability to independently deploy and scale services. Despite its benefits, ensuring the reliability and safety of a microservice system remains highly challenging. Existing anomaly detection algorithms based on a single data modality (i.e., metrics, logs, or traces) fail to fully account for the complex correlations and interactions between different modalities, leading to false negatives and false alarms, whereas incorporating more data modalities can offer opportunities for further performance gain. As a fresh attempt, we propose in this paper a semi-supervised graph-based anomaly detection method, MSTGAD, which seamlessly integrates all available data modalities via attentive multi-modal learning. First, we extract and normalize features from the three modalities, and further integrate them using a graph, namely MST (microservice system twin) graph, where each node represents a service instance and the edge indicates the scheduling relationship between different service instances. The MST graph provides a virtual representation of the status and scheduling relationships among service instances of a real-world microservice system. Second, we construct a transformer-based neural network with both spatial and temporal attention mechanisms to model the inter-correlations between different modalities and temporal dependencies between the data points. This enables us to detect anomalies automatically and accurately in real-time. The source code of MSTGAD is publicly available at https://github.com/alipay/microservice_system_twin_graph_based_anomaly_detection., Comment: This paper have been accepted by ASE 2023

Self-attention and position embedding are two key modules in transformer-based Large Language Models (LLMs). However, the potential relationship between them is far from well studied, especially for long context window extending. In fact, anomalous behaviors harming long context extrapolation exist between Rotary Position Embedding (RoPE) and vanilla self-attention unveiled by our work. To address this issue, we propose a novel attention mechanism, CoCA (Collinear Constrained Attention). Specifically, we enforce a collinear constraint between $Q$ and $K$ to seamlessly integrate RoPE and self-attention. While only adding minimal computational and spatial complexity, this integration significantly enhances long context window extrapolation ability. We provide an optimized implementation, making it a drop-in replacement for any existing transformer-based models. Extensive experiments show that CoCA performs extraordinarily well in extending context windows. A CoCA-based GPT model, trained with a context length of 512, can seamlessly extend the context window up to 32K (60$\times$), without any fine-tuning. Additionally, by dropping CoCA in LLaMA-7B, we achieve extrapolation up to 32K within only 2K training length. Our code is publicly available at: https://github.com/codefuse-ai/Collinear-Constrained-Attention, Comment: 16 pages, 7 figures