Recent advancements in large language models (LLMs) and AI systems have led to a paradigm shift in the design and optimization of complex AI workflows. By integrating multiple components, compound AI systems have become increasingly adept at performing sophisticated tasks. However, as these systems grow in complexity, new challenges arise in optimizing not only individual components but also their interactions. While traditional optimization methods such as supervised fine-tuning (SFT) and reinforcement learning (RL) remain foundational, the rise of natural language feedback introduces promising new approaches, especially for optimizing non-differentiable systems. This paper provides a systematic review of recent progress in optimizing compound AI systems, encompassing both numerical and language-based techniques. We formalize the notion of compound AI system optimization, classify existing methods along several key dimensions, and highlight open research challenges and future directions in this rapidly evolving field. A list of surveyed papers is publicly available at https://github.com/MiuLab/AISysOpt-Survey.

With the rapid development of artificial intelligence technology, its application in the optimization of complex computer systems is becoming more and more extensive. Edge computing is an efficient distributed computing architecture, and the health status of its nodes directly affects the performance and reliability of the entire system. In view of the lack of accuracy and interpretability of traditional methods in node health status judgment, this paper proposes a health status judgment method based on XGBoost and combines the SHAP method to analyze the interpretability of the model. Through experiments, it is verified that XGBoost has superior performance in processing complex features and nonlinear data of edge computing nodes, especially in capturing the impact of key features (such as response time and power consumption) on node status. SHAP value analysis further reveals the global and local importance of features, so that the model not only has high precision discrimination ability but also can provide intuitive explanations, providing data support for system optimization. Research shows that the combination of AI technology and computer system optimization can not only realize the intelligent monitoring of the health status of edge computing nodes but also provide a scientific basis for dynamic optimization scheduling, resource management and anomaly detection. In the future, with the in-depth development of AI technology, model dynamics, cross-node collaborative optimization and multimodal data fusion will become the focus of research, providing important support for the intelligent evolution of edge computing systems.

Text-based AI system optimization typically involves a feedback loop scheme where a single LLM generates an evaluation in natural language of the current output to improve the next iteration's output. However, in this work, we empirically demonstrate that for a practical and complex task (code generation) with multiple criteria to evaluate, utilizing only one LLM evaluator tends to let errors in generated code go undetected, thus leading to incorrect evaluations and ultimately suboptimal test case performance. Motivated by this failure case, we assume there exists an optimal evaluation policy that samples an evaluation between response and ground truth. We then theoretically prove that a linear combination of multiple evaluators can approximate this optimal policy. From this insight, we propose AI system optimization via Multiple LLM Evaluators (AIME). AIME is an evaluation protocol that utilizes multiple LLMs that each independently generate an evaluation on separate criteria and then combine them via concatenation. We provide an extensive empirical study showing AIME outperforming baseline methods in code generation tasks, with up to 62% higher error detection rate and up to 16% higher success rate than a single LLM evaluation protocol on LeetCodeHard and HumanEval datasets. We also show that the selection of the number of evaluators and which criteria to utilize is non-trivial as it can impact pact success rate by up to 12%. Pre-trained foundation models, such as Large Language Models (LLMs), have developed rapidly over the recent years (Achiam et al., 2023; Touvron et al., 2023). As the application complexity increases, the shift to AI systems containing multiple components such as LLM-based agents and web search (Xiong et al., 2024), will continue (Zaharia et al., 2024; Yuksekgonul et al., 2024).

Computation in a typical Transformer-based large language model (LLM) can be characterized by batch size, hidden dimension, number of layers, and sequence length. Until now, system works for accelerating LLM training have focused on the first three dimensions: data parallelism for batch size, tensor parallelism for hidden size and pipeline parallelism for model depth or layers. These widely studied forms of parallelism are not targeted or optimized for long sequence Transformer models. Given practical application needs for long sequence LLM, renewed attentions are being drawn to sequence parallelism. However, existing works in sequence parallelism are constrained by memory-communication inefficiency, limiting their scalability to long sequence large models. In this work, we introduce DeepSpeed-Ulysses, a novel, portable and effective methodology for enabling highly efficient and scalable LLM training with extremely long sequence length. DeepSpeed-Ulysses at its core partitions input data along the sequence dimension and employs an efficient all-to-all collective communication for attention computation. Theoretical communication analysis shows that whereas other methods incur communication overhead as sequence length increases, DeepSpeed-Ulysses maintains constant communication volume when sequence length and compute devices are increased proportionally. Furthermore, experimental evaluations show that DeepSpeed-Ulysses trains 2.5x faster with 4x longer sequence length than the existing method SOTA baseline.

To accelerate distributed training, many gradient compression methods have been proposed to alleviate the communication bottleneck in synchronous stochastic gradient descent (S-SGD), but their efficacy in real-world applications still remains unclear. In this work, we first evaluate the efficiency of three representative compression methods (quantization with Sign-SGD, sparsification with Top-k SGD, and low-rank with Power-SGD) on a 32-GPU cluster. The results show that they cannot always outperform well-optimized S-SGD or even worse due to their incompatibility with three key system optimization techniques (all-reduce, pipelining, and tensor fusion) in S-SGD. To this end, we propose a novel gradient compression method, called alternate compressed Power-SGD (ACP-SGD), which alternately compresses and communicates low-rank matrices. ACP-SGD not only significantly reduces the communication volume, but also enjoys the three system optimizations like S-SGD. Compared with Power-SGD, the optimized ACP-SGD can largely reduce the compression and communication overheads, while achieving similar model accuracy. In our experiments, ACP-SGD achieves an average of 4.06x and 1.43x speedups over S-SGD and Power-SGD, respectively, and it consistently outperforms other baselines across different setups (from 8 GPUs to 64 GPUs and from 1Gb/s Ethernet to 100Gb/s InfiniBand).

The Deep Learning (DL) open-source community has seen tremendous growth in the last few months. Incredibly powerful text generation models such as the Bloom 176B, or image generation model such as Stable Diffusion are now available to anyone with access to a handful or even a single GPU through platforms such as Hugging Face. While open sourcing has democratized access to AI capabilities, their application is still restricted by two critical factors: inference latency and cost. There has been significant progress in system optimizations for DL model inference that can drastically reduce both latency and cost, but those are not easily accessible. A main reason for this limited accessibility is that the DL model inference landscape is diverse with models varying in size, architecture, system performance characteristics, hardware requirements, etc. Identifying the appropriate set of system optimizations applicable to a given model and applying them correctly is often beyond the scope of most data scientists, making low latency and low-cost inference mostly inaccessible.

There is a growing need to deploy machine learning for different tasks on a wide array of new hardware platforms. Such deployment scenarios require tackling multiple challenges, including identifying a model architecture that can achieve a suitable predictive accuracy (architecture search), and finding an efficient implementation of the model to satisfy underlying hardware-specific systems constraints such as latency (system optimization search). Existing works treat architecture search and system optimization search as separate problems and solve them sequentially. In this paper, we instead propose to solve these problems jointly, and introduce a simple but effective baseline method called SONAR that interleaves these two search problems. SONAR aims to efficiently optimize for predictive accuracy and inference latency by applying early stopping to both search processes. Our experiments on multiple different hardware back-ends show that SONAR identifies nearly optimal architectures 30 times faster than a brute force approach.

Research in deep learning and AI is being revolutionized by large-scale models, which has resulted in significant advancements in numerous areas, including multilingual translation, creative text generation, and language interpretation. Nevertheless, the models' vast size results in latency and cost limits that make installing applications on top of them difficult, despite their impressive capabilities. The DeepSpeed team at Microsoft AI has been investigating system optimization and model compression advancements to meet these deployment problems. The DeepSpeed inference system was previously made available by the researchers as part of the Scale initiative. This system uses a variety of optimizations to increase the speed of model inference, such as highly optimized CUDA kernels and inference-adapted parallelism.

Over the past decade, machine learning (ML) has become a critical component of countless applications and services in a variety of domains. Fields ranging from healthcare to autonomous vehicles have been transformed by the use of ML techniques. Machine learning's increasing importance to real-world applications brought awareness of a new field focused on ML in practice - machine learning systems (or, as some call it, MLOps). This field acts as a bridging point between the domains of computer systems and machine learning, considering the new challenges of machine learning with a lens shaped by traditional systems research. So what are these "ML challenges"?

The introduction of unexpected system disturbances and new system dynamics does not allow initially selected static system and controller parameters to guarantee continued system stability and performance. In this research we present a novel approach for detecting early failure indicators of non-linear highly chaotic system and accordingly predict the best parameter calibrations to offset such instability using deep machine learning regression model. The approach proposed continuously monitors the system and controller signals. The Re-calibration of the system and controller parameters is triggered according to a set of conditions designed to maintain system stability without compromise to the system speed, intended outcome or required processing power. The deep neural model predicts the parameter values that would best counteract the expected system in-stability. To demonstrate the effectiveness of the proposed approach, it is applied to the non-linear complex combination of Duffing Van der pol oscillators. The approach is also tested under different scenarios the system and controller parameters are initially chosen incorrectly or the system parameters are changed while running or new system dynamics are introduced while running to measure effectiveness and reaction time.