The emergence of Superchips represents a significant advancement in next-generation AI hardware. These Superchips employ a tightly coupled heterogeneous architecture that integrates GPU and CPU on the same package, which offers unprecedented computational power. However, there has been scant research investigating how LLM training benefits from this new architecture. In this work, for the first time, we study LLM training solutions based on offloading for Superchips. We observe important differences between Superchips and traditional loosely-coupled GPU-CPU architecture, which necessitate revisiting prevailing assumptions about offloading. Based on that, we present SuperOffload, a Superchip-centric offloading system that simultaneously uses Hopper GPU, Grace CPU, and NVLink-C2C interconnect more efficiently. SuperOffload accomplishes this via a combination of techniques, such as adaptive weight offloading, bucketization repartitioning, Superchip-aware casting, speculative execution, and a highly optimized Adam optimizer for Grace CPUs. Our evaluation of SuperOffload on NVIDIA GH200 demonstrates up to 2.5× throughput improvement compared to state-of-the-art offloading-based systems, enabling training of up to 25B model on a single Superchip while achieving high training throughput. We also extend SuperOffload with ZeRO-style data parallelism and DeepSpeed-Ulysses sequence parallelism, enabling training of 13B model with sequence lengths up to 1 million tokens on 8 GH200 while achieving 55% MFU.

Misconfigurations are major causes of software failures. Existing practices rely on developer-written rules or test cases to validate configuration values, which are expensive. Machine learning (ML) for configuration validation is considered a promising direction, but has been facing challenges such as the need of large-scale field data and system-specific models. Recent advances in Large Language Models (LLMs) show promise in addressing some of the long-lasting limitations of ML-based configuration validation. We present the first analysis on the feasibility and effectiveness of using LLMs for configuration validation. We empirically evaluate LLMs as configuration validators by developing a generic LLM-based configuration validation framework, named Ciri. Ciri employs effective prompt engineering with few-shot learning based on both valid configuration and misconfiguration data. Ciri checks outputs from LLMs when producing results, addressing hallucination and nondeterminism of LLMs. We evaluate Ciri's validation effectiveness on eight popular LLMs using configuration data of ten widely deployed open-source systems. Our analysis (1) confirms the potential of using LLMs for configuration validation, (2) explores design space of LLM-based validators like Ciri, and (3) reveals open challenges such as ineffectiveness in detecting certain types of misconfigurations and biases towards popular configuration parameters.

Reconstructing 3D objects into editable programs is pivotal for applications like reverse engineering and shape editing. However, existing methods often rely on limited domain-specific languages (DSLs) and small-scale datasets, restricting their ability to model complex geometries and structures. To address these challenges, we introduce MeshLLM, a novel framework that reconstructs complex 3D objects from point clouds into editable Blender Python scripts. We develop a comprehensive set of expressive Blender Python APIs capable of synthesizing intricate geometries. Leveraging these APIs, we construct a large-scale paired object-code dataset, where the code for each object is decomposed into distinct semantic parts. Subsequently, we train a multimodal large language model (LLM) that translates 3D point cloud into executable Blender Python scripts. Our approach not only achieves superior performance in shape-to-code reconstruction tasks but also facilitates intuitive geometric and topological editing through convenient code modifications. Furthermore, our code-based representation enhances the reasoning capabilities of LLMs in 3D shape understanding tasks. Together, these contributions establish MeshLLM as a powerful and flexible solution for programmatic 3D shape reconstruction and understanding.

Today's software failures have two dominating root causes: code bugs and misconfigurations. To combat failure-inducing software changes, unified regression testing (URT) is needed to synergistically test the changed code and all changed production configurations for deployment reliability. However, URT could incur high cost, as it needs to run a large number of tests under multiple configurations. Regression test selection (RTS) can reduce regression testing cost. Unfortunately, no existing RTS technique reasons about code and configuration changes collectively. We introduce Unified Regression Test Selection (uRTS) to effectively reduce the cost of URT. uRTS supports project changes on 1) code only, 2) configurations only, and 3) both code and configurations. It selects regular tests and configuration tests with a unified selection algorithm. The uRTS algorithm analyzes code and configuration dependencies of each test across runs and across configurations. uRTS provides the same safety guarantee as the state-of-the-art RTS while selecting fewer tests and, more importantly, reducing the end-to-end testing time. We implemented uRTS on top of Ekstazi (a RTS tool for code changes) and Ctest (a configuration testing framework). We evaluate uRTS on hundreds of code revisions and dozens of configurations of five large projects. The results show that uRTS reduces the end-to-end testing time, on average, by 3.64X compared to executing all tests and 1.87X compared to a competitive reference solution that directly extends RTS for URT.

Reinforcement learning (RL) builds an effective agent that handles tasks in complex and uncertain environments by maximizing future reward. However, the efficiency is insufficient for practical use such as game AI and autonomous driving. An effective but selfish agent conflicts with other humans, and hence the demand of a human-like behavior arises. Imitation learning (IL) has been employed to train an agent to mimic the actions of expert behaviors provided as training data. However, IL tends to build an agent limited in performance by the expert skill, and even worse, the agent exhibits an inconsistent behavior since IL is not goal-oriented. In this paper, we propose a training scheme by mixing RL and IL for both discrete and continuous action space problems. The proposed scheme builds an agent that achieves a performance higher than an agent trained by only IL and exhibits a more human-like behavior than agents trained by RL or IL, validated by human sensitivity.

Artificial intelligence (AI) agent created with Deep Q-Networks (DQN) can defeat human agents in video games. Despite its high performance, DQN often exhibits odd behaviors, which could be immersion-breaking against the purpose of creating game AI. Moreover, DQN is capable of reacting to the game environment much faster than humans, making itself invincible (thus not fun to play with) in certain types of games. On the other hand, supervised learning framework trains an AI agent using historical play data of human agents as training data. Supervised learning agent exhibits a more human-like behavior than reinforcement learning agents because of imitating training data. However, its performance is often no better than human agents. The ultimate purpose of AI agents is to entertain human players. A good performance and a humanlike behavior are important factors of the AI agents, and both of them should be achieved simultaneously. This study proposes frameworks combining reinforcement learning and supervised learning and we call then separated network model and shared network model. We evaluated their performances by the game scores and behaviors by Turing test. The experimental results demonstrate that the proposed frameworks develop an AI agent of better performance than human agent and natural behavior than reinforcement learning agents.