This repository serves as a curated collection of papers and their associated code for AI systems, focusing on large language models (LLMs) and related technologies. It targets researchers and engineers working on optimizing LLM inference, training, evaluation, and deployment, providing a valuable resource for understanding the state-of-the-art in system design for AI.

How It Works

The repository categorizes papers and code across various domains within AI systems, including LLM serving frameworks, inference systems (both system and AI sides), evaluation platforms, RLHF, Mixture-of-Experts (MoE), LoRA, parallelism strategies, GPU cluster management, and optimization techniques. It aims to consolidate cutting-edge research, offering links to papers and code repositories for each entry.

Quick Start & Requirements

This repository is a collection of links and does not have a direct installation or execution command. Users are expected to navigate to the linked paper or code repositories for specific setup and requirements.

Highlighted Details

• Comprehensive coverage of LLM inference systems, with numerous entries from top-tier conferences (e.g., OSDI, ASPLOS, MLSys) published in 2023-2025.
• Extensive lists of papers and code for distributed training parallelism strategies, including pipeline, data, and tensor parallelism.
• Detailed sections on Mixture-of-Experts (MoE) and LoRA, covering both training and serving aspects.
• Includes research on GPU cluster management and resource scheduling for AI workloads.

Maintenance & Community

The repository is open for contributions via issues or pull requests. It lists numerous researchers from prominent universities, indicating a strong academic backing.

Licensing & Compatibility

The licensing of individual linked projects varies, as this repository itself is a collection of pointers. Users must consult the licenses of the respective code repositories for compatibility and usage restrictions.

Limitations & Caveats

This repository is a curated list and does not provide any executable code or unified framework. Users must independently locate, set up, and integrate the individual projects. The "paper and its code" claim may not always hold true for every entry, as some links might only point to papers.