This repository is a curated list of academic papers, articles, tutorials, and projects focused on Large Language Model (LLM) systems. It serves researchers and engineers working on optimizing LLM training, serving, efficiency, and related system-level challenges, providing a comprehensive overview of the state-of-the-art in this rapidly evolving field.

How It Works

The list is organized by key LLM system categories, including training, post-training (RLHF), fault tolerance, serving, compound AI systems, edge deployment, efficiency optimization, fine-tuning, multi-modal systems, LLM for systems applications, benchmarks, and frameworks. Each entry links to relevant research, offering a structured way to navigate and understand the system design considerations for LLMs.

Quick Start & Requirements

This is a curated list, not a software package. No installation or execution is required.

Highlighted Details

• Extensive coverage of LLM serving techniques, including KV cache management, quantization, speculative decoding, and multi-tenancy.
• Detailed sections on training optimizations like parallelism (data, tensor, pipeline, sequence), activation recomputation, and network architectures.
• Includes papers on fault tolerance, straggler mitigation, and efficient checkpointing for large-scale distributed training.
• Features benchmarks, leaderboards, and key LLM system frameworks (e.g., DeepSpeed, TensorRT-LLM, vLLM).

Maintenance & Community

The repository is maintained by AmberLJC. It acts as a community resource, with contributions likely coming from the broader LLM systems research community. Links to relevant courses and other "awesome" lists are provided for further exploration.

Licensing & Compatibility

The repository itself is not software and thus not subject to software licensing. The linked papers are subject to their respective publication licenses.

Limitations & Caveats

This is a reference list and does not provide code or implementations. The rapid pace of LLM research means the list may require frequent updates to remain fully comprehensive.