The reading list for making LLM inference cheaper and faster

What it does Awesome-LLM-Inference is a curated reading list that categorizes recent research papers on LLM and VLM inference optimization, pairing each paper with its open-source implementation. The repository covers everything from FlashAttention variants and PagedAttention to quantization schemes like WINT8/4 and AWQ, plus parallelism strategies and KV-cache compression. It also bundles a 500-page PDF primer aimed at beginners who want to understand the fundamentals without reading every original paper.

The interesting bit Instead of dumping links, the list slices the field into practitioner-sized topics—like “Disaggregating Prefill and Decoding” or “IO/FLOPs-Aware Sparse Attention”—so you can navigate directly to the technique you need. It also tracks the full DeepSeek inference stack (FlashMLA, DualPipe, DeepEP) alongside broader community work, making it a decent barometer for what is currently considered state-of-the-art in serving.

Key highlights

• Covers 20+ sub-topics including quantization, continuous batching, MoE inference, long-context attention, and non-transformer architectures.
• Each entry links directly to both the paper PDF and the corresponding code repository.
• Maintains a dedicated section for DeepSeek/MLA-related work and other trending topics like Star-Attention and MiniMax-01.
• Provides a 500-page “Awesome LLM Inference for Beginners” PDF covering foundational techniques from FlashAttention to SmoothQuant.
• Includes a small script to batch-download all referenced PDFs.

Caveats

• This is a bibliography, not a framework: there is no original inference code or benchmarks here, just links and categorization.
• Several high-profile entries—notably some DeepSeek papers and system overviews—are marked with a ⚠️ indicating missing or not-yet-available code.
• The PDF download helper was generated by an AI assistant (Doubao), so its reliability is unclear from the README alone.

Verdict Worth bookmarking if you are building or tuning an inference stack and need a quick, code-linked survey of the literature. Skip it if you are looking for a drop-in replacement for vLLM or TensorRT-LLM.