Awesome-LLM-Inference
A curated list of LLM and VLM inference research papers with accompanying code, covering Flash-Attention, Paged-Attention, quantization, parallelism and related inference-optimization techniques.
Awesome-LLM-Inference organizes papers and code by topic, including DeepSeek and MLA, multi-GPU and multi-node parallelism, disaggregated prefill and decoding, KV cache scheduling, weight and activation quantization, continuous batching, sparse attention, mixture-of-experts inference, and CPU, NPU, FPGA and mobile inference. It includes a downloadable 500-page PDF summarizing many of the underlying techniques and a script to fetch the cited PDFs.
It links to sibling lists for diffusion-model inference papers and CUDA learning notes, and each entry is dated and linked to its paper and, where available, its code repository.
It is written in Python and licensed under GPL-3.0.