liguodongiot/llm-action

Covers distributed training parallelism (data, tensor, pipeline), parameter-efficient fine-tuning (LoRA, QLoRA, P-Tuning v2), model compression (quantization, pruning, distillation), and inference optimization across frameworks like vLLM and LightLLM. Includes hands-on implementations spanning 6B to 65B models with tutorials on RLHF alignment, prompt engineering, and multi-modal applications, plus benchmarking tools for evaluation and performance analysis.

23,504 stars. Actively maintained with 4 commits in the last 30 days.