llamafactory

LlamaFactory is an open-source framework for fine-tuning large language models (LLMs) and multimodal models. The README highlights support for many model families, including LLaMA/LLaVA, Mistral, Mixtral, Qwen3/Qwen2.5-VL, DeepSeek, Gemma, GLM, Phi, GPT-OSS, InternVL, and others. It integrates a wide range of training methods: continued pre-training, supervised fine-tuning, reward modeling, and preference or reinforcement learning approaches such as PPO, DPO, KTO, ORPO, and SimPO. It also supports full-parameter, freeze, LoRA, and quantized QLoRA tuning, including 2/3/4/5/6/8-bit quantization via AQLM, AWQ, GPTQ, LLM.int8, HQQ, and EETQ. The project includes advanced optimizers and performance tricks: GaLore, BAdam, APOLLO, Adam-mini, Muon, OFT, DoRA, LongLoRA, PiSSA, FlashAttention-2, Unsloth, Liger Kernel, RoPE scaling, NEFTune, and rsLoRA. Recent changelog entries also mention a Megatron-core training backend, SGLang inference support, and fine-tuning support for GPT-OSS, Qwen2.5 Omni, GLM-4.1V-Thinking, and Intern-S1-mini models. LlamaFactory provides zero-code usage through a command-line interface and the LLaMA Board Web UI built with Gradio, plus a Colab notebook and cloud deployment templates for platforms such as PAI-DSW and Hugging Face Spaces. Fine-tuned checkpoints can be exported and served through an OpenAI-compatible API, a Gradio interface, or a CLI, with vLLM or SGLang available as faster inference backends. Experiment tracking is supported with TensorBoard, WandB, MLflow, SwanLab, and LlamaBoard. Models and datasets can be downloaded from Hugging Face, ModelScope, and Modelers Hub. The README also cites use by Amazon, NVIDIA, and Aliyun, and links to documentation, Docker images, sample notebooks, and the project's official blog.