LLaMA-Factory

Fine-tuning an open model usually means wiring transformers, PEFT, TRL and DeepSpeed together by hand; LlamaFactory hides that plumbing behind YAML configs, a llamafactory-cli command and a Gradio interface called LlamaBoard. It handles continued pre-training, supervised and multimodal fine-tuning, reward models, and preference methods like DPO, PPO, KTO and ORPO across families such as Llama, Qwen3-VL, Mistral, DeepSeek, Gemma, GLM and Phi, using full, freeze, LoRA or 2 to 8 bit QLoRA tuning plus extras like GaLore, Muon, DoRA, FlashAttention-2, Unsloth and Liger Kernel. Recent additions include a Megatron-core backend and KTransformers support for very large MoE models. After training, the same CLI merges and exports adapters, chats with the result, or serves an OpenAI-style API through vLLM or SGLang. You need Python 3.11 and PyTorch; the authors estimate about 16 GB of VRAM for LoRA on a 7B model and roughly 6 GB for 4-bit QLoRA, and Docker files cover CUDA, ROCm, Intel XPU and Ascend NPU. Code is Apache 2.0, while base models keep their own licenses. Ideal for repeatable fine-tunes without hand-written training loops.