Axolotl
Apache 2.0 LLM fine-tuning framework for local or cloud GPUs, supporting NVIDIA, AMD, LoRA, QLoRA and multimodal training.
Axolotl is an open-source LLM fine-tuning framework for developers, researchers, and teams training models on their own data. It runs on local hardware or cloud infrastructure you control, including Docker and Kubernetes environments. The framework uses Apache 2.0, which permits commercial use.
You can fully fine-tune models or use LoRA and QLoRA to reduce training memory needs. Supported families include GPT-OSS, Llama, Mistral, Mixtral, Qwen, and Gemma, alongside models available through Hugging Face Transformers. Training also covers vision-language and audio models such as Qwen2-VL, Pixtral, LLaVA, and Voxtral, with support for image, video, and audio data.
Beyond fine-tuning, Axolotl supports preference training with DPO, reinforcement learning with GRPO, and reward modelling. A shared configuration covers dataset preparation, training, evaluation, quantization, and inference. For larger workloads, it supports multiple GPUs and multiple machines through FSDP, DeepSpeed, Torchrun, and Ray. Attention optimizations and sequence packing help make better use of training hardware.
Axolotl requires an NVIDIA or AMD GPU. It accepts local datasets as well as data from Hugging Face and cloud storage, so training data doesn't have to go to an external AI service. Telemetry is enabled by default and can be disabled; it collects basic system information, model types, and error rates, excluding personal data and file paths.
