axolotl-ai-cloud/axolotl

What it solves

Axolotl simplifies the complex process of post-training and fine-tuning large language models (LLMs). It provides a unified framework that removes the need to write custom training scripts for every new model or technique, allowing users to configure their entire pipeline—from dataset preprocessing to inference—using a single YAML file.

How it works

Axolotl acts as a wrapper and orchestrator for various training libraries and optimizations. It supports a wide array of model architectures (including text, vision-language, and audio models) and integrates advanced training methods like LoRA, QLoRA, and RL (GRPO, GDPO). To maximize hardware efficiency, it incorporates performance kernels like Flash Attention and supports distributed training across multiple GPUs or nodes using FSDP2 and DeepSpeed.

Who it’s for

It is designed for developers and AI researchers who want to fine-tune state-of-the-art open-source models on custom datasets without managing the underlying boilerplate code for distributed training and memory optimization.

Highlights

• Extensive Model Support: Compatible with a vast range of models from the Hugging Face Hub, including LLaMA, Mistral, and multimodal VLMs.
• Cofiguration-Driven: Uses YAML files to manage datasets, training hyperparameters, and quantization settings.
• Advanced Training Techniques: Supports full fine-tuning, preference tuning (DPO, ORPO), and reward modeling (RM/PRM).
• High-Performance Optimizations: Includes multipacking, sequence parallelism, and various 4-bit and 8-bit quantization methods (FP8, NVFP4).
• Cloud and Local Flexibility: Provided as Docker images and PyPI packages for easy deployment across various cloud GPU providers.