xtuner

Why we included this project

Training a large model is usually more about wrestling with the training engine than the model itself, and XTuner exists to ease that pain. It's a PyTorch-native toolkit from the InternLM team covering pre-training, LoRA and QLoRA fine-tuning, and reinforcement learning for text and multimodal models. The V1 engine is built for the Mixture-of-Experts models now common in research and production, training 200B-parameter MoEs on 64k sequences without expert or sequence parallelism. It scales down just as well, fine-tuning a 7B model on a single 8GB GPU. Teams that fine-tune Qwen, DeepSeek, or similar open models get a practical, actively maintained alternative to Megatron and DeepSpeed.