fastNLP is a modular natural language processing framework developed at Fudan University designed to reduce engineering overhead in NLP projects. It provides convenient data processing with multi-process support, customizable training loops, and seamless switching between fp16 training, multi-GPU, and ZeRO optimization without code changes. The framework supports multiple deep learning frameworks as backends, primarily PyTorch, and includes loaders for Chinese NLP datasets like ChnSentiCorp.