FudanNLP/fastNLP

Health v2 · maintenance only

Adoption not part of the score

3141 stars · 449 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

fastNLP is a lightweight, modularized and extensible NLP framework in Python that reduces engineering boilerplate such as data processing loops, training loops, and multi-GPU execution. Version 1.0+ was redesigned to support multiple deep learning backends (primarily PyTorch) with features like fp16 switching, multi-GPU training, and ZeRO optimization.

Use cases

• train a text classification model with BERT in PyTorch
• process Chinese NLP datasets with tokenization and caching
• run multi-GPU training without changing my code
• avoid writing boilerplate training loops for NLP tasks
• apply functions over dataset fields with multiprocessing
• fine-tune pretrained transformers on Chinese sentiment data
• switch to fp16 or ZeRO optimization for faster training

When to choose

• you do NLP research or prototyping in Python, especially Chinese NLP
• you want a lightweight training/data-processing layer on top of PyTorch
• you need easy multi-GPU, fp16, or ZeRO support with minimal code changes

When to avoid

• you need a framework with a large active community and ecosystem like Hugging Face
• you depend on pre-1.0 fastNLP APIs, which are incompatible with 1.0+
• you need non-NLP general-purpose deep learning tooling

Facets

framework · maturity maintenance

machine-learning nlp deep-learning llm-training machine-learning deep-learning python chinese-nlp text-classification training-loop pytorch data-processing multi-gpu natural-language-processing

Member repositories

For agents

markdown · JSON · MCP: product_card(name="FudanNLP/fastNLP")