MindNLP is an open-source NLP and LLM framework built on MindSpore, designed for researchers and developers. It offers a user-friendly interface and high performance, aiming to simplify the construction and training of NLP models, with compatibility for Hugging Face models and datasets.

How It Works

MindNLP leverages MindSpore's capabilities to provide a unified dynamic and static graph execution environment. This allows for easy switching between modes with a single line of code (mindspore.jit), enabling rapid performance gains without sacrificing ease of use. It supports advanced features like distributed parallel inference for large models and quantization algorithms (SmoothQuant, int8) for efficient deployment across various hardware, including Ascend and GPU.

Quick Start & Requirements

• Install via pip: pip install mindnlp
• Source install: pip install git+https://github.com/mindspore-lab/mindnlp.git or clone and build.
• Python versions: 3.7.5-3.11, depending on MindNLP version.
• MindSpore versions: 1.8.1 to daily builds.
• Full platform support includes Ascend 910, Ascend 310B, GPU, and CPU.

Highlighted Details

• Supports 250+ pretrained models with Hugging Face transformers-like APIs.
• Achieves 85ms/token inference speed for Llama on Ascend (dynamic graph) and 45ms/token (static graph).
• Offers Sentence Transformer support for efficient RAG development.
• Includes comprehensive data processing tools and a simplified training engine.

Maintenance & Community

MindSpore NLP SIG is the main development team. Community engagement is encouraged via GitHub Issues.

Licensing & Compatibility

Released under the Apache 2.0 license, permitting commercial use and integration with closed-source projects.

Limitations & Caveats

The dynamic version is still under development. Version compatibility between MindNLP and MindSpore requires careful attention to the provided table.