PaddleNLP is a Python-based library for large language model development, training, and inference with a large model zoo and multi-device support. It provides tools for training, fine-tuning, and high-performance inference, plus a variety of supported models and documentation.
Collecting history — the radar snapshots this repo daily. The trend line appears after 3 days of data (1 so far).
What it is
PaddleNLP is an easy-to-use library for LLM and SLM development built on the PaddlePaddle framework. It includes model zoo access and tools for training, fine-tuning, and high-performance inference across multiple hardware backends.
How it works
The project offers a multi-hardware training and inference stack. It supports data-parallel, group-sharded data parallel, tensor model parallel, and pipeline parallel strategies via Trainer with configurable distributed policies. It includes a Unified Checkpoint tool for storage optimization and asynchronous saving, plus a high-performance inference path with operator fusion.
Getting started
Documentation is available at PaddleNLP Documentation. The README lists supported hardware (NVIDIA GPUs, Kunlun XPU, Ascend NPU, Suiyuan GCU, Haiguang DCU) and model families (LLaMA, Llama2, Llama3, Baichuan, Bloom, ChatGLM, Qwen, DeepSeek variants, etc.).
Recent releases
- rl-v1.0.0 Stable RL v1.0.0 (2025-05-21): GRPO、RF++ ready
- v3.0.0-beta4 (2025-03-12): DeepSeek R1 family, PP-UIE information extraction model, faster inference and 4-bit quantization discussed in notes
- v3.0.0-beta3 (2024-12-16): New Llama-3.2 and DeepSeekV2, tokenizer improvements, SFTTrainer rewrite, 7% training performance gain, Unified Checkpoint updates
- v3.0.0-beta2 (2024-10-08): Qwen2.5 and Mixtral 8*22B models, tokenizer upgrades, multi-GPU inference support, more model/config options
- v3.0.0-beta1 (2024-08-22): Introduction of Yuan, mamba, jamba models, faster tokenizer, MoE optimizer broadcasting, safety fixes, Windows mmap improvements
Traction
12965 stars, 3034 forks (as reported)
Behind the repo
PaddleNLP is developed by PaddlePaddle project and integrates large-model tooling, optimization features, and distributed training capabilities across multiple hardware platforms.
Caveats
License: Apache-2.0. Created in 2021-02-05. Last push: 2026-05-23. Issues: 358. Language: Python.
Installation and usage notes
"README" mentions installation and usage details, including hardware support and model zoo, but exact install commands are not quoted here.