PaddleNLP
A large-language-model development kit built on the PaddlePaddle framework, supporting efficient training across hardware types, lossless compression and high-performance inference.
PaddleNLP is a large-language-model development kit built on Baidu's PaddlePaddle deep-learning framework, supporting efficient training of large models across a range of hardware, lossless model compression, and high-performance inference, aimed at helping developers build production large-model applications.
It has added support for the Qwen3 model family, including six dense models and two mixture-of-experts models with dual thinking modes, and full support for the DeepSeek V3 and R1 series along with QwQ-32B, including FP8, INT8 and 4-bit quantized inference and multi-token-prediction speculative decoding for faster single-machine throughput. Other additions include a PP-UIE information-extraction model supporting long-document extraction, a MergeKit tool for merging models, large-scale embedding training, a rebuilt trainer for one-click supervised fine-tuning, and optimizer-state offloading and reloading for lower-memory training.
Its Unified Checkpoint system reduces model storage time and space through asynchronous, compressed saving and adaptive conversion across distributed training strategies, cutting storage needs by a large margin compared with earlier versions.