PaddlePaddle/PaddleFormers 1.1.1

PaddleFormers v1.1

Repository: PaddlePaddle/PaddleFormers

Tag: 1.1.1

Published: 2026-04-02T08:50:02Z

Prerelease: no

Release notes: PaddleFormers 1.1 is officially released! This release introduces several key features and improvements:

✨ New Features

1. Support multi-token prediction (MTP) training for GLM-4.5

We have added support for both single-step and multi-step MTP training for the GLM-4.5 model series. Leveraging the architectural advantages of MTP, developers can achieve significant improvements in inference efficiency. Additionally, for MTP module training scenarios, we have introduced a backbone network freezing toggle to flexibly meet the refined tuning requirements of various models.

2. Optimized VLM Model Performance

We have carried out in-depth optimizations on vision-language models. The Qwen3-VL 30B-A3B model delivers a 48% performance gain over the previous release and outperforms Megatron-LM by 6%.