Release Notes

vLLM-MindSpore Plugin 0.5.1 Release Notes

The vLLM-MindSpore Plugin 0.5.1 version is compatible with vLLM 0.11.0. Below are the new features and models supported in this release.

New Features

• Architecture Adaptation: Completed architectural upgrades and adapted to vLLM 0.11.0, fully inheriting all features of the older version.

• Performance Optimization: Enhanced the V1 architecture scheduling to improve inference performance in high-concurrency, long-input scenarios.

• Graph Capture Feature Support: AclGraph graph capture functionality is supported.

New Models

• Qwen3-VL Model Series

• GLM Model Series:

  • GLM-4.1V Native Model

Problem Fixes

• Encoder Memory Leak Issue

  • Fix for vLLM

  • Fix for vLLM-MindSpore Plugin

Contributors

Thanks to the following contributors for their efforts:

alien_0119, candyhong, can-gaa-hou, ccsszz, cs123abc, dayschan, Erpim, fary86, hangangqiang, horcam, huandong, huzhikun, i-robot, jiahaochen666, JingweiHuang, lijiakun, liu lili, lvhaoyu, lvhaoyu1, moran, nashturing, one_east, panshaowu, pengjingyou, r1chardf1d0, tongl, TrHan, tronzhang, TronZhang, twc, uh, w00521005, wangpingan2, WanYidong, WeiCheng Tan, wusimin, yangminghai, yyyyrf, zhaizhiqiang, zhangxuetong, zhang_xu_hao1230, zhanzhan1, zichun_ye, zlq2020

Contributions to the project in any form are welcome!