mindie2.0用2卡910B推理Qwen2.5-VL-3B正常，而切换为单卡推理报错_MindIE_昇腾论坛

# mindie2.0用2卡910B推理Qwen2.5-VL-3B正常，而切换为单卡推理报错_MindIE_昇腾论坛

## 概述

本文档基于昇腾社区论坛帖子生成的技术教程。

**原始链接**: https://www.hiascend.com/forum/thread-02117188546947472207-1-1.html

**生成时间**: 2025-08-27 10:33:58

---

## 问题描述

版本mindie2.0.T3CANN8.0 问题描述使用2卡启动mindie serviceQwen2.5-VL-3B可以正常推理图片切换到单卡修改 mindie-service/conf/config.json 中的 "npuDeviceIds": [[0,1]] 改为"npuDeviceIds": [[0]] "worldSize": 2 改为 "worldSize": 1 重新启动服务mindie-service/bin/mindieservice_daemon 使用curl请求服务报错信息如下 [llm][ERROR][acl_nn_operation.cpp:200] W16A16LinearNode call ExecuteAclNNOp fail, error:161001 [error][274652][runner.cpp:115] PluginRunner_101_0_0:1 Execute Failed. st: 12 [error][274652][graph_runner.cpp:906] LinearNoQuantRunner_101_0:0 node[0] execute fail, runner name:PluginRunner [error][274652][runner.cpp:115] LinearNoQuantRunner_101_0:1 Execute Failed. st: 12 [error][274652][graph_runner.cpp:906] LmHeadRunner_101:0 node[0] execute fail, runner name:LinearNoQuantRunner [error][274652][runner.cpp:115] LmHeadRunner_101:1 Execute Failed. st: 12 [error][274652][operation_base.cpp:726] LmHead_101 execute LmHeadRunner fail [error][274652][operation_base.cpp:764] LmHead_101 execute fail, error code: 12 使用 mindie2.1.T3的镜像在单卡910B上启动容器是可以正常...

## 完整内容

版本mindie2.0.T3CANN8.0 问题描述使用2卡启动mindie serviceQwen2.5-VL-3B可以正常推理图片切换到单卡修改 mindie-service/conf/config.json 中的 "npuDeviceIds": [[0,1]] 改为"npuDeviceIds": [[0]] "worldSize": 2 改为 "worldSize": 1 重新启动服务mindie-service/bin/mindieservice_daemon 使用curl请求服务报错信息如下 [llm][ERROR][acl_nn_operation.cpp:200] W16A16LinearNode call ExecuteAclNNOp fail, error:161001 [error][274652][runner.cpp:115] PluginRunner_101_0_0:1 Execute Failed. st: 12 [error][274652][graph_runner.cpp:906] LinearNoQuantRunner_101_0:0 node[0] execute fail, runner name:PluginRunner [error][274652][runner.cpp:115] LinearNoQuantRunner_101_0:1 Execute Failed. st: 12 [error][274652][graph_runner.cpp:906] LmHeadRunner_101:0 node[0] execute fail, runner name:LinearNoQuantRunner [error][274652][runner.cpp:115] LmHeadRunner_101:1 Execute Failed. st: 12 [error][274652][operation_base.cpp:726] LmHead_101 execute LmHeadRunner fail [error][274652][operation_base.cpp:764] LmHead_101 execute fail, error code: 12 使用 mindie2.1.T3的镜像在单卡910B上启动容器是可以正常推理 Qwen2.5-VL-3B 或 Qwen2.5-VL-7B模型

---

## 技术要点总结

基于以上内容，主要技术要点包括：

1. **问题类型**: 错误处理
2. **涉及技术**: NPU, CANN, AI, MindIE
3. **解决方案**: 请参考完整内容中的解决方案

## 相关资源

- 昇腾社区: https://www.hiascend.com/
- 昇腾论坛: https://www.hiascend.com/forum/

---

*本文档由AI自动生成，仅供参考。如有疑问，请参考原始帖子。*