300iduo，一张310P3 96G的卡，mindie可以跑qwen2.5vl吗？_MindIE_昇腾论坛

# 300iduo，一张310P3 96G的卡，mindie可以跑qwen2.5vl吗？_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-02116188640492076236-1-1.html

**生成时间**: 2025-08-27 10:33:57

---

## 问题描述

mindservice可以正常启动text格式也可以正常写入API但是image_url就报错报错内容为[error] [linear_operation.cpp:629] LinearOperation Error code: 8. value of inTensor0 k and value of inTensor1 k should be equal, inTensor0 k = 3424, inTensor1 k = 3420似乎是张量问题可以解决吗 我用mindie 试了也是这个错误然后用 vllm-ascend 部署是可以的 你好部署的教程或步骤有吗 自己改一改试试看 services: qwen2.5-vl-7b: container_name: qwen2.5-vl-7b image: quay.io/ascend/vllm-ascend:v0.9.2rc1-310p-openeuler command: > bash -c "export ASCEND_RT_VISIBLE_DEVICES=0 ; vllm serve /modelfile --host 0.0.0.0 --port 8089 --tensor-parallel-size 1 --max-model-len 16348 --max-num-batched-tokens 16384 --max-num-seqs 25 --block-size 128 --enable-prefix-caching --trust-remote-code --enforce-eager --dtype float16 --gpu-memory-utilization 0.8 --served-model-name qwen2.5-vl-7b" network_mode: host privileged: true shm_size: 1g devices: - /dev/davinci_manager - /dev/hisi_hdc - /dev/devmm_svm - /dev/davinci0 volumes: - /usr/local/Ascend/driver:/usr/local/Ascend/driver:ro - /usr/local/sbin:/usr/local/sbin:ro - ...

## 相关代码

### 代码示例 1

```
services:
qwen2.5-vl-7b:
container_name: qwen2.5-vl-7b
image: quay.io/ascend/vllm-ascend:v0.9.2rc1-310p-openeuler
command: >
bash -c "export ASCEND_RT_VISIBLE_DEVICES=0 ; vllm serve /modelfile --host 0.0.0.0 --port 8089 --tensor-parallel-size 1 --max-model-len 16348 --max-num-batched-tokens 16384 --max-num-seqs 25 --block-size 128 --enable-prefix-caching --trust-remote-code --enforce-eager  --dtype float16 --gpu-memory-utilization 0.8 --served-model-name qwen2.5-vl-7b"
network_mode: host
privileged: true
shm_size: 1g
devices:
- /dev/davinci_manager
- /dev/hisi_hdc
- /dev/devmm_svm
- /dev/davinci0
volumes:
- /usr/local/Ascend/driver:/usr/local/Ascend/driver:ro
- /usr/local/sbin:/usr/local/sbin:ro
- /PATH/models/Qwen/Qwen2___5-VL-7B-Instruct:/modelfile:ro
tty: true
stdin_open: true
restart: unless-stopped复制
```

### 代码示例 2

```
services:
qwen2.5-vl-7b:
container_name: qwen2.5-vl-7b
image: quay.io/ascend/vllm-ascend:v0.9.2rc1-310p-openeuler
command: >
bash -c "export ASCEND_RT_VISIBLE_DEVICES=0 ; vllm serve /modelfile --host 0.0.0.0 --port 8089 --tensor-parallel-size 1 --max-model-len 16348 --max-num-batched-tokens 16384 --max-num-seqs 25 --block-size 128 --enable-prefix-caching --trust-remote-code --enforce-eager  --dtype float16 --gpu-memory-utilization 0.8 --served-model-name qwen2.5-vl-7b"
network_mode: host
privileged: true
shm_size: 1g
devices:
- /dev/davinci_manager
- /dev/hisi_hdc
- /dev/devmm_svm
- /dev/davinci0
volumes:
- /usr/local/Ascend/driver:/usr/local/Ascend/driver:ro
- /usr/local/sbin:/usr/local/sbin:ro
- /PATH/models/Qwen/Qwen2___5-VL-7B-Instruct:/modelfile:ro
tty: true
stdin_open: true
restart: unless-stopped
```

## 完整内容

mindservice可以正常启动text格式也可以正常写入API但是image_url就报错报错内容为[error] [linear_operation.cpp:629] LinearOperation Error code: 8. value of inTensor0 k and value of inTensor1 k should be equal, inTensor0 k = 3424, inTensor1 k = 3420似乎是张量问题可以解决吗 我用mindie 试了也是这个错误然后用 vllm-ascend 部署是可以的 你好部署的教程或步骤有吗 自己改一改试试看 services: qwen2.5-vl-7b: container_name: qwen2.5-vl-7b image: quay.io/ascend/vllm-ascend:v0.9.2rc1-310p-openeuler command: > bash -c "export ASCEND_RT_VISIBLE_DEVICES=0 ; vllm serve /modelfile --host 0.0.0.0 --port 8089 --tensor-parallel-size 1 --max-model-len 16348 --max-num-batched-tokens 16384 --max-num-seqs 25 --block-size 128 --enable-prefix-caching --trust-remote-code --enforce-eager --dtype float16 --gpu-memory-utilization 0.8 --served-model-name qwen2.5-vl-7b" network_mode: host privileged: true shm_size: 1g devices: - /dev/davinci_manager - /dev/hisi_hdc - /dev/devmm_svm - /dev/davinci0 volumes: - /usr/local/Ascend/driver:/usr/local/Ascend/driver:ro - /usr/local/sbin:/usr/local/sbin:ro - /PATH/models/Qwen/Qwen2___5-VL-7B-Instruct:/modelfile:ro tty: true stdin_open: true restart: unless-stopped复制

---

## 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**: 昇腾, GPU, 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。*