[Bug]: 执行acl初始化操作后，创建NPU tensor触发abort-pytorch-AtomGit

已关闭

[Bug]: 执行acl初始化操作后，创建NPU tensor触发abort #1528

absoler创建于 1月29日关闭于 7月4日

1月29日创建

环境信息

- 操作系统 Huawei Cloud EulerOS release 2.0
- 昇腾硬件信息 910B
- CANN软件版本 8.3.RC1
- 安装的对应软件版本 torch/torch-npu 2.7.1

🐛 问题描述

python执行：

import acl
acl.init()
acl.rt.set_device(0)
acl.rt.create_context(0)
import torch
t = torch.tensor([1, 2], device='npu')

触发abort

free(): invalid pointer
Aborted (core dumped)

欢迎加入社区，感谢您对社区的贡献 🎉!

absoler

1月29日修改了issue 的描述

ascend-robot成员

1月29日添加了label：bug

nikita-merkulov

1月29日评论：

I think this might be an uninitialization race. Did you provide above the whole code, or do you also work with the tensor t after? Try to add acl.finalize() in the end if you don't have it.

absoler

2月3日评论：

I think this might be an uninitialization race. Did you provide above the whole code, or do you also work with the tensor t after? Try to add acl.finalize() in the end if you don't have it.

@nikita-mee

No, this is a case executed in a python console, after creation of t, the core dump will happen. Here's illustration:

nikita-merkulov

2月9日评论：

I think this might be an uninitialization race. Did you provide above the whole code, or do you also work with the tensor t after? Try to add acl.finalize() in the end if you don't have it.

@nikita-mee

No, this is a case executed in a python console, after creation of t, the core dump will happen. Here's illustration:

@absoler

Right, no connection to uninitialization then. Then maybe someone else would be able to help...

huangyunlong成员

4月10日评论：

Thank you for your feedback. You can upgrade to torch_npu versions after 7.3.0

chenrayray成员

7月4日评论：

您好，本Issue因长期无有效互动，现暂时关闭。如问题依然存在，请重新打开并提供最新复现信息，我们会继续跟进。感谢理解！

chenrayray成员

7月4日 issue状态由 TODO 改变为 DONE

chenrayray成员

7月4日关闭了 issue

ascend-robot成员

7月4日添加了label：resolved