如果使用国产卡的话，还需要安装nvidia-container-toolkit吗？

[janetat]

问题1：如果我只有国产卡，例如昇腾，还需要进行preparing-your-gpu-nodes操作吗？
问题2：如果我的集群是异构的，有NVIDIA卡的节点A，又有昇腾卡的节点B，那么全部节点都需要安装nvidia-container-toolkit吗？

[CoderTH]

Question1: if you are using ascend npu, you should refer to this document ：https://github.com/Project-HAMi/HAMi/blob/master/docs/ascend910b-support_cn.md
Question 2: only nvidia gpu nodes need to install nvidia-container-toolkit,ascend. Nodes need to install ascend-docker-runtime.

[janetat]

Thanks! @CoderTH

问题3：
一个节点只允许一个厂商的GPU么？能不能同时存在节点A既有NVIDIA的GPU，而且有昇腾的NPU？

问题4：
hami的nvidia device plugin Allocate接口返回信息，容器运行时根据这些信息来进行挂载设备文件、动态库。为什么还需要nvidia-container-runtime呢？原生的runc不也可以根据Allocate响应信息来进行挂载么？

[wawa0210]

@janetat

Q3: It is recommended that a single node use the same type of GPU. Heterogeneous nodes are not recommended.

[hzliangbin]

@janetat
In addition to handling device mounting and library files, the nvidia-container-runtime also needs to manage additional configurations related to the GPU (such as environment variables, CUDA paths, etc.), which runc does not provide.