I have found that the agent component achieves resource oversubscription and mixed online/offline deployment based on CPU and memory resources. Is it possible to achieve these based on GPU resources as well?

[sakuradiss]

Please describe your problem in detail

I have found that the agent component achieves resource oversubscription and mixed online/offline deployment based on CPU and memory resources. Is it possible to achieve these based on GPU resources as well?

Any other relevant information

No response

[Monokaix]

Can you describe your user case more detail？Currently volcano also support GPU virtualization, you can see https://github.com/volcano-sh/volcano/blob/master/docs/user-guide/how_to_use_gpu_number.md for more detail: )

[Monokaix]

/area agent

[Monokaix]

/area colocation

[sakuradiss]

Can you describe your user case more detail？Currently volcano also support GPU virtualization, you can see https://github.com/volcano-sh/volcano/blob/master/docs/user-guide/how_to_use_gpu_number.md for more detail: )你能更详细地描述一下你的用户案例吗？目前volcano也支持GPU虚拟化，你可以查看https://github.com/volcano-sh/volcano/blob/master/docs/user-guide/how_to_use_gpu_number.md了解更多细节：)

When I was reviewing the link https://github.com/volcano-sh/volcano/blob/master/docs/design/colocation/Overview.md, I noticed that their resource management and limitations are focused on CPU and memory. However, I would like to use GPU resources and memory for building a complex model training and inference system that supports both online and offline mixed deployment. Is this approach feasible? Thank you for your answer.

[Monokaix]

Can you describe your user case more detail？Currently volcano also support GPU virtualization, you can see https://github.com/volcano-sh/volcano/blob/master/docs/user-guide/how_to_use_gpu_number.md for more detail: )你能更详细地描述一下你的用户案例吗？目前volcano也支持GPU虚拟化，你可以查看https://github.com/volcano-sh/volcano/blob/master/docs/user-guide/how_to_use_gpu_number.md了解更多细节：)

When I was reviewing the link https://github.com/volcano-sh/volcano/blob/master/docs/design/colocation/Overview.md, I noticed that their resource management and limitations are focused on CPU and memory. However, I would like to use GPU resources and memory for building a complex model training and inference system that supports both online and offline mixed deployment. Is this approach feasible? Thank you for your answer.

Can GPU virtualization meet your need?