HierarchicalKV icon indicating copy to clipboard operation
HierarchicalKV copied to clipboard

How is the warmup function implemented?

Open imhandsome opened this issue 1 year ago • 3 comments

imhandsome avatar Jan 13 '25 03:01 imhandsome

Hi @imhandsome , the mechanism of warmup is under development. Any requirement input is encouraged, thanks!

jiashuy avatar Jan 13 '25 03:01 jiashuy

Excuse me,1. Is the follow-up warmup plan to keep the pin memory unchanged? (or use cudaMallocManaged UVM),2. warmup Provide users with a warmup interface to pre-exchange a portion of the hot embeding to the GPU? @jiashuy

imhandsome avatar Jan 13 '25 07:01 imhandsome

Excuse me,1. Is the follow-up warmup plan to keep the pin memory unchanged? (or use cudaMallocManaged UVM),2. warmup Provide users with a warmup interface to pre-exchange a portion of the hot embeding to the GPU? @jiashuy

For 1, pinned memory is given priority consideration as we can control the data movement more concretely; it will remain unchanged from my side.

For 2, for the mechanism of warmup, you can choose between an explicit interface and implicit exchange, for that users can hide it in application, and please refer to the next release version.

jiashuy avatar Jan 13 '25 14:01 jiashuy

Do we have any progress on warmup implementation?

jiafanou avatar May 06 '25 07:05 jiafanou

Hi @imhandsome we have just made a design about warmup, and will published soon. Thanks~

jiashuy avatar May 08 '25 05:05 jiashuy

Thanks @jiashuy . Do you have rough ETA of release?

jiafanou avatar May 08 '25 10:05 jiafanou

any update on this? @jiashuy Thanks.

EddieBurning avatar Jun 05 '25 13:06 EddieBurning