The GPU startup takes about 1 second. As a result, the delay of using the ray task is too long. If the ray actor mechanism is used, secondary scheduling may occur. Are there any better solutions in the community to solve this problem?
Could you give more detail about this? What do you mean by GPU startup?
import cudf takes about 1 second
I see. I think right now, we don’t have very good solution right now
We have some work going on speeding up/pre warming the workers, but not done yet.