For cuda slices

This is likely a non-issue because we are only using a single stream. Since each launch is guaranteed to happen in order, as long as the calling function had exclusive access, then so will the kernel (since only 1 happens at a time).

Additionally, only kernel functions can mutate these values, so it's impossible to mutate values on host & device at the same time.

Finally, even dropping values occurs on the stream, so you won't be able to use a cuda slice after it is freed.

For primitives

It is possible to mutate these on host device. Recommend to pass primitives by value instead of by reference into launch_async

from cudarc.