You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
In order to load CUDA into a VEC, we added support for faking the PID. (CUDA picks a port to use based off the current PID, so we can't load multiple copies of it without faking the PID). Currently we just make up a PID and hope for the best. It'd be nice if we could make this infrastructure more robust. For example, we could launch a separate empty process for each and then associate it's PID as the fake to use within a VEC. There probably are other options as well.
The text was updated successfully, but these errors were encountered:
We should probably ask the CUDA developers to fix things on their end, but I doubt we'll see any sort of quick response there (if any at all). Given that they break ABI with CUDA so frequently, being able to load it into a linker namespace seems like a reasonable ask.
I don't think an empty process is a good idea. The problem is that /proc/<pid> would be total lies at that point. What about using the thread ID of a empty thread? I think threads are in the same "ID space" as processes on LInux still. Though I could be wrong.
In order to load CUDA into a VEC, we added support for faking the PID. (CUDA picks a port to use based off the current PID, so we can't load multiple copies of it without faking the PID). Currently we just make up a PID and hope for the best. It'd be nice if we could make this infrastructure more robust. For example, we could launch a separate empty process for each and then associate it's PID as the fake to use within a VEC. There probably are other options as well.
The text was updated successfully, but these errors were encountered: