Bradley Dice
Bradley Dice
I am also considering upstreaming some of these docs to CCCL as the MR implementations there become stable. We’re on the same page for what the long term might look...
There are also changes in managed memory with CUDA 13, which we should explain here. The driver now has pools it can use for managed memory without requiring a large...
> The async mr leverages CUDA APIs that use a CUDA driver-managed pool under the hood, but we don't have any direct visibility into or control over that pool. As...
I'd like to gather a list of projects using RMM that require C++17 support and cannot move to C++20. Currently it sounds like `ucxx` might be on that list (cc:...
@trivialfis -- sounds like XGBoost may have a requirement for C++17. Please share any additional information you have.
My worry about PTDS is that it is hare to sync across different threads’ default streams when the default value has a different meaning on every thread. Thread 2 cannot...
I had some discussions related to this topic offline. I am updating my position here a bit. > Do we have a way for threads to give a handle or...
@Matt711 Please ping me when you want a review here. Seems pretty close but I have some ideas for polishing it up.
Yes, I think removing the `noexcept` here and adding safety checks would be good. This has bitten users enough times in unexpected ways, it seems reasonable to have stronger protections...
/ok to test