[QST] How to avoid cuda resource warning when using cargra? about raft HOT 4 CLOSED

Hi @FakeEnd

This warning has actually been removed in the latest nightlies for the 24.02 version because it's pretty much always recommended to use a pool allocator as the default allocator. We use RMM memory manager, which uses cudaMalloc and cudaFree by default but the vector search tutorial has instructions for setting a pool allocator as the default allocator. If you are using Python, you can find a similar example in the RMM readme (at https://github.com/rapidsai/rmm).

from raft.

here is the Python example for setting an RMM pool allocator.

from raft.

@FakeEnd since you are using PyTorch, another thing worth pointing out is that RMM supports PyTorch so the memory pool you create can be shared by both. We have found this to be a really useful feature when working across library boundaries.

from raft.

@FakeEnd since you are using PyTorch, another thing worth pointing out is that RMM supports PyTorch so the memory pool you create can be shared by both. We have found this to be a really useful feature when working across library boundaries.

Thank you:)

from raft.