GitLab

use a similar strategy as llama.cpp for deciding where tensors should be
allocated. this will be improved later to be aware of usable memory
before assigning the tensor