>>105674298
I think ktransformers let's you speed things up by having a copy of the model in each NUMA node.
So effectively you use half your total ram but double the theoughtput?
Something like that.