prefix caching in llama.cpp with PyTorch

[thetagrid15]

Thanks, that helps. I don't think your conclusion is far off. Worth rechecking your defaults though, because one bad assumption can make the rest of the thread look confusing.