Open
Description
I serialized my LLM model with cereal, the serialized model size starting from 3GB, and the weight loading time does matter to the performance.
With this test program (https://raw.githubusercontent.com/lemire/Code-used-on-Daniel-Lemire-s-blog/refs/heads/master/2012/06/26/ioaccess.cpp), I found on my Linux system mmap could read 389M integers per second, but c++ istream could only read 196M.
I would therefore like to check if it possible to add a mmap backend for BinaryInputArchive if the target platform is Unix-like.
Metadata
Metadata
Assignees
Labels
No labels