One of the issues with whisper.cpp is that it uses read() to copy the model into memory. It would be nice to be able to have it mmap() tensors into memory instead. It's most useful when using whisper as a command line utility, since it greatly reduces loading time.
One of the issues with whisper.cpp is that it uses read() to copy the model into memory. It would be nice to be able to have it mmap() tensors into memory instead. It's most useful when using whisper as a command line utility, since it greatly reduces loading time.