convert-pth-to-ggml.py has become very slow

[DJtheRedstoner]

A few weeks ago, convert-pth-to-ggml.py ran reasonably fast for me (still took a while to convert the larger models). Now (seemingly after the format changes to allow mmap) it is very slow. It will likely take over an hour to convert the 13B model. I'm running it in an anaconda environment on Windows using miniconda3. I can provide additional specs or debugging information if necessary.

Edit: looking at cProfiler trace, it looks like the file.seek() method is performing very poorly. Maybe this is because I'm on Windows and it performs better on linux?

[DJtheRedstoner]

By writing the models to my SSD instead of my HDD, converting them is much much faster. Maybe some room for improvement in the future (the naive solution if you have enough ram is just to store the entire thing in ram with BytesIO then write it to disk, but I don't have enough RAM for this), but random writes on HDDs are just not great.