Convert MPT checkpoints to FT format

[dskhudia]

A script to convert MPT checkpoints to FT format. A script to run with FasterTransformer is coming soon in a followup PR.

Partially fixes #67

Thanks @bheilbrun for sharing your code.

[hanlint]

To avoid nesting too many folders, can we remove the ft folder and just put this in the root directory of inference ?

Also, while this is a script, users may want to call the main conversion function progammatically. Can we rename that method something more clear, e.g. convert_to_faster_transformer and type/document the input variables?

[dskhudia]

To avoid nesting too many folders, can we remove the ft folder and just put this in the root directory of inference ?

@hanlint : I created a folder for two reasons 1) There will be at least one another script so all FT related stuff can stay in a single folder 2) Readme of this folder can explain all the FT related stuff in a single place.

[abhi-mosaic]

Left some small comments, but overall I think this would be great to merge soon to get community feedback. I would also love to have tests in this repo but I understand that the FT deps+installation is quite tedious. So as long as @dskhudia you have tested manually, no need to block on that.

I do agree with @hanlint that it makes sense to move these files up a level to scripts/inference, even if there end up being 2 or 3 scripts for FT. I'd like to have it all in one place until it grows big enough to necessitate a folder.

@dskhudia would it be ok to move this script a folder and append the short README.md here to llm/inference/README.md?

[dskhudia]

would it be ok to move this script a folder and append the short README.md here to llm/inference/README.md

moved it to the inference folder.

[bheilbrun]

Thanks in particular for the code quality improvements!