Skip to content
This repository has been archived by the owner on Jun 24, 2024. It is now read-only.

A tool to convert multipart models to singlepart models #150

Closed
philpax opened this issue Apr 22, 2023 · 4 comments
Closed

A tool to convert multipart models to singlepart models #150

philpax opened this issue Apr 22, 2023 · 4 comments
Assignees
Labels
issue:enhancement New feature or request
Milestone

Comments

@philpax
Copy link
Collaborator

philpax commented Apr 22, 2023

Our new loader does not support multipart models because they do unfun things with tensor strides. Instead of trying to support both loaders indefinitely, it's much simpler to build a tool that will convert a GGML/GGMF multipart model to a GGJT singlepart model. The multipart models are relatively rare, so not having native support shouldn't pose too much of an issue.

@philpax philpax added the issue:enhancement New feature or request label Apr 22, 2023
@philpax philpax self-assigned this Apr 22, 2023
@philpax
Copy link
Collaborator Author

philpax commented Apr 22, 2023

This will need to be built on top of #84.

@philpax
Copy link
Collaborator Author

philpax commented Apr 26, 2023

There are some changes in #141 around loading infrastructure. Wait for that one to be merged first.

@philpax
Copy link
Collaborator Author

philpax commented May 9, 2023

There are models floating around that aren't fully compliant with the format (e.g. missing scores). This tool could be used as a general-purpose solution for reading in those models with some kind of compatibility mode, fixing them up, and writing a fully compliant GGJT file.

@philpax philpax added this to the 0.2 milestone May 18, 2023
@philpax
Copy link
Collaborator Author

philpax commented May 22, 2023

Closing this because multipart GGML models are pretty rare now.

@philpax philpax closed this as completed May 22, 2023
Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.
Labels
issue:enhancement New feature or request
Projects
None yet
Development

No branches or pull requests

1 participant