Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Setting default alpha_value and fixing loading some newer DeepSeekCoder GGUFs #6111

Merged
merged 1 commit into from
Jun 27, 2024

Conversation

mefich
Copy link
Contributor

@mefich mefich commented Jun 9, 2024

  1. Ensures alpha_value resets back to 1 after selecting a new model. Previously if any model had a unusual alpha_value in saved settings or of it was changed manually, this value wouldn't reset after choosing another model. This could cause poor performance of a newly loaded model.
  2. Ensures rope scaling is loaded from metadata of newer GGUF models, notably from newer DeepSeekCoder finetunes where this value is now saved in rope.scaling.factor instead of rope.scale_linear where it used to be saved in old GGUFs. Though there might be needed additional check for rope.scaling.type: linear or yarn. But I couldn't find information how to properly load Yarn models in TGW to decide if it's needed to check the scaling type.

Checklist:

…ss_pos_emb for newer GGUFs

1. Ensures alpha_value resets to 1 after selecting a new model. Previously it didn't do it and it could potentially cause loading a newly selected model with sub-optimal settings.
2. Ensures rope scaling is loaded from GGUF metadata, notably from newer DeepSeekCoder finetunes where this value is saved in rope.scaling.factor instead of rope.scale_linear. 
Though there might be needed additional check for scaling type: linear or yarn.
But I couldn't find information how to properly load Yarn models properly in TGW to decide if it's even needed to check the scaling type.
@oobabooga
Copy link
Owner

Your changes are correct and I wasn't aware that rope.scale_linear had been renamed to rope.scaling.factor. Nice catch.

About scaling type, any metadata not explicitly passed while loading a model gets automatically detected by llama.cpp, so this is not an issue.

@oobabooga oobabooga merged commit a85749d into oobabooga:dev Jun 27, 2024
PoetOnTheRun pushed a commit to PoetOnTheRun/text-generation-webui that referenced this pull request Oct 22, 2024
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants