Skip to content

Support for adding custom architecture models #38

@HosseinMohammadi00427

Description

@HosseinMohammadi00427

Hi there,

I’m trying to add jinna-v3 embeddings in the MLX format. However, I’ve run into an issue due to the model’s customized architecture — it appears to differ significantly from standard architectures like xlm-roberta.

For example, one of the errors I encountered mentioned that around 400 layers were not found in the standard xlm-roberta model definition.

I’d like to better understand how to approach implementing such customized or non-standard architectures in MLX.

Specifically:
• Is there a recommended approach or best practice for integrating models with custom architectures?
• Are there any guides, cookbooks, or examples that show how to handle non-standard transformer-based models in MLX?

Thanks in advance for your help and guidance! 🙏

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions