[Feature] Request to Support Additional Embedding Models (e.g., DMetaSoul/Dmeta-embedding-zh)

[q77889966]

Checklist

• 1. If the issue you raised is not a feature but a question, please raise a discussion at https://github.com/sgl-project/sglang/discussions/new/choose Otherwise, it will be closed.
• 2. Please use English, otherwise it will be closed.

Motivation

Currently, I’m using the DMetaSoul/Dmeta-embedding-zh model from Hugging Face, and I noticed that its architecture is based on BertModel (see config.json). This model performs exceptionally well in Chinese NLP tasks, but it seems that it hasn’t been integrated into sglang yet.

Related resources

I’ve already tried the adaptation methods mentioned in the documentation to integrate this model into sglang, but unfortunately, I encountered issues and was unable to get it working. In my environment, sglang delivers the best performance for large models, and I’d prefer not to maintain two separate engines for different use cases. Supporting additional embedding models like DMetaSoul/Dmeta-embedding-zh would allow me to fully leverage sglang's capabilities while avoiding the overhead of managing multiple systems.
model link:https://huggingface.co/DMetaSoul/Dmeta-embedding-zh

[minleminzui]

thank for your request, all the embedding models we support is encoder only model, Sorry, we have no plans to support encoder only model at the moment.