Use is_matryoshka to denote whether a model is compatible with Matryoshka Representation Learning (MRL).
I come from the vllm community, and recently vllm has added support for Matryoshka Representation Learning (MRL).
Not all embeddings models support MRL. Changing the output dimension for models that do not support MRL will lead to poor results.
We hope that the open source community will adopt the terms “is_matryoshka ” or “matryoshka_dimensions ” to denote whether a model is compatible with Matryoshka Representation Learning (MRL).
hey @noooop0000 , thanks for the update on vllm's support for MRL. Can you advise on what terms should we use if we only officially support MRL at 256 dimensions? Same for other models you posted this under.
In vLLM, we simply check the existence of the fields is_matryoshka or matryoshka_dimensions inside config.json.
User can choose any output dimension for now.
https://github.com/vllm-project/vllm/blob/6b2427f995a81377e5758b34c8bb4c66db2f67bc/vllm/pooling_params.py#L31C1-L39C70
We might be able to control the user to only use certain dimensions with matryoshka_dimensions
I'm very sorry, I don't know how to modify this PR, whether it's web or local. So I proposed a new PR