How to set my value for the embedding size?

[aropb]

Description

The model supports setting the custom length of the embedded array. How do I set my value?

MRL Support indicates whether the embedding model supports custom dimensions for the final embedding.

[martindevans]

This sounds like something we don't currently support (I've certainly never heard of it). Could you link some things in the main llama.cpp repo related to this? That'll help inform development for the next version.

[aropb]

I didn't find it.

Qwen3 embedding and rerank:
ggml-org/llama.cpp#14029

[CISC]

You don't really set it, you just crop/ignore parts of the output, Matryoshka tuned models have the embedding weights sorted by importance.

[aropb]

I do not think so.
Then they would not have written about whether the model supports this mechanism (MRL Support) or not.

[CISC]

I do not think so. Then they would not have written about whether the model supports this mechanism (MRL Support) or not.

I see, then it's probably possible to prompt the model to change the output, I see no further mention of it anywhere though.

[aropb]

me too, which is weird :)

[aropb]

You were right, the vector is being clipped.

https://huggingface.co/blog/matryoshka#why-would-you-use-%F0%9F%AA%86-matryoshka-embedding-models