[Feature]: Support Multiple Tasks Per Model

[FurtherAI]

🚀 The feature, motivation and pitch

Requesting this for V1 #11862

The idea is pretty simple, it would be nice to be able to, e.g., get generations and embeddings out of a single model. An example use case is when you have a LoRA for generation and a LoRA for embedding on top of the same base model. Deploying two vLLM servers is really inefficient for accomplishing this.

Alternatives

A lesser feature would be one task per LoRA, but it's better to be general if possible.

Additional context

No response

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[YingxuH]

same need

[lstrozzi]

same need

[abid-jawad]

+1! Came here to create a feature request issue and found this. This would be a great feature indeed.

[Razos05]

+1 would be very useful as the Embeddings endpoint is available even when the embedding doesn't work and isn't activated :\

[yetanotherinbar]

+1 Please add this feature

[noamgat]

+1 This will be very useful for me

[biba10]

+1 I would also greatly appreciate this feature

[Torabi-srh]

+1 This will be very useful for me

[vaclcer]

Indeed. Typically encoder only models have often embeddings and ranking function.

[pauli31]

+1

[alphazyx]

+1. Came here for this.

[Soum-Soum]

+1

[YungTesla]

+1