Feature Request - Support for LLM2VEC Encoder-Decoder framework.

[HydrogenBombaklot]

I would like to use and fine-tune an Llm2VEC model using MLX.

LLM2VEC adapts decoder-only LLMs and transforms them into bidirectional encoder-decoder models.

https://github.com/McGill-NLP/llm2vec

Integrating LLM2VEC support into MLX would expand MLXs' capabilities and introduce a whole new type of modules into the MLX ecosystem.