ml-explore
Architecture Requests for Mamba
I would like support the following architectures:
- Mamba
- MambaByte
- Mamba-2
- Mamba-hybrid (mamba + transformer)
- Mamba-2-hybrid (mamba2 + transformer)
These architectures are becoming quite common now and are supported by most major LLM libraries.
We have Mamba in MLX LM already and there is a PR for Mamba 2 (#1009 ).
As for the others, it would be helpful if you could point to Hugging Face repos for each model type. We can consider adding them on an ongoing basis.
We have Mamba in MLX LM already and there is a PR for Mamba 2 (#1009 ).
As for the others, it would be helpful if you could point to Hugging Face repos for each model type. We can consider adding them on an ongoing basis.
Mamba: https://huggingface.co/tiiuae/falcon-mamba-7b Mamba-2: https://huggingface.co/state-spaces/mamba2-2.7b MambaByte: https://huggingface.co/JunxiongWang/MambaByte_Books Mamba-Hybrid: https://huggingface.co/Zyphra/Zamba-7B-v1 Mamba2-Hybrid: https://huggingface.co/Zyphra/Zamba2-2.7B-instruct
We have Mamba in MLX LM already and there is a PR for Mamba 2 (#1009 ).
As for the others, it would be helpful if you could point to Hugging Face repos for each model type. We can consider adding them on an ongoing basis.
Zamba2 7b was just released. One of the best models of its size, it outperforms Llama3.2 11b and Mistral 7b in almost every benchmark. It is a Mamba2-hybrid model. https://www.zyphra.com/post/zamba2-7b