[New Model]: Codestral Mamba

[K-Mistele]

The model to consider.

Mamba Codestral: https://huggingface.co/mistralai/mamba-codestral-7B-v0.1

Highlights:

• SOTA 7B code model
• theoretically unlimited context length; tested up to 256k
• inference is linear-complexity with respect to sequence length, compared to transformers which is quadratic-complexity

The closest model vllm already supports.

Jamba seems to be the closest model, since it is Mamba-based: https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/models/jamba.py

What's your difficulty of supporting the model you want?

Mamba is a non-transformer architecture, but there is already a mamba-based model supported, so it's unclear how difficult it would be to support.

[simon-mo]

cc @tlrmchlsmth who is working on it

[digantamisra98]

Any updates on this?

[tlrmchlsmth]

@digantamisra98 I have a branch where I have this working -- planning to land #6484 soon (possibly today) and will follow up with Mamba2 support afterwards (which will include support for Codestral Mamba)

[github-actions]

This issue has been automatically marked as stale because it has not had any activity within 90 days. It will be automatically closed if no further activity occurs within 30 days. Leave a comment if you feel this issue should remain open. Thank you!