Add Nvembed v2 model

[cdoko]

NV-Embed-v2 is a text embedding model that ranks No. 1 (as of Nov 25 2024) on the MTEB benchmark.

This implementation reproduces the example in the model card.

License: CC-BY-NC-4.0

Please let me know if any changes are required!

[LaurentMazare]

Thanks for adding this model, one tricky bit is that the candle-transformers crate has an apache/mit license so we cannot really include code that is CC-BY-NC whithin it - having models for which the code is apache/mit and the weights is CC-BY--NC would be fine though.
So if you can base your implementation on code that doesn't have the NC restriction that would make it possible to have this in candle-transformers.

[cdoko]

Hi, I'm new to open-source licensing, I want to clarify how it works. The NVEmbedV2 repository contains a CC-BY-NC-4.0 license in the README, but doesn't have explicit licensing information in individual files, including the implementation code. Is it generally assumed that the license applies to the entire repository, including both code and weights?

The README states 'This model should not be used for any commercial purpose. Refer the license for the detailed terms.' Does this language imply that the CC-BY-NC-4.0 license only applies to the model weights, or also to the accompanying implementation code?

To clarify, I added the CC-BY-NC-4.0 license to my implementation code, but I did this out of caution and uncertainty.

If necessary, I'd be happy to re-implement.

[LaurentMazare]

(disclaimer: not a lawyer neither a specialist of license)
The repo on huggingface has a license of "cc-by-nc-4.0" and unless it's explicitely stated otherwise in the files in the repo, my understanding is that it applies to them, so in particular it applies to the modeling_nvembed.py file and this cannot be used as a base for the implementation itself.
If you can re-implement the model based on the paper or an alternative open-source implementation with a more open source license that would be great.

[cdoko]

Thanks for the clarification on the licensing. I've updated the implementation to be based on the paper and weights. Looking forward to your feedback!

[LaurentMazare]

Looks all good to me, I've merged it, thanks!