support for gemma-2

[almugabo]

Could you please add fine-tuning support for gemma-2 ? It has good good multilingual capabilities and is a good candidate for fine-tuning for languages other than English.
Its different sizes also make it attractive for fine-tuning for different tasks.
I would gladly help but am not knowledgeable enough
Thank you

[krammnic]

Actually, in inspiration of one current kaggle competition - it is really good idea too add this pretty soon.

[ebsmothers]

Thanks @almugabo for creating the issue. I think this will be a bit of effort, quickly jotting down a couple of things I'm aware of that we'd need to support:

• Logit softcapping in attention layer
• Logit softcapping in output layer
• Sliding window attention
• Expose sliding window attention in configurable set of layers
• Post layernorm for both attention and FFN (hacky thing to do is just use attn_scale and mlp_scale)
• Model builders for 2B, 9B, 27B sizes

For logit softcapping and sliding window attention, I suspect we can use FlexAttention APIs. See this blog post where they give explicit examples of each.

[Optimox]

Hello, I have started the addition of gemma2, I will create now my PR in WIP mode. I haven't run any test yet but will do soon!

Edit: My PR is here: #1835

@ebsmothers it would be great if you could have a quick look to validate the choices I made in order to implement sliding windows, pre-post layer normalisation and softcapping... I would be happy to make things differently to keep the changes minimal (I tried as much as possible to keep all changes minimal).

[Optimox]

I didn't know about FlexAttention, I will look into it!

[joecummings]

ADDED!