Adding Qwen2.5

[ysjprojects]

see #1709

Qwen2.5

0.5B, 1.5B, 3B, 7B, 14B, 32B, 72B

Qwen2.5 Coder

0.5B, 1.5B, 3B, 7B, 14B, 32B

Both base and instruct models.

Motivation:

• Proven SOTA coding performance, especially on Qwen2.5-Coder series.
• One of more recent open-source LM releases with decent performance on general benchmarks that is competitive with proprietary models.
• Notably strong performance on Chinese benchmarks.
• SOTA model that goes as small as 0.5B, which is very where and will serve many use cases in small LMs.

[Andrei-Aksionov]

Hello @ysjprojects 👋

Thanks for the PR.
I believe we wanted to implement Qwen models for quite a while, but never did.

Overall, it looks wonderful.
I just added a couple of nits.

[Andrei-Aksionov]

One more thing: please update the description of the PR with more info about the model.

[Andrei-Aksionov]

I did a quick check of 0.5B and 1.5B instruct version (with the fix for conversion script).
Don't know about other languages. but in English even 0.5B performs surprisingly well 🙂.

@ysjprojects After you apply the fix that I've mentioned in the comment, I'll be happy to merge the PR.

[ysjprojects]

One more thing: please update the description of the PR with more info about the model.

is there some specific details that should be included?

[ysjprojects]

I did a quick check of 0.5B and 1.5B instruct version (with the fix for conversion script). Don't know about other languages. but in English even 0.5B performs surprisingly well 🙂.

@ysjprojects After you apply the fix that I've mentioned in the comment, I'll be happy to merge the PR.

fixed

[Andrei-Aksionov]

One more thing: please update the description of the PR with more info about the model.

is there some specific details that should be included?

Would be nice if you added why did you decide to add this exact model.
Coding abilities, excellent support of Chinese, ...
Just a bit of context.