Phi-1.5 support?

[spikespiegel]

It seems microsoft open-source phi-1.5 model (1.3B parameter model -> textbook is all you need) and put it in huggingface. (https://huggingface.co/microsoft/phi-1_5/tree/main), claiming comparable performance to 7B-13B models. Any chance to support this model?

[rayhern]

man oh man, im praying to the the llm gods now...

[chiefMarlin]

Any updates on this ?

[dmvaldman]

We can now skip straight to phi-2!

[kgoodridge]

Except Microsoft hasn't released the weights for phi-2, so that one would be extra challenging.

[kgoodridge]

Guess I'm wrong(ish)

https://twitter.com/SebastienBubeck/status/1734801722831413312?s=20

[liuzl]

https://huggingface.co/microsoft/phi-2

[aaryadav]

Quantized weights can be found here:
https://huggingface.co/radames/phi-2-quantized

Can be run using Candle (rust ML framework)
https://github.com/huggingface/candle/tree/main/candle-examples/examples/phi

Would be great if it was supported here as well.

[KnutJaegersberg]

phi-2 would be a great addition. properly the best model of this size.

[niutech]

The progress can be tracked in the issues #3146 and #4437.