Added Phi and Phi-2 support

[vigarov]

This PR aims to add support for Phi and Phi-2 models (both based on the PhiModel architecture).
I see in, #32

Phi 1.5 support has been attempted, but they have a very unusual model definition. Until it's been standardized, I am not sure I will support it.

Any particular reason for this? Phi's architecture is indeed unlike llama&co, but nothing too out of the ordinary imo

[casper-hansen]

Hi @vigarov, it's honestly been to long for me to remember what the issue was back then when I commented on it. However, it's good to see you come around to help implement the other Phi models.

Can you please run pip install -e .[eval] and benchmark the perplexity before and after using the script at examples/eval.py?

[vigarov]

Didn't realize the script was there, sorry would've done it in the first place!
Here are some results with phi-2:

	Ppl (wikitext)	Size
(Phi-2 base)	9.705	~5.6 GB
Quantization group size
G2	3761.297	4.6 GB
G8	3255.700	2.4 GB
G128	2893.224	1.8 GB
G512	5451.693	1.7 GB

Perplexity looks pretty bad :/ @casper-hansen any tips/pointers on what could usually be a symptom of this ?
I used GEMM, and zero_point=True for the rest of the config.

[casper-hansen]

@vigarov this is probably what I meant by being hard to support. To effectively quantize a model with AWQ means to observe how the inputs change as they pass through the layers of the models. Sometimes it can be hard when models have very unique model definitions.

[vigarov]

Debugged the quantization and everything seems to be running normally, following the paper/merlin's zero_point optimization.
For reference, BnB on phi-2 leads a PPL of 11.252 w/ 8-bit quantization and 11.692 w/ 4-bit

Marking this as draft until I can find where the issue comes from.