Add LLaMA 2 example for DirectML

[PatriceVignola]

This adds the LLaMA 2 optimizations for DirectML with examples, and a sample ChatApp that was inspired from this but stripped down to an MVP and customized for DirectML.

It also adds a "CompositePyTorchModel" which follows the same principle as the Composite Optimum model but for raw pytorch models instead.

[guotuofeng]

what's the difference between this one and OptimumModel? could we leverage OptimumModel or unify them? It seems the two class is similiar.

[trajepl]

For reference:
in the coming release of optimum, the optimum exporter can export only the merged models.
https://github.com/huggingface/optimum/pull/1257/files

Also, with nightly ort/optimum, Olive optimize llama2 as following:
https://github.com/microsoft/Olive/tree/main/examples/llama2

[PatriceVignola]

what's the difference between this one and OptimumModel? could we leverage OptimumModel or unify them? It seems the two class is similiar.

I decoupled OptimumModel and CompositePyTorch model. They look similar at first glance, but they don't have much in common aside from the model_components part. OptimumModel is a very lightweight class that only needs to return strings of the actual model names which get handled by optimum in its own conversion pass, whereas CompositePyTorch model is a container of multiple other PyTorchModels that get handled recursively in the main OnnxConversion pass.

[jambayk]

what's the difference between this one and OptimumModel? could we leverage OptimumModel or unify them? It seems the two class is similiar.

I decoupled OptimumModel and CompositePyTorch model. They look similar at first glance, but they don't have much in common aside from the model_components part. OptimumModel is a very lightweight class that only needs to return strings of the actual model names which get handled by optimum in its own conversion pass, whereas CompositePyTorch model is a container of multiple other PyTorchModels that get handled recursively in the main OnnxConversion pass.

That's what I thought too when looking at the code.
On another note: I am thinking of removing the OptimumModel class entirely later. It is only used for OptimumConversion but I don't see a reason why we cannot just use PyTorchModel + hf_config and OptimumConversion + model_components as pass config. This pass will be able return an ONNXModel or CompositeONNXModel base on the number of components.
There are other olive passes that also only support pytorch models with hf_config.