include nncf in openvino extras

[eaidova]

What does this PR do?

install nncf with openvino extras. role nncf in model optimization during openvino model export and loading is increased in last releases (e.g. export model with int8 or int4 weights compression for large models), it gives more benefits to users with having it from scratch

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[helena-intel]

it gives more benefits to users with having it from scratch

Users already get NNCF by default, unless they choose to not install the nncf extra, even though that is what we recommend in documentation (or if they don't read documentation, but then they would also not know to install the openvino extra)

Two downsides of including NNCF in [openvino]:

• NNCF adds 33 new dependencies (of dependencies). All these dependencies can have bugs, or incompatibilities. NNCF is imported on importing OVModelFor.... so these issues then can also affect users who do not even need NNCF (this happens from time to time).
• Having nncf installed increases import time (again because when it is available, it is imported). On my laptop (10th gen Core) from optimum.intel import OVModelForSequenceClassification taskes 4.5-5.5 seconds without nncf installed, and 7.5-10 seconds with nncf installed. Not a problem when you're exporting a model once and need NNCF, but when you're writing code and have to run an inference script or notebook cell many times, this is noticeable.

[AlexKoff88]

it gives more benefits to users with having it from scratch

Users already get NNCF by default, unless they choose to not install the nncf extra, even though that is what we recommend in documentation (or if they don't read documentation, but then they would also not know to install the openvino extra)

Two downsides of including NNCF in [openvino]:

• NNCF adds 33 new dependencies (of dependencies). All these dependencies can have bugs, or incompatibilities. NNCF is imported on importing OVModelFor.... so these issues then can also affect users who do not even need NNCF (this happens from time to time).
• Having nncf installed increases import time (again because when it is available, it is imported). On my laptop (10th gen Core) from optimum.intel import OVModelForSequenceClassification taskes 4.5-5.5 seconds without nncf installed, and 7.5-10 seconds with nncf installed. Not a problem when you're exporting a model once and need NNCF, but when you're writing code and have to run an inference script or notebook cell many times, this is noticeable.

Thanks, @helena-intel for providing your opinion here.
Regarding the additional 33 dependencies you mentioned, I would like to comment that originally it is much less (see https://github.com/openvinotoolkit/nncf/blob/481ce9d9624c4b699878aa337d2e9b78dc25a517/setup.py#L104). We are also validating compatibility of NNCF with OpenVINO on regular basis since this is a part of the product. Moreover, we rely only on official releases of NNCF in Optimum. If you know the cases where import of NNCF causes problems please share with reproducible step,
As for the import time comment, I don't think it is a stopper but we will investigate it. cc'ed @alexsu52, @l-bat please take a look.

Again, our motivation is to provide a smooth UX when working with LLMs and optimizing them. We are also enabling SD optimization in the separate PR. It involves NNCF as well.

[helena-intel]

Thanks @AlexKoff88 I agree, my comments are not intended as stopper. There is still a method to not install NNCF, so we can use that when needed.