Introduce OVPipelineQuantizationConfig

[nikita-savelyevv]

What does this PR do?

Changes:

• Introduced OVPipelineQuantizationConfig allowing to specify quantization parameters per model component. Add corresponding tests.
• Introduced a more advanced logic for inferring quantization config type from dictionary. Moved this logic to a separate function: _quantization_config_from_dict().
• Updated default int4 config for phi4-multimodal. WWB similarity: 85.30%.

For example the code below applies int8 PTQ to lm_model, int8 WC to text_embeddings_model and no optimization to vision_embeddings_model.

from optimum.intel import OVModelForVisualCausalLM
from optimum.intel import OVPipelineQuantizationConfig, OVQuantizationConfig, OVWeightQuantizationConfig

model_id = "OpenGVLab/InternVL2-1B"
model = OVModelForVisualCausalLM.from_pretrained(
    model_id,
    export=True,
    trust_remote_code=True,
    quantization_config=OVPipelineQuantizationConfig(
        quantization_configs={
            "lm_model": OVQuantizationConfig(bits=8),
            "text_embeddings_model": OVWeightQuantizationConfig(bits=8),
        },
        dataset="contextual",
        trust_remote_code=True,
    )
)

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.

[nikita-savelyevv]

@l-bat Could you please review this PR?

[IlyasMoutawwakil]

why is this value hidden here ?

[nikita-savelyevv]

128 was previously a default value for num_samples argument in OVQuantizationConfig. In this PR I've removed it so it is now None by default. That's why 128 has appeared here.

Ideally, we should transition to providing arguments via quantization_config.to_nncf_dict() here, as it is done for OV case. But I propose to do this in a separate PR.

[IlyasMoutawwakil]

LGTM !

[nikita-savelyevv]

Hi @echarlaix! Do you mind us merging this or you would like to take a look?

[echarlaix]

apologies for the delay @nikita-savelyevv, taking a look right now!

[echarlaix]

Looks great thanks @nikita-savelyevv !!