Bump transformers version

[nikita-malininn]

Changes

• Bumped transformers version to >=4.48.0

• Bumped optimum-intel version to >=1.22.0 (as it requires transformers)

• Bumped optimum version to >=1.24.0 (as it requires transformers)

• Updated statistics caching test with position_ids to fix:

  Exception from src/plugins/intel_cpu/src/node.cpp:769:
  [CPU] Concat node with name '__module.model.model.decoder.layers.0.self_attn/aten::cat/Concat' Check 'TRShape::merge_into(output_shape, in_copy)' failed at src/core/shape_inference/include/concat_shape_inference.hpp:43:
  While validating node 'opset1::Concat __module.model.model.decoder.layers.0.self_attn/aten::cat/Concat (opset1::Parameter Parameter_753753[0]:f32[?,4,?,4], opset1::Transpose __module.model.model.decoder.layers.0.self_attn/aten::transpose/Transpose_1[0]:f32[?,4,?,4]) -> (f32[?,4,?,4])' with friendly_name '__module.model.model.decoder.layers.0.self_attn/aten::cat/Concat':
  Shape inference input shapes {{1,4,0,4},{0,4,0,4}}
  Argument shapes are inconsistent; they must have the same rank, and must have equal dimension everywhere except on the concatenation axis (axis 2).

• Bumped accelerate version to >=1.1.0 to fix:

  NotImplementedError: data_seed requires Accelerate version accelerate >= 1.1.0. This is not supported and we recommend you to update your version.

• Added num_items_in_batch to compute_loss to fix:

  TypeError: CompressionTrainer.compute_loss() got an unexpected keyword argument 'num_items_in_batch'

• Bumped _ONNX_DEFAULT_OPSET from 13 to 14 to fix:

  torch.onnx.errors.UnsupportedOperatorError: Exporting the operator 'aten::scaled_dot_product_attention' to ONNX opset version 13 is not supported. Support for this operator was added in version 14, try exporting with this version.

• Updated sparsity/pruning/nas references for tests.

Reason for changes

• Security issues with transformers<4.48

Related tickets

• N/A

Tests

• install - https://github.com/openvinotoolkit/nncf/actions/runs/13393410660 - passed
• weight compression - https://github.com/openvinotoolkit/nncf/actions/runs/13393412141 - passed
• examples - https://github.com/openvinotoolkit/nncf/actions/runs/13393412919 - passed
• PTQ - manual/job/post_training_quantization/622/ - passed
• PTWC - manual/job/post_training_weight_compression/327/ - passed

[ljaljushkin]

I took a look at the references, and it's looks ok.
Most of the changes coming from changing attention implementation using SDPA.
Search for pruning group doesn't support it and finds much lower number of groups.

@nikita-malininn applied WA for the failed tests, since the issue in the deprecated components:
nikita-malininn#7

[ljaljushkin]

@AlexanderDokuchaev do we have some tests for torch export outside pre-commit scope?

[ljaljushkin]

@nikita-malininn I found PTQ conformance tests with export:
ImageClassificationTimm and ImageClassificationTorchvision test cases in https://github.com/openvinotoolkit/nncf/blob/develop/tests/post_training/model_scope.py
IMO, It's worth launching it