[Conformance] TorchFX/OV backends Alignment

[daniil-lyakhov]

Changes

• Constant folding is applied to all TorchFX models before the quantization
• Some torchvision models (swin_v2_s, vit_16_b) are exported by torch.export.export before ov conversation
• Moc transformations are applied to openvino compressed models after the compression

After the #2984

• Fixed _compress_qdq_constant_transformation for per tensor case

Reason for changes

• To align TorchFX/OV quantized models

Related tickets

#2766

Tests

post_training_quantization/504/ is finished successfully

[alexsu52]

Please rerun post_training_quantization build.

[daniil-lyakhov]

Please rerun post_training_quantization build.

/post_training_quantization/504/ is finished successfully

[AlexanderDokuchaev]

Please add unit tests for constant folding and docstrings.

[daniil-lyakhov]

Please add unit tests for constant folding and docstrings.

tested by test_models by swin_v2_s and vit_b_16 models. The same test approach is done for other transformation.
If you think we definitely need a unit test - let's create an issue for that

[AlexanderDokuchaev]

@alexsu52 i will not review PR that adding transformation that will applied for any model in nncf.quantize without unit tests and any documentation.

[daniil-lyakhov]

@AlexanderDokuchaev, please take a look

[anzr299]

In the case where the user gives a model with already inserted Quantize-Dequantize or Quantize-random_nodes-Dequantize, shouldn't it be ignored by constant folding?
cc @alexsu52

[daniil-lyakhov]

@anzr299, I fixed _compress_qdq_constant_transformation function in the per tensor case and tested it by test_compress_post_quantize_transformation, take a look

[anzr299]

I think we can use the same function for axis too.

Suggested change

	axis = nodes_map.get("axis")
	axis = get_const(nodes_map.get("axis"))

[daniil-lyakhov]

Nope, per tensor case has no "axis" key, [] raises key error. Intended axis value for per tensor case - None. .get returns None https://docs.python.org/3/library/stdtypes.html#dict.get

[anzr299]

Sorry about that, I have updated the suggestion. I meant to say that we can pass this also to the get_const funciton to keep it the same as others

[daniil-lyakhov]

Done

[anzr299]

@anzr299, I fixed _compress_qdq_constant_transformation function in the per tensor case and tested it by test_compress_post_quantize_transformation, take a look

Oh, yes that implementation did not consider per_tensor quantization. Thank you for pointing it out!