Fix onnx Attention and torch SDPA quantization handling

[ruro]

Changes

• added ONNXAttentionMetatype for the opset 23 Attention ONNX node
• fixed scaled_dot_product_attention quantization in torch2 for the case when Q, K and V are parallel edges coming from the same input node

Reason for changes

See #3750

Related tickets

Fixes #3750

Tests

• tests/onnx/quantization/test_graphs.py::test_synthetic_models_graph[AttentionModel]

• tests/torch2/function_hook/quantization/test_quantized_graphs.py::test_quantized_graphs[unbind_scaled_dot_product_attention_model]

[ruro]

Hm. iirc onnx added support for opset 23 in version 1.18.0. So the new test is currently failing in CI due to

onnx==1.17.0; python_version < '3.13'
onnx==1.18.0; python_version >= '3.13'

Do you have any preferences if I should mark this test as

@pytest.mark.skipif(
    version.parse(onnx.__version__) < version.parse("1.18.0"),
    reason="Opset 23 was added in onnx 1.18.0",
)

or bump the version or something else?

[andrey-churkin]

Hm. iirc onnx added support for opset 23 in version 1.18.0. So the new test is currently failing in CI due to

onnx==1.17.0; python_version < '3.13'
onnx==1.18.0; python_version >= '3.13'

Do you have any preferences if I should mark this test as

@pytest.mark.skipif(
    version.parse(onnx.__version__) < version.parse("1.18.0"),
    reason="Opset 23 was added in onnx 1.18.0",
)

or bump the version or something else?

Hi @ruro, thanks for your contribution. We currently support multiple versions of ONNX, and the Attention operator was added in opset 23, which corresponds to ONNX 1.18.0. I believe we should run this test only for ONNX versions >= 1.18.0.

[andrey-churkin]

@ruro Could you please briefly explain why these changes are necessary?

@daniil-lyakhov Please take a look

[ruro]

I've outlined my reasoning in the last two comments in #3750. The short version is that parallel edges aren't directly represented in the PTNNCFGraph (because it's not a Multi graph and doesn't allow repeated edges), but are instead stored in the parallel_input_port_ids property.

In this case, unbind has 3 outputs that are passed as q, k and v inputs of the sdpa node. Each of these 3 edges should be considered separately for the purposes of quantizer insertion/propagation, but the previous logic only added insertion points for "real" edges, ignoring any extra parallel edges.

Let me know if anything is unclear.

[daniil-lyakhov]

Great contribution! Could you please share a netron/ nncf graph visualization of the brand new supported subgraph? (nncf graph visualization api: https://github.com/openvinotoolkit/nncf/blob/develop/src/nncf/common/graph/graph.py#L611-L613)

This part of the code is the core logic of the NNCF, we need to figure out all possible side effects of this change

[ruro]

I am not sure, what you mean by "netron/nncf graph". The second image in the PR body is the expected graph for unbind+sdpa after applying quantization. Does that work?

[ruro]

Also, here are the before and after graphs, obtained by performing a torch.onnx.export (with opset_version=23) of a quantized timm.layers.attention.Attention module:

(The edge without the q/dq nodes is the V input of Attention as expected)