[ONNX FE] optimize GQA decomposition path

[Kotomi-Du]

Details:

SDPA in OV supports passing QKV information without explicit attention mask. Currently GQA decomposition manually generates causal mask which introduce unnecessary overhead (~7ms in Phi4 GPU). This PR is to replace the path with SDPA causal flag.

Besides, a conformance issue in GPU plugin is identified after enabling the path. Related fix is also applied in the PR.

Tickets:

• CVS-187874

AI Assistance:

• AI assistance used: yes
• AI helped review the code

[Copilot]

Pull request overview

This PR optimizes the ONNX Frontend (GQA) decomposition by relying on ScaledDotProductAttention’s built-in causal flag instead of materializing a causal attention mask, and includes a GPU kernel fix needed for correctness after enabling that path.

Changes:

• Replace manual causal mask construction in GroupQueryAttentionDecomposition with SDPA causal=true usage when no external attention mask is provided.
• Adjust GPU SDPA micro-kernel causal offset calculation to fix a conformance issue in a non-paged-attention path.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
src/common/transformations/src/transformations/op_conversions/group_query_attention_decomposition.cpp	Switches no-mask GQA decomposition to SDPA causal flag path (removes explicit causal mask construction).
src/plugins/intel_gpu/src/graph/impls/ocl_v2/sdpa_micro.cl	Fixes causal column offset computation to align masking behavior for non-paged attention.

[Copilot]

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

[mitruska]

The changes LGTM, but tests are missing.

[e-ddykim]

Could you explain when this change is necessary? Please add tests for the issue case.

[Kotomi-Du]

because SDPA non-PA path with causal as true will be failed without this change. Will update the PR by adding the tests.

[CuriousPanCake]

Do we need tests for these changes or are they already tested?

[Kotomi-Du]

Do we need tests for these changes or are they already tested?

I am preparing the tests, specifically for GPU. The onnx tests already include cases without mask. I also found even though SDPA as a general API to support is_casual as true, but it seems CPU inference failed in this case. I need to confirm if CPU plugin really supports.

Set the PR as Draft for now till I finish the tasks above.

[Kotomi-Du]

build_jenkins

[Kotomi-Du]

Open for CI-test, not ready for CPU plugin enabling yet.