[CPU]Add Support For GatedDeltaNet

[zhangYiIntel]

Details:

• Add Internal Op GatedDeltaNet
• Add GatedDeltaNet Fusion
• Add GatedDeltaNet CPU kernel

Tickets:

• CVS-180287
• CVS-180289
• CVS-180880

AI Assistance:

• AI assistance used: no
• If yes, summarize how AI was used and what human validation was performed (build/tests/manual checks).

[Copilot]

Pull request overview

This PR introduces CPU support for the GatedDeltaNet operation, a recurrent linear attention variant. It adds:

1. An internal OpenVINO GatedDeltaNet op (core) with a new ov::op::GatedDeltaNet class, shape inference, and Python bindings.
2. A pattern-matching fusion transformation (GatedDeltaNetFusion) that replaces a Loop-based subgraph with the fused op.
3. A CPU kernel (recurrent_linear_attn) with AVX512F/AVX2/scalar code paths, a CPU node wrapper, and functional tests.

Changes:

• New internal op GatedDeltaNet with validation, shape inference, config struct, and Python binding
• New GatedDeltaNetFusion transformation that fuses the Loop-based GDN subgraph into the internal op
• New CPU node and recurrent_linear_attn kernel (AVX512F + fallback), along with functional smoke tests

Reviewed changes

Copilot reviewed 24 out of 24 changed files in this pull request and generated 8 comments.

Show a summary per file

File	Description
src/core/dev_api/openvino/op/gated_delta_net.hpp	New internal op declaration with Config struct
src/core/src/op/gated_delta_net.cpp	Op implementation: validation, shape inference, clone_with_new_inputs
src/common/transformations/include/transformations/common_optimizations/fuse_gated_delta_net.hpp	Fusion pass declaration
src/common/transformations/src/transformations/common_optimizations/fuse_gated_delta_net.cpp	Fusion pass implementation (Loop → GatedDeltaNet)
src/plugins/intel_cpu/src/nodes/gated_delta_net.h	CPU node header
src/plugins/intel_cpu/src/nodes/gated_delta_net.cpp	CPU node execute logic
src/plugins/intel_cpu/src/nodes/kernels/linear_attn/recurrent_linear_attn.hpp	Kernel function declaration
src/plugins/intel_cpu/src/nodes/kernels/linear_attn/recurrent_linear_attn.cpp	AVX512F/AVX2/scalar kernel implementation
src/plugins/intel_cpu/src/transformations/transformation_pipeline.cpp	Registers fusion pass in PreLpt pipeline
src/plugins/intel_cpu/src/nodes_factory.cpp	Registers GatedDeltaNet CPU node (x64 only)
src/plugins/intel_cpu/src/graph_optimizer.cpp	Excludes GatedDeltaNet from tail precision optimization
src/plugins/intel_cpu/src/cpu_types.h	Adds GatedDeltaNet to Type enum
src/plugins/intel_cpu/src/cpu_types.cpp	Adds type name mapping for GatedDeltaNet
src/plugins/intel_cpu/src/extension.cpp	Registers op extension for serialization
src/plugins/intel_cpu/CMakeLists.txt	Adds cross-compiled build entry for kernel
src/bindings/python/src/pyopenvino/graph/ops/gated_delta_net.hpp	Python binding header
src/bindings/python/src/pyopenvino/graph/ops/gated_detla_net.cpp	Python binding implementation (typo in filename)
src/bindings/python/src/pyopenvino/pyopenvino.cpp	Registers Python binding class
src/bindings/python/src/openvino/op/__init__.pyi	Exports _GatedDeltaNet to Python stubs
src/bindings/python/src/openvino/_pyopenvino/op/__init__.pyi	Adds Python stub for _GatedDeltaNet
src/tests/functional/plugin/shared/include/shared_test_classes/subgraph/gated_delta_net.hpp	Test class declaration
src/tests/functional/plugin/shared/include/subgraph_tests/gated_delta_net.hpp	Test body with CompareWithRefs
src/tests/functional/plugin/shared/src/subgraph/gated_delta_net.cpp	Test setup (reference Loop model + fused model)
src/plugins/intel_cpu/tests/functional/shared_tests_instances/subgraph_tests/gated_delta_net.cpp	CPU-specific test instantiation

---

You can also share your feedback on Copilot code review. Take the survey.

[mryzhov]

When it is possible? sclae_node pattern is any_input().
Maybe it would be better to retrieve the scale node outside the helper function and check that it is not nullptr? For example, in that case we would not need to pass the default_scale_type matcher to the function.

[zhangYiIntel]

fixed

[mryzhov]

I would suggest to add the tests without Convert layers, with different Transposes on inputs and outputs, with different input and output shapes (when the inputs rank or output rank can be changed)

[zhangYiIntel]

The loop body are coupled with the transposes, the combination is not limited. Unlike SDPA case which has fixed order in QK matmul no matter q/k/v transposed or not, GDN's loop body is very diverse, I think in this release it's better to focus on current Qwen3Next pattern.

[Copilot]

Pull request overview

Copilot reviewed 25 out of 25 changed files in this pull request and generated 3 comments.

[Copilot]

The head-size mismatch validation compares k_head_size vs q_head_size, but the error message prints v_head_size as the second value (line 107). This will mislead users when debugging invalid models.

Please update the message to report q_head_size as the second operand.

Suggested change

	v_head_size,
	q_head_size,