The MVN fusion test suite in src/plugins/intel_gpu/tests/unit/fusions/mvn_fusion_test.cpp covers a single eltwise post-op on MVN and the combination of eltwise followed by quantize. Two cases are not tested.
First, the chained pattern of eltwise_prod(scale) followed by eltwise_sum(shift) with no quantize step. Second, the same pattern where scale and shift are runtime inputs rather and constant weights.
This pattern appears in AdaLayerNorm, which is used in DiT flow matching models. In OpenVINO IR it decomposes to exactly this MVN plus two eltwise subgraph.
The MVN fusion test suite in src/plugins/intel_gpu/tests/unit/fusions/mvn_fusion_test.cpp covers a single eltwise post-op on MVN and the combination of eltwise followed by quantize. Two cases are not tested.
First, the chained pattern of eltwise_prod(scale) followed by eltwise_sum(shift) with no quantize step. Second, the same pattern where scale and shift are runtime inputs rather and constant weights.
This pattern appears in AdaLayerNorm, which is used in DiT flow matching models. In OpenVINO IR it decomposes to exactly this MVN plus two eltwise subgraph.