Extend MOE support to more GPUs

- Update the `m_use_onednn` in `execution_config.cpp` to trigger for the new `ov::op::internal::MOE` and `ov::intel_gpu::op::MoERouterFused` types
- Update `transformations_pipeline.cpp` to guard router transformations (`FuseMoERouter`, `FuseMoERouterScale`) alongside the existing MoE optimizations under the `ENABLE_ONEDNN_FOR_GPU` block for `xe_lp` architecture and above.

Signed-off-by: Peter Chen <peter.chen@intel.com>

Author: Chuquan Chen

src/plugins/intel_gpu/src/plugin/transformations_pipeline.cpp

@@ -567,12 +567,13 @@ void TransformationsPipeline::apply(std::shared_ptr<ov::Model> func) {
             return false;
         }();
 
+#ifdef ENABLE_ONEDNN_FOR_GPU
         const bool disable_moe_opt = GPU_DEBUG_VALUE_OR(config.get_disable_moe_opt(), false);
 
         // MOE: TiledMoeBlock -> GatherMatmuls(compressed) -> MoeOp(compressed) -> MoeOpWithRouting(compressed).
-        // Gated on supports_immad (systolic-only) and oneDNN (required for expert GEMM dispatch).
-        // Note: even though we are already inside `if (supports_immad)`, oneDNN can still be explicitly disabled by the user.
-        if (device_info.supports_immad && config.get_use_onednn()) {
+        // Gated on oneDNN supports platforms.
+        // Note:  device_info.arch >= cldnn::gpu_arch::xe_lp are `oneDNN supports platforms`,  register all MOE3GEMM related pass-es.
+        if (device_info.arch >= cldnn::gpu_arch::xe_lp) {
             manager.register_pass<ov::pass::ConvertTiledMoeBlockToGatherMatmuls>();
 
             // f32 listed because this pass runs before ConvertPrecision (line ~588);
@@ -595,6 +596,7 @@ void TransformationsPipeline::apply(std::shared_ptr<ov::Model> func) {
                 manager.register_pass<ov::intel_gpu::FuseMOESharedExpert>();
             }
         }
+#endif // ENABLE_ONEDNN_FOR_GPU
         manager.register_pass<ov::pass::GatedDeltaNetFusion>();
         manager.register_pass<ov::pass::InitNodeInfo>();
         manager.register_pass<EinsumDecomposition>();

src/plugins/intel_gpu/src/runtime/execution_config.cpp

@@ -7,6 +7,10 @@
 
 #include "intel_gpu/op/indirect_sdpa.hpp"
 #include "intel_gpu/op/kv_cache.hpp"
+#ifdef ENABLE_ONEDNN_FOR_GPU
+#include "ov_ops/moe_compressed.hpp"
+#include "intel_gpu/op/moe_router_fused.hpp"
+#endif // ENABLE_ONEDNN_FOR_GPU
 #include "intel_gpu/op/sdpa.hpp"
 #include "intel_gpu/plugin/remote_context.hpp"
 #include "intel_gpu/primitives/paged_attention.hpp"
@@ -224,6 +228,21 @@ void ExecutionConfig::apply_model_specific_options(const IRemoteContext* context
             m_use_onednn = true;
         }
 
+#ifdef ENABLE_ONEDNN_FOR_GPU
+        // MOE and MoERouterFused uses oneDNN internally for matrix multiplications
+        // (onednn_linear wrappers in moe_3gemm_swiglu_opt.cpp), which requires:
+        //   1. use_onednn=true so create_onednn_engine() is called during program build
+        //      (see program.cpp: lo.enable_onednn_for<lstm_seq/gru_seq> path which makes
+        //       onednn_impls_optimization_attribute non-empty, triggering engine init).
+        //   2. in-order OCL command queue (finalize_impl sets this when use_onednn=true).
+        // Auto-enable this only on architectures with oneDNN support, consistent with
+        // the LSTM/GRU path above, to avoid initializing oneDNN on unsupported devices.
+        if ((ov::is_type<ov::op::internal::MOE>(op) || ov::is_type<ov::intel_gpu::op::MoERouterFused>(op)) &&
+            info.arch >= cldnn::gpu_arch::xe_lp) {
+            m_use_onednn = true;
+        }
+#endif //ENABLE_ONEDNN_FOR_GPU
+
         if (auto multi_subgraph_op = ov::as_type_ptr<ov::op::util::MultiSubGraphOp>(op)) {
             for (const auto& sub_graph : multi_subgraph_op->get_functions()) {
                 for (auto& sub_op : sub_graph->get_ops()) {