diff --git a/vllm/model_executor/layers/fused_moe/layer.py b/vllm/model_executor/layers/fused_moe/layer.py
old mode 100755
new mode 100644
index 3935fe374..c3be1be85
--- a/vllm/model_executor/layers/fused_moe/layer.py
+++ b/vllm/model_executor/layers/fused_moe/layer.py
@@ -335,6 +335,7 @@ class FusedMoE(CustomOp):
         expert_mapping: list[tuple[str, str, int, str]] | None = None,
         n_shared_experts: int | None = None,
         router_logits_dtype: torch.dtype | None = None,
+        has_shared_experts: bool = False,
     ):
         super().__init__()
 
@@ -564,7 +565,7 @@ class FusedMoE(CustomOp):
             device=vllm_config.device_config.device,
             routing_method=self.routing_method_type,
             # TODO: in_dtype == out_dtype?
-            disable_inplace=disable_inplace() or self.shared_experts is not None,
+            disable_inplace=disable_inplace() or has_shared_experts,
         )
         if self.use_mori_kernels:
             assert self.rocm_aiter_fmoe_enabled, (
diff --git a/vllm/model_executor/layers/fused_moe/shared_fused_moe.py b/vllm/model_executor/layers/fused_moe/shared_fused_moe.py
index cb601af70..937d13d34 100644
--- a/vllm/model_executor/layers/fused_moe/shared_fused_moe.py
+++ b/vllm/model_executor/layers/fused_moe/shared_fused_moe.py
@@ -26,6 +26,10 @@ class SharedFusedMoE(FusedMoE):
         routed_input_transform: torch.nn.Module | None = None,
         **kwargs,
     ):
+        # Pass has_shared_experts so FusedMoE.__init__ can set disable_inplace
+        # without accessing self.shared_experts (submodules cannot be set before
+        # Module.__init__()).
+        kwargs["has_shared_experts"] = shared_experts is not None
         super().__init__(**kwargs)
         self._shared_experts = shared_experts
         self._routed_input_transform = routed_input_transform