diff --git a/vllm/model_executor/layers/fused_moe/layer.py b/vllm/model_executor/layers/fused_moe/layer.py
index ef33852e3162..143123e577b1 100644
--- a/vllm/model_executor/layers/fused_moe/layer.py
+++ b/vllm/model_executor/layers/fused_moe/layer.py
@@ -309,7 +309,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
                                 expert_map=expert_map,
                                 renormalize=renormalize)
 
-    forward_native = forward_cuda
+    forward_native = forward_tpu if current_platform.is_tpu else forward_cuda
 
 
 def determine_expert_map(