diff --git a/cpp/tensorrt_llm/kernels/moeLoadBalance/moeLoadBalanceKernels.cu b/cpp/tensorrt_llm/kernels/moeLoadBalance/moeLoadBalanceKernels.cu
index ace4f135b63..1f74f0698dd 100644
--- a/cpp/tensorrt_llm/kernels/moeLoadBalance/moeLoadBalanceKernels.cu
+++ b/cpp/tensorrt_llm/kernels/moeLoadBalance/moeLoadBalanceKernels.cu
@@ -131,10 +131,6 @@ void moeSetSignalForCpuStageForTest(MoeLoadBalanceSingleLayerSignal* signal)
 template <typename TYPE>
 __global__ void zeroExpertTokenCountKernel(MoeLoadBalanceMetaInfo metaInfo, int* const enabled, int* expertTokenCount)
 {
-    if (*enabled == 0)
-    {
-        return;
-    }
     TYPE oldExpertTokenCount = {0};
     int* expertTokenCountPtr = expertTokenCount + metaInfo.expertCount * blockIdx.x;
     TYPE* typedExpertTokenCountPtr = reinterpret_cast<TYPE*>(expertTokenCountPtr);