diff --git a/paddle/phi/kernels/fusion/gpu/fused_linear_param_grad_add_kernel.cu b/paddle/phi/kernels/fusion/gpu/fused_linear_param_grad_add_kernel.cu
index f39cc8acdc31020b06bb4c56f92924ce59ecb57e..0840b6927a4bdd5007fd502bbd05acfcbcd1d585 100644
--- a/paddle/phi/kernels/fusion/gpu/fused_linear_param_grad_add_kernel.cu
+++ b/paddle/phi/kernels/fusion/gpu/fused_linear_param_grad_add_kernel.cu
@@ -44,7 +44,7 @@ void FusedLinearParamGradAddImpl(const Context &ctx,
                                  DenseTensor *dbias_out) {
   constexpr bool kIsMultiPrecision = !std::is_same<T, MT>::value;
 
-  const bool fuse_bias_grad = kIsMultiPrecision && dweight_out;
+  const bool fuse_bias_grad = false;  // kIsMultiPrecision && dweight_out;
   if (dweight_out) {
     phi::funcs::ComputeFusedGemmEpilogueBackward<T, T, MT>(
         ctx,