fix fused bias grad add nan (#53129)

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,