diff --git a/paddle/phi/kernels/gpu/embedding_grad_kernel.cu b/paddle/phi/kernels/gpu/embedding_grad_kernel.cu
index 0cfe2e43d1875ff2992cace2f9f49d6dfe62ba77..8bb00f075929c04710b5c5b24ab1d15388d38bb7 100644
--- a/paddle/phi/kernels/gpu/embedding_grad_kernel.cu
+++ b/paddle/phi/kernels/gpu/embedding_grad_kernel.cu
@@ -107,6 +107,7 @@ struct EmbeddingGradCUDAFunctor {
       if (FLAGS_cudnn_deterministic) {
         VLOG(2) << "Run grad kernel of embedding with single thread.";
         grids.x = 1;
+        threads.y = 1;
       }
       EmbeddingGrad<T, IdT><<<grids, threads, 0, dev_ctx_.stream()>>>(
           d_table, d_output, ids, N, K, D);