diff --git a/custom_ops/gpu_ops/append_attn/get_block_shape_and_split_kv_block.cu b/custom_ops/gpu_ops/append_attn/get_block_shape_and_split_kv_block.cu
index 2e2e8c7ba..d00e63875 100644
--- a/custom_ops/gpu_ops/append_attn/get_block_shape_and_split_kv_block.cu
+++ b/custom_ops/gpu_ops/append_attn/get_block_shape_and_split_kv_block.cu
@@ -331,7 +331,7 @@ void GetBlockShapeAndSplitKVBlock(
 
   // decoder
   if (max_dec_len_this_time > 0) {
-    const bool mla_use_tensorcore = GetMlaUseTensorcore();
+    const bool mla_use_tensorcore = true; //GetMlaUseTensorcore();
     if (mla_use_tensorcore && group_size <= 64) {
       const int set_chunk_size = get_mla_dec_chunk_size(bsz);
 
diff --git a/custom_ops/gpu_ops/multi_head_latent_attention.cu b/custom_ops/gpu_ops/multi_head_latent_attention.cu
index 126b014b8..6e804f3eb 100644
--- a/custom_ops/gpu_ops/multi_head_latent_attention.cu
+++ b/custom_ops/gpu_ops/multi_head_latent_attention.cu
@@ -66,7 +66,7 @@ std::vector<paddle::Tensor> MultiHeadLatentAttentionKernel(
   // int chunk_size = decoder_chunk_size_cpu.data<int>()[0];
   //
 
-  const bool mla_use_tensorcore = get_mla_use_tensorcore();
+  const bool mla_use_tensorcore = true; //get_mla_use_tensorcore();
   auto sm_version = GetSMVersion();
   if ((speculate_decoder || mla_use_tensorcore) && sm_version < 90) {
     PD_THROW("Please use speculate_decoder=0 and FLAGS_mla_use_tensorcore=0 when sm < 90.");