diff --git a/tf_adapter/kernels/aicpu/dataset_function.cc b/tf_adapter/kernels/aicpu/dataset_function.cc
index 5a01b4aea27d41e254c2a77e27b4366cae35d3e6..fb5edd01b6d90d565e8704825f552db662b680f4 100644
--- a/tf_adapter/kernels/aicpu/dataset_function.cc
+++ b/tf_adapter/kernels/aicpu/dataset_function.cc
@@ -368,7 +368,7 @@ void DatasetFunction::DestroyAclOutputDataset(aclmdlDataset *output, bool isFree
     aclDataBuffer* data_buffer = aclmdlGetDatasetBuffer(output, i);
     if (isFree) {
       void* data_addr = aclGetDataBufferAddr(data_buffer);
-      aclError ret = aclrtFree(data_addr);
+      ret = aclrtFree(data_addr);
       if (ret != ACL_ERROR_NONE) {
         ADP_LOG(ERROR) << "Free acl device memory failed.";
       }
diff --git a/tf_adapter/kernels/aicpu/map_and_batch_dataset_op.cc b/tf_adapter/kernels/aicpu/map_and_batch_dataset_op.cc
index 5387eef039776ab798030c6648b83c1fbb80a347..d42c7ed08b2fdb2850fdcf71fa34ce61c18efb95 100644
--- a/tf_adapter/kernels/aicpu/map_and_batch_dataset_op.cc
+++ b/tf_adapter/kernels/aicpu/map_and_batch_dataset_op.cc
@@ -396,6 +396,7 @@ class NpuMapAndBatchDatasetOp::Dataset : public DatasetBase {
           if (output_cpu_addr != nullptr) {
             delete[] output_cpu_addr;
             output_cpu_addr = nullptr;
+            output_cpu = nullptr;
           }
           ADP_LOG(EVENT) << "~BatchResultBase finish.";
         }
@@ -571,7 +572,8 @@ class NpuMapAndBatchDatasetOp::Dataset : public DatasetBase {
         batch_result.output_cpu = batch_result.output_cpu_addr;
 
         // reset start address for cpu memory when pass data to tensorflow
-        uint64_t offset = reinterpret_cast<uintptr_t>(batch_result.output_cpu_addr) % kTFTensorAlignment;
+        uint64_t offset = reinterpret_cast<uint64_t>(reinterpret_cast<uintptr_t>(batch_result.output_cpu_addr))
+                          % kTFTensorAlignment;
         if (offset != 0UL) {
           offset = kTFTensorAlignment - offset;
           batch_result.output_cpu = batch_result.output_cpu_addr + offset;
@@ -1130,6 +1132,10 @@ class NpuMapAndBatchDatasetOp::Dataset : public DatasetBase {
     explicit IteratorDyn(const Params& params)
         : IteratorMeBase(params) {
       max_batch_results_ = CeilDiv(GetParallelCallsNum(), params.dataset->batch_size_);
+      if(max_batch_results_ == 1ULL) {
+        // one result for consumer, one results for producer
+        max_batch_results_ += 1ULL;
+      }
       ADP_LOG(EVENT) << "IteratorDyn.";
     };
 
diff --git a/tf_adapter/kernels/aicpu/map_dataset_op.cc b/tf_adapter/kernels/aicpu/map_dataset_op.cc
index 9aa95e71515345cbd2c612b6227c84f0c6d3ff29..16af4ce8c6a2dda2794359e9f47784881ae63e8e 100644
--- a/tf_adapter/kernels/aicpu/map_dataset_op.cc
+++ b/tf_adapter/kernels/aicpu/map_dataset_op.cc
@@ -346,6 +346,7 @@ private:
         if (output_cpu_addr != nullptr) {
           delete[] output_cpu_addr;
           output_cpu_addr = nullptr;
+          output_cpu = nullptr;
         }
         ADP_LOG(EVENT) << "~OutputResultBase finish.";
       }
@@ -430,7 +431,8 @@ private:
       output_result.output_cpu = output_result.output_cpu_addr;
 
       // reset start address for cpu memory when pass data to tensorflow
-      uint64_t offset = reinterpret_cast<uintptr_t>(output_result.output_cpu_addr) % kTFTensorAlignment;
+      uint64_t offset = reinterpret_cast<uint64_t>(reinterpret_cast<uintptr_t>(output_result.output_cpu_addr))
+                        % kTFTensorAlignment;
       if (offset != 0UL) {
         offset = kTFTensorAlignment - offset;
         output_result.output_cpu = output_result.output_cpu_addr + offset;