diff --git a/ops/framework/ms_kernels_internal/pyboost/internal_pyboost_runner.cc b/ops/framework/ms_kernels_internal/pyboost/internal_pyboost_runner.cc
index f30de74da83a7e81ee2c9839eb4345bfdf6bac6d..fff0ee5941eee611b54e50dedad60977dd08fe70 100644
--- a/ops/framework/ms_kernels_internal/pyboost/internal_pyboost_runner.cc
+++ b/ops/framework/ms_kernels_internal/pyboost/internal_pyboost_runner.cc
@@ -26,27 +26,47 @@ size_t GetAlignedSize(size_t size) {
 
 namespace ms_custom_ops {
 void InternalPyboostRunner::GetOrCreateKernel(const TensorList &inputs, const TensorList &outputs) {
+  std::lock_guard<SimpleSpinLock> op_guard(internal_op_lock_);
+
   auto key = GetOrGenerateOpKey(op_key_);
-  auto it = hash_map_.find(key);
-  if (it != hash_map_.end()) {
-    internal_op_ = it->second;
-    MS_LOG(DEBUG) << "Internal Op [" << this->op_name() << "] hit cache";
-  } else {
+  internal_v2::InternalOpPtr cached_op = nullptr;
+
+  {
+    std::lock_guard<SimpleSpinLock> guard(hash_map_lock_);
+    auto it = hash_map_.find(key);
+    if (it != hash_map_.end()) {
+      cached_op = it->second;
+      MS_LOG(DEBUG) << "Internal Op [" << this->op_name() << "] hit cache";
+    }
+  }
+
+  if (cached_op == nullptr) {
     MS_LOG(DEBUG) << "Internal Op [" << this->op_name() << "] miss cache";
     TransDataType(inputs, outputs);
     UpdateArgImmutableInfo(&inputs_ii_, inputs, true);
     UpdateArgImmutableInfo(&outputs_ii_, outputs);
-    internal_op_ = CreateKernel(inputs_ii_, outputs_ii_);
-    MS_EXCEPTION_IF_NULL(internal_op_);
-    auto status = internal_op_->Init();
+
+    auto new_internal_op = CreateKernel(inputs_ii_, outputs_ii_);
+    MS_EXCEPTION_IF_NULL(new_internal_op);
+    auto status = new_internal_op->Init();
     if (status != mindspore::internal_v2::kInternalOk) {
-      internal_op_ = nullptr;
       MS_LOG(EXCEPTION) << "Init internal kernel failed, kernel_name: " << this->op_name();
-      return;
     }
-    hash_map_[key] = internal_op_;
+
+    {
+      std::lock_guard<SimpleSpinLock> guard(hash_map_lock_);
+      auto it = hash_map_.find(key);
+      if (it == hash_map_.end()) {
+        hash_map_[key] = new_internal_op;
+        cached_op = new_internal_op;
+      } else {
+        cached_op = it->second;
+      }
+    }
   }
 
+  internal_op_ = cached_op;
+
   internal_inputs_shape_.clear();
   internal_outputs_shape_.clear();
   internal_inputs_shape_.resize(inputs.size());
@@ -66,6 +86,7 @@ void InternalPyboostRunner::GetOrCreateKernel(const TensorList &inputs, const Te
 }
 
 size_t InternalPyboostRunner::CalcWorkspace() {
+  std::lock_guard<SimpleSpinLock> op_guard(internal_op_lock_);
   MS_EXCEPTION_IF_NULL(internal_op_);
   auto workspace_size_list = internal_op_->GetWorkspaceSize();
   // all workspace will be aligned, like graph_mode
@@ -190,6 +211,7 @@ void InternalPyboostRunner::UpdateArgImmutableInfo(std::vector<internal_v2::ArgI
 
 void InternalPyboostRunner::GetWorkspace(const internal_v2::InternalOpPtr &internal_op,
                                          internal_v2::WsAddrList *internal_wss_addr) {
+  std::lock_guard<SimpleSpinLock> op_guard(internal_op_lock_);
   auto workspace_ptr = this->workspace_ptr();
   if (workspace_ptr == nullptr) {
     return;
@@ -207,6 +229,7 @@ void InternalPyboostRunner::GetWorkspace(const internal_v2::InternalOpPtr &inter
 }
 
 void InternalPyboostRunner::LaunchKernel() {
+  std::lock_guard<SimpleSpinLock> op_guard(internal_op_lock_);
   MS_EXCEPTION_IF_NULL(tiling_cache_item_);
   MS_EXCEPTION_IF_NULL(internal_op_);
   internal_v2::InputsAddrList inputs_addr;
diff --git a/ops/framework/ms_kernels_internal/pyboost/internal_pyboost_runner.h b/ops/framework/ms_kernels_internal/pyboost/internal_pyboost_runner.h
index 0ed36aa843c6f34e970e822f86a23b2e61eb614b..1d8fab185eb66467a0a2bcc09912a0e347cad308 100644
--- a/ops/framework/ms_kernels_internal/pyboost/internal_pyboost_runner.h
+++ b/ops/framework/ms_kernels_internal/pyboost/internal_pyboost_runner.h
@@ -85,6 +85,8 @@ class InternalPyboostRunner : public ms::pynative::PyboostRunner {
   uint64_t tiling_key_{0};
   internal_v2::InternalOpPtr internal_op_{nullptr};
   inline static std::unordered_map<uint64_t, internal_v2::InternalOpPtr> hash_map_;
+  inline static SimpleSpinLock hash_map_lock_;
+  inline static SimpleSpinLock internal_op_lock_;
   internal_v2::DtypeInfoList internal_inputs_dtype_;
   internal_v2::DtypeInfoList internal_outputs_dtype_;
   internal_v2::ShapeInfoList internal_inputs_shape_;