diff --git a/paddle/fluid/memory/allocation/stream_safe_cuda_allocator.cc b/paddle/fluid/memory/allocation/stream_safe_cuda_allocator.cc
index 1967dd8502808835346fe5fd6fe84ac0404b1951..903f50d9a4cb7c69f09c2078dc3acfff83a24711 100644
--- a/paddle/fluid/memory/allocation/stream_safe_cuda_allocator.cc
+++ b/paddle/fluid/memory/allocation/stream_safe_cuda_allocator.cc
@@ -266,7 +266,7 @@ uint64_t StreamSafeCUDAAllocator::ProcessUnfreedAllocationsAndRelease() {
   return underlying_allocator_->Release(place_);
 }
 
-std::once_flag StreamSafeCUDAAllocation::once_flag_;
+thread_local std::once_flag StreamSafeCUDAAllocation::once_flag_;
 
 std::map<platform::Place, std::vector<StreamSafeCUDAAllocator*>>
     StreamSafeCUDAAllocator::allocator_map_;
diff --git a/paddle/fluid/memory/allocation/stream_safe_cuda_allocator.h b/paddle/fluid/memory/allocation/stream_safe_cuda_allocator.h
index 5f9b6208100d3904f6df491114cff9d222ecd412..08ecdd496973032c023699244ed407da92c8bac1 100644
--- a/paddle/fluid/memory/allocation/stream_safe_cuda_allocator.h
+++ b/paddle/fluid/memory/allocation/stream_safe_cuda_allocator.h
@@ -45,7 +45,7 @@ class StreamSafeCUDAAllocation : public Allocation {
   gpuStream_t GetOwningStream() const;
 
  private:
-  static std::once_flag once_flag_;
+  thread_local static std::once_flag once_flag_;
   void RecordGraphCapturingStreams();
   void RecordStreamWithNoGraphCapturing(gpuStream_t stream);
   DecoratedAllocationPtr underlying_allocation_;