TensorFlow C++ API中Session Run接口是否支持可重入？

Great question—this is a super important detail when building multi-threaded TensorFlow applications with the C++ API. Let’s clarify the difference between reentrancy and thread safety first (since they’re often confused), then dive into how this applies to Session::Run().

Key Definitions #

• Thread Safety: Multiple threads can safely call Session::Run() on the same Session instance at the same time. TensorFlow handles internal synchronization for this scenario.
• Reentrancy: A function is reentrant if it can be safely called again from the same thread before the first call completes (e.g., nested calls in a callback).

Behavior of Session::Run() #

• Thread Safe: Yes, you don’t need to add your own mutexes or synchronization when accessing the same Session from multiple threads. The TensorFlow runtime manages concurrent access to the Session’s internal state.
• Not Reentrant: Absolutely do NOT call Session::Run() on the same Session instance from within a running Run() call in the same thread. This leads to undefined behavior—think crashes, corrupted tensor data, or deadlocks. TensorFlow’s Session implementation doesn’t handle nested reentry of Run() in a single thread.

Practical Recommendations #

• For multi-threaded access across separate threads: Go ahead, no extra work needed. The API is designed for this use case.
• If you need to trigger a Run() from within an ongoing Run() (e.g., in a custom Op callback):
  • Either use a separate Session instance for the nested call (just make sure it’s initialized with the same graph if needed).
  • Or offload the nested Run() to a different thread to avoid reentry in the same thread context.

内容的提问来源于stack exchange，提问作者Ivan Jobs