[Code scan] Close TensorFlow sessions in C++ API destructors

[njzjz]

This issue comes from a Codex global scan of deepmodeling/deepmd-kit at commit 73de44b1f94471b2e3bdb6b11f57b34d7bc791bb.

Problem

Several TensorFlow C++ API wrappers create a tensorflow::Session with NewSession, but their destructors only delete the GraphDef.

• DeepPotTF creates the session here:

  deepmd-kit/source/api_cc/src/DeepPotTF.cc

  Line 458 in 73de44b

  check_status(NewSession(options, &session));

• Its destructor only deletes graph_def:

  deepmd-kit/source/api_cc/src/DeepPotTF.cc

  Line 419 in 73de44b

  DeepPotTF::~DeepPotTF() { delete graph_def; }

• The raw session pointer is stored in the class:

  deepmd-kit/source/api_cc/include/DeepPotTF.h

  Line 296 in 73de44b

  tensorflow::Session* session;

The same lifetime pattern appears in the other TensorFlow API wrappers:

• DeepSpinTF creates a session at

  deepmd-kit/source/api_cc/src/DeepSpinTF.cc

  Line 458 in 73de44b

  check_status(NewSession(options, &session));

  , but its destructor only deletes graph_def at

  deepmd-kit/source/api_cc/src/DeepSpinTF.cc

  Line 419 in 73de44b

  DeepSpinTF::~DeepSpinTF() { delete graph_def; }

  .
• DeepTensorTF creates a session at

  deepmd-kit/source/api_cc/src/DeepTensorTF.cc

  Line 57 in 73de44b

  deepmd::check_status(NewSession(options, &session));

  , but its destructor only deletes graph_def at

  deepmd-kit/source/api_cc/src/DeepTensorTF.cc

  Line 23 in 73de44b

  DeepTensorTF::~DeepTensorTF() { delete graph_def; }

  .
• DipoleChargeModifierTF creates a session at

  deepmd-kit/source/api_cc/src/DataModifierTF.cc

  Line 60 in 73de44b

  deepmd::check_status(NewSession(options, &session));

  , but its destructor only deletes graph_def at

  deepmd-kit/source/api_cc/src/DataModifierTF.cc

  Line 26 in 73de44b

  DipoleChargeModifierTF::~DipoleChargeModifierTF() { delete graph_def; };

  .

The constructor catch blocks also only delete graph_def if init() throws, so failures after NewSession can leak the partially constructed session before the destructor can run.

Impact

Repeated construction/destruction of TensorFlow C++ API objects can leak TensorFlow sessions, including associated CPU/GPU resources and runtime threads. This is especially relevant for long-running host applications, plugins, or test processes that load many models.

Suggested fix

Close and delete the TensorFlow session in each destructor, and clean it up on constructor failure after NewSession succeeds. Prefer RAII ownership for both session and graph_def so partial initialization and exceptions follow the same cleanup path.