bugfix: clean NPU process shutdown (bypass process-static destructors).

[Enguikong]

问题

SIGTERM 停服时,Ascend libopp_registry.so 的进程静态析构 + pybind GIL dec_ref 会在 run() 清理之后抛 terminate,导致进程非干净退出(core/terminate)。

方案

新增 exit_npu_process(int):flush 日志/stdio 后调用 std::_Exit,跳过进程静态析构,干净退出。main() 在 USE_NPU 下用它替代 return run()。

说明

• 与 DCP 功能正交,独立提交(不进 DCP PR feat: support Qwen3.5 Decode Context Parallel on Ascend NPU. #2125)。
• 已在 Ascend 910C 上卡验证:SIGTERM 停服哨兵日志出现、无 GIL/terminate/core、HBM/fd 归零。

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR changes the NPU build’s process-exit behavior to avoid crashes during shutdown caused by process-static destructors (notably Ascend libopp_registry.so) interacting poorly with pybind/Python teardown. It introduces an explicit NPU exit path that flushes logs/stdio and then terminates via std::_Exit(...) to bypass static destructors.

Changes:

• Add exit_npu_process(int) that shuts down verbose tracing, flushes glog and stdio streams, then calls std::_Exit(exit_code).
• Update main() to call exit_npu_process(run()) under USE_NPU instead of returning from main.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.