Attenton

500G 的显存为啥会无法完全加载 370G ？推理框架现在的事实标准应该都是 vllm 吧，试下 vllm 吧

@unidotnet 14b 即使是 fp16 也只是大概 28G 左右，int4 怎么可能要 40G

@geelaw 感谢

@geelaw 感谢，这个标准的定义可以在哪看？

@ysc3839 嗯嗯，我也找到了，感谢

If lock is called by a thread that already owns the mutex, the behavior is undefined: for example, the program may deadlock. An implementation that can detect the invalid usage is encouraged to throw a std::system_error with error condition resource_deadlock_would_occur instead of deadlocking.

If the mutex is currently locked by the same thread calling this function, it produces a deadlock (with undefined behavior).

是个 undefined behavior ，不能这样使用，结贴😓


Reference:
https://en.cppreference.com/w/cpp/thread/mutex/lock
https://cplusplus.com/reference/mutex/mutex/lock/

@iwdmb
(base) root@ubuntu:/home# ./a.out
n = 5
n = 4
n = 3
n = 2
n = 1
n = 0
(base) root@ubuntu:/home# g++ --version
g++ (Ubuntu 7.5.0-3ubuntu1~18.04) 7.5.0
Copyright (C) 2017 Free Software Foundation, Inc.
This is free software; see the source for copying conditions. There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

以上是我的执行结果，没搞懂是哪的问题

补充：CmakeLists.txt 里面没有添加任何编译选项，cmake 版本是 3.18 ， 用的 c++ 17