【解决：tensorrt c++编译gpu cuda报错：“error C2912: 显式专用化”或“不是函数模板的专用化”】

cuda/gpu报错，会提示类似如下错误：

error C2059: 语法错误:“volatile”
error C2912: 显式专用化
error C2065: “LOAD_CS”: 未声明的标识符
”不是函数模板的专用化

完整报错log部分截取如下：

Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\include\cub/thread/thread_load.cuh(291): error C2059: 语法错误:“volatile”
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\include\cub/thread/thread_load.cuh(291): error C2912: 显式专用化；“ulonglong2 cub::CUB_200802_SM_890::ThreadLoad<LOAD_CA,const ulonglong2*>(const ulonglong2 *)”不是函数模板的专用化
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\include\cub/thread/thread_load.cuh(291): error C2912: 显式专用化；“ushort4 cub::CUB_200802_SM_890::ThreadLoad<LOAD_CA,const ushort4*>(const ushort4 *)”不是函数模板的专用化
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\include\cub/thread/thread_load.cuh(291): error C2912: 显式专用化；“uint2 cub::CUB_200802_SM_890::ThreadLoad<LOAD_CA,const uint2*>(const uint2 *)”不是函数模板的专用化
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\include\cub/thread/thread_load.cuh(291): error C2912: 显式专用化；“unsigned __int64 cub::CUB_200802_SM_890::ThreadLoad<LOAD_CA,const unsigned __int64*>(const unsigned __int64 *)”不是函数模板的专用化
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\include\cub/thread/thread_load.cuh(291): error C2912: 显式专用化；“unsigned int cub::CUB_200802_SM_890::ThreadLoad<LOAD_CA,const unsigned int*>(const unsigned int *)”不是函数模板的专用化
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\include\cub/thread/thread_load.cuh(291): error C2912: 显式专用化；“unsigned short cub::CUB_200802_SM_890::ThreadLoad<LOAD_CA,const unsigned short*>(const unsigned short *)”不是函数模板的专用化
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\include\cub/thread/thread_load.cuh(291): error C2912: 显式专用化；“unsigned char cub::CUB_200802_SM_890::ThreadLoad<LOAD_CA,const unsigned char*>(const unsigned char *)”不是函数模板的专用化
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\include\cub/thread/thread_load.cuh(292): error C2065: “LOAD_CG”: 未声明的标识符
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\include\cub/thread/thread_load.cuh(292): error C2912: 显式专用化；“uint4 cub::CUB_200802_SM_890::ThreadLoad<LOAD_CG,const uint4*>(const uint4 *)”不是函数模板的专用化
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\include\cub/thread/thread_load.cuh(292): error C2059: 语法错误:“volatile”
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\include\cub/thread/thread_load.cuh(292): error C2912: 显式专用化；“ulonglong2 cub::CUB_200802_SM_890::ThreadLoad<LOAD_CG,const ulonglong2*>(const ulonglong2 *)”不是函数模板的专用化
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\include\cub/thread/thread_load.cuh(292): error C2912: 显式专用化；“ushort4 cub::CUB_200802_SM_890::ThreadLoad<LOAD_CG,const ushort4*>(const ushort4 *)”不是函数模板的专用化
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\include\cub/thread/thread_load.cuh(292): error C2912: 显式专用化；“uint2 cub::CUB_200802_SM_890::ThreadLoad<LOAD_CG,const uint2*>(const uint2 *)”不是函数模板的专用化
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\include\cub/thread/thread_load.cuh(292): error C2912: 显式专用化；“unsigned __int64 cub::CUB_200802_SM_890::ThreadLoad<LOAD_CG,const unsigned __int64*>(const unsigned __int64 *)”不是函数模板的专用化
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\include\cub/thread/thread_load.cuh(292): error C2912: 显式专用化；“unsigned int cub::CUB_200802_SM_890::ThreadLoad<LOAD_CG,const unsigned int*>(const unsigned int *)”不是函数模板的专用化
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\include\cub/thread/thread_load.cuh(292): error C2912: 显式专用化；“unsigned short cub::CUB_200802_SM_890::ThreadLoad<LOAD_CG,const unsigned short*>(const unsigned short *)”不是函数模板的专用化
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\include\cub/thread/thread_load.cuh(292): error C2912: 显式专用化；“unsigned char cub::CUB_200802_SM_890::ThreadLoad<LOAD_CG,const unsigned char*>(const unsigned char *)”不是函数模板的专用化
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\include\cub/thread/thread_load.cuh(293): error C2065: “LOAD_CS”: 未声明的标识符
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\include\cub/thread/thread_load.cuh(293): error C2912: 显式专用化；“uint4 cub::CUB_200802_SM_890::ThreadLoad<LOAD_CS,const uint4*>(const uint4 *)”不是函数模板的专用化
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\include\cub/thread/thread_load.cuh(293): error C2059: 语法错误:“volatile”
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\include\cub/thread/thread_load.cuh(293): error C2912: 显式专用化；“ulonglong2 cub::CUB_200802_SM_890::ThreadLoad<LOAD_CS,const ulonglong2*>(const ulonglong2 *)”不是函数模板的专用化
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\include\cub/thread/thread_load.cuh(293): error C2912: 显式专用化；“ushort4 cub::CUB_200802_SM_890::ThreadLoad<LOAD_CS,const ushort4*>(const ushort4 *)”不是函数模板的专用化
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\include\cub/thread/thread_load.cuh(293): error C2912: 显式专用化；“uint2 cub::CUB_200802_SM_890::ThreadLoad<LOAD_CS,const uint2*>(const uint2 *)”不是函数模板的专用化
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\include\cub/thread/thread_load.cuh(293): error C2912: 显式专用化；“unsigned __int64 cub::CUB_200802_SM_890::ThreadLoad<LOAD_CS,const unsigned __int64*>(const unsigned __int64 *)”不是函数模板的专用化
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\include\cub/thread/thread_load.cuh(293): error C2912: 显式专用化；“unsigned int cub::CUB_200802_SM_890::ThreadLoad<LOAD_CS,const unsigned int*>(const unsigned int *)”不是函数模板的专用化
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9\include\cub/thread/thread_load.cuh(293): fatal error C1003: 错误计数超过 100；正在停止编译

使用原因为我编译c++推理脚本，使用的是tensorrt10，换到另一台电脑上发现运行报错。两台电脑都是win11。
因为代码在上一台机器可以运行，那肯定就不是代码问题，而环境配置也都修改完毕，相关库也没有缺失，所以怀疑可能是cuda版本问题，对比发现原cuda版本为12.4，新电脑为12.9
查看cuda版本可以运行

nvcc --version
# 或者
nvcc -V

卸载cuda12.9，删除文件夹：C:\Program Files\NVIDIA GPU Computing Toolkit
然后重装cuda12.4，再配置cudnn，重新配置tensorrt10
此时再运行程序，完美解决！