pytorch 报错 RuntimeError: CUDA error: no kernel image is available for execution on the device

环境：

• python 3.8.3
• pytorch 1.7.1
• cuda10.1
• GPU: GeForce 920M

问题

用GPU运行代码时报错如下

因为我的cuda是配好了的，用tensorflow时就没问题，本来运行以下测试的时候发现也是没问题的。

In [2]: torch.cuda.is_available()
Out[2]: True

但是再测试一下其它的

In [3]: torch.empty(3,4,device='cuda')
或者
In [5]: torch.Tensor([1,2]).cuda()

就又会报错RuntimeError: CUDA error: no kernel image is available for execution on the device
说明pytorch的cuda环境还是没配好

解决方法

出现这个问题的原因应该是 cuda或pytorch 与显卡算力不区配

1.查看显卡算力

有下面两种方法
方法1: 终端进到装CUDA的目录C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.1\extras\demo_suite>
然后输入deviceQuery，如下所示

C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.1\extras\demo_suite>deviceQuery
deviceQuery Starting...

 CUDA Device Query (Runtime API) version (CUDART static linking)

Detected 1 CUDA Capable device(s)

Device 0: "GeForce 920M"
  CUDA Driver Version / Runtime Version          10.1 / 10.1
  CUDA Capability Major/Minor version number:    3.5
  Total amount of global memory:                 2048 MBytes (2147483648 bytes)
  ( 2) Multiprocessors, (192) CUDA Cores/MP:     384 CUDA Cores
  GPU Max Clock rate:                            954 MHz (0.95 GHz)
  Memory Clock rate:                             900 Mhz
  Memory Bus Width:                              64-bit
  L2 Cache Size:                                 524288 bytes
  Maximum Texture Dimension Size (x,y,z)         1D=(65536), 2D=(65536, 65536), 3D=(4096, 4096, 4096)
  Maximum Layered 1D Texture Size, (num) layers  1D=(16384), 2048 layers
  Maximum Layered 2D Texture Size, (num) layers  2D=(16384, 16384), 2048 layers
  Total amount of constant memory:               zu bytes
  Total amount of shared memory per block:       zu bytes
  Total number of registers available per block: 65536
  Warp size:                                     32
  Maximum number of threads per multiprocessor:  2048
  Maximum number of threads per block:           1024
  Max dimension size of a thread block (x,y,z): (1024, 1024, 64)
  Max dimension size of a grid size    (x,y,z): (2147483647, 65535, 65535)
  Maximum memory pitch:                          zu bytes
  Texture alignment:                             zu bytes
  Concurrent copy and kernel execution:          Yes with 1 copy engine(s)
  Run time limit on kernels:                     Yes
  Integrated GPU sharing Host Memory:            No
  Support host page-locked memory mapping:       Yes
  Alignment requirement for Surfaces:            Yes
  Device has ECC support:                        Disabled
  CUDA Device Driver Mode (TCC or WDDM):         WDDM (Windows Display Driver Model)
  Device supports Unified Addressing (UVA):      Yes
  Device supports Compute Preemption:            No
  Supports Cooperative Kernel Launch:            No
  Supports MultiDevice Co-op Kernel Launch:      No
  Device PCI Domain ID / Bus ID / location ID:   0 / 1 / 0
  Compute Mode:
     < Default (multiple host threads can use ::cudaSetDevice() with device simultaneously) >

deviceQuery, CUDA Driver = CUDART, CUDA Driver Version = 10.1, CUDA Runtime Version = 10.1, NumDevs = 1, Device0 = GeForce 920M
Result = PASS

这样就可以看到显卡信息了，上面第二行的3.5就是我的显卡算力了.
方法2: 进入这个链接,找到你对应的显卡就能看到了

2.找到问题

在python 里输入如下命令

In [1]: import torch

In [2]: torch.cuda.get_arch_list()
Out[2]: ['sm_37', 'sm_50', 'sm_60', 'sm_70']

（如果嫌麻烦下面这一步可以省略）

或者从这个gitee上下到collect_env.py
然后执行结果如下：

上面显示我的算力是3.5，但是下面提示我装的cuda或者pytorch最低支持3.7，所以我一用gpu它就会报错
因为新版的cuda或者pytorch对老显卡支持不太好，所以我们需要换一个老一点的版本。
那么到底应该是装旧版的cuda还是旧版的pytorch呢，我在一个装有pytorch1.7.1 + cuda9.2的服务上再测试刚才的colllect_env.py，结果如下：

可见，在同样的pytorch版本，不同的cuda下，它对算力的要求是一样的，所以问题出在pytorch上
但是cuda10以下好像不支持算力3.0以下了，所以其实与cuda也有关系

3.解决办法

解决方法具我所知应该有两个

方法1：安装老版本的pytorch

但具体要装的版本号是多少呢，
网好像并没有公布每个pytorch版本与显卡算力的对应关系，所以只能靠试了
以下是Pytoch的讨论区里找的别人试出来的，所以综合来看，cuda与pytorch版本都不能高了
原话是"I extracted the compute capabilities for which each pytorch package on conda is compiled:"

package	pyton	cuda	cudnn	architectures
pytorch-1.0.0	py3.7	cuda10.0.130	cudnn7.4.1_1	sm_30, sm_35, sm_50, sm_60, sm_61, sm_70, sm_75
pytorch-1.0.0	py3.7	cuda8.0.61	cudnn7.1.2_1	sm_20, sm_35, sm_37, sm_50, sm_52, sm_53, sm_60, sm_61
pytorch-1.0.0	py3.7	cuda9.0.176	cudnn7.4.1_1	sm_35, sm_37, sm_50, sm_52, sm_53, sm_60, sm_61, sm_70
pytorch-1.0.1	py3.7	cuda10.0.130	cudnn7.4.2_0	sm_35, sm_50, sm_60, sm_61, sm_70, sm_75
pytorch-1.0.1	py3.7	cuda10.0.130	cudnn7.4.2_2	sm_35, sm_50, sm_60, sm_61, sm_70, sm_75
pytorch-1.0.1	py3.7	cuda8.0.61	cudnn7.1.2_0	sm_35, sm_37, sm_50, sm_52, sm_53, sm_60, sm_61
pytorch-1.0.1	py3.7	cuda8.0.61	cudnn7.1.2_2	sm_35, sm_37, sm_50, sm_52, sm_53, sm_60, sm_61
pytorch-1.0.1	py3.7	cuda9.0.176	cudnn7.4.2_0	sm_35, sm_50, sm_60, sm_61, sm_70
pytorch-1.0.1	py3.7	cuda9.0.176	cudnn7.4.2_2	sm_35, sm_50, sm_60, sm_70
pytorch-1.1.0	py3.7	cuda10.0.130	cudnn7.5.1_0	sm_35, sm_50, sm_60, sm_61, sm_70, sm_75
pytorch-1.1.0	py3.7	cuda9.0.176	cudnn7.5.1_0	sm_35, sm_50, sm_60, sm_61, sm_70
pytorch-1.2.0	py3.7	cuda9.2.148	cudnn7.6.2_0	sm_35, sm_50, sm_60, sm_61, sm_70
pytorch-1.2.0	py3.7	cuda10.0.130	cudnn7.6.2_0	sm_35, sm_50, sm_60, sm_61, sm_70, sm_75
pytorch-1.2.0	py3.7	cuda9.2.148	cudnn7.6.2_0	sm_35, sm_50, sm_60, sm_61, sm_70
pytorch-1.3.0	py3.7	cuda10.0.130	cudnn7.6.3_0	sm_30, sm_35, sm_50, sm_60, sm_61, sm_70, sm_75
pytorch-1.3.0	py3.7	cuda10.1.243	cudnn7.6.3_0	sm_30, sm_35, sm_50, sm_60, sm_61, sm_70, sm_75
pytorch-1.3.0	py3.7	cuda9.2.148	cudnn7.6.3_0	sm_35, sm_50, sm_60, sm_61, sm_70
pytorch-1.3.1	py3.7	cuda10.0.130	cudnn7.6.3_0	sm_35, sm_37, sm_50, sm_60, sm_61, sm_70, sm_75
pytorch-1.3.1	py3.7	cuda10.1.243	cudnn7.6.3_0	sm_35, sm_37, sm_50, sm_60, sm_61, sm_70, sm_75
pytorch-1.3.1	py3.7	cuda9.2.148	cudnn7.6.3_0	sm_35, sm_37, sm_50, sm_60, sm_61, sm_70
pytorch-1.4.0	py3.7	cuda10.0.130	cudnn7.6.3_0	sm_35, sm_37, sm_50, sm_60, sm_61, sm_70, sm_75
pytorch-1.4.0	py3.7	cuda10.1.243	cudnn7.6.3_0	sm_35, sm_37, sm_50, sm_60, sm_61, sm_70, sm_75
pytorch-1.4.0	py3.7	cuda9.2.148	cudnn7.6.3_0	sm_35, sm_37, sm_50, sm_60, sm_61, sm_70
pytorch-1.5.0	py3.7	cuda10.1.243	cudnn7.6.3_0	sm_35, sm_37, sm_50, sm_60, sm_61, sm_70, sm_75
pytorch-1.5.0	py3.7	cuda10.2.89	cudnn7.6.5_0	sm_35, sm_37, sm_50, sm_60, sm_61, sm_70, sm_75
pytorch-1.5.0	py3.7	cuda9.2.148	cudnn7.6.3_0	sm_35, sm_37, sm_50, sm_60, sm_61, sm_70
pytorch-1.5.1	py3.7	cuda10.1.243	cudnn7.6.3_0	sm_35, sm_37, sm_50, sm_60, sm_61, sm_70, sm_75
pytorch-1.5.1	py3.7	cuda10.2.89	cudnn7.6.5_0	sm_35, sm_37, sm_50, sm_60, sm_61, sm_70, sm_75
pytorch-1.5.1	py3.7	cuda9.2.148	cudnn7.6.3_0	sm_35, sm_37, sm_50, sm_60, sm_61, sm_70
pytorch-1.6.0	py3.7	cuda10.1.243	cudnn7.6.3_0	sm_35, sm_37, sm_50, sm_60, sm_61, sm_70, sm_75
pytorch-1.6.0	py3.7	cuda10.2.89	cudnn7.6.5_0	sm_35, sm_37, sm_50, sm_60, sm_61, sm_70, sm_75
pytorch-1.6.0	py3.7	cuda9.2.148	cudnn7.6.3_0	sm_35, sm_37, sm_50, sm_60, sm_61, sm_70
pytorch-1.7.0	py3.7	cuda10.1.243	cudnn7.6.3_0	sm_35, sm_37, sm_50, sm_60, sm_61, sm_70, sm_75
pytorch-1.7.0	py3.7	cuda10.2.89	cudnn7.6.5_0	sm_35, sm_37, sm_50, sm_60, sm_61, sm_70, sm_75
pytorch-1.7.0	py3.7	cuda11.0.221	cudnn8.0.3_0	sm_37, sm_50, sm_60, sm_61, sm_70, sm_75, sm_80
pytorch-1.7.0	py3.7	cuda9.2.148	cudnn7.6.3_0	sm_35, sm_37, sm_50, sm_60, sm_61, sm_70
pytorch-1.7.1	py3.7	cuda10.1.243	cudnn7.6.3_0	sm_35, sm_37, sm_50, sm_60, sm_61, sm_70, sm_75
pytorch-1.7.1	py3.7	cuda10.2.89	cudnn7.6.5_0	sm_35, sm_37, sm_50, sm_60, sm_61, sm_70, sm_75
pytorch-1.7.1	py3.7	cuda11.0.221	cudnn8.0.5_0	sm_35, sm_37, sm_50, sm_60, sm_61, sm_70, sm_75, sm_80
pytorch-1.7.1	py3.7	cuda9.2.148	cudnn7.6.3_0	sm_37, sm_50, sm_60, sm_61, sm_70
pytorch-1.8.0	py3.7	cuda10.1	cudnn7.6.3_0	sm_35, sm_37, sm_50, sm_60, sm_61, sm_70, sm_75
pytorch-1.8.0	py3.7	cuda10.2	cudnn7.6.5_0	sm_35, sm_37, sm_50, sm_60, sm_61, sm_70, sm_75
pytorch-1.8.0	py3.7	cuda11.1	cudnn8.0.5_0	sm_35, sm_37, sm_50, sm_60, sm_61, sm_70, sm_75, sm_80, sm_86
pytorch-1.8.1	py3.7	cuda10.1	cudnn7.6.3_0	sm_35, sm_37, sm_50, sm_60, sm_61, sm_70, sm_75
pytorch-1.8.1	py3.7	cuda10.2	cudnn7.6.5_0	sm_35, sm_37, sm_50, sm_60, sm_61, sm_70, sm_75
pytorch-1.8.1	py3.7	cuda11.1	cudnn8.0.5_0	sm_35, sm_37, sm_50, sm_60, sm_61, sm_70, sm_75, sm_80, sm_86

但是上面的结果与我的有点不太一样，我在cuda10.1.243、CuDNN 7.6.4的情况下测试了pytorch1.7.1和1.6.0，所以就很怪
以下两个是我测的

pytorch1.7.1 : ['sm_37', 'sm_50', 'sm_60', 'sm_70']
pytorch1.6.0 ：['sm_37', 'sm_50', 'sm_60', 'sm_61', 'sm_70', 'sm_75', 'compute_37']

测试就是看一下这个的输出（好像从pytorch1.4 以下pytorch就没有这个属性）

In [2]: torch.cuda.get_arch_list()
Out[2]: ['sm_37', 'sm_50', 'sm_60', 'sm_70']

在cuda不变的情况下 我从1.7.1 装回1.6.0到1.4.0都不行，我又新开了个pyton3.6的虚拟环境装好了pytorch1.3.1，但仍然不行，如果换个低版本的cuda没准可以，但是我不想试了，因为我看到人家就可以，我的不行，不知道是为啥，我抽空去用源码安装一下试试

未完待续。。。

方法2：

从源码编译安装pytorch，然后在源码里面设置一下应该就好了，但可惜的是我不会，感觉有些麻烦，听说要弄好几个小时

还有的听说要在什么Makefile文件里面添加类似于下面这样的东西，但我一直没找到到底在添加在哪
-gencode arch=compute_30,code=[sm_30,compute_30]
-gencode arch=compute_35,code=[sm_35,compute_35] \

这个问题会继续探索，如果有新发现了我再更新