cudaCheckError() failed : no kernel image is available for execution on the device

1

Getting this error while running yolov2 https://github.com/longcw/yolo2-pytorch
Ubuntu 16.04
cuda 8.0
python 3.6
pytorch 0.3.0.post4
gcc 5.4.0
gpu Quadro M1200

Installed PyTorch Linux binary compiled with cuda8 , using pip

Any ideas on what could be done to fix it?

2

It seems the error is thrown from the compiled kernels using nvcc in YOLO2 (line of code), not PyTorch directly.
Could you try so get -arch to -arch=sm_61, as it seems your M1200 is built using the Pascal architecture (doc).

1 Like
3

I 'm facing the same error here,
I’m using Titan Xp GPUs to run neural-motifs .
I have both cuda-10.0 and cuda-9.0 installed, but the driver is for cuda-10.0
Any idea?

1 Like
4

I have been facing this problem for two days since i am leaning pytorch by the Official Pytorch Tutorial. I am stopped at
y = torch.ones_like(x, device=device)
when python3.6 reported this cuda error.
I searched the internet and got no anwser.
That’s disappointing …

I am on CentOS 7 with a GT650M gpu and CUDA 10.1
nvidia-smi information:

59%20PM
59%20PM1126Ă—584 45.1 KB

SO, H!E!L!P! M!E!

5

Your GT650M has compute capability 3.0 based on this table, which isn’t supported anymore using the prebuilt binaries. You could build from source following this instruction.

1 Like
6

Thanks a lot, however, the guide of “Install from source” said,
“Once you have Anaconda installed, here are the instructions.”

I can’t find where are the instructions…:pensive:

Maybe from source means from rpm?

7

The instructions are in the following sentences:

If you want to compile with CUDA support, install
...

Just follow this passage and it should walk you through the install process. :wink:

8

Hi Cao. Were you able to fix this?

9

Actually, it was the problem when I compiled some file written in cuda.
Please check the followings:

  1. Pytorch is compiled in the same version with system cuda version
  2. Choose the true architecture code for your GPUs. For example, in this line https://github.com/rowanz/neural-motifs/blob/d05a251b705cedaa51599bf2906cfa4653b7a761/lib/fpn/nms/src/cuda/Makefile#L2, the GPU architecture is 61 for Titan Xp. You must change it into 52 for Geforce GTX Titan, or 75 for Titan V.
    Hope this helps.