Pytorch CUDA runtime error

Hi,

I have built pytorch from source as shown at the official pytorch website.
I have a GPU with compute capability 3.0 and was advised to build pytorch which can support compute capability 3.0.
Even after building from source I am getting a warning as the GPU is too old, but the cuda_is_available returns true.
But when I am trying to run pytorch, I still have the following error.
File “waveglow/glow.py”, line 94, in forward
W_inverse = W_inverse.half()
RuntimeError: CUDA error: no kernel image is available for execution on the device
while trying to run the nvidia-tacotron2 demo

Hi,

I think float16 computations did not exist back then. So your GPU most likely does not support such operations.

1 Like

I also think this might be the reason.
If I’m not mistaken, FP16 operations were introduced for compute capability 5.3.

1 Like