CUDA runtime implicit initialization on GPU:0 failed Status: device kernel image is invalid. about dannce HOT 5 CLOSED

I have essentially the same cuda/tensorflow combo (my tensorflow is 2.3.1 but I doubt that matters). What operating system are you using? What GPU are you using? Does nvidia-smi produce a valid output? Can you run other tensorflow code?

Discussion around this tensorflow issue suggests that TF 2.3 might not be compatible with certain GPUs. Although it looks like downgrading to TF 2.2 might work.

from dannce.

@yuan0821 any luck?

from dannce.

Oh!! thanks a lot! and sorry for late reply! i have fixed it already. my computer system is Ubuntu 18.0, and the GPU is RTX3070. The cuda > 11., so the trick is to get the cudatoolkit 11.0.3 with pytorch 1.7. @spoonsso

from dannce.

Oh!! thanks a lot! and sorry for late reply! i have fixed it already. my computer system is Ubuntu 18.0, and the GPU is RTX3070. The cuda > 11., so the trick is to get the cudatoolkit 11.0.3 with pytorch 1.7. @spoonsso

Hi，Could you mind providing details about how to make it work on GPU RTX30xx.

I got a RTX3090 (30xx required tf2.4 and above), but it seems that DANNCE only works on tf 2.3.1.

How you make the DANNCE work on your hardware and CUDA11.

Thanks!

from dannce.

I am also having this issue with RTX3060, tf 2.3.1, cudatoolkit 10.1. Seems like it might be an RTX* issue since everyone in this thread has one? Did anyone find a combination of versions that works?

Trying to use the suggestion from @yuan0821 to use cudatoolkit 11.0.3 and pytorch 1.7 but having some trouble getting this to work with conda.

thanks!

from dannce.