A possible fix for "failed call to cuInit: CUDA_ERROR_NO_DEVICE: no CUDA-capable device"

gistfile1.txt

This is a dreaded error that seems pop up its ugly head again and again, in particular after upgrading CUDA or Tensorflow.
Typcially it looks like this:

2020-12-30 17:31:40.829615: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library libcudart.so.11.0
2020-12-30 17:31:42.149768: I tensorflow/compiler/jit/xla_cpu_device.cc:41] Not creating XLA devices, tf_xla_enable_xla_devices not set
2020-12-30 17:31:42.150368: I tensorflow/stream_executor/platform/default/dso_loader.cc:49] Successfully opened dynamic library libcuda.so.1
2020-12-30 17:31:42.176643: E tensorflow/stream_executor/cuda/cuda_driver.cc:328] failed call to cuInit: CUDA_ERROR_NO_DEVICE: no CUDA-capable device is detected

Here is a solution that currently seems to work on my system, 
with Cuda 11.0 and Tensorflow 2.4.0, you can try it if all the 
other 500 proposed fixes did not work for you:

Add these two lines at the very top of your script BEFORE any call to "import tensorflow":

from tensorflow.python.client import device_lib
device_lib.list_local_devices()

That's all suddenly it works. I have no clue why.

If that did not help, here are a few other things you can try:
https://victorwyee.com/engineering/failed-call-to-cuinit-cuda-error-unknown/

This helped me, thanks.

This fix seam to be silly, but still works perfectly. found this after a long struggle

I'm really guessing here but I have a thought on why it could work.
When I worked with serial devices in the past, sometimes the devices wouldn't talk unless I pinged them with a generic
query like, "what's your serial number?". Perhaps there's something similar happening with the GPU and the BUS that connects them.

LOL! It actually works. I didn't expect that at all. Thanks!

worked to me, thinks!