The record: could not create cublas handle: CUBLAS_STATUS_ALLOC

[size=16 px] nvidia 1660 s, the first to install tensorflow - gpu2.1.0, found that tf. Test_is_gpu_available () is True, but a run resnet18, CPU100 % usage, and the GPU is almost untouched, all kinds of search, to find the cause of the problem,

Before this problem bothering me for a long time, really thought it was a version is too high, from tensorflow - gpu2.10 to 2.0.0, or not, the corresponding cuda, cudnn version also is according to the corresponding official documents,

Had almost gave up, carefully check the error message:

The Loaded runtime CuDNN library: 7.4.1 but the source was compiled with: 7.6.0.

Then download the 7.6.0 cudnn, replace, there is a problem again:

Could not create cublas handle: CUBLAS_STATUS_ALLOC_FAILED

It doesn't matter, continue to search,

Online solutions
The config=tf. ConfigProto ()
The session=tf. The session (config=config)

Also not line, this is tensorflow1. X version of the code, so let's add tf.com. Pat v1, otherwise cannot find ConfigProto module,

To:

The config=tf.com pat. V1. ConfigProto ()
The session=tf. The session (config=config)

The results also not line,

Finally found less one line of code, as shown in figure:

We focus on the red box code, the rest of the code is to search out, but tried no matter use,

Hope for the same problem, help, [/size]