Force the CUDA runtime initialization before device creation.
This is to avoid silent failure and garbage results produced when launching two TensorFlow programs simultaneously in two different processes. PiperOrigin-RevId: 173456597
Loading
Please sign in to comment