I have installed TF from source and enabled XLA JIT compilation. When I run my training code in a session, it shows the following messages:
"XLA service 0x62bb180 executing computations on platform Host" and
"XLA service 0x62a43b0 executing computations on platform CUDA"
Does this mean that my training is with the XLA_GPU device? Because when I tried to explicitly wrap my session with tf.device("/device:XLA_GPU:0"), there is seemingly no performance gains at all (the training speed is still the same). And from the documentation I read, it seems that wrapping the session with tf.device is one of the ways to activate XLA JIT.
Is there a way to verify whether I am using XLA JIT in my training?