How are tensorflow ops lowered to XLA ops during for training ?
The files under tensorflow/compiler/tf2xla/kernels only define the forward pass / inference in terms of ops in ComputationBuilder.
This question is related to "Tensorflow - XLA | Passing tensors to external functions at runtime", in that I need TF ops that are lowered to XLA ops to pass the tensors during training and inference. Please share your thoughts on the other question as well.
Thanks
My understanding is that Tensorflow computes the derivatives needed for backpropagation before it passes the graph to tf2xla for lowering. All Tensorflow ops for training and inference are lowered to XLA using tf2xla; there is no other mechanism.