How does the XLA process when need to run on cpu op code?

[engineer1109]

Some pytorch code need CPU OP support.

How does the XLA process?

Build multi graphs?

[antkillerfarm]

The pipeline is below:
pytorch op -> xla op -> TIM-VX op

We don't support pytorch op directly.

There are two ways to support unofficial pytorch op:

• use official pytorch ops to construct your custom op.
• add op compiler in torch_xla to transform your custom op to XLA ops form.

[engineer1109]

It seems like lazy mode, not immediate mode.
Need write op to a graph.
When need to get output, then compile and run graph.
So all ops should be in the graph.