So far torch.fx supports symbolic trace only, therefore inevitably has some limitation. Is it going to support “scripting” mode to directly translate python code as torch.jit.script does?
Hi @yzhliu,
No, at this time the plan is to support tracing only. This stems from a desire to reduce the complexity of a) the front-end b) the represenatation and c) transforms and to keep FX accessible to a more general audience. True, this limits the type of models that can be transformed with FX, however we believe that the types of models folks work with (vision, NLP, etc) tend not to require sophisticated constructs