torch.fx.experimental.proxy_tensor.make_fx

torch.fx.experimental.proxy_tensor.make_fx(f, decomposition_table=None, tracing_mode='real', _allow_non_fake_inputs=False, *, pre_dispatch=False, record_module_stack=False, _allow_fake_constant=False, _error_on_data_dependent_ops=True, record_stack_traces=False, proxy_module_inputs=False, _disable_torch_fn_metadata_mode=False, dynamic_shapes=None)

Given a function f, return a new function which when executed with valid arguments to f, returns an FX GraphModule representing the set of operations that were executed during the course of execution.

If record_stack_traces is True, the stack trace will be preserved on node.meta[“stack_trace”]

By default, post-dispatch traces without an explicit decomposition_table rewrite detach to alias. Re-differentiating the returned graph does not preserve those detach calls as autograd gradient cuts; pass an explicit decomposition_table or use pre_dispatch=True if exact detach nodes are needed.

tracing_mode:

• "real": no fakification, traces with real tensors.

• "fake": tensors become FakeTensors. By default all dims are static (concrete); if dynamic_shapes is set, the spec controls which dims become unbacked symbolic sizes — everything not declared by the spec stays static.

• "symbolic": tensors become FakeTensors and every dim becomes a backed symbolic size. NOTE: using backed symbols this way is generally not recommended and not sound — backed symbols come with assumptions and can be constrained by guards silently. Kept for backward compatibility. Prefer "fake" + dynamic_shapes to declare dynamism explicitly with unbacked symbols.

Return type:

Callable[[…], GraphModule]