What's the difference between symbolic_trace and Tracer

torch.compile FX (Functional Transformations)
1

What’s the difference between torch.fx.symbolic_trace and torch.fx.Tracer().trace ?

import torch
from torchvision.models.resnet import resnet18

m = resnet18()

m1 = torch.fx.symbolic_trace(m)

m2 = torch.fx.Tracer().trace(m)

type(m1) <class ‘torch.fx.graph_module.GraphModule.new..GraphModuleImpl’>

type(m2) <class ‘torch.fx.graph.Graph’>

2
  • torch.fx.symbolic_trace is a wrapper around a call to fx.Tracer.trace to constructs a torch.fx.GraphModule that holds a torch.fx.Graph attribute.
  • torch.fx.Tracer().trace() calls the tracer only to construct only the FX graph:
graph : torch.fx.Graph = torch.fx.Tracer().trace(m)

See also torch.fx — PyTorch 2.0 documentation