Limitations

Torch tracing records the operations that dispatch through TracedTensor. The cases below are where the recorded graph cannot follow what the Python code does.

Data-dependent control flow

An exported graph is a static structure and cannot hold a branch that depends on tensor values. Using a traced tensor in a boolean context records only the branch taken during tracing and logs an error:

Attempted to use TracedTensor 'x' from node 'node' in a boolean context
(if/while/and/or/not) during tracing.
...
The traced graph is a static DAG and cannot represent dynamic branches.

Use torch.where() for element-wise choices, fixed iteration counts instead of value-driven loops, or split the node around the decision so the branch becomes the caller’s problem rather than the graph’s.

Extracting Python values

.item(), .tolist(), float(), and int() return ordinary Python values captured at trace time. They leave the graph, and anything computed from them afterwards is frozen at its traced value.

Torch gives no warning when this happens, unlike the NumPy path. A scalar pulled out mid-node and multiplied back in looks completely normal and exports a wrong model, so treat scalar extraction as a deliberate trace break.