pytorch runtime_error ai_generated true

RuntimeError: TorchScript supports tracing only for tensors. This function cannot be traced because it uses dynamic control flow (e.g., if/else or loops).

ID: pytorch/torchscript-dynamic-control-flow

Also available as: JSON · Markdown · 中文

90%Fix Rate

88%Confidence

1Evidence

2023-02-10First Seen

Version Compatibility

Version	Status	Introduced	Deprecated	Notes
torch 1.9.0	active	—	—	—
torch 2.0.0	active	—	—	—
python 3.9	active	—	—	—

Root Cause

TorchScript tracing (torch.jit.trace) cannot handle Python control flow that depends on tensor values; it records operations on specific inputs and fails when the path changes.

generic

中文

TorchScript 跟踪（torch.jit.trace）无法处理依赖于张量值的 Python 控制流；它记录特定输入上的操作，并在路径更改时失败。

Official Documentation

https://pytorch.org/docs/stable/jit.html#mixing-tracing-and-scripting

Workarounds

90% success Use torch.jit.script instead of trace for models with dynamic control flow: @torch.jit.script def forward(self, x): if x.sum() > 0: return x * 2 else: return x / 2 # Then script the model: scripted_model = torch.jit.script(model)
```
Use torch.jit.script instead of trace for models with dynamic control flow:
@torch.jit.script
def forward(self, x):
    if x.sum() > 0:
        return x * 2
    else:
        return x / 2
# Then script the model: scripted_model = torch.jit.script(model)
```
85% success Refactor the model to avoid dynamic control flow in the traced function, e.g., use torch.where or masked operations: def forward(self, x): mask = (x.sum(dim=1, keepdim=True) > 0).float() return x * (1 + mask) # equivalent to if-else
```
Refactor the model to avoid dynamic control flow in the traced function, e.g., use torch.where or masked operations:
def forward(self, x):
    mask = (x.sum(dim=1, keepdim=True) > 0).float()
    return x * (1 + mask)  # equivalent to if-else
```

中文步骤

Use torch.jit.script instead of trace for models with dynamic control flow:
@torch.jit.script
def forward(self, x):
    if x.sum() > 0:
        return x * 2
    else:
        return x / 2
# Then script the model: scripted_model = torch.jit.script(model)

Refactor the model to avoid dynamic control flow in the traced function, e.g., use torch.where or masked operations:
def forward(self, x):
    mask = (x.sum(dim=1, keepdim=True) > 0).float()
    return x * (1 + mask)  # equivalent to if-else

Dead Ends

Common approaches that don't work:

80% fail
Using torch.jit.trace with a large set of example inputs may still fail if control flow depends on runtime values not present in examples.
70% fail
Adding more concrete inputs to trace may cover some paths but not all, leading to incomplete traces.