pytorch runtime_error ai_generated true

RuntimeError: FX tracing failed: 'NoneType' object has no attribute 'shape' during symbolic tracing

ID: pytorch/fx-graph-tracing-non-tensor

Also available as: JSON · Markdown · 中文
78%Fix Rate
83%Confidence
1Evidence
2023-11-20First Seen

Version Compatibility

VersionStatusIntroducedDeprecatedNotes
torch>=1.8.0 active — — —
torch>=2.0.0 active — — —

Root Cause

During torch.fx symbolic tracing, a function or method returned None instead of a tensor, causing the tracer to fail when accessing .shape attribute.

generic

中文

在torch.fx符号追踪期间，函数或方法返回了None而不是张量，导致追踪器在访问.shape属性时失败。

Official Documentation

https://pytorch.org/docs/stable/fx.html#limitations

Workarounds

  1. 85% success Ensure all functions return tensors, use torch.zeros(1) as placeholder for None cases: if x is None: return torch.zeros(1) 
    Ensure all functions return tensors, use torch.zeros(1) as placeholder for None cases: if x is None: return torch.zeros(1)
  2. 75% success Decorate the problematic function with @torch.fx.wrap to skip tracing that part 
    Decorate the problematic function with @torch.fx.wrap to skip tracing that part
  3. 70% success Use torch.jit.trace instead of torch.fx.symbolic_trace for models with dynamic control flow 
    Use torch.jit.trace instead of torch.fx.symbolic_trace for models with dynamic control flow

中文步骤

  1. Ensure all functions return tensors, use torch.zeros(1) as placeholder for None cases: if x is None: return torch.zeros(1)
  2. Decorate the problematic function with @torch.fx.wrap to skip tracing that part
  3. Use torch.jit.trace instead of torch.fx.symbolic_trace for models with dynamic control flow

Dead Ends

Common approaches that don't work:

  1. 90% fail

    FX tracer executes the code symbolically; exceptions are propagated as tracer errors.

  2. 95% fail

    This flag doesn't exist; FX always traces shapes.