torch-npugraph-ex-compile-error-diagnosis

A graph break means TorchDynamo hit Python code it cannot compile and fell back to eager execution. Enable TORCH_LOGS=+dynamo,+graph_breaks or TORCH_COMPILE_DEBUG=1 to locate the exact break point, then check for unsupported control flow, dynamic data structures, or unsupported operations.

BackendCompilerFailed is usually a wrapper exception, not the root cause. Search the full traceback for the first concrete error, such as a Meta derivation failure, an unsupported operator during npu_fx_compiler passes, or an ACL graph capture failure, and diagnose from that point.

Meta derivation fails when a custom operator lacks a register_fake implementation or when its shape and dtype inference logic disagrees with the real computation. Check the operator's FakeTensor registration in torch_npu and align the Meta function with actual kernel behavior.

Compile-time errors occur before the first successful execution, during Dynamo tracing, AOT compilation, backend compilation, or graph capture. If the graph captured successfully and fails during replay, that is a runtime issue involving kernels, communication, or memory, and needs a different diagnostic path.

Verify in three stages: run the model in Eager mode first, then torch.compile with backend aot_eager, then with backend npugraph_ex. Eager failures point to model or operator issues, aot_eager failures point to Dynamo and AOT layers, and only npugraph_ex-only failures implicate the backend or graph capture.

No. Start by reproducing the failure and using debug_save to capture the minimal compilation artifacts. Only add data_dump or TORCH_LOGS options like +dynamo and +aot when initial evidence is insufficient, since enabling everything creates noise that hides the first causal error.