Troubleshoot common integration issues

Symptom	Likely cause	Action
torch.cuda.is_available() is false	CPU-only PyTorch, incompatible driver or unavailable device	Use the official PyTorch selector and confirm device access before importing the GPU API
CUDA input validation fails	Shape, dtype or device differs from preparation	Reuse the example shape and dtype or create a new executor
Nonfinite output raises ArithmeticError	Operator growth, coefficient size or arithmetic overflow	Check normalization and input magnitude; validate in FP64 and compare a small independent case
compile_ppr rejects the graph	Stored operator is not exactly symmetric or the resolvent enclosure fails	Check semantics and normalization; do not silently symmetrize a directed task
A parameter endpoint is rejected	Binary float lies outside an exact compiled interval	Use a consistent rational endpoint or an interior parameter value
Approximation is no faster	Full degree retained or launch/transfer overhead dominates	Inspect selected degree, bound method, preparation and complete latency
Captured execution is slower	Small task, capture overhead or workspace pressure	Measure eager capture=False and unpacked execution
Local propagation expands widely	The source reaches much of the graph within the degree	Account for graph locality; the backend can fall back to global evaluation
CPU tests pass but GPU tests skip	No CUDA device is accessible	Report the skips and run the GPU tests on a real supported device

Report a reproducible issue

Include the package version or commit, Python/library versions, graph dimensions and stored entry count, selected degree, dtype, settings, minimal nonprivate input, expected result and actual error. For performance issues, include all synchronized samples and full timing boundaries. Do not include credentials, private customer data or system paths containing personal information.

Use GitHub issues for ordinary bugs. Follow SECURITY.md for security-sensitive reports.