PyTea, a PyTorch tensor shape error analyzer, is designed to statically scan PyTorch applications and detect possible shape errors by analyzing execution paths and tensor shape constraints. However, PyTea may struggle with certain PyTorch APIs if they are not fully implemented or supported within its framework.
PyTea's analysis can raise false alarms if the code contains PyTorch or other third-party APIs that have not been implemented. This means that if PyTea encounters an unimplemented API call, it may incorrectly flag it as an error or issue a warning. The tool records each unimplemented API call in its logs, allowing users to identify which specific APIs are not supported[10].
Additionally, PyTea's effectiveness can be limited by the complexity of the code it analyzes. For instance, if the code involves complex operations or non-linear constraints that are difficult for the SMT solver (Z3) to handle, PyTea might classify these cases as "undecidable" and timeout[10]. However, specific PyTorch APIs that PyTea struggles with are not explicitly listed, but any unimplemented or overly complex operations could pose challenges.
Citations:[1] https://arxiv.org/pdf/2112.09037.pdf
[2] https://stackoverflow.com/questions/59158044/installing-a-specific-pytorch-build-f-e-cpu-only-with-poetry
[3] https://pytorch.org/docs/stable/func.ux_limitations.html
[4] https://pytorch.org/docs/stable/torch.compiler_fine_grain_apis.html
[5] https://pytorch.org/functorch/stable/ux_limitations.html
[6] https://discuss.pytorch.org/t/survey-what-are-you-using-the-c-api-for/55163
[7] https://docs.habana.ai/en/latest/PyTorch_CustomOp_API/page_index.html
[8] https://discuss.pytorch.org/t/is-there-an-elegant-way-to-calculate-integral-image-using-pytorch-api/22669
[9] https://www.altexsoft.com/blog/pytorch-library/
[10] https://github.com/ropas/pytea