pytorch - 💡(How to fix) Fix DISABLED test_hessian (__main__.FuncTorchHigherOrderOpTestsWithCompiledAutograd)

Traceback (most recent call last): File "/var/lib/jenkins/workspace/test/dynamo/test_higher_order_ops.py", line 3803, in test_hessian wrapped_gm = self._compile_check(wrapper_fn, (x,)) File "/var/lib/jenkins/workspace/test/dynamo/test_higher_order_ops.py", line 3789, in _compile_check expected = torch.compile(fn, backend=backend, fullgraph=fullgraph)(*inputs) File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/eval_frame.py", line 1051, in compile_wrapper raise e.with_traceback(None) from e.cause # User compiler error torch._dynamo.exc.Unsupported: Attempted to call a super() attribute that is not a function or method Explanation: Dynamo does not know how to trace the call super().unflatten() because super().unflatten is not a function or method attribute. Hint: Ensure the attribute accessed via super() is a standard method or function.

Developer debug context: call_method SuperVariable() unflatten

For more details about this graph break, please visit: https://meta-pytorch.github.io/compile-graph-break-site/gb/gb0006.html

from user code: File "/var/lib/jenkins/workspace/test/dynamo/test_higher_order_ops.py", line 3800, in wrapper_fn return torch.func.hessian(torch.sin)(x) File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_functorch/eager_transforms.py", line 1401, in wrapper_fn jac_outs_ins = tuple( File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_functorch/eager_transforms.py", line 1402, in <genexpr> tuple( File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_functorch/eager_transforms.py", line 1403, in <genexpr> safe_unflatten(jac_out_in, -1, primal.shape) File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_functorch/eager_transforms.py", line 1247, in safe_unflatten return tensor.unflatten(dim, shape) File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/utils/_device.py", line 122, in torch_function return func(*args, **kwargs) File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_tensor.py", line 1456, in unflatten return super().unflatten(dim, sizes)

Set TORCHDYNAMO_VERBOSE=1 for the internal stack trace (please do this especially if you're reporting a bug to PyTorch). For even more developer context, set TORCH_LOGS="+dynamo"

To execute this test, run the following from the base repo dir: python test/dynamo/test_higher_order_ops.py FuncTorchHigherOrderOpTestsWithCompiledAutograd.test_hessian

This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0

TL;DR

The test failure is likely due to an issue with PyTorch's Dynamo compiler, specifically with the unflatten method, and can be investigated by setting TORCHDYNAMO_VERBOSE=1 and examining the internal stack trace.

Guidance

• Set TORCHDYNAMO_VERBOSE=1 to enable verbose logging and obtain the internal stack trace, which may provide more insight into the issue.
• Examine the logs for the test_hessian test case, following the provided debugging instructions, to identify any patterns or clues that may indicate the root cause of the failure.
• Investigate the unflatten method in the PyTorch codebase, specifically the torch/_tensor.py file, to understand how it is being used and why it may be causing issues.
• Consider running the test locally using the provided command python test/dynamo/test_higher_order_ops.py FuncTorchHigherOrderOpTestsWithCompiledAutograd.test_hessian to reproduce the issue and gather more information.

Example

No code example is provided, as the issue is related to a specific PyTorch test case and requires a deeper understanding of the PyTorch codebase.

Notes

The issue appears to be related to a flaky test, which may be caused by a variety of factors, including issues with the test itself, the PyTorch codebase, or the testing environment. Further investigation is needed to determine the root cause of the failure.

Recommendation

Apply workaround by setting TORCHDYNAMO_VERBOSE=1 and examining the internal stack trace to gather more information about the issue, as this may provide valuable insights into the root cause of the failure.