Traceback (most recent call last): File "/var/lib/jenkins/workspace/test/export/test_export.py", line 1334, in test_dynamic_lstm self.assertEqual(eager_out_gru, ep_out_gru) ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/_dynamo/test_case.py", line 113, in assertEqual return super().assertEqual(x, y, *args, **kwargs) ~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/conda/envs/py_3.14/lib/python3.14/site-packages/torch/testing/_internal/common_utils.py", line 4365, in assertEqual raise error_metas.pop()[0].to_error( # type: ignore[index] ...<4 lines>... ) AssertionError: Tensor-likes are not close!

Mismatched elements: 563 / 262144 (0.2%) Greatest absolute difference: 3.3736228942871094e-05 at index (0, 8, 370) (up to 1e-05 allowed) Greatest relative difference: 0.13132105767726898 at index (0, 11, 149) (up to 1.3e-06 allowed)

To execute this test, run the following from the base repo dir: PYTORCH_TEST_WITH_CROSSREF=1 python test/export/test_export.py TestExport.test_dynamic_lstm

This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0

Fix Plan

Step 1: Identify the root cause of the flaky test

The flaky test is likely due to the assertion error in the test_dynamic_lstm function. The error message indicates that the tensor-likes are not close, with a mismatched element count and greatest absolute difference.

Step 2: Update the test to handle floating-point precision issues

To fix the flaky test, we need to update the test to handle floating-point precision issues. We can use the torch.testing.assert_close function instead of self.assertEqual to compare the tensor-likes.

import torch

# ...

def test_dynamic_lstm(self):
    # ...
    eager_out_gru = ...  # compute eager_out_gru
    ep_out_gru = ...  # compute ep_out_gru
    torch.testing.assert_close(eager_out_gru, ep_out_gru, atol=1e-5, rtol=1e-6)
    # ...

Step 3: Update the test to handle tensor-likes with different shapes

If the tensor-likes have different shapes, we need to update the test to handle this case. We can use the torch.testing.assert_close function with the rtol and atol arguments to specify the relative and absolute tolerance, respectively.

import torch

# ...

def test_dynamic_lstm(self):
    # ...
    eager_out_gru = ...  # compute eager_out_gru
    ep_out_gru = ...  # compute ep_out_gru
    torch.testing.assert_close(eager_out_gru, ep_out_gru, atol=1e-5, rtol=1e-6, equal_nan=True)
    # ...

Step 4: Run the test with the updated code

Run the test with the updated code to verify that the fix works.

PYTORCH_TEST_WITH_CROSSREF