pytorch - 💡(How to fix) Fix DISABLED test_mempool_no_split (__main__.TestMemPool) [2 comments, 1 participants]

Traceback (most recent call last): File "/var/lib/jenkins/workspace/test/test_cuda.py", line 6441, in test_mempool_no_split raise AssertionError( AssertionError: Expected no_split pool to have more segments, but got 1 vs 1

To execute this test, run the following from the base repo dir: python test/test_cuda.py TestMemPool.test_mempool_no_split

This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0

TL;DR

The most likely fix involves investigating and addressing the inconsistent behavior of the test_mempool_no_split test in the test_cuda.py file, potentially related to the expected number of segments in the no-split pool.

Guidance

• Investigate the test logs by following the provided debugging instructions to understand the context of the failures and successes.
• Run the test locally using the command python test/test_cuda.py TestMemPool.test_mempool_no_split to reproduce the issue and gather more information.
• Review the test case test_mempool_no_split in test_cuda.py to understand the assertion that's failing and why it's expecting more segments.
• Check the configuration and environment variables, such as PYTORCH_PRINT_REPRO_ON_FAILURE, to ensure they are not influencing the test behavior.

Example

No specific code snippet can be provided without further investigation, but the test case in question is test_mempool_no_split in test_cuda.py, which raises an AssertionError due to an unexpected number of segments.

Notes

The issue seems to be related to the flaky behavior of the test, which has both passed and failed in different workflow runs. The root cause might be related to the test's expectations or the environment in which it's run.

Recommendation

Apply workaround: Temporarily enable the test and run it multiple times to gather more logs and understand the pattern of failures and successes, which might help in identifying the root cause.