pytorch - 💡(How to fix) Fix DISABLED test_batch_invariance_float16_num_splits_1_window_size0_backend_fa2_cuda_float16 (__main__.TestVarlenAttentionCUDA) [1 comments, 2 participants]

TL;DR

• The test test_batch_invariance_float16_num_splits_1_window_size0_backend_fa2_cuda_float16 in test_varlen_attention.py needs to be investigated and fixed to resolve the failure on the main branch.

Guidance

• Review the recent failure examples on torch-ci.com to understand the nature of the test failure.
• Investigate the TestVarlenAttentionCUDA class, specifically the test_batch_invariance_float16_num_splits_1_window_size0_backend_fa2_cuda_float16 test method, to identify potential issues.
• Check the test's dependencies and environment, including the ROCm platform, to ensure they are correctly configured.
• Consider re-enabling the test and running it locally to gather more information about the failure.

Notes

• The issue seems to be specific to the ROCm platform and the test_varlen_attention.py test file.
• Without more information about the test failure, it's difficult to provide a more specific solution.

Recommendation

• Apply workaround: Investigate and fix the failing test to prevent it from causing issues on the main branch, as it is currently disabled due to failures.