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

Fix Plan

The fix involves re-enabling and updating the test_index_add_fallback test in test_cuda_repro.py to ensure it works correctly on the main branch.

Steps to Fix

• Re-enable the test by removing the decorator or comment that disabled it.
• Update the test to handle any changes in the CUDA reproducibility logic.
• Verify the test passes on the main branch.

Example Code

# test_cuda_repro.py
import unittest
import torch

class CudaReproTests(unittest.TestCase):
    # Re-enable the test
    def test_index_add_fallback(self):
        # Update the test to handle changes in CUDA reproducibility logic
        # For example, update the tensor creation to use the latest CUDA version
        tensor = torch.randn(10, device='cuda')
        # ... rest of the test code ...

Verification

Run the test using the command python test_cuda_repro.py and verify it passes without any errors.

Extra Tips

• Ensure the CUDA version is up-to-date and compatible with the PyTorch version.
• Run the test on multiple platforms to ensure reproducibility.
• Use torch.cuda.seed() and torch.cuda.manual_seed() to ensure reproducibility in CUDA operations.