pytorch - 💡(How to fix) Fix DISABLED test_dtensor_requires_grad_intermediate_backward (__main__.TestDTensorCompileWithCompiledAutograd) [2 comments, 3 participants]

Fix Plan

The fix involves modifying the test to handle the new dynamo graph break due to the test's backward using generators.

Steps to Fix

• Update the test_dtensor_requires_grad_intermediate_backward test in test_compiled_autograd.py to properly handle the generator in the backward pass.
• Use a torch.dynamo.reset context to ensure the dynamo graph is reset after each test.

Example Code

import torch
import torch.dynamo

class TestDTensorCompileWithCompiledAutograd:
    def test_dtensor_requires_grad_intermediate_backward(self):
        with torch.dynamo.reset():
            # existing test code here
            pass

Verification

Run the test again to verify that it passes without errors.

Extra Tips

• Make sure to update the test to handle any potential edge cases with the new dynamo graph break.
• Consider adding additional tests to ensure the fix is robust and doesn't introduce any regressions.