So clang-tidy warns about: out_labels[ndim++] = label; (line 1946) and i = out_label - output_labels; (line 2055) and axes[i] = match - labels; (line 2232) (narrowing casts). I have no idea if those are actually particularly problematic, but thought I would mention.

So clang-tidy warns about: out_labels[ndim++] = label; (line 1946) and i = out_label - output_labels; (line 2055) and axes[i] = match - labels; (line 2232) (narrowing casts). I have no idea if those are actually particularly problematic, but thought I would mention.

I will explore this tomorrow. Thank you for all your help!

Every fix should have a test case. I am not sure we need 4d4cc4d, can you add a test that fails before and passes after to prove it is required? The fix in 68861de is definitely needed for np.einsum('abab', x).

The test should be added to numpy/core/tests/test_einsum.py, maybe next to test_einsum_fixed_collapsingbug. Note that random.rand(10, 10, 10, 10) should be spelled random.random_sample((10, 10, 10, 10)). The test should mention the issues gh-14692 and gh-12689.

The test should be added to numpy/core/tests/test_einsum.py, maybe next to test_einsum_fixed_collapsingbug. Note that random.rand(10, 10, 10, 10) should be spelled random.random_sample((10, 10, 10, 10)). The test should mention the issues gh-14692 and gh-12689.

If it passes testing, I'll add that, thanks, @mattip

Note that random.rand(10, 10, 10, 10) should be spelled random.random_sample((10, 10, 10, 10)).

Hmm, this does not work:

tensor = np.random.random_sample(10, 10, 10, 10)
  File "mtrand.pyx", line 370, in numpy.random.mtrand.RandomState.random_sample
TypeError: random_sample() takes at most 1 positional argument (4 given)

Are you sure you used my code? It looks different to me.

The documented developer workflow is to add tests as part of the development process. I find it helpful to:

• reproduce a failure locally from the CLI with python runtests.py --python which will compile and use a HEAD version of numpy
• add a test that fails when run locally with python runtests.py -t path/to/changed-file
• add a fix and verify that the test suite runs locally with python runtests.py
• push the test and the fix together.

This makes sure your fix is correct and prevents unneeded CI runs. In this case, I needed to dive in with gdb to find out exactly what was going wrong so a test was critical to find the missing cast.

Are you sure you used my code? It looks different to me.

aha :) Lots of parentheses

w/r/t the test case, does it have to test some kind of result, or is it enough that the failure is that an exception is thrown?

Thanks @jwoehr