CI: failing slow runs fails to report properly

[tarekziade]

The following issue description was a red herring, the real problem is described at #44238 (comment)

Slow runs try to cat captured_info.txt which seems to not be always present

#43972 (comment)

the GH step that update the report fails after the runs finish.

I reproduced the issue in SSH with

PATCH_TESTING_METHODS_TO_COLLECT_OUTPUTS=yes _PATCHED_TESTING_METHODS_OUTPUT_DIR=. pytest -sxv tests/models/qwen3_vl/test_modeling_qwen3_vl.py::Qwen3VLModelTest::test_multi_gpu_data_parallel_forward --make-reports reports

and I can see that it does not hit the function that creates the file, most likely because it does not go through a patched assertion = here its StopIteration

maybe we could simply touch the file when the tests starts, or make the step more resilient in the absence of it.

[tarekziade]

cc @ydshieh

[tarekziade]

That error was a red herring, the real issue seems to be in check_bad_commit.py

It ran for 4 hours (12:29:27 → 16:29:28) , running git bisect — then the ls in "Show results" failed with exit code 2 because the output file was never written.

I think the root cause is a bug in check_bad_commit.py at line 234:

return None, "git bisect failed"   # ← returns a TUPLE, not a dict!

But the caller at line 360 does:

info.update(bad_commit_info)  # bad_commit_info = (None, "git bisect failed")

dict.update() tries to iterate over (None, "git bisect failed") as (key, value) pairs. The first element is None, which can't be unpacked → TypeError: cannot unpack non-iterable NoneType object.

The script crashes, never writes the output file, the ls fails with exit code 2, and the whole chain collapses to "Model CI failed to report results".

maybe we could fix with

 231      if "error: bisect run failed" in bash_result.stderr:
 232          error_msg = f"Error when running git bisect:\nbash error: {bash_result.stderr}\nbash output:\n{bash_result.stdout}\nset `bad_commit` to `None`."
 233          print(error_msg)
 234 -        return None, "git bisect failed"
 234 +        result["status"] = "git bisect failed"
 235 +        return result
 236
 237      pattern = r"(.+) is the first bad commit"
 238      commits = re

I think this was a miss in this PR,

0dda1ff

it converted all return statements in find_bad_commit from tuples to dicts but missed that occurence