[mihncalo]

Summary

Reviewed the original PR from mlcommons/chakra ( mlcommons#217 ) and adressed the comments left on that PR.

Test Plan

Environment used for testing

Python 3.13, torch 2.9.0+cu128, traceinsight 0.6.1, PARAM et_replay, networkx 3.6.1.

How the testing was done

• Updated pyproject.toml to use requires-python 3.10 and ruff target-version py310, to prepare the environment ( these changes will come from the PR for Issue Raise requires-python to >=3.10 mlcommons/chakra#224 )
• Created a clean environment using python 3.13.17, to test with both pandas 2.x and 3.x
• Followed the USER_GUIDE.md
• Ran python -m pytest tests: All 117 tests are passing, 1 warning: DeprecationWarning: Python 3.14 will, by default, filter extracted tar archives and reject files or modify their metadata. Use the filter argument to control this behavior
• Ran the simple_onerank.py and simple_multirank.py scripts in trace_collection/post_execution/ to generate the required traces
• Ran chakra_trace_link using both pandas 2.3.3 and pandas 3.0.6 and collected the logs
• Ran chakra_converter on the resulted linked traces to generate the .et files and collected the logs

Results

The logs show that:

• critical path runs: critical_path_analysis.py:1487 [INFO]: calculating critical path took 0.001714 seconds
• the host and device traces are linked successfully: Linking process successful. Output file is available at ...
• the linked traces are converted into .et files successfully: Conversion successful. Output file is available at ...

Additional Notes

• Pandas 3.x needs at least python 3.11, which means that requires-python will need to be raised in the future.
• A comment on the original PR refers to pinning networkx >=3.6.1. That would require a minimum Python version of 3.11, which requires-python doesn't enforce right now. I decided not to address the comment in this PR.

[theodorbadea]

the pr should target mlcommons/chakra instead of our fork