[razvcris]

Summary

Adds a script and CI job that prove a fresh install of Chakra works on every supported Python. Addresses #227.

scripts/check_clean_install.sh:

1. Builds the wheel (python -m build) from a copy of the git-tracked files.
2. Installs the wheel and PARAM et_replay in a fresh venv.
3. Runs the six tools with --help.
4. Imports chakra.schema.protobuf.et_def_pb2 and storage_pb2 in an isolated interpreter and checks that both .proto files are packaged.
5. Runs chakra_trace_link and chakra_converter on the fixture. For each tool it checks: exit code 0, the output exists and decodes (linked JSON parses; the .et reads with protolib and has more than 0 nodes), and no ERROR/CRITICAL line or Traceback in stdout, stderr or debug.log. It also fails on "Critical path analysis encountered an invalid graph structure". Warning lines are counted and reported, not fatal.
6. If a run fails, the script copies its logs, trace outputs and inputs (linked JSON, .et, debug.json and the input traces) to FAILURE_ARTIFACTS_DIR, and the workflow uploads them as an artifact (clean-install-<python version>), so a failure on one Python version can be investigated without rerunning it.

The script runs standalone (e.g. PYTHON=python3.12 scripts/check_clean_install.sh) and in CI (.github/workflows/clean_install_tests.yml).

Fixture

tests/data/simple_onerank_traces.tar.gz contains host.0.json and device.0.json generated by trace_collection/post_execution/simple_onerank.py on a CUDA machine.

Python versions

The workflow runs the script on Python 3.11, 3.12 and 3.13 (requires-python >= 3.11).

Test Plan

Run locally in a clean environment, as the workflow does (PYTHON=python scripts/check_clean_install.sh):

• Python 3.11.18 - passed (592 nodes in the .et, 10 WARNING lines)
• Python 3.12.16 - passed (592 nodes in the .et, 10 WARNING lines)
• Python 3.13.18 - passed (592 nodes in the .et, 10 WARNING lines)

The first real run on ubuntu-latest is this PR's CI.

Additional Notes

• GPU-collected traces only. I also tried a trace set collected on a CPU-only run. chakra_trace_link fails on it: load_sync_dependencies calls HTA/traceinsight critical_path_analysis, which expects GPU kernels and fails on a trace with none.
• The fixture has no cuda_sync events (the profiler disables them by default).
• closes clean-install check #227

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅

[ajbalogh]

A failed run discards all logs.

• Why it matters: The script deletes ${WORK} on exit through the trap. A CI failure on one Python version leaves only the console replay. There is no debug.log, linked trace or .et output to inspect.
• Suggested Fix: Keep the work directory on failure and upload it.

[ajbalogh]

The workflow does not restrict the GITHUB_TOKEN.

• Why it matters: The job runs third-party code from a pinned git dependency. Default token scopes can be broader than the job needs. CodeQL flags a missing permissions block.

[HanjiangWu313]

Tested locally on both standalone and with current main; all clean-install checks passed—looks good to me.

[theodorbadea]

This seems to be missing versions 3.14 and 3.15, any particular reason?

[theodorbadea]

should add 3.14 and 3.15