Add unit tests for TraceLens.TreePerf

[gabeweisz]

Summary

• Add tests/test_treeperf.py covering GPUEventAnalyser, JaxAnalyses, TreePerfAnalyzer, and JaxTreePerfAnalyzer with synthetic CPU-safe traces.
• Include optional @pytest.mark.gpu live-profile tests (skipped in CPU-only CI via importorskip + CUDA check).

Test plan

• pytest tests/test_treeperf.py -m "not gpu" (48 passed)
• black and ruff pass
• CI unit tests pass

Made with Cursor

After #905

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ All tests successful. No failed tests found.

📢 Thoughts on this report? Let us know!

[ajassani]

test_gpu_event_analyser_compute_metrics_on_live_trace is always skipped in the current unit-test job because PyTorch is not installed, so it contributes no CI coverage. Since the repository already has committed trace-based tests for the analyzer path, can we remove this test? If we want to preserve the current-PyTorch-profiler compatibility check, it would be more useful in a dedicated GPU integration suite that installs PyTorch and actually runs it.

[gabeweisz]

I was adding so it is ready when we actually do have GPUs for CI. But if you want we can remove it - it might interfere with the codecov

[ajassani]

Agreed—let’s keep it for future GPU CI. It shouldn’t materially affect Codecov because TraceLens post-processes profiler output, and the stored-trace tests already cover the same analyzer paths. When GPU CI arrives, we should centralize live-trace generation rather than duplicate profiler setup here; EventReplay already has the profile_resnet trace-generation function that we can reuse or generalize.

[gabeweisz]

Agreed—let’s keep it for future GPU CI. It shouldn’t materially affect Codecov because TraceLens post-processes profiler output, and the stored-trace tests already cover the same analyzer paths. When GPU CI arrives, we should centralize live-trace generation rather than duplicate profiler setup here; EventReplay already has the profile_resnet trace-generation function that we can reuse or generalize.

I already removed it but we can recreate later