[mihncalo]

Summary

Internalized et_replay from PARAM instead of adding git as a git dependency. It's no longer necessary to install PARAM separately, as the required code sits inside the chakra repo, thus removing the dependency.

Copied et_replay/ from PARAM into src/third_party/

__init__.py
execution_trace.py
utils.py
LICENSE

Removed unused or unncecessary functions from the execution_trace.py. Kept Node, NodeType, TensorNode, _CommArgs, ExecutionTrace and its schema-specific node constructors, LABEL_MARKERS, the process and thread annotation constants, and remove_dataloader_ops.
Changed import paths in the rest of the src/ files.
Updated pyproject.toml to incorporate the changes.
Removed PARAM from the workflows.

Test Plan

Created a fresh environment by runningpip install ..
Installed dev dependencies by running pip install -r requirements-dev.txt.
Ran the tests by running python -m pytest -vv tests. All 117 tests passed.
Used a sample input to test chakra_trace_link and chakra_converter. They run without issue and create byte-identical outputs as the current main.
Ruff checks passed without issues.

Additional Notes

There is still a lot that could be trimmed from the et_replay/ code, functions and attributes that we do not currently use. I intentionally decided against going any further for now to recreate exactly the behavior before these changes ( when PARAM was installed separately ). Further enhancements could be addressed in following issues.

Closes #237

[github-actions]

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

[theodorbadea]

Please start from looking into the converter parser because most probably that can be reused and extended so that it also works for the linker. Also, it probably would be for the best to reimplement only the logic that is used from execution_trace.py rather than copying the entire file with a lot of extra stuff.