Worked on stabilizing the test infrastructure for the facebookresearch/param repository by addressing a critical bug affecting CUDA stream handling in Triton kernel trace regeneration. Using Python and CUDA, focused on improving the accuracy of et_replay traces by correcting stream ID management and ensuring the raw CUDA stream was properly captured and propagated within the kernel run function. This targeted debugging and testing effort reduced flaky test runs and enhanced the reliability of automated tests. The work contributed to more robust trace tooling, enabling more accurate kernel execution validation and supporting ongoing development and maintenance of the repository’s testing framework.