Worked on the ROCm/rocMLIR repository to improve the stability and reliability of RDNA2 architecture support, focusing on backend development using C++ and MLIR. Addressed a bug in the AmdArchDb by correcting the minimum number of compute units (minNumCU) for RDNA2 GPUs, ensuring accurate hardware modeling. Enhanced test infrastructure by decoupling test expectations from architecture preset defaults, pinning specific compute unit values in test runs to reduce fragility. These changes improved the predictability of split-k selection-likelihood tests and contributed to more robust deployment for downstream users relying on ROCm/rocMLIR across various AMD GPU architectures. No new features were added.