Worked on stabilizing GPU memory usage and enhancing cross-architecture compatibility in the fla-org/flash-linear-attention repository, focusing on large-model training with limited shared memory on AMD RDNA GPUs. Addressed a gating bug by correcting CONST_TILING behavior and implemented shared memory guards along with autotuning safeguards for both forward and backward passes. Developed architecture-aware tiling logic to prevent invalid configurations across different GPU platforms, including RDNA, ADA, and Ampere/Hopper. Validated these improvements on RDNA4 hardware during Qwen3-Next-80B-A3B-Instruct training, resulting in reduced compilation and runtime failures. Utilized Python, deep learning frameworks, and GPU programming for robust performance optimization.