Worked on tracel-ai/cubecl to enhance GPU memory management by introducing a new allocation strategy supporting both asynchronous and synchronous allocations. Leveraged Rust and CUDA to extend GpuStorage, enabling it to handle multiple allocation types and ensuring correct deallocation based on the specific allocation method. Addressed a CUDA-version-specific issue where asynchronous memory was unavailable in vGPU environments, resolving a regression that affected cross-version compatibility. This work improved GPU resource utilization, reduced memory fragmentation, and increased workload stability across different CUDA variants. The engineering focused on robust resource management and compatibility, demonstrating depth in GPU programming and systems-level Rust development.