Worked on the pytorch/ao repository to address a critical bug affecting asynchronous CPU-to-GPU tensor transfers in TorchAOBaseTensor. Focused on backend development and performance optimization, the work restored proper non_blocking semantics, enabling overlap between data copy and compute operations in async offload workflows. This fix resolved real-world performance degradation in production diffusers integrations by ensuring the non_blocking argument was correctly propagated during tensor operations. The solution included comprehensive unit testing across multiple scenarios and cross-platform validation, including ROCm AMD GPUs and PyTorch 2.7. All changes maintained backward compatibility and improved throughput for GPU-accelerated workloads using Python.