Worked on the deepspeedai/DeepSpeed repository to enhance the stability and scalability of large tensor workloads in October 2025. Addressed a critical bug by updating TensorListMetadata to use int64_t instead of int, enabling support for tensor sizes exceeding INT_MAX and preventing illegal memory accesses in multi_tensor_apply.cuh. This change improved the robustness and correctness of large-scale tensor operations, supporting safer deployment of large models and batch workloads. Demonstrated expertise in C++, CUDA, and memory management by focusing on code safety and maintainability, ensuring that DeepSpeed could reliably handle increasingly large tensors in demanding machine learning environments.