Developed large-scale Tensor Parallelism for the tplr-ai/templar repository, enabling efficient multi-GPU distribution of large machine learning models. Leveraging Python, PyTorch, and advanced GPU programming, the implementation introduced DTensor-based model parallelization with configurable TP_DEGREE and support for mixed TP and FSDP configurations. The work addressed deadlocks and barrier synchronization, improved distributed communication, and enhanced gradient logging. Multi-dimensional parallelism modes such as dp_replicate, dp_shard, tp, pp, and cp were supported, all configurable via environment variables. The solution achieved substantial throughput and training speed improvements, validated at scale with robust gradient aggregation and no regressions in FSDP training.