Developed and integrated a Torch Inference Profiling Feature for the ModelTC/lightllm repository, focusing on enhancing observability and performance optimization in model inference workflows. Leveraging Python and PyTorch, the work introduced a torch_profile utility that wraps profiling logic and embeds it within the tppart_model_infer pipeline, covering both prefill and decode stages. This approach enabled comprehensive end-to-end profiling of forward-pass latency and resource usage. To ensure reliability and prevent regressions, dedicated test coverage was added for the profiling tooling. The resulting instrumentation supports data-driven optimization, allowing teams to diagnose latency hotspots and make informed deployment decisions for inference workloads.