Worked on stabilizing distributed deep learning workflows by addressing critical bugs in two major repositories. In kaiyux/TensorRT-LLM, resolved expert-mode routing issues in the WeightOnlyQuantRowLinear module by introducing the missing is_expert parameter, improving reliability in quantized inference across distributed systems. In JustinTong0323/sglang, corrected argument passing for the fused MoE Triton benchmark, eliminating runtime errors and ensuring accurate benchmarking results. Focused on Python-based solutions, leveraging skills in model quantization, benchmarking, and performance optimization. The work emphasized targeted bug fixes, robust testing, and maintaining code quality, resulting in more stable and dependable inference and benchmarking pipelines for downstream users.