Worked on the jeejeelee/vllm repository to enhance distributed system reliability by addressing a bug in the MNNVL backend’s all-reduce operation. Refactored the one-shot selection logic for FlashInfer all-reduce into a dedicated function, improving maintainability and testing. Introduced an MNNVL-specific bypass to resolve incorrect workspace size constraints, which reduced runtime errors and improved robustness during multi-backend training and inference. Leveraged Python and PyTorch to strengthen backend development, focusing on distributed communication paths. Demonstrated collaborative development practices through clear commit sign-offs and co-authorship, contributing to safer scaling and more stable distributed training workflows across multiple backends.