Worked on stabilizing distributed inference in the vllm-project/vllm-ascend repository by addressing a critical precision bug affecting DeepSeek-V4 token shards when FlashComm was enabled. Focused on improving reliability for MoE inference on Ascend hardware, the work involved resolving anomalies in reduce_scatter and TP AllGather code paths to ensure correct output alignment with upstream MoERunner behavior. Used Python and deep learning frameworks to implement and validate the fix, employing unit testing and end-to-end validation with curl-based tests. This engineering effort enhanced production readiness by reducing inference errors and ensuring traceability through updated release notes and linked upstream commits.