Worked on enhancing the stability and robustness of the Ascend sampling path in the vllm-ascend repository, focusing on deep learning and distributed systems using Python and PyTorch. Addressed a critical bug in the reduce sampling logic that previously allowed out-of-bounds errors when top_k and top_p parameters were unset, ensuring the sampler now safely returns gathered logits and a global identity index for unfiltered cases. Refactored the Ascend-specific sampling path to improve error handling and maintainability, reducing production risk. These improvements increased reliability and throughput for large-scale deployments on Ascend hardware, directly supporting safer and more robust model inference.