Worked on improving device compatibility and stability for video embedding workflows in the jeejeelee/vllm and vllm-project/vllm-ascend repositories. Addressed two critical bugs by ensuring tensors in the Qwen3-VL model are created on the same device as input video embeddings, preventing runtime errors on non-CPU devices. Enhanced the Efficient Video Sampling feature by adjusting device-compatible tensor operations across multiple vLLM versions, leading to more reliable EVS triggering without user-facing changes. Leveraged deep learning and model optimization expertise using PyTorch and Python, focusing on backend reliability and upstream compatibility for machine learning video processing pipelines.