Worked on enhancing distributed training scalability for vision models by implementing tensor parallelism support for the timm Vision Transformer (ViT) within Deepseek_vl2, part of the bytedance-iaas/vllm repository. Leveraged PyTorch and deep learning techniques to enable efficient multi-GPU training, focusing on optimizing model throughput and resource utilization for large-scale workloads. The approach involved integrating distributed computing strategies to support scalable deployments, addressing the challenges of training high-capacity models across multiple GPUs. This work laid the groundwork for improved performance in vision model training pipelines, emphasizing model optimization and robust engineering practices using Python and advanced parallelization methods.