Worked on the nvidia-cosmos/cosmos-rl repository to optimize the Qwen3VL model’s visual forward pass and batch synchronization for mixed-modality inference and training. Consolidated image and video feature extraction into a single Vision Transformer (ViT) pass, improving throughput for batches containing both modalities. For pure-text batches, introduced a dummy visual forward to maintain FSDP all-gather synchronization across distributed ranks. Implemented a transformers library version check to ensure compatibility and smoother upgrades. Leveraged Python and deep learning frameworks, focusing on model optimization and transformer architectures to enhance stability and efficiency in distributed, mixed-media machine learning workloads without addressing major bug fixes.