Worked on stabilizing dynamic shape exports in the pytorch/tensordict repository by addressing failures related to batch size parsing when scalar SymInt values were present. Focused on backend development using Python, the work involved refining the export path to correctly interpret batch dimensions and preserve shape metadata within the pytree context. This targeted bug fix improved the reliability of model workflows that depend on TensorDict for both training and inference, particularly in scenarios involving dynamic tensor shapes. Emphasized robust unit testing to ensure that pipelines relying on dynamic exports would function smoothly, reducing the likelihood of export-time errors in production environments.