Worked on enhancing stability and compatibility for dynamic shape workloads in the intel/neural-compressor repository, focusing on the Conv2d FP8 path to support PyTorch 2.5. Addressed a persistent shape-mismatch issue by ensuring that stride, padding, and dilation parameters for conv2d_fp8 are consistently represented as lists, introducing a helper function to normalize these parameters. This approach improved reliability for dynamic input workloads and streamlined deployment pipelines. The work centralized parameter normalization logic, contributing to maintainability and future-proofing of convolution operations. Utilized Python and deep learning expertise, with an emphasis on performance optimization and robust support for evolving PyTorch features.