Worked on the pytorch/FBGEMM repository to refactor embedding template code, focusing on simplifying the use of scalar types within code generation templates. The approach involved replacing calls to a scalar_type helper function with direct use of scalar type enums, as the arguments were already known to be scalar types. This change was applied across multiple embedding template files, reducing code complexity and improving maintainability without altering the API. Utilized C++ and template metaprogramming techniques to streamline the codebase, resulting in clearer, more maintainable code paths and lowering the future maintenance burden for the FBGEMM embedding templates.