Worked on the jeejeelee/vllm repository to enable speculative decoding across models with heterogeneous vocabularies. Developed the Token-Level Intersection (TLI) algorithm, which allows draft and target models with different tokenizers to collaborate during decoding. Introduced a configuration option to toggle support for heterogeneous vocabularies and implemented a token translation mapping utility to align tokens between disparate tokenizers. Leveraged Python and PyTorch to integrate these features, focusing on LLM optimization and tokenization challenges. The work broadened deployment scenarios by supporting flexible model composition in speculative decoding workflows, and included collaborative development practices with multiple sign-offs and co-authored contributions.