Developed Eagle3 speculative decoding support for the Mistral3ForConditionalGeneration model in the jeejeelee/vllm repository, focusing on enhancing multi-modal input processing and generation speed. The implementation, completed in Python, leveraged deep learning and model optimization techniques to reduce inference latency for multi-modal prompts. This work involved close collaboration with cross-team contributors, including thorough code reviews and co-authored commits, ensuring robust integration and platform readiness for downstream applications. By laying the foundation for future performance benchmarking and expanded multi-modal capabilities, the contribution addressed a key feature request and improved the efficiency of conditional generation workflows within the repository’s ecosystem.