Contributed to the NVIDIA-NeMo/Automodel repository by expanding Multi-Token Prediction (MTP) to support pre-fused input embeddings for multimodal models. This enhancement allowed continuous audio embeddings to be consumed directly, eliminating the need to re-embed token IDs and reducing preprocessing overhead for multimodal inference. The implementation leveraged Python and PyTorch, focusing on seamless integration with existing SALM and nemotron workflows. Comprehensive unit tests were developed to validate the new embedding path and ensure regression safety, including the resolution of a test-case regression. Code quality was maintained through collaborative reviews and signed-off commits, reflecting a methodical and quality-driven engineering approach.