The intent with the project is to explore Discrete Diffusion/Flow matching paradigms for LLMs in order to improve text generation speed during inference. The request is made in order to scale existing working prototypes of the models by utilizing multiple GPUs, first with torch.distributed, and later with NVSHMEM. The latter goal is to experiment with various losses, masking/noise schemes in order to figure out what works well with dLLMs at scale and why. Main Supervisor: Michael Doggett Associate Professor Lund University