How text diffusion models work, why they swap the memory-bound decode loop for a compute-bound one, and what the papers behind Mercury, LLaDA2.x and DiffusionGemma actually show.
TTFT, prefill vs decode, the KV-cache memory bill, quantization, multi-LoRA serving, and how vLLM, SGLang, llama.cpp and MLX run them across NVIDIA, AMD and Apple silicon.
Our IJETS 2024 paper — combining LLaVA-Med with Retrieval Augmented Generation to build a clinically grounded visual question answering assistant for CT and MRI imaging.