edge-dit.cpp
A native C/C++ inference engine, built on ggml, for running diffusion transformer image and video generation models locally on resource-constrained and edge devices.
edge-dit.cpp provides one runtime for text-to-image, image editing and video generation across major diffusion-transformer model families, including FLUX.1, FLUX.2 klein, Stable Diffusion 3 and 3.5, Qwen-Image and Wan, loading Diffusers directories, safetensors or GGUF files. It runs on CUDA, CPU, Vulkan or an experimental Metal backend with no Python or PyTorch needed at runtime.
An auto-fit flag picks a quantization level and per-component placement automatically to fit a model into limited VRAM, and a layered-offload mode streams the diffusion transformer one block at a time so that a model of 20 gigabytes or more can run on a 24 gigabyte card or smaller. A separate conversion tool produces a portable pre-quantized GGUF file offline, and few-step distilled models are auto-detected and given a shortened generation schedule.