Hey everyone!
I’ve been watching Wendell’s videos on Level1Techs for a long time and really appreciate how supportive this community is of open-source projects, so I wanted to share something I’ve been building and refining: SpawnDev.ILGPU.ML.
It is a neural network inference and training engine written entirely in C# that compiles C# compute kernels directly for whatever device is in front of it via SpawnDev.ILGPU.
The big kicker: No ONNX Runtime, no native inference binaries, and no server required. It runs entirely locally on the desktop (Windows, Linux, macOS via CUDA, OpenCL, or SIMD CPU) or right inside the browser via Blazor WebAssembly (WebGPU, WebGL, or WebAssembly).
One Codebase, Six Backends: WebGPU, WebGL, Wasm, CUDA, OpenCL, and SIMD CPU. Everything is tested across lanes.
Zero-Copy Architecture: Pixels, activations, and weights stay on the GPU. In the browser, data streams from fetch and OPFS straight into GPU buffers without bloating the .NET heap.
Model Formats Auto-Detected: Drop in ONNX, TFLite, GGUF, SafeTensors, TF GraphDef, PyTorch checkpoints, or Core ML. It builds a unified graph IR with a built-in graph optimizer (constant folding, fusion, dead-node elimination).
Generative AI & 3D Pipelines: Runs things like FLUX 3 / FLUX.2, Stable Diffusion (with LoRA, ControlNet, and ComfyUI patches), LTX-Video & Wan 2.2 for video gen, TRELLIS.2 for text/image-to-3D, SkinTokens/UniRig for auto-rigging, Kimodo for animation, and GGUF LLMs (with KV-cache decode and TurboQuant).
On-Device Training: Supports LoRA fine-tuning of GGUF language models directly on your hardware (e.g., fine-tuning llama3.1:8b on an RTX 4070).
If you want to kick the tires without installing anything, the live demo runs entirely client-side on your GPU (models download once and cache locally). You can find the demo on the GitHub repo page (this forum blocks links). github / LostBeard / SpawnDev.ILGPU.ML
I’d love to hear your thoughts, answer any architectural questions, or see what people think of running heavy generative and 3D workflows straight out of a browser tab via WebGPU!
Todd
aka LostBeard