VindexLLM
A GPU-accelerated LLM inference engine written entirely in Delphi that runs the full transformer forward pass for GGUF models on the GPU through Vulkan compute shaders, with no Python or CUDA toolkit.
VindexLLM loads standard GGUF model files and runs the complete transformer forward pass, embedding lookup, attention and feed-forward layers, normalization and sampling, entirely on the GPU using Vulkan compute shaders, needing only the vulkan-1.dll that already ships with modern GPU drivers. It requires no Python environment, no CUDA toolkit and no other external runtime.
All compute shaders are written in GLSL 450, compiled to SPIR-V at build time, and embedded as Windows resources directly into the executable, so there are no loose shader files or runtime shader compilation. GGUF files are memory-mapped and their weights uploaded to VRAM through staging buffers with no intermediate copies or format conversion, and the residual stream never leaves VRAM between layers; the only PCIe transfers are the initial token embedding and the final logits download.
An included example adds interactive chat with SQLite-backed persistent memory and retrieval-augmented generation across sessions. It is written in Object Pascal for Delphi on Windows.